xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/Local.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- Local.cpp - Functions to perform local transformations -------------===//
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 family of functions perform various local transformations to the
10 // program.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseMapInfo.h"
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/Hashing.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AssumeBundleQueries.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/DomTreeUpdater.h"
28 #include "llvm/Analysis/InstructionSimplify.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/MemorySSAUpdater.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/Analysis/VectorUtils.h"
34 #include "llvm/BinaryFormat/Dwarf.h"
35 #include "llvm/IR/Argument.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DIBuilder.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/DebugInfo.h"
45 #include "llvm/IR/DebugInfoMetadata.h"
46 #include "llvm/IR/DebugLoc.h"
47 #include "llvm/IR/DerivedTypes.h"
48 #include "llvm/IR/Dominators.h"
49 #include "llvm/IR/EHPersonalities.h"
50 #include "llvm/IR/Function.h"
51 #include "llvm/IR/GetElementPtrTypeIterator.h"
52 #include "llvm/IR/GlobalObject.h"
53 #include "llvm/IR/IRBuilder.h"
54 #include "llvm/IR/InstrTypes.h"
55 #include "llvm/IR/Instruction.h"
56 #include "llvm/IR/Instructions.h"
57 #include "llvm/IR/IntrinsicInst.h"
58 #include "llvm/IR/Intrinsics.h"
59 #include "llvm/IR/IntrinsicsWebAssembly.h"
60 #include "llvm/IR/LLVMContext.h"
61 #include "llvm/IR/MDBuilder.h"
62 #include "llvm/IR/MemoryModelRelaxationAnnotations.h"
63 #include "llvm/IR/Metadata.h"
64 #include "llvm/IR/Module.h"
65 #include "llvm/IR/PatternMatch.h"
66 #include "llvm/IR/ProfDataUtils.h"
67 #include "llvm/IR/Type.h"
68 #include "llvm/IR/Use.h"
69 #include "llvm/IR/User.h"
70 #include "llvm/IR/Value.h"
71 #include "llvm/IR/ValueHandle.h"
72 #include "llvm/Support/Casting.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/ErrorHandling.h"
76 #include "llvm/Support/KnownBits.h"
77 #include "llvm/Support/raw_ostream.h"
78 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
79 #include "llvm/Transforms/Utils/ValueMapper.h"
80 #include <algorithm>
81 #include <cassert>
82 #include <cstdint>
83 #include <iterator>
84 #include <map>
85 #include <optional>
86 #include <utility>
87 
88 using namespace llvm;
89 using namespace llvm::PatternMatch;
90 
91 extern cl::opt<bool> UseNewDbgInfoFormat;
92 
93 #define DEBUG_TYPE "local"
94 
95 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
96 STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd");
97 
98 static cl::opt<bool> PHICSEDebugHash(
99     "phicse-debug-hash",
100 #ifdef EXPENSIVE_CHECKS
101     cl::init(true),
102 #else
103     cl::init(false),
104 #endif
105     cl::Hidden,
106     cl::desc("Perform extra assertion checking to verify that PHINodes's hash "
107              "function is well-behaved w.r.t. its isEqual predicate"));
108 
109 static cl::opt<unsigned> PHICSENumPHISmallSize(
110     "phicse-num-phi-smallsize", cl::init(32), cl::Hidden,
111     cl::desc(
112         "When the basic block contains not more than this number of PHI nodes, "
113         "perform a (faster!) exhaustive search instead of set-driven one."));
114 
115 // Max recursion depth for collectBitParts used when detecting bswap and
116 // bitreverse idioms.
117 static const unsigned BitPartRecursionMaxDepth = 48;
118 
119 //===----------------------------------------------------------------------===//
120 //  Local constant propagation.
121 //
122 
123 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
124 /// constant value, convert it into an unconditional branch to the constant
125 /// destination.  This is a nontrivial operation because the successors of this
126 /// basic block must have their PHI nodes updated.
127 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
128 /// conditions and indirectbr addresses this might make dead if
129 /// DeleteDeadConditions is true.
ConstantFoldTerminator(BasicBlock * BB,bool DeleteDeadConditions,const TargetLibraryInfo * TLI,DomTreeUpdater * DTU)130 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
131                                   const TargetLibraryInfo *TLI,
132                                   DomTreeUpdater *DTU) {
133   Instruction *T = BB->getTerminator();
134   IRBuilder<> Builder(T);
135 
136   // Branch - See if we are conditional jumping on constant
137   if (auto *BI = dyn_cast<BranchInst>(T)) {
138     if (BI->isUnconditional()) return false;  // Can't optimize uncond branch
139 
140     BasicBlock *Dest1 = BI->getSuccessor(0);
141     BasicBlock *Dest2 = BI->getSuccessor(1);
142 
143     if (Dest2 == Dest1) {       // Conditional branch to same location?
144       // This branch matches something like this:
145       //     br bool %cond, label %Dest, label %Dest
146       // and changes it into:  br label %Dest
147 
148       // Let the basic block know that we are letting go of one copy of it.
149       assert(BI->getParent() && "Terminator not inserted in block!");
150       Dest1->removePredecessor(BI->getParent());
151 
152       // Replace the conditional branch with an unconditional one.
153       BranchInst *NewBI = Builder.CreateBr(Dest1);
154 
155       // Transfer the metadata to the new branch instruction.
156       NewBI->copyMetadata(*BI, {LLVMContext::MD_loop, LLVMContext::MD_dbg,
157                                 LLVMContext::MD_annotation});
158 
159       Value *Cond = BI->getCondition();
160       BI->eraseFromParent();
161       if (DeleteDeadConditions)
162         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
163       return true;
164     }
165 
166     if (auto *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
167       // Are we branching on constant?
168       // YES.  Change to unconditional branch...
169       BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
170       BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1;
171 
172       // Let the basic block know that we are letting go of it.  Based on this,
173       // it will adjust it's PHI nodes.
174       OldDest->removePredecessor(BB);
175 
176       // Replace the conditional branch with an unconditional one.
177       BranchInst *NewBI = Builder.CreateBr(Destination);
178 
179       // Transfer the metadata to the new branch instruction.
180       NewBI->copyMetadata(*BI, {LLVMContext::MD_loop, LLVMContext::MD_dbg,
181                                 LLVMContext::MD_annotation});
182 
183       BI->eraseFromParent();
184       if (DTU)
185         DTU->applyUpdates({{DominatorTree::Delete, BB, OldDest}});
186       return true;
187     }
188 
189     return false;
190   }
191 
192   if (auto *SI = dyn_cast<SwitchInst>(T)) {
193     // If we are switching on a constant, we can convert the switch to an
194     // unconditional branch.
195     auto *CI = dyn_cast<ConstantInt>(SI->getCondition());
196     BasicBlock *DefaultDest = SI->getDefaultDest();
197     BasicBlock *TheOnlyDest = DefaultDest;
198 
199     // If the default is unreachable, ignore it when searching for TheOnlyDest.
200     if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
201         SI->getNumCases() > 0) {
202       TheOnlyDest = SI->case_begin()->getCaseSuccessor();
203     }
204 
205     bool Changed = false;
206 
207     // Figure out which case it goes to.
208     for (auto It = SI->case_begin(), End = SI->case_end(); It != End;) {
209       // Found case matching a constant operand?
210       if (It->getCaseValue() == CI) {
211         TheOnlyDest = It->getCaseSuccessor();
212         break;
213       }
214 
215       // Check to see if this branch is going to the same place as the default
216       // dest.  If so, eliminate it as an explicit compare.
217       if (It->getCaseSuccessor() == DefaultDest) {
218         MDNode *MD = getValidBranchWeightMDNode(*SI);
219         unsigned NCases = SI->getNumCases();
220         // Fold the case metadata into the default if there will be any branches
221         // left, unless the metadata doesn't match the switch.
222         if (NCases > 1 && MD) {
223           // Collect branch weights into a vector.
224           SmallVector<uint32_t, 8> Weights;
225           extractBranchWeights(MD, Weights);
226 
227           // Merge weight of this case to the default weight.
228           unsigned Idx = It->getCaseIndex();
229           // TODO: Add overflow check.
230           Weights[0] += Weights[Idx + 1];
231           // Remove weight for this case.
232           std::swap(Weights[Idx + 1], Weights.back());
233           Weights.pop_back();
234           setBranchWeights(*SI, Weights, hasBranchWeightOrigin(MD));
235         }
236         // Remove this entry.
237         BasicBlock *ParentBB = SI->getParent();
238         DefaultDest->removePredecessor(ParentBB);
239         It = SI->removeCase(It);
240         End = SI->case_end();
241 
242         // Removing this case may have made the condition constant. In that
243         // case, update CI and restart iteration through the cases.
244         if (auto *NewCI = dyn_cast<ConstantInt>(SI->getCondition())) {
245           CI = NewCI;
246           It = SI->case_begin();
247         }
248 
249         Changed = true;
250         continue;
251       }
252 
253       // Otherwise, check to see if the switch only branches to one destination.
254       // We do this by reseting "TheOnlyDest" to null when we find two non-equal
255       // destinations.
256       if (It->getCaseSuccessor() != TheOnlyDest)
257         TheOnlyDest = nullptr;
258 
259       // Increment this iterator as we haven't removed the case.
260       ++It;
261     }
262 
263     if (CI && !TheOnlyDest) {
264       // Branching on a constant, but not any of the cases, go to the default
265       // successor.
266       TheOnlyDest = SI->getDefaultDest();
267     }
268 
269     // If we found a single destination that we can fold the switch into, do so
270     // now.
271     if (TheOnlyDest) {
272       // Insert the new branch.
273       Builder.CreateBr(TheOnlyDest);
274       BasicBlock *BB = SI->getParent();
275 
276       SmallSet<BasicBlock *, 8> RemovedSuccessors;
277 
278       // Remove entries from PHI nodes which we no longer branch to...
279       BasicBlock *SuccToKeep = TheOnlyDest;
280       for (BasicBlock *Succ : successors(SI)) {
281         if (DTU && Succ != TheOnlyDest)
282           RemovedSuccessors.insert(Succ);
283         // Found case matching a constant operand?
284         if (Succ == SuccToKeep) {
285           SuccToKeep = nullptr; // Don't modify the first branch to TheOnlyDest
286         } else {
287           Succ->removePredecessor(BB);
288         }
289       }
290 
291       // Delete the old switch.
292       Value *Cond = SI->getCondition();
293       SI->eraseFromParent();
294       if (DeleteDeadConditions)
295         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
296       if (DTU) {
297         std::vector<DominatorTree::UpdateType> Updates;
298         Updates.reserve(RemovedSuccessors.size());
299         for (auto *RemovedSuccessor : RemovedSuccessors)
300           Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
301         DTU->applyUpdates(Updates);
302       }
303       return true;
304     }
305 
306     if (SI->getNumCases() == 1) {
307       // Otherwise, we can fold this switch into a conditional branch
308       // instruction if it has only one non-default destination.
309       auto FirstCase = *SI->case_begin();
310       Value *Cond = Builder.CreateICmpEQ(SI->getCondition(),
311           FirstCase.getCaseValue(), "cond");
312 
313       // Insert the new branch.
314       BranchInst *NewBr = Builder.CreateCondBr(Cond,
315                                                FirstCase.getCaseSuccessor(),
316                                                SI->getDefaultDest());
317       SmallVector<uint32_t> Weights;
318       if (extractBranchWeights(*SI, Weights) && Weights.size() == 2) {
319         uint32_t DefWeight = Weights[0];
320         uint32_t CaseWeight = Weights[1];
321         // The TrueWeight should be the weight for the single case of SI.
322         NewBr->setMetadata(LLVMContext::MD_prof,
323                            MDBuilder(BB->getContext())
324                                .createBranchWeights(CaseWeight, DefWeight));
325       }
326 
327       // Update make.implicit metadata to the newly-created conditional branch.
328       MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit);
329       if (MakeImplicitMD)
330         NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD);
331 
332       // Delete the old switch.
333       SI->eraseFromParent();
334       return true;
335     }
336     return Changed;
337   }
338 
339   if (auto *IBI = dyn_cast<IndirectBrInst>(T)) {
340     // indirectbr blockaddress(@F, @BB) -> br label @BB
341     if (auto *BA =
342           dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) {
343       BasicBlock *TheOnlyDest = BA->getBasicBlock();
344       SmallSet<BasicBlock *, 8> RemovedSuccessors;
345 
346       // Insert the new branch.
347       Builder.CreateBr(TheOnlyDest);
348 
349       BasicBlock *SuccToKeep = TheOnlyDest;
350       for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
351         BasicBlock *DestBB = IBI->getDestination(i);
352         if (DTU && DestBB != TheOnlyDest)
353           RemovedSuccessors.insert(DestBB);
354         if (IBI->getDestination(i) == SuccToKeep) {
355           SuccToKeep = nullptr;
356         } else {
357           DestBB->removePredecessor(BB);
358         }
359       }
360       Value *Address = IBI->getAddress();
361       IBI->eraseFromParent();
362       if (DeleteDeadConditions)
363         // Delete pointer cast instructions.
364         RecursivelyDeleteTriviallyDeadInstructions(Address, TLI);
365 
366       // Also zap the blockaddress constant if there are no users remaining,
367       // otherwise the destination is still marked as having its address taken.
368       if (BA->use_empty())
369         BA->destroyConstant();
370 
371       // If we didn't find our destination in the IBI successor list, then we
372       // have undefined behavior.  Replace the unconditional branch with an
373       // 'unreachable' instruction.
374       if (SuccToKeep) {
375         BB->getTerminator()->eraseFromParent();
376         new UnreachableInst(BB->getContext(), BB);
377       }
378 
379       if (DTU) {
380         std::vector<DominatorTree::UpdateType> Updates;
381         Updates.reserve(RemovedSuccessors.size());
382         for (auto *RemovedSuccessor : RemovedSuccessors)
383           Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
384         DTU->applyUpdates(Updates);
385       }
386       return true;
387     }
388   }
389 
390   return false;
391 }
392 
393 //===----------------------------------------------------------------------===//
394 //  Local dead code elimination.
395 //
396 
397 /// isInstructionTriviallyDead - Return true if the result produced by the
398 /// instruction is not used, and the instruction has no side effects.
399 ///
isInstructionTriviallyDead(Instruction * I,const TargetLibraryInfo * TLI)400 bool llvm::isInstructionTriviallyDead(Instruction *I,
401                                       const TargetLibraryInfo *TLI) {
402   if (!I->use_empty())
403     return false;
404   return wouldInstructionBeTriviallyDead(I, TLI);
405 }
406 
wouldInstructionBeTriviallyDeadOnUnusedPaths(Instruction * I,const TargetLibraryInfo * TLI)407 bool llvm::wouldInstructionBeTriviallyDeadOnUnusedPaths(
408     Instruction *I, const TargetLibraryInfo *TLI) {
409   // Instructions that are "markers" and have implied meaning on code around
410   // them (without explicit uses), are not dead on unused paths.
411   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
412     if (II->getIntrinsicID() == Intrinsic::stacksave ||
413         II->getIntrinsicID() == Intrinsic::launder_invariant_group ||
414         II->isLifetimeStartOrEnd())
415       return false;
416   return wouldInstructionBeTriviallyDead(I, TLI);
417 }
418 
wouldInstructionBeTriviallyDead(const Instruction * I,const TargetLibraryInfo * TLI)419 bool llvm::wouldInstructionBeTriviallyDead(const Instruction *I,
420                                            const TargetLibraryInfo *TLI) {
421   if (I->isTerminator())
422     return false;
423 
424   // We don't want the landingpad-like instructions removed by anything this
425   // general.
