1 //===- SCCPSolver.cpp - SCCP Utility --------------------------- *- C++ -*-===//
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 // \file
10 // This file implements the Sparse Conditional Constant Propagation (SCCP)
11 // utility.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Transforms/Utils/SCCPSolver.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/ValueLattice.h"
19 #include "llvm/IR/InstVisitor.h"
20 #include "llvm/Support/Casting.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/raw_ostream.h"
24 #include <cassert>
25 #include <utility>
26 #include <vector>
27
28 using namespace llvm;
29
30 #define DEBUG_TYPE "sccp"
31
32 // The maximum number of range extensions allowed for operations requiring
33 // widening.
34 static const unsigned MaxNumRangeExtensions = 10;
35
36 /// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions.
getMaxWidenStepsOpts()37 static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() {
38 return ValueLatticeElement::MergeOptions().setMaxWidenSteps(
39 MaxNumRangeExtensions);
40 }
41
42 namespace {
43
44 // Helper to check if \p LV is either a constant or a constant
45 // range with a single element. This should cover exactly the same cases as the
46 // old ValueLatticeElement::isConstant() and is intended to be used in the
47 // transition to ValueLatticeElement.
isConstant(const ValueLatticeElement & LV)48 bool isConstant(const ValueLatticeElement &LV) {
49 return LV.isConstant() ||
50 (LV.isConstantRange() && LV.getConstantRange().isSingleElement());
51 }
52
53 // Helper to check if \p LV is either overdefined or a constant range with more
54 // than a single element. This should cover exactly the same cases as the old
55 // ValueLatticeElement::isOverdefined() and is intended to be used in the
56 // transition to ValueLatticeElement.
isOverdefined(const ValueLatticeElement & LV)57 bool isOverdefined(const ValueLatticeElement &LV) {
58 return !LV.isUnknownOrUndef() && !isConstant(LV);
59 }
60
61 } // namespace
62
63 namespace llvm {
64
65 /// Helper class for SCCPSolver. This implements the instruction visitor and
66 /// holds all the state.
67 class SCCPInstVisitor : public InstVisitor<SCCPInstVisitor> {
68 const DataLayout &DL;
69 std::function<const TargetLibraryInfo &(Function &)> GetTLI;
70 SmallPtrSet<BasicBlock *, 8> BBExecutable; // The BBs that are executable.
71 DenseMap<Value *, ValueLatticeElement>
72 ValueState; // The state each value is in.
73
74 /// StructValueState - This maintains ValueState for values that have
75 /// StructType, for example for formal arguments, calls, insertelement, etc.
76 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState;
77
78 /// GlobalValue - If we are tracking any values for the contents of a global
79 /// variable, we keep a mapping from the constant accessor to the element of
80 /// the global, to the currently known value. If the value becomes
81 /// overdefined, it's entry is simply removed from this map.
82 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals;
83
84 /// TrackedRetVals - If we are tracking arguments into and the return
85 /// value out of a function, it will have an entry in this map, indicating
86 /// what the known return value for the function is.
87 MapVector<Function *, ValueLatticeElement> TrackedRetVals;
88
89 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
90 /// that return multiple values.
91 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement>
92 TrackedMultipleRetVals;
93
94 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
95 /// represented here for efficient lookup.
96 SmallPtrSet<Function *, 16> MRVFunctionsTracked;
97
98 /// A list of functions whose return cannot be modified.
99 SmallPtrSet<Function *, 16> MustPreserveReturnsInFunctions;
100
101 /// TrackingIncomingArguments - This is the set of functions for whose
102 /// arguments we make optimistic assumptions about and try to prove as
103 /// constants.
104 SmallPtrSet<Function *, 16> TrackingIncomingArguments;
105
106 /// The reason for two worklists is that overdefined is the lowest state
107 /// on the lattice, and moving things to overdefined as fast as possible
108 /// makes SCCP converge much faster.
109 ///
110 /// By having a separate worklist, we accomplish this because everything
111 /// possibly overdefined will become overdefined at the soonest possible
112 /// point.
113 SmallVector<Value *, 64> OverdefinedInstWorkList;
114 SmallVector<Value *, 64> InstWorkList;
115
116 // The BasicBlock work list
117 SmallVector<BasicBlock *, 64> BBWorkList;
118
119 /// KnownFeasibleEdges - Entries in this set are edges which have already had
120 /// PHI nodes retriggered.
121 using Edge = std::pair<BasicBlock *, BasicBlock *>;
122 DenseSet<Edge> KnownFeasibleEdges;
123
124 DenseMap<Function *, AnalysisResultsForFn> AnalysisResults;
125 DenseMap<Value *, SmallPtrSet<User *, 2>> AdditionalUsers;
126
127 LLVMContext &Ctx;
128
129 private:
getConstantInt(const ValueLatticeElement & IV) const130 ConstantInt *getConstantInt(const ValueLatticeElement &IV) const {
131 return dyn_cast_or_null<ConstantInt>(getConstant(IV));
132 }
133
134 // pushToWorkList - Helper for markConstant/markOverdefined
135 void pushToWorkList(ValueLatticeElement &IV, Value *V);
136
137 // Helper to push \p V to the worklist, after updating it to \p IV. Also
138 // prints a debug message with the updated value.
139 void pushToWorkListMsg(ValueLatticeElement &IV, Value *V);
140
141 // markConstant - Make a value be marked as "constant". If the value
142 // is not already a constant, add it to the instruction work list so that
143 // the users of the instruction are updated later.
144 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C,
145 bool MayIncludeUndef = false);
146
markConstant(Value * V,Constant * C)147 bool markConstant(Value *V, Constant *C) {
148 assert(!V->getType()->isStructTy() && "structs should use mergeInValue");
149 return markConstant(ValueState[V], V, C);
150 }
151
152 // markOverdefined - Make a value be marked as "overdefined". If the
153 // value is not already overdefined, add it to the overdefined instruction
154 // work list so that the users of the instruction are updated later.
155 bool markOverdefined(ValueLatticeElement &IV, Value *V);
156
157 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV
158 /// changes.
159 bool mergeInValue(ValueLatticeElement &IV, Value *V,
160 ValueLatticeElement MergeWithV,
161 ValueLatticeElement::MergeOptions Opts = {
162 /*MayIncludeUndef=*/false, /*CheckWiden=*/false});
163
mergeInValue(Value * V,ValueLatticeElement MergeWithV,ValueLatticeElement::MergeOptions Opts={ false, false})164 bool mergeInValue(Value *V, ValueLatticeElement MergeWithV,
165 ValueLatticeElement::MergeOptions Opts = {
166 /*MayIncludeUndef=*/false, /*CheckWiden=*/false}) {
167 assert(!V->getType()->isStructTy() &&
168 "non-structs should use markConstant");
169 return mergeInValue(ValueState[V], V, MergeWithV, Opts);
170 }
171
172 /// getValueState - Return the ValueLatticeElement object that corresponds to
173 /// the value. This function handles the case when the value hasn't been seen
174 /// yet by properly seeding constants etc.
getValueState(Value * V)175 ValueLatticeElement &getValueState(Value *V) {
176 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
177
178 auto I = ValueState.insert(std::make_pair(V, ValueLatticeElement()));
179 ValueLatticeElement &LV = I.first->second;
180
181 if (!I.second)
182 return LV; // Common case, already in the map.
183
184 if (auto *C = dyn_cast<Constant>(V))
185 LV.markConstant(C); // Constants are constant
186
187 // All others are unknown by default.
188 return LV;
189 }
190
191 /// getStructValueState - Return the ValueLatticeElement object that
192 /// corresponds to the value/field pair. This function handles the case when
193 /// the value hasn't been seen yet by properly seeding constants etc.
getStructValueState(Value * V,unsigned i)194 ValueLatticeElement &getStructValueState(Value *V, unsigned i) {
195 assert(V->getType()->isStructTy() && "Should use getValueState");
196 assert(i < cast<StructType>(V->getType())->getNumElements() &&
197 "Invalid element #");
198
199 auto I = StructValueState.insert(
200 std::make_pair(std::make_pair(V, i), ValueLatticeElement()));
201 ValueLatticeElement &LV = I.first->second;
202
203 if (!I.second)
204 return LV; // Common case, already in the map.
205
206 if (auto *C = dyn_cast<Constant>(V)) {
207 Constant *Elt = C->getAggregateElement(i);
208
209 if (!Elt)
210 LV.markOverdefined(); // Unknown sort of constant.
211 else
212 LV.markConstant(Elt); // Constants are constant.
213 }
214
215 // All others are underdefined by default.
