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

refactor vermicelliDoubleMaskedExec()

Browse Source
This commit is contained in:
Konstantinos Margaritis 2021-11-02 22:30:21 +02:00
parent 16f3cca98b
commit 869d2bd53b
2 changed files with 106 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
src/nfa/vermicelli.hpp
View File

@ -35,6 +35,12 @@
#ifndef VERMICELLI_HPP #ifndef VERMICELLI_HPP
#define VERMICELLI_HPP #define VERMICELLI_HPP
#include "util/bitutils.h"
#ifdef HAVE_SVE2
#include "vermicelli_sve.h"
#endif
#ifdef __cplusplus #ifdef __cplusplus
extern "C" { extern "C" {
#endif #endif
@ -83,4 +89,12 @@ const u8 *rvermicelliDoubleExec(char c1, char c2, char nocase, const u8 *buf, co
} }
#endif #endif
#ifdef __cplusplus
extern "C" {
#endif
const u8 *vermicelliDoubleMaskedExec(char c1, char c2, char m1, char m2, const u8 *buf, const u8 *buf_end);
#ifdef __cplusplus
}
#endif
#endif /* VERMICELLI_HPP */ #endif /* VERMICELLI_HPP */

96
src/nfa/vermicelli_simd.cpp
View File

@ -111,6 +111,24 @@ const u8 *rvermicelliDoubleBlock(SuperVector<S> data, SuperVector<S> chars1, Sup
return last_non_zero_match<S>(buf, mask); return last_non_zero_match<S>(buf, mask);
} }
template <uint16_t S>
static really_inline
const u8 *vermicelliDoubleMaskedBlock(SuperVector<S> data, SuperVector<S> chars1, SuperVector<S> chars2,
SuperVector<S> mask1, SuperVector<S> mask2,
u8 const c1, u8 const c2, u8 const m1, u8 const m2, const u8 *buf) {
SuperVector<S> v1 = chars1.eq(data & mask1);
SuperVector<S> v2 = chars2.eq(data & mask2);
SuperVector<S> mask = v1 & (v2 >> 1);
DEBUG_PRINTF("rv[0] = %02hhx, rv[-1] = %02hhx\n", buf[0], buf[-1]);
bool partial_match = (((buf[0] & m1) == c2) && ((buf[-1] & m2) == c1));
DEBUG_PRINTF("partial = %d\n", partial_match);
if (partial_match) return buf - 1;
return first_non_zero_match<S>(buf, mask);
}
template <uint16_t S> template <uint16_t S>
static const u8 *vermicelliExecReal(SuperVector<S> const chars, SuperVector<S> const casemask, const u8 *buf, const u8 *buf_end) { static const u8 *vermicelliExecReal(SuperVector<S> const chars, SuperVector<S> const casemask, const u8 *buf, const u8 *buf_end) {
assert(buf && buf_end); assert(buf && buf_end);
@ -343,7 +361,7 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
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)) {
SuperVector<S> data = SuperVector<S>::loadu(d); SuperVector<S> data = SuperVector<S>::loadu(d);
rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, d);//, &lastmask1); rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, d);
if (rv) return rv; if (rv) return rv;
d = ROUNDUP_PTR(d, S); d = ROUNDUP_PTR(d, S);
} }
@ -352,7 +370,7 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
__builtin_prefetch(d + 64); __builtin_prefetch(d + 64);
DEBUG_PRINTF("d %p \n", d); DEBUG_PRINTF("d %p \n", d);
SuperVector<S> data = SuperVector<S>::load(d); SuperVector<S> data = SuperVector<S>::load(d);
rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, d);//, &lastmask1); rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, d);
if (rv) return rv; if (rv) return rv;
d += S; d += S;
} }
@ -363,7 +381,7 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
if (d != buf_end) { if (d != buf_end) {
SuperVector<S> data = SuperVector<S>::loadu_maskz(d, buf_end - d); SuperVector<S> data = SuperVector<S>::loadu_maskz(d, buf_end - d);
rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, d);//, buf_end - d); rv = vermicelliDoubleBlock(data, chars1, chars2, casemask, c1, c2, casechar, 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;
} }
@ -371,7 +389,6 @@ static const u8 *vermicelliDoubleExecReal(u8 const c1, u8 const c2, SuperVector<