426   if (I->isEHPad())
427     return false;
428 
429   // We don't want debug info removed by anything this general.
430   if (isa<DbgVariableIntrinsic>(I))
431     return false;
432 
433   if (const DbgLabelInst *DLI = dyn_cast<DbgLabelInst>(I)) {
434     if (DLI->getLabel())
435       return false;
436     return true;
437   }
438 
439   if (auto *CB = dyn_cast<CallBase>(I))
440     if (isRemovableAlloc(CB, TLI))
441       return true;
442 
443   if (!I->willReturn()) {
444     auto *II = dyn_cast<IntrinsicInst>(I);
445     if (!II)
446       return false;
447 
448     switch (II->getIntrinsicID()) {
449     case Intrinsic::experimental_guard: {
450       // Guards on true are operationally no-ops.  In the future we can
451       // consider more sophisticated tradeoffs for guards considering potential
452       // for check widening, but for now we keep things simple.
453       auto *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0));
454       return Cond && Cond->isOne();
455     }
456     // TODO: These intrinsics are not safe to remove, because this may remove
457     // a well-defined trap.
458     case Intrinsic::wasm_trunc_signed:
459     case Intrinsic::wasm_trunc_unsigned:
460     case Intrinsic::ptrauth_auth:
461     case Intrinsic::ptrauth_resign:
462       return true;
463     default:
464       return false;
465     }
466   }
467 
468   if (!I->mayHaveSideEffects())
469     return true;
470 
471   // Special case intrinsics that "may have side effects" but can be deleted
472   // when dead.
473   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
474     // Safe to delete llvm.stacksave and launder.invariant.group if dead.
475     if (II->getIntrinsicID() == Intrinsic::stacksave ||
476         II->getIntrinsicID() == Intrinsic::launder_invariant_group)
477       return true;
478 
479     // Intrinsics declare sideeffects to prevent them from moving, but they are
480     // nops without users.
481     if (II->getIntrinsicID() == Intrinsic::allow_runtime_check ||
482         II->getIntrinsicID() == Intrinsic::allow_ubsan_check)
483       return true;
484 
485     if (II->isLifetimeStartOrEnd()) {
486       auto *Arg = II->getArgOperand(1);
487       // Lifetime intrinsics are dead when their right-hand is undef.
488       if (isa<UndefValue>(Arg))
489         return true;
490       // If the right-hand is an alloc, global, or argument and the only uses
491       // are lifetime intrinsics then the intrinsics are dead.
492       if (isa<AllocaInst>(Arg) || isa<GlobalValue>(Arg) || isa<Argument>(Arg))
493         return llvm::all_of(Arg->uses(), [](Use &Use) {
494           if (IntrinsicInst *IntrinsicUse =
495                   dyn_cast<IntrinsicInst>(Use.getUser()))
496             return IntrinsicUse->isLifetimeStartOrEnd();
497           return false;
498         });
499       return false;
500     }
501 
502     // Assumptions are dead if their condition is trivially true.
503     if (II->getIntrinsicID() == Intrinsic::assume &&
504         isAssumeWithEmptyBundle(cast<AssumeInst>(*II))) {
505       if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
506         return !Cond->isZero();
507 
508       return false;
509     }
510 
511     if (auto *FPI = dyn_cast<ConstrainedFPIntrinsic>(I)) {
512       std::optional<fp::ExceptionBehavior> ExBehavior =
513           FPI->getExceptionBehavior();
514       return *ExBehavior != fp::ebStrict;
515     }
516   }
517 
518   if (auto *Call = dyn_cast<CallBase>(I)) {
519     if (Value *FreedOp = getFreedOperand(Call, TLI))
520       if (Constant *C = dyn_cast<Constant>(FreedOp))
521         return C->isNullValue() || isa<UndefValue>(C);
522     if (isMathLibCallNoop(Call, TLI))
523       return true;
524   }
525 
526   // Non-volatile atomic loads from constants can be removed.
527   if (auto *LI = dyn_cast<LoadInst>(I))
528     if (auto *GV = dyn_cast<GlobalVariable>(
529             LI->getPointerOperand()->stripPointerCasts()))
530       if (!LI->isVolatile() && GV->isConstant())
531         return true;
532 
533   return false;
534 }
535 
536 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
537 /// trivially dead instruction, delete it.  If that makes any of its operands
538 /// trivially dead, delete them too, recursively.  Return true if any
539 /// instructions were deleted.
RecursivelyDeleteTriviallyDeadInstructions(Value * V,const TargetLibraryInfo * TLI,MemorySSAUpdater * MSSAU,std::function<void (Value *)> AboutToDeleteCallback)540 bool llvm::RecursivelyDeleteTriviallyDeadInstructions(
541     Value *V, const TargetLibraryInfo *TLI, MemorySSAUpdater *MSSAU,
542     std::function<void(Value *)> AboutToDeleteCallback) {
543   Instruction *I = dyn_cast<Instruction>(V);
544   if (!I || !isInstructionTriviallyDead(I, TLI))
545     return false;
546 
547   SmallVector<WeakTrackingVH, 16> DeadInsts;
548   DeadInsts.push_back(I);
549   RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI, MSSAU,
550                                              AboutToDeleteCallback);
551 
552   return true;
553 }
554 
RecursivelyDeleteTriviallyDeadInstructionsPermissive(SmallVectorImpl<WeakTrackingVH> & DeadInsts,const TargetLibraryInfo * TLI,MemorySSAUpdater * MSSAU,std::function<void (Value *)> AboutToDeleteCallback)555 bool llvm::RecursivelyDeleteTriviallyDeadInstructionsPermissive(
556     SmallVectorImpl<WeakTrackingVH> &DeadInsts, const TargetLibraryInfo *TLI,
557     MemorySSAUpdater *MSSAU,
558     std::function<void(Value *)> AboutToDeleteCallback) {
559   unsigned S = 0, E = DeadInsts.size(), Alive = 0;
560   for (; S != E; ++S) {
561     auto *I = dyn_cast_or_null<Instruction>(DeadInsts[S]);
562     if (!I || !isInstructionTriviallyDead(I)) {
563       DeadInsts[S] = nullptr;
564       ++Alive;
565     }
566   }
567   if (Alive == E)
568     return false;
569   RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI, MSSAU,
570                                              AboutToDeleteCallback);
571   return true;
572 }
573 
RecursivelyDeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> & DeadInsts,const TargetLibraryInfo * TLI,MemorySSAUpdater * MSSAU,std::function<void (Value *)> AboutToDeleteCallback)574 void llvm::RecursivelyDeleteTriviallyDeadInstructions(
575     SmallVectorImpl<WeakTrackingVH> &DeadInsts, const TargetLibraryInfo *TLI,
576     MemorySSAUpdater *MSSAU,
577     std::function<void(Value *)> AboutToDeleteCallback) {
578   // Process the dead instruction list until empty.
579   while (!DeadInsts.empty()) {
580     Value *V = DeadInsts.pop_back_val();
581     Instruction *I = cast_or_null<Instruction>(V);
582     if (!I)
583       continue;
584     assert(isInstructionTriviallyDead(I, TLI) &&
585            "Live instruction found in dead worklist!");
586     assert(I->use_empty() && "Instructions with uses are not dead.");
587 
588     // Don't lose the debug info while deleting the instructions.
589     salvageDebugInfo(*I);
590 
591     if (AboutToDeleteCallback)
592       AboutToDeleteCallback(I);
593 
594     // Null out all of the instruction's operands to see if any operand becomes
595     // dead as we go.
596     for (Use &OpU : I->operands()) {
597       Value *OpV = OpU.get();
598       OpU.set(nullptr);
599 
600       if (!OpV->use_empty())
601         continue;
602 
603       // If the operand is an instruction that became dead as we nulled out the
604       // operand, and if it is 'trivially' dead, delete it in a future loop
605       // iteration.
606       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
607         if (isInstructionTriviallyDead(OpI, TLI))
608           DeadInsts.push_back(OpI);
609     }
610     if (MSSAU)
611       MSSAU->removeMemoryAccess(I);
612 
613     I->eraseFromParent();
614   }
615 }
616 
replaceDbgUsesWithUndef(Instruction * I)617 bool llvm::replaceDbgUsesWithUndef(Instruction *I) {
618   SmallVector<DbgVariableIntrinsic *, 1> DbgUsers;
619   SmallVector<DbgVariableRecord *, 1> DPUsers;
620   findDbgUsers(DbgUsers, I, &DPUsers);
621   for (auto *DII : DbgUsers)
622     DII->setKillLocation();
623   for (auto *DVR : DPUsers)
624     DVR->setKillLocation();
625   return !DbgUsers.empty() || !DPUsers.empty();
626 }
627 
628 /// areAllUsesEqual - Check whether the uses of a value are all the same.
629 /// This is similar to Instruction::hasOneUse() except this will also return
630 /// true when there are no uses or multiple uses that all refer to the same
631 /// value.
areAllUsesEqual(Instruction * I)632 static bool areAllUsesEqual(Instruction *I) {
633   Value::user_iterator UI = I->user_begin();
634   Value::user_iterator UE = I->user_end();
635   if (UI == UE)
636     return true;
637 
638   User *TheUse = *UI;
639   for (++UI; UI != UE; ++UI) {
640     if (*UI != TheUse)
641       return false;
642   }
643   return true;
644 }
645 
646 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
647 /// dead PHI node, due to being a def-use chain of single-use nodes that
648 /// either forms a cycle or is terminated by a trivially dead instruction,
649 /// delete it.  If that makes any of its operands trivially dead, delete them
650 /// too, recursively.  Return true if a change was made.
RecursivelyDeleteDeadPHINode(PHINode * PN,const TargetLibraryInfo * TLI,llvm::MemorySSAUpdater * MSSAU)651 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
652                                         const TargetLibraryInfo *TLI,
653                                         llvm::MemorySSAUpdater *MSSAU) {
654   SmallPtrSet<Instruction*, 4> Visited;
655   for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
656        I = cast<Instruction>(*I->user_begin())) {
657     if (I->use_empty())
658       return RecursivelyDeleteTriviallyDeadInstructions(I, TLI, MSSAU);
659 
660     // If we find an instruction more than once, we're on a cycle that
661     // won't prove fruitful.
662     if (!Visited.insert(I).second) {
663       // Break the cycle and delete the instruction and its operands.
664       I->replaceAllUsesWith(PoisonValue::get(I->getType()));
665       (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI, MSSAU);
666       return true;
667     }
668   }
669   return false;
670 }
671 
672 static bool
simplifyAndDCEInstruction(Instruction * I,SmallSetVector<Instruction *,16> & WorkList,const DataLayout & DL,const TargetLibraryInfo * TLI)673 simplifyAndDCEInstruction(Instruction *I,
674                           SmallSetVector<Instruction *, 16> &WorkList,
675                           const DataLayout &DL,
676                           const TargetLibraryInfo *TLI) {
677   if (isInstructionTriviallyDead(I, TLI)) {
678     salvageDebugInfo(*I);
679 
680     // Null out all of the instruction's operands to see if any operand becomes
681     // dead as we go.
682     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
683       Value *OpV = I->getOperand(i);
684       I->setOperand(i, nullptr);
685 
686       if (!OpV->use_empty() || I == OpV)
687         continue;
688 
689       // If the operand is an instruction that became dead as we nulled out the
690       // operand, and if it is 'trivially' dead, delete it in a future loop
691       // iteration.
692       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
693         if (isInstructionTriviallyDead(OpI, TLI))
694           WorkList.insert(OpI);
695     }
696 
697     I->eraseFromParent();
698 
699     return true;
700   }
701 
702   if (Value *SimpleV = simplifyInstruction(I, DL)) {
703     // Add the users to the worklist. CAREFUL: an instruction can use itself,
704     // in the case of a phi node.
705     for (User *U : I->users()) {
706       if (U != I) {
707         WorkList.insert(cast<Instruction>(U));
708       }
709     }
710 
711     // Replace the instruction with its simplified value.
712     bool Changed = false;
713     if (!I->use_empty()) {
714       I->replaceAllUsesWith(SimpleV);
715       Changed = true;
716     }
717     if (isInstructionTriviallyDead(I, TLI)) {
718       I->eraseFromParent();
719       Changed = true;
720     }
721     return Changed;
722   }
723   return false;
724 }
725 
726 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
727 /// simplify any instructions in it and recursively delete dead instructions.
728 ///
729 /// This returns true if it changed the code, note that it can delete
730 /// instructions in other blocks as well in this block.
SimplifyInstructionsInBlock(BasicBlock * BB,const TargetLibraryInfo * TLI)731 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
732                                        const TargetLibraryInfo *TLI) {
733   bool MadeChange = false;
734   const DataLayout &DL = BB->getDataLayout();
735 
736 #ifndef NDEBUG
737   // In debug builds, ensure that the terminator of the block is never replaced
738   // or deleted by these simplifications. The idea of simplification is that it
739   // cannot introduce new instructions, and there is no way to replace the
740   // terminator of a block without introducing a new instruction.
741   AssertingVH<Instruction> TerminatorVH(&BB->back());
742 #endif
743 
744   SmallSetVector<Instruction *, 16> WorkList;
745   // Iterate over the original function, only adding insts to the worklist
746   // if they actually need to be revisited. This avoids having to pre-init
747   // the worklist with the entire function's worth of instructions.
748   for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end());
749        BI != E;) {
750     assert(!BI->isTerminator());
751     Instruction *I = &*BI;
752     ++BI;
753 
754     // We're visiting this instruction now, so make sure it's not in the
755     // worklist from an earlier visit.
756     if (!WorkList.count(I))
757       MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
758   }
759 
760   while (!WorkList.empty()) {
761     Instruction *I = WorkList.pop_back_val();
762     MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
763   }
764   return MadeChange;
765 }
766 
767 //===----------------------------------------------------------------------===//
768 //  Control Flow Graph Restructuring.
769 //
770 
MergeBasicBlockIntoOnlyPred(BasicBlock * DestBB,DomTreeUpdater * DTU)771 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB,
772                                        DomTreeUpdater *DTU) {
773 
774   // If BB has single-entry PHI nodes, fold them.
775   while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
776     Value *NewVal = PN->getIncomingValue(0);
777     // Replace self referencing PHI with poison, it must be dead.
778     if (NewVal == PN) NewVal = PoisonValue::get(PN->getType());
779     PN->replaceAllUsesWith(NewVal);
780     PN->eraseFromParent();
781   }
782 
783   BasicBlock *PredBB = DestBB->getSinglePredecessor();
784   assert(PredBB && "Block doesn't have a single predecessor!");
785 
786   bool ReplaceEntryBB = PredBB->isEntryBlock();
787 
788   // DTU updates: Collect all the edges that enter
789   // PredBB. These dominator edges will be redirected to DestBB.
790   SmallVector<DominatorTree::UpdateType, 32> Updates;
791 
792   if (DTU) {
793     // To avoid processing the same predecessor more than once.
