github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix a couple of tests

Browse Source
This commit is contained in:
Danila Kutenin 2022-02-18 19:31:03 +00:00
parent b3e88e480f
commit 5f8729a085
1 changed files with 12 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
src/nfa/vermicelli_simd.cpp
View File

@ -310,7 +310,7 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
__builtin_prefetch(d + 4*64); __builtin_prefetch(d + 4*64);
DEBUG_PRINTF("start %p end %p \n", d, buf_end); DEBUG_PRINTF("start %p end %p \n", d, buf_end);
assert(d < buf_end); assert(d < buf_end);
if (d + S <= buf_end) { if (d + S < buf_end) {
// Reach vector aligned boundaries // Reach vector aligned boundaries
DEBUG_PRINTF("until aligned %p \n", ROUNDUP_PTR(d, S)); DEBUG_PRINTF("until aligned %p \n", ROUNDUP_PTR(d, S));
if (!ISALIGNED_N(d, S)) { if (!ISALIGNED_N(d, S)) {
@ -336,15 +336,12 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
if (d != buf_end) { if (d != buf_end) {
SuperVector<S> data = SuperVector<S>::Zeroes(); SuperVector<S> data = SuperVector<S>::Zeroes();
const u8* end_buf; if (buf_end - d < S) {
if (buf_end - buf < S) { memcpy(&data.u, d, buf_end - d);
memcpy(&data.u, buf, buf_end - buf);
end_buf = buf;
} else { } else {
data = SuperVector<S>::loadu(buf_end - S); data = SuperVector<S>::loadu(d);
end_buf = buf_end - S;
} }
rv = vermicelliDoubleBlock<S, false>(data, chars1, chars2, casemask, c1, c2, casechar, end_buf, buf_end - d); rv = vermicelliDoubleBlock<S, false>(data, chars1, chars2, casemask, c1, c2, casechar, d, buf_end - d);
DEBUG_PRINTF("rv %p \n", rv); DEBUG_PRINTF("rv %p \n", rv);
if (rv && rv < buf_end) return rv; if (rv && rv < buf_end) return rv;
} }
@ -383,7 +380,7 @@ const u8 *rvermicelliDoubleExecReal(char c1, char c2, SuperVector<S> const casem
__builtin_prefetch(d - 4*64); __builtin_prefetch(d - 4*64);
DEBUG_PRINTF("start %p end %p \n", buf, d); DEBUG_PRINTF("start %p end %p \n", buf, d);
assert(d > buf); assert(d > buf);
if (d - S >= buf) { if (d - S > buf) {
// Reach vector aligned boundaries // Reach vector aligned boundaries
DEBUG_PRINTF("until aligned %p \n", ROUNDDOWN_PTR(d, S)); DEBUG_PRINTF("until aligned %p \n", ROUNDDOWN_PTR(d, S));
if (!ISALIGNED_N(d, S)) { if (!ISALIGNED_N(d, S)) {
@ -395,7 +392,7 @@ const u8 *rvermicelliDoubleExecReal(char c1, char c2, SuperVector<S> const casem
d = d1; d = d1;
} }
while (d - S >= buf) { while (d - S > buf) {
DEBUG_PRINTF("aligned %p \n", d); DEBUG_PRINTF("aligned %p \n", d);
// On large packet buffers, this prefetch appears to get us about 2%. // On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d - 64); __builtin_prefetch(d - 64);
@ -447,7 +444,7 @@ static const u8 *vermicelliDoubleMaskedExecReal(u8 const c1, u8 const c2, u8 con
__builtin_prefetch(d + 4*64); __builtin_prefetch(d + 4*64);
DEBUG_PRINTF("start %p end %p \n", d, buf_end); DEBUG_PRINTF("start %p end %p \n", d, buf_end);
assert(d < buf_end); assert(d < buf_end);
if (d + S <= buf_end) { if (d + S < buf_end) {
// Reach vector aligned boundaries // Reach vector aligned boundaries
DEBUG_PRINTF("until aligned %p \n", ROUNDUP_PTR(d, S)); DEBUG_PRINTF("until aligned %p \n", ROUNDUP_PTR(d, S));
if (!ISALIGNED_N(d, S)) { if (!ISALIGNED_N(d, S)) {
@ -473,15 +470,12 @@ static const u8 *vermicelliDoubleMaskedExecReal(u8 const c1, u8 const c2, u8 con
if (d != buf_end) { if (d != buf_end) {
SuperVector<S> data = SuperVector<S>::Zeroes(); SuperVector<S> data = SuperVector<S>::Zeroes();
const u8* end_buf; if (buf_end - d < S) {
if (buf_end - buf < S) { memcpy(&data.u, d, buf_end - d);
memcpy(&data.u, buf, buf_end - buf);
end_buf = buf;
} else { } else {
data = SuperVector<S>::loadu(buf_end - S); data = SuperVector<S>::loadu(d);
end_buf = buf_end - S;
} }
rv = vermicelliDoubleMaskedBlock<S, false>(data, chars1, chars2, mask1, mask2, c1, c2, m1, m2, end_buf, buf_end - d); rv = vermicelliDoubleMaskedBlock<S, false>(data, chars1, chars2, mask1, mask2, c1, c2, m1, m2, d, buf_end - d);
DEBUG_PRINTF("rv %p \n", rv); DEBUG_PRINTF("rv %p \n", rv);
if (rv && rv < buf_end) return rv; if (rv && rv < buf_end) return rv;
} }