1//===- DFAEmitter.cpp - Finite state automaton emitter --------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This class can produce a generic deterministic finite state automaton (DFA),
10// given a set of possible states and transitions.
11//
12// The input transitions can be nondeterministic - this class will produce the
13// deterministic equivalent state machine.
14//
15// The generated code can run the DFA and produce an accepted / not accepted
16// state and also produce, given a sequence of transitions that results in an
17// accepted state, the sequence of intermediate states. This is useful if the
18// initial automaton was nondeterministic - it allows mapping back from the DFA
19// to the NFA.
20//
21//===----------------------------------------------------------------------===//
22
23#include "DFAEmitter.h"
24#include "Basic/SequenceToOffsetTable.h"
25#include "llvm/ADT/SmallVector.h"
26#include "llvm/ADT/StringExtras.h"
27#include "llvm/ADT/UniqueVector.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/Support/raw_ostream.h"
30#include "llvm/TableGen/Record.h"
31#include "llvm/TableGen/TableGenBackend.h"
32#include <cassert>
33#include <cstdint>
34#include <deque>
35#include <set>
36#include <string>
37#include <variant>
38#include <vector>
39
40#define DEBUG_TYPE "dfa-emitter"
41
42using namespace llvm;
43
44//===----------------------------------------------------------------------===//
45// DfaEmitter implementation. This is independent of the GenAutomaton backend.
46//===----------------------------------------------------------------------===//
47
48void DfaEmitter::addTransition(state_type From, state_type To, action_type A) {
49 Actions.insert(x: A);
50 NfaStates.insert(x: From);
51 NfaStates.insert(x: To);
52 NfaTransitions[{From, A}].push_back(x: To);
53 ++NumNfaTransitions;
54}
55
56void DfaEmitter::visitDfaState(const DfaState &DS) {
57 // For every possible action...
58 auto FromId = DfaStates.idFor(Entry: DS);
59 for (action_type A : Actions) {
60 DfaState NewStates;
61 DfaTransitionInfo TI;
62 // For every represented state, word pair in the original NFA...
63 for (state_type FromState : DS) {
64 // If this action is possible from this state add the transitioned-to
65 // states to NewStates.
66 auto I = NfaTransitions.find(x: {FromState, A});
67 if (I == NfaTransitions.end())
68 continue;
69 for (state_type &ToState : I->second) {
70 NewStates.push_back(Elt: ToState);
71 TI.emplace_back(Args&: FromState, Args&: ToState);
72 }
73 }
74 if (NewStates.empty())
75 continue;
76 // Sort and unique.
77 sort(C&: NewStates);
78 NewStates.erase(CS: llvm::unique(R&: NewStates), CE: NewStates.end());
79 sort(C&: TI);
80 TI.erase(CS: llvm::unique(R&: TI), CE: TI.end());
81 unsigned ToId = DfaStates.insert(Entry: NewStates);
82 DfaTransitions.emplace(args: std::pair(FromId, A), args: std::pair(ToId, TI));
83 }
84}
85
86void DfaEmitter::constructDfa() {
87 DfaState Initial(1, /*NFA initial state=*/0);
88 DfaStates.insert(Entry: Initial);
89
90 // Note that UniqueVector starts indices at 1, not zero.
91 unsigned DfaStateId = 1;
92 while (DfaStateId <= DfaStates.size()) {
93 DfaState S = DfaStates[DfaStateId];
94 visitDfaState(DS: S);
95 DfaStateId++;
96 }
97}
98
99void DfaEmitter::emit(StringRef Name, raw_ostream &OS) {
100 constructDfa();
101
102 OS << "// Input NFA has " << NfaStates.size() << " states with "
103 << NumNfaTransitions << " transitions.\n";
104 OS << "// Generated DFA has " << DfaStates.size() << " states with "
105 << DfaTransitions.size() << " transitions.\n\n";
106
107 // Implementation note: We don't bake a simple std::pair<> here as it requires
108 // significantly more effort to parse. A simple test with a large array of
109 // struct-pairs (N=100000) took clang-10 6s to parse. The same array of
110 // std::pair<uint64_t, uint64_t> took 242s. Instead we allow the user to
111 // define the pair type.
112 //
113 // FIXME: It may make sense to emit these as ULEB sequences instead of
114 // pairs of uint64_t.