794     SmallPtrSet<BasicBlock *, 2> SeenPreds;
795     Updates.reserve(Updates.size() + 2 * pred_size(PredBB) + 1);
796     for (BasicBlock *PredOfPredBB : predecessors(PredBB))
797       // This predecessor of PredBB may already have DestBB as a successor.
798       if (PredOfPredBB != PredBB)
799         if (SeenPreds.insert(PredOfPredBB).second)
800           Updates.push_back({DominatorTree::Insert, PredOfPredBB, DestBB});
801     SeenPreds.clear();
802     for (BasicBlock *PredOfPredBB : predecessors(PredBB))
803       if (SeenPreds.insert(PredOfPredBB).second)
804         Updates.push_back({DominatorTree::Delete, PredOfPredBB, PredBB});
805     Updates.push_back({DominatorTree::Delete, PredBB, DestBB});
806   }
807 
808   // Zap anything that took the address of DestBB.  Not doing this will give the
809   // address an invalid value.
810   if (DestBB->hasAddressTaken()) {
811     BlockAddress *BA = BlockAddress::get(DestBB);
812     Constant *Replacement =
813       ConstantInt::get(Type::getInt32Ty(BA->getContext()), 1);
814     BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
815                                                      BA->getType()));
816     BA->destroyConstant();
817   }
818 
819   // Anything that branched to PredBB now branches to DestBB.
820   PredBB->replaceAllUsesWith(DestBB);
821 
822   // Splice all the instructions from PredBB to DestBB.
823   PredBB->getTerminator()->eraseFromParent();
824   DestBB->splice(DestBB->begin(), PredBB);
825   new UnreachableInst(PredBB->getContext(), PredBB);
826 
827   // If the PredBB is the entry block of the function, move DestBB up to
828   // become the entry block after we erase PredBB.
829   if (ReplaceEntryBB)
830     DestBB->moveAfter(PredBB);
831 
832   if (DTU) {
833     assert(PredBB->size() == 1 &&
834            isa<UnreachableInst>(PredBB->getTerminator()) &&
835            "The successor list of PredBB isn't empty before "
836            "applying corresponding DTU updates.");
837     DTU->applyUpdatesPermissive(Updates);
838     DTU->deleteBB(PredBB);
839     // Recalculation of DomTree is needed when updating a forward DomTree and
840     // the Entry BB is replaced.
841     if (ReplaceEntryBB && DTU->hasDomTree()) {
842       // The entry block was removed and there is no external interface for
843       // the dominator tree to be notified of this change. In this corner-case
844       // we recalculate the entire tree.
845       DTU->recalculate(*(DestBB->getParent()));
846     }
847   }
848 
849   else {
850     PredBB->eraseFromParent(); // Nuke BB if DTU is nullptr.
851   }
852 }
853 
854 /// Return true if we can choose one of these values to use in place of the
855 /// other. Note that we will always choose the non-undef value to keep.
CanMergeValues(Value * First,Value * Second)856 static bool CanMergeValues(Value *First, Value *Second) {
857   return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
858 }
859 
860 /// Return true if we can fold BB, an almost-empty BB ending in an unconditional
861 /// branch to Succ, into Succ.
862 ///
863 /// Assumption: Succ is the single successor for BB.
864 static bool
CanPropagatePredecessorsForPHIs(BasicBlock * BB,BasicBlock * Succ,const SmallPtrSetImpl<BasicBlock * > & BBPreds)865 CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ,
866                                 const SmallPtrSetImpl<BasicBlock *> &BBPreds) {
867   assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
868 
869   LLVM_DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
870                     << Succ->getName() << "\n");
871   // Shortcut, if there is only a single predecessor it must be BB and merging
872   // is always safe
873   if (Succ->getSinglePredecessor())
874     return true;
875 
876   // Look at all the phi nodes in Succ, to see if they present a conflict when
877   // merging these blocks
878   for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
879     PHINode *PN = cast<PHINode>(I);
880 
881     // If the incoming value from BB is again a PHINode in
882     // BB which has the same incoming value for *PI as PN does, we can
883     // merge the phi nodes and then the blocks can still be merged
884     PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
885     if (BBPN && BBPN->getParent() == BB) {
886       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
887         BasicBlock *IBB = PN->getIncomingBlock(PI);
888         if (BBPreds.count(IBB) &&
889             !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
890                             PN->getIncomingValue(PI))) {
891           LLVM_DEBUG(dbgs()
892                      << "Can't fold, phi node " << PN->getName() << " in "
893                      << Succ->getName() << " is conflicting with "
894                      << BBPN->getName() << " with regard to common predecessor "
895                      << IBB->getName() << "\n");
896           return false;
897         }
898       }
899     } else {
900       Value* Val = PN->getIncomingValueForBlock(BB);
901       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
902         // See if the incoming value for the common predecessor is equal to the
903         // one for BB, in which case this phi node will not prevent the merging
904         // of the block.
905         BasicBlock *IBB = PN->getIncomingBlock(PI);
906         if (BBPreds.count(IBB) &&
907             !CanMergeValues(Val, PN->getIncomingValue(PI))) {
908           LLVM_DEBUG(dbgs() << "Can't fold, phi node " << PN->getName()
909                             << " in " << Succ->getName()
910                             << " is conflicting with regard to common "
911                             << "predecessor " << IBB->getName() << "\n");
912           return false;
913         }
914       }
915     }
916   }
917 
918   return true;
919 }
920 
921 using PredBlockVector = SmallVector<BasicBlock *, 16>;
922 using IncomingValueMap = DenseMap<BasicBlock *, Value *>;
923 
924 /// Determines the value to use as the phi node input for a block.
925 ///
926 /// Select between \p OldVal any value that we know flows from \p BB
927 /// to a particular phi on the basis of which one (if either) is not
928 /// undef. Update IncomingValues based on the selected value.
929 ///
930 /// \param OldVal The value we are considering selecting.
931 /// \param BB The block that the value flows in from.
932 /// \param IncomingValues A map from block-to-value for other phi inputs
933 /// that we have examined.
934 ///
935 /// \returns the selected value.
selectIncomingValueForBlock(Value * OldVal,BasicBlock * BB,IncomingValueMap & IncomingValues)936 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
937                                           IncomingValueMap &IncomingValues) {
938   if (!isa<UndefValue>(OldVal)) {
939     assert((!IncomingValues.count(BB) ||
940             IncomingValues.find(BB)->second == OldVal) &&
941            "Expected OldVal to match incoming value from BB!");
942 
943     IncomingValues.insert(std::make_pair(BB, OldVal));
944     return OldVal;
945   }
946 
947   IncomingValueMap::const_iterator It = IncomingValues.find(BB);
948   if (It != IncomingValues.end()) return It->second;
949 
950   return OldVal;
951 }
952 
953 /// Create a map from block to value for the operands of a
954 /// given phi.
955 ///
956 /// Create a map from block to value for each non-undef value flowing
957 /// into \p PN.
958 ///
959 /// \param PN The phi we are collecting the map for.
960 /// \param IncomingValues [out] The map from block to value for this phi.
gatherIncomingValuesToPhi(PHINode * PN,IncomingValueMap & IncomingValues)961 static void gatherIncomingValuesToPhi(PHINode *PN,
962                                       IncomingValueMap &IncomingValues) {
963   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
964     BasicBlock *BB = PN->getIncomingBlock(i);
965     Value *V = PN->getIncomingValue(i);
966 
967     if (!isa<UndefValue>(V))
968       IncomingValues.insert(std::make_pair(BB, V));
969   }
970 }
971 
972 /// Replace the incoming undef values to a phi with the values
973 /// from a block-to-value map.
974 ///
975 /// \param PN The phi we are replacing the undefs in.
976 /// \param IncomingValues A map from block to value.
replaceUndefValuesInPhi(PHINode * PN,const IncomingValueMap & IncomingValues)977 static void replaceUndefValuesInPhi(PHINode *PN,
978                                     const IncomingValueMap &IncomingValues) {
979   SmallVector<unsigned> TrueUndefOps;
980   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
981     Value *V = PN->getIncomingValue(i);
982 
983     if (!isa<UndefValue>(V)) continue;
984 
985     BasicBlock *BB = PN->getIncomingBlock(i);
986     IncomingValueMap::const_iterator It = IncomingValues.find(BB);
987 
988     // Keep track of undef/poison incoming values. Those must match, so we fix
989     // them up below if needed.
990     // Note: this is conservatively correct, but we could try harder and group
991     // the undef values per incoming basic block.
992     if (It == IncomingValues.end()) {
993       TrueUndefOps.push_back(i);
994       continue;
995     }
996 
997     // There is a defined value for this incoming block, so map this undef
998     // incoming value to the defined value.
999     PN->setIncomingValue(i, It->second);
1000   }
1001 
1002   // If there are both undef and poison values incoming, then convert those
1003   // values to undef. It is invalid to have different values for the same
1004   // incoming block.
1005   unsigned PoisonCount = count_if(TrueUndefOps, [&](unsigned i) {
1006     return isa<PoisonValue>(PN->getIncomingValue(i));
1007   });
1008   if (PoisonCount != 0 && PoisonCount != TrueUndefOps.size()) {
1009     for (unsigned i : TrueUndefOps)
1010       PN->setIncomingValue(i, UndefValue::get(PN->getType()));
1011   }
1012 }
1013 
1014 // Only when they shares a single common predecessor, return true.
1015 // Only handles cases when BB can't be merged while its predecessors can be
1016 // redirected.
1017 static bool
CanRedirectPredsOfEmptyBBToSucc(BasicBlock * BB,BasicBlock * Succ,const SmallPtrSetImpl<BasicBlock * > & BBPreds,const SmallPtrSetImpl<BasicBlock * > & SuccPreds,BasicBlock * & CommonPred)1018 CanRedirectPredsOfEmptyBBToSucc(BasicBlock *BB, BasicBlock *Succ,
1019                                 const SmallPtrSetImpl<BasicBlock *> &BBPreds,
1020                                 const SmallPtrSetImpl<BasicBlock *> &SuccPreds,
1021                                 BasicBlock *&CommonPred) {
1022 
1023   // There must be phis in BB, otherwise BB will be merged into Succ directly
1024   if (BB->phis().empty() || Succ->phis().empty())
1025     return false;
1026 
1027   // BB must have predecessors not shared that can be redirected to Succ
1028   if (!BB->hasNPredecessorsOrMore(2))
1029     return false;
1030 
1031   // Get single common predecessors of both BB and Succ
1032   for (BasicBlock *SuccPred : SuccPreds) {
1033     if (BBPreds.count(SuccPred)) {
1034       if (CommonPred)
1035         return false;
1036       CommonPred = SuccPred;
1037     }
1038   }
1039 
1040   return true;
1041 }
1042 
1043 /// Replace a value flowing from a block to a phi with
1044 /// potentially multiple instances of that value flowing from the
1045 /// block's predecessors to the phi.
1046 ///
1047 /// \param BB The block with the value flowing into the phi.
1048 /// \param BBPreds The predecessors of BB.
1049 /// \param PN The phi that we are updating.
1050 /// \param CommonPred The common predecessor of BB and PN's BasicBlock
redirectValuesFromPredecessorsToPhi(BasicBlock * BB,const PredBlockVector & BBPreds,PHINode * PN,BasicBlock * CommonPred)1051 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
1052                                                 const PredBlockVector &BBPreds,
1053                                                 PHINode *PN,
1054                                                 BasicBlock *CommonPred) {
1055   Value *OldVal = PN->removeIncomingValue(BB, false);
1056   assert(OldVal && "No entry in PHI for Pred BB!");
1057 
1058   IncomingValueMap IncomingValues;
1059 
1060   // We are merging two blocks - BB, and the block containing PN - and
1061   // as a result we need to redirect edges from the predecessors of BB
1062   // to go to the block containing PN, and update PN
1063   // accordingly. Since we allow merging blocks in the case where the
1064   // predecessor and successor blocks both share some predecessors,
1065   // and where some of those common predecessors might have undef
1066   // values flowing into PN, we want to rewrite those values to be
1067   // consistent with the non-undef values.
1068 
1069   gatherIncomingValuesToPhi(PN, IncomingValues);
1070 
1071   // If this incoming value is one of the PHI nodes in BB, the new entries
1072   // in the PHI node are the entries from the old PHI.
1073   if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
1074     PHINode *OldValPN = cast<PHINode>(OldVal);
1075     for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
1076       // Note that, since we are merging phi nodes and BB and Succ might
1077       // have common predecessors, we could end up with a phi node with
1078       // identical incoming branches. This will be cleaned up later (and
1079       // will trigger asserts if we try to clean it up now, without also
1080       // simplifying the corresponding conditional branch).
1081       BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
1082 
1083       if (PredBB == CommonPred)
1084         continue;
1085 
1086       Value *PredVal = OldValPN->getIncomingValue(i);
1087       Value *Selected =
1088           selectIncomingValueForBlock(PredVal, PredBB, IncomingValues);
1089 
1090       // And add a new incoming value for this predecessor for the
1091       // newly retargeted branch.
1092       PN->addIncoming(Selected, PredBB);
1093     }
1094     if (CommonPred)
1095       PN->addIncoming(OldValPN->getIncomingValueForBlock(CommonPred), BB);
1096 
1097   } else {
1098     for (BasicBlock *PredBB : BBPreds) {
1099       // Update existing incoming values in PN for this
1100       // predecessor of BB.
1101       if (PredBB == CommonPred)
1102         continue;
1103 
1104       Value *Selected =
1105           selectIncomingValueForBlock(OldVal, PredBB, IncomingValues);
1106 
1107       // And add a new incoming value for this predecessor for the
1108       // newly retargeted branch.
1109       PN->addIncoming(Selected, PredBB);
1110     }
1111     if (CommonPred)
1112       PN->addIncoming(OldVal, BB);
1113   }
1114 
1115   replaceUndefValuesInPhi(PN, IncomingValues);
1116 }
1117 
TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock * BB,DomTreeUpdater * DTU)1118 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB,
1119                                                    DomTreeUpdater *DTU) {
1120   assert(BB != &BB->getParent()->getEntryBlock() &&
1121          "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
1122 
1123   // We can't simplify infinite loops.
1124   BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
1125   if (BB == Succ)
1126     return false;
1127 
1128   SmallPtrSet<BasicBlock *, 16> BBPreds(pred_begin(BB), pred_end(BB));
1129   SmallPtrSet<BasicBlock *, 16> SuccPreds(pred_begin(Succ), pred_end(Succ));
1130 
1131   // The single common predecessor of BB and Succ when BB cannot be killed
1132   BasicBlock *CommonPred = nullptr;
1133 
1134   bool BBKillable = CanPropagatePredecessorsForPHIs(BB, Succ, BBPreds);
1135 
1136   // Even if we can not fold bB into Succ, we may be able to redirect the
1137   // predecessors of BB to Succ.
1138   bool BBPhisMergeable =
1139       BBKillable ||
1140       CanRedirectPredsOfEmptyBBToSucc(BB, Succ, BBPreds, SuccPreds, CommonPred);
1141 
1142   if (!BBKillable && !BBPhisMergeable)
1143     return false;
1144 
1145   // Check to see if merging these blocks/phis would cause conflicts for any of
1146   // the phi nodes in BB or Succ. If not, we can safely merge.
