prefetch works best when addresses are 64-byte aligned

This commit is contained in:
Konstantinos Margaritis
2021-03-12 10:10:53 +02:00
parent 68b92f338d
commit acb542a5be
6 changed files with 26 additions and 22 deletions

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);