DEBUG_PRINTF("real tail d %p e %p \n", d, buf_end); DEBUG_PRINTF("real tail d %p e %p \n", d, buf_end);
/* check for partial match at end */ /* check for partial match at end */
u8 mask = casemask.u.u8[0]; u8 mask = casemask.u.u8[0];
// u8 c1 = chars1.u.u8[0];
if ((buf_end[-1] & mask) == (u8)c1) { if ((buf_end[-1] & mask) == (u8)c1) {
DEBUG_PRINTF("partial!!!\n"); DEBUG_PRINTF("partial!!!\n");
return buf_end - 1; return buf_end - 1;
@ -439,6 +456,68 @@ const u8 *rvermicelliDoubleExecReal(char c1, char c2, SuperVector<S> const casem
return buf - 1; return buf - 1;
} }
template <uint16_t S>
static const u8 *vermicelliDoubleMaskedExecReal(u8 const c1, u8 const c2, u8 const m1, u8 const m2,
const u8 *buf, const u8 *buf_end) {
assert(buf && buf_end);
assert(buf < buf_end);
DEBUG_PRINTF("verm %p len %zu\n", buf, buf_end - buf);
DEBUG_PRINTF("b %s\n", buf);
const u8 *d = buf;
const u8 *rv;
// SuperVector<S> lastmask1{0};
const SuperVector<VECTORSIZE> chars1 = SuperVector<VECTORSIZE>::dup_u8(c1);
const SuperVector<VECTORSIZE> chars2 = SuperVector<VECTORSIZE>::dup_u8(c2);
const SuperVector<VECTORSIZE> mask1 = SuperVector<VECTORSIZE>::dup_u8(m1);
const SuperVector<VECTORSIZE> mask2 = SuperVector<VECTORSIZE>::dup_u8(m2);
__builtin_prefetch(d + 64);
__builtin_prefetch(d + 2*64);
__builtin_prefetch(d + 3*64);
__builtin_prefetch(d + 4*64);
DEBUG_PRINTF("start %p end %p \n", d, buf_end);
assert(d < buf_end);
if (d + S <= buf_end) {
// Reach vector aligned boundaries
DEBUG_PRINTF("until aligned %p \n", ROUNDUP_PTR(d, S));
if (!ISALIGNED_N(d, S)) {
SuperVector<S> data = SuperVector<S>::loadu(d);
rv = vermicelliDoubleMaskedBlock(data, chars1, chars2, mask1, mask2, c1, c2, m1, m2, d);
if (rv) return rv;
d = ROUNDUP_PTR(d, S);
}
while(d + S <= buf_end) {
__builtin_prefetch(d + 64);
DEBUG_PRINTF("d %p \n", d);
SuperVector<S> data = SuperVector<S>::load(d);
rv = vermicelliDoubleMaskedBlock(data, chars1, chars2, mask1, mask2, c1, c2, m1, m2, d);
if (rv) return rv;
d += S;
}
}
DEBUG_PRINTF("tail d %p e %p \n", d, buf_end);
// finish off tail
if (d != buf_end) {
SuperVector<S> data = SuperVector<S>::loadu_maskz(d, buf_end - d);
rv = vermicelliDoubleMaskedBlock(data, chars1, chars2, mask1, mask2, c1, c2, m1, m2, d);
DEBUG_PRINTF("rv %p \n", rv);
if (rv && rv < buf_end) return rv;
}
DEBUG_PRINTF("real tail d %p e %p \n", d, buf_end);
/* check for partial match at end */
if ((buf_end[-1] & m1) == (u8)c1) {
DEBUG_PRINTF("partial!!!\n");
return buf_end - 1;
}
return buf_end;
}
extern "C" const u8 *vermicelliExec(char c, char nocase, const u8 *buf, const u8 *buf_end) { extern "C" const u8 *vermicelliExec(char c, char nocase, const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("verm scan %s\\x%02hhx over %zu bytes\n", DEBUG_PRINTF("verm scan %s\\x%02hhx over %zu bytes\n",
nocase ? "nocase " : "", c, (size_t)(buf_end - buf)); nocase ? "nocase " : "", c, (size_t)(buf_end - buf));
@ -503,4 +582,13 @@ extern "C" const u8 *rvermicelliDoubleExec(char c1, char c2, char nocase, const
const SuperVector<VECTORSIZE> casemask{nocase ? getCaseMask<VECTORSIZE>() : SuperVector<VECTORSIZE>::Ones()}; const SuperVector<VECTORSIZE> casemask{nocase ? getCaseMask<VECTORSIZE>() : SuperVector<VECTORSIZE>::Ones()};
return rvermicelliDoubleExecReal<VECTORSIZE>(c1, c2, casemask, buf, buf_end); return rvermicelliDoubleExecReal<VECTORSIZE>(c1, c2, casemask, buf, buf_end);
}
extern "C" const u8 *vermicelliDoubleMaskedExec(char c1, char c2, char m1, char m2,
const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("double verm scan (\\x%02hhx&\\x%02hhx)(\\x%02hhx&\\x%02hhx) "
"over %zu bytes\n", c1, m1, c2, m2, (size_t)(buf_end - buf));
assert(buf < buf_end);
return vermicelliDoubleMaskedExecReal<VECTORSIZE>(c1, c2, m1, m2, buf, buf_end);
} }