1147 
1148   // Check for cases where Succ has multiple predecessors and a PHI node in BB
1149   // has uses which will not disappear when the PHI nodes are merged.  It is
1150   // possible to handle such cases, but difficult: it requires checking whether
1151   // BB dominates Succ, which is non-trivial to calculate in the case where
1152   // Succ has multiple predecessors.  Also, it requires checking whether
1153   // constructing the necessary self-referential PHI node doesn't introduce any
1154   // conflicts; this isn't too difficult, but the previous code for doing this
1155   // was incorrect.
1156   //
1157   // Note that if this check finds a live use, BB dominates Succ, so BB is
1158   // something like a loop pre-header (or rarely, a part of an irreducible CFG);
1159   // folding the branch isn't profitable in that case anyway.
1160   if (!Succ->getSinglePredecessor()) {
1161     BasicBlock::iterator BBI = BB->begin();
1162     while (isa<PHINode>(*BBI)) {
1163       for (Use &U : BBI->uses()) {
1164         if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
1165           if (PN->getIncomingBlock(U) != BB)
1166             return false;
1167         } else {
1168           return false;
1169         }
1170       }
1171       ++BBI;
1172     }
1173   }
1174 
1175   if (BBPhisMergeable && CommonPred)
1176     LLVM_DEBUG(dbgs() << "Found Common Predecessor between: " << BB->getName()
1177                       << " and " << Succ->getName() << " : "
1178                       << CommonPred->getName() << "\n");
1179 
1180   // 'BB' and 'BB->Pred' are loop latches, bail out to presrve inner loop
1181   // metadata.
1182   //
1183   // FIXME: This is a stop-gap solution to preserve inner-loop metadata given
1184   // current status (that loop metadata is implemented as metadata attached to
1185   // the branch instruction in the loop latch block). To quote from review
1186   // comments, "the current representation of loop metadata (using a loop latch
1187   // terminator attachment) is known to be fundamentally broken. Loop latches
1188   // are not uniquely associated with loops (both in that a latch can be part of
1189   // multiple loops and a loop may have multiple latches). Loop headers are. The
1190   // solution to this problem is also known: Add support for basic block
1191   // metadata, and attach loop metadata to the loop header."
1192   //
1193   // Why bail out:
1194   // In this case, we expect 'BB' is the latch for outer-loop and 'BB->Pred' is
1195   // the latch for inner-loop (see reason below), so bail out to prerserve
1196   // inner-loop metadata rather than eliminating 'BB' and attaching its metadata
1197   // to this inner-loop.
1198   // - The reason we believe 'BB' and 'BB->Pred' have different inner-most
1199   // loops: assuming 'BB' and 'BB->Pred' are from the same inner-most loop L,
1200   // then 'BB' is the header and latch of 'L' and thereby 'L' must consist of
1201   // one self-looping basic block, which is contradictory with the assumption.
1202   //
1203   // To illustrate how inner-loop metadata is dropped:
1204   //
1205   // CFG Before
1206   //
1207   // BB is while.cond.exit, attached with loop metdata md2.
1208   // BB->Pred is for.body, attached with loop metadata md1.
1209   //
1210   //      entry
1211   //        |
1212   //        v
1213   // ---> while.cond   ------------->  while.end
1214   // |       |
1215   // |       v
1216   // |   while.body
1217   // |       |
1218   // |       v
1219   // |    for.body <---- (md1)
1220   // |       |  |______|
1221   // |       v
1222   // |    while.cond.exit (md2)
1223   // |       |
1224   // |_______|
1225   //
1226   // CFG After
1227   //
1228   // while.cond1 is the merge of while.cond.exit and while.cond above.
1229   // for.body is attached with md2, and md1 is dropped.
1230   // If LoopSimplify runs later (as a part of loop pass), it could create
1231   // dedicated exits for inner-loop (essentially adding `while.cond.exit`
1232   // back), but won't it won't see 'md1' nor restore it for the inner-loop.
1233   //
1234   //       entry
1235   //         |
1236   //         v
1237   // ---> while.cond1  ------------->  while.end
1238   // |       |
1239   // |       v
1240   // |   while.body
1241   // |       |
1242   // |       v
1243   // |    for.body <---- (md2)
1244   // |_______|  |______|
1245   if (Instruction *TI = BB->getTerminator())
1246     if (TI->hasMetadata(LLVMContext::MD_loop))
1247       for (BasicBlock *Pred : predecessors(BB))
1248         if (Instruction *PredTI = Pred->getTerminator())
1249           if (PredTI->hasMetadata(LLVMContext::MD_loop))
1250             return false;
1251 
1252   if (BBKillable)
1253     LLVM_DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
1254   else if (BBPhisMergeable)
1255     LLVM_DEBUG(dbgs() << "Merge Phis in Trivial BB: \n" << *BB);
1256 
1257   SmallVector<DominatorTree::UpdateType, 32> Updates;
1258 
1259   if (DTU) {
1260     // To avoid processing the same predecessor more than once.
1261     SmallPtrSet<BasicBlock *, 8> SeenPreds;
1262     // All predecessors of BB (except the common predecessor) will be moved to
1263     // Succ.
1264     Updates.reserve(Updates.size() + 2 * pred_size(BB) + 1);
1265 
1266     for (auto *PredOfBB : predecessors(BB)) {
1267       // Do not modify those common predecessors of BB and Succ
1268       if (!SuccPreds.contains(PredOfBB))
1269         if (SeenPreds.insert(PredOfBB).second)
1270           Updates.push_back({DominatorTree::Insert, PredOfBB, Succ});
1271     }
1272 
1273     SeenPreds.clear();
1274 
1275     for (auto *PredOfBB : predecessors(BB))
1276       // When BB cannot be killed, do not remove the edge between BB and
1277       // CommonPred.
1278       if (SeenPreds.insert(PredOfBB).second && PredOfBB != CommonPred)
1279         Updates.push_back({DominatorTree::Delete, PredOfBB, BB});
1280 
1281     if (BBKillable)
1282       Updates.push_back({DominatorTree::Delete, BB, Succ});
1283   }
1284 
1285   if (isa<PHINode>(Succ->begin())) {
1286     // If there is more than one pred of succ, and there are PHI nodes in
1287     // the successor, then we need to add incoming edges for the PHI nodes
1288     //
1289     const PredBlockVector BBPreds(predecessors(BB));
1290 
1291     // Loop over all of the PHI nodes in the successor of BB.
1292     for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
1293       PHINode *PN = cast<PHINode>(I);
1294       redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN, CommonPred);
1295     }
1296   }
1297 
1298   if (Succ->getSinglePredecessor()) {
1299     // BB is the only predecessor of Succ, so Succ will end up with exactly
1300     // the same predecessors BB had.
1301     // Copy over any phi, debug or lifetime instruction.
1302     BB->getTerminator()->eraseFromParent();
1303     Succ->splice(Succ->getFirstNonPHIIt(), BB);
1304   } else {
1305     while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
1306       // We explicitly check for such uses for merging phis.
1307       assert(PN->use_empty() && "There shouldn't be any uses here!");
1308       PN->eraseFromParent();
1309     }
1310   }
1311 
1312   // If the unconditional branch we replaced contains llvm.loop metadata, we
1313   // add the metadata to the branch instructions in the predecessors.
1314   if (Instruction *TI = BB->getTerminator())
1315     if (MDNode *LoopMD = TI->getMetadata(LLVMContext::MD_loop))
1316       for (BasicBlock *Pred : predecessors(BB))
1317         Pred->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopMD);
1318 
1319   if (BBKillable) {
1320     // Everything that jumped to BB now goes to Succ.
1321     BB->replaceAllUsesWith(Succ);
1322 
1323     if (!Succ->hasName())
1324       Succ->takeName(BB);
1325 
1326     // Clear the successor list of BB to match updates applying to DTU later.
1327     if (BB->getTerminator())
1328       BB->back().eraseFromParent();
1329 
1330     new UnreachableInst(BB->getContext(), BB);
1331     assert(succ_empty(BB) && "The successor list of BB isn't empty before "
1332                              "applying corresponding DTU updates.");
1333   } else if (BBPhisMergeable) {
1334     //  Everything except CommonPred that jumped to BB now goes to Succ.
1335     BB->replaceUsesWithIf(Succ, [BBPreds, CommonPred](Use &U) -> bool {
1336       if (Instruction *UseInst = dyn_cast<Instruction>(U.getUser()))
1337         return UseInst->getParent() != CommonPred &&
1338                BBPreds.contains(UseInst->getParent());
1339       return false;
1340     });
1341   }
1342 
1343   if (DTU)
1344     DTU->applyUpdates(Updates);
1345 
1346   if (BBKillable)
1347     DeleteDeadBlock(BB, DTU);
1348 
1349   return true;
1350 }
1351 
1352 static bool
EliminateDuplicatePHINodesNaiveImpl(BasicBlock * BB,SmallPtrSetImpl<PHINode * > & ToRemove)1353 EliminateDuplicatePHINodesNaiveImpl(BasicBlock *BB,
1354                                     SmallPtrSetImpl<PHINode *> &ToRemove) {
1355   // This implementation doesn't currently consider undef operands
1356   // specially. Theoretically, two phis which are identical except for
1357   // one having an undef where the other doesn't could be collapsed.
1358 
1359   bool Changed = false;
1360 
1361   // Examine each PHI.
1362   // Note that increment of I must *NOT* be in the iteration_expression, since
1363   // we don't want to immediately advance when we restart from the beginning.
1364   for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I);) {
1365     ++I;
1366     // Is there an identical PHI node in this basic block?
1367     // Note that we only look in the upper square's triangle,
1368     // we already checked that the lower triangle PHI's aren't identical.
1369     for (auto J = I; PHINode *DuplicatePN = dyn_cast<PHINode>(J); ++J) {
1370       if (ToRemove.contains(DuplicatePN))
1371         continue;
1372       if (!DuplicatePN->isIdenticalToWhenDefined(PN))
1373         continue;
1374       // A duplicate. Replace this PHI with the base PHI.
1375       ++NumPHICSEs;
1376       DuplicatePN->replaceAllUsesWith(PN);
1377       ToRemove.insert(DuplicatePN);
1378       Changed = true;
1379 
1380       // The RAUW can change PHIs that we already visited.
1381       I = BB->begin();
1382       break; // Start over from the beginning.
1383     }
1384   }
1385   return Changed;
1386 }
1387 
1388 static bool
EliminateDuplicatePHINodesSetBasedImpl(BasicBlock * BB,SmallPtrSetImpl<PHINode * > & ToRemove)1389 EliminateDuplicatePHINodesSetBasedImpl(BasicBlock *BB,
1390                                        SmallPtrSetImpl<PHINode *> &ToRemove) {
1391   // This implementation doesn't currently consider undef operands
1392   // specially. Theoretically, two phis which are identical except for
1393   // one having an undef where the other doesn't could be collapsed.
1394 
1395   struct PHIDenseMapInfo {
1396     static PHINode *getEmptyKey() {
1397       return DenseMapInfo<PHINode *>::getEmptyKey();
1398     }
1399 
1400     static PHINode *getTombstoneKey() {
1401       return DenseMapInfo<PHINode *>::getTombstoneKey();
1402     }
1403 
1404     static bool isSentinel(PHINode *PN) {
1405       return PN == getEmptyKey() || PN == getTombstoneKey();
1406     }
1407 
1408     // WARNING: this logic must be kept in sync with
1409     //          Instruction::isIdenticalToWhenDefined()!
1410     static unsigned getHashValueImpl(PHINode *PN) {
1411       // Compute a hash value on the operands. Instcombine will likely have
1412       // sorted them, which helps expose duplicates, but we have to check all
1413       // the operands to be safe in case instcombine hasn't run.
1414       return static_cast<unsigned>(hash_combine(
1415           hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
1416           hash_combine_range(PN->block_begin(), PN->block_end())));
1417     }
1418 
1419     static unsigned getHashValue(PHINode *PN) {
1420 #ifndef NDEBUG
1421       // If -phicse-debug-hash was specified, return a constant -- this
1422       // will force all hashing to collide, so we'll exhaustively search
1423       // the table for a match, and the assertion in isEqual will fire if
1424       // there's a bug causing equal keys to hash differently.
1425       if (PHICSEDebugHash)
1426         return 0;
1427 #endif
1428       return getHashValueImpl(PN);
1429     }
1430 
1431     static bool isEqualImpl(PHINode *LHS, PHINode *RHS) {
1432       if (isSentinel(LHS) || isSentinel(RHS))
1433         return LHS == RHS;
1434       return LHS->isIdenticalTo(RHS);
1435     }
1436 
1437     static bool isEqual(PHINode *LHS, PHINode *RHS) {
1438       // These comparisons are nontrivial, so assert that equality implies
1439       // hash equality (DenseMap demands this as an invariant).
1440       bool Result = isEqualImpl(LHS, RHS);
1441       assert(!Result || (isSentinel(LHS) && LHS == RHS) ||
1442              getHashValueImpl(LHS) == getHashValueImpl(RHS));
1443       return Result;
1444     }
1445   };
1446 
1447   // Set of unique PHINodes.
1448   DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
1449   PHISet.reserve(4 * PHICSENumPHISmallSize);
1450 
1451   // Examine each PHI.
1452   bool Changed = false;
1453   for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
1454     if (ToRemove.contains(PN))
1455       continue;
1456     auto Inserted = PHISet.insert(PN);
1457     if (!Inserted.second) {
1458       // A duplicate. Replace this PHI with its duplicate.
1459       ++NumPHICSEs;
1460       PN->replaceAllUsesWith(*Inserted.first);
1461       ToRemove.insert(PN);
1462       Changed = true;
1463 
1464       // The RAUW can change PHIs that we already visited. Start over from the
1465       // beginning.
1466       PHISet.clear();
1467       I = BB->begin();
1468     }
1469   }
1470 
1471   return Changed;
1472 }
1473 
EliminateDuplicatePHINodes(BasicBlock * BB,SmallPtrSetImpl<PHINode * > & ToRemove)1474 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB,
1475                                       SmallPtrSetImpl<PHINode *> &ToRemove) {
1476   if (
1477 #ifndef NDEBUG
1478       !PHICSEDebugHash &&
1479 #endif
1480       hasNItemsOrLess(BB->phis(), PHICSENumPHISmallSize))
1481     return EliminateDuplicatePHINodesNaiveImpl(BB, ToRemove);
1482   return EliminateDuplicatePHINodesSetBasedImpl(BB, ToRemove);
1483 }
1484 
EliminateDuplicatePHINodes(BasicBlock * BB)1485 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
1486   SmallPtrSet<PHINode *, 8> ToRemove;
1487   bool Changed = EliminateDuplicatePHINodes(BB, ToRemove);
1488   for (PHINode *PN : ToRemove)
1489     PN->eraseFromParent();
1490   return Changed;
1491 }
1492 
tryEnforceAlignment(Value * V,Align PrefAlign,const DataLayout & DL)1493 Align llvm::tryEnforceAlignment(Value *V, Align PrefAlign,
1494                                 const DataLayout &DL) {
1495   V = V->stripPointerCasts();
1496 
1497   if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1498     // TODO: Ideally, this function would not be called if PrefAlign is smaller
1499     // than the current alignment, as the known bits calculation should have
1500     // already taken it into account. However, this is not always the case,
1501     // as computeKnownBits() has a depth limit, while stripPointerCasts()
1502     // doesn't.
