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

raise the limit of strings in double shufti

Browse Source
This commit is contained in:
Alex Coyte 2016-04-04 15:54:09 +10:00 committed by Matthew Barr
parent c0a5b037a1
commit ed3ef5b997
5 changed files with 22 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
src/nfa/accelcompile.cpp
View File

@ -207,17 +207,14 @@ void buildAccelDouble(const AccelInfo &info, AccelAux *aux) {
} }
} }
if (outs1 + outs2 <= 8) { if (outs1 < outs2 && outs1 <= 2) { // Heuristic from UE-438.
if (outs1 < outs2 && outs1 <= 2) { // Heuristic from UE-438. DEBUG_PRINTF("building double-shufti for %zu one-byte and %zu"
DEBUG_PRINTF("building double-shufti for %zu one-byte and %zu" " two-byte literals\n", outs1, outs2);
" two-byte literals\n", outs1, outs2); aux->accel_type = ACCEL_DSHUFTI;
aux->accel_type = ACCEL_DSHUFTI; aux->dshufti.offset = offset;
aux->dshufti.offset = offset; if (shuftiBuildDoubleMasks(info.double_stop1, info.double_stop2,
shuftiBuildDoubleMasks(info.double_stop1, info.double_stop2, &aux->dshufti.lo1, &aux->dshufti.hi1,
&aux->dshufti.lo1, &aux->dshufti.lo2, &aux->dshufti.hi2)) {
&aux->dshufti.hi1,
&aux->dshufti.lo2,
&aux->dshufti.hi2);
return; return;
} }
} }

11
src/nfa/mcclellancompile.cpp
View File

@ -130,7 +130,6 @@ mstate_aux *getAux(NFA *n, dstate_id_t i) {
static static
bool double_byte_ok(const escape_info &info) { bool double_byte_ok(const escape_info &info) {
return !info.outs2_broken return !info.outs2_broken
&& info.outs2_single.count() + info.outs2.size() <= 8
&& info.outs2_single.count() < info.outs2.size() && info.outs2_single.count() < info.outs2.size()
&& info.outs2_single.count() <= 2 && !info.outs2.empty(); && info.outs2_single.count() <= 2 && !info.outs2.empty();
} }
@ -256,14 +255,12 @@ void mcclellan_build_strat::buildAccel(UNUSED dstate_id_t this_idx,
} }
} }
if (double_byte_ok(info)) { if (double_byte_ok(info)
&& shuftiBuildDoubleMasks(info.outs2_single, info.outs2,
&accel->dshufti.lo1, &accel->dshufti.hi1,
&accel->dshufti.lo2, &accel->dshufti.hi2)) {
accel->accel_type = ACCEL_DSHUFTI; accel->accel_type = ACCEL_DSHUFTI;
accel->dshufti.offset = verify_u8(info.outs2_offset); accel->dshufti.offset = verify_u8(info.outs2_offset);
shuftiBuildDoubleMasks(info.outs2_single, info.outs2,
&accel->dshufti.lo1,
&accel->dshufti.hi1,
&accel->dshufti.lo2,
&accel->dshufti.hi2);
DEBUG_PRINTF("state %hu is double shufti\n", this_idx); DEBUG_PRINTF("state %hu is double shufti\n", this_idx);
return; return;
} }

1
src/nfa/mcclellancompile_accel.cpp
View File

@ -334,7 +334,6 @@ map<dstate_id_t, escape_info> populateAccelerationInfo(const raw_dfa &rdfa,
static static
bool double_byte_ok(const escape_info &info) { bool double_byte_ok(const escape_info &info) {
return !info.outs2_broken return !info.outs2_broken
&& info.outs2_single.count() + info.outs2.size() <= 8
&& info.outs2_single.count() < info.outs2.size() && info.outs2_single.count() < info.outs2.size()
&& info.outs2_single.count() <= 2 && !info.outs2.empty(); && info.outs2_single.count() <= 2 && !info.outs2.empty();
} }

7
src/nfagraph/ng_limex_accel.cpp
View File

@ -284,7 +284,7 @@ AccelScheme make_double_accel(AccelScheme as, CharReach cr_1,
return as; return as;
} }
if (two_count > 8) { if (two_count > DOUBLE_SHUFTI_LIMIT) {
if (cr_2.count() < cr_1.count()) { if (cr_2.count() < cr_1.count()) {
as.double_cr |= cr_2; as.double_cr |= cr_2;
offset = offset_in + 1; offset = offset_in + 1;
@ -513,7 +513,7 @@ AccelScheme findBestAccelScheme(vector<vector<CharReach> > paths,
best.offset = offset; best.offset = offset;
/* merge best single and best double */ /* merge best single and best double */
if (!da.double_byte.empty() && da.double_byte.size() <= 8 if (!da.double_byte.empty() && da.double_byte.size() <= DOUBLE_SHUFTI_LIMIT
&& da.double_cr.count() < best.cr.count()) { && da.double_cr.count() < best.cr.count()) {
best.double_byte = da.double_byte; best.double_byte = da.double_byte;
best.double_cr = da.double_cr; best.double_cr = da.double_cr;
@ -857,7 +857,8 @@ depth_done:
// literals) // literals)
if (depth > 1) { if (depth > 1) {
for (unsigned int i = 0; i < (depth - 1); i++) { for (unsigned int i = 0; i < (depth - 1); i++) {
if (depthReach[i].count()*depthReach[i+1].count() <= 8) { if (depthReach[i].count() * depthReach[i+1].count()
<= DOUBLE_SHUFTI_LIMIT) {
DEBUG_PRINTF("two-byte shufti, depth %u\n", i); DEBUG_PRINTF("two-byte shufti, depth %u\n", i);
*as = AccelScheme(CharReach::dot(), i); *as = AccelScheme(CharReach::dot(), i);
return true; return true;

10
src/nfagraph/ng_limex_accel.h
View File

@ -63,6 +63,8 @@ void findAccelFriends(const NGHolder &g, NFAVertex v,
u32 offset, u32 offset,
ue2::flat_set<NFAVertex> *friends); ue2::flat_set<NFAVertex> *friends);
#define DOUBLE_SHUFTI_LIMIT 20
struct AccelScheme { struct AccelScheme {
AccelScheme(const CharReach &cr_in, u32 offset_in) AccelScheme(const CharReach &cr_in, u32 offset_in)
: cr(cr_in), offset(offset_in) { : cr(cr_in), offset(offset_in) {
@ -78,10 +80,10 @@ struct AccelScheme {
size_t a_dcount = double_cr.count(); size_t a_dcount = double_cr.count();
size_t b_dcount = b.double_cr.count(); size_t b_dcount = b.double_cr.count();
bool feasible_double_a bool feasible_double_a = !a.double_byte.empty()
= !a.double_byte.empty() && a.double_byte.size() <= 8; && a.double_byte.size() <= DOUBLE_SHUFTI_LIMIT;
bool feasible_double_b bool feasible_double_b = !b.double_byte.empty()
= !b.double_byte.empty() && b.double_byte.size() <= 8; && b.double_byte.size() <= DOUBLE_SHUFTI_LIMIT;
if (feasible_double_a != feasible_double_b) { if (feasible_double_a != feasible_double_b) {
return feasible_double_a > feasible_double_b; return feasible_double_a > feasible_double_b;