/** Some squash states are clearly not advantageous in the NFA, as they do
* incur the cost of an exception:
* -# acyclic states
* -# squash only a few acyclic states
*/
void filterSquashers(const NGHolder &g,
map<NFAVertex, NFAStateSet> &squash) {
DEBUG_PRINTF("filtering\n");
map<u32, NFAVertex> rev; /* vertex_index -> vertex */
for (auto v : vertices_range(g)) {
rev[g[v].index] = v;
}
for (auto v : vertices_range(g)) {
if (!contains(squash, v)) {
continue;
}
DEBUG_PRINTF("looking at squash set for vertex %u\n",
g[v].index);
if (!hasSelfLoop(v, g)) {
DEBUG_PRINTF("acyclic\n");
squash.erase(v);
continue;
}
NFAStateSet squashed = squash[v];
squashed.flip(); /* default sense for mask of survivors */
for (NFAStateSet::size_type sq = squashed.find_first();
sq != squashed.npos; sq = squashed.find_next(sq)) {
NFAVertex u = rev[sq];
if (hasSelfLoop(u, g)) {
DEBUG_PRINTF("squashing a cyclic (%zu) is always good\n", sq);
goto next_vertex;
}
}
if (squashed.count() < MIN_PURE_ACYCLIC_SQUASH) {
DEBUG_PRINTF("squash set too small\n");
squash.erase(v);
continue;
}
next_vertex:;
DEBUG_PRINTF("squash set ok\n");
}
}
/**
* Builds a squash mask based on the pdom tree of v and the given char reach.
* The built squash mask is a bit conservative for non-dot cases and could
* be improved with a bit of thought.
*/
static
void buildSquashMask(NFAStateSet &mask, const NGHolder &g, NFAVertex v,
const CharReach &cr, const NFAStateSet &init,
const vector<NFAVertex> &vByIndex, const PostDomTree &tree,
som_type som, const vector<DepthMinMax> &som_depths,
const ue2::unordered_map<NFAVertex, u32> ®ion_map,
smgb_cache &cache) {
DEBUG_PRINTF("build base squash mask for vertex %u)\n",
g[v].index);
vector<NFAVertex> q;
PostDomTree::const_iterator it = tree.find(v);
if (it != tree.end()) {
q.insert(q.end(), it->second.begin(), it->second.end());
}
const u32 v_index = g[v].index;
while (!q.empty()) {
NFAVertex u = q.back();
q.pop_back();
const CharReach &cru = g[u].char_reach;
if ((cru & ~cr).any()) {
/* bail: bad cr on vertex u */
/* TODO: this could be better
*
* we still need to ensure that we record any paths leading to u.
* Hence all vertices R which can reach u must be excluded from the
* squash mask. Note: R != pdom(u) and there may exist an x in (R -
* pdom(u)) which is in pdom(y) where y is in q. Clear ?
*/
mask.set();
return;
}
const u32 u_index = g[u].index;
if (som) {
/* We cannot add a state u to the squash mask of v if it may have an
* earlier start of match offset. ie for us to add a state u to v
* maxSomDist(u) <= minSomDist(v)
*/
const depth &max_som_dist_u = som_depths[u_index].max;
const depth &min_som_dist_v = som_depths[v_index].min;
if (max_som_dist_u.is_infinite()) {
/* it is hard to tell due to the INF if u can actually store an
* earlier SOM than w (state we are building the squash mask
* for) - need to think more deeply
*/
if (mustBeSetBefore(u, v, g, cache)
&& !somMayGoBackwards(u, g, region_map, cache)) {
DEBUG_PRINTF("u %u v %u\n", u_index, v_index);
goto squash_ok;
}
}
if (max_som_dist_u > min_som_dist_v) {
/* u can't be squashed as it may be storing an earlier SOM */
goto add_children_to_queue;
}
}
squash_ok:
mask.set(u_index);
DEBUG_PRINTF("pdom'ed %u\n", u_index);
add_children_to_queue:
it = tree.find(u);
if (it != tree.end()) {
q.insert(q.end(), it->second.begin(), it->second.end());
}
}
if (cr.all()) {
/* the init states aren't in the pdom tree. If all their succ states
* are set (or v), we can consider them post dominated */
/* Note: init states will always result in a later som */
for (size_t i = init.find_first(); i != init.npos;
i = init.find_next(i)) {
/* Yes vacuous patterns do exist */
NFAVertex iv = vByIndex[i];
for (auto w : adjacent_vertices_range(iv, g)) {
if (w == g.accept || w == g.acceptEod) {
DEBUG_PRINTF("skipping %zu due to vacuous accept\n", i);
goto next_init_state;
}
u32 vert_id = g[w].index;
if (w != iv && w != v && !mask.test(vert_id)) {
DEBUG_PRINTF("skipping %zu due to %u\n", i, vert_id);
goto next_init_state;
}
}
DEBUG_PRINTF("pdom'ed %zu\n", i);
mask.set(i);
next_init_state:;
}
}
mask.flip();
}