1503     Align CurrentAlign = AI->getAlign();
1504     if (PrefAlign <= CurrentAlign)
1505       return CurrentAlign;
1506 
1507     // If the preferred alignment is greater than the natural stack alignment
1508     // then don't round up. This avoids dynamic stack realignment.
1509     if (DL.exceedsNaturalStackAlignment(PrefAlign))
1510       return CurrentAlign;
1511     AI->setAlignment(PrefAlign);
1512     return PrefAlign;
1513   }
1514 
1515   if (auto *GO = dyn_cast<GlobalObject>(V)) {
1516     // TODO: as above, this shouldn't be necessary.
1517     Align CurrentAlign = GO->getPointerAlignment(DL);
1518     if (PrefAlign <= CurrentAlign)
1519       return CurrentAlign;
1520 
1521     // If there is a large requested alignment and we can, bump up the alignment
1522     // of the global.  If the memory we set aside for the global may not be the
1523     // memory used by the final program then it is impossible for us to reliably
1524     // enforce the preferred alignment.
1525     if (!GO->canIncreaseAlignment())
1526       return CurrentAlign;
1527 
1528     if (GO->isThreadLocal()) {
1529       unsigned MaxTLSAlign = GO->getParent()->getMaxTLSAlignment() / CHAR_BIT;
1530       if (MaxTLSAlign && PrefAlign > Align(MaxTLSAlign))
1531         PrefAlign = Align(MaxTLSAlign);
1532     }
1533 
1534     GO->setAlignment(PrefAlign);
1535     return PrefAlign;
1536   }
1537 
1538   return Align(1);
1539 }
1540 
getOrEnforceKnownAlignment(Value * V,MaybeAlign PrefAlign,const DataLayout & DL,const Instruction * CxtI,AssumptionCache * AC,const DominatorTree * DT)1541 Align llvm::getOrEnforceKnownAlignment(Value *V, MaybeAlign PrefAlign,
1542                                        const DataLayout &DL,
1543                                        const Instruction *CxtI,
1544                                        AssumptionCache *AC,
1545                                        const DominatorTree *DT) {
1546   assert(V->getType()->isPointerTy() &&
1547          "getOrEnforceKnownAlignment expects a pointer!");
1548 
1549   KnownBits Known = computeKnownBits(V, DL, 0, AC, CxtI, DT);
1550   unsigned TrailZ = Known.countMinTrailingZeros();
1551 
1552   // Avoid trouble with ridiculously large TrailZ values, such as
1553   // those computed from a null pointer.
1554   // LLVM doesn't support alignments larger than (1 << MaxAlignmentExponent).
1555   TrailZ = std::min(TrailZ, +Value::MaxAlignmentExponent);
1556 
1557   Align Alignment = Align(1ull << std::min(Known.getBitWidth() - 1, TrailZ));
1558 
1559   if (PrefAlign && *PrefAlign > Alignment)
1560     Alignment = std::max(Alignment, tryEnforceAlignment(V, *PrefAlign, DL));
1561 
1562   // We don't need to make any adjustment.
1563   return Alignment;
1564 }
1565 
1566 ///===---------------------------------------------------------------------===//
1567 ///  Dbg Intrinsic utilities
1568 ///
1569 
1570 /// See if there is a dbg.value intrinsic for DIVar for the PHI node.
PhiHasDebugValue(DILocalVariable * DIVar,DIExpression * DIExpr,PHINode * APN)1571 static bool PhiHasDebugValue(DILocalVariable *DIVar,
1572                              DIExpression *DIExpr,
1573                              PHINode *APN) {
1574   // Since we can't guarantee that the original dbg.declare intrinsic
1575   // is removed by LowerDbgDeclare(), we need to make sure that we are
1576   // not inserting the same dbg.value intrinsic over and over.
1577   SmallVector<DbgValueInst *, 1> DbgValues;
1578   SmallVector<DbgVariableRecord *, 1> DbgVariableRecords;
1579   findDbgValues(DbgValues, APN, &DbgVariableRecords);
1580   for (auto *DVI : DbgValues) {
1581     assert(is_contained(DVI->getValues(), APN));
1582     if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr))
1583       return true;
1584   }
1585   for (auto *DVR : DbgVariableRecords) {
1586     assert(is_contained(DVR->location_ops(), APN));
1587     if ((DVR->getVariable() == DIVar) && (DVR->getExpression() == DIExpr))
1588       return true;
1589   }
1590   return false;
1591 }
1592 
1593 /// Check if the alloc size of \p ValTy is large enough to cover the variable
1594 /// (or fragment of the variable) described by \p DII.
1595 ///
1596 /// This is primarily intended as a helper for the different
1597 /// ConvertDebugDeclareToDebugValue functions. The dbg.declare that is converted
1598 /// describes an alloca'd variable, so we need to use the alloc size of the
1599 /// value when doing the comparison. E.g. an i1 value will be identified as
1600 /// covering an n-bit fragment, if the store size of i1 is at least n bits.
valueCoversEntireFragment(Type * ValTy,DbgVariableIntrinsic * DII)1601 static bool valueCoversEntireFragment(Type *ValTy, DbgVariableIntrinsic *DII) {
1602   const DataLayout &DL = DII->getDataLayout();
1603   TypeSize ValueSize = DL.getTypeAllocSizeInBits(ValTy);
1604   if (std::optional<uint64_t> FragmentSize =
1605           DII->getExpression()->getActiveBits(DII->getVariable()))
1606     return TypeSize::isKnownGE(ValueSize, TypeSize::getFixed(*FragmentSize));
1607 
1608   // We can't always calculate the size of the DI variable (e.g. if it is a
1609   // VLA). Try to use the size of the alloca that the dbg intrinsic describes
1610   // intead.
1611   if (DII->isAddressOfVariable()) {
1612     // DII should have exactly 1 location when it is an address.
1613     assert(DII->getNumVariableLocationOps() == 1 &&
1614            "address of variable must have exactly 1 location operand.");
1615     if (auto *AI =
1616             dyn_cast_or_null<AllocaInst>(DII->getVariableLocationOp(0))) {
1617       if (std::optional<TypeSize> FragmentSize =
1618               AI->getAllocationSizeInBits(DL)) {
1619         return TypeSize::isKnownGE(ValueSize, *FragmentSize);
1620       }
1621     }
1622   }
1623   // Could not determine size of variable. Conservatively return false.
1624   return false;
1625 }
1626 // RemoveDIs: duplicate implementation of the above, using DbgVariableRecords,
1627 // the replacement for dbg.values.
valueCoversEntireFragment(Type * ValTy,DbgVariableRecord * DVR)1628 static bool valueCoversEntireFragment(Type *ValTy, DbgVariableRecord *DVR) {
1629   const DataLayout &DL = DVR->getModule()->getDataLayout();
1630   TypeSize ValueSize = DL.getTypeAllocSizeInBits(ValTy);
1631   if (std::optional<uint64_t> FragmentSize =
1632           DVR->getExpression()->getActiveBits(DVR->getVariable()))
1633     return TypeSize::isKnownGE(ValueSize, TypeSize::getFixed(*FragmentSize));
1634 
1635   // We can't always calculate the size of the DI variable (e.g. if it is a
1636   // VLA). Try to use the size of the alloca that the dbg intrinsic describes
1637   // intead.
1638   if (DVR->isAddressOfVariable()) {
1639     // DVR should have exactly 1 location when it is an address.
1640     assert(DVR->getNumVariableLocationOps() == 1 &&
1641            "address of variable must have exactly 1 location operand.");
1642     if (auto *AI =
1643             dyn_cast_or_null<AllocaInst>(DVR->getVariableLocationOp(0))) {
1644       if (std::optional<TypeSize> FragmentSize = AI->getAllocationSizeInBits(DL)) {
1645         return TypeSize::isKnownGE(ValueSize, *FragmentSize);
1646       }
1647     }
1648   }
1649   // Could not determine size of variable. Conservatively return false.
1650   return false;
1651 }
1652 
insertDbgValueOrDbgVariableRecord(DIBuilder & Builder,Value * DV,DILocalVariable * DIVar,DIExpression * DIExpr,const DebugLoc & NewLoc,BasicBlock::iterator Instr)1653 static void insertDbgValueOrDbgVariableRecord(DIBuilder &Builder, Value *DV,
1654                                               DILocalVariable *DIVar,
1655                                               DIExpression *DIExpr,
1656                                               const DebugLoc &NewLoc,
1657                                               BasicBlock::iterator Instr) {
1658   if (!UseNewDbgInfoFormat) {
1659     auto DbgVal = Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, NewLoc,
1660                                                   (Instruction *)nullptr);
1661     DbgVal.get<Instruction *>()->insertBefore(Instr);
1662   } else {
1663     // RemoveDIs: if we're using the new debug-info format, allocate a
1664     // DbgVariableRecord directly instead of a dbg.value intrinsic.
1665     ValueAsMetadata *DVAM = ValueAsMetadata::get(DV);
1666     DbgVariableRecord *DV =
1667         new DbgVariableRecord(DVAM, DIVar, DIExpr, NewLoc.get());
1668     Instr->getParent()->insertDbgRecordBefore(DV, Instr);
1669   }
1670 }
1671 
insertDbgValueOrDbgVariableRecordAfter(DIBuilder & Builder,Value * DV,DILocalVariable * DIVar,DIExpression * DIExpr,const DebugLoc & NewLoc,BasicBlock::iterator Instr)1672 static void insertDbgValueOrDbgVariableRecordAfter(
1673     DIBuilder &Builder, Value *DV, DILocalVariable *DIVar, DIExpression *DIExpr,
1674     const DebugLoc &NewLoc, BasicBlock::iterator Instr) {
1675   if (!UseNewDbgInfoFormat) {
1676     auto DbgVal = Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, NewLoc,
1677                                                   (Instruction *)nullptr);
1678     DbgVal.get<Instruction *>()->insertAfter(&*Instr);
1679   } else {
1680     // RemoveDIs: if we're using the new debug-info format, allocate a
1681     // DbgVariableRecord directly instead of a dbg.value intrinsic.
1682     ValueAsMetadata *DVAM = ValueAsMetadata::get(DV);
1683     DbgVariableRecord *DV =
1684         new DbgVariableRecord(DVAM, DIVar, DIExpr, NewLoc.get());
1685     Instr->getParent()->insertDbgRecordAfter(DV, &*Instr);
1686   }
1687 }
1688 
1689 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1690 /// that has an associated llvm.dbg.declare intrinsic.
ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic * DII,StoreInst * SI,DIBuilder & Builder)1691 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic *DII,
1692                                            StoreInst *SI, DIBuilder &Builder) {
1693   assert(DII->isAddressOfVariable() || isa<DbgAssignIntrinsic>(DII));
1694   auto *DIVar = DII->getVariable();
1695   assert(DIVar && "Missing variable");
1696   auto *DIExpr = DII->getExpression();
1697   Value *DV = SI->getValueOperand();
1698 
1699   DebugLoc NewLoc = getDebugValueLoc(DII);
1700 
1701   // If the alloca describes the variable itself, i.e. the expression in the
1702   // dbg.declare doesn't start with a dereference, we can perform the
1703   // conversion if the value covers the entire fragment of DII.
1704   // If the alloca describes the *address* of DIVar, i.e. DIExpr is
1705   // *just* a DW_OP_deref, we use DV as is for the dbg.value.
1706   // We conservatively ignore other dereferences, because the following two are
1707   // not equivalent:
1708   //     dbg.declare(alloca, ..., !Expr(deref, plus_uconstant, 2))
1709   //     dbg.value(DV, ..., !Expr(deref, plus_uconstant, 2))
1710   // The former is adding 2 to the address of the variable, whereas the latter
1711   // is adding 2 to the value of the variable. As such, we insist on just a
1712   // deref expression.
1713   bool CanConvert =
1714       DIExpr->isDeref() || (!DIExpr->startsWithDeref() &&
1715                             valueCoversEntireFragment(DV->getType(), DII));
1716   if (CanConvert) {
1717     insertDbgValueOrDbgVariableRecord(Builder, DV, DIVar, DIExpr, NewLoc,
1718                                       SI->getIterator());
1719     return;
1720   }
1721 
1722   // FIXME: If storing to a part of the variable described by the dbg.declare,
1723   // then we want to insert a dbg.value for the corresponding fragment.
1724   LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: " << *DII
1725                     << '\n');
1726   // For now, when there is a store to parts of the variable (but we do not
1727   // know which part) we insert an dbg.value intrinsic to indicate that we
1728   // know nothing about the variable's content.
1729   DV = UndefValue::get(DV->getType());
1730   insertDbgValueOrDbgVariableRecord(Builder, DV, DIVar, DIExpr, NewLoc,
1731                                     SI->getIterator());
1732 }
1733 
1734 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1735 /// that has an associated llvm.dbg.declare intrinsic.
ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic * DII,LoadInst * LI,DIBuilder & Builder)1736 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic *DII,
1737                                            LoadInst *LI, DIBuilder &Builder) {
1738   auto *DIVar = DII->getVariable();
1739   auto *DIExpr = DII->getExpression();
1740   assert(DIVar && "Missing variable");
1741 
1742   if (!valueCoversEntireFragment(LI->getType(), DII)) {
1743     // FIXME: If only referring to a part of the variable described by the
1744     // dbg.declare, then we want to insert a dbg.value for the corresponding
1745     // fragment.
1746     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: "
1747                       << *DII << '\n');
1748     return;
1749   }
1750 
1751   DebugLoc NewLoc = getDebugValueLoc(DII);
1752 
1753   // We are now tracking the loaded value instead of the address. In the
1754   // future if multi-location support is added to the IR, it might be
1755   // preferable to keep tracking both the loaded value and the original
1756   // address in case the alloca can not be elided.
1757   insertDbgValueOrDbgVariableRecordAfter(Builder, LI, DIVar, DIExpr, NewLoc,
1758                                          LI->getIterator());
1759 }
1760 
ConvertDebugDeclareToDebugValue(DbgVariableRecord * DVR,StoreInst * SI,DIBuilder & Builder)1761 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableRecord *DVR,
1762                                            StoreInst *SI, DIBuilder &Builder) {
1763   assert(DVR->isAddressOfVariable() || DVR->isDbgAssign());
1764   auto *DIVar = DVR->getVariable();
1765   assert(DIVar && "Missing variable");
1766   auto *DIExpr = DVR->getExpression();
1767   Value *DV = SI->getValueOperand();
1768 
1769   DebugLoc NewLoc = getDebugValueLoc(DVR);
1770 
1771   // If the alloca describes the variable itself, i.e. the expression in the
1772   // dbg.declare doesn't start with a dereference, we can perform the
1773   // conversion if the value covers the entire fragment of DII.
1774   // If the alloca describes the *address* of DIVar, i.e. DIExpr is
1775   // *just* a DW_OP_deref, we use DV as is for the dbg.value.
1776   // We conservatively ignore other dereferences, because the following two are
1777   // not equivalent:
1778   //     dbg.declare(alloca, ..., !Expr(deref, plus_uconstant, 2))
1779   //     dbg.value(DV, ..., !Expr(deref, plus_uconstant, 2))
1780   // The former is adding 2 to the address of the variable, whereas the latter
1781   // is adding 2 to the value of the variable. As such, we insist on just a
1782   // deref expression.
