From e1fe3e755b54c0dd8ccb5c6a0bc3a7f48f4c9693 Mon Sep 17 00:00:00 2001 From: jsteube Date: Fri, 1 Mar 2019 14:49:00 +0100 Subject: [PATCH] Optimize some switch_buffer_* functions for generic OpenCL devices (CPU, various OSX, ...) --- OpenCL/inc_common.cl | 16608 ++++++++++++++++---------------- OpenCL/inc_types.cl | 112 +- OpenCL/m00500-optimized.cl | 38 +- OpenCL/m01600-optimized.cl | 38 +- OpenCL/m05800-optimized.cl | 17 +- OpenCL/m06300-optimized.cl | 38 +- OpenCL/m07400-optimized.cl | 58 +- OpenCL/m10700-optimized.cl | 18 +- OpenCL/m11600-pure.cl | 10 +- OpenCL/m13800_a0-optimized.cl | 34 +- OpenCL/m13800_a1-optimized.cl | 34 +- OpenCL/m13800_a3-optimized.cl | 34 +- 12 files changed, 8345 insertions(+), 8694 deletions(-) diff --git a/OpenCL/inc_common.cl b/OpenCL/inc_common.cl index 00ceb29b0..743eb4a86 100644 --- a/OpenCL/inc_common.cl +++ b/OpenCL/inc_common.cl @@ -724,23 +724,6 @@ DECLSPEC void switch_buffer_by_offset_le (u32x *w0, u32x *w1, u32x *w2, u32x *w3 const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); - switch (offset_switch) { case 0: @@ -1063,23 +1046,6 @@ DECLSPEC void switch_buffer_by_offset_le (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; } - - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -1414,7 +1380,6 @@ DECLSPEC void switch_buffer_by_offset_le (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; } - #endif } @@ -1427,23 +1392,6 @@ DECLSPEC void switch_buffer_by_offset_carry_le (u32x *w0, u32x *w1, u32x *w2, u3 const int offset_switch = offset / 4; #if defined IS_AMD || defined IS_GENERIC - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); - switch (offset_switch) { case 0: @@ -1902,39 +1850,6 @@ DECLSPEC void switch_buffer_by_offset_carry_le (u32x *w0, u32x *w1, u32x *w2, u3 break; } - - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); - c0[0] = swap32 (c0[0]); - c0[1] = swap32 (c0[1]); - c0[2] = swap32 (c0[2]); - c0[3] = swap32 (c0[3]); - c1[0] = swap32 (c1[0]); - c1[1] = swap32 (c1[1]); - c1[2] = swap32 (c1[2]); - c1[3] = swap32 (c1[3]); - c2[0] = swap32 (c2[0]); - c2[1] = swap32 (c2[1]); - c2[2] = swap32 (c2[2]); - c2[3] = swap32 (c2[3]); - c3[0] = swap32 (c3[0]); - c3[1] = swap32 (c3[1]); - c3[2] = swap32 (c3[2]); - c3[3] = swap32 (c3[3]); #endif #ifdef IS_NV @@ -2741,83 +2656,82 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - switch (offset_switch) { case 0: - w3[3] = hc_bytealign (w3[2], w3[3], offset); - w3[2] = hc_bytealign (w3[1], w3[2], offset); - w3[1] = hc_bytealign (w3[0], w3[1], offset); - w3[0] = hc_bytealign (w2[3], w3[0], offset); - w2[3] = hc_bytealign (w2[2], w2[3], offset); - w2[2] = hc_bytealign (w2[1], w2[2], offset); - w2[1] = hc_bytealign (w2[0], w2[1], offset); - w2[0] = hc_bytealign (w1[3], w2[0], offset); - w1[3] = hc_bytealign (w1[2], w1[3], offset); - w1[2] = hc_bytealign (w1[1], w1[2], offset); - w1[1] = hc_bytealign (w1[0], w1[1], offset); - w1[0] = hc_bytealign (w0[3], w1[0], offset); - w0[3] = hc_bytealign (w0[2], w0[3], offset); - w0[2] = hc_bytealign (w0[1], w0[2], offset); - w0[1] = hc_bytealign (w0[0], w0[1], offset); - w0[0] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be ( 0, w0[0], offset); break; case 1: - w3[3] = hc_bytealign (w3[1], w3[2], offset); - w3[2] = hc_bytealign (w3[0], w3[1], offset); - w3[1] = hc_bytealign (w2[3], w3[0], offset); - w3[0] = hc_bytealign (w2[2], w2[3], offset); - w2[3] = hc_bytealign (w2[1], w2[2], offset); - w2[2] = hc_bytealign (w2[0], w2[1], offset); - w2[1] = hc_bytealign (w1[3], w2[0], offset); - w2[0] = hc_bytealign (w1[2], w1[3], offset); - w1[3] = hc_bytealign (w1[1], w1[2], offset); - w1[2] = hc_bytealign (w1[0], w1[1], offset); - w1[1] = hc_bytealign (w0[3], w1[0], offset); - w1[0] = hc_bytealign (w0[2], w0[3], offset); - w0[3] = hc_bytealign (w0[1], w0[2], offset); - w0[2] = hc_bytealign (w0[0], w0[1], offset); - w0[1] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be ( 0, w0[0], offset); w0[0] = 0; break; case 2: - w3[3] = hc_bytealign (w3[0], w3[1], offset); - w3[2] = hc_bytealign (w2[3], w3[0], offset); - w3[1] = hc_bytealign (w2[2], w2[3], offset); - w3[0] = hc_bytealign (w2[1], w2[2], offset); - w2[3] = hc_bytealign (w2[0], w2[1], offset); - w2[2] = hc_bytealign (w1[3], w2[0], offset); - w2[1] = hc_bytealign (w1[2], w1[3], offset); - w2[0] = hc_bytealign (w1[1], w1[2], offset); - w1[3] = hc_bytealign (w1[0], w1[1], offset); - w1[2] = hc_bytealign (w0[3], w1[0], offset); - w1[1] = hc_bytealign (w0[2], w0[3], offset); - w1[0] = hc_bytealign (w0[1], w0[2], offset); - w0[3] = hc_bytealign (w0[0], w0[1], offset); - w0[2] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - w3[3] = hc_bytealign (w2[3], w3[0], offset); - w3[2] = hc_bytealign (w2[2], w2[3], offset); - w3[1] = hc_bytealign (w2[1], w2[2], offset); - w3[0] = hc_bytealign (w2[0], w2[1], offset); - w2[3] = hc_bytealign (w1[3], w2[0], offset); - w2[2] = hc_bytealign (w1[2], w1[3], offset); - w2[1] = hc_bytealign (w1[1], w1[2], offset); - w2[0] = hc_bytealign (w1[0], w1[1], offset); - w1[3] = hc_bytealign (w0[3], w1[0], offset); - w1[2] = hc_bytealign (w0[2], w0[3], offset); - w1[1] = hc_bytealign (w0[1], w0[2], offset); - w1[0] = hc_bytealign (w0[0], w0[1], offset); - w0[3] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -2825,18 +2739,18 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 4: - w3[3] = hc_bytealign (w2[2], w2[3], offset); - w3[2] = hc_bytealign (w2[1], w2[2], offset); - w3[1] = hc_bytealign (w2[0], w2[1], offset); - w3[0] = hc_bytealign (w1[3], w2[0], offset); - w2[3] = hc_bytealign (w1[2], w1[3], offset); - w2[2] = hc_bytealign (w1[1], w1[2], offset); - w2[1] = hc_bytealign (w1[0], w1[1], offset); - w2[0] = hc_bytealign (w0[3], w1[0], offset); - w1[3] = hc_bytealign (w0[2], w0[3], offset); - w1[2] = hc_bytealign (w0[1], w0[2], offset); - w1[1] = hc_bytealign (w0[0], w0[1], offset); - w1[0] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -2845,17 +2759,17 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 5: - w3[3] = hc_bytealign (w2[1], w2[2], offset); - w3[2] = hc_bytealign (w2[0], w2[1], offset); - w3[1] = hc_bytealign (w1[3], w2[0], offset); - w3[0] = hc_bytealign (w1[2], w1[3], offset); - w2[3] = hc_bytealign (w1[1], w1[2], offset); - w2[2] = hc_bytealign (w1[0], w1[1], offset); - w2[1] = hc_bytealign (w0[3], w1[0], offset); - w2[0] = hc_bytealign (w0[2], w0[3], offset); - w1[3] = hc_bytealign (w0[1], w0[2], offset); - w1[2] = hc_bytealign (w0[0], w0[1], offset); - w1[1] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -2865,16 +2779,16 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 6: - w3[3] = hc_bytealign (w2[0], w2[1], offset); - w3[2] = hc_bytealign (w1[3], w2[0], offset); - w3[1] = hc_bytealign (w1[2], w1[3], offset); - w3[0] = hc_bytealign (w1[1], w1[2], offset); - w2[3] = hc_bytealign (w1[0], w1[1], offset); - w2[2] = hc_bytealign (w0[3], w1[0], offset); - w2[1] = hc_bytealign (w0[2], w0[3], offset); - w2[0] = hc_bytealign (w0[1], w0[2], offset); - w1[3] = hc_bytealign (w0[0], w0[1], offset); - w1[2] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -2885,15 +2799,15 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 7: - w3[3] = hc_bytealign (w1[3], w2[0], offset); - w3[2] = hc_bytealign (w1[2], w1[3], offset); - w3[1] = hc_bytealign (w1[1], w1[2], offset); - w3[0] = hc_bytealign (w1[0], w1[1], offset); - w2[3] = hc_bytealign (w0[3], w1[0], offset); - w2[2] = hc_bytealign (w0[2], w0[3], offset); - w2[1] = hc_bytealign (w0[1], w0[2], offset); - w2[0] = hc_bytealign (w0[0], w0[1], offset); - w1[3] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -2905,14 +2819,14 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 8: - w3[3] = hc_bytealign (w1[2], w1[3], offset); - w3[2] = hc_bytealign (w1[1], w1[2], offset); - w3[1] = hc_bytealign (w1[0], w1[1], offset); - w3[0] = hc_bytealign (w0[3], w1[0], offset); - w2[3] = hc_bytealign (w0[2], w0[3], offset); - w2[2] = hc_bytealign (w0[1], w0[2], offset); - w2[1] = hc_bytealign (w0[0], w0[1], offset); - w2[0] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -2925,13 +2839,13 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 9: - w3[3] = hc_bytealign (w1[1], w1[2], offset); - w3[2] = hc_bytealign (w1[0], w1[1], offset); - w3[1] = hc_bytealign (w0[3], w1[0], offset); - w3[0] = hc_bytealign (w0[2], w0[3], offset); - w2[3] = hc_bytealign (w0[1], w0[2], offset); - w2[2] = hc_bytealign (w0[0], w0[1], offset); - w2[1] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -2945,12 +2859,12 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 10: - w3[3] = hc_bytealign (w1[0], w1[1], offset); - w3[2] = hc_bytealign (w0[3], w1[0], offset); - w3[1] = hc_bytealign (w0[2], w0[3], offset); - w3[0] = hc_bytealign (w0[1], w0[2], offset); - w2[3] = hc_bytealign (w0[0], w0[1], offset); - w2[2] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -2965,11 +2879,11 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 11: - w3[3] = hc_bytealign (w0[3], w1[0], offset); - w3[2] = hc_bytealign (w0[2], w0[3], offset); - w3[1] = hc_bytealign (w0[1], w0[2], offset); - w3[0] = hc_bytealign (w0[0], w0[1], offset); - w2[3] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -2985,10 +2899,10 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 12: - w3[3] = hc_bytealign (w0[2], w0[3], offset); - w3[2] = hc_bytealign (w0[1], w0[2], offset); - w3[1] = hc_bytealign (w0[0], w0[1], offset); - w3[0] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -3005,9 +2919,9 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 13: - w3[3] = hc_bytealign (w0[1], w0[2], offset); - w3[2] = hc_bytealign (w0[0], w0[1], offset); - w3[1] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -3025,8 +2939,8 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 14: - w3[3] = hc_bytealign (w0[0], w0[1], offset); - w3[2] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -3045,7 +2959,7 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; case 15: - w3[3] = hc_bytealign ( 0, w0[0], offset); + w3[3] = hc_bytealign_be ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -3064,7 +2978,6 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; } - #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -3399,7 +3312,6 @@ DECLSPEC void switch_buffer_by_offset_be (u32x *w0, u32x *w1, u32x *w2, u32x *w3 break; } - #endif } @@ -3411,89 +3323,89 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 switch (offset_switch) { case 0: - c0[0] = hc_bytealign (w3[3], 0, offset); - w3[3] = hc_bytealign (w3[2], w3[3], offset); - w3[2] = hc_bytealign (w3[1], w3[2], offset); - w3[1] = hc_bytealign (w3[0], w3[1], offset); - w3[0] = hc_bytealign (w2[3], w3[0], offset); - w2[3] = hc_bytealign (w2[2], w2[3], offset); - w2[2] = hc_bytealign (w2[1], w2[2], offset); - w2[1] = hc_bytealign (w2[0], w2[1], offset); - w2[0] = hc_bytealign (w1[3], w2[0], offset); - w1[3] = hc_bytealign (w1[2], w1[3], offset); - w1[2] = hc_bytealign (w1[1], w1[2], offset); - w1[1] = hc_bytealign (w1[0], w1[1], offset); - w1[0] = hc_bytealign (w0[3], w1[0], offset); - w0[3] = hc_bytealign (w0[2], w0[3], offset); - w0[2] = hc_bytealign (w0[1], w0[2], offset); - w0[1] = hc_bytealign (w0[0], w0[1], offset); - w0[0] = hc_bytealign ( 0, w0[0], offset); + c0[0] = hc_bytealign_be (w3[3], 0, offset); + w3[3] = hc_bytealign_be (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be ( 0, w0[0], offset); break; case 1: - c0[1] = hc_bytealign (w3[3], 0, offset); - c0[0] = hc_bytealign (w3[2], w3[3], offset); - w3[3] = hc_bytealign (w3[1], w3[2], offset); - w3[2] = hc_bytealign (w3[0], w3[1], offset); - w3[1] = hc_bytealign (w2[3], w3[0], offset); - w3[0] = hc_bytealign (w2[2], w2[3], offset); - w2[3] = hc_bytealign (w2[1], w2[2], offset); - w2[2] = hc_bytealign (w2[0], w2[1], offset); - w2[1] = hc_bytealign (w1[3], w2[0], offset); - w2[0] = hc_bytealign (w1[2], w1[3], offset); - w1[3] = hc_bytealign (w1[1], w1[2], offset); - w1[2] = hc_bytealign (w1[0], w1[1], offset); - w1[1] = hc_bytealign (w0[3], w1[0], offset); - w1[0] = hc_bytealign (w0[2], w0[3], offset); - w0[3] = hc_bytealign (w0[1], w0[2], offset); - w0[2] = hc_bytealign (w0[0], w0[1], offset); - w0[1] = hc_bytealign ( 0, w0[0], offset); + c0[1] = hc_bytealign_be (w3[3], 0, offset); + c0[0] = hc_bytealign_be (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be ( 0, w0[0], offset); w0[0] = 0; break; case 2: - c0[2] = hc_bytealign (w3[3], 0, offset); - c0[1] = hc_bytealign (w3[2], w3[3], offset); - c0[0] = hc_bytealign (w3[1], w3[2], offset); - w3[3] = hc_bytealign (w3[0], w3[1], offset); - w3[2] = hc_bytealign (w2[3], w3[0], offset); - w3[1] = hc_bytealign (w2[2], w2[3], offset); - w3[0] = hc_bytealign (w2[1], w2[2], offset); - w2[3] = hc_bytealign (w2[0], w2[1], offset); - w2[2] = hc_bytealign (w1[3], w2[0], offset); - w2[1] = hc_bytealign (w1[2], w1[3], offset); - w2[0] = hc_bytealign (w1[1], w1[2], offset); - w1[3] = hc_bytealign (w1[0], w1[1], offset); - w1[2] = hc_bytealign (w0[3], w1[0], offset); - w1[1] = hc_bytealign (w0[2], w0[3], offset); - w1[0] = hc_bytealign (w0[1], w0[2], offset); - w0[3] = hc_bytealign (w0[0], w0[1], offset); - w0[2] = hc_bytealign ( 0, w0[0], offset); + c0[2] = hc_bytealign_be (w3[3], 0, offset); + c0[1] = hc_bytealign_be (w3[2], w3[3], offset); + c0[0] = hc_bytealign_be (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - c0[3] = hc_bytealign (w3[3], 0, offset); - c0[2] = hc_bytealign (w3[2], w3[3], offset); - c0[1] = hc_bytealign (w3[1], w3[2], offset); - c0[0] = hc_bytealign (w3[0], w3[1], offset); - w3[3] = hc_bytealign (w2[3], w3[0], offset); - w3[2] = hc_bytealign (w2[2], w2[3], offset); - w3[1] = hc_bytealign (w2[1], w2[2], offset); - w3[0] = hc_bytealign (w2[0], w2[1], offset); - w2[3] = hc_bytealign (w1[3], w2[0], offset); - w2[2] = hc_bytealign (w1[2], w1[3], offset); - w2[1] = hc_bytealign (w1[1], w1[2], offset); - w2[0] = hc_bytealign (w1[0], w1[1], offset); - w1[3] = hc_bytealign (w0[3], w1[0], offset); - w1[2] = hc_bytealign (w0[2], w0[3], offset); - w1[1] = hc_bytealign (w0[1], w0[2], offset); - w1[0] = hc_bytealign (w0[0], w0[1], offset); - w0[3] = hc_bytealign ( 0, w0[0], offset); + c0[3] = hc_bytealign_be (w3[3], 0, offset); + c0[2] = hc_bytealign_be (w3[2], w3[3], offset); + c0[1] = hc_bytealign_be (w3[1], w3[2], offset); + c0[0] = hc_bytealign_be (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -3501,23 +3413,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 4: - c1[0] = hc_bytealign (w3[3], 0, offset); - c0[3] = hc_bytealign (w3[2], w3[3], offset); - c0[2] = hc_bytealign (w3[1], w3[2], offset); - c0[1] = hc_bytealign (w3[0], w3[1], offset); - c0[0] = hc_bytealign (w2[3], w3[0], offset); - w3[3] = hc_bytealign (w2[2], w2[3], offset); - w3[2] = hc_bytealign (w2[1], w2[2], offset); - w3[1] = hc_bytealign (w2[0], w2[1], offset); - w3[0] = hc_bytealign (w1[3], w2[0], offset); - w2[3] = hc_bytealign (w1[2], w1[3], offset); - w2[2] = hc_bytealign (w1[1], w1[2], offset); - w2[1] = hc_bytealign (w1[0], w1[1], offset); - w2[0] = hc_bytealign (w0[3], w1[0], offset); - w1[3] = hc_bytealign (w0[2], w0[3], offset); - w1[2] = hc_bytealign (w0[1], w0[2], offset); - w1[1] = hc_bytealign (w0[0], w0[1], offset); - w1[0] = hc_bytealign ( 0, w0[0], offset); + c1[0] = hc_bytealign_be (w3[3], 0, offset); + c0[3] = hc_bytealign_be (w3[2], w3[3], offset); + c0[2] = hc_bytealign_be (w3[1], w3[2], offset); + c0[1] = hc_bytealign_be (w3[0], w3[1], offset); + c0[0] = hc_bytealign_be (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -3526,23 +3438,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 5: - c1[1] = hc_bytealign (w3[3], 0, offset); - c1[0] = hc_bytealign (w3[2], w3[3], offset); - c0[3] = hc_bytealign (w3[1], w3[2], offset); - c0[2] = hc_bytealign (w3[0], w3[1], offset); - c0[1] = hc_bytealign (w2[3], w3[0], offset); - c0[0] = hc_bytealign (w2[2], w2[3], offset); - w3[3] = hc_bytealign (w2[1], w2[2], offset); - w3[2] = hc_bytealign (w2[0], w2[1], offset); - w3[1] = hc_bytealign (w1[3], w2[0], offset); - w3[0] = hc_bytealign (w1[2], w1[3], offset); - w2[3] = hc_bytealign (w1[1], w1[2], offset); - w2[2] = hc_bytealign (w1[0], w1[1], offset); - w2[1] = hc_bytealign (w0[3], w1[0], offset); - w2[0] = hc_bytealign (w0[2], w0[3], offset); - w1[3] = hc_bytealign (w0[1], w0[2], offset); - w1[2] = hc_bytealign (w0[0], w0[1], offset); - w1[1] = hc_bytealign ( 0, w0[0], offset); + c1[1] = hc_bytealign_be (w3[3], 0, offset); + c1[0] = hc_bytealign_be (w3[2], w3[3], offset); + c0[3] = hc_bytealign_be (w3[1], w3[2], offset); + c0[2] = hc_bytealign_be (w3[0], w3[1], offset); + c0[1] = hc_bytealign_be (w2[3], w3[0], offset); + c0[0] = hc_bytealign_be (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -3552,23 +3464,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 6: - c1[2] = hc_bytealign (w3[3], 0, offset); - c1[1] = hc_bytealign (w3[2], w3[3], offset); - c1[0] = hc_bytealign (w3[1], w3[2], offset); - c0[3] = hc_bytealign (w3[0], w3[1], offset); - c0[2] = hc_bytealign (w2[3], w3[0], offset); - c0[1] = hc_bytealign (w2[2], w2[3], offset); - c0[0] = hc_bytealign (w2[1], w2[2], offset); - w3[3] = hc_bytealign (w2[0], w2[1], offset); - w3[2] = hc_bytealign (w1[3], w2[0], offset); - w3[1] = hc_bytealign (w1[2], w1[3], offset); - w3[0] = hc_bytealign (w1[1], w1[2], offset); - w2[3] = hc_bytealign (w1[0], w1[1], offset); - w2[2] = hc_bytealign (w0[3], w1[0], offset); - w2[1] = hc_bytealign (w0[2], w0[3], offset); - w2[0] = hc_bytealign (w0[1], w0[2], offset); - w1[3] = hc_bytealign (w0[0], w0[1], offset); - w1[2] = hc_bytealign ( 0, w0[0], offset); + c1[2] = hc_bytealign_be (w3[3], 0, offset); + c1[1] = hc_bytealign_be (w3[2], w3[3], offset); + c1[0] = hc_bytealign_be (w3[1], w3[2], offset); + c0[3] = hc_bytealign_be (w3[0], w3[1], offset); + c0[2] = hc_bytealign_be (w2[3], w3[0], offset); + c0[1] = hc_bytealign_be (w2[2], w2[3], offset); + c0[0] = hc_bytealign_be (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -3579,23 +3491,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 7: - c1[3] = hc_bytealign (w3[3], 0, offset); - c1[2] = hc_bytealign (w3[2], w3[3], offset); - c1[1] = hc_bytealign (w3[1], w3[2], offset); - c1[0] = hc_bytealign (w3[0], w3[1], offset); - c0[3] = hc_bytealign (w2[3], w3[0], offset); - c0[2] = hc_bytealign (w2[2], w2[3], offset); - c0[1] = hc_bytealign (w2[1], w2[2], offset); - c0[0] = hc_bytealign (w2[0], w2[1], offset); - w3[3] = hc_bytealign (w1[3], w2[0], offset); - w3[2] = hc_bytealign (w1[2], w1[3], offset); - w3[1] = hc_bytealign (w1[1], w1[2], offset); - w3[0] = hc_bytealign (w1[0], w1[1], offset); - w2[3] = hc_bytealign (w0[3], w1[0], offset); - w2[2] = hc_bytealign (w0[2], w0[3], offset); - w2[1] = hc_bytealign (w0[1], w0[2], offset); - w2[0] = hc_bytealign (w0[0], w0[1], offset); - w1[3] = hc_bytealign ( 0, w0[0], offset); + c1[3] = hc_bytealign_be (w3[3], 0, offset); + c1[2] = hc_bytealign_be (w3[2], w3[3], offset); + c1[1] = hc_bytealign_be (w3[1], w3[2], offset); + c1[0] = hc_bytealign_be (w3[0], w3[1], offset); + c0[3] = hc_bytealign_be (w2[3], w3[0], offset); + c0[2] = hc_bytealign_be (w2[2], w2[3], offset); + c0[1] = hc_bytealign_be (w2[1], w2[2], offset); + c0[0] = hc_bytealign_be (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -3607,23 +3519,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 8: - c2[0] = hc_bytealign (w3[3], 0, offset); - c1[3] = hc_bytealign (w3[2], w3[3], offset); - c1[2] = hc_bytealign (w3[1], w3[2], offset); - c1[1] = hc_bytealign (w3[0], w3[1], offset); - c1[0] = hc_bytealign (w2[3], w3[0], offset); - c0[3] = hc_bytealign (w2[2], w2[3], offset); - c0[2] = hc_bytealign (w2[1], w2[2], offset); - c0[1] = hc_bytealign (w2[0], w2[1], offset); - c0[0] = hc_bytealign (w1[3], w2[0], offset); - w3[3] = hc_bytealign (w1[2], w1[3], offset); - w3[2] = hc_bytealign (w1[1], w1[2], offset); - w3[1] = hc_bytealign (w1[0], w1[1], offset); - w3[0] = hc_bytealign (w0[3], w1[0], offset); - w2[3] = hc_bytealign (w0[2], w0[3], offset); - w2[2] = hc_bytealign (w0[1], w0[2], offset); - w2[1] = hc_bytealign (w0[0], w0[1], offset); - w2[0] = hc_bytealign ( 0, w0[0], offset); + c2[0] = hc_bytealign_be (w3[3], 0, offset); + c1[3] = hc_bytealign_be (w3[2], w3[3], offset); + c1[2] = hc_bytealign_be (w3[1], w3[2], offset); + c1[1] = hc_bytealign_be (w3[0], w3[1], offset); + c1[0] = hc_bytealign_be (w2[3], w3[0], offset); + c0[3] = hc_bytealign_be (w2[2], w2[3], offset); + c0[2] = hc_bytealign_be (w2[1], w2[2], offset); + c0[1] = hc_bytealign_be (w2[0], w2[1], offset); + c0[0] = hc_bytealign_be (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -3636,23 +3548,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 9: - c2[1] = hc_bytealign (w3[3], 0, offset); - c2[0] = hc_bytealign (w3[2], w3[3], offset); - c1[3] = hc_bytealign (w3[1], w3[2], offset); - c1[2] = hc_bytealign (w3[0], w3[1], offset); - c1[1] = hc_bytealign (w2[3], w3[0], offset); - c1[0] = hc_bytealign (w2[2], w2[3], offset); - c0[3] = hc_bytealign (w2[1], w2[2], offset); - c0[2] = hc_bytealign (w2[0], w2[1], offset); - c0[1] = hc_bytealign (w1[3], w2[0], offset); - c0[0] = hc_bytealign (w1[2], w1[3], offset); - w3[3] = hc_bytealign (w1[1], w1[2], offset); - w3[2] = hc_bytealign (w1[0], w1[1], offset); - w3[1] = hc_bytealign (w0[3], w1[0], offset); - w3[0] = hc_bytealign (w0[2], w0[3], offset); - w2[3] = hc_bytealign (w0[1], w0[2], offset); - w2[2] = hc_bytealign (w0[0], w0[1], offset); - w2[1] = hc_bytealign ( 0, w0[0], offset); + c2[1] = hc_bytealign_be (w3[3], 0, offset); + c2[0] = hc_bytealign_be (w3[2], w3[3], offset); + c1[3] = hc_bytealign_be (w3[1], w3[2], offset); + c1[2] = hc_bytealign_be (w3[0], w3[1], offset); + c1[1] = hc_bytealign_be (w2[3], w3[0], offset); + c1[0] = hc_bytealign_be (w2[2], w2[3], offset); + c0[3] = hc_bytealign_be (w2[1], w2[2], offset); + c0[2] = hc_bytealign_be (w2[0], w2[1], offset); + c0[1] = hc_bytealign_be (w1[3], w2[0], offset); + c0[0] = hc_bytealign_be (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -3666,23 +3578,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 10: - c2[2] = hc_bytealign (w3[3], 0, offset); - c2[1] = hc_bytealign (w3[2], w3[3], offset); - c2[0] = hc_bytealign (w3[1], w3[2], offset); - c1[3] = hc_bytealign (w3[0], w3[1], offset); - c1[2] = hc_bytealign (w2[3], w3[0], offset); - c1[1] = hc_bytealign (w2[2], w2[3], offset); - c1[0] = hc_bytealign (w2[1], w2[2], offset); - c0[3] = hc_bytealign (w2[0], w2[1], offset); - c0[2] = hc_bytealign (w1[3], w2[0], offset); - c0[1] = hc_bytealign (w1[2], w1[3], offset); - c0[0] = hc_bytealign (w1[1], w1[2], offset); - w3[3] = hc_bytealign (w1[0], w1[1], offset); - w3[2] = hc_bytealign (w0[3], w1[0], offset); - w3[1] = hc_bytealign (w0[2], w0[3], offset); - w3[0] = hc_bytealign (w0[1], w0[2], offset); - w2[3] = hc_bytealign (w0[0], w0[1], offset); - w2[2] = hc_bytealign ( 0, w0[0], offset); + c2[2] = hc_bytealign_be (w3[3], 0, offset); + c2[1] = hc_bytealign_be (w3[2], w3[3], offset); + c2[0] = hc_bytealign_be (w3[1], w3[2], offset); + c1[3] = hc_bytealign_be (w3[0], w3[1], offset); + c1[2] = hc_bytealign_be (w2[3], w3[0], offset); + c1[1] = hc_bytealign_be (w2[2], w2[3], offset); + c1[0] = hc_bytealign_be (w2[1], w2[2], offset); + c0[3] = hc_bytealign_be (w2[0], w2[1], offset); + c0[2] = hc_bytealign_be (w1[3], w2[0], offset); + c0[1] = hc_bytealign_be (w1[2], w1[3], offset); + c0[0] = hc_bytealign_be (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -3697,23 +3609,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 11: - c2[3] = hc_bytealign (w3[3], 0, offset); - c2[2] = hc_bytealign (w3[2], w3[3], offset); - c2[1] = hc_bytealign (w3[1], w3[2], offset); - c2[0] = hc_bytealign (w3[0], w3[1], offset); - c1[3] = hc_bytealign (w2[3], w3[0], offset); - c1[2] = hc_bytealign (w2[2], w2[3], offset); - c1[1] = hc_bytealign (w2[1], w2[2], offset); - c1[0] = hc_bytealign (w2[0], w2[1], offset); - c0[3] = hc_bytealign (w1[3], w2[0], offset); - c0[2] = hc_bytealign (w1[2], w1[3], offset); - c0[1] = hc_bytealign (w1[1], w1[2], offset); - c0[0] = hc_bytealign (w1[0], w1[1], offset); - w3[3] = hc_bytealign (w0[3], w1[0], offset); - w3[2] = hc_bytealign (w0[2], w0[3], offset); - w3[1] = hc_bytealign (w0[1], w0[2], offset); - w3[0] = hc_bytealign (w0[0], w0[1], offset); - w2[3] = hc_bytealign ( 0, w0[0], offset); + c2[3] = hc_bytealign_be (w3[3], 0, offset); + c2[2] = hc_bytealign_be (w3[2], w3[3], offset); + c2[1] = hc_bytealign_be (w3[1], w3[2], offset); + c2[0] = hc_bytealign_be (w3[0], w3[1], offset); + c1[3] = hc_bytealign_be (w2[3], w3[0], offset); + c1[2] = hc_bytealign_be (w2[2], w2[3], offset); + c1[1] = hc_bytealign_be (w2[1], w2[2], offset); + c1[0] = hc_bytealign_be (w2[0], w2[1], offset); + c0[3] = hc_bytealign_be (w1[3], w2[0], offset); + c0[2] = hc_bytealign_be (w1[2], w1[3], offset); + c0[1] = hc_bytealign_be (w1[1], w1[2], offset); + c0[0] = hc_bytealign_be (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -3729,23 +3641,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 12: - c3[0] = hc_bytealign (w3[3], 0, offset); - c2[3] = hc_bytealign (w3[2], w3[3], offset); - c2[2] = hc_bytealign (w3[1], w3[2], offset); - c2[1] = hc_bytealign (w3[0], w3[1], offset); - c2[0] = hc_bytealign (w2[3], w3[0], offset); - c1[3] = hc_bytealign (w2[2], w2[3], offset); - c1[2] = hc_bytealign (w2[1], w2[2], offset); - c1[1] = hc_bytealign (w2[0], w2[1], offset); - c1[0] = hc_bytealign (w1[3], w2[0], offset); - c0[3] = hc_bytealign (w1[2], w1[3], offset); - c0[2] = hc_bytealign (w1[1], w1[2], offset); - c0[1] = hc_bytealign (w1[0], w1[1], offset); - c0[0] = hc_bytealign (w0[3], w1[0], offset); - w3[3] = hc_bytealign (w0[2], w0[3], offset); - w3[2] = hc_bytealign (w0[1], w0[2], offset); - w3[1] = hc_bytealign (w0[0], w0[1], offset); - w3[0] = hc_bytealign ( 0, w0[0], offset); + c3[0] = hc_bytealign_be (w3[3], 0, offset); + c2[3] = hc_bytealign_be (w3[2], w3[3], offset); + c2[2] = hc_bytealign_be (w3[1], w3[2], offset); + c2[1] = hc_bytealign_be (w3[0], w3[1], offset); + c2[0] = hc_bytealign_be (w2[3], w3[0], offset); + c1[3] = hc_bytealign_be (w2[2], w2[3], offset); + c1[2] = hc_bytealign_be (w2[1], w2[2], offset); + c1[1] = hc_bytealign_be (w2[0], w2[1], offset); + c1[0] = hc_bytealign_be (w1[3], w2[0], offset); + c0[3] = hc_bytealign_be (w1[2], w1[3], offset); + c0[2] = hc_bytealign_be (w1[1], w1[2], offset); + c0[1] = hc_bytealign_be (w1[0], w1[1], offset); + c0[0] = hc_bytealign_be (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -3762,23 +3674,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 13: - c3[1] = hc_bytealign (w3[3], 0, offset); - c3[0] = hc_bytealign (w3[2], w3[3], offset); - c2[3] = hc_bytealign (w3[1], w3[2], offset); - c2[2] = hc_bytealign (w3[0], w3[1], offset); - c2[1] = hc_bytealign (w2[3], w3[0], offset); - c2[0] = hc_bytealign (w2[2], w2[3], offset); - c1[3] = hc_bytealign (w2[1], w2[2], offset); - c1[2] = hc_bytealign (w2[0], w2[1], offset); - c1[1] = hc_bytealign (w1[3], w2[0], offset); - c1[0] = hc_bytealign (w1[2], w1[3], offset); - c0[3] = hc_bytealign (w1[1], w1[2], offset); - c0[2] = hc_bytealign (w1[0], w1[1], offset); - c0[1] = hc_bytealign (w0[3], w1[0], offset); - c0[0] = hc_bytealign (w0[2], w0[3], offset); - w3[3] = hc_bytealign (w0[1], w0[2], offset); - w3[2] = hc_bytealign (w0[0], w0[1], offset); - w3[1] = hc_bytealign ( 0, w0[0], offset); + c3[1] = hc_bytealign_be (w3[3], 0, offset); + c3[0] = hc_bytealign_be (w3[2], w3[3], offset); + c2[3] = hc_bytealign_be (w3[1], w3[2], offset); + c2[2] = hc_bytealign_be (w3[0], w3[1], offset); + c2[1] = hc_bytealign_be (w2[3], w3[0], offset); + c2[0] = hc_bytealign_be (w2[2], w2[3], offset); + c1[3] = hc_bytealign_be (w2[1], w2[2], offset); + c1[2] = hc_bytealign_be (w2[0], w2[1], offset); + c1[1] = hc_bytealign_be (w1[3], w2[0], offset); + c1[0] = hc_bytealign_be (w1[2], w1[3], offset); + c0[3] = hc_bytealign_be (w1[1], w1[2], offset); + c0[2] = hc_bytealign_be (w1[0], w1[1], offset); + c0[1] = hc_bytealign_be (w0[3], w1[0], offset); + c0[0] = hc_bytealign_be (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -3796,23 +3708,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 14: - c3[2] = hc_bytealign (w3[3], 0, offset); - c3[1] = hc_bytealign (w3[2], w3[3], offset); - c3[0] = hc_bytealign (w3[1], w3[2], offset); - c2[3] = hc_bytealign (w3[0], w3[1], offset); - c2[2] = hc_bytealign (w2[3], w3[0], offset); - c2[1] = hc_bytealign (w2[2], w2[3], offset); - c2[0] = hc_bytealign (w2[1], w2[2], offset); - c1[3] = hc_bytealign (w2[0], w2[1], offset); - c1[2] = hc_bytealign (w1[3], w2[0], offset); - c1[1] = hc_bytealign (w1[2], w1[3], offset); - c1[0] = hc_bytealign (w1[1], w1[2], offset); - c0[3] = hc_bytealign (w1[0], w1[1], offset); - c0[2] = hc_bytealign (w0[3], w1[0], offset); - c0[1] = hc_bytealign (w0[2], w0[3], offset); - c0[0] = hc_bytealign (w0[1], w0[2], offset); - w3[3] = hc_bytealign (w0[0], w0[1], offset); - w3[2] = hc_bytealign ( 0, w0[0], offset); + c3[2] = hc_bytealign_be (w3[3], 0, offset); + c3[1] = hc_bytealign_be (w3[2], w3[3], offset); + c3[0] = hc_bytealign_be (w3[1], w3[2], offset); + c2[3] = hc_bytealign_be (w3[0], w3[1], offset); + c2[2] = hc_bytealign_be (w2[3], w3[0], offset); + c2[1] = hc_bytealign_be (w2[2], w2[3], offset); + c2[0] = hc_bytealign_be (w2[1], w2[2], offset); + c1[3] = hc_bytealign_be (w2[0], w2[1], offset); + c1[2] = hc_bytealign_be (w1[3], w2[0], offset); + c1[1] = hc_bytealign_be (w1[2], w1[3], offset); + c1[0] = hc_bytealign_be (w1[1], w1[2], offset); + c0[3] = hc_bytealign_be (w1[0], w1[1], offset); + c0[2] = hc_bytealign_be (w0[3], w1[0], offset); + c0[1] = hc_bytealign_be (w0[2], w0[3], offset); + c0[0] = hc_bytealign_be (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -3831,23 +3743,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be (u32x *w0, u32x *w1, u32x *w2, u3 break; case 15: - c3[3] = hc_bytealign (w3[3], 0, offset); - c3[2] = hc_bytealign (w3[2], w3[3], offset); - c3[1] = hc_bytealign (w3[1], w3[2], offset); - c3[0] = hc_bytealign (w3[0], w3[1], offset); - c2[3] = hc_bytealign (w2[3], w3[0], offset); - c2[2] = hc_bytealign (w2[2], w2[3], offset); - c2[1] = hc_bytealign (w2[1], w2[2], offset); - c2[0] = hc_bytealign (w2[0], w2[1], offset); - c1[3] = hc_bytealign (w1[3], w2[0], offset); - c1[2] = hc_bytealign (w1[2], w1[3], offset); - c1[1] = hc_bytealign (w1[1], w1[2], offset); - c1[0] = hc_bytealign (w1[0], w1[1], offset); - c0[3] = hc_bytealign (w0[3], w1[0], offset); - c0[2] = hc_bytealign (w0[2], w0[3], offset); - c0[1] = hc_bytealign (w0[1], w0[2], offset); - c0[0] = hc_bytealign (w0[0], w0[1], offset); - w3[3] = hc_bytealign ( 0, w0[0], offset); + c3[3] = hc_bytealign_be (w3[3], 0, offset); + c3[2] = hc_bytealign_be (w3[2], w3[3], offset); + c3[1] = hc_bytealign_be (w3[1], w3[2], offset); + c3[0] = hc_bytealign_be (w3[0], w3[1], offset); + c2[3] = hc_bytealign_be (w2[3], w3[0], offset); + c2[2] = hc_bytealign_be (w2[2], w2[3], offset); + c2[1] = hc_bytealign_be (w2[1], w2[2], offset); + c2[0] = hc_bytealign_be (w2[0], w2[1], offset); + c1[3] = hc_bytealign_be (w1[3], w2[0], offset); + c1[2] = hc_bytealign_be (w1[2], w1[3], offset); + c1[1] = hc_bytealign_be (w1[1], w1[2], offset); + c1[0] = hc_bytealign_be (w1[0], w1[1], offset); + c0[3] = hc_bytealign_be (w0[3], w1[0], offset); + c0[2] = hc_bytealign_be (w0[2], w0[3], offset); + c0[1] = hc_bytealign_be (w0[1], w0[2], offset); + c0[0] = hc_bytealign_be (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -4348,39 +4260,6 @@ DECLSPEC void switch_buffer_by_offset_8x4_le (u32x *w0, u32x *w1, u32x *w2, u32x const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); - w4[0] = swap32 (w4[0]); - w4[1] = swap32 (w4[1]); - w4[2] = swap32 (w4[2]); - w4[3] = swap32 (w4[3]); - w5[0] = swap32 (w5[0]); - w5[1] = swap32 (w5[1]); - w5[2] = swap32 (w5[2]); - w5[3] = swap32 (w5[3]); - w6[0] = swap32 (w6[0]); - w6[1] = swap32 (w6[1]); - w6[2] = swap32 (w6[2]); - w6[3] = swap32 (w6[3]); - w7[0] = swap32 (w7[0]); - w7[1] = swap32 (w7[1]); - w7[2] = swap32 (w7[2]); - w7[3] = swap32 (w7[3]); - switch (offset_switch) { case 0: @@ -5535,39 +5414,6 @@ DECLSPEC void switch_buffer_by_offset_8x4_le (u32x *w0, u32x *w1, u32x *w2, u32x break; } - - w0[0] = swap32 (w0[0]); - w0[1] = swap32 (w0[1]); - w0[2] = swap32 (w0[2]); - w0[3] = swap32 (w0[3]); - w1[0] = swap32 (w1[0]); - w1[1] = swap32 (w1[1]); - w1[2] = swap32 (w1[2]); - w1[3] = swap32 (w1[3]); - w2[0] = swap32 (w2[0]); - w2[1] = swap32 (w2[1]); - w2[2] = swap32 (w2[2]); - w2[3] = swap32 (w2[3]); - w3[0] = swap32 (w3[0]); - w3[1] = swap32 (w3[1]); - w3[2] = swap32 (w3[2]); - w3[3] = swap32 (w3[3]); - w4[0] = swap32 (w4[0]); - w4[1] = swap32 (w4[1]); - w4[2] = swap32 (w4[2]); - w4[3] = swap32 (w4[3]); - w5[0] = swap32 (w5[0]); - w5[1] = swap32 (w5[1]); - w5[2] = swap32 (w5[2]); - w5[3] = swap32 (w5[3]); - w6[0] = swap32 (w6[0]); - w6[1] = swap32 (w6[1]); - w6[2] = swap32 (w6[2]); - w6[3] = swap32 (w6[3]); - w7[0] = swap32 (w7[0]); - w7[1] = swap32 (w7[1]); - w7[2] = swap32 (w7[2]); - w7[3] = swap32 (w7[3]); #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -6153,143 +5999,143 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x switch (offset_switch) { case 0: - w7[3] = hc_bytealign (w7[2], w7[3], offset); - w7[2] = hc_bytealign (w7[1], w7[2], offset); - w7[1] = hc_bytealign (w7[0], w7[1], offset); - w7[0] = hc_bytealign (w6[3], w7[0], offset); - w6[3] = hc_bytealign (w6[2], w6[3], offset); - w6[2] = hc_bytealign (w6[1], w6[2], offset); - w6[1] = hc_bytealign (w6[0], w6[1], offset); - w6[0] = hc_bytealign (w5[3], w6[0], offset); - w5[3] = hc_bytealign (w5[2], w5[3], offset); - w5[2] = hc_bytealign (w5[1], w5[2], offset); - w5[1] = hc_bytealign (w5[0], w5[1], offset); - w5[0] = hc_bytealign (w4[3], w5[0], offset); - w4[3] = hc_bytealign (w4[2], w4[3], offset); - w4[2] = hc_bytealign (w4[1], w4[2], offset); - w4[1] = hc_bytealign (w4[0], w4[1], offset); - w4[0] = hc_bytealign (w3[3], w4[0], offset); - w3[3] = hc_bytealign (w3[2], w3[3], offset); - w3[2] = hc_bytealign (w3[1], w3[2], offset); - w3[1] = hc_bytealign (w3[0], w3[1], offset); - w3[0] = hc_bytealign (w2[3], w3[0], offset); - w2[3] = hc_bytealign (w2[2], w2[3], offset); - w2[2] = hc_bytealign (w2[1], w2[2], offset); - w2[1] = hc_bytealign (w2[0], w2[1], offset); - w2[0] = hc_bytealign (w1[3], w2[0], offset); - w1[3] = hc_bytealign (w1[2], w1[3], offset); - w1[2] = hc_bytealign (w1[1], w1[2], offset); - w1[1] = hc_bytealign (w1[0], w1[1], offset); - w1[0] = hc_bytealign (w0[3], w1[0], offset); - w0[3] = hc_bytealign (w0[2], w0[3], offset); - w0[2] = hc_bytealign (w0[1], w0[2], offset); - w0[1] = hc_bytealign (w0[0], w0[1], offset); - w0[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w7[2], w7[3], offset); + w7[2] = hc_bytealign_be (w7[1], w7[2], offset); + w7[1] = hc_bytealign_be (w7[0], w7[1], offset); + w7[0] = hc_bytealign_be (w6[3], w7[0], offset); + w6[3] = hc_bytealign_be (w6[2], w6[3], offset); + w6[2] = hc_bytealign_be (w6[1], w6[2], offset); + w6[1] = hc_bytealign_be (w6[0], w6[1], offset); + w6[0] = hc_bytealign_be (w5[3], w6[0], offset); + w5[3] = hc_bytealign_be (w5[2], w5[3], offset); + w5[2] = hc_bytealign_be (w5[1], w5[2], offset); + w5[1] = hc_bytealign_be (w5[0], w5[1], offset); + w5[0] = hc_bytealign_be (w4[3], w5[0], offset); + w4[3] = hc_bytealign_be (w4[2], w4[3], offset); + w4[2] = hc_bytealign_be (w4[1], w4[2], offset); + w4[1] = hc_bytealign_be (w4[0], w4[1], offset); + w4[0] = hc_bytealign_be (w3[3], w4[0], offset); + w3[3] = hc_bytealign_be (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be ( 0, w0[0], offset); break; case 1: - w7[3] = hc_bytealign (w7[1], w7[2], offset); - w7[2] = hc_bytealign (w7[0], w7[1], offset); - w7[1] = hc_bytealign (w6[3], w7[0], offset); - w7[0] = hc_bytealign (w6[2], w6[3], offset); - w6[3] = hc_bytealign (w6[1], w6[2], offset); - w6[2] = hc_bytealign (w6[0], w6[1], offset); - w6[1] = hc_bytealign (w5[3], w6[0], offset); - w6[0] = hc_bytealign (w5[2], w5[3], offset); - w5[3] = hc_bytealign (w5[1], w5[2], offset); - w5[2] = hc_bytealign (w5[0], w5[1], offset); - w5[1] = hc_bytealign (w4[3], w5[0], offset); - w5[0] = hc_bytealign (w4[2], w4[3], offset); - w4[3] = hc_bytealign (w4[1], w4[2], offset); - w4[2] = hc_bytealign (w4[0], w4[1], offset); - w4[1] = hc_bytealign (w3[3], w4[0], offset); - w4[0] = hc_bytealign (w3[2], w3[3], offset); - w3[3] = hc_bytealign (w3[1], w3[2], offset); - w3[2] = hc_bytealign (w3[0], w3[1], offset); - w3[1] = hc_bytealign (w2[3], w3[0], offset); - w3[0] = hc_bytealign (w2[2], w2[3], offset); - w2[3] = hc_bytealign (w2[1], w2[2], offset); - w2[2] = hc_bytealign (w2[0], w2[1], offset); - w2[1] = hc_bytealign (w1[3], w2[0], offset); - w2[0] = hc_bytealign (w1[2], w1[3], offset); - w1[3] = hc_bytealign (w1[1], w1[2], offset); - w1[2] = hc_bytealign (w1[0], w1[1], offset); - w1[1] = hc_bytealign (w0[3], w1[0], offset); - w1[0] = hc_bytealign (w0[2], w0[3], offset); - w0[3] = hc_bytealign (w0[1], w0[2], offset); - w0[2] = hc_bytealign (w0[0], w0[1], offset); - w0[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w7[1], w7[2], offset); + w7[2] = hc_bytealign_be (w7[0], w7[1], offset); + w7[1] = hc_bytealign_be (w6[3], w7[0], offset); + w7[0] = hc_bytealign_be (w6[2], w6[3], offset); + w6[3] = hc_bytealign_be (w6[1], w6[2], offset); + w6[2] = hc_bytealign_be (w6[0], w6[1], offset); + w6[1] = hc_bytealign_be (w5[3], w6[0], offset); + w6[0] = hc_bytealign_be (w5[2], w5[3], offset); + w5[3] = hc_bytealign_be (w5[1], w5[2], offset); + w5[2] = hc_bytealign_be (w5[0], w5[1], offset); + w5[1] = hc_bytealign_be (w4[3], w5[0], offset); + w5[0] = hc_bytealign_be (w4[2], w4[3], offset); + w4[3] = hc_bytealign_be (w4[1], w4[2], offset); + w4[2] = hc_bytealign_be (w4[0], w4[1], offset); + w4[1] = hc_bytealign_be (w3[3], w4[0], offset); + w4[0] = hc_bytealign_be (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be ( 0, w0[0], offset); w0[0] = 0; break; case 2: - w7[3] = hc_bytealign (w7[0], w7[1], offset); - w7[2] = hc_bytealign (w6[3], w7[0], offset); - w7[1] = hc_bytealign (w6[2], w6[3], offset); - w7[0] = hc_bytealign (w6[1], w6[2], offset); - w6[3] = hc_bytealign (w6[0], w6[1], offset); - w6[2] = hc_bytealign (w5[3], w6[0], offset); - w6[1] = hc_bytealign (w5[2], w5[3], offset); - w6[0] = hc_bytealign (w5[1], w5[2], offset); - w5[3] = hc_bytealign (w5[0], w5[1], offset); - w5[2] = hc_bytealign (w4[3], w5[0], offset); - w5[1] = hc_bytealign (w4[2], w4[3], offset); - w5[0] = hc_bytealign (w4[1], w4[2], offset); - w4[3] = hc_bytealign (w4[0], w4[1], offset); - w4[2] = hc_bytealign (w3[3], w4[0], offset); - w4[1] = hc_bytealign (w3[2], w3[3], offset); - w4[0] = hc_bytealign (w3[1], w3[2], offset); - w3[3] = hc_bytealign (w3[0], w3[1], offset); - w3[2] = hc_bytealign (w2[3], w3[0], offset); - w3[1] = hc_bytealign (w2[2], w2[3], offset); - w3[0] = hc_bytealign (w2[1], w2[2], offset); - w2[3] = hc_bytealign (w2[0], w2[1], offset); - w2[2] = hc_bytealign (w1[3], w2[0], offset); - w2[1] = hc_bytealign (w1[2], w1[3], offset); - w2[0] = hc_bytealign (w1[1], w1[2], offset); - w1[3] = hc_bytealign (w1[0], w1[1], offset); - w1[2] = hc_bytealign (w0[3], w1[0], offset); - w1[1] = hc_bytealign (w0[2], w0[3], offset); - w1[0] = hc_bytealign (w0[1], w0[2], offset); - w0[3] = hc_bytealign (w0[0], w0[1], offset); - w0[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w7[0], w7[1], offset); + w7[2] = hc_bytealign_be (w6[3], w7[0], offset); + w7[1] = hc_bytealign_be (w6[2], w6[3], offset); + w7[0] = hc_bytealign_be (w6[1], w6[2], offset); + w6[3] = hc_bytealign_be (w6[0], w6[1], offset); + w6[2] = hc_bytealign_be (w5[3], w6[0], offset); + w6[1] = hc_bytealign_be (w5[2], w5[3], offset); + w6[0] = hc_bytealign_be (w5[1], w5[2], offset); + w5[3] = hc_bytealign_be (w5[0], w5[1], offset); + w5[2] = hc_bytealign_be (w4[3], w5[0], offset); + w5[1] = hc_bytealign_be (w4[2], w4[3], offset); + w5[0] = hc_bytealign_be (w4[1], w4[2], offset); + w4[3] = hc_bytealign_be (w4[0], w4[1], offset); + w4[2] = hc_bytealign_be (w3[3], w4[0], offset); + w4[1] = hc_bytealign_be (w3[2], w3[3], offset); + w4[0] = hc_bytealign_be (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - w7[3] = hc_bytealign (w6[3], w7[0], offset); - w7[2] = hc_bytealign (w6[2], w6[3], offset); - w7[1] = hc_bytealign (w6[1], w6[2], offset); - w7[0] = hc_bytealign (w6[0], w6[1], offset); - w6[3] = hc_bytealign (w5[3], w6[0], offset); - w6[2] = hc_bytealign (w5[2], w5[3], offset); - w6[1] = hc_bytealign (w5[1], w5[2], offset); - w6[0] = hc_bytealign (w5[0], w5[1], offset); - w5[3] = hc_bytealign (w4[3], w5[0], offset); - w5[2] = hc_bytealign (w4[2], w4[3], offset); - w5[1] = hc_bytealign (w4[1], w4[2], offset); - w5[0] = hc_bytealign (w4[0], w4[1], offset); - w4[3] = hc_bytealign (w3[3], w4[0], offset); - w4[2] = hc_bytealign (w3[2], w3[3], offset); - w4[1] = hc_bytealign (w3[1], w3[2], offset); - w4[0] = hc_bytealign (w3[0], w3[1], offset); - w3[3] = hc_bytealign (w2[3], w3[0], offset); - w3[2] = hc_bytealign (w2[2], w2[3], offset); - w3[1] = hc_bytealign (w2[1], w2[2], offset); - w3[0] = hc_bytealign (w2[0], w2[1], offset); - w2[3] = hc_bytealign (w1[3], w2[0], offset); - w2[2] = hc_bytealign (w1[2], w1[3], offset); - w2[1] = hc_bytealign (w1[1], w1[2], offset); - w2[0] = hc_bytealign (w1[0], w1[1], offset); - w1[3] = hc_bytealign (w0[3], w1[0], offset); - w1[2] = hc_bytealign (w0[2], w0[3], offset); - w1[1] = hc_bytealign (w0[1], w0[2], offset); - w1[0] = hc_bytealign (w0[0], w0[1], offset); - w0[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w6[3], w7[0], offset); + w7[2] = hc_bytealign_be (w6[2], w6[3], offset); + w7[1] = hc_bytealign_be (w6[1], w6[2], offset); + w7[0] = hc_bytealign_be (w6[0], w6[1], offset); + w6[3] = hc_bytealign_be (w5[3], w6[0], offset); + w6[2] = hc_bytealign_be (w5[2], w5[3], offset); + w6[1] = hc_bytealign_be (w5[1], w5[2], offset); + w6[0] = hc_bytealign_be (w5[0], w5[1], offset); + w5[3] = hc_bytealign_be (w4[3], w5[0], offset); + w5[2] = hc_bytealign_be (w4[2], w4[3], offset); + w5[1] = hc_bytealign_be (w4[1], w4[2], offset); + w5[0] = hc_bytealign_be (w4[0], w4[1], offset); + w4[3] = hc_bytealign_be (w3[3], w4[0], offset); + w4[2] = hc_bytealign_be (w3[2], w3[3], offset); + w4[1] = hc_bytealign_be (w3[1], w3[2], offset); + w4[0] = hc_bytealign_be (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -6297,34 +6143,34 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 4: - w7[3] = hc_bytealign (w6[2], w6[3], offset); - w7[2] = hc_bytealign (w6[1], w6[2], offset); - w7[1] = hc_bytealign (w6[0], w6[1], offset); - w7[0] = hc_bytealign (w5[3], w6[0], offset); - w6[3] = hc_bytealign (w5[2], w5[3], offset); - w6[2] = hc_bytealign (w5[1], w5[2], offset); - w6[1] = hc_bytealign (w5[0], w5[1], offset); - w6[0] = hc_bytealign (w4[3], w5[0], offset); - w5[3] = hc_bytealign (w4[2], w4[3], offset); - w5[2] = hc_bytealign (w4[1], w4[2], offset); - w5[1] = hc_bytealign (w4[0], w4[1], offset); - w5[0] = hc_bytealign (w3[3], w4[0], offset); - w4[3] = hc_bytealign (w3[2], w3[3], offset); - w4[2] = hc_bytealign (w3[1], w3[2], offset); - w4[1] = hc_bytealign (w3[0], w3[1], offset); - w4[0] = hc_bytealign (w2[3], w3[0], offset); - w3[3] = hc_bytealign (w2[2], w2[3], offset); - w3[2] = hc_bytealign (w2[1], w2[2], offset); - w3[1] = hc_bytealign (w2[0], w2[1], offset); - w3[0] = hc_bytealign (w1[3], w2[0], offset); - w2[3] = hc_bytealign (w1[2], w1[3], offset); - w2[2] = hc_bytealign (w1[1], w1[2], offset); - w2[1] = hc_bytealign (w1[0], w1[1], offset); - w2[0] = hc_bytealign (w0[3], w1[0], offset); - w1[3] = hc_bytealign (w0[2], w0[3], offset); - w1[2] = hc_bytealign (w0[1], w0[2], offset); - w1[1] = hc_bytealign (w0[0], w0[1], offset); - w1[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w6[2], w6[3], offset); + w7[2] = hc_bytealign_be (w6[1], w6[2], offset); + w7[1] = hc_bytealign_be (w6[0], w6[1], offset); + w7[0] = hc_bytealign_be (w5[3], w6[0], offset); + w6[3] = hc_bytealign_be (w5[2], w5[3], offset); + w6[2] = hc_bytealign_be (w5[1], w5[2], offset); + w6[1] = hc_bytealign_be (w5[0], w5[1], offset); + w6[0] = hc_bytealign_be (w4[3], w5[0], offset); + w5[3] = hc_bytealign_be (w4[2], w4[3], offset); + w5[2] = hc_bytealign_be (w4[1], w4[2], offset); + w5[1] = hc_bytealign_be (w4[0], w4[1], offset); + w5[0] = hc_bytealign_be (w3[3], w4[0], offset); + w4[3] = hc_bytealign_be (w3[2], w3[3], offset); + w4[2] = hc_bytealign_be (w3[1], w3[2], offset); + w4[1] = hc_bytealign_be (w3[0], w3[1], offset); + w4[0] = hc_bytealign_be (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -6333,33 +6179,33 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 5: - w7[3] = hc_bytealign (w6[1], w6[2], offset); - w7[2] = hc_bytealign (w6[0], w6[1], offset); - w7[1] = hc_bytealign (w5[3], w6[0], offset); - w7[0] = hc_bytealign (w5[2], w5[3], offset); - w6[3] = hc_bytealign (w5[1], w5[2], offset); - w6[2] = hc_bytealign (w5[0], w5[1], offset); - w6[1] = hc_bytealign (w4[3], w5[0], offset); - w6[0] = hc_bytealign (w4[2], w4[3], offset); - w5[3] = hc_bytealign (w4[1], w4[2], offset); - w5[2] = hc_bytealign (w4[0], w4[1], offset); - w5[1] = hc_bytealign (w3[3], w4[0], offset); - w5[0] = hc_bytealign (w3[2], w3[3], offset); - w4[3] = hc_bytealign (w3[1], w3[2], offset); - w4[2] = hc_bytealign (w3[0], w3[1], offset); - w4[1] = hc_bytealign (w2[3], w3[0], offset); - w4[0] = hc_bytealign (w2[2], w2[3], offset); - w3[3] = hc_bytealign (w2[1], w2[2], offset); - w3[2] = hc_bytealign (w2[0], w2[1], offset); - w3[1] = hc_bytealign (w1[3], w2[0], offset); - w3[0] = hc_bytealign (w1[2], w1[3], offset); - w2[3] = hc_bytealign (w1[1], w1[2], offset); - w2[2] = hc_bytealign (w1[0], w1[1], offset); - w2[1] = hc_bytealign (w0[3], w1[0], offset); - w2[0] = hc_bytealign (w0[2], w0[3], offset); - w1[3] = hc_bytealign (w0[1], w0[2], offset); - w1[2] = hc_bytealign (w0[0], w0[1], offset); - w1[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w6[1], w6[2], offset); + w7[2] = hc_bytealign_be (w6[0], w6[1], offset); + w7[1] = hc_bytealign_be (w5[3], w6[0], offset); + w7[0] = hc_bytealign_be (w5[2], w5[3], offset); + w6[3] = hc_bytealign_be (w5[1], w5[2], offset); + w6[2] = hc_bytealign_be (w5[0], w5[1], offset); + w6[1] = hc_bytealign_be (w4[3], w5[0], offset); + w6[0] = hc_bytealign_be (w4[2], w4[3], offset); + w5[3] = hc_bytealign_be (w4[1], w4[2], offset); + w5[2] = hc_bytealign_be (w4[0], w4[1], offset); + w5[1] = hc_bytealign_be (w3[3], w4[0], offset); + w5[0] = hc_bytealign_be (w3[2], w3[3], offset); + w4[3] = hc_bytealign_be (w3[1], w3[2], offset); + w4[2] = hc_bytealign_be (w3[0], w3[1], offset); + w4[1] = hc_bytealign_be (w2[3], w3[0], offset); + w4[0] = hc_bytealign_be (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -6369,32 +6215,32 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 6: - w7[3] = hc_bytealign (w6[0], w6[1], offset); - w7[2] = hc_bytealign (w5[3], w6[0], offset); - w7[1] = hc_bytealign (w5[2], w5[3], offset); - w7[0] = hc_bytealign (w5[1], w5[2], offset); - w6[3] = hc_bytealign (w5[0], w5[1], offset); - w6[2] = hc_bytealign (w4[3], w5[0], offset); - w6[1] = hc_bytealign (w4[2], w4[3], offset); - w6[0] = hc_bytealign (w4[1], w4[2], offset); - w5[3] = hc_bytealign (w4[0], w4[1], offset); - w5[2] = hc_bytealign (w3[3], w4[0], offset); - w5[1] = hc_bytealign (w3[2], w3[3], offset); - w5[0] = hc_bytealign (w3[1], w3[2], offset); - w4[3] = hc_bytealign (w3[0], w3[1], offset); - w4[2] = hc_bytealign (w2[3], w3[0], offset); - w4[1] = hc_bytealign (w2[2], w2[3], offset); - w4[0] = hc_bytealign (w2[1], w2[2], offset); - w3[3] = hc_bytealign (w2[0], w2[1], offset); - w3[2] = hc_bytealign (w1[3], w2[0], offset); - w3[1] = hc_bytealign (w1[2], w1[3], offset); - w3[0] = hc_bytealign (w1[1], w1[2], offset); - w2[3] = hc_bytealign (w1[0], w1[1], offset); - w2[2] = hc_bytealign (w0[3], w1[0], offset); - w2[1] = hc_bytealign (w0[2], w0[3], offset); - w2[0] = hc_bytealign (w0[1], w0[2], offset); - w1[3] = hc_bytealign (w0[0], w0[1], offset); - w1[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w6[0], w6[1], offset); + w7[2] = hc_bytealign_be (w5[3], w6[0], offset); + w7[1] = hc_bytealign_be (w5[2], w5[3], offset); + w7[0] = hc_bytealign_be (w5[1], w5[2], offset); + w6[3] = hc_bytealign_be (w5[0], w5[1], offset); + w6[2] = hc_bytealign_be (w4[3], w5[0], offset); + w6[1] = hc_bytealign_be (w4[2], w4[3], offset); + w6[0] = hc_bytealign_be (w4[1], w4[2], offset); + w5[3] = hc_bytealign_be (w4[0], w4[1], offset); + w5[2] = hc_bytealign_be (w3[3], w4[0], offset); + w5[1] = hc_bytealign_be (w3[2], w3[3], offset); + w5[0] = hc_bytealign_be (w3[1], w3[2], offset); + w4[3] = hc_bytealign_be (w3[0], w3[1], offset); + w4[2] = hc_bytealign_be (w2[3], w3[0], offset); + w4[1] = hc_bytealign_be (w2[2], w2[3], offset); + w4[0] = hc_bytealign_be (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -6405,31 +6251,31 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 7: - w7[3] = hc_bytealign (w5[3], w6[0], offset); - w7[2] = hc_bytealign (w5[2], w5[3], offset); - w7[1] = hc_bytealign (w5[1], w5[2], offset); - w7[0] = hc_bytealign (w5[0], w5[1], offset); - w6[3] = hc_bytealign (w4[3], w5[0], offset); - w6[2] = hc_bytealign (w4[2], w4[3], offset); - w6[1] = hc_bytealign (w4[1], w4[2], offset); - w6[0] = hc_bytealign (w4[0], w4[1], offset); - w5[3] = hc_bytealign (w3[3], w4[0], offset); - w5[2] = hc_bytealign (w3[2], w3[3], offset); - w5[1] = hc_bytealign (w3[1], w3[2], offset); - w5[0] = hc_bytealign (w3[0], w3[1], offset); - w4[3] = hc_bytealign (w2[3], w3[0], offset); - w4[2] = hc_bytealign (w2[2], w2[3], offset); - w4[1] = hc_bytealign (w2[1], w2[2], offset); - w4[0] = hc_bytealign (w2[0], w2[1], offset); - w3[3] = hc_bytealign (w1[3], w2[0], offset); - w3[2] = hc_bytealign (w1[2], w1[3], offset); - w3[1] = hc_bytealign (w1[1], w1[2], offset); - w3[0] = hc_bytealign (w1[0], w1[1], offset); - w2[3] = hc_bytealign (w0[3], w1[0], offset); - w2[2] = hc_bytealign (w0[2], w0[3], offset); - w2[1] = hc_bytealign (w0[1], w0[2], offset); - w2[0] = hc_bytealign (w0[0], w0[1], offset); - w1[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w5[3], w6[0], offset); + w7[2] = hc_bytealign_be (w5[2], w5[3], offset); + w7[1] = hc_bytealign_be (w5[1], w5[2], offset); + w7[0] = hc_bytealign_be (w5[0], w5[1], offset); + w6[3] = hc_bytealign_be (w4[3], w5[0], offset); + w6[2] = hc_bytealign_be (w4[2], w4[3], offset); + w6[1] = hc_bytealign_be (w4[1], w4[2], offset); + w6[0] = hc_bytealign_be (w4[0], w4[1], offset); + w5[3] = hc_bytealign_be (w3[3], w4[0], offset); + w5[2] = hc_bytealign_be (w3[2], w3[3], offset); + w5[1] = hc_bytealign_be (w3[1], w3[2], offset); + w5[0] = hc_bytealign_be (w3[0], w3[1], offset); + w4[3] = hc_bytealign_be (w2[3], w3[0], offset); + w4[2] = hc_bytealign_be (w2[2], w2[3], offset); + w4[1] = hc_bytealign_be (w2[1], w2[2], offset); + w4[0] = hc_bytealign_be (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -6441,30 +6287,30 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 8: - w7[3] = hc_bytealign (w5[2], w5[3], offset); - w7[2] = hc_bytealign (w5[1], w5[2], offset); - w7[1] = hc_bytealign (w5[0], w5[1], offset); - w7[0] = hc_bytealign (w4[3], w5[0], offset); - w6[3] = hc_bytealign (w4[2], w4[3], offset); - w6[2] = hc_bytealign (w4[1], w4[2], offset); - w6[1] = hc_bytealign (w4[0], w4[1], offset); - w6[0] = hc_bytealign (w3[3], w4[0], offset); - w5[3] = hc_bytealign (w3[2], w3[3], offset); - w5[2] = hc_bytealign (w3[1], w3[2], offset); - w5[1] = hc_bytealign (w3[0], w3[1], offset); - w5[0] = hc_bytealign (w2[3], w3[0], offset); - w4[3] = hc_bytealign (w2[2], w2[3], offset); - w4[2] = hc_bytealign (w2[1], w2[2], offset); - w4[1] = hc_bytealign (w2[0], w2[1], offset); - w4[0] = hc_bytealign (w1[3], w2[0], offset); - w3[3] = hc_bytealign (w1[2], w1[3], offset); - w3[2] = hc_bytealign (w1[1], w1[2], offset); - w3[1] = hc_bytealign (w1[0], w1[1], offset); - w3[0] = hc_bytealign (w0[3], w1[0], offset); - w2[3] = hc_bytealign (w0[2], w0[3], offset); - w2[2] = hc_bytealign (w0[1], w0[2], offset); - w2[1] = hc_bytealign (w0[0], w0[1], offset); - w2[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w5[2], w5[3], offset); + w7[2] = hc_bytealign_be (w5[1], w5[2], offset); + w7[1] = hc_bytealign_be (w5[0], w5[1], offset); + w7[0] = hc_bytealign_be (w4[3], w5[0], offset); + w6[3] = hc_bytealign_be (w4[2], w4[3], offset); + w6[2] = hc_bytealign_be (w4[1], w4[2], offset); + w6[1] = hc_bytealign_be (w4[0], w4[1], offset); + w6[0] = hc_bytealign_be (w3[3], w4[0], offset); + w5[3] = hc_bytealign_be (w3[2], w3[3], offset); + w5[2] = hc_bytealign_be (w3[1], w3[2], offset); + w5[1] = hc_bytealign_be (w3[0], w3[1], offset); + w5[0] = hc_bytealign_be (w2[3], w3[0], offset); + w4[3] = hc_bytealign_be (w2[2], w2[3], offset); + w4[2] = hc_bytealign_be (w2[1], w2[2], offset); + w4[1] = hc_bytealign_be (w2[0], w2[1], offset); + w4[0] = hc_bytealign_be (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -6477,29 +6323,29 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 9: - w7[3] = hc_bytealign (w5[1], w5[2], offset); - w7[2] = hc_bytealign (w5[0], w5[1], offset); - w7[1] = hc_bytealign (w4[3], w5[0], offset); - w7[0] = hc_bytealign (w4[2], w4[3], offset); - w6[3] = hc_bytealign (w4[1], w4[2], offset); - w6[2] = hc_bytealign (w4[0], w4[1], offset); - w6[1] = hc_bytealign (w3[3], w4[0], offset); - w6[0] = hc_bytealign (w3[2], w3[3], offset); - w5[3] = hc_bytealign (w3[1], w3[2], offset); - w5[2] = hc_bytealign (w3[0], w3[1], offset); - w5[1] = hc_bytealign (w2[3], w3[0], offset); - w5[0] = hc_bytealign (w2[2], w2[3], offset); - w4[3] = hc_bytealign (w2[1], w2[2], offset); - w4[2] = hc_bytealign (w2[0], w2[1], offset); - w4[1] = hc_bytealign (w1[3], w2[0], offset); - w4[0] = hc_bytealign (w1[2], w1[3], offset); - w3[3] = hc_bytealign (w1[1], w1[2], offset); - w3[2] = hc_bytealign (w1[0], w1[1], offset); - w3[1] = hc_bytealign (w0[3], w1[0], offset); - w3[0] = hc_bytealign (w0[2], w0[3], offset); - w2[3] = hc_bytealign (w0[1], w0[2], offset); - w2[2] = hc_bytealign (w0[0], w0[1], offset); - w2[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w5[1], w5[2], offset); + w7[2] = hc_bytealign_be (w5[0], w5[1], offset); + w7[1] = hc_bytealign_be (w4[3], w5[0], offset); + w7[0] = hc_bytealign_be (w4[2], w4[3], offset); + w6[3] = hc_bytealign_be (w4[1], w4[2], offset); + w6[2] = hc_bytealign_be (w4[0], w4[1], offset); + w6[1] = hc_bytealign_be (w3[3], w4[0], offset); + w6[0] = hc_bytealign_be (w3[2], w3[3], offset); + w5[3] = hc_bytealign_be (w3[1], w3[2], offset); + w5[2] = hc_bytealign_be (w3[0], w3[1], offset); + w5[1] = hc_bytealign_be (w2[3], w3[0], offset); + w5[0] = hc_bytealign_be (w2[2], w2[3], offset); + w4[3] = hc_bytealign_be (w2[1], w2[2], offset); + w4[2] = hc_bytealign_be (w2[0], w2[1], offset); + w4[1] = hc_bytealign_be (w1[3], w2[0], offset); + w4[0] = hc_bytealign_be (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -6513,28 +6359,28 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 10: - w7[3] = hc_bytealign (w5[0], w5[1], offset); - w7[2] = hc_bytealign (w4[3], w5[0], offset); - w7[1] = hc_bytealign (w4[2], w4[3], offset); - w7[0] = hc_bytealign (w4[1], w4[2], offset); - w6[3] = hc_bytealign (w4[0], w4[1], offset); - w6[2] = hc_bytealign (w3[3], w4[0], offset); - w6[1] = hc_bytealign (w3[2], w3[3], offset); - w6[0] = hc_bytealign (w3[1], w3[2], offset); - w5[3] = hc_bytealign (w3[0], w3[1], offset); - w5[2] = hc_bytealign (w2[3], w3[0], offset); - w5[1] = hc_bytealign (w2[2], w2[3], offset); - w5[0] = hc_bytealign (w2[1], w2[2], offset); - w4[3] = hc_bytealign (w2[0], w2[1], offset); - w4[2] = hc_bytealign (w1[3], w2[0], offset); - w4[1] = hc_bytealign (w1[2], w1[3], offset); - w4[0] = hc_bytealign (w1[1], w1[2], offset); - w3[3] = hc_bytealign (w1[0], w1[1], offset); - w3[2] = hc_bytealign (w0[3], w1[0], offset); - w3[1] = hc_bytealign (w0[2], w0[3], offset); - w3[0] = hc_bytealign (w0[1], w0[2], offset); - w2[3] = hc_bytealign (w0[0], w0[1], offset); - w2[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w5[0], w5[1], offset); + w7[2] = hc_bytealign_be (w4[3], w5[0], offset); + w7[1] = hc_bytealign_be (w4[2], w4[3], offset); + w7[0] = hc_bytealign_be (w4[1], w4[2], offset); + w6[3] = hc_bytealign_be (w4[0], w4[1], offset); + w6[2] = hc_bytealign_be (w3[3], w4[0], offset); + w6[1] = hc_bytealign_be (w3[2], w3[3], offset); + w6[0] = hc_bytealign_be (w3[1], w3[2], offset); + w5[3] = hc_bytealign_be (w3[0], w3[1], offset); + w5[2] = hc_bytealign_be (w2[3], w3[0], offset); + w5[1] = hc_bytealign_be (w2[2], w2[3], offset); + w5[0] = hc_bytealign_be (w2[1], w2[2], offset); + w4[3] = hc_bytealign_be (w2[0], w2[1], offset); + w4[2] = hc_bytealign_be (w1[3], w2[0], offset); + w4[1] = hc_bytealign_be (w1[2], w1[3], offset); + w4[0] = hc_bytealign_be (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -6549,27 +6395,27 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 11: - w7[3] = hc_bytealign (w4[3], w5[0], offset); - w7[2] = hc_bytealign (w4[2], w4[3], offset); - w7[1] = hc_bytealign (w4[1], w4[2], offset); - w7[0] = hc_bytealign (w4[0], w4[1], offset); - w6[3] = hc_bytealign (w3[3], w4[0], offset); - w6[2] = hc_bytealign (w3[2], w3[3], offset); - w6[1] = hc_bytealign (w3[1], w3[2], offset); - w6[0] = hc_bytealign (w3[0], w3[1], offset); - w5[3] = hc_bytealign (w2[3], w3[0], offset); - w5[2] = hc_bytealign (w2[2], w2[3], offset); - w5[1] = hc_bytealign (w2[1], w2[2], offset); - w5[0] = hc_bytealign (w2[0], w2[1], offset); - w4[3] = hc_bytealign (w1[3], w2[0], offset); - w4[2] = hc_bytealign (w1[2], w1[3], offset); - w4[1] = hc_bytealign (w1[1], w1[2], offset); - w4[0] = hc_bytealign (w1[0], w1[1], offset); - w3[3] = hc_bytealign (w0[3], w1[0], offset); - w3[2] = hc_bytealign (w0[2], w0[3], offset); - w3[1] = hc_bytealign (w0[1], w0[2], offset); - w3[0] = hc_bytealign (w0[0], w0[1], offset); - w2[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w4[3], w5[0], offset); + w7[2] = hc_bytealign_be (w4[2], w4[3], offset); + w7[1] = hc_bytealign_be (w4[1], w4[2], offset); + w7[0] = hc_bytealign_be (w4[0], w4[1], offset); + w6[3] = hc_bytealign_be (w3[3], w4[0], offset); + w6[2] = hc_bytealign_be (w3[2], w3[3], offset); + w6[1] = hc_bytealign_be (w3[1], w3[2], offset); + w6[0] = hc_bytealign_be (w3[0], w3[1], offset); + w5[3] = hc_bytealign_be (w2[3], w3[0], offset); + w5[2] = hc_bytealign_be (w2[2], w2[3], offset); + w5[1] = hc_bytealign_be (w2[1], w2[2], offset); + w5[0] = hc_bytealign_be (w2[0], w2[1], offset); + w4[3] = hc_bytealign_be (w1[3], w2[0], offset); + w4[2] = hc_bytealign_be (w1[2], w1[3], offset); + w4[1] = hc_bytealign_be (w1[1], w1[2], offset); + w4[0] = hc_bytealign_be (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -6585,26 +6431,26 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 12: - w7[3] = hc_bytealign (w4[2], w4[3], offset); - w7[2] = hc_bytealign (w4[1], w4[2], offset); - w7[1] = hc_bytealign (w4[0], w4[1], offset); - w7[0] = hc_bytealign (w3[3], w4[0], offset); - w6[3] = hc_bytealign (w3[2], w3[3], offset); - w6[2] = hc_bytealign (w3[1], w3[2], offset); - w6[1] = hc_bytealign (w3[0], w3[1], offset); - w6[0] = hc_bytealign (w2[3], w3[0], offset); - w5[3] = hc_bytealign (w2[2], w2[3], offset); - w5[2] = hc_bytealign (w2[1], w2[2], offset); - w5[1] = hc_bytealign (w2[0], w2[1], offset); - w5[0] = hc_bytealign (w1[3], w2[0], offset); - w4[3] = hc_bytealign (w1[2], w1[3], offset); - w4[2] = hc_bytealign (w1[1], w1[2], offset); - w4[1] = hc_bytealign (w1[0], w1[1], offset); - w4[0] = hc_bytealign (w0[3], w1[0], offset); - w3[3] = hc_bytealign (w0[2], w0[3], offset); - w3[2] = hc_bytealign (w0[1], w0[2], offset); - w3[1] = hc_bytealign (w0[0], w0[1], offset); - w3[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w4[2], w4[3], offset); + w7[2] = hc_bytealign_be (w4[1], w4[2], offset); + w7[1] = hc_bytealign_be (w4[0], w4[1], offset); + w7[0] = hc_bytealign_be (w3[3], w4[0], offset); + w6[3] = hc_bytealign_be (w3[2], w3[3], offset); + w6[2] = hc_bytealign_be (w3[1], w3[2], offset); + w6[1] = hc_bytealign_be (w3[0], w3[1], offset); + w6[0] = hc_bytealign_be (w2[3], w3[0], offset); + w5[3] = hc_bytealign_be (w2[2], w2[3], offset); + w5[2] = hc_bytealign_be (w2[1], w2[2], offset); + w5[1] = hc_bytealign_be (w2[0], w2[1], offset); + w5[0] = hc_bytealign_be (w1[3], w2[0], offset); + w4[3] = hc_bytealign_be (w1[2], w1[3], offset); + w4[2] = hc_bytealign_be (w1[1], w1[2], offset); + w4[1] = hc_bytealign_be (w1[0], w1[1], offset); + w4[0] = hc_bytealign_be (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -6621,25 +6467,25 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 13: - w7[3] = hc_bytealign (w4[1], w4[2], offset); - w7[2] = hc_bytealign (w4[0], w4[1], offset); - w7[1] = hc_bytealign (w3[3], w4[0], offset); - w7[0] = hc_bytealign (w3[2], w3[3], offset); - w6[3] = hc_bytealign (w3[1], w3[2], offset); - w6[2] = hc_bytealign (w3[0], w3[1], offset); - w6[1] = hc_bytealign (w2[3], w3[0], offset); - w6[0] = hc_bytealign (w2[2], w2[3], offset); - w5[3] = hc_bytealign (w2[1], w2[2], offset); - w5[2] = hc_bytealign (w2[0], w2[1], offset); - w5[1] = hc_bytealign (w1[3], w2[0], offset); - w5[0] = hc_bytealign (w1[2], w1[3], offset); - w4[3] = hc_bytealign (w1[1], w1[2], offset); - w4[2] = hc_bytealign (w1[0], w1[1], offset); - w4[1] = hc_bytealign (w0[3], w1[0], offset); - w4[0] = hc_bytealign (w0[2], w0[3], offset); - w3[3] = hc_bytealign (w0[1], w0[2], offset); - w3[2] = hc_bytealign (w0[0], w0[1], offset); - w3[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w4[1], w4[2], offset); + w7[2] = hc_bytealign_be (w4[0], w4[1], offset); + w7[1] = hc_bytealign_be (w3[3], w4[0], offset); + w7[0] = hc_bytealign_be (w3[2], w3[3], offset); + w6[3] = hc_bytealign_be (w3[1], w3[2], offset); + w6[2] = hc_bytealign_be (w3[0], w3[1], offset); + w6[1] = hc_bytealign_be (w2[3], w3[0], offset); + w6[0] = hc_bytealign_be (w2[2], w2[3], offset); + w5[3] = hc_bytealign_be (w2[1], w2[2], offset); + w5[2] = hc_bytealign_be (w2[0], w2[1], offset); + w5[1] = hc_bytealign_be (w1[3], w2[0], offset); + w5[0] = hc_bytealign_be (w1[2], w1[3], offset); + w4[3] = hc_bytealign_be (w1[1], w1[2], offset); + w4[2] = hc_bytealign_be (w1[0], w1[1], offset); + w4[1] = hc_bytealign_be (w0[3], w1[0], offset); + w4[0] = hc_bytealign_be (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -6657,24 +6503,24 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 14: - w7[3] = hc_bytealign (w4[0], w4[1], offset); - w7[2] = hc_bytealign (w3[3], w4[0], offset); - w7[1] = hc_bytealign (w3[2], w3[3], offset); - w7[0] = hc_bytealign (w3[1], w3[2], offset); - w6[3] = hc_bytealign (w3[0], w3[1], offset); - w6[2] = hc_bytealign (w2[3], w3[0], offset); - w6[1] = hc_bytealign (w2[2], w2[3], offset); - w6[0] = hc_bytealign (w2[1], w2[2], offset); - w5[3] = hc_bytealign (w2[0], w2[1], offset); - w5[2] = hc_bytealign (w1[3], w2[0], offset); - w5[1] = hc_bytealign (w1[2], w1[3], offset); - w5[0] = hc_bytealign (w1[1], w1[2], offset); - w4[3] = hc_bytealign (w1[0], w1[1], offset); - w4[2] = hc_bytealign (w0[3], w1[0], offset); - w4[1] = hc_bytealign (w0[2], w0[3], offset); - w4[0] = hc_bytealign (w0[1], w0[2], offset); - w3[3] = hc_bytealign (w0[0], w0[1], offset); - w3[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w4[0], w4[1], offset); + w7[2] = hc_bytealign_be (w3[3], w4[0], offset); + w7[1] = hc_bytealign_be (w3[2], w3[3], offset); + w7[0] = hc_bytealign_be (w3[1], w3[2], offset); + w6[3] = hc_bytealign_be (w3[0], w3[1], offset); + w6[2] = hc_bytealign_be (w2[3], w3[0], offset); + w6[1] = hc_bytealign_be (w2[2], w2[3], offset); + w6[0] = hc_bytealign_be (w2[1], w2[2], offset); + w5[3] = hc_bytealign_be (w2[0], w2[1], offset); + w5[2] = hc_bytealign_be (w1[3], w2[0], offset); + w5[1] = hc_bytealign_be (w1[2], w1[3], offset); + w5[0] = hc_bytealign_be (w1[1], w1[2], offset); + w4[3] = hc_bytealign_be (w1[0], w1[1], offset); + w4[2] = hc_bytealign_be (w0[3], w1[0], offset); + w4[1] = hc_bytealign_be (w0[2], w0[3], offset); + w4[0] = hc_bytealign_be (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -6693,23 +6539,23 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 15: - w7[3] = hc_bytealign (w3[3], w4[0], offset); - w7[2] = hc_bytealign (w3[2], w3[3], offset); - w7[1] = hc_bytealign (w3[1], w3[2], offset); - w7[0] = hc_bytealign (w3[0], w3[1], offset); - w6[3] = hc_bytealign (w2[3], w3[0], offset); - w6[2] = hc_bytealign (w2[2], w2[3], offset); - w6[1] = hc_bytealign (w2[1], w2[2], offset); - w6[0] = hc_bytealign (w2[0], w2[1], offset); - w5[3] = hc_bytealign (w1[3], w2[0], offset); - w5[2] = hc_bytealign (w1[2], w1[3], offset); - w5[1] = hc_bytealign (w1[1], w1[2], offset); - w5[0] = hc_bytealign (w1[0], w1[1], offset); - w4[3] = hc_bytealign (w0[3], w1[0], offset); - w4[2] = hc_bytealign (w0[2], w0[3], offset); - w4[1] = hc_bytealign (w0[1], w0[2], offset); - w4[0] = hc_bytealign (w0[0], w0[1], offset); - w3[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w3[3], w4[0], offset); + w7[2] = hc_bytealign_be (w3[2], w3[3], offset); + w7[1] = hc_bytealign_be (w3[1], w3[2], offset); + w7[0] = hc_bytealign_be (w3[0], w3[1], offset); + w6[3] = hc_bytealign_be (w2[3], w3[0], offset); + w6[2] = hc_bytealign_be (w2[2], w2[3], offset); + w6[1] = hc_bytealign_be (w2[1], w2[2], offset); + w6[0] = hc_bytealign_be (w2[0], w2[1], offset); + w5[3] = hc_bytealign_be (w1[3], w2[0], offset); + w5[2] = hc_bytealign_be (w1[2], w1[3], offset); + w5[1] = hc_bytealign_be (w1[1], w1[2], offset); + w5[0] = hc_bytealign_be (w1[0], w1[1], offset); + w4[3] = hc_bytealign_be (w0[3], w1[0], offset); + w4[2] = hc_bytealign_be (w0[2], w0[3], offset); + w4[1] = hc_bytealign_be (w0[1], w0[2], offset); + w4[0] = hc_bytealign_be (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -6729,22 +6575,22 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 16: - w7[3] = hc_bytealign (w3[2], w3[3], offset); - w7[2] = hc_bytealign (w3[1], w3[2], offset); - w7[1] = hc_bytealign (w3[0], w3[1], offset); - w7[0] = hc_bytealign (w2[3], w3[0], offset); - w6[3] = hc_bytealign (w2[2], w2[3], offset); - w6[2] = hc_bytealign (w2[1], w2[2], offset); - w6[1] = hc_bytealign (w2[0], w2[1], offset); - w6[0] = hc_bytealign (w1[3], w2[0], offset); - w5[3] = hc_bytealign (w1[2], w1[3], offset); - w5[2] = hc_bytealign (w1[1], w1[2], offset); - w5[1] = hc_bytealign (w1[0], w1[1], offset); - w5[0] = hc_bytealign (w0[3], w1[0], offset); - w4[3] = hc_bytealign (w0[2], w0[3], offset); - w4[2] = hc_bytealign (w0[1], w0[2], offset); - w4[1] = hc_bytealign (w0[0], w0[1], offset); - w4[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w3[2], w3[3], offset); + w7[2] = hc_bytealign_be (w3[1], w3[2], offset); + w7[1] = hc_bytealign_be (w3[0], w3[1], offset); + w7[0] = hc_bytealign_be (w2[3], w3[0], offset); + w6[3] = hc_bytealign_be (w2[2], w2[3], offset); + w6[2] = hc_bytealign_be (w2[1], w2[2], offset); + w6[1] = hc_bytealign_be (w2[0], w2[1], offset); + w6[0] = hc_bytealign_be (w1[3], w2[0], offset); + w5[3] = hc_bytealign_be (w1[2], w1[3], offset); + w5[2] = hc_bytealign_be (w1[1], w1[2], offset); + w5[1] = hc_bytealign_be (w1[0], w1[1], offset); + w5[0] = hc_bytealign_be (w0[3], w1[0], offset); + w4[3] = hc_bytealign_be (w0[2], w0[3], offset); + w4[2] = hc_bytealign_be (w0[1], w0[2], offset); + w4[1] = hc_bytealign_be (w0[0], w0[1], offset); + w4[0] = hc_bytealign_be ( 0, w0[0], offset); w3[3] = 0; w3[2] = 0; w3[1] = 0; @@ -6765,21 +6611,21 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 17: - w7[3] = hc_bytealign (w3[1], w3[2], offset); - w7[2] = hc_bytealign (w3[0], w3[1], offset); - w7[1] = hc_bytealign (w2[3], w3[0], offset); - w7[0] = hc_bytealign (w2[2], w2[3], offset); - w6[3] = hc_bytealign (w2[1], w2[2], offset); - w6[2] = hc_bytealign (w2[0], w2[1], offset); - w6[1] = hc_bytealign (w1[3], w2[0], offset); - w6[0] = hc_bytealign (w1[2], w1[3], offset); - w5[3] = hc_bytealign (w1[1], w1[2], offset); - w5[2] = hc_bytealign (w1[0], w1[1], offset); - w5[1] = hc_bytealign (w0[3], w1[0], offset); - w5[0] = hc_bytealign (w0[2], w0[3], offset); - w4[3] = hc_bytealign (w0[1], w0[2], offset); - w4[2] = hc_bytealign (w0[0], w0[1], offset); - w4[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w3[1], w3[2], offset); + w7[2] = hc_bytealign_be (w3[0], w3[1], offset); + w7[1] = hc_bytealign_be (w2[3], w3[0], offset); + w7[0] = hc_bytealign_be (w2[2], w2[3], offset); + w6[3] = hc_bytealign_be (w2[1], w2[2], offset); + w6[2] = hc_bytealign_be (w2[0], w2[1], offset); + w6[1] = hc_bytealign_be (w1[3], w2[0], offset); + w6[0] = hc_bytealign_be (w1[2], w1[3], offset); + w5[3] = hc_bytealign_be (w1[1], w1[2], offset); + w5[2] = hc_bytealign_be (w1[0], w1[1], offset); + w5[1] = hc_bytealign_be (w0[3], w1[0], offset); + w5[0] = hc_bytealign_be (w0[2], w0[3], offset); + w4[3] = hc_bytealign_be (w0[1], w0[2], offset); + w4[2] = hc_bytealign_be (w0[0], w0[1], offset); + w4[1] = hc_bytealign_be ( 0, w0[0], offset); w4[0] = 0; w3[3] = 0; w3[2] = 0; @@ -6801,20 +6647,20 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 18: - w7[3] = hc_bytealign (w3[0], w3[1], offset); - w7[2] = hc_bytealign (w2[3], w3[0], offset); - w7[1] = hc_bytealign (w2[2], w2[3], offset); - w7[0] = hc_bytealign (w2[1], w2[2], offset); - w6[3] = hc_bytealign (w2[0], w2[1], offset); - w6[2] = hc_bytealign (w1[3], w2[0], offset); - w6[1] = hc_bytealign (w1[2], w1[3], offset); - w6[0] = hc_bytealign (w1[1], w1[2], offset); - w5[3] = hc_bytealign (w1[0], w1[1], offset); - w5[2] = hc_bytealign (w0[3], w1[0], offset); - w5[1] = hc_bytealign (w0[2], w0[3], offset); - w5[0] = hc_bytealign (w0[1], w0[2], offset); - w4[3] = hc_bytealign (w0[0], w0[1], offset); - w4[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w3[0], w3[1], offset); + w7[2] = hc_bytealign_be (w2[3], w3[0], offset); + w7[1] = hc_bytealign_be (w2[2], w2[3], offset); + w7[0] = hc_bytealign_be (w2[1], w2[2], offset); + w6[3] = hc_bytealign_be (w2[0], w2[1], offset); + w6[2] = hc_bytealign_be (w1[3], w2[0], offset); + w6[1] = hc_bytealign_be (w1[2], w1[3], offset); + w6[0] = hc_bytealign_be (w1[1], w1[2], offset); + w5[3] = hc_bytealign_be (w1[0], w1[1], offset); + w5[2] = hc_bytealign_be (w0[3], w1[0], offset); + w5[1] = hc_bytealign_be (w0[2], w0[3], offset); + w5[0] = hc_bytealign_be (w0[1], w0[2], offset); + w4[3] = hc_bytealign_be (w0[0], w0[1], offset); + w4[2] = hc_bytealign_be ( 0, w0[0], offset); w4[1] = 0; w4[0] = 0; w3[3] = 0; @@ -6837,19 +6683,19 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 19: - w7[3] = hc_bytealign (w2[3], w3[0], offset); - w7[2] = hc_bytealign (w2[2], w2[3], offset); - w7[1] = hc_bytealign (w2[1], w2[2], offset); - w7[0] = hc_bytealign (w2[0], w2[1], offset); - w6[3] = hc_bytealign (w1[3], w2[0], offset); - w6[2] = hc_bytealign (w1[2], w1[3], offset); - w6[1] = hc_bytealign (w1[1], w1[2], offset); - w6[0] = hc_bytealign (w1[0], w1[1], offset); - w5[3] = hc_bytealign (w0[3], w1[0], offset); - w5[2] = hc_bytealign (w0[2], w0[3], offset); - w5[1] = hc_bytealign (w0[1], w0[2], offset); - w5[0] = hc_bytealign (w0[0], w0[1], offset); - w4[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w2[3], w3[0], offset); + w7[2] = hc_bytealign_be (w2[2], w2[3], offset); + w7[1] = hc_bytealign_be (w2[1], w2[2], offset); + w7[0] = hc_bytealign_be (w2[0], w2[1], offset); + w6[3] = hc_bytealign_be (w1[3], w2[0], offset); + w6[2] = hc_bytealign_be (w1[2], w1[3], offset); + w6[1] = hc_bytealign_be (w1[1], w1[2], offset); + w6[0] = hc_bytealign_be (w1[0], w1[1], offset); + w5[3] = hc_bytealign_be (w0[3], w1[0], offset); + w5[2] = hc_bytealign_be (w0[2], w0[3], offset); + w5[1] = hc_bytealign_be (w0[1], w0[2], offset); + w5[0] = hc_bytealign_be (w0[0], w0[1], offset); + w4[3] = hc_bytealign_be ( 0, w0[0], offset); w4[2] = 0; w4[1] = 0; w4[0] = 0; @@ -6873,18 +6719,18 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 20: - w7[3] = hc_bytealign (w2[2], w2[3], offset); - w7[2] = hc_bytealign (w2[1], w2[2], offset); - w7[1] = hc_bytealign (w2[0], w2[1], offset); - w7[0] = hc_bytealign (w1[3], w2[0], offset); - w6[3] = hc_bytealign (w1[2], w1[3], offset); - w6[2] = hc_bytealign (w1[1], w1[2], offset); - w6[1] = hc_bytealign (w1[0], w1[1], offset); - w6[0] = hc_bytealign (w0[3], w1[0], offset); - w5[3] = hc_bytealign (w0[2], w0[3], offset); - w5[2] = hc_bytealign (w0[1], w0[2], offset); - w5[1] = hc_bytealign (w0[0], w0[1], offset); - w5[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w2[2], w2[3], offset); + w7[2] = hc_bytealign_be (w2[1], w2[2], offset); + w7[1] = hc_bytealign_be (w2[0], w2[1], offset); + w7[0] = hc_bytealign_be (w1[3], w2[0], offset); + w6[3] = hc_bytealign_be (w1[2], w1[3], offset); + w6[2] = hc_bytealign_be (w1[1], w1[2], offset); + w6[1] = hc_bytealign_be (w1[0], w1[1], offset); + w6[0] = hc_bytealign_be (w0[3], w1[0], offset); + w5[3] = hc_bytealign_be (w0[2], w0[3], offset); + w5[2] = hc_bytealign_be (w0[1], w0[2], offset); + w5[1] = hc_bytealign_be (w0[0], w0[1], offset); + w5[0] = hc_bytealign_be ( 0, w0[0], offset); w4[3] = 0; w4[2] = 0; w4[1] = 0; @@ -6909,17 +6755,17 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 21: - w7[3] = hc_bytealign (w2[1], w2[2], offset); - w7[2] = hc_bytealign (w2[0], w2[1], offset); - w7[1] = hc_bytealign (w1[3], w2[0], offset); - w7[0] = hc_bytealign (w1[2], w1[3], offset); - w6[3] = hc_bytealign (w1[1], w1[2], offset); - w6[2] = hc_bytealign (w1[0], w1[1], offset); - w6[1] = hc_bytealign (w0[3], w1[0], offset); - w6[0] = hc_bytealign (w0[2], w0[3], offset); - w5[3] = hc_bytealign (w0[1], w0[2], offset); - w5[2] = hc_bytealign (w0[0], w0[1], offset); - w5[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w2[1], w2[2], offset); + w7[2] = hc_bytealign_be (w2[0], w2[1], offset); + w7[1] = hc_bytealign_be (w1[3], w2[0], offset); + w7[0] = hc_bytealign_be (w1[2], w1[3], offset); + w6[3] = hc_bytealign_be (w1[1], w1[2], offset); + w6[2] = hc_bytealign_be (w1[0], w1[1], offset); + w6[1] = hc_bytealign_be (w0[3], w1[0], offset); + w6[0] = hc_bytealign_be (w0[2], w0[3], offset); + w5[3] = hc_bytealign_be (w0[1], w0[2], offset); + w5[2] = hc_bytealign_be (w0[0], w0[1], offset); + w5[1] = hc_bytealign_be ( 0, w0[0], offset); w5[0] = 0; w4[3] = 0; w4[2] = 0; @@ -6945,16 +6791,16 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 22: - w7[3] = hc_bytealign (w2[0], w2[1], offset); - w7[2] = hc_bytealign (w1[3], w2[0], offset); - w7[1] = hc_bytealign (w1[2], w1[3], offset); - w7[0] = hc_bytealign (w1[1], w1[2], offset); - w6[3] = hc_bytealign (w1[0], w1[1], offset); - w6[2] = hc_bytealign (w0[3], w1[0], offset); - w6[1] = hc_bytealign (w0[2], w0[3], offset); - w6[0] = hc_bytealign (w0[1], w0[2], offset); - w5[3] = hc_bytealign (w0[0], w0[1], offset); - w5[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w2[0], w2[1], offset); + w7[2] = hc_bytealign_be (w1[3], w2[0], offset); + w7[1] = hc_bytealign_be (w1[2], w1[3], offset); + w7[0] = hc_bytealign_be (w1[1], w1[2], offset); + w6[3] = hc_bytealign_be (w1[0], w1[1], offset); + w6[2] = hc_bytealign_be (w0[3], w1[0], offset); + w6[1] = hc_bytealign_be (w0[2], w0[3], offset); + w6[0] = hc_bytealign_be (w0[1], w0[2], offset); + w5[3] = hc_bytealign_be (w0[0], w0[1], offset); + w5[2] = hc_bytealign_be ( 0, w0[0], offset); w5[1] = 0; w5[0] = 0; w4[3] = 0; @@ -6981,15 +6827,15 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 23: - w7[3] = hc_bytealign (w1[3], w2[0], offset); - w7[2] = hc_bytealign (w1[2], w1[3], offset); - w7[1] = hc_bytealign (w1[1], w1[2], offset); - w7[0] = hc_bytealign (w1[0], w1[1], offset); - w6[3] = hc_bytealign (w0[3], w1[0], offset); - w6[2] = hc_bytealign (w0[2], w0[3], offset); - w6[1] = hc_bytealign (w0[1], w0[2], offset); - w6[0] = hc_bytealign (w0[0], w0[1], offset); - w5[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w1[3], w2[0], offset); + w7[2] = hc_bytealign_be (w1[2], w1[3], offset); + w7[1] = hc_bytealign_be (w1[1], w1[2], offset); + w7[0] = hc_bytealign_be (w1[0], w1[1], offset); + w6[3] = hc_bytealign_be (w0[3], w1[0], offset); + w6[2] = hc_bytealign_be (w0[2], w0[3], offset); + w6[1] = hc_bytealign_be (w0[1], w0[2], offset); + w6[0] = hc_bytealign_be (w0[0], w0[1], offset); + w5[3] = hc_bytealign_be ( 0, w0[0], offset); w5[2] = 0; w5[1] = 0; w5[0] = 0; @@ -7017,14 +6863,14 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 24: - w7[3] = hc_bytealign (w1[2], w1[3], offset); - w7[2] = hc_bytealign (w1[1], w1[2], offset); - w7[1] = hc_bytealign (w1[0], w1[1], offset); - w7[0] = hc_bytealign (w0[3], w1[0], offset); - w6[3] = hc_bytealign (w0[2], w0[3], offset); - w6[2] = hc_bytealign (w0[1], w0[2], offset); - w6[1] = hc_bytealign (w0[0], w0[1], offset); - w6[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w1[2], w1[3], offset); + w7[2] = hc_bytealign_be (w1[1], w1[2], offset); + w7[1] = hc_bytealign_be (w1[0], w1[1], offset); + w7[0] = hc_bytealign_be (w0[3], w1[0], offset); + w6[3] = hc_bytealign_be (w0[2], w0[3], offset); + w6[2] = hc_bytealign_be (w0[1], w0[2], offset); + w6[1] = hc_bytealign_be (w0[0], w0[1], offset); + w6[0] = hc_bytealign_be ( 0, w0[0], offset); w5[3] = 0; w5[2] = 0; w5[1] = 0; @@ -7053,13 +6899,13 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 25: - w7[3] = hc_bytealign (w1[1], w1[2], offset); - w7[2] = hc_bytealign (w1[0], w1[1], offset); - w7[1] = hc_bytealign (w0[3], w1[0], offset); - w7[0] = hc_bytealign (w0[2], w0[3], offset); - w6[3] = hc_bytealign (w0[1], w0[2], offset); - w6[2] = hc_bytealign (w0[0], w0[1], offset); - w6[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w1[1], w1[2], offset); + w7[2] = hc_bytealign_be (w1[0], w1[1], offset); + w7[1] = hc_bytealign_be (w0[3], w1[0], offset); + w7[0] = hc_bytealign_be (w0[2], w0[3], offset); + w6[3] = hc_bytealign_be (w0[1], w0[2], offset); + w6[2] = hc_bytealign_be (w0[0], w0[1], offset); + w6[1] = hc_bytealign_be ( 0, w0[0], offset); w6[0] = 0; w5[3] = 0; w5[2] = 0; @@ -7089,12 +6935,12 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 26: - w7[3] = hc_bytealign (w1[0], w1[1], offset); - w7[2] = hc_bytealign (w0[3], w1[0], offset); - w7[1] = hc_bytealign (w0[2], w0[3], offset); - w7[0] = hc_bytealign (w0[1], w0[2], offset); - w6[3] = hc_bytealign (w0[0], w0[1], offset); - w6[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w1[0], w1[1], offset); + w7[2] = hc_bytealign_be (w0[3], w1[0], offset); + w7[1] = hc_bytealign_be (w0[2], w0[3], offset); + w7[0] = hc_bytealign_be (w0[1], w0[2], offset); + w6[3] = hc_bytealign_be (w0[0], w0[1], offset); + w6[2] = hc_bytealign_be ( 0, w0[0], offset); w6[1] = 0; w6[0] = 0; w5[3] = 0; @@ -7125,11 +6971,11 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 27: - w7[3] = hc_bytealign (w0[3], w1[0], offset); - w7[2] = hc_bytealign (w0[2], w0[3], offset); - w7[1] = hc_bytealign (w0[1], w0[2], offset); - w7[0] = hc_bytealign (w0[0], w0[1], offset); - w6[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w0[3], w1[0], offset); + w7[2] = hc_bytealign_be (w0[2], w0[3], offset); + w7[1] = hc_bytealign_be (w0[1], w0[2], offset); + w7[0] = hc_bytealign_be (w0[0], w0[1], offset); + w6[3] = hc_bytealign_be ( 0, w0[0], offset); w6[2] = 0; w6[1] = 0; w6[0] = 0; @@ -7161,10 +7007,10 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 28: - w7[3] = hc_bytealign (w0[2], w0[3], offset); - w7[2] = hc_bytealign (w0[1], w0[2], offset); - w7[1] = hc_bytealign (w0[0], w0[1], offset); - w7[0] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w0[2], w0[3], offset); + w7[2] = hc_bytealign_be (w0[1], w0[2], offset); + w7[1] = hc_bytealign_be (w0[0], w0[1], offset); + w7[0] = hc_bytealign_be ( 0, w0[0], offset); w6[3] = 0; w6[2] = 0; w6[1] = 0; @@ -7197,9 +7043,9 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 29: - w7[3] = hc_bytealign (w0[1], w0[2], offset); - w7[2] = hc_bytealign (w0[0], w0[1], offset); - w7[1] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w0[1], w0[2], offset); + w7[2] = hc_bytealign_be (w0[0], w0[1], offset); + w7[1] = hc_bytealign_be ( 0, w0[0], offset); w7[0] = 0; w6[3] = 0; w6[2] = 0; @@ -7233,8 +7079,8 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 30: - w7[3] = hc_bytealign (w0[0], w0[1], offset); - w7[2] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be (w0[0], w0[1], offset); + w7[2] = hc_bytealign_be ( 0, w0[0], offset); w7[1] = 0; w7[0] = 0; w6[3] = 0; @@ -7269,7 +7115,7 @@ DECLSPEC void switch_buffer_by_offset_8x4_be (u32x *w0, u32x *w1, u32x *w2, u32x break; case 31: - w7[3] = hc_bytealign ( 0, w0[0], offset); + w7[3] = hc_bytealign_be ( 0, w0[0], offset); w7[2] = 0; w7[1] = 0; w7[0] = 0; @@ -8481,153 +8327,153 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 switch (offset_switch) { case 0: - c0[0] = hc_bytealign (w7[3], 0, offset); - w7[3] = hc_bytealign (w7[2], w7[3], offset); - w7[2] = hc_bytealign (w7[1], w7[2], offset); - w7[1] = hc_bytealign (w7[0], w7[1], offset); - w7[0] = hc_bytealign (w6[3], w7[0], offset); - w6[3] = hc_bytealign (w6[2], w6[3], offset); - w6[2] = hc_bytealign (w6[1], w6[2], offset); - w6[1] = hc_bytealign (w6[0], w6[1], offset); - w6[0] = hc_bytealign (w5[3], w6[0], offset); - w5[3] = hc_bytealign (w5[2], w5[3], offset); - w5[2] = hc_bytealign (w5[1], w5[2], offset); - w5[1] = hc_bytealign (w5[0], w5[1], offset); - w5[0] = hc_bytealign (w4[3], w5[0], offset); - w4[3] = hc_bytealign (w4[2], w4[3], offset); - w4[2] = hc_bytealign (w4[1], w4[2], offset); - w4[1] = hc_bytealign (w4[0], w4[1], offset); - w4[0] = hc_bytealign (w3[3], w4[0], offset); - w3[3] = hc_bytealign (w3[2], w3[3], offset); - w3[2] = hc_bytealign (w3[1], w3[2], offset); - w3[1] = hc_bytealign (w3[0], w3[1], offset); - w3[0] = hc_bytealign (w2[3], w3[0], offset); - w2[3] = hc_bytealign (w2[2], w2[3], offset); - w2[2] = hc_bytealign (w2[1], w2[2], offset); - w2[1] = hc_bytealign (w2[0], w2[1], offset); - w2[0] = hc_bytealign (w1[3], w2[0], offset); - w1[3] = hc_bytealign (w1[2], w1[3], offset); - w1[2] = hc_bytealign (w1[1], w1[2], offset); - w1[1] = hc_bytealign (w1[0], w1[1], offset); - w1[0] = hc_bytealign (w0[3], w1[0], offset); - w0[3] = hc_bytealign (w0[2], w0[3], offset); - w0[2] = hc_bytealign (w0[1], w0[2], offset); - w0[1] = hc_bytealign (w0[0], w0[1], offset); - w0[0] = hc_bytealign ( 0, w0[0], offset); + c0[0] = hc_bytealign_be (w7[3], 0, offset); + w7[3] = hc_bytealign_be (w7[2], w7[3], offset); + w7[2] = hc_bytealign_be (w7[1], w7[2], offset); + w7[1] = hc_bytealign_be (w7[0], w7[1], offset); + w7[0] = hc_bytealign_be (w6[3], w7[0], offset); + w6[3] = hc_bytealign_be (w6[2], w6[3], offset); + w6[2] = hc_bytealign_be (w6[1], w6[2], offset); + w6[1] = hc_bytealign_be (w6[0], w6[1], offset); + w6[0] = hc_bytealign_be (w5[3], w6[0], offset); + w5[3] = hc_bytealign_be (w5[2], w5[3], offset); + w5[2] = hc_bytealign_be (w5[1], w5[2], offset); + w5[1] = hc_bytealign_be (w5[0], w5[1], offset); + w5[0] = hc_bytealign_be (w4[3], w5[0], offset); + w4[3] = hc_bytealign_be (w4[2], w4[3], offset); + w4[2] = hc_bytealign_be (w4[1], w4[2], offset); + w4[1] = hc_bytealign_be (w4[0], w4[1], offset); + w4[0] = hc_bytealign_be (w3[3], w4[0], offset); + w3[3] = hc_bytealign_be (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be ( 0, w0[0], offset); break; case 1: - c0[1] = hc_bytealign (w7[3], 0, offset); - c0[0] = hc_bytealign (w7[2], w7[3], offset); - w7[3] = hc_bytealign (w7[1], w7[2], offset); - w7[2] = hc_bytealign (w7[0], w7[1], offset); - w7[1] = hc_bytealign (w6[3], w7[0], offset); - w7[0] = hc_bytealign (w6[2], w6[3], offset); - w6[3] = hc_bytealign (w6[1], w6[2], offset); - w6[2] = hc_bytealign (w6[0], w6[1], offset); - w6[1] = hc_bytealign (w5[3], w6[0], offset); - w6[0] = hc_bytealign (w5[2], w5[3], offset); - w5[3] = hc_bytealign (w5[1], w5[2], offset); - w5[2] = hc_bytealign (w5[0], w5[1], offset); - w5[1] = hc_bytealign (w4[3], w5[0], offset); - w5[0] = hc_bytealign (w4[2], w4[3], offset); - w4[3] = hc_bytealign (w4[1], w4[2], offset); - w4[2] = hc_bytealign (w4[0], w4[1], offset); - w4[1] = hc_bytealign (w3[3], w4[0], offset); - w4[0] = hc_bytealign (w3[2], w3[3], offset); - w3[3] = hc_bytealign (w3[1], w3[2], offset); - w3[2] = hc_bytealign (w3[0], w3[1], offset); - w3[1] = hc_bytealign (w2[3], w3[0], offset); - w3[0] = hc_bytealign (w2[2], w2[3], offset); - w2[3] = hc_bytealign (w2[1], w2[2], offset); - w2[2] = hc_bytealign (w2[0], w2[1], offset); - w2[1] = hc_bytealign (w1[3], w2[0], offset); - w2[0] = hc_bytealign (w1[2], w1[3], offset); - w1[3] = hc_bytealign (w1[1], w1[2], offset); - w1[2] = hc_bytealign (w1[0], w1[1], offset); - w1[1] = hc_bytealign (w0[3], w1[0], offset); - w1[0] = hc_bytealign (w0[2], w0[3], offset); - w0[3] = hc_bytealign (w0[1], w0[2], offset); - w0[2] = hc_bytealign (w0[0], w0[1], offset); - w0[1] = hc_bytealign ( 0, w0[0], offset); + c0[1] = hc_bytealign_be (w7[3], 0, offset); + c0[0] = hc_bytealign_be (w7[2], w7[3], offset); + w7[3] = hc_bytealign_be (w7[1], w7[2], offset); + w7[2] = hc_bytealign_be (w7[0], w7[1], offset); + w7[1] = hc_bytealign_be (w6[3], w7[0], offset); + w7[0] = hc_bytealign_be (w6[2], w6[3], offset); + w6[3] = hc_bytealign_be (w6[1], w6[2], offset); + w6[2] = hc_bytealign_be (w6[0], w6[1], offset); + w6[1] = hc_bytealign_be (w5[3], w6[0], offset); + w6[0] = hc_bytealign_be (w5[2], w5[3], offset); + w5[3] = hc_bytealign_be (w5[1], w5[2], offset); + w5[2] = hc_bytealign_be (w5[0], w5[1], offset); + w5[1] = hc_bytealign_be (w4[3], w5[0], offset); + w5[0] = hc_bytealign_be (w4[2], w4[3], offset); + w4[3] = hc_bytealign_be (w4[1], w4[2], offset); + w4[2] = hc_bytealign_be (w4[0], w4[1], offset); + w4[1] = hc_bytealign_be (w3[3], w4[0], offset); + w4[0] = hc_bytealign_be (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be ( 0, w0[0], offset); w0[0] = 0; break; case 2: - c0[2] = hc_bytealign (w7[3], 0, offset); - c0[1] = hc_bytealign (w7[2], w7[3], offset); - c0[0] = hc_bytealign (w7[1], w7[2], offset); - w7[3] = hc_bytealign (w7[0], w7[1], offset); - w7[2] = hc_bytealign (w6[3], w7[0], offset); - w7[1] = hc_bytealign (w6[2], w6[3], offset); - w7[0] = hc_bytealign (w6[1], w6[2], offset); - w6[3] = hc_bytealign (w6[0], w6[1], offset); - w6[2] = hc_bytealign (w5[3], w6[0], offset); - w6[1] = hc_bytealign (w5[2], w5[3], offset); - w6[0] = hc_bytealign (w5[1], w5[2], offset); - w5[3] = hc_bytealign (w5[0], w5[1], offset); - w5[2] = hc_bytealign (w4[3], w5[0], offset); - w5[1] = hc_bytealign (w4[2], w4[3], offset); - w5[0] = hc_bytealign (w4[1], w4[2], offset); - w4[3] = hc_bytealign (w4[0], w4[1], offset); - w4[2] = hc_bytealign (w3[3], w4[0], offset); - w4[1] = hc_bytealign (w3[2], w3[3], offset); - w4[0] = hc_bytealign (w3[1], w3[2], offset); - w3[3] = hc_bytealign (w3[0], w3[1], offset); - w3[2] = hc_bytealign (w2[3], w3[0], offset); - w3[1] = hc_bytealign (w2[2], w2[3], offset); - w3[0] = hc_bytealign (w2[1], w2[2], offset); - w2[3] = hc_bytealign (w2[0], w2[1], offset); - w2[2] = hc_bytealign (w1[3], w2[0], offset); - w2[1] = hc_bytealign (w1[2], w1[3], offset); - w2[0] = hc_bytealign (w1[1], w1[2], offset); - w1[3] = hc_bytealign (w1[0], w1[1], offset); - w1[2] = hc_bytealign (w0[3], w1[0], offset); - w1[1] = hc_bytealign (w0[2], w0[3], offset); - w1[0] = hc_bytealign (w0[1], w0[2], offset); - w0[3] = hc_bytealign (w0[0], w0[1], offset); - w0[2] = hc_bytealign ( 0, w0[0], offset); + c0[2] = hc_bytealign_be (w7[3], 0, offset); + c0[1] = hc_bytealign_be (w7[2], w7[3], offset); + c0[0] = hc_bytealign_be (w7[1], w7[2], offset); + w7[3] = hc_bytealign_be (w7[0], w7[1], offset); + w7[2] = hc_bytealign_be (w6[3], w7[0], offset); + w7[1] = hc_bytealign_be (w6[2], w6[3], offset); + w7[0] = hc_bytealign_be (w6[1], w6[2], offset); + w6[3] = hc_bytealign_be (w6[0], w6[1], offset); + w6[2] = hc_bytealign_be (w5[3], w6[0], offset); + w6[1] = hc_bytealign_be (w5[2], w5[3], offset); + w6[0] = hc_bytealign_be (w5[1], w5[2], offset); + w5[3] = hc_bytealign_be (w5[0], w5[1], offset); + w5[2] = hc_bytealign_be (w4[3], w5[0], offset); + w5[1] = hc_bytealign_be (w4[2], w4[3], offset); + w5[0] = hc_bytealign_be (w4[1], w4[2], offset); + w4[3] = hc_bytealign_be (w4[0], w4[1], offset); + w4[2] = hc_bytealign_be (w3[3], w4[0], offset); + w4[1] = hc_bytealign_be (w3[2], w3[3], offset); + w4[0] = hc_bytealign_be (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - c0[3] = hc_bytealign (w7[3], 0, offset); - c0[2] = hc_bytealign (w7[2], w7[3], offset); - c0[1] = hc_bytealign (w7[1], w7[2], offset); - c0[0] = hc_bytealign (w7[0], w7[1], offset); - w7[3] = hc_bytealign (w6[3], w7[0], offset); - w7[2] = hc_bytealign (w6[2], w6[3], offset); - w7[1] = hc_bytealign (w6[1], w6[2], offset); - w7[0] = hc_bytealign (w6[0], w6[1], offset); - w6[3] = hc_bytealign (w5[3], w6[0], offset); - w6[2] = hc_bytealign (w5[2], w5[3], offset); - w6[1] = hc_bytealign (w5[1], w5[2], offset); - w6[0] = hc_bytealign (w5[0], w5[1], offset); - w5[3] = hc_bytealign (w4[3], w5[0], offset); - w5[2] = hc_bytealign (w4[2], w4[3], offset); - w5[1] = hc_bytealign (w4[1], w4[2], offset); - w5[0] = hc_bytealign (w4[0], w4[1], offset); - w4[3] = hc_bytealign (w3[3], w4[0], offset); - w4[2] = hc_bytealign (w3[2], w3[3], offset); - w4[1] = hc_bytealign (w3[1], w3[2], offset); - w4[0] = hc_bytealign (w3[0], w3[1], offset); - w3[3] = hc_bytealign (w2[3], w3[0], offset); - w3[2] = hc_bytealign (w2[2], w2[3], offset); - w3[1] = hc_bytealign (w2[1], w2[2], offset); - w3[0] = hc_bytealign (w2[0], w2[1], offset); - w2[3] = hc_bytealign (w1[3], w2[0], offset); - w2[2] = hc_bytealign (w1[2], w1[3], offset); - w2[1] = hc_bytealign (w1[1], w1[2], offset); - w2[0] = hc_bytealign (w1[0], w1[1], offset); - w1[3] = hc_bytealign (w0[3], w1[0], offset); - w1[2] = hc_bytealign (w0[2], w0[3], offset); - w1[1] = hc_bytealign (w0[1], w0[2], offset); - w1[0] = hc_bytealign (w0[0], w0[1], offset); - w0[3] = hc_bytealign ( 0, w0[0], offset); + c0[3] = hc_bytealign_be (w7[3], 0, offset); + c0[2] = hc_bytealign_be (w7[2], w7[3], offset); + c0[1] = hc_bytealign_be (w7[1], w7[2], offset); + c0[0] = hc_bytealign_be (w7[0], w7[1], offset); + w7[3] = hc_bytealign_be (w6[3], w7[0], offset); + w7[2] = hc_bytealign_be (w6[2], w6[3], offset); + w7[1] = hc_bytealign_be (w6[1], w6[2], offset); + w7[0] = hc_bytealign_be (w6[0], w6[1], offset); + w6[3] = hc_bytealign_be (w5[3], w6[0], offset); + w6[2] = hc_bytealign_be (w5[2], w5[3], offset); + w6[1] = hc_bytealign_be (w5[1], w5[2], offset); + w6[0] = hc_bytealign_be (w5[0], w5[1], offset); + w5[3] = hc_bytealign_be (w4[3], w5[0], offset); + w5[2] = hc_bytealign_be (w4[2], w4[3], offset); + w5[1] = hc_bytealign_be (w4[1], w4[2], offset); + w5[0] = hc_bytealign_be (w4[0], w4[1], offset); + w4[3] = hc_bytealign_be (w3[3], w4[0], offset); + w4[2] = hc_bytealign_be (w3[2], w3[3], offset); + w4[1] = hc_bytealign_be (w3[1], w3[2], offset); + w4[0] = hc_bytealign_be (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -8635,39 +8481,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 4: - c1[0] = hc_bytealign (w7[3], 0, offset); - c0[3] = hc_bytealign (w7[2], w7[3], offset); - c0[2] = hc_bytealign (w7[1], w7[2], offset); - c0[1] = hc_bytealign (w7[0], w7[1], offset); - c0[0] = hc_bytealign (w6[3], w7[0], offset); - w7[3] = hc_bytealign (w6[2], w6[3], offset); - w7[2] = hc_bytealign (w6[1], w6[2], offset); - w7[1] = hc_bytealign (w6[0], w6[1], offset); - w7[0] = hc_bytealign (w5[3], w6[0], offset); - w6[3] = hc_bytealign (w5[2], w5[3], offset); - w6[2] = hc_bytealign (w5[1], w5[2], offset); - w6[1] = hc_bytealign (w5[0], w5[1], offset); - w6[0] = hc_bytealign (w4[3], w5[0], offset); - w5[3] = hc_bytealign (w4[2], w4[3], offset); - w5[2] = hc_bytealign (w4[1], w4[2], offset); - w5[1] = hc_bytealign (w4[0], w4[1], offset); - w5[0] = hc_bytealign (w3[3], w4[0], offset); - w4[3] = hc_bytealign (w3[2], w3[3], offset); - w4[2] = hc_bytealign (w3[1], w3[2], offset); - w4[1] = hc_bytealign (w3[0], w3[1], offset); - w4[0] = hc_bytealign (w2[3], w3[0], offset); - w3[3] = hc_bytealign (w2[2], w2[3], offset); - w3[2] = hc_bytealign (w2[1], w2[2], offset); - w3[1] = hc_bytealign (w2[0], w2[1], offset); - w3[0] = hc_bytealign (w1[3], w2[0], offset); - w2[3] = hc_bytealign (w1[2], w1[3], offset); - w2[2] = hc_bytealign (w1[1], w1[2], offset); - w2[1] = hc_bytealign (w1[0], w1[1], offset); - w2[0] = hc_bytealign (w0[3], w1[0], offset); - w1[3] = hc_bytealign (w0[2], w0[3], offset); - w1[2] = hc_bytealign (w0[1], w0[2], offset); - w1[1] = hc_bytealign (w0[0], w0[1], offset); - w1[0] = hc_bytealign ( 0, w0[0], offset); + c1[0] = hc_bytealign_be (w7[3], 0, offset); + c0[3] = hc_bytealign_be (w7[2], w7[3], offset); + c0[2] = hc_bytealign_be (w7[1], w7[2], offset); + c0[1] = hc_bytealign_be (w7[0], w7[1], offset); + c0[0] = hc_bytealign_be (w6[3], w7[0], offset); + w7[3] = hc_bytealign_be (w6[2], w6[3], offset); + w7[2] = hc_bytealign_be (w6[1], w6[2], offset); + w7[1] = hc_bytealign_be (w6[0], w6[1], offset); + w7[0] = hc_bytealign_be (w5[3], w6[0], offset); + w6[3] = hc_bytealign_be (w5[2], w5[3], offset); + w6[2] = hc_bytealign_be (w5[1], w5[2], offset); + w6[1] = hc_bytealign_be (w5[0], w5[1], offset); + w6[0] = hc_bytealign_be (w4[3], w5[0], offset); + w5[3] = hc_bytealign_be (w4[2], w4[3], offset); + w5[2] = hc_bytealign_be (w4[1], w4[2], offset); + w5[1] = hc_bytealign_be (w4[0], w4[1], offset); + w5[0] = hc_bytealign_be (w3[3], w4[0], offset); + w4[3] = hc_bytealign_be (w3[2], w3[3], offset); + w4[2] = hc_bytealign_be (w3[1], w3[2], offset); + w4[1] = hc_bytealign_be (w3[0], w3[1], offset); + w4[0] = hc_bytealign_be (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -8676,39 +8522,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 5: - c1[1] = hc_bytealign (w7[3], 0, offset); - c1[0] = hc_bytealign (w7[2], w7[3], offset); - c0[3] = hc_bytealign (w7[1], w7[2], offset); - c0[2] = hc_bytealign (w7[0], w7[1], offset); - c0[1] = hc_bytealign (w6[3], w7[0], offset); - c0[0] = hc_bytealign (w6[2], w6[3], offset); - w7[3] = hc_bytealign (w6[1], w6[2], offset); - w7[2] = hc_bytealign (w6[0], w6[1], offset); - w7[1] = hc_bytealign (w5[3], w6[0], offset); - w7[0] = hc_bytealign (w5[2], w5[3], offset); - w6[3] = hc_bytealign (w5[1], w5[2], offset); - w6[2] = hc_bytealign (w5[0], w5[1], offset); - w6[1] = hc_bytealign (w4[3], w5[0], offset); - w6[0] = hc_bytealign (w4[2], w4[3], offset); - w5[3] = hc_bytealign (w4[1], w4[2], offset); - w5[2] = hc_bytealign (w4[0], w4[1], offset); - w5[1] = hc_bytealign (w3[3], w4[0], offset); - w5[0] = hc_bytealign (w3[2], w3[3], offset); - w4[3] = hc_bytealign (w3[1], w3[2], offset); - w4[2] = hc_bytealign (w3[0], w3[1], offset); - w4[1] = hc_bytealign (w2[3], w3[0], offset); - w4[0] = hc_bytealign (w2[2], w2[3], offset); - w3[3] = hc_bytealign (w2[1], w2[2], offset); - w3[2] = hc_bytealign (w2[0], w2[1], offset); - w3[1] = hc_bytealign (w1[3], w2[0], offset); - w3[0] = hc_bytealign (w1[2], w1[3], offset); - w2[3] = hc_bytealign (w1[1], w1[2], offset); - w2[2] = hc_bytealign (w1[0], w1[1], offset); - w2[1] = hc_bytealign (w0[3], w1[0], offset); - w2[0] = hc_bytealign (w0[2], w0[3], offset); - w1[3] = hc_bytealign (w0[1], w0[2], offset); - w1[2] = hc_bytealign (w0[0], w0[1], offset); - w1[1] = hc_bytealign ( 0, w0[0], offset); + c1[1] = hc_bytealign_be (w7[3], 0, offset); + c1[0] = hc_bytealign_be (w7[2], w7[3], offset); + c0[3] = hc_bytealign_be (w7[1], w7[2], offset); + c0[2] = hc_bytealign_be (w7[0], w7[1], offset); + c0[1] = hc_bytealign_be (w6[3], w7[0], offset); + c0[0] = hc_bytealign_be (w6[2], w6[3], offset); + w7[3] = hc_bytealign_be (w6[1], w6[2], offset); + w7[2] = hc_bytealign_be (w6[0], w6[1], offset); + w7[1] = hc_bytealign_be (w5[3], w6[0], offset); + w7[0] = hc_bytealign_be (w5[2], w5[3], offset); + w6[3] = hc_bytealign_be (w5[1], w5[2], offset); + w6[2] = hc_bytealign_be (w5[0], w5[1], offset); + w6[1] = hc_bytealign_be (w4[3], w5[0], offset); + w6[0] = hc_bytealign_be (w4[2], w4[3], offset); + w5[3] = hc_bytealign_be (w4[1], w4[2], offset); + w5[2] = hc_bytealign_be (w4[0], w4[1], offset); + w5[1] = hc_bytealign_be (w3[3], w4[0], offset); + w5[0] = hc_bytealign_be (w3[2], w3[3], offset); + w4[3] = hc_bytealign_be (w3[1], w3[2], offset); + w4[2] = hc_bytealign_be (w3[0], w3[1], offset); + w4[1] = hc_bytealign_be (w2[3], w3[0], offset); + w4[0] = hc_bytealign_be (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -8718,39 +8564,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 6: - c1[2] = hc_bytealign (w7[3], 0, offset); - c1[1] = hc_bytealign (w7[2], w7[3], offset); - c1[0] = hc_bytealign (w7[1], w7[2], offset); - c0[3] = hc_bytealign (w7[0], w7[1], offset); - c0[2] = hc_bytealign (w6[3], w7[0], offset); - c0[1] = hc_bytealign (w6[2], w6[3], offset); - c0[0] = hc_bytealign (w6[1], w6[2], offset); - w7[3] = hc_bytealign (w6[0], w6[1], offset); - w7[2] = hc_bytealign (w5[3], w6[0], offset); - w7[1] = hc_bytealign (w5[2], w5[3], offset); - w7[0] = hc_bytealign (w5[1], w5[2], offset); - w6[3] = hc_bytealign (w5[0], w5[1], offset); - w6[2] = hc_bytealign (w4[3], w5[0], offset); - w6[1] = hc_bytealign (w4[2], w4[3], offset); - w6[0] = hc_bytealign (w4[1], w4[2], offset); - w5[3] = hc_bytealign (w4[0], w4[1], offset); - w5[2] = hc_bytealign (w3[3], w4[0], offset); - w5[1] = hc_bytealign (w3[2], w3[3], offset); - w5[0] = hc_bytealign (w3[1], w3[2], offset); - w4[3] = hc_bytealign (w3[0], w3[1], offset); - w4[2] = hc_bytealign (w2[3], w3[0], offset); - w4[1] = hc_bytealign (w2[2], w2[3], offset); - w4[0] = hc_bytealign (w2[1], w2[2], offset); - w3[3] = hc_bytealign (w2[0], w2[1], offset); - w3[2] = hc_bytealign (w1[3], w2[0], offset); - w3[1] = hc_bytealign (w1[2], w1[3], offset); - w3[0] = hc_bytealign (w1[1], w1[2], offset); - w2[3] = hc_bytealign (w1[0], w1[1], offset); - w2[2] = hc_bytealign (w0[3], w1[0], offset); - w2[1] = hc_bytealign (w0[2], w0[3], offset); - w2[0] = hc_bytealign (w0[1], w0[2], offset); - w1[3] = hc_bytealign (w0[0], w0[1], offset); - w1[2] = hc_bytealign ( 0, w0[0], offset); + c1[2] = hc_bytealign_be (w7[3], 0, offset); + c1[1] = hc_bytealign_be (w7[2], w7[3], offset); + c1[0] = hc_bytealign_be (w7[1], w7[2], offset); + c0[3] = hc_bytealign_be (w7[0], w7[1], offset); + c0[2] = hc_bytealign_be (w6[3], w7[0], offset); + c0[1] = hc_bytealign_be (w6[2], w6[3], offset); + c0[0] = hc_bytealign_be (w6[1], w6[2], offset); + w7[3] = hc_bytealign_be (w6[0], w6[1], offset); + w7[2] = hc_bytealign_be (w5[3], w6[0], offset); + w7[1] = hc_bytealign_be (w5[2], w5[3], offset); + w7[0] = hc_bytealign_be (w5[1], w5[2], offset); + w6[3] = hc_bytealign_be (w5[0], w5[1], offset); + w6[2] = hc_bytealign_be (w4[3], w5[0], offset); + w6[1] = hc_bytealign_be (w4[2], w4[3], offset); + w6[0] = hc_bytealign_be (w4[1], w4[2], offset); + w5[3] = hc_bytealign_be (w4[0], w4[1], offset); + w5[2] = hc_bytealign_be (w3[3], w4[0], offset); + w5[1] = hc_bytealign_be (w3[2], w3[3], offset); + w5[0] = hc_bytealign_be (w3[1], w3[2], offset); + w4[3] = hc_bytealign_be (w3[0], w3[1], offset); + w4[2] = hc_bytealign_be (w2[3], w3[0], offset); + w4[1] = hc_bytealign_be (w2[2], w2[3], offset); + w4[0] = hc_bytealign_be (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -8761,39 +8607,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 7: - c1[3] = hc_bytealign (w7[3], 0, offset); - c1[2] = hc_bytealign (w7[2], w7[3], offset); - c1[1] = hc_bytealign (w7[1], w7[2], offset); - c1[0] = hc_bytealign (w7[0], w7[1], offset); - c0[3] = hc_bytealign (w6[3], w7[0], offset); - c0[2] = hc_bytealign (w6[2], w6[3], offset); - c0[1] = hc_bytealign (w6[1], w6[2], offset); - c0[0] = hc_bytealign (w6[0], w6[1], offset); - w7[3] = hc_bytealign (w5[3], w6[0], offset); - w7[2] = hc_bytealign (w5[2], w5[3], offset); - w7[1] = hc_bytealign (w5[1], w5[2], offset); - w7[0] = hc_bytealign (w5[0], w5[1], offset); - w6[3] = hc_bytealign (w4[3], w5[0], offset); - w6[2] = hc_bytealign (w4[2], w4[3], offset); - w6[1] = hc_bytealign (w4[1], w4[2], offset); - w6[0] = hc_bytealign (w4[0], w4[1], offset); - w5[3] = hc_bytealign (w3[3], w4[0], offset); - w5[2] = hc_bytealign (w3[2], w3[3], offset); - w5[1] = hc_bytealign (w3[1], w3[2], offset); - w5[0] = hc_bytealign (w3[0], w3[1], offset); - w4[3] = hc_bytealign (w2[3], w3[0], offset); - w4[2] = hc_bytealign (w2[2], w2[3], offset); - w4[1] = hc_bytealign (w2[1], w2[2], offset); - w4[0] = hc_bytealign (w2[0], w2[1], offset); - w3[3] = hc_bytealign (w1[3], w2[0], offset); - w3[2] = hc_bytealign (w1[2], w1[3], offset); - w3[1] = hc_bytealign (w1[1], w1[2], offset); - w3[0] = hc_bytealign (w1[0], w1[1], offset); - w2[3] = hc_bytealign (w0[3], w1[0], offset); - w2[2] = hc_bytealign (w0[2], w0[3], offset); - w2[1] = hc_bytealign (w0[1], w0[2], offset); - w2[0] = hc_bytealign (w0[0], w0[1], offset); - w1[3] = hc_bytealign ( 0, w0[0], offset); + c1[3] = hc_bytealign_be (w7[3], 0, offset); + c1[2] = hc_bytealign_be (w7[2], w7[3], offset); + c1[1] = hc_bytealign_be (w7[1], w7[2], offset); + c1[0] = hc_bytealign_be (w7[0], w7[1], offset); + c0[3] = hc_bytealign_be (w6[3], w7[0], offset); + c0[2] = hc_bytealign_be (w6[2], w6[3], offset); + c0[1] = hc_bytealign_be (w6[1], w6[2], offset); + c0[0] = hc_bytealign_be (w6[0], w6[1], offset); + w7[3] = hc_bytealign_be (w5[3], w6[0], offset); + w7[2] = hc_bytealign_be (w5[2], w5[3], offset); + w7[1] = hc_bytealign_be (w5[1], w5[2], offset); + w7[0] = hc_bytealign_be (w5[0], w5[1], offset); + w6[3] = hc_bytealign_be (w4[3], w5[0], offset); + w6[2] = hc_bytealign_be (w4[2], w4[3], offset); + w6[1] = hc_bytealign_be (w4[1], w4[2], offset); + w6[0] = hc_bytealign_be (w4[0], w4[1], offset); + w5[3] = hc_bytealign_be (w3[3], w4[0], offset); + w5[2] = hc_bytealign_be (w3[2], w3[3], offset); + w5[1] = hc_bytealign_be (w3[1], w3[2], offset); + w5[0] = hc_bytealign_be (w3[0], w3[1], offset); + w4[3] = hc_bytealign_be (w2[3], w3[0], offset); + w4[2] = hc_bytealign_be (w2[2], w2[3], offset); + w4[1] = hc_bytealign_be (w2[1], w2[2], offset); + w4[0] = hc_bytealign_be (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -8805,39 +8651,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 8: - c2[0] = hc_bytealign (w7[3], 0, offset); - c1[3] = hc_bytealign (w7[2], w7[3], offset); - c1[2] = hc_bytealign (w7[1], w7[2], offset); - c1[1] = hc_bytealign (w7[0], w7[1], offset); - c1[0] = hc_bytealign (w6[3], w7[0], offset); - c0[3] = hc_bytealign (w6[2], w6[3], offset); - c0[2] = hc_bytealign (w6[1], w6[2], offset); - c0[1] = hc_bytealign (w6[0], w6[1], offset); - c0[0] = hc_bytealign (w5[3], w6[0], offset); - w7[3] = hc_bytealign (w5[2], w5[3], offset); - w7[2] = hc_bytealign (w5[1], w5[2], offset); - w7[1] = hc_bytealign (w5[0], w5[1], offset); - w7[0] = hc_bytealign (w4[3], w5[0], offset); - w6[3] = hc_bytealign (w4[2], w4[3], offset); - w6[2] = hc_bytealign (w4[1], w4[2], offset); - w6[1] = hc_bytealign (w4[0], w4[1], offset); - w6[0] = hc_bytealign (w3[3], w4[0], offset); - w5[3] = hc_bytealign (w3[2], w3[3], offset); - w5[2] = hc_bytealign (w3[1], w3[2], offset); - w5[1] = hc_bytealign (w3[0], w3[1], offset); - w5[0] = hc_bytealign (w2[3], w3[0], offset); - w4[3] = hc_bytealign (w2[2], w2[3], offset); - w4[2] = hc_bytealign (w2[1], w2[2], offset); - w4[1] = hc_bytealign (w2[0], w2[1], offset); - w4[0] = hc_bytealign (w1[3], w2[0], offset); - w3[3] = hc_bytealign (w1[2], w1[3], offset); - w3[2] = hc_bytealign (w1[1], w1[2], offset); - w3[1] = hc_bytealign (w1[0], w1[1], offset); - w3[0] = hc_bytealign (w0[3], w1[0], offset); - w2[3] = hc_bytealign (w0[2], w0[3], offset); - w2[2] = hc_bytealign (w0[1], w0[2], offset); - w2[1] = hc_bytealign (w0[0], w0[1], offset); - w2[0] = hc_bytealign ( 0, w0[0], offset); + c2[0] = hc_bytealign_be (w7[3], 0, offset); + c1[3] = hc_bytealign_be (w7[2], w7[3], offset); + c1[2] = hc_bytealign_be (w7[1], w7[2], offset); + c1[1] = hc_bytealign_be (w7[0], w7[1], offset); + c1[0] = hc_bytealign_be (w6[3], w7[0], offset); + c0[3] = hc_bytealign_be (w6[2], w6[3], offset); + c0[2] = hc_bytealign_be (w6[1], w6[2], offset); + c0[1] = hc_bytealign_be (w6[0], w6[1], offset); + c0[0] = hc_bytealign_be (w5[3], w6[0], offset); + w7[3] = hc_bytealign_be (w5[2], w5[3], offset); + w7[2] = hc_bytealign_be (w5[1], w5[2], offset); + w7[1] = hc_bytealign_be (w5[0], w5[1], offset); + w7[0] = hc_bytealign_be (w4[3], w5[0], offset); + w6[3] = hc_bytealign_be (w4[2], w4[3], offset); + w6[2] = hc_bytealign_be (w4[1], w4[2], offset); + w6[1] = hc_bytealign_be (w4[0], w4[1], offset); + w6[0] = hc_bytealign_be (w3[3], w4[0], offset); + w5[3] = hc_bytealign_be (w3[2], w3[3], offset); + w5[2] = hc_bytealign_be (w3[1], w3[2], offset); + w5[1] = hc_bytealign_be (w3[0], w3[1], offset); + w5[0] = hc_bytealign_be (w2[3], w3[0], offset); + w4[3] = hc_bytealign_be (w2[2], w2[3], offset); + w4[2] = hc_bytealign_be (w2[1], w2[2], offset); + w4[1] = hc_bytealign_be (w2[0], w2[1], offset); + w4[0] = hc_bytealign_be (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -8850,39 +8696,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 9: - c2[1] = hc_bytealign (w7[3], 0, offset); - c2[0] = hc_bytealign (w7[2], w7[3], offset); - c1[3] = hc_bytealign (w7[1], w7[2], offset); - c1[2] = hc_bytealign (w7[0], w7[1], offset); - c1[1] = hc_bytealign (w6[3], w7[0], offset); - c1[0] = hc_bytealign (w6[2], w6[3], offset); - c0[3] = hc_bytealign (w6[1], w6[2], offset); - c0[2] = hc_bytealign (w6[0], w6[1], offset); - c0[1] = hc_bytealign (w5[3], w6[0], offset); - c0[0] = hc_bytealign (w5[2], w5[3], offset); - w7[3] = hc_bytealign (w5[1], w5[2], offset); - w7[2] = hc_bytealign (w5[0], w5[1], offset); - w7[1] = hc_bytealign (w4[3], w5[0], offset); - w7[0] = hc_bytealign (w4[2], w4[3], offset); - w6[3] = hc_bytealign (w4[1], w4[2], offset); - w6[2] = hc_bytealign (w4[0], w4[1], offset); - w6[1] = hc_bytealign (w3[3], w4[0], offset); - w6[0] = hc_bytealign (w3[2], w3[3], offset); - w5[3] = hc_bytealign (w3[1], w3[2], offset); - w5[2] = hc_bytealign (w3[0], w3[1], offset); - w5[1] = hc_bytealign (w2[3], w3[0], offset); - w5[0] = hc_bytealign (w2[2], w2[3], offset); - w4[3] = hc_bytealign (w2[1], w2[2], offset); - w4[2] = hc_bytealign (w2[0], w2[1], offset); - w4[1] = hc_bytealign (w1[3], w2[0], offset); - w4[0] = hc_bytealign (w1[2], w1[3], offset); - w3[3] = hc_bytealign (w1[1], w1[2], offset); - w3[2] = hc_bytealign (w1[0], w1[1], offset); - w3[1] = hc_bytealign (w0[3], w1[0], offset); - w3[0] = hc_bytealign (w0[2], w0[3], offset); - w2[3] = hc_bytealign (w0[1], w0[2], offset); - w2[2] = hc_bytealign (w0[0], w0[1], offset); - w2[1] = hc_bytealign ( 0, w0[0], offset); + c2[1] = hc_bytealign_be (w7[3], 0, offset); + c2[0] = hc_bytealign_be (w7[2], w7[3], offset); + c1[3] = hc_bytealign_be (w7[1], w7[2], offset); + c1[2] = hc_bytealign_be (w7[0], w7[1], offset); + c1[1] = hc_bytealign_be (w6[3], w7[0], offset); + c1[0] = hc_bytealign_be (w6[2], w6[3], offset); + c0[3] = hc_bytealign_be (w6[1], w6[2], offset); + c0[2] = hc_bytealign_be (w6[0], w6[1], offset); + c0[1] = hc_bytealign_be (w5[3], w6[0], offset); + c0[0] = hc_bytealign_be (w5[2], w5[3], offset); + w7[3] = hc_bytealign_be (w5[1], w5[2], offset); + w7[2] = hc_bytealign_be (w5[0], w5[1], offset); + w7[1] = hc_bytealign_be (w4[3], w5[0], offset); + w7[0] = hc_bytealign_be (w4[2], w4[3], offset); + w6[3] = hc_bytealign_be (w4[1], w4[2], offset); + w6[2] = hc_bytealign_be (w4[0], w4[1], offset); + w6[1] = hc_bytealign_be (w3[3], w4[0], offset); + w6[0] = hc_bytealign_be (w3[2], w3[3], offset); + w5[3] = hc_bytealign_be (w3[1], w3[2], offset); + w5[2] = hc_bytealign_be (w3[0], w3[1], offset); + w5[1] = hc_bytealign_be (w2[3], w3[0], offset); + w5[0] = hc_bytealign_be (w2[2], w2[3], offset); + w4[3] = hc_bytealign_be (w2[1], w2[2], offset); + w4[2] = hc_bytealign_be (w2[0], w2[1], offset); + w4[1] = hc_bytealign_be (w1[3], w2[0], offset); + w4[0] = hc_bytealign_be (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -8896,39 +8742,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 10: - c2[2] = hc_bytealign (w7[3], 0, offset); - c2[1] = hc_bytealign (w7[2], w7[3], offset); - c2[0] = hc_bytealign (w7[1], w7[2], offset); - c1[3] = hc_bytealign (w7[0], w7[1], offset); - c1[2] = hc_bytealign (w6[3], w7[0], offset); - c1[1] = hc_bytealign (w6[2], w6[3], offset); - c1[0] = hc_bytealign (w6[1], w6[2], offset); - c0[3] = hc_bytealign (w6[0], w6[1], offset); - c0[2] = hc_bytealign (w5[3], w6[0], offset); - c0[1] = hc_bytealign (w5[2], w5[3], offset); - c0[0] = hc_bytealign (w5[1], w5[2], offset); - w7[3] = hc_bytealign (w5[0], w5[1], offset); - w7[2] = hc_bytealign (w4[3], w5[0], offset); - w7[1] = hc_bytealign (w4[2], w4[3], offset); - w7[0] = hc_bytealign (w4[1], w4[2], offset); - w6[3] = hc_bytealign (w4[0], w4[1], offset); - w6[2] = hc_bytealign (w3[3], w4[0], offset); - w6[1] = hc_bytealign (w3[2], w3[3], offset); - w6[0] = hc_bytealign (w3[1], w3[2], offset); - w5[3] = hc_bytealign (w3[0], w3[1], offset); - w5[2] = hc_bytealign (w2[3], w3[0], offset); - w5[1] = hc_bytealign (w2[2], w2[3], offset); - w5[0] = hc_bytealign (w2[1], w2[2], offset); - w4[3] = hc_bytealign (w2[0], w2[1], offset); - w4[2] = hc_bytealign (w1[3], w2[0], offset); - w4[1] = hc_bytealign (w1[2], w1[3], offset); - w4[0] = hc_bytealign (w1[1], w1[2], offset); - w3[3] = hc_bytealign (w1[0], w1[1], offset); - w3[2] = hc_bytealign (w0[3], w1[0], offset); - w3[1] = hc_bytealign (w0[2], w0[3], offset); - w3[0] = hc_bytealign (w0[1], w0[2], offset); - w2[3] = hc_bytealign (w0[0], w0[1], offset); - w2[2] = hc_bytealign ( 0, w0[0], offset); + c2[2] = hc_bytealign_be (w7[3], 0, offset); + c2[1] = hc_bytealign_be (w7[2], w7[3], offset); + c2[0] = hc_bytealign_be (w7[1], w7[2], offset); + c1[3] = hc_bytealign_be (w7[0], w7[1], offset); + c1[2] = hc_bytealign_be (w6[3], w7[0], offset); + c1[1] = hc_bytealign_be (w6[2], w6[3], offset); + c1[0] = hc_bytealign_be (w6[1], w6[2], offset); + c0[3] = hc_bytealign_be (w6[0], w6[1], offset); + c0[2] = hc_bytealign_be (w5[3], w6[0], offset); + c0[1] = hc_bytealign_be (w5[2], w5[3], offset); + c0[0] = hc_bytealign_be (w5[1], w5[2], offset); + w7[3] = hc_bytealign_be (w5[0], w5[1], offset); + w7[2] = hc_bytealign_be (w4[3], w5[0], offset); + w7[1] = hc_bytealign_be (w4[2], w4[3], offset); + w7[0] = hc_bytealign_be (w4[1], w4[2], offset); + w6[3] = hc_bytealign_be (w4[0], w4[1], offset); + w6[2] = hc_bytealign_be (w3[3], w4[0], offset); + w6[1] = hc_bytealign_be (w3[2], w3[3], offset); + w6[0] = hc_bytealign_be (w3[1], w3[2], offset); + w5[3] = hc_bytealign_be (w3[0], w3[1], offset); + w5[2] = hc_bytealign_be (w2[3], w3[0], offset); + w5[1] = hc_bytealign_be (w2[2], w2[3], offset); + w5[0] = hc_bytealign_be (w2[1], w2[2], offset); + w4[3] = hc_bytealign_be (w2[0], w2[1], offset); + w4[2] = hc_bytealign_be (w1[3], w2[0], offset); + w4[1] = hc_bytealign_be (w1[2], w1[3], offset); + w4[0] = hc_bytealign_be (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -8943,39 +8789,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 11: - c2[3] = hc_bytealign (w7[3], 0, offset); - c2[2] = hc_bytealign (w7[2], w7[3], offset); - c2[1] = hc_bytealign (w7[1], w7[2], offset); - c2[0] = hc_bytealign (w7[0], w7[1], offset); - c1[3] = hc_bytealign (w6[3], w7[0], offset); - c1[2] = hc_bytealign (w6[2], w6[3], offset); - c1[1] = hc_bytealign (w6[1], w6[2], offset); - c1[0] = hc_bytealign (w6[0], w6[1], offset); - c0[3] = hc_bytealign (w5[3], w6[0], offset); - c0[2] = hc_bytealign (w5[2], w5[3], offset); - c0[1] = hc_bytealign (w5[1], w5[2], offset); - c0[0] = hc_bytealign (w5[0], w5[1], offset); - w7[3] = hc_bytealign (w4[3], w5[0], offset); - w7[2] = hc_bytealign (w4[2], w4[3], offset); - w7[1] = hc_bytealign (w4[1], w4[2], offset); - w7[0] = hc_bytealign (w4[0], w4[1], offset); - w6[3] = hc_bytealign (w3[3], w4[0], offset); - w6[2] = hc_bytealign (w3[2], w3[3], offset); - w6[1] = hc_bytealign (w3[1], w3[2], offset); - w6[0] = hc_bytealign (w3[0], w3[1], offset); - w5[3] = hc_bytealign (w2[3], w3[0], offset); - w5[2] = hc_bytealign (w2[2], w2[3], offset); - w5[1] = hc_bytealign (w2[1], w2[2], offset); - w5[0] = hc_bytealign (w2[0], w2[1], offset); - w4[3] = hc_bytealign (w1[3], w2[0], offset); - w4[2] = hc_bytealign (w1[2], w1[3], offset); - w4[1] = hc_bytealign (w1[1], w1[2], offset); - w4[0] = hc_bytealign (w1[0], w1[1], offset); - w3[3] = hc_bytealign (w0[3], w1[0], offset); - w3[2] = hc_bytealign (w0[2], w0[3], offset); - w3[1] = hc_bytealign (w0[1], w0[2], offset); - w3[0] = hc_bytealign (w0[0], w0[1], offset); - w2[3] = hc_bytealign ( 0, w0[0], offset); + c2[3] = hc_bytealign_be (w7[3], 0, offset); + c2[2] = hc_bytealign_be (w7[2], w7[3], offset); + c2[1] = hc_bytealign_be (w7[1], w7[2], offset); + c2[0] = hc_bytealign_be (w7[0], w7[1], offset); + c1[3] = hc_bytealign_be (w6[3], w7[0], offset); + c1[2] = hc_bytealign_be (w6[2], w6[3], offset); + c1[1] = hc_bytealign_be (w6[1], w6[2], offset); + c1[0] = hc_bytealign_be (w6[0], w6[1], offset); + c0[3] = hc_bytealign_be (w5[3], w6[0], offset); + c0[2] = hc_bytealign_be (w5[2], w5[3], offset); + c0[1] = hc_bytealign_be (w5[1], w5[2], offset); + c0[0] = hc_bytealign_be (w5[0], w5[1], offset); + w7[3] = hc_bytealign_be (w4[3], w5[0], offset); + w7[2] = hc_bytealign_be (w4[2], w4[3], offset); + w7[1] = hc_bytealign_be (w4[1], w4[2], offset); + w7[0] = hc_bytealign_be (w4[0], w4[1], offset); + w6[3] = hc_bytealign_be (w3[3], w4[0], offset); + w6[2] = hc_bytealign_be (w3[2], w3[3], offset); + w6[1] = hc_bytealign_be (w3[1], w3[2], offset); + w6[0] = hc_bytealign_be (w3[0], w3[1], offset); + w5[3] = hc_bytealign_be (w2[3], w3[0], offset); + w5[2] = hc_bytealign_be (w2[2], w2[3], offset); + w5[1] = hc_bytealign_be (w2[1], w2[2], offset); + w5[0] = hc_bytealign_be (w2[0], w2[1], offset); + w4[3] = hc_bytealign_be (w1[3], w2[0], offset); + w4[2] = hc_bytealign_be (w1[2], w1[3], offset); + w4[1] = hc_bytealign_be (w1[1], w1[2], offset); + w4[0] = hc_bytealign_be (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -8991,39 +8837,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 12: - c3[0] = hc_bytealign (w7[3], 0, offset); - c2[3] = hc_bytealign (w7[2], w7[3], offset); - c2[2] = hc_bytealign (w7[1], w7[2], offset); - c2[1] = hc_bytealign (w7[0], w7[1], offset); - c2[0] = hc_bytealign (w6[3], w7[0], offset); - c1[3] = hc_bytealign (w6[2], w6[3], offset); - c1[2] = hc_bytealign (w6[1], w6[2], offset); - c1[1] = hc_bytealign (w6[0], w6[1], offset); - c1[0] = hc_bytealign (w5[3], w6[0], offset); - c0[3] = hc_bytealign (w5[2], w5[3], offset); - c0[2] = hc_bytealign (w5[1], w5[2], offset); - c0[1] = hc_bytealign (w5[0], w5[1], offset); - c0[0] = hc_bytealign (w4[3], w5[0], offset); - w7[3] = hc_bytealign (w4[2], w4[3], offset); - w7[2] = hc_bytealign (w4[1], w4[2], offset); - w7[1] = hc_bytealign (w4[0], w4[1], offset); - w7[0] = hc_bytealign (w3[3], w4[0], offset); - w6[3] = hc_bytealign (w3[2], w3[3], offset); - w6[2] = hc_bytealign (w3[1], w3[2], offset); - w6[1] = hc_bytealign (w3[0], w3[1], offset); - w6[0] = hc_bytealign (w2[3], w3[0], offset); - w5[3] = hc_bytealign (w2[2], w2[3], offset); - w5[2] = hc_bytealign (w2[1], w2[2], offset); - w5[1] = hc_bytealign (w2[0], w2[1], offset); - w5[0] = hc_bytealign (w1[3], w2[0], offset); - w4[3] = hc_bytealign (w1[2], w1[3], offset); - w4[2] = hc_bytealign (w1[1], w1[2], offset); - w4[1] = hc_bytealign (w1[0], w1[1], offset); - w4[0] = hc_bytealign (w0[3], w1[0], offset); - w3[3] = hc_bytealign (w0[2], w0[3], offset); - w3[2] = hc_bytealign (w0[1], w0[2], offset); - w3[1] = hc_bytealign (w0[0], w0[1], offset); - w3[0] = hc_bytealign ( 0, w0[0], offset); + c3[0] = hc_bytealign_be (w7[3], 0, offset); + c2[3] = hc_bytealign_be (w7[2], w7[3], offset); + c2[2] = hc_bytealign_be (w7[1], w7[2], offset); + c2[1] = hc_bytealign_be (w7[0], w7[1], offset); + c2[0] = hc_bytealign_be (w6[3], w7[0], offset); + c1[3] = hc_bytealign_be (w6[2], w6[3], offset); + c1[2] = hc_bytealign_be (w6[1], w6[2], offset); + c1[1] = hc_bytealign_be (w6[0], w6[1], offset); + c1[0] = hc_bytealign_be (w5[3], w6[0], offset); + c0[3] = hc_bytealign_be (w5[2], w5[3], offset); + c0[2] = hc_bytealign_be (w5[1], w5[2], offset); + c0[1] = hc_bytealign_be (w5[0], w5[1], offset); + c0[0] = hc_bytealign_be (w4[3], w5[0], offset); + w7[3] = hc_bytealign_be (w4[2], w4[3], offset); + w7[2] = hc_bytealign_be (w4[1], w4[2], offset); + w7[1] = hc_bytealign_be (w4[0], w4[1], offset); + w7[0] = hc_bytealign_be (w3[3], w4[0], offset); + w6[3] = hc_bytealign_be (w3[2], w3[3], offset); + w6[2] = hc_bytealign_be (w3[1], w3[2], offset); + w6[1] = hc_bytealign_be (w3[0], w3[1], offset); + w6[0] = hc_bytealign_be (w2[3], w3[0], offset); + w5[3] = hc_bytealign_be (w2[2], w2[3], offset); + w5[2] = hc_bytealign_be (w2[1], w2[2], offset); + w5[1] = hc_bytealign_be (w2[0], w2[1], offset); + w5[0] = hc_bytealign_be (w1[3], w2[0], offset); + w4[3] = hc_bytealign_be (w1[2], w1[3], offset); + w4[2] = hc_bytealign_be (w1[1], w1[2], offset); + w4[1] = hc_bytealign_be (w1[0], w1[1], offset); + w4[0] = hc_bytealign_be (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -9040,39 +8886,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 13: - c3[1] = hc_bytealign (w7[3], 0, offset); - c3[0] = hc_bytealign (w7[2], w7[3], offset); - c2[3] = hc_bytealign (w7[1], w7[2], offset); - c2[2] = hc_bytealign (w7[0], w7[1], offset); - c2[1] = hc_bytealign (w6[3], w7[0], offset); - c2[0] = hc_bytealign (w6[2], w6[3], offset); - c1[3] = hc_bytealign (w6[1], w6[2], offset); - c1[2] = hc_bytealign (w6[0], w6[1], offset); - c1[1] = hc_bytealign (w5[3], w6[0], offset); - c1[0] = hc_bytealign (w5[2], w5[3], offset); - c0[3] = hc_bytealign (w5[1], w5[2], offset); - c0[2] = hc_bytealign (w5[0], w5[1], offset); - c0[1] = hc_bytealign (w4[3], w5[0], offset); - c0[0] = hc_bytealign (w4[2], w4[3], offset); - w7[3] = hc_bytealign (w4[1], w4[2], offset); - w7[2] = hc_bytealign (w4[0], w4[1], offset); - w7[1] = hc_bytealign (w3[3], w4[0], offset); - w7[0] = hc_bytealign (w3[2], w3[3], offset); - w6[3] = hc_bytealign (w3[1], w3[2], offset); - w6[2] = hc_bytealign (w3[0], w3[1], offset); - w6[1] = hc_bytealign (w2[3], w3[0], offset); - w6[0] = hc_bytealign (w2[2], w2[3], offset); - w5[3] = hc_bytealign (w2[1], w2[2], offset); - w5[2] = hc_bytealign (w2[0], w2[1], offset); - w5[1] = hc_bytealign (w1[3], w2[0], offset); - w5[0] = hc_bytealign (w1[2], w1[3], offset); - w4[3] = hc_bytealign (w1[1], w1[2], offset); - w4[2] = hc_bytealign (w1[0], w1[1], offset); - w4[1] = hc_bytealign (w0[3], w1[0], offset); - w4[0] = hc_bytealign (w0[2], w0[3], offset); - w3[3] = hc_bytealign (w0[1], w0[2], offset); - w3[2] = hc_bytealign (w0[0], w0[1], offset); - w3[1] = hc_bytealign ( 0, w0[0], offset); + c3[1] = hc_bytealign_be (w7[3], 0, offset); + c3[0] = hc_bytealign_be (w7[2], w7[3], offset); + c2[3] = hc_bytealign_be (w7[1], w7[2], offset); + c2[2] = hc_bytealign_be (w7[0], w7[1], offset); + c2[1] = hc_bytealign_be (w6[3], w7[0], offset); + c2[0] = hc_bytealign_be (w6[2], w6[3], offset); + c1[3] = hc_bytealign_be (w6[1], w6[2], offset); + c1[2] = hc_bytealign_be (w6[0], w6[1], offset); + c1[1] = hc_bytealign_be (w5[3], w6[0], offset); + c1[0] = hc_bytealign_be (w5[2], w5[3], offset); + c0[3] = hc_bytealign_be (w5[1], w5[2], offset); + c0[2] = hc_bytealign_be (w5[0], w5[1], offset); + c0[1] = hc_bytealign_be (w4[3], w5[0], offset); + c0[0] = hc_bytealign_be (w4[2], w4[3], offset); + w7[3] = hc_bytealign_be (w4[1], w4[2], offset); + w7[2] = hc_bytealign_be (w4[0], w4[1], offset); + w7[1] = hc_bytealign_be (w3[3], w4[0], offset); + w7[0] = hc_bytealign_be (w3[2], w3[3], offset); + w6[3] = hc_bytealign_be (w3[1], w3[2], offset); + w6[2] = hc_bytealign_be (w3[0], w3[1], offset); + w6[1] = hc_bytealign_be (w2[3], w3[0], offset); + w6[0] = hc_bytealign_be (w2[2], w2[3], offset); + w5[3] = hc_bytealign_be (w2[1], w2[2], offset); + w5[2] = hc_bytealign_be (w2[0], w2[1], offset); + w5[1] = hc_bytealign_be (w1[3], w2[0], offset); + w5[0] = hc_bytealign_be (w1[2], w1[3], offset); + w4[3] = hc_bytealign_be (w1[1], w1[2], offset); + w4[2] = hc_bytealign_be (w1[0], w1[1], offset); + w4[1] = hc_bytealign_be (w0[3], w1[0], offset); + w4[0] = hc_bytealign_be (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -9090,39 +8936,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 14: - c3[2] = hc_bytealign (w7[3], 0, offset); - c3[1] = hc_bytealign (w7[2], w7[3], offset); - c3[0] = hc_bytealign (w7[1], w7[2], offset); - c2[3] = hc_bytealign (w7[0], w7[1], offset); - c2[2] = hc_bytealign (w6[3], w7[0], offset); - c2[1] = hc_bytealign (w6[2], w6[3], offset); - c2[0] = hc_bytealign (w6[1], w6[2], offset); - c1[3] = hc_bytealign (w6[0], w6[1], offset); - c1[2] = hc_bytealign (w5[3], w6[0], offset); - c1[1] = hc_bytealign (w5[2], w5[3], offset); - c1[0] = hc_bytealign (w5[1], w5[2], offset); - c0[3] = hc_bytealign (w5[0], w5[1], offset); - c0[2] = hc_bytealign (w4[3], w5[0], offset); - c0[1] = hc_bytealign (w4[2], w4[3], offset); - c0[0] = hc_bytealign (w4[1], w4[2], offset); - w7[3] = hc_bytealign (w4[0], w4[1], offset); - w7[2] = hc_bytealign (w3[3], w4[0], offset); - w7[1] = hc_bytealign (w3[2], w3[3], offset); - w7[0] = hc_bytealign (w3[1], w3[2], offset); - w6[3] = hc_bytealign (w3[0], w3[1], offset); - w6[2] = hc_bytealign (w2[3], w3[0], offset); - w6[1] = hc_bytealign (w2[2], w2[3], offset); - w6[0] = hc_bytealign (w2[1], w2[2], offset); - w5[3] = hc_bytealign (w2[0], w2[1], offset); - w5[2] = hc_bytealign (w1[3], w2[0], offset); - w5[1] = hc_bytealign (w1[2], w1[3], offset); - w5[0] = hc_bytealign (w1[1], w1[2], offset); - w4[3] = hc_bytealign (w1[0], w1[1], offset); - w4[2] = hc_bytealign (w0[3], w1[0], offset); - w4[1] = hc_bytealign (w0[2], w0[3], offset); - w4[0] = hc_bytealign (w0[1], w0[2], offset); - w3[3] = hc_bytealign (w0[0], w0[1], offset); - w3[2] = hc_bytealign ( 0, w0[0], offset); + c3[2] = hc_bytealign_be (w7[3], 0, offset); + c3[1] = hc_bytealign_be (w7[2], w7[3], offset); + c3[0] = hc_bytealign_be (w7[1], w7[2], offset); + c2[3] = hc_bytealign_be (w7[0], w7[1], offset); + c2[2] = hc_bytealign_be (w6[3], w7[0], offset); + c2[1] = hc_bytealign_be (w6[2], w6[3], offset); + c2[0] = hc_bytealign_be (w6[1], w6[2], offset); + c1[3] = hc_bytealign_be (w6[0], w6[1], offset); + c1[2] = hc_bytealign_be (w5[3], w6[0], offset); + c1[1] = hc_bytealign_be (w5[2], w5[3], offset); + c1[0] = hc_bytealign_be (w5[1], w5[2], offset); + c0[3] = hc_bytealign_be (w5[0], w5[1], offset); + c0[2] = hc_bytealign_be (w4[3], w5[0], offset); + c0[1] = hc_bytealign_be (w4[2], w4[3], offset); + c0[0] = hc_bytealign_be (w4[1], w4[2], offset); + w7[3] = hc_bytealign_be (w4[0], w4[1], offset); + w7[2] = hc_bytealign_be (w3[3], w4[0], offset); + w7[1] = hc_bytealign_be (w3[2], w3[3], offset); + w7[0] = hc_bytealign_be (w3[1], w3[2], offset); + w6[3] = hc_bytealign_be (w3[0], w3[1], offset); + w6[2] = hc_bytealign_be (w2[3], w3[0], offset); + w6[1] = hc_bytealign_be (w2[2], w2[3], offset); + w6[0] = hc_bytealign_be (w2[1], w2[2], offset); + w5[3] = hc_bytealign_be (w2[0], w2[1], offset); + w5[2] = hc_bytealign_be (w1[3], w2[0], offset); + w5[1] = hc_bytealign_be (w1[2], w1[3], offset); + w5[0] = hc_bytealign_be (w1[1], w1[2], offset); + w4[3] = hc_bytealign_be (w1[0], w1[1], offset); + w4[2] = hc_bytealign_be (w0[3], w1[0], offset); + w4[1] = hc_bytealign_be (w0[2], w0[3], offset); + w4[0] = hc_bytealign_be (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -9141,39 +8987,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 15: - c3[3] = hc_bytealign (w7[3], 0, offset); - c3[2] = hc_bytealign (w7[2], w7[3], offset); - c3[1] = hc_bytealign (w7[1], w7[2], offset); - c3[0] = hc_bytealign (w7[0], w7[1], offset); - c2[3] = hc_bytealign (w6[3], w7[0], offset); - c2[2] = hc_bytealign (w6[2], w6[3], offset); - c2[1] = hc_bytealign (w6[1], w6[2], offset); - c2[0] = hc_bytealign (w6[0], w6[1], offset); - c1[3] = hc_bytealign (w5[3], w6[0], offset); - c1[2] = hc_bytealign (w5[2], w5[3], offset); - c1[1] = hc_bytealign (w5[1], w5[2], offset); - c1[0] = hc_bytealign (w5[0], w5[1], offset); - c0[3] = hc_bytealign (w4[3], w5[0], offset); - c0[2] = hc_bytealign (w4[2], w4[3], offset); - c0[1] = hc_bytealign (w4[1], w4[2], offset); - c0[0] = hc_bytealign (w4[0], w4[1], offset); - w7[3] = hc_bytealign (w3[3], w4[0], offset); - w7[2] = hc_bytealign (w3[2], w3[3], offset); - w7[1] = hc_bytealign (w3[1], w3[2], offset); - w7[0] = hc_bytealign (w3[0], w3[1], offset); - w6[3] = hc_bytealign (w2[3], w3[0], offset); - w6[2] = hc_bytealign (w2[2], w2[3], offset); - w6[1] = hc_bytealign (w2[1], w2[2], offset); - w6[0] = hc_bytealign (w2[0], w2[1], offset); - w5[3] = hc_bytealign (w1[3], w2[0], offset); - w5[2] = hc_bytealign (w1[2], w1[3], offset); - w5[1] = hc_bytealign (w1[1], w1[2], offset); - w5[0] = hc_bytealign (w1[0], w1[1], offset); - w4[3] = hc_bytealign (w0[3], w1[0], offset); - w4[2] = hc_bytealign (w0[2], w0[3], offset); - w4[1] = hc_bytealign (w0[1], w0[2], offset); - w4[0] = hc_bytealign (w0[0], w0[1], offset); - w3[3] = hc_bytealign ( 0, w0[0], offset); + c3[3] = hc_bytealign_be (w7[3], 0, offset); + c3[2] = hc_bytealign_be (w7[2], w7[3], offset); + c3[1] = hc_bytealign_be (w7[1], w7[2], offset); + c3[0] = hc_bytealign_be (w7[0], w7[1], offset); + c2[3] = hc_bytealign_be (w6[3], w7[0], offset); + c2[2] = hc_bytealign_be (w6[2], w6[3], offset); + c2[1] = hc_bytealign_be (w6[1], w6[2], offset); + c2[0] = hc_bytealign_be (w6[0], w6[1], offset); + c1[3] = hc_bytealign_be (w5[3], w6[0], offset); + c1[2] = hc_bytealign_be (w5[2], w5[3], offset); + c1[1] = hc_bytealign_be (w5[1], w5[2], offset); + c1[0] = hc_bytealign_be (w5[0], w5[1], offset); + c0[3] = hc_bytealign_be (w4[3], w5[0], offset); + c0[2] = hc_bytealign_be (w4[2], w4[3], offset); + c0[1] = hc_bytealign_be (w4[1], w4[2], offset); + c0[0] = hc_bytealign_be (w4[0], w4[1], offset); + w7[3] = hc_bytealign_be (w3[3], w4[0], offset); + w7[2] = hc_bytealign_be (w3[2], w3[3], offset); + w7[1] = hc_bytealign_be (w3[1], w3[2], offset); + w7[0] = hc_bytealign_be (w3[0], w3[1], offset); + w6[3] = hc_bytealign_be (w2[3], w3[0], offset); + w6[2] = hc_bytealign_be (w2[2], w2[3], offset); + w6[1] = hc_bytealign_be (w2[1], w2[2], offset); + w6[0] = hc_bytealign_be (w2[0], w2[1], offset); + w5[3] = hc_bytealign_be (w1[3], w2[0], offset); + w5[2] = hc_bytealign_be (w1[2], w1[3], offset); + w5[1] = hc_bytealign_be (w1[1], w1[2], offset); + w5[0] = hc_bytealign_be (w1[0], w1[1], offset); + w4[3] = hc_bytealign_be (w0[3], w1[0], offset); + w4[2] = hc_bytealign_be (w0[2], w0[3], offset); + w4[1] = hc_bytealign_be (w0[1], w0[2], offset); + w4[0] = hc_bytealign_be (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -9193,39 +9039,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 16: - c4[0] = hc_bytealign (w7[3], 0, offset); - c3[3] = hc_bytealign (w7[2], w7[3], offset); - c3[2] = hc_bytealign (w7[1], w7[2], offset); - c3[1] = hc_bytealign (w7[0], w7[1], offset); - c3[0] = hc_bytealign (w6[3], w7[0], offset); - c2[3] = hc_bytealign (w6[2], w6[3], offset); - c2[2] = hc_bytealign (w6[1], w6[2], offset); - c2[1] = hc_bytealign (w6[0], w6[1], offset); - c2[0] = hc_bytealign (w5[3], w6[0], offset); - c1[3] = hc_bytealign (w5[2], w5[3], offset); - c1[2] = hc_bytealign (w5[1], w5[2], offset); - c1[1] = hc_bytealign (w5[0], w5[1], offset); - c1[0] = hc_bytealign (w4[3], w5[0], offset); - c0[3] = hc_bytealign (w4[2], w4[3], offset); - c0[2] = hc_bytealign (w4[1], w4[2], offset); - c0[1] = hc_bytealign (w4[0], w4[1], offset); - c0[0] = hc_bytealign (w3[3], w4[0], offset); - w7[3] = hc_bytealign (w3[2], w3[3], offset); - w7[2] = hc_bytealign (w3[1], w3[2], offset); - w7[1] = hc_bytealign (w3[0], w3[1], offset); - w7[0] = hc_bytealign (w2[3], w3[0], offset); - w6[3] = hc_bytealign (w2[2], w2[3], offset); - w6[2] = hc_bytealign (w2[1], w2[2], offset); - w6[1] = hc_bytealign (w2[0], w2[1], offset); - w6[0] = hc_bytealign (w1[3], w2[0], offset); - w5[3] = hc_bytealign (w1[2], w1[3], offset); - w5[2] = hc_bytealign (w1[1], w1[2], offset); - w5[1] = hc_bytealign (w1[0], w1[1], offset); - w5[0] = hc_bytealign (w0[3], w1[0], offset); - w4[3] = hc_bytealign (w0[2], w0[3], offset); - w4[2] = hc_bytealign (w0[1], w0[2], offset); - w4[1] = hc_bytealign (w0[0], w0[1], offset); - w4[0] = hc_bytealign ( 0, w0[0], offset); + c4[0] = hc_bytealign_be (w7[3], 0, offset); + c3[3] = hc_bytealign_be (w7[2], w7[3], offset); + c3[2] = hc_bytealign_be (w7[1], w7[2], offset); + c3[1] = hc_bytealign_be (w7[0], w7[1], offset); + c3[0] = hc_bytealign_be (w6[3], w7[0], offset); + c2[3] = hc_bytealign_be (w6[2], w6[3], offset); + c2[2] = hc_bytealign_be (w6[1], w6[2], offset); + c2[1] = hc_bytealign_be (w6[0], w6[1], offset); + c2[0] = hc_bytealign_be (w5[3], w6[0], offset); + c1[3] = hc_bytealign_be (w5[2], w5[3], offset); + c1[2] = hc_bytealign_be (w5[1], w5[2], offset); + c1[1] = hc_bytealign_be (w5[0], w5[1], offset); + c1[0] = hc_bytealign_be (w4[3], w5[0], offset); + c0[3] = hc_bytealign_be (w4[2], w4[3], offset); + c0[2] = hc_bytealign_be (w4[1], w4[2], offset); + c0[1] = hc_bytealign_be (w4[0], w4[1], offset); + c0[0] = hc_bytealign_be (w3[3], w4[0], offset); + w7[3] = hc_bytealign_be (w3[2], w3[3], offset); + w7[2] = hc_bytealign_be (w3[1], w3[2], offset); + w7[1] = hc_bytealign_be (w3[0], w3[1], offset); + w7[0] = hc_bytealign_be (w2[3], w3[0], offset); + w6[3] = hc_bytealign_be (w2[2], w2[3], offset); + w6[2] = hc_bytealign_be (w2[1], w2[2], offset); + w6[1] = hc_bytealign_be (w2[0], w2[1], offset); + w6[0] = hc_bytealign_be (w1[3], w2[0], offset); + w5[3] = hc_bytealign_be (w1[2], w1[3], offset); + w5[2] = hc_bytealign_be (w1[1], w1[2], offset); + w5[1] = hc_bytealign_be (w1[0], w1[1], offset); + w5[0] = hc_bytealign_be (w0[3], w1[0], offset); + w4[3] = hc_bytealign_be (w0[2], w0[3], offset); + w4[2] = hc_bytealign_be (w0[1], w0[2], offset); + w4[1] = hc_bytealign_be (w0[0], w0[1], offset); + w4[0] = hc_bytealign_be ( 0, w0[0], offset); w3[3] = 0; w3[2] = 0; w3[1] = 0; @@ -9246,39 +9092,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 17: - c4[1] = hc_bytealign (w7[3], 0, offset); - c4[0] = hc_bytealign (w7[2], w7[3], offset); - c3[3] = hc_bytealign (w7[1], w7[2], offset); - c3[2] = hc_bytealign (w7[0], w7[1], offset); - c3[1] = hc_bytealign (w6[3], w7[0], offset); - c3[0] = hc_bytealign (w6[2], w6[3], offset); - c2[3] = hc_bytealign (w6[1], w6[2], offset); - c2[2] = hc_bytealign (w6[0], w6[1], offset); - c2[1] = hc_bytealign (w5[3], w6[0], offset); - c2[0] = hc_bytealign (w5[2], w5[3], offset); - c1[3] = hc_bytealign (w5[1], w5[2], offset); - c1[2] = hc_bytealign (w5[0], w5[1], offset); - c1[1] = hc_bytealign (w4[3], w5[0], offset); - c1[0] = hc_bytealign (w4[2], w4[3], offset); - c0[3] = hc_bytealign (w4[1], w4[2], offset); - c0[2] = hc_bytealign (w4[0], w4[1], offset); - c0[1] = hc_bytealign (w3[3], w4[0], offset); - c0[0] = hc_bytealign (w3[2], w3[3], offset); - w7[3] = hc_bytealign (w3[1], w3[2], offset); - w7[2] = hc_bytealign (w3[0], w3[1], offset); - w7[1] = hc_bytealign (w2[3], w3[0], offset); - w7[0] = hc_bytealign (w2[2], w2[3], offset); - w6[3] = hc_bytealign (w2[1], w2[2], offset); - w6[2] = hc_bytealign (w2[0], w2[1], offset); - w6[1] = hc_bytealign (w1[3], w2[0], offset); - w6[0] = hc_bytealign (w1[2], w1[3], offset); - w5[3] = hc_bytealign (w1[1], w1[2], offset); - w5[2] = hc_bytealign (w1[0], w1[1], offset); - w5[1] = hc_bytealign (w0[3], w1[0], offset); - w5[0] = hc_bytealign (w0[2], w0[3], offset); - w4[3] = hc_bytealign (w0[1], w0[2], offset); - w4[2] = hc_bytealign (w0[0], w0[1], offset); - w4[1] = hc_bytealign ( 0, w0[0], offset); + c4[1] = hc_bytealign_be (w7[3], 0, offset); + c4[0] = hc_bytealign_be (w7[2], w7[3], offset); + c3[3] = hc_bytealign_be (w7[1], w7[2], offset); + c3[2] = hc_bytealign_be (w7[0], w7[1], offset); + c3[1] = hc_bytealign_be (w6[3], w7[0], offset); + c3[0] = hc_bytealign_be (w6[2], w6[3], offset); + c2[3] = hc_bytealign_be (w6[1], w6[2], offset); + c2[2] = hc_bytealign_be (w6[0], w6[1], offset); + c2[1] = hc_bytealign_be (w5[3], w6[0], offset); + c2[0] = hc_bytealign_be (w5[2], w5[3], offset); + c1[3] = hc_bytealign_be (w5[1], w5[2], offset); + c1[2] = hc_bytealign_be (w5[0], w5[1], offset); + c1[1] = hc_bytealign_be (w4[3], w5[0], offset); + c1[0] = hc_bytealign_be (w4[2], w4[3], offset); + c0[3] = hc_bytealign_be (w4[1], w4[2], offset); + c0[2] = hc_bytealign_be (w4[0], w4[1], offset); + c0[1] = hc_bytealign_be (w3[3], w4[0], offset); + c0[0] = hc_bytealign_be (w3[2], w3[3], offset); + w7[3] = hc_bytealign_be (w3[1], w3[2], offset); + w7[2] = hc_bytealign_be (w3[0], w3[1], offset); + w7[1] = hc_bytealign_be (w2[3], w3[0], offset); + w7[0] = hc_bytealign_be (w2[2], w2[3], offset); + w6[3] = hc_bytealign_be (w2[1], w2[2], offset); + w6[2] = hc_bytealign_be (w2[0], w2[1], offset); + w6[1] = hc_bytealign_be (w1[3], w2[0], offset); + w6[0] = hc_bytealign_be (w1[2], w1[3], offset); + w5[3] = hc_bytealign_be (w1[1], w1[2], offset); + w5[2] = hc_bytealign_be (w1[0], w1[1], offset); + w5[1] = hc_bytealign_be (w0[3], w1[0], offset); + w5[0] = hc_bytealign_be (w0[2], w0[3], offset); + w4[3] = hc_bytealign_be (w0[1], w0[2], offset); + w4[2] = hc_bytealign_be (w0[0], w0[1], offset); + w4[1] = hc_bytealign_be ( 0, w0[0], offset); w4[0] = 0; w3[3] = 0; w3[2] = 0; @@ -9300,39 +9146,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 18: - c4[2] = hc_bytealign (w7[3], 0, offset); - c4[1] = hc_bytealign (w7[2], w7[3], offset); - c4[0] = hc_bytealign (w7[1], w7[2], offset); - c3[3] = hc_bytealign (w7[0], w7[1], offset); - c3[2] = hc_bytealign (w6[3], w7[0], offset); - c3[1] = hc_bytealign (w6[2], w6[3], offset); - c3[0] = hc_bytealign (w6[1], w6[2], offset); - c2[3] = hc_bytealign (w6[0], w6[1], offset); - c2[2] = hc_bytealign (w5[3], w6[0], offset); - c2[1] = hc_bytealign (w5[2], w5[3], offset); - c2[0] = hc_bytealign (w5[1], w5[2], offset); - c1[3] = hc_bytealign (w5[0], w5[1], offset); - c1[2] = hc_bytealign (w4[3], w5[0], offset); - c1[1] = hc_bytealign (w4[2], w4[3], offset); - c1[0] = hc_bytealign (w4[1], w4[2], offset); - c0[3] = hc_bytealign (w4[0], w4[1], offset); - c0[2] = hc_bytealign (w3[3], w4[0], offset); - c0[1] = hc_bytealign (w3[2], w3[3], offset); - c0[0] = hc_bytealign (w3[1], w3[2], offset); - w7[3] = hc_bytealign (w3[0], w3[1], offset); - w7[2] = hc_bytealign (w2[3], w3[0], offset); - w7[1] = hc_bytealign (w2[2], w2[3], offset); - w7[0] = hc_bytealign (w2[1], w2[2], offset); - w6[3] = hc_bytealign (w2[0], w2[1], offset); - w6[2] = hc_bytealign (w1[3], w2[0], offset); - w6[1] = hc_bytealign (w1[2], w1[3], offset); - w6[0] = hc_bytealign (w1[1], w1[2], offset); - w5[3] = hc_bytealign (w1[0], w1[1], offset); - w5[2] = hc_bytealign (w0[3], w1[0], offset); - w5[1] = hc_bytealign (w0[2], w0[3], offset); - w5[0] = hc_bytealign (w0[1], w0[2], offset); - w4[3] = hc_bytealign (w0[0], w0[1], offset); - w4[2] = hc_bytealign ( 0, w0[0], offset); + c4[2] = hc_bytealign_be (w7[3], 0, offset); + c4[1] = hc_bytealign_be (w7[2], w7[3], offset); + c4[0] = hc_bytealign_be (w7[1], w7[2], offset); + c3[3] = hc_bytealign_be (w7[0], w7[1], offset); + c3[2] = hc_bytealign_be (w6[3], w7[0], offset); + c3[1] = hc_bytealign_be (w6[2], w6[3], offset); + c3[0] = hc_bytealign_be (w6[1], w6[2], offset); + c2[3] = hc_bytealign_be (w6[0], w6[1], offset); + c2[2] = hc_bytealign_be (w5[3], w6[0], offset); + c2[1] = hc_bytealign_be (w5[2], w5[3], offset); + c2[0] = hc_bytealign_be (w5[1], w5[2], offset); + c1[3] = hc_bytealign_be (w5[0], w5[1], offset); + c1[2] = hc_bytealign_be (w4[3], w5[0], offset); + c1[1] = hc_bytealign_be (w4[2], w4[3], offset); + c1[0] = hc_bytealign_be (w4[1], w4[2], offset); + c0[3] = hc_bytealign_be (w4[0], w4[1], offset); + c0[2] = hc_bytealign_be (w3[3], w4[0], offset); + c0[1] = hc_bytealign_be (w3[2], w3[3], offset); + c0[0] = hc_bytealign_be (w3[1], w3[2], offset); + w7[3] = hc_bytealign_be (w3[0], w3[1], offset); + w7[2] = hc_bytealign_be (w2[3], w3[0], offset); + w7[1] = hc_bytealign_be (w2[2], w2[3], offset); + w7[0] = hc_bytealign_be (w2[1], w2[2], offset); + w6[3] = hc_bytealign_be (w2[0], w2[1], offset); + w6[2] = hc_bytealign_be (w1[3], w2[0], offset); + w6[1] = hc_bytealign_be (w1[2], w1[3], offset); + w6[0] = hc_bytealign_be (w1[1], w1[2], offset); + w5[3] = hc_bytealign_be (w1[0], w1[1], offset); + w5[2] = hc_bytealign_be (w0[3], w1[0], offset); + w5[1] = hc_bytealign_be (w0[2], w0[3], offset); + w5[0] = hc_bytealign_be (w0[1], w0[2], offset); + w4[3] = hc_bytealign_be (w0[0], w0[1], offset); + w4[2] = hc_bytealign_be ( 0, w0[0], offset); w4[1] = 0; w4[0] = 0; w3[3] = 0; @@ -9355,39 +9201,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 19: - c4[3] = hc_bytealign (w7[3], 0, offset); - c4[2] = hc_bytealign (w7[2], w7[3], offset); - c4[1] = hc_bytealign (w7[1], w7[2], offset); - c4[0] = hc_bytealign (w7[0], w7[1], offset); - c3[3] = hc_bytealign (w6[3], w7[0], offset); - c3[2] = hc_bytealign (w6[2], w6[3], offset); - c3[1] = hc_bytealign (w6[1], w6[2], offset); - c3[0] = hc_bytealign (w6[0], w6[1], offset); - c2[3] = hc_bytealign (w5[3], w6[0], offset); - c2[2] = hc_bytealign (w5[2], w5[3], offset); - c2[1] = hc_bytealign (w5[1], w5[2], offset); - c2[0] = hc_bytealign (w5[0], w5[1], offset); - c1[3] = hc_bytealign (w4[3], w5[0], offset); - c1[2] = hc_bytealign (w4[2], w4[3], offset); - c1[1] = hc_bytealign (w4[1], w4[2], offset); - c1[0] = hc_bytealign (w4[0], w4[1], offset); - c0[3] = hc_bytealign (w3[3], w4[0], offset); - c0[2] = hc_bytealign (w3[2], w3[3], offset); - c0[1] = hc_bytealign (w3[1], w3[2], offset); - c0[0] = hc_bytealign (w3[0], w3[1], offset); - w7[3] = hc_bytealign (w2[3], w3[0], offset); - w7[2] = hc_bytealign (w2[2], w2[3], offset); - w7[1] = hc_bytealign (w2[1], w2[2], offset); - w7[0] = hc_bytealign (w2[0], w2[1], offset); - w6[3] = hc_bytealign (w1[3], w2[0], offset); - w6[2] = hc_bytealign (w1[2], w1[3], offset); - w6[1] = hc_bytealign (w1[1], w1[2], offset); - w6[0] = hc_bytealign (w1[0], w1[1], offset); - w5[3] = hc_bytealign (w0[3], w1[0], offset); - w5[2] = hc_bytealign (w0[2], w0[3], offset); - w5[1] = hc_bytealign (w0[1], w0[2], offset); - w5[0] = hc_bytealign (w0[0], w0[1], offset); - w4[3] = hc_bytealign ( 0, w0[0], offset); + c4[3] = hc_bytealign_be (w7[3], 0, offset); + c4[2] = hc_bytealign_be (w7[2], w7[3], offset); + c4[1] = hc_bytealign_be (w7[1], w7[2], offset); + c4[0] = hc_bytealign_be (w7[0], w7[1], offset); + c3[3] = hc_bytealign_be (w6[3], w7[0], offset); + c3[2] = hc_bytealign_be (w6[2], w6[3], offset); + c3[1] = hc_bytealign_be (w6[1], w6[2], offset); + c3[0] = hc_bytealign_be (w6[0], w6[1], offset); + c2[3] = hc_bytealign_be (w5[3], w6[0], offset); + c2[2] = hc_bytealign_be (w5[2], w5[3], offset); + c2[1] = hc_bytealign_be (w5[1], w5[2], offset); + c2[0] = hc_bytealign_be (w5[0], w5[1], offset); + c1[3] = hc_bytealign_be (w4[3], w5[0], offset); + c1[2] = hc_bytealign_be (w4[2], w4[3], offset); + c1[1] = hc_bytealign_be (w4[1], w4[2], offset); + c1[0] = hc_bytealign_be (w4[0], w4[1], offset); + c0[3] = hc_bytealign_be (w3[3], w4[0], offset); + c0[2] = hc_bytealign_be (w3[2], w3[3], offset); + c0[1] = hc_bytealign_be (w3[1], w3[2], offset); + c0[0] = hc_bytealign_be (w3[0], w3[1], offset); + w7[3] = hc_bytealign_be (w2[3], w3[0], offset); + w7[2] = hc_bytealign_be (w2[2], w2[3], offset); + w7[1] = hc_bytealign_be (w2[1], w2[2], offset); + w7[0] = hc_bytealign_be (w2[0], w2[1], offset); + w6[3] = hc_bytealign_be (w1[3], w2[0], offset); + w6[2] = hc_bytealign_be (w1[2], w1[3], offset); + w6[1] = hc_bytealign_be (w1[1], w1[2], offset); + w6[0] = hc_bytealign_be (w1[0], w1[1], offset); + w5[3] = hc_bytealign_be (w0[3], w1[0], offset); + w5[2] = hc_bytealign_be (w0[2], w0[3], offset); + w5[1] = hc_bytealign_be (w0[1], w0[2], offset); + w5[0] = hc_bytealign_be (w0[0], w0[1], offset); + w4[3] = hc_bytealign_be ( 0, w0[0], offset); w4[2] = 0; w4[1] = 0; w4[0] = 0; @@ -9411,39 +9257,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 20: - c5[0] = hc_bytealign (w7[3], 0, offset); - c4[3] = hc_bytealign (w7[2], w7[3], offset); - c4[2] = hc_bytealign (w7[1], w7[2], offset); - c4[1] = hc_bytealign (w7[0], w7[1], offset); - c4[0] = hc_bytealign (w6[3], w7[0], offset); - c3[3] = hc_bytealign (w6[2], w6[3], offset); - c3[2] = hc_bytealign (w6[1], w6[2], offset); - c3[1] = hc_bytealign (w6[0], w6[1], offset); - c3[0] = hc_bytealign (w5[3], w6[0], offset); - c2[3] = hc_bytealign (w5[2], w5[3], offset); - c2[2] = hc_bytealign (w5[1], w5[2], offset); - c2[1] = hc_bytealign (w5[0], w5[1], offset); - c2[0] = hc_bytealign (w4[3], w5[0], offset); - c1[3] = hc_bytealign (w4[2], w4[3], offset); - c1[2] = hc_bytealign (w4[1], w4[2], offset); - c1[1] = hc_bytealign (w4[0], w4[1], offset); - c1[0] = hc_bytealign (w3[3], w4[0], offset); - c0[3] = hc_bytealign (w3[2], w3[3], offset); - c0[2] = hc_bytealign (w3[1], w3[2], offset); - c0[1] = hc_bytealign (w3[0], w3[1], offset); - c0[0] = hc_bytealign (w2[3], w3[0], offset); - w7[3] = hc_bytealign (w2[2], w2[3], offset); - w7[2] = hc_bytealign (w2[1], w2[2], offset); - w7[1] = hc_bytealign (w2[0], w2[1], offset); - w7[0] = hc_bytealign (w1[3], w2[0], offset); - w6[3] = hc_bytealign (w1[2], w1[3], offset); - w6[2] = hc_bytealign (w1[1], w1[2], offset); - w6[1] = hc_bytealign (w1[0], w1[1], offset); - w6[0] = hc_bytealign (w0[3], w1[0], offset); - w5[3] = hc_bytealign (w0[2], w0[3], offset); - w5[2] = hc_bytealign (w0[1], w0[2], offset); - w5[1] = hc_bytealign (w0[0], w0[1], offset); - w5[0] = hc_bytealign ( 0, w0[0], offset); + c5[0] = hc_bytealign_be (w7[3], 0, offset); + c4[3] = hc_bytealign_be (w7[2], w7[3], offset); + c4[2] = hc_bytealign_be (w7[1], w7[2], offset); + c4[1] = hc_bytealign_be (w7[0], w7[1], offset); + c4[0] = hc_bytealign_be (w6[3], w7[0], offset); + c3[3] = hc_bytealign_be (w6[2], w6[3], offset); + c3[2] = hc_bytealign_be (w6[1], w6[2], offset); + c3[1] = hc_bytealign_be (w6[0], w6[1], offset); + c3[0] = hc_bytealign_be (w5[3], w6[0], offset); + c2[3] = hc_bytealign_be (w5[2], w5[3], offset); + c2[2] = hc_bytealign_be (w5[1], w5[2], offset); + c2[1] = hc_bytealign_be (w5[0], w5[1], offset); + c2[0] = hc_bytealign_be (w4[3], w5[0], offset); + c1[3] = hc_bytealign_be (w4[2], w4[3], offset); + c1[2] = hc_bytealign_be (w4[1], w4[2], offset); + c1[1] = hc_bytealign_be (w4[0], w4[1], offset); + c1[0] = hc_bytealign_be (w3[3], w4[0], offset); + c0[3] = hc_bytealign_be (w3[2], w3[3], offset); + c0[2] = hc_bytealign_be (w3[1], w3[2], offset); + c0[1] = hc_bytealign_be (w3[0], w3[1], offset); + c0[0] = hc_bytealign_be (w2[3], w3[0], offset); + w7[3] = hc_bytealign_be (w2[2], w2[3], offset); + w7[2] = hc_bytealign_be (w2[1], w2[2], offset); + w7[1] = hc_bytealign_be (w2[0], w2[1], offset); + w7[0] = hc_bytealign_be (w1[3], w2[0], offset); + w6[3] = hc_bytealign_be (w1[2], w1[3], offset); + w6[2] = hc_bytealign_be (w1[1], w1[2], offset); + w6[1] = hc_bytealign_be (w1[0], w1[1], offset); + w6[0] = hc_bytealign_be (w0[3], w1[0], offset); + w5[3] = hc_bytealign_be (w0[2], w0[3], offset); + w5[2] = hc_bytealign_be (w0[1], w0[2], offset); + w5[1] = hc_bytealign_be (w0[0], w0[1], offset); + w5[0] = hc_bytealign_be ( 0, w0[0], offset); w4[3] = 0; w4[2] = 0; w4[1] = 0; @@ -9468,39 +9314,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 21: - c5[1] = hc_bytealign (w7[3], 0, offset); - c5[0] = hc_bytealign (w7[2], w7[3], offset); - c4[3] = hc_bytealign (w7[1], w7[2], offset); - c4[2] = hc_bytealign (w7[0], w7[1], offset); - c4[1] = hc_bytealign (w6[3], w7[0], offset); - c4[0] = hc_bytealign (w6[2], w6[3], offset); - c3[3] = hc_bytealign (w6[1], w6[2], offset); - c3[2] = hc_bytealign (w6[0], w6[1], offset); - c3[1] = hc_bytealign (w5[3], w6[0], offset); - c3[0] = hc_bytealign (w5[2], w5[3], offset); - c2[3] = hc_bytealign (w5[1], w5[2], offset); - c2[2] = hc_bytealign (w5[0], w5[1], offset); - c2[1] = hc_bytealign (w4[3], w5[0], offset); - c2[0] = hc_bytealign (w4[2], w4[3], offset); - c1[3] = hc_bytealign (w4[1], w4[2], offset); - c1[2] = hc_bytealign (w4[0], w4[1], offset); - c1[1] = hc_bytealign (w3[3], w4[0], offset); - c1[0] = hc_bytealign (w3[2], w3[3], offset); - c0[3] = hc_bytealign (w3[1], w3[2], offset); - c0[2] = hc_bytealign (w3[0], w3[1], offset); - c0[1] = hc_bytealign (w2[3], w3[0], offset); - c0[0] = hc_bytealign (w2[2], w2[3], offset); - w7[3] = hc_bytealign (w2[1], w2[2], offset); - w7[2] = hc_bytealign (w2[0], w2[1], offset); - w7[1] = hc_bytealign (w1[3], w2[0], offset); - w7[0] = hc_bytealign (w1[2], w1[3], offset); - w6[3] = hc_bytealign (w1[1], w1[2], offset); - w6[2] = hc_bytealign (w1[0], w1[1], offset); - w6[1] = hc_bytealign (w0[3], w1[0], offset); - w6[0] = hc_bytealign (w0[2], w0[3], offset); - w5[3] = hc_bytealign (w0[1], w0[2], offset); - w5[2] = hc_bytealign (w0[0], w0[1], offset); - w5[1] = hc_bytealign ( 0, w0[0], offset); + c5[1] = hc_bytealign_be (w7[3], 0, offset); + c5[0] = hc_bytealign_be (w7[2], w7[3], offset); + c4[3] = hc_bytealign_be (w7[1], w7[2], offset); + c4[2] = hc_bytealign_be (w7[0], w7[1], offset); + c4[1] = hc_bytealign_be (w6[3], w7[0], offset); + c4[0] = hc_bytealign_be (w6[2], w6[3], offset); + c3[3] = hc_bytealign_be (w6[1], w6[2], offset); + c3[2] = hc_bytealign_be (w6[0], w6[1], offset); + c3[1] = hc_bytealign_be (w5[3], w6[0], offset); + c3[0] = hc_bytealign_be (w5[2], w5[3], offset); + c2[3] = hc_bytealign_be (w5[1], w5[2], offset); + c2[2] = hc_bytealign_be (w5[0], w5[1], offset); + c2[1] = hc_bytealign_be (w4[3], w5[0], offset); + c2[0] = hc_bytealign_be (w4[2], w4[3], offset); + c1[3] = hc_bytealign_be (w4[1], w4[2], offset); + c1[2] = hc_bytealign_be (w4[0], w4[1], offset); + c1[1] = hc_bytealign_be (w3[3], w4[0], offset); + c1[0] = hc_bytealign_be (w3[2], w3[3], offset); + c0[3] = hc_bytealign_be (w3[1], w3[2], offset); + c0[2] = hc_bytealign_be (w3[0], w3[1], offset); + c0[1] = hc_bytealign_be (w2[3], w3[0], offset); + c0[0] = hc_bytealign_be (w2[2], w2[3], offset); + w7[3] = hc_bytealign_be (w2[1], w2[2], offset); + w7[2] = hc_bytealign_be (w2[0], w2[1], offset); + w7[1] = hc_bytealign_be (w1[3], w2[0], offset); + w7[0] = hc_bytealign_be (w1[2], w1[3], offset); + w6[3] = hc_bytealign_be (w1[1], w1[2], offset); + w6[2] = hc_bytealign_be (w1[0], w1[1], offset); + w6[1] = hc_bytealign_be (w0[3], w1[0], offset); + w6[0] = hc_bytealign_be (w0[2], w0[3], offset); + w5[3] = hc_bytealign_be (w0[1], w0[2], offset); + w5[2] = hc_bytealign_be (w0[0], w0[1], offset); + w5[1] = hc_bytealign_be ( 0, w0[0], offset); w5[0] = 0; w4[3] = 0; w4[2] = 0; @@ -9526,39 +9372,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 22: - c5[2] = hc_bytealign (w7[3], 0, offset); - c5[1] = hc_bytealign (w7[2], w7[3], offset); - c5[0] = hc_bytealign (w7[1], w7[2], offset); - c4[3] = hc_bytealign (w7[0], w7[1], offset); - c4[2] = hc_bytealign (w6[3], w7[0], offset); - c4[1] = hc_bytealign (w6[2], w6[3], offset); - c4[0] = hc_bytealign (w6[1], w6[2], offset); - c3[3] = hc_bytealign (w6[0], w6[1], offset); - c3[2] = hc_bytealign (w5[3], w6[0], offset); - c3[1] = hc_bytealign (w5[2], w5[3], offset); - c3[0] = hc_bytealign (w5[1], w5[2], offset); - c2[3] = hc_bytealign (w5[0], w5[1], offset); - c2[2] = hc_bytealign (w4[3], w5[0], offset); - c2[1] = hc_bytealign (w4[2], w4[3], offset); - c2[0] = hc_bytealign (w4[1], w4[2], offset); - c1[3] = hc_bytealign (w4[0], w4[1], offset); - c1[2] = hc_bytealign (w3[3], w4[0], offset); - c1[1] = hc_bytealign (w3[2], w3[3], offset); - c1[0] = hc_bytealign (w3[1], w3[2], offset); - c0[3] = hc_bytealign (w3[0], w3[1], offset); - c0[2] = hc_bytealign (w2[3], w3[0], offset); - c0[1] = hc_bytealign (w2[2], w2[3], offset); - c0[0] = hc_bytealign (w2[1], w2[2], offset); - w7[3] = hc_bytealign (w2[0], w2[1], offset); - w7[2] = hc_bytealign (w1[3], w2[0], offset); - w7[1] = hc_bytealign (w1[2], w1[3], offset); - w7[0] = hc_bytealign (w1[1], w1[2], offset); - w6[3] = hc_bytealign (w1[0], w1[1], offset); - w6[2] = hc_bytealign (w0[3], w1[0], offset); - w6[1] = hc_bytealign (w0[2], w0[3], offset); - w6[0] = hc_bytealign (w0[1], w0[2], offset); - w5[3] = hc_bytealign (w0[0], w0[1], offset); - w5[2] = hc_bytealign ( 0, w0[0], offset); + c5[2] = hc_bytealign_be (w7[3], 0, offset); + c5[1] = hc_bytealign_be (w7[2], w7[3], offset); + c5[0] = hc_bytealign_be (w7[1], w7[2], offset); + c4[3] = hc_bytealign_be (w7[0], w7[1], offset); + c4[2] = hc_bytealign_be (w6[3], w7[0], offset); + c4[1] = hc_bytealign_be (w6[2], w6[3], offset); + c4[0] = hc_bytealign_be (w6[1], w6[2], offset); + c3[3] = hc_bytealign_be (w6[0], w6[1], offset); + c3[2] = hc_bytealign_be (w5[3], w6[0], offset); + c3[1] = hc_bytealign_be (w5[2], w5[3], offset); + c3[0] = hc_bytealign_be (w5[1], w5[2], offset); + c2[3] = hc_bytealign_be (w5[0], w5[1], offset); + c2[2] = hc_bytealign_be (w4[3], w5[0], offset); + c2[1] = hc_bytealign_be (w4[2], w4[3], offset); + c2[0] = hc_bytealign_be (w4[1], w4[2], offset); + c1[3] = hc_bytealign_be (w4[0], w4[1], offset); + c1[2] = hc_bytealign_be (w3[3], w4[0], offset); + c1[1] = hc_bytealign_be (w3[2], w3[3], offset); + c1[0] = hc_bytealign_be (w3[1], w3[2], offset); + c0[3] = hc_bytealign_be (w3[0], w3[1], offset); + c0[2] = hc_bytealign_be (w2[3], w3[0], offset); + c0[1] = hc_bytealign_be (w2[2], w2[3], offset); + c0[0] = hc_bytealign_be (w2[1], w2[2], offset); + w7[3] = hc_bytealign_be (w2[0], w2[1], offset); + w7[2] = hc_bytealign_be (w1[3], w2[0], offset); + w7[1] = hc_bytealign_be (w1[2], w1[3], offset); + w7[0] = hc_bytealign_be (w1[1], w1[2], offset); + w6[3] = hc_bytealign_be (w1[0], w1[1], offset); + w6[2] = hc_bytealign_be (w0[3], w1[0], offset); + w6[1] = hc_bytealign_be (w0[2], w0[3], offset); + w6[0] = hc_bytealign_be (w0[1], w0[2], offset); + w5[3] = hc_bytealign_be (w0[0], w0[1], offset); + w5[2] = hc_bytealign_be ( 0, w0[0], offset); w5[1] = 0; w5[0] = 0; w4[3] = 0; @@ -9585,39 +9431,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 23: - c5[3] = hc_bytealign (w7[3], 0, offset); - c5[2] = hc_bytealign (w7[2], w7[3], offset); - c5[1] = hc_bytealign (w7[1], w7[2], offset); - c5[0] = hc_bytealign (w7[0], w7[1], offset); - c4[3] = hc_bytealign (w6[3], w7[0], offset); - c4[2] = hc_bytealign (w6[2], w6[3], offset); - c4[1] = hc_bytealign (w6[1], w6[2], offset); - c4[0] = hc_bytealign (w6[0], w6[1], offset); - c3[3] = hc_bytealign (w5[3], w6[0], offset); - c3[2] = hc_bytealign (w5[2], w5[3], offset); - c3[1] = hc_bytealign (w5[1], w5[2], offset); - c3[0] = hc_bytealign (w5[0], w5[1], offset); - c2[3] = hc_bytealign (w4[3], w5[0], offset); - c2[2] = hc_bytealign (w4[2], w4[3], offset); - c2[1] = hc_bytealign (w4[1], w4[2], offset); - c2[0] = hc_bytealign (w4[0], w4[1], offset); - c1[3] = hc_bytealign (w3[3], w4[0], offset); - c1[2] = hc_bytealign (w3[2], w3[3], offset); - c1[1] = hc_bytealign (w3[1], w3[2], offset); - c1[0] = hc_bytealign (w3[0], w3[1], offset); - c0[3] = hc_bytealign (w2[3], w3[0], offset); - c0[2] = hc_bytealign (w2[2], w2[3], offset); - c0[1] = hc_bytealign (w2[1], w2[2], offset); - c0[0] = hc_bytealign (w2[0], w2[1], offset); - w7[3] = hc_bytealign (w1[3], w2[0], offset); - w7[2] = hc_bytealign (w1[2], w1[3], offset); - w7[1] = hc_bytealign (w1[1], w1[2], offset); - w7[0] = hc_bytealign (w1[0], w1[1], offset); - w6[3] = hc_bytealign (w0[3], w1[0], offset); - w6[2] = hc_bytealign (w0[2], w0[3], offset); - w6[1] = hc_bytealign (w0[1], w0[2], offset); - w6[0] = hc_bytealign (w0[0], w0[1], offset); - w5[3] = hc_bytealign ( 0, w0[0], offset); + c5[3] = hc_bytealign_be (w7[3], 0, offset); + c5[2] = hc_bytealign_be (w7[2], w7[3], offset); + c5[1] = hc_bytealign_be (w7[1], w7[2], offset); + c5[0] = hc_bytealign_be (w7[0], w7[1], offset); + c4[3] = hc_bytealign_be (w6[3], w7[0], offset); + c4[2] = hc_bytealign_be (w6[2], w6[3], offset); + c4[1] = hc_bytealign_be (w6[1], w6[2], offset); + c4[0] = hc_bytealign_be (w6[0], w6[1], offset); + c3[3] = hc_bytealign_be (w5[3], w6[0], offset); + c3[2] = hc_bytealign_be (w5[2], w5[3], offset); + c3[1] = hc_bytealign_be (w5[1], w5[2], offset); + c3[0] = hc_bytealign_be (w5[0], w5[1], offset); + c2[3] = hc_bytealign_be (w4[3], w5[0], offset); + c2[2] = hc_bytealign_be (w4[2], w4[3], offset); + c2[1] = hc_bytealign_be (w4[1], w4[2], offset); + c2[0] = hc_bytealign_be (w4[0], w4[1], offset); + c1[3] = hc_bytealign_be (w3[3], w4[0], offset); + c1[2] = hc_bytealign_be (w3[2], w3[3], offset); + c1[1] = hc_bytealign_be (w3[1], w3[2], offset); + c1[0] = hc_bytealign_be (w3[0], w3[1], offset); + c0[3] = hc_bytealign_be (w2[3], w3[0], offset); + c0[2] = hc_bytealign_be (w2[2], w2[3], offset); + c0[1] = hc_bytealign_be (w2[1], w2[2], offset); + c0[0] = hc_bytealign_be (w2[0], w2[1], offset); + w7[3] = hc_bytealign_be (w1[3], w2[0], offset); + w7[2] = hc_bytealign_be (w1[2], w1[3], offset); + w7[1] = hc_bytealign_be (w1[1], w1[2], offset); + w7[0] = hc_bytealign_be (w1[0], w1[1], offset); + w6[3] = hc_bytealign_be (w0[3], w1[0], offset); + w6[2] = hc_bytealign_be (w0[2], w0[3], offset); + w6[1] = hc_bytealign_be (w0[1], w0[2], offset); + w6[0] = hc_bytealign_be (w0[0], w0[1], offset); + w5[3] = hc_bytealign_be ( 0, w0[0], offset); w5[2] = 0; w5[1] = 0; w5[0] = 0; @@ -9645,39 +9491,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 24: - c6[0] = hc_bytealign (w7[3], 0, offset); - c5[3] = hc_bytealign (w7[2], w7[3], offset); - c5[2] = hc_bytealign (w7[1], w7[2], offset); - c5[1] = hc_bytealign (w7[0], w7[1], offset); - c5[0] = hc_bytealign (w6[3], w7[0], offset); - c4[3] = hc_bytealign (w6[2], w6[3], offset); - c4[2] = hc_bytealign (w6[1], w6[2], offset); - c4[1] = hc_bytealign (w6[0], w6[1], offset); - c4[0] = hc_bytealign (w5[3], w6[0], offset); - c3[3] = hc_bytealign (w5[2], w5[3], offset); - c3[2] = hc_bytealign (w5[1], w5[2], offset); - c3[1] = hc_bytealign (w5[0], w5[1], offset); - c3[0] = hc_bytealign (w4[3], w5[0], offset); - c2[3] = hc_bytealign (w4[2], w4[3], offset); - c2[2] = hc_bytealign (w4[1], w4[2], offset); - c2[1] = hc_bytealign (w4[0], w4[1], offset); - c2[0] = hc_bytealign (w3[3], w4[0], offset); - c1[3] = hc_bytealign (w3[2], w3[3], offset); - c1[2] = hc_bytealign (w3[1], w3[2], offset); - c1[1] = hc_bytealign (w3[0], w3[1], offset); - c1[0] = hc_bytealign (w2[3], w3[0], offset); - c0[3] = hc_bytealign (w2[2], w2[3], offset); - c0[2] = hc_bytealign (w2[1], w2[2], offset); - c0[1] = hc_bytealign (w2[0], w2[1], offset); - c0[0] = hc_bytealign (w1[3], w2[0], offset); - w7[3] = hc_bytealign (w1[2], w1[3], offset); - w7[2] = hc_bytealign (w1[1], w1[2], offset); - w7[1] = hc_bytealign (w1[0], w1[1], offset); - w7[0] = hc_bytealign (w0[3], w1[0], offset); - w6[3] = hc_bytealign (w0[2], w0[3], offset); - w6[2] = hc_bytealign (w0[1], w0[2], offset); - w6[1] = hc_bytealign (w0[0], w0[1], offset); - w6[0] = hc_bytealign ( 0, w0[0], offset); + c6[0] = hc_bytealign_be (w7[3], 0, offset); + c5[3] = hc_bytealign_be (w7[2], w7[3], offset); + c5[2] = hc_bytealign_be (w7[1], w7[2], offset); + c5[1] = hc_bytealign_be (w7[0], w7[1], offset); + c5[0] = hc_bytealign_be (w6[3], w7[0], offset); + c4[3] = hc_bytealign_be (w6[2], w6[3], offset); + c4[2] = hc_bytealign_be (w6[1], w6[2], offset); + c4[1] = hc_bytealign_be (w6[0], w6[1], offset); + c4[0] = hc_bytealign_be (w5[3], w6[0], offset); + c3[3] = hc_bytealign_be (w5[2], w5[3], offset); + c3[2] = hc_bytealign_be (w5[1], w5[2], offset); + c3[1] = hc_bytealign_be (w5[0], w5[1], offset); + c3[0] = hc_bytealign_be (w4[3], w5[0], offset); + c2[3] = hc_bytealign_be (w4[2], w4[3], offset); + c2[2] = hc_bytealign_be (w4[1], w4[2], offset); + c2[1] = hc_bytealign_be (w4[0], w4[1], offset); + c2[0] = hc_bytealign_be (w3[3], w4[0], offset); + c1[3] = hc_bytealign_be (w3[2], w3[3], offset); + c1[2] = hc_bytealign_be (w3[1], w3[2], offset); + c1[1] = hc_bytealign_be (w3[0], w3[1], offset); + c1[0] = hc_bytealign_be (w2[3], w3[0], offset); + c0[3] = hc_bytealign_be (w2[2], w2[3], offset); + c0[2] = hc_bytealign_be (w2[1], w2[2], offset); + c0[1] = hc_bytealign_be (w2[0], w2[1], offset); + c0[0] = hc_bytealign_be (w1[3], w2[0], offset); + w7[3] = hc_bytealign_be (w1[2], w1[3], offset); + w7[2] = hc_bytealign_be (w1[1], w1[2], offset); + w7[1] = hc_bytealign_be (w1[0], w1[1], offset); + w7[0] = hc_bytealign_be (w0[3], w1[0], offset); + w6[3] = hc_bytealign_be (w0[2], w0[3], offset); + w6[2] = hc_bytealign_be (w0[1], w0[2], offset); + w6[1] = hc_bytealign_be (w0[0], w0[1], offset); + w6[0] = hc_bytealign_be ( 0, w0[0], offset); w5[3] = 0; w5[2] = 0; w5[1] = 0; @@ -9706,39 +9552,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 25: - c6[1] = hc_bytealign (w7[3], 0, offset); - c6[0] = hc_bytealign (w7[2], w7[3], offset); - c5[3] = hc_bytealign (w7[1], w7[2], offset); - c5[2] = hc_bytealign (w7[0], w7[1], offset); - c5[1] = hc_bytealign (w6[3], w7[0], offset); - c5[0] = hc_bytealign (w6[2], w6[3], offset); - c4[3] = hc_bytealign (w6[1], w6[2], offset); - c4[2] = hc_bytealign (w6[0], w6[1], offset); - c4[1] = hc_bytealign (w5[3], w6[0], offset); - c4[0] = hc_bytealign (w5[2], w5[3], offset); - c3[3] = hc_bytealign (w5[1], w5[2], offset); - c3[2] = hc_bytealign (w5[0], w5[1], offset); - c3[1] = hc_bytealign (w4[3], w5[0], offset); - c3[0] = hc_bytealign (w4[2], w4[3], offset); - c2[3] = hc_bytealign (w4[1], w4[2], offset); - c2[2] = hc_bytealign (w4[0], w4[1], offset); - c2[1] = hc_bytealign (w3[3], w4[0], offset); - c2[0] = hc_bytealign (w3[2], w3[3], offset); - c1[3] = hc_bytealign (w3[1], w3[2], offset); - c1[2] = hc_bytealign (w3[0], w3[1], offset); - c1[1] = hc_bytealign (w2[3], w3[0], offset); - c1[0] = hc_bytealign (w2[2], w2[3], offset); - c0[3] = hc_bytealign (w2[1], w2[2], offset); - c0[2] = hc_bytealign (w2[0], w2[1], offset); - c0[1] = hc_bytealign (w1[3], w2[0], offset); - c0[0] = hc_bytealign (w1[2], w1[3], offset); - w7[3] = hc_bytealign (w1[1], w1[2], offset); - w7[2] = hc_bytealign (w1[0], w1[1], offset); - w7[1] = hc_bytealign (w0[3], w1[0], offset); - w7[0] = hc_bytealign (w0[2], w0[3], offset); - w6[3] = hc_bytealign (w0[1], w0[2], offset); - w6[2] = hc_bytealign (w0[0], w0[1], offset); - w6[1] = hc_bytealign ( 0, w0[0], offset); + c6[1] = hc_bytealign_be (w7[3], 0, offset); + c6[0] = hc_bytealign_be (w7[2], w7[3], offset); + c5[3] = hc_bytealign_be (w7[1], w7[2], offset); + c5[2] = hc_bytealign_be (w7[0], w7[1], offset); + c5[1] = hc_bytealign_be (w6[3], w7[0], offset); + c5[0] = hc_bytealign_be (w6[2], w6[3], offset); + c4[3] = hc_bytealign_be (w6[1], w6[2], offset); + c4[2] = hc_bytealign_be (w6[0], w6[1], offset); + c4[1] = hc_bytealign_be (w5[3], w6[0], offset); + c4[0] = hc_bytealign_be (w5[2], w5[3], offset); + c3[3] = hc_bytealign_be (w5[1], w5[2], offset); + c3[2] = hc_bytealign_be (w5[0], w5[1], offset); + c3[1] = hc_bytealign_be (w4[3], w5[0], offset); + c3[0] = hc_bytealign_be (w4[2], w4[3], offset); + c2[3] = hc_bytealign_be (w4[1], w4[2], offset); + c2[2] = hc_bytealign_be (w4[0], w4[1], offset); + c2[1] = hc_bytealign_be (w3[3], w4[0], offset); + c2[0] = hc_bytealign_be (w3[2], w3[3], offset); + c1[3] = hc_bytealign_be (w3[1], w3[2], offset); + c1[2] = hc_bytealign_be (w3[0], w3[1], offset); + c1[1] = hc_bytealign_be (w2[3], w3[0], offset); + c1[0] = hc_bytealign_be (w2[2], w2[3], offset); + c0[3] = hc_bytealign_be (w2[1], w2[2], offset); + c0[2] = hc_bytealign_be (w2[0], w2[1], offset); + c0[1] = hc_bytealign_be (w1[3], w2[0], offset); + c0[0] = hc_bytealign_be (w1[2], w1[3], offset); + w7[3] = hc_bytealign_be (w1[1], w1[2], offset); + w7[2] = hc_bytealign_be (w1[0], w1[1], offset); + w7[1] = hc_bytealign_be (w0[3], w1[0], offset); + w7[0] = hc_bytealign_be (w0[2], w0[3], offset); + w6[3] = hc_bytealign_be (w0[1], w0[2], offset); + w6[2] = hc_bytealign_be (w0[0], w0[1], offset); + w6[1] = hc_bytealign_be ( 0, w0[0], offset); w6[0] = 0; w5[3] = 0; w5[2] = 0; @@ -9768,39 +9614,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 26: - c6[2] = hc_bytealign (w7[3], 0, offset); - c6[1] = hc_bytealign (w7[2], w7[3], offset); - c6[0] = hc_bytealign (w7[1], w7[2], offset); - c5[3] = hc_bytealign (w7[0], w7[1], offset); - c5[2] = hc_bytealign (w6[3], w7[0], offset); - c5[1] = hc_bytealign (w6[2], w6[3], offset); - c5[0] = hc_bytealign (w6[1], w6[2], offset); - c4[3] = hc_bytealign (w6[0], w6[1], offset); - c4[2] = hc_bytealign (w5[3], w6[0], offset); - c4[1] = hc_bytealign (w5[2], w5[3], offset); - c4[0] = hc_bytealign (w5[1], w5[2], offset); - c3[3] = hc_bytealign (w5[0], w5[1], offset); - c3[2] = hc_bytealign (w4[3], w5[0], offset); - c3[1] = hc_bytealign (w4[2], w4[3], offset); - c3[0] = hc_bytealign (w4[1], w4[2], offset); - c2[3] = hc_bytealign (w4[0], w4[1], offset); - c2[2] = hc_bytealign (w3[3], w4[0], offset); - c2[1] = hc_bytealign (w3[2], w3[3], offset); - c2[0] = hc_bytealign (w3[1], w3[2], offset); - c1[3] = hc_bytealign (w3[0], w3[1], offset); - c1[2] = hc_bytealign (w2[3], w3[0], offset); - c1[1] = hc_bytealign (w2[2], w2[3], offset); - c1[0] = hc_bytealign (w2[1], w2[2], offset); - c0[3] = hc_bytealign (w2[0], w2[1], offset); - c0[2] = hc_bytealign (w1[3], w2[0], offset); - c0[1] = hc_bytealign (w1[2], w1[3], offset); - c0[0] = hc_bytealign (w1[1], w1[2], offset); - w7[3] = hc_bytealign (w1[0], w1[1], offset); - w7[2] = hc_bytealign (w0[3], w1[0], offset); - w7[1] = hc_bytealign (w0[2], w0[3], offset); - w7[0] = hc_bytealign (w0[1], w0[2], offset); - w6[3] = hc_bytealign (w0[0], w0[1], offset); - w6[2] = hc_bytealign ( 0, w0[0], offset); + c6[2] = hc_bytealign_be (w7[3], 0, offset); + c6[1] = hc_bytealign_be (w7[2], w7[3], offset); + c6[0] = hc_bytealign_be (w7[1], w7[2], offset); + c5[3] = hc_bytealign_be (w7[0], w7[1], offset); + c5[2] = hc_bytealign_be (w6[3], w7[0], offset); + c5[1] = hc_bytealign_be (w6[2], w6[3], offset); + c5[0] = hc_bytealign_be (w6[1], w6[2], offset); + c4[3] = hc_bytealign_be (w6[0], w6[1], offset); + c4[2] = hc_bytealign_be (w5[3], w6[0], offset); + c4[1] = hc_bytealign_be (w5[2], w5[3], offset); + c4[0] = hc_bytealign_be (w5[1], w5[2], offset); + c3[3] = hc_bytealign_be (w5[0], w5[1], offset); + c3[2] = hc_bytealign_be (w4[3], w5[0], offset); + c3[1] = hc_bytealign_be (w4[2], w4[3], offset); + c3[0] = hc_bytealign_be (w4[1], w4[2], offset); + c2[3] = hc_bytealign_be (w4[0], w4[1], offset); + c2[2] = hc_bytealign_be (w3[3], w4[0], offset); + c2[1] = hc_bytealign_be (w3[2], w3[3], offset); + c2[0] = hc_bytealign_be (w3[1], w3[2], offset); + c1[3] = hc_bytealign_be (w3[0], w3[1], offset); + c1[2] = hc_bytealign_be (w2[3], w3[0], offset); + c1[1] = hc_bytealign_be (w2[2], w2[3], offset); + c1[0] = hc_bytealign_be (w2[1], w2[2], offset); + c0[3] = hc_bytealign_be (w2[0], w2[1], offset); + c0[2] = hc_bytealign_be (w1[3], w2[0], offset); + c0[1] = hc_bytealign_be (w1[2], w1[3], offset); + c0[0] = hc_bytealign_be (w1[1], w1[2], offset); + w7[3] = hc_bytealign_be (w1[0], w1[1], offset); + w7[2] = hc_bytealign_be (w0[3], w1[0], offset); + w7[1] = hc_bytealign_be (w0[2], w0[3], offset); + w7[0] = hc_bytealign_be (w0[1], w0[2], offset); + w6[3] = hc_bytealign_be (w0[0], w0[1], offset); + w6[2] = hc_bytealign_be ( 0, w0[0], offset); w6[1] = 0; w6[0] = 0; w5[3] = 0; @@ -9831,39 +9677,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 27: - c6[3] = hc_bytealign (w7[3], 0, offset); - c6[2] = hc_bytealign (w7[2], w7[3], offset); - c6[1] = hc_bytealign (w7[1], w7[2], offset); - c6[0] = hc_bytealign (w7[0], w7[1], offset); - c5[3] = hc_bytealign (w6[3], w7[0], offset); - c5[2] = hc_bytealign (w6[2], w6[3], offset); - c5[1] = hc_bytealign (w6[1], w6[2], offset); - c5[0] = hc_bytealign (w6[0], w6[1], offset); - c4[3] = hc_bytealign (w5[3], w6[0], offset); - c4[2] = hc_bytealign (w5[2], w5[3], offset); - c4[1] = hc_bytealign (w5[1], w5[2], offset); - c4[0] = hc_bytealign (w5[0], w5[1], offset); - c3[3] = hc_bytealign (w4[3], w5[0], offset); - c3[2] = hc_bytealign (w4[2], w4[3], offset); - c3[1] = hc_bytealign (w4[1], w4[2], offset); - c3[0] = hc_bytealign (w4[0], w4[1], offset); - c2[3] = hc_bytealign (w3[3], w4[0], offset); - c2[2] = hc_bytealign (w3[2], w3[3], offset); - c2[1] = hc_bytealign (w3[1], w3[2], offset); - c2[0] = hc_bytealign (w3[0], w3[1], offset); - c1[3] = hc_bytealign (w2[3], w3[0], offset); - c1[2] = hc_bytealign (w2[2], w2[3], offset); - c1[1] = hc_bytealign (w2[1], w2[2], offset); - c1[0] = hc_bytealign (w2[0], w2[1], offset); - c0[3] = hc_bytealign (w1[3], w2[0], offset); - c0[2] = hc_bytealign (w1[2], w1[3], offset); - c0[1] = hc_bytealign (w1[1], w1[2], offset); - c0[0] = hc_bytealign (w1[0], w1[1], offset); - w7[3] = hc_bytealign (w0[3], w1[0], offset); - w7[2] = hc_bytealign (w0[2], w0[3], offset); - w7[1] = hc_bytealign (w0[1], w0[2], offset); - w7[0] = hc_bytealign (w0[0], w0[1], offset); - w6[3] = hc_bytealign ( 0, w0[0], offset); + c6[3] = hc_bytealign_be (w7[3], 0, offset); + c6[2] = hc_bytealign_be (w7[2], w7[3], offset); + c6[1] = hc_bytealign_be (w7[1], w7[2], offset); + c6[0] = hc_bytealign_be (w7[0], w7[1], offset); + c5[3] = hc_bytealign_be (w6[3], w7[0], offset); + c5[2] = hc_bytealign_be (w6[2], w6[3], offset); + c5[1] = hc_bytealign_be (w6[1], w6[2], offset); + c5[0] = hc_bytealign_be (w6[0], w6[1], offset); + c4[3] = hc_bytealign_be (w5[3], w6[0], offset); + c4[2] = hc_bytealign_be (w5[2], w5[3], offset); + c4[1] = hc_bytealign_be (w5[1], w5[2], offset); + c4[0] = hc_bytealign_be (w5[0], w5[1], offset); + c3[3] = hc_bytealign_be (w4[3], w5[0], offset); + c3[2] = hc_bytealign_be (w4[2], w4[3], offset); + c3[1] = hc_bytealign_be (w4[1], w4[2], offset); + c3[0] = hc_bytealign_be (w4[0], w4[1], offset); + c2[3] = hc_bytealign_be (w3[3], w4[0], offset); + c2[2] = hc_bytealign_be (w3[2], w3[3], offset); + c2[1] = hc_bytealign_be (w3[1], w3[2], offset); + c2[0] = hc_bytealign_be (w3[0], w3[1], offset); + c1[3] = hc_bytealign_be (w2[3], w3[0], offset); + c1[2] = hc_bytealign_be (w2[2], w2[3], offset); + c1[1] = hc_bytealign_be (w2[1], w2[2], offset); + c1[0] = hc_bytealign_be (w2[0], w2[1], offset); + c0[3] = hc_bytealign_be (w1[3], w2[0], offset); + c0[2] = hc_bytealign_be (w1[2], w1[3], offset); + c0[1] = hc_bytealign_be (w1[1], w1[2], offset); + c0[0] = hc_bytealign_be (w1[0], w1[1], offset); + w7[3] = hc_bytealign_be (w0[3], w1[0], offset); + w7[2] = hc_bytealign_be (w0[2], w0[3], offset); + w7[1] = hc_bytealign_be (w0[1], w0[2], offset); + w7[0] = hc_bytealign_be (w0[0], w0[1], offset); + w6[3] = hc_bytealign_be ( 0, w0[0], offset); w6[2] = 0; w6[1] = 0; w6[0] = 0; @@ -9895,39 +9741,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 28: - c7[0] = hc_bytealign (w7[3], 0, offset); - c6[3] = hc_bytealign (w7[2], w7[3], offset); - c6[2] = hc_bytealign (w7[1], w7[2], offset); - c6[1] = hc_bytealign (w7[0], w7[1], offset); - c6[0] = hc_bytealign (w6[3], w7[0], offset); - c5[3] = hc_bytealign (w6[2], w6[3], offset); - c5[2] = hc_bytealign (w6[1], w6[2], offset); - c5[1] = hc_bytealign (w6[0], w6[1], offset); - c5[0] = hc_bytealign (w5[3], w6[0], offset); - c4[3] = hc_bytealign (w5[2], w5[3], offset); - c4[2] = hc_bytealign (w5[1], w5[2], offset); - c4[1] = hc_bytealign (w5[0], w5[1], offset); - c4[0] = hc_bytealign (w4[3], w5[0], offset); - c3[3] = hc_bytealign (w4[2], w4[3], offset); - c3[2] = hc_bytealign (w4[1], w4[2], offset); - c3[1] = hc_bytealign (w4[0], w4[1], offset); - c3[0] = hc_bytealign (w3[3], w4[0], offset); - c2[3] = hc_bytealign (w3[2], w3[3], offset); - c2[2] = hc_bytealign (w3[1], w3[2], offset); - c2[1] = hc_bytealign (w3[0], w3[1], offset); - c2[0] = hc_bytealign (w2[3], w3[0], offset); - c1[3] = hc_bytealign (w2[2], w2[3], offset); - c1[2] = hc_bytealign (w2[1], w2[2], offset); - c1[1] = hc_bytealign (w2[0], w2[1], offset); - c1[0] = hc_bytealign (w1[3], w2[0], offset); - c0[3] = hc_bytealign (w1[2], w1[3], offset); - c0[2] = hc_bytealign (w1[1], w1[2], offset); - c0[1] = hc_bytealign (w1[0], w1[1], offset); - c0[0] = hc_bytealign (w0[3], w1[0], offset); - w7[3] = hc_bytealign (w0[2], w0[3], offset); - w7[2] = hc_bytealign (w0[1], w0[2], offset); - w7[1] = hc_bytealign (w0[0], w0[1], offset); - w7[0] = hc_bytealign ( 0, w0[0], offset); + c7[0] = hc_bytealign_be (w7[3], 0, offset); + c6[3] = hc_bytealign_be (w7[2], w7[3], offset); + c6[2] = hc_bytealign_be (w7[1], w7[2], offset); + c6[1] = hc_bytealign_be (w7[0], w7[1], offset); + c6[0] = hc_bytealign_be (w6[3], w7[0], offset); + c5[3] = hc_bytealign_be (w6[2], w6[3], offset); + c5[2] = hc_bytealign_be (w6[1], w6[2], offset); + c5[1] = hc_bytealign_be (w6[0], w6[1], offset); + c5[0] = hc_bytealign_be (w5[3], w6[0], offset); + c4[3] = hc_bytealign_be (w5[2], w5[3], offset); + c4[2] = hc_bytealign_be (w5[1], w5[2], offset); + c4[1] = hc_bytealign_be (w5[0], w5[1], offset); + c4[0] = hc_bytealign_be (w4[3], w5[0], offset); + c3[3] = hc_bytealign_be (w4[2], w4[3], offset); + c3[2] = hc_bytealign_be (w4[1], w4[2], offset); + c3[1] = hc_bytealign_be (w4[0], w4[1], offset); + c3[0] = hc_bytealign_be (w3[3], w4[0], offset); + c2[3] = hc_bytealign_be (w3[2], w3[3], offset); + c2[2] = hc_bytealign_be (w3[1], w3[2], offset); + c2[1] = hc_bytealign_be (w3[0], w3[1], offset); + c2[0] = hc_bytealign_be (w2[3], w3[0], offset); + c1[3] = hc_bytealign_be (w2[2], w2[3], offset); + c1[2] = hc_bytealign_be (w2[1], w2[2], offset); + c1[1] = hc_bytealign_be (w2[0], w2[1], offset); + c1[0] = hc_bytealign_be (w1[3], w2[0], offset); + c0[3] = hc_bytealign_be (w1[2], w1[3], offset); + c0[2] = hc_bytealign_be (w1[1], w1[2], offset); + c0[1] = hc_bytealign_be (w1[0], w1[1], offset); + c0[0] = hc_bytealign_be (w0[3], w1[0], offset); + w7[3] = hc_bytealign_be (w0[2], w0[3], offset); + w7[2] = hc_bytealign_be (w0[1], w0[2], offset); + w7[1] = hc_bytealign_be (w0[0], w0[1], offset); + w7[0] = hc_bytealign_be ( 0, w0[0], offset); w6[3] = 0; w6[2] = 0; w6[1] = 0; @@ -9960,39 +9806,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 29: - c7[1] = hc_bytealign (w7[3], 0, offset); - c7[0] = hc_bytealign (w7[2], w7[3], offset); - c6[3] = hc_bytealign (w7[1], w7[2], offset); - c6[2] = hc_bytealign (w7[0], w7[1], offset); - c6[1] = hc_bytealign (w6[3], w7[0], offset); - c6[0] = hc_bytealign (w6[2], w6[3], offset); - c5[3] = hc_bytealign (w6[1], w6[2], offset); - c5[2] = hc_bytealign (w6[0], w6[1], offset); - c5[1] = hc_bytealign (w5[3], w6[0], offset); - c5[0] = hc_bytealign (w5[2], w5[3], offset); - c4[3] = hc_bytealign (w5[1], w5[2], offset); - c4[2] = hc_bytealign (w5[0], w5[1], offset); - c4[1] = hc_bytealign (w4[3], w5[0], offset); - c4[0] = hc_bytealign (w4[2], w4[3], offset); - c3[3] = hc_bytealign (w4[1], w4[2], offset); - c3[2] = hc_bytealign (w4[0], w4[1], offset); - c3[1] = hc_bytealign (w3[3], w4[0], offset); - c3[0] = hc_bytealign (w3[2], w3[3], offset); - c2[3] = hc_bytealign (w3[1], w3[2], offset); - c2[2] = hc_bytealign (w3[0], w3[1], offset); - c2[1] = hc_bytealign (w2[3], w3[0], offset); - c2[0] = hc_bytealign (w2[2], w2[3], offset); - c1[3] = hc_bytealign (w2[1], w2[2], offset); - c1[2] = hc_bytealign (w2[0], w2[1], offset); - c1[1] = hc_bytealign (w1[3], w2[0], offset); - c1[0] = hc_bytealign (w1[2], w1[3], offset); - c0[3] = hc_bytealign (w1[1], w1[2], offset); - c0[2] = hc_bytealign (w1[0], w1[1], offset); - c0[1] = hc_bytealign (w0[3], w1[0], offset); - c0[0] = hc_bytealign (w0[2], w0[3], offset); - w7[3] = hc_bytealign (w0[1], w0[2], offset); - w7[2] = hc_bytealign (w0[0], w0[1], offset); - w7[1] = hc_bytealign ( 0, w0[0], offset); + c7[1] = hc_bytealign_be (w7[3], 0, offset); + c7[0] = hc_bytealign_be (w7[2], w7[3], offset); + c6[3] = hc_bytealign_be (w7[1], w7[2], offset); + c6[2] = hc_bytealign_be (w7[0], w7[1], offset); + c6[1] = hc_bytealign_be (w6[3], w7[0], offset); + c6[0] = hc_bytealign_be (w6[2], w6[3], offset); + c5[3] = hc_bytealign_be (w6[1], w6[2], offset); + c5[2] = hc_bytealign_be (w6[0], w6[1], offset); + c5[1] = hc_bytealign_be (w5[3], w6[0], offset); + c5[0] = hc_bytealign_be (w5[2], w5[3], offset); + c4[3] = hc_bytealign_be (w5[1], w5[2], offset); + c4[2] = hc_bytealign_be (w5[0], w5[1], offset); + c4[1] = hc_bytealign_be (w4[3], w5[0], offset); + c4[0] = hc_bytealign_be (w4[2], w4[3], offset); + c3[3] = hc_bytealign_be (w4[1], w4[2], offset); + c3[2] = hc_bytealign_be (w4[0], w4[1], offset); + c3[1] = hc_bytealign_be (w3[3], w4[0], offset); + c3[0] = hc_bytealign_be (w3[2], w3[3], offset); + c2[3] = hc_bytealign_be (w3[1], w3[2], offset); + c2[2] = hc_bytealign_be (w3[0], w3[1], offset); + c2[1] = hc_bytealign_be (w2[3], w3[0], offset); + c2[0] = hc_bytealign_be (w2[2], w2[3], offset); + c1[3] = hc_bytealign_be (w2[1], w2[2], offset); + c1[2] = hc_bytealign_be (w2[0], w2[1], offset); + c1[1] = hc_bytealign_be (w1[3], w2[0], offset); + c1[0] = hc_bytealign_be (w1[2], w1[3], offset); + c0[3] = hc_bytealign_be (w1[1], w1[2], offset); + c0[2] = hc_bytealign_be (w1[0], w1[1], offset); + c0[1] = hc_bytealign_be (w0[3], w1[0], offset); + c0[0] = hc_bytealign_be (w0[2], w0[3], offset); + w7[3] = hc_bytealign_be (w0[1], w0[2], offset); + w7[2] = hc_bytealign_be (w0[0], w0[1], offset); + w7[1] = hc_bytealign_be ( 0, w0[0], offset); w7[0] = 0; w6[3] = 0; w6[2] = 0; @@ -10026,39 +9872,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 30: - c7[2] = hc_bytealign (w7[3], 0, offset); - c7[1] = hc_bytealign (w7[2], w7[3], offset); - c7[0] = hc_bytealign (w7[1], w7[2], offset); - c6[3] = hc_bytealign (w7[0], w7[1], offset); - c6[2] = hc_bytealign (w6[3], w7[0], offset); - c6[1] = hc_bytealign (w6[2], w6[3], offset); - c6[0] = hc_bytealign (w6[1], w6[2], offset); - c5[3] = hc_bytealign (w6[0], w6[1], offset); - c5[2] = hc_bytealign (w5[3], w6[0], offset); - c5[1] = hc_bytealign (w5[2], w5[3], offset); - c5[0] = hc_bytealign (w5[1], w5[2], offset); - c4[3] = hc_bytealign (w5[0], w5[1], offset); - c4[2] = hc_bytealign (w4[3], w5[0], offset); - c4[1] = hc_bytealign (w4[2], w4[3], offset); - c4[0] = hc_bytealign (w4[1], w4[2], offset); - c3[3] = hc_bytealign (w4[0], w4[1], offset); - c3[2] = hc_bytealign (w3[3], w4[0], offset); - c3[1] = hc_bytealign (w3[2], w3[3], offset); - c3[0] = hc_bytealign (w3[1], w3[2], offset); - c2[3] = hc_bytealign (w3[0], w3[1], offset); - c2[2] = hc_bytealign (w2[3], w3[0], offset); - c2[1] = hc_bytealign (w2[2], w2[3], offset); - c2[0] = hc_bytealign (w2[1], w2[2], offset); - c1[3] = hc_bytealign (w2[0], w2[1], offset); - c1[2] = hc_bytealign (w1[3], w2[0], offset); - c1[1] = hc_bytealign (w1[2], w1[3], offset); - c1[0] = hc_bytealign (w1[1], w1[2], offset); - c0[3] = hc_bytealign (w1[0], w1[1], offset); - c0[2] = hc_bytealign (w0[3], w1[0], offset); - c0[1] = hc_bytealign (w0[2], w0[3], offset); - c0[0] = hc_bytealign (w0[1], w0[2], offset); - w7[3] = hc_bytealign (w0[0], w0[1], offset); - w7[2] = hc_bytealign ( 0, w0[0], offset); + c7[2] = hc_bytealign_be (w7[3], 0, offset); + c7[1] = hc_bytealign_be (w7[2], w7[3], offset); + c7[0] = hc_bytealign_be (w7[1], w7[2], offset); + c6[3] = hc_bytealign_be (w7[0], w7[1], offset); + c6[2] = hc_bytealign_be (w6[3], w7[0], offset); + c6[1] = hc_bytealign_be (w6[2], w6[3], offset); + c6[0] = hc_bytealign_be (w6[1], w6[2], offset); + c5[3] = hc_bytealign_be (w6[0], w6[1], offset); + c5[2] = hc_bytealign_be (w5[3], w6[0], offset); + c5[1] = hc_bytealign_be (w5[2], w5[3], offset); + c5[0] = hc_bytealign_be (w5[1], w5[2], offset); + c4[3] = hc_bytealign_be (w5[0], w5[1], offset); + c4[2] = hc_bytealign_be (w4[3], w5[0], offset); + c4[1] = hc_bytealign_be (w4[2], w4[3], offset); + c4[0] = hc_bytealign_be (w4[1], w4[2], offset); + c3[3] = hc_bytealign_be (w4[0], w4[1], offset); + c3[2] = hc_bytealign_be (w3[3], w4[0], offset); + c3[1] = hc_bytealign_be (w3[2], w3[3], offset); + c3[0] = hc_bytealign_be (w3[1], w3[2], offset); + c2[3] = hc_bytealign_be (w3[0], w3[1], offset); + c2[2] = hc_bytealign_be (w2[3], w3[0], offset); + c2[1] = hc_bytealign_be (w2[2], w2[3], offset); + c2[0] = hc_bytealign_be (w2[1], w2[2], offset); + c1[3] = hc_bytealign_be (w2[0], w2[1], offset); + c1[2] = hc_bytealign_be (w1[3], w2[0], offset); + c1[1] = hc_bytealign_be (w1[2], w1[3], offset); + c1[0] = hc_bytealign_be (w1[1], w1[2], offset); + c0[3] = hc_bytealign_be (w1[0], w1[1], offset); + c0[2] = hc_bytealign_be (w0[3], w1[0], offset); + c0[1] = hc_bytealign_be (w0[2], w0[3], offset); + c0[0] = hc_bytealign_be (w0[1], w0[2], offset); + w7[3] = hc_bytealign_be (w0[0], w0[1], offset); + w7[2] = hc_bytealign_be ( 0, w0[0], offset); w7[1] = 0; w7[0] = 0; w6[3] = 0; @@ -10093,39 +9939,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be (u32x *w0, u32x *w1, u32x *w2 break; case 31: - c7[3] = hc_bytealign (w7[3], 0, offset); - c7[2] = hc_bytealign (w7[2], w7[3], offset); - c7[1] = hc_bytealign (w7[1], w7[2], offset); - c7[0] = hc_bytealign (w7[0], w7[1], offset); - c6[3] = hc_bytealign (w6[3], w7[0], offset); - c6[2] = hc_bytealign (w6[2], w6[3], offset); - c6[1] = hc_bytealign (w6[1], w6[2], offset); - c6[0] = hc_bytealign (w6[0], w6[1], offset); - c5[3] = hc_bytealign (w5[3], w6[0], offset); - c5[2] = hc_bytealign (w5[2], w5[3], offset); - c5[1] = hc_bytealign (w5[1], w5[2], offset); - c5[0] = hc_bytealign (w5[0], w5[1], offset); - c4[3] = hc_bytealign (w4[3], w5[0], offset); - c4[2] = hc_bytealign (w4[2], w4[3], offset); - c4[1] = hc_bytealign (w4[1], w4[2], offset); - c4[0] = hc_bytealign (w4[0], w4[1], offset); - c3[3] = hc_bytealign (w3[3], w4[0], offset); - c3[2] = hc_bytealign (w3[2], w3[3], offset); - c3[1] = hc_bytealign (w3[1], w3[2], offset); - c3[0] = hc_bytealign (w3[0], w3[1], offset); - c2[3] = hc_bytealign (w2[3], w3[0], offset); - c2[2] = hc_bytealign (w2[2], w2[3], offset); - c2[1] = hc_bytealign (w2[1], w2[2], offset); - c2[0] = hc_bytealign (w2[0], w2[1], offset); - c1[3] = hc_bytealign (w1[3], w2[0], offset); - c1[2] = hc_bytealign (w1[2], w1[3], offset); - c1[1] = hc_bytealign (w1[1], w1[2], offset); - c1[0] = hc_bytealign (w1[0], w1[1], offset); - c0[3] = hc_bytealign (w0[3], w1[0], offset); - c0[2] = hc_bytealign (w0[2], w0[3], offset); - c0[1] = hc_bytealign (w0[1], w0[2], offset); - c0[0] = hc_bytealign (w0[0], w0[1], offset); - w7[3] = hc_bytealign ( 0, w0[0], offset); + c7[3] = hc_bytealign_be (w7[3], 0, offset); + c7[2] = hc_bytealign_be (w7[2], w7[3], offset); + c7[1] = hc_bytealign_be (w7[1], w7[2], offset); + c7[0] = hc_bytealign_be (w7[0], w7[1], offset); + c6[3] = hc_bytealign_be (w6[3], w7[0], offset); + c6[2] = hc_bytealign_be (w6[2], w6[3], offset); + c6[1] = hc_bytealign_be (w6[1], w6[2], offset); + c6[0] = hc_bytealign_be (w6[0], w6[1], offset); + c5[3] = hc_bytealign_be (w5[3], w6[0], offset); + c5[2] = hc_bytealign_be (w5[2], w5[3], offset); + c5[1] = hc_bytealign_be (w5[1], w5[2], offset); + c5[0] = hc_bytealign_be (w5[0], w5[1], offset); + c4[3] = hc_bytealign_be (w4[3], w5[0], offset); + c4[2] = hc_bytealign_be (w4[2], w4[3], offset); + c4[1] = hc_bytealign_be (w4[1], w4[2], offset); + c4[0] = hc_bytealign_be (w4[0], w4[1], offset); + c3[3] = hc_bytealign_be (w3[3], w4[0], offset); + c3[2] = hc_bytealign_be (w3[2], w3[3], offset); + c3[1] = hc_bytealign_be (w3[1], w3[2], offset); + c3[0] = hc_bytealign_be (w3[0], w3[1], offset); + c2[3] = hc_bytealign_be (w2[3], w3[0], offset); + c2[2] = hc_bytealign_be (w2[2], w2[3], offset); + c2[1] = hc_bytealign_be (w2[1], w2[2], offset); + c2[0] = hc_bytealign_be (w2[0], w2[1], offset); + c1[3] = hc_bytealign_be (w1[3], w2[0], offset); + c1[2] = hc_bytealign_be (w1[2], w1[3], offset); + c1[1] = hc_bytealign_be (w1[1], w1[2], offset); + c1[0] = hc_bytealign_be (w1[0], w1[1], offset); + c0[3] = hc_bytealign_be (w0[3], w1[0], offset); + c0[2] = hc_bytealign_be (w0[2], w0[3], offset); + c0[1] = hc_bytealign_be (w0[1], w0[2], offset); + c0[0] = hc_bytealign_be (w0[0], w0[1], offset); + w7[3] = hc_bytealign_be ( 0, w0[0], offset); w7[2] = 0; w7[1] = 0; w7[0] = 0; @@ -11866,10 +11712,6 @@ DECLSPEC void switch_buffer_by_offset_1x64_le (u32x *w, const u32 offset) const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - - #pragma unroll - for (int i = 0; i < 64; i++) w[i] = swap32 (w[i]); - switch (offset_switch) { case 0: @@ -16224,10 +16066,6 @@ DECLSPEC void switch_buffer_by_offset_1x64_le (u32x *w, const u32 offset) break; } - - #pragma unroll - for (int i = 0; i < 64; i++) w[i] = swap32 (w[i]); - #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -20605,271 +20443,271 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) switch (offset_switch) { case 0: - w[63] = hc_bytealign (w[62], w[63], offset); - w[62] = hc_bytealign (w[61], w[62], offset); - w[61] = hc_bytealign (w[60], w[61], offset); - w[60] = hc_bytealign (w[59], w[60], offset); - w[59] = hc_bytealign (w[58], w[59], offset); - w[58] = hc_bytealign (w[57], w[58], offset); - w[57] = hc_bytealign (w[56], w[57], offset); - w[56] = hc_bytealign (w[55], w[56], offset); - w[55] = hc_bytealign (w[54], w[55], offset); - w[54] = hc_bytealign (w[53], w[54], offset); - w[53] = hc_bytealign (w[52], w[53], offset); - w[52] = hc_bytealign (w[51], w[52], offset); - w[51] = hc_bytealign (w[50], w[51], offset); - w[50] = hc_bytealign (w[49], w[50], offset); - w[49] = hc_bytealign (w[48], w[49], offset); - w[48] = hc_bytealign (w[47], w[48], offset); - w[47] = hc_bytealign (w[46], w[47], offset); - w[46] = hc_bytealign (w[45], w[46], offset); - w[45] = hc_bytealign (w[44], w[45], offset); - w[44] = hc_bytealign (w[43], w[44], offset); - w[43] = hc_bytealign (w[42], w[43], offset); - w[42] = hc_bytealign (w[41], w[42], offset); - w[41] = hc_bytealign (w[40], w[41], offset); - w[40] = hc_bytealign (w[39], w[40], offset); - w[39] = hc_bytealign (w[38], w[39], offset); - w[38] = hc_bytealign (w[37], w[38], offset); - w[37] = hc_bytealign (w[36], w[37], offset); - w[36] = hc_bytealign (w[35], w[36], offset); - w[35] = hc_bytealign (w[34], w[35], offset); - w[34] = hc_bytealign (w[33], w[34], offset); - w[33] = hc_bytealign (w[32], w[33], offset); - w[32] = hc_bytealign (w[31], w[32], offset); - w[31] = hc_bytealign (w[30], w[31], offset); - w[30] = hc_bytealign (w[29], w[30], offset); - w[29] = hc_bytealign (w[28], w[29], offset); - w[28] = hc_bytealign (w[27], w[28], offset); - w[27] = hc_bytealign (w[26], w[27], offset); - w[26] = hc_bytealign (w[25], w[26], offset); - w[25] = hc_bytealign (w[24], w[25], offset); - w[24] = hc_bytealign (w[23], w[24], offset); - w[23] = hc_bytealign (w[22], w[23], offset); - w[22] = hc_bytealign (w[21], w[22], offset); - w[21] = hc_bytealign (w[20], w[21], offset); - w[20] = hc_bytealign (w[19], w[20], offset); - w[19] = hc_bytealign (w[18], w[19], offset); - w[18] = hc_bytealign (w[17], w[18], offset); - w[17] = hc_bytealign (w[16], w[17], offset); - w[16] = hc_bytealign (w[15], w[16], offset); - w[15] = hc_bytealign (w[14], w[15], offset); - w[14] = hc_bytealign (w[13], w[14], offset); - w[13] = hc_bytealign (w[12], w[13], offset); - w[12] = hc_bytealign (w[11], w[12], offset); - w[11] = hc_bytealign (w[10], w[11], offset); - w[10] = hc_bytealign (w[ 9], w[10], offset); - w[ 9] = hc_bytealign (w[ 8], w[ 9], offset); - w[ 8] = hc_bytealign (w[ 7], w[ 8], offset); - w[ 7] = hc_bytealign (w[ 6], w[ 7], offset); - w[ 6] = hc_bytealign (w[ 5], w[ 6], offset); - w[ 5] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 4] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 3] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 2] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 1] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 0] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[62], w[63], offset); + w[62] = hc_bytealign_be (w[61], w[62], offset); + w[61] = hc_bytealign_be (w[60], w[61], offset); + w[60] = hc_bytealign_be (w[59], w[60], offset); + w[59] = hc_bytealign_be (w[58], w[59], offset); + w[58] = hc_bytealign_be (w[57], w[58], offset); + w[57] = hc_bytealign_be (w[56], w[57], offset); + w[56] = hc_bytealign_be (w[55], w[56], offset); + w[55] = hc_bytealign_be (w[54], w[55], offset); + w[54] = hc_bytealign_be (w[53], w[54], offset); + w[53] = hc_bytealign_be (w[52], w[53], offset); + w[52] = hc_bytealign_be (w[51], w[52], offset); + w[51] = hc_bytealign_be (w[50], w[51], offset); + w[50] = hc_bytealign_be (w[49], w[50], offset); + w[49] = hc_bytealign_be (w[48], w[49], offset); + w[48] = hc_bytealign_be (w[47], w[48], offset); + w[47] = hc_bytealign_be (w[46], w[47], offset); + w[46] = hc_bytealign_be (w[45], w[46], offset); + w[45] = hc_bytealign_be (w[44], w[45], offset); + w[44] = hc_bytealign_be (w[43], w[44], offset); + w[43] = hc_bytealign_be (w[42], w[43], offset); + w[42] = hc_bytealign_be (w[41], w[42], offset); + w[41] = hc_bytealign_be (w[40], w[41], offset); + w[40] = hc_bytealign_be (w[39], w[40], offset); + w[39] = hc_bytealign_be (w[38], w[39], offset); + w[38] = hc_bytealign_be (w[37], w[38], offset); + w[37] = hc_bytealign_be (w[36], w[37], offset); + w[36] = hc_bytealign_be (w[35], w[36], offset); + w[35] = hc_bytealign_be (w[34], w[35], offset); + w[34] = hc_bytealign_be (w[33], w[34], offset); + w[33] = hc_bytealign_be (w[32], w[33], offset); + w[32] = hc_bytealign_be (w[31], w[32], offset); + w[31] = hc_bytealign_be (w[30], w[31], offset); + w[30] = hc_bytealign_be (w[29], w[30], offset); + w[29] = hc_bytealign_be (w[28], w[29], offset); + w[28] = hc_bytealign_be (w[27], w[28], offset); + w[27] = hc_bytealign_be (w[26], w[27], offset); + w[26] = hc_bytealign_be (w[25], w[26], offset); + w[25] = hc_bytealign_be (w[24], w[25], offset); + w[24] = hc_bytealign_be (w[23], w[24], offset); + w[23] = hc_bytealign_be (w[22], w[23], offset); + w[22] = hc_bytealign_be (w[21], w[22], offset); + w[21] = hc_bytealign_be (w[20], w[21], offset); + w[20] = hc_bytealign_be (w[19], w[20], offset); + w[19] = hc_bytealign_be (w[18], w[19], offset); + w[18] = hc_bytealign_be (w[17], w[18], offset); + w[17] = hc_bytealign_be (w[16], w[17], offset); + w[16] = hc_bytealign_be (w[15], w[16], offset); + w[15] = hc_bytealign_be (w[14], w[15], offset); + w[14] = hc_bytealign_be (w[13], w[14], offset); + w[13] = hc_bytealign_be (w[12], w[13], offset); + w[12] = hc_bytealign_be (w[11], w[12], offset); + w[11] = hc_bytealign_be (w[10], w[11], offset); + w[10] = hc_bytealign_be (w[ 9], w[10], offset); + w[ 9] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[ 8] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[ 7] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[ 6] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[ 5] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 4] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 3] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 2] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 1] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 0] = hc_bytealign_be ( 0, w[ 0], offset); break; case 1: - w[63] = hc_bytealign (w[61], w[62], offset); - w[62] = hc_bytealign (w[60], w[61], offset); - w[61] = hc_bytealign (w[59], w[60], offset); - w[60] = hc_bytealign (w[58], w[59], offset); - w[59] = hc_bytealign (w[57], w[58], offset); - w[58] = hc_bytealign (w[56], w[57], offset); - w[57] = hc_bytealign (w[55], w[56], offset); - w[56] = hc_bytealign (w[54], w[55], offset); - w[55] = hc_bytealign (w[53], w[54], offset); - w[54] = hc_bytealign (w[52], w[53], offset); - w[53] = hc_bytealign (w[51], w[52], offset); - w[52] = hc_bytealign (w[50], w[51], offset); - w[51] = hc_bytealign (w[49], w[50], offset); - w[50] = hc_bytealign (w[48], w[49], offset); - w[49] = hc_bytealign (w[47], w[48], offset); - w[48] = hc_bytealign (w[46], w[47], offset); - w[47] = hc_bytealign (w[45], w[46], offset); - w[46] = hc_bytealign (w[44], w[45], offset); - w[45] = hc_bytealign (w[43], w[44], offset); - w[44] = hc_bytealign (w[42], w[43], offset); - w[43] = hc_bytealign (w[41], w[42], offset); - w[42] = hc_bytealign (w[40], w[41], offset); - w[41] = hc_bytealign (w[39], w[40], offset); - w[40] = hc_bytealign (w[38], w[39], offset); - w[39] = hc_bytealign (w[37], w[38], offset); - w[38] = hc_bytealign (w[36], w[37], offset); - w[37] = hc_bytealign (w[35], w[36], offset); - w[36] = hc_bytealign (w[34], w[35], offset); - w[35] = hc_bytealign (w[33], w[34], offset); - w[34] = hc_bytealign (w[32], w[33], offset); - w[33] = hc_bytealign (w[31], w[32], offset); - w[32] = hc_bytealign (w[30], w[31], offset); - w[31] = hc_bytealign (w[29], w[30], offset); - w[30] = hc_bytealign (w[28], w[29], offset); - w[29] = hc_bytealign (w[27], w[28], offset); - w[28] = hc_bytealign (w[26], w[27], offset); - w[27] = hc_bytealign (w[25], w[26], offset); - w[26] = hc_bytealign (w[24], w[25], offset); - w[25] = hc_bytealign (w[23], w[24], offset); - w[24] = hc_bytealign (w[22], w[23], offset); - w[23] = hc_bytealign (w[21], w[22], offset); - w[22] = hc_bytealign (w[20], w[21], offset); - w[21] = hc_bytealign (w[19], w[20], offset); - w[20] = hc_bytealign (w[18], w[19], offset); - w[19] = hc_bytealign (w[17], w[18], offset); - w[18] = hc_bytealign (w[16], w[17], offset); - w[17] = hc_bytealign (w[15], w[16], offset); - w[16] = hc_bytealign (w[14], w[15], offset); - w[15] = hc_bytealign (w[13], w[14], offset); - w[14] = hc_bytealign (w[12], w[13], offset); - w[13] = hc_bytealign (w[11], w[12], offset); - w[12] = hc_bytealign (w[10], w[11], offset); - w[11] = hc_bytealign (w[ 9], w[10], offset); - w[10] = hc_bytealign (w[ 8], w[ 9], offset); - w[ 9] = hc_bytealign (w[ 7], w[ 8], offset); - w[ 8] = hc_bytealign (w[ 6], w[ 7], offset); - w[ 7] = hc_bytealign (w[ 5], w[ 6], offset); - w[ 6] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 5] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 4] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 3] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 2] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 1] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[61], w[62], offset); + w[62] = hc_bytealign_be (w[60], w[61], offset); + w[61] = hc_bytealign_be (w[59], w[60], offset); + w[60] = hc_bytealign_be (w[58], w[59], offset); + w[59] = hc_bytealign_be (w[57], w[58], offset); + w[58] = hc_bytealign_be (w[56], w[57], offset); + w[57] = hc_bytealign_be (w[55], w[56], offset); + w[56] = hc_bytealign_be (w[54], w[55], offset); + w[55] = hc_bytealign_be (w[53], w[54], offset); + w[54] = hc_bytealign_be (w[52], w[53], offset); + w[53] = hc_bytealign_be (w[51], w[52], offset); + w[52] = hc_bytealign_be (w[50], w[51], offset); + w[51] = hc_bytealign_be (w[49], w[50], offset); + w[50] = hc_bytealign_be (w[48], w[49], offset); + w[49] = hc_bytealign_be (w[47], w[48], offset); + w[48] = hc_bytealign_be (w[46], w[47], offset); + w[47] = hc_bytealign_be (w[45], w[46], offset); + w[46] = hc_bytealign_be (w[44], w[45], offset); + w[45] = hc_bytealign_be (w[43], w[44], offset); + w[44] = hc_bytealign_be (w[42], w[43], offset); + w[43] = hc_bytealign_be (w[41], w[42], offset); + w[42] = hc_bytealign_be (w[40], w[41], offset); + w[41] = hc_bytealign_be (w[39], w[40], offset); + w[40] = hc_bytealign_be (w[38], w[39], offset); + w[39] = hc_bytealign_be (w[37], w[38], offset); + w[38] = hc_bytealign_be (w[36], w[37], offset); + w[37] = hc_bytealign_be (w[35], w[36], offset); + w[36] = hc_bytealign_be (w[34], w[35], offset); + w[35] = hc_bytealign_be (w[33], w[34], offset); + w[34] = hc_bytealign_be (w[32], w[33], offset); + w[33] = hc_bytealign_be (w[31], w[32], offset); + w[32] = hc_bytealign_be (w[30], w[31], offset); + w[31] = hc_bytealign_be (w[29], w[30], offset); + w[30] = hc_bytealign_be (w[28], w[29], offset); + w[29] = hc_bytealign_be (w[27], w[28], offset); + w[28] = hc_bytealign_be (w[26], w[27], offset); + w[27] = hc_bytealign_be (w[25], w[26], offset); + w[26] = hc_bytealign_be (w[24], w[25], offset); + w[25] = hc_bytealign_be (w[23], w[24], offset); + w[24] = hc_bytealign_be (w[22], w[23], offset); + w[23] = hc_bytealign_be (w[21], w[22], offset); + w[22] = hc_bytealign_be (w[20], w[21], offset); + w[21] = hc_bytealign_be (w[19], w[20], offset); + w[20] = hc_bytealign_be (w[18], w[19], offset); + w[19] = hc_bytealign_be (w[17], w[18], offset); + w[18] = hc_bytealign_be (w[16], w[17], offset); + w[17] = hc_bytealign_be (w[15], w[16], offset); + w[16] = hc_bytealign_be (w[14], w[15], offset); + w[15] = hc_bytealign_be (w[13], w[14], offset); + w[14] = hc_bytealign_be (w[12], w[13], offset); + w[13] = hc_bytealign_be (w[11], w[12], offset); + w[12] = hc_bytealign_be (w[10], w[11], offset); + w[11] = hc_bytealign_be (w[ 9], w[10], offset); + w[10] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[ 9] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[ 8] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[ 7] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[ 6] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 5] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 4] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 3] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 2] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 1] = hc_bytealign_be ( 0, w[ 0], offset); w[ 0] = 0; break; case 2: - w[63] = hc_bytealign (w[60], w[61], offset); - w[62] = hc_bytealign (w[59], w[60], offset); - w[61] = hc_bytealign (w[58], w[59], offset); - w[60] = hc_bytealign (w[57], w[58], offset); - w[59] = hc_bytealign (w[56], w[57], offset); - w[58] = hc_bytealign (w[55], w[56], offset); - w[57] = hc_bytealign (w[54], w[55], offset); - w[56] = hc_bytealign (w[53], w[54], offset); - w[55] = hc_bytealign (w[52], w[53], offset); - w[54] = hc_bytealign (w[51], w[52], offset); - w[53] = hc_bytealign (w[50], w[51], offset); - w[52] = hc_bytealign (w[49], w[50], offset); - w[51] = hc_bytealign (w[48], w[49], offset); - w[50] = hc_bytealign (w[47], w[48], offset); - w[49] = hc_bytealign (w[46], w[47], offset); - w[48] = hc_bytealign (w[45], w[46], offset); - w[47] = hc_bytealign (w[44], w[45], offset); - w[46] = hc_bytealign (w[43], w[44], offset); - w[45] = hc_bytealign (w[42], w[43], offset); - w[44] = hc_bytealign (w[41], w[42], offset); - w[43] = hc_bytealign (w[40], w[41], offset); - w[42] = hc_bytealign (w[39], w[40], offset); - w[41] = hc_bytealign (w[38], w[39], offset); - w[40] = hc_bytealign (w[37], w[38], offset); - w[39] = hc_bytealign (w[36], w[37], offset); - w[38] = hc_bytealign (w[35], w[36], offset); - w[37] = hc_bytealign (w[34], w[35], offset); - w[36] = hc_bytealign (w[33], w[34], offset); - w[35] = hc_bytealign (w[32], w[33], offset); - w[34] = hc_bytealign (w[31], w[32], offset); - w[33] = hc_bytealign (w[30], w[31], offset); - w[32] = hc_bytealign (w[29], w[30], offset); - w[31] = hc_bytealign (w[28], w[29], offset); - w[30] = hc_bytealign (w[27], w[28], offset); - w[29] = hc_bytealign (w[26], w[27], offset); - w[28] = hc_bytealign (w[25], w[26], offset); - w[27] = hc_bytealign (w[24], w[25], offset); - w[26] = hc_bytealign (w[23], w[24], offset); - w[25] = hc_bytealign (w[22], w[23], offset); - w[24] = hc_bytealign (w[21], w[22], offset); - w[23] = hc_bytealign (w[20], w[21], offset); - w[22] = hc_bytealign (w[19], w[20], offset); - w[21] = hc_bytealign (w[18], w[19], offset); - w[20] = hc_bytealign (w[17], w[18], offset); - w[19] = hc_bytealign (w[16], w[17], offset); - w[18] = hc_bytealign (w[15], w[16], offset); - w[17] = hc_bytealign (w[14], w[15], offset); - w[16] = hc_bytealign (w[13], w[14], offset); - w[15] = hc_bytealign (w[12], w[13], offset); - w[14] = hc_bytealign (w[11], w[12], offset); - w[13] = hc_bytealign (w[10], w[11], offset); - w[12] = hc_bytealign (w[ 9], w[10], offset); - w[11] = hc_bytealign (w[ 8], w[ 9], offset); - w[10] = hc_bytealign (w[ 7], w[ 8], offset); - w[ 9] = hc_bytealign (w[ 6], w[ 7], offset); - w[ 8] = hc_bytealign (w[ 5], w[ 6], offset); - w[ 7] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 6] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 5] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 4] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 3] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 2] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[60], w[61], offset); + w[62] = hc_bytealign_be (w[59], w[60], offset); + w[61] = hc_bytealign_be (w[58], w[59], offset); + w[60] = hc_bytealign_be (w[57], w[58], offset); + w[59] = hc_bytealign_be (w[56], w[57], offset); + w[58] = hc_bytealign_be (w[55], w[56], offset); + w[57] = hc_bytealign_be (w[54], w[55], offset); + w[56] = hc_bytealign_be (w[53], w[54], offset); + w[55] = hc_bytealign_be (w[52], w[53], offset); + w[54] = hc_bytealign_be (w[51], w[52], offset); + w[53] = hc_bytealign_be (w[50], w[51], offset); + w[52] = hc_bytealign_be (w[49], w[50], offset); + w[51] = hc_bytealign_be (w[48], w[49], offset); + w[50] = hc_bytealign_be (w[47], w[48], offset); + w[49] = hc_bytealign_be (w[46], w[47], offset); + w[48] = hc_bytealign_be (w[45], w[46], offset); + w[47] = hc_bytealign_be (w[44], w[45], offset); + w[46] = hc_bytealign_be (w[43], w[44], offset); + w[45] = hc_bytealign_be (w[42], w[43], offset); + w[44] = hc_bytealign_be (w[41], w[42], offset); + w[43] = hc_bytealign_be (w[40], w[41], offset); + w[42] = hc_bytealign_be (w[39], w[40], offset); + w[41] = hc_bytealign_be (w[38], w[39], offset); + w[40] = hc_bytealign_be (w[37], w[38], offset); + w[39] = hc_bytealign_be (w[36], w[37], offset); + w[38] = hc_bytealign_be (w[35], w[36], offset); + w[37] = hc_bytealign_be (w[34], w[35], offset); + w[36] = hc_bytealign_be (w[33], w[34], offset); + w[35] = hc_bytealign_be (w[32], w[33], offset); + w[34] = hc_bytealign_be (w[31], w[32], offset); + w[33] = hc_bytealign_be (w[30], w[31], offset); + w[32] = hc_bytealign_be (w[29], w[30], offset); + w[31] = hc_bytealign_be (w[28], w[29], offset); + w[30] = hc_bytealign_be (w[27], w[28], offset); + w[29] = hc_bytealign_be (w[26], w[27], offset); + w[28] = hc_bytealign_be (w[25], w[26], offset); + w[27] = hc_bytealign_be (w[24], w[25], offset); + w[26] = hc_bytealign_be (w[23], w[24], offset); + w[25] = hc_bytealign_be (w[22], w[23], offset); + w[24] = hc_bytealign_be (w[21], w[22], offset); + w[23] = hc_bytealign_be (w[20], w[21], offset); + w[22] = hc_bytealign_be (w[19], w[20], offset); + w[21] = hc_bytealign_be (w[18], w[19], offset); + w[20] = hc_bytealign_be (w[17], w[18], offset); + w[19] = hc_bytealign_be (w[16], w[17], offset); + w[18] = hc_bytealign_be (w[15], w[16], offset); + w[17] = hc_bytealign_be (w[14], w[15], offset); + w[16] = hc_bytealign_be (w[13], w[14], offset); + w[15] = hc_bytealign_be (w[12], w[13], offset); + w[14] = hc_bytealign_be (w[11], w[12], offset); + w[13] = hc_bytealign_be (w[10], w[11], offset); + w[12] = hc_bytealign_be (w[ 9], w[10], offset); + w[11] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[10] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[ 9] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[ 8] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[ 7] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 6] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 5] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 4] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 3] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 2] = hc_bytealign_be ( 0, w[ 0], offset); w[ 1] = 0; w[ 0] = 0; break; case 3: - w[63] = hc_bytealign (w[59], w[60], offset); - w[62] = hc_bytealign (w[58], w[59], offset); - w[61] = hc_bytealign (w[57], w[58], offset); - w[60] = hc_bytealign (w[56], w[57], offset); - w[59] = hc_bytealign (w[55], w[56], offset); - w[58] = hc_bytealign (w[54], w[55], offset); - w[57] = hc_bytealign (w[53], w[54], offset); - w[56] = hc_bytealign (w[52], w[53], offset); - w[55] = hc_bytealign (w[51], w[52], offset); - w[54] = hc_bytealign (w[50], w[51], offset); - w[53] = hc_bytealign (w[49], w[50], offset); - w[52] = hc_bytealign (w[48], w[49], offset); - w[51] = hc_bytealign (w[47], w[48], offset); - w[50] = hc_bytealign (w[46], w[47], offset); - w[49] = hc_bytealign (w[45], w[46], offset); - w[48] = hc_bytealign (w[44], w[45], offset); - w[47] = hc_bytealign (w[43], w[44], offset); - w[46] = hc_bytealign (w[42], w[43], offset); - w[45] = hc_bytealign (w[41], w[42], offset); - w[44] = hc_bytealign (w[40], w[41], offset); - w[43] = hc_bytealign (w[39], w[40], offset); - w[42] = hc_bytealign (w[38], w[39], offset); - w[41] = hc_bytealign (w[37], w[38], offset); - w[40] = hc_bytealign (w[36], w[37], offset); - w[39] = hc_bytealign (w[35], w[36], offset); - w[38] = hc_bytealign (w[34], w[35], offset); - w[37] = hc_bytealign (w[33], w[34], offset); - w[36] = hc_bytealign (w[32], w[33], offset); - w[35] = hc_bytealign (w[31], w[32], offset); - w[34] = hc_bytealign (w[30], w[31], offset); - w[33] = hc_bytealign (w[29], w[30], offset); - w[32] = hc_bytealign (w[28], w[29], offset); - w[31] = hc_bytealign (w[27], w[28], offset); - w[30] = hc_bytealign (w[26], w[27], offset); - w[29] = hc_bytealign (w[25], w[26], offset); - w[28] = hc_bytealign (w[24], w[25], offset); - w[27] = hc_bytealign (w[23], w[24], offset); - w[26] = hc_bytealign (w[22], w[23], offset); - w[25] = hc_bytealign (w[21], w[22], offset); - w[24] = hc_bytealign (w[20], w[21], offset); - w[23] = hc_bytealign (w[19], w[20], offset); - w[22] = hc_bytealign (w[18], w[19], offset); - w[21] = hc_bytealign (w[17], w[18], offset); - w[20] = hc_bytealign (w[16], w[17], offset); - w[19] = hc_bytealign (w[15], w[16], offset); - w[18] = hc_bytealign (w[14], w[15], offset); - w[17] = hc_bytealign (w[13], w[14], offset); - w[16] = hc_bytealign (w[12], w[13], offset); - w[15] = hc_bytealign (w[11], w[12], offset); - w[14] = hc_bytealign (w[10], w[11], offset); - w[13] = hc_bytealign (w[ 9], w[10], offset); - w[12] = hc_bytealign (w[ 8], w[ 9], offset); - w[11] = hc_bytealign (w[ 7], w[ 8], offset); - w[10] = hc_bytealign (w[ 6], w[ 7], offset); - w[ 9] = hc_bytealign (w[ 5], w[ 6], offset); - w[ 8] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 7] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 6] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 5] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 4] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 3] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[59], w[60], offset); + w[62] = hc_bytealign_be (w[58], w[59], offset); + w[61] = hc_bytealign_be (w[57], w[58], offset); + w[60] = hc_bytealign_be (w[56], w[57], offset); + w[59] = hc_bytealign_be (w[55], w[56], offset); + w[58] = hc_bytealign_be (w[54], w[55], offset); + w[57] = hc_bytealign_be (w[53], w[54], offset); + w[56] = hc_bytealign_be (w[52], w[53], offset); + w[55] = hc_bytealign_be (w[51], w[52], offset); + w[54] = hc_bytealign_be (w[50], w[51], offset); + w[53] = hc_bytealign_be (w[49], w[50], offset); + w[52] = hc_bytealign_be (w[48], w[49], offset); + w[51] = hc_bytealign_be (w[47], w[48], offset); + w[50] = hc_bytealign_be (w[46], w[47], offset); + w[49] = hc_bytealign_be (w[45], w[46], offset); + w[48] = hc_bytealign_be (w[44], w[45], offset); + w[47] = hc_bytealign_be (w[43], w[44], offset); + w[46] = hc_bytealign_be (w[42], w[43], offset); + w[45] = hc_bytealign_be (w[41], w[42], offset); + w[44] = hc_bytealign_be (w[40], w[41], offset); + w[43] = hc_bytealign_be (w[39], w[40], offset); + w[42] = hc_bytealign_be (w[38], w[39], offset); + w[41] = hc_bytealign_be (w[37], w[38], offset); + w[40] = hc_bytealign_be (w[36], w[37], offset); + w[39] = hc_bytealign_be (w[35], w[36], offset); + w[38] = hc_bytealign_be (w[34], w[35], offset); + w[37] = hc_bytealign_be (w[33], w[34], offset); + w[36] = hc_bytealign_be (w[32], w[33], offset); + w[35] = hc_bytealign_be (w[31], w[32], offset); + w[34] = hc_bytealign_be (w[30], w[31], offset); + w[33] = hc_bytealign_be (w[29], w[30], offset); + w[32] = hc_bytealign_be (w[28], w[29], offset); + w[31] = hc_bytealign_be (w[27], w[28], offset); + w[30] = hc_bytealign_be (w[26], w[27], offset); + w[29] = hc_bytealign_be (w[25], w[26], offset); + w[28] = hc_bytealign_be (w[24], w[25], offset); + w[27] = hc_bytealign_be (w[23], w[24], offset); + w[26] = hc_bytealign_be (w[22], w[23], offset); + w[25] = hc_bytealign_be (w[21], w[22], offset); + w[24] = hc_bytealign_be (w[20], w[21], offset); + w[23] = hc_bytealign_be (w[19], w[20], offset); + w[22] = hc_bytealign_be (w[18], w[19], offset); + w[21] = hc_bytealign_be (w[17], w[18], offset); + w[20] = hc_bytealign_be (w[16], w[17], offset); + w[19] = hc_bytealign_be (w[15], w[16], offset); + w[18] = hc_bytealign_be (w[14], w[15], offset); + w[17] = hc_bytealign_be (w[13], w[14], offset); + w[16] = hc_bytealign_be (w[12], w[13], offset); + w[15] = hc_bytealign_be (w[11], w[12], offset); + w[14] = hc_bytealign_be (w[10], w[11], offset); + w[13] = hc_bytealign_be (w[ 9], w[10], offset); + w[12] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[11] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[10] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[ 9] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[ 8] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 7] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 6] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 5] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 4] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 3] = hc_bytealign_be ( 0, w[ 0], offset); w[ 2] = 0; w[ 1] = 0; w[ 0] = 0; @@ -20877,66 +20715,66 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 4: - w[63] = hc_bytealign (w[58], w[59], offset); - w[62] = hc_bytealign (w[57], w[58], offset); - w[61] = hc_bytealign (w[56], w[57], offset); - w[60] = hc_bytealign (w[55], w[56], offset); - w[59] = hc_bytealign (w[54], w[55], offset); - w[58] = hc_bytealign (w[53], w[54], offset); - w[57] = hc_bytealign (w[52], w[53], offset); - w[56] = hc_bytealign (w[51], w[52], offset); - w[55] = hc_bytealign (w[50], w[51], offset); - w[54] = hc_bytealign (w[49], w[50], offset); - w[53] = hc_bytealign (w[48], w[49], offset); - w[52] = hc_bytealign (w[47], w[48], offset); - w[51] = hc_bytealign (w[46], w[47], offset); - w[50] = hc_bytealign (w[45], w[46], offset); - w[49] = hc_bytealign (w[44], w[45], offset); - w[48] = hc_bytealign (w[43], w[44], offset); - w[47] = hc_bytealign (w[42], w[43], offset); - w[46] = hc_bytealign (w[41], w[42], offset); - w[45] = hc_bytealign (w[40], w[41], offset); - w[44] = hc_bytealign (w[39], w[40], offset); - w[43] = hc_bytealign (w[38], w[39], offset); - w[42] = hc_bytealign (w[37], w[38], offset); - w[41] = hc_bytealign (w[36], w[37], offset); - w[40] = hc_bytealign (w[35], w[36], offset); - w[39] = hc_bytealign (w[34], w[35], offset); - w[38] = hc_bytealign (w[33], w[34], offset); - w[37] = hc_bytealign (w[32], w[33], offset); - w[36] = hc_bytealign (w[31], w[32], offset); - w[35] = hc_bytealign (w[30], w[31], offset); - w[34] = hc_bytealign (w[29], w[30], offset); - w[33] = hc_bytealign (w[28], w[29], offset); - w[32] = hc_bytealign (w[27], w[28], offset); - w[31] = hc_bytealign (w[26], w[27], offset); - w[30] = hc_bytealign (w[25], w[26], offset); - w[29] = hc_bytealign (w[24], w[25], offset); - w[28] = hc_bytealign (w[23], w[24], offset); - w[27] = hc_bytealign (w[22], w[23], offset); - w[26] = hc_bytealign (w[21], w[22], offset); - w[25] = hc_bytealign (w[20], w[21], offset); - w[24] = hc_bytealign (w[19], w[20], offset); - w[23] = hc_bytealign (w[18], w[19], offset); - w[22] = hc_bytealign (w[17], w[18], offset); - w[21] = hc_bytealign (w[16], w[17], offset); - w[20] = hc_bytealign (w[15], w[16], offset); - w[19] = hc_bytealign (w[14], w[15], offset); - w[18] = hc_bytealign (w[13], w[14], offset); - w[17] = hc_bytealign (w[12], w[13], offset); - w[16] = hc_bytealign (w[11], w[12], offset); - w[15] = hc_bytealign (w[10], w[11], offset); - w[14] = hc_bytealign (w[ 9], w[10], offset); - w[13] = hc_bytealign (w[ 8], w[ 9], offset); - w[12] = hc_bytealign (w[ 7], w[ 8], offset); - w[11] = hc_bytealign (w[ 6], w[ 7], offset); - w[10] = hc_bytealign (w[ 5], w[ 6], offset); - w[ 9] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 8] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 7] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 6] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 5] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 4] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[58], w[59], offset); + w[62] = hc_bytealign_be (w[57], w[58], offset); + w[61] = hc_bytealign_be (w[56], w[57], offset); + w[60] = hc_bytealign_be (w[55], w[56], offset); + w[59] = hc_bytealign_be (w[54], w[55], offset); + w[58] = hc_bytealign_be (w[53], w[54], offset); + w[57] = hc_bytealign_be (w[52], w[53], offset); + w[56] = hc_bytealign_be (w[51], w[52], offset); + w[55] = hc_bytealign_be (w[50], w[51], offset); + w[54] = hc_bytealign_be (w[49], w[50], offset); + w[53] = hc_bytealign_be (w[48], w[49], offset); + w[52] = hc_bytealign_be (w[47], w[48], offset); + w[51] = hc_bytealign_be (w[46], w[47], offset); + w[50] = hc_bytealign_be (w[45], w[46], offset); + w[49] = hc_bytealign_be (w[44], w[45], offset); + w[48] = hc_bytealign_be (w[43], w[44], offset); + w[47] = hc_bytealign_be (w[42], w[43], offset); + w[46] = hc_bytealign_be (w[41], w[42], offset); + w[45] = hc_bytealign_be (w[40], w[41], offset); + w[44] = hc_bytealign_be (w[39], w[40], offset); + w[43] = hc_bytealign_be (w[38], w[39], offset); + w[42] = hc_bytealign_be (w[37], w[38], offset); + w[41] = hc_bytealign_be (w[36], w[37], offset); + w[40] = hc_bytealign_be (w[35], w[36], offset); + w[39] = hc_bytealign_be (w[34], w[35], offset); + w[38] = hc_bytealign_be (w[33], w[34], offset); + w[37] = hc_bytealign_be (w[32], w[33], offset); + w[36] = hc_bytealign_be (w[31], w[32], offset); + w[35] = hc_bytealign_be (w[30], w[31], offset); + w[34] = hc_bytealign_be (w[29], w[30], offset); + w[33] = hc_bytealign_be (w[28], w[29], offset); + w[32] = hc_bytealign_be (w[27], w[28], offset); + w[31] = hc_bytealign_be (w[26], w[27], offset); + w[30] = hc_bytealign_be (w[25], w[26], offset); + w[29] = hc_bytealign_be (w[24], w[25], offset); + w[28] = hc_bytealign_be (w[23], w[24], offset); + w[27] = hc_bytealign_be (w[22], w[23], offset); + w[26] = hc_bytealign_be (w[21], w[22], offset); + w[25] = hc_bytealign_be (w[20], w[21], offset); + w[24] = hc_bytealign_be (w[19], w[20], offset); + w[23] = hc_bytealign_be (w[18], w[19], offset); + w[22] = hc_bytealign_be (w[17], w[18], offset); + w[21] = hc_bytealign_be (w[16], w[17], offset); + w[20] = hc_bytealign_be (w[15], w[16], offset); + w[19] = hc_bytealign_be (w[14], w[15], offset); + w[18] = hc_bytealign_be (w[13], w[14], offset); + w[17] = hc_bytealign_be (w[12], w[13], offset); + w[16] = hc_bytealign_be (w[11], w[12], offset); + w[15] = hc_bytealign_be (w[10], w[11], offset); + w[14] = hc_bytealign_be (w[ 9], w[10], offset); + w[13] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[12] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[11] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[10] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[ 9] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 8] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 7] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 6] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 5] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 4] = hc_bytealign_be ( 0, w[ 0], offset); w[ 3] = 0; w[ 2] = 0; w[ 1] = 0; @@ -20945,65 +20783,65 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 5: - w[63] = hc_bytealign (w[57], w[58], offset); - w[62] = hc_bytealign (w[56], w[57], offset); - w[61] = hc_bytealign (w[55], w[56], offset); - w[60] = hc_bytealign (w[54], w[55], offset); - w[59] = hc_bytealign (w[53], w[54], offset); - w[58] = hc_bytealign (w[52], w[53], offset); - w[57] = hc_bytealign (w[51], w[52], offset); - w[56] = hc_bytealign (w[50], w[51], offset); - w[55] = hc_bytealign (w[49], w[50], offset); - w[54] = hc_bytealign (w[48], w[49], offset); - w[53] = hc_bytealign (w[47], w[48], offset); - w[52] = hc_bytealign (w[46], w[47], offset); - w[51] = hc_bytealign (w[45], w[46], offset); - w[50] = hc_bytealign (w[44], w[45], offset); - w[49] = hc_bytealign (w[43], w[44], offset); - w[48] = hc_bytealign (w[42], w[43], offset); - w[47] = hc_bytealign (w[41], w[42], offset); - w[46] = hc_bytealign (w[40], w[41], offset); - w[45] = hc_bytealign (w[39], w[40], offset); - w[44] = hc_bytealign (w[38], w[39], offset); - w[43] = hc_bytealign (w[37], w[38], offset); - w[42] = hc_bytealign (w[36], w[37], offset); - w[41] = hc_bytealign (w[35], w[36], offset); - w[40] = hc_bytealign (w[34], w[35], offset); - w[39] = hc_bytealign (w[33], w[34], offset); - w[38] = hc_bytealign (w[32], w[33], offset); - w[37] = hc_bytealign (w[31], w[32], offset); - w[36] = hc_bytealign (w[30], w[31], offset); - w[35] = hc_bytealign (w[29], w[30], offset); - w[34] = hc_bytealign (w[28], w[29], offset); - w[33] = hc_bytealign (w[27], w[28], offset); - w[32] = hc_bytealign (w[26], w[27], offset); - w[31] = hc_bytealign (w[25], w[26], offset); - w[30] = hc_bytealign (w[24], w[25], offset); - w[29] = hc_bytealign (w[23], w[24], offset); - w[28] = hc_bytealign (w[22], w[23], offset); - w[27] = hc_bytealign (w[21], w[22], offset); - w[26] = hc_bytealign (w[20], w[21], offset); - w[25] = hc_bytealign (w[19], w[20], offset); - w[24] = hc_bytealign (w[18], w[19], offset); - w[23] = hc_bytealign (w[17], w[18], offset); - w[22] = hc_bytealign (w[16], w[17], offset); - w[21] = hc_bytealign (w[15], w[16], offset); - w[20] = hc_bytealign (w[14], w[15], offset); - w[19] = hc_bytealign (w[13], w[14], offset); - w[18] = hc_bytealign (w[12], w[13], offset); - w[17] = hc_bytealign (w[11], w[12], offset); - w[16] = hc_bytealign (w[10], w[11], offset); - w[15] = hc_bytealign (w[ 9], w[10], offset); - w[14] = hc_bytealign (w[ 8], w[ 9], offset); - w[13] = hc_bytealign (w[ 7], w[ 8], offset); - w[12] = hc_bytealign (w[ 6], w[ 7], offset); - w[11] = hc_bytealign (w[ 5], w[ 6], offset); - w[10] = hc_bytealign (w[ 4], w[ 5], offset); - w[ 9] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 8] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 7] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 6] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 5] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[57], w[58], offset); + w[62] = hc_bytealign_be (w[56], w[57], offset); + w[61] = hc_bytealign_be (w[55], w[56], offset); + w[60] = hc_bytealign_be (w[54], w[55], offset); + w[59] = hc_bytealign_be (w[53], w[54], offset); + w[58] = hc_bytealign_be (w[52], w[53], offset); + w[57] = hc_bytealign_be (w[51], w[52], offset); + w[56] = hc_bytealign_be (w[50], w[51], offset); + w[55] = hc_bytealign_be (w[49], w[50], offset); + w[54] = hc_bytealign_be (w[48], w[49], offset); + w[53] = hc_bytealign_be (w[47], w[48], offset); + w[52] = hc_bytealign_be (w[46], w[47], offset); + w[51] = hc_bytealign_be (w[45], w[46], offset); + w[50] = hc_bytealign_be (w[44], w[45], offset); + w[49] = hc_bytealign_be (w[43], w[44], offset); + w[48] = hc_bytealign_be (w[42], w[43], offset); + w[47] = hc_bytealign_be (w[41], w[42], offset); + w[46] = hc_bytealign_be (w[40], w[41], offset); + w[45] = hc_bytealign_be (w[39], w[40], offset); + w[44] = hc_bytealign_be (w[38], w[39], offset); + w[43] = hc_bytealign_be (w[37], w[38], offset); + w[42] = hc_bytealign_be (w[36], w[37], offset); + w[41] = hc_bytealign_be (w[35], w[36], offset); + w[40] = hc_bytealign_be (w[34], w[35], offset); + w[39] = hc_bytealign_be (w[33], w[34], offset); + w[38] = hc_bytealign_be (w[32], w[33], offset); + w[37] = hc_bytealign_be (w[31], w[32], offset); + w[36] = hc_bytealign_be (w[30], w[31], offset); + w[35] = hc_bytealign_be (w[29], w[30], offset); + w[34] = hc_bytealign_be (w[28], w[29], offset); + w[33] = hc_bytealign_be (w[27], w[28], offset); + w[32] = hc_bytealign_be (w[26], w[27], offset); + w[31] = hc_bytealign_be (w[25], w[26], offset); + w[30] = hc_bytealign_be (w[24], w[25], offset); + w[29] = hc_bytealign_be (w[23], w[24], offset); + w[28] = hc_bytealign_be (w[22], w[23], offset); + w[27] = hc_bytealign_be (w[21], w[22], offset); + w[26] = hc_bytealign_be (w[20], w[21], offset); + w[25] = hc_bytealign_be (w[19], w[20], offset); + w[24] = hc_bytealign_be (w[18], w[19], offset); + w[23] = hc_bytealign_be (w[17], w[18], offset); + w[22] = hc_bytealign_be (w[16], w[17], offset); + w[21] = hc_bytealign_be (w[15], w[16], offset); + w[20] = hc_bytealign_be (w[14], w[15], offset); + w[19] = hc_bytealign_be (w[13], w[14], offset); + w[18] = hc_bytealign_be (w[12], w[13], offset); + w[17] = hc_bytealign_be (w[11], w[12], offset); + w[16] = hc_bytealign_be (w[10], w[11], offset); + w[15] = hc_bytealign_be (w[ 9], w[10], offset); + w[14] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[13] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[12] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[11] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[10] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[ 9] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 8] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 7] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 6] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 5] = hc_bytealign_be ( 0, w[ 0], offset); w[ 4] = 0; w[ 3] = 0; w[ 2] = 0; @@ -21013,64 +20851,64 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 6: - w[63] = hc_bytealign (w[56], w[57], offset); - w[62] = hc_bytealign (w[55], w[56], offset); - w[61] = hc_bytealign (w[54], w[55], offset); - w[60] = hc_bytealign (w[53], w[54], offset); - w[59] = hc_bytealign (w[52], w[53], offset); - w[58] = hc_bytealign (w[51], w[52], offset); - w[57] = hc_bytealign (w[50], w[51], offset); - w[56] = hc_bytealign (w[49], w[50], offset); - w[55] = hc_bytealign (w[48], w[49], offset); - w[54] = hc_bytealign (w[47], w[48], offset); - w[53] = hc_bytealign (w[46], w[47], offset); - w[52] = hc_bytealign (w[45], w[46], offset); - w[51] = hc_bytealign (w[44], w[45], offset); - w[50] = hc_bytealign (w[43], w[44], offset); - w[49] = hc_bytealign (w[42], w[43], offset); - w[48] = hc_bytealign (w[41], w[42], offset); - w[47] = hc_bytealign (w[40], w[41], offset); - w[46] = hc_bytealign (w[39], w[40], offset); - w[45] = hc_bytealign (w[38], w[39], offset); - w[44] = hc_bytealign (w[37], w[38], offset); - w[43] = hc_bytealign (w[36], w[37], offset); - w[42] = hc_bytealign (w[35], w[36], offset); - w[41] = hc_bytealign (w[34], w[35], offset); - w[40] = hc_bytealign (w[33], w[34], offset); - w[39] = hc_bytealign (w[32], w[33], offset); - w[38] = hc_bytealign (w[31], w[32], offset); - w[37] = hc_bytealign (w[30], w[31], offset); - w[36] = hc_bytealign (w[29], w[30], offset); - w[35] = hc_bytealign (w[28], w[29], offset); - w[34] = hc_bytealign (w[27], w[28], offset); - w[33] = hc_bytealign (w[26], w[27], offset); - w[32] = hc_bytealign (w[25], w[26], offset); - w[31] = hc_bytealign (w[24], w[25], offset); - w[30] = hc_bytealign (w[23], w[24], offset); - w[29] = hc_bytealign (w[22], w[23], offset); - w[28] = hc_bytealign (w[21], w[22], offset); - w[27] = hc_bytealign (w[20], w[21], offset); - w[26] = hc_bytealign (w[19], w[20], offset); - w[25] = hc_bytealign (w[18], w[19], offset); - w[24] = hc_bytealign (w[17], w[18], offset); - w[23] = hc_bytealign (w[16], w[17], offset); - w[22] = hc_bytealign (w[15], w[16], offset); - w[21] = hc_bytealign (w[14], w[15], offset); - w[20] = hc_bytealign (w[13], w[14], offset); - w[19] = hc_bytealign (w[12], w[13], offset); - w[18] = hc_bytealign (w[11], w[12], offset); - w[17] = hc_bytealign (w[10], w[11], offset); - w[16] = hc_bytealign (w[ 9], w[10], offset); - w[15] = hc_bytealign (w[ 8], w[ 9], offset); - w[14] = hc_bytealign (w[ 7], w[ 8], offset); - w[13] = hc_bytealign (w[ 6], w[ 7], offset); - w[12] = hc_bytealign (w[ 5], w[ 6], offset); - w[11] = hc_bytealign (w[ 4], w[ 5], offset); - w[10] = hc_bytealign (w[ 3], w[ 4], offset); - w[ 9] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 8] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 7] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 6] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[56], w[57], offset); + w[62] = hc_bytealign_be (w[55], w[56], offset); + w[61] = hc_bytealign_be (w[54], w[55], offset); + w[60] = hc_bytealign_be (w[53], w[54], offset); + w[59] = hc_bytealign_be (w[52], w[53], offset); + w[58] = hc_bytealign_be (w[51], w[52], offset); + w[57] = hc_bytealign_be (w[50], w[51], offset); + w[56] = hc_bytealign_be (w[49], w[50], offset); + w[55] = hc_bytealign_be (w[48], w[49], offset); + w[54] = hc_bytealign_be (w[47], w[48], offset); + w[53] = hc_bytealign_be (w[46], w[47], offset); + w[52] = hc_bytealign_be (w[45], w[46], offset); + w[51] = hc_bytealign_be (w[44], w[45], offset); + w[50] = hc_bytealign_be (w[43], w[44], offset); + w[49] = hc_bytealign_be (w[42], w[43], offset); + w[48] = hc_bytealign_be (w[41], w[42], offset); + w[47] = hc_bytealign_be (w[40], w[41], offset); + w[46] = hc_bytealign_be (w[39], w[40], offset); + w[45] = hc_bytealign_be (w[38], w[39], offset); + w[44] = hc_bytealign_be (w[37], w[38], offset); + w[43] = hc_bytealign_be (w[36], w[37], offset); + w[42] = hc_bytealign_be (w[35], w[36], offset); + w[41] = hc_bytealign_be (w[34], w[35], offset); + w[40] = hc_bytealign_be (w[33], w[34], offset); + w[39] = hc_bytealign_be (w[32], w[33], offset); + w[38] = hc_bytealign_be (w[31], w[32], offset); + w[37] = hc_bytealign_be (w[30], w[31], offset); + w[36] = hc_bytealign_be (w[29], w[30], offset); + w[35] = hc_bytealign_be (w[28], w[29], offset); + w[34] = hc_bytealign_be (w[27], w[28], offset); + w[33] = hc_bytealign_be (w[26], w[27], offset); + w[32] = hc_bytealign_be (w[25], w[26], offset); + w[31] = hc_bytealign_be (w[24], w[25], offset); + w[30] = hc_bytealign_be (w[23], w[24], offset); + w[29] = hc_bytealign_be (w[22], w[23], offset); + w[28] = hc_bytealign_be (w[21], w[22], offset); + w[27] = hc_bytealign_be (w[20], w[21], offset); + w[26] = hc_bytealign_be (w[19], w[20], offset); + w[25] = hc_bytealign_be (w[18], w[19], offset); + w[24] = hc_bytealign_be (w[17], w[18], offset); + w[23] = hc_bytealign_be (w[16], w[17], offset); + w[22] = hc_bytealign_be (w[15], w[16], offset); + w[21] = hc_bytealign_be (w[14], w[15], offset); + w[20] = hc_bytealign_be (w[13], w[14], offset); + w[19] = hc_bytealign_be (w[12], w[13], offset); + w[18] = hc_bytealign_be (w[11], w[12], offset); + w[17] = hc_bytealign_be (w[10], w[11], offset); + w[16] = hc_bytealign_be (w[ 9], w[10], offset); + w[15] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[14] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[13] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[12] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[11] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[10] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[ 9] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 8] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 7] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 6] = hc_bytealign_be ( 0, w[ 0], offset); w[ 5] = 0; w[ 4] = 0; w[ 3] = 0; @@ -21081,63 +20919,63 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 7: - w[63] = hc_bytealign (w[55], w[56], offset); - w[62] = hc_bytealign (w[54], w[55], offset); - w[61] = hc_bytealign (w[53], w[54], offset); - w[60] = hc_bytealign (w[52], w[53], offset); - w[59] = hc_bytealign (w[51], w[52], offset); - w[58] = hc_bytealign (w[50], w[51], offset); - w[57] = hc_bytealign (w[49], w[50], offset); - w[56] = hc_bytealign (w[48], w[49], offset); - w[55] = hc_bytealign (w[47], w[48], offset); - w[54] = hc_bytealign (w[46], w[47], offset); - w[53] = hc_bytealign (w[45], w[46], offset); - w[52] = hc_bytealign (w[44], w[45], offset); - w[51] = hc_bytealign (w[43], w[44], offset); - w[50] = hc_bytealign (w[42], w[43], offset); - w[49] = hc_bytealign (w[41], w[42], offset); - w[48] = hc_bytealign (w[40], w[41], offset); - w[47] = hc_bytealign (w[39], w[40], offset); - w[46] = hc_bytealign (w[38], w[39], offset); - w[45] = hc_bytealign (w[37], w[38], offset); - w[44] = hc_bytealign (w[36], w[37], offset); - w[43] = hc_bytealign (w[35], w[36], offset); - w[42] = hc_bytealign (w[34], w[35], offset); - w[41] = hc_bytealign (w[33], w[34], offset); - w[40] = hc_bytealign (w[32], w[33], offset); - w[39] = hc_bytealign (w[31], w[32], offset); - w[38] = hc_bytealign (w[30], w[31], offset); - w[37] = hc_bytealign (w[29], w[30], offset); - w[36] = hc_bytealign (w[28], w[29], offset); - w[35] = hc_bytealign (w[27], w[28], offset); - w[34] = hc_bytealign (w[26], w[27], offset); - w[33] = hc_bytealign (w[25], w[26], offset); - w[32] = hc_bytealign (w[24], w[25], offset); - w[31] = hc_bytealign (w[23], w[24], offset); - w[30] = hc_bytealign (w[22], w[23], offset); - w[29] = hc_bytealign (w[21], w[22], offset); - w[28] = hc_bytealign (w[20], w[21], offset); - w[27] = hc_bytealign (w[19], w[20], offset); - w[26] = hc_bytealign (w[18], w[19], offset); - w[25] = hc_bytealign (w[17], w[18], offset); - w[24] = hc_bytealign (w[16], w[17], offset); - w[23] = hc_bytealign (w[15], w[16], offset); - w[22] = hc_bytealign (w[14], w[15], offset); - w[21] = hc_bytealign (w[13], w[14], offset); - w[20] = hc_bytealign (w[12], w[13], offset); - w[19] = hc_bytealign (w[11], w[12], offset); - w[18] = hc_bytealign (w[10], w[11], offset); - w[17] = hc_bytealign (w[ 9], w[10], offset); - w[16] = hc_bytealign (w[ 8], w[ 9], offset); - w[15] = hc_bytealign (w[ 7], w[ 8], offset); - w[14] = hc_bytealign (w[ 6], w[ 7], offset); - w[13] = hc_bytealign (w[ 5], w[ 6], offset); - w[12] = hc_bytealign (w[ 4], w[ 5], offset); - w[11] = hc_bytealign (w[ 3], w[ 4], offset); - w[10] = hc_bytealign (w[ 2], w[ 3], offset); - w[ 9] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 8] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 7] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[55], w[56], offset); + w[62] = hc_bytealign_be (w[54], w[55], offset); + w[61] = hc_bytealign_be (w[53], w[54], offset); + w[60] = hc_bytealign_be (w[52], w[53], offset); + w[59] = hc_bytealign_be (w[51], w[52], offset); + w[58] = hc_bytealign_be (w[50], w[51], offset); + w[57] = hc_bytealign_be (w[49], w[50], offset); + w[56] = hc_bytealign_be (w[48], w[49], offset); + w[55] = hc_bytealign_be (w[47], w[48], offset); + w[54] = hc_bytealign_be (w[46], w[47], offset); + w[53] = hc_bytealign_be (w[45], w[46], offset); + w[52] = hc_bytealign_be (w[44], w[45], offset); + w[51] = hc_bytealign_be (w[43], w[44], offset); + w[50] = hc_bytealign_be (w[42], w[43], offset); + w[49] = hc_bytealign_be (w[41], w[42], offset); + w[48] = hc_bytealign_be (w[40], w[41], offset); + w[47] = hc_bytealign_be (w[39], w[40], offset); + w[46] = hc_bytealign_be (w[38], w[39], offset); + w[45] = hc_bytealign_be (w[37], w[38], offset); + w[44] = hc_bytealign_be (w[36], w[37], offset); + w[43] = hc_bytealign_be (w[35], w[36], offset); + w[42] = hc_bytealign_be (w[34], w[35], offset); + w[41] = hc_bytealign_be (w[33], w[34], offset); + w[40] = hc_bytealign_be (w[32], w[33], offset); + w[39] = hc_bytealign_be (w[31], w[32], offset); + w[38] = hc_bytealign_be (w[30], w[31], offset); + w[37] = hc_bytealign_be (w[29], w[30], offset); + w[36] = hc_bytealign_be (w[28], w[29], offset); + w[35] = hc_bytealign_be (w[27], w[28], offset); + w[34] = hc_bytealign_be (w[26], w[27], offset); + w[33] = hc_bytealign_be (w[25], w[26], offset); + w[32] = hc_bytealign_be (w[24], w[25], offset); + w[31] = hc_bytealign_be (w[23], w[24], offset); + w[30] = hc_bytealign_be (w[22], w[23], offset); + w[29] = hc_bytealign_be (w[21], w[22], offset); + w[28] = hc_bytealign_be (w[20], w[21], offset); + w[27] = hc_bytealign_be (w[19], w[20], offset); + w[26] = hc_bytealign_be (w[18], w[19], offset); + w[25] = hc_bytealign_be (w[17], w[18], offset); + w[24] = hc_bytealign_be (w[16], w[17], offset); + w[23] = hc_bytealign_be (w[15], w[16], offset); + w[22] = hc_bytealign_be (w[14], w[15], offset); + w[21] = hc_bytealign_be (w[13], w[14], offset); + w[20] = hc_bytealign_be (w[12], w[13], offset); + w[19] = hc_bytealign_be (w[11], w[12], offset); + w[18] = hc_bytealign_be (w[10], w[11], offset); + w[17] = hc_bytealign_be (w[ 9], w[10], offset); + w[16] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[15] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[14] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[13] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[12] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[11] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[10] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[ 9] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 8] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 7] = hc_bytealign_be ( 0, w[ 0], offset); w[ 6] = 0; w[ 5] = 0; w[ 4] = 0; @@ -21149,62 +20987,62 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 8: - w[63] = hc_bytealign (w[54], w[55], offset); - w[62] = hc_bytealign (w[53], w[54], offset); - w[61] = hc_bytealign (w[52], w[53], offset); - w[60] = hc_bytealign (w[51], w[52], offset); - w[59] = hc_bytealign (w[50], w[51], offset); - w[58] = hc_bytealign (w[49], w[50], offset); - w[57] = hc_bytealign (w[48], w[49], offset); - w[56] = hc_bytealign (w[47], w[48], offset); - w[55] = hc_bytealign (w[46], w[47], offset); - w[54] = hc_bytealign (w[45], w[46], offset); - w[53] = hc_bytealign (w[44], w[45], offset); - w[52] = hc_bytealign (w[43], w[44], offset); - w[51] = hc_bytealign (w[42], w[43], offset); - w[50] = hc_bytealign (w[41], w[42], offset); - w[49] = hc_bytealign (w[40], w[41], offset); - w[48] = hc_bytealign (w[39], w[40], offset); - w[47] = hc_bytealign (w[38], w[39], offset); - w[46] = hc_bytealign (w[37], w[38], offset); - w[45] = hc_bytealign (w[36], w[37], offset); - w[44] = hc_bytealign (w[35], w[36], offset); - w[43] = hc_bytealign (w[34], w[35], offset); - w[42] = hc_bytealign (w[33], w[34], offset); - w[41] = hc_bytealign (w[32], w[33], offset); - w[40] = hc_bytealign (w[31], w[32], offset); - w[39] = hc_bytealign (w[30], w[31], offset); - w[38] = hc_bytealign (w[29], w[30], offset); - w[37] = hc_bytealign (w[28], w[29], offset); - w[36] = hc_bytealign (w[27], w[28], offset); - w[35] = hc_bytealign (w[26], w[27], offset); - w[34] = hc_bytealign (w[25], w[26], offset); - w[33] = hc_bytealign (w[24], w[25], offset); - w[32] = hc_bytealign (w[23], w[24], offset); - w[31] = hc_bytealign (w[22], w[23], offset); - w[30] = hc_bytealign (w[21], w[22], offset); - w[29] = hc_bytealign (w[20], w[21], offset); - w[28] = hc_bytealign (w[19], w[20], offset); - w[27] = hc_bytealign (w[18], w[19], offset); - w[26] = hc_bytealign (w[17], w[18], offset); - w[25] = hc_bytealign (w[16], w[17], offset); - w[24] = hc_bytealign (w[15], w[16], offset); - w[23] = hc_bytealign (w[14], w[15], offset); - w[22] = hc_bytealign (w[13], w[14], offset); - w[21] = hc_bytealign (w[12], w[13], offset); - w[20] = hc_bytealign (w[11], w[12], offset); - w[19] = hc_bytealign (w[10], w[11], offset); - w[18] = hc_bytealign (w[ 9], w[10], offset); - w[17] = hc_bytealign (w[ 8], w[ 9], offset); - w[16] = hc_bytealign (w[ 7], w[ 8], offset); - w[15] = hc_bytealign (w[ 6], w[ 7], offset); - w[14] = hc_bytealign (w[ 5], w[ 6], offset); - w[13] = hc_bytealign (w[ 4], w[ 5], offset); - w[12] = hc_bytealign (w[ 3], w[ 4], offset); - w[11] = hc_bytealign (w[ 2], w[ 3], offset); - w[10] = hc_bytealign (w[ 1], w[ 2], offset); - w[ 9] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 8] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[54], w[55], offset); + w[62] = hc_bytealign_be (w[53], w[54], offset); + w[61] = hc_bytealign_be (w[52], w[53], offset); + w[60] = hc_bytealign_be (w[51], w[52], offset); + w[59] = hc_bytealign_be (w[50], w[51], offset); + w[58] = hc_bytealign_be (w[49], w[50], offset); + w[57] = hc_bytealign_be (w[48], w[49], offset); + w[56] = hc_bytealign_be (w[47], w[48], offset); + w[55] = hc_bytealign_be (w[46], w[47], offset); + w[54] = hc_bytealign_be (w[45], w[46], offset); + w[53] = hc_bytealign_be (w[44], w[45], offset); + w[52] = hc_bytealign_be (w[43], w[44], offset); + w[51] = hc_bytealign_be (w[42], w[43], offset); + w[50] = hc_bytealign_be (w[41], w[42], offset); + w[49] = hc_bytealign_be (w[40], w[41], offset); + w[48] = hc_bytealign_be (w[39], w[40], offset); + w[47] = hc_bytealign_be (w[38], w[39], offset); + w[46] = hc_bytealign_be (w[37], w[38], offset); + w[45] = hc_bytealign_be (w[36], w[37], offset); + w[44] = hc_bytealign_be (w[35], w[36], offset); + w[43] = hc_bytealign_be (w[34], w[35], offset); + w[42] = hc_bytealign_be (w[33], w[34], offset); + w[41] = hc_bytealign_be (w[32], w[33], offset); + w[40] = hc_bytealign_be (w[31], w[32], offset); + w[39] = hc_bytealign_be (w[30], w[31], offset); + w[38] = hc_bytealign_be (w[29], w[30], offset); + w[37] = hc_bytealign_be (w[28], w[29], offset); + w[36] = hc_bytealign_be (w[27], w[28], offset); + w[35] = hc_bytealign_be (w[26], w[27], offset); + w[34] = hc_bytealign_be (w[25], w[26], offset); + w[33] = hc_bytealign_be (w[24], w[25], offset); + w[32] = hc_bytealign_be (w[23], w[24], offset); + w[31] = hc_bytealign_be (w[22], w[23], offset); + w[30] = hc_bytealign_be (w[21], w[22], offset); + w[29] = hc_bytealign_be (w[20], w[21], offset); + w[28] = hc_bytealign_be (w[19], w[20], offset); + w[27] = hc_bytealign_be (w[18], w[19], offset); + w[26] = hc_bytealign_be (w[17], w[18], offset); + w[25] = hc_bytealign_be (w[16], w[17], offset); + w[24] = hc_bytealign_be (w[15], w[16], offset); + w[23] = hc_bytealign_be (w[14], w[15], offset); + w[22] = hc_bytealign_be (w[13], w[14], offset); + w[21] = hc_bytealign_be (w[12], w[13], offset); + w[20] = hc_bytealign_be (w[11], w[12], offset); + w[19] = hc_bytealign_be (w[10], w[11], offset); + w[18] = hc_bytealign_be (w[ 9], w[10], offset); + w[17] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[16] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[15] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[14] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[13] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[12] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[11] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[10] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[ 9] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 8] = hc_bytealign_be ( 0, w[ 0], offset); w[ 7] = 0; w[ 6] = 0; w[ 5] = 0; @@ -21217,61 +21055,61 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 9: - w[63] = hc_bytealign (w[53], w[54], offset); - w[62] = hc_bytealign (w[52], w[53], offset); - w[61] = hc_bytealign (w[51], w[52], offset); - w[60] = hc_bytealign (w[50], w[51], offset); - w[59] = hc_bytealign (w[49], w[50], offset); - w[58] = hc_bytealign (w[48], w[49], offset); - w[57] = hc_bytealign (w[47], w[48], offset); - w[56] = hc_bytealign (w[46], w[47], offset); - w[55] = hc_bytealign (w[45], w[46], offset); - w[54] = hc_bytealign (w[44], w[45], offset); - w[53] = hc_bytealign (w[43], w[44], offset); - w[52] = hc_bytealign (w[42], w[43], offset); - w[51] = hc_bytealign (w[41], w[42], offset); - w[50] = hc_bytealign (w[40], w[41], offset); - w[49] = hc_bytealign (w[39], w[40], offset); - w[48] = hc_bytealign (w[38], w[39], offset); - w[47] = hc_bytealign (w[37], w[38], offset); - w[46] = hc_bytealign (w[36], w[37], offset); - w[45] = hc_bytealign (w[35], w[36], offset); - w[44] = hc_bytealign (w[34], w[35], offset); - w[43] = hc_bytealign (w[33], w[34], offset); - w[42] = hc_bytealign (w[32], w[33], offset); - w[41] = hc_bytealign (w[31], w[32], offset); - w[40] = hc_bytealign (w[30], w[31], offset); - w[39] = hc_bytealign (w[29], w[30], offset); - w[38] = hc_bytealign (w[28], w[29], offset); - w[37] = hc_bytealign (w[27], w[28], offset); - w[36] = hc_bytealign (w[26], w[27], offset); - w[35] = hc_bytealign (w[25], w[26], offset); - w[34] = hc_bytealign (w[24], w[25], offset); - w[33] = hc_bytealign (w[23], w[24], offset); - w[32] = hc_bytealign (w[22], w[23], offset); - w[31] = hc_bytealign (w[21], w[22], offset); - w[30] = hc_bytealign (w[20], w[21], offset); - w[29] = hc_bytealign (w[19], w[20], offset); - w[28] = hc_bytealign (w[18], w[19], offset); - w[27] = hc_bytealign (w[17], w[18], offset); - w[26] = hc_bytealign (w[16], w[17], offset); - w[25] = hc_bytealign (w[15], w[16], offset); - w[24] = hc_bytealign (w[14], w[15], offset); - w[23] = hc_bytealign (w[13], w[14], offset); - w[22] = hc_bytealign (w[12], w[13], offset); - w[21] = hc_bytealign (w[11], w[12], offset); - w[20] = hc_bytealign (w[10], w[11], offset); - w[19] = hc_bytealign (w[ 9], w[10], offset); - w[18] = hc_bytealign (w[ 8], w[ 9], offset); - w[17] = hc_bytealign (w[ 7], w[ 8], offset); - w[16] = hc_bytealign (w[ 6], w[ 7], offset); - w[15] = hc_bytealign (w[ 5], w[ 6], offset); - w[14] = hc_bytealign (w[ 4], w[ 5], offset); - w[13] = hc_bytealign (w[ 3], w[ 4], offset); - w[12] = hc_bytealign (w[ 2], w[ 3], offset); - w[11] = hc_bytealign (w[ 1], w[ 2], offset); - w[10] = hc_bytealign (w[ 0], w[ 1], offset); - w[ 9] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[53], w[54], offset); + w[62] = hc_bytealign_be (w[52], w[53], offset); + w[61] = hc_bytealign_be (w[51], w[52], offset); + w[60] = hc_bytealign_be (w[50], w[51], offset); + w[59] = hc_bytealign_be (w[49], w[50], offset); + w[58] = hc_bytealign_be (w[48], w[49], offset); + w[57] = hc_bytealign_be (w[47], w[48], offset); + w[56] = hc_bytealign_be (w[46], w[47], offset); + w[55] = hc_bytealign_be (w[45], w[46], offset); + w[54] = hc_bytealign_be (w[44], w[45], offset); + w[53] = hc_bytealign_be (w[43], w[44], offset); + w[52] = hc_bytealign_be (w[42], w[43], offset); + w[51] = hc_bytealign_be (w[41], w[42], offset); + w[50] = hc_bytealign_be (w[40], w[41], offset); + w[49] = hc_bytealign_be (w[39], w[40], offset); + w[48] = hc_bytealign_be (w[38], w[39], offset); + w[47] = hc_bytealign_be (w[37], w[38], offset); + w[46] = hc_bytealign_be (w[36], w[37], offset); + w[45] = hc_bytealign_be (w[35], w[36], offset); + w[44] = hc_bytealign_be (w[34], w[35], offset); + w[43] = hc_bytealign_be (w[33], w[34], offset); + w[42] = hc_bytealign_be (w[32], w[33], offset); + w[41] = hc_bytealign_be (w[31], w[32], offset); + w[40] = hc_bytealign_be (w[30], w[31], offset); + w[39] = hc_bytealign_be (w[29], w[30], offset); + w[38] = hc_bytealign_be (w[28], w[29], offset); + w[37] = hc_bytealign_be (w[27], w[28], offset); + w[36] = hc_bytealign_be (w[26], w[27], offset); + w[35] = hc_bytealign_be (w[25], w[26], offset); + w[34] = hc_bytealign_be (w[24], w[25], offset); + w[33] = hc_bytealign_be (w[23], w[24], offset); + w[32] = hc_bytealign_be (w[22], w[23], offset); + w[31] = hc_bytealign_be (w[21], w[22], offset); + w[30] = hc_bytealign_be (w[20], w[21], offset); + w[29] = hc_bytealign_be (w[19], w[20], offset); + w[28] = hc_bytealign_be (w[18], w[19], offset); + w[27] = hc_bytealign_be (w[17], w[18], offset); + w[26] = hc_bytealign_be (w[16], w[17], offset); + w[25] = hc_bytealign_be (w[15], w[16], offset); + w[24] = hc_bytealign_be (w[14], w[15], offset); + w[23] = hc_bytealign_be (w[13], w[14], offset); + w[22] = hc_bytealign_be (w[12], w[13], offset); + w[21] = hc_bytealign_be (w[11], w[12], offset); + w[20] = hc_bytealign_be (w[10], w[11], offset); + w[19] = hc_bytealign_be (w[ 9], w[10], offset); + w[18] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[17] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[16] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[15] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[14] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[13] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[12] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[11] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[10] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[ 9] = hc_bytealign_be ( 0, w[ 0], offset); w[ 8] = 0; w[ 7] = 0; w[ 6] = 0; @@ -21285,60 +21123,60 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 10: - w[63] = hc_bytealign (w[52], w[53], offset); - w[62] = hc_bytealign (w[51], w[52], offset); - w[61] = hc_bytealign (w[50], w[51], offset); - w[60] = hc_bytealign (w[49], w[50], offset); - w[59] = hc_bytealign (w[48], w[49], offset); - w[58] = hc_bytealign (w[47], w[48], offset); - w[57] = hc_bytealign (w[46], w[47], offset); - w[56] = hc_bytealign (w[45], w[46], offset); - w[55] = hc_bytealign (w[44], w[45], offset); - w[54] = hc_bytealign (w[43], w[44], offset); - w[53] = hc_bytealign (w[42], w[43], offset); - w[52] = hc_bytealign (w[41], w[42], offset); - w[51] = hc_bytealign (w[40], w[41], offset); - w[50] = hc_bytealign (w[39], w[40], offset); - w[49] = hc_bytealign (w[38], w[39], offset); - w[48] = hc_bytealign (w[37], w[38], offset); - w[47] = hc_bytealign (w[36], w[37], offset); - w[46] = hc_bytealign (w[35], w[36], offset); - w[45] = hc_bytealign (w[34], w[35], offset); - w[44] = hc_bytealign (w[33], w[34], offset); - w[43] = hc_bytealign (w[32], w[33], offset); - w[42] = hc_bytealign (w[31], w[32], offset); - w[41] = hc_bytealign (w[30], w[31], offset); - w[40] = hc_bytealign (w[29], w[30], offset); - w[39] = hc_bytealign (w[28], w[29], offset); - w[38] = hc_bytealign (w[27], w[28], offset); - w[37] = hc_bytealign (w[26], w[27], offset); - w[36] = hc_bytealign (w[25], w[26], offset); - w[35] = hc_bytealign (w[24], w[25], offset); - w[34] = hc_bytealign (w[23], w[24], offset); - w[33] = hc_bytealign (w[22], w[23], offset); - w[32] = hc_bytealign (w[21], w[22], offset); - w[31] = hc_bytealign (w[20], w[21], offset); - w[30] = hc_bytealign (w[19], w[20], offset); - w[29] = hc_bytealign (w[18], w[19], offset); - w[28] = hc_bytealign (w[17], w[18], offset); - w[27] = hc_bytealign (w[16], w[17], offset); - w[26] = hc_bytealign (w[15], w[16], offset); - w[25] = hc_bytealign (w[14], w[15], offset); - w[24] = hc_bytealign (w[13], w[14], offset); - w[23] = hc_bytealign (w[12], w[13], offset); - w[22] = hc_bytealign (w[11], w[12], offset); - w[21] = hc_bytealign (w[10], w[11], offset); - w[20] = hc_bytealign (w[ 9], w[10], offset); - w[19] = hc_bytealign (w[ 8], w[ 9], offset); - w[18] = hc_bytealign (w[ 7], w[ 8], offset); - w[17] = hc_bytealign (w[ 6], w[ 7], offset); - w[16] = hc_bytealign (w[ 5], w[ 6], offset); - w[15] = hc_bytealign (w[ 4], w[ 5], offset); - w[14] = hc_bytealign (w[ 3], w[ 4], offset); - w[13] = hc_bytealign (w[ 2], w[ 3], offset); - w[12] = hc_bytealign (w[ 1], w[ 2], offset); - w[11] = hc_bytealign (w[ 0], w[ 1], offset); - w[10] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[52], w[53], offset); + w[62] = hc_bytealign_be (w[51], w[52], offset); + w[61] = hc_bytealign_be (w[50], w[51], offset); + w[60] = hc_bytealign_be (w[49], w[50], offset); + w[59] = hc_bytealign_be (w[48], w[49], offset); + w[58] = hc_bytealign_be (w[47], w[48], offset); + w[57] = hc_bytealign_be (w[46], w[47], offset); + w[56] = hc_bytealign_be (w[45], w[46], offset); + w[55] = hc_bytealign_be (w[44], w[45], offset); + w[54] = hc_bytealign_be (w[43], w[44], offset); + w[53] = hc_bytealign_be (w[42], w[43], offset); + w[52] = hc_bytealign_be (w[41], w[42], offset); + w[51] = hc_bytealign_be (w[40], w[41], offset); + w[50] = hc_bytealign_be (w[39], w[40], offset); + w[49] = hc_bytealign_be (w[38], w[39], offset); + w[48] = hc_bytealign_be (w[37], w[38], offset); + w[47] = hc_bytealign_be (w[36], w[37], offset); + w[46] = hc_bytealign_be (w[35], w[36], offset); + w[45] = hc_bytealign_be (w[34], w[35], offset); + w[44] = hc_bytealign_be (w[33], w[34], offset); + w[43] = hc_bytealign_be (w[32], w[33], offset); + w[42] = hc_bytealign_be (w[31], w[32], offset); + w[41] = hc_bytealign_be (w[30], w[31], offset); + w[40] = hc_bytealign_be (w[29], w[30], offset); + w[39] = hc_bytealign_be (w[28], w[29], offset); + w[38] = hc_bytealign_be (w[27], w[28], offset); + w[37] = hc_bytealign_be (w[26], w[27], offset); + w[36] = hc_bytealign_be (w[25], w[26], offset); + w[35] = hc_bytealign_be (w[24], w[25], offset); + w[34] = hc_bytealign_be (w[23], w[24], offset); + w[33] = hc_bytealign_be (w[22], w[23], offset); + w[32] = hc_bytealign_be (w[21], w[22], offset); + w[31] = hc_bytealign_be (w[20], w[21], offset); + w[30] = hc_bytealign_be (w[19], w[20], offset); + w[29] = hc_bytealign_be (w[18], w[19], offset); + w[28] = hc_bytealign_be (w[17], w[18], offset); + w[27] = hc_bytealign_be (w[16], w[17], offset); + w[26] = hc_bytealign_be (w[15], w[16], offset); + w[25] = hc_bytealign_be (w[14], w[15], offset); + w[24] = hc_bytealign_be (w[13], w[14], offset); + w[23] = hc_bytealign_be (w[12], w[13], offset); + w[22] = hc_bytealign_be (w[11], w[12], offset); + w[21] = hc_bytealign_be (w[10], w[11], offset); + w[20] = hc_bytealign_be (w[ 9], w[10], offset); + w[19] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[18] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[17] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[16] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[15] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[14] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[13] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[12] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[11] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[10] = hc_bytealign_be ( 0, w[ 0], offset); w[ 9] = 0; w[ 8] = 0; w[ 7] = 0; @@ -21353,59 +21191,59 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 11: - w[63] = hc_bytealign (w[51], w[52], offset); - w[62] = hc_bytealign (w[50], w[51], offset); - w[61] = hc_bytealign (w[49], w[50], offset); - w[60] = hc_bytealign (w[48], w[49], offset); - w[59] = hc_bytealign (w[47], w[48], offset); - w[58] = hc_bytealign (w[46], w[47], offset); - w[57] = hc_bytealign (w[45], w[46], offset); - w[56] = hc_bytealign (w[44], w[45], offset); - w[55] = hc_bytealign (w[43], w[44], offset); - w[54] = hc_bytealign (w[42], w[43], offset); - w[53] = hc_bytealign (w[41], w[42], offset); - w[52] = hc_bytealign (w[40], w[41], offset); - w[51] = hc_bytealign (w[39], w[40], offset); - w[50] = hc_bytealign (w[38], w[39], offset); - w[49] = hc_bytealign (w[37], w[38], offset); - w[48] = hc_bytealign (w[36], w[37], offset); - w[47] = hc_bytealign (w[35], w[36], offset); - w[46] = hc_bytealign (w[34], w[35], offset); - w[45] = hc_bytealign (w[33], w[34], offset); - w[44] = hc_bytealign (w[32], w[33], offset); - w[43] = hc_bytealign (w[31], w[32], offset); - w[42] = hc_bytealign (w[30], w[31], offset); - w[41] = hc_bytealign (w[29], w[30], offset); - w[40] = hc_bytealign (w[28], w[29], offset); - w[39] = hc_bytealign (w[27], w[28], offset); - w[38] = hc_bytealign (w[26], w[27], offset); - w[37] = hc_bytealign (w[25], w[26], offset); - w[36] = hc_bytealign (w[24], w[25], offset); - w[35] = hc_bytealign (w[23], w[24], offset); - w[34] = hc_bytealign (w[22], w[23], offset); - w[33] = hc_bytealign (w[21], w[22], offset); - w[32] = hc_bytealign (w[20], w[21], offset); - w[31] = hc_bytealign (w[19], w[20], offset); - w[30] = hc_bytealign (w[18], w[19], offset); - w[29] = hc_bytealign (w[17], w[18], offset); - w[28] = hc_bytealign (w[16], w[17], offset); - w[27] = hc_bytealign (w[15], w[16], offset); - w[26] = hc_bytealign (w[14], w[15], offset); - w[25] = hc_bytealign (w[13], w[14], offset); - w[24] = hc_bytealign (w[12], w[13], offset); - w[23] = hc_bytealign (w[11], w[12], offset); - w[22] = hc_bytealign (w[10], w[11], offset); - w[21] = hc_bytealign (w[ 9], w[10], offset); - w[20] = hc_bytealign (w[ 8], w[ 9], offset); - w[19] = hc_bytealign (w[ 7], w[ 8], offset); - w[18] = hc_bytealign (w[ 6], w[ 7], offset); - w[17] = hc_bytealign (w[ 5], w[ 6], offset); - w[16] = hc_bytealign (w[ 4], w[ 5], offset); - w[15] = hc_bytealign (w[ 3], w[ 4], offset); - w[14] = hc_bytealign (w[ 2], w[ 3], offset); - w[13] = hc_bytealign (w[ 1], w[ 2], offset); - w[12] = hc_bytealign (w[ 0], w[ 1], offset); - w[11] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[51], w[52], offset); + w[62] = hc_bytealign_be (w[50], w[51], offset); + w[61] = hc_bytealign_be (w[49], w[50], offset); + w[60] = hc_bytealign_be (w[48], w[49], offset); + w[59] = hc_bytealign_be (w[47], w[48], offset); + w[58] = hc_bytealign_be (w[46], w[47], offset); + w[57] = hc_bytealign_be (w[45], w[46], offset); + w[56] = hc_bytealign_be (w[44], w[45], offset); + w[55] = hc_bytealign_be (w[43], w[44], offset); + w[54] = hc_bytealign_be (w[42], w[43], offset); + w[53] = hc_bytealign_be (w[41], w[42], offset); + w[52] = hc_bytealign_be (w[40], w[41], offset); + w[51] = hc_bytealign_be (w[39], w[40], offset); + w[50] = hc_bytealign_be (w[38], w[39], offset); + w[49] = hc_bytealign_be (w[37], w[38], offset); + w[48] = hc_bytealign_be (w[36], w[37], offset); + w[47] = hc_bytealign_be (w[35], w[36], offset); + w[46] = hc_bytealign_be (w[34], w[35], offset); + w[45] = hc_bytealign_be (w[33], w[34], offset); + w[44] = hc_bytealign_be (w[32], w[33], offset); + w[43] = hc_bytealign_be (w[31], w[32], offset); + w[42] = hc_bytealign_be (w[30], w[31], offset); + w[41] = hc_bytealign_be (w[29], w[30], offset); + w[40] = hc_bytealign_be (w[28], w[29], offset); + w[39] = hc_bytealign_be (w[27], w[28], offset); + w[38] = hc_bytealign_be (w[26], w[27], offset); + w[37] = hc_bytealign_be (w[25], w[26], offset); + w[36] = hc_bytealign_be (w[24], w[25], offset); + w[35] = hc_bytealign_be (w[23], w[24], offset); + w[34] = hc_bytealign_be (w[22], w[23], offset); + w[33] = hc_bytealign_be (w[21], w[22], offset); + w[32] = hc_bytealign_be (w[20], w[21], offset); + w[31] = hc_bytealign_be (w[19], w[20], offset); + w[30] = hc_bytealign_be (w[18], w[19], offset); + w[29] = hc_bytealign_be (w[17], w[18], offset); + w[28] = hc_bytealign_be (w[16], w[17], offset); + w[27] = hc_bytealign_be (w[15], w[16], offset); + w[26] = hc_bytealign_be (w[14], w[15], offset); + w[25] = hc_bytealign_be (w[13], w[14], offset); + w[24] = hc_bytealign_be (w[12], w[13], offset); + w[23] = hc_bytealign_be (w[11], w[12], offset); + w[22] = hc_bytealign_be (w[10], w[11], offset); + w[21] = hc_bytealign_be (w[ 9], w[10], offset); + w[20] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[19] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[18] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[17] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[16] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[15] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[14] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[13] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[12] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[11] = hc_bytealign_be ( 0, w[ 0], offset); w[10] = 0; w[ 9] = 0; w[ 8] = 0; @@ -21421,58 +21259,58 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 12: - w[63] = hc_bytealign (w[50], w[51], offset); - w[62] = hc_bytealign (w[49], w[50], offset); - w[61] = hc_bytealign (w[48], w[49], offset); - w[60] = hc_bytealign (w[47], w[48], offset); - w[59] = hc_bytealign (w[46], w[47], offset); - w[58] = hc_bytealign (w[45], w[46], offset); - w[57] = hc_bytealign (w[44], w[45], offset); - w[56] = hc_bytealign (w[43], w[44], offset); - w[55] = hc_bytealign (w[42], w[43], offset); - w[54] = hc_bytealign (w[41], w[42], offset); - w[53] = hc_bytealign (w[40], w[41], offset); - w[52] = hc_bytealign (w[39], w[40], offset); - w[51] = hc_bytealign (w[38], w[39], offset); - w[50] = hc_bytealign (w[37], w[38], offset); - w[49] = hc_bytealign (w[36], w[37], offset); - w[48] = hc_bytealign (w[35], w[36], offset); - w[47] = hc_bytealign (w[34], w[35], offset); - w[46] = hc_bytealign (w[33], w[34], offset); - w[45] = hc_bytealign (w[32], w[33], offset); - w[44] = hc_bytealign (w[31], w[32], offset); - w[43] = hc_bytealign (w[30], w[31], offset); - w[42] = hc_bytealign (w[29], w[30], offset); - w[41] = hc_bytealign (w[28], w[29], offset); - w[40] = hc_bytealign (w[27], w[28], offset); - w[39] = hc_bytealign (w[26], w[27], offset); - w[38] = hc_bytealign (w[25], w[26], offset); - w[37] = hc_bytealign (w[24], w[25], offset); - w[36] = hc_bytealign (w[23], w[24], offset); - w[35] = hc_bytealign (w[22], w[23], offset); - w[34] = hc_bytealign (w[21], w[22], offset); - w[33] = hc_bytealign (w[20], w[21], offset); - w[32] = hc_bytealign (w[19], w[20], offset); - w[31] = hc_bytealign (w[18], w[19], offset); - w[30] = hc_bytealign (w[17], w[18], offset); - w[29] = hc_bytealign (w[16], w[17], offset); - w[28] = hc_bytealign (w[15], w[16], offset); - w[27] = hc_bytealign (w[14], w[15], offset); - w[26] = hc_bytealign (w[13], w[14], offset); - w[25] = hc_bytealign (w[12], w[13], offset); - w[24] = hc_bytealign (w[11], w[12], offset); - w[23] = hc_bytealign (w[10], w[11], offset); - w[22] = hc_bytealign (w[ 9], w[10], offset); - w[21] = hc_bytealign (w[ 8], w[ 9], offset); - w[20] = hc_bytealign (w[ 7], w[ 8], offset); - w[19] = hc_bytealign (w[ 6], w[ 7], offset); - w[18] = hc_bytealign (w[ 5], w[ 6], offset); - w[17] = hc_bytealign (w[ 4], w[ 5], offset); - w[16] = hc_bytealign (w[ 3], w[ 4], offset); - w[15] = hc_bytealign (w[ 2], w[ 3], offset); - w[14] = hc_bytealign (w[ 1], w[ 2], offset); - w[13] = hc_bytealign (w[ 0], w[ 1], offset); - w[12] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[50], w[51], offset); + w[62] = hc_bytealign_be (w[49], w[50], offset); + w[61] = hc_bytealign_be (w[48], w[49], offset); + w[60] = hc_bytealign_be (w[47], w[48], offset); + w[59] = hc_bytealign_be (w[46], w[47], offset); + w[58] = hc_bytealign_be (w[45], w[46], offset); + w[57] = hc_bytealign_be (w[44], w[45], offset); + w[56] = hc_bytealign_be (w[43], w[44], offset); + w[55] = hc_bytealign_be (w[42], w[43], offset); + w[54] = hc_bytealign_be (w[41], w[42], offset); + w[53] = hc_bytealign_be (w[40], w[41], offset); + w[52] = hc_bytealign_be (w[39], w[40], offset); + w[51] = hc_bytealign_be (w[38], w[39], offset); + w[50] = hc_bytealign_be (w[37], w[38], offset); + w[49] = hc_bytealign_be (w[36], w[37], offset); + w[48] = hc_bytealign_be (w[35], w[36], offset); + w[47] = hc_bytealign_be (w[34], w[35], offset); + w[46] = hc_bytealign_be (w[33], w[34], offset); + w[45] = hc_bytealign_be (w[32], w[33], offset); + w[44] = hc_bytealign_be (w[31], w[32], offset); + w[43] = hc_bytealign_be (w[30], w[31], offset); + w[42] = hc_bytealign_be (w[29], w[30], offset); + w[41] = hc_bytealign_be (w[28], w[29], offset); + w[40] = hc_bytealign_be (w[27], w[28], offset); + w[39] = hc_bytealign_be (w[26], w[27], offset); + w[38] = hc_bytealign_be (w[25], w[26], offset); + w[37] = hc_bytealign_be (w[24], w[25], offset); + w[36] = hc_bytealign_be (w[23], w[24], offset); + w[35] = hc_bytealign_be (w[22], w[23], offset); + w[34] = hc_bytealign_be (w[21], w[22], offset); + w[33] = hc_bytealign_be (w[20], w[21], offset); + w[32] = hc_bytealign_be (w[19], w[20], offset); + w[31] = hc_bytealign_be (w[18], w[19], offset); + w[30] = hc_bytealign_be (w[17], w[18], offset); + w[29] = hc_bytealign_be (w[16], w[17], offset); + w[28] = hc_bytealign_be (w[15], w[16], offset); + w[27] = hc_bytealign_be (w[14], w[15], offset); + w[26] = hc_bytealign_be (w[13], w[14], offset); + w[25] = hc_bytealign_be (w[12], w[13], offset); + w[24] = hc_bytealign_be (w[11], w[12], offset); + w[23] = hc_bytealign_be (w[10], w[11], offset); + w[22] = hc_bytealign_be (w[ 9], w[10], offset); + w[21] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[20] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[19] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[18] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[17] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[16] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[15] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[14] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[13] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[12] = hc_bytealign_be ( 0, w[ 0], offset); w[11] = 0; w[10] = 0; w[ 9] = 0; @@ -21489,57 +21327,57 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 13: - w[63] = hc_bytealign (w[49], w[50], offset); - w[62] = hc_bytealign (w[48], w[49], offset); - w[61] = hc_bytealign (w[47], w[48], offset); - w[60] = hc_bytealign (w[46], w[47], offset); - w[59] = hc_bytealign (w[45], w[46], offset); - w[58] = hc_bytealign (w[44], w[45], offset); - w[57] = hc_bytealign (w[43], w[44], offset); - w[56] = hc_bytealign (w[42], w[43], offset); - w[55] = hc_bytealign (w[41], w[42], offset); - w[54] = hc_bytealign (w[40], w[41], offset); - w[53] = hc_bytealign (w[39], w[40], offset); - w[52] = hc_bytealign (w[38], w[39], offset); - w[51] = hc_bytealign (w[37], w[38], offset); - w[50] = hc_bytealign (w[36], w[37], offset); - w[49] = hc_bytealign (w[35], w[36], offset); - w[48] = hc_bytealign (w[34], w[35], offset); - w[47] = hc_bytealign (w[33], w[34], offset); - w[46] = hc_bytealign (w[32], w[33], offset); - w[45] = hc_bytealign (w[31], w[32], offset); - w[44] = hc_bytealign (w[30], w[31], offset); - w[43] = hc_bytealign (w[29], w[30], offset); - w[42] = hc_bytealign (w[28], w[29], offset); - w[41] = hc_bytealign (w[27], w[28], offset); - w[40] = hc_bytealign (w[26], w[27], offset); - w[39] = hc_bytealign (w[25], w[26], offset); - w[38] = hc_bytealign (w[24], w[25], offset); - w[37] = hc_bytealign (w[23], w[24], offset); - w[36] = hc_bytealign (w[22], w[23], offset); - w[35] = hc_bytealign (w[21], w[22], offset); - w[34] = hc_bytealign (w[20], w[21], offset); - w[33] = hc_bytealign (w[19], w[20], offset); - w[32] = hc_bytealign (w[18], w[19], offset); - w[31] = hc_bytealign (w[17], w[18], offset); - w[30] = hc_bytealign (w[16], w[17], offset); - w[29] = hc_bytealign (w[15], w[16], offset); - w[28] = hc_bytealign (w[14], w[15], offset); - w[27] = hc_bytealign (w[13], w[14], offset); - w[26] = hc_bytealign (w[12], w[13], offset); - w[25] = hc_bytealign (w[11], w[12], offset); - w[24] = hc_bytealign (w[10], w[11], offset); - w[23] = hc_bytealign (w[ 9], w[10], offset); - w[22] = hc_bytealign (w[ 8], w[ 9], offset); - w[21] = hc_bytealign (w[ 7], w[ 8], offset); - w[20] = hc_bytealign (w[ 6], w[ 7], offset); - w[19] = hc_bytealign (w[ 5], w[ 6], offset); - w[18] = hc_bytealign (w[ 4], w[ 5], offset); - w[17] = hc_bytealign (w[ 3], w[ 4], offset); - w[16] = hc_bytealign (w[ 2], w[ 3], offset); - w[15] = hc_bytealign (w[ 1], w[ 2], offset); - w[14] = hc_bytealign (w[ 0], w[ 1], offset); - w[13] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[49], w[50], offset); + w[62] = hc_bytealign_be (w[48], w[49], offset); + w[61] = hc_bytealign_be (w[47], w[48], offset); + w[60] = hc_bytealign_be (w[46], w[47], offset); + w[59] = hc_bytealign_be (w[45], w[46], offset); + w[58] = hc_bytealign_be (w[44], w[45], offset); + w[57] = hc_bytealign_be (w[43], w[44], offset); + w[56] = hc_bytealign_be (w[42], w[43], offset); + w[55] = hc_bytealign_be (w[41], w[42], offset); + w[54] = hc_bytealign_be (w[40], w[41], offset); + w[53] = hc_bytealign_be (w[39], w[40], offset); + w[52] = hc_bytealign_be (w[38], w[39], offset); + w[51] = hc_bytealign_be (w[37], w[38], offset); + w[50] = hc_bytealign_be (w[36], w[37], offset); + w[49] = hc_bytealign_be (w[35], w[36], offset); + w[48] = hc_bytealign_be (w[34], w[35], offset); + w[47] = hc_bytealign_be (w[33], w[34], offset); + w[46] = hc_bytealign_be (w[32], w[33], offset); + w[45] = hc_bytealign_be (w[31], w[32], offset); + w[44] = hc_bytealign_be (w[30], w[31], offset); + w[43] = hc_bytealign_be (w[29], w[30], offset); + w[42] = hc_bytealign_be (w[28], w[29], offset); + w[41] = hc_bytealign_be (w[27], w[28], offset); + w[40] = hc_bytealign_be (w[26], w[27], offset); + w[39] = hc_bytealign_be (w[25], w[26], offset); + w[38] = hc_bytealign_be (w[24], w[25], offset); + w[37] = hc_bytealign_be (w[23], w[24], offset); + w[36] = hc_bytealign_be (w[22], w[23], offset); + w[35] = hc_bytealign_be (w[21], w[22], offset); + w[34] = hc_bytealign_be (w[20], w[21], offset); + w[33] = hc_bytealign_be (w[19], w[20], offset); + w[32] = hc_bytealign_be (w[18], w[19], offset); + w[31] = hc_bytealign_be (w[17], w[18], offset); + w[30] = hc_bytealign_be (w[16], w[17], offset); + w[29] = hc_bytealign_be (w[15], w[16], offset); + w[28] = hc_bytealign_be (w[14], w[15], offset); + w[27] = hc_bytealign_be (w[13], w[14], offset); + w[26] = hc_bytealign_be (w[12], w[13], offset); + w[25] = hc_bytealign_be (w[11], w[12], offset); + w[24] = hc_bytealign_be (w[10], w[11], offset); + w[23] = hc_bytealign_be (w[ 9], w[10], offset); + w[22] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[21] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[20] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[19] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[18] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[17] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[16] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[15] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[14] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[13] = hc_bytealign_be ( 0, w[ 0], offset); w[12] = 0; w[11] = 0; w[10] = 0; @@ -21557,56 +21395,56 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 14: - w[63] = hc_bytealign (w[48], w[49], offset); - w[62] = hc_bytealign (w[47], w[48], offset); - w[61] = hc_bytealign (w[46], w[47], offset); - w[60] = hc_bytealign (w[45], w[46], offset); - w[59] = hc_bytealign (w[44], w[45], offset); - w[58] = hc_bytealign (w[43], w[44], offset); - w[57] = hc_bytealign (w[42], w[43], offset); - w[56] = hc_bytealign (w[41], w[42], offset); - w[55] = hc_bytealign (w[40], w[41], offset); - w[54] = hc_bytealign (w[39], w[40], offset); - w[53] = hc_bytealign (w[38], w[39], offset); - w[52] = hc_bytealign (w[37], w[38], offset); - w[51] = hc_bytealign (w[36], w[37], offset); - w[50] = hc_bytealign (w[35], w[36], offset); - w[49] = hc_bytealign (w[34], w[35], offset); - w[48] = hc_bytealign (w[33], w[34], offset); - w[47] = hc_bytealign (w[32], w[33], offset); - w[46] = hc_bytealign (w[31], w[32], offset); - w[45] = hc_bytealign (w[30], w[31], offset); - w[44] = hc_bytealign (w[29], w[30], offset); - w[43] = hc_bytealign (w[28], w[29], offset); - w[42] = hc_bytealign (w[27], w[28], offset); - w[41] = hc_bytealign (w[26], w[27], offset); - w[40] = hc_bytealign (w[25], w[26], offset); - w[39] = hc_bytealign (w[24], w[25], offset); - w[38] = hc_bytealign (w[23], w[24], offset); - w[37] = hc_bytealign (w[22], w[23], offset); - w[36] = hc_bytealign (w[21], w[22], offset); - w[35] = hc_bytealign (w[20], w[21], offset); - w[34] = hc_bytealign (w[19], w[20], offset); - w[33] = hc_bytealign (w[18], w[19], offset); - w[32] = hc_bytealign (w[17], w[18], offset); - w[31] = hc_bytealign (w[16], w[17], offset); - w[30] = hc_bytealign (w[15], w[16], offset); - w[29] = hc_bytealign (w[14], w[15], offset); - w[28] = hc_bytealign (w[13], w[14], offset); - w[27] = hc_bytealign (w[12], w[13], offset); - w[26] = hc_bytealign (w[11], w[12], offset); - w[25] = hc_bytealign (w[10], w[11], offset); - w[24] = hc_bytealign (w[ 9], w[10], offset); - w[23] = hc_bytealign (w[ 8], w[ 9], offset); - w[22] = hc_bytealign (w[ 7], w[ 8], offset); - w[21] = hc_bytealign (w[ 6], w[ 7], offset); - w[20] = hc_bytealign (w[ 5], w[ 6], offset); - w[19] = hc_bytealign (w[ 4], w[ 5], offset); - w[18] = hc_bytealign (w[ 3], w[ 4], offset); - w[17] = hc_bytealign (w[ 2], w[ 3], offset); - w[16] = hc_bytealign (w[ 1], w[ 2], offset); - w[15] = hc_bytealign (w[ 0], w[ 1], offset); - w[14] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[48], w[49], offset); + w[62] = hc_bytealign_be (w[47], w[48], offset); + w[61] = hc_bytealign_be (w[46], w[47], offset); + w[60] = hc_bytealign_be (w[45], w[46], offset); + w[59] = hc_bytealign_be (w[44], w[45], offset); + w[58] = hc_bytealign_be (w[43], w[44], offset); + w[57] = hc_bytealign_be (w[42], w[43], offset); + w[56] = hc_bytealign_be (w[41], w[42], offset); + w[55] = hc_bytealign_be (w[40], w[41], offset); + w[54] = hc_bytealign_be (w[39], w[40], offset); + w[53] = hc_bytealign_be (w[38], w[39], offset); + w[52] = hc_bytealign_be (w[37], w[38], offset); + w[51] = hc_bytealign_be (w[36], w[37], offset); + w[50] = hc_bytealign_be (w[35], w[36], offset); + w[49] = hc_bytealign_be (w[34], w[35], offset); + w[48] = hc_bytealign_be (w[33], w[34], offset); + w[47] = hc_bytealign_be (w[32], w[33], offset); + w[46] = hc_bytealign_be (w[31], w[32], offset); + w[45] = hc_bytealign_be (w[30], w[31], offset); + w[44] = hc_bytealign_be (w[29], w[30], offset); + w[43] = hc_bytealign_be (w[28], w[29], offset); + w[42] = hc_bytealign_be (w[27], w[28], offset); + w[41] = hc_bytealign_be (w[26], w[27], offset); + w[40] = hc_bytealign_be (w[25], w[26], offset); + w[39] = hc_bytealign_be (w[24], w[25], offset); + w[38] = hc_bytealign_be (w[23], w[24], offset); + w[37] = hc_bytealign_be (w[22], w[23], offset); + w[36] = hc_bytealign_be (w[21], w[22], offset); + w[35] = hc_bytealign_be (w[20], w[21], offset); + w[34] = hc_bytealign_be (w[19], w[20], offset); + w[33] = hc_bytealign_be (w[18], w[19], offset); + w[32] = hc_bytealign_be (w[17], w[18], offset); + w[31] = hc_bytealign_be (w[16], w[17], offset); + w[30] = hc_bytealign_be (w[15], w[16], offset); + w[29] = hc_bytealign_be (w[14], w[15], offset); + w[28] = hc_bytealign_be (w[13], w[14], offset); + w[27] = hc_bytealign_be (w[12], w[13], offset); + w[26] = hc_bytealign_be (w[11], w[12], offset); + w[25] = hc_bytealign_be (w[10], w[11], offset); + w[24] = hc_bytealign_be (w[ 9], w[10], offset); + w[23] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[22] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[21] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[20] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[19] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[18] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[17] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[16] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[15] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[14] = hc_bytealign_be ( 0, w[ 0], offset); w[13] = 0; w[12] = 0; w[11] = 0; @@ -21625,55 +21463,55 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 15: - w[63] = hc_bytealign (w[47], w[48], offset); - w[62] = hc_bytealign (w[46], w[47], offset); - w[61] = hc_bytealign (w[45], w[46], offset); - w[60] = hc_bytealign (w[44], w[45], offset); - w[59] = hc_bytealign (w[43], w[44], offset); - w[58] = hc_bytealign (w[42], w[43], offset); - w[57] = hc_bytealign (w[41], w[42], offset); - w[56] = hc_bytealign (w[40], w[41], offset); - w[55] = hc_bytealign (w[39], w[40], offset); - w[54] = hc_bytealign (w[38], w[39], offset); - w[53] = hc_bytealign (w[37], w[38], offset); - w[52] = hc_bytealign (w[36], w[37], offset); - w[51] = hc_bytealign (w[35], w[36], offset); - w[50] = hc_bytealign (w[34], w[35], offset); - w[49] = hc_bytealign (w[33], w[34], offset); - w[48] = hc_bytealign (w[32], w[33], offset); - w[47] = hc_bytealign (w[31], w[32], offset); - w[46] = hc_bytealign (w[30], w[31], offset); - w[45] = hc_bytealign (w[29], w[30], offset); - w[44] = hc_bytealign (w[28], w[29], offset); - w[43] = hc_bytealign (w[27], w[28], offset); - w[42] = hc_bytealign (w[26], w[27], offset); - w[41] = hc_bytealign (w[25], w[26], offset); - w[40] = hc_bytealign (w[24], w[25], offset); - w[39] = hc_bytealign (w[23], w[24], offset); - w[38] = hc_bytealign (w[22], w[23], offset); - w[37] = hc_bytealign (w[21], w[22], offset); - w[36] = hc_bytealign (w[20], w[21], offset); - w[35] = hc_bytealign (w[19], w[20], offset); - w[34] = hc_bytealign (w[18], w[19], offset); - w[33] = hc_bytealign (w[17], w[18], offset); - w[32] = hc_bytealign (w[16], w[17], offset); - w[31] = hc_bytealign (w[15], w[16], offset); - w[30] = hc_bytealign (w[14], w[15], offset); - w[29] = hc_bytealign (w[13], w[14], offset); - w[28] = hc_bytealign (w[12], w[13], offset); - w[27] = hc_bytealign (w[11], w[12], offset); - w[26] = hc_bytealign (w[10], w[11], offset); - w[25] = hc_bytealign (w[ 9], w[10], offset); - w[24] = hc_bytealign (w[ 8], w[ 9], offset); - w[23] = hc_bytealign (w[ 7], w[ 8], offset); - w[22] = hc_bytealign (w[ 6], w[ 7], offset); - w[21] = hc_bytealign (w[ 5], w[ 6], offset); - w[20] = hc_bytealign (w[ 4], w[ 5], offset); - w[19] = hc_bytealign (w[ 3], w[ 4], offset); - w[18] = hc_bytealign (w[ 2], w[ 3], offset); - w[17] = hc_bytealign (w[ 1], w[ 2], offset); - w[16] = hc_bytealign (w[ 0], w[ 1], offset); - w[15] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[47], w[48], offset); + w[62] = hc_bytealign_be (w[46], w[47], offset); + w[61] = hc_bytealign_be (w[45], w[46], offset); + w[60] = hc_bytealign_be (w[44], w[45], offset); + w[59] = hc_bytealign_be (w[43], w[44], offset); + w[58] = hc_bytealign_be (w[42], w[43], offset); + w[57] = hc_bytealign_be (w[41], w[42], offset); + w[56] = hc_bytealign_be (w[40], w[41], offset); + w[55] = hc_bytealign_be (w[39], w[40], offset); + w[54] = hc_bytealign_be (w[38], w[39], offset); + w[53] = hc_bytealign_be (w[37], w[38], offset); + w[52] = hc_bytealign_be (w[36], w[37], offset); + w[51] = hc_bytealign_be (w[35], w[36], offset); + w[50] = hc_bytealign_be (w[34], w[35], offset); + w[49] = hc_bytealign_be (w[33], w[34], offset); + w[48] = hc_bytealign_be (w[32], w[33], offset); + w[47] = hc_bytealign_be (w[31], w[32], offset); + w[46] = hc_bytealign_be (w[30], w[31], offset); + w[45] = hc_bytealign_be (w[29], w[30], offset); + w[44] = hc_bytealign_be (w[28], w[29], offset); + w[43] = hc_bytealign_be (w[27], w[28], offset); + w[42] = hc_bytealign_be (w[26], w[27], offset); + w[41] = hc_bytealign_be (w[25], w[26], offset); + w[40] = hc_bytealign_be (w[24], w[25], offset); + w[39] = hc_bytealign_be (w[23], w[24], offset); + w[38] = hc_bytealign_be (w[22], w[23], offset); + w[37] = hc_bytealign_be (w[21], w[22], offset); + w[36] = hc_bytealign_be (w[20], w[21], offset); + w[35] = hc_bytealign_be (w[19], w[20], offset); + w[34] = hc_bytealign_be (w[18], w[19], offset); + w[33] = hc_bytealign_be (w[17], w[18], offset); + w[32] = hc_bytealign_be (w[16], w[17], offset); + w[31] = hc_bytealign_be (w[15], w[16], offset); + w[30] = hc_bytealign_be (w[14], w[15], offset); + w[29] = hc_bytealign_be (w[13], w[14], offset); + w[28] = hc_bytealign_be (w[12], w[13], offset); + w[27] = hc_bytealign_be (w[11], w[12], offset); + w[26] = hc_bytealign_be (w[10], w[11], offset); + w[25] = hc_bytealign_be (w[ 9], w[10], offset); + w[24] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[23] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[22] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[21] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[20] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[19] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[18] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[17] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[16] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[15] = hc_bytealign_be ( 0, w[ 0], offset); w[14] = 0; w[13] = 0; w[12] = 0; @@ -21693,54 +21531,54 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 16: - w[63] = hc_bytealign (w[46], w[47], offset); - w[62] = hc_bytealign (w[45], w[46], offset); - w[61] = hc_bytealign (w[44], w[45], offset); - w[60] = hc_bytealign (w[43], w[44], offset); - w[59] = hc_bytealign (w[42], w[43], offset); - w[58] = hc_bytealign (w[41], w[42], offset); - w[57] = hc_bytealign (w[40], w[41], offset); - w[56] = hc_bytealign (w[39], w[40], offset); - w[55] = hc_bytealign (w[38], w[39], offset); - w[54] = hc_bytealign (w[37], w[38], offset); - w[53] = hc_bytealign (w[36], w[37], offset); - w[52] = hc_bytealign (w[35], w[36], offset); - w[51] = hc_bytealign (w[34], w[35], offset); - w[50] = hc_bytealign (w[33], w[34], offset); - w[49] = hc_bytealign (w[32], w[33], offset); - w[48] = hc_bytealign (w[31], w[32], offset); - w[47] = hc_bytealign (w[30], w[31], offset); - w[46] = hc_bytealign (w[29], w[30], offset); - w[45] = hc_bytealign (w[28], w[29], offset); - w[44] = hc_bytealign (w[27], w[28], offset); - w[43] = hc_bytealign (w[26], w[27], offset); - w[42] = hc_bytealign (w[25], w[26], offset); - w[41] = hc_bytealign (w[24], w[25], offset); - w[40] = hc_bytealign (w[23], w[24], offset); - w[39] = hc_bytealign (w[22], w[23], offset); - w[38] = hc_bytealign (w[21], w[22], offset); - w[37] = hc_bytealign (w[20], w[21], offset); - w[36] = hc_bytealign (w[19], w[20], offset); - w[35] = hc_bytealign (w[18], w[19], offset); - w[34] = hc_bytealign (w[17], w[18], offset); - w[33] = hc_bytealign (w[16], w[17], offset); - w[32] = hc_bytealign (w[15], w[16], offset); - w[31] = hc_bytealign (w[14], w[15], offset); - w[30] = hc_bytealign (w[13], w[14], offset); - w[29] = hc_bytealign (w[12], w[13], offset); - w[28] = hc_bytealign (w[11], w[12], offset); - w[27] = hc_bytealign (w[10], w[11], offset); - w[26] = hc_bytealign (w[ 9], w[10], offset); - w[25] = hc_bytealign (w[ 8], w[ 9], offset); - w[24] = hc_bytealign (w[ 7], w[ 8], offset); - w[23] = hc_bytealign (w[ 6], w[ 7], offset); - w[22] = hc_bytealign (w[ 5], w[ 6], offset); - w[21] = hc_bytealign (w[ 4], w[ 5], offset); - w[20] = hc_bytealign (w[ 3], w[ 4], offset); - w[19] = hc_bytealign (w[ 2], w[ 3], offset); - w[18] = hc_bytealign (w[ 1], w[ 2], offset); - w[17] = hc_bytealign (w[ 0], w[ 1], offset); - w[16] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[46], w[47], offset); + w[62] = hc_bytealign_be (w[45], w[46], offset); + w[61] = hc_bytealign_be (w[44], w[45], offset); + w[60] = hc_bytealign_be (w[43], w[44], offset); + w[59] = hc_bytealign_be (w[42], w[43], offset); + w[58] = hc_bytealign_be (w[41], w[42], offset); + w[57] = hc_bytealign_be (w[40], w[41], offset); + w[56] = hc_bytealign_be (w[39], w[40], offset); + w[55] = hc_bytealign_be (w[38], w[39], offset); + w[54] = hc_bytealign_be (w[37], w[38], offset); + w[53] = hc_bytealign_be (w[36], w[37], offset); + w[52] = hc_bytealign_be (w[35], w[36], offset); + w[51] = hc_bytealign_be (w[34], w[35], offset); + w[50] = hc_bytealign_be (w[33], w[34], offset); + w[49] = hc_bytealign_be (w[32], w[33], offset); + w[48] = hc_bytealign_be (w[31], w[32], offset); + w[47] = hc_bytealign_be (w[30], w[31], offset); + w[46] = hc_bytealign_be (w[29], w[30], offset); + w[45] = hc_bytealign_be (w[28], w[29], offset); + w[44] = hc_bytealign_be (w[27], w[28], offset); + w[43] = hc_bytealign_be (w[26], w[27], offset); + w[42] = hc_bytealign_be (w[25], w[26], offset); + w[41] = hc_bytealign_be (w[24], w[25], offset); + w[40] = hc_bytealign_be (w[23], w[24], offset); + w[39] = hc_bytealign_be (w[22], w[23], offset); + w[38] = hc_bytealign_be (w[21], w[22], offset); + w[37] = hc_bytealign_be (w[20], w[21], offset); + w[36] = hc_bytealign_be (w[19], w[20], offset); + w[35] = hc_bytealign_be (w[18], w[19], offset); + w[34] = hc_bytealign_be (w[17], w[18], offset); + w[33] = hc_bytealign_be (w[16], w[17], offset); + w[32] = hc_bytealign_be (w[15], w[16], offset); + w[31] = hc_bytealign_be (w[14], w[15], offset); + w[30] = hc_bytealign_be (w[13], w[14], offset); + w[29] = hc_bytealign_be (w[12], w[13], offset); + w[28] = hc_bytealign_be (w[11], w[12], offset); + w[27] = hc_bytealign_be (w[10], w[11], offset); + w[26] = hc_bytealign_be (w[ 9], w[10], offset); + w[25] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[24] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[23] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[22] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[21] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[20] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[19] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[18] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[17] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[16] = hc_bytealign_be ( 0, w[ 0], offset); w[15] = 0; w[14] = 0; w[13] = 0; @@ -21761,53 +21599,53 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 17: - w[63] = hc_bytealign (w[45], w[46], offset); - w[62] = hc_bytealign (w[44], w[45], offset); - w[61] = hc_bytealign (w[43], w[44], offset); - w[60] = hc_bytealign (w[42], w[43], offset); - w[59] = hc_bytealign (w[41], w[42], offset); - w[58] = hc_bytealign (w[40], w[41], offset); - w[57] = hc_bytealign (w[39], w[40], offset); - w[56] = hc_bytealign (w[38], w[39], offset); - w[55] = hc_bytealign (w[37], w[38], offset); - w[54] = hc_bytealign (w[36], w[37], offset); - w[53] = hc_bytealign (w[35], w[36], offset); - w[52] = hc_bytealign (w[34], w[35], offset); - w[51] = hc_bytealign (w[33], w[34], offset); - w[50] = hc_bytealign (w[32], w[33], offset); - w[49] = hc_bytealign (w[31], w[32], offset); - w[48] = hc_bytealign (w[30], w[31], offset); - w[47] = hc_bytealign (w[29], w[30], offset); - w[46] = hc_bytealign (w[28], w[29], offset); - w[45] = hc_bytealign (w[27], w[28], offset); - w[44] = hc_bytealign (w[26], w[27], offset); - w[43] = hc_bytealign (w[25], w[26], offset); - w[42] = hc_bytealign (w[24], w[25], offset); - w[41] = hc_bytealign (w[23], w[24], offset); - w[40] = hc_bytealign (w[22], w[23], offset); - w[39] = hc_bytealign (w[21], w[22], offset); - w[38] = hc_bytealign (w[20], w[21], offset); - w[37] = hc_bytealign (w[19], w[20], offset); - w[36] = hc_bytealign (w[18], w[19], offset); - w[35] = hc_bytealign (w[17], w[18], offset); - w[34] = hc_bytealign (w[16], w[17], offset); - w[33] = hc_bytealign (w[15], w[16], offset); - w[32] = hc_bytealign (w[14], w[15], offset); - w[31] = hc_bytealign (w[13], w[14], offset); - w[30] = hc_bytealign (w[12], w[13], offset); - w[29] = hc_bytealign (w[11], w[12], offset); - w[28] = hc_bytealign (w[10], w[11], offset); - w[27] = hc_bytealign (w[ 9], w[10], offset); - w[26] = hc_bytealign (w[ 8], w[ 9], offset); - w[25] = hc_bytealign (w[ 7], w[ 8], offset); - w[24] = hc_bytealign (w[ 6], w[ 7], offset); - w[23] = hc_bytealign (w[ 5], w[ 6], offset); - w[22] = hc_bytealign (w[ 4], w[ 5], offset); - w[21] = hc_bytealign (w[ 3], w[ 4], offset); - w[20] = hc_bytealign (w[ 2], w[ 3], offset); - w[19] = hc_bytealign (w[ 1], w[ 2], offset); - w[18] = hc_bytealign (w[ 0], w[ 1], offset); - w[17] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[45], w[46], offset); + w[62] = hc_bytealign_be (w[44], w[45], offset); + w[61] = hc_bytealign_be (w[43], w[44], offset); + w[60] = hc_bytealign_be (w[42], w[43], offset); + w[59] = hc_bytealign_be (w[41], w[42], offset); + w[58] = hc_bytealign_be (w[40], w[41], offset); + w[57] = hc_bytealign_be (w[39], w[40], offset); + w[56] = hc_bytealign_be (w[38], w[39], offset); + w[55] = hc_bytealign_be (w[37], w[38], offset); + w[54] = hc_bytealign_be (w[36], w[37], offset); + w[53] = hc_bytealign_be (w[35], w[36], offset); + w[52] = hc_bytealign_be (w[34], w[35], offset); + w[51] = hc_bytealign_be (w[33], w[34], offset); + w[50] = hc_bytealign_be (w[32], w[33], offset); + w[49] = hc_bytealign_be (w[31], w[32], offset); + w[48] = hc_bytealign_be (w[30], w[31], offset); + w[47] = hc_bytealign_be (w[29], w[30], offset); + w[46] = hc_bytealign_be (w[28], w[29], offset); + w[45] = hc_bytealign_be (w[27], w[28], offset); + w[44] = hc_bytealign_be (w[26], w[27], offset); + w[43] = hc_bytealign_be (w[25], w[26], offset); + w[42] = hc_bytealign_be (w[24], w[25], offset); + w[41] = hc_bytealign_be (w[23], w[24], offset); + w[40] = hc_bytealign_be (w[22], w[23], offset); + w[39] = hc_bytealign_be (w[21], w[22], offset); + w[38] = hc_bytealign_be (w[20], w[21], offset); + w[37] = hc_bytealign_be (w[19], w[20], offset); + w[36] = hc_bytealign_be (w[18], w[19], offset); + w[35] = hc_bytealign_be (w[17], w[18], offset); + w[34] = hc_bytealign_be (w[16], w[17], offset); + w[33] = hc_bytealign_be (w[15], w[16], offset); + w[32] = hc_bytealign_be (w[14], w[15], offset); + w[31] = hc_bytealign_be (w[13], w[14], offset); + w[30] = hc_bytealign_be (w[12], w[13], offset); + w[29] = hc_bytealign_be (w[11], w[12], offset); + w[28] = hc_bytealign_be (w[10], w[11], offset); + w[27] = hc_bytealign_be (w[ 9], w[10], offset); + w[26] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[25] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[24] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[23] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[22] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[21] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[20] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[19] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[18] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[17] = hc_bytealign_be ( 0, w[ 0], offset); w[16] = 0; w[15] = 0; w[14] = 0; @@ -21829,52 +21667,52 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 18: - w[63] = hc_bytealign (w[44], w[45], offset); - w[62] = hc_bytealign (w[43], w[44], offset); - w[61] = hc_bytealign (w[42], w[43], offset); - w[60] = hc_bytealign (w[41], w[42], offset); - w[59] = hc_bytealign (w[40], w[41], offset); - w[58] = hc_bytealign (w[39], w[40], offset); - w[57] = hc_bytealign (w[38], w[39], offset); - w[56] = hc_bytealign (w[37], w[38], offset); - w[55] = hc_bytealign (w[36], w[37], offset); - w[54] = hc_bytealign (w[35], w[36], offset); - w[53] = hc_bytealign (w[34], w[35], offset); - w[52] = hc_bytealign (w[33], w[34], offset); - w[51] = hc_bytealign (w[32], w[33], offset); - w[50] = hc_bytealign (w[31], w[32], offset); - w[49] = hc_bytealign (w[30], w[31], offset); - w[48] = hc_bytealign (w[29], w[30], offset); - w[47] = hc_bytealign (w[28], w[29], offset); - w[46] = hc_bytealign (w[27], w[28], offset); - w[45] = hc_bytealign (w[26], w[27], offset); - w[44] = hc_bytealign (w[25], w[26], offset); - w[43] = hc_bytealign (w[24], w[25], offset); - w[42] = hc_bytealign (w[23], w[24], offset); - w[41] = hc_bytealign (w[22], w[23], offset); - w[40] = hc_bytealign (w[21], w[22], offset); - w[39] = hc_bytealign (w[20], w[21], offset); - w[38] = hc_bytealign (w[19], w[20], offset); - w[37] = hc_bytealign (w[18], w[19], offset); - w[36] = hc_bytealign (w[17], w[18], offset); - w[35] = hc_bytealign (w[16], w[17], offset); - w[34] = hc_bytealign (w[15], w[16], offset); - w[33] = hc_bytealign (w[14], w[15], offset); - w[32] = hc_bytealign (w[13], w[14], offset); - w[31] = hc_bytealign (w[12], w[13], offset); - w[30] = hc_bytealign (w[11], w[12], offset); - w[29] = hc_bytealign (w[10], w[11], offset); - w[28] = hc_bytealign (w[ 9], w[10], offset); - w[27] = hc_bytealign (w[ 8], w[ 9], offset); - w[26] = hc_bytealign (w[ 7], w[ 8], offset); - w[25] = hc_bytealign (w[ 6], w[ 7], offset); - w[24] = hc_bytealign (w[ 5], w[ 6], offset); - w[23] = hc_bytealign (w[ 4], w[ 5], offset); - w[22] = hc_bytealign (w[ 3], w[ 4], offset); - w[21] = hc_bytealign (w[ 2], w[ 3], offset); - w[20] = hc_bytealign (w[ 1], w[ 2], offset); - w[19] = hc_bytealign (w[ 0], w[ 1], offset); - w[18] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[44], w[45], offset); + w[62] = hc_bytealign_be (w[43], w[44], offset); + w[61] = hc_bytealign_be (w[42], w[43], offset); + w[60] = hc_bytealign_be (w[41], w[42], offset); + w[59] = hc_bytealign_be (w[40], w[41], offset); + w[58] = hc_bytealign_be (w[39], w[40], offset); + w[57] = hc_bytealign_be (w[38], w[39], offset); + w[56] = hc_bytealign_be (w[37], w[38], offset); + w[55] = hc_bytealign_be (w[36], w[37], offset); + w[54] = hc_bytealign_be (w[35], w[36], offset); + w[53] = hc_bytealign_be (w[34], w[35], offset); + w[52] = hc_bytealign_be (w[33], w[34], offset); + w[51] = hc_bytealign_be (w[32], w[33], offset); + w[50] = hc_bytealign_be (w[31], w[32], offset); + w[49] = hc_bytealign_be (w[30], w[31], offset); + w[48] = hc_bytealign_be (w[29], w[30], offset); + w[47] = hc_bytealign_be (w[28], w[29], offset); + w[46] = hc_bytealign_be (w[27], w[28], offset); + w[45] = hc_bytealign_be (w[26], w[27], offset); + w[44] = hc_bytealign_be (w[25], w[26], offset); + w[43] = hc_bytealign_be (w[24], w[25], offset); + w[42] = hc_bytealign_be (w[23], w[24], offset); + w[41] = hc_bytealign_be (w[22], w[23], offset); + w[40] = hc_bytealign_be (w[21], w[22], offset); + w[39] = hc_bytealign_be (w[20], w[21], offset); + w[38] = hc_bytealign_be (w[19], w[20], offset); + w[37] = hc_bytealign_be (w[18], w[19], offset); + w[36] = hc_bytealign_be (w[17], w[18], offset); + w[35] = hc_bytealign_be (w[16], w[17], offset); + w[34] = hc_bytealign_be (w[15], w[16], offset); + w[33] = hc_bytealign_be (w[14], w[15], offset); + w[32] = hc_bytealign_be (w[13], w[14], offset); + w[31] = hc_bytealign_be (w[12], w[13], offset); + w[30] = hc_bytealign_be (w[11], w[12], offset); + w[29] = hc_bytealign_be (w[10], w[11], offset); + w[28] = hc_bytealign_be (w[ 9], w[10], offset); + w[27] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[26] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[25] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[24] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[23] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[22] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[21] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[20] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[19] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[18] = hc_bytealign_be ( 0, w[ 0], offset); w[17] = 0; w[16] = 0; w[15] = 0; @@ -21897,51 +21735,51 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 19: - w[63] = hc_bytealign (w[43], w[44], offset); - w[62] = hc_bytealign (w[42], w[43], offset); - w[61] = hc_bytealign (w[41], w[42], offset); - w[60] = hc_bytealign (w[40], w[41], offset); - w[59] = hc_bytealign (w[39], w[40], offset); - w[58] = hc_bytealign (w[38], w[39], offset); - w[57] = hc_bytealign (w[37], w[38], offset); - w[56] = hc_bytealign (w[36], w[37], offset); - w[55] = hc_bytealign (w[35], w[36], offset); - w[54] = hc_bytealign (w[34], w[35], offset); - w[53] = hc_bytealign (w[33], w[34], offset); - w[52] = hc_bytealign (w[32], w[33], offset); - w[51] = hc_bytealign (w[31], w[32], offset); - w[50] = hc_bytealign (w[30], w[31], offset); - w[49] = hc_bytealign (w[29], w[30], offset); - w[48] = hc_bytealign (w[28], w[29], offset); - w[47] = hc_bytealign (w[27], w[28], offset); - w[46] = hc_bytealign (w[26], w[27], offset); - w[45] = hc_bytealign (w[25], w[26], offset); - w[44] = hc_bytealign (w[24], w[25], offset); - w[43] = hc_bytealign (w[23], w[24], offset); - w[42] = hc_bytealign (w[22], w[23], offset); - w[41] = hc_bytealign (w[21], w[22], offset); - w[40] = hc_bytealign (w[20], w[21], offset); - w[39] = hc_bytealign (w[19], w[20], offset); - w[38] = hc_bytealign (w[18], w[19], offset); - w[37] = hc_bytealign (w[17], w[18], offset); - w[36] = hc_bytealign (w[16], w[17], offset); - w[35] = hc_bytealign (w[15], w[16], offset); - w[34] = hc_bytealign (w[14], w[15], offset); - w[33] = hc_bytealign (w[13], w[14], offset); - w[32] = hc_bytealign (w[12], w[13], offset); - w[31] = hc_bytealign (w[11], w[12], offset); - w[30] = hc_bytealign (w[10], w[11], offset); - w[29] = hc_bytealign (w[ 9], w[10], offset); - w[28] = hc_bytealign (w[ 8], w[ 9], offset); - w[27] = hc_bytealign (w[ 7], w[ 8], offset); - w[26] = hc_bytealign (w[ 6], w[ 7], offset); - w[25] = hc_bytealign (w[ 5], w[ 6], offset); - w[24] = hc_bytealign (w[ 4], w[ 5], offset); - w[23] = hc_bytealign (w[ 3], w[ 4], offset); - w[22] = hc_bytealign (w[ 2], w[ 3], offset); - w[21] = hc_bytealign (w[ 1], w[ 2], offset); - w[20] = hc_bytealign (w[ 0], w[ 1], offset); - w[19] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[43], w[44], offset); + w[62] = hc_bytealign_be (w[42], w[43], offset); + w[61] = hc_bytealign_be (w[41], w[42], offset); + w[60] = hc_bytealign_be (w[40], w[41], offset); + w[59] = hc_bytealign_be (w[39], w[40], offset); + w[58] = hc_bytealign_be (w[38], w[39], offset); + w[57] = hc_bytealign_be (w[37], w[38], offset); + w[56] = hc_bytealign_be (w[36], w[37], offset); + w[55] = hc_bytealign_be (w[35], w[36], offset); + w[54] = hc_bytealign_be (w[34], w[35], offset); + w[53] = hc_bytealign_be (w[33], w[34], offset); + w[52] = hc_bytealign_be (w[32], w[33], offset); + w[51] = hc_bytealign_be (w[31], w[32], offset); + w[50] = hc_bytealign_be (w[30], w[31], offset); + w[49] = hc_bytealign_be (w[29], w[30], offset); + w[48] = hc_bytealign_be (w[28], w[29], offset); + w[47] = hc_bytealign_be (w[27], w[28], offset); + w[46] = hc_bytealign_be (w[26], w[27], offset); + w[45] = hc_bytealign_be (w[25], w[26], offset); + w[44] = hc_bytealign_be (w[24], w[25], offset); + w[43] = hc_bytealign_be (w[23], w[24], offset); + w[42] = hc_bytealign_be (w[22], w[23], offset); + w[41] = hc_bytealign_be (w[21], w[22], offset); + w[40] = hc_bytealign_be (w[20], w[21], offset); + w[39] = hc_bytealign_be (w[19], w[20], offset); + w[38] = hc_bytealign_be (w[18], w[19], offset); + w[37] = hc_bytealign_be (w[17], w[18], offset); + w[36] = hc_bytealign_be (w[16], w[17], offset); + w[35] = hc_bytealign_be (w[15], w[16], offset); + w[34] = hc_bytealign_be (w[14], w[15], offset); + w[33] = hc_bytealign_be (w[13], w[14], offset); + w[32] = hc_bytealign_be (w[12], w[13], offset); + w[31] = hc_bytealign_be (w[11], w[12], offset); + w[30] = hc_bytealign_be (w[10], w[11], offset); + w[29] = hc_bytealign_be (w[ 9], w[10], offset); + w[28] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[27] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[26] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[25] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[24] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[23] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[22] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[21] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[20] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[19] = hc_bytealign_be ( 0, w[ 0], offset); w[18] = 0; w[17] = 0; w[16] = 0; @@ -21965,50 +21803,50 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 20: - w[63] = hc_bytealign (w[42], w[43], offset); - w[62] = hc_bytealign (w[41], w[42], offset); - w[61] = hc_bytealign (w[40], w[41], offset); - w[60] = hc_bytealign (w[39], w[40], offset); - w[59] = hc_bytealign (w[38], w[39], offset); - w[58] = hc_bytealign (w[37], w[38], offset); - w[57] = hc_bytealign (w[36], w[37], offset); - w[56] = hc_bytealign (w[35], w[36], offset); - w[55] = hc_bytealign (w[34], w[35], offset); - w[54] = hc_bytealign (w[33], w[34], offset); - w[53] = hc_bytealign (w[32], w[33], offset); - w[52] = hc_bytealign (w[31], w[32], offset); - w[51] = hc_bytealign (w[30], w[31], offset); - w[50] = hc_bytealign (w[29], w[30], offset); - w[49] = hc_bytealign (w[28], w[29], offset); - w[48] = hc_bytealign (w[27], w[28], offset); - w[47] = hc_bytealign (w[26], w[27], offset); - w[46] = hc_bytealign (w[25], w[26], offset); - w[45] = hc_bytealign (w[24], w[25], offset); - w[44] = hc_bytealign (w[23], w[24], offset); - w[43] = hc_bytealign (w[22], w[23], offset); - w[42] = hc_bytealign (w[21], w[22], offset); - w[41] = hc_bytealign (w[20], w[21], offset); - w[40] = hc_bytealign (w[19], w[20], offset); - w[39] = hc_bytealign (w[18], w[19], offset); - w[38] = hc_bytealign (w[17], w[18], offset); - w[37] = hc_bytealign (w[16], w[17], offset); - w[36] = hc_bytealign (w[15], w[16], offset); - w[35] = hc_bytealign (w[14], w[15], offset); - w[34] = hc_bytealign (w[13], w[14], offset); - w[33] = hc_bytealign (w[12], w[13], offset); - w[32] = hc_bytealign (w[11], w[12], offset); - w[31] = hc_bytealign (w[10], w[11], offset); - w[30] = hc_bytealign (w[ 9], w[10], offset); - w[29] = hc_bytealign (w[ 8], w[ 9], offset); - w[28] = hc_bytealign (w[ 7], w[ 8], offset); - w[27] = hc_bytealign (w[ 6], w[ 7], offset); - w[26] = hc_bytealign (w[ 5], w[ 6], offset); - w[25] = hc_bytealign (w[ 4], w[ 5], offset); - w[24] = hc_bytealign (w[ 3], w[ 4], offset); - w[23] = hc_bytealign (w[ 2], w[ 3], offset); - w[22] = hc_bytealign (w[ 1], w[ 2], offset); - w[21] = hc_bytealign (w[ 0], w[ 1], offset); - w[20] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[42], w[43], offset); + w[62] = hc_bytealign_be (w[41], w[42], offset); + w[61] = hc_bytealign_be (w[40], w[41], offset); + w[60] = hc_bytealign_be (w[39], w[40], offset); + w[59] = hc_bytealign_be (w[38], w[39], offset); + w[58] = hc_bytealign_be (w[37], w[38], offset); + w[57] = hc_bytealign_be (w[36], w[37], offset); + w[56] = hc_bytealign_be (w[35], w[36], offset); + w[55] = hc_bytealign_be (w[34], w[35], offset); + w[54] = hc_bytealign_be (w[33], w[34], offset); + w[53] = hc_bytealign_be (w[32], w[33], offset); + w[52] = hc_bytealign_be (w[31], w[32], offset); + w[51] = hc_bytealign_be (w[30], w[31], offset); + w[50] = hc_bytealign_be (w[29], w[30], offset); + w[49] = hc_bytealign_be (w[28], w[29], offset); + w[48] = hc_bytealign_be (w[27], w[28], offset); + w[47] = hc_bytealign_be (w[26], w[27], offset); + w[46] = hc_bytealign_be (w[25], w[26], offset); + w[45] = hc_bytealign_be (w[24], w[25], offset); + w[44] = hc_bytealign_be (w[23], w[24], offset); + w[43] = hc_bytealign_be (w[22], w[23], offset); + w[42] = hc_bytealign_be (w[21], w[22], offset); + w[41] = hc_bytealign_be (w[20], w[21], offset); + w[40] = hc_bytealign_be (w[19], w[20], offset); + w[39] = hc_bytealign_be (w[18], w[19], offset); + w[38] = hc_bytealign_be (w[17], w[18], offset); + w[37] = hc_bytealign_be (w[16], w[17], offset); + w[36] = hc_bytealign_be (w[15], w[16], offset); + w[35] = hc_bytealign_be (w[14], w[15], offset); + w[34] = hc_bytealign_be (w[13], w[14], offset); + w[33] = hc_bytealign_be (w[12], w[13], offset); + w[32] = hc_bytealign_be (w[11], w[12], offset); + w[31] = hc_bytealign_be (w[10], w[11], offset); + w[30] = hc_bytealign_be (w[ 9], w[10], offset); + w[29] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[28] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[27] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[26] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[25] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[24] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[23] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[22] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[21] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[20] = hc_bytealign_be ( 0, w[ 0], offset); w[19] = 0; w[18] = 0; w[17] = 0; @@ -22033,49 +21871,49 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 21: - w[63] = hc_bytealign (w[41], w[42], offset); - w[62] = hc_bytealign (w[40], w[41], offset); - w[61] = hc_bytealign (w[39], w[40], offset); - w[60] = hc_bytealign (w[38], w[39], offset); - w[59] = hc_bytealign (w[37], w[38], offset); - w[58] = hc_bytealign (w[36], w[37], offset); - w[57] = hc_bytealign (w[35], w[36], offset); - w[56] = hc_bytealign (w[34], w[35], offset); - w[55] = hc_bytealign (w[33], w[34], offset); - w[54] = hc_bytealign (w[32], w[33], offset); - w[53] = hc_bytealign (w[31], w[32], offset); - w[52] = hc_bytealign (w[30], w[31], offset); - w[51] = hc_bytealign (w[29], w[30], offset); - w[50] = hc_bytealign (w[28], w[29], offset); - w[49] = hc_bytealign (w[27], w[28], offset); - w[48] = hc_bytealign (w[26], w[27], offset); - w[47] = hc_bytealign (w[25], w[26], offset); - w[46] = hc_bytealign (w[24], w[25], offset); - w[45] = hc_bytealign (w[23], w[24], offset); - w[44] = hc_bytealign (w[22], w[23], offset); - w[43] = hc_bytealign (w[21], w[22], offset); - w[42] = hc_bytealign (w[20], w[21], offset); - w[41] = hc_bytealign (w[19], w[20], offset); - w[40] = hc_bytealign (w[18], w[19], offset); - w[39] = hc_bytealign (w[17], w[18], offset); - w[38] = hc_bytealign (w[16], w[17], offset); - w[37] = hc_bytealign (w[15], w[16], offset); - w[36] = hc_bytealign (w[14], w[15], offset); - w[35] = hc_bytealign (w[13], w[14], offset); - w[34] = hc_bytealign (w[12], w[13], offset); - w[33] = hc_bytealign (w[11], w[12], offset); - w[32] = hc_bytealign (w[10], w[11], offset); - w[31] = hc_bytealign (w[ 9], w[10], offset); - w[30] = hc_bytealign (w[ 8], w[ 9], offset); - w[29] = hc_bytealign (w[ 7], w[ 8], offset); - w[28] = hc_bytealign (w[ 6], w[ 7], offset); - w[27] = hc_bytealign (w[ 5], w[ 6], offset); - w[26] = hc_bytealign (w[ 4], w[ 5], offset); - w[25] = hc_bytealign (w[ 3], w[ 4], offset); - w[24] = hc_bytealign (w[ 2], w[ 3], offset); - w[23] = hc_bytealign (w[ 1], w[ 2], offset); - w[22] = hc_bytealign (w[ 0], w[ 1], offset); - w[21] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[41], w[42], offset); + w[62] = hc_bytealign_be (w[40], w[41], offset); + w[61] = hc_bytealign_be (w[39], w[40], offset); + w[60] = hc_bytealign_be (w[38], w[39], offset); + w[59] = hc_bytealign_be (w[37], w[38], offset); + w[58] = hc_bytealign_be (w[36], w[37], offset); + w[57] = hc_bytealign_be (w[35], w[36], offset); + w[56] = hc_bytealign_be (w[34], w[35], offset); + w[55] = hc_bytealign_be (w[33], w[34], offset); + w[54] = hc_bytealign_be (w[32], w[33], offset); + w[53] = hc_bytealign_be (w[31], w[32], offset); + w[52] = hc_bytealign_be (w[30], w[31], offset); + w[51] = hc_bytealign_be (w[29], w[30], offset); + w[50] = hc_bytealign_be (w[28], w[29], offset); + w[49] = hc_bytealign_be (w[27], w[28], offset); + w[48] = hc_bytealign_be (w[26], w[27], offset); + w[47] = hc_bytealign_be (w[25], w[26], offset); + w[46] = hc_bytealign_be (w[24], w[25], offset); + w[45] = hc_bytealign_be (w[23], w[24], offset); + w[44] = hc_bytealign_be (w[22], w[23], offset); + w[43] = hc_bytealign_be (w[21], w[22], offset); + w[42] = hc_bytealign_be (w[20], w[21], offset); + w[41] = hc_bytealign_be (w[19], w[20], offset); + w[40] = hc_bytealign_be (w[18], w[19], offset); + w[39] = hc_bytealign_be (w[17], w[18], offset); + w[38] = hc_bytealign_be (w[16], w[17], offset); + w[37] = hc_bytealign_be (w[15], w[16], offset); + w[36] = hc_bytealign_be (w[14], w[15], offset); + w[35] = hc_bytealign_be (w[13], w[14], offset); + w[34] = hc_bytealign_be (w[12], w[13], offset); + w[33] = hc_bytealign_be (w[11], w[12], offset); + w[32] = hc_bytealign_be (w[10], w[11], offset); + w[31] = hc_bytealign_be (w[ 9], w[10], offset); + w[30] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[29] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[28] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[27] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[26] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[25] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[24] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[23] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[22] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[21] = hc_bytealign_be ( 0, w[ 0], offset); w[20] = 0; w[19] = 0; w[18] = 0; @@ -22101,48 +21939,48 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 22: - w[63] = hc_bytealign (w[40], w[41], offset); - w[62] = hc_bytealign (w[39], w[40], offset); - w[61] = hc_bytealign (w[38], w[39], offset); - w[60] = hc_bytealign (w[37], w[38], offset); - w[59] = hc_bytealign (w[36], w[37], offset); - w[58] = hc_bytealign (w[35], w[36], offset); - w[57] = hc_bytealign (w[34], w[35], offset); - w[56] = hc_bytealign (w[33], w[34], offset); - w[55] = hc_bytealign (w[32], w[33], offset); - w[54] = hc_bytealign (w[31], w[32], offset); - w[53] = hc_bytealign (w[30], w[31], offset); - w[52] = hc_bytealign (w[29], w[30], offset); - w[51] = hc_bytealign (w[28], w[29], offset); - w[50] = hc_bytealign (w[27], w[28], offset); - w[49] = hc_bytealign (w[26], w[27], offset); - w[48] = hc_bytealign (w[25], w[26], offset); - w[47] = hc_bytealign (w[24], w[25], offset); - w[46] = hc_bytealign (w[23], w[24], offset); - w[45] = hc_bytealign (w[22], w[23], offset); - w[44] = hc_bytealign (w[21], w[22], offset); - w[43] = hc_bytealign (w[20], w[21], offset); - w[42] = hc_bytealign (w[19], w[20], offset); - w[41] = hc_bytealign (w[18], w[19], offset); - w[40] = hc_bytealign (w[17], w[18], offset); - w[39] = hc_bytealign (w[16], w[17], offset); - w[38] = hc_bytealign (w[15], w[16], offset); - w[37] = hc_bytealign (w[14], w[15], offset); - w[36] = hc_bytealign (w[13], w[14], offset); - w[35] = hc_bytealign (w[12], w[13], offset); - w[34] = hc_bytealign (w[11], w[12], offset); - w[33] = hc_bytealign (w[10], w[11], offset); - w[32] = hc_bytealign (w[ 9], w[10], offset); - w[31] = hc_bytealign (w[ 8], w[ 9], offset); - w[30] = hc_bytealign (w[ 7], w[ 8], offset); - w[29] = hc_bytealign (w[ 6], w[ 7], offset); - w[28] = hc_bytealign (w[ 5], w[ 6], offset); - w[27] = hc_bytealign (w[ 4], w[ 5], offset); - w[26] = hc_bytealign (w[ 3], w[ 4], offset); - w[25] = hc_bytealign (w[ 2], w[ 3], offset); - w[24] = hc_bytealign (w[ 1], w[ 2], offset); - w[23] = hc_bytealign (w[ 0], w[ 1], offset); - w[22] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[40], w[41], offset); + w[62] = hc_bytealign_be (w[39], w[40], offset); + w[61] = hc_bytealign_be (w[38], w[39], offset); + w[60] = hc_bytealign_be (w[37], w[38], offset); + w[59] = hc_bytealign_be (w[36], w[37], offset); + w[58] = hc_bytealign_be (w[35], w[36], offset); + w[57] = hc_bytealign_be (w[34], w[35], offset); + w[56] = hc_bytealign_be (w[33], w[34], offset); + w[55] = hc_bytealign_be (w[32], w[33], offset); + w[54] = hc_bytealign_be (w[31], w[32], offset); + w[53] = hc_bytealign_be (w[30], w[31], offset); + w[52] = hc_bytealign_be (w[29], w[30], offset); + w[51] = hc_bytealign_be (w[28], w[29], offset); + w[50] = hc_bytealign_be (w[27], w[28], offset); + w[49] = hc_bytealign_be (w[26], w[27], offset); + w[48] = hc_bytealign_be (w[25], w[26], offset); + w[47] = hc_bytealign_be (w[24], w[25], offset); + w[46] = hc_bytealign_be (w[23], w[24], offset); + w[45] = hc_bytealign_be (w[22], w[23], offset); + w[44] = hc_bytealign_be (w[21], w[22], offset); + w[43] = hc_bytealign_be (w[20], w[21], offset); + w[42] = hc_bytealign_be (w[19], w[20], offset); + w[41] = hc_bytealign_be (w[18], w[19], offset); + w[40] = hc_bytealign_be (w[17], w[18], offset); + w[39] = hc_bytealign_be (w[16], w[17], offset); + w[38] = hc_bytealign_be (w[15], w[16], offset); + w[37] = hc_bytealign_be (w[14], w[15], offset); + w[36] = hc_bytealign_be (w[13], w[14], offset); + w[35] = hc_bytealign_be (w[12], w[13], offset); + w[34] = hc_bytealign_be (w[11], w[12], offset); + w[33] = hc_bytealign_be (w[10], w[11], offset); + w[32] = hc_bytealign_be (w[ 9], w[10], offset); + w[31] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[30] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[29] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[28] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[27] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[26] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[25] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[24] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[23] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[22] = hc_bytealign_be ( 0, w[ 0], offset); w[21] = 0; w[20] = 0; w[19] = 0; @@ -22169,47 +22007,47 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 23: - w[63] = hc_bytealign (w[39], w[40], offset); - w[62] = hc_bytealign (w[38], w[39], offset); - w[61] = hc_bytealign (w[37], w[38], offset); - w[60] = hc_bytealign (w[36], w[37], offset); - w[59] = hc_bytealign (w[35], w[36], offset); - w[58] = hc_bytealign (w[34], w[35], offset); - w[57] = hc_bytealign (w[33], w[34], offset); - w[56] = hc_bytealign (w[32], w[33], offset); - w[55] = hc_bytealign (w[31], w[32], offset); - w[54] = hc_bytealign (w[30], w[31], offset); - w[53] = hc_bytealign (w[29], w[30], offset); - w[52] = hc_bytealign (w[28], w[29], offset); - w[51] = hc_bytealign (w[27], w[28], offset); - w[50] = hc_bytealign (w[26], w[27], offset); - w[49] = hc_bytealign (w[25], w[26], offset); - w[48] = hc_bytealign (w[24], w[25], offset); - w[47] = hc_bytealign (w[23], w[24], offset); - w[46] = hc_bytealign (w[22], w[23], offset); - w[45] = hc_bytealign (w[21], w[22], offset); - w[44] = hc_bytealign (w[20], w[21], offset); - w[43] = hc_bytealign (w[19], w[20], offset); - w[42] = hc_bytealign (w[18], w[19], offset); - w[41] = hc_bytealign (w[17], w[18], offset); - w[40] = hc_bytealign (w[16], w[17], offset); - w[39] = hc_bytealign (w[15], w[16], offset); - w[38] = hc_bytealign (w[14], w[15], offset); - w[37] = hc_bytealign (w[13], w[14], offset); - w[36] = hc_bytealign (w[12], w[13], offset); - w[35] = hc_bytealign (w[11], w[12], offset); - w[34] = hc_bytealign (w[10], w[11], offset); - w[33] = hc_bytealign (w[ 9], w[10], offset); - w[32] = hc_bytealign (w[ 8], w[ 9], offset); - w[31] = hc_bytealign (w[ 7], w[ 8], offset); - w[30] = hc_bytealign (w[ 6], w[ 7], offset); - w[29] = hc_bytealign (w[ 5], w[ 6], offset); - w[28] = hc_bytealign (w[ 4], w[ 5], offset); - w[27] = hc_bytealign (w[ 3], w[ 4], offset); - w[26] = hc_bytealign (w[ 2], w[ 3], offset); - w[25] = hc_bytealign (w[ 1], w[ 2], offset); - w[24] = hc_bytealign (w[ 0], w[ 1], offset); - w[23] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[39], w[40], offset); + w[62] = hc_bytealign_be (w[38], w[39], offset); + w[61] = hc_bytealign_be (w[37], w[38], offset); + w[60] = hc_bytealign_be (w[36], w[37], offset); + w[59] = hc_bytealign_be (w[35], w[36], offset); + w[58] = hc_bytealign_be (w[34], w[35], offset); + w[57] = hc_bytealign_be (w[33], w[34], offset); + w[56] = hc_bytealign_be (w[32], w[33], offset); + w[55] = hc_bytealign_be (w[31], w[32], offset); + w[54] = hc_bytealign_be (w[30], w[31], offset); + w[53] = hc_bytealign_be (w[29], w[30], offset); + w[52] = hc_bytealign_be (w[28], w[29], offset); + w[51] = hc_bytealign_be (w[27], w[28], offset); + w[50] = hc_bytealign_be (w[26], w[27], offset); + w[49] = hc_bytealign_be (w[25], w[26], offset); + w[48] = hc_bytealign_be (w[24], w[25], offset); + w[47] = hc_bytealign_be (w[23], w[24], offset); + w[46] = hc_bytealign_be (w[22], w[23], offset); + w[45] = hc_bytealign_be (w[21], w[22], offset); + w[44] = hc_bytealign_be (w[20], w[21], offset); + w[43] = hc_bytealign_be (w[19], w[20], offset); + w[42] = hc_bytealign_be (w[18], w[19], offset); + w[41] = hc_bytealign_be (w[17], w[18], offset); + w[40] = hc_bytealign_be (w[16], w[17], offset); + w[39] = hc_bytealign_be (w[15], w[16], offset); + w[38] = hc_bytealign_be (w[14], w[15], offset); + w[37] = hc_bytealign_be (w[13], w[14], offset); + w[36] = hc_bytealign_be (w[12], w[13], offset); + w[35] = hc_bytealign_be (w[11], w[12], offset); + w[34] = hc_bytealign_be (w[10], w[11], offset); + w[33] = hc_bytealign_be (w[ 9], w[10], offset); + w[32] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[31] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[30] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[29] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[28] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[27] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[26] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[25] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[24] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[23] = hc_bytealign_be ( 0, w[ 0], offset); w[22] = 0; w[21] = 0; w[20] = 0; @@ -22237,46 +22075,46 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 24: - w[63] = hc_bytealign (w[38], w[39], offset); - w[62] = hc_bytealign (w[37], w[38], offset); - w[61] = hc_bytealign (w[36], w[37], offset); - w[60] = hc_bytealign (w[35], w[36], offset); - w[59] = hc_bytealign (w[34], w[35], offset); - w[58] = hc_bytealign (w[33], w[34], offset); - w[57] = hc_bytealign (w[32], w[33], offset); - w[56] = hc_bytealign (w[31], w[32], offset); - w[55] = hc_bytealign (w[30], w[31], offset); - w[54] = hc_bytealign (w[29], w[30], offset); - w[53] = hc_bytealign (w[28], w[29], offset); - w[52] = hc_bytealign (w[27], w[28], offset); - w[51] = hc_bytealign (w[26], w[27], offset); - w[50] = hc_bytealign (w[25], w[26], offset); - w[49] = hc_bytealign (w[24], w[25], offset); - w[48] = hc_bytealign (w[23], w[24], offset); - w[47] = hc_bytealign (w[22], w[23], offset); - w[46] = hc_bytealign (w[21], w[22], offset); - w[45] = hc_bytealign (w[20], w[21], offset); - w[44] = hc_bytealign (w[19], w[20], offset); - w[43] = hc_bytealign (w[18], w[19], offset); - w[42] = hc_bytealign (w[17], w[18], offset); - w[41] = hc_bytealign (w[16], w[17], offset); - w[40] = hc_bytealign (w[15], w[16], offset); - w[39] = hc_bytealign (w[14], w[15], offset); - w[38] = hc_bytealign (w[13], w[14], offset); - w[37] = hc_bytealign (w[12], w[13], offset); - w[36] = hc_bytealign (w[11], w[12], offset); - w[35] = hc_bytealign (w[10], w[11], offset); - w[34] = hc_bytealign (w[ 9], w[10], offset); - w[33] = hc_bytealign (w[ 8], w[ 9], offset); - w[32] = hc_bytealign (w[ 7], w[ 8], offset); - w[31] = hc_bytealign (w[ 6], w[ 7], offset); - w[30] = hc_bytealign (w[ 5], w[ 6], offset); - w[29] = hc_bytealign (w[ 4], w[ 5], offset); - w[28] = hc_bytealign (w[ 3], w[ 4], offset); - w[27] = hc_bytealign (w[ 2], w[ 3], offset); - w[26] = hc_bytealign (w[ 1], w[ 2], offset); - w[25] = hc_bytealign (w[ 0], w[ 1], offset); - w[24] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[38], w[39], offset); + w[62] = hc_bytealign_be (w[37], w[38], offset); + w[61] = hc_bytealign_be (w[36], w[37], offset); + w[60] = hc_bytealign_be (w[35], w[36], offset); + w[59] = hc_bytealign_be (w[34], w[35], offset); + w[58] = hc_bytealign_be (w[33], w[34], offset); + w[57] = hc_bytealign_be (w[32], w[33], offset); + w[56] = hc_bytealign_be (w[31], w[32], offset); + w[55] = hc_bytealign_be (w[30], w[31], offset); + w[54] = hc_bytealign_be (w[29], w[30], offset); + w[53] = hc_bytealign_be (w[28], w[29], offset); + w[52] = hc_bytealign_be (w[27], w[28], offset); + w[51] = hc_bytealign_be (w[26], w[27], offset); + w[50] = hc_bytealign_be (w[25], w[26], offset); + w[49] = hc_bytealign_be (w[24], w[25], offset); + w[48] = hc_bytealign_be (w[23], w[24], offset); + w[47] = hc_bytealign_be (w[22], w[23], offset); + w[46] = hc_bytealign_be (w[21], w[22], offset); + w[45] = hc_bytealign_be (w[20], w[21], offset); + w[44] = hc_bytealign_be (w[19], w[20], offset); + w[43] = hc_bytealign_be (w[18], w[19], offset); + w[42] = hc_bytealign_be (w[17], w[18], offset); + w[41] = hc_bytealign_be (w[16], w[17], offset); + w[40] = hc_bytealign_be (w[15], w[16], offset); + w[39] = hc_bytealign_be (w[14], w[15], offset); + w[38] = hc_bytealign_be (w[13], w[14], offset); + w[37] = hc_bytealign_be (w[12], w[13], offset); + w[36] = hc_bytealign_be (w[11], w[12], offset); + w[35] = hc_bytealign_be (w[10], w[11], offset); + w[34] = hc_bytealign_be (w[ 9], w[10], offset); + w[33] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[32] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[31] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[30] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[29] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[28] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[27] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[26] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[25] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[24] = hc_bytealign_be ( 0, w[ 0], offset); w[23] = 0; w[22] = 0; w[21] = 0; @@ -22305,45 +22143,45 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 25: - w[63] = hc_bytealign (w[37], w[38], offset); - w[62] = hc_bytealign (w[36], w[37], offset); - w[61] = hc_bytealign (w[35], w[36], offset); - w[60] = hc_bytealign (w[34], w[35], offset); - w[59] = hc_bytealign (w[33], w[34], offset); - w[58] = hc_bytealign (w[32], w[33], offset); - w[57] = hc_bytealign (w[31], w[32], offset); - w[56] = hc_bytealign (w[30], w[31], offset); - w[55] = hc_bytealign (w[29], w[30], offset); - w[54] = hc_bytealign (w[28], w[29], offset); - w[53] = hc_bytealign (w[27], w[28], offset); - w[52] = hc_bytealign (w[26], w[27], offset); - w[51] = hc_bytealign (w[25], w[26], offset); - w[50] = hc_bytealign (w[24], w[25], offset); - w[49] = hc_bytealign (w[23], w[24], offset); - w[48] = hc_bytealign (w[22], w[23], offset); - w[47] = hc_bytealign (w[21], w[22], offset); - w[46] = hc_bytealign (w[20], w[21], offset); - w[45] = hc_bytealign (w[19], w[20], offset); - w[44] = hc_bytealign (w[18], w[19], offset); - w[43] = hc_bytealign (w[17], w[18], offset); - w[42] = hc_bytealign (w[16], w[17], offset); - w[41] = hc_bytealign (w[15], w[16], offset); - w[40] = hc_bytealign (w[14], w[15], offset); - w[39] = hc_bytealign (w[13], w[14], offset); - w[38] = hc_bytealign (w[12], w[13], offset); - w[37] = hc_bytealign (w[11], w[12], offset); - w[36] = hc_bytealign (w[10], w[11], offset); - w[35] = hc_bytealign (w[ 9], w[10], offset); - w[34] = hc_bytealign (w[ 8], w[ 9], offset); - w[33] = hc_bytealign (w[ 7], w[ 8], offset); - w[32] = hc_bytealign (w[ 6], w[ 7], offset); - w[31] = hc_bytealign (w[ 5], w[ 6], offset); - w[30] = hc_bytealign (w[ 4], w[ 5], offset); - w[29] = hc_bytealign (w[ 3], w[ 4], offset); - w[28] = hc_bytealign (w[ 2], w[ 3], offset); - w[27] = hc_bytealign (w[ 1], w[ 2], offset); - w[26] = hc_bytealign (w[ 0], w[ 1], offset); - w[25] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[37], w[38], offset); + w[62] = hc_bytealign_be (w[36], w[37], offset); + w[61] = hc_bytealign_be (w[35], w[36], offset); + w[60] = hc_bytealign_be (w[34], w[35], offset); + w[59] = hc_bytealign_be (w[33], w[34], offset); + w[58] = hc_bytealign_be (w[32], w[33], offset); + w[57] = hc_bytealign_be (w[31], w[32], offset); + w[56] = hc_bytealign_be (w[30], w[31], offset); + w[55] = hc_bytealign_be (w[29], w[30], offset); + w[54] = hc_bytealign_be (w[28], w[29], offset); + w[53] = hc_bytealign_be (w[27], w[28], offset); + w[52] = hc_bytealign_be (w[26], w[27], offset); + w[51] = hc_bytealign_be (w[25], w[26], offset); + w[50] = hc_bytealign_be (w[24], w[25], offset); + w[49] = hc_bytealign_be (w[23], w[24], offset); + w[48] = hc_bytealign_be (w[22], w[23], offset); + w[47] = hc_bytealign_be (w[21], w[22], offset); + w[46] = hc_bytealign_be (w[20], w[21], offset); + w[45] = hc_bytealign_be (w[19], w[20], offset); + w[44] = hc_bytealign_be (w[18], w[19], offset); + w[43] = hc_bytealign_be (w[17], w[18], offset); + w[42] = hc_bytealign_be (w[16], w[17], offset); + w[41] = hc_bytealign_be (w[15], w[16], offset); + w[40] = hc_bytealign_be (w[14], w[15], offset); + w[39] = hc_bytealign_be (w[13], w[14], offset); + w[38] = hc_bytealign_be (w[12], w[13], offset); + w[37] = hc_bytealign_be (w[11], w[12], offset); + w[36] = hc_bytealign_be (w[10], w[11], offset); + w[35] = hc_bytealign_be (w[ 9], w[10], offset); + w[34] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[33] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[32] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[31] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[30] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[29] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[28] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[27] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[26] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[25] = hc_bytealign_be ( 0, w[ 0], offset); w[24] = 0; w[23] = 0; w[22] = 0; @@ -22373,44 +22211,44 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 26: - w[63] = hc_bytealign (w[36], w[37], offset); - w[62] = hc_bytealign (w[35], w[36], offset); - w[61] = hc_bytealign (w[34], w[35], offset); - w[60] = hc_bytealign (w[33], w[34], offset); - w[59] = hc_bytealign (w[32], w[33], offset); - w[58] = hc_bytealign (w[31], w[32], offset); - w[57] = hc_bytealign (w[30], w[31], offset); - w[56] = hc_bytealign (w[29], w[30], offset); - w[55] = hc_bytealign (w[28], w[29], offset); - w[54] = hc_bytealign (w[27], w[28], offset); - w[53] = hc_bytealign (w[26], w[27], offset); - w[52] = hc_bytealign (w[25], w[26], offset); - w[51] = hc_bytealign (w[24], w[25], offset); - w[50] = hc_bytealign (w[23], w[24], offset); - w[49] = hc_bytealign (w[22], w[23], offset); - w[48] = hc_bytealign (w[21], w[22], offset); - w[47] = hc_bytealign (w[20], w[21], offset); - w[46] = hc_bytealign (w[19], w[20], offset); - w[45] = hc_bytealign (w[18], w[19], offset); - w[44] = hc_bytealign (w[17], w[18], offset); - w[43] = hc_bytealign (w[16], w[17], offset); - w[42] = hc_bytealign (w[15], w[16], offset); - w[41] = hc_bytealign (w[14], w[15], offset); - w[40] = hc_bytealign (w[13], w[14], offset); - w[39] = hc_bytealign (w[12], w[13], offset); - w[38] = hc_bytealign (w[11], w[12], offset); - w[37] = hc_bytealign (w[10], w[11], offset); - w[36] = hc_bytealign (w[ 9], w[10], offset); - w[35] = hc_bytealign (w[ 8], w[ 9], offset); - w[34] = hc_bytealign (w[ 7], w[ 8], offset); - w[33] = hc_bytealign (w[ 6], w[ 7], offset); - w[32] = hc_bytealign (w[ 5], w[ 6], offset); - w[31] = hc_bytealign (w[ 4], w[ 5], offset); - w[30] = hc_bytealign (w[ 3], w[ 4], offset); - w[29] = hc_bytealign (w[ 2], w[ 3], offset); - w[28] = hc_bytealign (w[ 1], w[ 2], offset); - w[27] = hc_bytealign (w[ 0], w[ 1], offset); - w[26] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[36], w[37], offset); + w[62] = hc_bytealign_be (w[35], w[36], offset); + w[61] = hc_bytealign_be (w[34], w[35], offset); + w[60] = hc_bytealign_be (w[33], w[34], offset); + w[59] = hc_bytealign_be (w[32], w[33], offset); + w[58] = hc_bytealign_be (w[31], w[32], offset); + w[57] = hc_bytealign_be (w[30], w[31], offset); + w[56] = hc_bytealign_be (w[29], w[30], offset); + w[55] = hc_bytealign_be (w[28], w[29], offset); + w[54] = hc_bytealign_be (w[27], w[28], offset); + w[53] = hc_bytealign_be (w[26], w[27], offset); + w[52] = hc_bytealign_be (w[25], w[26], offset); + w[51] = hc_bytealign_be (w[24], w[25], offset); + w[50] = hc_bytealign_be (w[23], w[24], offset); + w[49] = hc_bytealign_be (w[22], w[23], offset); + w[48] = hc_bytealign_be (w[21], w[22], offset); + w[47] = hc_bytealign_be (w[20], w[21], offset); + w[46] = hc_bytealign_be (w[19], w[20], offset); + w[45] = hc_bytealign_be (w[18], w[19], offset); + w[44] = hc_bytealign_be (w[17], w[18], offset); + w[43] = hc_bytealign_be (w[16], w[17], offset); + w[42] = hc_bytealign_be (w[15], w[16], offset); + w[41] = hc_bytealign_be (w[14], w[15], offset); + w[40] = hc_bytealign_be (w[13], w[14], offset); + w[39] = hc_bytealign_be (w[12], w[13], offset); + w[38] = hc_bytealign_be (w[11], w[12], offset); + w[37] = hc_bytealign_be (w[10], w[11], offset); + w[36] = hc_bytealign_be (w[ 9], w[10], offset); + w[35] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[34] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[33] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[32] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[31] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[30] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[29] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[28] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[27] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[26] = hc_bytealign_be ( 0, w[ 0], offset); w[25] = 0; w[24] = 0; w[23] = 0; @@ -22441,43 +22279,43 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 27: - w[63] = hc_bytealign (w[35], w[36], offset); - w[62] = hc_bytealign (w[34], w[35], offset); - w[61] = hc_bytealign (w[33], w[34], offset); - w[60] = hc_bytealign (w[32], w[33], offset); - w[59] = hc_bytealign (w[31], w[32], offset); - w[58] = hc_bytealign (w[30], w[31], offset); - w[57] = hc_bytealign (w[29], w[30], offset); - w[56] = hc_bytealign (w[28], w[29], offset); - w[55] = hc_bytealign (w[27], w[28], offset); - w[54] = hc_bytealign (w[26], w[27], offset); - w[53] = hc_bytealign (w[25], w[26], offset); - w[52] = hc_bytealign (w[24], w[25], offset); - w[51] = hc_bytealign (w[23], w[24], offset); - w[50] = hc_bytealign (w[22], w[23], offset); - w[49] = hc_bytealign (w[21], w[22], offset); - w[48] = hc_bytealign (w[20], w[21], offset); - w[47] = hc_bytealign (w[19], w[20], offset); - w[46] = hc_bytealign (w[18], w[19], offset); - w[45] = hc_bytealign (w[17], w[18], offset); - w[44] = hc_bytealign (w[16], w[17], offset); - w[43] = hc_bytealign (w[15], w[16], offset); - w[42] = hc_bytealign (w[14], w[15], offset); - w[41] = hc_bytealign (w[13], w[14], offset); - w[40] = hc_bytealign (w[12], w[13], offset); - w[39] = hc_bytealign (w[11], w[12], offset); - w[38] = hc_bytealign (w[10], w[11], offset); - w[37] = hc_bytealign (w[ 9], w[10], offset); - w[36] = hc_bytealign (w[ 8], w[ 9], offset); - w[35] = hc_bytealign (w[ 7], w[ 8], offset); - w[34] = hc_bytealign (w[ 6], w[ 7], offset); - w[33] = hc_bytealign (w[ 5], w[ 6], offset); - w[32] = hc_bytealign (w[ 4], w[ 5], offset); - w[31] = hc_bytealign (w[ 3], w[ 4], offset); - w[30] = hc_bytealign (w[ 2], w[ 3], offset); - w[29] = hc_bytealign (w[ 1], w[ 2], offset); - w[28] = hc_bytealign (w[ 0], w[ 1], offset); - w[27] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[35], w[36], offset); + w[62] = hc_bytealign_be (w[34], w[35], offset); + w[61] = hc_bytealign_be (w[33], w[34], offset); + w[60] = hc_bytealign_be (w[32], w[33], offset); + w[59] = hc_bytealign_be (w[31], w[32], offset); + w[58] = hc_bytealign_be (w[30], w[31], offset); + w[57] = hc_bytealign_be (w[29], w[30], offset); + w[56] = hc_bytealign_be (w[28], w[29], offset); + w[55] = hc_bytealign_be (w[27], w[28], offset); + w[54] = hc_bytealign_be (w[26], w[27], offset); + w[53] = hc_bytealign_be (w[25], w[26], offset); + w[52] = hc_bytealign_be (w[24], w[25], offset); + w[51] = hc_bytealign_be (w[23], w[24], offset); + w[50] = hc_bytealign_be (w[22], w[23], offset); + w[49] = hc_bytealign_be (w[21], w[22], offset); + w[48] = hc_bytealign_be (w[20], w[21], offset); + w[47] = hc_bytealign_be (w[19], w[20], offset); + w[46] = hc_bytealign_be (w[18], w[19], offset); + w[45] = hc_bytealign_be (w[17], w[18], offset); + w[44] = hc_bytealign_be (w[16], w[17], offset); + w[43] = hc_bytealign_be (w[15], w[16], offset); + w[42] = hc_bytealign_be (w[14], w[15], offset); + w[41] = hc_bytealign_be (w[13], w[14], offset); + w[40] = hc_bytealign_be (w[12], w[13], offset); + w[39] = hc_bytealign_be (w[11], w[12], offset); + w[38] = hc_bytealign_be (w[10], w[11], offset); + w[37] = hc_bytealign_be (w[ 9], w[10], offset); + w[36] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[35] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[34] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[33] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[32] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[31] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[30] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[29] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[28] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[27] = hc_bytealign_be ( 0, w[ 0], offset); w[26] = 0; w[25] = 0; w[24] = 0; @@ -22509,42 +22347,42 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 28: - w[63] = hc_bytealign (w[34], w[35], offset); - w[62] = hc_bytealign (w[33], w[34], offset); - w[61] = hc_bytealign (w[32], w[33], offset); - w[60] = hc_bytealign (w[31], w[32], offset); - w[59] = hc_bytealign (w[30], w[31], offset); - w[58] = hc_bytealign (w[29], w[30], offset); - w[57] = hc_bytealign (w[28], w[29], offset); - w[56] = hc_bytealign (w[27], w[28], offset); - w[55] = hc_bytealign (w[26], w[27], offset); - w[54] = hc_bytealign (w[25], w[26], offset); - w[53] = hc_bytealign (w[24], w[25], offset); - w[52] = hc_bytealign (w[23], w[24], offset); - w[51] = hc_bytealign (w[22], w[23], offset); - w[50] = hc_bytealign (w[21], w[22], offset); - w[49] = hc_bytealign (w[20], w[21], offset); - w[48] = hc_bytealign (w[19], w[20], offset); - w[47] = hc_bytealign (w[18], w[19], offset); - w[46] = hc_bytealign (w[17], w[18], offset); - w[45] = hc_bytealign (w[16], w[17], offset); - w[44] = hc_bytealign (w[15], w[16], offset); - w[43] = hc_bytealign (w[14], w[15], offset); - w[42] = hc_bytealign (w[13], w[14], offset); - w[41] = hc_bytealign (w[12], w[13], offset); - w[40] = hc_bytealign (w[11], w[12], offset); - w[39] = hc_bytealign (w[10], w[11], offset); - w[38] = hc_bytealign (w[ 9], w[10], offset); - w[37] = hc_bytealign (w[ 8], w[ 9], offset); - w[36] = hc_bytealign (w[ 7], w[ 8], offset); - w[35] = hc_bytealign (w[ 6], w[ 7], offset); - w[34] = hc_bytealign (w[ 5], w[ 6], offset); - w[33] = hc_bytealign (w[ 4], w[ 5], offset); - w[32] = hc_bytealign (w[ 3], w[ 4], offset); - w[31] = hc_bytealign (w[ 2], w[ 3], offset); - w[30] = hc_bytealign (w[ 1], w[ 2], offset); - w[29] = hc_bytealign (w[ 0], w[ 1], offset); - w[28] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[34], w[35], offset); + w[62] = hc_bytealign_be (w[33], w[34], offset); + w[61] = hc_bytealign_be (w[32], w[33], offset); + w[60] = hc_bytealign_be (w[31], w[32], offset); + w[59] = hc_bytealign_be (w[30], w[31], offset); + w[58] = hc_bytealign_be (w[29], w[30], offset); + w[57] = hc_bytealign_be (w[28], w[29], offset); + w[56] = hc_bytealign_be (w[27], w[28], offset); + w[55] = hc_bytealign_be (w[26], w[27], offset); + w[54] = hc_bytealign_be (w[25], w[26], offset); + w[53] = hc_bytealign_be (w[24], w[25], offset); + w[52] = hc_bytealign_be (w[23], w[24], offset); + w[51] = hc_bytealign_be (w[22], w[23], offset); + w[50] = hc_bytealign_be (w[21], w[22], offset); + w[49] = hc_bytealign_be (w[20], w[21], offset); + w[48] = hc_bytealign_be (w[19], w[20], offset); + w[47] = hc_bytealign_be (w[18], w[19], offset); + w[46] = hc_bytealign_be (w[17], w[18], offset); + w[45] = hc_bytealign_be (w[16], w[17], offset); + w[44] = hc_bytealign_be (w[15], w[16], offset); + w[43] = hc_bytealign_be (w[14], w[15], offset); + w[42] = hc_bytealign_be (w[13], w[14], offset); + w[41] = hc_bytealign_be (w[12], w[13], offset); + w[40] = hc_bytealign_be (w[11], w[12], offset); + w[39] = hc_bytealign_be (w[10], w[11], offset); + w[38] = hc_bytealign_be (w[ 9], w[10], offset); + w[37] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[36] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[35] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[34] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[33] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[32] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[31] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[30] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[29] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[28] = hc_bytealign_be ( 0, w[ 0], offset); w[27] = 0; w[26] = 0; w[25] = 0; @@ -22577,41 +22415,41 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 29: - w[63] = hc_bytealign (w[33], w[34], offset); - w[62] = hc_bytealign (w[32], w[33], offset); - w[61] = hc_bytealign (w[31], w[32], offset); - w[60] = hc_bytealign (w[30], w[31], offset); - w[59] = hc_bytealign (w[29], w[30], offset); - w[58] = hc_bytealign (w[28], w[29], offset); - w[57] = hc_bytealign (w[27], w[28], offset); - w[56] = hc_bytealign (w[26], w[27], offset); - w[55] = hc_bytealign (w[25], w[26], offset); - w[54] = hc_bytealign (w[24], w[25], offset); - w[53] = hc_bytealign (w[23], w[24], offset); - w[52] = hc_bytealign (w[22], w[23], offset); - w[51] = hc_bytealign (w[21], w[22], offset); - w[50] = hc_bytealign (w[20], w[21], offset); - w[49] = hc_bytealign (w[19], w[20], offset); - w[48] = hc_bytealign (w[18], w[19], offset); - w[47] = hc_bytealign (w[17], w[18], offset); - w[46] = hc_bytealign (w[16], w[17], offset); - w[45] = hc_bytealign (w[15], w[16], offset); - w[44] = hc_bytealign (w[14], w[15], offset); - w[43] = hc_bytealign (w[13], w[14], offset); - w[42] = hc_bytealign (w[12], w[13], offset); - w[41] = hc_bytealign (w[11], w[12], offset); - w[40] = hc_bytealign (w[10], w[11], offset); - w[39] = hc_bytealign (w[ 9], w[10], offset); - w[38] = hc_bytealign (w[ 8], w[ 9], offset); - w[37] = hc_bytealign (w[ 7], w[ 8], offset); - w[36] = hc_bytealign (w[ 6], w[ 7], offset); - w[35] = hc_bytealign (w[ 5], w[ 6], offset); - w[34] = hc_bytealign (w[ 4], w[ 5], offset); - w[33] = hc_bytealign (w[ 3], w[ 4], offset); - w[32] = hc_bytealign (w[ 2], w[ 3], offset); - w[31] = hc_bytealign (w[ 1], w[ 2], offset); - w[30] = hc_bytealign (w[ 0], w[ 1], offset); - w[29] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[33], w[34], offset); + w[62] = hc_bytealign_be (w[32], w[33], offset); + w[61] = hc_bytealign_be (w[31], w[32], offset); + w[60] = hc_bytealign_be (w[30], w[31], offset); + w[59] = hc_bytealign_be (w[29], w[30], offset); + w[58] = hc_bytealign_be (w[28], w[29], offset); + w[57] = hc_bytealign_be (w[27], w[28], offset); + w[56] = hc_bytealign_be (w[26], w[27], offset); + w[55] = hc_bytealign_be (w[25], w[26], offset); + w[54] = hc_bytealign_be (w[24], w[25], offset); + w[53] = hc_bytealign_be (w[23], w[24], offset); + w[52] = hc_bytealign_be (w[22], w[23], offset); + w[51] = hc_bytealign_be (w[21], w[22], offset); + w[50] = hc_bytealign_be (w[20], w[21], offset); + w[49] = hc_bytealign_be (w[19], w[20], offset); + w[48] = hc_bytealign_be (w[18], w[19], offset); + w[47] = hc_bytealign_be (w[17], w[18], offset); + w[46] = hc_bytealign_be (w[16], w[17], offset); + w[45] = hc_bytealign_be (w[15], w[16], offset); + w[44] = hc_bytealign_be (w[14], w[15], offset); + w[43] = hc_bytealign_be (w[13], w[14], offset); + w[42] = hc_bytealign_be (w[12], w[13], offset); + w[41] = hc_bytealign_be (w[11], w[12], offset); + w[40] = hc_bytealign_be (w[10], w[11], offset); + w[39] = hc_bytealign_be (w[ 9], w[10], offset); + w[38] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[37] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[36] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[35] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[34] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[33] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[32] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[31] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[30] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[29] = hc_bytealign_be ( 0, w[ 0], offset); w[28] = 0; w[27] = 0; w[26] = 0; @@ -22645,40 +22483,40 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 30: - w[63] = hc_bytealign (w[32], w[33], offset); - w[62] = hc_bytealign (w[31], w[32], offset); - w[61] = hc_bytealign (w[30], w[31], offset); - w[60] = hc_bytealign (w[29], w[30], offset); - w[59] = hc_bytealign (w[28], w[29], offset); - w[58] = hc_bytealign (w[27], w[28], offset); - w[57] = hc_bytealign (w[26], w[27], offset); - w[56] = hc_bytealign (w[25], w[26], offset); - w[55] = hc_bytealign (w[24], w[25], offset); - w[54] = hc_bytealign (w[23], w[24], offset); - w[53] = hc_bytealign (w[22], w[23], offset); - w[52] = hc_bytealign (w[21], w[22], offset); - w[51] = hc_bytealign (w[20], w[21], offset); - w[50] = hc_bytealign (w[19], w[20], offset); - w[49] = hc_bytealign (w[18], w[19], offset); - w[48] = hc_bytealign (w[17], w[18], offset); - w[47] = hc_bytealign (w[16], w[17], offset); - w[46] = hc_bytealign (w[15], w[16], offset); - w[45] = hc_bytealign (w[14], w[15], offset); - w[44] = hc_bytealign (w[13], w[14], offset); - w[43] = hc_bytealign (w[12], w[13], offset); - w[42] = hc_bytealign (w[11], w[12], offset); - w[41] = hc_bytealign (w[10], w[11], offset); - w[40] = hc_bytealign (w[ 9], w[10], offset); - w[39] = hc_bytealign (w[ 8], w[ 9], offset); - w[38] = hc_bytealign (w[ 7], w[ 8], offset); - w[37] = hc_bytealign (w[ 6], w[ 7], offset); - w[36] = hc_bytealign (w[ 5], w[ 6], offset); - w[35] = hc_bytealign (w[ 4], w[ 5], offset); - w[34] = hc_bytealign (w[ 3], w[ 4], offset); - w[33] = hc_bytealign (w[ 2], w[ 3], offset); - w[32] = hc_bytealign (w[ 1], w[ 2], offset); - w[31] = hc_bytealign (w[ 0], w[ 1], offset); - w[30] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[32], w[33], offset); + w[62] = hc_bytealign_be (w[31], w[32], offset); + w[61] = hc_bytealign_be (w[30], w[31], offset); + w[60] = hc_bytealign_be (w[29], w[30], offset); + w[59] = hc_bytealign_be (w[28], w[29], offset); + w[58] = hc_bytealign_be (w[27], w[28], offset); + w[57] = hc_bytealign_be (w[26], w[27], offset); + w[56] = hc_bytealign_be (w[25], w[26], offset); + w[55] = hc_bytealign_be (w[24], w[25], offset); + w[54] = hc_bytealign_be (w[23], w[24], offset); + w[53] = hc_bytealign_be (w[22], w[23], offset); + w[52] = hc_bytealign_be (w[21], w[22], offset); + w[51] = hc_bytealign_be (w[20], w[21], offset); + w[50] = hc_bytealign_be (w[19], w[20], offset); + w[49] = hc_bytealign_be (w[18], w[19], offset); + w[48] = hc_bytealign_be (w[17], w[18], offset); + w[47] = hc_bytealign_be (w[16], w[17], offset); + w[46] = hc_bytealign_be (w[15], w[16], offset); + w[45] = hc_bytealign_be (w[14], w[15], offset); + w[44] = hc_bytealign_be (w[13], w[14], offset); + w[43] = hc_bytealign_be (w[12], w[13], offset); + w[42] = hc_bytealign_be (w[11], w[12], offset); + w[41] = hc_bytealign_be (w[10], w[11], offset); + w[40] = hc_bytealign_be (w[ 9], w[10], offset); + w[39] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[38] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[37] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[36] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[35] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[34] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[33] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[32] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[31] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[30] = hc_bytealign_be ( 0, w[ 0], offset); w[29] = 0; w[28] = 0; w[27] = 0; @@ -22713,39 +22551,39 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 31: - w[63] = hc_bytealign (w[31], w[32], offset); - w[62] = hc_bytealign (w[30], w[31], offset); - w[61] = hc_bytealign (w[29], w[30], offset); - w[60] = hc_bytealign (w[28], w[29], offset); - w[59] = hc_bytealign (w[27], w[28], offset); - w[58] = hc_bytealign (w[26], w[27], offset); - w[57] = hc_bytealign (w[25], w[26], offset); - w[56] = hc_bytealign (w[24], w[25], offset); - w[55] = hc_bytealign (w[23], w[24], offset); - w[54] = hc_bytealign (w[22], w[23], offset); - w[53] = hc_bytealign (w[21], w[22], offset); - w[52] = hc_bytealign (w[20], w[21], offset); - w[51] = hc_bytealign (w[19], w[20], offset); - w[50] = hc_bytealign (w[18], w[19], offset); - w[49] = hc_bytealign (w[17], w[18], offset); - w[48] = hc_bytealign (w[16], w[17], offset); - w[47] = hc_bytealign (w[15], w[16], offset); - w[46] = hc_bytealign (w[14], w[15], offset); - w[45] = hc_bytealign (w[13], w[14], offset); - w[44] = hc_bytealign (w[12], w[13], offset); - w[43] = hc_bytealign (w[11], w[12], offset); - w[42] = hc_bytealign (w[10], w[11], offset); - w[41] = hc_bytealign (w[ 9], w[10], offset); - w[40] = hc_bytealign (w[ 8], w[ 9], offset); - w[39] = hc_bytealign (w[ 7], w[ 8], offset); - w[38] = hc_bytealign (w[ 6], w[ 7], offset); - w[37] = hc_bytealign (w[ 5], w[ 6], offset); - w[36] = hc_bytealign (w[ 4], w[ 5], offset); - w[35] = hc_bytealign (w[ 3], w[ 4], offset); - w[34] = hc_bytealign (w[ 2], w[ 3], offset); - w[33] = hc_bytealign (w[ 1], w[ 2], offset); - w[32] = hc_bytealign (w[ 0], w[ 1], offset); - w[31] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[31], w[32], offset); + w[62] = hc_bytealign_be (w[30], w[31], offset); + w[61] = hc_bytealign_be (w[29], w[30], offset); + w[60] = hc_bytealign_be (w[28], w[29], offset); + w[59] = hc_bytealign_be (w[27], w[28], offset); + w[58] = hc_bytealign_be (w[26], w[27], offset); + w[57] = hc_bytealign_be (w[25], w[26], offset); + w[56] = hc_bytealign_be (w[24], w[25], offset); + w[55] = hc_bytealign_be (w[23], w[24], offset); + w[54] = hc_bytealign_be (w[22], w[23], offset); + w[53] = hc_bytealign_be (w[21], w[22], offset); + w[52] = hc_bytealign_be (w[20], w[21], offset); + w[51] = hc_bytealign_be (w[19], w[20], offset); + w[50] = hc_bytealign_be (w[18], w[19], offset); + w[49] = hc_bytealign_be (w[17], w[18], offset); + w[48] = hc_bytealign_be (w[16], w[17], offset); + w[47] = hc_bytealign_be (w[15], w[16], offset); + w[46] = hc_bytealign_be (w[14], w[15], offset); + w[45] = hc_bytealign_be (w[13], w[14], offset); + w[44] = hc_bytealign_be (w[12], w[13], offset); + w[43] = hc_bytealign_be (w[11], w[12], offset); + w[42] = hc_bytealign_be (w[10], w[11], offset); + w[41] = hc_bytealign_be (w[ 9], w[10], offset); + w[40] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[39] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[38] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[37] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[36] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[35] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[34] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[33] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[32] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[31] = hc_bytealign_be ( 0, w[ 0], offset); w[30] = 0; w[29] = 0; w[28] = 0; @@ -22781,38 +22619,38 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 32: - w[63] = hc_bytealign (w[30], w[31], offset); - w[62] = hc_bytealign (w[29], w[30], offset); - w[61] = hc_bytealign (w[28], w[29], offset); - w[60] = hc_bytealign (w[27], w[28], offset); - w[59] = hc_bytealign (w[26], w[27], offset); - w[58] = hc_bytealign (w[25], w[26], offset); - w[57] = hc_bytealign (w[24], w[25], offset); - w[56] = hc_bytealign (w[23], w[24], offset); - w[55] = hc_bytealign (w[22], w[23], offset); - w[54] = hc_bytealign (w[21], w[22], offset); - w[53] = hc_bytealign (w[20], w[21], offset); - w[52] = hc_bytealign (w[19], w[20], offset); - w[51] = hc_bytealign (w[18], w[19], offset); - w[50] = hc_bytealign (w[17], w[18], offset); - w[49] = hc_bytealign (w[16], w[17], offset); - w[48] = hc_bytealign (w[15], w[16], offset); - w[47] = hc_bytealign (w[14], w[15], offset); - w[46] = hc_bytealign (w[13], w[14], offset); - w[45] = hc_bytealign (w[12], w[13], offset); - w[44] = hc_bytealign (w[11], w[12], offset); - w[43] = hc_bytealign (w[10], w[11], offset); - w[42] = hc_bytealign (w[ 9], w[10], offset); - w[41] = hc_bytealign (w[ 8], w[ 9], offset); - w[40] = hc_bytealign (w[ 7], w[ 8], offset); - w[39] = hc_bytealign (w[ 6], w[ 7], offset); - w[38] = hc_bytealign (w[ 5], w[ 6], offset); - w[37] = hc_bytealign (w[ 4], w[ 5], offset); - w[36] = hc_bytealign (w[ 3], w[ 4], offset); - w[35] = hc_bytealign (w[ 2], w[ 3], offset); - w[34] = hc_bytealign (w[ 1], w[ 2], offset); - w[33] = hc_bytealign (w[ 0], w[ 1], offset); - w[32] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[30], w[31], offset); + w[62] = hc_bytealign_be (w[29], w[30], offset); + w[61] = hc_bytealign_be (w[28], w[29], offset); + w[60] = hc_bytealign_be (w[27], w[28], offset); + w[59] = hc_bytealign_be (w[26], w[27], offset); + w[58] = hc_bytealign_be (w[25], w[26], offset); + w[57] = hc_bytealign_be (w[24], w[25], offset); + w[56] = hc_bytealign_be (w[23], w[24], offset); + w[55] = hc_bytealign_be (w[22], w[23], offset); + w[54] = hc_bytealign_be (w[21], w[22], offset); + w[53] = hc_bytealign_be (w[20], w[21], offset); + w[52] = hc_bytealign_be (w[19], w[20], offset); + w[51] = hc_bytealign_be (w[18], w[19], offset); + w[50] = hc_bytealign_be (w[17], w[18], offset); + w[49] = hc_bytealign_be (w[16], w[17], offset); + w[48] = hc_bytealign_be (w[15], w[16], offset); + w[47] = hc_bytealign_be (w[14], w[15], offset); + w[46] = hc_bytealign_be (w[13], w[14], offset); + w[45] = hc_bytealign_be (w[12], w[13], offset); + w[44] = hc_bytealign_be (w[11], w[12], offset); + w[43] = hc_bytealign_be (w[10], w[11], offset); + w[42] = hc_bytealign_be (w[ 9], w[10], offset); + w[41] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[40] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[39] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[38] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[37] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[36] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[35] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[34] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[33] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[32] = hc_bytealign_be ( 0, w[ 0], offset); w[31] = 0; w[30] = 0; w[29] = 0; @@ -22849,37 +22687,37 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 33: - w[63] = hc_bytealign (w[29], w[30], offset); - w[62] = hc_bytealign (w[28], w[29], offset); - w[61] = hc_bytealign (w[27], w[28], offset); - w[60] = hc_bytealign (w[26], w[27], offset); - w[59] = hc_bytealign (w[25], w[26], offset); - w[58] = hc_bytealign (w[24], w[25], offset); - w[57] = hc_bytealign (w[23], w[24], offset); - w[56] = hc_bytealign (w[22], w[23], offset); - w[55] = hc_bytealign (w[21], w[22], offset); - w[54] = hc_bytealign (w[20], w[21], offset); - w[53] = hc_bytealign (w[19], w[20], offset); - w[52] = hc_bytealign (w[18], w[19], offset); - w[51] = hc_bytealign (w[17], w[18], offset); - w[50] = hc_bytealign (w[16], w[17], offset); - w[49] = hc_bytealign (w[15], w[16], offset); - w[48] = hc_bytealign (w[14], w[15], offset); - w[47] = hc_bytealign (w[13], w[14], offset); - w[46] = hc_bytealign (w[12], w[13], offset); - w[45] = hc_bytealign (w[11], w[12], offset); - w[44] = hc_bytealign (w[10], w[11], offset); - w[43] = hc_bytealign (w[ 9], w[10], offset); - w[42] = hc_bytealign (w[ 8], w[ 9], offset); - w[41] = hc_bytealign (w[ 7], w[ 8], offset); - w[40] = hc_bytealign (w[ 6], w[ 7], offset); - w[39] = hc_bytealign (w[ 5], w[ 6], offset); - w[38] = hc_bytealign (w[ 4], w[ 5], offset); - w[37] = hc_bytealign (w[ 3], w[ 4], offset); - w[36] = hc_bytealign (w[ 2], w[ 3], offset); - w[35] = hc_bytealign (w[ 1], w[ 2], offset); - w[34] = hc_bytealign (w[ 0], w[ 1], offset); - w[33] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[29], w[30], offset); + w[62] = hc_bytealign_be (w[28], w[29], offset); + w[61] = hc_bytealign_be (w[27], w[28], offset); + w[60] = hc_bytealign_be (w[26], w[27], offset); + w[59] = hc_bytealign_be (w[25], w[26], offset); + w[58] = hc_bytealign_be (w[24], w[25], offset); + w[57] = hc_bytealign_be (w[23], w[24], offset); + w[56] = hc_bytealign_be (w[22], w[23], offset); + w[55] = hc_bytealign_be (w[21], w[22], offset); + w[54] = hc_bytealign_be (w[20], w[21], offset); + w[53] = hc_bytealign_be (w[19], w[20], offset); + w[52] = hc_bytealign_be (w[18], w[19], offset); + w[51] = hc_bytealign_be (w[17], w[18], offset); + w[50] = hc_bytealign_be (w[16], w[17], offset); + w[49] = hc_bytealign_be (w[15], w[16], offset); + w[48] = hc_bytealign_be (w[14], w[15], offset); + w[47] = hc_bytealign_be (w[13], w[14], offset); + w[46] = hc_bytealign_be (w[12], w[13], offset); + w[45] = hc_bytealign_be (w[11], w[12], offset); + w[44] = hc_bytealign_be (w[10], w[11], offset); + w[43] = hc_bytealign_be (w[ 9], w[10], offset); + w[42] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[41] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[40] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[39] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[38] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[37] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[36] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[35] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[34] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[33] = hc_bytealign_be ( 0, w[ 0], offset); w[32] = 0; w[31] = 0; w[30] = 0; @@ -22917,36 +22755,36 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 34: - w[63] = hc_bytealign (w[28], w[29], offset); - w[62] = hc_bytealign (w[27], w[28], offset); - w[61] = hc_bytealign (w[26], w[27], offset); - w[60] = hc_bytealign (w[25], w[26], offset); - w[59] = hc_bytealign (w[24], w[25], offset); - w[58] = hc_bytealign (w[23], w[24], offset); - w[57] = hc_bytealign (w[22], w[23], offset); - w[56] = hc_bytealign (w[21], w[22], offset); - w[55] = hc_bytealign (w[20], w[21], offset); - w[54] = hc_bytealign (w[19], w[20], offset); - w[53] = hc_bytealign (w[18], w[19], offset); - w[52] = hc_bytealign (w[17], w[18], offset); - w[51] = hc_bytealign (w[16], w[17], offset); - w[50] = hc_bytealign (w[15], w[16], offset); - w[49] = hc_bytealign (w[14], w[15], offset); - w[48] = hc_bytealign (w[13], w[14], offset); - w[47] = hc_bytealign (w[12], w[13], offset); - w[46] = hc_bytealign (w[11], w[12], offset); - w[45] = hc_bytealign (w[10], w[11], offset); - w[44] = hc_bytealign (w[ 9], w[10], offset); - w[43] = hc_bytealign (w[ 8], w[ 9], offset); - w[42] = hc_bytealign (w[ 7], w[ 8], offset); - w[41] = hc_bytealign (w[ 6], w[ 7], offset); - w[40] = hc_bytealign (w[ 5], w[ 6], offset); - w[39] = hc_bytealign (w[ 4], w[ 5], offset); - w[38] = hc_bytealign (w[ 3], w[ 4], offset); - w[37] = hc_bytealign (w[ 2], w[ 3], offset); - w[36] = hc_bytealign (w[ 1], w[ 2], offset); - w[35] = hc_bytealign (w[ 0], w[ 1], offset); - w[34] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[28], w[29], offset); + w[62] = hc_bytealign_be (w[27], w[28], offset); + w[61] = hc_bytealign_be (w[26], w[27], offset); + w[60] = hc_bytealign_be (w[25], w[26], offset); + w[59] = hc_bytealign_be (w[24], w[25], offset); + w[58] = hc_bytealign_be (w[23], w[24], offset); + w[57] = hc_bytealign_be (w[22], w[23], offset); + w[56] = hc_bytealign_be (w[21], w[22], offset); + w[55] = hc_bytealign_be (w[20], w[21], offset); + w[54] = hc_bytealign_be (w[19], w[20], offset); + w[53] = hc_bytealign_be (w[18], w[19], offset); + w[52] = hc_bytealign_be (w[17], w[18], offset); + w[51] = hc_bytealign_be (w[16], w[17], offset); + w[50] = hc_bytealign_be (w[15], w[16], offset); + w[49] = hc_bytealign_be (w[14], w[15], offset); + w[48] = hc_bytealign_be (w[13], w[14], offset); + w[47] = hc_bytealign_be (w[12], w[13], offset); + w[46] = hc_bytealign_be (w[11], w[12], offset); + w[45] = hc_bytealign_be (w[10], w[11], offset); + w[44] = hc_bytealign_be (w[ 9], w[10], offset); + w[43] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[42] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[41] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[40] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[39] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[38] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[37] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[36] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[35] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[34] = hc_bytealign_be ( 0, w[ 0], offset); w[33] = 0; w[32] = 0; w[31] = 0; @@ -22985,35 +22823,35 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 35: - w[63] = hc_bytealign (w[27], w[28], offset); - w[62] = hc_bytealign (w[26], w[27], offset); - w[61] = hc_bytealign (w[25], w[26], offset); - w[60] = hc_bytealign (w[24], w[25], offset); - w[59] = hc_bytealign (w[23], w[24], offset); - w[58] = hc_bytealign (w[22], w[23], offset); - w[57] = hc_bytealign (w[21], w[22], offset); - w[56] = hc_bytealign (w[20], w[21], offset); - w[55] = hc_bytealign (w[19], w[20], offset); - w[54] = hc_bytealign (w[18], w[19], offset); - w[53] = hc_bytealign (w[17], w[18], offset); - w[52] = hc_bytealign (w[16], w[17], offset); - w[51] = hc_bytealign (w[15], w[16], offset); - w[50] = hc_bytealign (w[14], w[15], offset); - w[49] = hc_bytealign (w[13], w[14], offset); - w[48] = hc_bytealign (w[12], w[13], offset); - w[47] = hc_bytealign (w[11], w[12], offset); - w[46] = hc_bytealign (w[10], w[11], offset); - w[45] = hc_bytealign (w[ 9], w[10], offset); - w[44] = hc_bytealign (w[ 8], w[ 9], offset); - w[43] = hc_bytealign (w[ 7], w[ 8], offset); - w[42] = hc_bytealign (w[ 6], w[ 7], offset); - w[41] = hc_bytealign (w[ 5], w[ 6], offset); - w[40] = hc_bytealign (w[ 4], w[ 5], offset); - w[39] = hc_bytealign (w[ 3], w[ 4], offset); - w[38] = hc_bytealign (w[ 2], w[ 3], offset); - w[37] = hc_bytealign (w[ 1], w[ 2], offset); - w[36] = hc_bytealign (w[ 0], w[ 1], offset); - w[35] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[27], w[28], offset); + w[62] = hc_bytealign_be (w[26], w[27], offset); + w[61] = hc_bytealign_be (w[25], w[26], offset); + w[60] = hc_bytealign_be (w[24], w[25], offset); + w[59] = hc_bytealign_be (w[23], w[24], offset); + w[58] = hc_bytealign_be (w[22], w[23], offset); + w[57] = hc_bytealign_be (w[21], w[22], offset); + w[56] = hc_bytealign_be (w[20], w[21], offset); + w[55] = hc_bytealign_be (w[19], w[20], offset); + w[54] = hc_bytealign_be (w[18], w[19], offset); + w[53] = hc_bytealign_be (w[17], w[18], offset); + w[52] = hc_bytealign_be (w[16], w[17], offset); + w[51] = hc_bytealign_be (w[15], w[16], offset); + w[50] = hc_bytealign_be (w[14], w[15], offset); + w[49] = hc_bytealign_be (w[13], w[14], offset); + w[48] = hc_bytealign_be (w[12], w[13], offset); + w[47] = hc_bytealign_be (w[11], w[12], offset); + w[46] = hc_bytealign_be (w[10], w[11], offset); + w[45] = hc_bytealign_be (w[ 9], w[10], offset); + w[44] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[43] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[42] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[41] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[40] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[39] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[38] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[37] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[36] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[35] = hc_bytealign_be ( 0, w[ 0], offset); w[34] = 0; w[33] = 0; w[32] = 0; @@ -23053,34 +22891,34 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 36: - w[63] = hc_bytealign (w[26], w[27], offset); - w[62] = hc_bytealign (w[25], w[26], offset); - w[61] = hc_bytealign (w[24], w[25], offset); - w[60] = hc_bytealign (w[23], w[24], offset); - w[59] = hc_bytealign (w[22], w[23], offset); - w[58] = hc_bytealign (w[21], w[22], offset); - w[57] = hc_bytealign (w[20], w[21], offset); - w[56] = hc_bytealign (w[19], w[20], offset); - w[55] = hc_bytealign (w[18], w[19], offset); - w[54] = hc_bytealign (w[17], w[18], offset); - w[53] = hc_bytealign (w[16], w[17], offset); - w[52] = hc_bytealign (w[15], w[16], offset); - w[51] = hc_bytealign (w[14], w[15], offset); - w[50] = hc_bytealign (w[13], w[14], offset); - w[49] = hc_bytealign (w[12], w[13], offset); - w[48] = hc_bytealign (w[11], w[12], offset); - w[47] = hc_bytealign (w[10], w[11], offset); - w[46] = hc_bytealign (w[ 9], w[10], offset); - w[45] = hc_bytealign (w[ 8], w[ 9], offset); - w[44] = hc_bytealign (w[ 7], w[ 8], offset); - w[43] = hc_bytealign (w[ 6], w[ 7], offset); - w[42] = hc_bytealign (w[ 5], w[ 6], offset); - w[41] = hc_bytealign (w[ 4], w[ 5], offset); - w[40] = hc_bytealign (w[ 3], w[ 4], offset); - w[39] = hc_bytealign (w[ 2], w[ 3], offset); - w[38] = hc_bytealign (w[ 1], w[ 2], offset); - w[37] = hc_bytealign (w[ 0], w[ 1], offset); - w[36] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[26], w[27], offset); + w[62] = hc_bytealign_be (w[25], w[26], offset); + w[61] = hc_bytealign_be (w[24], w[25], offset); + w[60] = hc_bytealign_be (w[23], w[24], offset); + w[59] = hc_bytealign_be (w[22], w[23], offset); + w[58] = hc_bytealign_be (w[21], w[22], offset); + w[57] = hc_bytealign_be (w[20], w[21], offset); + w[56] = hc_bytealign_be (w[19], w[20], offset); + w[55] = hc_bytealign_be (w[18], w[19], offset); + w[54] = hc_bytealign_be (w[17], w[18], offset); + w[53] = hc_bytealign_be (w[16], w[17], offset); + w[52] = hc_bytealign_be (w[15], w[16], offset); + w[51] = hc_bytealign_be (w[14], w[15], offset); + w[50] = hc_bytealign_be (w[13], w[14], offset); + w[49] = hc_bytealign_be (w[12], w[13], offset); + w[48] = hc_bytealign_be (w[11], w[12], offset); + w[47] = hc_bytealign_be (w[10], w[11], offset); + w[46] = hc_bytealign_be (w[ 9], w[10], offset); + w[45] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[44] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[43] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[42] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[41] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[40] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[39] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[38] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[37] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[36] = hc_bytealign_be ( 0, w[ 0], offset); w[35] = 0; w[34] = 0; w[33] = 0; @@ -23121,33 +22959,33 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 37: - w[63] = hc_bytealign (w[25], w[26], offset); - w[62] = hc_bytealign (w[24], w[25], offset); - w[61] = hc_bytealign (w[23], w[24], offset); - w[60] = hc_bytealign (w[22], w[23], offset); - w[59] = hc_bytealign (w[21], w[22], offset); - w[58] = hc_bytealign (w[20], w[21], offset); - w[57] = hc_bytealign (w[19], w[20], offset); - w[56] = hc_bytealign (w[18], w[19], offset); - w[55] = hc_bytealign (w[17], w[18], offset); - w[54] = hc_bytealign (w[16], w[17], offset); - w[53] = hc_bytealign (w[15], w[16], offset); - w[52] = hc_bytealign (w[14], w[15], offset); - w[51] = hc_bytealign (w[13], w[14], offset); - w[50] = hc_bytealign (w[12], w[13], offset); - w[49] = hc_bytealign (w[11], w[12], offset); - w[48] = hc_bytealign (w[10], w[11], offset); - w[47] = hc_bytealign (w[ 9], w[10], offset); - w[46] = hc_bytealign (w[ 8], w[ 9], offset); - w[45] = hc_bytealign (w[ 7], w[ 8], offset); - w[44] = hc_bytealign (w[ 6], w[ 7], offset); - w[43] = hc_bytealign (w[ 5], w[ 6], offset); - w[42] = hc_bytealign (w[ 4], w[ 5], offset); - w[41] = hc_bytealign (w[ 3], w[ 4], offset); - w[40] = hc_bytealign (w[ 2], w[ 3], offset); - w[39] = hc_bytealign (w[ 1], w[ 2], offset); - w[38] = hc_bytealign (w[ 0], w[ 1], offset); - w[37] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[25], w[26], offset); + w[62] = hc_bytealign_be (w[24], w[25], offset); + w[61] = hc_bytealign_be (w[23], w[24], offset); + w[60] = hc_bytealign_be (w[22], w[23], offset); + w[59] = hc_bytealign_be (w[21], w[22], offset); + w[58] = hc_bytealign_be (w[20], w[21], offset); + w[57] = hc_bytealign_be (w[19], w[20], offset); + w[56] = hc_bytealign_be (w[18], w[19], offset); + w[55] = hc_bytealign_be (w[17], w[18], offset); + w[54] = hc_bytealign_be (w[16], w[17], offset); + w[53] = hc_bytealign_be (w[15], w[16], offset); + w[52] = hc_bytealign_be (w[14], w[15], offset); + w[51] = hc_bytealign_be (w[13], w[14], offset); + w[50] = hc_bytealign_be (w[12], w[13], offset); + w[49] = hc_bytealign_be (w[11], w[12], offset); + w[48] = hc_bytealign_be (w[10], w[11], offset); + w[47] = hc_bytealign_be (w[ 9], w[10], offset); + w[46] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[45] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[44] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[43] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[42] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[41] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[40] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[39] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[38] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[37] = hc_bytealign_be ( 0, w[ 0], offset); w[36] = 0; w[35] = 0; w[34] = 0; @@ -23189,32 +23027,32 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 38: - w[63] = hc_bytealign (w[24], w[25], offset); - w[62] = hc_bytealign (w[23], w[24], offset); - w[61] = hc_bytealign (w[22], w[23], offset); - w[60] = hc_bytealign (w[21], w[22], offset); - w[59] = hc_bytealign (w[20], w[21], offset); - w[58] = hc_bytealign (w[19], w[20], offset); - w[57] = hc_bytealign (w[18], w[19], offset); - w[56] = hc_bytealign (w[17], w[18], offset); - w[55] = hc_bytealign (w[16], w[17], offset); - w[54] = hc_bytealign (w[15], w[16], offset); - w[53] = hc_bytealign (w[14], w[15], offset); - w[52] = hc_bytealign (w[13], w[14], offset); - w[51] = hc_bytealign (w[12], w[13], offset); - w[50] = hc_bytealign (w[11], w[12], offset); - w[49] = hc_bytealign (w[10], w[11], offset); - w[48] = hc_bytealign (w[ 9], w[10], offset); - w[47] = hc_bytealign (w[ 8], w[ 9], offset); - w[46] = hc_bytealign (w[ 7], w[ 8], offset); - w[45] = hc_bytealign (w[ 6], w[ 7], offset); - w[44] = hc_bytealign (w[ 5], w[ 6], offset); - w[43] = hc_bytealign (w[ 4], w[ 5], offset); - w[42] = hc_bytealign (w[ 3], w[ 4], offset); - w[41] = hc_bytealign (w[ 2], w[ 3], offset); - w[40] = hc_bytealign (w[ 1], w[ 2], offset); - w[39] = hc_bytealign (w[ 0], w[ 1], offset); - w[38] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[24], w[25], offset); + w[62] = hc_bytealign_be (w[23], w[24], offset); + w[61] = hc_bytealign_be (w[22], w[23], offset); + w[60] = hc_bytealign_be (w[21], w[22], offset); + w[59] = hc_bytealign_be (w[20], w[21], offset); + w[58] = hc_bytealign_be (w[19], w[20], offset); + w[57] = hc_bytealign_be (w[18], w[19], offset); + w[56] = hc_bytealign_be (w[17], w[18], offset); + w[55] = hc_bytealign_be (w[16], w[17], offset); + w[54] = hc_bytealign_be (w[15], w[16], offset); + w[53] = hc_bytealign_be (w[14], w[15], offset); + w[52] = hc_bytealign_be (w[13], w[14], offset); + w[51] = hc_bytealign_be (w[12], w[13], offset); + w[50] = hc_bytealign_be (w[11], w[12], offset); + w[49] = hc_bytealign_be (w[10], w[11], offset); + w[48] = hc_bytealign_be (w[ 9], w[10], offset); + w[47] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[46] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[45] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[44] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[43] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[42] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[41] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[40] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[39] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[38] = hc_bytealign_be ( 0, w[ 0], offset); w[37] = 0; w[36] = 0; w[35] = 0; @@ -23257,31 +23095,31 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 39: - w[63] = hc_bytealign (w[23], w[24], offset); - w[62] = hc_bytealign (w[22], w[23], offset); - w[61] = hc_bytealign (w[21], w[22], offset); - w[60] = hc_bytealign (w[20], w[21], offset); - w[59] = hc_bytealign (w[19], w[20], offset); - w[58] = hc_bytealign (w[18], w[19], offset); - w[57] = hc_bytealign (w[17], w[18], offset); - w[56] = hc_bytealign (w[16], w[17], offset); - w[55] = hc_bytealign (w[15], w[16], offset); - w[54] = hc_bytealign (w[14], w[15], offset); - w[53] = hc_bytealign (w[13], w[14], offset); - w[52] = hc_bytealign (w[12], w[13], offset); - w[51] = hc_bytealign (w[11], w[12], offset); - w[50] = hc_bytealign (w[10], w[11], offset); - w[49] = hc_bytealign (w[ 9], w[10], offset); - w[48] = hc_bytealign (w[ 8], w[ 9], offset); - w[47] = hc_bytealign (w[ 7], w[ 8], offset); - w[46] = hc_bytealign (w[ 6], w[ 7], offset); - w[45] = hc_bytealign (w[ 5], w[ 6], offset); - w[44] = hc_bytealign (w[ 4], w[ 5], offset); - w[43] = hc_bytealign (w[ 3], w[ 4], offset); - w[42] = hc_bytealign (w[ 2], w[ 3], offset); - w[41] = hc_bytealign (w[ 1], w[ 2], offset); - w[40] = hc_bytealign (w[ 0], w[ 1], offset); - w[39] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[23], w[24], offset); + w[62] = hc_bytealign_be (w[22], w[23], offset); + w[61] = hc_bytealign_be (w[21], w[22], offset); + w[60] = hc_bytealign_be (w[20], w[21], offset); + w[59] = hc_bytealign_be (w[19], w[20], offset); + w[58] = hc_bytealign_be (w[18], w[19], offset); + w[57] = hc_bytealign_be (w[17], w[18], offset); + w[56] = hc_bytealign_be (w[16], w[17], offset); + w[55] = hc_bytealign_be (w[15], w[16], offset); + w[54] = hc_bytealign_be (w[14], w[15], offset); + w[53] = hc_bytealign_be (w[13], w[14], offset); + w[52] = hc_bytealign_be (w[12], w[13], offset); + w[51] = hc_bytealign_be (w[11], w[12], offset); + w[50] = hc_bytealign_be (w[10], w[11], offset); + w[49] = hc_bytealign_be (w[ 9], w[10], offset); + w[48] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[47] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[46] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[45] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[44] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[43] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[42] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[41] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[40] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[39] = hc_bytealign_be ( 0, w[ 0], offset); w[38] = 0; w[37] = 0; w[36] = 0; @@ -23325,30 +23163,30 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 40: - w[63] = hc_bytealign (w[22], w[23], offset); - w[62] = hc_bytealign (w[21], w[22], offset); - w[61] = hc_bytealign (w[20], w[21], offset); - w[60] = hc_bytealign (w[19], w[20], offset); - w[59] = hc_bytealign (w[18], w[19], offset); - w[58] = hc_bytealign (w[17], w[18], offset); - w[57] = hc_bytealign (w[16], w[17], offset); - w[56] = hc_bytealign (w[15], w[16], offset); - w[55] = hc_bytealign (w[14], w[15], offset); - w[54] = hc_bytealign (w[13], w[14], offset); - w[53] = hc_bytealign (w[12], w[13], offset); - w[52] = hc_bytealign (w[11], w[12], offset); - w[51] = hc_bytealign (w[10], w[11], offset); - w[50] = hc_bytealign (w[ 9], w[10], offset); - w[49] = hc_bytealign (w[ 8], w[ 9], offset); - w[48] = hc_bytealign (w[ 7], w[ 8], offset); - w[47] = hc_bytealign (w[ 6], w[ 7], offset); - w[46] = hc_bytealign (w[ 5], w[ 6], offset); - w[45] = hc_bytealign (w[ 4], w[ 5], offset); - w[44] = hc_bytealign (w[ 3], w[ 4], offset); - w[43] = hc_bytealign (w[ 2], w[ 3], offset); - w[42] = hc_bytealign (w[ 1], w[ 2], offset); - w[41] = hc_bytealign (w[ 0], w[ 1], offset); - w[40] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[22], w[23], offset); + w[62] = hc_bytealign_be (w[21], w[22], offset); + w[61] = hc_bytealign_be (w[20], w[21], offset); + w[60] = hc_bytealign_be (w[19], w[20], offset); + w[59] = hc_bytealign_be (w[18], w[19], offset); + w[58] = hc_bytealign_be (w[17], w[18], offset); + w[57] = hc_bytealign_be (w[16], w[17], offset); + w[56] = hc_bytealign_be (w[15], w[16], offset); + w[55] = hc_bytealign_be (w[14], w[15], offset); + w[54] = hc_bytealign_be (w[13], w[14], offset); + w[53] = hc_bytealign_be (w[12], w[13], offset); + w[52] = hc_bytealign_be (w[11], w[12], offset); + w[51] = hc_bytealign_be (w[10], w[11], offset); + w[50] = hc_bytealign_be (w[ 9], w[10], offset); + w[49] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[48] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[47] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[46] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[45] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[44] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[43] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[42] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[41] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[40] = hc_bytealign_be ( 0, w[ 0], offset); w[39] = 0; w[38] = 0; w[37] = 0; @@ -23393,29 +23231,29 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 41: - w[63] = hc_bytealign (w[21], w[22], offset); - w[62] = hc_bytealign (w[20], w[21], offset); - w[61] = hc_bytealign (w[19], w[20], offset); - w[60] = hc_bytealign (w[18], w[19], offset); - w[59] = hc_bytealign (w[17], w[18], offset); - w[58] = hc_bytealign (w[16], w[17], offset); - w[57] = hc_bytealign (w[15], w[16], offset); - w[56] = hc_bytealign (w[14], w[15], offset); - w[55] = hc_bytealign (w[13], w[14], offset); - w[54] = hc_bytealign (w[12], w[13], offset); - w[53] = hc_bytealign (w[11], w[12], offset); - w[52] = hc_bytealign (w[10], w[11], offset); - w[51] = hc_bytealign (w[ 9], w[10], offset); - w[50] = hc_bytealign (w[ 8], w[ 9], offset); - w[49] = hc_bytealign (w[ 7], w[ 8], offset); - w[48] = hc_bytealign (w[ 6], w[ 7], offset); - w[47] = hc_bytealign (w[ 5], w[ 6], offset); - w[46] = hc_bytealign (w[ 4], w[ 5], offset); - w[45] = hc_bytealign (w[ 3], w[ 4], offset); - w[44] = hc_bytealign (w[ 2], w[ 3], offset); - w[43] = hc_bytealign (w[ 1], w[ 2], offset); - w[42] = hc_bytealign (w[ 0], w[ 1], offset); - w[41] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[21], w[22], offset); + w[62] = hc_bytealign_be (w[20], w[21], offset); + w[61] = hc_bytealign_be (w[19], w[20], offset); + w[60] = hc_bytealign_be (w[18], w[19], offset); + w[59] = hc_bytealign_be (w[17], w[18], offset); + w[58] = hc_bytealign_be (w[16], w[17], offset); + w[57] = hc_bytealign_be (w[15], w[16], offset); + w[56] = hc_bytealign_be (w[14], w[15], offset); + w[55] = hc_bytealign_be (w[13], w[14], offset); + w[54] = hc_bytealign_be (w[12], w[13], offset); + w[53] = hc_bytealign_be (w[11], w[12], offset); + w[52] = hc_bytealign_be (w[10], w[11], offset); + w[51] = hc_bytealign_be (w[ 9], w[10], offset); + w[50] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[49] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[48] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[47] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[46] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[45] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[44] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[43] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[42] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[41] = hc_bytealign_be ( 0, w[ 0], offset); w[40] = 0; w[39] = 0; w[38] = 0; @@ -23461,28 +23299,28 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 42: - w[63] = hc_bytealign (w[20], w[21], offset); - w[62] = hc_bytealign (w[19], w[20], offset); - w[61] = hc_bytealign (w[18], w[19], offset); - w[60] = hc_bytealign (w[17], w[18], offset); - w[59] = hc_bytealign (w[16], w[17], offset); - w[58] = hc_bytealign (w[15], w[16], offset); - w[57] = hc_bytealign (w[14], w[15], offset); - w[56] = hc_bytealign (w[13], w[14], offset); - w[55] = hc_bytealign (w[12], w[13], offset); - w[54] = hc_bytealign (w[11], w[12], offset); - w[53] = hc_bytealign (w[10], w[11], offset); - w[52] = hc_bytealign (w[ 9], w[10], offset); - w[51] = hc_bytealign (w[ 8], w[ 9], offset); - w[50] = hc_bytealign (w[ 7], w[ 8], offset); - w[49] = hc_bytealign (w[ 6], w[ 7], offset); - w[48] = hc_bytealign (w[ 5], w[ 6], offset); - w[47] = hc_bytealign (w[ 4], w[ 5], offset); - w[46] = hc_bytealign (w[ 3], w[ 4], offset); - w[45] = hc_bytealign (w[ 2], w[ 3], offset); - w[44] = hc_bytealign (w[ 1], w[ 2], offset); - w[43] = hc_bytealign (w[ 0], w[ 1], offset); - w[42] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[20], w[21], offset); + w[62] = hc_bytealign_be (w[19], w[20], offset); + w[61] = hc_bytealign_be (w[18], w[19], offset); + w[60] = hc_bytealign_be (w[17], w[18], offset); + w[59] = hc_bytealign_be (w[16], w[17], offset); + w[58] = hc_bytealign_be (w[15], w[16], offset); + w[57] = hc_bytealign_be (w[14], w[15], offset); + w[56] = hc_bytealign_be (w[13], w[14], offset); + w[55] = hc_bytealign_be (w[12], w[13], offset); + w[54] = hc_bytealign_be (w[11], w[12], offset); + w[53] = hc_bytealign_be (w[10], w[11], offset); + w[52] = hc_bytealign_be (w[ 9], w[10], offset); + w[51] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[50] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[49] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[48] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[47] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[46] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[45] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[44] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[43] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[42] = hc_bytealign_be ( 0, w[ 0], offset); w[41] = 0; w[40] = 0; w[39] = 0; @@ -23529,27 +23367,27 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 43: - w[63] = hc_bytealign (w[19], w[20], offset); - w[62] = hc_bytealign (w[18], w[19], offset); - w[61] = hc_bytealign (w[17], w[18], offset); - w[60] = hc_bytealign (w[16], w[17], offset); - w[59] = hc_bytealign (w[15], w[16], offset); - w[58] = hc_bytealign (w[14], w[15], offset); - w[57] = hc_bytealign (w[13], w[14], offset); - w[56] = hc_bytealign (w[12], w[13], offset); - w[55] = hc_bytealign (w[11], w[12], offset); - w[54] = hc_bytealign (w[10], w[11], offset); - w[53] = hc_bytealign (w[ 9], w[10], offset); - w[52] = hc_bytealign (w[ 8], w[ 9], offset); - w[51] = hc_bytealign (w[ 7], w[ 8], offset); - w[50] = hc_bytealign (w[ 6], w[ 7], offset); - w[49] = hc_bytealign (w[ 5], w[ 6], offset); - w[48] = hc_bytealign (w[ 4], w[ 5], offset); - w[47] = hc_bytealign (w[ 3], w[ 4], offset); - w[46] = hc_bytealign (w[ 2], w[ 3], offset); - w[45] = hc_bytealign (w[ 1], w[ 2], offset); - w[44] = hc_bytealign (w[ 0], w[ 1], offset); - w[43] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[19], w[20], offset); + w[62] = hc_bytealign_be (w[18], w[19], offset); + w[61] = hc_bytealign_be (w[17], w[18], offset); + w[60] = hc_bytealign_be (w[16], w[17], offset); + w[59] = hc_bytealign_be (w[15], w[16], offset); + w[58] = hc_bytealign_be (w[14], w[15], offset); + w[57] = hc_bytealign_be (w[13], w[14], offset); + w[56] = hc_bytealign_be (w[12], w[13], offset); + w[55] = hc_bytealign_be (w[11], w[12], offset); + w[54] = hc_bytealign_be (w[10], w[11], offset); + w[53] = hc_bytealign_be (w[ 9], w[10], offset); + w[52] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[51] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[50] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[49] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[48] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[47] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[46] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[45] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[44] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[43] = hc_bytealign_be ( 0, w[ 0], offset); w[42] = 0; w[41] = 0; w[40] = 0; @@ -23597,26 +23435,26 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 44: - w[63] = hc_bytealign (w[18], w[19], offset); - w[62] = hc_bytealign (w[17], w[18], offset); - w[61] = hc_bytealign (w[16], w[17], offset); - w[60] = hc_bytealign (w[15], w[16], offset); - w[59] = hc_bytealign (w[14], w[15], offset); - w[58] = hc_bytealign (w[13], w[14], offset); - w[57] = hc_bytealign (w[12], w[13], offset); - w[56] = hc_bytealign (w[11], w[12], offset); - w[55] = hc_bytealign (w[10], w[11], offset); - w[54] = hc_bytealign (w[ 9], w[10], offset); - w[53] = hc_bytealign (w[ 8], w[ 9], offset); - w[52] = hc_bytealign (w[ 7], w[ 8], offset); - w[51] = hc_bytealign (w[ 6], w[ 7], offset); - w[50] = hc_bytealign (w[ 5], w[ 6], offset); - w[49] = hc_bytealign (w[ 4], w[ 5], offset); - w[48] = hc_bytealign (w[ 3], w[ 4], offset); - w[47] = hc_bytealign (w[ 2], w[ 3], offset); - w[46] = hc_bytealign (w[ 1], w[ 2], offset); - w[45] = hc_bytealign (w[ 0], w[ 1], offset); - w[44] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[18], w[19], offset); + w[62] = hc_bytealign_be (w[17], w[18], offset); + w[61] = hc_bytealign_be (w[16], w[17], offset); + w[60] = hc_bytealign_be (w[15], w[16], offset); + w[59] = hc_bytealign_be (w[14], w[15], offset); + w[58] = hc_bytealign_be (w[13], w[14], offset); + w[57] = hc_bytealign_be (w[12], w[13], offset); + w[56] = hc_bytealign_be (w[11], w[12], offset); + w[55] = hc_bytealign_be (w[10], w[11], offset); + w[54] = hc_bytealign_be (w[ 9], w[10], offset); + w[53] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[52] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[51] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[50] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[49] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[48] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[47] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[46] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[45] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[44] = hc_bytealign_be ( 0, w[ 0], offset); w[43] = 0; w[42] = 0; w[41] = 0; @@ -23665,25 +23503,25 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 45: - w[63] = hc_bytealign (w[17], w[18], offset); - w[62] = hc_bytealign (w[16], w[17], offset); - w[61] = hc_bytealign (w[15], w[16], offset); - w[60] = hc_bytealign (w[14], w[15], offset); - w[59] = hc_bytealign (w[13], w[14], offset); - w[58] = hc_bytealign (w[12], w[13], offset); - w[57] = hc_bytealign (w[11], w[12], offset); - w[56] = hc_bytealign (w[10], w[11], offset); - w[55] = hc_bytealign (w[ 9], w[10], offset); - w[54] = hc_bytealign (w[ 8], w[ 9], offset); - w[53] = hc_bytealign (w[ 7], w[ 8], offset); - w[52] = hc_bytealign (w[ 6], w[ 7], offset); - w[51] = hc_bytealign (w[ 5], w[ 6], offset); - w[50] = hc_bytealign (w[ 4], w[ 5], offset); - w[49] = hc_bytealign (w[ 3], w[ 4], offset); - w[48] = hc_bytealign (w[ 2], w[ 3], offset); - w[47] = hc_bytealign (w[ 1], w[ 2], offset); - w[46] = hc_bytealign (w[ 0], w[ 1], offset); - w[45] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[17], w[18], offset); + w[62] = hc_bytealign_be (w[16], w[17], offset); + w[61] = hc_bytealign_be (w[15], w[16], offset); + w[60] = hc_bytealign_be (w[14], w[15], offset); + w[59] = hc_bytealign_be (w[13], w[14], offset); + w[58] = hc_bytealign_be (w[12], w[13], offset); + w[57] = hc_bytealign_be (w[11], w[12], offset); + w[56] = hc_bytealign_be (w[10], w[11], offset); + w[55] = hc_bytealign_be (w[ 9], w[10], offset); + w[54] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[53] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[52] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[51] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[50] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[49] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[48] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[47] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[46] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[45] = hc_bytealign_be ( 0, w[ 0], offset); w[44] = 0; w[43] = 0; w[42] = 0; @@ -23733,24 +23571,24 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 46: - w[63] = hc_bytealign (w[16], w[17], offset); - w[62] = hc_bytealign (w[15], w[16], offset); - w[61] = hc_bytealign (w[14], w[15], offset); - w[60] = hc_bytealign (w[13], w[14], offset); - w[59] = hc_bytealign (w[12], w[13], offset); - w[58] = hc_bytealign (w[11], w[12], offset); - w[57] = hc_bytealign (w[10], w[11], offset); - w[56] = hc_bytealign (w[ 9], w[10], offset); - w[55] = hc_bytealign (w[ 8], w[ 9], offset); - w[54] = hc_bytealign (w[ 7], w[ 8], offset); - w[53] = hc_bytealign (w[ 6], w[ 7], offset); - w[52] = hc_bytealign (w[ 5], w[ 6], offset); - w[51] = hc_bytealign (w[ 4], w[ 5], offset); - w[50] = hc_bytealign (w[ 3], w[ 4], offset); - w[49] = hc_bytealign (w[ 2], w[ 3], offset); - w[48] = hc_bytealign (w[ 1], w[ 2], offset); - w[47] = hc_bytealign (w[ 0], w[ 1], offset); - w[46] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[16], w[17], offset); + w[62] = hc_bytealign_be (w[15], w[16], offset); + w[61] = hc_bytealign_be (w[14], w[15], offset); + w[60] = hc_bytealign_be (w[13], w[14], offset); + w[59] = hc_bytealign_be (w[12], w[13], offset); + w[58] = hc_bytealign_be (w[11], w[12], offset); + w[57] = hc_bytealign_be (w[10], w[11], offset); + w[56] = hc_bytealign_be (w[ 9], w[10], offset); + w[55] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[54] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[53] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[52] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[51] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[50] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[49] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[48] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[47] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[46] = hc_bytealign_be ( 0, w[ 0], offset); w[45] = 0; w[44] = 0; w[43] = 0; @@ -23801,23 +23639,23 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 47: - w[63] = hc_bytealign (w[15], w[16], offset); - w[62] = hc_bytealign (w[14], w[15], offset); - w[61] = hc_bytealign (w[13], w[14], offset); - w[60] = hc_bytealign (w[12], w[13], offset); - w[59] = hc_bytealign (w[11], w[12], offset); - w[58] = hc_bytealign (w[10], w[11], offset); - w[57] = hc_bytealign (w[ 9], w[10], offset); - w[56] = hc_bytealign (w[ 8], w[ 9], offset); - w[55] = hc_bytealign (w[ 7], w[ 8], offset); - w[54] = hc_bytealign (w[ 6], w[ 7], offset); - w[53] = hc_bytealign (w[ 5], w[ 6], offset); - w[52] = hc_bytealign (w[ 4], w[ 5], offset); - w[51] = hc_bytealign (w[ 3], w[ 4], offset); - w[50] = hc_bytealign (w[ 2], w[ 3], offset); - w[49] = hc_bytealign (w[ 1], w[ 2], offset); - w[48] = hc_bytealign (w[ 0], w[ 1], offset); - w[47] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[15], w[16], offset); + w[62] = hc_bytealign_be (w[14], w[15], offset); + w[61] = hc_bytealign_be (w[13], w[14], offset); + w[60] = hc_bytealign_be (w[12], w[13], offset); + w[59] = hc_bytealign_be (w[11], w[12], offset); + w[58] = hc_bytealign_be (w[10], w[11], offset); + w[57] = hc_bytealign_be (w[ 9], w[10], offset); + w[56] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[55] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[54] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[53] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[52] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[51] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[50] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[49] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[48] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[47] = hc_bytealign_be ( 0, w[ 0], offset); w[46] = 0; w[45] = 0; w[44] = 0; @@ -23869,22 +23707,22 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 48: - w[63] = hc_bytealign (w[14], w[15], offset); - w[62] = hc_bytealign (w[13], w[14], offset); - w[61] = hc_bytealign (w[12], w[13], offset); - w[60] = hc_bytealign (w[11], w[12], offset); - w[59] = hc_bytealign (w[10], w[11], offset); - w[58] = hc_bytealign (w[ 9], w[10], offset); - w[57] = hc_bytealign (w[ 8], w[ 9], offset); - w[56] = hc_bytealign (w[ 7], w[ 8], offset); - w[55] = hc_bytealign (w[ 6], w[ 7], offset); - w[54] = hc_bytealign (w[ 5], w[ 6], offset); - w[53] = hc_bytealign (w[ 4], w[ 5], offset); - w[52] = hc_bytealign (w[ 3], w[ 4], offset); - w[51] = hc_bytealign (w[ 2], w[ 3], offset); - w[50] = hc_bytealign (w[ 1], w[ 2], offset); - w[49] = hc_bytealign (w[ 0], w[ 1], offset); - w[48] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[14], w[15], offset); + w[62] = hc_bytealign_be (w[13], w[14], offset); + w[61] = hc_bytealign_be (w[12], w[13], offset); + w[60] = hc_bytealign_be (w[11], w[12], offset); + w[59] = hc_bytealign_be (w[10], w[11], offset); + w[58] = hc_bytealign_be (w[ 9], w[10], offset); + w[57] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[56] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[55] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[54] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[53] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[52] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[51] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[50] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[49] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[48] = hc_bytealign_be ( 0, w[ 0], offset); w[47] = 0; w[46] = 0; w[45] = 0; @@ -23937,21 +23775,21 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 49: - w[63] = hc_bytealign (w[13], w[14], offset); - w[62] = hc_bytealign (w[12], w[13], offset); - w[61] = hc_bytealign (w[11], w[12], offset); - w[60] = hc_bytealign (w[10], w[11], offset); - w[59] = hc_bytealign (w[ 9], w[10], offset); - w[58] = hc_bytealign (w[ 8], w[ 9], offset); - w[57] = hc_bytealign (w[ 7], w[ 8], offset); - w[56] = hc_bytealign (w[ 6], w[ 7], offset); - w[55] = hc_bytealign (w[ 5], w[ 6], offset); - w[54] = hc_bytealign (w[ 4], w[ 5], offset); - w[53] = hc_bytealign (w[ 3], w[ 4], offset); - w[52] = hc_bytealign (w[ 2], w[ 3], offset); - w[51] = hc_bytealign (w[ 1], w[ 2], offset); - w[50] = hc_bytealign (w[ 0], w[ 1], offset); - w[49] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[13], w[14], offset); + w[62] = hc_bytealign_be (w[12], w[13], offset); + w[61] = hc_bytealign_be (w[11], w[12], offset); + w[60] = hc_bytealign_be (w[10], w[11], offset); + w[59] = hc_bytealign_be (w[ 9], w[10], offset); + w[58] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[57] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[56] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[55] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[54] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[53] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[52] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[51] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[50] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[49] = hc_bytealign_be ( 0, w[ 0], offset); w[48] = 0; w[47] = 0; w[46] = 0; @@ -24005,20 +23843,20 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 50: - w[63] = hc_bytealign (w[12], w[13], offset); - w[62] = hc_bytealign (w[11], w[12], offset); - w[61] = hc_bytealign (w[10], w[11], offset); - w[60] = hc_bytealign (w[ 9], w[10], offset); - w[59] = hc_bytealign (w[ 8], w[ 9], offset); - w[58] = hc_bytealign (w[ 7], w[ 8], offset); - w[57] = hc_bytealign (w[ 6], w[ 7], offset); - w[56] = hc_bytealign (w[ 5], w[ 6], offset); - w[55] = hc_bytealign (w[ 4], w[ 5], offset); - w[54] = hc_bytealign (w[ 3], w[ 4], offset); - w[53] = hc_bytealign (w[ 2], w[ 3], offset); - w[52] = hc_bytealign (w[ 1], w[ 2], offset); - w[51] = hc_bytealign (w[ 0], w[ 1], offset); - w[50] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[12], w[13], offset); + w[62] = hc_bytealign_be (w[11], w[12], offset); + w[61] = hc_bytealign_be (w[10], w[11], offset); + w[60] = hc_bytealign_be (w[ 9], w[10], offset); + w[59] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[58] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[57] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[56] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[55] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[54] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[53] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[52] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[51] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[50] = hc_bytealign_be ( 0, w[ 0], offset); w[49] = 0; w[48] = 0; w[47] = 0; @@ -24073,19 +23911,19 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 51: - w[63] = hc_bytealign (w[11], w[12], offset); - w[62] = hc_bytealign (w[10], w[11], offset); - w[61] = hc_bytealign (w[ 9], w[10], offset); - w[60] = hc_bytealign (w[ 8], w[ 9], offset); - w[59] = hc_bytealign (w[ 7], w[ 8], offset); - w[58] = hc_bytealign (w[ 6], w[ 7], offset); - w[57] = hc_bytealign (w[ 5], w[ 6], offset); - w[56] = hc_bytealign (w[ 4], w[ 5], offset); - w[55] = hc_bytealign (w[ 3], w[ 4], offset); - w[54] = hc_bytealign (w[ 2], w[ 3], offset); - w[53] = hc_bytealign (w[ 1], w[ 2], offset); - w[52] = hc_bytealign (w[ 0], w[ 1], offset); - w[51] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[11], w[12], offset); + w[62] = hc_bytealign_be (w[10], w[11], offset); + w[61] = hc_bytealign_be (w[ 9], w[10], offset); + w[60] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[59] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[58] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[57] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[56] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[55] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[54] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[53] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[52] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[51] = hc_bytealign_be ( 0, w[ 0], offset); w[50] = 0; w[49] = 0; w[48] = 0; @@ -24141,18 +23979,18 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 52: - w[63] = hc_bytealign (w[10], w[11], offset); - w[62] = hc_bytealign (w[ 9], w[10], offset); - w[61] = hc_bytealign (w[ 8], w[ 9], offset); - w[60] = hc_bytealign (w[ 7], w[ 8], offset); - w[59] = hc_bytealign (w[ 6], w[ 7], offset); - w[58] = hc_bytealign (w[ 5], w[ 6], offset); - w[57] = hc_bytealign (w[ 4], w[ 5], offset); - w[56] = hc_bytealign (w[ 3], w[ 4], offset); - w[55] = hc_bytealign (w[ 2], w[ 3], offset); - w[54] = hc_bytealign (w[ 1], w[ 2], offset); - w[53] = hc_bytealign (w[ 0], w[ 1], offset); - w[52] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[10], w[11], offset); + w[62] = hc_bytealign_be (w[ 9], w[10], offset); + w[61] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[60] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[59] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[58] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[57] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[56] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[55] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[54] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[53] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[52] = hc_bytealign_be ( 0, w[ 0], offset); w[51] = 0; w[50] = 0; w[49] = 0; @@ -24209,17 +24047,17 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 53: - w[63] = hc_bytealign (w[ 9], w[10], offset); - w[62] = hc_bytealign (w[ 8], w[ 9], offset); - w[61] = hc_bytealign (w[ 7], w[ 8], offset); - w[60] = hc_bytealign (w[ 6], w[ 7], offset); - w[59] = hc_bytealign (w[ 5], w[ 6], offset); - w[58] = hc_bytealign (w[ 4], w[ 5], offset); - w[57] = hc_bytealign (w[ 3], w[ 4], offset); - w[56] = hc_bytealign (w[ 2], w[ 3], offset); - w[55] = hc_bytealign (w[ 1], w[ 2], offset); - w[54] = hc_bytealign (w[ 0], w[ 1], offset); - w[53] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 9], w[10], offset); + w[62] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[61] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[60] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[59] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[58] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[57] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[56] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[55] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[54] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[53] = hc_bytealign_be ( 0, w[ 0], offset); w[52] = 0; w[51] = 0; w[50] = 0; @@ -24277,16 +24115,16 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 54: - w[63] = hc_bytealign (w[ 8], w[ 9], offset); - w[62] = hc_bytealign (w[ 7], w[ 8], offset); - w[61] = hc_bytealign (w[ 6], w[ 7], offset); - w[60] = hc_bytealign (w[ 5], w[ 6], offset); - w[59] = hc_bytealign (w[ 4], w[ 5], offset); - w[58] = hc_bytealign (w[ 3], w[ 4], offset); - w[57] = hc_bytealign (w[ 2], w[ 3], offset); - w[56] = hc_bytealign (w[ 1], w[ 2], offset); - w[55] = hc_bytealign (w[ 0], w[ 1], offset); - w[54] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 8], w[ 9], offset); + w[62] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[61] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[60] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[59] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[58] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[57] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[56] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[55] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[54] = hc_bytealign_be ( 0, w[ 0], offset); w[53] = 0; w[52] = 0; w[51] = 0; @@ -24345,15 +24183,15 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 55: - w[63] = hc_bytealign (w[ 7], w[ 8], offset); - w[62] = hc_bytealign (w[ 6], w[ 7], offset); - w[61] = hc_bytealign (w[ 5], w[ 6], offset); - w[60] = hc_bytealign (w[ 4], w[ 5], offset); - w[59] = hc_bytealign (w[ 3], w[ 4], offset); - w[58] = hc_bytealign (w[ 2], w[ 3], offset); - w[57] = hc_bytealign (w[ 1], w[ 2], offset); - w[56] = hc_bytealign (w[ 0], w[ 1], offset); - w[55] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 7], w[ 8], offset); + w[62] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[61] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[60] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[59] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[58] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[57] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[56] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[55] = hc_bytealign_be ( 0, w[ 0], offset); w[54] = 0; w[53] = 0; w[52] = 0; @@ -24413,14 +24251,14 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 56: - w[63] = hc_bytealign (w[ 6], w[ 7], offset); - w[62] = hc_bytealign (w[ 5], w[ 6], offset); - w[61] = hc_bytealign (w[ 4], w[ 5], offset); - w[60] = hc_bytealign (w[ 3], w[ 4], offset); - w[59] = hc_bytealign (w[ 2], w[ 3], offset); - w[58] = hc_bytealign (w[ 1], w[ 2], offset); - w[57] = hc_bytealign (w[ 0], w[ 1], offset); - w[56] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 6], w[ 7], offset); + w[62] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[61] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[60] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[59] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[58] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[57] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[56] = hc_bytealign_be ( 0, w[ 0], offset); w[55] = 0; w[54] = 0; w[53] = 0; @@ -24481,13 +24319,13 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 57: - w[63] = hc_bytealign (w[ 5], w[ 6], offset); - w[62] = hc_bytealign (w[ 4], w[ 5], offset); - w[61] = hc_bytealign (w[ 3], w[ 4], offset); - w[60] = hc_bytealign (w[ 2], w[ 3], offset); - w[59] = hc_bytealign (w[ 1], w[ 2], offset); - w[58] = hc_bytealign (w[ 0], w[ 1], offset); - w[57] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 5], w[ 6], offset); + w[62] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[61] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[60] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[59] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[58] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[57] = hc_bytealign_be ( 0, w[ 0], offset); w[56] = 0; w[55] = 0; w[54] = 0; @@ -24549,12 +24387,12 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 58: - w[63] = hc_bytealign (w[ 4], w[ 5], offset); - w[62] = hc_bytealign (w[ 3], w[ 4], offset); - w[61] = hc_bytealign (w[ 2], w[ 3], offset); - w[60] = hc_bytealign (w[ 1], w[ 2], offset); - w[59] = hc_bytealign (w[ 0], w[ 1], offset); - w[58] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 4], w[ 5], offset); + w[62] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[61] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[60] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[59] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[58] = hc_bytealign_be ( 0, w[ 0], offset); w[57] = 0; w[56] = 0; w[55] = 0; @@ -24617,11 +24455,11 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 59: - w[63] = hc_bytealign (w[ 3], w[ 4], offset); - w[62] = hc_bytealign (w[ 2], w[ 3], offset); - w[61] = hc_bytealign (w[ 1], w[ 2], offset); - w[60] = hc_bytealign (w[ 0], w[ 1], offset); - w[59] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 3], w[ 4], offset); + w[62] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[61] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[60] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[59] = hc_bytealign_be ( 0, w[ 0], offset); w[58] = 0; w[57] = 0; w[56] = 0; @@ -24685,10 +24523,10 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 60: - w[63] = hc_bytealign (w[ 2], w[ 3], offset); - w[62] = hc_bytealign (w[ 1], w[ 2], offset); - w[61] = hc_bytealign (w[ 0], w[ 1], offset); - w[60] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 2], w[ 3], offset); + w[62] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[61] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[60] = hc_bytealign_be ( 0, w[ 0], offset); w[59] = 0; w[58] = 0; w[57] = 0; @@ -24753,9 +24591,9 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 61: - w[63] = hc_bytealign (w[ 1], w[ 2], offset); - w[62] = hc_bytealign (w[ 0], w[ 1], offset); - w[61] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 1], w[ 2], offset); + w[62] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[61] = hc_bytealign_be ( 0, w[ 0], offset); w[60] = 0; w[59] = 0; w[58] = 0; @@ -24821,8 +24659,8 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 62: - w[63] = hc_bytealign (w[ 0], w[ 1], offset); - w[62] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be (w[ 0], w[ 1], offset); + w[62] = hc_bytealign_be ( 0, w[ 0], offset); w[61] = 0; w[60] = 0; w[59] = 0; @@ -24889,7 +24727,7 @@ DECLSPEC void switch_buffer_by_offset_1x64_be (u32x *w, const u32 offset) break; case 63: - w[63] = hc_bytealign ( 0, w[ 0], offset); + w[63] = hc_bytealign_be ( 0, w[ 0], offset); w[62] = 0; w[61] = 0; w[60] = 0; @@ -31429,23 +31267,6 @@ DECLSPEC void switch_buffer_by_offset_le_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); - switch (offset_switch) { case 0: @@ -31768,23 +31589,6 @@ DECLSPEC void switch_buffer_by_offset_le_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; } - - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -32131,23 +31935,6 @@ DECLSPEC void switch_buffer_by_offset_carry_le_S (u32 *w0, u32 *w1, u32 *w2, u32 const int offset_switch = offset / 4; #if defined IS_AMD || defined IS_GENERIC - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); - switch (offset_switch) { case 0: @@ -32606,39 +32393,6 @@ DECLSPEC void switch_buffer_by_offset_carry_le_S (u32 *w0, u32 *w1, u32 *w2, u32 break; } - - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); - c0[0] = swap32_S (c0[0]); - c0[1] = swap32_S (c0[1]); - c0[2] = swap32_S (c0[2]); - c0[3] = swap32_S (c0[3]); - c1[0] = swap32_S (c1[0]); - c1[1] = swap32_S (c1[1]); - c1[2] = swap32_S (c1[2]); - c1[3] = swap32_S (c1[3]); - c2[0] = swap32_S (c2[0]); - c2[1] = swap32_S (c2[1]); - c2[2] = swap32_S (c2[2]); - c2[3] = swap32_S (c2[3]); - c3[0] = swap32_S (c3[0]); - c3[1] = swap32_S (c3[1]); - c3[2] = swap32_S (c3[2]); - c3[3] = swap32_S (c3[3]); #endif #ifdef IS_NV @@ -33448,79 +33202,79 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, switch (offset_switch) { case 0: - w3[3] = hc_bytealign_S (w3[2], w3[3], offset); - w3[2] = hc_bytealign_S (w3[1], w3[2], offset); - w3[1] = hc_bytealign_S (w3[0], w3[1], offset); - w3[0] = hc_bytealign_S (w2[3], w3[0], offset); - w2[3] = hc_bytealign_S (w2[2], w2[3], offset); - w2[2] = hc_bytealign_S (w2[1], w2[2], offset); - w2[1] = hc_bytealign_S (w2[0], w2[1], offset); - w2[0] = hc_bytealign_S (w1[3], w2[0], offset); - w1[3] = hc_bytealign_S (w1[2], w1[3], offset); - w1[2] = hc_bytealign_S (w1[1], w1[2], offset); - w1[1] = hc_bytealign_S (w1[0], w1[1], offset); - w1[0] = hc_bytealign_S (w0[3], w1[0], offset); - w0[3] = hc_bytealign_S (w0[2], w0[3], offset); - w0[2] = hc_bytealign_S (w0[1], w0[2], offset); - w0[1] = hc_bytealign_S (w0[0], w0[1], offset); - w0[0] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be_S ( 0, w0[0], offset); break; case 1: - w3[3] = hc_bytealign_S (w3[1], w3[2], offset); - w3[2] = hc_bytealign_S (w3[0], w3[1], offset); - w3[1] = hc_bytealign_S (w2[3], w3[0], offset); - w3[0] = hc_bytealign_S (w2[2], w2[3], offset); - w2[3] = hc_bytealign_S (w2[1], w2[2], offset); - w2[2] = hc_bytealign_S (w2[0], w2[1], offset); - w2[1] = hc_bytealign_S (w1[3], w2[0], offset); - w2[0] = hc_bytealign_S (w1[2], w1[3], offset); - w1[3] = hc_bytealign_S (w1[1], w1[2], offset); - w1[2] = hc_bytealign_S (w1[0], w1[1], offset); - w1[1] = hc_bytealign_S (w0[3], w1[0], offset); - w1[0] = hc_bytealign_S (w0[2], w0[3], offset); - w0[3] = hc_bytealign_S (w0[1], w0[2], offset); - w0[2] = hc_bytealign_S (w0[0], w0[1], offset); - w0[1] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be_S ( 0, w0[0], offset); w0[0] = 0; break; case 2: - w3[3] = hc_bytealign_S (w3[0], w3[1], offset); - w3[2] = hc_bytealign_S (w2[3], w3[0], offset); - w3[1] = hc_bytealign_S (w2[2], w2[3], offset); - w3[0] = hc_bytealign_S (w2[1], w2[2], offset); - w2[3] = hc_bytealign_S (w2[0], w2[1], offset); - w2[2] = hc_bytealign_S (w1[3], w2[0], offset); - w2[1] = hc_bytealign_S (w1[2], w1[3], offset); - w2[0] = hc_bytealign_S (w1[1], w1[2], offset); - w1[3] = hc_bytealign_S (w1[0], w1[1], offset); - w1[2] = hc_bytealign_S (w0[3], w1[0], offset); - w1[1] = hc_bytealign_S (w0[2], w0[3], offset); - w1[0] = hc_bytealign_S (w0[1], w0[2], offset); - w0[3] = hc_bytealign_S (w0[0], w0[1], offset); - w0[2] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be_S ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - w3[3] = hc_bytealign_S (w2[3], w3[0], offset); - w3[2] = hc_bytealign_S (w2[2], w2[3], offset); - w3[1] = hc_bytealign_S (w2[1], w2[2], offset); - w3[0] = hc_bytealign_S (w2[0], w2[1], offset); - w2[3] = hc_bytealign_S (w1[3], w2[0], offset); - w2[2] = hc_bytealign_S (w1[2], w1[3], offset); - w2[1] = hc_bytealign_S (w1[1], w1[2], offset); - w2[0] = hc_bytealign_S (w1[0], w1[1], offset); - w1[3] = hc_bytealign_S (w0[3], w1[0], offset); - w1[2] = hc_bytealign_S (w0[2], w0[3], offset); - w1[1] = hc_bytealign_S (w0[1], w0[2], offset); - w1[0] = hc_bytealign_S (w0[0], w0[1], offset); - w0[3] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be_S ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -33528,18 +33282,18 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 4: - w3[3] = hc_bytealign_S (w2[2], w2[3], offset); - w3[2] = hc_bytealign_S (w2[1], w2[2], offset); - w3[1] = hc_bytealign_S (w2[0], w2[1], offset); - w3[0] = hc_bytealign_S (w1[3], w2[0], offset); - w2[3] = hc_bytealign_S (w1[2], w1[3], offset); - w2[2] = hc_bytealign_S (w1[1], w1[2], offset); - w2[1] = hc_bytealign_S (w1[0], w1[1], offset); - w2[0] = hc_bytealign_S (w0[3], w1[0], offset); - w1[3] = hc_bytealign_S (w0[2], w0[3], offset); - w1[2] = hc_bytealign_S (w0[1], w0[2], offset); - w1[1] = hc_bytealign_S (w0[0], w0[1], offset); - w1[0] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be_S ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -33548,17 +33302,17 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 5: - w3[3] = hc_bytealign_S (w2[1], w2[2], offset); - w3[2] = hc_bytealign_S (w2[0], w2[1], offset); - w3[1] = hc_bytealign_S (w1[3], w2[0], offset); - w3[0] = hc_bytealign_S (w1[2], w1[3], offset); - w2[3] = hc_bytealign_S (w1[1], w1[2], offset); - w2[2] = hc_bytealign_S (w1[0], w1[1], offset); - w2[1] = hc_bytealign_S (w0[3], w1[0], offset); - w2[0] = hc_bytealign_S (w0[2], w0[3], offset); - w1[3] = hc_bytealign_S (w0[1], w0[2], offset); - w1[2] = hc_bytealign_S (w0[0], w0[1], offset); - w1[1] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be_S ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -33568,16 +33322,16 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 6: - w3[3] = hc_bytealign_S (w2[0], w2[1], offset); - w3[2] = hc_bytealign_S (w1[3], w2[0], offset); - w3[1] = hc_bytealign_S (w1[2], w1[3], offset); - w3[0] = hc_bytealign_S (w1[1], w1[2], offset); - w2[3] = hc_bytealign_S (w1[0], w1[1], offset); - w2[2] = hc_bytealign_S (w0[3], w1[0], offset); - w2[1] = hc_bytealign_S (w0[2], w0[3], offset); - w2[0] = hc_bytealign_S (w0[1], w0[2], offset); - w1[3] = hc_bytealign_S (w0[0], w0[1], offset); - w1[2] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be_S ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -33588,15 +33342,15 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 7: - w3[3] = hc_bytealign_S (w1[3], w2[0], offset); - w3[2] = hc_bytealign_S (w1[2], w1[3], offset); - w3[1] = hc_bytealign_S (w1[1], w1[2], offset); - w3[0] = hc_bytealign_S (w1[0], w1[1], offset); - w2[3] = hc_bytealign_S (w0[3], w1[0], offset); - w2[2] = hc_bytealign_S (w0[2], w0[3], offset); - w2[1] = hc_bytealign_S (w0[1], w0[2], offset); - w2[0] = hc_bytealign_S (w0[0], w0[1], offset); - w1[3] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be_S ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -33608,14 +33362,14 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 8: - w3[3] = hc_bytealign_S (w1[2], w1[3], offset); - w3[2] = hc_bytealign_S (w1[1], w1[2], offset); - w3[1] = hc_bytealign_S (w1[0], w1[1], offset); - w3[0] = hc_bytealign_S (w0[3], w1[0], offset); - w2[3] = hc_bytealign_S (w0[2], w0[3], offset); - w2[2] = hc_bytealign_S (w0[1], w0[2], offset); - w2[1] = hc_bytealign_S (w0[0], w0[1], offset); - w2[0] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be_S ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -33628,13 +33382,13 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 9: - w3[3] = hc_bytealign_S (w1[1], w1[2], offset); - w3[2] = hc_bytealign_S (w1[0], w1[1], offset); - w3[1] = hc_bytealign_S (w0[3], w1[0], offset); - w3[0] = hc_bytealign_S (w0[2], w0[3], offset); - w2[3] = hc_bytealign_S (w0[1], w0[2], offset); - w2[2] = hc_bytealign_S (w0[0], w0[1], offset); - w2[1] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be_S ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -33648,12 +33402,12 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 10: - w3[3] = hc_bytealign_S (w1[0], w1[1], offset); - w3[2] = hc_bytealign_S (w0[3], w1[0], offset); - w3[1] = hc_bytealign_S (w0[2], w0[3], offset); - w3[0] = hc_bytealign_S (w0[1], w0[2], offset); - w2[3] = hc_bytealign_S (w0[0], w0[1], offset); - w2[2] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be_S ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -33668,11 +33422,11 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 11: - w3[3] = hc_bytealign_S (w0[3], w1[0], offset); - w3[2] = hc_bytealign_S (w0[2], w0[3], offset); - w3[1] = hc_bytealign_S (w0[1], w0[2], offset); - w3[0] = hc_bytealign_S (w0[0], w0[1], offset); - w2[3] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be_S ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -33688,10 +33442,10 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 12: - w3[3] = hc_bytealign_S (w0[2], w0[3], offset); - w3[2] = hc_bytealign_S (w0[1], w0[2], offset); - w3[1] = hc_bytealign_S (w0[0], w0[1], offset); - w3[0] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be_S ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -33708,9 +33462,9 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 13: - w3[3] = hc_bytealign_S (w0[1], w0[2], offset); - w3[2] = hc_bytealign_S (w0[0], w0[1], offset); - w3[1] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be_S ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -33728,8 +33482,8 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 14: - w3[3] = hc_bytealign_S (w0[0], w0[1], offset); - w3[2] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be_S ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -33748,7 +33502,7 @@ DECLSPEC void switch_buffer_by_offset_be_S (u32 *w0, u32 *w1, u32 *w2, u32 *w3, break; case 15: - w3[3] = hc_bytealign_S ( 0, w0[0], offset); + w3[3] = hc_bytealign_be_S ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -34112,89 +33866,89 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 switch (offset_switch) { case 0: - c0[0] = hc_bytealign_S (w3[3], 0, offset); - w3[3] = hc_bytealign_S (w3[2], w3[3], offset); - w3[2] = hc_bytealign_S (w3[1], w3[2], offset); - w3[1] = hc_bytealign_S (w3[0], w3[1], offset); - w3[0] = hc_bytealign_S (w2[3], w3[0], offset); - w2[3] = hc_bytealign_S (w2[2], w2[3], offset); - w2[2] = hc_bytealign_S (w2[1], w2[2], offset); - w2[1] = hc_bytealign_S (w2[0], w2[1], offset); - w2[0] = hc_bytealign_S (w1[3], w2[0], offset); - w1[3] = hc_bytealign_S (w1[2], w1[3], offset); - w1[2] = hc_bytealign_S (w1[1], w1[2], offset); - w1[1] = hc_bytealign_S (w1[0], w1[1], offset); - w1[0] = hc_bytealign_S (w0[3], w1[0], offset); - w0[3] = hc_bytealign_S (w0[2], w0[3], offset); - w0[2] = hc_bytealign_S (w0[1], w0[2], offset); - w0[1] = hc_bytealign_S (w0[0], w0[1], offset); - w0[0] = hc_bytealign_S ( 0, w0[0], offset); + c0[0] = hc_bytealign_be_S (w3[3], 0, offset); + w3[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be_S ( 0, w0[0], offset); break; case 1: - c0[1] = hc_bytealign_S (w3[3], 0, offset); - c0[0] = hc_bytealign_S (w3[2], w3[3], offset); - w3[3] = hc_bytealign_S (w3[1], w3[2], offset); - w3[2] = hc_bytealign_S (w3[0], w3[1], offset); - w3[1] = hc_bytealign_S (w2[3], w3[0], offset); - w3[0] = hc_bytealign_S (w2[2], w2[3], offset); - w2[3] = hc_bytealign_S (w2[1], w2[2], offset); - w2[2] = hc_bytealign_S (w2[0], w2[1], offset); - w2[1] = hc_bytealign_S (w1[3], w2[0], offset); - w2[0] = hc_bytealign_S (w1[2], w1[3], offset); - w1[3] = hc_bytealign_S (w1[1], w1[2], offset); - w1[2] = hc_bytealign_S (w1[0], w1[1], offset); - w1[1] = hc_bytealign_S (w0[3], w1[0], offset); - w1[0] = hc_bytealign_S (w0[2], w0[3], offset); - w0[3] = hc_bytealign_S (w0[1], w0[2], offset); - w0[2] = hc_bytealign_S (w0[0], w0[1], offset); - w0[1] = hc_bytealign_S ( 0, w0[0], offset); + c0[1] = hc_bytealign_be_S (w3[3], 0, offset); + c0[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be_S ( 0, w0[0], offset); w0[0] = 0; break; case 2: - c0[2] = hc_bytealign_S (w3[3], 0, offset); - c0[1] = hc_bytealign_S (w3[2], w3[3], offset); - c0[0] = hc_bytealign_S (w3[1], w3[2], offset); - w3[3] = hc_bytealign_S (w3[0], w3[1], offset); - w3[2] = hc_bytealign_S (w2[3], w3[0], offset); - w3[1] = hc_bytealign_S (w2[2], w2[3], offset); - w3[0] = hc_bytealign_S (w2[1], w2[2], offset); - w2[3] = hc_bytealign_S (w2[0], w2[1], offset); - w2[2] = hc_bytealign_S (w1[3], w2[0], offset); - w2[1] = hc_bytealign_S (w1[2], w1[3], offset); - w2[0] = hc_bytealign_S (w1[1], w1[2], offset); - w1[3] = hc_bytealign_S (w1[0], w1[1], offset); - w1[2] = hc_bytealign_S (w0[3], w1[0], offset); - w1[1] = hc_bytealign_S (w0[2], w0[3], offset); - w1[0] = hc_bytealign_S (w0[1], w0[2], offset); - w0[3] = hc_bytealign_S (w0[0], w0[1], offset); - w0[2] = hc_bytealign_S ( 0, w0[0], offset); + c0[2] = hc_bytealign_be_S (w3[3], 0, offset); + c0[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be_S ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - c0[3] = hc_bytealign_S (w3[3], 0, offset); - c0[2] = hc_bytealign_S (w3[2], w3[3], offset); - c0[1] = hc_bytealign_S (w3[1], w3[2], offset); - c0[0] = hc_bytealign_S (w3[0], w3[1], offset); - w3[3] = hc_bytealign_S (w2[3], w3[0], offset); - w3[2] = hc_bytealign_S (w2[2], w2[3], offset); - w3[1] = hc_bytealign_S (w2[1], w2[2], offset); - w3[0] = hc_bytealign_S (w2[0], w2[1], offset); - w2[3] = hc_bytealign_S (w1[3], w2[0], offset); - w2[2] = hc_bytealign_S (w1[2], w1[3], offset); - w2[1] = hc_bytealign_S (w1[1], w1[2], offset); - w2[0] = hc_bytealign_S (w1[0], w1[1], offset); - w1[3] = hc_bytealign_S (w0[3], w1[0], offset); - w1[2] = hc_bytealign_S (w0[2], w0[3], offset); - w1[1] = hc_bytealign_S (w0[1], w0[2], offset); - w1[0] = hc_bytealign_S (w0[0], w0[1], offset); - w0[3] = hc_bytealign_S ( 0, w0[0], offset); + c0[3] = hc_bytealign_be_S (w3[3], 0, offset); + c0[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be_S ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -34202,23 +33956,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 4: - c1[0] = hc_bytealign_S (w3[3], 0, offset); - c0[3] = hc_bytealign_S (w3[2], w3[3], offset); - c0[2] = hc_bytealign_S (w3[1], w3[2], offset); - c0[1] = hc_bytealign_S (w3[0], w3[1], offset); - c0[0] = hc_bytealign_S (w2[3], w3[0], offset); - w3[3] = hc_bytealign_S (w2[2], w2[3], offset); - w3[2] = hc_bytealign_S (w2[1], w2[2], offset); - w3[1] = hc_bytealign_S (w2[0], w2[1], offset); - w3[0] = hc_bytealign_S (w1[3], w2[0], offset); - w2[3] = hc_bytealign_S (w1[2], w1[3], offset); - w2[2] = hc_bytealign_S (w1[1], w1[2], offset); - w2[1] = hc_bytealign_S (w1[0], w1[1], offset); - w2[0] = hc_bytealign_S (w0[3], w1[0], offset); - w1[3] = hc_bytealign_S (w0[2], w0[3], offset); - w1[2] = hc_bytealign_S (w0[1], w0[2], offset); - w1[1] = hc_bytealign_S (w0[0], w0[1], offset); - w1[0] = hc_bytealign_S ( 0, w0[0], offset); + c1[0] = hc_bytealign_be_S (w3[3], 0, offset); + c0[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be_S ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -34227,23 +33981,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 5: - c1[1] = hc_bytealign_S (w3[3], 0, offset); - c1[0] = hc_bytealign_S (w3[2], w3[3], offset); - c0[3] = hc_bytealign_S (w3[1], w3[2], offset); - c0[2] = hc_bytealign_S (w3[0], w3[1], offset); - c0[1] = hc_bytealign_S (w2[3], w3[0], offset); - c0[0] = hc_bytealign_S (w2[2], w2[3], offset); - w3[3] = hc_bytealign_S (w2[1], w2[2], offset); - w3[2] = hc_bytealign_S (w2[0], w2[1], offset); - w3[1] = hc_bytealign_S (w1[3], w2[0], offset); - w3[0] = hc_bytealign_S (w1[2], w1[3], offset); - w2[3] = hc_bytealign_S (w1[1], w1[2], offset); - w2[2] = hc_bytealign_S (w1[0], w1[1], offset); - w2[1] = hc_bytealign_S (w0[3], w1[0], offset); - w2[0] = hc_bytealign_S (w0[2], w0[3], offset); - w1[3] = hc_bytealign_S (w0[1], w0[2], offset); - w1[2] = hc_bytealign_S (w0[0], w0[1], offset); - w1[1] = hc_bytealign_S ( 0, w0[0], offset); + c1[1] = hc_bytealign_be_S (w3[3], 0, offset); + c1[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be_S ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -34253,23 +34007,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 6: - c1[2] = hc_bytealign_S (w3[3], 0, offset); - c1[1] = hc_bytealign_S (w3[2], w3[3], offset); - c1[0] = hc_bytealign_S (w3[1], w3[2], offset); - c0[3] = hc_bytealign_S (w3[0], w3[1], offset); - c0[2] = hc_bytealign_S (w2[3], w3[0], offset); - c0[1] = hc_bytealign_S (w2[2], w2[3], offset); - c0[0] = hc_bytealign_S (w2[1], w2[2], offset); - w3[3] = hc_bytealign_S (w2[0], w2[1], offset); - w3[2] = hc_bytealign_S (w1[3], w2[0], offset); - w3[1] = hc_bytealign_S (w1[2], w1[3], offset); - w3[0] = hc_bytealign_S (w1[1], w1[2], offset); - w2[3] = hc_bytealign_S (w1[0], w1[1], offset); - w2[2] = hc_bytealign_S (w0[3], w1[0], offset); - w2[1] = hc_bytealign_S (w0[2], w0[3], offset); - w2[0] = hc_bytealign_S (w0[1], w0[2], offset); - w1[3] = hc_bytealign_S (w0[0], w0[1], offset); - w1[2] = hc_bytealign_S ( 0, w0[0], offset); + c1[2] = hc_bytealign_be_S (w3[3], 0, offset); + c1[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be_S ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -34280,23 +34034,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 7: - c1[3] = hc_bytealign_S (w3[3], 0, offset); - c1[2] = hc_bytealign_S (w3[2], w3[3], offset); - c1[1] = hc_bytealign_S (w3[1], w3[2], offset); - c1[0] = hc_bytealign_S (w3[0], w3[1], offset); - c0[3] = hc_bytealign_S (w2[3], w3[0], offset); - c0[2] = hc_bytealign_S (w2[2], w2[3], offset); - c0[1] = hc_bytealign_S (w2[1], w2[2], offset); - c0[0] = hc_bytealign_S (w2[0], w2[1], offset); - w3[3] = hc_bytealign_S (w1[3], w2[0], offset); - w3[2] = hc_bytealign_S (w1[2], w1[3], offset); - w3[1] = hc_bytealign_S (w1[1], w1[2], offset); - w3[0] = hc_bytealign_S (w1[0], w1[1], offset); - w2[3] = hc_bytealign_S (w0[3], w1[0], offset); - w2[2] = hc_bytealign_S (w0[2], w0[3], offset); - w2[1] = hc_bytealign_S (w0[1], w0[2], offset); - w2[0] = hc_bytealign_S (w0[0], w0[1], offset); - w1[3] = hc_bytealign_S ( 0, w0[0], offset); + c1[3] = hc_bytealign_be_S (w3[3], 0, offset); + c1[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be_S ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -34308,23 +34062,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 8: - c2[0] = hc_bytealign_S (w3[3], 0, offset); - c1[3] = hc_bytealign_S (w3[2], w3[3], offset); - c1[2] = hc_bytealign_S (w3[1], w3[2], offset); - c1[1] = hc_bytealign_S (w3[0], w3[1], offset); - c1[0] = hc_bytealign_S (w2[3], w3[0], offset); - c0[3] = hc_bytealign_S (w2[2], w2[3], offset); - c0[2] = hc_bytealign_S (w2[1], w2[2], offset); - c0[1] = hc_bytealign_S (w2[0], w2[1], offset); - c0[0] = hc_bytealign_S (w1[3], w2[0], offset); - w3[3] = hc_bytealign_S (w1[2], w1[3], offset); - w3[2] = hc_bytealign_S (w1[1], w1[2], offset); - w3[1] = hc_bytealign_S (w1[0], w1[1], offset); - w3[0] = hc_bytealign_S (w0[3], w1[0], offset); - w2[3] = hc_bytealign_S (w0[2], w0[3], offset); - w2[2] = hc_bytealign_S (w0[1], w0[2], offset); - w2[1] = hc_bytealign_S (w0[0], w0[1], offset); - w2[0] = hc_bytealign_S ( 0, w0[0], offset); + c2[0] = hc_bytealign_be_S (w3[3], 0, offset); + c1[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be_S ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -34337,23 +34091,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 9: - c2[1] = hc_bytealign_S (w3[3], 0, offset); - c2[0] = hc_bytealign_S (w3[2], w3[3], offset); - c1[3] = hc_bytealign_S (w3[1], w3[2], offset); - c1[2] = hc_bytealign_S (w3[0], w3[1], offset); - c1[1] = hc_bytealign_S (w2[3], w3[0], offset); - c1[0] = hc_bytealign_S (w2[2], w2[3], offset); - c0[3] = hc_bytealign_S (w2[1], w2[2], offset); - c0[2] = hc_bytealign_S (w2[0], w2[1], offset); - c0[1] = hc_bytealign_S (w1[3], w2[0], offset); - c0[0] = hc_bytealign_S (w1[2], w1[3], offset); - w3[3] = hc_bytealign_S (w1[1], w1[2], offset); - w3[2] = hc_bytealign_S (w1[0], w1[1], offset); - w3[1] = hc_bytealign_S (w0[3], w1[0], offset); - w3[0] = hc_bytealign_S (w0[2], w0[3], offset); - w2[3] = hc_bytealign_S (w0[1], w0[2], offset); - w2[2] = hc_bytealign_S (w0[0], w0[1], offset); - w2[1] = hc_bytealign_S ( 0, w0[0], offset); + c2[1] = hc_bytealign_be_S (w3[3], 0, offset); + c2[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be_S ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -34367,23 +34121,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 10: - c2[2] = hc_bytealign_S (w3[3], 0, offset); - c2[1] = hc_bytealign_S (w3[2], w3[3], offset); - c2[0] = hc_bytealign_S (w3[1], w3[2], offset); - c1[3] = hc_bytealign_S (w3[0], w3[1], offset); - c1[2] = hc_bytealign_S (w2[3], w3[0], offset); - c1[1] = hc_bytealign_S (w2[2], w2[3], offset); - c1[0] = hc_bytealign_S (w2[1], w2[2], offset); - c0[3] = hc_bytealign_S (w2[0], w2[1], offset); - c0[2] = hc_bytealign_S (w1[3], w2[0], offset); - c0[1] = hc_bytealign_S (w1[2], w1[3], offset); - c0[0] = hc_bytealign_S (w1[1], w1[2], offset); - w3[3] = hc_bytealign_S (w1[0], w1[1], offset); - w3[2] = hc_bytealign_S (w0[3], w1[0], offset); - w3[1] = hc_bytealign_S (w0[2], w0[3], offset); - w3[0] = hc_bytealign_S (w0[1], w0[2], offset); - w2[3] = hc_bytealign_S (w0[0], w0[1], offset); - w2[2] = hc_bytealign_S ( 0, w0[0], offset); + c2[2] = hc_bytealign_be_S (w3[3], 0, offset); + c2[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be_S ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -34398,23 +34152,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 11: - c2[3] = hc_bytealign_S (w3[3], 0, offset); - c2[2] = hc_bytealign_S (w3[2], w3[3], offset); - c2[1] = hc_bytealign_S (w3[1], w3[2], offset); - c2[0] = hc_bytealign_S (w3[0], w3[1], offset); - c1[3] = hc_bytealign_S (w2[3], w3[0], offset); - c1[2] = hc_bytealign_S (w2[2], w2[3], offset); - c1[1] = hc_bytealign_S (w2[1], w2[2], offset); - c1[0] = hc_bytealign_S (w2[0], w2[1], offset); - c0[3] = hc_bytealign_S (w1[3], w2[0], offset); - c0[2] = hc_bytealign_S (w1[2], w1[3], offset); - c0[1] = hc_bytealign_S (w1[1], w1[2], offset); - c0[0] = hc_bytealign_S (w1[0], w1[1], offset); - w3[3] = hc_bytealign_S (w0[3], w1[0], offset); - w3[2] = hc_bytealign_S (w0[2], w0[3], offset); - w3[1] = hc_bytealign_S (w0[1], w0[2], offset); - w3[0] = hc_bytealign_S (w0[0], w0[1], offset); - w2[3] = hc_bytealign_S ( 0, w0[0], offset); + c2[3] = hc_bytealign_be_S (w3[3], 0, offset); + c2[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be_S ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -34430,23 +34184,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 12: - c3[0] = hc_bytealign_S (w3[3], 0, offset); - c2[3] = hc_bytealign_S (w3[2], w3[3], offset); - c2[2] = hc_bytealign_S (w3[1], w3[2], offset); - c2[1] = hc_bytealign_S (w3[0], w3[1], offset); - c2[0] = hc_bytealign_S (w2[3], w3[0], offset); - c1[3] = hc_bytealign_S (w2[2], w2[3], offset); - c1[2] = hc_bytealign_S (w2[1], w2[2], offset); - c1[1] = hc_bytealign_S (w2[0], w2[1], offset); - c1[0] = hc_bytealign_S (w1[3], w2[0], offset); - c0[3] = hc_bytealign_S (w1[2], w1[3], offset); - c0[2] = hc_bytealign_S (w1[1], w1[2], offset); - c0[1] = hc_bytealign_S (w1[0], w1[1], offset); - c0[0] = hc_bytealign_S (w0[3], w1[0], offset); - w3[3] = hc_bytealign_S (w0[2], w0[3], offset); - w3[2] = hc_bytealign_S (w0[1], w0[2], offset); - w3[1] = hc_bytealign_S (w0[0], w0[1], offset); - w3[0] = hc_bytealign_S ( 0, w0[0], offset); + c3[0] = hc_bytealign_be_S (w3[3], 0, offset); + c2[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be_S ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -34463,23 +34217,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 13: - c3[1] = hc_bytealign_S (w3[3], 0, offset); - c3[0] = hc_bytealign_S (w3[2], w3[3], offset); - c2[3] = hc_bytealign_S (w3[1], w3[2], offset); - c2[2] = hc_bytealign_S (w3[0], w3[1], offset); - c2[1] = hc_bytealign_S (w2[3], w3[0], offset); - c2[0] = hc_bytealign_S (w2[2], w2[3], offset); - c1[3] = hc_bytealign_S (w2[1], w2[2], offset); - c1[2] = hc_bytealign_S (w2[0], w2[1], offset); - c1[1] = hc_bytealign_S (w1[3], w2[0], offset); - c1[0] = hc_bytealign_S (w1[2], w1[3], offset); - c0[3] = hc_bytealign_S (w1[1], w1[2], offset); - c0[2] = hc_bytealign_S (w1[0], w1[1], offset); - c0[1] = hc_bytealign_S (w0[3], w1[0], offset); - c0[0] = hc_bytealign_S (w0[2], w0[3], offset); - w3[3] = hc_bytealign_S (w0[1], w0[2], offset); - w3[2] = hc_bytealign_S (w0[0], w0[1], offset); - w3[1] = hc_bytealign_S ( 0, w0[0], offset); + c3[1] = hc_bytealign_be_S (w3[3], 0, offset); + c3[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be_S ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -34497,23 +34251,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 14: - c3[2] = hc_bytealign_S (w3[3], 0, offset); - c3[1] = hc_bytealign_S (w3[2], w3[3], offset); - c3[0] = hc_bytealign_S (w3[1], w3[2], offset); - c2[3] = hc_bytealign_S (w3[0], w3[1], offset); - c2[2] = hc_bytealign_S (w2[3], w3[0], offset); - c2[1] = hc_bytealign_S (w2[2], w2[3], offset); - c2[0] = hc_bytealign_S (w2[1], w2[2], offset); - c1[3] = hc_bytealign_S (w2[0], w2[1], offset); - c1[2] = hc_bytealign_S (w1[3], w2[0], offset); - c1[1] = hc_bytealign_S (w1[2], w1[3], offset); - c1[0] = hc_bytealign_S (w1[1], w1[2], offset); - c0[3] = hc_bytealign_S (w1[0], w1[1], offset); - c0[2] = hc_bytealign_S (w0[3], w1[0], offset); - c0[1] = hc_bytealign_S (w0[2], w0[3], offset); - c0[0] = hc_bytealign_S (w0[1], w0[2], offset); - w3[3] = hc_bytealign_S (w0[0], w0[1], offset); - w3[2] = hc_bytealign_S ( 0, w0[0], offset); + c3[2] = hc_bytealign_be_S (w3[3], 0, offset); + c3[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c3[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c2[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + c1[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + c0[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be_S ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -34532,23 +34286,23 @@ DECLSPEC void switch_buffer_by_offset_carry_be_S (u32 *w0, u32 *w1, u32 *w2, u32 break; case 15: - c3[3] = hc_bytealign_S (w3[3], 0, offset); - c3[2] = hc_bytealign_S (w3[2], w3[3], offset); - c3[1] = hc_bytealign_S (w3[1], w3[2], offset); - c3[0] = hc_bytealign_S (w3[0], w3[1], offset); - c2[3] = hc_bytealign_S (w2[3], w3[0], offset); - c2[2] = hc_bytealign_S (w2[2], w2[3], offset); - c2[1] = hc_bytealign_S (w2[1], w2[2], offset); - c2[0] = hc_bytealign_S (w2[0], w2[1], offset); - c1[3] = hc_bytealign_S (w1[3], w2[0], offset); - c1[2] = hc_bytealign_S (w1[2], w1[3], offset); - c1[1] = hc_bytealign_S (w1[1], w1[2], offset); - c1[0] = hc_bytealign_S (w1[0], w1[1], offset); - c0[3] = hc_bytealign_S (w0[3], w1[0], offset); - c0[2] = hc_bytealign_S (w0[2], w0[3], offset); - c0[1] = hc_bytealign_S (w0[1], w0[2], offset); - c0[0] = hc_bytealign_S (w0[0], w0[1], offset); - w3[3] = hc_bytealign_S ( 0, w0[0], offset); + c3[3] = hc_bytealign_be_S (w3[3], 0, offset); + c3[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c3[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c3[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c2[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c2[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + c1[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + c1[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + c0[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + c0[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be_S ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -35049,39 +34803,6 @@ DECLSPEC void switch_buffer_by_offset_8x4_le_S (u32 *w0, u32 *w1, u32 *w2, u32 * const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); - w4[0] = swap32_S (w4[0]); - w4[1] = swap32_S (w4[1]); - w4[2] = swap32_S (w4[2]); - w4[3] = swap32_S (w4[3]); - w5[0] = swap32_S (w5[0]); - w5[1] = swap32_S (w5[1]); - w5[2] = swap32_S (w5[2]); - w5[3] = swap32_S (w5[3]); - w6[0] = swap32_S (w6[0]); - w6[1] = swap32_S (w6[1]); - w6[2] = swap32_S (w6[2]); - w6[3] = swap32_S (w6[3]); - w7[0] = swap32_S (w7[0]); - w7[1] = swap32_S (w7[1]); - w7[2] = swap32_S (w7[2]); - w7[3] = swap32_S (w7[3]); - switch (offset_switch) { case 0: @@ -36236,39 +35957,6 @@ DECLSPEC void switch_buffer_by_offset_8x4_le_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; } - - w0[0] = swap32_S (w0[0]); - w0[1] = swap32_S (w0[1]); - w0[2] = swap32_S (w0[2]); - w0[3] = swap32_S (w0[3]); - w1[0] = swap32_S (w1[0]); - w1[1] = swap32_S (w1[1]); - w1[2] = swap32_S (w1[2]); - w1[3] = swap32_S (w1[3]); - w2[0] = swap32_S (w2[0]); - w2[1] = swap32_S (w2[1]); - w2[2] = swap32_S (w2[2]); - w2[3] = swap32_S (w2[3]); - w3[0] = swap32_S (w3[0]); - w3[1] = swap32_S (w3[1]); - w3[2] = swap32_S (w3[2]); - w3[3] = swap32_S (w3[3]); - w4[0] = swap32_S (w4[0]); - w4[1] = swap32_S (w4[1]); - w4[2] = swap32_S (w4[2]); - w4[3] = swap32_S (w4[3]); - w5[0] = swap32_S (w5[0]); - w5[1] = swap32_S (w5[1]); - w5[2] = swap32_S (w5[2]); - w5[3] = swap32_S (w5[3]); - w6[0] = swap32_S (w6[0]); - w6[1] = swap32_S (w6[1]); - w6[2] = swap32_S (w6[2]); - w6[3] = swap32_S (w6[3]); - w7[0] = swap32_S (w7[0]); - w7[1] = swap32_S (w7[1]); - w7[2] = swap32_S (w7[2]); - w7[3] = swap32_S (w7[3]); #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -36854,143 +36542,143 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * switch (offset_switch) { case 0: - w7[3] = hc_bytealign_S (w7[2], w7[3], offset); - w7[2] = hc_bytealign_S (w7[1], w7[2], offset); - w7[1] = hc_bytealign_S (w7[0], w7[1], offset); - w7[0] = hc_bytealign_S (w6[3], w7[0], offset); - w6[3] = hc_bytealign_S (w6[2], w6[3], offset); - w6[2] = hc_bytealign_S (w6[1], w6[2], offset); - w6[1] = hc_bytealign_S (w6[0], w6[1], offset); - w6[0] = hc_bytealign_S (w5[3], w6[0], offset); - w5[3] = hc_bytealign_S (w5[2], w5[3], offset); - w5[2] = hc_bytealign_S (w5[1], w5[2], offset); - w5[1] = hc_bytealign_S (w5[0], w5[1], offset); - w5[0] = hc_bytealign_S (w4[3], w5[0], offset); - w4[3] = hc_bytealign_S (w4[2], w4[3], offset); - w4[2] = hc_bytealign_S (w4[1], w4[2], offset); - w4[1] = hc_bytealign_S (w4[0], w4[1], offset); - w4[0] = hc_bytealign_S (w3[3], w4[0], offset); - w3[3] = hc_bytealign_S (w3[2], w3[3], offset); - w3[2] = hc_bytealign_S (w3[1], w3[2], offset); - w3[1] = hc_bytealign_S (w3[0], w3[1], offset); - w3[0] = hc_bytealign_S (w2[3], w3[0], offset); - w2[3] = hc_bytealign_S (w2[2], w2[3], offset); - w2[2] = hc_bytealign_S (w2[1], w2[2], offset); - w2[1] = hc_bytealign_S (w2[0], w2[1], offset); - w2[0] = hc_bytealign_S (w1[3], w2[0], offset); - w1[3] = hc_bytealign_S (w1[2], w1[3], offset); - w1[2] = hc_bytealign_S (w1[1], w1[2], offset); - w1[1] = hc_bytealign_S (w1[0], w1[1], offset); - w1[0] = hc_bytealign_S (w0[3], w1[0], offset); - w0[3] = hc_bytealign_S (w0[2], w0[3], offset); - w0[2] = hc_bytealign_S (w0[1], w0[2], offset); - w0[1] = hc_bytealign_S (w0[0], w0[1], offset); - w0[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + w7[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + w7[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + w6[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + w6[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + w5[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w5[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w4[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w4[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w3[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be_S ( 0, w0[0], offset); break; case 1: - w7[3] = hc_bytealign_S (w7[1], w7[2], offset); - w7[2] = hc_bytealign_S (w7[0], w7[1], offset); - w7[1] = hc_bytealign_S (w6[3], w7[0], offset); - w7[0] = hc_bytealign_S (w6[2], w6[3], offset); - w6[3] = hc_bytealign_S (w6[1], w6[2], offset); - w6[2] = hc_bytealign_S (w6[0], w6[1], offset); - w6[1] = hc_bytealign_S (w5[3], w6[0], offset); - w6[0] = hc_bytealign_S (w5[2], w5[3], offset); - w5[3] = hc_bytealign_S (w5[1], w5[2], offset); - w5[2] = hc_bytealign_S (w5[0], w5[1], offset); - w5[1] = hc_bytealign_S (w4[3], w5[0], offset); - w5[0] = hc_bytealign_S (w4[2], w4[3], offset); - w4[3] = hc_bytealign_S (w4[1], w4[2], offset); - w4[2] = hc_bytealign_S (w4[0], w4[1], offset); - w4[1] = hc_bytealign_S (w3[3], w4[0], offset); - w4[0] = hc_bytealign_S (w3[2], w3[3], offset); - w3[3] = hc_bytealign_S (w3[1], w3[2], offset); - w3[2] = hc_bytealign_S (w3[0], w3[1], offset); - w3[1] = hc_bytealign_S (w2[3], w3[0], offset); - w3[0] = hc_bytealign_S (w2[2], w2[3], offset); - w2[3] = hc_bytealign_S (w2[1], w2[2], offset); - w2[2] = hc_bytealign_S (w2[0], w2[1], offset); - w2[1] = hc_bytealign_S (w1[3], w2[0], offset); - w2[0] = hc_bytealign_S (w1[2], w1[3], offset); - w1[3] = hc_bytealign_S (w1[1], w1[2], offset); - w1[2] = hc_bytealign_S (w1[0], w1[1], offset); - w1[1] = hc_bytealign_S (w0[3], w1[0], offset); - w1[0] = hc_bytealign_S (w0[2], w0[3], offset); - w0[3] = hc_bytealign_S (w0[1], w0[2], offset); - w0[2] = hc_bytealign_S (w0[0], w0[1], offset); - w0[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + w7[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + w6[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + w5[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w4[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be_S ( 0, w0[0], offset); w0[0] = 0; break; case 2: - w7[3] = hc_bytealign_S (w7[0], w7[1], offset); - w7[2] = hc_bytealign_S (w6[3], w7[0], offset); - w7[1] = hc_bytealign_S (w6[2], w6[3], offset); - w7[0] = hc_bytealign_S (w6[1], w6[2], offset); - w6[3] = hc_bytealign_S (w6[0], w6[1], offset); - w6[2] = hc_bytealign_S (w5[3], w6[0], offset); - w6[1] = hc_bytealign_S (w5[2], w5[3], offset); - w6[0] = hc_bytealign_S (w5[1], w5[2], offset); - w5[3] = hc_bytealign_S (w5[0], w5[1], offset); - w5[2] = hc_bytealign_S (w4[3], w5[0], offset); - w5[1] = hc_bytealign_S (w4[2], w4[3], offset); - w5[0] = hc_bytealign_S (w4[1], w4[2], offset); - w4[3] = hc_bytealign_S (w4[0], w4[1], offset); - w4[2] = hc_bytealign_S (w3[3], w4[0], offset); - w4[1] = hc_bytealign_S (w3[2], w3[3], offset); - w4[0] = hc_bytealign_S (w3[1], w3[2], offset); - w3[3] = hc_bytealign_S (w3[0], w3[1], offset); - w3[2] = hc_bytealign_S (w2[3], w3[0], offset); - w3[1] = hc_bytealign_S (w2[2], w2[3], offset); - w3[0] = hc_bytealign_S (w2[1], w2[2], offset); - w2[3] = hc_bytealign_S (w2[0], w2[1], offset); - w2[2] = hc_bytealign_S (w1[3], w2[0], offset); - w2[1] = hc_bytealign_S (w1[2], w1[3], offset); - w2[0] = hc_bytealign_S (w1[1], w1[2], offset); - w1[3] = hc_bytealign_S (w1[0], w1[1], offset); - w1[2] = hc_bytealign_S (w0[3], w1[0], offset); - w1[1] = hc_bytealign_S (w0[2], w0[3], offset); - w1[0] = hc_bytealign_S (w0[1], w0[2], offset); - w0[3] = hc_bytealign_S (w0[0], w0[1], offset); - w0[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be_S ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - w7[3] = hc_bytealign_S (w6[3], w7[0], offset); - w7[2] = hc_bytealign_S (w6[2], w6[3], offset); - w7[1] = hc_bytealign_S (w6[1], w6[2], offset); - w7[0] = hc_bytealign_S (w6[0], w6[1], offset); - w6[3] = hc_bytealign_S (w5[3], w6[0], offset); - w6[2] = hc_bytealign_S (w5[2], w5[3], offset); - w6[1] = hc_bytealign_S (w5[1], w5[2], offset); - w6[0] = hc_bytealign_S (w5[0], w5[1], offset); - w5[3] = hc_bytealign_S (w4[3], w5[0], offset); - w5[2] = hc_bytealign_S (w4[2], w4[3], offset); - w5[1] = hc_bytealign_S (w4[1], w4[2], offset); - w5[0] = hc_bytealign_S (w4[0], w4[1], offset); - w4[3] = hc_bytealign_S (w3[3], w4[0], offset); - w4[2] = hc_bytealign_S (w3[2], w3[3], offset); - w4[1] = hc_bytealign_S (w3[1], w3[2], offset); - w4[0] = hc_bytealign_S (w3[0], w3[1], offset); - w3[3] = hc_bytealign_S (w2[3], w3[0], offset); - w3[2] = hc_bytealign_S (w2[2], w2[3], offset); - w3[1] = hc_bytealign_S (w2[1], w2[2], offset); - w3[0] = hc_bytealign_S (w2[0], w2[1], offset); - w2[3] = hc_bytealign_S (w1[3], w2[0], offset); - w2[2] = hc_bytealign_S (w1[2], w1[3], offset); - w2[1] = hc_bytealign_S (w1[1], w1[2], offset); - w2[0] = hc_bytealign_S (w1[0], w1[1], offset); - w1[3] = hc_bytealign_S (w0[3], w1[0], offset); - w1[2] = hc_bytealign_S (w0[2], w0[3], offset); - w1[1] = hc_bytealign_S (w0[1], w0[2], offset); - w1[0] = hc_bytealign_S (w0[0], w0[1], offset); - w0[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be_S ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -36998,34 +36686,34 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 4: - w7[3] = hc_bytealign_S (w6[2], w6[3], offset); - w7[2] = hc_bytealign_S (w6[1], w6[2], offset); - w7[1] = hc_bytealign_S (w6[0], w6[1], offset); - w7[0] = hc_bytealign_S (w5[3], w6[0], offset); - w6[3] = hc_bytealign_S (w5[2], w5[3], offset); - w6[2] = hc_bytealign_S (w5[1], w5[2], offset); - w6[1] = hc_bytealign_S (w5[0], w5[1], offset); - w6[0] = hc_bytealign_S (w4[3], w5[0], offset); - w5[3] = hc_bytealign_S (w4[2], w4[3], offset); - w5[2] = hc_bytealign_S (w4[1], w4[2], offset); - w5[1] = hc_bytealign_S (w4[0], w4[1], offset); - w5[0] = hc_bytealign_S (w3[3], w4[0], offset); - w4[3] = hc_bytealign_S (w3[2], w3[3], offset); - w4[2] = hc_bytealign_S (w3[1], w3[2], offset); - w4[1] = hc_bytealign_S (w3[0], w3[1], offset); - w4[0] = hc_bytealign_S (w2[3], w3[0], offset); - w3[3] = hc_bytealign_S (w2[2], w2[3], offset); - w3[2] = hc_bytealign_S (w2[1], w2[2], offset); - w3[1] = hc_bytealign_S (w2[0], w2[1], offset); - w3[0] = hc_bytealign_S (w1[3], w2[0], offset); - w2[3] = hc_bytealign_S (w1[2], w1[3], offset); - w2[2] = hc_bytealign_S (w1[1], w1[2], offset); - w2[1] = hc_bytealign_S (w1[0], w1[1], offset); - w2[0] = hc_bytealign_S (w0[3], w1[0], offset); - w1[3] = hc_bytealign_S (w0[2], w0[3], offset); - w1[2] = hc_bytealign_S (w0[1], w0[2], offset); - w1[1] = hc_bytealign_S (w0[0], w0[1], offset); - w1[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be_S ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -37034,33 +36722,33 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 5: - w7[3] = hc_bytealign_S (w6[1], w6[2], offset); - w7[2] = hc_bytealign_S (w6[0], w6[1], offset); - w7[1] = hc_bytealign_S (w5[3], w6[0], offset); - w7[0] = hc_bytealign_S (w5[2], w5[3], offset); - w6[3] = hc_bytealign_S (w5[1], w5[2], offset); - w6[2] = hc_bytealign_S (w5[0], w5[1], offset); - w6[1] = hc_bytealign_S (w4[3], w5[0], offset); - w6[0] = hc_bytealign_S (w4[2], w4[3], offset); - w5[3] = hc_bytealign_S (w4[1], w4[2], offset); - w5[2] = hc_bytealign_S (w4[0], w4[1], offset); - w5[1] = hc_bytealign_S (w3[3], w4[0], offset); - w5[0] = hc_bytealign_S (w3[2], w3[3], offset); - w4[3] = hc_bytealign_S (w3[1], w3[2], offset); - w4[2] = hc_bytealign_S (w3[0], w3[1], offset); - w4[1] = hc_bytealign_S (w2[3], w3[0], offset); - w4[0] = hc_bytealign_S (w2[2], w2[3], offset); - w3[3] = hc_bytealign_S (w2[1], w2[2], offset); - w3[2] = hc_bytealign_S (w2[0], w2[1], offset); - w3[1] = hc_bytealign_S (w1[3], w2[0], offset); - w3[0] = hc_bytealign_S (w1[2], w1[3], offset); - w2[3] = hc_bytealign_S (w1[1], w1[2], offset); - w2[2] = hc_bytealign_S (w1[0], w1[1], offset); - w2[1] = hc_bytealign_S (w0[3], w1[0], offset); - w2[0] = hc_bytealign_S (w0[2], w0[3], offset); - w1[3] = hc_bytealign_S (w0[1], w0[2], offset); - w1[2] = hc_bytealign_S (w0[0], w0[1], offset); - w1[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be_S ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -37070,32 +36758,32 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 6: - w7[3] = hc_bytealign_S (w6[0], w6[1], offset); - w7[2] = hc_bytealign_S (w5[3], w6[0], offset); - w7[1] = hc_bytealign_S (w5[2], w5[3], offset); - w7[0] = hc_bytealign_S (w5[1], w5[2], offset); - w6[3] = hc_bytealign_S (w5[0], w5[1], offset); - w6[2] = hc_bytealign_S (w4[3], w5[0], offset); - w6[1] = hc_bytealign_S (w4[2], w4[3], offset); - w6[0] = hc_bytealign_S (w4[1], w4[2], offset); - w5[3] = hc_bytealign_S (w4[0], w4[1], offset); - w5[2] = hc_bytealign_S (w3[3], w4[0], offset); - w5[1] = hc_bytealign_S (w3[2], w3[3], offset); - w5[0] = hc_bytealign_S (w3[1], w3[2], offset); - w4[3] = hc_bytealign_S (w3[0], w3[1], offset); - w4[2] = hc_bytealign_S (w2[3], w3[0], offset); - w4[1] = hc_bytealign_S (w2[2], w2[3], offset); - w4[0] = hc_bytealign_S (w2[1], w2[2], offset); - w3[3] = hc_bytealign_S (w2[0], w2[1], offset); - w3[2] = hc_bytealign_S (w1[3], w2[0], offset); - w3[1] = hc_bytealign_S (w1[2], w1[3], offset); - w3[0] = hc_bytealign_S (w1[1], w1[2], offset); - w2[3] = hc_bytealign_S (w1[0], w1[1], offset); - w2[2] = hc_bytealign_S (w0[3], w1[0], offset); - w2[1] = hc_bytealign_S (w0[2], w0[3], offset); - w2[0] = hc_bytealign_S (w0[1], w0[2], offset); - w1[3] = hc_bytealign_S (w0[0], w0[1], offset); - w1[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be_S ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -37106,31 +36794,31 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 7: - w7[3] = hc_bytealign_S (w5[3], w6[0], offset); - w7[2] = hc_bytealign_S (w5[2], w5[3], offset); - w7[1] = hc_bytealign_S (w5[1], w5[2], offset); - w7[0] = hc_bytealign_S (w5[0], w5[1], offset); - w6[3] = hc_bytealign_S (w4[3], w5[0], offset); - w6[2] = hc_bytealign_S (w4[2], w4[3], offset); - w6[1] = hc_bytealign_S (w4[1], w4[2], offset); - w6[0] = hc_bytealign_S (w4[0], w4[1], offset); - w5[3] = hc_bytealign_S (w3[3], w4[0], offset); - w5[2] = hc_bytealign_S (w3[2], w3[3], offset); - w5[1] = hc_bytealign_S (w3[1], w3[2], offset); - w5[0] = hc_bytealign_S (w3[0], w3[1], offset); - w4[3] = hc_bytealign_S (w2[3], w3[0], offset); - w4[2] = hc_bytealign_S (w2[2], w2[3], offset); - w4[1] = hc_bytealign_S (w2[1], w2[2], offset); - w4[0] = hc_bytealign_S (w2[0], w2[1], offset); - w3[3] = hc_bytealign_S (w1[3], w2[0], offset); - w3[2] = hc_bytealign_S (w1[2], w1[3], offset); - w3[1] = hc_bytealign_S (w1[1], w1[2], offset); - w3[0] = hc_bytealign_S (w1[0], w1[1], offset); - w2[3] = hc_bytealign_S (w0[3], w1[0], offset); - w2[2] = hc_bytealign_S (w0[2], w0[3], offset); - w2[1] = hc_bytealign_S (w0[1], w0[2], offset); - w2[0] = hc_bytealign_S (w0[0], w0[1], offset); - w1[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be_S ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -37142,30 +36830,30 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 8: - w7[3] = hc_bytealign_S (w5[2], w5[3], offset); - w7[2] = hc_bytealign_S (w5[1], w5[2], offset); - w7[1] = hc_bytealign_S (w5[0], w5[1], offset); - w7[0] = hc_bytealign_S (w4[3], w5[0], offset); - w6[3] = hc_bytealign_S (w4[2], w4[3], offset); - w6[2] = hc_bytealign_S (w4[1], w4[2], offset); - w6[1] = hc_bytealign_S (w4[0], w4[1], offset); - w6[0] = hc_bytealign_S (w3[3], w4[0], offset); - w5[3] = hc_bytealign_S (w3[2], w3[3], offset); - w5[2] = hc_bytealign_S (w3[1], w3[2], offset); - w5[1] = hc_bytealign_S (w3[0], w3[1], offset); - w5[0] = hc_bytealign_S (w2[3], w3[0], offset); - w4[3] = hc_bytealign_S (w2[2], w2[3], offset); - w4[2] = hc_bytealign_S (w2[1], w2[2], offset); - w4[1] = hc_bytealign_S (w2[0], w2[1], offset); - w4[0] = hc_bytealign_S (w1[3], w2[0], offset); - w3[3] = hc_bytealign_S (w1[2], w1[3], offset); - w3[2] = hc_bytealign_S (w1[1], w1[2], offset); - w3[1] = hc_bytealign_S (w1[0], w1[1], offset); - w3[0] = hc_bytealign_S (w0[3], w1[0], offset); - w2[3] = hc_bytealign_S (w0[2], w0[3], offset); - w2[2] = hc_bytealign_S (w0[1], w0[2], offset); - w2[1] = hc_bytealign_S (w0[0], w0[1], offset); - w2[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be_S ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -37178,29 +36866,29 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 9: - w7[3] = hc_bytealign_S (w5[1], w5[2], offset); - w7[2] = hc_bytealign_S (w5[0], w5[1], offset); - w7[1] = hc_bytealign_S (w4[3], w5[0], offset); - w7[0] = hc_bytealign_S (w4[2], w4[3], offset); - w6[3] = hc_bytealign_S (w4[1], w4[2], offset); - w6[2] = hc_bytealign_S (w4[0], w4[1], offset); - w6[1] = hc_bytealign_S (w3[3], w4[0], offset); - w6[0] = hc_bytealign_S (w3[2], w3[3], offset); - w5[3] = hc_bytealign_S (w3[1], w3[2], offset); - w5[2] = hc_bytealign_S (w3[0], w3[1], offset); - w5[1] = hc_bytealign_S (w2[3], w3[0], offset); - w5[0] = hc_bytealign_S (w2[2], w2[3], offset); - w4[3] = hc_bytealign_S (w2[1], w2[2], offset); - w4[2] = hc_bytealign_S (w2[0], w2[1], offset); - w4[1] = hc_bytealign_S (w1[3], w2[0], offset); - w4[0] = hc_bytealign_S (w1[2], w1[3], offset); - w3[3] = hc_bytealign_S (w1[1], w1[2], offset); - w3[2] = hc_bytealign_S (w1[0], w1[1], offset); - w3[1] = hc_bytealign_S (w0[3], w1[0], offset); - w3[0] = hc_bytealign_S (w0[2], w0[3], offset); - w2[3] = hc_bytealign_S (w0[1], w0[2], offset); - w2[2] = hc_bytealign_S (w0[0], w0[1], offset); - w2[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be_S ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -37214,28 +36902,28 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 10: - w7[3] = hc_bytealign_S (w5[0], w5[1], offset); - w7[2] = hc_bytealign_S (w4[3], w5[0], offset); - w7[1] = hc_bytealign_S (w4[2], w4[3], offset); - w7[0] = hc_bytealign_S (w4[1], w4[2], offset); - w6[3] = hc_bytealign_S (w4[0], w4[1], offset); - w6[2] = hc_bytealign_S (w3[3], w4[0], offset); - w6[1] = hc_bytealign_S (w3[2], w3[3], offset); - w6[0] = hc_bytealign_S (w3[1], w3[2], offset); - w5[3] = hc_bytealign_S (w3[0], w3[1], offset); - w5[2] = hc_bytealign_S (w2[3], w3[0], offset); - w5[1] = hc_bytealign_S (w2[2], w2[3], offset); - w5[0] = hc_bytealign_S (w2[1], w2[2], offset); - w4[3] = hc_bytealign_S (w2[0], w2[1], offset); - w4[2] = hc_bytealign_S (w1[3], w2[0], offset); - w4[1] = hc_bytealign_S (w1[2], w1[3], offset); - w4[0] = hc_bytealign_S (w1[1], w1[2], offset); - w3[3] = hc_bytealign_S (w1[0], w1[1], offset); - w3[2] = hc_bytealign_S (w0[3], w1[0], offset); - w3[1] = hc_bytealign_S (w0[2], w0[3], offset); - w3[0] = hc_bytealign_S (w0[1], w0[2], offset); - w2[3] = hc_bytealign_S (w0[0], w0[1], offset); - w2[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be_S ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -37250,27 +36938,27 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 11: - w7[3] = hc_bytealign_S (w4[3], w5[0], offset); - w7[2] = hc_bytealign_S (w4[2], w4[3], offset); - w7[1] = hc_bytealign_S (w4[1], w4[2], offset); - w7[0] = hc_bytealign_S (w4[0], w4[1], offset); - w6[3] = hc_bytealign_S (w3[3], w4[0], offset); - w6[2] = hc_bytealign_S (w3[2], w3[3], offset); - w6[1] = hc_bytealign_S (w3[1], w3[2], offset); - w6[0] = hc_bytealign_S (w3[0], w3[1], offset); - w5[3] = hc_bytealign_S (w2[3], w3[0], offset); - w5[2] = hc_bytealign_S (w2[2], w2[3], offset); - w5[1] = hc_bytealign_S (w2[1], w2[2], offset); - w5[0] = hc_bytealign_S (w2[0], w2[1], offset); - w4[3] = hc_bytealign_S (w1[3], w2[0], offset); - w4[2] = hc_bytealign_S (w1[2], w1[3], offset); - w4[1] = hc_bytealign_S (w1[1], w1[2], offset); - w4[0] = hc_bytealign_S (w1[0], w1[1], offset); - w3[3] = hc_bytealign_S (w0[3], w1[0], offset); - w3[2] = hc_bytealign_S (w0[2], w0[3], offset); - w3[1] = hc_bytealign_S (w0[1], w0[2], offset); - w3[0] = hc_bytealign_S (w0[0], w0[1], offset); - w2[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be_S ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -37286,26 +36974,26 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 12: - w7[3] = hc_bytealign_S (w4[2], w4[3], offset); - w7[2] = hc_bytealign_S (w4[1], w4[2], offset); - w7[1] = hc_bytealign_S (w4[0], w4[1], offset); - w7[0] = hc_bytealign_S (w3[3], w4[0], offset); - w6[3] = hc_bytealign_S (w3[2], w3[3], offset); - w6[2] = hc_bytealign_S (w3[1], w3[2], offset); - w6[1] = hc_bytealign_S (w3[0], w3[1], offset); - w6[0] = hc_bytealign_S (w2[3], w3[0], offset); - w5[3] = hc_bytealign_S (w2[2], w2[3], offset); - w5[2] = hc_bytealign_S (w2[1], w2[2], offset); - w5[1] = hc_bytealign_S (w2[0], w2[1], offset); - w5[0] = hc_bytealign_S (w1[3], w2[0], offset); - w4[3] = hc_bytealign_S (w1[2], w1[3], offset); - w4[2] = hc_bytealign_S (w1[1], w1[2], offset); - w4[1] = hc_bytealign_S (w1[0], w1[1], offset); - w4[0] = hc_bytealign_S (w0[3], w1[0], offset); - w3[3] = hc_bytealign_S (w0[2], w0[3], offset); - w3[2] = hc_bytealign_S (w0[1], w0[2], offset); - w3[1] = hc_bytealign_S (w0[0], w0[1], offset); - w3[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be_S ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -37322,25 +37010,25 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 13: - w7[3] = hc_bytealign_S (w4[1], w4[2], offset); - w7[2] = hc_bytealign_S (w4[0], w4[1], offset); - w7[1] = hc_bytealign_S (w3[3], w4[0], offset); - w7[0] = hc_bytealign_S (w3[2], w3[3], offset); - w6[3] = hc_bytealign_S (w3[1], w3[2], offset); - w6[2] = hc_bytealign_S (w3[0], w3[1], offset); - w6[1] = hc_bytealign_S (w2[3], w3[0], offset); - w6[0] = hc_bytealign_S (w2[2], w2[3], offset); - w5[3] = hc_bytealign_S (w2[1], w2[2], offset); - w5[2] = hc_bytealign_S (w2[0], w2[1], offset); - w5[1] = hc_bytealign_S (w1[3], w2[0], offset); - w5[0] = hc_bytealign_S (w1[2], w1[3], offset); - w4[3] = hc_bytealign_S (w1[1], w1[2], offset); - w4[2] = hc_bytealign_S (w1[0], w1[1], offset); - w4[1] = hc_bytealign_S (w0[3], w1[0], offset); - w4[0] = hc_bytealign_S (w0[2], w0[3], offset); - w3[3] = hc_bytealign_S (w0[1], w0[2], offset); - w3[2] = hc_bytealign_S (w0[0], w0[1], offset); - w3[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be_S ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -37358,24 +37046,24 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 14: - w7[3] = hc_bytealign_S (w4[0], w4[1], offset); - w7[2] = hc_bytealign_S (w3[3], w4[0], offset); - w7[1] = hc_bytealign_S (w3[2], w3[3], offset); - w7[0] = hc_bytealign_S (w3[1], w3[2], offset); - w6[3] = hc_bytealign_S (w3[0], w3[1], offset); - w6[2] = hc_bytealign_S (w2[3], w3[0], offset); - w6[1] = hc_bytealign_S (w2[2], w2[3], offset); - w6[0] = hc_bytealign_S (w2[1], w2[2], offset); - w5[3] = hc_bytealign_S (w2[0], w2[1], offset); - w5[2] = hc_bytealign_S (w1[3], w2[0], offset); - w5[1] = hc_bytealign_S (w1[2], w1[3], offset); - w5[0] = hc_bytealign_S (w1[1], w1[2], offset); - w4[3] = hc_bytealign_S (w1[0], w1[1], offset); - w4[2] = hc_bytealign_S (w0[3], w1[0], offset); - w4[1] = hc_bytealign_S (w0[2], w0[3], offset); - w4[0] = hc_bytealign_S (w0[1], w0[2], offset); - w3[3] = hc_bytealign_S (w0[0], w0[1], offset); - w3[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be_S ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -37394,23 +37082,23 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 15: - w7[3] = hc_bytealign_S (w3[3], w4[0], offset); - w7[2] = hc_bytealign_S (w3[2], w3[3], offset); - w7[1] = hc_bytealign_S (w3[1], w3[2], offset); - w7[0] = hc_bytealign_S (w3[0], w3[1], offset); - w6[3] = hc_bytealign_S (w2[3], w3[0], offset); - w6[2] = hc_bytealign_S (w2[2], w2[3], offset); - w6[1] = hc_bytealign_S (w2[1], w2[2], offset); - w6[0] = hc_bytealign_S (w2[0], w2[1], offset); - w5[3] = hc_bytealign_S (w1[3], w2[0], offset); - w5[2] = hc_bytealign_S (w1[2], w1[3], offset); - w5[1] = hc_bytealign_S (w1[1], w1[2], offset); - w5[0] = hc_bytealign_S (w1[0], w1[1], offset); - w4[3] = hc_bytealign_S (w0[3], w1[0], offset); - w4[2] = hc_bytealign_S (w0[2], w0[3], offset); - w4[1] = hc_bytealign_S (w0[1], w0[2], offset); - w4[0] = hc_bytealign_S (w0[0], w0[1], offset); - w3[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be_S ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -37430,22 +37118,22 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 16: - w7[3] = hc_bytealign_S (w3[2], w3[3], offset); - w7[2] = hc_bytealign_S (w3[1], w3[2], offset); - w7[1] = hc_bytealign_S (w3[0], w3[1], offset); - w7[0] = hc_bytealign_S (w2[3], w3[0], offset); - w6[3] = hc_bytealign_S (w2[2], w2[3], offset); - w6[2] = hc_bytealign_S (w2[1], w2[2], offset); - w6[1] = hc_bytealign_S (w2[0], w2[1], offset); - w6[0] = hc_bytealign_S (w1[3], w2[0], offset); - w5[3] = hc_bytealign_S (w1[2], w1[3], offset); - w5[2] = hc_bytealign_S (w1[1], w1[2], offset); - w5[1] = hc_bytealign_S (w1[0], w1[1], offset); - w5[0] = hc_bytealign_S (w0[3], w1[0], offset); - w4[3] = hc_bytealign_S (w0[2], w0[3], offset); - w4[2] = hc_bytealign_S (w0[1], w0[2], offset); - w4[1] = hc_bytealign_S (w0[0], w0[1], offset); - w4[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[0] = hc_bytealign_be_S ( 0, w0[0], offset); w3[3] = 0; w3[2] = 0; w3[1] = 0; @@ -37466,21 +37154,21 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 17: - w7[3] = hc_bytealign_S (w3[1], w3[2], offset); - w7[2] = hc_bytealign_S (w3[0], w3[1], offset); - w7[1] = hc_bytealign_S (w2[3], w3[0], offset); - w7[0] = hc_bytealign_S (w2[2], w2[3], offset); - w6[3] = hc_bytealign_S (w2[1], w2[2], offset); - w6[2] = hc_bytealign_S (w2[0], w2[1], offset); - w6[1] = hc_bytealign_S (w1[3], w2[0], offset); - w6[0] = hc_bytealign_S (w1[2], w1[3], offset); - w5[3] = hc_bytealign_S (w1[1], w1[2], offset); - w5[2] = hc_bytealign_S (w1[0], w1[1], offset); - w5[1] = hc_bytealign_S (w0[3], w1[0], offset); - w5[0] = hc_bytealign_S (w0[2], w0[3], offset); - w4[3] = hc_bytealign_S (w0[1], w0[2], offset); - w4[2] = hc_bytealign_S (w0[0], w0[1], offset); - w4[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[1] = hc_bytealign_be_S ( 0, w0[0], offset); w4[0] = 0; w3[3] = 0; w3[2] = 0; @@ -37502,20 +37190,20 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 18: - w7[3] = hc_bytealign_S (w3[0], w3[1], offset); - w7[2] = hc_bytealign_S (w2[3], w3[0], offset); - w7[1] = hc_bytealign_S (w2[2], w2[3], offset); - w7[0] = hc_bytealign_S (w2[1], w2[2], offset); - w6[3] = hc_bytealign_S (w2[0], w2[1], offset); - w6[2] = hc_bytealign_S (w1[3], w2[0], offset); - w6[1] = hc_bytealign_S (w1[2], w1[3], offset); - w6[0] = hc_bytealign_S (w1[1], w1[2], offset); - w5[3] = hc_bytealign_S (w1[0], w1[1], offset); - w5[2] = hc_bytealign_S (w0[3], w1[0], offset); - w5[1] = hc_bytealign_S (w0[2], w0[3], offset); - w5[0] = hc_bytealign_S (w0[1], w0[2], offset); - w4[3] = hc_bytealign_S (w0[0], w0[1], offset); - w4[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[2] = hc_bytealign_be_S ( 0, w0[0], offset); w4[1] = 0; w4[0] = 0; w3[3] = 0; @@ -37538,19 +37226,19 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 19: - w7[3] = hc_bytealign_S (w2[3], w3[0], offset); - w7[2] = hc_bytealign_S (w2[2], w2[3], offset); - w7[1] = hc_bytealign_S (w2[1], w2[2], offset); - w7[0] = hc_bytealign_S (w2[0], w2[1], offset); - w6[3] = hc_bytealign_S (w1[3], w2[0], offset); - w6[2] = hc_bytealign_S (w1[2], w1[3], offset); - w6[1] = hc_bytealign_S (w1[1], w1[2], offset); - w6[0] = hc_bytealign_S (w1[0], w1[1], offset); - w5[3] = hc_bytealign_S (w0[3], w1[0], offset); - w5[2] = hc_bytealign_S (w0[2], w0[3], offset); - w5[1] = hc_bytealign_S (w0[1], w0[2], offset); - w5[0] = hc_bytealign_S (w0[0], w0[1], offset); - w4[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[3] = hc_bytealign_be_S ( 0, w0[0], offset); w4[2] = 0; w4[1] = 0; w4[0] = 0; @@ -37574,18 +37262,18 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 20: - w7[3] = hc_bytealign_S (w2[2], w2[3], offset); - w7[2] = hc_bytealign_S (w2[1], w2[2], offset); - w7[1] = hc_bytealign_S (w2[0], w2[1], offset); - w7[0] = hc_bytealign_S (w1[3], w2[0], offset); - w6[3] = hc_bytealign_S (w1[2], w1[3], offset); - w6[2] = hc_bytealign_S (w1[1], w1[2], offset); - w6[1] = hc_bytealign_S (w1[0], w1[1], offset); - w6[0] = hc_bytealign_S (w0[3], w1[0], offset); - w5[3] = hc_bytealign_S (w0[2], w0[3], offset); - w5[2] = hc_bytealign_S (w0[1], w0[2], offset); - w5[1] = hc_bytealign_S (w0[0], w0[1], offset); - w5[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[0] = hc_bytealign_be_S ( 0, w0[0], offset); w4[3] = 0; w4[2] = 0; w4[1] = 0; @@ -37610,17 +37298,17 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 21: - w7[3] = hc_bytealign_S (w2[1], w2[2], offset); - w7[2] = hc_bytealign_S (w2[0], w2[1], offset); - w7[1] = hc_bytealign_S (w1[3], w2[0], offset); - w7[0] = hc_bytealign_S (w1[2], w1[3], offset); - w6[3] = hc_bytealign_S (w1[1], w1[2], offset); - w6[2] = hc_bytealign_S (w1[0], w1[1], offset); - w6[1] = hc_bytealign_S (w0[3], w1[0], offset); - w6[0] = hc_bytealign_S (w0[2], w0[3], offset); - w5[3] = hc_bytealign_S (w0[1], w0[2], offset); - w5[2] = hc_bytealign_S (w0[0], w0[1], offset); - w5[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[1] = hc_bytealign_be_S ( 0, w0[0], offset); w5[0] = 0; w4[3] = 0; w4[2] = 0; @@ -37646,16 +37334,16 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 22: - w7[3] = hc_bytealign_S (w2[0], w2[1], offset); - w7[2] = hc_bytealign_S (w1[3], w2[0], offset); - w7[1] = hc_bytealign_S (w1[2], w1[3], offset); - w7[0] = hc_bytealign_S (w1[1], w1[2], offset); - w6[3] = hc_bytealign_S (w1[0], w1[1], offset); - w6[2] = hc_bytealign_S (w0[3], w1[0], offset); - w6[1] = hc_bytealign_S (w0[2], w0[3], offset); - w6[0] = hc_bytealign_S (w0[1], w0[2], offset); - w5[3] = hc_bytealign_S (w0[0], w0[1], offset); - w5[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[2] = hc_bytealign_be_S ( 0, w0[0], offset); w5[1] = 0; w5[0] = 0; w4[3] = 0; @@ -37682,15 +37370,15 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 23: - w7[3] = hc_bytealign_S (w1[3], w2[0], offset); - w7[2] = hc_bytealign_S (w1[2], w1[3], offset); - w7[1] = hc_bytealign_S (w1[1], w1[2], offset); - w7[0] = hc_bytealign_S (w1[0], w1[1], offset); - w6[3] = hc_bytealign_S (w0[3], w1[0], offset); - w6[2] = hc_bytealign_S (w0[2], w0[3], offset); - w6[1] = hc_bytealign_S (w0[1], w0[2], offset); - w6[0] = hc_bytealign_S (w0[0], w0[1], offset); - w5[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[3] = hc_bytealign_be_S ( 0, w0[0], offset); w5[2] = 0; w5[1] = 0; w5[0] = 0; @@ -37718,14 +37406,14 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 24: - w7[3] = hc_bytealign_S (w1[2], w1[3], offset); - w7[2] = hc_bytealign_S (w1[1], w1[2], offset); - w7[1] = hc_bytealign_S (w1[0], w1[1], offset); - w7[0] = hc_bytealign_S (w0[3], w1[0], offset); - w6[3] = hc_bytealign_S (w0[2], w0[3], offset); - w6[2] = hc_bytealign_S (w0[1], w0[2], offset); - w6[1] = hc_bytealign_S (w0[0], w0[1], offset); - w6[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[0] = hc_bytealign_be_S ( 0, w0[0], offset); w5[3] = 0; w5[2] = 0; w5[1] = 0; @@ -37754,13 +37442,13 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 25: - w7[3] = hc_bytealign_S (w1[1], w1[2], offset); - w7[2] = hc_bytealign_S (w1[0], w1[1], offset); - w7[1] = hc_bytealign_S (w0[3], w1[0], offset); - w7[0] = hc_bytealign_S (w0[2], w0[3], offset); - w6[3] = hc_bytealign_S (w0[1], w0[2], offset); - w6[2] = hc_bytealign_S (w0[0], w0[1], offset); - w6[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[1] = hc_bytealign_be_S ( 0, w0[0], offset); w6[0] = 0; w5[3] = 0; w5[2] = 0; @@ -37790,12 +37478,12 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 26: - w7[3] = hc_bytealign_S (w1[0], w1[1], offset); - w7[2] = hc_bytealign_S (w0[3], w1[0], offset); - w7[1] = hc_bytealign_S (w0[2], w0[3], offset); - w7[0] = hc_bytealign_S (w0[1], w0[2], offset); - w6[3] = hc_bytealign_S (w0[0], w0[1], offset); - w6[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[2] = hc_bytealign_be_S ( 0, w0[0], offset); w6[1] = 0; w6[0] = 0; w5[3] = 0; @@ -37826,11 +37514,11 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 27: - w7[3] = hc_bytealign_S (w0[3], w1[0], offset); - w7[2] = hc_bytealign_S (w0[2], w0[3], offset); - w7[1] = hc_bytealign_S (w0[1], w0[2], offset); - w7[0] = hc_bytealign_S (w0[0], w0[1], offset); - w6[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[3] = hc_bytealign_be_S ( 0, w0[0], offset); w6[2] = 0; w6[1] = 0; w6[0] = 0; @@ -37862,10 +37550,10 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 28: - w7[3] = hc_bytealign_S (w0[2], w0[3], offset); - w7[2] = hc_bytealign_S (w0[1], w0[2], offset); - w7[1] = hc_bytealign_S (w0[0], w0[1], offset); - w7[0] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[0] = hc_bytealign_be_S ( 0, w0[0], offset); w6[3] = 0; w6[2] = 0; w6[1] = 0; @@ -37898,9 +37586,9 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 29: - w7[3] = hc_bytealign_S (w0[1], w0[2], offset); - w7[2] = hc_bytealign_S (w0[0], w0[1], offset); - w7[1] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[1] = hc_bytealign_be_S ( 0, w0[0], offset); w7[0] = 0; w6[3] = 0; w6[2] = 0; @@ -37934,8 +37622,8 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 30: - w7[3] = hc_bytealign_S (w0[0], w0[1], offset); - w7[2] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[2] = hc_bytealign_be_S ( 0, w0[0], offset); w7[1] = 0; w7[0] = 0; w6[3] = 0; @@ -37970,7 +37658,7 @@ DECLSPEC void switch_buffer_by_offset_8x4_be_S (u32 *w0, u32 *w1, u32 *w2, u32 * break; case 31: - w7[3] = hc_bytealign_S ( 0, w0[0], offset); + w7[3] = hc_bytealign_be_S ( 0, w0[0], offset); w7[2] = 0; w7[1] = 0; w7[0] = 0; @@ -39182,153 +38870,153 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, switch (offset_switch) { case 0: - c0[0] = hc_bytealign_S (w7[3], 0, offset); - w7[3] = hc_bytealign_S (w7[2], w7[3], offset); - w7[2] = hc_bytealign_S (w7[1], w7[2], offset); - w7[1] = hc_bytealign_S (w7[0], w7[1], offset); - w7[0] = hc_bytealign_S (w6[3], w7[0], offset); - w6[3] = hc_bytealign_S (w6[2], w6[3], offset); - w6[2] = hc_bytealign_S (w6[1], w6[2], offset); - w6[1] = hc_bytealign_S (w6[0], w6[1], offset); - w6[0] = hc_bytealign_S (w5[3], w6[0], offset); - w5[3] = hc_bytealign_S (w5[2], w5[3], offset); - w5[2] = hc_bytealign_S (w5[1], w5[2], offset); - w5[1] = hc_bytealign_S (w5[0], w5[1], offset); - w5[0] = hc_bytealign_S (w4[3], w5[0], offset); - w4[3] = hc_bytealign_S (w4[2], w4[3], offset); - w4[2] = hc_bytealign_S (w4[1], w4[2], offset); - w4[1] = hc_bytealign_S (w4[0], w4[1], offset); - w4[0] = hc_bytealign_S (w3[3], w4[0], offset); - w3[3] = hc_bytealign_S (w3[2], w3[3], offset); - w3[2] = hc_bytealign_S (w3[1], w3[2], offset); - w3[1] = hc_bytealign_S (w3[0], w3[1], offset); - w3[0] = hc_bytealign_S (w2[3], w3[0], offset); - w2[3] = hc_bytealign_S (w2[2], w2[3], offset); - w2[2] = hc_bytealign_S (w2[1], w2[2], offset); - w2[1] = hc_bytealign_S (w2[0], w2[1], offset); - w2[0] = hc_bytealign_S (w1[3], w2[0], offset); - w1[3] = hc_bytealign_S (w1[2], w1[3], offset); - w1[2] = hc_bytealign_S (w1[1], w1[2], offset); - w1[1] = hc_bytealign_S (w1[0], w1[1], offset); - w1[0] = hc_bytealign_S (w0[3], w1[0], offset); - w0[3] = hc_bytealign_S (w0[2], w0[3], offset); - w0[2] = hc_bytealign_S (w0[1], w0[2], offset); - w0[1] = hc_bytealign_S (w0[0], w0[1], offset); - w0[0] = hc_bytealign_S ( 0, w0[0], offset); + c0[0] = hc_bytealign_be_S (w7[3], 0, offset); + w7[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + w7[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + w7[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + w6[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + w6[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + w5[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w5[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w4[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w4[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w3[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w2[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w1[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w0[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[0] = hc_bytealign_be_S ( 0, w0[0], offset); break; case 1: - c0[1] = hc_bytealign_S (w7[3], 0, offset); - c0[0] = hc_bytealign_S (w7[2], w7[3], offset); - w7[3] = hc_bytealign_S (w7[1], w7[2], offset); - w7[2] = hc_bytealign_S (w7[0], w7[1], offset); - w7[1] = hc_bytealign_S (w6[3], w7[0], offset); - w7[0] = hc_bytealign_S (w6[2], w6[3], offset); - w6[3] = hc_bytealign_S (w6[1], w6[2], offset); - w6[2] = hc_bytealign_S (w6[0], w6[1], offset); - w6[1] = hc_bytealign_S (w5[3], w6[0], offset); - w6[0] = hc_bytealign_S (w5[2], w5[3], offset); - w5[3] = hc_bytealign_S (w5[1], w5[2], offset); - w5[2] = hc_bytealign_S (w5[0], w5[1], offset); - w5[1] = hc_bytealign_S (w4[3], w5[0], offset); - w5[0] = hc_bytealign_S (w4[2], w4[3], offset); - w4[3] = hc_bytealign_S (w4[1], w4[2], offset); - w4[2] = hc_bytealign_S (w4[0], w4[1], offset); - w4[1] = hc_bytealign_S (w3[3], w4[0], offset); - w4[0] = hc_bytealign_S (w3[2], w3[3], offset); - w3[3] = hc_bytealign_S (w3[1], w3[2], offset); - w3[2] = hc_bytealign_S (w3[0], w3[1], offset); - w3[1] = hc_bytealign_S (w2[3], w3[0], offset); - w3[0] = hc_bytealign_S (w2[2], w2[3], offset); - w2[3] = hc_bytealign_S (w2[1], w2[2], offset); - w2[2] = hc_bytealign_S (w2[0], w2[1], offset); - w2[1] = hc_bytealign_S (w1[3], w2[0], offset); - w2[0] = hc_bytealign_S (w1[2], w1[3], offset); - w1[3] = hc_bytealign_S (w1[1], w1[2], offset); - w1[2] = hc_bytealign_S (w1[0], w1[1], offset); - w1[1] = hc_bytealign_S (w0[3], w1[0], offset); - w1[0] = hc_bytealign_S (w0[2], w0[3], offset); - w0[3] = hc_bytealign_S (w0[1], w0[2], offset); - w0[2] = hc_bytealign_S (w0[0], w0[1], offset); - w0[1] = hc_bytealign_S ( 0, w0[0], offset); + c0[1] = hc_bytealign_be_S (w7[3], 0, offset); + c0[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + w7[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + w7[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + w6[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + w5[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w4[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w3[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w2[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w1[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w0[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[1] = hc_bytealign_be_S ( 0, w0[0], offset); w0[0] = 0; break; case 2: - c0[2] = hc_bytealign_S (w7[3], 0, offset); - c0[1] = hc_bytealign_S (w7[2], w7[3], offset); - c0[0] = hc_bytealign_S (w7[1], w7[2], offset); - w7[3] = hc_bytealign_S (w7[0], w7[1], offset); - w7[2] = hc_bytealign_S (w6[3], w7[0], offset); - w7[1] = hc_bytealign_S (w6[2], w6[3], offset); - w7[0] = hc_bytealign_S (w6[1], w6[2], offset); - w6[3] = hc_bytealign_S (w6[0], w6[1], offset); - w6[2] = hc_bytealign_S (w5[3], w6[0], offset); - w6[1] = hc_bytealign_S (w5[2], w5[3], offset); - w6[0] = hc_bytealign_S (w5[1], w5[2], offset); - w5[3] = hc_bytealign_S (w5[0], w5[1], offset); - w5[2] = hc_bytealign_S (w4[3], w5[0], offset); - w5[1] = hc_bytealign_S (w4[2], w4[3], offset); - w5[0] = hc_bytealign_S (w4[1], w4[2], offset); - w4[3] = hc_bytealign_S (w4[0], w4[1], offset); - w4[2] = hc_bytealign_S (w3[3], w4[0], offset); - w4[1] = hc_bytealign_S (w3[2], w3[3], offset); - w4[0] = hc_bytealign_S (w3[1], w3[2], offset); - w3[3] = hc_bytealign_S (w3[0], w3[1], offset); - w3[2] = hc_bytealign_S (w2[3], w3[0], offset); - w3[1] = hc_bytealign_S (w2[2], w2[3], offset); - w3[0] = hc_bytealign_S (w2[1], w2[2], offset); - w2[3] = hc_bytealign_S (w2[0], w2[1], offset); - w2[2] = hc_bytealign_S (w1[3], w2[0], offset); - w2[1] = hc_bytealign_S (w1[2], w1[3], offset); - w2[0] = hc_bytealign_S (w1[1], w1[2], offset); - w1[3] = hc_bytealign_S (w1[0], w1[1], offset); - w1[2] = hc_bytealign_S (w0[3], w1[0], offset); - w1[1] = hc_bytealign_S (w0[2], w0[3], offset); - w1[0] = hc_bytealign_S (w0[1], w0[2], offset); - w0[3] = hc_bytealign_S (w0[0], w0[1], offset); - w0[2] = hc_bytealign_S ( 0, w0[0], offset); + c0[2] = hc_bytealign_be_S (w7[3], 0, offset); + c0[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c0[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + w7[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + w6[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + w5[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w4[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w3[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w2[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w1[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w0[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[2] = hc_bytealign_be_S ( 0, w0[0], offset); w0[1] = 0; w0[0] = 0; break; case 3: - c0[3] = hc_bytealign_S (w7[3], 0, offset); - c0[2] = hc_bytealign_S (w7[2], w7[3], offset); - c0[1] = hc_bytealign_S (w7[1], w7[2], offset); - c0[0] = hc_bytealign_S (w7[0], w7[1], offset); - w7[3] = hc_bytealign_S (w6[3], w7[0], offset); - w7[2] = hc_bytealign_S (w6[2], w6[3], offset); - w7[1] = hc_bytealign_S (w6[1], w6[2], offset); - w7[0] = hc_bytealign_S (w6[0], w6[1], offset); - w6[3] = hc_bytealign_S (w5[3], w6[0], offset); - w6[2] = hc_bytealign_S (w5[2], w5[3], offset); - w6[1] = hc_bytealign_S (w5[1], w5[2], offset); - w6[0] = hc_bytealign_S (w5[0], w5[1], offset); - w5[3] = hc_bytealign_S (w4[3], w5[0], offset); - w5[2] = hc_bytealign_S (w4[2], w4[3], offset); - w5[1] = hc_bytealign_S (w4[1], w4[2], offset); - w5[0] = hc_bytealign_S (w4[0], w4[1], offset); - w4[3] = hc_bytealign_S (w3[3], w4[0], offset); - w4[2] = hc_bytealign_S (w3[2], w3[3], offset); - w4[1] = hc_bytealign_S (w3[1], w3[2], offset); - w4[0] = hc_bytealign_S (w3[0], w3[1], offset); - w3[3] = hc_bytealign_S (w2[3], w3[0], offset); - w3[2] = hc_bytealign_S (w2[2], w2[3], offset); - w3[1] = hc_bytealign_S (w2[1], w2[2], offset); - w3[0] = hc_bytealign_S (w2[0], w2[1], offset); - w2[3] = hc_bytealign_S (w1[3], w2[0], offset); - w2[2] = hc_bytealign_S (w1[2], w1[3], offset); - w2[1] = hc_bytealign_S (w1[1], w1[2], offset); - w2[0] = hc_bytealign_S (w1[0], w1[1], offset); - w1[3] = hc_bytealign_S (w0[3], w1[0], offset); - w1[2] = hc_bytealign_S (w0[2], w0[3], offset); - w1[1] = hc_bytealign_S (w0[1], w0[2], offset); - w1[0] = hc_bytealign_S (w0[0], w0[1], offset); - w0[3] = hc_bytealign_S ( 0, w0[0], offset); + c0[3] = hc_bytealign_be_S (w7[3], 0, offset); + c0[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c0[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c0[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + w7[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + w6[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + w5[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w4[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w3[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w2[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w1[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w0[3] = hc_bytealign_be_S ( 0, w0[0], offset); w0[2] = 0; w0[1] = 0; w0[0] = 0; @@ -39336,39 +39024,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 4: - c1[0] = hc_bytealign_S (w7[3], 0, offset); - c0[3] = hc_bytealign_S (w7[2], w7[3], offset); - c0[2] = hc_bytealign_S (w7[1], w7[2], offset); - c0[1] = hc_bytealign_S (w7[0], w7[1], offset); - c0[0] = hc_bytealign_S (w6[3], w7[0], offset); - w7[3] = hc_bytealign_S (w6[2], w6[3], offset); - w7[2] = hc_bytealign_S (w6[1], w6[2], offset); - w7[1] = hc_bytealign_S (w6[0], w6[1], offset); - w7[0] = hc_bytealign_S (w5[3], w6[0], offset); - w6[3] = hc_bytealign_S (w5[2], w5[3], offset); - w6[2] = hc_bytealign_S (w5[1], w5[2], offset); - w6[1] = hc_bytealign_S (w5[0], w5[1], offset); - w6[0] = hc_bytealign_S (w4[3], w5[0], offset); - w5[3] = hc_bytealign_S (w4[2], w4[3], offset); - w5[2] = hc_bytealign_S (w4[1], w4[2], offset); - w5[1] = hc_bytealign_S (w4[0], w4[1], offset); - w5[0] = hc_bytealign_S (w3[3], w4[0], offset); - w4[3] = hc_bytealign_S (w3[2], w3[3], offset); - w4[2] = hc_bytealign_S (w3[1], w3[2], offset); - w4[1] = hc_bytealign_S (w3[0], w3[1], offset); - w4[0] = hc_bytealign_S (w2[3], w3[0], offset); - w3[3] = hc_bytealign_S (w2[2], w2[3], offset); - w3[2] = hc_bytealign_S (w2[1], w2[2], offset); - w3[1] = hc_bytealign_S (w2[0], w2[1], offset); - w3[0] = hc_bytealign_S (w1[3], w2[0], offset); - w2[3] = hc_bytealign_S (w1[2], w1[3], offset); - w2[2] = hc_bytealign_S (w1[1], w1[2], offset); - w2[1] = hc_bytealign_S (w1[0], w1[1], offset); - w2[0] = hc_bytealign_S (w0[3], w1[0], offset); - w1[3] = hc_bytealign_S (w0[2], w0[3], offset); - w1[2] = hc_bytealign_S (w0[1], w0[2], offset); - w1[1] = hc_bytealign_S (w0[0], w0[1], offset); - w1[0] = hc_bytealign_S ( 0, w0[0], offset); + c1[0] = hc_bytealign_be_S (w7[3], 0, offset); + c0[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c0[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c0[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c0[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + w7[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + w6[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w5[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w4[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w3[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w2[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w1[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[0] = hc_bytealign_be_S ( 0, w0[0], offset); w0[3] = 0; w0[2] = 0; w0[1] = 0; @@ -39377,39 +39065,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 5: - c1[1] = hc_bytealign_S (w7[3], 0, offset); - c1[0] = hc_bytealign_S (w7[2], w7[3], offset); - c0[3] = hc_bytealign_S (w7[1], w7[2], offset); - c0[2] = hc_bytealign_S (w7[0], w7[1], offset); - c0[1] = hc_bytealign_S (w6[3], w7[0], offset); - c0[0] = hc_bytealign_S (w6[2], w6[3], offset); - w7[3] = hc_bytealign_S (w6[1], w6[2], offset); - w7[2] = hc_bytealign_S (w6[0], w6[1], offset); - w7[1] = hc_bytealign_S (w5[3], w6[0], offset); - w7[0] = hc_bytealign_S (w5[2], w5[3], offset); - w6[3] = hc_bytealign_S (w5[1], w5[2], offset); - w6[2] = hc_bytealign_S (w5[0], w5[1], offset); - w6[1] = hc_bytealign_S (w4[3], w5[0], offset); - w6[0] = hc_bytealign_S (w4[2], w4[3], offset); - w5[3] = hc_bytealign_S (w4[1], w4[2], offset); - w5[2] = hc_bytealign_S (w4[0], w4[1], offset); - w5[1] = hc_bytealign_S (w3[3], w4[0], offset); - w5[0] = hc_bytealign_S (w3[2], w3[3], offset); - w4[3] = hc_bytealign_S (w3[1], w3[2], offset); - w4[2] = hc_bytealign_S (w3[0], w3[1], offset); - w4[1] = hc_bytealign_S (w2[3], w3[0], offset); - w4[0] = hc_bytealign_S (w2[2], w2[3], offset); - w3[3] = hc_bytealign_S (w2[1], w2[2], offset); - w3[2] = hc_bytealign_S (w2[0], w2[1], offset); - w3[1] = hc_bytealign_S (w1[3], w2[0], offset); - w3[0] = hc_bytealign_S (w1[2], w1[3], offset); - w2[3] = hc_bytealign_S (w1[1], w1[2], offset); - w2[2] = hc_bytealign_S (w1[0], w1[1], offset); - w2[1] = hc_bytealign_S (w0[3], w1[0], offset); - w2[0] = hc_bytealign_S (w0[2], w0[3], offset); - w1[3] = hc_bytealign_S (w0[1], w0[2], offset); - w1[2] = hc_bytealign_S (w0[0], w0[1], offset); - w1[1] = hc_bytealign_S ( 0, w0[0], offset); + c1[1] = hc_bytealign_be_S (w7[3], 0, offset); + c1[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c0[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c0[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c0[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c0[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + w7[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + w6[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w5[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w4[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w3[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w2[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w1[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[1] = hc_bytealign_be_S ( 0, w0[0], offset); w1[0] = 0; w0[3] = 0; w0[2] = 0; @@ -39419,39 +39107,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 6: - c1[2] = hc_bytealign_S (w7[3], 0, offset); - c1[1] = hc_bytealign_S (w7[2], w7[3], offset); - c1[0] = hc_bytealign_S (w7[1], w7[2], offset); - c0[3] = hc_bytealign_S (w7[0], w7[1], offset); - c0[2] = hc_bytealign_S (w6[3], w7[0], offset); - c0[1] = hc_bytealign_S (w6[2], w6[3], offset); - c0[0] = hc_bytealign_S (w6[1], w6[2], offset); - w7[3] = hc_bytealign_S (w6[0], w6[1], offset); - w7[2] = hc_bytealign_S (w5[3], w6[0], offset); - w7[1] = hc_bytealign_S (w5[2], w5[3], offset); - w7[0] = hc_bytealign_S (w5[1], w5[2], offset); - w6[3] = hc_bytealign_S (w5[0], w5[1], offset); - w6[2] = hc_bytealign_S (w4[3], w5[0], offset); - w6[1] = hc_bytealign_S (w4[2], w4[3], offset); - w6[0] = hc_bytealign_S (w4[1], w4[2], offset); - w5[3] = hc_bytealign_S (w4[0], w4[1], offset); - w5[2] = hc_bytealign_S (w3[3], w4[0], offset); - w5[1] = hc_bytealign_S (w3[2], w3[3], offset); - w5[0] = hc_bytealign_S (w3[1], w3[2], offset); - w4[3] = hc_bytealign_S (w3[0], w3[1], offset); - w4[2] = hc_bytealign_S (w2[3], w3[0], offset); - w4[1] = hc_bytealign_S (w2[2], w2[3], offset); - w4[0] = hc_bytealign_S (w2[1], w2[2], offset); - w3[3] = hc_bytealign_S (w2[0], w2[1], offset); - w3[2] = hc_bytealign_S (w1[3], w2[0], offset); - w3[1] = hc_bytealign_S (w1[2], w1[3], offset); - w3[0] = hc_bytealign_S (w1[1], w1[2], offset); - w2[3] = hc_bytealign_S (w1[0], w1[1], offset); - w2[2] = hc_bytealign_S (w0[3], w1[0], offset); - w2[1] = hc_bytealign_S (w0[2], w0[3], offset); - w2[0] = hc_bytealign_S (w0[1], w0[2], offset); - w1[3] = hc_bytealign_S (w0[0], w0[1], offset); - w1[2] = hc_bytealign_S ( 0, w0[0], offset); + c1[2] = hc_bytealign_be_S (w7[3], 0, offset); + c1[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c1[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c0[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c0[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c0[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c0[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + w7[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + w6[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w5[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w4[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w3[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w2[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w1[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[2] = hc_bytealign_be_S ( 0, w0[0], offset); w1[1] = 0; w1[0] = 0; w0[3] = 0; @@ -39462,39 +39150,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 7: - c1[3] = hc_bytealign_S (w7[3], 0, offset); - c1[2] = hc_bytealign_S (w7[2], w7[3], offset); - c1[1] = hc_bytealign_S (w7[1], w7[2], offset); - c1[0] = hc_bytealign_S (w7[0], w7[1], offset); - c0[3] = hc_bytealign_S (w6[3], w7[0], offset); - c0[2] = hc_bytealign_S (w6[2], w6[3], offset); - c0[1] = hc_bytealign_S (w6[1], w6[2], offset); - c0[0] = hc_bytealign_S (w6[0], w6[1], offset); - w7[3] = hc_bytealign_S (w5[3], w6[0], offset); - w7[2] = hc_bytealign_S (w5[2], w5[3], offset); - w7[1] = hc_bytealign_S (w5[1], w5[2], offset); - w7[0] = hc_bytealign_S (w5[0], w5[1], offset); - w6[3] = hc_bytealign_S (w4[3], w5[0], offset); - w6[2] = hc_bytealign_S (w4[2], w4[3], offset); - w6[1] = hc_bytealign_S (w4[1], w4[2], offset); - w6[0] = hc_bytealign_S (w4[0], w4[1], offset); - w5[3] = hc_bytealign_S (w3[3], w4[0], offset); - w5[2] = hc_bytealign_S (w3[2], w3[3], offset); - w5[1] = hc_bytealign_S (w3[1], w3[2], offset); - w5[0] = hc_bytealign_S (w3[0], w3[1], offset); - w4[3] = hc_bytealign_S (w2[3], w3[0], offset); - w4[2] = hc_bytealign_S (w2[2], w2[3], offset); - w4[1] = hc_bytealign_S (w2[1], w2[2], offset); - w4[0] = hc_bytealign_S (w2[0], w2[1], offset); - w3[3] = hc_bytealign_S (w1[3], w2[0], offset); - w3[2] = hc_bytealign_S (w1[2], w1[3], offset); - w3[1] = hc_bytealign_S (w1[1], w1[2], offset); - w3[0] = hc_bytealign_S (w1[0], w1[1], offset); - w2[3] = hc_bytealign_S (w0[3], w1[0], offset); - w2[2] = hc_bytealign_S (w0[2], w0[3], offset); - w2[1] = hc_bytealign_S (w0[1], w0[2], offset); - w2[0] = hc_bytealign_S (w0[0], w0[1], offset); - w1[3] = hc_bytealign_S ( 0, w0[0], offset); + c1[3] = hc_bytealign_be_S (w7[3], 0, offset); + c1[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c1[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c1[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c0[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c0[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c0[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c0[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + w7[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + w6[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w5[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w4[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w3[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w2[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w1[3] = hc_bytealign_be_S ( 0, w0[0], offset); w1[2] = 0; w1[1] = 0; w1[0] = 0; @@ -39506,39 +39194,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 8: - c2[0] = hc_bytealign_S (w7[3], 0, offset); - c1[3] = hc_bytealign_S (w7[2], w7[3], offset); - c1[2] = hc_bytealign_S (w7[1], w7[2], offset); - c1[1] = hc_bytealign_S (w7[0], w7[1], offset); - c1[0] = hc_bytealign_S (w6[3], w7[0], offset); - c0[3] = hc_bytealign_S (w6[2], w6[3], offset); - c0[2] = hc_bytealign_S (w6[1], w6[2], offset); - c0[1] = hc_bytealign_S (w6[0], w6[1], offset); - c0[0] = hc_bytealign_S (w5[3], w6[0], offset); - w7[3] = hc_bytealign_S (w5[2], w5[3], offset); - w7[2] = hc_bytealign_S (w5[1], w5[2], offset); - w7[1] = hc_bytealign_S (w5[0], w5[1], offset); - w7[0] = hc_bytealign_S (w4[3], w5[0], offset); - w6[3] = hc_bytealign_S (w4[2], w4[3], offset); - w6[2] = hc_bytealign_S (w4[1], w4[2], offset); - w6[1] = hc_bytealign_S (w4[0], w4[1], offset); - w6[0] = hc_bytealign_S (w3[3], w4[0], offset); - w5[3] = hc_bytealign_S (w3[2], w3[3], offset); - w5[2] = hc_bytealign_S (w3[1], w3[2], offset); - w5[1] = hc_bytealign_S (w3[0], w3[1], offset); - w5[0] = hc_bytealign_S (w2[3], w3[0], offset); - w4[3] = hc_bytealign_S (w2[2], w2[3], offset); - w4[2] = hc_bytealign_S (w2[1], w2[2], offset); - w4[1] = hc_bytealign_S (w2[0], w2[1], offset); - w4[0] = hc_bytealign_S (w1[3], w2[0], offset); - w3[3] = hc_bytealign_S (w1[2], w1[3], offset); - w3[2] = hc_bytealign_S (w1[1], w1[2], offset); - w3[1] = hc_bytealign_S (w1[0], w1[1], offset); - w3[0] = hc_bytealign_S (w0[3], w1[0], offset); - w2[3] = hc_bytealign_S (w0[2], w0[3], offset); - w2[2] = hc_bytealign_S (w0[1], w0[2], offset); - w2[1] = hc_bytealign_S (w0[0], w0[1], offset); - w2[0] = hc_bytealign_S ( 0, w0[0], offset); + c2[0] = hc_bytealign_be_S (w7[3], 0, offset); + c1[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c1[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c1[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c1[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c0[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c0[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c0[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c0[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + w7[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w6[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w5[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w4[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w3[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w2[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[0] = hc_bytealign_be_S ( 0, w0[0], offset); w1[3] = 0; w1[2] = 0; w1[1] = 0; @@ -39551,39 +39239,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 9: - c2[1] = hc_bytealign_S (w7[3], 0, offset); - c2[0] = hc_bytealign_S (w7[2], w7[3], offset); - c1[3] = hc_bytealign_S (w7[1], w7[2], offset); - c1[2] = hc_bytealign_S (w7[0], w7[1], offset); - c1[1] = hc_bytealign_S (w6[3], w7[0], offset); - c1[0] = hc_bytealign_S (w6[2], w6[3], offset); - c0[3] = hc_bytealign_S (w6[1], w6[2], offset); - c0[2] = hc_bytealign_S (w6[0], w6[1], offset); - c0[1] = hc_bytealign_S (w5[3], w6[0], offset); - c0[0] = hc_bytealign_S (w5[2], w5[3], offset); - w7[3] = hc_bytealign_S (w5[1], w5[2], offset); - w7[2] = hc_bytealign_S (w5[0], w5[1], offset); - w7[1] = hc_bytealign_S (w4[3], w5[0], offset); - w7[0] = hc_bytealign_S (w4[2], w4[3], offset); - w6[3] = hc_bytealign_S (w4[1], w4[2], offset); - w6[2] = hc_bytealign_S (w4[0], w4[1], offset); - w6[1] = hc_bytealign_S (w3[3], w4[0], offset); - w6[0] = hc_bytealign_S (w3[2], w3[3], offset); - w5[3] = hc_bytealign_S (w3[1], w3[2], offset); - w5[2] = hc_bytealign_S (w3[0], w3[1], offset); - w5[1] = hc_bytealign_S (w2[3], w3[0], offset); - w5[0] = hc_bytealign_S (w2[2], w2[3], offset); - w4[3] = hc_bytealign_S (w2[1], w2[2], offset); - w4[2] = hc_bytealign_S (w2[0], w2[1], offset); - w4[1] = hc_bytealign_S (w1[3], w2[0], offset); - w4[0] = hc_bytealign_S (w1[2], w1[3], offset); - w3[3] = hc_bytealign_S (w1[1], w1[2], offset); - w3[2] = hc_bytealign_S (w1[0], w1[1], offset); - w3[1] = hc_bytealign_S (w0[3], w1[0], offset); - w3[0] = hc_bytealign_S (w0[2], w0[3], offset); - w2[3] = hc_bytealign_S (w0[1], w0[2], offset); - w2[2] = hc_bytealign_S (w0[0], w0[1], offset); - w2[1] = hc_bytealign_S ( 0, w0[0], offset); + c2[1] = hc_bytealign_be_S (w7[3], 0, offset); + c2[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c1[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c1[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c1[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c1[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c0[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c0[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c0[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c0[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + w7[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w6[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w5[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w4[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w3[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w2[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[1] = hc_bytealign_be_S ( 0, w0[0], offset); w2[0] = 0; w1[3] = 0; w1[2] = 0; @@ -39597,39 +39285,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 10: - c2[2] = hc_bytealign_S (w7[3], 0, offset); - c2[1] = hc_bytealign_S (w7[2], w7[3], offset); - c2[0] = hc_bytealign_S (w7[1], w7[2], offset); - c1[3] = hc_bytealign_S (w7[0], w7[1], offset); - c1[2] = hc_bytealign_S (w6[3], w7[0], offset); - c1[1] = hc_bytealign_S (w6[2], w6[3], offset); - c1[0] = hc_bytealign_S (w6[1], w6[2], offset); - c0[3] = hc_bytealign_S (w6[0], w6[1], offset); - c0[2] = hc_bytealign_S (w5[3], w6[0], offset); - c0[1] = hc_bytealign_S (w5[2], w5[3], offset); - c0[0] = hc_bytealign_S (w5[1], w5[2], offset); - w7[3] = hc_bytealign_S (w5[0], w5[1], offset); - w7[2] = hc_bytealign_S (w4[3], w5[0], offset); - w7[1] = hc_bytealign_S (w4[2], w4[3], offset); - w7[0] = hc_bytealign_S (w4[1], w4[2], offset); - w6[3] = hc_bytealign_S (w4[0], w4[1], offset); - w6[2] = hc_bytealign_S (w3[3], w4[0], offset); - w6[1] = hc_bytealign_S (w3[2], w3[3], offset); - w6[0] = hc_bytealign_S (w3[1], w3[2], offset); - w5[3] = hc_bytealign_S (w3[0], w3[1], offset); - w5[2] = hc_bytealign_S (w2[3], w3[0], offset); - w5[1] = hc_bytealign_S (w2[2], w2[3], offset); - w5[0] = hc_bytealign_S (w2[1], w2[2], offset); - w4[3] = hc_bytealign_S (w2[0], w2[1], offset); - w4[2] = hc_bytealign_S (w1[3], w2[0], offset); - w4[1] = hc_bytealign_S (w1[2], w1[3], offset); - w4[0] = hc_bytealign_S (w1[1], w1[2], offset); - w3[3] = hc_bytealign_S (w1[0], w1[1], offset); - w3[2] = hc_bytealign_S (w0[3], w1[0], offset); - w3[1] = hc_bytealign_S (w0[2], w0[3], offset); - w3[0] = hc_bytealign_S (w0[1], w0[2], offset); - w2[3] = hc_bytealign_S (w0[0], w0[1], offset); - w2[2] = hc_bytealign_S ( 0, w0[0], offset); + c2[2] = hc_bytealign_be_S (w7[3], 0, offset); + c2[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c2[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c1[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c1[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c1[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c1[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c0[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c0[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c0[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c0[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + w7[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w6[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w5[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w4[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w3[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w2[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[2] = hc_bytealign_be_S ( 0, w0[0], offset); w2[1] = 0; w2[0] = 0; w1[3] = 0; @@ -39644,39 +39332,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 11: - c2[3] = hc_bytealign_S (w7[3], 0, offset); - c2[2] = hc_bytealign_S (w7[2], w7[3], offset); - c2[1] = hc_bytealign_S (w7[1], w7[2], offset); - c2[0] = hc_bytealign_S (w7[0], w7[1], offset); - c1[3] = hc_bytealign_S (w6[3], w7[0], offset); - c1[2] = hc_bytealign_S (w6[2], w6[3], offset); - c1[1] = hc_bytealign_S (w6[1], w6[2], offset); - c1[0] = hc_bytealign_S (w6[0], w6[1], offset); - c0[3] = hc_bytealign_S (w5[3], w6[0], offset); - c0[2] = hc_bytealign_S (w5[2], w5[3], offset); - c0[1] = hc_bytealign_S (w5[1], w5[2], offset); - c0[0] = hc_bytealign_S (w5[0], w5[1], offset); - w7[3] = hc_bytealign_S (w4[3], w5[0], offset); - w7[2] = hc_bytealign_S (w4[2], w4[3], offset); - w7[1] = hc_bytealign_S (w4[1], w4[2], offset); - w7[0] = hc_bytealign_S (w4[0], w4[1], offset); - w6[3] = hc_bytealign_S (w3[3], w4[0], offset); - w6[2] = hc_bytealign_S (w3[2], w3[3], offset); - w6[1] = hc_bytealign_S (w3[1], w3[2], offset); - w6[0] = hc_bytealign_S (w3[0], w3[1], offset); - w5[3] = hc_bytealign_S (w2[3], w3[0], offset); - w5[2] = hc_bytealign_S (w2[2], w2[3], offset); - w5[1] = hc_bytealign_S (w2[1], w2[2], offset); - w5[0] = hc_bytealign_S (w2[0], w2[1], offset); - w4[3] = hc_bytealign_S (w1[3], w2[0], offset); - w4[2] = hc_bytealign_S (w1[2], w1[3], offset); - w4[1] = hc_bytealign_S (w1[1], w1[2], offset); - w4[0] = hc_bytealign_S (w1[0], w1[1], offset); - w3[3] = hc_bytealign_S (w0[3], w1[0], offset); - w3[2] = hc_bytealign_S (w0[2], w0[3], offset); - w3[1] = hc_bytealign_S (w0[1], w0[2], offset); - w3[0] = hc_bytealign_S (w0[0], w0[1], offset); - w2[3] = hc_bytealign_S ( 0, w0[0], offset); + c2[3] = hc_bytealign_be_S (w7[3], 0, offset); + c2[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c2[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c2[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c1[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c1[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c1[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c1[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c0[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c0[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c0[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c0[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + w7[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w6[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w5[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w4[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w3[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w2[3] = hc_bytealign_be_S ( 0, w0[0], offset); w2[2] = 0; w2[1] = 0; w2[0] = 0; @@ -39692,39 +39380,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 12: - c3[0] = hc_bytealign_S (w7[3], 0, offset); - c2[3] = hc_bytealign_S (w7[2], w7[3], offset); - c2[2] = hc_bytealign_S (w7[1], w7[2], offset); - c2[1] = hc_bytealign_S (w7[0], w7[1], offset); - c2[0] = hc_bytealign_S (w6[3], w7[0], offset); - c1[3] = hc_bytealign_S (w6[2], w6[3], offset); - c1[2] = hc_bytealign_S (w6[1], w6[2], offset); - c1[1] = hc_bytealign_S (w6[0], w6[1], offset); - c1[0] = hc_bytealign_S (w5[3], w6[0], offset); - c0[3] = hc_bytealign_S (w5[2], w5[3], offset); - c0[2] = hc_bytealign_S (w5[1], w5[2], offset); - c0[1] = hc_bytealign_S (w5[0], w5[1], offset); - c0[0] = hc_bytealign_S (w4[3], w5[0], offset); - w7[3] = hc_bytealign_S (w4[2], w4[3], offset); - w7[2] = hc_bytealign_S (w4[1], w4[2], offset); - w7[1] = hc_bytealign_S (w4[0], w4[1], offset); - w7[0] = hc_bytealign_S (w3[3], w4[0], offset); - w6[3] = hc_bytealign_S (w3[2], w3[3], offset); - w6[2] = hc_bytealign_S (w3[1], w3[2], offset); - w6[1] = hc_bytealign_S (w3[0], w3[1], offset); - w6[0] = hc_bytealign_S (w2[3], w3[0], offset); - w5[3] = hc_bytealign_S (w2[2], w2[3], offset); - w5[2] = hc_bytealign_S (w2[1], w2[2], offset); - w5[1] = hc_bytealign_S (w2[0], w2[1], offset); - w5[0] = hc_bytealign_S (w1[3], w2[0], offset); - w4[3] = hc_bytealign_S (w1[2], w1[3], offset); - w4[2] = hc_bytealign_S (w1[1], w1[2], offset); - w4[1] = hc_bytealign_S (w1[0], w1[1], offset); - w4[0] = hc_bytealign_S (w0[3], w1[0], offset); - w3[3] = hc_bytealign_S (w0[2], w0[3], offset); - w3[2] = hc_bytealign_S (w0[1], w0[2], offset); - w3[1] = hc_bytealign_S (w0[0], w0[1], offset); - w3[0] = hc_bytealign_S ( 0, w0[0], offset); + c3[0] = hc_bytealign_be_S (w7[3], 0, offset); + c2[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c2[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c2[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c2[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c1[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c1[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c1[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c1[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + c0[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + c0[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + c0[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + c0[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + w7[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w6[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w5[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w4[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w3[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[0] = hc_bytealign_be_S ( 0, w0[0], offset); w2[3] = 0; w2[2] = 0; w2[1] = 0; @@ -39741,39 +39429,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 13: - c3[1] = hc_bytealign_S (w7[3], 0, offset); - c3[0] = hc_bytealign_S (w7[2], w7[3], offset); - c2[3] = hc_bytealign_S (w7[1], w7[2], offset); - c2[2] = hc_bytealign_S (w7[0], w7[1], offset); - c2[1] = hc_bytealign_S (w6[3], w7[0], offset); - c2[0] = hc_bytealign_S (w6[2], w6[3], offset); - c1[3] = hc_bytealign_S (w6[1], w6[2], offset); - c1[2] = hc_bytealign_S (w6[0], w6[1], offset); - c1[1] = hc_bytealign_S (w5[3], w6[0], offset); - c1[0] = hc_bytealign_S (w5[2], w5[3], offset); - c0[3] = hc_bytealign_S (w5[1], w5[2], offset); - c0[2] = hc_bytealign_S (w5[0], w5[1], offset); - c0[1] = hc_bytealign_S (w4[3], w5[0], offset); - c0[0] = hc_bytealign_S (w4[2], w4[3], offset); - w7[3] = hc_bytealign_S (w4[1], w4[2], offset); - w7[2] = hc_bytealign_S (w4[0], w4[1], offset); - w7[1] = hc_bytealign_S (w3[3], w4[0], offset); - w7[0] = hc_bytealign_S (w3[2], w3[3], offset); - w6[3] = hc_bytealign_S (w3[1], w3[2], offset); - w6[2] = hc_bytealign_S (w3[0], w3[1], offset); - w6[1] = hc_bytealign_S (w2[3], w3[0], offset); - w6[0] = hc_bytealign_S (w2[2], w2[3], offset); - w5[3] = hc_bytealign_S (w2[1], w2[2], offset); - w5[2] = hc_bytealign_S (w2[0], w2[1], offset); - w5[1] = hc_bytealign_S (w1[3], w2[0], offset); - w5[0] = hc_bytealign_S (w1[2], w1[3], offset); - w4[3] = hc_bytealign_S (w1[1], w1[2], offset); - w4[2] = hc_bytealign_S (w1[0], w1[1], offset); - w4[1] = hc_bytealign_S (w0[3], w1[0], offset); - w4[0] = hc_bytealign_S (w0[2], w0[3], offset); - w3[3] = hc_bytealign_S (w0[1], w0[2], offset); - w3[2] = hc_bytealign_S (w0[0], w0[1], offset); - w3[1] = hc_bytealign_S ( 0, w0[0], offset); + c3[1] = hc_bytealign_be_S (w7[3], 0, offset); + c3[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c2[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c2[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c2[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c2[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c1[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c1[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c1[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c1[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + c0[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + c0[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + c0[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + c0[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + w7[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w6[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w5[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w4[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w3[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[1] = hc_bytealign_be_S ( 0, w0[0], offset); w3[0] = 0; w2[3] = 0; w2[2] = 0; @@ -39791,39 +39479,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 14: - c3[2] = hc_bytealign_S (w7[3], 0, offset); - c3[1] = hc_bytealign_S (w7[2], w7[3], offset); - c3[0] = hc_bytealign_S (w7[1], w7[2], offset); - c2[3] = hc_bytealign_S (w7[0], w7[1], offset); - c2[2] = hc_bytealign_S (w6[3], w7[0], offset); - c2[1] = hc_bytealign_S (w6[2], w6[3], offset); - c2[0] = hc_bytealign_S (w6[1], w6[2], offset); - c1[3] = hc_bytealign_S (w6[0], w6[1], offset); - c1[2] = hc_bytealign_S (w5[3], w6[0], offset); - c1[1] = hc_bytealign_S (w5[2], w5[3], offset); - c1[0] = hc_bytealign_S (w5[1], w5[2], offset); - c0[3] = hc_bytealign_S (w5[0], w5[1], offset); - c0[2] = hc_bytealign_S (w4[3], w5[0], offset); - c0[1] = hc_bytealign_S (w4[2], w4[3], offset); - c0[0] = hc_bytealign_S (w4[1], w4[2], offset); - w7[3] = hc_bytealign_S (w4[0], w4[1], offset); - w7[2] = hc_bytealign_S (w3[3], w4[0], offset); - w7[1] = hc_bytealign_S (w3[2], w3[3], offset); - w7[0] = hc_bytealign_S (w3[1], w3[2], offset); - w6[3] = hc_bytealign_S (w3[0], w3[1], offset); - w6[2] = hc_bytealign_S (w2[3], w3[0], offset); - w6[1] = hc_bytealign_S (w2[2], w2[3], offset); - w6[0] = hc_bytealign_S (w2[1], w2[2], offset); - w5[3] = hc_bytealign_S (w2[0], w2[1], offset); - w5[2] = hc_bytealign_S (w1[3], w2[0], offset); - w5[1] = hc_bytealign_S (w1[2], w1[3], offset); - w5[0] = hc_bytealign_S (w1[1], w1[2], offset); - w4[3] = hc_bytealign_S (w1[0], w1[1], offset); - w4[2] = hc_bytealign_S (w0[3], w1[0], offset); - w4[1] = hc_bytealign_S (w0[2], w0[3], offset); - w4[0] = hc_bytealign_S (w0[1], w0[2], offset); - w3[3] = hc_bytealign_S (w0[0], w0[1], offset); - w3[2] = hc_bytealign_S ( 0, w0[0], offset); + c3[2] = hc_bytealign_be_S (w7[3], 0, offset); + c3[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c3[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c2[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c2[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c2[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c2[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c1[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c1[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c1[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c1[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + c0[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + c0[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + c0[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + c0[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + w7[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w6[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w5[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w4[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w3[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[2] = hc_bytealign_be_S ( 0, w0[0], offset); w3[1] = 0; w3[0] = 0; w2[3] = 0; @@ -39842,39 +39530,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 15: - c3[3] = hc_bytealign_S (w7[3], 0, offset); - c3[2] = hc_bytealign_S (w7[2], w7[3], offset); - c3[1] = hc_bytealign_S (w7[1], w7[2], offset); - c3[0] = hc_bytealign_S (w7[0], w7[1], offset); - c2[3] = hc_bytealign_S (w6[3], w7[0], offset); - c2[2] = hc_bytealign_S (w6[2], w6[3], offset); - c2[1] = hc_bytealign_S (w6[1], w6[2], offset); - c2[0] = hc_bytealign_S (w6[0], w6[1], offset); - c1[3] = hc_bytealign_S (w5[3], w6[0], offset); - c1[2] = hc_bytealign_S (w5[2], w5[3], offset); - c1[1] = hc_bytealign_S (w5[1], w5[2], offset); - c1[0] = hc_bytealign_S (w5[0], w5[1], offset); - c0[3] = hc_bytealign_S (w4[3], w5[0], offset); - c0[2] = hc_bytealign_S (w4[2], w4[3], offset); - c0[1] = hc_bytealign_S (w4[1], w4[2], offset); - c0[0] = hc_bytealign_S (w4[0], w4[1], offset); - w7[3] = hc_bytealign_S (w3[3], w4[0], offset); - w7[2] = hc_bytealign_S (w3[2], w3[3], offset); - w7[1] = hc_bytealign_S (w3[1], w3[2], offset); - w7[0] = hc_bytealign_S (w3[0], w3[1], offset); - w6[3] = hc_bytealign_S (w2[3], w3[0], offset); - w6[2] = hc_bytealign_S (w2[2], w2[3], offset); - w6[1] = hc_bytealign_S (w2[1], w2[2], offset); - w6[0] = hc_bytealign_S (w2[0], w2[1], offset); - w5[3] = hc_bytealign_S (w1[3], w2[0], offset); - w5[2] = hc_bytealign_S (w1[2], w1[3], offset); - w5[1] = hc_bytealign_S (w1[1], w1[2], offset); - w5[0] = hc_bytealign_S (w1[0], w1[1], offset); - w4[3] = hc_bytealign_S (w0[3], w1[0], offset); - w4[2] = hc_bytealign_S (w0[2], w0[3], offset); - w4[1] = hc_bytealign_S (w0[1], w0[2], offset); - w4[0] = hc_bytealign_S (w0[0], w0[1], offset); - w3[3] = hc_bytealign_S ( 0, w0[0], offset); + c3[3] = hc_bytealign_be_S (w7[3], 0, offset); + c3[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c3[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c3[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c2[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c2[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c2[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c2[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c1[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c1[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c1[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c1[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + c0[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + c0[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + c0[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + c0[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + w7[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w6[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w5[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w4[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w3[3] = hc_bytealign_be_S ( 0, w0[0], offset); w3[2] = 0; w3[1] = 0; w3[0] = 0; @@ -39894,39 +39582,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 16: - c4[0] = hc_bytealign_S (w7[3], 0, offset); - c3[3] = hc_bytealign_S (w7[2], w7[3], offset); - c3[2] = hc_bytealign_S (w7[1], w7[2], offset); - c3[1] = hc_bytealign_S (w7[0], w7[1], offset); - c3[0] = hc_bytealign_S (w6[3], w7[0], offset); - c2[3] = hc_bytealign_S (w6[2], w6[3], offset); - c2[2] = hc_bytealign_S (w6[1], w6[2], offset); - c2[1] = hc_bytealign_S (w6[0], w6[1], offset); - c2[0] = hc_bytealign_S (w5[3], w6[0], offset); - c1[3] = hc_bytealign_S (w5[2], w5[3], offset); - c1[2] = hc_bytealign_S (w5[1], w5[2], offset); - c1[1] = hc_bytealign_S (w5[0], w5[1], offset); - c1[0] = hc_bytealign_S (w4[3], w5[0], offset); - c0[3] = hc_bytealign_S (w4[2], w4[3], offset); - c0[2] = hc_bytealign_S (w4[1], w4[2], offset); - c0[1] = hc_bytealign_S (w4[0], w4[1], offset); - c0[0] = hc_bytealign_S (w3[3], w4[0], offset); - w7[3] = hc_bytealign_S (w3[2], w3[3], offset); - w7[2] = hc_bytealign_S (w3[1], w3[2], offset); - w7[1] = hc_bytealign_S (w3[0], w3[1], offset); - w7[0] = hc_bytealign_S (w2[3], w3[0], offset); - w6[3] = hc_bytealign_S (w2[2], w2[3], offset); - w6[2] = hc_bytealign_S (w2[1], w2[2], offset); - w6[1] = hc_bytealign_S (w2[0], w2[1], offset); - w6[0] = hc_bytealign_S (w1[3], w2[0], offset); - w5[3] = hc_bytealign_S (w1[2], w1[3], offset); - w5[2] = hc_bytealign_S (w1[1], w1[2], offset); - w5[1] = hc_bytealign_S (w1[0], w1[1], offset); - w5[0] = hc_bytealign_S (w0[3], w1[0], offset); - w4[3] = hc_bytealign_S (w0[2], w0[3], offset); - w4[2] = hc_bytealign_S (w0[1], w0[2], offset); - w4[1] = hc_bytealign_S (w0[0], w0[1], offset); - w4[0] = hc_bytealign_S ( 0, w0[0], offset); + c4[0] = hc_bytealign_be_S (w7[3], 0, offset); + c3[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c3[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c3[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c3[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c2[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c2[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c2[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c2[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + c1[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + c1[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + c1[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + c1[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + c0[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + c0[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + c0[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + c0[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + w7[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w6[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w5[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w4[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[0] = hc_bytealign_be_S ( 0, w0[0], offset); w3[3] = 0; w3[2] = 0; w3[1] = 0; @@ -39947,39 +39635,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 17: - c4[1] = hc_bytealign_S (w7[3], 0, offset); - c4[0] = hc_bytealign_S (w7[2], w7[3], offset); - c3[3] = hc_bytealign_S (w7[1], w7[2], offset); - c3[2] = hc_bytealign_S (w7[0], w7[1], offset); - c3[1] = hc_bytealign_S (w6[3], w7[0], offset); - c3[0] = hc_bytealign_S (w6[2], w6[3], offset); - c2[3] = hc_bytealign_S (w6[1], w6[2], offset); - c2[2] = hc_bytealign_S (w6[0], w6[1], offset); - c2[1] = hc_bytealign_S (w5[3], w6[0], offset); - c2[0] = hc_bytealign_S (w5[2], w5[3], offset); - c1[3] = hc_bytealign_S (w5[1], w5[2], offset); - c1[2] = hc_bytealign_S (w5[0], w5[1], offset); - c1[1] = hc_bytealign_S (w4[3], w5[0], offset); - c1[0] = hc_bytealign_S (w4[2], w4[3], offset); - c0[3] = hc_bytealign_S (w4[1], w4[2], offset); - c0[2] = hc_bytealign_S (w4[0], w4[1], offset); - c0[1] = hc_bytealign_S (w3[3], w4[0], offset); - c0[0] = hc_bytealign_S (w3[2], w3[3], offset); - w7[3] = hc_bytealign_S (w3[1], w3[2], offset); - w7[2] = hc_bytealign_S (w3[0], w3[1], offset); - w7[1] = hc_bytealign_S (w2[3], w3[0], offset); - w7[0] = hc_bytealign_S (w2[2], w2[3], offset); - w6[3] = hc_bytealign_S (w2[1], w2[2], offset); - w6[2] = hc_bytealign_S (w2[0], w2[1], offset); - w6[1] = hc_bytealign_S (w1[3], w2[0], offset); - w6[0] = hc_bytealign_S (w1[2], w1[3], offset); - w5[3] = hc_bytealign_S (w1[1], w1[2], offset); - w5[2] = hc_bytealign_S (w1[0], w1[1], offset); - w5[1] = hc_bytealign_S (w0[3], w1[0], offset); - w5[0] = hc_bytealign_S (w0[2], w0[3], offset); - w4[3] = hc_bytealign_S (w0[1], w0[2], offset); - w4[2] = hc_bytealign_S (w0[0], w0[1], offset); - w4[1] = hc_bytealign_S ( 0, w0[0], offset); + c4[1] = hc_bytealign_be_S (w7[3], 0, offset); + c4[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c3[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c3[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c3[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c3[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c2[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c2[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c2[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c2[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + c1[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + c1[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + c1[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + c1[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + c0[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + c0[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + c0[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + c0[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + w7[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w6[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w5[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w4[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[1] = hc_bytealign_be_S ( 0, w0[0], offset); w4[0] = 0; w3[3] = 0; w3[2] = 0; @@ -40001,39 +39689,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 18: - c4[2] = hc_bytealign_S (w7[3], 0, offset); - c4[1] = hc_bytealign_S (w7[2], w7[3], offset); - c4[0] = hc_bytealign_S (w7[1], w7[2], offset); - c3[3] = hc_bytealign_S (w7[0], w7[1], offset); - c3[2] = hc_bytealign_S (w6[3], w7[0], offset); - c3[1] = hc_bytealign_S (w6[2], w6[3], offset); - c3[0] = hc_bytealign_S (w6[1], w6[2], offset); - c2[3] = hc_bytealign_S (w6[0], w6[1], offset); - c2[2] = hc_bytealign_S (w5[3], w6[0], offset); - c2[1] = hc_bytealign_S (w5[2], w5[3], offset); - c2[0] = hc_bytealign_S (w5[1], w5[2], offset); - c1[3] = hc_bytealign_S (w5[0], w5[1], offset); - c1[2] = hc_bytealign_S (w4[3], w5[0], offset); - c1[1] = hc_bytealign_S (w4[2], w4[3], offset); - c1[0] = hc_bytealign_S (w4[1], w4[2], offset); - c0[3] = hc_bytealign_S (w4[0], w4[1], offset); - c0[2] = hc_bytealign_S (w3[3], w4[0], offset); - c0[1] = hc_bytealign_S (w3[2], w3[3], offset); - c0[0] = hc_bytealign_S (w3[1], w3[2], offset); - w7[3] = hc_bytealign_S (w3[0], w3[1], offset); - w7[2] = hc_bytealign_S (w2[3], w3[0], offset); - w7[1] = hc_bytealign_S (w2[2], w2[3], offset); - w7[0] = hc_bytealign_S (w2[1], w2[2], offset); - w6[3] = hc_bytealign_S (w2[0], w2[1], offset); - w6[2] = hc_bytealign_S (w1[3], w2[0], offset); - w6[1] = hc_bytealign_S (w1[2], w1[3], offset); - w6[0] = hc_bytealign_S (w1[1], w1[2], offset); - w5[3] = hc_bytealign_S (w1[0], w1[1], offset); - w5[2] = hc_bytealign_S (w0[3], w1[0], offset); - w5[1] = hc_bytealign_S (w0[2], w0[3], offset); - w5[0] = hc_bytealign_S (w0[1], w0[2], offset); - w4[3] = hc_bytealign_S (w0[0], w0[1], offset); - w4[2] = hc_bytealign_S ( 0, w0[0], offset); + c4[2] = hc_bytealign_be_S (w7[3], 0, offset); + c4[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c4[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c3[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c3[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c3[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c3[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c2[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c2[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c2[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c2[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + c1[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + c1[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + c1[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + c1[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + c0[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + c0[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + c0[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + w7[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w6[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w5[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w4[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[2] = hc_bytealign_be_S ( 0, w0[0], offset); w4[1] = 0; w4[0] = 0; w3[3] = 0; @@ -40056,39 +39744,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 19: - c4[3] = hc_bytealign_S (w7[3], 0, offset); - c4[2] = hc_bytealign_S (w7[2], w7[3], offset); - c4[1] = hc_bytealign_S (w7[1], w7[2], offset); - c4[0] = hc_bytealign_S (w7[0], w7[1], offset); - c3[3] = hc_bytealign_S (w6[3], w7[0], offset); - c3[2] = hc_bytealign_S (w6[2], w6[3], offset); - c3[1] = hc_bytealign_S (w6[1], w6[2], offset); - c3[0] = hc_bytealign_S (w6[0], w6[1], offset); - c2[3] = hc_bytealign_S (w5[3], w6[0], offset); - c2[2] = hc_bytealign_S (w5[2], w5[3], offset); - c2[1] = hc_bytealign_S (w5[1], w5[2], offset); - c2[0] = hc_bytealign_S (w5[0], w5[1], offset); - c1[3] = hc_bytealign_S (w4[3], w5[0], offset); - c1[2] = hc_bytealign_S (w4[2], w4[3], offset); - c1[1] = hc_bytealign_S (w4[1], w4[2], offset); - c1[0] = hc_bytealign_S (w4[0], w4[1], offset); - c0[3] = hc_bytealign_S (w3[3], w4[0], offset); - c0[2] = hc_bytealign_S (w3[2], w3[3], offset); - c0[1] = hc_bytealign_S (w3[1], w3[2], offset); - c0[0] = hc_bytealign_S (w3[0], w3[1], offset); - w7[3] = hc_bytealign_S (w2[3], w3[0], offset); - w7[2] = hc_bytealign_S (w2[2], w2[3], offset); - w7[1] = hc_bytealign_S (w2[1], w2[2], offset); - w7[0] = hc_bytealign_S (w2[0], w2[1], offset); - w6[3] = hc_bytealign_S (w1[3], w2[0], offset); - w6[2] = hc_bytealign_S (w1[2], w1[3], offset); - w6[1] = hc_bytealign_S (w1[1], w1[2], offset); - w6[0] = hc_bytealign_S (w1[0], w1[1], offset); - w5[3] = hc_bytealign_S (w0[3], w1[0], offset); - w5[2] = hc_bytealign_S (w0[2], w0[3], offset); - w5[1] = hc_bytealign_S (w0[1], w0[2], offset); - w5[0] = hc_bytealign_S (w0[0], w0[1], offset); - w4[3] = hc_bytealign_S ( 0, w0[0], offset); + c4[3] = hc_bytealign_be_S (w7[3], 0, offset); + c4[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c4[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c4[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c3[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c3[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c3[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c3[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c2[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c2[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c2[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c2[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + c1[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + c1[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + c1[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + c1[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + c0[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + c0[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + w7[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w6[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w5[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w4[3] = hc_bytealign_be_S ( 0, w0[0], offset); w4[2] = 0; w4[1] = 0; w4[0] = 0; @@ -40112,39 +39800,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 20: - c5[0] = hc_bytealign_S (w7[3], 0, offset); - c4[3] = hc_bytealign_S (w7[2], w7[3], offset); - c4[2] = hc_bytealign_S (w7[1], w7[2], offset); - c4[1] = hc_bytealign_S (w7[0], w7[1], offset); - c4[0] = hc_bytealign_S (w6[3], w7[0], offset); - c3[3] = hc_bytealign_S (w6[2], w6[3], offset); - c3[2] = hc_bytealign_S (w6[1], w6[2], offset); - c3[1] = hc_bytealign_S (w6[0], w6[1], offset); - c3[0] = hc_bytealign_S (w5[3], w6[0], offset); - c2[3] = hc_bytealign_S (w5[2], w5[3], offset); - c2[2] = hc_bytealign_S (w5[1], w5[2], offset); - c2[1] = hc_bytealign_S (w5[0], w5[1], offset); - c2[0] = hc_bytealign_S (w4[3], w5[0], offset); - c1[3] = hc_bytealign_S (w4[2], w4[3], offset); - c1[2] = hc_bytealign_S (w4[1], w4[2], offset); - c1[1] = hc_bytealign_S (w4[0], w4[1], offset); - c1[0] = hc_bytealign_S (w3[3], w4[0], offset); - c0[3] = hc_bytealign_S (w3[2], w3[3], offset); - c0[2] = hc_bytealign_S (w3[1], w3[2], offset); - c0[1] = hc_bytealign_S (w3[0], w3[1], offset); - c0[0] = hc_bytealign_S (w2[3], w3[0], offset); - w7[3] = hc_bytealign_S (w2[2], w2[3], offset); - w7[2] = hc_bytealign_S (w2[1], w2[2], offset); - w7[1] = hc_bytealign_S (w2[0], w2[1], offset); - w7[0] = hc_bytealign_S (w1[3], w2[0], offset); - w6[3] = hc_bytealign_S (w1[2], w1[3], offset); - w6[2] = hc_bytealign_S (w1[1], w1[2], offset); - w6[1] = hc_bytealign_S (w1[0], w1[1], offset); - w6[0] = hc_bytealign_S (w0[3], w1[0], offset); - w5[3] = hc_bytealign_S (w0[2], w0[3], offset); - w5[2] = hc_bytealign_S (w0[1], w0[2], offset); - w5[1] = hc_bytealign_S (w0[0], w0[1], offset); - w5[0] = hc_bytealign_S ( 0, w0[0], offset); + c5[0] = hc_bytealign_be_S (w7[3], 0, offset); + c4[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c4[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c4[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c4[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c3[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c3[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c3[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c3[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + c2[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + c2[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + c2[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + c2[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + c1[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + c1[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + c1[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + c1[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + c0[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + w7[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w6[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w5[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[0] = hc_bytealign_be_S ( 0, w0[0], offset); w4[3] = 0; w4[2] = 0; w4[1] = 0; @@ -40169,39 +39857,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 21: - c5[1] = hc_bytealign_S (w7[3], 0, offset); - c5[0] = hc_bytealign_S (w7[2], w7[3], offset); - c4[3] = hc_bytealign_S (w7[1], w7[2], offset); - c4[2] = hc_bytealign_S (w7[0], w7[1], offset); - c4[1] = hc_bytealign_S (w6[3], w7[0], offset); - c4[0] = hc_bytealign_S (w6[2], w6[3], offset); - c3[3] = hc_bytealign_S (w6[1], w6[2], offset); - c3[2] = hc_bytealign_S (w6[0], w6[1], offset); - c3[1] = hc_bytealign_S (w5[3], w6[0], offset); - c3[0] = hc_bytealign_S (w5[2], w5[3], offset); - c2[3] = hc_bytealign_S (w5[1], w5[2], offset); - c2[2] = hc_bytealign_S (w5[0], w5[1], offset); - c2[1] = hc_bytealign_S (w4[3], w5[0], offset); - c2[0] = hc_bytealign_S (w4[2], w4[3], offset); - c1[3] = hc_bytealign_S (w4[1], w4[2], offset); - c1[2] = hc_bytealign_S (w4[0], w4[1], offset); - c1[1] = hc_bytealign_S (w3[3], w4[0], offset); - c1[0] = hc_bytealign_S (w3[2], w3[3], offset); - c0[3] = hc_bytealign_S (w3[1], w3[2], offset); - c0[2] = hc_bytealign_S (w3[0], w3[1], offset); - c0[1] = hc_bytealign_S (w2[3], w3[0], offset); - c0[0] = hc_bytealign_S (w2[2], w2[3], offset); - w7[3] = hc_bytealign_S (w2[1], w2[2], offset); - w7[2] = hc_bytealign_S (w2[0], w2[1], offset); - w7[1] = hc_bytealign_S (w1[3], w2[0], offset); - w7[0] = hc_bytealign_S (w1[2], w1[3], offset); - w6[3] = hc_bytealign_S (w1[1], w1[2], offset); - w6[2] = hc_bytealign_S (w1[0], w1[1], offset); - w6[1] = hc_bytealign_S (w0[3], w1[0], offset); - w6[0] = hc_bytealign_S (w0[2], w0[3], offset); - w5[3] = hc_bytealign_S (w0[1], w0[2], offset); - w5[2] = hc_bytealign_S (w0[0], w0[1], offset); - w5[1] = hc_bytealign_S ( 0, w0[0], offset); + c5[1] = hc_bytealign_be_S (w7[3], 0, offset); + c5[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c4[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c4[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c4[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c4[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c3[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c3[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c3[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c3[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + c2[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + c2[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + c2[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + c2[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + c1[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + c1[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + c1[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + c1[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c0[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + w7[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w6[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w5[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[1] = hc_bytealign_be_S ( 0, w0[0], offset); w5[0] = 0; w4[3] = 0; w4[2] = 0; @@ -40227,39 +39915,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 22: - c5[2] = hc_bytealign_S (w7[3], 0, offset); - c5[1] = hc_bytealign_S (w7[2], w7[3], offset); - c5[0] = hc_bytealign_S (w7[1], w7[2], offset); - c4[3] = hc_bytealign_S (w7[0], w7[1], offset); - c4[2] = hc_bytealign_S (w6[3], w7[0], offset); - c4[1] = hc_bytealign_S (w6[2], w6[3], offset); - c4[0] = hc_bytealign_S (w6[1], w6[2], offset); - c3[3] = hc_bytealign_S (w6[0], w6[1], offset); - c3[2] = hc_bytealign_S (w5[3], w6[0], offset); - c3[1] = hc_bytealign_S (w5[2], w5[3], offset); - c3[0] = hc_bytealign_S (w5[1], w5[2], offset); - c2[3] = hc_bytealign_S (w5[0], w5[1], offset); - c2[2] = hc_bytealign_S (w4[3], w5[0], offset); - c2[1] = hc_bytealign_S (w4[2], w4[3], offset); - c2[0] = hc_bytealign_S (w4[1], w4[2], offset); - c1[3] = hc_bytealign_S (w4[0], w4[1], offset); - c1[2] = hc_bytealign_S (w3[3], w4[0], offset); - c1[1] = hc_bytealign_S (w3[2], w3[3], offset); - c1[0] = hc_bytealign_S (w3[1], w3[2], offset); - c0[3] = hc_bytealign_S (w3[0], w3[1], offset); - c0[2] = hc_bytealign_S (w2[3], w3[0], offset); - c0[1] = hc_bytealign_S (w2[2], w2[3], offset); - c0[0] = hc_bytealign_S (w2[1], w2[2], offset); - w7[3] = hc_bytealign_S (w2[0], w2[1], offset); - w7[2] = hc_bytealign_S (w1[3], w2[0], offset); - w7[1] = hc_bytealign_S (w1[2], w1[3], offset); - w7[0] = hc_bytealign_S (w1[1], w1[2], offset); - w6[3] = hc_bytealign_S (w1[0], w1[1], offset); - w6[2] = hc_bytealign_S (w0[3], w1[0], offset); - w6[1] = hc_bytealign_S (w0[2], w0[3], offset); - w6[0] = hc_bytealign_S (w0[1], w0[2], offset); - w5[3] = hc_bytealign_S (w0[0], w0[1], offset); - w5[2] = hc_bytealign_S ( 0, w0[0], offset); + c5[2] = hc_bytealign_be_S (w7[3], 0, offset); + c5[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c5[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c4[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c4[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c4[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c4[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c3[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c3[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c3[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c3[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + c2[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + c2[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + c2[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + c2[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + c1[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + c1[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + c1[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c0[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + w7[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w6[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w5[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[2] = hc_bytealign_be_S ( 0, w0[0], offset); w5[1] = 0; w5[0] = 0; w4[3] = 0; @@ -40286,39 +39974,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 23: - c5[3] = hc_bytealign_S (w7[3], 0, offset); - c5[2] = hc_bytealign_S (w7[2], w7[3], offset); - c5[1] = hc_bytealign_S (w7[1], w7[2], offset); - c5[0] = hc_bytealign_S (w7[0], w7[1], offset); - c4[3] = hc_bytealign_S (w6[3], w7[0], offset); - c4[2] = hc_bytealign_S (w6[2], w6[3], offset); - c4[1] = hc_bytealign_S (w6[1], w6[2], offset); - c4[0] = hc_bytealign_S (w6[0], w6[1], offset); - c3[3] = hc_bytealign_S (w5[3], w6[0], offset); - c3[2] = hc_bytealign_S (w5[2], w5[3], offset); - c3[1] = hc_bytealign_S (w5[1], w5[2], offset); - c3[0] = hc_bytealign_S (w5[0], w5[1], offset); - c2[3] = hc_bytealign_S (w4[3], w5[0], offset); - c2[2] = hc_bytealign_S (w4[2], w4[3], offset); - c2[1] = hc_bytealign_S (w4[1], w4[2], offset); - c2[0] = hc_bytealign_S (w4[0], w4[1], offset); - c1[3] = hc_bytealign_S (w3[3], w4[0], offset); - c1[2] = hc_bytealign_S (w3[2], w3[3], offset); - c1[1] = hc_bytealign_S (w3[1], w3[2], offset); - c1[0] = hc_bytealign_S (w3[0], w3[1], offset); - c0[3] = hc_bytealign_S (w2[3], w3[0], offset); - c0[2] = hc_bytealign_S (w2[2], w2[3], offset); - c0[1] = hc_bytealign_S (w2[1], w2[2], offset); - c0[0] = hc_bytealign_S (w2[0], w2[1], offset); - w7[3] = hc_bytealign_S (w1[3], w2[0], offset); - w7[2] = hc_bytealign_S (w1[2], w1[3], offset); - w7[1] = hc_bytealign_S (w1[1], w1[2], offset); - w7[0] = hc_bytealign_S (w1[0], w1[1], offset); - w6[3] = hc_bytealign_S (w0[3], w1[0], offset); - w6[2] = hc_bytealign_S (w0[2], w0[3], offset); - w6[1] = hc_bytealign_S (w0[1], w0[2], offset); - w6[0] = hc_bytealign_S (w0[0], w0[1], offset); - w5[3] = hc_bytealign_S ( 0, w0[0], offset); + c5[3] = hc_bytealign_be_S (w7[3], 0, offset); + c5[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c5[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c5[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c4[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c4[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c4[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c4[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c3[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c3[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c3[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c3[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + c2[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + c2[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + c2[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + c2[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + c1[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + c1[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c0[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + w7[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w6[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w5[3] = hc_bytealign_be_S ( 0, w0[0], offset); w5[2] = 0; w5[1] = 0; w5[0] = 0; @@ -40346,39 +40034,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 24: - c6[0] = hc_bytealign_S (w7[3], 0, offset); - c5[3] = hc_bytealign_S (w7[2], w7[3], offset); - c5[2] = hc_bytealign_S (w7[1], w7[2], offset); - c5[1] = hc_bytealign_S (w7[0], w7[1], offset); - c5[0] = hc_bytealign_S (w6[3], w7[0], offset); - c4[3] = hc_bytealign_S (w6[2], w6[3], offset); - c4[2] = hc_bytealign_S (w6[1], w6[2], offset); - c4[1] = hc_bytealign_S (w6[0], w6[1], offset); - c4[0] = hc_bytealign_S (w5[3], w6[0], offset); - c3[3] = hc_bytealign_S (w5[2], w5[3], offset); - c3[2] = hc_bytealign_S (w5[1], w5[2], offset); - c3[1] = hc_bytealign_S (w5[0], w5[1], offset); - c3[0] = hc_bytealign_S (w4[3], w5[0], offset); - c2[3] = hc_bytealign_S (w4[2], w4[3], offset); - c2[2] = hc_bytealign_S (w4[1], w4[2], offset); - c2[1] = hc_bytealign_S (w4[0], w4[1], offset); - c2[0] = hc_bytealign_S (w3[3], w4[0], offset); - c1[3] = hc_bytealign_S (w3[2], w3[3], offset); - c1[2] = hc_bytealign_S (w3[1], w3[2], offset); - c1[1] = hc_bytealign_S (w3[0], w3[1], offset); - c1[0] = hc_bytealign_S (w2[3], w3[0], offset); - c0[3] = hc_bytealign_S (w2[2], w2[3], offset); - c0[2] = hc_bytealign_S (w2[1], w2[2], offset); - c0[1] = hc_bytealign_S (w2[0], w2[1], offset); - c0[0] = hc_bytealign_S (w1[3], w2[0], offset); - w7[3] = hc_bytealign_S (w1[2], w1[3], offset); - w7[2] = hc_bytealign_S (w1[1], w1[2], offset); - w7[1] = hc_bytealign_S (w1[0], w1[1], offset); - w7[0] = hc_bytealign_S (w0[3], w1[0], offset); - w6[3] = hc_bytealign_S (w0[2], w0[3], offset); - w6[2] = hc_bytealign_S (w0[1], w0[2], offset); - w6[1] = hc_bytealign_S (w0[0], w0[1], offset); - w6[0] = hc_bytealign_S ( 0, w0[0], offset); + c6[0] = hc_bytealign_be_S (w7[3], 0, offset); + c5[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c5[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c5[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c5[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c4[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c4[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c4[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c4[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + c3[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + c3[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + c3[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + c3[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + c2[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + c2[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + c2[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + c2[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + c1[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + c0[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + w7[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w6[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[0] = hc_bytealign_be_S ( 0, w0[0], offset); w5[3] = 0; w5[2] = 0; w5[1] = 0; @@ -40407,39 +40095,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 25: - c6[1] = hc_bytealign_S (w7[3], 0, offset); - c6[0] = hc_bytealign_S (w7[2], w7[3], offset); - c5[3] = hc_bytealign_S (w7[1], w7[2], offset); - c5[2] = hc_bytealign_S (w7[0], w7[1], offset); - c5[1] = hc_bytealign_S (w6[3], w7[0], offset); - c5[0] = hc_bytealign_S (w6[2], w6[3], offset); - c4[3] = hc_bytealign_S (w6[1], w6[2], offset); - c4[2] = hc_bytealign_S (w6[0], w6[1], offset); - c4[1] = hc_bytealign_S (w5[3], w6[0], offset); - c4[0] = hc_bytealign_S (w5[2], w5[3], offset); - c3[3] = hc_bytealign_S (w5[1], w5[2], offset); - c3[2] = hc_bytealign_S (w5[0], w5[1], offset); - c3[1] = hc_bytealign_S (w4[3], w5[0], offset); - c3[0] = hc_bytealign_S (w4[2], w4[3], offset); - c2[3] = hc_bytealign_S (w4[1], w4[2], offset); - c2[2] = hc_bytealign_S (w4[0], w4[1], offset); - c2[1] = hc_bytealign_S (w3[3], w4[0], offset); - c2[0] = hc_bytealign_S (w3[2], w3[3], offset); - c1[3] = hc_bytealign_S (w3[1], w3[2], offset); - c1[2] = hc_bytealign_S (w3[0], w3[1], offset); - c1[1] = hc_bytealign_S (w2[3], w3[0], offset); - c1[0] = hc_bytealign_S (w2[2], w2[3], offset); - c0[3] = hc_bytealign_S (w2[1], w2[2], offset); - c0[2] = hc_bytealign_S (w2[0], w2[1], offset); - c0[1] = hc_bytealign_S (w1[3], w2[0], offset); - c0[0] = hc_bytealign_S (w1[2], w1[3], offset); - w7[3] = hc_bytealign_S (w1[1], w1[2], offset); - w7[2] = hc_bytealign_S (w1[0], w1[1], offset); - w7[1] = hc_bytealign_S (w0[3], w1[0], offset); - w7[0] = hc_bytealign_S (w0[2], w0[3], offset); - w6[3] = hc_bytealign_S (w0[1], w0[2], offset); - w6[2] = hc_bytealign_S (w0[0], w0[1], offset); - w6[1] = hc_bytealign_S ( 0, w0[0], offset); + c6[1] = hc_bytealign_be_S (w7[3], 0, offset); + c6[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c5[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c5[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c5[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c5[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c4[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c4[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c4[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c4[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + c3[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + c3[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + c3[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + c3[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + c2[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + c2[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + c2[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + c2[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c1[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + c0[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + w7[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w6[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[1] = hc_bytealign_be_S ( 0, w0[0], offset); w6[0] = 0; w5[3] = 0; w5[2] = 0; @@ -40469,39 +40157,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 26: - c6[2] = hc_bytealign_S (w7[3], 0, offset); - c6[1] = hc_bytealign_S (w7[2], w7[3], offset); - c6[0] = hc_bytealign_S (w7[1], w7[2], offset); - c5[3] = hc_bytealign_S (w7[0], w7[1], offset); - c5[2] = hc_bytealign_S (w6[3], w7[0], offset); - c5[1] = hc_bytealign_S (w6[2], w6[3], offset); - c5[0] = hc_bytealign_S (w6[1], w6[2], offset); - c4[3] = hc_bytealign_S (w6[0], w6[1], offset); - c4[2] = hc_bytealign_S (w5[3], w6[0], offset); - c4[1] = hc_bytealign_S (w5[2], w5[3], offset); - c4[0] = hc_bytealign_S (w5[1], w5[2], offset); - c3[3] = hc_bytealign_S (w5[0], w5[1], offset); - c3[2] = hc_bytealign_S (w4[3], w5[0], offset); - c3[1] = hc_bytealign_S (w4[2], w4[3], offset); - c3[0] = hc_bytealign_S (w4[1], w4[2], offset); - c2[3] = hc_bytealign_S (w4[0], w4[1], offset); - c2[2] = hc_bytealign_S (w3[3], w4[0], offset); - c2[1] = hc_bytealign_S (w3[2], w3[3], offset); - c2[0] = hc_bytealign_S (w3[1], w3[2], offset); - c1[3] = hc_bytealign_S (w3[0], w3[1], offset); - c1[2] = hc_bytealign_S (w2[3], w3[0], offset); - c1[1] = hc_bytealign_S (w2[2], w2[3], offset); - c1[0] = hc_bytealign_S (w2[1], w2[2], offset); - c0[3] = hc_bytealign_S (w2[0], w2[1], offset); - c0[2] = hc_bytealign_S (w1[3], w2[0], offset); - c0[1] = hc_bytealign_S (w1[2], w1[3], offset); - c0[0] = hc_bytealign_S (w1[1], w1[2], offset); - w7[3] = hc_bytealign_S (w1[0], w1[1], offset); - w7[2] = hc_bytealign_S (w0[3], w1[0], offset); - w7[1] = hc_bytealign_S (w0[2], w0[3], offset); - w7[0] = hc_bytealign_S (w0[1], w0[2], offset); - w6[3] = hc_bytealign_S (w0[0], w0[1], offset); - w6[2] = hc_bytealign_S ( 0, w0[0], offset); + c6[2] = hc_bytealign_be_S (w7[3], 0, offset); + c6[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c6[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c5[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c5[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c5[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c5[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c4[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c4[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c4[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c4[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + c3[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + c3[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + c3[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + c3[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + c2[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + c2[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + c2[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c1[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + c0[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + w7[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w6[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[2] = hc_bytealign_be_S ( 0, w0[0], offset); w6[1] = 0; w6[0] = 0; w5[3] = 0; @@ -40532,39 +40220,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 27: - c6[3] = hc_bytealign_S (w7[3], 0, offset); - c6[2] = hc_bytealign_S (w7[2], w7[3], offset); - c6[1] = hc_bytealign_S (w7[1], w7[2], offset); - c6[0] = hc_bytealign_S (w7[0], w7[1], offset); - c5[3] = hc_bytealign_S (w6[3], w7[0], offset); - c5[2] = hc_bytealign_S (w6[2], w6[3], offset); - c5[1] = hc_bytealign_S (w6[1], w6[2], offset); - c5[0] = hc_bytealign_S (w6[0], w6[1], offset); - c4[3] = hc_bytealign_S (w5[3], w6[0], offset); - c4[2] = hc_bytealign_S (w5[2], w5[3], offset); - c4[1] = hc_bytealign_S (w5[1], w5[2], offset); - c4[0] = hc_bytealign_S (w5[0], w5[1], offset); - c3[3] = hc_bytealign_S (w4[3], w5[0], offset); - c3[2] = hc_bytealign_S (w4[2], w4[3], offset); - c3[1] = hc_bytealign_S (w4[1], w4[2], offset); - c3[0] = hc_bytealign_S (w4[0], w4[1], offset); - c2[3] = hc_bytealign_S (w3[3], w4[0], offset); - c2[2] = hc_bytealign_S (w3[2], w3[3], offset); - c2[1] = hc_bytealign_S (w3[1], w3[2], offset); - c2[0] = hc_bytealign_S (w3[0], w3[1], offset); - c1[3] = hc_bytealign_S (w2[3], w3[0], offset); - c1[2] = hc_bytealign_S (w2[2], w2[3], offset); - c1[1] = hc_bytealign_S (w2[1], w2[2], offset); - c1[0] = hc_bytealign_S (w2[0], w2[1], offset); - c0[3] = hc_bytealign_S (w1[3], w2[0], offset); - c0[2] = hc_bytealign_S (w1[2], w1[3], offset); - c0[1] = hc_bytealign_S (w1[1], w1[2], offset); - c0[0] = hc_bytealign_S (w1[0], w1[1], offset); - w7[3] = hc_bytealign_S (w0[3], w1[0], offset); - w7[2] = hc_bytealign_S (w0[2], w0[3], offset); - w7[1] = hc_bytealign_S (w0[1], w0[2], offset); - w7[0] = hc_bytealign_S (w0[0], w0[1], offset); - w6[3] = hc_bytealign_S ( 0, w0[0], offset); + c6[3] = hc_bytealign_be_S (w7[3], 0, offset); + c6[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c6[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c6[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c5[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c5[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c5[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c5[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c4[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c4[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c4[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c4[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + c3[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + c3[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + c3[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + c3[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + c2[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + c2[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c1[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + c0[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + w7[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w6[3] = hc_bytealign_be_S ( 0, w0[0], offset); w6[2] = 0; w6[1] = 0; w6[0] = 0; @@ -40596,39 +40284,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 28: - c7[0] = hc_bytealign_S (w7[3], 0, offset); - c6[3] = hc_bytealign_S (w7[2], w7[3], offset); - c6[2] = hc_bytealign_S (w7[1], w7[2], offset); - c6[1] = hc_bytealign_S (w7[0], w7[1], offset); - c6[0] = hc_bytealign_S (w6[3], w7[0], offset); - c5[3] = hc_bytealign_S (w6[2], w6[3], offset); - c5[2] = hc_bytealign_S (w6[1], w6[2], offset); - c5[1] = hc_bytealign_S (w6[0], w6[1], offset); - c5[0] = hc_bytealign_S (w5[3], w6[0], offset); - c4[3] = hc_bytealign_S (w5[2], w5[3], offset); - c4[2] = hc_bytealign_S (w5[1], w5[2], offset); - c4[1] = hc_bytealign_S (w5[0], w5[1], offset); - c4[0] = hc_bytealign_S (w4[3], w5[0], offset); - c3[3] = hc_bytealign_S (w4[2], w4[3], offset); - c3[2] = hc_bytealign_S (w4[1], w4[2], offset); - c3[1] = hc_bytealign_S (w4[0], w4[1], offset); - c3[0] = hc_bytealign_S (w3[3], w4[0], offset); - c2[3] = hc_bytealign_S (w3[2], w3[3], offset); - c2[2] = hc_bytealign_S (w3[1], w3[2], offset); - c2[1] = hc_bytealign_S (w3[0], w3[1], offset); - c2[0] = hc_bytealign_S (w2[3], w3[0], offset); - c1[3] = hc_bytealign_S (w2[2], w2[3], offset); - c1[2] = hc_bytealign_S (w2[1], w2[2], offset); - c1[1] = hc_bytealign_S (w2[0], w2[1], offset); - c1[0] = hc_bytealign_S (w1[3], w2[0], offset); - c0[3] = hc_bytealign_S (w1[2], w1[3], offset); - c0[2] = hc_bytealign_S (w1[1], w1[2], offset); - c0[1] = hc_bytealign_S (w1[0], w1[1], offset); - c0[0] = hc_bytealign_S (w0[3], w1[0], offset); - w7[3] = hc_bytealign_S (w0[2], w0[3], offset); - w7[2] = hc_bytealign_S (w0[1], w0[2], offset); - w7[1] = hc_bytealign_S (w0[0], w0[1], offset); - w7[0] = hc_bytealign_S ( 0, w0[0], offset); + c7[0] = hc_bytealign_be_S (w7[3], 0, offset); + c6[3] = hc_bytealign_be_S (w7[2], w7[3], offset); + c6[2] = hc_bytealign_be_S (w7[1], w7[2], offset); + c6[1] = hc_bytealign_be_S (w7[0], w7[1], offset); + c6[0] = hc_bytealign_be_S (w6[3], w7[0], offset); + c5[3] = hc_bytealign_be_S (w6[2], w6[3], offset); + c5[2] = hc_bytealign_be_S (w6[1], w6[2], offset); + c5[1] = hc_bytealign_be_S (w6[0], w6[1], offset); + c5[0] = hc_bytealign_be_S (w5[3], w6[0], offset); + c4[3] = hc_bytealign_be_S (w5[2], w5[3], offset); + c4[2] = hc_bytealign_be_S (w5[1], w5[2], offset); + c4[1] = hc_bytealign_be_S (w5[0], w5[1], offset); + c4[0] = hc_bytealign_be_S (w4[3], w5[0], offset); + c3[3] = hc_bytealign_be_S (w4[2], w4[3], offset); + c3[2] = hc_bytealign_be_S (w4[1], w4[2], offset); + c3[1] = hc_bytealign_be_S (w4[0], w4[1], offset); + c3[0] = hc_bytealign_be_S (w3[3], w4[0], offset); + c2[3] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[2] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[1] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[0] = hc_bytealign_be_S (w2[3], w3[0], offset); + c1[3] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[2] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[1] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[0] = hc_bytealign_be_S (w1[3], w2[0], offset); + c0[3] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[2] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[1] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[0] = hc_bytealign_be_S (w0[3], w1[0], offset); + w7[3] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[2] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[1] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[0] = hc_bytealign_be_S ( 0, w0[0], offset); w6[3] = 0; w6[2] = 0; w6[1] = 0; @@ -40661,39 +40349,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 29: - c7[1] = hc_bytealign_S (w7[3], 0, offset); - c7[0] = hc_bytealign_S (w7[2], w7[3], offset); - c6[3] = hc_bytealign_S (w7[1], w7[2], offset); - c6[2] = hc_bytealign_S (w7[0], w7[1], offset); - c6[1] = hc_bytealign_S (w6[3], w7[0], offset); - c6[0] = hc_bytealign_S (w6[2], w6[3], offset); - c5[3] = hc_bytealign_S (w6[1], w6[2], offset); - c5[2] = hc_bytealign_S (w6[0], w6[1], offset); - c5[1] = hc_bytealign_S (w5[3], w6[0], offset); - c5[0] = hc_bytealign_S (w5[2], w5[3], offset); - c4[3] = hc_bytealign_S (w5[1], w5[2], offset); - c4[2] = hc_bytealign_S (w5[0], w5[1], offset); - c4[1] = hc_bytealign_S (w4[3], w5[0], offset); - c4[0] = hc_bytealign_S (w4[2], w4[3], offset); - c3[3] = hc_bytealign_S (w4[1], w4[2], offset); - c3[2] = hc_bytealign_S (w4[0], w4[1], offset); - c3[1] = hc_bytealign_S (w3[3], w4[0], offset); - c3[0] = hc_bytealign_S (w3[2], w3[3], offset); - c2[3] = hc_bytealign_S (w3[1], w3[2], offset); - c2[2] = hc_bytealign_S (w3[0], w3[1], offset); - c2[1] = hc_bytealign_S (w2[3], w3[0], offset); - c2[0] = hc_bytealign_S (w2[2], w2[3], offset); - c1[3] = hc_bytealign_S (w2[1], w2[2], offset); - c1[2] = hc_bytealign_S (w2[0], w2[1], offset); - c1[1] = hc_bytealign_S (w1[3], w2[0], offset); - c1[0] = hc_bytealign_S (w1[2], w1[3], offset); - c0[3] = hc_bytealign_S (w1[1], w1[2], offset); - c0[2] = hc_bytealign_S (w1[0], w1[1], offset); - c0[1] = hc_bytealign_S (w0[3], w1[0], offset); - c0[0] = hc_bytealign_S (w0[2], w0[3], offset); - w7[3] = hc_bytealign_S (w0[1], w0[2], offset); - w7[2] = hc_bytealign_S (w0[0], w0[1], offset); - w7[1] = hc_bytealign_S ( 0, w0[0], offset); + c7[1] = hc_bytealign_be_S (w7[3], 0, offset); + c7[0] = hc_bytealign_be_S (w7[2], w7[3], offset); + c6[3] = hc_bytealign_be_S (w7[1], w7[2], offset); + c6[2] = hc_bytealign_be_S (w7[0], w7[1], offset); + c6[1] = hc_bytealign_be_S (w6[3], w7[0], offset); + c6[0] = hc_bytealign_be_S (w6[2], w6[3], offset); + c5[3] = hc_bytealign_be_S (w6[1], w6[2], offset); + c5[2] = hc_bytealign_be_S (w6[0], w6[1], offset); + c5[1] = hc_bytealign_be_S (w5[3], w6[0], offset); + c5[0] = hc_bytealign_be_S (w5[2], w5[3], offset); + c4[3] = hc_bytealign_be_S (w5[1], w5[2], offset); + c4[2] = hc_bytealign_be_S (w5[0], w5[1], offset); + c4[1] = hc_bytealign_be_S (w4[3], w5[0], offset); + c4[0] = hc_bytealign_be_S (w4[2], w4[3], offset); + c3[3] = hc_bytealign_be_S (w4[1], w4[2], offset); + c3[2] = hc_bytealign_be_S (w4[0], w4[1], offset); + c3[1] = hc_bytealign_be_S (w3[3], w4[0], offset); + c3[0] = hc_bytealign_be_S (w3[2], w3[3], offset); + c2[3] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[2] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[1] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[0] = hc_bytealign_be_S (w2[2], w2[3], offset); + c1[3] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[2] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[1] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[0] = hc_bytealign_be_S (w1[2], w1[3], offset); + c0[3] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[2] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[1] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[0] = hc_bytealign_be_S (w0[2], w0[3], offset); + w7[3] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[2] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[1] = hc_bytealign_be_S ( 0, w0[0], offset); w7[0] = 0; w6[3] = 0; w6[2] = 0; @@ -40727,39 +40415,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 30: - c7[2] = hc_bytealign_S (w7[3], 0, offset); - c7[1] = hc_bytealign_S (w7[2], w7[3], offset); - c7[0] = hc_bytealign_S (w7[1], w7[2], offset); - c6[3] = hc_bytealign_S (w7[0], w7[1], offset); - c6[2] = hc_bytealign_S (w6[3], w7[0], offset); - c6[1] = hc_bytealign_S (w6[2], w6[3], offset); - c6[0] = hc_bytealign_S (w6[1], w6[2], offset); - c5[3] = hc_bytealign_S (w6[0], w6[1], offset); - c5[2] = hc_bytealign_S (w5[3], w6[0], offset); - c5[1] = hc_bytealign_S (w5[2], w5[3], offset); - c5[0] = hc_bytealign_S (w5[1], w5[2], offset); - c4[3] = hc_bytealign_S (w5[0], w5[1], offset); - c4[2] = hc_bytealign_S (w4[3], w5[0], offset); - c4[1] = hc_bytealign_S (w4[2], w4[3], offset); - c4[0] = hc_bytealign_S (w4[1], w4[2], offset); - c3[3] = hc_bytealign_S (w4[0], w4[1], offset); - c3[2] = hc_bytealign_S (w3[3], w4[0], offset); - c3[1] = hc_bytealign_S (w3[2], w3[3], offset); - c3[0] = hc_bytealign_S (w3[1], w3[2], offset); - c2[3] = hc_bytealign_S (w3[0], w3[1], offset); - c2[2] = hc_bytealign_S (w2[3], w3[0], offset); - c2[1] = hc_bytealign_S (w2[2], w2[3], offset); - c2[0] = hc_bytealign_S (w2[1], w2[2], offset); - c1[3] = hc_bytealign_S (w2[0], w2[1], offset); - c1[2] = hc_bytealign_S (w1[3], w2[0], offset); - c1[1] = hc_bytealign_S (w1[2], w1[3], offset); - c1[0] = hc_bytealign_S (w1[1], w1[2], offset); - c0[3] = hc_bytealign_S (w1[0], w1[1], offset); - c0[2] = hc_bytealign_S (w0[3], w1[0], offset); - c0[1] = hc_bytealign_S (w0[2], w0[3], offset); - c0[0] = hc_bytealign_S (w0[1], w0[2], offset); - w7[3] = hc_bytealign_S (w0[0], w0[1], offset); - w7[2] = hc_bytealign_S ( 0, w0[0], offset); + c7[2] = hc_bytealign_be_S (w7[3], 0, offset); + c7[1] = hc_bytealign_be_S (w7[2], w7[3], offset); + c7[0] = hc_bytealign_be_S (w7[1], w7[2], offset); + c6[3] = hc_bytealign_be_S (w7[0], w7[1], offset); + c6[2] = hc_bytealign_be_S (w6[3], w7[0], offset); + c6[1] = hc_bytealign_be_S (w6[2], w6[3], offset); + c6[0] = hc_bytealign_be_S (w6[1], w6[2], offset); + c5[3] = hc_bytealign_be_S (w6[0], w6[1], offset); + c5[2] = hc_bytealign_be_S (w5[3], w6[0], offset); + c5[1] = hc_bytealign_be_S (w5[2], w5[3], offset); + c5[0] = hc_bytealign_be_S (w5[1], w5[2], offset); + c4[3] = hc_bytealign_be_S (w5[0], w5[1], offset); + c4[2] = hc_bytealign_be_S (w4[3], w5[0], offset); + c4[1] = hc_bytealign_be_S (w4[2], w4[3], offset); + c4[0] = hc_bytealign_be_S (w4[1], w4[2], offset); + c3[3] = hc_bytealign_be_S (w4[0], w4[1], offset); + c3[2] = hc_bytealign_be_S (w3[3], w4[0], offset); + c3[1] = hc_bytealign_be_S (w3[2], w3[3], offset); + c3[0] = hc_bytealign_be_S (w3[1], w3[2], offset); + c2[3] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[2] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[1] = hc_bytealign_be_S (w2[2], w2[3], offset); + c2[0] = hc_bytealign_be_S (w2[1], w2[2], offset); + c1[3] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[2] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[1] = hc_bytealign_be_S (w1[2], w1[3], offset); + c1[0] = hc_bytealign_be_S (w1[1], w1[2], offset); + c0[3] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[2] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[1] = hc_bytealign_be_S (w0[2], w0[3], offset); + c0[0] = hc_bytealign_be_S (w0[1], w0[2], offset); + w7[3] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[2] = hc_bytealign_be_S ( 0, w0[0], offset); w7[1] = 0; w7[0] = 0; w6[3] = 0; @@ -40794,39 +40482,39 @@ DECLSPEC void switch_buffer_by_offset_8x4_carry_be_S (u32 *w0, u32 *w1, u32 *w2, break; case 31: - c7[3] = hc_bytealign_S (w7[3], 0, offset); - c7[2] = hc_bytealign_S (w7[2], w7[3], offset); - c7[1] = hc_bytealign_S (w7[1], w7[2], offset); - c7[0] = hc_bytealign_S (w7[0], w7[1], offset); - c6[3] = hc_bytealign_S (w6[3], w7[0], offset); - c6[2] = hc_bytealign_S (w6[2], w6[3], offset); - c6[1] = hc_bytealign_S (w6[1], w6[2], offset); - c6[0] = hc_bytealign_S (w6[0], w6[1], offset); - c5[3] = hc_bytealign_S (w5[3], w6[0], offset); - c5[2] = hc_bytealign_S (w5[2], w5[3], offset); - c5[1] = hc_bytealign_S (w5[1], w5[2], offset); - c5[0] = hc_bytealign_S (w5[0], w5[1], offset); - c4[3] = hc_bytealign_S (w4[3], w5[0], offset); - c4[2] = hc_bytealign_S (w4[2], w4[3], offset); - c4[1] = hc_bytealign_S (w4[1], w4[2], offset); - c4[0] = hc_bytealign_S (w4[0], w4[1], offset); - c3[3] = hc_bytealign_S (w3[3], w4[0], offset); - c3[2] = hc_bytealign_S (w3[2], w3[3], offset); - c3[1] = hc_bytealign_S (w3[1], w3[2], offset); - c3[0] = hc_bytealign_S (w3[0], w3[1], offset); - c2[3] = hc_bytealign_S (w2[3], w3[0], offset); - c2[2] = hc_bytealign_S (w2[2], w2[3], offset); - c2[1] = hc_bytealign_S (w2[1], w2[2], offset); - c2[0] = hc_bytealign_S (w2[0], w2[1], offset); - c1[3] = hc_bytealign_S (w1[3], w2[0], offset); - c1[2] = hc_bytealign_S (w1[2], w1[3], offset); - c1[1] = hc_bytealign_S (w1[1], w1[2], offset); - c1[0] = hc_bytealign_S (w1[0], w1[1], offset); - c0[3] = hc_bytealign_S (w0[3], w1[0], offset); - c0[2] = hc_bytealign_S (w0[2], w0[3], offset); - c0[1] = hc_bytealign_S (w0[1], w0[2], offset); - c0[0] = hc_bytealign_S (w0[0], w0[1], offset); - w7[3] = hc_bytealign_S ( 0, w0[0], offset); + c7[3] = hc_bytealign_be_S (w7[3], 0, offset); + c7[2] = hc_bytealign_be_S (w7[2], w7[3], offset); + c7[1] = hc_bytealign_be_S (w7[1], w7[2], offset); + c7[0] = hc_bytealign_be_S (w7[0], w7[1], offset); + c6[3] = hc_bytealign_be_S (w6[3], w7[0], offset); + c6[2] = hc_bytealign_be_S (w6[2], w6[3], offset); + c6[1] = hc_bytealign_be_S (w6[1], w6[2], offset); + c6[0] = hc_bytealign_be_S (w6[0], w6[1], offset); + c5[3] = hc_bytealign_be_S (w5[3], w6[0], offset); + c5[2] = hc_bytealign_be_S (w5[2], w5[3], offset); + c5[1] = hc_bytealign_be_S (w5[1], w5[2], offset); + c5[0] = hc_bytealign_be_S (w5[0], w5[1], offset); + c4[3] = hc_bytealign_be_S (w4[3], w5[0], offset); + c4[2] = hc_bytealign_be_S (w4[2], w4[3], offset); + c4[1] = hc_bytealign_be_S (w4[1], w4[2], offset); + c4[0] = hc_bytealign_be_S (w4[0], w4[1], offset); + c3[3] = hc_bytealign_be_S (w3[3], w4[0], offset); + c3[2] = hc_bytealign_be_S (w3[2], w3[3], offset); + c3[1] = hc_bytealign_be_S (w3[1], w3[2], offset); + c3[0] = hc_bytealign_be_S (w3[0], w3[1], offset); + c2[3] = hc_bytealign_be_S (w2[3], w3[0], offset); + c2[2] = hc_bytealign_be_S (w2[2], w2[3], offset); + c2[1] = hc_bytealign_be_S (w2[1], w2[2], offset); + c2[0] = hc_bytealign_be_S (w2[0], w2[1], offset); + c1[3] = hc_bytealign_be_S (w1[3], w2[0], offset); + c1[2] = hc_bytealign_be_S (w1[2], w1[3], offset); + c1[1] = hc_bytealign_be_S (w1[1], w1[2], offset); + c1[0] = hc_bytealign_be_S (w1[0], w1[1], offset); + c0[3] = hc_bytealign_be_S (w0[3], w1[0], offset); + c0[2] = hc_bytealign_be_S (w0[2], w0[3], offset); + c0[1] = hc_bytealign_be_S (w0[1], w0[2], offset); + c0[0] = hc_bytealign_be_S (w0[0], w0[1], offset); + w7[3] = hc_bytealign_be_S ( 0, w0[0], offset); w7[2] = 0; w7[1] = 0; w7[0] = 0; @@ -42567,10 +42255,6 @@ DECLSPEC void switch_buffer_by_offset_1x64_le_S (u32 *w, const u32 offset) const int offset_switch = offset / 4; #if (defined IS_AMD && HAS_VPERM == 0) || defined IS_GENERIC - - #pragma unroll - for (int i = 0; i < 64; i++) w[i] = swap32_S (w[i]); - switch (offset_switch) { case 0: @@ -46925,10 +46609,6 @@ DECLSPEC void switch_buffer_by_offset_1x64_le_S (u32 *w, const u32 offset) break; } - - #pragma unroll - for (int i = 0; i < 64; i++) w[i] = swap32_S (w[i]); - #endif #if (defined IS_AMD && HAS_VPERM == 1) || defined IS_NV @@ -51306,271 +50986,271 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) switch (offset_switch) { case 0: - w[63] = hc_bytealign_S (w[62], w[63], offset); - w[62] = hc_bytealign_S (w[61], w[62], offset); - w[61] = hc_bytealign_S (w[60], w[61], offset); - w[60] = hc_bytealign_S (w[59], w[60], offset); - w[59] = hc_bytealign_S (w[58], w[59], offset); - w[58] = hc_bytealign_S (w[57], w[58], offset); - w[57] = hc_bytealign_S (w[56], w[57], offset); - w[56] = hc_bytealign_S (w[55], w[56], offset); - w[55] = hc_bytealign_S (w[54], w[55], offset); - w[54] = hc_bytealign_S (w[53], w[54], offset); - w[53] = hc_bytealign_S (w[52], w[53], offset); - w[52] = hc_bytealign_S (w[51], w[52], offset); - w[51] = hc_bytealign_S (w[50], w[51], offset); - w[50] = hc_bytealign_S (w[49], w[50], offset); - w[49] = hc_bytealign_S (w[48], w[49], offset); - w[48] = hc_bytealign_S (w[47], w[48], offset); - w[47] = hc_bytealign_S (w[46], w[47], offset); - w[46] = hc_bytealign_S (w[45], w[46], offset); - w[45] = hc_bytealign_S (w[44], w[45], offset); - w[44] = hc_bytealign_S (w[43], w[44], offset); - w[43] = hc_bytealign_S (w[42], w[43], offset); - w[42] = hc_bytealign_S (w[41], w[42], offset); - w[41] = hc_bytealign_S (w[40], w[41], offset); - w[40] = hc_bytealign_S (w[39], w[40], offset); - w[39] = hc_bytealign_S (w[38], w[39], offset); - w[38] = hc_bytealign_S (w[37], w[38], offset); - w[37] = hc_bytealign_S (w[36], w[37], offset); - w[36] = hc_bytealign_S (w[35], w[36], offset); - w[35] = hc_bytealign_S (w[34], w[35], offset); - w[34] = hc_bytealign_S (w[33], w[34], offset); - w[33] = hc_bytealign_S (w[32], w[33], offset); - w[32] = hc_bytealign_S (w[31], w[32], offset); - w[31] = hc_bytealign_S (w[30], w[31], offset); - w[30] = hc_bytealign_S (w[29], w[30], offset); - w[29] = hc_bytealign_S (w[28], w[29], offset); - w[28] = hc_bytealign_S (w[27], w[28], offset); - w[27] = hc_bytealign_S (w[26], w[27], offset); - w[26] = hc_bytealign_S (w[25], w[26], offset); - w[25] = hc_bytealign_S (w[24], w[25], offset); - w[24] = hc_bytealign_S (w[23], w[24], offset); - w[23] = hc_bytealign_S (w[22], w[23], offset); - w[22] = hc_bytealign_S (w[21], w[22], offset); - w[21] = hc_bytealign_S (w[20], w[21], offset); - w[20] = hc_bytealign_S (w[19], w[20], offset); - w[19] = hc_bytealign_S (w[18], w[19], offset); - w[18] = hc_bytealign_S (w[17], w[18], offset); - w[17] = hc_bytealign_S (w[16], w[17], offset); - w[16] = hc_bytealign_S (w[15], w[16], offset); - w[15] = hc_bytealign_S (w[14], w[15], offset); - w[14] = hc_bytealign_S (w[13], w[14], offset); - w[13] = hc_bytealign_S (w[12], w[13], offset); - w[12] = hc_bytealign_S (w[11], w[12], offset); - w[11] = hc_bytealign_S (w[10], w[11], offset); - w[10] = hc_bytealign_S (w[ 9], w[10], offset); - w[ 9] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[ 8] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[ 7] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[ 6] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[ 5] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 4] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 3] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 2] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 1] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 0] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[62], w[63], offset); + w[62] = hc_bytealign_be_S (w[61], w[62], offset); + w[61] = hc_bytealign_be_S (w[60], w[61], offset); + w[60] = hc_bytealign_be_S (w[59], w[60], offset); + w[59] = hc_bytealign_be_S (w[58], w[59], offset); + w[58] = hc_bytealign_be_S (w[57], w[58], offset); + w[57] = hc_bytealign_be_S (w[56], w[57], offset); + w[56] = hc_bytealign_be_S (w[55], w[56], offset); + w[55] = hc_bytealign_be_S (w[54], w[55], offset); + w[54] = hc_bytealign_be_S (w[53], w[54], offset); + w[53] = hc_bytealign_be_S (w[52], w[53], offset); + w[52] = hc_bytealign_be_S (w[51], w[52], offset); + w[51] = hc_bytealign_be_S (w[50], w[51], offset); + w[50] = hc_bytealign_be_S (w[49], w[50], offset); + w[49] = hc_bytealign_be_S (w[48], w[49], offset); + w[48] = hc_bytealign_be_S (w[47], w[48], offset); + w[47] = hc_bytealign_be_S (w[46], w[47], offset); + w[46] = hc_bytealign_be_S (w[45], w[46], offset); + w[45] = hc_bytealign_be_S (w[44], w[45], offset); + w[44] = hc_bytealign_be_S (w[43], w[44], offset); + w[43] = hc_bytealign_be_S (w[42], w[43], offset); + w[42] = hc_bytealign_be_S (w[41], w[42], offset); + w[41] = hc_bytealign_be_S (w[40], w[41], offset); + w[40] = hc_bytealign_be_S (w[39], w[40], offset); + w[39] = hc_bytealign_be_S (w[38], w[39], offset); + w[38] = hc_bytealign_be_S (w[37], w[38], offset); + w[37] = hc_bytealign_be_S (w[36], w[37], offset); + w[36] = hc_bytealign_be_S (w[35], w[36], offset); + w[35] = hc_bytealign_be_S (w[34], w[35], offset); + w[34] = hc_bytealign_be_S (w[33], w[34], offset); + w[33] = hc_bytealign_be_S (w[32], w[33], offset); + w[32] = hc_bytealign_be_S (w[31], w[32], offset); + w[31] = hc_bytealign_be_S (w[30], w[31], offset); + w[30] = hc_bytealign_be_S (w[29], w[30], offset); + w[29] = hc_bytealign_be_S (w[28], w[29], offset); + w[28] = hc_bytealign_be_S (w[27], w[28], offset); + w[27] = hc_bytealign_be_S (w[26], w[27], offset); + w[26] = hc_bytealign_be_S (w[25], w[26], offset); + w[25] = hc_bytealign_be_S (w[24], w[25], offset); + w[24] = hc_bytealign_be_S (w[23], w[24], offset); + w[23] = hc_bytealign_be_S (w[22], w[23], offset); + w[22] = hc_bytealign_be_S (w[21], w[22], offset); + w[21] = hc_bytealign_be_S (w[20], w[21], offset); + w[20] = hc_bytealign_be_S (w[19], w[20], offset); + w[19] = hc_bytealign_be_S (w[18], w[19], offset); + w[18] = hc_bytealign_be_S (w[17], w[18], offset); + w[17] = hc_bytealign_be_S (w[16], w[17], offset); + w[16] = hc_bytealign_be_S (w[15], w[16], offset); + w[15] = hc_bytealign_be_S (w[14], w[15], offset); + w[14] = hc_bytealign_be_S (w[13], w[14], offset); + w[13] = hc_bytealign_be_S (w[12], w[13], offset); + w[12] = hc_bytealign_be_S (w[11], w[12], offset); + w[11] = hc_bytealign_be_S (w[10], w[11], offset); + w[10] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[ 9] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[ 8] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[ 7] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[ 6] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[ 5] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 4] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 3] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 2] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 1] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 0] = hc_bytealign_be_S ( 0, w[ 0], offset); break; case 1: - w[63] = hc_bytealign_S (w[61], w[62], offset); - w[62] = hc_bytealign_S (w[60], w[61], offset); - w[61] = hc_bytealign_S (w[59], w[60], offset); - w[60] = hc_bytealign_S (w[58], w[59], offset); - w[59] = hc_bytealign_S (w[57], w[58], offset); - w[58] = hc_bytealign_S (w[56], w[57], offset); - w[57] = hc_bytealign_S (w[55], w[56], offset); - w[56] = hc_bytealign_S (w[54], w[55], offset); - w[55] = hc_bytealign_S (w[53], w[54], offset); - w[54] = hc_bytealign_S (w[52], w[53], offset); - w[53] = hc_bytealign_S (w[51], w[52], offset); - w[52] = hc_bytealign_S (w[50], w[51], offset); - w[51] = hc_bytealign_S (w[49], w[50], offset); - w[50] = hc_bytealign_S (w[48], w[49], offset); - w[49] = hc_bytealign_S (w[47], w[48], offset); - w[48] = hc_bytealign_S (w[46], w[47], offset); - w[47] = hc_bytealign_S (w[45], w[46], offset); - w[46] = hc_bytealign_S (w[44], w[45], offset); - w[45] = hc_bytealign_S (w[43], w[44], offset); - w[44] = hc_bytealign_S (w[42], w[43], offset); - w[43] = hc_bytealign_S (w[41], w[42], offset); - w[42] = hc_bytealign_S (w[40], w[41], offset); - w[41] = hc_bytealign_S (w[39], w[40], offset); - w[40] = hc_bytealign_S (w[38], w[39], offset); - w[39] = hc_bytealign_S (w[37], w[38], offset); - w[38] = hc_bytealign_S (w[36], w[37], offset); - w[37] = hc_bytealign_S (w[35], w[36], offset); - w[36] = hc_bytealign_S (w[34], w[35], offset); - w[35] = hc_bytealign_S (w[33], w[34], offset); - w[34] = hc_bytealign_S (w[32], w[33], offset); - w[33] = hc_bytealign_S (w[31], w[32], offset); - w[32] = hc_bytealign_S (w[30], w[31], offset); - w[31] = hc_bytealign_S (w[29], w[30], offset); - w[30] = hc_bytealign_S (w[28], w[29], offset); - w[29] = hc_bytealign_S (w[27], w[28], offset); - w[28] = hc_bytealign_S (w[26], w[27], offset); - w[27] = hc_bytealign_S (w[25], w[26], offset); - w[26] = hc_bytealign_S (w[24], w[25], offset); - w[25] = hc_bytealign_S (w[23], w[24], offset); - w[24] = hc_bytealign_S (w[22], w[23], offset); - w[23] = hc_bytealign_S (w[21], w[22], offset); - w[22] = hc_bytealign_S (w[20], w[21], offset); - w[21] = hc_bytealign_S (w[19], w[20], offset); - w[20] = hc_bytealign_S (w[18], w[19], offset); - w[19] = hc_bytealign_S (w[17], w[18], offset); - w[18] = hc_bytealign_S (w[16], w[17], offset); - w[17] = hc_bytealign_S (w[15], w[16], offset); - w[16] = hc_bytealign_S (w[14], w[15], offset); - w[15] = hc_bytealign_S (w[13], w[14], offset); - w[14] = hc_bytealign_S (w[12], w[13], offset); - w[13] = hc_bytealign_S (w[11], w[12], offset); - w[12] = hc_bytealign_S (w[10], w[11], offset); - w[11] = hc_bytealign_S (w[ 9], w[10], offset); - w[10] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[ 9] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[ 8] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[ 7] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[ 6] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 5] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 4] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 3] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 2] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 1] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[61], w[62], offset); + w[62] = hc_bytealign_be_S (w[60], w[61], offset); + w[61] = hc_bytealign_be_S (w[59], w[60], offset); + w[60] = hc_bytealign_be_S (w[58], w[59], offset); + w[59] = hc_bytealign_be_S (w[57], w[58], offset); + w[58] = hc_bytealign_be_S (w[56], w[57], offset); + w[57] = hc_bytealign_be_S (w[55], w[56], offset); + w[56] = hc_bytealign_be_S (w[54], w[55], offset); + w[55] = hc_bytealign_be_S (w[53], w[54], offset); + w[54] = hc_bytealign_be_S (w[52], w[53], offset); + w[53] = hc_bytealign_be_S (w[51], w[52], offset); + w[52] = hc_bytealign_be_S (w[50], w[51], offset); + w[51] = hc_bytealign_be_S (w[49], w[50], offset); + w[50] = hc_bytealign_be_S (w[48], w[49], offset); + w[49] = hc_bytealign_be_S (w[47], w[48], offset); + w[48] = hc_bytealign_be_S (w[46], w[47], offset); + w[47] = hc_bytealign_be_S (w[45], w[46], offset); + w[46] = hc_bytealign_be_S (w[44], w[45], offset); + w[45] = hc_bytealign_be_S (w[43], w[44], offset); + w[44] = hc_bytealign_be_S (w[42], w[43], offset); + w[43] = hc_bytealign_be_S (w[41], w[42], offset); + w[42] = hc_bytealign_be_S (w[40], w[41], offset); + w[41] = hc_bytealign_be_S (w[39], w[40], offset); + w[40] = hc_bytealign_be_S (w[38], w[39], offset); + w[39] = hc_bytealign_be_S (w[37], w[38], offset); + w[38] = hc_bytealign_be_S (w[36], w[37], offset); + w[37] = hc_bytealign_be_S (w[35], w[36], offset); + w[36] = hc_bytealign_be_S (w[34], w[35], offset); + w[35] = hc_bytealign_be_S (w[33], w[34], offset); + w[34] = hc_bytealign_be_S (w[32], w[33], offset); + w[33] = hc_bytealign_be_S (w[31], w[32], offset); + w[32] = hc_bytealign_be_S (w[30], w[31], offset); + w[31] = hc_bytealign_be_S (w[29], w[30], offset); + w[30] = hc_bytealign_be_S (w[28], w[29], offset); + w[29] = hc_bytealign_be_S (w[27], w[28], offset); + w[28] = hc_bytealign_be_S (w[26], w[27], offset); + w[27] = hc_bytealign_be_S (w[25], w[26], offset); + w[26] = hc_bytealign_be_S (w[24], w[25], offset); + w[25] = hc_bytealign_be_S (w[23], w[24], offset); + w[24] = hc_bytealign_be_S (w[22], w[23], offset); + w[23] = hc_bytealign_be_S (w[21], w[22], offset); + w[22] = hc_bytealign_be_S (w[20], w[21], offset); + w[21] = hc_bytealign_be_S (w[19], w[20], offset); + w[20] = hc_bytealign_be_S (w[18], w[19], offset); + w[19] = hc_bytealign_be_S (w[17], w[18], offset); + w[18] = hc_bytealign_be_S (w[16], w[17], offset); + w[17] = hc_bytealign_be_S (w[15], w[16], offset); + w[16] = hc_bytealign_be_S (w[14], w[15], offset); + w[15] = hc_bytealign_be_S (w[13], w[14], offset); + w[14] = hc_bytealign_be_S (w[12], w[13], offset); + w[13] = hc_bytealign_be_S (w[11], w[12], offset); + w[12] = hc_bytealign_be_S (w[10], w[11], offset); + w[11] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[10] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[ 9] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[ 8] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[ 7] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[ 6] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 5] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 4] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 3] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 2] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 1] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 0] = 0; break; case 2: - w[63] = hc_bytealign_S (w[60], w[61], offset); - w[62] = hc_bytealign_S (w[59], w[60], offset); - w[61] = hc_bytealign_S (w[58], w[59], offset); - w[60] = hc_bytealign_S (w[57], w[58], offset); - w[59] = hc_bytealign_S (w[56], w[57], offset); - w[58] = hc_bytealign_S (w[55], w[56], offset); - w[57] = hc_bytealign_S (w[54], w[55], offset); - w[56] = hc_bytealign_S (w[53], w[54], offset); - w[55] = hc_bytealign_S (w[52], w[53], offset); - w[54] = hc_bytealign_S (w[51], w[52], offset); - w[53] = hc_bytealign_S (w[50], w[51], offset); - w[52] = hc_bytealign_S (w[49], w[50], offset); - w[51] = hc_bytealign_S (w[48], w[49], offset); - w[50] = hc_bytealign_S (w[47], w[48], offset); - w[49] = hc_bytealign_S (w[46], w[47], offset); - w[48] = hc_bytealign_S (w[45], w[46], offset); - w[47] = hc_bytealign_S (w[44], w[45], offset); - w[46] = hc_bytealign_S (w[43], w[44], offset); - w[45] = hc_bytealign_S (w[42], w[43], offset); - w[44] = hc_bytealign_S (w[41], w[42], offset); - w[43] = hc_bytealign_S (w[40], w[41], offset); - w[42] = hc_bytealign_S (w[39], w[40], offset); - w[41] = hc_bytealign_S (w[38], w[39], offset); - w[40] = hc_bytealign_S (w[37], w[38], offset); - w[39] = hc_bytealign_S (w[36], w[37], offset); - w[38] = hc_bytealign_S (w[35], w[36], offset); - w[37] = hc_bytealign_S (w[34], w[35], offset); - w[36] = hc_bytealign_S (w[33], w[34], offset); - w[35] = hc_bytealign_S (w[32], w[33], offset); - w[34] = hc_bytealign_S (w[31], w[32], offset); - w[33] = hc_bytealign_S (w[30], w[31], offset); - w[32] = hc_bytealign_S (w[29], w[30], offset); - w[31] = hc_bytealign_S (w[28], w[29], offset); - w[30] = hc_bytealign_S (w[27], w[28], offset); - w[29] = hc_bytealign_S (w[26], w[27], offset); - w[28] = hc_bytealign_S (w[25], w[26], offset); - w[27] = hc_bytealign_S (w[24], w[25], offset); - w[26] = hc_bytealign_S (w[23], w[24], offset); - w[25] = hc_bytealign_S (w[22], w[23], offset); - w[24] = hc_bytealign_S (w[21], w[22], offset); - w[23] = hc_bytealign_S (w[20], w[21], offset); - w[22] = hc_bytealign_S (w[19], w[20], offset); - w[21] = hc_bytealign_S (w[18], w[19], offset); - w[20] = hc_bytealign_S (w[17], w[18], offset); - w[19] = hc_bytealign_S (w[16], w[17], offset); - w[18] = hc_bytealign_S (w[15], w[16], offset); - w[17] = hc_bytealign_S (w[14], w[15], offset); - w[16] = hc_bytealign_S (w[13], w[14], offset); - w[15] = hc_bytealign_S (w[12], w[13], offset); - w[14] = hc_bytealign_S (w[11], w[12], offset); - w[13] = hc_bytealign_S (w[10], w[11], offset); - w[12] = hc_bytealign_S (w[ 9], w[10], offset); - w[11] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[10] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[ 9] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[ 8] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[ 7] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 6] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 5] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 4] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 3] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 2] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[60], w[61], offset); + w[62] = hc_bytealign_be_S (w[59], w[60], offset); + w[61] = hc_bytealign_be_S (w[58], w[59], offset); + w[60] = hc_bytealign_be_S (w[57], w[58], offset); + w[59] = hc_bytealign_be_S (w[56], w[57], offset); + w[58] = hc_bytealign_be_S (w[55], w[56], offset); + w[57] = hc_bytealign_be_S (w[54], w[55], offset); + w[56] = hc_bytealign_be_S (w[53], w[54], offset); + w[55] = hc_bytealign_be_S (w[52], w[53], offset); + w[54] = hc_bytealign_be_S (w[51], w[52], offset); + w[53] = hc_bytealign_be_S (w[50], w[51], offset); + w[52] = hc_bytealign_be_S (w[49], w[50], offset); + w[51] = hc_bytealign_be_S (w[48], w[49], offset); + w[50] = hc_bytealign_be_S (w[47], w[48], offset); + w[49] = hc_bytealign_be_S (w[46], w[47], offset); + w[48] = hc_bytealign_be_S (w[45], w[46], offset); + w[47] = hc_bytealign_be_S (w[44], w[45], offset); + w[46] = hc_bytealign_be_S (w[43], w[44], offset); + w[45] = hc_bytealign_be_S (w[42], w[43], offset); + w[44] = hc_bytealign_be_S (w[41], w[42], offset); + w[43] = hc_bytealign_be_S (w[40], w[41], offset); + w[42] = hc_bytealign_be_S (w[39], w[40], offset); + w[41] = hc_bytealign_be_S (w[38], w[39], offset); + w[40] = hc_bytealign_be_S (w[37], w[38], offset); + w[39] = hc_bytealign_be_S (w[36], w[37], offset); + w[38] = hc_bytealign_be_S (w[35], w[36], offset); + w[37] = hc_bytealign_be_S (w[34], w[35], offset); + w[36] = hc_bytealign_be_S (w[33], w[34], offset); + w[35] = hc_bytealign_be_S (w[32], w[33], offset); + w[34] = hc_bytealign_be_S (w[31], w[32], offset); + w[33] = hc_bytealign_be_S (w[30], w[31], offset); + w[32] = hc_bytealign_be_S (w[29], w[30], offset); + w[31] = hc_bytealign_be_S (w[28], w[29], offset); + w[30] = hc_bytealign_be_S (w[27], w[28], offset); + w[29] = hc_bytealign_be_S (w[26], w[27], offset); + w[28] = hc_bytealign_be_S (w[25], w[26], offset); + w[27] = hc_bytealign_be_S (w[24], w[25], offset); + w[26] = hc_bytealign_be_S (w[23], w[24], offset); + w[25] = hc_bytealign_be_S (w[22], w[23], offset); + w[24] = hc_bytealign_be_S (w[21], w[22], offset); + w[23] = hc_bytealign_be_S (w[20], w[21], offset); + w[22] = hc_bytealign_be_S (w[19], w[20], offset); + w[21] = hc_bytealign_be_S (w[18], w[19], offset); + w[20] = hc_bytealign_be_S (w[17], w[18], offset); + w[19] = hc_bytealign_be_S (w[16], w[17], offset); + w[18] = hc_bytealign_be_S (w[15], w[16], offset); + w[17] = hc_bytealign_be_S (w[14], w[15], offset); + w[16] = hc_bytealign_be_S (w[13], w[14], offset); + w[15] = hc_bytealign_be_S (w[12], w[13], offset); + w[14] = hc_bytealign_be_S (w[11], w[12], offset); + w[13] = hc_bytealign_be_S (w[10], w[11], offset); + w[12] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[11] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[10] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[ 9] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[ 8] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[ 7] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 6] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 5] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 4] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 3] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 2] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 1] = 0; w[ 0] = 0; break; case 3: - w[63] = hc_bytealign_S (w[59], w[60], offset); - w[62] = hc_bytealign_S (w[58], w[59], offset); - w[61] = hc_bytealign_S (w[57], w[58], offset); - w[60] = hc_bytealign_S (w[56], w[57], offset); - w[59] = hc_bytealign_S (w[55], w[56], offset); - w[58] = hc_bytealign_S (w[54], w[55], offset); - w[57] = hc_bytealign_S (w[53], w[54], offset); - w[56] = hc_bytealign_S (w[52], w[53], offset); - w[55] = hc_bytealign_S (w[51], w[52], offset); - w[54] = hc_bytealign_S (w[50], w[51], offset); - w[53] = hc_bytealign_S (w[49], w[50], offset); - w[52] = hc_bytealign_S (w[48], w[49], offset); - w[51] = hc_bytealign_S (w[47], w[48], offset); - w[50] = hc_bytealign_S (w[46], w[47], offset); - w[49] = hc_bytealign_S (w[45], w[46], offset); - w[48] = hc_bytealign_S (w[44], w[45], offset); - w[47] = hc_bytealign_S (w[43], w[44], offset); - w[46] = hc_bytealign_S (w[42], w[43], offset); - w[45] = hc_bytealign_S (w[41], w[42], offset); - w[44] = hc_bytealign_S (w[40], w[41], offset); - w[43] = hc_bytealign_S (w[39], w[40], offset); - w[42] = hc_bytealign_S (w[38], w[39], offset); - w[41] = hc_bytealign_S (w[37], w[38], offset); - w[40] = hc_bytealign_S (w[36], w[37], offset); - w[39] = hc_bytealign_S (w[35], w[36], offset); - w[38] = hc_bytealign_S (w[34], w[35], offset); - w[37] = hc_bytealign_S (w[33], w[34], offset); - w[36] = hc_bytealign_S (w[32], w[33], offset); - w[35] = hc_bytealign_S (w[31], w[32], offset); - w[34] = hc_bytealign_S (w[30], w[31], offset); - w[33] = hc_bytealign_S (w[29], w[30], offset); - w[32] = hc_bytealign_S (w[28], w[29], offset); - w[31] = hc_bytealign_S (w[27], w[28], offset); - w[30] = hc_bytealign_S (w[26], w[27], offset); - w[29] = hc_bytealign_S (w[25], w[26], offset); - w[28] = hc_bytealign_S (w[24], w[25], offset); - w[27] = hc_bytealign_S (w[23], w[24], offset); - w[26] = hc_bytealign_S (w[22], w[23], offset); - w[25] = hc_bytealign_S (w[21], w[22], offset); - w[24] = hc_bytealign_S (w[20], w[21], offset); - w[23] = hc_bytealign_S (w[19], w[20], offset); - w[22] = hc_bytealign_S (w[18], w[19], offset); - w[21] = hc_bytealign_S (w[17], w[18], offset); - w[20] = hc_bytealign_S (w[16], w[17], offset); - w[19] = hc_bytealign_S (w[15], w[16], offset); - w[18] = hc_bytealign_S (w[14], w[15], offset); - w[17] = hc_bytealign_S (w[13], w[14], offset); - w[16] = hc_bytealign_S (w[12], w[13], offset); - w[15] = hc_bytealign_S (w[11], w[12], offset); - w[14] = hc_bytealign_S (w[10], w[11], offset); - w[13] = hc_bytealign_S (w[ 9], w[10], offset); - w[12] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[11] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[10] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[ 9] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[ 8] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 7] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 6] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 5] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 4] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 3] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[59], w[60], offset); + w[62] = hc_bytealign_be_S (w[58], w[59], offset); + w[61] = hc_bytealign_be_S (w[57], w[58], offset); + w[60] = hc_bytealign_be_S (w[56], w[57], offset); + w[59] = hc_bytealign_be_S (w[55], w[56], offset); + w[58] = hc_bytealign_be_S (w[54], w[55], offset); + w[57] = hc_bytealign_be_S (w[53], w[54], offset); + w[56] = hc_bytealign_be_S (w[52], w[53], offset); + w[55] = hc_bytealign_be_S (w[51], w[52], offset); + w[54] = hc_bytealign_be_S (w[50], w[51], offset); + w[53] = hc_bytealign_be_S (w[49], w[50], offset); + w[52] = hc_bytealign_be_S (w[48], w[49], offset); + w[51] = hc_bytealign_be_S (w[47], w[48], offset); + w[50] = hc_bytealign_be_S (w[46], w[47], offset); + w[49] = hc_bytealign_be_S (w[45], w[46], offset); + w[48] = hc_bytealign_be_S (w[44], w[45], offset); + w[47] = hc_bytealign_be_S (w[43], w[44], offset); + w[46] = hc_bytealign_be_S (w[42], w[43], offset); + w[45] = hc_bytealign_be_S (w[41], w[42], offset); + w[44] = hc_bytealign_be_S (w[40], w[41], offset); + w[43] = hc_bytealign_be_S (w[39], w[40], offset); + w[42] = hc_bytealign_be_S (w[38], w[39], offset); + w[41] = hc_bytealign_be_S (w[37], w[38], offset); + w[40] = hc_bytealign_be_S (w[36], w[37], offset); + w[39] = hc_bytealign_be_S (w[35], w[36], offset); + w[38] = hc_bytealign_be_S (w[34], w[35], offset); + w[37] = hc_bytealign_be_S (w[33], w[34], offset); + w[36] = hc_bytealign_be_S (w[32], w[33], offset); + w[35] = hc_bytealign_be_S (w[31], w[32], offset); + w[34] = hc_bytealign_be_S (w[30], w[31], offset); + w[33] = hc_bytealign_be_S (w[29], w[30], offset); + w[32] = hc_bytealign_be_S (w[28], w[29], offset); + w[31] = hc_bytealign_be_S (w[27], w[28], offset); + w[30] = hc_bytealign_be_S (w[26], w[27], offset); + w[29] = hc_bytealign_be_S (w[25], w[26], offset); + w[28] = hc_bytealign_be_S (w[24], w[25], offset); + w[27] = hc_bytealign_be_S (w[23], w[24], offset); + w[26] = hc_bytealign_be_S (w[22], w[23], offset); + w[25] = hc_bytealign_be_S (w[21], w[22], offset); + w[24] = hc_bytealign_be_S (w[20], w[21], offset); + w[23] = hc_bytealign_be_S (w[19], w[20], offset); + w[22] = hc_bytealign_be_S (w[18], w[19], offset); + w[21] = hc_bytealign_be_S (w[17], w[18], offset); + w[20] = hc_bytealign_be_S (w[16], w[17], offset); + w[19] = hc_bytealign_be_S (w[15], w[16], offset); + w[18] = hc_bytealign_be_S (w[14], w[15], offset); + w[17] = hc_bytealign_be_S (w[13], w[14], offset); + w[16] = hc_bytealign_be_S (w[12], w[13], offset); + w[15] = hc_bytealign_be_S (w[11], w[12], offset); + w[14] = hc_bytealign_be_S (w[10], w[11], offset); + w[13] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[12] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[11] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[10] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[ 9] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[ 8] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 7] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 6] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 5] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 4] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 3] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 2] = 0; w[ 1] = 0; w[ 0] = 0; @@ -51578,66 +51258,66 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 4: - w[63] = hc_bytealign_S (w[58], w[59], offset); - w[62] = hc_bytealign_S (w[57], w[58], offset); - w[61] = hc_bytealign_S (w[56], w[57], offset); - w[60] = hc_bytealign_S (w[55], w[56], offset); - w[59] = hc_bytealign_S (w[54], w[55], offset); - w[58] = hc_bytealign_S (w[53], w[54], offset); - w[57] = hc_bytealign_S (w[52], w[53], offset); - w[56] = hc_bytealign_S (w[51], w[52], offset); - w[55] = hc_bytealign_S (w[50], w[51], offset); - w[54] = hc_bytealign_S (w[49], w[50], offset); - w[53] = hc_bytealign_S (w[48], w[49], offset); - w[52] = hc_bytealign_S (w[47], w[48], offset); - w[51] = hc_bytealign_S (w[46], w[47], offset); - w[50] = hc_bytealign_S (w[45], w[46], offset); - w[49] = hc_bytealign_S (w[44], w[45], offset); - w[48] = hc_bytealign_S (w[43], w[44], offset); - w[47] = hc_bytealign_S (w[42], w[43], offset); - w[46] = hc_bytealign_S (w[41], w[42], offset); - w[45] = hc_bytealign_S (w[40], w[41], offset); - w[44] = hc_bytealign_S (w[39], w[40], offset); - w[43] = hc_bytealign_S (w[38], w[39], offset); - w[42] = hc_bytealign_S (w[37], w[38], offset); - w[41] = hc_bytealign_S (w[36], w[37], offset); - w[40] = hc_bytealign_S (w[35], w[36], offset); - w[39] = hc_bytealign_S (w[34], w[35], offset); - w[38] = hc_bytealign_S (w[33], w[34], offset); - w[37] = hc_bytealign_S (w[32], w[33], offset); - w[36] = hc_bytealign_S (w[31], w[32], offset); - w[35] = hc_bytealign_S (w[30], w[31], offset); - w[34] = hc_bytealign_S (w[29], w[30], offset); - w[33] = hc_bytealign_S (w[28], w[29], offset); - w[32] = hc_bytealign_S (w[27], w[28], offset); - w[31] = hc_bytealign_S (w[26], w[27], offset); - w[30] = hc_bytealign_S (w[25], w[26], offset); - w[29] = hc_bytealign_S (w[24], w[25], offset); - w[28] = hc_bytealign_S (w[23], w[24], offset); - w[27] = hc_bytealign_S (w[22], w[23], offset); - w[26] = hc_bytealign_S (w[21], w[22], offset); - w[25] = hc_bytealign_S (w[20], w[21], offset); - w[24] = hc_bytealign_S (w[19], w[20], offset); - w[23] = hc_bytealign_S (w[18], w[19], offset); - w[22] = hc_bytealign_S (w[17], w[18], offset); - w[21] = hc_bytealign_S (w[16], w[17], offset); - w[20] = hc_bytealign_S (w[15], w[16], offset); - w[19] = hc_bytealign_S (w[14], w[15], offset); - w[18] = hc_bytealign_S (w[13], w[14], offset); - w[17] = hc_bytealign_S (w[12], w[13], offset); - w[16] = hc_bytealign_S (w[11], w[12], offset); - w[15] = hc_bytealign_S (w[10], w[11], offset); - w[14] = hc_bytealign_S (w[ 9], w[10], offset); - w[13] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[12] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[11] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[10] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[ 9] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 8] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 7] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 6] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 5] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 4] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[58], w[59], offset); + w[62] = hc_bytealign_be_S (w[57], w[58], offset); + w[61] = hc_bytealign_be_S (w[56], w[57], offset); + w[60] = hc_bytealign_be_S (w[55], w[56], offset); + w[59] = hc_bytealign_be_S (w[54], w[55], offset); + w[58] = hc_bytealign_be_S (w[53], w[54], offset); + w[57] = hc_bytealign_be_S (w[52], w[53], offset); + w[56] = hc_bytealign_be_S (w[51], w[52], offset); + w[55] = hc_bytealign_be_S (w[50], w[51], offset); + w[54] = hc_bytealign_be_S (w[49], w[50], offset); + w[53] = hc_bytealign_be_S (w[48], w[49], offset); + w[52] = hc_bytealign_be_S (w[47], w[48], offset); + w[51] = hc_bytealign_be_S (w[46], w[47], offset); + w[50] = hc_bytealign_be_S (w[45], w[46], offset); + w[49] = hc_bytealign_be_S (w[44], w[45], offset); + w[48] = hc_bytealign_be_S (w[43], w[44], offset); + w[47] = hc_bytealign_be_S (w[42], w[43], offset); + w[46] = hc_bytealign_be_S (w[41], w[42], offset); + w[45] = hc_bytealign_be_S (w[40], w[41], offset); + w[44] = hc_bytealign_be_S (w[39], w[40], offset); + w[43] = hc_bytealign_be_S (w[38], w[39], offset); + w[42] = hc_bytealign_be_S (w[37], w[38], offset); + w[41] = hc_bytealign_be_S (w[36], w[37], offset); + w[40] = hc_bytealign_be_S (w[35], w[36], offset); + w[39] = hc_bytealign_be_S (w[34], w[35], offset); + w[38] = hc_bytealign_be_S (w[33], w[34], offset); + w[37] = hc_bytealign_be_S (w[32], w[33], offset); + w[36] = hc_bytealign_be_S (w[31], w[32], offset); + w[35] = hc_bytealign_be_S (w[30], w[31], offset); + w[34] = hc_bytealign_be_S (w[29], w[30], offset); + w[33] = hc_bytealign_be_S (w[28], w[29], offset); + w[32] = hc_bytealign_be_S (w[27], w[28], offset); + w[31] = hc_bytealign_be_S (w[26], w[27], offset); + w[30] = hc_bytealign_be_S (w[25], w[26], offset); + w[29] = hc_bytealign_be_S (w[24], w[25], offset); + w[28] = hc_bytealign_be_S (w[23], w[24], offset); + w[27] = hc_bytealign_be_S (w[22], w[23], offset); + w[26] = hc_bytealign_be_S (w[21], w[22], offset); + w[25] = hc_bytealign_be_S (w[20], w[21], offset); + w[24] = hc_bytealign_be_S (w[19], w[20], offset); + w[23] = hc_bytealign_be_S (w[18], w[19], offset); + w[22] = hc_bytealign_be_S (w[17], w[18], offset); + w[21] = hc_bytealign_be_S (w[16], w[17], offset); + w[20] = hc_bytealign_be_S (w[15], w[16], offset); + w[19] = hc_bytealign_be_S (w[14], w[15], offset); + w[18] = hc_bytealign_be_S (w[13], w[14], offset); + w[17] = hc_bytealign_be_S (w[12], w[13], offset); + w[16] = hc_bytealign_be_S (w[11], w[12], offset); + w[15] = hc_bytealign_be_S (w[10], w[11], offset); + w[14] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[13] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[12] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[11] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[10] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[ 9] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 8] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 7] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 6] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 5] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 4] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 3] = 0; w[ 2] = 0; w[ 1] = 0; @@ -51646,65 +51326,65 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 5: - w[63] = hc_bytealign_S (w[57], w[58], offset); - w[62] = hc_bytealign_S (w[56], w[57], offset); - w[61] = hc_bytealign_S (w[55], w[56], offset); - w[60] = hc_bytealign_S (w[54], w[55], offset); - w[59] = hc_bytealign_S (w[53], w[54], offset); - w[58] = hc_bytealign_S (w[52], w[53], offset); - w[57] = hc_bytealign_S (w[51], w[52], offset); - w[56] = hc_bytealign_S (w[50], w[51], offset); - w[55] = hc_bytealign_S (w[49], w[50], offset); - w[54] = hc_bytealign_S (w[48], w[49], offset); - w[53] = hc_bytealign_S (w[47], w[48], offset); - w[52] = hc_bytealign_S (w[46], w[47], offset); - w[51] = hc_bytealign_S (w[45], w[46], offset); - w[50] = hc_bytealign_S (w[44], w[45], offset); - w[49] = hc_bytealign_S (w[43], w[44], offset); - w[48] = hc_bytealign_S (w[42], w[43], offset); - w[47] = hc_bytealign_S (w[41], w[42], offset); - w[46] = hc_bytealign_S (w[40], w[41], offset); - w[45] = hc_bytealign_S (w[39], w[40], offset); - w[44] = hc_bytealign_S (w[38], w[39], offset); - w[43] = hc_bytealign_S (w[37], w[38], offset); - w[42] = hc_bytealign_S (w[36], w[37], offset); - w[41] = hc_bytealign_S (w[35], w[36], offset); - w[40] = hc_bytealign_S (w[34], w[35], offset); - w[39] = hc_bytealign_S (w[33], w[34], offset); - w[38] = hc_bytealign_S (w[32], w[33], offset); - w[37] = hc_bytealign_S (w[31], w[32], offset); - w[36] = hc_bytealign_S (w[30], w[31], offset); - w[35] = hc_bytealign_S (w[29], w[30], offset); - w[34] = hc_bytealign_S (w[28], w[29], offset); - w[33] = hc_bytealign_S (w[27], w[28], offset); - w[32] = hc_bytealign_S (w[26], w[27], offset); - w[31] = hc_bytealign_S (w[25], w[26], offset); - w[30] = hc_bytealign_S (w[24], w[25], offset); - w[29] = hc_bytealign_S (w[23], w[24], offset); - w[28] = hc_bytealign_S (w[22], w[23], offset); - w[27] = hc_bytealign_S (w[21], w[22], offset); - w[26] = hc_bytealign_S (w[20], w[21], offset); - w[25] = hc_bytealign_S (w[19], w[20], offset); - w[24] = hc_bytealign_S (w[18], w[19], offset); - w[23] = hc_bytealign_S (w[17], w[18], offset); - w[22] = hc_bytealign_S (w[16], w[17], offset); - w[21] = hc_bytealign_S (w[15], w[16], offset); - w[20] = hc_bytealign_S (w[14], w[15], offset); - w[19] = hc_bytealign_S (w[13], w[14], offset); - w[18] = hc_bytealign_S (w[12], w[13], offset); - w[17] = hc_bytealign_S (w[11], w[12], offset); - w[16] = hc_bytealign_S (w[10], w[11], offset); - w[15] = hc_bytealign_S (w[ 9], w[10], offset); - w[14] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[13] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[12] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[11] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[10] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[ 9] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 8] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 7] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 6] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 5] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[57], w[58], offset); + w[62] = hc_bytealign_be_S (w[56], w[57], offset); + w[61] = hc_bytealign_be_S (w[55], w[56], offset); + w[60] = hc_bytealign_be_S (w[54], w[55], offset); + w[59] = hc_bytealign_be_S (w[53], w[54], offset); + w[58] = hc_bytealign_be_S (w[52], w[53], offset); + w[57] = hc_bytealign_be_S (w[51], w[52], offset); + w[56] = hc_bytealign_be_S (w[50], w[51], offset); + w[55] = hc_bytealign_be_S (w[49], w[50], offset); + w[54] = hc_bytealign_be_S (w[48], w[49], offset); + w[53] = hc_bytealign_be_S (w[47], w[48], offset); + w[52] = hc_bytealign_be_S (w[46], w[47], offset); + w[51] = hc_bytealign_be_S (w[45], w[46], offset); + w[50] = hc_bytealign_be_S (w[44], w[45], offset); + w[49] = hc_bytealign_be_S (w[43], w[44], offset); + w[48] = hc_bytealign_be_S (w[42], w[43], offset); + w[47] = hc_bytealign_be_S (w[41], w[42], offset); + w[46] = hc_bytealign_be_S (w[40], w[41], offset); + w[45] = hc_bytealign_be_S (w[39], w[40], offset); + w[44] = hc_bytealign_be_S (w[38], w[39], offset); + w[43] = hc_bytealign_be_S (w[37], w[38], offset); + w[42] = hc_bytealign_be_S (w[36], w[37], offset); + w[41] = hc_bytealign_be_S (w[35], w[36], offset); + w[40] = hc_bytealign_be_S (w[34], w[35], offset); + w[39] = hc_bytealign_be_S (w[33], w[34], offset); + w[38] = hc_bytealign_be_S (w[32], w[33], offset); + w[37] = hc_bytealign_be_S (w[31], w[32], offset); + w[36] = hc_bytealign_be_S (w[30], w[31], offset); + w[35] = hc_bytealign_be_S (w[29], w[30], offset); + w[34] = hc_bytealign_be_S (w[28], w[29], offset); + w[33] = hc_bytealign_be_S (w[27], w[28], offset); + w[32] = hc_bytealign_be_S (w[26], w[27], offset); + w[31] = hc_bytealign_be_S (w[25], w[26], offset); + w[30] = hc_bytealign_be_S (w[24], w[25], offset); + w[29] = hc_bytealign_be_S (w[23], w[24], offset); + w[28] = hc_bytealign_be_S (w[22], w[23], offset); + w[27] = hc_bytealign_be_S (w[21], w[22], offset); + w[26] = hc_bytealign_be_S (w[20], w[21], offset); + w[25] = hc_bytealign_be_S (w[19], w[20], offset); + w[24] = hc_bytealign_be_S (w[18], w[19], offset); + w[23] = hc_bytealign_be_S (w[17], w[18], offset); + w[22] = hc_bytealign_be_S (w[16], w[17], offset); + w[21] = hc_bytealign_be_S (w[15], w[16], offset); + w[20] = hc_bytealign_be_S (w[14], w[15], offset); + w[19] = hc_bytealign_be_S (w[13], w[14], offset); + w[18] = hc_bytealign_be_S (w[12], w[13], offset); + w[17] = hc_bytealign_be_S (w[11], w[12], offset); + w[16] = hc_bytealign_be_S (w[10], w[11], offset); + w[15] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[14] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[13] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[12] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[11] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[10] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[ 9] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 8] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 7] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 6] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 5] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 4] = 0; w[ 3] = 0; w[ 2] = 0; @@ -51714,64 +51394,64 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 6: - w[63] = hc_bytealign_S (w[56], w[57], offset); - w[62] = hc_bytealign_S (w[55], w[56], offset); - w[61] = hc_bytealign_S (w[54], w[55], offset); - w[60] = hc_bytealign_S (w[53], w[54], offset); - w[59] = hc_bytealign_S (w[52], w[53], offset); - w[58] = hc_bytealign_S (w[51], w[52], offset); - w[57] = hc_bytealign_S (w[50], w[51], offset); - w[56] = hc_bytealign_S (w[49], w[50], offset); - w[55] = hc_bytealign_S (w[48], w[49], offset); - w[54] = hc_bytealign_S (w[47], w[48], offset); - w[53] = hc_bytealign_S (w[46], w[47], offset); - w[52] = hc_bytealign_S (w[45], w[46], offset); - w[51] = hc_bytealign_S (w[44], w[45], offset); - w[50] = hc_bytealign_S (w[43], w[44], offset); - w[49] = hc_bytealign_S (w[42], w[43], offset); - w[48] = hc_bytealign_S (w[41], w[42], offset); - w[47] = hc_bytealign_S (w[40], w[41], offset); - w[46] = hc_bytealign_S (w[39], w[40], offset); - w[45] = hc_bytealign_S (w[38], w[39], offset); - w[44] = hc_bytealign_S (w[37], w[38], offset); - w[43] = hc_bytealign_S (w[36], w[37], offset); - w[42] = hc_bytealign_S (w[35], w[36], offset); - w[41] = hc_bytealign_S (w[34], w[35], offset); - w[40] = hc_bytealign_S (w[33], w[34], offset); - w[39] = hc_bytealign_S (w[32], w[33], offset); - w[38] = hc_bytealign_S (w[31], w[32], offset); - w[37] = hc_bytealign_S (w[30], w[31], offset); - w[36] = hc_bytealign_S (w[29], w[30], offset); - w[35] = hc_bytealign_S (w[28], w[29], offset); - w[34] = hc_bytealign_S (w[27], w[28], offset); - w[33] = hc_bytealign_S (w[26], w[27], offset); - w[32] = hc_bytealign_S (w[25], w[26], offset); - w[31] = hc_bytealign_S (w[24], w[25], offset); - w[30] = hc_bytealign_S (w[23], w[24], offset); - w[29] = hc_bytealign_S (w[22], w[23], offset); - w[28] = hc_bytealign_S (w[21], w[22], offset); - w[27] = hc_bytealign_S (w[20], w[21], offset); - w[26] = hc_bytealign_S (w[19], w[20], offset); - w[25] = hc_bytealign_S (w[18], w[19], offset); - w[24] = hc_bytealign_S (w[17], w[18], offset); - w[23] = hc_bytealign_S (w[16], w[17], offset); - w[22] = hc_bytealign_S (w[15], w[16], offset); - w[21] = hc_bytealign_S (w[14], w[15], offset); - w[20] = hc_bytealign_S (w[13], w[14], offset); - w[19] = hc_bytealign_S (w[12], w[13], offset); - w[18] = hc_bytealign_S (w[11], w[12], offset); - w[17] = hc_bytealign_S (w[10], w[11], offset); - w[16] = hc_bytealign_S (w[ 9], w[10], offset); - w[15] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[14] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[13] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[12] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[11] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[10] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[ 9] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 8] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 7] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 6] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[56], w[57], offset); + w[62] = hc_bytealign_be_S (w[55], w[56], offset); + w[61] = hc_bytealign_be_S (w[54], w[55], offset); + w[60] = hc_bytealign_be_S (w[53], w[54], offset); + w[59] = hc_bytealign_be_S (w[52], w[53], offset); + w[58] = hc_bytealign_be_S (w[51], w[52], offset); + w[57] = hc_bytealign_be_S (w[50], w[51], offset); + w[56] = hc_bytealign_be_S (w[49], w[50], offset); + w[55] = hc_bytealign_be_S (w[48], w[49], offset); + w[54] = hc_bytealign_be_S (w[47], w[48], offset); + w[53] = hc_bytealign_be_S (w[46], w[47], offset); + w[52] = hc_bytealign_be_S (w[45], w[46], offset); + w[51] = hc_bytealign_be_S (w[44], w[45], offset); + w[50] = hc_bytealign_be_S (w[43], w[44], offset); + w[49] = hc_bytealign_be_S (w[42], w[43], offset); + w[48] = hc_bytealign_be_S (w[41], w[42], offset); + w[47] = hc_bytealign_be_S (w[40], w[41], offset); + w[46] = hc_bytealign_be_S (w[39], w[40], offset); + w[45] = hc_bytealign_be_S (w[38], w[39], offset); + w[44] = hc_bytealign_be_S (w[37], w[38], offset); + w[43] = hc_bytealign_be_S (w[36], w[37], offset); + w[42] = hc_bytealign_be_S (w[35], w[36], offset); + w[41] = hc_bytealign_be_S (w[34], w[35], offset); + w[40] = hc_bytealign_be_S (w[33], w[34], offset); + w[39] = hc_bytealign_be_S (w[32], w[33], offset); + w[38] = hc_bytealign_be_S (w[31], w[32], offset); + w[37] = hc_bytealign_be_S (w[30], w[31], offset); + w[36] = hc_bytealign_be_S (w[29], w[30], offset); + w[35] = hc_bytealign_be_S (w[28], w[29], offset); + w[34] = hc_bytealign_be_S (w[27], w[28], offset); + w[33] = hc_bytealign_be_S (w[26], w[27], offset); + w[32] = hc_bytealign_be_S (w[25], w[26], offset); + w[31] = hc_bytealign_be_S (w[24], w[25], offset); + w[30] = hc_bytealign_be_S (w[23], w[24], offset); + w[29] = hc_bytealign_be_S (w[22], w[23], offset); + w[28] = hc_bytealign_be_S (w[21], w[22], offset); + w[27] = hc_bytealign_be_S (w[20], w[21], offset); + w[26] = hc_bytealign_be_S (w[19], w[20], offset); + w[25] = hc_bytealign_be_S (w[18], w[19], offset); + w[24] = hc_bytealign_be_S (w[17], w[18], offset); + w[23] = hc_bytealign_be_S (w[16], w[17], offset); + w[22] = hc_bytealign_be_S (w[15], w[16], offset); + w[21] = hc_bytealign_be_S (w[14], w[15], offset); + w[20] = hc_bytealign_be_S (w[13], w[14], offset); + w[19] = hc_bytealign_be_S (w[12], w[13], offset); + w[18] = hc_bytealign_be_S (w[11], w[12], offset); + w[17] = hc_bytealign_be_S (w[10], w[11], offset); + w[16] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[15] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[14] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[13] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[12] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[11] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[10] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[ 9] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 8] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 7] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 6] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 5] = 0; w[ 4] = 0; w[ 3] = 0; @@ -51782,63 +51462,63 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 7: - w[63] = hc_bytealign_S (w[55], w[56], offset); - w[62] = hc_bytealign_S (w[54], w[55], offset); - w[61] = hc_bytealign_S (w[53], w[54], offset); - w[60] = hc_bytealign_S (w[52], w[53], offset); - w[59] = hc_bytealign_S (w[51], w[52], offset); - w[58] = hc_bytealign_S (w[50], w[51], offset); - w[57] = hc_bytealign_S (w[49], w[50], offset); - w[56] = hc_bytealign_S (w[48], w[49], offset); - w[55] = hc_bytealign_S (w[47], w[48], offset); - w[54] = hc_bytealign_S (w[46], w[47], offset); - w[53] = hc_bytealign_S (w[45], w[46], offset); - w[52] = hc_bytealign_S (w[44], w[45], offset); - w[51] = hc_bytealign_S (w[43], w[44], offset); - w[50] = hc_bytealign_S (w[42], w[43], offset); - w[49] = hc_bytealign_S (w[41], w[42], offset); - w[48] = hc_bytealign_S (w[40], w[41], offset); - w[47] = hc_bytealign_S (w[39], w[40], offset); - w[46] = hc_bytealign_S (w[38], w[39], offset); - w[45] = hc_bytealign_S (w[37], w[38], offset); - w[44] = hc_bytealign_S (w[36], w[37], offset); - w[43] = hc_bytealign_S (w[35], w[36], offset); - w[42] = hc_bytealign_S (w[34], w[35], offset); - w[41] = hc_bytealign_S (w[33], w[34], offset); - w[40] = hc_bytealign_S (w[32], w[33], offset); - w[39] = hc_bytealign_S (w[31], w[32], offset); - w[38] = hc_bytealign_S (w[30], w[31], offset); - w[37] = hc_bytealign_S (w[29], w[30], offset); - w[36] = hc_bytealign_S (w[28], w[29], offset); - w[35] = hc_bytealign_S (w[27], w[28], offset); - w[34] = hc_bytealign_S (w[26], w[27], offset); - w[33] = hc_bytealign_S (w[25], w[26], offset); - w[32] = hc_bytealign_S (w[24], w[25], offset); - w[31] = hc_bytealign_S (w[23], w[24], offset); - w[30] = hc_bytealign_S (w[22], w[23], offset); - w[29] = hc_bytealign_S (w[21], w[22], offset); - w[28] = hc_bytealign_S (w[20], w[21], offset); - w[27] = hc_bytealign_S (w[19], w[20], offset); - w[26] = hc_bytealign_S (w[18], w[19], offset); - w[25] = hc_bytealign_S (w[17], w[18], offset); - w[24] = hc_bytealign_S (w[16], w[17], offset); - w[23] = hc_bytealign_S (w[15], w[16], offset); - w[22] = hc_bytealign_S (w[14], w[15], offset); - w[21] = hc_bytealign_S (w[13], w[14], offset); - w[20] = hc_bytealign_S (w[12], w[13], offset); - w[19] = hc_bytealign_S (w[11], w[12], offset); - w[18] = hc_bytealign_S (w[10], w[11], offset); - w[17] = hc_bytealign_S (w[ 9], w[10], offset); - w[16] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[15] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[14] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[13] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[12] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[11] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[10] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[ 9] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 8] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 7] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[55], w[56], offset); + w[62] = hc_bytealign_be_S (w[54], w[55], offset); + w[61] = hc_bytealign_be_S (w[53], w[54], offset); + w[60] = hc_bytealign_be_S (w[52], w[53], offset); + w[59] = hc_bytealign_be_S (w[51], w[52], offset); + w[58] = hc_bytealign_be_S (w[50], w[51], offset); + w[57] = hc_bytealign_be_S (w[49], w[50], offset); + w[56] = hc_bytealign_be_S (w[48], w[49], offset); + w[55] = hc_bytealign_be_S (w[47], w[48], offset); + w[54] = hc_bytealign_be_S (w[46], w[47], offset); + w[53] = hc_bytealign_be_S (w[45], w[46], offset); + w[52] = hc_bytealign_be_S (w[44], w[45], offset); + w[51] = hc_bytealign_be_S (w[43], w[44], offset); + w[50] = hc_bytealign_be_S (w[42], w[43], offset); + w[49] = hc_bytealign_be_S (w[41], w[42], offset); + w[48] = hc_bytealign_be_S (w[40], w[41], offset); + w[47] = hc_bytealign_be_S (w[39], w[40], offset); + w[46] = hc_bytealign_be_S (w[38], w[39], offset); + w[45] = hc_bytealign_be_S (w[37], w[38], offset); + w[44] = hc_bytealign_be_S (w[36], w[37], offset); + w[43] = hc_bytealign_be_S (w[35], w[36], offset); + w[42] = hc_bytealign_be_S (w[34], w[35], offset); + w[41] = hc_bytealign_be_S (w[33], w[34], offset); + w[40] = hc_bytealign_be_S (w[32], w[33], offset); + w[39] = hc_bytealign_be_S (w[31], w[32], offset); + w[38] = hc_bytealign_be_S (w[30], w[31], offset); + w[37] = hc_bytealign_be_S (w[29], w[30], offset); + w[36] = hc_bytealign_be_S (w[28], w[29], offset); + w[35] = hc_bytealign_be_S (w[27], w[28], offset); + w[34] = hc_bytealign_be_S (w[26], w[27], offset); + w[33] = hc_bytealign_be_S (w[25], w[26], offset); + w[32] = hc_bytealign_be_S (w[24], w[25], offset); + w[31] = hc_bytealign_be_S (w[23], w[24], offset); + w[30] = hc_bytealign_be_S (w[22], w[23], offset); + w[29] = hc_bytealign_be_S (w[21], w[22], offset); + w[28] = hc_bytealign_be_S (w[20], w[21], offset); + w[27] = hc_bytealign_be_S (w[19], w[20], offset); + w[26] = hc_bytealign_be_S (w[18], w[19], offset); + w[25] = hc_bytealign_be_S (w[17], w[18], offset); + w[24] = hc_bytealign_be_S (w[16], w[17], offset); + w[23] = hc_bytealign_be_S (w[15], w[16], offset); + w[22] = hc_bytealign_be_S (w[14], w[15], offset); + w[21] = hc_bytealign_be_S (w[13], w[14], offset); + w[20] = hc_bytealign_be_S (w[12], w[13], offset); + w[19] = hc_bytealign_be_S (w[11], w[12], offset); + w[18] = hc_bytealign_be_S (w[10], w[11], offset); + w[17] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[16] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[15] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[14] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[13] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[12] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[11] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[10] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[ 9] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 8] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 7] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 6] = 0; w[ 5] = 0; w[ 4] = 0; @@ -51850,62 +51530,62 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 8: - w[63] = hc_bytealign_S (w[54], w[55], offset); - w[62] = hc_bytealign_S (w[53], w[54], offset); - w[61] = hc_bytealign_S (w[52], w[53], offset); - w[60] = hc_bytealign_S (w[51], w[52], offset); - w[59] = hc_bytealign_S (w[50], w[51], offset); - w[58] = hc_bytealign_S (w[49], w[50], offset); - w[57] = hc_bytealign_S (w[48], w[49], offset); - w[56] = hc_bytealign_S (w[47], w[48], offset); - w[55] = hc_bytealign_S (w[46], w[47], offset); - w[54] = hc_bytealign_S (w[45], w[46], offset); - w[53] = hc_bytealign_S (w[44], w[45], offset); - w[52] = hc_bytealign_S (w[43], w[44], offset); - w[51] = hc_bytealign_S (w[42], w[43], offset); - w[50] = hc_bytealign_S (w[41], w[42], offset); - w[49] = hc_bytealign_S (w[40], w[41], offset); - w[48] = hc_bytealign_S (w[39], w[40], offset); - w[47] = hc_bytealign_S (w[38], w[39], offset); - w[46] = hc_bytealign_S (w[37], w[38], offset); - w[45] = hc_bytealign_S (w[36], w[37], offset); - w[44] = hc_bytealign_S (w[35], w[36], offset); - w[43] = hc_bytealign_S (w[34], w[35], offset); - w[42] = hc_bytealign_S (w[33], w[34], offset); - w[41] = hc_bytealign_S (w[32], w[33], offset); - w[40] = hc_bytealign_S (w[31], w[32], offset); - w[39] = hc_bytealign_S (w[30], w[31], offset); - w[38] = hc_bytealign_S (w[29], w[30], offset); - w[37] = hc_bytealign_S (w[28], w[29], offset); - w[36] = hc_bytealign_S (w[27], w[28], offset); - w[35] = hc_bytealign_S (w[26], w[27], offset); - w[34] = hc_bytealign_S (w[25], w[26], offset); - w[33] = hc_bytealign_S (w[24], w[25], offset); - w[32] = hc_bytealign_S (w[23], w[24], offset); - w[31] = hc_bytealign_S (w[22], w[23], offset); - w[30] = hc_bytealign_S (w[21], w[22], offset); - w[29] = hc_bytealign_S (w[20], w[21], offset); - w[28] = hc_bytealign_S (w[19], w[20], offset); - w[27] = hc_bytealign_S (w[18], w[19], offset); - w[26] = hc_bytealign_S (w[17], w[18], offset); - w[25] = hc_bytealign_S (w[16], w[17], offset); - w[24] = hc_bytealign_S (w[15], w[16], offset); - w[23] = hc_bytealign_S (w[14], w[15], offset); - w[22] = hc_bytealign_S (w[13], w[14], offset); - w[21] = hc_bytealign_S (w[12], w[13], offset); - w[20] = hc_bytealign_S (w[11], w[12], offset); - w[19] = hc_bytealign_S (w[10], w[11], offset); - w[18] = hc_bytealign_S (w[ 9], w[10], offset); - w[17] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[16] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[15] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[14] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[13] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[12] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[11] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[10] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[ 9] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 8] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[54], w[55], offset); + w[62] = hc_bytealign_be_S (w[53], w[54], offset); + w[61] = hc_bytealign_be_S (w[52], w[53], offset); + w[60] = hc_bytealign_be_S (w[51], w[52], offset); + w[59] = hc_bytealign_be_S (w[50], w[51], offset); + w[58] = hc_bytealign_be_S (w[49], w[50], offset); + w[57] = hc_bytealign_be_S (w[48], w[49], offset); + w[56] = hc_bytealign_be_S (w[47], w[48], offset); + w[55] = hc_bytealign_be_S (w[46], w[47], offset); + w[54] = hc_bytealign_be_S (w[45], w[46], offset); + w[53] = hc_bytealign_be_S (w[44], w[45], offset); + w[52] = hc_bytealign_be_S (w[43], w[44], offset); + w[51] = hc_bytealign_be_S (w[42], w[43], offset); + w[50] = hc_bytealign_be_S (w[41], w[42], offset); + w[49] = hc_bytealign_be_S (w[40], w[41], offset); + w[48] = hc_bytealign_be_S (w[39], w[40], offset); + w[47] = hc_bytealign_be_S (w[38], w[39], offset); + w[46] = hc_bytealign_be_S (w[37], w[38], offset); + w[45] = hc_bytealign_be_S (w[36], w[37], offset); + w[44] = hc_bytealign_be_S (w[35], w[36], offset); + w[43] = hc_bytealign_be_S (w[34], w[35], offset); + w[42] = hc_bytealign_be_S (w[33], w[34], offset); + w[41] = hc_bytealign_be_S (w[32], w[33], offset); + w[40] = hc_bytealign_be_S (w[31], w[32], offset); + w[39] = hc_bytealign_be_S (w[30], w[31], offset); + w[38] = hc_bytealign_be_S (w[29], w[30], offset); + w[37] = hc_bytealign_be_S (w[28], w[29], offset); + w[36] = hc_bytealign_be_S (w[27], w[28], offset); + w[35] = hc_bytealign_be_S (w[26], w[27], offset); + w[34] = hc_bytealign_be_S (w[25], w[26], offset); + w[33] = hc_bytealign_be_S (w[24], w[25], offset); + w[32] = hc_bytealign_be_S (w[23], w[24], offset); + w[31] = hc_bytealign_be_S (w[22], w[23], offset); + w[30] = hc_bytealign_be_S (w[21], w[22], offset); + w[29] = hc_bytealign_be_S (w[20], w[21], offset); + w[28] = hc_bytealign_be_S (w[19], w[20], offset); + w[27] = hc_bytealign_be_S (w[18], w[19], offset); + w[26] = hc_bytealign_be_S (w[17], w[18], offset); + w[25] = hc_bytealign_be_S (w[16], w[17], offset); + w[24] = hc_bytealign_be_S (w[15], w[16], offset); + w[23] = hc_bytealign_be_S (w[14], w[15], offset); + w[22] = hc_bytealign_be_S (w[13], w[14], offset); + w[21] = hc_bytealign_be_S (w[12], w[13], offset); + w[20] = hc_bytealign_be_S (w[11], w[12], offset); + w[19] = hc_bytealign_be_S (w[10], w[11], offset); + w[18] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[17] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[16] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[15] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[14] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[13] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[12] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[11] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[10] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[ 9] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 8] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 7] = 0; w[ 6] = 0; w[ 5] = 0; @@ -51918,61 +51598,61 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 9: - w[63] = hc_bytealign_S (w[53], w[54], offset); - w[62] = hc_bytealign_S (w[52], w[53], offset); - w[61] = hc_bytealign_S (w[51], w[52], offset); - w[60] = hc_bytealign_S (w[50], w[51], offset); - w[59] = hc_bytealign_S (w[49], w[50], offset); - w[58] = hc_bytealign_S (w[48], w[49], offset); - w[57] = hc_bytealign_S (w[47], w[48], offset); - w[56] = hc_bytealign_S (w[46], w[47], offset); - w[55] = hc_bytealign_S (w[45], w[46], offset); - w[54] = hc_bytealign_S (w[44], w[45], offset); - w[53] = hc_bytealign_S (w[43], w[44], offset); - w[52] = hc_bytealign_S (w[42], w[43], offset); - w[51] = hc_bytealign_S (w[41], w[42], offset); - w[50] = hc_bytealign_S (w[40], w[41], offset); - w[49] = hc_bytealign_S (w[39], w[40], offset); - w[48] = hc_bytealign_S (w[38], w[39], offset); - w[47] = hc_bytealign_S (w[37], w[38], offset); - w[46] = hc_bytealign_S (w[36], w[37], offset); - w[45] = hc_bytealign_S (w[35], w[36], offset); - w[44] = hc_bytealign_S (w[34], w[35], offset); - w[43] = hc_bytealign_S (w[33], w[34], offset); - w[42] = hc_bytealign_S (w[32], w[33], offset); - w[41] = hc_bytealign_S (w[31], w[32], offset); - w[40] = hc_bytealign_S (w[30], w[31], offset); - w[39] = hc_bytealign_S (w[29], w[30], offset); - w[38] = hc_bytealign_S (w[28], w[29], offset); - w[37] = hc_bytealign_S (w[27], w[28], offset); - w[36] = hc_bytealign_S (w[26], w[27], offset); - w[35] = hc_bytealign_S (w[25], w[26], offset); - w[34] = hc_bytealign_S (w[24], w[25], offset); - w[33] = hc_bytealign_S (w[23], w[24], offset); - w[32] = hc_bytealign_S (w[22], w[23], offset); - w[31] = hc_bytealign_S (w[21], w[22], offset); - w[30] = hc_bytealign_S (w[20], w[21], offset); - w[29] = hc_bytealign_S (w[19], w[20], offset); - w[28] = hc_bytealign_S (w[18], w[19], offset); - w[27] = hc_bytealign_S (w[17], w[18], offset); - w[26] = hc_bytealign_S (w[16], w[17], offset); - w[25] = hc_bytealign_S (w[15], w[16], offset); - w[24] = hc_bytealign_S (w[14], w[15], offset); - w[23] = hc_bytealign_S (w[13], w[14], offset); - w[22] = hc_bytealign_S (w[12], w[13], offset); - w[21] = hc_bytealign_S (w[11], w[12], offset); - w[20] = hc_bytealign_S (w[10], w[11], offset); - w[19] = hc_bytealign_S (w[ 9], w[10], offset); - w[18] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[17] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[16] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[15] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[14] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[13] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[12] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[11] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[10] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[ 9] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[53], w[54], offset); + w[62] = hc_bytealign_be_S (w[52], w[53], offset); + w[61] = hc_bytealign_be_S (w[51], w[52], offset); + w[60] = hc_bytealign_be_S (w[50], w[51], offset); + w[59] = hc_bytealign_be_S (w[49], w[50], offset); + w[58] = hc_bytealign_be_S (w[48], w[49], offset); + w[57] = hc_bytealign_be_S (w[47], w[48], offset); + w[56] = hc_bytealign_be_S (w[46], w[47], offset); + w[55] = hc_bytealign_be_S (w[45], w[46], offset); + w[54] = hc_bytealign_be_S (w[44], w[45], offset); + w[53] = hc_bytealign_be_S (w[43], w[44], offset); + w[52] = hc_bytealign_be_S (w[42], w[43], offset); + w[51] = hc_bytealign_be_S (w[41], w[42], offset); + w[50] = hc_bytealign_be_S (w[40], w[41], offset); + w[49] = hc_bytealign_be_S (w[39], w[40], offset); + w[48] = hc_bytealign_be_S (w[38], w[39], offset); + w[47] = hc_bytealign_be_S (w[37], w[38], offset); + w[46] = hc_bytealign_be_S (w[36], w[37], offset); + w[45] = hc_bytealign_be_S (w[35], w[36], offset); + w[44] = hc_bytealign_be_S (w[34], w[35], offset); + w[43] = hc_bytealign_be_S (w[33], w[34], offset); + w[42] = hc_bytealign_be_S (w[32], w[33], offset); + w[41] = hc_bytealign_be_S (w[31], w[32], offset); + w[40] = hc_bytealign_be_S (w[30], w[31], offset); + w[39] = hc_bytealign_be_S (w[29], w[30], offset); + w[38] = hc_bytealign_be_S (w[28], w[29], offset); + w[37] = hc_bytealign_be_S (w[27], w[28], offset); + w[36] = hc_bytealign_be_S (w[26], w[27], offset); + w[35] = hc_bytealign_be_S (w[25], w[26], offset); + w[34] = hc_bytealign_be_S (w[24], w[25], offset); + w[33] = hc_bytealign_be_S (w[23], w[24], offset); + w[32] = hc_bytealign_be_S (w[22], w[23], offset); + w[31] = hc_bytealign_be_S (w[21], w[22], offset); + w[30] = hc_bytealign_be_S (w[20], w[21], offset); + w[29] = hc_bytealign_be_S (w[19], w[20], offset); + w[28] = hc_bytealign_be_S (w[18], w[19], offset); + w[27] = hc_bytealign_be_S (w[17], w[18], offset); + w[26] = hc_bytealign_be_S (w[16], w[17], offset); + w[25] = hc_bytealign_be_S (w[15], w[16], offset); + w[24] = hc_bytealign_be_S (w[14], w[15], offset); + w[23] = hc_bytealign_be_S (w[13], w[14], offset); + w[22] = hc_bytealign_be_S (w[12], w[13], offset); + w[21] = hc_bytealign_be_S (w[11], w[12], offset); + w[20] = hc_bytealign_be_S (w[10], w[11], offset); + w[19] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[18] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[17] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[16] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[15] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[14] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[13] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[12] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[11] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[10] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[ 9] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 8] = 0; w[ 7] = 0; w[ 6] = 0; @@ -51986,60 +51666,60 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 10: - w[63] = hc_bytealign_S (w[52], w[53], offset); - w[62] = hc_bytealign_S (w[51], w[52], offset); - w[61] = hc_bytealign_S (w[50], w[51], offset); - w[60] = hc_bytealign_S (w[49], w[50], offset); - w[59] = hc_bytealign_S (w[48], w[49], offset); - w[58] = hc_bytealign_S (w[47], w[48], offset); - w[57] = hc_bytealign_S (w[46], w[47], offset); - w[56] = hc_bytealign_S (w[45], w[46], offset); - w[55] = hc_bytealign_S (w[44], w[45], offset); - w[54] = hc_bytealign_S (w[43], w[44], offset); - w[53] = hc_bytealign_S (w[42], w[43], offset); - w[52] = hc_bytealign_S (w[41], w[42], offset); - w[51] = hc_bytealign_S (w[40], w[41], offset); - w[50] = hc_bytealign_S (w[39], w[40], offset); - w[49] = hc_bytealign_S (w[38], w[39], offset); - w[48] = hc_bytealign_S (w[37], w[38], offset); - w[47] = hc_bytealign_S (w[36], w[37], offset); - w[46] = hc_bytealign_S (w[35], w[36], offset); - w[45] = hc_bytealign_S (w[34], w[35], offset); - w[44] = hc_bytealign_S (w[33], w[34], offset); - w[43] = hc_bytealign_S (w[32], w[33], offset); - w[42] = hc_bytealign_S (w[31], w[32], offset); - w[41] = hc_bytealign_S (w[30], w[31], offset); - w[40] = hc_bytealign_S (w[29], w[30], offset); - w[39] = hc_bytealign_S (w[28], w[29], offset); - w[38] = hc_bytealign_S (w[27], w[28], offset); - w[37] = hc_bytealign_S (w[26], w[27], offset); - w[36] = hc_bytealign_S (w[25], w[26], offset); - w[35] = hc_bytealign_S (w[24], w[25], offset); - w[34] = hc_bytealign_S (w[23], w[24], offset); - w[33] = hc_bytealign_S (w[22], w[23], offset); - w[32] = hc_bytealign_S (w[21], w[22], offset); - w[31] = hc_bytealign_S (w[20], w[21], offset); - w[30] = hc_bytealign_S (w[19], w[20], offset); - w[29] = hc_bytealign_S (w[18], w[19], offset); - w[28] = hc_bytealign_S (w[17], w[18], offset); - w[27] = hc_bytealign_S (w[16], w[17], offset); - w[26] = hc_bytealign_S (w[15], w[16], offset); - w[25] = hc_bytealign_S (w[14], w[15], offset); - w[24] = hc_bytealign_S (w[13], w[14], offset); - w[23] = hc_bytealign_S (w[12], w[13], offset); - w[22] = hc_bytealign_S (w[11], w[12], offset); - w[21] = hc_bytealign_S (w[10], w[11], offset); - w[20] = hc_bytealign_S (w[ 9], w[10], offset); - w[19] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[18] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[17] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[16] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[15] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[14] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[13] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[12] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[11] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[10] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[52], w[53], offset); + w[62] = hc_bytealign_be_S (w[51], w[52], offset); + w[61] = hc_bytealign_be_S (w[50], w[51], offset); + w[60] = hc_bytealign_be_S (w[49], w[50], offset); + w[59] = hc_bytealign_be_S (w[48], w[49], offset); + w[58] = hc_bytealign_be_S (w[47], w[48], offset); + w[57] = hc_bytealign_be_S (w[46], w[47], offset); + w[56] = hc_bytealign_be_S (w[45], w[46], offset); + w[55] = hc_bytealign_be_S (w[44], w[45], offset); + w[54] = hc_bytealign_be_S (w[43], w[44], offset); + w[53] = hc_bytealign_be_S (w[42], w[43], offset); + w[52] = hc_bytealign_be_S (w[41], w[42], offset); + w[51] = hc_bytealign_be_S (w[40], w[41], offset); + w[50] = hc_bytealign_be_S (w[39], w[40], offset); + w[49] = hc_bytealign_be_S (w[38], w[39], offset); + w[48] = hc_bytealign_be_S (w[37], w[38], offset); + w[47] = hc_bytealign_be_S (w[36], w[37], offset); + w[46] = hc_bytealign_be_S (w[35], w[36], offset); + w[45] = hc_bytealign_be_S (w[34], w[35], offset); + w[44] = hc_bytealign_be_S (w[33], w[34], offset); + w[43] = hc_bytealign_be_S (w[32], w[33], offset); + w[42] = hc_bytealign_be_S (w[31], w[32], offset); + w[41] = hc_bytealign_be_S (w[30], w[31], offset); + w[40] = hc_bytealign_be_S (w[29], w[30], offset); + w[39] = hc_bytealign_be_S (w[28], w[29], offset); + w[38] = hc_bytealign_be_S (w[27], w[28], offset); + w[37] = hc_bytealign_be_S (w[26], w[27], offset); + w[36] = hc_bytealign_be_S (w[25], w[26], offset); + w[35] = hc_bytealign_be_S (w[24], w[25], offset); + w[34] = hc_bytealign_be_S (w[23], w[24], offset); + w[33] = hc_bytealign_be_S (w[22], w[23], offset); + w[32] = hc_bytealign_be_S (w[21], w[22], offset); + w[31] = hc_bytealign_be_S (w[20], w[21], offset); + w[30] = hc_bytealign_be_S (w[19], w[20], offset); + w[29] = hc_bytealign_be_S (w[18], w[19], offset); + w[28] = hc_bytealign_be_S (w[17], w[18], offset); + w[27] = hc_bytealign_be_S (w[16], w[17], offset); + w[26] = hc_bytealign_be_S (w[15], w[16], offset); + w[25] = hc_bytealign_be_S (w[14], w[15], offset); + w[24] = hc_bytealign_be_S (w[13], w[14], offset); + w[23] = hc_bytealign_be_S (w[12], w[13], offset); + w[22] = hc_bytealign_be_S (w[11], w[12], offset); + w[21] = hc_bytealign_be_S (w[10], w[11], offset); + w[20] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[19] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[18] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[17] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[16] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[15] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[14] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[13] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[12] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[11] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[10] = hc_bytealign_be_S ( 0, w[ 0], offset); w[ 9] = 0; w[ 8] = 0; w[ 7] = 0; @@ -52054,59 +51734,59 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 11: - w[63] = hc_bytealign_S (w[51], w[52], offset); - w[62] = hc_bytealign_S (w[50], w[51], offset); - w[61] = hc_bytealign_S (w[49], w[50], offset); - w[60] = hc_bytealign_S (w[48], w[49], offset); - w[59] = hc_bytealign_S (w[47], w[48], offset); - w[58] = hc_bytealign_S (w[46], w[47], offset); - w[57] = hc_bytealign_S (w[45], w[46], offset); - w[56] = hc_bytealign_S (w[44], w[45], offset); - w[55] = hc_bytealign_S (w[43], w[44], offset); - w[54] = hc_bytealign_S (w[42], w[43], offset); - w[53] = hc_bytealign_S (w[41], w[42], offset); - w[52] = hc_bytealign_S (w[40], w[41], offset); - w[51] = hc_bytealign_S (w[39], w[40], offset); - w[50] = hc_bytealign_S (w[38], w[39], offset); - w[49] = hc_bytealign_S (w[37], w[38], offset); - w[48] = hc_bytealign_S (w[36], w[37], offset); - w[47] = hc_bytealign_S (w[35], w[36], offset); - w[46] = hc_bytealign_S (w[34], w[35], offset); - w[45] = hc_bytealign_S (w[33], w[34], offset); - w[44] = hc_bytealign_S (w[32], w[33], offset); - w[43] = hc_bytealign_S (w[31], w[32], offset); - w[42] = hc_bytealign_S (w[30], w[31], offset); - w[41] = hc_bytealign_S (w[29], w[30], offset); - w[40] = hc_bytealign_S (w[28], w[29], offset); - w[39] = hc_bytealign_S (w[27], w[28], offset); - w[38] = hc_bytealign_S (w[26], w[27], offset); - w[37] = hc_bytealign_S (w[25], w[26], offset); - w[36] = hc_bytealign_S (w[24], w[25], offset); - w[35] = hc_bytealign_S (w[23], w[24], offset); - w[34] = hc_bytealign_S (w[22], w[23], offset); - w[33] = hc_bytealign_S (w[21], w[22], offset); - w[32] = hc_bytealign_S (w[20], w[21], offset); - w[31] = hc_bytealign_S (w[19], w[20], offset); - w[30] = hc_bytealign_S (w[18], w[19], offset); - w[29] = hc_bytealign_S (w[17], w[18], offset); - w[28] = hc_bytealign_S (w[16], w[17], offset); - w[27] = hc_bytealign_S (w[15], w[16], offset); - w[26] = hc_bytealign_S (w[14], w[15], offset); - w[25] = hc_bytealign_S (w[13], w[14], offset); - w[24] = hc_bytealign_S (w[12], w[13], offset); - w[23] = hc_bytealign_S (w[11], w[12], offset); - w[22] = hc_bytealign_S (w[10], w[11], offset); - w[21] = hc_bytealign_S (w[ 9], w[10], offset); - w[20] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[19] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[18] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[17] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[16] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[15] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[14] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[13] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[12] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[11] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[51], w[52], offset); + w[62] = hc_bytealign_be_S (w[50], w[51], offset); + w[61] = hc_bytealign_be_S (w[49], w[50], offset); + w[60] = hc_bytealign_be_S (w[48], w[49], offset); + w[59] = hc_bytealign_be_S (w[47], w[48], offset); + w[58] = hc_bytealign_be_S (w[46], w[47], offset); + w[57] = hc_bytealign_be_S (w[45], w[46], offset); + w[56] = hc_bytealign_be_S (w[44], w[45], offset); + w[55] = hc_bytealign_be_S (w[43], w[44], offset); + w[54] = hc_bytealign_be_S (w[42], w[43], offset); + w[53] = hc_bytealign_be_S (w[41], w[42], offset); + w[52] = hc_bytealign_be_S (w[40], w[41], offset); + w[51] = hc_bytealign_be_S (w[39], w[40], offset); + w[50] = hc_bytealign_be_S (w[38], w[39], offset); + w[49] = hc_bytealign_be_S (w[37], w[38], offset); + w[48] = hc_bytealign_be_S (w[36], w[37], offset); + w[47] = hc_bytealign_be_S (w[35], w[36], offset); + w[46] = hc_bytealign_be_S (w[34], w[35], offset); + w[45] = hc_bytealign_be_S (w[33], w[34], offset); + w[44] = hc_bytealign_be_S (w[32], w[33], offset); + w[43] = hc_bytealign_be_S (w[31], w[32], offset); + w[42] = hc_bytealign_be_S (w[30], w[31], offset); + w[41] = hc_bytealign_be_S (w[29], w[30], offset); + w[40] = hc_bytealign_be_S (w[28], w[29], offset); + w[39] = hc_bytealign_be_S (w[27], w[28], offset); + w[38] = hc_bytealign_be_S (w[26], w[27], offset); + w[37] = hc_bytealign_be_S (w[25], w[26], offset); + w[36] = hc_bytealign_be_S (w[24], w[25], offset); + w[35] = hc_bytealign_be_S (w[23], w[24], offset); + w[34] = hc_bytealign_be_S (w[22], w[23], offset); + w[33] = hc_bytealign_be_S (w[21], w[22], offset); + w[32] = hc_bytealign_be_S (w[20], w[21], offset); + w[31] = hc_bytealign_be_S (w[19], w[20], offset); + w[30] = hc_bytealign_be_S (w[18], w[19], offset); + w[29] = hc_bytealign_be_S (w[17], w[18], offset); + w[28] = hc_bytealign_be_S (w[16], w[17], offset); + w[27] = hc_bytealign_be_S (w[15], w[16], offset); + w[26] = hc_bytealign_be_S (w[14], w[15], offset); + w[25] = hc_bytealign_be_S (w[13], w[14], offset); + w[24] = hc_bytealign_be_S (w[12], w[13], offset); + w[23] = hc_bytealign_be_S (w[11], w[12], offset); + w[22] = hc_bytealign_be_S (w[10], w[11], offset); + w[21] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[20] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[19] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[18] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[17] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[16] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[15] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[14] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[13] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[12] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[11] = hc_bytealign_be_S ( 0, w[ 0], offset); w[10] = 0; w[ 9] = 0; w[ 8] = 0; @@ -52122,58 +51802,58 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 12: - w[63] = hc_bytealign_S (w[50], w[51], offset); - w[62] = hc_bytealign_S (w[49], w[50], offset); - w[61] = hc_bytealign_S (w[48], w[49], offset); - w[60] = hc_bytealign_S (w[47], w[48], offset); - w[59] = hc_bytealign_S (w[46], w[47], offset); - w[58] = hc_bytealign_S (w[45], w[46], offset); - w[57] = hc_bytealign_S (w[44], w[45], offset); - w[56] = hc_bytealign_S (w[43], w[44], offset); - w[55] = hc_bytealign_S (w[42], w[43], offset); - w[54] = hc_bytealign_S (w[41], w[42], offset); - w[53] = hc_bytealign_S (w[40], w[41], offset); - w[52] = hc_bytealign_S (w[39], w[40], offset); - w[51] = hc_bytealign_S (w[38], w[39], offset); - w[50] = hc_bytealign_S (w[37], w[38], offset); - w[49] = hc_bytealign_S (w[36], w[37], offset); - w[48] = hc_bytealign_S (w[35], w[36], offset); - w[47] = hc_bytealign_S (w[34], w[35], offset); - w[46] = hc_bytealign_S (w[33], w[34], offset); - w[45] = hc_bytealign_S (w[32], w[33], offset); - w[44] = hc_bytealign_S (w[31], w[32], offset); - w[43] = hc_bytealign_S (w[30], w[31], offset); - w[42] = hc_bytealign_S (w[29], w[30], offset); - w[41] = hc_bytealign_S (w[28], w[29], offset); - w[40] = hc_bytealign_S (w[27], w[28], offset); - w[39] = hc_bytealign_S (w[26], w[27], offset); - w[38] = hc_bytealign_S (w[25], w[26], offset); - w[37] = hc_bytealign_S (w[24], w[25], offset); - w[36] = hc_bytealign_S (w[23], w[24], offset); - w[35] = hc_bytealign_S (w[22], w[23], offset); - w[34] = hc_bytealign_S (w[21], w[22], offset); - w[33] = hc_bytealign_S (w[20], w[21], offset); - w[32] = hc_bytealign_S (w[19], w[20], offset); - w[31] = hc_bytealign_S (w[18], w[19], offset); - w[30] = hc_bytealign_S (w[17], w[18], offset); - w[29] = hc_bytealign_S (w[16], w[17], offset); - w[28] = hc_bytealign_S (w[15], w[16], offset); - w[27] = hc_bytealign_S (w[14], w[15], offset); - w[26] = hc_bytealign_S (w[13], w[14], offset); - w[25] = hc_bytealign_S (w[12], w[13], offset); - w[24] = hc_bytealign_S (w[11], w[12], offset); - w[23] = hc_bytealign_S (w[10], w[11], offset); - w[22] = hc_bytealign_S (w[ 9], w[10], offset); - w[21] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[20] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[19] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[18] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[17] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[16] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[15] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[14] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[13] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[12] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[50], w[51], offset); + w[62] = hc_bytealign_be_S (w[49], w[50], offset); + w[61] = hc_bytealign_be_S (w[48], w[49], offset); + w[60] = hc_bytealign_be_S (w[47], w[48], offset); + w[59] = hc_bytealign_be_S (w[46], w[47], offset); + w[58] = hc_bytealign_be_S (w[45], w[46], offset); + w[57] = hc_bytealign_be_S (w[44], w[45], offset); + w[56] = hc_bytealign_be_S (w[43], w[44], offset); + w[55] = hc_bytealign_be_S (w[42], w[43], offset); + w[54] = hc_bytealign_be_S (w[41], w[42], offset); + w[53] = hc_bytealign_be_S (w[40], w[41], offset); + w[52] = hc_bytealign_be_S (w[39], w[40], offset); + w[51] = hc_bytealign_be_S (w[38], w[39], offset); + w[50] = hc_bytealign_be_S (w[37], w[38], offset); + w[49] = hc_bytealign_be_S (w[36], w[37], offset); + w[48] = hc_bytealign_be_S (w[35], w[36], offset); + w[47] = hc_bytealign_be_S (w[34], w[35], offset); + w[46] = hc_bytealign_be_S (w[33], w[34], offset); + w[45] = hc_bytealign_be_S (w[32], w[33], offset); + w[44] = hc_bytealign_be_S (w[31], w[32], offset); + w[43] = hc_bytealign_be_S (w[30], w[31], offset); + w[42] = hc_bytealign_be_S (w[29], w[30], offset); + w[41] = hc_bytealign_be_S (w[28], w[29], offset); + w[40] = hc_bytealign_be_S (w[27], w[28], offset); + w[39] = hc_bytealign_be_S (w[26], w[27], offset); + w[38] = hc_bytealign_be_S (w[25], w[26], offset); + w[37] = hc_bytealign_be_S (w[24], w[25], offset); + w[36] = hc_bytealign_be_S (w[23], w[24], offset); + w[35] = hc_bytealign_be_S (w[22], w[23], offset); + w[34] = hc_bytealign_be_S (w[21], w[22], offset); + w[33] = hc_bytealign_be_S (w[20], w[21], offset); + w[32] = hc_bytealign_be_S (w[19], w[20], offset); + w[31] = hc_bytealign_be_S (w[18], w[19], offset); + w[30] = hc_bytealign_be_S (w[17], w[18], offset); + w[29] = hc_bytealign_be_S (w[16], w[17], offset); + w[28] = hc_bytealign_be_S (w[15], w[16], offset); + w[27] = hc_bytealign_be_S (w[14], w[15], offset); + w[26] = hc_bytealign_be_S (w[13], w[14], offset); + w[25] = hc_bytealign_be_S (w[12], w[13], offset); + w[24] = hc_bytealign_be_S (w[11], w[12], offset); + w[23] = hc_bytealign_be_S (w[10], w[11], offset); + w[22] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[21] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[20] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[19] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[18] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[17] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[16] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[15] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[14] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[13] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[12] = hc_bytealign_be_S ( 0, w[ 0], offset); w[11] = 0; w[10] = 0; w[ 9] = 0; @@ -52190,57 +51870,57 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 13: - w[63] = hc_bytealign_S (w[49], w[50], offset); - w[62] = hc_bytealign_S (w[48], w[49], offset); - w[61] = hc_bytealign_S (w[47], w[48], offset); - w[60] = hc_bytealign_S (w[46], w[47], offset); - w[59] = hc_bytealign_S (w[45], w[46], offset); - w[58] = hc_bytealign_S (w[44], w[45], offset); - w[57] = hc_bytealign_S (w[43], w[44], offset); - w[56] = hc_bytealign_S (w[42], w[43], offset); - w[55] = hc_bytealign_S (w[41], w[42], offset); - w[54] = hc_bytealign_S (w[40], w[41], offset); - w[53] = hc_bytealign_S (w[39], w[40], offset); - w[52] = hc_bytealign_S (w[38], w[39], offset); - w[51] = hc_bytealign_S (w[37], w[38], offset); - w[50] = hc_bytealign_S (w[36], w[37], offset); - w[49] = hc_bytealign_S (w[35], w[36], offset); - w[48] = hc_bytealign_S (w[34], w[35], offset); - w[47] = hc_bytealign_S (w[33], w[34], offset); - w[46] = hc_bytealign_S (w[32], w[33], offset); - w[45] = hc_bytealign_S (w[31], w[32], offset); - w[44] = hc_bytealign_S (w[30], w[31], offset); - w[43] = hc_bytealign_S (w[29], w[30], offset); - w[42] = hc_bytealign_S (w[28], w[29], offset); - w[41] = hc_bytealign_S (w[27], w[28], offset); - w[40] = hc_bytealign_S (w[26], w[27], offset); - w[39] = hc_bytealign_S (w[25], w[26], offset); - w[38] = hc_bytealign_S (w[24], w[25], offset); - w[37] = hc_bytealign_S (w[23], w[24], offset); - w[36] = hc_bytealign_S (w[22], w[23], offset); - w[35] = hc_bytealign_S (w[21], w[22], offset); - w[34] = hc_bytealign_S (w[20], w[21], offset); - w[33] = hc_bytealign_S (w[19], w[20], offset); - w[32] = hc_bytealign_S (w[18], w[19], offset); - w[31] = hc_bytealign_S (w[17], w[18], offset); - w[30] = hc_bytealign_S (w[16], w[17], offset); - w[29] = hc_bytealign_S (w[15], w[16], offset); - w[28] = hc_bytealign_S (w[14], w[15], offset); - w[27] = hc_bytealign_S (w[13], w[14], offset); - w[26] = hc_bytealign_S (w[12], w[13], offset); - w[25] = hc_bytealign_S (w[11], w[12], offset); - w[24] = hc_bytealign_S (w[10], w[11], offset); - w[23] = hc_bytealign_S (w[ 9], w[10], offset); - w[22] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[21] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[20] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[19] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[18] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[17] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[16] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[15] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[14] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[13] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[49], w[50], offset); + w[62] = hc_bytealign_be_S (w[48], w[49], offset); + w[61] = hc_bytealign_be_S (w[47], w[48], offset); + w[60] = hc_bytealign_be_S (w[46], w[47], offset); + w[59] = hc_bytealign_be_S (w[45], w[46], offset); + w[58] = hc_bytealign_be_S (w[44], w[45], offset); + w[57] = hc_bytealign_be_S (w[43], w[44], offset); + w[56] = hc_bytealign_be_S (w[42], w[43], offset); + w[55] = hc_bytealign_be_S (w[41], w[42], offset); + w[54] = hc_bytealign_be_S (w[40], w[41], offset); + w[53] = hc_bytealign_be_S (w[39], w[40], offset); + w[52] = hc_bytealign_be_S (w[38], w[39], offset); + w[51] = hc_bytealign_be_S (w[37], w[38], offset); + w[50] = hc_bytealign_be_S (w[36], w[37], offset); + w[49] = hc_bytealign_be_S (w[35], w[36], offset); + w[48] = hc_bytealign_be_S (w[34], w[35], offset); + w[47] = hc_bytealign_be_S (w[33], w[34], offset); + w[46] = hc_bytealign_be_S (w[32], w[33], offset); + w[45] = hc_bytealign_be_S (w[31], w[32], offset); + w[44] = hc_bytealign_be_S (w[30], w[31], offset); + w[43] = hc_bytealign_be_S (w[29], w[30], offset); + w[42] = hc_bytealign_be_S (w[28], w[29], offset); + w[41] = hc_bytealign_be_S (w[27], w[28], offset); + w[40] = hc_bytealign_be_S (w[26], w[27], offset); + w[39] = hc_bytealign_be_S (w[25], w[26], offset); + w[38] = hc_bytealign_be_S (w[24], w[25], offset); + w[37] = hc_bytealign_be_S (w[23], w[24], offset); + w[36] = hc_bytealign_be_S (w[22], w[23], offset); + w[35] = hc_bytealign_be_S (w[21], w[22], offset); + w[34] = hc_bytealign_be_S (w[20], w[21], offset); + w[33] = hc_bytealign_be_S (w[19], w[20], offset); + w[32] = hc_bytealign_be_S (w[18], w[19], offset); + w[31] = hc_bytealign_be_S (w[17], w[18], offset); + w[30] = hc_bytealign_be_S (w[16], w[17], offset); + w[29] = hc_bytealign_be_S (w[15], w[16], offset); + w[28] = hc_bytealign_be_S (w[14], w[15], offset); + w[27] = hc_bytealign_be_S (w[13], w[14], offset); + w[26] = hc_bytealign_be_S (w[12], w[13], offset); + w[25] = hc_bytealign_be_S (w[11], w[12], offset); + w[24] = hc_bytealign_be_S (w[10], w[11], offset); + w[23] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[22] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[21] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[20] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[19] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[18] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[17] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[16] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[15] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[14] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[13] = hc_bytealign_be_S ( 0, w[ 0], offset); w[12] = 0; w[11] = 0; w[10] = 0; @@ -52258,56 +51938,56 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 14: - w[63] = hc_bytealign_S (w[48], w[49], offset); - w[62] = hc_bytealign_S (w[47], w[48], offset); - w[61] = hc_bytealign_S (w[46], w[47], offset); - w[60] = hc_bytealign_S (w[45], w[46], offset); - w[59] = hc_bytealign_S (w[44], w[45], offset); - w[58] = hc_bytealign_S (w[43], w[44], offset); - w[57] = hc_bytealign_S (w[42], w[43], offset); - w[56] = hc_bytealign_S (w[41], w[42], offset); - w[55] = hc_bytealign_S (w[40], w[41], offset); - w[54] = hc_bytealign_S (w[39], w[40], offset); - w[53] = hc_bytealign_S (w[38], w[39], offset); - w[52] = hc_bytealign_S (w[37], w[38], offset); - w[51] = hc_bytealign_S (w[36], w[37], offset); - w[50] = hc_bytealign_S (w[35], w[36], offset); - w[49] = hc_bytealign_S (w[34], w[35], offset); - w[48] = hc_bytealign_S (w[33], w[34], offset); - w[47] = hc_bytealign_S (w[32], w[33], offset); - w[46] = hc_bytealign_S (w[31], w[32], offset); - w[45] = hc_bytealign_S (w[30], w[31], offset); - w[44] = hc_bytealign_S (w[29], w[30], offset); - w[43] = hc_bytealign_S (w[28], w[29], offset); - w[42] = hc_bytealign_S (w[27], w[28], offset); - w[41] = hc_bytealign_S (w[26], w[27], offset); - w[40] = hc_bytealign_S (w[25], w[26], offset); - w[39] = hc_bytealign_S (w[24], w[25], offset); - w[38] = hc_bytealign_S (w[23], w[24], offset); - w[37] = hc_bytealign_S (w[22], w[23], offset); - w[36] = hc_bytealign_S (w[21], w[22], offset); - w[35] = hc_bytealign_S (w[20], w[21], offset); - w[34] = hc_bytealign_S (w[19], w[20], offset); - w[33] = hc_bytealign_S (w[18], w[19], offset); - w[32] = hc_bytealign_S (w[17], w[18], offset); - w[31] = hc_bytealign_S (w[16], w[17], offset); - w[30] = hc_bytealign_S (w[15], w[16], offset); - w[29] = hc_bytealign_S (w[14], w[15], offset); - w[28] = hc_bytealign_S (w[13], w[14], offset); - w[27] = hc_bytealign_S (w[12], w[13], offset); - w[26] = hc_bytealign_S (w[11], w[12], offset); - w[25] = hc_bytealign_S (w[10], w[11], offset); - w[24] = hc_bytealign_S (w[ 9], w[10], offset); - w[23] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[22] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[21] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[20] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[19] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[18] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[17] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[16] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[15] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[14] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[48], w[49], offset); + w[62] = hc_bytealign_be_S (w[47], w[48], offset); + w[61] = hc_bytealign_be_S (w[46], w[47], offset); + w[60] = hc_bytealign_be_S (w[45], w[46], offset); + w[59] = hc_bytealign_be_S (w[44], w[45], offset); + w[58] = hc_bytealign_be_S (w[43], w[44], offset); + w[57] = hc_bytealign_be_S (w[42], w[43], offset); + w[56] = hc_bytealign_be_S (w[41], w[42], offset); + w[55] = hc_bytealign_be_S (w[40], w[41], offset); + w[54] = hc_bytealign_be_S (w[39], w[40], offset); + w[53] = hc_bytealign_be_S (w[38], w[39], offset); + w[52] = hc_bytealign_be_S (w[37], w[38], offset); + w[51] = hc_bytealign_be_S (w[36], w[37], offset); + w[50] = hc_bytealign_be_S (w[35], w[36], offset); + w[49] = hc_bytealign_be_S (w[34], w[35], offset); + w[48] = hc_bytealign_be_S (w[33], w[34], offset); + w[47] = hc_bytealign_be_S (w[32], w[33], offset); + w[46] = hc_bytealign_be_S (w[31], w[32], offset); + w[45] = hc_bytealign_be_S (w[30], w[31], offset); + w[44] = hc_bytealign_be_S (w[29], w[30], offset); + w[43] = hc_bytealign_be_S (w[28], w[29], offset); + w[42] = hc_bytealign_be_S (w[27], w[28], offset); + w[41] = hc_bytealign_be_S (w[26], w[27], offset); + w[40] = hc_bytealign_be_S (w[25], w[26], offset); + w[39] = hc_bytealign_be_S (w[24], w[25], offset); + w[38] = hc_bytealign_be_S (w[23], w[24], offset); + w[37] = hc_bytealign_be_S (w[22], w[23], offset); + w[36] = hc_bytealign_be_S (w[21], w[22], offset); + w[35] = hc_bytealign_be_S (w[20], w[21], offset); + w[34] = hc_bytealign_be_S (w[19], w[20], offset); + w[33] = hc_bytealign_be_S (w[18], w[19], offset); + w[32] = hc_bytealign_be_S (w[17], w[18], offset); + w[31] = hc_bytealign_be_S (w[16], w[17], offset); + w[30] = hc_bytealign_be_S (w[15], w[16], offset); + w[29] = hc_bytealign_be_S (w[14], w[15], offset); + w[28] = hc_bytealign_be_S (w[13], w[14], offset); + w[27] = hc_bytealign_be_S (w[12], w[13], offset); + w[26] = hc_bytealign_be_S (w[11], w[12], offset); + w[25] = hc_bytealign_be_S (w[10], w[11], offset); + w[24] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[23] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[22] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[21] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[20] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[19] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[18] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[17] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[16] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[15] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[14] = hc_bytealign_be_S ( 0, w[ 0], offset); w[13] = 0; w[12] = 0; w[11] = 0; @@ -52326,55 +52006,55 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 15: - w[63] = hc_bytealign_S (w[47], w[48], offset); - w[62] = hc_bytealign_S (w[46], w[47], offset); - w[61] = hc_bytealign_S (w[45], w[46], offset); - w[60] = hc_bytealign_S (w[44], w[45], offset); - w[59] = hc_bytealign_S (w[43], w[44], offset); - w[58] = hc_bytealign_S (w[42], w[43], offset); - w[57] = hc_bytealign_S (w[41], w[42], offset); - w[56] = hc_bytealign_S (w[40], w[41], offset); - w[55] = hc_bytealign_S (w[39], w[40], offset); - w[54] = hc_bytealign_S (w[38], w[39], offset); - w[53] = hc_bytealign_S (w[37], w[38], offset); - w[52] = hc_bytealign_S (w[36], w[37], offset); - w[51] = hc_bytealign_S (w[35], w[36], offset); - w[50] = hc_bytealign_S (w[34], w[35], offset); - w[49] = hc_bytealign_S (w[33], w[34], offset); - w[48] = hc_bytealign_S (w[32], w[33], offset); - w[47] = hc_bytealign_S (w[31], w[32], offset); - w[46] = hc_bytealign_S (w[30], w[31], offset); - w[45] = hc_bytealign_S (w[29], w[30], offset); - w[44] = hc_bytealign_S (w[28], w[29], offset); - w[43] = hc_bytealign_S (w[27], w[28], offset); - w[42] = hc_bytealign_S (w[26], w[27], offset); - w[41] = hc_bytealign_S (w[25], w[26], offset); - w[40] = hc_bytealign_S (w[24], w[25], offset); - w[39] = hc_bytealign_S (w[23], w[24], offset); - w[38] = hc_bytealign_S (w[22], w[23], offset); - w[37] = hc_bytealign_S (w[21], w[22], offset); - w[36] = hc_bytealign_S (w[20], w[21], offset); - w[35] = hc_bytealign_S (w[19], w[20], offset); - w[34] = hc_bytealign_S (w[18], w[19], offset); - w[33] = hc_bytealign_S (w[17], w[18], offset); - w[32] = hc_bytealign_S (w[16], w[17], offset); - w[31] = hc_bytealign_S (w[15], w[16], offset); - w[30] = hc_bytealign_S (w[14], w[15], offset); - w[29] = hc_bytealign_S (w[13], w[14], offset); - w[28] = hc_bytealign_S (w[12], w[13], offset); - w[27] = hc_bytealign_S (w[11], w[12], offset); - w[26] = hc_bytealign_S (w[10], w[11], offset); - w[25] = hc_bytealign_S (w[ 9], w[10], offset); - w[24] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[23] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[22] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[21] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[20] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[19] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[18] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[17] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[16] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[15] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[47], w[48], offset); + w[62] = hc_bytealign_be_S (w[46], w[47], offset); + w[61] = hc_bytealign_be_S (w[45], w[46], offset); + w[60] = hc_bytealign_be_S (w[44], w[45], offset); + w[59] = hc_bytealign_be_S (w[43], w[44], offset); + w[58] = hc_bytealign_be_S (w[42], w[43], offset); + w[57] = hc_bytealign_be_S (w[41], w[42], offset); + w[56] = hc_bytealign_be_S (w[40], w[41], offset); + w[55] = hc_bytealign_be_S (w[39], w[40], offset); + w[54] = hc_bytealign_be_S (w[38], w[39], offset); + w[53] = hc_bytealign_be_S (w[37], w[38], offset); + w[52] = hc_bytealign_be_S (w[36], w[37], offset); + w[51] = hc_bytealign_be_S (w[35], w[36], offset); + w[50] = hc_bytealign_be_S (w[34], w[35], offset); + w[49] = hc_bytealign_be_S (w[33], w[34], offset); + w[48] = hc_bytealign_be_S (w[32], w[33], offset); + w[47] = hc_bytealign_be_S (w[31], w[32], offset); + w[46] = hc_bytealign_be_S (w[30], w[31], offset); + w[45] = hc_bytealign_be_S (w[29], w[30], offset); + w[44] = hc_bytealign_be_S (w[28], w[29], offset); + w[43] = hc_bytealign_be_S (w[27], w[28], offset); + w[42] = hc_bytealign_be_S (w[26], w[27], offset); + w[41] = hc_bytealign_be_S (w[25], w[26], offset); + w[40] = hc_bytealign_be_S (w[24], w[25], offset); + w[39] = hc_bytealign_be_S (w[23], w[24], offset); + w[38] = hc_bytealign_be_S (w[22], w[23], offset); + w[37] = hc_bytealign_be_S (w[21], w[22], offset); + w[36] = hc_bytealign_be_S (w[20], w[21], offset); + w[35] = hc_bytealign_be_S (w[19], w[20], offset); + w[34] = hc_bytealign_be_S (w[18], w[19], offset); + w[33] = hc_bytealign_be_S (w[17], w[18], offset); + w[32] = hc_bytealign_be_S (w[16], w[17], offset); + w[31] = hc_bytealign_be_S (w[15], w[16], offset); + w[30] = hc_bytealign_be_S (w[14], w[15], offset); + w[29] = hc_bytealign_be_S (w[13], w[14], offset); + w[28] = hc_bytealign_be_S (w[12], w[13], offset); + w[27] = hc_bytealign_be_S (w[11], w[12], offset); + w[26] = hc_bytealign_be_S (w[10], w[11], offset); + w[25] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[24] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[23] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[22] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[21] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[20] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[19] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[18] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[17] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[16] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[15] = hc_bytealign_be_S ( 0, w[ 0], offset); w[14] = 0; w[13] = 0; w[12] = 0; @@ -52394,54 +52074,54 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 16: - w[63] = hc_bytealign_S (w[46], w[47], offset); - w[62] = hc_bytealign_S (w[45], w[46], offset); - w[61] = hc_bytealign_S (w[44], w[45], offset); - w[60] = hc_bytealign_S (w[43], w[44], offset); - w[59] = hc_bytealign_S (w[42], w[43], offset); - w[58] = hc_bytealign_S (w[41], w[42], offset); - w[57] = hc_bytealign_S (w[40], w[41], offset); - w[56] = hc_bytealign_S (w[39], w[40], offset); - w[55] = hc_bytealign_S (w[38], w[39], offset); - w[54] = hc_bytealign_S (w[37], w[38], offset); - w[53] = hc_bytealign_S (w[36], w[37], offset); - w[52] = hc_bytealign_S (w[35], w[36], offset); - w[51] = hc_bytealign_S (w[34], w[35], offset); - w[50] = hc_bytealign_S (w[33], w[34], offset); - w[49] = hc_bytealign_S (w[32], w[33], offset); - w[48] = hc_bytealign_S (w[31], w[32], offset); - w[47] = hc_bytealign_S (w[30], w[31], offset); - w[46] = hc_bytealign_S (w[29], w[30], offset); - w[45] = hc_bytealign_S (w[28], w[29], offset); - w[44] = hc_bytealign_S (w[27], w[28], offset); - w[43] = hc_bytealign_S (w[26], w[27], offset); - w[42] = hc_bytealign_S (w[25], w[26], offset); - w[41] = hc_bytealign_S (w[24], w[25], offset); - w[40] = hc_bytealign_S (w[23], w[24], offset); - w[39] = hc_bytealign_S (w[22], w[23], offset); - w[38] = hc_bytealign_S (w[21], w[22], offset); - w[37] = hc_bytealign_S (w[20], w[21], offset); - w[36] = hc_bytealign_S (w[19], w[20], offset); - w[35] = hc_bytealign_S (w[18], w[19], offset); - w[34] = hc_bytealign_S (w[17], w[18], offset); - w[33] = hc_bytealign_S (w[16], w[17], offset); - w[32] = hc_bytealign_S (w[15], w[16], offset); - w[31] = hc_bytealign_S (w[14], w[15], offset); - w[30] = hc_bytealign_S (w[13], w[14], offset); - w[29] = hc_bytealign_S (w[12], w[13], offset); - w[28] = hc_bytealign_S (w[11], w[12], offset); - w[27] = hc_bytealign_S (w[10], w[11], offset); - w[26] = hc_bytealign_S (w[ 9], w[10], offset); - w[25] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[24] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[23] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[22] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[21] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[20] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[19] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[18] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[17] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[16] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[46], w[47], offset); + w[62] = hc_bytealign_be_S (w[45], w[46], offset); + w[61] = hc_bytealign_be_S (w[44], w[45], offset); + w[60] = hc_bytealign_be_S (w[43], w[44], offset); + w[59] = hc_bytealign_be_S (w[42], w[43], offset); + w[58] = hc_bytealign_be_S (w[41], w[42], offset); + w[57] = hc_bytealign_be_S (w[40], w[41], offset); + w[56] = hc_bytealign_be_S (w[39], w[40], offset); + w[55] = hc_bytealign_be_S (w[38], w[39], offset); + w[54] = hc_bytealign_be_S (w[37], w[38], offset); + w[53] = hc_bytealign_be_S (w[36], w[37], offset); + w[52] = hc_bytealign_be_S (w[35], w[36], offset); + w[51] = hc_bytealign_be_S (w[34], w[35], offset); + w[50] = hc_bytealign_be_S (w[33], w[34], offset); + w[49] = hc_bytealign_be_S (w[32], w[33], offset); + w[48] = hc_bytealign_be_S (w[31], w[32], offset); + w[47] = hc_bytealign_be_S (w[30], w[31], offset); + w[46] = hc_bytealign_be_S (w[29], w[30], offset); + w[45] = hc_bytealign_be_S (w[28], w[29], offset); + w[44] = hc_bytealign_be_S (w[27], w[28], offset); + w[43] = hc_bytealign_be_S (w[26], w[27], offset); + w[42] = hc_bytealign_be_S (w[25], w[26], offset); + w[41] = hc_bytealign_be_S (w[24], w[25], offset); + w[40] = hc_bytealign_be_S (w[23], w[24], offset); + w[39] = hc_bytealign_be_S (w[22], w[23], offset); + w[38] = hc_bytealign_be_S (w[21], w[22], offset); + w[37] = hc_bytealign_be_S (w[20], w[21], offset); + w[36] = hc_bytealign_be_S (w[19], w[20], offset); + w[35] = hc_bytealign_be_S (w[18], w[19], offset); + w[34] = hc_bytealign_be_S (w[17], w[18], offset); + w[33] = hc_bytealign_be_S (w[16], w[17], offset); + w[32] = hc_bytealign_be_S (w[15], w[16], offset); + w[31] = hc_bytealign_be_S (w[14], w[15], offset); + w[30] = hc_bytealign_be_S (w[13], w[14], offset); + w[29] = hc_bytealign_be_S (w[12], w[13], offset); + w[28] = hc_bytealign_be_S (w[11], w[12], offset); + w[27] = hc_bytealign_be_S (w[10], w[11], offset); + w[26] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[25] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[24] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[23] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[22] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[21] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[20] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[19] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[18] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[17] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[16] = hc_bytealign_be_S ( 0, w[ 0], offset); w[15] = 0; w[14] = 0; w[13] = 0; @@ -52462,53 +52142,53 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 17: - w[63] = hc_bytealign_S (w[45], w[46], offset); - w[62] = hc_bytealign_S (w[44], w[45], offset); - w[61] = hc_bytealign_S (w[43], w[44], offset); - w[60] = hc_bytealign_S (w[42], w[43], offset); - w[59] = hc_bytealign_S (w[41], w[42], offset); - w[58] = hc_bytealign_S (w[40], w[41], offset); - w[57] = hc_bytealign_S (w[39], w[40], offset); - w[56] = hc_bytealign_S (w[38], w[39], offset); - w[55] = hc_bytealign_S (w[37], w[38], offset); - w[54] = hc_bytealign_S (w[36], w[37], offset); - w[53] = hc_bytealign_S (w[35], w[36], offset); - w[52] = hc_bytealign_S (w[34], w[35], offset); - w[51] = hc_bytealign_S (w[33], w[34], offset); - w[50] = hc_bytealign_S (w[32], w[33], offset); - w[49] = hc_bytealign_S (w[31], w[32], offset); - w[48] = hc_bytealign_S (w[30], w[31], offset); - w[47] = hc_bytealign_S (w[29], w[30], offset); - w[46] = hc_bytealign_S (w[28], w[29], offset); - w[45] = hc_bytealign_S (w[27], w[28], offset); - w[44] = hc_bytealign_S (w[26], w[27], offset); - w[43] = hc_bytealign_S (w[25], w[26], offset); - w[42] = hc_bytealign_S (w[24], w[25], offset); - w[41] = hc_bytealign_S (w[23], w[24], offset); - w[40] = hc_bytealign_S (w[22], w[23], offset); - w[39] = hc_bytealign_S (w[21], w[22], offset); - w[38] = hc_bytealign_S (w[20], w[21], offset); - w[37] = hc_bytealign_S (w[19], w[20], offset); - w[36] = hc_bytealign_S (w[18], w[19], offset); - w[35] = hc_bytealign_S (w[17], w[18], offset); - w[34] = hc_bytealign_S (w[16], w[17], offset); - w[33] = hc_bytealign_S (w[15], w[16], offset); - w[32] = hc_bytealign_S (w[14], w[15], offset); - w[31] = hc_bytealign_S (w[13], w[14], offset); - w[30] = hc_bytealign_S (w[12], w[13], offset); - w[29] = hc_bytealign_S (w[11], w[12], offset); - w[28] = hc_bytealign_S (w[10], w[11], offset); - w[27] = hc_bytealign_S (w[ 9], w[10], offset); - w[26] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[25] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[24] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[23] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[22] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[21] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[20] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[19] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[18] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[17] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[45], w[46], offset); + w[62] = hc_bytealign_be_S (w[44], w[45], offset); + w[61] = hc_bytealign_be_S (w[43], w[44], offset); + w[60] = hc_bytealign_be_S (w[42], w[43], offset); + w[59] = hc_bytealign_be_S (w[41], w[42], offset); + w[58] = hc_bytealign_be_S (w[40], w[41], offset); + w[57] = hc_bytealign_be_S (w[39], w[40], offset); + w[56] = hc_bytealign_be_S (w[38], w[39], offset); + w[55] = hc_bytealign_be_S (w[37], w[38], offset); + w[54] = hc_bytealign_be_S (w[36], w[37], offset); + w[53] = hc_bytealign_be_S (w[35], w[36], offset); + w[52] = hc_bytealign_be_S (w[34], w[35], offset); + w[51] = hc_bytealign_be_S (w[33], w[34], offset); + w[50] = hc_bytealign_be_S (w[32], w[33], offset); + w[49] = hc_bytealign_be_S (w[31], w[32], offset); + w[48] = hc_bytealign_be_S (w[30], w[31], offset); + w[47] = hc_bytealign_be_S (w[29], w[30], offset); + w[46] = hc_bytealign_be_S (w[28], w[29], offset); + w[45] = hc_bytealign_be_S (w[27], w[28], offset); + w[44] = hc_bytealign_be_S (w[26], w[27], offset); + w[43] = hc_bytealign_be_S (w[25], w[26], offset); + w[42] = hc_bytealign_be_S (w[24], w[25], offset); + w[41] = hc_bytealign_be_S (w[23], w[24], offset); + w[40] = hc_bytealign_be_S (w[22], w[23], offset); + w[39] = hc_bytealign_be_S (w[21], w[22], offset); + w[38] = hc_bytealign_be_S (w[20], w[21], offset); + w[37] = hc_bytealign_be_S (w[19], w[20], offset); + w[36] = hc_bytealign_be_S (w[18], w[19], offset); + w[35] = hc_bytealign_be_S (w[17], w[18], offset); + w[34] = hc_bytealign_be_S (w[16], w[17], offset); + w[33] = hc_bytealign_be_S (w[15], w[16], offset); + w[32] = hc_bytealign_be_S (w[14], w[15], offset); + w[31] = hc_bytealign_be_S (w[13], w[14], offset); + w[30] = hc_bytealign_be_S (w[12], w[13], offset); + w[29] = hc_bytealign_be_S (w[11], w[12], offset); + w[28] = hc_bytealign_be_S (w[10], w[11], offset); + w[27] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[26] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[25] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[24] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[23] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[22] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[21] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[20] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[19] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[18] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[17] = hc_bytealign_be_S ( 0, w[ 0], offset); w[16] = 0; w[15] = 0; w[14] = 0; @@ -52530,52 +52210,52 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 18: - w[63] = hc_bytealign_S (w[44], w[45], offset); - w[62] = hc_bytealign_S (w[43], w[44], offset); - w[61] = hc_bytealign_S (w[42], w[43], offset); - w[60] = hc_bytealign_S (w[41], w[42], offset); - w[59] = hc_bytealign_S (w[40], w[41], offset); - w[58] = hc_bytealign_S (w[39], w[40], offset); - w[57] = hc_bytealign_S (w[38], w[39], offset); - w[56] = hc_bytealign_S (w[37], w[38], offset); - w[55] = hc_bytealign_S (w[36], w[37], offset); - w[54] = hc_bytealign_S (w[35], w[36], offset); - w[53] = hc_bytealign_S (w[34], w[35], offset); - w[52] = hc_bytealign_S (w[33], w[34], offset); - w[51] = hc_bytealign_S (w[32], w[33], offset); - w[50] = hc_bytealign_S (w[31], w[32], offset); - w[49] = hc_bytealign_S (w[30], w[31], offset); - w[48] = hc_bytealign_S (w[29], w[30], offset); - w[47] = hc_bytealign_S (w[28], w[29], offset); - w[46] = hc_bytealign_S (w[27], w[28], offset); - w[45] = hc_bytealign_S (w[26], w[27], offset); - w[44] = hc_bytealign_S (w[25], w[26], offset); - w[43] = hc_bytealign_S (w[24], w[25], offset); - w[42] = hc_bytealign_S (w[23], w[24], offset); - w[41] = hc_bytealign_S (w[22], w[23], offset); - w[40] = hc_bytealign_S (w[21], w[22], offset); - w[39] = hc_bytealign_S (w[20], w[21], offset); - w[38] = hc_bytealign_S (w[19], w[20], offset); - w[37] = hc_bytealign_S (w[18], w[19], offset); - w[36] = hc_bytealign_S (w[17], w[18], offset); - w[35] = hc_bytealign_S (w[16], w[17], offset); - w[34] = hc_bytealign_S (w[15], w[16], offset); - w[33] = hc_bytealign_S (w[14], w[15], offset); - w[32] = hc_bytealign_S (w[13], w[14], offset); - w[31] = hc_bytealign_S (w[12], w[13], offset); - w[30] = hc_bytealign_S (w[11], w[12], offset); - w[29] = hc_bytealign_S (w[10], w[11], offset); - w[28] = hc_bytealign_S (w[ 9], w[10], offset); - w[27] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[26] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[25] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[24] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[23] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[22] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[21] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[20] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[19] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[18] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[44], w[45], offset); + w[62] = hc_bytealign_be_S (w[43], w[44], offset); + w[61] = hc_bytealign_be_S (w[42], w[43], offset); + w[60] = hc_bytealign_be_S (w[41], w[42], offset); + w[59] = hc_bytealign_be_S (w[40], w[41], offset); + w[58] = hc_bytealign_be_S (w[39], w[40], offset); + w[57] = hc_bytealign_be_S (w[38], w[39], offset); + w[56] = hc_bytealign_be_S (w[37], w[38], offset); + w[55] = hc_bytealign_be_S (w[36], w[37], offset); + w[54] = hc_bytealign_be_S (w[35], w[36], offset); + w[53] = hc_bytealign_be_S (w[34], w[35], offset); + w[52] = hc_bytealign_be_S (w[33], w[34], offset); + w[51] = hc_bytealign_be_S (w[32], w[33], offset); + w[50] = hc_bytealign_be_S (w[31], w[32], offset); + w[49] = hc_bytealign_be_S (w[30], w[31], offset); + w[48] = hc_bytealign_be_S (w[29], w[30], offset); + w[47] = hc_bytealign_be_S (w[28], w[29], offset); + w[46] = hc_bytealign_be_S (w[27], w[28], offset); + w[45] = hc_bytealign_be_S (w[26], w[27], offset); + w[44] = hc_bytealign_be_S (w[25], w[26], offset); + w[43] = hc_bytealign_be_S (w[24], w[25], offset); + w[42] = hc_bytealign_be_S (w[23], w[24], offset); + w[41] = hc_bytealign_be_S (w[22], w[23], offset); + w[40] = hc_bytealign_be_S (w[21], w[22], offset); + w[39] = hc_bytealign_be_S (w[20], w[21], offset); + w[38] = hc_bytealign_be_S (w[19], w[20], offset); + w[37] = hc_bytealign_be_S (w[18], w[19], offset); + w[36] = hc_bytealign_be_S (w[17], w[18], offset); + w[35] = hc_bytealign_be_S (w[16], w[17], offset); + w[34] = hc_bytealign_be_S (w[15], w[16], offset); + w[33] = hc_bytealign_be_S (w[14], w[15], offset); + w[32] = hc_bytealign_be_S (w[13], w[14], offset); + w[31] = hc_bytealign_be_S (w[12], w[13], offset); + w[30] = hc_bytealign_be_S (w[11], w[12], offset); + w[29] = hc_bytealign_be_S (w[10], w[11], offset); + w[28] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[27] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[26] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[25] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[24] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[23] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[22] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[21] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[20] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[19] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[18] = hc_bytealign_be_S ( 0, w[ 0], offset); w[17] = 0; w[16] = 0; w[15] = 0; @@ -52598,51 +52278,51 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 19: - w[63] = hc_bytealign_S (w[43], w[44], offset); - w[62] = hc_bytealign_S (w[42], w[43], offset); - w[61] = hc_bytealign_S (w[41], w[42], offset); - w[60] = hc_bytealign_S (w[40], w[41], offset); - w[59] = hc_bytealign_S (w[39], w[40], offset); - w[58] = hc_bytealign_S (w[38], w[39], offset); - w[57] = hc_bytealign_S (w[37], w[38], offset); - w[56] = hc_bytealign_S (w[36], w[37], offset); - w[55] = hc_bytealign_S (w[35], w[36], offset); - w[54] = hc_bytealign_S (w[34], w[35], offset); - w[53] = hc_bytealign_S (w[33], w[34], offset); - w[52] = hc_bytealign_S (w[32], w[33], offset); - w[51] = hc_bytealign_S (w[31], w[32], offset); - w[50] = hc_bytealign_S (w[30], w[31], offset); - w[49] = hc_bytealign_S (w[29], w[30], offset); - w[48] = hc_bytealign_S (w[28], w[29], offset); - w[47] = hc_bytealign_S (w[27], w[28], offset); - w[46] = hc_bytealign_S (w[26], w[27], offset); - w[45] = hc_bytealign_S (w[25], w[26], offset); - w[44] = hc_bytealign_S (w[24], w[25], offset); - w[43] = hc_bytealign_S (w[23], w[24], offset); - w[42] = hc_bytealign_S (w[22], w[23], offset); - w[41] = hc_bytealign_S (w[21], w[22], offset); - w[40] = hc_bytealign_S (w[20], w[21], offset); - w[39] = hc_bytealign_S (w[19], w[20], offset); - w[38] = hc_bytealign_S (w[18], w[19], offset); - w[37] = hc_bytealign_S (w[17], w[18], offset); - w[36] = hc_bytealign_S (w[16], w[17], offset); - w[35] = hc_bytealign_S (w[15], w[16], offset); - w[34] = hc_bytealign_S (w[14], w[15], offset); - w[33] = hc_bytealign_S (w[13], w[14], offset); - w[32] = hc_bytealign_S (w[12], w[13], offset); - w[31] = hc_bytealign_S (w[11], w[12], offset); - w[30] = hc_bytealign_S (w[10], w[11], offset); - w[29] = hc_bytealign_S (w[ 9], w[10], offset); - w[28] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[27] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[26] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[25] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[24] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[23] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[22] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[21] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[20] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[19] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[43], w[44], offset); + w[62] = hc_bytealign_be_S (w[42], w[43], offset); + w[61] = hc_bytealign_be_S (w[41], w[42], offset); + w[60] = hc_bytealign_be_S (w[40], w[41], offset); + w[59] = hc_bytealign_be_S (w[39], w[40], offset); + w[58] = hc_bytealign_be_S (w[38], w[39], offset); + w[57] = hc_bytealign_be_S (w[37], w[38], offset); + w[56] = hc_bytealign_be_S (w[36], w[37], offset); + w[55] = hc_bytealign_be_S (w[35], w[36], offset); + w[54] = hc_bytealign_be_S (w[34], w[35], offset); + w[53] = hc_bytealign_be_S (w[33], w[34], offset); + w[52] = hc_bytealign_be_S (w[32], w[33], offset); + w[51] = hc_bytealign_be_S (w[31], w[32], offset); + w[50] = hc_bytealign_be_S (w[30], w[31], offset); + w[49] = hc_bytealign_be_S (w[29], w[30], offset); + w[48] = hc_bytealign_be_S (w[28], w[29], offset); + w[47] = hc_bytealign_be_S (w[27], w[28], offset); + w[46] = hc_bytealign_be_S (w[26], w[27], offset); + w[45] = hc_bytealign_be_S (w[25], w[26], offset); + w[44] = hc_bytealign_be_S (w[24], w[25], offset); + w[43] = hc_bytealign_be_S (w[23], w[24], offset); + w[42] = hc_bytealign_be_S (w[22], w[23], offset); + w[41] = hc_bytealign_be_S (w[21], w[22], offset); + w[40] = hc_bytealign_be_S (w[20], w[21], offset); + w[39] = hc_bytealign_be_S (w[19], w[20], offset); + w[38] = hc_bytealign_be_S (w[18], w[19], offset); + w[37] = hc_bytealign_be_S (w[17], w[18], offset); + w[36] = hc_bytealign_be_S (w[16], w[17], offset); + w[35] = hc_bytealign_be_S (w[15], w[16], offset); + w[34] = hc_bytealign_be_S (w[14], w[15], offset); + w[33] = hc_bytealign_be_S (w[13], w[14], offset); + w[32] = hc_bytealign_be_S (w[12], w[13], offset); + w[31] = hc_bytealign_be_S (w[11], w[12], offset); + w[30] = hc_bytealign_be_S (w[10], w[11], offset); + w[29] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[28] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[27] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[26] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[25] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[24] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[23] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[22] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[21] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[20] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[19] = hc_bytealign_be_S ( 0, w[ 0], offset); w[18] = 0; w[17] = 0; w[16] = 0; @@ -52666,50 +52346,50 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 20: - w[63] = hc_bytealign_S (w[42], w[43], offset); - w[62] = hc_bytealign_S (w[41], w[42], offset); - w[61] = hc_bytealign_S (w[40], w[41], offset); - w[60] = hc_bytealign_S (w[39], w[40], offset); - w[59] = hc_bytealign_S (w[38], w[39], offset); - w[58] = hc_bytealign_S (w[37], w[38], offset); - w[57] = hc_bytealign_S (w[36], w[37], offset); - w[56] = hc_bytealign_S (w[35], w[36], offset); - w[55] = hc_bytealign_S (w[34], w[35], offset); - w[54] = hc_bytealign_S (w[33], w[34], offset); - w[53] = hc_bytealign_S (w[32], w[33], offset); - w[52] = hc_bytealign_S (w[31], w[32], offset); - w[51] = hc_bytealign_S (w[30], w[31], offset); - w[50] = hc_bytealign_S (w[29], w[30], offset); - w[49] = hc_bytealign_S (w[28], w[29], offset); - w[48] = hc_bytealign_S (w[27], w[28], offset); - w[47] = hc_bytealign_S (w[26], w[27], offset); - w[46] = hc_bytealign_S (w[25], w[26], offset); - w[45] = hc_bytealign_S (w[24], w[25], offset); - w[44] = hc_bytealign_S (w[23], w[24], offset); - w[43] = hc_bytealign_S (w[22], w[23], offset); - w[42] = hc_bytealign_S (w[21], w[22], offset); - w[41] = hc_bytealign_S (w[20], w[21], offset); - w[40] = hc_bytealign_S (w[19], w[20], offset); - w[39] = hc_bytealign_S (w[18], w[19], offset); - w[38] = hc_bytealign_S (w[17], w[18], offset); - w[37] = hc_bytealign_S (w[16], w[17], offset); - w[36] = hc_bytealign_S (w[15], w[16], offset); - w[35] = hc_bytealign_S (w[14], w[15], offset); - w[34] = hc_bytealign_S (w[13], w[14], offset); - w[33] = hc_bytealign_S (w[12], w[13], offset); - w[32] = hc_bytealign_S (w[11], w[12], offset); - w[31] = hc_bytealign_S (w[10], w[11], offset); - w[30] = hc_bytealign_S (w[ 9], w[10], offset); - w[29] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[28] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[27] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[26] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[25] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[24] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[23] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[22] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[21] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[20] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[42], w[43], offset); + w[62] = hc_bytealign_be_S (w[41], w[42], offset); + w[61] = hc_bytealign_be_S (w[40], w[41], offset); + w[60] = hc_bytealign_be_S (w[39], w[40], offset); + w[59] = hc_bytealign_be_S (w[38], w[39], offset); + w[58] = hc_bytealign_be_S (w[37], w[38], offset); + w[57] = hc_bytealign_be_S (w[36], w[37], offset); + w[56] = hc_bytealign_be_S (w[35], w[36], offset); + w[55] = hc_bytealign_be_S (w[34], w[35], offset); + w[54] = hc_bytealign_be_S (w[33], w[34], offset); + w[53] = hc_bytealign_be_S (w[32], w[33], offset); + w[52] = hc_bytealign_be_S (w[31], w[32], offset); + w[51] = hc_bytealign_be_S (w[30], w[31], offset); + w[50] = hc_bytealign_be_S (w[29], w[30], offset); + w[49] = hc_bytealign_be_S (w[28], w[29], offset); + w[48] = hc_bytealign_be_S (w[27], w[28], offset); + w[47] = hc_bytealign_be_S (w[26], w[27], offset); + w[46] = hc_bytealign_be_S (w[25], w[26], offset); + w[45] = hc_bytealign_be_S (w[24], w[25], offset); + w[44] = hc_bytealign_be_S (w[23], w[24], offset); + w[43] = hc_bytealign_be_S (w[22], w[23], offset); + w[42] = hc_bytealign_be_S (w[21], w[22], offset); + w[41] = hc_bytealign_be_S (w[20], w[21], offset); + w[40] = hc_bytealign_be_S (w[19], w[20], offset); + w[39] = hc_bytealign_be_S (w[18], w[19], offset); + w[38] = hc_bytealign_be_S (w[17], w[18], offset); + w[37] = hc_bytealign_be_S (w[16], w[17], offset); + w[36] = hc_bytealign_be_S (w[15], w[16], offset); + w[35] = hc_bytealign_be_S (w[14], w[15], offset); + w[34] = hc_bytealign_be_S (w[13], w[14], offset); + w[33] = hc_bytealign_be_S (w[12], w[13], offset); + w[32] = hc_bytealign_be_S (w[11], w[12], offset); + w[31] = hc_bytealign_be_S (w[10], w[11], offset); + w[30] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[29] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[28] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[27] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[26] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[25] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[24] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[23] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[22] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[21] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[20] = hc_bytealign_be_S ( 0, w[ 0], offset); w[19] = 0; w[18] = 0; w[17] = 0; @@ -52734,49 +52414,49 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 21: - w[63] = hc_bytealign_S (w[41], w[42], offset); - w[62] = hc_bytealign_S (w[40], w[41], offset); - w[61] = hc_bytealign_S (w[39], w[40], offset); - w[60] = hc_bytealign_S (w[38], w[39], offset); - w[59] = hc_bytealign_S (w[37], w[38], offset); - w[58] = hc_bytealign_S (w[36], w[37], offset); - w[57] = hc_bytealign_S (w[35], w[36], offset); - w[56] = hc_bytealign_S (w[34], w[35], offset); - w[55] = hc_bytealign_S (w[33], w[34], offset); - w[54] = hc_bytealign_S (w[32], w[33], offset); - w[53] = hc_bytealign_S (w[31], w[32], offset); - w[52] = hc_bytealign_S (w[30], w[31], offset); - w[51] = hc_bytealign_S (w[29], w[30], offset); - w[50] = hc_bytealign_S (w[28], w[29], offset); - w[49] = hc_bytealign_S (w[27], w[28], offset); - w[48] = hc_bytealign_S (w[26], w[27], offset); - w[47] = hc_bytealign_S (w[25], w[26], offset); - w[46] = hc_bytealign_S (w[24], w[25], offset); - w[45] = hc_bytealign_S (w[23], w[24], offset); - w[44] = hc_bytealign_S (w[22], w[23], offset); - w[43] = hc_bytealign_S (w[21], w[22], offset); - w[42] = hc_bytealign_S (w[20], w[21], offset); - w[41] = hc_bytealign_S (w[19], w[20], offset); - w[40] = hc_bytealign_S (w[18], w[19], offset); - w[39] = hc_bytealign_S (w[17], w[18], offset); - w[38] = hc_bytealign_S (w[16], w[17], offset); - w[37] = hc_bytealign_S (w[15], w[16], offset); - w[36] = hc_bytealign_S (w[14], w[15], offset); - w[35] = hc_bytealign_S (w[13], w[14], offset); - w[34] = hc_bytealign_S (w[12], w[13], offset); - w[33] = hc_bytealign_S (w[11], w[12], offset); - w[32] = hc_bytealign_S (w[10], w[11], offset); - w[31] = hc_bytealign_S (w[ 9], w[10], offset); - w[30] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[29] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[28] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[27] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[26] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[25] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[24] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[23] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[22] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[21] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[41], w[42], offset); + w[62] = hc_bytealign_be_S (w[40], w[41], offset); + w[61] = hc_bytealign_be_S (w[39], w[40], offset); + w[60] = hc_bytealign_be_S (w[38], w[39], offset); + w[59] = hc_bytealign_be_S (w[37], w[38], offset); + w[58] = hc_bytealign_be_S (w[36], w[37], offset); + w[57] = hc_bytealign_be_S (w[35], w[36], offset); + w[56] = hc_bytealign_be_S (w[34], w[35], offset); + w[55] = hc_bytealign_be_S (w[33], w[34], offset); + w[54] = hc_bytealign_be_S (w[32], w[33], offset); + w[53] = hc_bytealign_be_S (w[31], w[32], offset); + w[52] = hc_bytealign_be_S (w[30], w[31], offset); + w[51] = hc_bytealign_be_S (w[29], w[30], offset); + w[50] = hc_bytealign_be_S (w[28], w[29], offset); + w[49] = hc_bytealign_be_S (w[27], w[28], offset); + w[48] = hc_bytealign_be_S (w[26], w[27], offset); + w[47] = hc_bytealign_be_S (w[25], w[26], offset); + w[46] = hc_bytealign_be_S (w[24], w[25], offset); + w[45] = hc_bytealign_be_S (w[23], w[24], offset); + w[44] = hc_bytealign_be_S (w[22], w[23], offset); + w[43] = hc_bytealign_be_S (w[21], w[22], offset); + w[42] = hc_bytealign_be_S (w[20], w[21], offset); + w[41] = hc_bytealign_be_S (w[19], w[20], offset); + w[40] = hc_bytealign_be_S (w[18], w[19], offset); + w[39] = hc_bytealign_be_S (w[17], w[18], offset); + w[38] = hc_bytealign_be_S (w[16], w[17], offset); + w[37] = hc_bytealign_be_S (w[15], w[16], offset); + w[36] = hc_bytealign_be_S (w[14], w[15], offset); + w[35] = hc_bytealign_be_S (w[13], w[14], offset); + w[34] = hc_bytealign_be_S (w[12], w[13], offset); + w[33] = hc_bytealign_be_S (w[11], w[12], offset); + w[32] = hc_bytealign_be_S (w[10], w[11], offset); + w[31] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[30] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[29] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[28] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[27] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[26] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[25] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[24] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[23] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[22] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[21] = hc_bytealign_be_S ( 0, w[ 0], offset); w[20] = 0; w[19] = 0; w[18] = 0; @@ -52802,48 +52482,48 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 22: - w[63] = hc_bytealign_S (w[40], w[41], offset); - w[62] = hc_bytealign_S (w[39], w[40], offset); - w[61] = hc_bytealign_S (w[38], w[39], offset); - w[60] = hc_bytealign_S (w[37], w[38], offset); - w[59] = hc_bytealign_S (w[36], w[37], offset); - w[58] = hc_bytealign_S (w[35], w[36], offset); - w[57] = hc_bytealign_S (w[34], w[35], offset); - w[56] = hc_bytealign_S (w[33], w[34], offset); - w[55] = hc_bytealign_S (w[32], w[33], offset); - w[54] = hc_bytealign_S (w[31], w[32], offset); - w[53] = hc_bytealign_S (w[30], w[31], offset); - w[52] = hc_bytealign_S (w[29], w[30], offset); - w[51] = hc_bytealign_S (w[28], w[29], offset); - w[50] = hc_bytealign_S (w[27], w[28], offset); - w[49] = hc_bytealign_S (w[26], w[27], offset); - w[48] = hc_bytealign_S (w[25], w[26], offset); - w[47] = hc_bytealign_S (w[24], w[25], offset); - w[46] = hc_bytealign_S (w[23], w[24], offset); - w[45] = hc_bytealign_S (w[22], w[23], offset); - w[44] = hc_bytealign_S (w[21], w[22], offset); - w[43] = hc_bytealign_S (w[20], w[21], offset); - w[42] = hc_bytealign_S (w[19], w[20], offset); - w[41] = hc_bytealign_S (w[18], w[19], offset); - w[40] = hc_bytealign_S (w[17], w[18], offset); - w[39] = hc_bytealign_S (w[16], w[17], offset); - w[38] = hc_bytealign_S (w[15], w[16], offset); - w[37] = hc_bytealign_S (w[14], w[15], offset); - w[36] = hc_bytealign_S (w[13], w[14], offset); - w[35] = hc_bytealign_S (w[12], w[13], offset); - w[34] = hc_bytealign_S (w[11], w[12], offset); - w[33] = hc_bytealign_S (w[10], w[11], offset); - w[32] = hc_bytealign_S (w[ 9], w[10], offset); - w[31] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[30] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[29] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[28] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[27] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[26] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[25] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[24] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[23] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[22] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[40], w[41], offset); + w[62] = hc_bytealign_be_S (w[39], w[40], offset); + w[61] = hc_bytealign_be_S (w[38], w[39], offset); + w[60] = hc_bytealign_be_S (w[37], w[38], offset); + w[59] = hc_bytealign_be_S (w[36], w[37], offset); + w[58] = hc_bytealign_be_S (w[35], w[36], offset); + w[57] = hc_bytealign_be_S (w[34], w[35], offset); + w[56] = hc_bytealign_be_S (w[33], w[34], offset); + w[55] = hc_bytealign_be_S (w[32], w[33], offset); + w[54] = hc_bytealign_be_S (w[31], w[32], offset); + w[53] = hc_bytealign_be_S (w[30], w[31], offset); + w[52] = hc_bytealign_be_S (w[29], w[30], offset); + w[51] = hc_bytealign_be_S (w[28], w[29], offset); + w[50] = hc_bytealign_be_S (w[27], w[28], offset); + w[49] = hc_bytealign_be_S (w[26], w[27], offset); + w[48] = hc_bytealign_be_S (w[25], w[26], offset); + w[47] = hc_bytealign_be_S (w[24], w[25], offset); + w[46] = hc_bytealign_be_S (w[23], w[24], offset); + w[45] = hc_bytealign_be_S (w[22], w[23], offset); + w[44] = hc_bytealign_be_S (w[21], w[22], offset); + w[43] = hc_bytealign_be_S (w[20], w[21], offset); + w[42] = hc_bytealign_be_S (w[19], w[20], offset); + w[41] = hc_bytealign_be_S (w[18], w[19], offset); + w[40] = hc_bytealign_be_S (w[17], w[18], offset); + w[39] = hc_bytealign_be_S (w[16], w[17], offset); + w[38] = hc_bytealign_be_S (w[15], w[16], offset); + w[37] = hc_bytealign_be_S (w[14], w[15], offset); + w[36] = hc_bytealign_be_S (w[13], w[14], offset); + w[35] = hc_bytealign_be_S (w[12], w[13], offset); + w[34] = hc_bytealign_be_S (w[11], w[12], offset); + w[33] = hc_bytealign_be_S (w[10], w[11], offset); + w[32] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[31] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[30] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[29] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[28] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[27] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[26] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[25] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[24] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[23] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[22] = hc_bytealign_be_S ( 0, w[ 0], offset); w[21] = 0; w[20] = 0; w[19] = 0; @@ -52870,47 +52550,47 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 23: - w[63] = hc_bytealign_S (w[39], w[40], offset); - w[62] = hc_bytealign_S (w[38], w[39], offset); - w[61] = hc_bytealign_S (w[37], w[38], offset); - w[60] = hc_bytealign_S (w[36], w[37], offset); - w[59] = hc_bytealign_S (w[35], w[36], offset); - w[58] = hc_bytealign_S (w[34], w[35], offset); - w[57] = hc_bytealign_S (w[33], w[34], offset); - w[56] = hc_bytealign_S (w[32], w[33], offset); - w[55] = hc_bytealign_S (w[31], w[32], offset); - w[54] = hc_bytealign_S (w[30], w[31], offset); - w[53] = hc_bytealign_S (w[29], w[30], offset); - w[52] = hc_bytealign_S (w[28], w[29], offset); - w[51] = hc_bytealign_S (w[27], w[28], offset); - w[50] = hc_bytealign_S (w[26], w[27], offset); - w[49] = hc_bytealign_S (w[25], w[26], offset); - w[48] = hc_bytealign_S (w[24], w[25], offset); - w[47] = hc_bytealign_S (w[23], w[24], offset); - w[46] = hc_bytealign_S (w[22], w[23], offset); - w[45] = hc_bytealign_S (w[21], w[22], offset); - w[44] = hc_bytealign_S (w[20], w[21], offset); - w[43] = hc_bytealign_S (w[19], w[20], offset); - w[42] = hc_bytealign_S (w[18], w[19], offset); - w[41] = hc_bytealign_S (w[17], w[18], offset); - w[40] = hc_bytealign_S (w[16], w[17], offset); - w[39] = hc_bytealign_S (w[15], w[16], offset); - w[38] = hc_bytealign_S (w[14], w[15], offset); - w[37] = hc_bytealign_S (w[13], w[14], offset); - w[36] = hc_bytealign_S (w[12], w[13], offset); - w[35] = hc_bytealign_S (w[11], w[12], offset); - w[34] = hc_bytealign_S (w[10], w[11], offset); - w[33] = hc_bytealign_S (w[ 9], w[10], offset); - w[32] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[31] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[30] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[29] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[28] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[27] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[26] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[25] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[24] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[23] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[39], w[40], offset); + w[62] = hc_bytealign_be_S (w[38], w[39], offset); + w[61] = hc_bytealign_be_S (w[37], w[38], offset); + w[60] = hc_bytealign_be_S (w[36], w[37], offset); + w[59] = hc_bytealign_be_S (w[35], w[36], offset); + w[58] = hc_bytealign_be_S (w[34], w[35], offset); + w[57] = hc_bytealign_be_S (w[33], w[34], offset); + w[56] = hc_bytealign_be_S (w[32], w[33], offset); + w[55] = hc_bytealign_be_S (w[31], w[32], offset); + w[54] = hc_bytealign_be_S (w[30], w[31], offset); + w[53] = hc_bytealign_be_S (w[29], w[30], offset); + w[52] = hc_bytealign_be_S (w[28], w[29], offset); + w[51] = hc_bytealign_be_S (w[27], w[28], offset); + w[50] = hc_bytealign_be_S (w[26], w[27], offset); + w[49] = hc_bytealign_be_S (w[25], w[26], offset); + w[48] = hc_bytealign_be_S (w[24], w[25], offset); + w[47] = hc_bytealign_be_S (w[23], w[24], offset); + w[46] = hc_bytealign_be_S (w[22], w[23], offset); + w[45] = hc_bytealign_be_S (w[21], w[22], offset); + w[44] = hc_bytealign_be_S (w[20], w[21], offset); + w[43] = hc_bytealign_be_S (w[19], w[20], offset); + w[42] = hc_bytealign_be_S (w[18], w[19], offset); + w[41] = hc_bytealign_be_S (w[17], w[18], offset); + w[40] = hc_bytealign_be_S (w[16], w[17], offset); + w[39] = hc_bytealign_be_S (w[15], w[16], offset); + w[38] = hc_bytealign_be_S (w[14], w[15], offset); + w[37] = hc_bytealign_be_S (w[13], w[14], offset); + w[36] = hc_bytealign_be_S (w[12], w[13], offset); + w[35] = hc_bytealign_be_S (w[11], w[12], offset); + w[34] = hc_bytealign_be_S (w[10], w[11], offset); + w[33] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[32] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[31] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[30] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[29] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[28] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[27] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[26] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[25] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[24] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[23] = hc_bytealign_be_S ( 0, w[ 0], offset); w[22] = 0; w[21] = 0; w[20] = 0; @@ -52938,46 +52618,46 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 24: - w[63] = hc_bytealign_S (w[38], w[39], offset); - w[62] = hc_bytealign_S (w[37], w[38], offset); - w[61] = hc_bytealign_S (w[36], w[37], offset); - w[60] = hc_bytealign_S (w[35], w[36], offset); - w[59] = hc_bytealign_S (w[34], w[35], offset); - w[58] = hc_bytealign_S (w[33], w[34], offset); - w[57] = hc_bytealign_S (w[32], w[33], offset); - w[56] = hc_bytealign_S (w[31], w[32], offset); - w[55] = hc_bytealign_S (w[30], w[31], offset); - w[54] = hc_bytealign_S (w[29], w[30], offset); - w[53] = hc_bytealign_S (w[28], w[29], offset); - w[52] = hc_bytealign_S (w[27], w[28], offset); - w[51] = hc_bytealign_S (w[26], w[27], offset); - w[50] = hc_bytealign_S (w[25], w[26], offset); - w[49] = hc_bytealign_S (w[24], w[25], offset); - w[48] = hc_bytealign_S (w[23], w[24], offset); - w[47] = hc_bytealign_S (w[22], w[23], offset); - w[46] = hc_bytealign_S (w[21], w[22], offset); - w[45] = hc_bytealign_S (w[20], w[21], offset); - w[44] = hc_bytealign_S (w[19], w[20], offset); - w[43] = hc_bytealign_S (w[18], w[19], offset); - w[42] = hc_bytealign_S (w[17], w[18], offset); - w[41] = hc_bytealign_S (w[16], w[17], offset); - w[40] = hc_bytealign_S (w[15], w[16], offset); - w[39] = hc_bytealign_S (w[14], w[15], offset); - w[38] = hc_bytealign_S (w[13], w[14], offset); - w[37] = hc_bytealign_S (w[12], w[13], offset); - w[36] = hc_bytealign_S (w[11], w[12], offset); - w[35] = hc_bytealign_S (w[10], w[11], offset); - w[34] = hc_bytealign_S (w[ 9], w[10], offset); - w[33] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[32] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[31] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[30] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[29] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[28] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[27] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[26] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[25] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[24] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[38], w[39], offset); + w[62] = hc_bytealign_be_S (w[37], w[38], offset); + w[61] = hc_bytealign_be_S (w[36], w[37], offset); + w[60] = hc_bytealign_be_S (w[35], w[36], offset); + w[59] = hc_bytealign_be_S (w[34], w[35], offset); + w[58] = hc_bytealign_be_S (w[33], w[34], offset); + w[57] = hc_bytealign_be_S (w[32], w[33], offset); + w[56] = hc_bytealign_be_S (w[31], w[32], offset); + w[55] = hc_bytealign_be_S (w[30], w[31], offset); + w[54] = hc_bytealign_be_S (w[29], w[30], offset); + w[53] = hc_bytealign_be_S (w[28], w[29], offset); + w[52] = hc_bytealign_be_S (w[27], w[28], offset); + w[51] = hc_bytealign_be_S (w[26], w[27], offset); + w[50] = hc_bytealign_be_S (w[25], w[26], offset); + w[49] = hc_bytealign_be_S (w[24], w[25], offset); + w[48] = hc_bytealign_be_S (w[23], w[24], offset); + w[47] = hc_bytealign_be_S (w[22], w[23], offset); + w[46] = hc_bytealign_be_S (w[21], w[22], offset); + w[45] = hc_bytealign_be_S (w[20], w[21], offset); + w[44] = hc_bytealign_be_S (w[19], w[20], offset); + w[43] = hc_bytealign_be_S (w[18], w[19], offset); + w[42] = hc_bytealign_be_S (w[17], w[18], offset); + w[41] = hc_bytealign_be_S (w[16], w[17], offset); + w[40] = hc_bytealign_be_S (w[15], w[16], offset); + w[39] = hc_bytealign_be_S (w[14], w[15], offset); + w[38] = hc_bytealign_be_S (w[13], w[14], offset); + w[37] = hc_bytealign_be_S (w[12], w[13], offset); + w[36] = hc_bytealign_be_S (w[11], w[12], offset); + w[35] = hc_bytealign_be_S (w[10], w[11], offset); + w[34] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[33] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[32] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[31] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[30] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[29] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[28] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[27] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[26] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[25] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[24] = hc_bytealign_be_S ( 0, w[ 0], offset); w[23] = 0; w[22] = 0; w[21] = 0; @@ -53006,45 +52686,45 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 25: - w[63] = hc_bytealign_S (w[37], w[38], offset); - w[62] = hc_bytealign_S (w[36], w[37], offset); - w[61] = hc_bytealign_S (w[35], w[36], offset); - w[60] = hc_bytealign_S (w[34], w[35], offset); - w[59] = hc_bytealign_S (w[33], w[34], offset); - w[58] = hc_bytealign_S (w[32], w[33], offset); - w[57] = hc_bytealign_S (w[31], w[32], offset); - w[56] = hc_bytealign_S (w[30], w[31], offset); - w[55] = hc_bytealign_S (w[29], w[30], offset); - w[54] = hc_bytealign_S (w[28], w[29], offset); - w[53] = hc_bytealign_S (w[27], w[28], offset); - w[52] = hc_bytealign_S (w[26], w[27], offset); - w[51] = hc_bytealign_S (w[25], w[26], offset); - w[50] = hc_bytealign_S (w[24], w[25], offset); - w[49] = hc_bytealign_S (w[23], w[24], offset); - w[48] = hc_bytealign_S (w[22], w[23], offset); - w[47] = hc_bytealign_S (w[21], w[22], offset); - w[46] = hc_bytealign_S (w[20], w[21], offset); - w[45] = hc_bytealign_S (w[19], w[20], offset); - w[44] = hc_bytealign_S (w[18], w[19], offset); - w[43] = hc_bytealign_S (w[17], w[18], offset); - w[42] = hc_bytealign_S (w[16], w[17], offset); - w[41] = hc_bytealign_S (w[15], w[16], offset); - w[40] = hc_bytealign_S (w[14], w[15], offset); - w[39] = hc_bytealign_S (w[13], w[14], offset); - w[38] = hc_bytealign_S (w[12], w[13], offset); - w[37] = hc_bytealign_S (w[11], w[12], offset); - w[36] = hc_bytealign_S (w[10], w[11], offset); - w[35] = hc_bytealign_S (w[ 9], w[10], offset); - w[34] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[33] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[32] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[31] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[30] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[29] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[28] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[27] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[26] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[25] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[37], w[38], offset); + w[62] = hc_bytealign_be_S (w[36], w[37], offset); + w[61] = hc_bytealign_be_S (w[35], w[36], offset); + w[60] = hc_bytealign_be_S (w[34], w[35], offset); + w[59] = hc_bytealign_be_S (w[33], w[34], offset); + w[58] = hc_bytealign_be_S (w[32], w[33], offset); + w[57] = hc_bytealign_be_S (w[31], w[32], offset); + w[56] = hc_bytealign_be_S (w[30], w[31], offset); + w[55] = hc_bytealign_be_S (w[29], w[30], offset); + w[54] = hc_bytealign_be_S (w[28], w[29], offset); + w[53] = hc_bytealign_be_S (w[27], w[28], offset); + w[52] = hc_bytealign_be_S (w[26], w[27], offset); + w[51] = hc_bytealign_be_S (w[25], w[26], offset); + w[50] = hc_bytealign_be_S (w[24], w[25], offset); + w[49] = hc_bytealign_be_S (w[23], w[24], offset); + w[48] = hc_bytealign_be_S (w[22], w[23], offset); + w[47] = hc_bytealign_be_S (w[21], w[22], offset); + w[46] = hc_bytealign_be_S (w[20], w[21], offset); + w[45] = hc_bytealign_be_S (w[19], w[20], offset); + w[44] = hc_bytealign_be_S (w[18], w[19], offset); + w[43] = hc_bytealign_be_S (w[17], w[18], offset); + w[42] = hc_bytealign_be_S (w[16], w[17], offset); + w[41] = hc_bytealign_be_S (w[15], w[16], offset); + w[40] = hc_bytealign_be_S (w[14], w[15], offset); + w[39] = hc_bytealign_be_S (w[13], w[14], offset); + w[38] = hc_bytealign_be_S (w[12], w[13], offset); + w[37] = hc_bytealign_be_S (w[11], w[12], offset); + w[36] = hc_bytealign_be_S (w[10], w[11], offset); + w[35] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[34] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[33] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[32] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[31] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[30] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[29] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[28] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[27] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[26] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[25] = hc_bytealign_be_S ( 0, w[ 0], offset); w[24] = 0; w[23] = 0; w[22] = 0; @@ -53074,44 +52754,44 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 26: - w[63] = hc_bytealign_S (w[36], w[37], offset); - w[62] = hc_bytealign_S (w[35], w[36], offset); - w[61] = hc_bytealign_S (w[34], w[35], offset); - w[60] = hc_bytealign_S (w[33], w[34], offset); - w[59] = hc_bytealign_S (w[32], w[33], offset); - w[58] = hc_bytealign_S (w[31], w[32], offset); - w[57] = hc_bytealign_S (w[30], w[31], offset); - w[56] = hc_bytealign_S (w[29], w[30], offset); - w[55] = hc_bytealign_S (w[28], w[29], offset); - w[54] = hc_bytealign_S (w[27], w[28], offset); - w[53] = hc_bytealign_S (w[26], w[27], offset); - w[52] = hc_bytealign_S (w[25], w[26], offset); - w[51] = hc_bytealign_S (w[24], w[25], offset); - w[50] = hc_bytealign_S (w[23], w[24], offset); - w[49] = hc_bytealign_S (w[22], w[23], offset); - w[48] = hc_bytealign_S (w[21], w[22], offset); - w[47] = hc_bytealign_S (w[20], w[21], offset); - w[46] = hc_bytealign_S (w[19], w[20], offset); - w[45] = hc_bytealign_S (w[18], w[19], offset); - w[44] = hc_bytealign_S (w[17], w[18], offset); - w[43] = hc_bytealign_S (w[16], w[17], offset); - w[42] = hc_bytealign_S (w[15], w[16], offset); - w[41] = hc_bytealign_S (w[14], w[15], offset); - w[40] = hc_bytealign_S (w[13], w[14], offset); - w[39] = hc_bytealign_S (w[12], w[13], offset); - w[38] = hc_bytealign_S (w[11], w[12], offset); - w[37] = hc_bytealign_S (w[10], w[11], offset); - w[36] = hc_bytealign_S (w[ 9], w[10], offset); - w[35] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[34] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[33] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[32] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[31] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[30] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[29] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[28] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[27] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[26] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[36], w[37], offset); + w[62] = hc_bytealign_be_S (w[35], w[36], offset); + w[61] = hc_bytealign_be_S (w[34], w[35], offset); + w[60] = hc_bytealign_be_S (w[33], w[34], offset); + w[59] = hc_bytealign_be_S (w[32], w[33], offset); + w[58] = hc_bytealign_be_S (w[31], w[32], offset); + w[57] = hc_bytealign_be_S (w[30], w[31], offset); + w[56] = hc_bytealign_be_S (w[29], w[30], offset); + w[55] = hc_bytealign_be_S (w[28], w[29], offset); + w[54] = hc_bytealign_be_S (w[27], w[28], offset); + w[53] = hc_bytealign_be_S (w[26], w[27], offset); + w[52] = hc_bytealign_be_S (w[25], w[26], offset); + w[51] = hc_bytealign_be_S (w[24], w[25], offset); + w[50] = hc_bytealign_be_S (w[23], w[24], offset); + w[49] = hc_bytealign_be_S (w[22], w[23], offset); + w[48] = hc_bytealign_be_S (w[21], w[22], offset); + w[47] = hc_bytealign_be_S (w[20], w[21], offset); + w[46] = hc_bytealign_be_S (w[19], w[20], offset); + w[45] = hc_bytealign_be_S (w[18], w[19], offset); + w[44] = hc_bytealign_be_S (w[17], w[18], offset); + w[43] = hc_bytealign_be_S (w[16], w[17], offset); + w[42] = hc_bytealign_be_S (w[15], w[16], offset); + w[41] = hc_bytealign_be_S (w[14], w[15], offset); + w[40] = hc_bytealign_be_S (w[13], w[14], offset); + w[39] = hc_bytealign_be_S (w[12], w[13], offset); + w[38] = hc_bytealign_be_S (w[11], w[12], offset); + w[37] = hc_bytealign_be_S (w[10], w[11], offset); + w[36] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[35] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[34] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[33] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[32] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[31] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[30] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[29] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[28] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[27] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[26] = hc_bytealign_be_S ( 0, w[ 0], offset); w[25] = 0; w[24] = 0; w[23] = 0; @@ -53142,43 +52822,43 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 27: - w[63] = hc_bytealign_S (w[35], w[36], offset); - w[62] = hc_bytealign_S (w[34], w[35], offset); - w[61] = hc_bytealign_S (w[33], w[34], offset); - w[60] = hc_bytealign_S (w[32], w[33], offset); - w[59] = hc_bytealign_S (w[31], w[32], offset); - w[58] = hc_bytealign_S (w[30], w[31], offset); - w[57] = hc_bytealign_S (w[29], w[30], offset); - w[56] = hc_bytealign_S (w[28], w[29], offset); - w[55] = hc_bytealign_S (w[27], w[28], offset); - w[54] = hc_bytealign_S (w[26], w[27], offset); - w[53] = hc_bytealign_S (w[25], w[26], offset); - w[52] = hc_bytealign_S (w[24], w[25], offset); - w[51] = hc_bytealign_S (w[23], w[24], offset); - w[50] = hc_bytealign_S (w[22], w[23], offset); - w[49] = hc_bytealign_S (w[21], w[22], offset); - w[48] = hc_bytealign_S (w[20], w[21], offset); - w[47] = hc_bytealign_S (w[19], w[20], offset); - w[46] = hc_bytealign_S (w[18], w[19], offset); - w[45] = hc_bytealign_S (w[17], w[18], offset); - w[44] = hc_bytealign_S (w[16], w[17], offset); - w[43] = hc_bytealign_S (w[15], w[16], offset); - w[42] = hc_bytealign_S (w[14], w[15], offset); - w[41] = hc_bytealign_S (w[13], w[14], offset); - w[40] = hc_bytealign_S (w[12], w[13], offset); - w[39] = hc_bytealign_S (w[11], w[12], offset); - w[38] = hc_bytealign_S (w[10], w[11], offset); - w[37] = hc_bytealign_S (w[ 9], w[10], offset); - w[36] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[35] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[34] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[33] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[32] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[31] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[30] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[29] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[28] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[27] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[35], w[36], offset); + w[62] = hc_bytealign_be_S (w[34], w[35], offset); + w[61] = hc_bytealign_be_S (w[33], w[34], offset); + w[60] = hc_bytealign_be_S (w[32], w[33], offset); + w[59] = hc_bytealign_be_S (w[31], w[32], offset); + w[58] = hc_bytealign_be_S (w[30], w[31], offset); + w[57] = hc_bytealign_be_S (w[29], w[30], offset); + w[56] = hc_bytealign_be_S (w[28], w[29], offset); + w[55] = hc_bytealign_be_S (w[27], w[28], offset); + w[54] = hc_bytealign_be_S (w[26], w[27], offset); + w[53] = hc_bytealign_be_S (w[25], w[26], offset); + w[52] = hc_bytealign_be_S (w[24], w[25], offset); + w[51] = hc_bytealign_be_S (w[23], w[24], offset); + w[50] = hc_bytealign_be_S (w[22], w[23], offset); + w[49] = hc_bytealign_be_S (w[21], w[22], offset); + w[48] = hc_bytealign_be_S (w[20], w[21], offset); + w[47] = hc_bytealign_be_S (w[19], w[20], offset); + w[46] = hc_bytealign_be_S (w[18], w[19], offset); + w[45] = hc_bytealign_be_S (w[17], w[18], offset); + w[44] = hc_bytealign_be_S (w[16], w[17], offset); + w[43] = hc_bytealign_be_S (w[15], w[16], offset); + w[42] = hc_bytealign_be_S (w[14], w[15], offset); + w[41] = hc_bytealign_be_S (w[13], w[14], offset); + w[40] = hc_bytealign_be_S (w[12], w[13], offset); + w[39] = hc_bytealign_be_S (w[11], w[12], offset); + w[38] = hc_bytealign_be_S (w[10], w[11], offset); + w[37] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[36] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[35] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[34] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[33] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[32] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[31] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[30] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[29] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[28] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[27] = hc_bytealign_be_S ( 0, w[ 0], offset); w[26] = 0; w[25] = 0; w[24] = 0; @@ -53210,42 +52890,42 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 28: - w[63] = hc_bytealign_S (w[34], w[35], offset); - w[62] = hc_bytealign_S (w[33], w[34], offset); - w[61] = hc_bytealign_S (w[32], w[33], offset); - w[60] = hc_bytealign_S (w[31], w[32], offset); - w[59] = hc_bytealign_S (w[30], w[31], offset); - w[58] = hc_bytealign_S (w[29], w[30], offset); - w[57] = hc_bytealign_S (w[28], w[29], offset); - w[56] = hc_bytealign_S (w[27], w[28], offset); - w[55] = hc_bytealign_S (w[26], w[27], offset); - w[54] = hc_bytealign_S (w[25], w[26], offset); - w[53] = hc_bytealign_S (w[24], w[25], offset); - w[52] = hc_bytealign_S (w[23], w[24], offset); - w[51] = hc_bytealign_S (w[22], w[23], offset); - w[50] = hc_bytealign_S (w[21], w[22], offset); - w[49] = hc_bytealign_S (w[20], w[21], offset); - w[48] = hc_bytealign_S (w[19], w[20], offset); - w[47] = hc_bytealign_S (w[18], w[19], offset); - w[46] = hc_bytealign_S (w[17], w[18], offset); - w[45] = hc_bytealign_S (w[16], w[17], offset); - w[44] = hc_bytealign_S (w[15], w[16], offset); - w[43] = hc_bytealign_S (w[14], w[15], offset); - w[42] = hc_bytealign_S (w[13], w[14], offset); - w[41] = hc_bytealign_S (w[12], w[13], offset); - w[40] = hc_bytealign_S (w[11], w[12], offset); - w[39] = hc_bytealign_S (w[10], w[11], offset); - w[38] = hc_bytealign_S (w[ 9], w[10], offset); - w[37] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[36] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[35] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[34] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[33] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[32] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[31] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[30] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[29] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[28] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[34], w[35], offset); + w[62] = hc_bytealign_be_S (w[33], w[34], offset); + w[61] = hc_bytealign_be_S (w[32], w[33], offset); + w[60] = hc_bytealign_be_S (w[31], w[32], offset); + w[59] = hc_bytealign_be_S (w[30], w[31], offset); + w[58] = hc_bytealign_be_S (w[29], w[30], offset); + w[57] = hc_bytealign_be_S (w[28], w[29], offset); + w[56] = hc_bytealign_be_S (w[27], w[28], offset); + w[55] = hc_bytealign_be_S (w[26], w[27], offset); + w[54] = hc_bytealign_be_S (w[25], w[26], offset); + w[53] = hc_bytealign_be_S (w[24], w[25], offset); + w[52] = hc_bytealign_be_S (w[23], w[24], offset); + w[51] = hc_bytealign_be_S (w[22], w[23], offset); + w[50] = hc_bytealign_be_S (w[21], w[22], offset); + w[49] = hc_bytealign_be_S (w[20], w[21], offset); + w[48] = hc_bytealign_be_S (w[19], w[20], offset); + w[47] = hc_bytealign_be_S (w[18], w[19], offset); + w[46] = hc_bytealign_be_S (w[17], w[18], offset); + w[45] = hc_bytealign_be_S (w[16], w[17], offset); + w[44] = hc_bytealign_be_S (w[15], w[16], offset); + w[43] = hc_bytealign_be_S (w[14], w[15], offset); + w[42] = hc_bytealign_be_S (w[13], w[14], offset); + w[41] = hc_bytealign_be_S (w[12], w[13], offset); + w[40] = hc_bytealign_be_S (w[11], w[12], offset); + w[39] = hc_bytealign_be_S (w[10], w[11], offset); + w[38] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[37] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[36] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[35] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[34] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[33] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[32] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[31] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[30] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[29] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[28] = hc_bytealign_be_S ( 0, w[ 0], offset); w[27] = 0; w[26] = 0; w[25] = 0; @@ -53278,41 +52958,41 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 29: - w[63] = hc_bytealign_S (w[33], w[34], offset); - w[62] = hc_bytealign_S (w[32], w[33], offset); - w[61] = hc_bytealign_S (w[31], w[32], offset); - w[60] = hc_bytealign_S (w[30], w[31], offset); - w[59] = hc_bytealign_S (w[29], w[30], offset); - w[58] = hc_bytealign_S (w[28], w[29], offset); - w[57] = hc_bytealign_S (w[27], w[28], offset); - w[56] = hc_bytealign_S (w[26], w[27], offset); - w[55] = hc_bytealign_S (w[25], w[26], offset); - w[54] = hc_bytealign_S (w[24], w[25], offset); - w[53] = hc_bytealign_S (w[23], w[24], offset); - w[52] = hc_bytealign_S (w[22], w[23], offset); - w[51] = hc_bytealign_S (w[21], w[22], offset); - w[50] = hc_bytealign_S (w[20], w[21], offset); - w[49] = hc_bytealign_S (w[19], w[20], offset); - w[48] = hc_bytealign_S (w[18], w[19], offset); - w[47] = hc_bytealign_S (w[17], w[18], offset); - w[46] = hc_bytealign_S (w[16], w[17], offset); - w[45] = hc_bytealign_S (w[15], w[16], offset); - w[44] = hc_bytealign_S (w[14], w[15], offset); - w[43] = hc_bytealign_S (w[13], w[14], offset); - w[42] = hc_bytealign_S (w[12], w[13], offset); - w[41] = hc_bytealign_S (w[11], w[12], offset); - w[40] = hc_bytealign_S (w[10], w[11], offset); - w[39] = hc_bytealign_S (w[ 9], w[10], offset); - w[38] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[37] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[36] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[35] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[34] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[33] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[32] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[31] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[30] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[29] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[33], w[34], offset); + w[62] = hc_bytealign_be_S (w[32], w[33], offset); + w[61] = hc_bytealign_be_S (w[31], w[32], offset); + w[60] = hc_bytealign_be_S (w[30], w[31], offset); + w[59] = hc_bytealign_be_S (w[29], w[30], offset); + w[58] = hc_bytealign_be_S (w[28], w[29], offset); + w[57] = hc_bytealign_be_S (w[27], w[28], offset); + w[56] = hc_bytealign_be_S (w[26], w[27], offset); + w[55] = hc_bytealign_be_S (w[25], w[26], offset); + w[54] = hc_bytealign_be_S (w[24], w[25], offset); + w[53] = hc_bytealign_be_S (w[23], w[24], offset); + w[52] = hc_bytealign_be_S (w[22], w[23], offset); + w[51] = hc_bytealign_be_S (w[21], w[22], offset); + w[50] = hc_bytealign_be_S (w[20], w[21], offset); + w[49] = hc_bytealign_be_S (w[19], w[20], offset); + w[48] = hc_bytealign_be_S (w[18], w[19], offset); + w[47] = hc_bytealign_be_S (w[17], w[18], offset); + w[46] = hc_bytealign_be_S (w[16], w[17], offset); + w[45] = hc_bytealign_be_S (w[15], w[16], offset); + w[44] = hc_bytealign_be_S (w[14], w[15], offset); + w[43] = hc_bytealign_be_S (w[13], w[14], offset); + w[42] = hc_bytealign_be_S (w[12], w[13], offset); + w[41] = hc_bytealign_be_S (w[11], w[12], offset); + w[40] = hc_bytealign_be_S (w[10], w[11], offset); + w[39] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[38] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[37] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[36] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[35] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[34] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[33] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[32] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[31] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[30] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[29] = hc_bytealign_be_S ( 0, w[ 0], offset); w[28] = 0; w[27] = 0; w[26] = 0; @@ -53346,40 +53026,40 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 30: - w[63] = hc_bytealign_S (w[32], w[33], offset); - w[62] = hc_bytealign_S (w[31], w[32], offset); - w[61] = hc_bytealign_S (w[30], w[31], offset); - w[60] = hc_bytealign_S (w[29], w[30], offset); - w[59] = hc_bytealign_S (w[28], w[29], offset); - w[58] = hc_bytealign_S (w[27], w[28], offset); - w[57] = hc_bytealign_S (w[26], w[27], offset); - w[56] = hc_bytealign_S (w[25], w[26], offset); - w[55] = hc_bytealign_S (w[24], w[25], offset); - w[54] = hc_bytealign_S (w[23], w[24], offset); - w[53] = hc_bytealign_S (w[22], w[23], offset); - w[52] = hc_bytealign_S (w[21], w[22], offset); - w[51] = hc_bytealign_S (w[20], w[21], offset); - w[50] = hc_bytealign_S (w[19], w[20], offset); - w[49] = hc_bytealign_S (w[18], w[19], offset); - w[48] = hc_bytealign_S (w[17], w[18], offset); - w[47] = hc_bytealign_S (w[16], w[17], offset); - w[46] = hc_bytealign_S (w[15], w[16], offset); - w[45] = hc_bytealign_S (w[14], w[15], offset); - w[44] = hc_bytealign_S (w[13], w[14], offset); - w[43] = hc_bytealign_S (w[12], w[13], offset); - w[42] = hc_bytealign_S (w[11], w[12], offset); - w[41] = hc_bytealign_S (w[10], w[11], offset); - w[40] = hc_bytealign_S (w[ 9], w[10], offset); - w[39] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[38] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[37] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[36] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[35] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[34] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[33] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[32] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[31] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[30] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[32], w[33], offset); + w[62] = hc_bytealign_be_S (w[31], w[32], offset); + w[61] = hc_bytealign_be_S (w[30], w[31], offset); + w[60] = hc_bytealign_be_S (w[29], w[30], offset); + w[59] = hc_bytealign_be_S (w[28], w[29], offset); + w[58] = hc_bytealign_be_S (w[27], w[28], offset); + w[57] = hc_bytealign_be_S (w[26], w[27], offset); + w[56] = hc_bytealign_be_S (w[25], w[26], offset); + w[55] = hc_bytealign_be_S (w[24], w[25], offset); + w[54] = hc_bytealign_be_S (w[23], w[24], offset); + w[53] = hc_bytealign_be_S (w[22], w[23], offset); + w[52] = hc_bytealign_be_S (w[21], w[22], offset); + w[51] = hc_bytealign_be_S (w[20], w[21], offset); + w[50] = hc_bytealign_be_S (w[19], w[20], offset); + w[49] = hc_bytealign_be_S (w[18], w[19], offset); + w[48] = hc_bytealign_be_S (w[17], w[18], offset); + w[47] = hc_bytealign_be_S (w[16], w[17], offset); + w[46] = hc_bytealign_be_S (w[15], w[16], offset); + w[45] = hc_bytealign_be_S (w[14], w[15], offset); + w[44] = hc_bytealign_be_S (w[13], w[14], offset); + w[43] = hc_bytealign_be_S (w[12], w[13], offset); + w[42] = hc_bytealign_be_S (w[11], w[12], offset); + w[41] = hc_bytealign_be_S (w[10], w[11], offset); + w[40] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[39] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[38] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[37] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[36] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[35] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[34] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[33] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[32] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[31] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[30] = hc_bytealign_be_S ( 0, w[ 0], offset); w[29] = 0; w[28] = 0; w[27] = 0; @@ -53414,39 +53094,39 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 31: - w[63] = hc_bytealign_S (w[31], w[32], offset); - w[62] = hc_bytealign_S (w[30], w[31], offset); - w[61] = hc_bytealign_S (w[29], w[30], offset); - w[60] = hc_bytealign_S (w[28], w[29], offset); - w[59] = hc_bytealign_S (w[27], w[28], offset); - w[58] = hc_bytealign_S (w[26], w[27], offset); - w[57] = hc_bytealign_S (w[25], w[26], offset); - w[56] = hc_bytealign_S (w[24], w[25], offset); - w[55] = hc_bytealign_S (w[23], w[24], offset); - w[54] = hc_bytealign_S (w[22], w[23], offset); - w[53] = hc_bytealign_S (w[21], w[22], offset); - w[52] = hc_bytealign_S (w[20], w[21], offset); - w[51] = hc_bytealign_S (w[19], w[20], offset); - w[50] = hc_bytealign_S (w[18], w[19], offset); - w[49] = hc_bytealign_S (w[17], w[18], offset); - w[48] = hc_bytealign_S (w[16], w[17], offset); - w[47] = hc_bytealign_S (w[15], w[16], offset); - w[46] = hc_bytealign_S (w[14], w[15], offset); - w[45] = hc_bytealign_S (w[13], w[14], offset); - w[44] = hc_bytealign_S (w[12], w[13], offset); - w[43] = hc_bytealign_S (w[11], w[12], offset); - w[42] = hc_bytealign_S (w[10], w[11], offset); - w[41] = hc_bytealign_S (w[ 9], w[10], offset); - w[40] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[39] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[38] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[37] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[36] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[35] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[34] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[33] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[32] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[31] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[31], w[32], offset); + w[62] = hc_bytealign_be_S (w[30], w[31], offset); + w[61] = hc_bytealign_be_S (w[29], w[30], offset); + w[60] = hc_bytealign_be_S (w[28], w[29], offset); + w[59] = hc_bytealign_be_S (w[27], w[28], offset); + w[58] = hc_bytealign_be_S (w[26], w[27], offset); + w[57] = hc_bytealign_be_S (w[25], w[26], offset); + w[56] = hc_bytealign_be_S (w[24], w[25], offset); + w[55] = hc_bytealign_be_S (w[23], w[24], offset); + w[54] = hc_bytealign_be_S (w[22], w[23], offset); + w[53] = hc_bytealign_be_S (w[21], w[22], offset); + w[52] = hc_bytealign_be_S (w[20], w[21], offset); + w[51] = hc_bytealign_be_S (w[19], w[20], offset); + w[50] = hc_bytealign_be_S (w[18], w[19], offset); + w[49] = hc_bytealign_be_S (w[17], w[18], offset); + w[48] = hc_bytealign_be_S (w[16], w[17], offset); + w[47] = hc_bytealign_be_S (w[15], w[16], offset); + w[46] = hc_bytealign_be_S (w[14], w[15], offset); + w[45] = hc_bytealign_be_S (w[13], w[14], offset); + w[44] = hc_bytealign_be_S (w[12], w[13], offset); + w[43] = hc_bytealign_be_S (w[11], w[12], offset); + w[42] = hc_bytealign_be_S (w[10], w[11], offset); + w[41] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[40] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[39] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[38] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[37] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[36] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[35] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[34] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[33] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[32] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[31] = hc_bytealign_be_S ( 0, w[ 0], offset); w[30] = 0; w[29] = 0; w[28] = 0; @@ -53482,38 +53162,38 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 32: - w[63] = hc_bytealign_S (w[30], w[31], offset); - w[62] = hc_bytealign_S (w[29], w[30], offset); - w[61] = hc_bytealign_S (w[28], w[29], offset); - w[60] = hc_bytealign_S (w[27], w[28], offset); - w[59] = hc_bytealign_S (w[26], w[27], offset); - w[58] = hc_bytealign_S (w[25], w[26], offset); - w[57] = hc_bytealign_S (w[24], w[25], offset); - w[56] = hc_bytealign_S (w[23], w[24], offset); - w[55] = hc_bytealign_S (w[22], w[23], offset); - w[54] = hc_bytealign_S (w[21], w[22], offset); - w[53] = hc_bytealign_S (w[20], w[21], offset); - w[52] = hc_bytealign_S (w[19], w[20], offset); - w[51] = hc_bytealign_S (w[18], w[19], offset); - w[50] = hc_bytealign_S (w[17], w[18], offset); - w[49] = hc_bytealign_S (w[16], w[17], offset); - w[48] = hc_bytealign_S (w[15], w[16], offset); - w[47] = hc_bytealign_S (w[14], w[15], offset); - w[46] = hc_bytealign_S (w[13], w[14], offset); - w[45] = hc_bytealign_S (w[12], w[13], offset); - w[44] = hc_bytealign_S (w[11], w[12], offset); - w[43] = hc_bytealign_S (w[10], w[11], offset); - w[42] = hc_bytealign_S (w[ 9], w[10], offset); - w[41] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[40] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[39] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[38] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[37] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[36] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[35] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[34] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[33] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[32] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[30], w[31], offset); + w[62] = hc_bytealign_be_S (w[29], w[30], offset); + w[61] = hc_bytealign_be_S (w[28], w[29], offset); + w[60] = hc_bytealign_be_S (w[27], w[28], offset); + w[59] = hc_bytealign_be_S (w[26], w[27], offset); + w[58] = hc_bytealign_be_S (w[25], w[26], offset); + w[57] = hc_bytealign_be_S (w[24], w[25], offset); + w[56] = hc_bytealign_be_S (w[23], w[24], offset); + w[55] = hc_bytealign_be_S (w[22], w[23], offset); + w[54] = hc_bytealign_be_S (w[21], w[22], offset); + w[53] = hc_bytealign_be_S (w[20], w[21], offset); + w[52] = hc_bytealign_be_S (w[19], w[20], offset); + w[51] = hc_bytealign_be_S (w[18], w[19], offset); + w[50] = hc_bytealign_be_S (w[17], w[18], offset); + w[49] = hc_bytealign_be_S (w[16], w[17], offset); + w[48] = hc_bytealign_be_S (w[15], w[16], offset); + w[47] = hc_bytealign_be_S (w[14], w[15], offset); + w[46] = hc_bytealign_be_S (w[13], w[14], offset); + w[45] = hc_bytealign_be_S (w[12], w[13], offset); + w[44] = hc_bytealign_be_S (w[11], w[12], offset); + w[43] = hc_bytealign_be_S (w[10], w[11], offset); + w[42] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[41] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[40] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[39] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[38] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[37] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[36] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[35] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[34] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[33] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[32] = hc_bytealign_be_S ( 0, w[ 0], offset); w[31] = 0; w[30] = 0; w[29] = 0; @@ -53550,37 +53230,37 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 33: - w[63] = hc_bytealign_S (w[29], w[30], offset); - w[62] = hc_bytealign_S (w[28], w[29], offset); - w[61] = hc_bytealign_S (w[27], w[28], offset); - w[60] = hc_bytealign_S (w[26], w[27], offset); - w[59] = hc_bytealign_S (w[25], w[26], offset); - w[58] = hc_bytealign_S (w[24], w[25], offset); - w[57] = hc_bytealign_S (w[23], w[24], offset); - w[56] = hc_bytealign_S (w[22], w[23], offset); - w[55] = hc_bytealign_S (w[21], w[22], offset); - w[54] = hc_bytealign_S (w[20], w[21], offset); - w[53] = hc_bytealign_S (w[19], w[20], offset); - w[52] = hc_bytealign_S (w[18], w[19], offset); - w[51] = hc_bytealign_S (w[17], w[18], offset); - w[50] = hc_bytealign_S (w[16], w[17], offset); - w[49] = hc_bytealign_S (w[15], w[16], offset); - w[48] = hc_bytealign_S (w[14], w[15], offset); - w[47] = hc_bytealign_S (w[13], w[14], offset); - w[46] = hc_bytealign_S (w[12], w[13], offset); - w[45] = hc_bytealign_S (w[11], w[12], offset); - w[44] = hc_bytealign_S (w[10], w[11], offset); - w[43] = hc_bytealign_S (w[ 9], w[10], offset); - w[42] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[41] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[40] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[39] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[38] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[37] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[36] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[35] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[34] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[33] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[29], w[30], offset); + w[62] = hc_bytealign_be_S (w[28], w[29], offset); + w[61] = hc_bytealign_be_S (w[27], w[28], offset); + w[60] = hc_bytealign_be_S (w[26], w[27], offset); + w[59] = hc_bytealign_be_S (w[25], w[26], offset); + w[58] = hc_bytealign_be_S (w[24], w[25], offset); + w[57] = hc_bytealign_be_S (w[23], w[24], offset); + w[56] = hc_bytealign_be_S (w[22], w[23], offset); + w[55] = hc_bytealign_be_S (w[21], w[22], offset); + w[54] = hc_bytealign_be_S (w[20], w[21], offset); + w[53] = hc_bytealign_be_S (w[19], w[20], offset); + w[52] = hc_bytealign_be_S (w[18], w[19], offset); + w[51] = hc_bytealign_be_S (w[17], w[18], offset); + w[50] = hc_bytealign_be_S (w[16], w[17], offset); + w[49] = hc_bytealign_be_S (w[15], w[16], offset); + w[48] = hc_bytealign_be_S (w[14], w[15], offset); + w[47] = hc_bytealign_be_S (w[13], w[14], offset); + w[46] = hc_bytealign_be_S (w[12], w[13], offset); + w[45] = hc_bytealign_be_S (w[11], w[12], offset); + w[44] = hc_bytealign_be_S (w[10], w[11], offset); + w[43] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[42] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[41] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[40] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[39] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[38] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[37] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[36] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[35] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[34] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[33] = hc_bytealign_be_S ( 0, w[ 0], offset); w[32] = 0; w[31] = 0; w[30] = 0; @@ -53618,36 +53298,36 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 34: - w[63] = hc_bytealign_S (w[28], w[29], offset); - w[62] = hc_bytealign_S (w[27], w[28], offset); - w[61] = hc_bytealign_S (w[26], w[27], offset); - w[60] = hc_bytealign_S (w[25], w[26], offset); - w[59] = hc_bytealign_S (w[24], w[25], offset); - w[58] = hc_bytealign_S (w[23], w[24], offset); - w[57] = hc_bytealign_S (w[22], w[23], offset); - w[56] = hc_bytealign_S (w[21], w[22], offset); - w[55] = hc_bytealign_S (w[20], w[21], offset); - w[54] = hc_bytealign_S (w[19], w[20], offset); - w[53] = hc_bytealign_S (w[18], w[19], offset); - w[52] = hc_bytealign_S (w[17], w[18], offset); - w[51] = hc_bytealign_S (w[16], w[17], offset); - w[50] = hc_bytealign_S (w[15], w[16], offset); - w[49] = hc_bytealign_S (w[14], w[15], offset); - w[48] = hc_bytealign_S (w[13], w[14], offset); - w[47] = hc_bytealign_S (w[12], w[13], offset); - w[46] = hc_bytealign_S (w[11], w[12], offset); - w[45] = hc_bytealign_S (w[10], w[11], offset); - w[44] = hc_bytealign_S (w[ 9], w[10], offset); - w[43] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[42] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[41] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[40] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[39] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[38] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[37] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[36] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[35] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[34] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[28], w[29], offset); + w[62] = hc_bytealign_be_S (w[27], w[28], offset); + w[61] = hc_bytealign_be_S (w[26], w[27], offset); + w[60] = hc_bytealign_be_S (w[25], w[26], offset); + w[59] = hc_bytealign_be_S (w[24], w[25], offset); + w[58] = hc_bytealign_be_S (w[23], w[24], offset); + w[57] = hc_bytealign_be_S (w[22], w[23], offset); + w[56] = hc_bytealign_be_S (w[21], w[22], offset); + w[55] = hc_bytealign_be_S (w[20], w[21], offset); + w[54] = hc_bytealign_be_S (w[19], w[20], offset); + w[53] = hc_bytealign_be_S (w[18], w[19], offset); + w[52] = hc_bytealign_be_S (w[17], w[18], offset); + w[51] = hc_bytealign_be_S (w[16], w[17], offset); + w[50] = hc_bytealign_be_S (w[15], w[16], offset); + w[49] = hc_bytealign_be_S (w[14], w[15], offset); + w[48] = hc_bytealign_be_S (w[13], w[14], offset); + w[47] = hc_bytealign_be_S (w[12], w[13], offset); + w[46] = hc_bytealign_be_S (w[11], w[12], offset); + w[45] = hc_bytealign_be_S (w[10], w[11], offset); + w[44] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[43] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[42] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[41] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[40] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[39] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[38] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[37] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[36] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[35] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[34] = hc_bytealign_be_S ( 0, w[ 0], offset); w[33] = 0; w[32] = 0; w[31] = 0; @@ -53686,35 +53366,35 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 35: - w[63] = hc_bytealign_S (w[27], w[28], offset); - w[62] = hc_bytealign_S (w[26], w[27], offset); - w[61] = hc_bytealign_S (w[25], w[26], offset); - w[60] = hc_bytealign_S (w[24], w[25], offset); - w[59] = hc_bytealign_S (w[23], w[24], offset); - w[58] = hc_bytealign_S (w[22], w[23], offset); - w[57] = hc_bytealign_S (w[21], w[22], offset); - w[56] = hc_bytealign_S (w[20], w[21], offset); - w[55] = hc_bytealign_S (w[19], w[20], offset); - w[54] = hc_bytealign_S (w[18], w[19], offset); - w[53] = hc_bytealign_S (w[17], w[18], offset); - w[52] = hc_bytealign_S (w[16], w[17], offset); - w[51] = hc_bytealign_S (w[15], w[16], offset); - w[50] = hc_bytealign_S (w[14], w[15], offset); - w[49] = hc_bytealign_S (w[13], w[14], offset); - w[48] = hc_bytealign_S (w[12], w[13], offset); - w[47] = hc_bytealign_S (w[11], w[12], offset); - w[46] = hc_bytealign_S (w[10], w[11], offset); - w[45] = hc_bytealign_S (w[ 9], w[10], offset); - w[44] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[43] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[42] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[41] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[40] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[39] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[38] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[37] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[36] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[35] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[27], w[28], offset); + w[62] = hc_bytealign_be_S (w[26], w[27], offset); + w[61] = hc_bytealign_be_S (w[25], w[26], offset); + w[60] = hc_bytealign_be_S (w[24], w[25], offset); + w[59] = hc_bytealign_be_S (w[23], w[24], offset); + w[58] = hc_bytealign_be_S (w[22], w[23], offset); + w[57] = hc_bytealign_be_S (w[21], w[22], offset); + w[56] = hc_bytealign_be_S (w[20], w[21], offset); + w[55] = hc_bytealign_be_S (w[19], w[20], offset); + w[54] = hc_bytealign_be_S (w[18], w[19], offset); + w[53] = hc_bytealign_be_S (w[17], w[18], offset); + w[52] = hc_bytealign_be_S (w[16], w[17], offset); + w[51] = hc_bytealign_be_S (w[15], w[16], offset); + w[50] = hc_bytealign_be_S (w[14], w[15], offset); + w[49] = hc_bytealign_be_S (w[13], w[14], offset); + w[48] = hc_bytealign_be_S (w[12], w[13], offset); + w[47] = hc_bytealign_be_S (w[11], w[12], offset); + w[46] = hc_bytealign_be_S (w[10], w[11], offset); + w[45] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[44] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[43] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[42] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[41] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[40] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[39] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[38] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[37] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[36] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[35] = hc_bytealign_be_S ( 0, w[ 0], offset); w[34] = 0; w[33] = 0; w[32] = 0; @@ -53754,34 +53434,34 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 36: - w[63] = hc_bytealign_S (w[26], w[27], offset); - w[62] = hc_bytealign_S (w[25], w[26], offset); - w[61] = hc_bytealign_S (w[24], w[25], offset); - w[60] = hc_bytealign_S (w[23], w[24], offset); - w[59] = hc_bytealign_S (w[22], w[23], offset); - w[58] = hc_bytealign_S (w[21], w[22], offset); - w[57] = hc_bytealign_S (w[20], w[21], offset); - w[56] = hc_bytealign_S (w[19], w[20], offset); - w[55] = hc_bytealign_S (w[18], w[19], offset); - w[54] = hc_bytealign_S (w[17], w[18], offset); - w[53] = hc_bytealign_S (w[16], w[17], offset); - w[52] = hc_bytealign_S (w[15], w[16], offset); - w[51] = hc_bytealign_S (w[14], w[15], offset); - w[50] = hc_bytealign_S (w[13], w[14], offset); - w[49] = hc_bytealign_S (w[12], w[13], offset); - w[48] = hc_bytealign_S (w[11], w[12], offset); - w[47] = hc_bytealign_S (w[10], w[11], offset); - w[46] = hc_bytealign_S (w[ 9], w[10], offset); - w[45] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[44] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[43] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[42] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[41] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[40] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[39] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[38] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[37] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[36] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[26], w[27], offset); + w[62] = hc_bytealign_be_S (w[25], w[26], offset); + w[61] = hc_bytealign_be_S (w[24], w[25], offset); + w[60] = hc_bytealign_be_S (w[23], w[24], offset); + w[59] = hc_bytealign_be_S (w[22], w[23], offset); + w[58] = hc_bytealign_be_S (w[21], w[22], offset); + w[57] = hc_bytealign_be_S (w[20], w[21], offset); + w[56] = hc_bytealign_be_S (w[19], w[20], offset); + w[55] = hc_bytealign_be_S (w[18], w[19], offset); + w[54] = hc_bytealign_be_S (w[17], w[18], offset); + w[53] = hc_bytealign_be_S (w[16], w[17], offset); + w[52] = hc_bytealign_be_S (w[15], w[16], offset); + w[51] = hc_bytealign_be_S (w[14], w[15], offset); + w[50] = hc_bytealign_be_S (w[13], w[14], offset); + w[49] = hc_bytealign_be_S (w[12], w[13], offset); + w[48] = hc_bytealign_be_S (w[11], w[12], offset); + w[47] = hc_bytealign_be_S (w[10], w[11], offset); + w[46] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[45] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[44] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[43] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[42] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[41] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[40] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[39] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[38] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[37] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[36] = hc_bytealign_be_S ( 0, w[ 0], offset); w[35] = 0; w[34] = 0; w[33] = 0; @@ -53822,33 +53502,33 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 37: - w[63] = hc_bytealign_S (w[25], w[26], offset); - w[62] = hc_bytealign_S (w[24], w[25], offset); - w[61] = hc_bytealign_S (w[23], w[24], offset); - w[60] = hc_bytealign_S (w[22], w[23], offset); - w[59] = hc_bytealign_S (w[21], w[22], offset); - w[58] = hc_bytealign_S (w[20], w[21], offset); - w[57] = hc_bytealign_S (w[19], w[20], offset); - w[56] = hc_bytealign_S (w[18], w[19], offset); - w[55] = hc_bytealign_S (w[17], w[18], offset); - w[54] = hc_bytealign_S (w[16], w[17], offset); - w[53] = hc_bytealign_S (w[15], w[16], offset); - w[52] = hc_bytealign_S (w[14], w[15], offset); - w[51] = hc_bytealign_S (w[13], w[14], offset); - w[50] = hc_bytealign_S (w[12], w[13], offset); - w[49] = hc_bytealign_S (w[11], w[12], offset); - w[48] = hc_bytealign_S (w[10], w[11], offset); - w[47] = hc_bytealign_S (w[ 9], w[10], offset); - w[46] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[45] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[44] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[43] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[42] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[41] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[40] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[39] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[38] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[37] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[25], w[26], offset); + w[62] = hc_bytealign_be_S (w[24], w[25], offset); + w[61] = hc_bytealign_be_S (w[23], w[24], offset); + w[60] = hc_bytealign_be_S (w[22], w[23], offset); + w[59] = hc_bytealign_be_S (w[21], w[22], offset); + w[58] = hc_bytealign_be_S (w[20], w[21], offset); + w[57] = hc_bytealign_be_S (w[19], w[20], offset); + w[56] = hc_bytealign_be_S (w[18], w[19], offset); + w[55] = hc_bytealign_be_S (w[17], w[18], offset); + w[54] = hc_bytealign_be_S (w[16], w[17], offset); + w[53] = hc_bytealign_be_S (w[15], w[16], offset); + w[52] = hc_bytealign_be_S (w[14], w[15], offset); + w[51] = hc_bytealign_be_S (w[13], w[14], offset); + w[50] = hc_bytealign_be_S (w[12], w[13], offset); + w[49] = hc_bytealign_be_S (w[11], w[12], offset); + w[48] = hc_bytealign_be_S (w[10], w[11], offset); + w[47] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[46] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[45] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[44] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[43] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[42] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[41] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[40] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[39] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[38] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[37] = hc_bytealign_be_S ( 0, w[ 0], offset); w[36] = 0; w[35] = 0; w[34] = 0; @@ -53890,32 +53570,32 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 38: - w[63] = hc_bytealign_S (w[24], w[25], offset); - w[62] = hc_bytealign_S (w[23], w[24], offset); - w[61] = hc_bytealign_S (w[22], w[23], offset); - w[60] = hc_bytealign_S (w[21], w[22], offset); - w[59] = hc_bytealign_S (w[20], w[21], offset); - w[58] = hc_bytealign_S (w[19], w[20], offset); - w[57] = hc_bytealign_S (w[18], w[19], offset); - w[56] = hc_bytealign_S (w[17], w[18], offset); - w[55] = hc_bytealign_S (w[16], w[17], offset); - w[54] = hc_bytealign_S (w[15], w[16], offset); - w[53] = hc_bytealign_S (w[14], w[15], offset); - w[52] = hc_bytealign_S (w[13], w[14], offset); - w[51] = hc_bytealign_S (w[12], w[13], offset); - w[50] = hc_bytealign_S (w[11], w[12], offset); - w[49] = hc_bytealign_S (w[10], w[11], offset); - w[48] = hc_bytealign_S (w[ 9], w[10], offset); - w[47] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[46] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[45] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[44] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[43] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[42] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[41] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[40] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[39] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[38] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[24], w[25], offset); + w[62] = hc_bytealign_be_S (w[23], w[24], offset); + w[61] = hc_bytealign_be_S (w[22], w[23], offset); + w[60] = hc_bytealign_be_S (w[21], w[22], offset); + w[59] = hc_bytealign_be_S (w[20], w[21], offset); + w[58] = hc_bytealign_be_S (w[19], w[20], offset); + w[57] = hc_bytealign_be_S (w[18], w[19], offset); + w[56] = hc_bytealign_be_S (w[17], w[18], offset); + w[55] = hc_bytealign_be_S (w[16], w[17], offset); + w[54] = hc_bytealign_be_S (w[15], w[16], offset); + w[53] = hc_bytealign_be_S (w[14], w[15], offset); + w[52] = hc_bytealign_be_S (w[13], w[14], offset); + w[51] = hc_bytealign_be_S (w[12], w[13], offset); + w[50] = hc_bytealign_be_S (w[11], w[12], offset); + w[49] = hc_bytealign_be_S (w[10], w[11], offset); + w[48] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[47] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[46] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[45] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[44] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[43] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[42] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[41] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[40] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[39] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[38] = hc_bytealign_be_S ( 0, w[ 0], offset); w[37] = 0; w[36] = 0; w[35] = 0; @@ -53958,31 +53638,31 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 39: - w[63] = hc_bytealign_S (w[23], w[24], offset); - w[62] = hc_bytealign_S (w[22], w[23], offset); - w[61] = hc_bytealign_S (w[21], w[22], offset); - w[60] = hc_bytealign_S (w[20], w[21], offset); - w[59] = hc_bytealign_S (w[19], w[20], offset); - w[58] = hc_bytealign_S (w[18], w[19], offset); - w[57] = hc_bytealign_S (w[17], w[18], offset); - w[56] = hc_bytealign_S (w[16], w[17], offset); - w[55] = hc_bytealign_S (w[15], w[16], offset); - w[54] = hc_bytealign_S (w[14], w[15], offset); - w[53] = hc_bytealign_S (w[13], w[14], offset); - w[52] = hc_bytealign_S (w[12], w[13], offset); - w[51] = hc_bytealign_S (w[11], w[12], offset); - w[50] = hc_bytealign_S (w[10], w[11], offset); - w[49] = hc_bytealign_S (w[ 9], w[10], offset); - w[48] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[47] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[46] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[45] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[44] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[43] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[42] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[41] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[40] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[39] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[23], w[24], offset); + w[62] = hc_bytealign_be_S (w[22], w[23], offset); + w[61] = hc_bytealign_be_S (w[21], w[22], offset); + w[60] = hc_bytealign_be_S (w[20], w[21], offset); + w[59] = hc_bytealign_be_S (w[19], w[20], offset); + w[58] = hc_bytealign_be_S (w[18], w[19], offset); + w[57] = hc_bytealign_be_S (w[17], w[18], offset); + w[56] = hc_bytealign_be_S (w[16], w[17], offset); + w[55] = hc_bytealign_be_S (w[15], w[16], offset); + w[54] = hc_bytealign_be_S (w[14], w[15], offset); + w[53] = hc_bytealign_be_S (w[13], w[14], offset); + w[52] = hc_bytealign_be_S (w[12], w[13], offset); + w[51] = hc_bytealign_be_S (w[11], w[12], offset); + w[50] = hc_bytealign_be_S (w[10], w[11], offset); + w[49] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[48] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[47] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[46] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[45] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[44] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[43] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[42] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[41] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[40] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[39] = hc_bytealign_be_S ( 0, w[ 0], offset); w[38] = 0; w[37] = 0; w[36] = 0; @@ -54026,30 +53706,30 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 40: - w[63] = hc_bytealign_S (w[22], w[23], offset); - w[62] = hc_bytealign_S (w[21], w[22], offset); - w[61] = hc_bytealign_S (w[20], w[21], offset); - w[60] = hc_bytealign_S (w[19], w[20], offset); - w[59] = hc_bytealign_S (w[18], w[19], offset); - w[58] = hc_bytealign_S (w[17], w[18], offset); - w[57] = hc_bytealign_S (w[16], w[17], offset); - w[56] = hc_bytealign_S (w[15], w[16], offset); - w[55] = hc_bytealign_S (w[14], w[15], offset); - w[54] = hc_bytealign_S (w[13], w[14], offset); - w[53] = hc_bytealign_S (w[12], w[13], offset); - w[52] = hc_bytealign_S (w[11], w[12], offset); - w[51] = hc_bytealign_S (w[10], w[11], offset); - w[50] = hc_bytealign_S (w[ 9], w[10], offset); - w[49] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[48] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[47] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[46] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[45] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[44] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[43] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[42] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[41] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[40] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[22], w[23], offset); + w[62] = hc_bytealign_be_S (w[21], w[22], offset); + w[61] = hc_bytealign_be_S (w[20], w[21], offset); + w[60] = hc_bytealign_be_S (w[19], w[20], offset); + w[59] = hc_bytealign_be_S (w[18], w[19], offset); + w[58] = hc_bytealign_be_S (w[17], w[18], offset); + w[57] = hc_bytealign_be_S (w[16], w[17], offset); + w[56] = hc_bytealign_be_S (w[15], w[16], offset); + w[55] = hc_bytealign_be_S (w[14], w[15], offset); + w[54] = hc_bytealign_be_S (w[13], w[14], offset); + w[53] = hc_bytealign_be_S (w[12], w[13], offset); + w[52] = hc_bytealign_be_S (w[11], w[12], offset); + w[51] = hc_bytealign_be_S (w[10], w[11], offset); + w[50] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[49] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[48] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[47] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[46] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[45] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[44] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[43] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[42] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[41] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[40] = hc_bytealign_be_S ( 0, w[ 0], offset); w[39] = 0; w[38] = 0; w[37] = 0; @@ -54094,29 +53774,29 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 41: - w[63] = hc_bytealign_S (w[21], w[22], offset); - w[62] = hc_bytealign_S (w[20], w[21], offset); - w[61] = hc_bytealign_S (w[19], w[20], offset); - w[60] = hc_bytealign_S (w[18], w[19], offset); - w[59] = hc_bytealign_S (w[17], w[18], offset); - w[58] = hc_bytealign_S (w[16], w[17], offset); - w[57] = hc_bytealign_S (w[15], w[16], offset); - w[56] = hc_bytealign_S (w[14], w[15], offset); - w[55] = hc_bytealign_S (w[13], w[14], offset); - w[54] = hc_bytealign_S (w[12], w[13], offset); - w[53] = hc_bytealign_S (w[11], w[12], offset); - w[52] = hc_bytealign_S (w[10], w[11], offset); - w[51] = hc_bytealign_S (w[ 9], w[10], offset); - w[50] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[49] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[48] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[47] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[46] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[45] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[44] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[43] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[42] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[41] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[21], w[22], offset); + w[62] = hc_bytealign_be_S (w[20], w[21], offset); + w[61] = hc_bytealign_be_S (w[19], w[20], offset); + w[60] = hc_bytealign_be_S (w[18], w[19], offset); + w[59] = hc_bytealign_be_S (w[17], w[18], offset); + w[58] = hc_bytealign_be_S (w[16], w[17], offset); + w[57] = hc_bytealign_be_S (w[15], w[16], offset); + w[56] = hc_bytealign_be_S (w[14], w[15], offset); + w[55] = hc_bytealign_be_S (w[13], w[14], offset); + w[54] = hc_bytealign_be_S (w[12], w[13], offset); + w[53] = hc_bytealign_be_S (w[11], w[12], offset); + w[52] = hc_bytealign_be_S (w[10], w[11], offset); + w[51] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[50] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[49] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[48] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[47] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[46] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[45] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[44] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[43] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[42] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[41] = hc_bytealign_be_S ( 0, w[ 0], offset); w[40] = 0; w[39] = 0; w[38] = 0; @@ -54162,28 +53842,28 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 42: - w[63] = hc_bytealign_S (w[20], w[21], offset); - w[62] = hc_bytealign_S (w[19], w[20], offset); - w[61] = hc_bytealign_S (w[18], w[19], offset); - w[60] = hc_bytealign_S (w[17], w[18], offset); - w[59] = hc_bytealign_S (w[16], w[17], offset); - w[58] = hc_bytealign_S (w[15], w[16], offset); - w[57] = hc_bytealign_S (w[14], w[15], offset); - w[56] = hc_bytealign_S (w[13], w[14], offset); - w[55] = hc_bytealign_S (w[12], w[13], offset); - w[54] = hc_bytealign_S (w[11], w[12], offset); - w[53] = hc_bytealign_S (w[10], w[11], offset); - w[52] = hc_bytealign_S (w[ 9], w[10], offset); - w[51] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[50] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[49] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[48] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[47] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[46] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[45] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[44] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[43] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[42] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[20], w[21], offset); + w[62] = hc_bytealign_be_S (w[19], w[20], offset); + w[61] = hc_bytealign_be_S (w[18], w[19], offset); + w[60] = hc_bytealign_be_S (w[17], w[18], offset); + w[59] = hc_bytealign_be_S (w[16], w[17], offset); + w[58] = hc_bytealign_be_S (w[15], w[16], offset); + w[57] = hc_bytealign_be_S (w[14], w[15], offset); + w[56] = hc_bytealign_be_S (w[13], w[14], offset); + w[55] = hc_bytealign_be_S (w[12], w[13], offset); + w[54] = hc_bytealign_be_S (w[11], w[12], offset); + w[53] = hc_bytealign_be_S (w[10], w[11], offset); + w[52] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[51] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[50] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[49] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[48] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[47] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[46] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[45] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[44] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[43] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[42] = hc_bytealign_be_S ( 0, w[ 0], offset); w[41] = 0; w[40] = 0; w[39] = 0; @@ -54230,27 +53910,27 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 43: - w[63] = hc_bytealign_S (w[19], w[20], offset); - w[62] = hc_bytealign_S (w[18], w[19], offset); - w[61] = hc_bytealign_S (w[17], w[18], offset); - w[60] = hc_bytealign_S (w[16], w[17], offset); - w[59] = hc_bytealign_S (w[15], w[16], offset); - w[58] = hc_bytealign_S (w[14], w[15], offset); - w[57] = hc_bytealign_S (w[13], w[14], offset); - w[56] = hc_bytealign_S (w[12], w[13], offset); - w[55] = hc_bytealign_S (w[11], w[12], offset); - w[54] = hc_bytealign_S (w[10], w[11], offset); - w[53] = hc_bytealign_S (w[ 9], w[10], offset); - w[52] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[51] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[50] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[49] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[48] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[47] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[46] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[45] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[44] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[43] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[19], w[20], offset); + w[62] = hc_bytealign_be_S (w[18], w[19], offset); + w[61] = hc_bytealign_be_S (w[17], w[18], offset); + w[60] = hc_bytealign_be_S (w[16], w[17], offset); + w[59] = hc_bytealign_be_S (w[15], w[16], offset); + w[58] = hc_bytealign_be_S (w[14], w[15], offset); + w[57] = hc_bytealign_be_S (w[13], w[14], offset); + w[56] = hc_bytealign_be_S (w[12], w[13], offset); + w[55] = hc_bytealign_be_S (w[11], w[12], offset); + w[54] = hc_bytealign_be_S (w[10], w[11], offset); + w[53] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[52] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[51] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[50] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[49] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[48] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[47] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[46] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[45] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[44] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[43] = hc_bytealign_be_S ( 0, w[ 0], offset); w[42] = 0; w[41] = 0; w[40] = 0; @@ -54298,26 +53978,26 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 44: - w[63] = hc_bytealign_S (w[18], w[19], offset); - w[62] = hc_bytealign_S (w[17], w[18], offset); - w[61] = hc_bytealign_S (w[16], w[17], offset); - w[60] = hc_bytealign_S (w[15], w[16], offset); - w[59] = hc_bytealign_S (w[14], w[15], offset); - w[58] = hc_bytealign_S (w[13], w[14], offset); - w[57] = hc_bytealign_S (w[12], w[13], offset); - w[56] = hc_bytealign_S (w[11], w[12], offset); - w[55] = hc_bytealign_S (w[10], w[11], offset); - w[54] = hc_bytealign_S (w[ 9], w[10], offset); - w[53] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[52] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[51] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[50] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[49] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[48] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[47] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[46] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[45] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[44] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[18], w[19], offset); + w[62] = hc_bytealign_be_S (w[17], w[18], offset); + w[61] = hc_bytealign_be_S (w[16], w[17], offset); + w[60] = hc_bytealign_be_S (w[15], w[16], offset); + w[59] = hc_bytealign_be_S (w[14], w[15], offset); + w[58] = hc_bytealign_be_S (w[13], w[14], offset); + w[57] = hc_bytealign_be_S (w[12], w[13], offset); + w[56] = hc_bytealign_be_S (w[11], w[12], offset); + w[55] = hc_bytealign_be_S (w[10], w[11], offset); + w[54] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[53] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[52] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[51] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[50] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[49] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[48] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[47] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[46] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[45] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[44] = hc_bytealign_be_S ( 0, w[ 0], offset); w[43] = 0; w[42] = 0; w[41] = 0; @@ -54366,25 +54046,25 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 45: - w[63] = hc_bytealign_S (w[17], w[18], offset); - w[62] = hc_bytealign_S (w[16], w[17], offset); - w[61] = hc_bytealign_S (w[15], w[16], offset); - w[60] = hc_bytealign_S (w[14], w[15], offset); - w[59] = hc_bytealign_S (w[13], w[14], offset); - w[58] = hc_bytealign_S (w[12], w[13], offset); - w[57] = hc_bytealign_S (w[11], w[12], offset); - w[56] = hc_bytealign_S (w[10], w[11], offset); - w[55] = hc_bytealign_S (w[ 9], w[10], offset); - w[54] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[53] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[52] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[51] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[50] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[49] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[48] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[47] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[46] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[45] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[17], w[18], offset); + w[62] = hc_bytealign_be_S (w[16], w[17], offset); + w[61] = hc_bytealign_be_S (w[15], w[16], offset); + w[60] = hc_bytealign_be_S (w[14], w[15], offset); + w[59] = hc_bytealign_be_S (w[13], w[14], offset); + w[58] = hc_bytealign_be_S (w[12], w[13], offset); + w[57] = hc_bytealign_be_S (w[11], w[12], offset); + w[56] = hc_bytealign_be_S (w[10], w[11], offset); + w[55] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[54] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[53] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[52] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[51] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[50] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[49] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[48] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[47] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[46] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[45] = hc_bytealign_be_S ( 0, w[ 0], offset); w[44] = 0; w[43] = 0; w[42] = 0; @@ -54434,24 +54114,24 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 46: - w[63] = hc_bytealign_S (w[16], w[17], offset); - w[62] = hc_bytealign_S (w[15], w[16], offset); - w[61] = hc_bytealign_S (w[14], w[15], offset); - w[60] = hc_bytealign_S (w[13], w[14], offset); - w[59] = hc_bytealign_S (w[12], w[13], offset); - w[58] = hc_bytealign_S (w[11], w[12], offset); - w[57] = hc_bytealign_S (w[10], w[11], offset); - w[56] = hc_bytealign_S (w[ 9], w[10], offset); - w[55] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[54] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[53] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[52] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[51] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[50] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[49] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[48] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[47] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[46] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[16], w[17], offset); + w[62] = hc_bytealign_be_S (w[15], w[16], offset); + w[61] = hc_bytealign_be_S (w[14], w[15], offset); + w[60] = hc_bytealign_be_S (w[13], w[14], offset); + w[59] = hc_bytealign_be_S (w[12], w[13], offset); + w[58] = hc_bytealign_be_S (w[11], w[12], offset); + w[57] = hc_bytealign_be_S (w[10], w[11], offset); + w[56] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[55] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[54] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[53] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[52] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[51] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[50] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[49] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[48] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[47] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[46] = hc_bytealign_be_S ( 0, w[ 0], offset); w[45] = 0; w[44] = 0; w[43] = 0; @@ -54502,23 +54182,23 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 47: - w[63] = hc_bytealign_S (w[15], w[16], offset); - w[62] = hc_bytealign_S (w[14], w[15], offset); - w[61] = hc_bytealign_S (w[13], w[14], offset); - w[60] = hc_bytealign_S (w[12], w[13], offset); - w[59] = hc_bytealign_S (w[11], w[12], offset); - w[58] = hc_bytealign_S (w[10], w[11], offset); - w[57] = hc_bytealign_S (w[ 9], w[10], offset); - w[56] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[55] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[54] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[53] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[52] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[51] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[50] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[49] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[48] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[47] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[15], w[16], offset); + w[62] = hc_bytealign_be_S (w[14], w[15], offset); + w[61] = hc_bytealign_be_S (w[13], w[14], offset); + w[60] = hc_bytealign_be_S (w[12], w[13], offset); + w[59] = hc_bytealign_be_S (w[11], w[12], offset); + w[58] = hc_bytealign_be_S (w[10], w[11], offset); + w[57] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[56] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[55] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[54] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[53] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[52] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[51] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[50] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[49] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[48] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[47] = hc_bytealign_be_S ( 0, w[ 0], offset); w[46] = 0; w[45] = 0; w[44] = 0; @@ -54570,22 +54250,22 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 48: - w[63] = hc_bytealign_S (w[14], w[15], offset); - w[62] = hc_bytealign_S (w[13], w[14], offset); - w[61] = hc_bytealign_S (w[12], w[13], offset); - w[60] = hc_bytealign_S (w[11], w[12], offset); - w[59] = hc_bytealign_S (w[10], w[11], offset); - w[58] = hc_bytealign_S (w[ 9], w[10], offset); - w[57] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[56] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[55] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[54] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[53] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[52] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[51] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[50] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[49] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[48] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[14], w[15], offset); + w[62] = hc_bytealign_be_S (w[13], w[14], offset); + w[61] = hc_bytealign_be_S (w[12], w[13], offset); + w[60] = hc_bytealign_be_S (w[11], w[12], offset); + w[59] = hc_bytealign_be_S (w[10], w[11], offset); + w[58] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[57] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[56] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[55] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[54] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[53] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[52] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[51] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[50] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[49] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[48] = hc_bytealign_be_S ( 0, w[ 0], offset); w[47] = 0; w[46] = 0; w[45] = 0; @@ -54638,21 +54318,21 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 49: - w[63] = hc_bytealign_S (w[13], w[14], offset); - w[62] = hc_bytealign_S (w[12], w[13], offset); - w[61] = hc_bytealign_S (w[11], w[12], offset); - w[60] = hc_bytealign_S (w[10], w[11], offset); - w[59] = hc_bytealign_S (w[ 9], w[10], offset); - w[58] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[57] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[56] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[55] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[54] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[53] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[52] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[51] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[50] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[49] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[13], w[14], offset); + w[62] = hc_bytealign_be_S (w[12], w[13], offset); + w[61] = hc_bytealign_be_S (w[11], w[12], offset); + w[60] = hc_bytealign_be_S (w[10], w[11], offset); + w[59] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[58] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[57] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[56] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[55] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[54] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[53] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[52] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[51] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[50] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[49] = hc_bytealign_be_S ( 0, w[ 0], offset); w[48] = 0; w[47] = 0; w[46] = 0; @@ -54706,20 +54386,20 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 50: - w[63] = hc_bytealign_S (w[12], w[13], offset); - w[62] = hc_bytealign_S (w[11], w[12], offset); - w[61] = hc_bytealign_S (w[10], w[11], offset); - w[60] = hc_bytealign_S (w[ 9], w[10], offset); - w[59] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[58] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[57] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[56] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[55] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[54] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[53] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[52] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[51] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[50] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[12], w[13], offset); + w[62] = hc_bytealign_be_S (w[11], w[12], offset); + w[61] = hc_bytealign_be_S (w[10], w[11], offset); + w[60] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[59] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[58] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[57] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[56] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[55] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[54] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[53] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[52] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[51] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[50] = hc_bytealign_be_S ( 0, w[ 0], offset); w[49] = 0; w[48] = 0; w[47] = 0; @@ -54774,19 +54454,19 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 51: - w[63] = hc_bytealign_S (w[11], w[12], offset); - w[62] = hc_bytealign_S (w[10], w[11], offset); - w[61] = hc_bytealign_S (w[ 9], w[10], offset); - w[60] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[59] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[58] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[57] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[56] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[55] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[54] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[53] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[52] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[51] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[11], w[12], offset); + w[62] = hc_bytealign_be_S (w[10], w[11], offset); + w[61] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[60] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[59] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[58] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[57] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[56] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[55] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[54] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[53] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[52] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[51] = hc_bytealign_be_S ( 0, w[ 0], offset); w[50] = 0; w[49] = 0; w[48] = 0; @@ -54842,18 +54522,18 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 52: - w[63] = hc_bytealign_S (w[10], w[11], offset); - w[62] = hc_bytealign_S (w[ 9], w[10], offset); - w[61] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[60] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[59] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[58] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[57] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[56] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[55] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[54] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[53] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[52] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[10], w[11], offset); + w[62] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[61] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[60] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[59] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[58] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[57] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[56] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[55] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[54] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[53] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[52] = hc_bytealign_be_S ( 0, w[ 0], offset); w[51] = 0; w[50] = 0; w[49] = 0; @@ -54910,17 +54590,17 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 53: - w[63] = hc_bytealign_S (w[ 9], w[10], offset); - w[62] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[61] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[60] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[59] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[58] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[57] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[56] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[55] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[54] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[53] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 9], w[10], offset); + w[62] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[61] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[60] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[59] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[58] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[57] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[56] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[55] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[54] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[53] = hc_bytealign_be_S ( 0, w[ 0], offset); w[52] = 0; w[51] = 0; w[50] = 0; @@ -54978,16 +54658,16 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 54: - w[63] = hc_bytealign_S (w[ 8], w[ 9], offset); - w[62] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[61] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[60] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[59] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[58] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[57] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[56] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[55] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[54] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 8], w[ 9], offset); + w[62] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[61] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[60] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[59] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[58] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[57] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[56] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[55] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[54] = hc_bytealign_be_S ( 0, w[ 0], offset); w[53] = 0; w[52] = 0; w[51] = 0; @@ -55046,15 +54726,15 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 55: - w[63] = hc_bytealign_S (w[ 7], w[ 8], offset); - w[62] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[61] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[60] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[59] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[58] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[57] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[56] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[55] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 7], w[ 8], offset); + w[62] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[61] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[60] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[59] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[58] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[57] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[56] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[55] = hc_bytealign_be_S ( 0, w[ 0], offset); w[54] = 0; w[53] = 0; w[52] = 0; @@ -55114,14 +54794,14 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 56: - w[63] = hc_bytealign_S (w[ 6], w[ 7], offset); - w[62] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[61] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[60] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[59] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[58] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[57] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[56] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 6], w[ 7], offset); + w[62] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[61] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[60] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[59] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[58] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[57] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[56] = hc_bytealign_be_S ( 0, w[ 0], offset); w[55] = 0; w[54] = 0; w[53] = 0; @@ -55182,13 +54862,13 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 57: - w[63] = hc_bytealign_S (w[ 5], w[ 6], offset); - w[62] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[61] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[60] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[59] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[58] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[57] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 5], w[ 6], offset); + w[62] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[61] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[60] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[59] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[58] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[57] = hc_bytealign_be_S ( 0, w[ 0], offset); w[56] = 0; w[55] = 0; w[54] = 0; @@ -55250,12 +54930,12 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 58: - w[63] = hc_bytealign_S (w[ 4], w[ 5], offset); - w[62] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[61] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[60] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[59] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[58] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 4], w[ 5], offset); + w[62] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[61] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[60] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[59] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[58] = hc_bytealign_be_S ( 0, w[ 0], offset); w[57] = 0; w[56] = 0; w[55] = 0; @@ -55318,11 +54998,11 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 59: - w[63] = hc_bytealign_S (w[ 3], w[ 4], offset); - w[62] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[61] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[60] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[59] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 3], w[ 4], offset); + w[62] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[61] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[60] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[59] = hc_bytealign_be_S ( 0, w[ 0], offset); w[58] = 0; w[57] = 0; w[56] = 0; @@ -55386,10 +55066,10 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 60: - w[63] = hc_bytealign_S (w[ 2], w[ 3], offset); - w[62] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[61] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[60] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 2], w[ 3], offset); + w[62] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[61] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[60] = hc_bytealign_be_S ( 0, w[ 0], offset); w[59] = 0; w[58] = 0; w[57] = 0; @@ -55454,9 +55134,9 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 61: - w[63] = hc_bytealign_S (w[ 1], w[ 2], offset); - w[62] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[61] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 1], w[ 2], offset); + w[62] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[61] = hc_bytealign_be_S ( 0, w[ 0], offset); w[60] = 0; w[59] = 0; w[58] = 0; @@ -55522,8 +55202,8 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 62: - w[63] = hc_bytealign_S (w[ 0], w[ 1], offset); - w[62] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S (w[ 0], w[ 1], offset); + w[62] = hc_bytealign_be_S ( 0, w[ 0], offset); w[61] = 0; w[60] = 0; w[59] = 0; @@ -55590,7 +55270,7 @@ DECLSPEC void switch_buffer_by_offset_1x64_be_S (u32 *w, const u32 offset) break; case 63: - w[63] = hc_bytealign_S ( 0, w[ 0], offset); + w[63] = hc_bytealign_be_S ( 0, w[ 0], offset); w[62] = 0; w[61] = 0; w[60] = 0; diff --git a/OpenCL/inc_types.cl b/OpenCL/inc_types.cl index f8bf43e62..32ca1d109 100644 --- a/OpenCL/inc_types.cl +++ b/OpenCL/inc_types.cl @@ -361,16 +361,46 @@ DECLSPEC u32 hc_bfe_S (const u32 a, const u32 b, const u32 c) return amd_bfe (a, b, c); } -DECLSPEC u32x hc_bytealign (const u32x a, const u32x b, const u32x c) +DECLSPEC u32x hc_bytealign_be (const u32x a, const u32x b, const u32 c) { return amd_bytealign (a, b, c); } -DECLSPEC u32 hc_bytealign_S (const u32 a, const u32 b, const u32 c) +DECLSPEC u32 hc_bytealign_be_S (const u32 a, const u32 b, const u32 c) { return amd_bytealign (a, b, c); } +DECLSPEC u32x hc_bytealign (const u32x a, const u32x b, const u32 c) +{ + u32x r; + + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a >> 24) | (b << 8); break; + case 2: r = (a >> 16) | (b << 16); break; + case 3: r = (a >> 8) | (b << 24); break; + } + + return r; +} + +DECLSPEC u32 hc_bytealign_S (const u32 a, const u32 b, const u32 c) +{ + u32 r; + + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a >> 24) | (b << 8); break; + case 2: r = (a >> 16) | (b << 16); break; + case 3: r = (a >> 8) | (b << 24); break; + } + + return r; +} + #if HAS_VPERM DECLSPEC u32x hc_byte_perm (const u32x a, const u32x b, const u32x c) { @@ -1089,44 +1119,64 @@ DECLSPEC u32 hc_bfe_S (const u32 a, const u32 b, const u32 c) #undef BFE } +DECLSPEC u32x hc_bytealign_be (const u32x a, const u32x b, const u32 c) +{ + u32x r; + + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a << 24) | (b >> 8); break; + case 2: r = (a << 16) | (b >> 16); break; + case 3: r = (a << 8) | (b >> 24); break; + } + + return r; +} + +DECLSPEC u32 hc_bytealign_be_S (const u32 a, const u32 b, const u32 c) +{ + u32 r; + + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a << 24) | (b >> 8); break; + case 2: r = (a << 16) | (b >> 16); break; + case 3: r = (a << 8) | (b >> 24); break; + } + + return r; +} + DECLSPEC u32x hc_bytealign (const u32x a, const u32x b, const u32 c) { - #if VECT_SIZE == 1 - const u64x tmp = ((((u64x) (a)) << 32) | ((u64x) (b))) >> ((c & 3) * 8); + u32x r; - return (u32x) (tmp); - #endif + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a >> 24) | (b << 8); break; + case 2: r = (a >> 16) | (b << 16); break; + case 3: r = (a >> 8) | (b << 24); break; + } - #if VECT_SIZE == 2 - const u64x tmp = ((((u64x) (a.s0, a.s1)) << 32) | ((u64x) (b.s0, b.s1))) >> ((c & 3) * 8); - - return (u32x) (tmp.s0, tmp.s1); - #endif - - #if VECT_SIZE == 4 - const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3))) >> ((c & 3) * 8); - - return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3); - #endif - - #if VECT_SIZE == 8 - const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3, a.s4, a.s5, a.s6, a.s7)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3, b.s4, b.s5, b.s6, b.s7))) >> ((c & 3) * 8); - - return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7); - #endif - - #if VECT_SIZE == 16 - const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3, a.s4, a.s5, a.s6, a.s7, a.s8, a.s9, a.sa, a.sb, a.sc, a.sd, a.se, a.sf)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3, b.s4, b.s5, b.s6, b.s7, b.s8, b.s9, b.sa, b.sb, b.sc, b.sd, b.se, b.sf))) >> ((c & 3) * 8); - - return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7, tmp.s8, tmp.s9, tmp.sa, tmp.sb, tmp.sc, tmp.sd, tmp.se, tmp.sf); - #endif + return r; } DECLSPEC u32 hc_bytealign_S (const u32 a, const u32 b, const u32 c) { - const u64 tmp = ((((u64) a) << 32) | ((u64) b)) >> ((c & 3) * 8); + u32 r; - return (u32) (tmp); + switch (c & 3) + { + case 0: r = b; break; + case 1: r = (a >> 24) | (b << 8); break; + case 2: r = (a >> 16) | (b << 16); break; + case 3: r = (a >> 8) | (b << 24); break; + } + + return r; } DECLSPEC u32x hc_add3 (const u32x a, const u32x b, const u32x c) diff --git a/OpenCL/m00500-optimized.cl b/OpenCL/m00500-optimized.cl index d28a3994c..e2a39b401 100644 --- a/OpenCL/m00500-optimized.cl +++ b/OpenCL/m00500-optimized.cl @@ -30,22 +30,16 @@ DECLSPEC void memcat16 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, cons const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -143,23 +137,17 @@ DECLSPEC void memcat16_x80 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); - u32 in4 = 0x80000000; + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; + u32 in4 = 0x80; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, in4, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -256,16 +244,12 @@ DECLSPEC void memcat8 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); + u32 in0 = append[0]; + u32 in1 = append[1]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); #endif #ifdef IS_NV diff --git a/OpenCL/m01600-optimized.cl b/OpenCL/m01600-optimized.cl index e0927a056..e5a09c5b3 100644 --- a/OpenCL/m01600-optimized.cl +++ b/OpenCL/m01600-optimized.cl @@ -29,22 +29,16 @@ DECLSPEC void memcat16 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, cons const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -142,23 +136,17 @@ DECLSPEC void memcat16_x80 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); - u32 in4 = 0x80000000; + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; + u32 in4 = 0x80; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, in4, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -255,16 +243,12 @@ DECLSPEC void memcat8 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); + u32 in0 = append[0]; + u32 in1 = append[1]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); #endif #ifdef IS_NV diff --git a/OpenCL/m05800-optimized.cl b/OpenCL/m05800-optimized.cl index 7fd955917..cdfb8295d 100644 --- a/OpenCL/m05800-optimized.cl +++ b/OpenCL/m05800-optimized.cl @@ -2117,11 +2117,11 @@ DECLSPEC void append_salt (u32 *w0, u32 *w1, u32 *w2, const u32 *append, const u const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); - u32 in4 = swap32_S (append[4]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; + u32 in4 = append[4]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); @@ -2129,13 +2129,6 @@ DECLSPEC void append_salt (u32 *w0, u32 *w1, u32 *w2, const u32 *append, const u tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, in4, offset); tmp5 = hc_bytealign (in4, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); - tmp5 = swap32_S (tmp5); #endif #ifdef IS_NV diff --git a/OpenCL/m06300-optimized.cl b/OpenCL/m06300-optimized.cl index 8972d64b0..1f5aceedc 100644 --- a/OpenCL/m06300-optimized.cl +++ b/OpenCL/m06300-optimized.cl @@ -26,22 +26,16 @@ DECLSPEC void memcat16 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, cons const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -139,23 +133,17 @@ DECLSPEC void memcat16_x80 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); - u32 in4 = 0x80000000; + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; + u32 in4 = 0x80; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, in4, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -252,16 +240,12 @@ DECLSPEC void memcat8 (u32 *block0, u32 *block1, u32 *block2, u32 *block3, const const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); + u32 in0 = append[0]; + u32 in1 = append[1]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); #endif #ifdef IS_NV diff --git a/OpenCL/m07400-optimized.cl b/OpenCL/m07400-optimized.cl index 9bb671c84..0ee8ad118 100644 --- a/OpenCL/m07400-optimized.cl +++ b/OpenCL/m07400-optimized.cl @@ -97,22 +97,16 @@ DECLSPEC u32 memcat16 (u32 *block, const u32 offset, const u32 *append, const u3 const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -238,22 +232,16 @@ DECLSPEC u32 memcat16c (u32 *block, const u32 offset, const u32 *append, const u const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -405,22 +393,16 @@ DECLSPEC u32 memcat20 (u32 *block, const u32 offset, const u32 *append, const u3 const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, 0, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV @@ -554,23 +536,17 @@ DECLSPEC u32 memcat20_x80 (u32 *block, const u32 offset, const u32 *append, cons const int offset_minus_4 = 4 - offset_mod_4; #if defined IS_AMD || defined IS_GENERIC - u32 in0 = swap32_S (append[0]); - u32 in1 = swap32_S (append[1]); - u32 in2 = swap32_S (append[2]); - u32 in3 = swap32_S (append[3]); - u32 in4 = 0x80000000; + u32 in0 = append[0]; + u32 in1 = append[1]; + u32 in2 = append[2]; + u32 in3 = append[3]; + u32 in4 = 0x80; tmp0 = hc_bytealign ( 0, in0, offset); tmp1 = hc_bytealign (in0, in1, offset); tmp2 = hc_bytealign (in1, in2, offset); tmp3 = hc_bytealign (in2, in3, offset); tmp4 = hc_bytealign (in3, in4, offset); - - tmp0 = swap32_S (tmp0); - tmp1 = swap32_S (tmp1); - tmp2 = swap32_S (tmp2); - tmp3 = swap32_S (tmp3); - tmp4 = swap32_S (tmp4); #endif #ifdef IS_NV diff --git a/OpenCL/m10700-optimized.cl b/OpenCL/m10700-optimized.cl index 491fd0dee..813979c58 100644 --- a/OpenCL/m10700-optimized.cl +++ b/OpenCL/m10700-optimized.cl @@ -200,11 +200,11 @@ DECLSPEC void make_sc (u32 *sc, const u32 *pw, const u32 pw_len, const u32 *bl, #if defined IS_AMD || defined IS_GENERIC for (i = 0; i < pd; i++) sc[idx++] = pw[i]; sc[idx++] = pw[i] - | hc_bytealign (bl[0], 0, pm4); - for (i = 1; i < bd; i++) sc[idx++] = hc_bytealign (bl[i], bl[i - 1], pm4); - sc[idx++] = hc_bytealign (sc[0], bl[i - 1], pm4); - for (i = 1; i < 4; i++) sc[idx++] = hc_bytealign (sc[i], sc[i - 1], pm4); - sc[idx++] = hc_bytealign ( 0, sc[i - 1], pm4); + | hc_bytealign_be (bl[0], 0, pm4); + for (i = 1; i < bd; i++) sc[idx++] = hc_bytealign_be (bl[i], bl[i - 1], pm4); + sc[idx++] = hc_bytealign_be (sc[0], bl[i - 1], pm4); + for (i = 1; i < 4; i++) sc[idx++] = hc_bytealign_be (sc[i], sc[i - 1], pm4); + sc[idx++] = hc_bytealign_be ( 0, sc[i - 1], pm4); #endif #ifdef IS_NV @@ -229,10 +229,10 @@ DECLSPEC void make_pt_with_offset (u32 *pt, const u32 offset, const u32 *sc, con const u32 od = m / 4; #if defined IS_AMD || defined IS_GENERIC - pt[0] = hc_bytealign (sc[od + 1], sc[od + 0], om); - pt[1] = hc_bytealign (sc[od + 2], sc[od + 1], om); - pt[2] = hc_bytealign (sc[od + 3], sc[od + 2], om); - pt[3] = hc_bytealign (sc[od + 4], sc[od + 3], om); + pt[0] = hc_bytealign_be (sc[od + 1], sc[od + 0], om); + pt[1] = hc_bytealign_be (sc[od + 2], sc[od + 1], om); + pt[2] = hc_bytealign_be (sc[od + 3], sc[od + 2], om); + pt[3] = hc_bytealign_be (sc[od + 4], sc[od + 3], om); #endif #ifdef IS_NV diff --git a/OpenCL/m11600-pure.cl b/OpenCL/m11600-pure.cl index c22df6156..2f14a40c6 100644 --- a/OpenCL/m11600-pure.cl +++ b/OpenCL/m11600-pure.cl @@ -20,6 +20,11 @@ DECLSPEC void memcat8c_be (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 len, co u32 tmp0; u32 tmp1; + #if defined IS_AMD || defined IS_GENERIC + tmp0 = hc_bytealign_be (0, append, func_len); + tmp1 = hc_bytealign_be (append, 0, func_len); + #endif + #ifdef IS_NV const int selector = (0x76543210 >> ((func_len & 3) * 4)) & 0xffff; @@ -27,11 +32,6 @@ DECLSPEC void memcat8c_be (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 len, co tmp1 = hc_byte_perm (0, append, selector); #endif - #if defined IS_AMD || defined IS_GENERIC - tmp0 = hc_bytealign (0, append, func_len); - tmp1 = hc_bytealign (append, 0, func_len); - #endif - u32 carry = 0; switch (div) diff --git a/OpenCL/m13800_a0-optimized.cl b/OpenCL/m13800_a0-optimized.cl index ce2aae680..7d88d9f7f 100644 --- a/OpenCL/m13800_a0-optimized.cl +++ b/OpenCL/m13800_a0-optimized.cl @@ -45,23 +45,23 @@ DECLSPEC void memcat64c_be (u32x *block, const u32 offset, u32x *carry) u32x tmp16; #if defined IS_AMD || defined IS_GENERIC - tmp00 = hc_bytealign ( 0, carry[ 0], offset); - tmp01 = hc_bytealign (carry[ 0], carry[ 1], offset); - tmp02 = hc_bytealign (carry[ 1], carry[ 2], offset); - tmp03 = hc_bytealign (carry[ 2], carry[ 3], offset); - tmp04 = hc_bytealign (carry[ 3], carry[ 4], offset); - tmp05 = hc_bytealign (carry[ 4], carry[ 5], offset); - tmp06 = hc_bytealign (carry[ 5], carry[ 6], offset); - tmp07 = hc_bytealign (carry[ 6], carry[ 7], offset); - tmp08 = hc_bytealign (carry[ 7], carry[ 8], offset); - tmp09 = hc_bytealign (carry[ 8], carry[ 9], offset); - tmp10 = hc_bytealign (carry[ 9], carry[10], offset); - tmp11 = hc_bytealign (carry[10], carry[11], offset); - tmp12 = hc_bytealign (carry[11], carry[12], offset); - tmp13 = hc_bytealign (carry[12], carry[13], offset); - tmp14 = hc_bytealign (carry[13], carry[14], offset); - tmp15 = hc_bytealign (carry[14], carry[15], offset); - tmp16 = hc_bytealign (carry[15], 0, offset); + tmp00 = hc_bytealign_be ( 0, carry[ 0], offset); + tmp01 = hc_bytealign_be (carry[ 0], carry[ 1], offset); + tmp02 = hc_bytealign_be (carry[ 1], carry[ 2], offset); + tmp03 = hc_bytealign_be (carry[ 2], carry[ 3], offset); + tmp04 = hc_bytealign_be (carry[ 3], carry[ 4], offset); + tmp05 = hc_bytealign_be (carry[ 4], carry[ 5], offset); + tmp06 = hc_bytealign_be (carry[ 5], carry[ 6], offset); + tmp07 = hc_bytealign_be (carry[ 6], carry[ 7], offset); + tmp08 = hc_bytealign_be (carry[ 7], carry[ 8], offset); + tmp09 = hc_bytealign_be (carry[ 8], carry[ 9], offset); + tmp10 = hc_bytealign_be (carry[ 9], carry[10], offset); + tmp11 = hc_bytealign_be (carry[10], carry[11], offset); + tmp12 = hc_bytealign_be (carry[11], carry[12], offset); + tmp13 = hc_bytealign_be (carry[12], carry[13], offset); + tmp14 = hc_bytealign_be (carry[13], carry[14], offset); + tmp15 = hc_bytealign_be (carry[14], carry[15], offset); + tmp16 = hc_bytealign_be (carry[15], 0, offset); #endif #ifdef IS_NV diff --git a/OpenCL/m13800_a1-optimized.cl b/OpenCL/m13800_a1-optimized.cl index 99a15bf70..06367a01f 100644 --- a/OpenCL/m13800_a1-optimized.cl +++ b/OpenCL/m13800_a1-optimized.cl @@ -43,23 +43,23 @@ DECLSPEC void memcat64c_be (u32x *block, const u32 offset, u32x *carry) u32x tmp16; #if defined IS_AMD || defined IS_GENERIC - tmp00 = hc_bytealign ( 0, carry[ 0], offset); - tmp01 = hc_bytealign (carry[ 0], carry[ 1], offset); - tmp02 = hc_bytealign (carry[ 1], carry[ 2], offset); - tmp03 = hc_bytealign (carry[ 2], carry[ 3], offset); - tmp04 = hc_bytealign (carry[ 3], carry[ 4], offset); - tmp05 = hc_bytealign (carry[ 4], carry[ 5], offset); - tmp06 = hc_bytealign (carry[ 5], carry[ 6], offset); - tmp07 = hc_bytealign (carry[ 6], carry[ 7], offset); - tmp08 = hc_bytealign (carry[ 7], carry[ 8], offset); - tmp09 = hc_bytealign (carry[ 8], carry[ 9], offset); - tmp10 = hc_bytealign (carry[ 9], carry[10], offset); - tmp11 = hc_bytealign (carry[10], carry[11], offset); - tmp12 = hc_bytealign (carry[11], carry[12], offset); - tmp13 = hc_bytealign (carry[12], carry[13], offset); - tmp14 = hc_bytealign (carry[13], carry[14], offset); - tmp15 = hc_bytealign (carry[14], carry[15], offset); - tmp16 = hc_bytealign (carry[15], 0, offset); + tmp00 = hc_bytealign_be ( 0, carry[ 0], offset); + tmp01 = hc_bytealign_be (carry[ 0], carry[ 1], offset); + tmp02 = hc_bytealign_be (carry[ 1], carry[ 2], offset); + tmp03 = hc_bytealign_be (carry[ 2], carry[ 3], offset); + tmp04 = hc_bytealign_be (carry[ 3], carry[ 4], offset); + tmp05 = hc_bytealign_be (carry[ 4], carry[ 5], offset); + tmp06 = hc_bytealign_be (carry[ 5], carry[ 6], offset); + tmp07 = hc_bytealign_be (carry[ 6], carry[ 7], offset); + tmp08 = hc_bytealign_be (carry[ 7], carry[ 8], offset); + tmp09 = hc_bytealign_be (carry[ 8], carry[ 9], offset); + tmp10 = hc_bytealign_be (carry[ 9], carry[10], offset); + tmp11 = hc_bytealign_be (carry[10], carry[11], offset); + tmp12 = hc_bytealign_be (carry[11], carry[12], offset); + tmp13 = hc_bytealign_be (carry[12], carry[13], offset); + tmp14 = hc_bytealign_be (carry[13], carry[14], offset); + tmp15 = hc_bytealign_be (carry[14], carry[15], offset); + tmp16 = hc_bytealign_be (carry[15], 0, offset); #endif #ifdef IS_NV diff --git a/OpenCL/m13800_a3-optimized.cl b/OpenCL/m13800_a3-optimized.cl index 74b17e559..341e97f57 100644 --- a/OpenCL/m13800_a3-optimized.cl +++ b/OpenCL/m13800_a3-optimized.cl @@ -42,23 +42,23 @@ DECLSPEC void memcat64c_be (u32x *block, const u32 offset, u32x *carry) u32x tmp16; #if defined IS_AMD || defined IS_GENERIC - tmp00 = hc_bytealign ( 0, carry[ 0], offset); - tmp01 = hc_bytealign (carry[ 0], carry[ 1], offset); - tmp02 = hc_bytealign (carry[ 1], carry[ 2], offset); - tmp03 = hc_bytealign (carry[ 2], carry[ 3], offset); - tmp04 = hc_bytealign (carry[ 3], carry[ 4], offset); - tmp05 = hc_bytealign (carry[ 4], carry[ 5], offset); - tmp06 = hc_bytealign (carry[ 5], carry[ 6], offset); - tmp07 = hc_bytealign (carry[ 6], carry[ 7], offset); - tmp08 = hc_bytealign (carry[ 7], carry[ 8], offset); - tmp09 = hc_bytealign (carry[ 8], carry[ 9], offset); - tmp10 = hc_bytealign (carry[ 9], carry[10], offset); - tmp11 = hc_bytealign (carry[10], carry[11], offset); - tmp12 = hc_bytealign (carry[11], carry[12], offset); - tmp13 = hc_bytealign (carry[12], carry[13], offset); - tmp14 = hc_bytealign (carry[13], carry[14], offset); - tmp15 = hc_bytealign (carry[14], carry[15], offset); - tmp16 = hc_bytealign (carry[15], 0, offset); + tmp00 = hc_bytealign_be ( 0, carry[ 0], offset); + tmp01 = hc_bytealign_be (carry[ 0], carry[ 1], offset); + tmp02 = hc_bytealign_be (carry[ 1], carry[ 2], offset); + tmp03 = hc_bytealign_be (carry[ 2], carry[ 3], offset); + tmp04 = hc_bytealign_be (carry[ 3], carry[ 4], offset); + tmp05 = hc_bytealign_be (carry[ 4], carry[ 5], offset); + tmp06 = hc_bytealign_be (carry[ 5], carry[ 6], offset); + tmp07 = hc_bytealign_be (carry[ 6], carry[ 7], offset); + tmp08 = hc_bytealign_be (carry[ 7], carry[ 8], offset); + tmp09 = hc_bytealign_be (carry[ 8], carry[ 9], offset); + tmp10 = hc_bytealign_be (carry[ 9], carry[10], offset); + tmp11 = hc_bytealign_be (carry[10], carry[11], offset); + tmp12 = hc_bytealign_be (carry[11], carry[12], offset); + tmp13 = hc_bytealign_be (carry[12], carry[13], offset); + tmp14 = hc_bytealign_be (carry[13], carry[14], offset); + tmp15 = hc_bytealign_be (carry[14], carry[15], offset); + tmp16 = hc_bytealign_be (carry[15], 0, offset); #endif #ifdef IS_NV