216 return LV;
217 }
218
219 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
220 /// work list if it is not already executable.
221 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest);
222
223 // getFeasibleSuccessors - Return a vector of booleans to indicate which
224 // successors are reachable from a given terminator instruction.
225 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs);
226
227 // OperandChangedState - This method is invoked on all of the users of an
228 // instruction that was just changed state somehow. Based on this
229 // information, we need to update the specified user of this instruction.
operandChangedState(Instruction * I)230 void operandChangedState(Instruction *I) {
231 if (BBExecutable.count(I->getParent())) // Inst is executable?
232 visit(*I);
233 }
234
235 // Add U as additional user of V.
addAdditionalUser(Value * V,User * U)236 void addAdditionalUser(Value *V, User *U) {
237 auto Iter = AdditionalUsers.insert({V, {}});
238 Iter.first->second.insert(U);
239 }
240
241 // Mark I's users as changed, including AdditionalUsers.
markUsersAsChanged(Value * I)242 void markUsersAsChanged(Value *I) {
243 // Functions include their arguments in the use-list. Changed function
244 // values mean that the result of the function changed. We only need to
245 // update the call sites with the new function result and do not have to
246 // propagate the call arguments.
247 if (isa<Function>(I)) {
248 for (User *U : I->users()) {
249 if (auto *CB = dyn_cast<CallBase>(U))
250 handleCallResult(*CB);
251 }
252 } else {
253 for (User *U : I->users())
254 if (auto *UI = dyn_cast<Instruction>(U))
255 operandChangedState(UI);
256 }
257
258 auto Iter = AdditionalUsers.find(I);
259 if (Iter != AdditionalUsers.end()) {
260 // Copy additional users before notifying them of changes, because new
261 // users may be added, potentially invalidating the iterator.
262 SmallVector<Instruction *, 2> ToNotify;
263 for (User *U : Iter->second)
264 if (auto *UI = dyn_cast<Instruction>(U))
265 ToNotify.push_back(UI);
266 for (Instruction *UI : ToNotify)
267 operandChangedState(UI);
268 }
269 }
270 void handleCallOverdefined(CallBase &CB);
271 void handleCallResult(CallBase &CB);
272 void handleCallArguments(CallBase &CB);
273
274 private:
275 friend class InstVisitor<SCCPInstVisitor>;
276
277 // visit implementations - Something changed in this instruction. Either an
278 // operand made a transition, or the instruction is newly executable. Change
279 // the value type of I to reflect these changes if appropriate.
280 void visitPHINode(PHINode &I);
281
282 // Terminators
283
284 void visitReturnInst(ReturnInst &I);
285 void visitTerminator(Instruction &TI);
286
287 void visitCastInst(CastInst &I);
288 void visitSelectInst(SelectInst &I);
289 void visitUnaryOperator(Instruction &I);
290 void visitBinaryOperator(Instruction &I);
291 void visitCmpInst(CmpInst &I);
292 void visitExtractValueInst(ExtractValueInst &EVI);
293 void visitInsertValueInst(InsertValueInst &IVI);
294
visitCatchSwitchInst(CatchSwitchInst & CPI)295 void visitCatchSwitchInst(CatchSwitchInst &CPI) {
296 markOverdefined(&CPI);
297 visitTerminator(CPI);
298 }
299
300 // Instructions that cannot be folded away.
301
302 void visitStoreInst(StoreInst &I);
303 void visitLoadInst(LoadInst &I);
304 void visitGetElementPtrInst(GetElementPtrInst &I);
305
visitInvokeInst(InvokeInst & II)306 void visitInvokeInst(InvokeInst &II) {
307 visitCallBase(II);
308 visitTerminator(II);
309 }
310
visitCallBrInst(CallBrInst & CBI)311 void visitCallBrInst(CallBrInst &CBI) {
312 visitCallBase(CBI);
313 visitTerminator(CBI);
314 }
315
316 void visitCallBase(CallBase &CB);
visitResumeInst(ResumeInst & I)317 void visitResumeInst(ResumeInst &I) { /*returns void*/
318 }
visitUnreachableInst(UnreachableInst & I)319 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/
320 }
visitFenceInst(FenceInst & I)321 void visitFenceInst(FenceInst &I) { /*returns void*/
322 }
323
324 void visitInstruction(Instruction &I);
325
326 public:
addAnalysis(Function & F,AnalysisResultsForFn A)327 void addAnalysis(Function &F, AnalysisResultsForFn A) {
328 AnalysisResults.insert({&F, std::move(A)});
329 }
330
visitCallInst(CallInst & I)331 void visitCallInst(CallInst &I) { visitCallBase(I); }
332
333 bool markBlockExecutable(BasicBlock *BB);
334
getPredicateInfoFor(Instruction * I)335 const PredicateBase *getPredicateInfoFor(Instruction *I) {
336 auto A = AnalysisResults.find(I->getParent()->getParent());
337 if (A == AnalysisResults.end())
338 return nullptr;
339 return A->second.PredInfo->getPredicateInfoFor(I);
340 }
341
getDTU(Function & F)342 DomTreeUpdater getDTU(Function &F) {
343 auto A = AnalysisResults.find(&F);
344 assert(A != AnalysisResults.end() && "Need analysis results for function.");
345 return {A->second.DT, A->second.PDT, DomTreeUpdater::UpdateStrategy::Lazy};
346 }
347
SCCPInstVisitor(const DataLayout & DL,std::function<const TargetLibraryInfo & (Function &)> GetTLI,LLVMContext & Ctx)348 SCCPInstVisitor(const DataLayout &DL,
349 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
350 LLVMContext &Ctx)
351 : DL(DL), GetTLI(GetTLI), Ctx(Ctx) {}
352
trackValueOfGlobalVariable(GlobalVariable * GV)353 void trackValueOfGlobalVariable(GlobalVariable *GV) {
354 // We only track the contents of scalar globals.
355 if (GV->getValueType()->isSingleValueType()) {
356 ValueLatticeElement &IV = TrackedGlobals[GV];
357 IV.markConstant(GV->getInitializer());
358 }
359 }
360
addTrackedFunction(Function * F)361 void addTrackedFunction(Function *F) {
362 // Add an entry, F -> undef.
363 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) {
364 MRVFunctionsTracked.insert(F);
365 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
366 TrackedMultipleRetVals.insert(
367 std::make_pair(std::make_pair(F, i), ValueLatticeElement()));
368 } else if (!F->getReturnType()->isVoidTy())
369 TrackedRetVals.insert(std::make_pair(F, ValueLatticeElement()));
370 }
371
addToMustPreserveReturnsInFunctions(Function * F)372 void addToMustPreserveReturnsInFunctions(Function *F) {
373 MustPreserveReturnsInFunctions.insert(F);
374 }
375
mustPreserveReturn(Function * F)376 bool mustPreserveReturn(Function *F) {
377 return MustPreserveReturnsInFunctions.count(F);
378 }
379
addArgumentTrackedFunction(Function * F)380 void addArgumentTrackedFunction(Function *F) {
381 TrackingIncomingArguments.insert(F);
382 }
383
isArgumentTrackedFunction(Function * F)384 bool isArgumentTrackedFunction(Function *F) {
385 return TrackingIncomingArguments.count(F);
386 }
387
388 void solve();
389
390 bool resolvedUndefsIn(Function &F);
391
isBlockExecutable(BasicBlock * BB) const392 bool isBlockExecutable(BasicBlock *BB) const {
393 return BBExecutable.count(BB);
394 }
395
396 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const;
397
getStructLatticeValueFor(Value * V) const398 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const {
399 std::vector<ValueLatticeElement> StructValues;
400 auto *STy = dyn_cast<StructType>(V->getType());
401 assert(STy && "getStructLatticeValueFor() can be called only on structs");
402 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
403 auto I = StructValueState.find(std::make_pair(V, i));
404 assert(I != StructValueState.end() && "Value not in valuemap!");
405 StructValues.push_back(I->second);
406 }
407 return StructValues;
408 }
409
removeLatticeValueFor(Value * V)410 void removeLatticeValueFor(Value *V) { ValueState.erase(V); }
411
getLatticeValueFor(Value * V) const412 const ValueLatticeElement &getLatticeValueFor(Value *V) const {
413 assert(!V->getType()->isStructTy() &&
414 "Should use getStructLatticeValueFor");
415 DenseMap<Value *, ValueLatticeElement>::const_iterator I =
416 ValueState.find(V);
417 assert(I != ValueState.end() &&
418 "V not found in ValueState nor Paramstate map!");
419 return I->second;
420 }
421
getTrackedRetVals()422 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() {
423 return TrackedRetVals;
424 }
425
getTrackedGlobals()426 const DenseMap<GlobalVariable *, ValueLatticeElement> &getTrackedGlobals() {
427 return TrackedGlobals;
428 }
429
getMRVFunctionsTracked()430 const SmallPtrSet<Function *, 16> getMRVFunctionsTracked() {
431 return MRVFunctionsTracked;
432 }
433
markOverdefined(Value * V)434 void markOverdefined(Value *V) {
435 if (auto *STy = dyn_cast<StructType>(V->getType()))
436 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
437 markOverdefined(getStructValueState(V, i), V);
438 else
439 markOverdefined(ValueState[V], V);
440 }
441
442 bool isStructLatticeConstant(Function *F, StructType *STy);
443
444 Constant *getConstant(const ValueLatticeElement &LV) const;
445
getArgumentTrackedFunctions()446 SmallPtrSetImpl<Function *> &getArgumentTrackedFunctions() {
447 return TrackingIncomingArguments;
448 }
449
450 void markArgInFuncSpecialization(Function *F,
451 const SmallVectorImpl<ArgInfo> &Args);
452
markFunctionUnreachable(Function * F)453 void markFunctionUnreachable(Function *F) {
454 for (auto &BB : *F)
455 BBExecutable.erase(&BB);
456 }
457 };
458
459 } // namespace llvm
460
markBlockExecutable(BasicBlock * BB)461 bool SCCPInstVisitor::markBlockExecutable(BasicBlock *BB) {
462 if (!BBExecutable.insert(BB).second)
463 return false;
464 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n');
