/** Returns the minimum width in bytes of an input that will match the given
* graph. */
depth findMinWidth(const NGHolder &h) {
depth startDepth = findMinWidth(h, h.start);
depth dotstarDepth = findMinWidth(h, h.startDs);
DEBUG_PRINTF("startDepth=%s, dotstarDepth=%s\n", startDepth.str().c_str(),
dotstarDepth.str().c_str());
if (startDepth.is_unreachable()) {
assert(dotstarDepth.is_finite());
return dotstarDepth;
} else if (dotstarDepth.is_unreachable()) {
assert(startDepth.is_finite());
return startDepth;
} else {
assert(min(startDepth, dotstarDepth).is_finite());
return min(startDepth, dotstarDepth);
}
}
void SmallWriteBuildImpl::add(const NGWrapper &w) {
// If the graph is poisoned (i.e. we can't build a SmallWrite version),
// we don't even try.
if (poisoned) {
return;
}
if (w.som || w.min_length || isVacuous(w)) { /* cannot support in smwr */
poisoned = true;
return;
}
DEBUG_PRINTF("w=%p\n", &w);
// make a copy of the graph so that we can modify it for our purposes
unique_ptr<NGHolder> h = cloneHolder(w);
reduceGraph(*h, SOM_NONE, w.utf8, cc);
// If the earliest match location is outside the small write region,
// then we don't need to build a SmallWrite version.
// However, we don't poison this case either, since it is simply a case,
// where we know the resulting graph won't match.
if (findMinWidth(*h) > depth(cc.grey.smallWriteLargestBuffer)) {
return;
}
// Now we can actually build the McClellan DFA
assert(h->kind == NFA_OUTFIX);
auto r = buildMcClellan(*h, &rm, cc.grey);
// If we couldn't build a McClellan DFA for this portion, we won't be able
// build a smwr which represents the pattern set
if (!r) {
DEBUG_PRINTF("failed to determinise\n");
poisoned = true;
return;
}
prune_overlong(*r, cc.grey.smallWriteLargestBuffer);
if (rdfa) {
// do a merge of the new dfa with the existing dfa
auto merged = mergeTwoDfas(rdfa.get(), r.get(), DFA_MERGE_MAX_STATES,
&rm, cc.grey);
if (!merged) {
DEBUG_PRINTF("merge failed\n");
poisoned = true;
return;
}
DEBUG_PRINTF("merge succeeded, built %p\n", merged.get());
rdfa = move(merged);
} else {
rdfa = move(r);
}
}
void handleExtendedParams(ReportManager &rm, NGWrapper &g,
UNUSED const CompileContext &cc) {
if (!hasExtParams(g)) {
return;
}
depth minWidth = findMinWidth(g);
depth maxWidth = findMaxWidth(g);
bool is_anchored = !has_proper_successor(g.startDs, g)
&& out_degree(g.start, g);
bool has_offset_adj = hasOffsetAdjustments(rm, g);
DEBUG_PRINTF("minWidth=%s, maxWidth=%s, anchored=%d, offset_adj=%d\n",
minWidth.str().c_str(), maxWidth.str().c_str(), is_anchored,
has_offset_adj);
DepthMinMax match_depths = findMatchLengths(rm, g);
DEBUG_PRINTF("match depths %s\n", match_depths.str().c_str());
if (is_anchored && maxWidth.is_finite() && g.min_offset > maxWidth) {
ostringstream oss;
oss << "Expression is anchored and cannot satisfy min_offset="
<< g.min_offset << " as it can only produce matches of length "
<< maxWidth << " bytes at most.";
throw CompileError(g.expressionIndex, oss.str());
}
if (minWidth > g.max_offset) {
ostringstream oss;
oss << "Expression has max_offset=" << g.max_offset << " but requires "
<< minWidth << " bytes to match.";
throw CompileError(g.expressionIndex, oss.str());
}
if (maxWidth.is_finite() && match_depths.max < g.min_length) {
ostringstream oss;
oss << "Expression has min_length=" << g.min_length << " but can "
"only produce matches of length " << match_depths.max <<
" bytes at most.";
throw CompileError(g.expressionIndex, oss.str());
}
if (g.min_length && g.min_length <= match_depths.min) {
DEBUG_PRINTF("min_length=%llu constraint is unnecessary\n",
g.min_length);
g.min_length = 0;
}
if (!hasExtParams(g)) {
return;
}
pruneVacuousEdges(g);
pruneUnmatchable(g, rm);
if (!has_offset_adj) {
pruneExtUnreachable(g);
}
// We may have removed all the edges to accept, in which case this
// expression cannot match.
if (in_degree(g.accept, g) == 0 && in_degree(g.acceptEod, g) == 1) {
throw CompileError(g.expressionIndex, "Extended parameter "
"constraints can not be satisfied for any match from "
"this expression.");
}
// Remove reports on vertices without an edge to accept (which have been
// pruned above).