1783   bool CanConvert =
1784       DIExpr->isDeref() || (!DIExpr->startsWithDeref() &&
1785                             valueCoversEntireFragment(DV->getType(), DVR));
1786   if (CanConvert) {
1787     insertDbgValueOrDbgVariableRecord(Builder, DV, DIVar, DIExpr, NewLoc,
1788                                       SI->getIterator());
1789     return;
1790   }
1791 
1792   // FIXME: If storing to a part of the variable described by the dbg.declare,
1793   // then we want to insert a dbg.value for the corresponding fragment.
1794   LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: " << *DVR
1795                     << '\n');
1796   assert(UseNewDbgInfoFormat);
1797 
1798   // For now, when there is a store to parts of the variable (but we do not
1799   // know which part) we insert an dbg.value intrinsic to indicate that we
1800   // know nothing about the variable's content.
1801   DV = UndefValue::get(DV->getType());
1802   ValueAsMetadata *DVAM = ValueAsMetadata::get(DV);
1803   DbgVariableRecord *NewDVR =
1804       new DbgVariableRecord(DVAM, DIVar, DIExpr, NewLoc.get());
1805   SI->getParent()->insertDbgRecordBefore(NewDVR, SI->getIterator());
1806 }
1807 
1808 /// Inserts a llvm.dbg.value intrinsic after a phi that has an associated
1809 /// llvm.dbg.declare intrinsic.
ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic * DII,PHINode * APN,DIBuilder & Builder)1810 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic *DII,
1811                                            PHINode *APN, DIBuilder &Builder) {
1812   auto *DIVar = DII->getVariable();
1813   auto *DIExpr = DII->getExpression();
1814   assert(DIVar && "Missing variable");
1815 
1816   if (PhiHasDebugValue(DIVar, DIExpr, APN))
1817     return;
1818 
1819   if (!valueCoversEntireFragment(APN->getType(), DII)) {
1820     // FIXME: If only referring to a part of the variable described by the
1821     // dbg.declare, then we want to insert a dbg.value for the corresponding
1822     // fragment.
1823     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: "
1824                       << *DII << '\n');
1825     return;
1826   }
1827 
1828   BasicBlock *BB = APN->getParent();
1829   auto InsertionPt = BB->getFirstInsertionPt();
1830 
1831   DebugLoc NewLoc = getDebugValueLoc(DII);
1832 
1833   // The block may be a catchswitch block, which does not have a valid
1834   // insertion point.
1835   // FIXME: Insert dbg.value markers in the successors when appropriate.
1836   if (InsertionPt != BB->end()) {
1837     insertDbgValueOrDbgVariableRecord(Builder, APN, DIVar, DIExpr, NewLoc,
1838                                       InsertionPt);
1839   }
1840 }
1841 
ConvertDebugDeclareToDebugValue(DbgVariableRecord * DVR,LoadInst * LI,DIBuilder & Builder)1842 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableRecord *DVR, LoadInst *LI,
1843                                            DIBuilder &Builder) {
1844   auto *DIVar = DVR->getVariable();
1845   auto *DIExpr = DVR->getExpression();
1846   assert(DIVar && "Missing variable");
1847 
1848   if (!valueCoversEntireFragment(LI->getType(), DVR)) {
1849     // FIXME: If only referring to a part of the variable described by the
1850     // dbg.declare, then we want to insert a DbgVariableRecord for the
1851     // corresponding fragment.
1852     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to DbgVariableRecord: "
1853                       << *DVR << '\n');
1854     return;
1855   }
1856 
1857   DebugLoc NewLoc = getDebugValueLoc(DVR);
1858 
1859   // We are now tracking the loaded value instead of the address. In the
1860   // future if multi-location support is added to the IR, it might be
1861   // preferable to keep tracking both the loaded value and the original
1862   // address in case the alloca can not be elided.
1863   assert(UseNewDbgInfoFormat);
1864 
1865   // Create a DbgVariableRecord directly and insert.
1866   ValueAsMetadata *LIVAM = ValueAsMetadata::get(LI);
1867   DbgVariableRecord *DV =
1868       new DbgVariableRecord(LIVAM, DIVar, DIExpr, NewLoc.get());
1869   LI->getParent()->insertDbgRecordAfter(DV, LI);
1870 }
1871 
1872 /// Determine whether this alloca is either a VLA or an array.
isArray(AllocaInst * AI)1873 static bool isArray(AllocaInst *AI) {
1874   return AI->isArrayAllocation() ||
1875          (AI->getAllocatedType() && AI->getAllocatedType()->isArrayTy());
1876 }
1877 
1878 /// Determine whether this alloca is a structure.
isStructure(AllocaInst * AI)1879 static bool isStructure(AllocaInst *AI) {
1880   return AI->getAllocatedType() && AI->getAllocatedType()->isStructTy();
1881 }
ConvertDebugDeclareToDebugValue(DbgVariableRecord * DVR,PHINode * APN,DIBuilder & Builder)1882 void llvm::ConvertDebugDeclareToDebugValue(DbgVariableRecord *DVR, PHINode *APN,
1883                                            DIBuilder &Builder) {
1884   auto *DIVar = DVR->getVariable();
1885   auto *DIExpr = DVR->getExpression();
1886   assert(DIVar && "Missing variable");
1887 
1888   if (PhiHasDebugValue(DIVar, DIExpr, APN))
1889     return;
1890 
1891   if (!valueCoversEntireFragment(APN->getType(), DVR)) {
1892     // FIXME: If only referring to a part of the variable described by the
1893     // dbg.declare, then we want to insert a DbgVariableRecord for the
1894     // corresponding fragment.
1895     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to DbgVariableRecord: "
1896                       << *DVR << '\n');
1897     return;
1898   }
1899 
1900   BasicBlock *BB = APN->getParent();
1901   auto InsertionPt = BB->getFirstInsertionPt();
1902 
1903   DebugLoc NewLoc = getDebugValueLoc(DVR);
1904 
1905   // The block may be a catchswitch block, which does not have a valid
1906   // insertion point.
1907   // FIXME: Insert DbgVariableRecord markers in the successors when appropriate.
1908   if (InsertionPt != BB->end()) {
1909     insertDbgValueOrDbgVariableRecord(Builder, APN, DIVar, DIExpr, NewLoc,
1910                                       InsertionPt);
1911   }
1912 }
1913 
1914 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1915 /// of llvm.dbg.value intrinsics.
LowerDbgDeclare(Function & F)1916 bool llvm::LowerDbgDeclare(Function &F) {
1917   bool Changed = false;
1918   DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1919   SmallVector<DbgDeclareInst *, 4> Dbgs;
1920   SmallVector<DbgVariableRecord *> DVRs;
1921   for (auto &FI : F) {
1922     for (Instruction &BI : FI) {
1923       if (auto *DDI = dyn_cast<DbgDeclareInst>(&BI))
1924         Dbgs.push_back(DDI);
1925       for (DbgVariableRecord &DVR : filterDbgVars(BI.getDbgRecordRange())) {
1926         if (DVR.getType() == DbgVariableRecord::LocationType::Declare)
1927           DVRs.push_back(&DVR);
1928       }
1929     }
1930   }
1931 
1932   if (Dbgs.empty() && DVRs.empty())
1933     return Changed;
1934 
1935   auto LowerOne = [&](auto *DDI) {
1936     AllocaInst *AI =
1937         dyn_cast_or_null<AllocaInst>(DDI->getVariableLocationOp(0));
1938     // If this is an alloca for a scalar variable, insert a dbg.value
1939     // at each load and store to the alloca and erase the dbg.declare.
1940     // The dbg.values allow tracking a variable even if it is not
1941     // stored on the stack, while the dbg.declare can only describe
1942     // the stack slot (and at a lexical-scope granularity). Later
1943     // passes will attempt to elide the stack slot.
1944     if (!AI || isArray(AI) || isStructure(AI))
1945       return;
1946 
1947     // A volatile load/store means that the alloca can't be elided anyway.
1948     if (llvm::any_of(AI->users(), [](User *U) -> bool {
1949           if (LoadInst *LI = dyn_cast<LoadInst>(U))
1950             return LI->isVolatile();
1951           if (StoreInst *SI = dyn_cast<StoreInst>(U))
1952             return SI->isVolatile();
1953           return false;
1954         }))
1955       return;
1956 
1957     SmallVector<const Value *, 8> WorkList;
1958     WorkList.push_back(AI);
1959     while (!WorkList.empty()) {
1960       const Value *V = WorkList.pop_back_val();
1961       for (const auto &AIUse : V->uses()) {
1962         User *U = AIUse.getUser();
1963         if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1964           if (AIUse.getOperandNo() == 1)
1965             ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1966         } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1967           ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1968         } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1969           // This is a call by-value or some other instruction that takes a
1970           // pointer to the variable. Insert a *value* intrinsic that describes
1971           // the variable by dereferencing the alloca.
1972           if (!CI->isLifetimeStartOrEnd()) {
1973             DebugLoc NewLoc = getDebugValueLoc(DDI);
1974             auto *DerefExpr =
1975                 DIExpression::append(DDI->getExpression(), dwarf::DW_OP_deref);
1976             insertDbgValueOrDbgVariableRecord(DIB, AI, DDI->getVariable(),
1977                                               DerefExpr, NewLoc,
1978                                               CI->getIterator());
1979           }
1980         } else if (BitCastInst *BI = dyn_cast<BitCastInst>(U)) {
1981           if (BI->getType()->isPointerTy())
1982             WorkList.push_back(BI);
1983         }
1984       }
1985     }
1986     DDI->eraseFromParent();
1987     Changed = true;
1988   };
1989 
1990   for_each(Dbgs, LowerOne);
1991   for_each(DVRs, LowerOne);
1992 
1993   if (Changed)
1994     for (BasicBlock &BB : F)
1995       RemoveRedundantDbgInstrs(&BB);
1996 
1997   return Changed;
1998 }
1999 
2000 // RemoveDIs: re-implementation of insertDebugValuesForPHIs, but which pulls the
2001 // debug-info out of the block's DbgVariableRecords rather than dbg.value
2002 // intrinsics.
2003 static void
insertDbgVariableRecordsForPHIs(BasicBlock * BB,SmallVectorImpl<PHINode * > & InsertedPHIs)2004 insertDbgVariableRecordsForPHIs(BasicBlock *BB,
2005                                 SmallVectorImpl<PHINode *> &InsertedPHIs) {
2006   assert(BB && "No BasicBlock to clone DbgVariableRecord(s) from.");
2007   if (InsertedPHIs.size() == 0)
2008     return;
2009 
2010   // Map existing PHI nodes to their DbgVariableRecords.
2011   DenseMap<Value *, DbgVariableRecord *> DbgValueMap;
2012   for (auto &I : *BB) {
2013     for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange())) {
2014       for (Value *V : DVR.location_ops())
2015         if (auto *Loc = dyn_cast_or_null<PHINode>(V))
2016           DbgValueMap.insert({Loc, &DVR});
2017     }
2018   }
2019   if (DbgValueMap.size() == 0)
2020     return;
2021 
2022   // Map a pair of the destination BB and old DbgVariableRecord to the new
2023   // DbgVariableRecord, so that if a DbgVariableRecord is being rewritten to use
2024   // more than one of the inserted PHIs in the same destination BB, we can
2025   // update the same DbgVariableRecord with all the new PHIs instead of creating
2026   // one copy for each.
2027   MapVector<std::pair<BasicBlock *, DbgVariableRecord *>, DbgVariableRecord *>
2028       NewDbgValueMap;
2029   // Then iterate through the new PHIs and look to see if they use one of the
2030   // previously mapped PHIs. If so, create a new DbgVariableRecord that will
2031   // propagate the info through the new PHI. If we use more than one new PHI in
2032   // a single destination BB with the same old dbg.value, merge the updates so
2033   // that we get a single new DbgVariableRecord with all the new PHIs.
2034   for (auto PHI : InsertedPHIs) {
2035     BasicBlock *Parent = PHI->getParent();
2036     // Avoid inserting a debug-info record into an EH block.
2037     if (Parent->getFirstNonPHI()->isEHPad())
2038       continue;
2039     for (auto VI : PHI->operand_values()) {
2040       auto V = DbgValueMap.find(VI);
2041       if (V != DbgValueMap.end()) {
2042         DbgVariableRecord *DbgII = cast<DbgVariableRecord>(V->second);
2043         auto NewDI = NewDbgValueMap.find({Parent, DbgII});
2044         if (NewDI == NewDbgValueMap.end()) {
2045           DbgVariableRecord *NewDbgII = DbgII->clone();
2046           NewDI = NewDbgValueMap.insert({{Parent, DbgII}, NewDbgII}).first;
2047         }
2048         DbgVariableRecord *NewDbgII = NewDI->second;
2049         // If PHI contains VI as an operand more than once, we may
2050         // replaced it in NewDbgII; confirm that it is present.
2051         if (is_contained(NewDbgII->location_ops(), VI))
2052           NewDbgII->replaceVariableLocationOp(VI, PHI);
2053       }
2054     }
2055   }
2056   // Insert the new DbgVariableRecords into their destination blocks.
2057   for (auto DI : NewDbgValueMap) {
2058     BasicBlock *Parent = DI.first.first;
2059     DbgVariableRecord *NewDbgII = DI.second;
2060     auto InsertionPt = Parent->getFirstInsertionPt();
2061     assert(InsertionPt != Parent->end() && "Ill-formed basic block");
2062 
2063     Parent->insertDbgRecordBefore(NewDbgII, InsertionPt);
2064   }
2065 }
2066 
2067 /// Propagate dbg.value intrinsics through the newly inserted PHIs.
insertDebugValuesForPHIs(BasicBlock * BB,SmallVectorImpl<PHINode * > & InsertedPHIs)2068 void llvm::insertDebugValuesForPHIs(BasicBlock *BB,
2069                                     SmallVectorImpl<PHINode *> &InsertedPHIs) {
2070   assert(BB && "No BasicBlock to clone dbg.value(s) from.");
2071   if (InsertedPHIs.size() == 0)
2072     return;
2073 
2074   insertDbgVariableRecordsForPHIs(BB, InsertedPHIs);
2075 
2076   // Map existing PHI nodes to their dbg.values.
2077   ValueToValueMapTy DbgValueMap;
2078   for (auto &I : *BB) {
2079     if (auto DbgII = dyn_cast<DbgVariableIntrinsic>(&I)) {
2080       for (Value *V : DbgII->location_ops())
2081         if (auto *Loc = dyn_cast_or_null<PHINode>(V))
2082           DbgValueMap.insert({Loc, DbgII});
2083     }
2084   }
2085   if (DbgValueMap.size() == 0)
2086     return;
2087 
2088   // Map a pair of the destination BB and old dbg.value to the new dbg.value,
2089   // so that if a dbg.value is being rewritten to use more than one of the
2090   // inserted PHIs in the same destination BB, we can update the same dbg.value
2091   // with all the new PHIs instead of creating one copy for each.
2092   MapVector<std::pair<BasicBlock *, DbgVariableIntrinsic *>,
2093             DbgVariableIntrinsic *>
2094       NewDbgValueMap;
2095   // Then iterate through the new PHIs and look to see if they use one of the
2096   // previously mapped PHIs. If so, create a new dbg.value intrinsic that will
2097   // propagate the info through the new PHI. If we use more than one new PHI in
2098   // a single destination BB with the same old dbg.value, merge the updates so
2099   // that we get a single new dbg.value with all the new PHIs.