465 BBWorkList.push_back(BB); // Add the block to the work list!
466 return true;
467 }
468
pushToWorkList(ValueLatticeElement & IV,Value * V)469 void SCCPInstVisitor::pushToWorkList(ValueLatticeElement &IV, Value *V) {
470 if (IV.isOverdefined())
471 return OverdefinedInstWorkList.push_back(V);
472 InstWorkList.push_back(V);
473 }
474
pushToWorkListMsg(ValueLatticeElement & IV,Value * V)475 void SCCPInstVisitor::pushToWorkListMsg(ValueLatticeElement &IV, Value *V) {
476 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n');
477 pushToWorkList(IV, V);
478 }
479
markConstant(ValueLatticeElement & IV,Value * V,Constant * C,bool MayIncludeUndef)480 bool SCCPInstVisitor::markConstant(ValueLatticeElement &IV, Value *V,
481 Constant *C, bool MayIncludeUndef) {
482 if (!IV.markConstant(C, MayIncludeUndef))
483 return false;
484 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
485 pushToWorkList(IV, V);
486 return true;
487 }
488
markOverdefined(ValueLatticeElement & IV,Value * V)489 bool SCCPInstVisitor::markOverdefined(ValueLatticeElement &IV, Value *V) {
490 if (!IV.markOverdefined())
491 return false;
492
493 LLVM_DEBUG(dbgs() << "markOverdefined: ";
494 if (auto *F = dyn_cast<Function>(V)) dbgs()
495 << "Function '" << F->getName() << "'\n";
496 else dbgs() << *V << '\n');
497 // Only instructions go on the work list
498 pushToWorkList(IV, V);
499 return true;
500 }
501
isStructLatticeConstant(Function * F,StructType * STy)502 bool SCCPInstVisitor::isStructLatticeConstant(Function *F, StructType *STy) {
503 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
504 const auto &It = TrackedMultipleRetVals.find(std::make_pair(F, i));
505 assert(It != TrackedMultipleRetVals.end());
506 ValueLatticeElement LV = It->second;
507 if (!isConstant(LV))
508 return false;
509 }
510 return true;
511 }
512
getConstant(const ValueLatticeElement & LV) const513 Constant *SCCPInstVisitor::getConstant(const ValueLatticeElement &LV) const {
514 if (LV.isConstant())
515 return LV.getConstant();
516
517 if (LV.isConstantRange()) {
518 const auto &CR = LV.getConstantRange();
519 if (CR.getSingleElement())
520 return ConstantInt::get(Ctx, *CR.getSingleElement());
521 }
522 return nullptr;
523 }
524
markArgInFuncSpecialization(Function * F,const SmallVectorImpl<ArgInfo> & Args)525 void SCCPInstVisitor::markArgInFuncSpecialization(
526 Function *F, const SmallVectorImpl<ArgInfo> &Args) {
527 assert(!Args.empty() && "Specialization without arguments");
528 assert(F->arg_size() == Args[0].Formal->getParent()->arg_size() &&
529 "Functions should have the same number of arguments");
530
531 auto Iter = Args.begin();
532 Argument *NewArg = F->arg_begin();
533 Argument *OldArg = Args[0].Formal->getParent()->arg_begin();
534 for (auto End = F->arg_end(); NewArg != End; ++NewArg, ++OldArg) {
535
536 LLVM_DEBUG(dbgs() << "SCCP: Marking argument "
537 << NewArg->getNameOrAsOperand() << "\n");
538
539 if (Iter != Args.end() && OldArg == Iter->Formal) {
540 // Mark the argument constants in the new function.
541 markConstant(NewArg, Iter->Actual);
542 ++Iter;
543 } else if (ValueState.count(OldArg)) {
544 // For the remaining arguments in the new function, copy the lattice state
545 // over from the old function.
546 //
547 // Note: This previously looked like this:
548 // ValueState[NewArg] = ValueState[OldArg];
549 // This is incorrect because the DenseMap class may resize the underlying
550 // memory when inserting `NewArg`, which will invalidate the reference to
551 // `OldArg`. Instead, we make sure `NewArg` exists before setting it.
552 auto &NewValue = ValueState[NewArg];
553 NewValue = ValueState[OldArg];
554 pushToWorkList(NewValue, NewArg);
555 }
556 }
557 }
558
visitInstruction(Instruction & I)559 void SCCPInstVisitor::visitInstruction(Instruction &I) {
560 // All the instructions we don't do any special handling for just
561 // go to overdefined.
562 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n');
563 markOverdefined(&I);
564 }
565
mergeInValue(ValueLatticeElement & IV,Value * V,ValueLatticeElement MergeWithV,ValueLatticeElement::MergeOptions Opts)566 bool SCCPInstVisitor::mergeInValue(ValueLatticeElement &IV, Value *V,
567 ValueLatticeElement MergeWithV,
568 ValueLatticeElement::MergeOptions Opts) {
569 if (IV.mergeIn(MergeWithV, Opts)) {
570 pushToWorkList(IV, V);
571 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : "
572 << IV << "\n");
573 return true;
574 }
575 return false;
576 }
577
markEdgeExecutable(BasicBlock * Source,BasicBlock * Dest)578 bool SCCPInstVisitor::markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
579 if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
580 return false; // This edge is already known to be executable!
581
582 if (!markBlockExecutable(Dest)) {
583 // If the destination is already executable, we just made an *edge*
584 // feasible that wasn't before. Revisit the PHI nodes in the block
585 // because they have potentially new operands.
586 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
587 << " -> " << Dest->getName() << '\n');
588
589 for (PHINode &PN : Dest->phis())
590 visitPHINode(PN);
591 }
592 return true;
593 }
594
595 // getFeasibleSuccessors - Return a vector of booleans to indicate which
596 // successors are reachable from a given terminator instruction.
getFeasibleSuccessors(Instruction & TI,SmallVectorImpl<bool> & Succs)597 void SCCPInstVisitor::getFeasibleSuccessors(Instruction &TI,
598 SmallVectorImpl<bool> &Succs) {
599 Succs.resize(TI.getNumSuccessors());
600 if (auto *BI = dyn_cast<BranchInst>(&TI)) {
601 if (BI->isUnconditional()) {
602 Succs[0] = true;
603 return;
604 }
605
606 ValueLatticeElement BCValue = getValueState(BI->getCondition());
607 ConstantInt *CI = getConstantInt(BCValue);
608 if (!CI) {
609 // Overdefined condition variables, and branches on unfoldable constant
610 // conditions, mean the branch could go either way.
611 if (!BCValue.isUnknownOrUndef())
612 Succs[0] = Succs[1] = true;
613 return;
614 }
615
616 // Constant condition variables mean the branch can only go a single way.
617 Succs[CI->isZero()] = true;
618 return;
619 }
620
621 // Unwinding instructions successors are always executable.
