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

prefetch works best when addresses are 64-byte aligned

Browse Source
This commit is contained in:
Konstantinos Margaritis 2021-03-12 10:10:53 +02:00
parent 521f233cfd
commit d3ff893871
6 changed files with 26 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/fdr/fdr.c
View File

@ -665,7 +665,7 @@ size_t prepareZones(const u8 *buf, size_t len, const u8 *hend,
const u8 *tryFloodDetect = zz->floodPtr; \
const u8 *start_ptr = zz->start; \
const u8 *end_ptr = zz->end; \
for (const u8 *itPtr = start_ptr; itPtr + 4*ITER_BYTES <= end_ptr; \
for (const u8 *itPtr = ROUNDDOWN_PTR(start_ptr, 64); itPtr + 4*ITER_BYTES <= end_ptr; \
itPtr += 4*ITER_BYTES) { \
__builtin_prefetch(itPtr); \
} \

4
src/hwlm/noodle_engine_avx2.c
View File

@ -95,7 +95,7 @@ hwlm_error_t scanSingleFast(const struct noodTable *n, const u8 *buf,
u32 z = movemask256(eq256(mask1, v));
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
hwlm_error_t result = single_zscan(n, d, buf, z, len, cbi);
if (unlikely(result != HWLM_SUCCESS))
@ -126,7 +126,7 @@ hwlm_error_t scanDoubleFast(const struct noodTable *n, const u8 *buf,
lastz0 = z0 >> 31;
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
hwlm_error_t result = double_zscan(n, d, buf, z, len, cbi);
if (unlikely(result != HWLM_SUCCESS))

4
src/hwlm/noodle_engine_sse.c
View File

@ -91,7 +91,7 @@ hwlm_error_t scanSingleFast(const struct noodTable *n, const u8 *buf,
u32 z = movemask128(eq128(mask1, v));
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
DEBUG_PRINTF("z 0x%08x\n", z);
hwlm_error_t result = single_zscan(n, d, buf, z, len, cbi);
@ -118,7 +118,7 @@ hwlm_error_t scanDoubleFast(const struct noodTable *n, const u8 *buf,
lastz1 = z1;
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
DEBUG_PRINTF("z 0x%08x\n", z);
hwlm_error_t result = double_zscan(n, d, buf, z, len, cbi);

18
src/nfa/mcclellan.c
View File

@ -634,10 +634,11 @@ char nfaExecMcClellan16_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
assert(ISALIGNED_N(q->state, 2));
u32 s = *(u16 *)q->state;
__builtin_prefetch(&m->remap[0]);
__builtin_prefetch(&m->remap[64]);
__builtin_prefetch(&m->remap[128]);
__builtin_prefetch(&m->remap[192]);
const u8 *base = ROUNDDOWN_PTR(&m->remap[0], 64);
__builtin_prefetch(base);
__builtin_prefetch(base + 64);
__builtin_prefetch(base + 128);
__builtin_prefetch(base + 192);
if (q->report_current) {
assert(s);
@ -795,10 +796,11 @@ char nfaExecMcClellan8_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
u32 s = *(u8 *)q->state;
__builtin_prefetch(&m->remap[0]);
__builtin_prefetch(&m->remap[64]);
__builtin_prefetch(&m->remap[128]);
__builtin_prefetch(&m->remap[192]);
const u8 *base = ROUNDDOWN_PTR(&m->remap[0], 64);
__builtin_prefetch(base);
__builtin_prefetch(base + 64);
__builtin_prefetch(base + 128);
__builtin_prefetch(base + 192);
if (q->report_current) {
assert(s);

18
src/nfa/mcsheng.c
View File

@ -889,10 +889,11 @@ char nfaExecMcSheng16_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
return MO_ALIVE;
}
__builtin_prefetch(&m->remap[0]);
__builtin_prefetch(&m->remap[64]);
__builtin_prefetch(&m->remap[128]);
__builtin_prefetch(&m->remap[192]);
const u8 *base = ROUNDDOWN_PTR(&m->remap[0], 64);
__builtin_prefetch(base);
__builtin_prefetch(base + 64);
__builtin_prefetch(base + 128);
__builtin_prefetch(base + 192);
while (1) {
assert(q->cur < q->end);
@ -1022,10 +1023,11 @@ char nfaExecMcSheng8_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
return MO_ALIVE;
}
__builtin_prefetch(&m->remap[0]);
__builtin_prefetch(&m->remap[64]);
__builtin_prefetch(&m->remap[128]);
__builtin_prefetch(&m->remap[192]);
const u8 *base = ROUNDDOWN_PTR(&m->remap[0], 64);
__builtin_prefetch(base);
__builtin_prefetch(base + 64);
__builtin_prefetch(base + 128);
__builtin_prefetch(base + 192);
while (1) {
DEBUG_PRINTF("%s @ %llu\n", q->items[q->cur].type == MQE_TOP ? "TOP" :

2
src/nfa/shufti.c
View File

@ -179,7 +179,7 @@ const u8 *shuftiExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
const u8 *last_block = buf_end - 16;
for (const u8 *itPtr = buf; itPtr + 4*16 <= last_block; itPtr += 4*16) {
for (const u8 *itPtr = ROUNDDOWN_PTR(buf, 64); itPtr + 4*16 <= last_block; itPtr += 4*16) {
__builtin_prefetch(itPtr);
}
while (buf < last_block) {