github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-16 09:21:52 +03:00
Code Issues Packages Projects Releases Wiki Activity

De-multiaccel

Browse Source
This commit is contained in:
Matthew Barr
2017-03-31 10:38:03 +11:00
parent 2b1a7da188
commit 423569ec82
38 changed files with 217 additions and 5293 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
src/nfagraph/ng_limex_accel.cpp
View File

@@ -37,7 +37,6 @@
#include "ue2common.h"
#include "nfa/accel.h"
#include "nfa/multiaccel_compilehelper.h"
#include "util/bitutils.h" // for CASE_CLEAR
#include "util/charreach.h"
@@ -677,134 +676,6 @@ NFAVertex get_sds_or_proxy(const NGHolder &g) {
return g.startDs;
}
static
NFAVertex find_next(const NFAVertex v, const NGHolder &g) {
NFAVertex res = NGHolder::null_vertex();
for (NFAVertex u : adjacent_vertices_range(v, g)) {
if (u != v) {
res = u;
break;
}
}
return res;
}
/** \brief Check if vertex \a v is a multi accelerable state (for a limex NFA). */
MultibyteAccelInfo nfaCheckMultiAccel(const NGHolder &g,
const vector<NFAVertex> &states,
const CompileContext &cc) {
// For a set of states to be accelerable, we basically have to have only
// one state to accelerate.
if (states.size() != 1) {
DEBUG_PRINTF("can't accelerate multiple states\n");
return MultibyteAccelInfo();
}
// Get our base vertex
NFAVertex v = states[0];
// We need the base vertex to be a self-looping dotall leading to exactly
// one vertex.
if (!hasSelfLoop(v, g)) {
DEBUG_PRINTF("base vertex has self-loop\n");
return MultibyteAccelInfo();
}
if (!g[v].char_reach.all()) {
DEBUG_PRINTF("can't accelerate anything but dot\n");
return MultibyteAccelInfo();
}
if (proper_out_degree(v, g) != 1) {
DEBUG_PRINTF("can't accelerate states with multiple successors\n");
return MultibyteAccelInfo();
}
// find our start vertex
NFAVertex cur = find_next(v, g);
if (cur == NGHolder::null_vertex()) {
DEBUG_PRINTF("invalid start vertex\n");
return MultibyteAccelInfo();
}
bool has_offset = false;
u32 offset = 0;
CharReach cr = g[cur].char_reach;
// if we start with a dot, we have an offset, so defer figuring out the
// real CharReach for this accel scheme
if (cr == CharReach::dot()) {
has_offset = true;
offset = 1;
}
// figure out our offset
while (has_offset) {
// vertices have to have no self loops
if (hasSelfLoop(cur, g)) {
DEBUG_PRINTF("can't have self-loops\n");
return MultibyteAccelInfo();
}
// we have to have exactly 1 successor to have this acceleration scheme
if (out_degree(cur, g) != 1) {
DEBUG_PRINTF("can't have multiple successors\n");
return MultibyteAccelInfo();
}
cur = *adjacent_vertices(cur, g).first;
// if we met a special vertex, bail out
if (is_special(cur, g)) {
DEBUG_PRINTF("can't have special vertices\n");
return MultibyteAccelInfo();
}
// now, get the real char reach
if (g[cur].char_reach != CharReach::dot()) {
cr = g[cur].char_reach;
has_offset = false;
} else {
offset++;
}
}
// now, fire up the compilation machinery
target_t ti = cc.target_info;
unsigned max_len = ti.has_avx2() ? MULTIACCEL_MAX_LEN_AVX2 : MULTIACCEL_MAX_LEN_SSE;
MultiaccelCompileHelper mac(cr, offset, max_len);
while (mac.canAdvance()) {
// vertices have to have no self loops
if (hasSelfLoop(cur, g)) {
break;
}
// we have to have exactly 1 successor to have this acceleration scheme
if (out_degree(cur, g) != 1) {
break;
}
cur = *adjacent_vertices(cur, g).first;
// if we met a special vertex, bail out
if (is_special(cur, g)) {
break;
}
mac.advance(g[cur].char_reach);
}
MultibyteAccelInfo mai = mac.getBestScheme();
#ifdef DEBUG
DEBUG_PRINTF("Multibyte acceleration scheme: type: %u offset: %u lengths: %u,%u\n",
mai.type, mai.offset, mai.len1, mai.len2);
for (size_t c = mai.cr.find_first(); c != CharReach::npos; c = mai.cr.find_next(c)) {
DEBUG_PRINTF("multibyte accel char: %zu\n", c);
}
#endif
return mai;
}
/** \brief Check if vertex \a v is an accelerable state (for a limex NFA). */
bool nfaCheckAccel(const NGHolder &g, NFAVertex v,
const vector<CharReach> &refined_cr,