622 if (TI.isExceptionalTerminator()) {
623 Succs.assign(TI.getNumSuccessors(), true);
624 return;
625 }
626
627 if (auto *SI = dyn_cast<SwitchInst>(&TI)) {
628 if (!SI->getNumCases()) {
629 Succs[0] = true;
630 return;
631 }
632 const ValueLatticeElement &SCValue = getValueState(SI->getCondition());
633 if (ConstantInt *CI = getConstantInt(SCValue)) {
634 Succs[SI->findCaseValue(CI)->getSuccessorIndex()] = true;
635 return;
636 }
637
638 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM
639 // is ready.
640 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) {
641 const ConstantRange &Range = SCValue.getConstantRange();
642 for (const auto &Case : SI->cases()) {
643 const APInt &CaseValue = Case.getCaseValue()->getValue();
644 if (Range.contains(CaseValue))
645 Succs[Case.getSuccessorIndex()] = true;
646 }
647
648 // TODO: Determine whether default case is reachable.
649 Succs[SI->case_default()->getSuccessorIndex()] = true;
650 return;
651 }
652
653 // Overdefined or unknown condition? All destinations are executable!
654 if (!SCValue.isUnknownOrUndef())
655 Succs.assign(TI.getNumSuccessors(), true);
656 return;
657 }
658
659 // In case of indirect branch and its address is a blockaddress, we mark
660 // the target as executable.
661 if (auto *IBR = dyn_cast<IndirectBrInst>(&TI)) {
662 // Casts are folded by visitCastInst.
663 ValueLatticeElement IBRValue = getValueState(IBR->getAddress());
664 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(getConstant(IBRValue));
665 if (!Addr) { // Overdefined or unknown condition?
666 // All destinations are executable!
667 if (!IBRValue.isUnknownOrUndef())
668 Succs.assign(TI.getNumSuccessors(), true);
669 return;
670 }
671
672 BasicBlock *T = Addr->getBasicBlock();
673 assert(Addr->getFunction() == T->getParent() &&
674 "Block address of a different function ?");
675 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) {
676 // This is the target.
677 if (IBR->getDestination(i) == T) {
678 Succs[i] = true;
679 return;
680 }
681 }
682
683 // If we didn't find our destination in the IBR successor list, then we
684 // have undefined behavior. Its ok to assume no successor is executable.
685 return;
686 }
687
688 // In case of callbr, we pessimistically assume that all successors are
689 // feasible.
690 if (isa<CallBrInst>(&TI)) {
691 Succs.assign(TI.getNumSuccessors(), true);
692 return;
693 }
694
695 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n');
696 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
697 }
698
699 // isEdgeFeasible - Return true if the control flow edge from the 'From' basic
700 // block to the 'To' basic block is currently feasible.
isEdgeFeasible(BasicBlock * From,BasicBlock * To) const701 bool SCCPInstVisitor::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const {
702 // Check if we've called markEdgeExecutable on the edge yet. (We could
703 // be more aggressive and try to consider edges which haven't been marked
704 // yet, but there isn't any need.)
705 return KnownFeasibleEdges.count(Edge(From, To));
706 }
707
708 // visit Implementations - Something changed in this instruction, either an
709 // operand made a transition, or the instruction is newly executable. Change
710 // the value type of I to reflect these changes if appropriate. This method
711 // makes sure to do the following actions:
712 //
713 // 1. If a phi node merges two constants in, and has conflicting value coming
714 // from different branches, or if the PHI node merges in an overdefined
715 // value, then the PHI node becomes overdefined.
716 // 2. If a phi node merges only constants in, and they all agree on value, the
717 // PHI node becomes a constant value equal to that.
718 // 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
719 // 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
720 // 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
721 // 6. If a conditional branch has a value that is constant, make the selected
722 // destination executable
723 // 7. If a conditional branch has a value that is overdefined, make all
724 // successors executable.
visitPHINode(PHINode & PN)725 void SCCPInstVisitor::visitPHINode(PHINode &PN) {
726 // If this PN returns a struct, just mark the result overdefined.
727 // TODO: We could do a lot better than this if code actually uses this.
728 if (PN.getType()->isStructTy())
729 return (void)markOverdefined(&PN);
730
731 if (getValueState(&PN).isOverdefined())
732 return; // Quick exit
733
734 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
735 // and slow us down a lot. Just mark them overdefined.
736 if (PN.getNumIncomingValues() > 64)
737 return (void)markOverdefined(&PN);
738
739 unsigned NumActiveIncoming = 0;
740
741 // Look at all of the executable operands of the PHI node. If any of them
742 // are overdefined, the PHI becomes overdefined as well. If they are all
743 // constant, and they agree with each other, the PHI becomes the identical
744 // constant. If they are constant and don't agree, the PHI is a constant
745 // range. If there are no executable operands, the PHI remains unknown.
746 ValueLatticeElement PhiState = getValueState(&PN);
747 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
748 if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent()))
749 continue;
750
751 ValueLatticeElement IV = getValueState(PN.getIncomingValue(i));
752 PhiState.mergeIn(IV);
753 NumActiveIncoming++;
754 if (PhiState.isOverdefined())
755 break;
756 }
757
758 // We allow up to 1 range extension per active incoming value and one
759 // additional extension. Note that we manually adjust the number of range
760 // extensions to match the number of active incoming values. This helps to
761 // limit multiple extensions caused by the same incoming value, if other
762 // incoming values are equal.
763 mergeInValue(&PN, PhiState,
764 ValueLatticeElement::MergeOptions().setMaxWidenSteps(
765 NumActiveIncoming + 1));
766 ValueLatticeElement &PhiStateRef = getValueState(&PN);
767 PhiStateRef.setNumRangeExtensions(
768 std::max(NumActiveIncoming, PhiStateRef.getNumRangeExtensions()));
769 }
770
visitReturnInst(ReturnInst & I)771 void SCCPInstVisitor::visitReturnInst(ReturnInst &I) {
772 if (I.getNumOperands() == 0)
773 return; // ret void
774
775 Function *F = I.getParent()->getParent();
776 Value *ResultOp = I.getOperand(0);
777
778 // If we are tracking the return value of this function, merge it in.
779 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
780 auto TFRVI = TrackedRetVals.find(F);
781 if (TFRVI != TrackedRetVals.end()) {
782 mergeInValue(TFRVI->second, F, getValueState(ResultOp));
783 return;
784 }
785 }
786
787 // Handle functions that return multiple values.
788 if (!TrackedMultipleRetVals.empty()) {
789 if (auto *STy = dyn_cast<StructType>(ResultOp->getType()))
790 if (MRVFunctionsTracked.count(F))
791 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
792 mergeInValue(TrackedMultipleRetVals[std::make_pair(F, i)], F,
793 getStructValueState(ResultOp, i));
794 }
795 }
796
visitTerminator(Instruction & TI)797 void SCCPInstVisitor::visitTerminator(Instruction &TI) {
798 SmallVector<bool, 16> SuccFeasible;
799 getFeasibleSuccessors(TI, SuccFeasible);
800
801 BasicBlock *BB = TI.getParent();
802
803 // Mark all feasible successors executable.
804 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
805 if (SuccFeasible[i])
806 markEdgeExecutable(BB, TI.getSuccessor(i));
807 }
808
visitCastInst(CastInst & I)809 void SCCPInstVisitor::visitCastInst(CastInst &I) {
810 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would
811 // discover a concrete value later.
812 if (ValueState[&I].isOverdefined())
813 return;
814
815 ValueLatticeElement OpSt = getValueState(I.getOperand(0));
816 if (OpSt.isUnknownOrUndef())
817 return;
818
819 if (Constant *OpC = getConstant(OpSt)) {
820 // Fold the constant as we build.
821 Constant *C = ConstantFoldCastOperand(I.getOpcode(), OpC, I.getType(), DL);
822 markConstant(&I, C);
823 } else if (I.getDestTy()->isIntegerTy()) {
824 auto &LV = getValueState(&I);
825 ConstantRange OpRange =
826 OpSt.isConstantRange()
827 ? OpSt.getConstantRange()
828 : ConstantRange::getFull(
829 I.getOperand(0)->getType()->getScalarSizeInBits());
830
831 Type *DestTy = I.getDestTy();
832 // Vectors where all elements have the same known constant range are treated
833 // as a single constant range in the lattice. When bitcasting such vectors,
834 // there is a mis-match between the width of the lattice value (single
835 // constant range) and the original operands (vector). Go to overdefined in
836 // that case.
