From ec874c1d594daab45eecd9a3aa6b8ab9a8fdf060 Mon Sep 17 00:00:00 2001 From: jsteube Date: Sun, 13 Aug 2017 16:43:46 +0200 Subject: [PATCH] Optimized the following pure kernel rule engine functions: - mangle_lrest() - mangle_lrest_ufirst() - mangle_urest() - mangle_urest_lfirst() - mangle_trest() - mangle_toggle_at() - mangle_reverse() - mangle_dupeword() - mangle_reflect() - mangle_rotate_left() - mangle_rotate_right() - mangle_switch_first() - mangle_switch_last() - mangle_switch_at() - mangle_title_sep() - mangle_title_sep() Added some helper functions: - generate_cmask() - append_four_byte() - append_three_byte() - append_two_byte() - append_one_byte() - append_block() - exchange_byte() Removed some helper functions: - upper_at() - lower_at() - toggle_at() - mangle_switch() NOTE: Changes need to be backported to CPU when finished --- OpenCL/inc_rp.cl | 449 ++++++++++++++++++++++++++++++++--------------- 1 file changed, 311 insertions(+), 138 deletions(-) diff --git a/OpenCL/inc_rp.cl b/OpenCL/inc_rp.cl index be8c6e8cc..8e0a69d56 100644 --- a/OpenCL/inc_rp.cl +++ b/OpenCL/inc_rp.cl @@ -5,104 +5,275 @@ #define MAYBE_UNUSED -static void upper_at (u8 *buf, const int pos) +static u32 generate_cmask (const u32 value) { - const u8 c = buf[pos]; + const u32 rmask = ((value & 0x40404040u) >> 1u) + & ~((value & 0x80808080u) >> 2u); - if ((c >= 'a') && (c <= 'z')) buf[pos] ^= 0x20; + const u32 hmask = (value & 0x1f1f1f1fu) + 0x05050505u; + const u32 lmask = (value & 0x1f1f1f1fu) + 0x1f1f1f1fu; + + return rmask & ~hmask & lmask; } -static void lower_at (u8 *buf, const int pos) +static void append_four_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst) { - const u8 c = buf[pos]; + const int sd = off_src / 4; + const int sm = off_src & 3; + const int sm8 = sm * 8; - if ((c >= 'A') && (c <= 'Z')) buf[pos] ^= 0x20; + const int dd = off_dst / 4; + const int dm = off_dst & 3; + const int dm8 = dm * 8; + + u64 t64 = hl32_to_64 (buf_src[sd + 1], buf_src[sd + 0]); + + t64 >>= sm8; + t64 <<= dm8; + + const u32 t0 = l32_from_64 (t64); + const u32 t1 = h32_from_64 (t64); + + buf_dst[dd + 0] |= t0; + buf_dst[dd + 1] |= t1; } -static void toggle_at (u8 *buf, const int pos) +static void append_three_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst) { - const u8 c = buf[pos]; + const int sd = off_src / 4; + const int sm = off_src & 3; + const int sm8 = sm * 8; - if ((c >= 'a') && (c <= 'z')) buf[pos] ^= 0x20; - if ((c >= 'A') && (c <= 'Z')) buf[pos] ^= 0x20; + const int dd = off_dst / 4; + const int dm = off_dst & 3; + const int dm8 = dm * 8; + + u64 t64 = hl32_to_64 (buf_src[sd + 1], buf_src[sd + 0]); + + t64 >>= sm8; + t64 &= 0x00ffffff; + t64 <<= dm8; + + const u32 t0 = l32_from_64 (t64); + const u32 t1 = h32_from_64 (t64); + + buf_dst[dd + 0] |= t0; + buf_dst[dd + 1] |= t1; } -static void mangle_switch (u8 *buf, const int l, const int r) +static void append_two_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst) { - const u8 c = buf[r]; - buf[r] = buf[l]; - buf[l] = c; + const int sd = off_src / 4; + const int sm = off_src & 3; + const int sm8 = sm * 8; + + const int dd = off_dst / 4; + const int dm = off_dst & 3; + const int dm8 = dm * 8; + + u64 t64 = hl32_to_64 (buf_src[sd + 1], buf_src[sd + 0]); + + t64 >>= sm8; + t64 &= 0x0000ffff; + t64 <<= dm8; + + const u32 t0 = l32_from_64 (t64); + const u32 t1 = h32_from_64 (t64); + + buf_dst[dd + 0] |= t0; + buf_dst[dd + 1] |= t1; } -static int mangle_lrest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static void append_one_byte (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst) { - for (int pos = 0; pos < len; pos++) lower_at (buf, pos); + const int sd = off_src / 4; + const int sm = off_src & 3; + const int sm8 = sm * 8; - return (len); + const int dd = off_dst / 4; + const int dm = off_dst & 3; + const int dm8 = dm * 8; + + u32 t = buf_src[sd]; + + t >>= sm8; + t &= 0xff; + t <<= dm8; + + buf_dst[dd] |= t; } -static int mangle_lrest_ufirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static void append_block (const u32 *buf_src, const int off_src, u32 *buf_dst, const int off_dst, const int len) { - for (int pos = 0; pos < len; pos++) lower_at (buf, pos); + int i; - upper_at (buf, 0); - - return (len); -} - -static int mangle_urest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) -{ - for (int pos = 0; pos < len; pos++) upper_at (buf, pos); - - return (len); -} - -static int mangle_urest_lfirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) -{ - for (int pos = 0; pos < len; pos++) upper_at (buf, pos); - - lower_at (buf, 0); - - return (len); -} - -static int mangle_trest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) -{ - for (int pos = 0; pos < len; pos++) toggle_at (buf, pos); - - return (len); -} - -static int mangle_toggle_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) -{ - if (p0 >= len) return (len); - - toggle_at (buf, p0); - - return (len); -} - -static int mangle_reverse (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) -{ - for (int l = 0; l < len / 2; l++) + for (i = 0; i < len - 4; i += 4) { - const int r = len - 1 - l; + append_four_byte (buf_src, off_src + i, buf_dst, off_dst + i); + } - mangle_switch (buf, l, r); + const int left = len - i; + + switch (left) + { + case 3: append_three_byte (buf_src, off_src + i, buf_dst, off_dst + i); break; + case 2: append_two_byte (buf_src, off_src + i, buf_dst, off_dst + i); break; + case 1: append_one_byte (buf_src, off_src + i, buf_dst, off_dst + i); break; + } +} + +static void exchange_byte (u32 *buf, const int off_src, const int off_dst) +{ + u8 *ptr = (u8 *) buf; + + const u8 tmp = ptr[off_src]; + + ptr[off_src] = ptr[off_dst]; + ptr[off_dst] = tmp; + + /* + something tells me we do this faster + + const int sd = off_src / 4; + const int sm = off_src & 3; + const int sm8 = sm * 8; + + const int dd = off_dst / 4; + const int dm = off_dst & 3; + const int dm8 = dm * 8; + + u32 ts = buf[sd]; + u32 td = buf[dd]; + + ts >>= sm8; + td >>= dm8; + + ts &= 0xff; + td &= 0xff; + + const u32 x = ts ^ td; + + const u32 xs = x << sm8; + const u32 xd = x << dm8; + + buf[sd] ^= xs; + buf[dd] ^= xd; + */ +} + +static int mangle_lrest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + const int lenv = ceil ((float) len / 4); + + for (int i = 0; i < lenv; i++) + { + const u32 t = buf[i]; + + buf[i] = t | generate_cmask (t); } return (len); } -static int mangle_dupeword (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_lrest_ufirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + const int lenv = ceil ((float) len / 4); + + for (int i = 0; i < lenv; i++) + { + const u32 t = buf[i]; + + buf[i] = t | generate_cmask (t); + } + + const u32 t = buf[0]; + + buf[0] = t & ~(0x00000020 & generate_cmask (t)); + + return (len); +} + +static int mangle_urest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + const int lenv = ceil ((float) len / 4); + + for (int i = 0; i < lenv; i++) + { + const u32 t = buf[i]; + + buf[i] = t & ~(generate_cmask (t)); + } + + return (len); +} + +static int mangle_urest_lfirst (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + const int lenv = ceil ((float) len / 4); + + for (int i = 0; i < lenv; i++) + { + const u32 t = buf[i]; + + buf[i] = t & ~(generate_cmask (t)); + } + + const u32 t = buf[0]; + + buf[0] = t | (0x00000020 & generate_cmask (t)); + + return (len); +} + +static int mangle_trest (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + const int lenv = ceil ((float) len / 4); + + for (int i = 0; i < lenv; i++) + { + const u32 t = buf[i]; + + buf[i] = t ^ generate_cmask (t); + } + + return (len); +} + +static int mangle_toggle_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + if (p0 >= len) return (len); + + const u8 p0d = p0 / 4; + const u8 p0m = p0 & 3; + + const u32 tmp = 0x20u << (p0m * 8); + + const u32 t = buf[p0d]; + + buf[p0d] = t ^ (generate_cmask (t) & tmp); + + return (len); +} + +static int mangle_reverse (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) +{ + for (int l = 0; l < len / 2; l++) + { + const int r = len - 1 - l; + + exchange_byte (buf, l, r); + } + + return (len); +} + +static int mangle_dupeword (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { const int out_len = len * 2; if (out_len >= RP_PASSWORD_SIZE) return (len); - u8 *out = buf + len; - - for (int i = 0; i < len; i++) *out++ = *buf++; + append_block (buf, 0, buf, len, len); return (out_len); } @@ -120,15 +291,20 @@ static int mangle_dupeword_times (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u return (out_len); } -static int mangle_reflect (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_reflect (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { const int out_len = len * 2; if (out_len >= RP_PASSWORD_SIZE) return (len); - mangle_dupeword (p0, p1, buf, len); + append_block (buf, 0, buf, len, len); - mangle_reverse (p0, p1, buf + len, len); + for (int l = 0; l < len / 2; l++) + { + const int r = len - 1 - l; + + exchange_byte (buf, len + l, len + r); + } return out_len; } @@ -160,21 +336,21 @@ static int mangle_prepend (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u return (out_len); } -static int mangle_rotate_left (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_rotate_left (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { for (int l = 0, r = len - 1; r > l; r--) { - mangle_switch (buf, l, r); + exchange_byte (buf, l, r); } return (len); } -static int mangle_rotate_right (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_rotate_right (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { for (int l = 0, r = len - 1; l < r; l++) { - mangle_switch (buf, l, r); + exchange_byte (buf, l, r); } return (len); @@ -366,30 +542,30 @@ static int mangle_dupechar_all (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 return (out_len); } -static int mangle_switch_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_switch_first (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { if (len < 2) return (len); - mangle_switch (buf, 0, 1); + exchange_byte (buf, 0, 1); return (len); } -static int mangle_switch_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_switch_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { if (len < 2) return (len); - mangle_switch (buf, len - 2, len - 1); + exchange_byte (buf, len - 2, len - 1); return (len); } -static int mangle_switch_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_switch_at (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { if (p0 >= len) return (len); if (p1 >= len) return (len); - mangle_switch (buf, p0, p1); + exchange_byte (buf, p0, p1); return (len); } @@ -486,86 +662,83 @@ static int mangle_dupeblock_last (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u return (out_len); } -static int mangle_title_sep (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int len) +static int mangle_title_sep (MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int len) { - int upper_next = 1; + if ((len + 4) >= RP_PASSWORD_SIZE) return (len); // cheap way to not need to check for overflow of i + 1 - for (int pos = 0; pos < len; pos++) + mangle_lrest_ufirst (0, 0, buf, len); + + for (int i = 0; i < ceil ((float) len / 4); i++) { - if (buf[pos] == p0) - { - upper_next = 1; + const u32 v = buf[i]; - continue; - } + u32 out0 = 0; + u32 out1 = 0; - if (upper_next) - { - upper_next = 0; + if (((v >> 0) & 0xff) == p0) out0 |= 0x0000ff00; + if (((v >> 8) & 0xff) == p0) out0 |= 0x00ff0000; + if (((v >> 16) & 0xff) == p0) out0 |= 0xff000000; + if (((v >> 24) & 0xff) == p0) out1 |= 0x000000ff; - upper_at (buf, pos); - } - else - { - lower_at (buf, pos); - } + buf[i + 0] &= ~(generate_cmask (buf[i + 0]) & out0); + buf[i + 1] &= ~(generate_cmask (buf[i + 1]) & out1); } return (len); } -static int apply_rule (const u32 name, MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u8 *buf, const int in_len) +static int apply_rule (const u32 name, MAYBE_UNUSED const u8 p0, MAYBE_UNUSED const u8 p1, u32 *buf, const int in_len) { int out_len = in_len; switch (name) { - case RULE_OP_MANGLE_LREST: out_len = mangle_lrest (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_LREST_UFIRST: out_len = mangle_lrest_ufirst (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_UREST: out_len = mangle_urest (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_UREST_LFIRST: out_len = mangle_urest_lfirst (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_TREST: out_len = mangle_trest (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_TOGGLE_AT: out_len = mangle_toggle_at (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_REVERSE: out_len = mangle_reverse (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPEWORD: out_len = mangle_dupeword (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPEWORD_TIMES: out_len = mangle_dupeword_times (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_REFLECT: out_len = mangle_reflect (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_APPEND: out_len = mangle_append (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_PREPEND: out_len = mangle_prepend (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_ROTATE_LEFT: out_len = mangle_rotate_left (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_ROTATE_RIGHT: out_len = mangle_rotate_right (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DELETE_FIRST: out_len = mangle_delete_first (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DELETE_LAST: out_len = mangle_delete_last (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DELETE_AT: out_len = mangle_delete_at (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_EXTRACT: out_len = mangle_extract (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_OMIT: out_len = mangle_omit (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_INSERT: out_len = mangle_insert (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_OVERSTRIKE: out_len = mangle_overstrike (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_TRUNCATE_AT: out_len = mangle_truncate_at (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_REPLACE: out_len = mangle_replace (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_PURGECHAR: out_len = mangle_purgechar (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPECHAR_FIRST: out_len = mangle_dupechar_first (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPECHAR_LAST: out_len = mangle_dupechar_last (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPECHAR_ALL: out_len = mangle_dupechar_all (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_SWITCH_FIRST: out_len = mangle_switch_first (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_SWITCH_LAST: out_len = mangle_switch_last (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_SWITCH_AT: out_len = mangle_switch_at (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_CHR_SHIFTL: out_len = mangle_chr_shiftl (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_CHR_SHIFTR: out_len = mangle_chr_shiftr (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_CHR_INCR: out_len = mangle_chr_incr (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_CHR_DECR: out_len = mangle_chr_decr (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_REPLACE_NP1: out_len = mangle_replace_np1 (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_REPLACE_NM1: out_len = mangle_replace_nm1 (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPEBLOCK_FIRST: out_len = mangle_dupeblock_first (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_DUPEBLOCK_LAST: out_len = mangle_dupeblock_last (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_TITLE_SEP: out_len = mangle_title_sep (p0, p1, buf, out_len); break; - case RULE_OP_MANGLE_TITLE: out_len = mangle_title_sep (' ', p1, buf, out_len); break; + case RULE_OP_MANGLE_LREST: out_len = mangle_lrest (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_LREST_UFIRST: out_len = mangle_lrest_ufirst (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_UREST: out_len = mangle_urest (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_UREST_LFIRST: out_len = mangle_urest_lfirst (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_TREST: out_len = mangle_trest (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_TOGGLE_AT: out_len = mangle_toggle_at (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_REVERSE: out_len = mangle_reverse (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_DUPEWORD: out_len = mangle_dupeword (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_DUPEWORD_TIMES: out_len = mangle_dupeword_times (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_REFLECT: out_len = mangle_reflect (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_APPEND: out_len = mangle_append (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_PREPEND: out_len = mangle_prepend (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_ROTATE_LEFT: out_len = mangle_rotate_left (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_ROTATE_RIGHT: out_len = mangle_rotate_right (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_DELETE_FIRST: out_len = mangle_delete_first (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DELETE_LAST: out_len = mangle_delete_last (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DELETE_AT: out_len = mangle_delete_at (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_EXTRACT: out_len = mangle_extract (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_OMIT: out_len = mangle_omit (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_INSERT: out_len = mangle_insert (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_OVERSTRIKE: out_len = mangle_overstrike (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_TRUNCATE_AT: out_len = mangle_truncate_at (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_REPLACE: out_len = mangle_replace (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_PURGECHAR: out_len = mangle_purgechar (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DUPECHAR_FIRST: out_len = mangle_dupechar_first (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DUPECHAR_LAST: out_len = mangle_dupechar_last (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DUPECHAR_ALL: out_len = mangle_dupechar_all (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_SWITCH_FIRST: out_len = mangle_switch_first (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_SWITCH_LAST: out_len = mangle_switch_last (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_SWITCH_AT: out_len = mangle_switch_at (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_CHR_SHIFTL: out_len = mangle_chr_shiftl (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_CHR_SHIFTR: out_len = mangle_chr_shiftr (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_CHR_INCR: out_len = mangle_chr_incr (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_CHR_DECR: out_len = mangle_chr_decr (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_REPLACE_NP1: out_len = mangle_replace_np1 (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_REPLACE_NM1: out_len = mangle_replace_nm1 (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DUPEBLOCK_FIRST: out_len = mangle_dupeblock_first (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_DUPEBLOCK_LAST: out_len = mangle_dupeblock_last (p0, p1, (u8 *) buf, out_len); break; + case RULE_OP_MANGLE_TITLE_SEP: out_len = mangle_title_sep (p0, p1, buf, out_len); break; + case RULE_OP_MANGLE_TITLE: out_len = mangle_title_sep (' ', p1, buf, out_len); break; } return out_len; } -int apply_rules (__global const u32 *cmds, u32 buf[64], const int in_len) +int apply_rules (__global const u32 *cmds, u32 *buf, const int in_len) { int out_len = in_len; @@ -577,7 +750,7 @@ int apply_rules (__global const u32 *cmds, u32 buf[64], const int in_len) const u8 p0 = (cmd >> 8) & 0xff; const u8 p1 = (cmd >> 16) & 0xff; - out_len = apply_rule (name, p0, p1, (u8 *) buf, out_len); + out_len = apply_rule (name, p0, p1, buf, out_len); } return out_len;