blob: 3c9038006deaef9f5453157423ac90de5424fddc [file] [log] [blame]
/*
* Copyright 2017 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#ifndef NFAtoDFA_DEFINED
#define NFAtoDFA_DEFINED
#include "src/sksl/lex/DFA.h"
#include "src/sksl/lex/DFAState.h"
#include "src/sksl/lex/NFA.h"
#include "src/sksl/lex/NFAState.h"
#include <algorithm>
#include <climits>
#include <memory>
#include <unordered_map>
#include <set>
#include <vector>
/**
* Converts a nondeterministic finite automaton to a deterministic finite automaton. Since NFAs and
* DFAs differ only in that an NFA allows multiple states at the same time, we can find each
* possible combination of simultaneous NFA states and give this combination a label. These labelled
* nodes are our DFA nodes, since we can only be in one such unique set of NFA states at a time.
*
* As an NFA can end up in multiple accept states at the same time (for instance, the token "while"
* is valid for both WHILE and IDENTIFIER), we disambiguate by preferring the first matching regex
* (in terms of the order in which they were added to the NFA).
*/
class NFAtoDFA {
public:
static constexpr char START_CHAR = 9;
static constexpr char END_CHAR = 126;
NFAtoDFA(NFA* nfa)
: fNFA(*nfa) {}
/**
* Returns a DFA created from the NFA.
*/
DFA convert() {
// create state 0, the "reject" state
getState(DFAState::Label({}));
// create a state representing being in all of the NFA's start states at once
std::vector<int> startStates = fNFA.fStartStates;
std::sort(startStates.begin(), startStates.end());
// this becomes state 1, our start state
DFAState* start = getState(DFAState::Label(startStates));
this->scanState(start);
this->computeMappings();
int stateCount = 0;
for (const auto& row : fTransitions) {
stateCount = std::max(stateCount, (int) row.size());
}
return DFA(fCharMappings, fTransitions, fAccepts);
}
private:
/**
* Returns an existing state with the given label, or creates a new one and returns it.
*/
DFAState* getState(DFAState::Label label) {
auto found = fStates.find(label);
if (found == fStates.end()) {
int id = fStates.size();
fStates[label] = std::unique_ptr<DFAState>(new DFAState(id, label));
return fStates[label].get();
}
return found->second.get();
}
void add(int nfaState, std::vector<int>* states) {
NFAState state = fNFA.fStates[nfaState];
if (state.fKind == NFAState::kRemapped_Kind) {
for (int next : state.fData) {
this->add(next, states);
}
} else {
for (int state : *states) {
if (nfaState == state) {
return;
}
}
states->push_back(nfaState);
}
}
void addTransition(char c, int start, int next) {
while (fTransitions.size() <= (size_t) c) {
fTransitions.push_back(std::vector<int>());
}
std::vector<int>& row = fTransitions[c];
while (row.size() <= (size_t) start) {
row.push_back(INVALID);
}
row[start] = next;
}
void scanState(DFAState* state) {
state->fIsScanned = true;
for (char c = START_CHAR; c <= END_CHAR; ++c) {
std::vector<int> next;
int bestAccept = INT_MAX;
for (int idx : state->fLabel.fStates) {
const NFAState& nfaState = fNFA.fStates[idx];
if (nfaState.accept(c)) {
for (int nextState : nfaState.fNext) {
if (fNFA.fStates[nextState].fKind == NFAState::kAccept_Kind) {
bestAccept = std::min(bestAccept, fNFA.fStates[nextState].fData[0]);
}
this->add(nextState, &next);
}
}
}
std::sort(next.begin(), next.end());
DFAState* nextState = this->getState(DFAState::Label(next));
this->addTransition(c, state->fId, nextState->fId);
if (bestAccept != INT_MAX) {
while (fAccepts.size() <= (size_t) nextState->fId) {
fAccepts.push_back(INVALID);
}
fAccepts[nextState->fId] = bestAccept;
}
if (!nextState->fIsScanned) {
this->scanState(nextState);
}
}
}
// collapse rows with the same transitions to a single row. This is common, as each row
// represents a character and often there are many characters for which all transitions are
// identical (e.g. [0-9] are treated the same way by all lexer rules)
void computeMappings() {
// mappings[<input row>] = <output row>
std::vector<std::vector<int>*> uniques;
// this could be done more efficiently, but O(n^2) is plenty fast for our purposes
for (size_t i = 0; i < fTransitions.size(); ++i) {
int found = -1;
for (size_t j = 0; j < uniques.size(); ++j) {
if (*uniques[j] == fTransitions[i]) {
found = j;
break;
}
}
if (found == -1) {
found = (int) uniques.size();
uniques.push_back(&fTransitions[i]);
}
fCharMappings.push_back(found);
}
std::vector<std::vector<int>> newTransitions;
for (std::vector<int>* row : uniques) {
newTransitions.push_back(*row);
}
fTransitions = newTransitions;
}
const NFA& fNFA;
std::unordered_map<DFAState::Label, std::unique_ptr<DFAState>> fStates;
std::vector<std::vector<int>> fTransitions;
std::vector<int> fCharMappings;
std::vector<int> fAccepts;
};
#endif // NFAtoDFA_DEFINED