837 if (I.getOpcode() == Instruction::BitCast &&
838 I.getOperand(0)->getType()->isVectorTy() &&
839 OpRange.getBitWidth() < DL.getTypeSizeInBits(DestTy))
840 return (void)markOverdefined(&I);
841
842 ConstantRange Res =
843 OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy));
844 mergeInValue(LV, &I, ValueLatticeElement::getRange(Res));
845 } else
846 markOverdefined(&I);
847 }
848
visitExtractValueInst(ExtractValueInst & EVI)849 void SCCPInstVisitor::visitExtractValueInst(ExtractValueInst &EVI) {
850 // If this returns a struct, mark all elements over defined, we don't track
851 // structs in structs.
852 if (EVI.getType()->isStructTy())
853 return (void)markOverdefined(&EVI);
854
855 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
856 // discover a concrete value later.
857 if (ValueState[&EVI].isOverdefined())
858 return (void)markOverdefined(&EVI);
859
860 // If this is extracting from more than one level of struct, we don't know.
861 if (EVI.getNumIndices() != 1)
862 return (void)markOverdefined(&EVI);
863
864 Value *AggVal = EVI.getAggregateOperand();
865 if (AggVal->getType()->isStructTy()) {
866 unsigned i = *EVI.idx_begin();
867 ValueLatticeElement EltVal = getStructValueState(AggVal, i);
868 mergeInValue(getValueState(&EVI), &EVI, EltVal);
869 } else {
870 // Otherwise, must be extracting from an array.
871 return (void)markOverdefined(&EVI);
872 }
873 }
874
visitInsertValueInst(InsertValueInst & IVI)875 void SCCPInstVisitor::visitInsertValueInst(InsertValueInst &IVI) {
876 auto *STy = dyn_cast<StructType>(IVI.getType());
877 if (!STy)
878 return (void)markOverdefined(&IVI);
879
880 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
881 // discover a concrete value later.
882 if (isOverdefined(ValueState[&IVI]))
883 return (void)markOverdefined(&IVI);
884
885 // If this has more than one index, we can't handle it, drive all results to
886 // undef.
887 if (IVI.getNumIndices() != 1)
888 return (void)markOverdefined(&IVI);
889
890 Value *Aggr = IVI.getAggregateOperand();
891 unsigned Idx = *IVI.idx_begin();
892
893 // Compute the result based on what we're inserting.
894 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
895 // This passes through all values that aren't the inserted element.
896 if (i != Idx) {
897 ValueLatticeElement EltVal = getStructValueState(Aggr, i);
898 mergeInValue(getStructValueState(&IVI, i), &IVI, EltVal);
899 continue;
900 }
901
902 Value *Val = IVI.getInsertedValueOperand();
903 if (Val->getType()->isStructTy())
904 // We don't track structs in structs.
905 markOverdefined(getStructValueState(&IVI, i), &IVI);
906 else {
907 ValueLatticeElement InVal = getValueState(Val);
908 mergeInValue(getStructValueState(&IVI, i), &IVI, InVal);
909 }
910 }
911 }
912
visitSelectInst(SelectInst & I)913 void SCCPInstVisitor::visitSelectInst(SelectInst &I) {
914 // If this select returns a struct, just mark the result overdefined.
915 // TODO: We could do a lot better than this if code actually uses this.
916 if (I.getType()->isStructTy())
917 return (void)markOverdefined(&I);
918
919 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
920 // discover a concrete value later.
921 if (ValueState[&I].isOverdefined())
922 return (void)markOverdefined(&I);
923
924 ValueLatticeElement CondValue = getValueState(I.getCondition());
925 if (CondValue.isUnknownOrUndef())
926 return;
927
928 if (ConstantInt *CondCB = getConstantInt(CondValue)) {
929 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
930 mergeInValue(&I, getValueState(OpVal));
931 return;
932 }
933
934 // Otherwise, the condition is overdefined or a constant we can't evaluate.
935 // See if we can produce something better than overdefined based on the T/F
936 // value.
937 ValueLatticeElement TVal = getValueState(I.getTrueValue());
938 ValueLatticeElement FVal = getValueState(I.getFalseValue());
939
940 bool Changed = ValueState[&I].mergeIn(TVal);
941 Changed |= ValueState[&I].mergeIn(FVal);
942 if (Changed)
943 pushToWorkListMsg(ValueState[&I], &I);
944 }
945
946 // Handle Unary Operators.
visitUnaryOperator(Instruction & I)947 void SCCPInstVisitor::visitUnaryOperator(Instruction &I) {
948 ValueLatticeElement V0State = getValueState(I.getOperand(0));
949
950 ValueLatticeElement &IV = ValueState[&I];
951 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
952 // discover a concrete value later.
953 if (isOverdefined(IV))
954 return (void)markOverdefined(&I);
955
956 // If something is unknown/undef, wait for it to resolve.
957 if (V0State.isUnknownOrUndef())
958 return;
959
960 if (isConstant(V0State))
961 if (Constant *C = ConstantFoldUnaryOpOperand(I.getOpcode(),
962 getConstant(V0State), DL))
963 return (void)markConstant(IV, &I, C);
964
965 markOverdefined(&I);
966 }
967
968 // Handle Binary Operators.
visitBinaryOperator(Instruction & I)969 void SCCPInstVisitor::visitBinaryOperator(Instruction &I) {
970 ValueLatticeElement V1State = getValueState(I.getOperand(0));
971 ValueLatticeElement V2State = getValueState(I.getOperand(1));
972
973 ValueLatticeElement &IV = ValueState[&I];
974 if (IV.isOverdefined())
975 return;
976
977 // If something is undef, wait for it to resolve.
978 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef())
979 return;
980
981 if (V1State.isOverdefined() && V2State.isOverdefined())
982 return (void)markOverdefined(&I);
983
984 // If either of the operands is a constant, try to fold it to a constant.
985 // TODO: Use information from notconstant better.
986 if ((V1State.isConstant() || V2State.isConstant())) {
987 Value *V1 = isConstant(V1State) ? getConstant(V1State) : I.getOperand(0);
988 Value *V2 = isConstant(V2State) ? getConstant(V2State) : I.getOperand(1);
989 Value *R = simplifyBinOp(I.getOpcode(), V1, V2, SimplifyQuery(DL));
990 auto *C = dyn_cast_or_null<Constant>(R);
991 if (C) {
992 // Conservatively assume that the result may be based on operands that may
993 // be undef. Note that we use mergeInValue to combine the constant with
994 // the existing lattice value for I, as different constants might be found
995 // after one of the operands go to overdefined, e.g. due to one operand
996 // being a special floating value.
997 ValueLatticeElement NewV;
998 NewV.markConstant(C, /*MayIncludeUndef=*/true);
999 return (void)mergeInValue(&I, NewV);
1000 }
1001 }
1002
1003 // Only use ranges for binary operators on integers.
1004 if (!I.getType()->isIntegerTy())
1005 return markOverdefined(&I);
1006
1007 // Try to simplify to a constant range.
1008 ConstantRange A = ConstantRange::getFull(I.getType()->getScalarSizeInBits());
1009 ConstantRange B = ConstantRange::getFull(I.getType()->getScalarSizeInBits());
1010 if (V1State.isConstantRange())
1011 A = V1State.getConstantRange();
1012 if (V2State.isConstantRange())
1013 B = V2State.getConstantRange();
1014
1015 ConstantRange R = A.binaryOp(cast<BinaryOperator>(&I)->getOpcode(), B);
1016 mergeInValue(&I, ValueLatticeElement::getRange(R));
1017
1018 // TODO: Currently we do not exploit special values that produce something
1019 // better than overdefined with an overdefined operand for vector or floating
1020 // point types, like and <4 x i32> overdefined, zeroinitializer.
1021 }
1022
1023 // Handle ICmpInst instruction.
visitCmpInst(CmpInst & I)1024 void SCCPInstVisitor::visitCmpInst(CmpInst &I) {
1025 // Do not cache this lookup, getValueState calls later in the function might
1026 // invalidate the reference.
1027 if (isOverdefined(ValueState[&I]))
1028 return (void)markOverdefined(&I);
1029
1030 Value *Op1 = I.getOperand(0);
1031 Value *Op2 = I.getOperand(1);
1032
1033 // For parameters, use ParamState which includes constant range info if
1034 // available.
1035 auto V1State = getValueState(Op1);
1036 auto V2State = getValueState(Op2);
1037
1038 Constant *C = V1State.getCompare(I.getPredicate(), I.getType(), V2State);
1039 if (C) {
1040 // TODO: getCompare() currently has incorrect handling for unknown/undef.