clearReports(g);
// Recalc.
minWidth = findMinWidth(g);
maxWidth = findMaxWidth(g);
is_anchored = proper_out_degree(g.startDs, g) == 0 &&
out_degree(g.start, g);
has_offset_adj = hasOffsetAdjustments(rm, g);
// If the pattern is completely anchored and has a min_length set, this can
// be converted to a min_offset.
if (g.min_length && (g.min_offset <= g.min_length) && is_anchored) {
DEBUG_PRINTF("converting min_length to min_offset=%llu for "
"anchored case\n", g.min_length);
g.min_offset = g.min_length;
g.min_length = 0;
}
if (g.min_offset && g.min_offset <= minWidth && !has_offset_adj) {
DEBUG_PRINTF("min_offset=%llu constraint is unnecessary\n",
g.min_offset);
g.min_offset = 0;
}
if (!hasExtParams(g)) {
return;
}
// If the pattern has a min_length and is of "ratchet" form with one
// unbounded repeat, that repeat can become a bounded repeat.
// e.g. /foo.*bar/{min_length=100} --> /foo.{94,}bar/
if (g.min_length && transformMinLengthToRepeat(rm, g)) {
DEBUG_PRINTF("converted min_length to bounded repeat\n");
// recalc
minWidth = findMinWidth(g);
}
// If the pattern is unanchored, has a max_offset and has not asked for
// SOM, we can use that knowledge to anchor it which will limit its
// lifespan. Note that we can't use this transformation if there's a
// min_length, as it's currently handled using "sly SOM".
// Note that it is possible to handle graphs that have a combination of
// anchored and unanchored paths, but it's too tricky for the moment.
if (g.max_offset != MAX_OFFSET && !g.som && !g.min_length &&
!has_offset_adj && isUnanchored(g)) {
if (anchorPatternWithBoundedRepeat(g, minWidth, maxWidth)) {
DEBUG_PRINTF("minWidth=%s, maxWidth=%s\n", minWidth.str().c_str(),
maxWidth.str().c_str());
if (minWidth == maxWidth) {
// For a fixed width pattern, we can retire the offsets as they
// are implicit in the graph now.
g.min_offset = 0;
g.max_offset = MAX_OFFSET;
}
}
}
//dumpGraph("final.dot", g.g);
if (!hasExtParams(g)) {
return;
}
set<NFAVertex> done;
updateReportBounds(rm, g, g.accept, done);
updateReportBounds(rm, g, g.acceptEod, done);
}
depth findMinWidth(const NGHolder &h, u32 top) {
return findMinWidth(h, SpecialEdgeFilter(h, top));
}
depth findMinWidth(const NGHolder &h) {
return findMinWidth(h, SpecialEdgeFilter(h));
}
static
depth findMaxWidth(const NGHolder &h, const SpecialEdgeFilter &filter,
NFAVertex src) {
if (isLeafNode(src, h.g)) {
return depth::unreachable();
}
if (hasReachableCycle(h, src)) {
// There's a cycle reachable from this src, so we have inf width.
return depth::infinity();
}
boost::filtered_graph<NFAGraph, SpecialEdgeFilter> g(h.g, filter);
assert(hasCorrectlyNumberedVertices(h));
const size_t num = num_vertices(h);
vector<int> distance(num);
vector<boost::default_color_type> colors(num);
auto index_map = get(&NFAGraphVertexProps::index, g);
// DAG shortest paths with negative edge weights.
dag_shortest_paths(
g, src,
distance_map(make_iterator_property_map(distance.begin(), index_map))
.weight_map(boost::make_constant_property<NFAEdge>(-1))
.vertex_index_map(index_map)
.color_map(make_iterator_property_map(colors.begin(), index_map)));
depth acceptDepth, acceptEodDepth;
if (colors.at(NODE_ACCEPT) == boost::white_color) {
acceptDepth = depth::unreachable();
} else {
acceptDepth = -1 * distance.at(NODE_ACCEPT);
}
if (colors.at(NODE_ACCEPT_EOD) == boost::white_color) {
acceptEodDepth = depth::unreachable();
} else {
acceptEodDepth = -1 * distance.at(NODE_ACCEPT_EOD);
}
depth d;
if (acceptDepth.is_unreachable()) {
d = acceptEodDepth;
} else if (acceptEodDepth.is_unreachable()) {
d = acceptDepth;
} else {
d = max(acceptDepth, acceptEodDepth);
}
if (d.is_unreachable()) {
// If we're actually reachable, we'll have a min width, so we can
// return infinity in this case.
if (findMinWidth(h, filter, src).is_reachable()) {
return depth::infinity();
}
return d;
}
// Invert sign and subtract one for start transition.
assert(d.is_finite() && d > depth(0));
return d - depth(1);
}