115 OS << "// A zero-terminated sequence of NFA state transitions. Every DFA\n";
116 OS << "// transition implies a set of NFA transitions. These are referred\n";
117 OS << "// to by index in " << Name << "Transitions[].\n";
118
119 SequenceToOffsetTable<DfaTransitionInfo> Table;
120 for (auto &T : DfaTransitions)
121 Table.add(Seq: T.second.second);
122 Table.layout();
123 OS << "const std::array<NfaStatePair, " << Table.size() << "> " << Name
124 << "TransitionInfo = {{\n";
125 Table.emit(OS, Print: [](raw_ostream &OS, std::pair<uint64_t, uint64_t> P) {
126 OS << "{" << P.first << ", " << P.second << "}";
127 });
128
129 OS << "}};\n\n";
130
131 OS << "// A transition in the generated " << Name << " DFA.\n";
132 OS << "struct " << Name << "Transition {\n";
133 OS << " unsigned FromDfaState; // The transitioned-from DFA state.\n";
134 OS << " ";
135 printActionType(OS);
136 OS << " Action; // The input symbol that causes this transition.\n";
137 OS << " unsigned ToDfaState; // The transitioned-to DFA state.\n";
138 OS << " unsigned InfoIdx; // Start index into " << Name
139 << "TransitionInfo.\n";
140 OS << "};\n\n";
141
142 OS << "// A table of DFA transitions, ordered by {FromDfaState, Action}.\n";
143 OS << "// The initial state is 1, not zero.\n";
144 OS << "const std::array<" << Name << "Transition, " << DfaTransitions.size()
145 << "> " << Name << "Transitions = {{\n";
146 for (auto &KV : DfaTransitions) {
147 dfa_state_type From = KV.first.first;
148 dfa_state_type To = KV.second.first;
149 action_type A = KV.first.second;
150 unsigned InfoIdx = Table.get(Seq: KV.second.second);
151 OS << " {" << From << ", ";
152 printActionValue(A, OS);
153 OS << ", " << To << ", " << InfoIdx << "},\n";
154 }
155 OS << "\n}};\n\n";
156}
157
158void DfaEmitter::printActionType(raw_ostream &OS) { OS << "uint64_t"; }
159
160void DfaEmitter::printActionValue(action_type A, raw_ostream &OS) { OS << A; }
161
162//===----------------------------------------------------------------------===//
163// AutomatonEmitter implementation
164//===----------------------------------------------------------------------===//
165
166namespace {
167
168using Action = std::variant<const Record *, unsigned, std::string>;
169using ActionTuple = std::vector<Action>;
170class Automaton;
171
172class Transition {
173 uint64_t NewState;
174 // The tuple of actions that causes this transition.
175 ActionTuple Actions;
176 // The types of the actions; this is the same across all transitions.
177 SmallVector<std::string, 4> Types;
178
179public:
180 Transition(const Record *R, Automaton *Parent);
181 const ActionTuple &getActions() { return Actions; }
182 SmallVector<std::string, 4> getTypes() { return Types; }
183
184 bool canTransitionFrom(uint64_t State);
185 uint64_t transitionFrom(uint64_t State);
186};
187
188class Automaton {
189 const RecordKeeper &Records;
190 const Record *R;
191 std::vector<Transition> Transitions;
192 /// All possible action tuples, uniqued.
193 UniqueVector<ActionTuple> Actions;
194 /// The fields within each Transition object to find the action symbols.
195 std::vector<StringRef> ActionSymbolFields;
196
197public:
198 Automaton(const RecordKeeper &Records, const Record *R);
199 void emit(raw_ostream &OS);
200
201 ArrayRef<StringRef> getActionSymbolFields() { return ActionSymbolFields; }
202 /// If the type of action A has been overridden (there exists a field
203 /// "TypeOf_A") return that, otherwise return the empty string.
204 StringRef getActionSymbolType(StringRef A);
205};
206
207class AutomatonEmitter {
208 const RecordKeeper &Records;
209
210public:
211 AutomatonEmitter(const RecordKeeper &R) : Records(R) {}
212 void run(raw_ostream &OS);
213};
214
215/// A DfaEmitter implementation that can print our variant action type.
216class CustomDfaEmitter : public DfaEmitter {
217 const UniqueVector<ActionTuple> &Actions;
218 std::string TypeName;
219
220public:
221 CustomDfaEmitter(const UniqueVector<ActionTuple> &Actions, StringRef TypeName)
222 : Actions(Actions), TypeName(TypeName) {}
223
224 void printActionType(raw_ostream &OS) override;
225 void printActionValue(action_type A, raw_ostream &OS) override;
226};
227} // namespace
228
229void AutomatonEmitter::run(raw_ostream &OS) {
230 for (const Record *R : Records.getAllDerivedDefinitions(ClassName: "GenericAutomaton")) {
231 Automaton A(Records, R);
232 OS << "#ifdef GET_" << R->getName() << "_DECL\n";
233 A.emit(OS);
234 OS << "#endif // GET_" << R->getName() << "_DECL\n";
235 }
236}
237
238Automaton::Automaton(const RecordKeeper &Records, const Record *R)
239 : Records(Records), R(R) {
240 LLVM_DEBUG(dbgs() << "Emitting automaton for " << R->getName() << "\n");
241 ActionSymbolFields = R->getValueAsListOfStrings(FieldName: "SymbolFields");
242}
243
244void Automaton::emit(raw_ostream &OS) {
245 StringRef TransitionClass = R->getValueAsString(FieldName: "TransitionClass");
246 for (const Record *T : Records.getAllDerivedDefinitions(ClassName: TransitionClass)) {
247 assert(T->isSubClassOf("Transition"));
248 Transitions.emplace_back(args&: T, args: this);
249 Actions.insert(Entry: Transitions.back().getActions());
250 }
251
252 LLVM_DEBUG(dbgs() << " Action alphabet cardinality: " << Actions.size()
253 << "\n");
254 LLVM_DEBUG(dbgs() << " Each state has " << Transitions.size()
255 << " potential transitions.\n");
256
257 StringRef Name = R->getName();
258
259 CustomDfaEmitter Emitter(Actions, Name.str() + "Action");
260 // Starting from the initial state, build up a list of possible states and
261 // transitions.