1041 if (isa<UndefValue>(C))
1042 return;
1043 ValueLatticeElement CV;
1044 CV.markConstant(C);
1045 mergeInValue(&I, CV);
1046 return;
1047 }
1048
1049 // If operands are still unknown, wait for it to resolve.
1050 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) &&
1051 !isConstant(ValueState[&I]))
1052 return;
1053
1054 markOverdefined(&I);
1055 }
1056
1057 // Handle getelementptr instructions. If all operands are constants then we
1058 // can turn this into a getelementptr ConstantExpr.
visitGetElementPtrInst(GetElementPtrInst & I)1059 void SCCPInstVisitor::visitGetElementPtrInst(GetElementPtrInst &I) {
1060 if (isOverdefined(ValueState[&I]))
1061 return (void)markOverdefined(&I);
1062
1063 SmallVector<Constant *, 8> Operands;
1064 Operands.reserve(I.getNumOperands());
1065
1066 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1067 ValueLatticeElement State = getValueState(I.getOperand(i));
1068 if (State.isUnknownOrUndef())
1069 return; // Operands are not resolved yet.
1070
1071 if (isOverdefined(State))
1072 return (void)markOverdefined(&I);
1073
1074 if (Constant *C = getConstant(State)) {
1075 Operands.push_back(C);
1076 continue;
1077 }
1078
1079 return (void)markOverdefined(&I);
1080 }
1081
1082 Constant *Ptr = Operands[0];
1083 auto Indices = makeArrayRef(Operands.begin() + 1, Operands.end());
1084 Constant *C =
1085 ConstantExpr::getGetElementPtr(I.getSourceElementType(), Ptr, Indices);
1086 markConstant(&I, C);
1087 }
1088
visitStoreInst(StoreInst & SI)1089 void SCCPInstVisitor::visitStoreInst(StoreInst &SI) {
1090 // If this store is of a struct, ignore it.
1091 if (SI.getOperand(0)->getType()->isStructTy())
1092 return;
1093
1094 if (TrackedGlobals.empty() || !isa<GlobalVariable>(SI.getOperand(1)))
1095 return;
1096
1097 GlobalVariable *GV = cast<GlobalVariable>(SI.getOperand(1));
1098 auto I = TrackedGlobals.find(GV);
1099 if (I == TrackedGlobals.end())
1100 return;
1101
1102 // Get the value we are storing into the global, then merge it.
1103 mergeInValue(I->second, GV, getValueState(SI.getOperand(0)),
1104 ValueLatticeElement::MergeOptions().setCheckWiden(false));
1105 if (I->second.isOverdefined())
1106 TrackedGlobals.erase(I); // No need to keep tracking this!
1107 }
1108
getValueFromMetadata(const Instruction * I)1109 static ValueLatticeElement getValueFromMetadata(const Instruction *I) {
1110 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range))
1111 if (I->getType()->isIntegerTy())
1112 return ValueLatticeElement::getRange(
1113 getConstantRangeFromMetadata(*Ranges));
1114 if (I->hasMetadata(LLVMContext::MD_nonnull))
1115 return ValueLatticeElement::getNot(
1116 ConstantPointerNull::get(cast<PointerType>(I->getType())));
1117 return ValueLatticeElement::getOverdefined();
1118 }
1119
1120 // Handle load instructions. If the operand is a constant pointer to a constant
1121 // global, we can replace the load with the loaded constant value!
visitLoadInst(LoadInst & I)1122 void SCCPInstVisitor::visitLoadInst(LoadInst &I) {
1123 // If this load is of a struct or the load is volatile, just mark the result
1124 // as overdefined.
1125 if (I.getType()->isStructTy() || I.isVolatile())
1126 return (void)markOverdefined(&I);
1127
1128 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1129 // discover a concrete value later.
1130 if (ValueState[&I].isOverdefined())
1131 return (void)markOverdefined(&I);
1132
1133 ValueLatticeElement PtrVal = getValueState(I.getOperand(0));
1134 if (PtrVal.isUnknownOrUndef())
1135 return; // The pointer is not resolved yet!
1136
1137 ValueLatticeElement &IV = ValueState[&I];
1138
1139 if (isConstant(PtrVal)) {
1140 Constant *Ptr = getConstant(PtrVal);
1141
1142 // load null is undefined.
1143 if (isa<ConstantPointerNull>(Ptr)) {
1144 if (NullPointerIsDefined(I.getFunction(), I.getPointerAddressSpace()))
1145 return (void)markOverdefined(IV, &I);
1146 else
1147 return;
1148 }
1149
1150 // Transform load (constant global) into the value loaded.
1151 if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) {
1152 if (!TrackedGlobals.empty()) {
1153 // If we are tracking this global, merge in the known value for it.
1154 auto It = TrackedGlobals.find(GV);
1155 if (It != TrackedGlobals.end()) {
1156 mergeInValue(IV, &I, It->second, getMaxWidenStepsOpts());
1157 return;
1158 }
1159 }
1160 }
1161
1162 // Transform load from a constant into a constant if possible.
1163 if (Constant *C = ConstantFoldLoadFromConstPtr(Ptr, I.getType(), DL))
1164 return (void)markConstant(IV, &I, C);
1165 }
1166
1167 // Fall back to metadata.
1168 mergeInValue(&I, getValueFromMetadata(&I));
1169 }
1170
visitCallBase(CallBase & CB)1171 void SCCPInstVisitor::visitCallBase(CallBase &CB) {
1172 handleCallResult(CB);
1173 handleCallArguments(CB);
1174 }
1175
handleCallOverdefined(CallBase & CB)1176 void SCCPInstVisitor::handleCallOverdefined(CallBase &CB) {
1177 Function *F = CB.getCalledFunction();
1178
1179 // Void return and not tracking callee, just bail.
1180 if (CB.getType()->isVoidTy())
1181 return;
1182
1183 // Always mark struct return as overdefined.
1184 if (CB.getType()->isStructTy())
1185 return (void)markOverdefined(&CB);
1186
1187 // Otherwise, if we have a single return value case, and if the function is
1188 // a declaration, maybe we can constant fold it.
1189 if (F && F->isDeclaration() && canConstantFoldCallTo(&CB, F)) {
1190 SmallVector<Constant *, 8> Operands;
1191 for (const Use &A : CB.args()) {
1192 if (A.get()->getType()->isStructTy())
1193 return markOverdefined(&CB); // Can't handle struct args.
1194 ValueLatticeElement State = getValueState(A);
1195
1196 if (State.isUnknownOrUndef())
1197 return; // Operands are not resolved yet.
1198 if (isOverdefined(State))
1199 return (void)markOverdefined(&CB);
1200 assert(isConstant(State) && "Unknown state!");
1201 Operands.push_back(getConstant(State));
1202 }
1203
1204 if (isOverdefined(getValueState(&CB)))
1205 return (void)markOverdefined(&CB);
1206
1207 // If we can constant fold this, mark the result of the call as a
1208 // constant.
1209 if (Constant *C = ConstantFoldCall(&CB, F, Operands, &GetTLI(*F)))
1210 return (void)markConstant(&CB, C);
1211 }
1212
1213 // Fall back to metadata.
1214 mergeInValue(&CB, getValueFromMetadata(&CB));
1215 }
1216
handleCallArguments(CallBase & CB)1217 void SCCPInstVisitor::handleCallArguments(CallBase &CB) {
1218 Function *F = CB.getCalledFunction();
1219 // If this is a local function that doesn't have its address taken, mark its
1220 // entry block executable and merge in the actual arguments to the call into
1221 // the formal arguments of the function.
1222 if (!TrackingIncomingArguments.empty() &&
1223 TrackingIncomingArguments.count(F)) {
1224 markBlockExecutable(&F->front());
1225
1226 // Propagate information from this call site into the callee.
1227 auto CAI = CB.arg_begin();
1228 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1229 ++AI, ++CAI) {
1230 // If this argument is byval, and if the function is not readonly, there
1231 // will be an implicit copy formed of the input aggregate.
