github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-10-09 07:42:21 +03:00
Code Issues Packages Projects Releases Wiki Activity

Add SVE2 support for dvermicelli

Browse Source 
Change-Id: I056ef15e162ab6fb1f78964321ce893f4096367e
This commit is contained in:
George Wort
2021-06-23 14:14:28 +01:00
committed by Konstantinos Margaritis
parent 455789db9f
commit 185c45263b
6 changed files with 532 additions and 317 deletions
Download Patch File Download Diff File Expand all files Collapse all files

176
src/nfa/vermicelli_common.h
View File

@@ -37,51 +37,20 @@
#define VERM_TYPE m128
#define VERM_SET_FN set1_16x8
// returns NULL if not found
static really_inline
const u8 *lastMatchOffset(const u8 *buf_end, u32 z) {
assert(z);
return buf_end - 16 + 31 - clz32(z);
}
const u8 *dvermPreconditionMasked(m128 chars1, m128 chars2,
m128 mask1, m128 mask2, const u8 *buf) {
m128 data = loadu128(buf); // unaligned
m128 v1 = eq128(chars1, and128(data, mask1));
m128 v2 = eq128(chars2, and128(data, mask2));
u32 z = movemask128(and128(v1, rshiftbyte_m128(v2, 1)));
static really_inline
const u8 *dvermSearchAligned(m128 chars1, m128 chars2, u8 c1, u8 c2,
const u8 *buf, const u8 *buf_end) {
for (; buf + 16 < buf_end; buf += 16) {
m128 data = load128(buf);
u32 z = movemask128(and128(eq128(chars1, data),
rshiftbyte_m128(eq128(chars2, data), 1)));
if (buf[15] == c1 && buf[16] == c2) {
z |= (1 << 15);
}
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
return NULL;
}
static really_inline
const u8 *dvermSearchAlignedNocase(m128 chars1, m128 chars2, u8 c1, u8 c2,
const u8 *buf, const u8 *buf_end) {
assert((size_t)buf % 16 == 0);
m128 casemask = set1_16x8(CASE_CLEAR);
for (; buf + 16 < buf_end; buf += 16) {
m128 data = load128(buf);
m128 v = and128(casemask, data);
u32 z = movemask128(and128(eq128(chars1, v),
rshiftbyte_m128(eq128(chars2, v), 1)));
if ((buf[15] & CASE_CLEAR) == c1 && (buf[16] & CASE_CLEAR) == c2) {
z |= (1 << 15);
}
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
}
return NULL;
}
@@ -106,128 +75,5 @@ const u8 *dvermSearchAlignedMasked(m128 chars1, m128 chars2,
}
}
return NULL;
}
// returns NULL if not found
static really_inline
const u8 *dvermPrecondition(m128 chars1, m128 chars2, const u8 *buf) {
m128 data = loadu128(buf); // unaligned
u32 z = movemask128(and128(eq128(chars1, data),
rshiftbyte_m128(eq128(chars2, data), 1)));
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
return NULL;
}
// returns NULL if not found
static really_inline
const u8 *dvermPreconditionNocase(m128 chars1, m128 chars2, const u8 *buf) {
/* due to laziness, nonalphas and nocase having interesting behaviour */
m128 casemask = set1_16x8(CASE_CLEAR);
m128 data = loadu128(buf); // unaligned
m128 v = and128(casemask, data);
u32 z = movemask128(and128(eq128(chars1, v),
rshiftbyte_m128(eq128(chars2, v), 1)));
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
return NULL;
}
// returns NULL if not found
static really_inline
const u8 *dvermPreconditionMasked(m128 chars1, m128 chars2,
m128 mask1, m128 mask2, const u8 *buf) {
m128 data = loadu128(buf); // unaligned
m128 v1 = eq128(chars1, and128(data, mask1));
m128 v2 = eq128(chars2, and128(data, mask2));
u32 z = movemask128(and128(v1, rshiftbyte_m128(v2, 1)));
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
u32 pos = ctz32(z);
return buf + pos;
}
return NULL;
}
static really_inline
const u8 *rdvermSearchAligned(m128 chars1, m128 chars2, u8 c1, u8 c2,
const u8 *buf, const u8 *buf_end) {
assert((size_t)buf_end % 16 == 0);
for (; buf + 16 < buf_end; buf_end -= 16) {
m128 data = load128(buf_end - 16);
u32 z = movemask128(and128(eq128(chars2, data),
lshiftbyte_m128(eq128(chars1, data), 1)));
if (buf_end[-17] == c1 && buf_end[-16] == c2) {
z |= 1;
}
if (unlikely(z)) {
return lastMatchOffset(buf_end, z);
}
}
return buf_end;
}
static really_inline
const u8 *rdvermSearchAlignedNocase(m128 chars1, m128 chars2, u8 c1, u8 c2,
const u8 *buf, const u8 *buf_end) {
assert((size_t)buf_end % 16 == 0);
m128 casemask = set1_16x8(CASE_CLEAR);
for (; buf + 16 < buf_end; buf_end -= 16) {
m128 data = load128(buf_end - 16);
m128 v = and128(casemask, data);
u32 z = movemask128(and128(eq128(chars2, v),
lshiftbyte_m128(eq128(chars1, v), 1)));
if ((buf_end[-17] & CASE_CLEAR) == c1
&& (buf_end[-16] & CASE_CLEAR) == c2) {
z |= 1;
}
if (unlikely(z)) {
return lastMatchOffset(buf_end, z);
}
}
return buf_end;
}
// returns NULL if not found
static really_inline
const u8 *rdvermPrecondition(m128 chars1, m128 chars2, const u8 *buf) {
m128 data = loadu128(buf);
u32 z = movemask128(and128(eq128(chars2, data),
lshiftbyte_m128(eq128(chars1, data), 1)));
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
return lastMatchOffset(buf + 16, z);
}
return NULL;
}
// returns NULL if not found
static really_inline
const u8 *rdvermPreconditionNocase(m128 chars1, m128 chars2, const u8 *buf) {
/* due to laziness, nonalphas and nocase having interesting behaviour */
m128 casemask = set1_16x8(CASE_CLEAR);
m128 data = loadu128(buf);
m128 v = and128(casemask, data);
u32 z = movemask128(and128(eq128(chars2, v),
lshiftbyte_m128(eq128(chars1, v), 1)));
/* no fixup of the boundary required - the aligned run will pick it up */
if (unlikely(z)) {
return lastMatchOffset(buf + 16, z);
}
return NULL;
}