github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-19 02:30:35 +03:00
Code Issues Packages Projects Releases Wiki Activity

De-multiaccel

Browse Source
This commit is contained in:
Matthew Barr
2017-03-31 10:38:03 +11:00
parent 2b1a7da188
commit 423569ec82
38 changed files with 217 additions and 5293 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
src/nfa/shufti.c
View File

@@ -39,7 +39,52 @@
#include "util/simd_utils.h"
#include "util/unaligned.h"
#include "shufti_common.h"
#ifdef DEBUG
#include <ctype.h>
#define DUMP_MSK(_t) \
static UNUSED \
void dumpMsk##_t(m##_t msk) { \
u8 * mskAsU8 = (u8 *)&msk; \
for (unsigned i = 0; i < sizeof(msk); i++) { \
u8 c = mskAsU8[i]; \
for (int j = 0; j < 8; j++) { \
if ((c >> (7-j)) & 0x1) \
printf("1"); \
else \
printf("0"); \
} \
printf(" "); \
} \
} \
static UNUSED \
void dumpMsk##_t##AsChars(m##_t msk) { \
u8 * mskAsU8 = (u8 *)&msk; \
for (unsigned i = 0; i < sizeof(msk); i++) { \
u8 c = mskAsU8[i]; \
if (isprint(c)) \
printf("%c",c); \
else \
printf("."); \
} \
}
#endif
/** \brief Naive byte-by-byte implementation. */
static really_inline
const u8 *shuftiFwdSlow(const u8 *lo, const u8 *hi, const u8 *buf,
const u8 *buf_end) {
assert(buf < buf_end);
for (; buf < buf_end; ++buf) {
u8 c = *buf;
if (lo[c & 0xf] & hi[c >> 4]) {
break;
}
}
return buf;
}
/** \brief Naive byte-by-byte implementation. */
static really_inline
@@ -59,6 +104,30 @@ const u8 *shuftiRevSlow(const u8 *lo, const u8 *hi, const u8 *buf,
#if !defined(HAVE_AVX2)
/* Normal SSSE3 shufti */
#ifdef DEBUG
DUMP_MSK(128)
#endif
#define GET_LO_4(chars) and128(chars, low4bits)
#define GET_HI_4(chars) rshift64_m128(andnot128(low4bits, chars), 4)
static really_inline
u32 block(m128 mask_lo, m128 mask_hi, m128 chars, const m128 low4bits,
const m128 compare) {
m128 c_lo = pshufb(mask_lo, GET_LO_4(chars));
m128 c_hi = pshufb(mask_hi, GET_HI_4(chars));
m128 t = and128(c_lo, c_hi);
#ifdef DEBUG
DEBUG_PRINTF(" chars: "); dumpMsk128AsChars(chars); printf("\n");
DEBUG_PRINTF(" char: "); dumpMsk128(chars); printf("\n");
DEBUG_PRINTF(" c_lo: "); dumpMsk128(c_lo); printf("\n");
DEBUG_PRINTF(" c_hi: "); dumpMsk128(c_hi); printf("\n");
DEBUG_PRINTF(" t: "); dumpMsk128(t); printf("\n");
#endif
return movemask128(eq128(t, compare));
}
static really_inline
const u8 *firstMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffff)) {
@@ -293,6 +362,31 @@ const u8 *shuftiDoubleExec(m128 mask1_lo, m128 mask1_hi,
#else // AVX2 - 256 wide shuftis
#ifdef DEBUG
DUMP_MSK(256)
#endif
#define GET_LO_4(chars) and256(chars, low4bits)
#define GET_HI_4(chars) rshift64_m256(andnot256(low4bits, chars), 4)
static really_inline
u32 block(m256 mask_lo, m256 mask_hi, m256 chars, const m256 low4bits,
const m256 compare) {
m256 c_lo = vpshufb(mask_lo, GET_LO_4(chars));
m256 c_hi = vpshufb(mask_hi, GET_HI_4(chars));
m256 t = and256(c_lo, c_hi);
#ifdef DEBUG
DEBUG_PRINTF(" chars: "); dumpMsk256AsChars(chars); printf("\n");
DEBUG_PRINTF(" char: "); dumpMsk256(chars); printf("\n");
DEBUG_PRINTF(" c_lo: "); dumpMsk256(c_lo); printf("\n");
DEBUG_PRINTF(" c_hi: "); dumpMsk256(c_hi); printf("\n");
DEBUG_PRINTF(" t: "); dumpMsk256(t); printf("\n");
#endif
return movemask256(eq256(t, compare));
}
static really_inline
const u8 *firstMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffffffff)) {