1232 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
1233 markOverdefined(&*AI);
1234 continue;
1235 }
1236
1237 if (auto *STy = dyn_cast<StructType>(AI->getType())) {
1238 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1239 ValueLatticeElement CallArg = getStructValueState(*CAI, i);
1240 mergeInValue(getStructValueState(&*AI, i), &*AI, CallArg,
1241 getMaxWidenStepsOpts());
1242 }
1243 } else
1244 mergeInValue(&*AI, getValueState(*CAI), getMaxWidenStepsOpts());
1245 }
1246 }
1247 }
1248
handleCallResult(CallBase & CB)1249 void SCCPInstVisitor::handleCallResult(CallBase &CB) {
1250 Function *F = CB.getCalledFunction();
1251
1252 if (auto *II = dyn_cast<IntrinsicInst>(&CB)) {
1253 if (II->getIntrinsicID() == Intrinsic::ssa_copy) {
1254 if (ValueState[&CB].isOverdefined())
1255 return;
1256
1257 Value *CopyOf = CB.getOperand(0);
1258 ValueLatticeElement CopyOfVal = getValueState(CopyOf);
1259 const auto *PI = getPredicateInfoFor(&CB);
1260 assert(PI && "Missing predicate info for ssa.copy");
1261
1262 const Optional<PredicateConstraint> &Constraint = PI->getConstraint();
1263 if (!Constraint) {
1264 mergeInValue(ValueState[&CB], &CB, CopyOfVal);
1265 return;
1266 }
1267
1268 CmpInst::Predicate Pred = Constraint->Predicate;
1269 Value *OtherOp = Constraint->OtherOp;
1270
1271 // Wait until OtherOp is resolved.
1272 if (getValueState(OtherOp).isUnknown()) {
1273 addAdditionalUser(OtherOp, &CB);
1274 return;
1275 }
1276
1277 ValueLatticeElement CondVal = getValueState(OtherOp);
1278 ValueLatticeElement &IV = ValueState[&CB];
1279 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) {
1280 auto ImposedCR =
1281 ConstantRange::getFull(DL.getTypeSizeInBits(CopyOf->getType()));
1282
1283 // Get the range imposed by the condition.
1284 if (CondVal.isConstantRange())
1285 ImposedCR = ConstantRange::makeAllowedICmpRegion(
1286 Pred, CondVal.getConstantRange());
1287
1288 // Combine range info for the original value with the new range from the
1289 // condition.
1290 auto CopyOfCR = CopyOfVal.isConstantRange()
1291 ? CopyOfVal.getConstantRange()
1292 : ConstantRange::getFull(
1293 DL.getTypeSizeInBits(CopyOf->getType()));
1294 auto NewCR = ImposedCR.intersectWith(CopyOfCR);
1295 // If the existing information is != x, do not use the information from
1296 // a chained predicate, as the != x information is more likely to be
1297 // helpful in practice.
1298 if (!CopyOfCR.contains(NewCR) && CopyOfCR.getSingleMissingElement())
1299 NewCR = CopyOfCR;
1300
1301 // The new range is based on a branch condition. That guarantees that
1302 // neither of the compare operands can be undef in the branch targets,
1303 // unless we have conditions that are always true/false (e.g. icmp ule
1304 // i32, %a, i32_max). For the latter overdefined/empty range will be
1305 // inferred, but the branch will get folded accordingly anyways.
1306 addAdditionalUser(OtherOp, &CB);
1307 mergeInValue(
1308 IV, &CB,
1309 ValueLatticeElement::getRange(NewCR, /*MayIncludeUndef*/ false));
1310 return;
1311 } else if (Pred == CmpInst::ICMP_EQ && CondVal.isConstant()) {
1312 // For non-integer values or integer constant expressions, only
1313 // propagate equal constants.
1314 addAdditionalUser(OtherOp, &CB);
1315 mergeInValue(IV, &CB, CondVal);
1316 return;
1317 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant()) {
1318 // Propagate inequalities.
1319 addAdditionalUser(OtherOp, &CB);
1320 mergeInValue(IV, &CB,
1321 ValueLatticeElement::getNot(CondVal.getConstant()));
1322 return;
1323 }
1324
1325 return (void)mergeInValue(IV, &CB, CopyOfVal);
1326 }
1327
1328 if (ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) {
1329 // Compute result range for intrinsics supported by ConstantRange.
1330 // Do this even if we don't know a range for all operands, as we may
1331 // still know something about the result range, e.g. of abs(x).
1332 SmallVector<ConstantRange, 2> OpRanges;
1333 for (Value *Op : II->args()) {
1334 const ValueLatticeElement &State = getValueState(Op);
1335 if (State.isConstantRange())
1336 OpRanges.push_back(State.getConstantRange());
1337 else
1338 OpRanges.push_back(
1339 ConstantRange::getFull(Op->getType()->getScalarSizeInBits()));
1340 }
1341
1342 ConstantRange Result =
1343 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges);
1344 return (void)mergeInValue(II, ValueLatticeElement::getRange(Result));
1345 }
1346 }
1347
1348 // The common case is that we aren't tracking the callee, either because we
1349 // are not doing interprocedural analysis or the callee is indirect, or is
1350 // external. Handle these cases first.
1351 if (!F || F->isDeclaration())
1352 return handleCallOverdefined(CB);
1353
1354 // If this is a single/zero retval case, see if we're tracking the function.
1355 if (auto *STy = dyn_cast<StructType>(F->getReturnType())) {
1356 if (!MRVFunctionsTracked.count(F))
1357 return handleCallOverdefined(CB); // Not tracking this callee.
1358
1359 // If we are tracking this callee, propagate the result of the function
1360 // into this call site.
1361 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1362 mergeInValue(getStructValueState(&CB, i), &CB,
1363 TrackedMultipleRetVals[std::make_pair(F, i)],
1364 getMaxWidenStepsOpts());
1365 } else {
1366 auto TFRVI = TrackedRetVals.find(F);
1367 if (TFRVI == TrackedRetVals.end())
1368 return handleCallOverdefined(CB); // Not tracking this callee.
1369
1370 // If so, propagate the return value of the callee into this call result.
1371 mergeInValue(&CB, TFRVI->second, getMaxWidenStepsOpts());
1372 }
1373 }
1374
solve()1375 void SCCPInstVisitor::solve() {
1376 // Process the work lists until they are empty!
1377 while (!BBWorkList.empty() || !InstWorkList.empty() ||
1378 !OverdefinedInstWorkList.empty()) {
1379 // Process the overdefined instruction's work list first, which drives other
1380 // things to overdefined more quickly.
1381 while (!OverdefinedInstWorkList.empty()) {
1382 Value *I = OverdefinedInstWorkList.pop_back_val();
1383
1384 LLVM_DEBUG(dbgs() << "\nPopped off OI-WL: " << *I << '\n');
1385
1386 // "I" got into the work list because it either made the transition from
1387 // bottom to constant, or to overdefined.
1388 //
1389 // Anything on this worklist that is overdefined need not be visited
1390 // since all of its users will have already been marked as overdefined
1391 // Update all of the users of this instruction's value.
1392 //
1393 markUsersAsChanged(I);
1394 }
1395
1396 // Process the instruction work list.
1397 while (!InstWorkList.empty()) {
1398 Value *I = InstWorkList.pop_back_val();
1399
1400 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
1401
1402 // "I" got into the work list because it made the transition from undef to
1403 // constant.
1404 //
1405 // Anything on this worklist that is overdefined need not be visited
1406 // since all of its users will have already been marked as overdefined.
1407 // Update all of the users of this instruction's value.
1408 //
1409 if (I->getType()->isStructTy() || !getValueState(I).isOverdefined())
1410 markUsersAsChanged(I);
1411 }
1412
1413 // Process the basic block work list.
1414 while (!BBWorkList.empty()) {
1415 BasicBlock *BB = BBWorkList.pop_back_val();
1416
1417 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
1418
1419 // Notify all instructions in this basic block that they are newly
1420 // executable.
1421 visit(BB);
1422 }
1423 }
1424 }
1425
1426 /// While solving the dataflow for a function, we don't compute a result for
1427 /// operations with an undef operand, to allow undef to be lowered to a
1428 /// constant later. For example, constant folding of "zext i8 undef to i16"
1429 /// would result in "i16 0", and if undef is later lowered to "i8 1", then the
1430 /// zext result would become "i16 1" and would result into an overdefined
1431 /// lattice value once merged with the previous result. Not computing the
1432 /// result of the zext (treating undef the same as unknown) allows us to handle
1433 /// a later undef->constant lowering more optimally.