262 std::deque<uint64_t> Worklist(1, 0);
263 std::set<uint64_t> SeenStates;
264 unsigned NumTransitions = 0;
265 SeenStates.insert(x: Worklist.front());
266 while (!Worklist.empty()) {
267 uint64_t State = Worklist.front();
268 Worklist.pop_front();
269 for (Transition &T : Transitions) {
270 if (!T.canTransitionFrom(State))
271 continue;
272 uint64_t NewState = T.transitionFrom(State);
273 if (SeenStates.emplace(args&: NewState).second)
274 Worklist.emplace_back(args&: NewState);
275 ++NumTransitions;
276 Emitter.addTransition(From: State, To: NewState, A: Actions.idFor(Entry: T.getActions()));
277 }
278 }
279 LLVM_DEBUG(dbgs() << " NFA automaton has " << SeenStates.size()
280 << " states with " << NumTransitions << " transitions.\n");
281 (void)NumTransitions;
282
283 const auto &ActionTypes = Transitions.back().getTypes();
284 OS << "// The type of an action in the " << Name << " automaton.\n";
285 if (ActionTypes.size() == 1) {
286 OS << "using " << Name << "Action = " << ActionTypes[0] << ";\n";
287 } else {
288 OS << "using " << Name << "Action = std::tuple<" << join(R: ActionTypes, Separator: ", ")
289 << ">;\n";
290 }
291 OS << "\n";
292
293 Emitter.emit(Name, OS);
294}
295
296StringRef Automaton::getActionSymbolType(StringRef A) {
297 Twine Ty = "TypeOf_" + A;
298 if (!R->getValue(Name: Ty.str()))
299 return "";
300 return R->getValueAsString(FieldName: Ty.str());
301}
302
303Transition::Transition(const Record *R, Automaton *Parent) {
304 const BitsInit *NewStateInit = R->getValueAsBitsInit(FieldName: "NewState");
305 assert(NewStateInit->getNumBits() <= sizeof(uint64_t) * 8 &&
306 "State cannot be represented in 64 bits!");
307 NewState = NewStateInit->convertKnownBitsToInt();
308 for (StringRef A : Parent->getActionSymbolFields()) {
309 const RecordVal *SymbolV = R->getValue(Name: A);
310 if (const auto *Ty = dyn_cast<RecordRecTy>(Val: SymbolV->getType())) {
311 Actions.emplace_back(args: R->getValueAsDef(FieldName: A));
312 Types.emplace_back(Args: Ty->getAsString());
313 } else if (isa<IntRecTy>(Val: SymbolV->getType())) {
314 Actions.emplace_back(args: static_cast<unsigned>(R->getValueAsInt(FieldName: A)));
315 Types.emplace_back(Args: "unsigned");
316 } else if (isa<StringRecTy>(Val: SymbolV->getType())) {
317 Actions.emplace_back(args: R->getValueAsString(FieldName: A).str());
318 Types.emplace_back(Args: "std::string");
319 } else {
320 report_fatal_error(reason: "Unhandled symbol type!");
321 }
322
323 StringRef TypeOverride = Parent->getActionSymbolType(A);
324 if (!TypeOverride.empty())
325 Types.back() = TypeOverride.str();
326 }
327}
328
329bool Transition::canTransitionFrom(uint64_t State) {
330 if ((State & NewState) == 0)
331 // The bits we want to set are not set;
332 return true;
333 return false;
334}
335
336uint64_t Transition::transitionFrom(uint64_t State) { return State | NewState; }
337
338void CustomDfaEmitter::printActionType(raw_ostream &OS) { OS << TypeName; }
339
340void CustomDfaEmitter::printActionValue(action_type A, raw_ostream &OS) {
341 const ActionTuple &AT = Actions[A];
342 if (AT.size() > 1)
343 OS << "{";
344 ListSeparator LS;
345 for (const auto &SingleAction : AT) {
346 OS << LS;
347 if (const auto *R = std::get_if<const Record *>(ptr: &SingleAction))
348 OS << (*R)->getName();
349 else if (const auto *S = std::get_if<std::string>(ptr: &SingleAction))
350 OS << '"' << *S << '"';
351 else
352 OS << std::get<unsigned>(v: SingleAction);
353 }
354 if (AT.size() > 1)
355 OS << "}";
356}
357
358static TableGen::Emitter::OptClass<AutomatonEmitter>
359 X("gen-automata", "Generate generic automata");
360