github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2026-01-02 22:54:43 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix double shufti's vector end false positive (#325)

Browse Source 
* Add regression test for double shufti

It tests for false positive at vector edges.

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix double shufti reporting false positives

Double shufti used to offset one vector, resulting in losing one character
at the end of every vector. This was replaced by a magic value indicating a
match. This meant that if the first char of a pattern fell on the last char of
a vector, double shufti would assume the second character is present and
report a match.
This patch fixes it by keeping the previous vector and feeding its data to the
new one when we shift it, preventing any loss of data.

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* vshl() will call the correct implementation

* implement missing vshr_512_imm(), simplifies caller x86 code

* Fix x86 case, use alignr instead

* it's the reverse, the avx512 alignr is incorrect, need to fix

* Make shufti's OR reduce size agnostic

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix test's array size

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix AVX2/AVX512 alignr implementations and unit tests

* Fix Power VSX alignr

---------

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>
Co-authored-by: Konstantinos Margaritis <konstantinos@vectorcamp.gr>
This commit is contained in:
ypicchi-arm
2025-06-11 16:55:10 +01:00
committed by GitHub
parent 8a44a6eda2
commit ca70a3d9be
9 changed files with 240 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

101
src/nfa/shufti_sve.hpp
View File

@@ -153,7 +153,7 @@ const u8 *rshuftiExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
static really_inline
svbool_t doubleMatched(svuint8_t mask1_lo, svuint8_t mask1_hi,
svuint8_t mask2_lo, svuint8_t mask2_hi,
const u8 *buf, const svbool_t pg) {
svuint8_t* inout_t1, const u8 *buf, const svbool_t pg) {
svuint8_t vec = svld1_u8(pg, buf);
svuint8_t chars_lo = svand_x(svptrue_b8(), vec, (uint8_t)0xf);
@@ -161,38 +161,59 @@ svbool_t doubleMatched(svuint8_t mask1_lo, svuint8_t mask1_hi,
svuint8_t c1_lo = svtbl(mask1_lo, chars_lo);
svuint8_t c1_hi = svtbl(mask1_hi, chars_hi);
svuint8_t t1 = svorr_x(svptrue_b8(), c1_lo, c1_hi);
svuint8_t new_t1 = svorr_z(svptrue_b8(), c1_lo, c1_hi);
svuint8_t c2_lo = svtbl(mask2_lo, chars_lo);
svuint8_t c2_hi = svtbl(mask2_hi, chars_hi);
svuint8_t t2 = svext(svorr_z(pg, c2_lo, c2_hi), svdup_u8(0), 1);
svuint8_t t2 = svorr_x(svptrue_b8(), c2_lo, c2_hi);
svuint8_t t = svorr_x(svptrue_b8(), t1, t2);
// shift t1 left by one and feeds in the last element from the previous t1
uint8_t last_elem = svlastb(svptrue_b8(), *inout_t1);
svuint8_t merged_t1 = svinsr(new_t1, last_elem);
svuint8_t t = svorr_x(svptrue_b8(), merged_t1, t2);
*inout_t1 = new_t1;
return svnot_z(svptrue_b8(), svcmpeq(svptrue_b8(), t, (uint8_t)0xff));
}
static really_inline
const u8 *check_last_byte(svuint8_t mask2_lo, svuint8_t mask2_hi,
uint8_t last_elem, const u8 *buf_end) {
uint8_t wild_lo = svorv(svptrue_b8(), mask2_lo);
uint8_t wild_hi = svorv(svptrue_b8(), mask2_hi);
uint8_t match_inverted = wild_lo | wild_hi | last_elem;
int match = match_inverted != 0xff;
if(match) {
return buf_end - 1;
}
return NULL;
}
static really_inline
const u8 *dshuftiOnce(svuint8_t mask1_lo, svuint8_t mask1_hi,
svuint8_t mask2_lo, svuint8_t mask2_hi,
const u8 *buf, const u8 *buf_end) {