1434 ///
1435 /// However, if the operand remains undef when the solver returns, we do need
1436 /// to assign some result to the instruction (otherwise we would treat it as
1437 /// unreachable). For simplicity, we mark any instructions that are still
1438 /// unknown as overdefined.
resolvedUndefsIn(Function & F)1439 bool SCCPInstVisitor::resolvedUndefsIn(Function &F) {
1440 bool MadeChange = false;
1441 for (BasicBlock &BB : F) {
1442 if (!BBExecutable.count(&BB))
1443 continue;
1444
1445 for (Instruction &I : BB) {
1446 // Look for instructions which produce undef values.
1447 if (I.getType()->isVoidTy())
1448 continue;
1449
1450 if (auto *STy = dyn_cast<StructType>(I.getType())) {
1451 // Only a few things that can be structs matter for undef.
1452
1453 // Tracked calls must never be marked overdefined in resolvedUndefsIn.
1454 if (auto *CB = dyn_cast<CallBase>(&I))
1455 if (Function *F = CB->getCalledFunction())
1456 if (MRVFunctionsTracked.count(F))
1457 continue;
1458
1459 // extractvalue and insertvalue don't need to be marked; they are
1460 // tracked as precisely as their operands.
1461 if (isa<ExtractValueInst>(I) || isa<InsertValueInst>(I))
1462 continue;
1463 // Send the results of everything else to overdefined. We could be
1464 // more precise than this but it isn't worth bothering.
1465 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1466 ValueLatticeElement &LV = getStructValueState(&I, i);
1467 if (LV.isUnknown()) {
1468 markOverdefined(LV, &I);
1469 MadeChange = true;
1470 }
1471 }
1472 continue;
1473 }
1474
1475 ValueLatticeElement &LV = getValueState(&I);
1476 if (!LV.isUnknown())
1477 continue;
1478
1479 // There are two reasons a call can have an undef result
1480 // 1. It could be tracked.
1481 // 2. It could be constant-foldable.
1482 // Because of the way we solve return values, tracked calls must
1483 // never be marked overdefined in resolvedUndefsIn.
1484 if (auto *CB = dyn_cast<CallBase>(&I))
1485 if (Function *F = CB->getCalledFunction())
1486 if (TrackedRetVals.count(F))
1487 continue;
1488
1489 if (isa<LoadInst>(I)) {
1490 // A load here means one of two things: a load of undef from a global,
1491 // a load from an unknown pointer. Either way, having it return undef
1492 // is okay.
1493 continue;
1494 }
1495
1496 markOverdefined(&I);
1497 MadeChange = true;
1498 }
1499 }
1500
1501 return MadeChange;
1502 }
1503
1504 //===----------------------------------------------------------------------===//
1505 //
1506 // SCCPSolver implementations
1507 //
SCCPSolver(const DataLayout & DL,std::function<const TargetLibraryInfo & (Function &)> GetTLI,LLVMContext & Ctx)1508 SCCPSolver::SCCPSolver(
1509 const DataLayout &DL,
1510 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
1511 LLVMContext &Ctx)
1512 : Visitor(new SCCPInstVisitor(DL, std::move(GetTLI), Ctx)) {}
1513
1514 SCCPSolver::~SCCPSolver() = default;
1515
addAnalysis(Function & F,AnalysisResultsForFn A)1516 void SCCPSolver::addAnalysis(Function &F, AnalysisResultsForFn A) {
1517 return Visitor->addAnalysis(F, std::move(A));
1518 }
1519
markBlockExecutable(BasicBlock * BB)1520 bool SCCPSolver::markBlockExecutable(BasicBlock *BB) {
1521 return Visitor->markBlockExecutable(BB);
1522 }
1523
getPredicateInfoFor(Instruction * I)1524 const PredicateBase *SCCPSolver::getPredicateInfoFor(Instruction *I) {
1525 return Visitor->getPredicateInfoFor(I);
1526 }
1527
getDTU(Function & F)1528 DomTreeUpdater SCCPSolver::getDTU(Function &F) { return Visitor->getDTU(F); }
1529
trackValueOfGlobalVariable(GlobalVariable * GV)1530 void SCCPSolver::trackValueOfGlobalVariable(GlobalVariable *GV) {
1531 Visitor->trackValueOfGlobalVariable(GV);
1532 }
1533
addTrackedFunction(Function * F)1534 void SCCPSolver::addTrackedFunction(Function *F) {
1535 Visitor->addTrackedFunction(F);
1536 }
1537
addToMustPreserveReturnsInFunctions(Function * F)1538 void SCCPSolver::addToMustPreserveReturnsInFunctions(Function *F) {
1539 Visitor->addToMustPreserveReturnsInFunctions(F);
1540 }
1541
mustPreserveReturn(Function * F)1542 bool SCCPSolver::mustPreserveReturn(Function *F) {
1543 return Visitor->mustPreserveReturn(F);
1544 }
1545
addArgumentTrackedFunction(Function * F)1546 void SCCPSolver::addArgumentTrackedFunction(Function *F) {
1547 Visitor->addArgumentTrackedFunction(F);
1548 }
1549
isArgumentTrackedFunction(Function * F)1550 bool SCCPSolver::isArgumentTrackedFunction(Function *F) {
1551 return Visitor->isArgumentTrackedFunction(F);
1552 }
1553
solve()1554 void SCCPSolver::solve() { Visitor->solve(); }
1555
resolvedUndefsIn(Function & F)1556 bool SCCPSolver::resolvedUndefsIn(Function &F) {
1557 return Visitor->resolvedUndefsIn(F);
1558 }
1559
isBlockExecutable(BasicBlock * BB) const1560 bool SCCPSolver::isBlockExecutable(BasicBlock *BB) const {
1561 return Visitor->isBlockExecutable(BB);
1562 }
1563
isEdgeFeasible(BasicBlock * From,BasicBlock * To) const1564 bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const {
1565 return Visitor->isEdgeFeasible(From, To);
1566 }
1567
1568 std::vector<ValueLatticeElement>
getStructLatticeValueFor(Value * V) const1569 SCCPSolver::getStructLatticeValueFor(Value *V) const {
1570 return Visitor->getStructLatticeValueFor(V);
1571 }
1572
removeLatticeValueFor(Value * V)1573 void SCCPSolver::removeLatticeValueFor(Value *V) {
1574 return Visitor->removeLatticeValueFor(V);
1575 }
1576
getLatticeValueFor(Value * V) const1577 const ValueLatticeElement &SCCPSolver::getLatticeValueFor(Value *V) const {
1578 return Visitor->getLatticeValueFor(V);
1579 }
1580
1581 const MapVector<Function *, ValueLatticeElement> &
getTrackedRetVals()1582 SCCPSolver::getTrackedRetVals() {
1583 return Visitor->getTrackedRetVals();
1584 }
1585
1586 const DenseMap<GlobalVariable *, ValueLatticeElement> &
getTrackedGlobals()1587 SCCPSolver::getTrackedGlobals() {
1588 return Visitor->getTrackedGlobals();
1589 }
1590
getMRVFunctionsTracked()1591 const SmallPtrSet<Function *, 16> SCCPSolver::getMRVFunctionsTracked() {
1592 return Visitor->getMRVFunctionsTracked();
1593 }
1594
markOverdefined(Value * V)1595 void SCCPSolver::markOverdefined(Value *V) { Visitor->markOverdefined(V); }
1596
isStructLatticeConstant(Function * F,StructType * STy)1597 bool SCCPSolver::isStructLatticeConstant(Function *F, StructType *STy) {
1598 return Visitor->isStructLatticeConstant(F, STy);
1599 }
1600
getConstant(const ValueLatticeElement & LV) const1601 Constant *SCCPSolver::getConstant(const ValueLatticeElement &LV) const {
1602 return Visitor->getConstant(LV);
1603 }
1604
getArgumentTrackedFunctions()1605 SmallPtrSetImpl<Function *> &SCCPSolver::getArgumentTrackedFunctions() {
1606 return Visitor->getArgumentTrackedFunctions();
1607 }
1608
markArgInFuncSpecialization(Function * F,const SmallVectorImpl<ArgInfo> & Args)1609 void SCCPSolver::markArgInFuncSpecialization(
1610 Function *F, const SmallVectorImpl<ArgInfo> &Args) {
1611 Visitor->markArgInFuncSpecialization(F, Args);
1612 }
1613
markFunctionUnreachable(Function * F)1614 void SCCPSolver::markFunctionUnreachable(Function *F) {
1615 Visitor->markFunctionUnreachable(F);
1616 }
1617
visit(Instruction * I)1618 void SCCPSolver::visit(Instruction *I) { Visitor->visit(I); }
1619
visitCall(CallInst & I)1620 void SCCPSolver::visitCall(CallInst &I) { Visitor->visitCall(I); }
1621