2100   for (auto *PHI : InsertedPHIs) {
2101     BasicBlock *Parent = PHI->getParent();
2102     // Avoid inserting an intrinsic into an EH block.
2103     if (Parent->getFirstNonPHI()->isEHPad())
2104       continue;
2105     for (auto *VI : PHI->operand_values()) {
2106       auto V = DbgValueMap.find(VI);
2107       if (V != DbgValueMap.end()) {
2108         auto *DbgII = cast<DbgVariableIntrinsic>(V->second);
2109         auto NewDI = NewDbgValueMap.find({Parent, DbgII});
2110         if (NewDI == NewDbgValueMap.end()) {
2111           auto *NewDbgII = cast<DbgVariableIntrinsic>(DbgII->clone());
2112           NewDI = NewDbgValueMap.insert({{Parent, DbgII}, NewDbgII}).first;
2113         }
2114         DbgVariableIntrinsic *NewDbgII = NewDI->second;
2115         // If PHI contains VI as an operand more than once, we may
2116         // replaced it in NewDbgII; confirm that it is present.
2117         if (is_contained(NewDbgII->location_ops(), VI))
2118           NewDbgII->replaceVariableLocationOp(VI, PHI);
2119       }
2120     }
2121   }
2122   // Insert thew new dbg.values into their destination blocks.
2123   for (auto DI : NewDbgValueMap) {
2124     BasicBlock *Parent = DI.first.first;
2125     auto *NewDbgII = DI.second;
2126     auto InsertionPt = Parent->getFirstInsertionPt();
2127     assert(InsertionPt != Parent->end() && "Ill-formed basic block");
2128     NewDbgII->insertBefore(&*InsertionPt);
2129   }
2130 }
2131 
replaceDbgDeclare(Value * Address,Value * NewAddress,DIBuilder & Builder,uint8_t DIExprFlags,int Offset)2132 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
2133                              DIBuilder &Builder, uint8_t DIExprFlags,
2134                              int Offset) {
2135   TinyPtrVector<DbgDeclareInst *> DbgDeclares = findDbgDeclares(Address);
2136   TinyPtrVector<DbgVariableRecord *> DVRDeclares = findDVRDeclares(Address);
2137 
2138   auto ReplaceOne = [&](auto *DII) {
2139     assert(DII->getVariable() && "Missing variable");
2140     auto *DIExpr = DII->getExpression();
2141     DIExpr = DIExpression::prepend(DIExpr, DIExprFlags, Offset);
2142     DII->setExpression(DIExpr);
2143     DII->replaceVariableLocationOp(Address, NewAddress);
2144   };
2145 
2146   for_each(DbgDeclares, ReplaceOne);
2147   for_each(DVRDeclares, ReplaceOne);
2148 
2149   return !DbgDeclares.empty() || !DVRDeclares.empty();
2150 }
2151 
updateOneDbgValueForAlloca(const DebugLoc & Loc,DILocalVariable * DIVar,DIExpression * DIExpr,Value * NewAddress,DbgValueInst * DVI,DbgVariableRecord * DVR,DIBuilder & Builder,int Offset)2152 static void updateOneDbgValueForAlloca(const DebugLoc &Loc,
2153                                        DILocalVariable *DIVar,
2154                                        DIExpression *DIExpr, Value *NewAddress,
2155                                        DbgValueInst *DVI,
2156                                        DbgVariableRecord *DVR,
2157                                        DIBuilder &Builder, int Offset) {
2158   assert(DIVar && "Missing variable");
2159 
2160   // This is an alloca-based dbg.value/DbgVariableRecord. The first thing it
2161   // should do with the alloca pointer is dereference it. Otherwise we don't
2162   // know how to handle it and give up.
2163   if (!DIExpr || DIExpr->getNumElements() < 1 ||
2164       DIExpr->getElement(0) != dwarf::DW_OP_deref)
2165     return;
2166 
2167   // Insert the offset before the first deref.
2168   if (Offset)
2169     DIExpr = DIExpression::prepend(DIExpr, 0, Offset);
2170 
2171   if (DVI) {
2172     DVI->setExpression(DIExpr);
2173     DVI->replaceVariableLocationOp(0u, NewAddress);
2174   } else {
2175     assert(DVR);
2176     DVR->setExpression(DIExpr);
2177     DVR->replaceVariableLocationOp(0u, NewAddress);
2178   }
2179 }
2180 
replaceDbgValueForAlloca(AllocaInst * AI,Value * NewAllocaAddress,DIBuilder & Builder,int Offset)2181 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
2182                                     DIBuilder &Builder, int Offset) {
2183   SmallVector<DbgValueInst *, 1> DbgUsers;
2184   SmallVector<DbgVariableRecord *, 1> DPUsers;
2185   findDbgValues(DbgUsers, AI, &DPUsers);
2186 
2187   // Attempt to replace dbg.values that use this alloca.
2188   for (auto *DVI : DbgUsers)
2189     updateOneDbgValueForAlloca(DVI->getDebugLoc(), DVI->getVariable(),
2190                                DVI->getExpression(), NewAllocaAddress, DVI,
2191                                nullptr, Builder, Offset);
2192 
2193   // Replace any DbgVariableRecords that use this alloca.
2194   for (DbgVariableRecord *DVR : DPUsers)
2195     updateOneDbgValueForAlloca(DVR->getDebugLoc(), DVR->getVariable(),
2196                                DVR->getExpression(), NewAllocaAddress, nullptr,
2197                                DVR, Builder, Offset);
2198 }
2199 
2200 /// Where possible to salvage debug information for \p I do so.
2201 /// If not possible mark undef.
salvageDebugInfo(Instruction & I)2202 void llvm::salvageDebugInfo(Instruction &I) {
2203   SmallVector<DbgVariableIntrinsic *, 1> DbgUsers;
2204   SmallVector<DbgVariableRecord *, 1> DPUsers;
2205   findDbgUsers(DbgUsers, &I, &DPUsers);
2206   salvageDebugInfoForDbgValues(I, DbgUsers, DPUsers);
2207 }
2208 
salvageDbgAssignAddress(T * Assign)2209 template <typename T> static void salvageDbgAssignAddress(T *Assign) {
2210   Instruction *I = dyn_cast<Instruction>(Assign->getAddress());
2211   // Only instructions can be salvaged at the moment.
2212   if (!I)
2213     return;
2214 
2215   assert(!Assign->getAddressExpression()->getFragmentInfo().has_value() &&
2216          "address-expression shouldn't have fragment info");
2217 
2218   // The address component of a dbg.assign cannot be variadic.
2219   uint64_t CurrentLocOps = 0;
2220   SmallVector<Value *, 4> AdditionalValues;
2221   SmallVector<uint64_t, 16> Ops;
2222   Value *NewV = salvageDebugInfoImpl(*I, CurrentLocOps, Ops, AdditionalValues);
2223 
2224   // Check if the salvage failed.
2225   if (!NewV)
2226     return;
2227 
2228   DIExpression *SalvagedExpr = DIExpression::appendOpsToArg(
2229       Assign->getAddressExpression(), Ops, 0, /*StackValue=*/false);
2230   assert(!SalvagedExpr->getFragmentInfo().has_value() &&
2231          "address-expression shouldn't have fragment info");
2232 
2233   SalvagedExpr = SalvagedExpr->foldConstantMath();
2234 
2235   // Salvage succeeds if no additional values are required.
2236   if (AdditionalValues.empty()) {
2237     Assign->setAddress(NewV);
2238     Assign->setAddressExpression(SalvagedExpr);
2239   } else {
2240     Assign->setKillAddress();
2241   }
2242 }
2243 
salvageDebugInfoForDbgValues(Instruction & I,ArrayRef<DbgVariableIntrinsic * > DbgUsers,ArrayRef<DbgVariableRecord * > DPUsers)2244 void llvm::salvageDebugInfoForDbgValues(
2245     Instruction &I, ArrayRef<DbgVariableIntrinsic *> DbgUsers,
2246     ArrayRef<DbgVariableRecord *> DPUsers) {
2247   // These are arbitrary chosen limits on the maximum number of values and the
2248   // maximum size of a debug expression we can salvage up to, used for
2249   // performance reasons.
2250   const unsigned MaxDebugArgs = 16;
2251   const unsigned MaxExpressionSize = 128;
2252   bool Salvaged = false;
2253 
2254   for (auto *DII : DbgUsers) {
2255     if (auto *DAI = dyn_cast<DbgAssignIntrinsic>(DII)) {
2256       if (DAI->getAddress() == &I) {
2257         salvageDbgAssignAddress(DAI);
2258         Salvaged = true;
2259       }
2260       if (DAI->getValue() != &I)
2261         continue;
2262     }
2263 
2264     // Do not add DW_OP_stack_value for DbgDeclare, because they are implicitly
2265     // pointing out the value as a DWARF memory location description.
2266     bool StackValue = isa<DbgValueInst>(DII);
2267     auto DIILocation = DII->location_ops();
2268     assert(
2269         is_contained(DIILocation, &I) &&
2270         "DbgVariableIntrinsic must use salvaged instruction as its location");
2271     SmallVector<Value *, 4> AdditionalValues;
2272     // `I` may appear more than once in DII's location ops, and each use of `I`
2273     // must be updated in the DIExpression and potentially have additional
2274     // values added; thus we call salvageDebugInfoImpl for each `I` instance in
2275     // DIILocation.
2276     Value *Op0 = nullptr;
2277     DIExpression *SalvagedExpr = DII->getExpression();
2278     auto LocItr = find(DIILocation, &I);
2279     while (SalvagedExpr && LocItr != DIILocation.end()) {
2280       SmallVector<uint64_t, 16> Ops;
2281       unsigned LocNo = std::distance(DIILocation.begin(), LocItr);
2282       uint64_t CurrentLocOps = SalvagedExpr->getNumLocationOperands();
2283       Op0 = salvageDebugInfoImpl(I, CurrentLocOps, Ops, AdditionalValues);
2284       if (!Op0)
2285         break;
2286       SalvagedExpr =
2287           DIExpression::appendOpsToArg(SalvagedExpr, Ops, LocNo, StackValue);
2288       LocItr = std::find(++LocItr, DIILocation.end(), &I);
2289     }
2290     // salvageDebugInfoImpl should fail on examining the first element of
2291     // DbgUsers, or none of them.
2292     if (!Op0)
2293       break;
2294 
2295     SalvagedExpr = SalvagedExpr->foldConstantMath();
2296     DII->replaceVariableLocationOp(&I, Op0);
2297     bool IsValidSalvageExpr = SalvagedExpr->getNumElements() <= MaxExpressionSize;
2298     if (AdditionalValues.empty() && IsValidSalvageExpr) {
2299       DII->setExpression(SalvagedExpr);
2300     } else if (isa<DbgValueInst>(DII) && IsValidSalvageExpr &&
2301                DII->getNumVariableLocationOps() + AdditionalValues.size() <=
2302                    MaxDebugArgs) {
2303       DII->addVariableLocationOps(AdditionalValues, SalvagedExpr);
2304     } else {
2305       // Do not salvage using DIArgList for dbg.declare, as it is not currently
2306       // supported in those instructions. Also do not salvage if the resulting
2307       // DIArgList would contain an unreasonably large number of values.
2308       DII->setKillLocation();
2309     }
2310     LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n');
2311     Salvaged = true;
2312   }
2313   // Duplicate of above block for DbgVariableRecords.
2314   for (auto *DVR : DPUsers) {
2315     if (DVR->isDbgAssign()) {
2316       if (DVR->getAddress() == &I) {
2317         salvageDbgAssignAddress(DVR);
2318         Salvaged = true;
2319       }
2320       if (DVR->getValue() != &I)
2321         continue;
2322     }
2323 
2324     // Do not add DW_OP_stack_value for DbgDeclare and DbgAddr, because they
2325     // are implicitly pointing out the value as a DWARF memory location
2326     // description.
2327     bool StackValue =
2328         DVR->getType() != DbgVariableRecord::LocationType::Declare;
2329     auto DVRLocation = DVR->location_ops();
2330     assert(
2331         is_contained(DVRLocation, &I) &&
2332         "DbgVariableIntrinsic must use salvaged instruction as its location");
2333     SmallVector<Value *, 4> AdditionalValues;
2334     // 'I' may appear more than once in DVR's location ops, and each use of 'I'
2335     // must be updated in the DIExpression and potentially have additional
2336     // values added; thus we call salvageDebugInfoImpl for each 'I' instance in
2337     // DVRLocation.
2338     Value *Op0 = nullptr;
2339     DIExpression *SalvagedExpr = DVR->getExpression();
2340     auto LocItr = find(DVRLocation, &I);
2341     while (SalvagedExpr && LocItr != DVRLocation.end()) {
2342       SmallVector<uint64_t, 16> Ops;
2343       unsigned LocNo = std::distance(DVRLocation.begin(), LocItr);
2344       uint64_t CurrentLocOps = SalvagedExpr->getNumLocationOperands();
2345       Op0 = salvageDebugInfoImpl(I, CurrentLocOps, Ops, AdditionalValues);
2346       if (!Op0)
2347         break;
2348       SalvagedExpr =
2349           DIExpression::appendOpsToArg(SalvagedExpr, Ops, LocNo, StackValue);
2350       LocItr = std::find(++LocItr, DVRLocation.end(), &I);
2351     }
2352     // salvageDebugInfoImpl should fail on examining the first element of
2353     // DbgUsers, or none of them.
2354     if (!Op0)
2355       break;
2356 
2357     SalvagedExpr = SalvagedExpr->foldConstantMath();
2358     DVR->replaceVariableLocationOp(&I, Op0);
2359     bool IsValidSalvageExpr =
2360         SalvagedExpr->getNumElements() <= MaxExpressionSize;
2361     if (AdditionalValues.empty() && IsValidSalvageExpr) {
2362       DVR->setExpression(SalvagedExpr);
2363     } else if (DVR->getType() != DbgVariableRecord::LocationType::Declare &&
2364                IsValidSalvageExpr &&
2365                DVR->getNumVariableLocationOps() + AdditionalValues.size() <=
2366                    MaxDebugArgs) {
2367       DVR->addVariableLocationOps(AdditionalValues, SalvagedExpr);
2368     } else {
2369       // Do not salvage using DIArgList for dbg.addr/dbg.declare, as it is
2370       // currently only valid for stack value expressions.
2371       // Also do not salvage if the resulting DIArgList would contain an
2372       // unreasonably large number of values.
2373       DVR->setKillLocation();
2374     }
2375     LLVM_DEBUG(dbgs() << "SALVAGE: " << DVR << '\n');
2376     Salvaged = true;
2377   }
2378 
2379   if (Salvaged)
2380     return;
2381 
2382   for (auto *DII : DbgUsers)
2383     DII->setKillLocation();
2384 
2385   for (auto *DVR : DPUsers)
2386     DVR->setKillLocation();
2387 }
2388 
getSalvageOpsForGEP(GetElementPtrInst * GEP,const DataLayout & DL,uint64_t CurrentLocOps,SmallVectorImpl<uint64_t> & Opcodes,SmallVectorImpl<Value * > & AdditionalValues)2389 Value *getSalvageOpsForGEP(GetElementPtrInst *GEP, const DataLayout &DL,
2390                            uint64_t CurrentLocOps,
2391                            SmallVectorImpl<uint64_t> &Opcodes,
2392                            SmallVectorImpl<Value *> &AdditionalValues) {
2393   unsigned BitWidth = DL.getIndexSizeInBits(GEP->getPointerAddressSpace());
2394   // Rewrite a GEP into a DIExpression.