svuint8_t *inout_t1, const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("start %p end %p\n", buf, buf_end);
assert(buf < buf_end);
DEBUG_PRINTF("l = %td\n", buf_end - buf);
svbool_t pg = svwhilelt_b8_s64(0, buf_end - buf);
svbool_t matched = doubleMatched(mask1_lo, mask1_hi, mask2_lo, mask2_hi,
buf, pg);
return accelSearchCheckMatched(buf, matched);
inout_t1, buf, pg);
// doubleMatched return match position of the second char, but here we
// return the position of the first char, hence the buffer offset
return accelSearchCheckMatched(buf - 1, matched);
}
static really_inline
const u8 *dshuftiLoopBody(svuint8_t mask1_lo, svuint8_t mask1_hi,
svuint8_t mask2_lo, svuint8_t mask2_hi,
const u8 *buf) {
svuint8_t *inout_t1, const u8 *buf) {
DEBUG_PRINTF("start %p end %p\n", buf, buf + svcntb());
svbool_t matched = doubleMatched(mask1_lo, mask1_hi, mask2_lo, mask2_hi,
buf, svptrue_b8());
return accelSearchCheckMatched(buf, matched);
inout_t1, buf, svptrue_b8());
// doubleMatched return match position of the second char, but here we
// return the position of the first char, hence the buffer offset
return accelSearchCheckMatched(buf - 1, matched);
}
static really_inline
@@ -200,31 +221,47 @@ const u8 *dshuftiSearch(svuint8_t mask1_lo, svuint8_t mask1_hi,
svuint8_t mask2_lo, svuint8_t mask2_hi,
const u8 *buf, const u8 *buf_end) {
assert(buf < buf_end);
svuint8_t inout_t1 = svdup_u8(0xff);
size_t len = buf_end - buf;
if (len <= svcntb()) {
return dshuftiOnce(mask1_lo, mask1_hi,
mask2_lo, mask2_hi, buf, buf_end);
}
// peel off first part to align to the vector size
const u8 *aligned_buf = ROUNDUP_PTR(buf, svcntb_pat(SV_POW2));
assert(aligned_buf < buf_end);
if (buf != aligned_buf) {
const u8 *ptr = dshuftiLoopBody(mask1_lo, mask1_hi,
mask2_lo, mask2_hi, buf);
if (ptr) return ptr;
}
buf = aligned_buf;
size_t loops = (buf_end - buf) / svcntb();
DEBUG_PRINTF("loops %zu \n", loops);
for (size_t i = 0; i < loops; i++, buf += svcntb()) {
const u8 *ptr = dshuftiLoopBody(mask1_lo, mask1_hi,
mask2_lo, mask2_hi, buf);
if (ptr) return ptr;
if (len > svcntb()) {
// peel off first part to align to the vector size
const u8 *aligned_buf = ROUNDUP_PTR(buf, svcntb_pat(SV_POW2));
assert(aligned_buf < buf_end);
if (buf != aligned_buf) {
const u8 *ptr = dshuftiLoopBody(mask1_lo, mask1_hi, mask2_lo,
mask2_hi, &inout_t1, buf);
if (ptr) return ptr;
// The last match in inout won't line up with the next round as we
// use an overlap. We need to set inout according to the last
// unique-searched char.
size_t offset = aligned_buf - buf;
uint8_t last_unique_elem =
svlastb(svwhilelt_b8(0UL, offset), inout_t1);
inout_t1 = svdup_u8(last_unique_elem);
}
buf = aligned_buf;
size_t loops = (buf_end - buf) / svcntb();
DEBUG_PRINTF("loops %zu \n", loops);
for (size_t i = 0; i < loops; i++, buf += svcntb()) {
const u8 *ptr = dshuftiLoopBody(mask1_lo, mask1_hi, mask2_lo,
mask2_hi, &inout_t1, buf);
if (ptr) return ptr;
}
if (buf == buf_end) {
uint8_t last_elem = svlastb(svptrue_b8(), inout_t1);
return check_last_byte(mask2_lo, mask2_hi, last_elem, buf_end);
}
}
DEBUG_PRINTF("buf %p buf_end %p \n", buf, buf_end);
return buf == buf_end ? NULL : dshuftiLoopBody(mask1_lo, mask1_hi,
mask2_lo, mask2_hi,
buf_end - svcntb());
len = buf_end - buf;
const u8 *ptr = dshuftiOnce(mask1_lo, mask1_hi,
mask2_lo, mask2_hi, &inout_t1, buf, buf_end);
if (ptr) return ptr;
uint8_t last_elem =
svlastb(svwhilelt_b8(0UL, len), inout_t1);
return check_last_byte(mask2_lo, mask2_hi, last_elem, buf_end);
}
const u8 *shuftiDoubleExec(m128 mask1_lo, m128 mask1_hi,