2395   MapVector<Value *, APInt> VariableOffsets;
2396   APInt ConstantOffset(BitWidth, 0);
2397   if (!GEP->collectOffset(DL, BitWidth, VariableOffsets, ConstantOffset))
2398     return nullptr;
2399   if (!VariableOffsets.empty() && !CurrentLocOps) {
2400     Opcodes.insert(Opcodes.begin(), {dwarf::DW_OP_LLVM_arg, 0});
2401     CurrentLocOps = 1;
2402   }
2403   for (const auto &Offset : VariableOffsets) {
2404     AdditionalValues.push_back(Offset.first);
2405     assert(Offset.second.isStrictlyPositive() &&
2406            "Expected strictly positive multiplier for offset.");
2407     Opcodes.append({dwarf::DW_OP_LLVM_arg, CurrentLocOps++, dwarf::DW_OP_constu,
2408                     Offset.second.getZExtValue(), dwarf::DW_OP_mul,
2409                     dwarf::DW_OP_plus});
2410   }
2411   DIExpression::appendOffset(Opcodes, ConstantOffset.getSExtValue());
2412   return GEP->getOperand(0);
2413 }
2414 
getDwarfOpForBinOp(Instruction::BinaryOps Opcode)2415 uint64_t getDwarfOpForBinOp(Instruction::BinaryOps Opcode) {
2416   switch (Opcode) {
2417   case Instruction::Add:
2418     return dwarf::DW_OP_plus;
2419   case Instruction::Sub:
2420     return dwarf::DW_OP_minus;
2421   case Instruction::Mul:
2422     return dwarf::DW_OP_mul;
2423   case Instruction::SDiv:
2424     return dwarf::DW_OP_div;
2425   case Instruction::SRem:
2426     return dwarf::DW_OP_mod;
2427   case Instruction::Or:
2428     return dwarf::DW_OP_or;
2429   case Instruction::And:
2430     return dwarf::DW_OP_and;
2431   case Instruction::Xor:
2432     return dwarf::DW_OP_xor;
2433   case Instruction::Shl:
2434     return dwarf::DW_OP_shl;
2435   case Instruction::LShr:
2436     return dwarf::DW_OP_shr;
2437   case Instruction::AShr:
2438     return dwarf::DW_OP_shra;
2439   default:
2440     // TODO: Salvage from each kind of binop we know about.
2441     return 0;
2442   }
2443 }
2444 
handleSSAValueOperands(uint64_t CurrentLocOps,SmallVectorImpl<uint64_t> & Opcodes,SmallVectorImpl<Value * > & AdditionalValues,Instruction * I)2445 static void handleSSAValueOperands(uint64_t CurrentLocOps,
2446                                    SmallVectorImpl<uint64_t> &Opcodes,
2447                                    SmallVectorImpl<Value *> &AdditionalValues,
2448                                    Instruction *I) {
2449   if (!CurrentLocOps) {
2450     Opcodes.append({dwarf::DW_OP_LLVM_arg, 0});
2451     CurrentLocOps = 1;
2452   }
2453   Opcodes.append({dwarf::DW_OP_LLVM_arg, CurrentLocOps});
2454   AdditionalValues.push_back(I->getOperand(1));
2455 }
2456 
getSalvageOpsForBinOp(BinaryOperator * BI,uint64_t CurrentLocOps,SmallVectorImpl<uint64_t> & Opcodes,SmallVectorImpl<Value * > & AdditionalValues)2457 Value *getSalvageOpsForBinOp(BinaryOperator *BI, uint64_t CurrentLocOps,
2458                              SmallVectorImpl<uint64_t> &Opcodes,
2459                              SmallVectorImpl<Value *> &AdditionalValues) {
2460   // Handle binary operations with constant integer operands as a special case.
2461   auto *ConstInt = dyn_cast<ConstantInt>(BI->getOperand(1));
2462   // Values wider than 64 bits cannot be represented within a DIExpression.
2463   if (ConstInt && ConstInt->getBitWidth() > 64)
2464     return nullptr;
2465 
2466   Instruction::BinaryOps BinOpcode = BI->getOpcode();
2467   // Push any Constant Int operand onto the expression stack.
2468   if (ConstInt) {
2469     uint64_t Val = ConstInt->getSExtValue();
2470     // Add or Sub Instructions with a constant operand can potentially be
2471     // simplified.
2472     if (BinOpcode == Instruction::Add || BinOpcode == Instruction::Sub) {
2473       uint64_t Offset = BinOpcode == Instruction::Add ? Val : -int64_t(Val);
2474       DIExpression::appendOffset(Opcodes, Offset);
2475       return BI->getOperand(0);
2476     }
2477     Opcodes.append({dwarf::DW_OP_constu, Val});
2478   } else {
2479     handleSSAValueOperands(CurrentLocOps, Opcodes, AdditionalValues, BI);
2480   }
2481 
2482   // Add salvaged binary operator to expression stack, if it has a valid
2483   // representation in a DIExpression.
2484   uint64_t DwarfBinOp = getDwarfOpForBinOp(BinOpcode);
2485   if (!DwarfBinOp)
2486     return nullptr;
2487   Opcodes.push_back(DwarfBinOp);
2488   return BI->getOperand(0);
2489 }
2490 
getDwarfOpForIcmpPred(CmpInst::Predicate Pred)2491 uint64_t getDwarfOpForIcmpPred(CmpInst::Predicate Pred) {
2492   // The signedness of the operation is implicit in the typed stack, signed and
2493   // unsigned instructions map to the same DWARF opcode.
2494   switch (Pred) {
2495   case CmpInst::ICMP_EQ:
2496     return dwarf::DW_OP_eq;
2497   case CmpInst::ICMP_NE:
2498     return dwarf::DW_OP_ne;
2499   case CmpInst::ICMP_UGT:
2500   case CmpInst::ICMP_SGT:
2501     return dwarf::DW_OP_gt;
2502   case CmpInst::ICMP_UGE:
2503   case CmpInst::ICMP_SGE:
2504     return dwarf::DW_OP_ge;
2505   case CmpInst::ICMP_ULT:
2506   case CmpInst::ICMP_SLT:
2507     return dwarf::DW_OP_lt;
2508   case CmpInst::ICMP_ULE:
2509   case CmpInst::ICMP_SLE:
2510     return dwarf::DW_OP_le;
2511   default:
2512     return 0;
2513   }
2514 }
2515 
getSalvageOpsForIcmpOp(ICmpInst * Icmp,uint64_t CurrentLocOps,SmallVectorImpl<uint64_t> & Opcodes,SmallVectorImpl<Value * > & AdditionalValues)2516 Value *getSalvageOpsForIcmpOp(ICmpInst *Icmp, uint64_t CurrentLocOps,
2517                               SmallVectorImpl<uint64_t> &Opcodes,
2518                               SmallVectorImpl<Value *> &AdditionalValues) {
2519   // Handle icmp operations with constant integer operands as a special case.
2520   auto *ConstInt = dyn_cast<ConstantInt>(Icmp->getOperand(1));
2521   // Values wider than 64 bits cannot be represented within a DIExpression.
2522   if (ConstInt && ConstInt->getBitWidth() > 64)
2523     return nullptr;
2524   // Push any Constant Int operand onto the expression stack.
2525   if (ConstInt) {
2526     if (Icmp->isSigned())
2527       Opcodes.push_back(dwarf::DW_OP_consts);
2528     else
2529       Opcodes.push_back(dwarf::DW_OP_constu);
2530     uint64_t Val = ConstInt->getSExtValue();
2531     Opcodes.push_back(Val);
2532   } else {
2533     handleSSAValueOperands(CurrentLocOps, Opcodes, AdditionalValues, Icmp);
2534   }
2535 
2536   // Add salvaged binary operator to expression stack, if it has a valid
2537   // representation in a DIExpression.
2538   uint64_t DwarfIcmpOp = getDwarfOpForIcmpPred(Icmp->getPredicate());
2539   if (!DwarfIcmpOp)
2540     return nullptr;
2541   Opcodes.push_back(DwarfIcmpOp);
2542   return Icmp->getOperand(0);
2543 }
2544 
salvageDebugInfoImpl(Instruction & I,uint64_t CurrentLocOps,SmallVectorImpl<uint64_t> & Ops,SmallVectorImpl<Value * > & AdditionalValues)2545 Value *llvm::salvageDebugInfoImpl(Instruction &I, uint64_t CurrentLocOps,
2546                                   SmallVectorImpl<uint64_t> &Ops,
2547                                   SmallVectorImpl<Value *> &AdditionalValues) {
2548   auto &M = *I.getModule();
2549   auto &DL = M.getDataLayout();
2550 
2551   if (auto *CI = dyn_cast<CastInst>(&I)) {
2552     Value *FromValue = CI->getOperand(0);
2553     // No-op casts are irrelevant for debug info.
2554     if (CI->isNoopCast(DL)) {
2555       return FromValue;
2556     }
2557 
2558     Type *Type = CI->getType();
2559     if (Type->isPointerTy())
2560       Type = DL.getIntPtrType(Type);
2561     // Casts other than Trunc, SExt, or ZExt to scalar types cannot be salvaged.
2562     if (Type->isVectorTy() ||
2563         !(isa<TruncInst>(&I) || isa<SExtInst>(&I) || isa<ZExtInst>(&I) ||
2564           isa<IntToPtrInst>(&I) || isa<PtrToIntInst>(&I)))
2565       return nullptr;
2566 
2567     llvm::Type *FromType = FromValue->getType();
2568     if (FromType->isPointerTy())
2569       FromType = DL.getIntPtrType(FromType);
2570 
2571     unsigned FromTypeBitSize = FromType->getScalarSizeInBits();
2572     unsigned ToTypeBitSize = Type->getScalarSizeInBits();
2573 
2574     auto ExtOps = DIExpression::getExtOps(FromTypeBitSize, ToTypeBitSize,
2575                                           isa<SExtInst>(&I));
2576     Ops.append(ExtOps.begin(), ExtOps.end());
2577     return FromValue;
2578   }
2579 
2580   if (auto *GEP = dyn_cast<GetElementPtrInst>(&I))
2581     return getSalvageOpsForGEP(GEP, DL, CurrentLocOps, Ops, AdditionalValues);
2582   if (auto *BI = dyn_cast<BinaryOperator>(&I))
2583     return getSalvageOpsForBinOp(BI, CurrentLocOps, Ops, AdditionalValues);
2584   if (auto *IC = dyn_cast<ICmpInst>(&I))
2585     return getSalvageOpsForIcmpOp(IC, CurrentLocOps, Ops, AdditionalValues);
2586 
2587   // *Not* to do: we should not attempt to salvage load instructions,
2588   // because the validity and lifetime of a dbg.value containing
2589   // DW_OP_deref becomes difficult to analyze. See PR40628 for examples.
2590   return nullptr;
2591 }
2592 
2593 /// A replacement for a dbg.value expression.
2594 using DbgValReplacement = std::optional<DIExpression *>;
2595 
2596 /// Point debug users of \p From to \p To using exprs given by \p RewriteExpr,
2597 /// possibly moving/undefing users to prevent use-before-def. Returns true if
2598 /// changes are made.
rewriteDebugUsers(Instruction & From,Value & To,Instruction & DomPoint,DominatorTree & DT,function_ref<DbgValReplacement (DbgVariableIntrinsic & DII)> RewriteExpr,function_ref<DbgValReplacement (DbgVariableRecord & DVR)> RewriteDVRExpr)2599 static bool rewriteDebugUsers(
2600     Instruction &From, Value &To, Instruction &DomPoint, DominatorTree &DT,
2601     function_ref<DbgValReplacement(DbgVariableIntrinsic &DII)> RewriteExpr,
2602     function_ref<DbgValReplacement(DbgVariableRecord &DVR)> RewriteDVRExpr) {
2603   // Find debug users of From.
2604   SmallVector<DbgVariableIntrinsic *, 1> Users;
2605   SmallVector<DbgVariableRecord *, 1> DPUsers;
2606   findDbgUsers(Users, &From, &DPUsers);
2607   if (Users.empty() && DPUsers.empty())
2608     return false;
2609 
2610   // Prevent use-before-def of To.
2611   bool Changed = false;
2612 
2613   SmallPtrSet<DbgVariableIntrinsic *, 1> UndefOrSalvage;
2614   SmallPtrSet<DbgVariableRecord *, 1> UndefOrSalvageDVR;
2615   if (isa<Instruction>(&To)) {
2616     bool DomPointAfterFrom = From.getNextNonDebugInstruction() == &DomPoint;
2617 
2618     for (auto *DII : Users) {
2619       // It's common to see a debug user between From and DomPoint. Move it
2620       // after DomPoint to preserve the variable update without any reordering.
2621       if (DomPointAfterFrom && DII->getNextNonDebugInstruction() == &DomPoint) {
2622         LLVM_DEBUG(dbgs() << "MOVE:  " << *DII << '\n');
2623         DII->moveAfter(&DomPoint);
2624         Changed = true;
2625 
2626       // Users which otherwise aren't dominated by the replacement value must
2627       // be salvaged or deleted.
2628       } else if (!DT.dominates(&DomPoint, DII)) {
2629         UndefOrSalvage.insert(DII);
2630       }
2631     }
2632 
2633     // DbgVariableRecord implementation of the above.
2634     for (auto *DVR : DPUsers) {
2635       Instruction *MarkedInstr = DVR->getMarker()->MarkedInstr;
2636       Instruction *NextNonDebug = MarkedInstr;
2637       // The next instruction might still be a dbg.declare, skip over it.
2638       if (isa<DbgVariableIntrinsic>(NextNonDebug))
2639         NextNonDebug = NextNonDebug->getNextNonDebugInstruction();
2640 
2641       if (DomPointAfterFrom && NextNonDebug == &DomPoint) {
2642         LLVM_DEBUG(dbgs() << "MOVE:  " << *DVR << '\n');
2643         DVR->removeFromParent();
2644         // Ensure there's a marker.
2645         DomPoint.getParent()->insertDbgRecordAfter(DVR, &DomPoint);
2646         Changed = true;
2647       } else if (!DT.dominates(&DomPoint, MarkedInstr)) {
2648         UndefOrSalvageDVR.insert(DVR);
2649       }
2650     }
2651   }
2652 
2653   // Update debug users without use-before-def risk.
2654   for (auto *DII : Users) {
2655     if (UndefOrSalvage.count(DII))
2656       continue;
2657 
2658     DbgValReplacement DVRepl = RewriteExpr(*DII);
2659     if (!DVRepl)
2660       continue;
2661 
2662     DII->replaceVariableLocationOp(&From, &To);
2663     DII->setExpression(*DVRepl);
2664     LLVM_DEBUG(dbgs() << "REWRITE:  " << *DII << '\n');
2665     Changed = true;
2666   }
2667   for (auto *DVR : DPUsers) {
2668     if (UndefOrSalvageDVR.count(DVR))
2669       continue;
2670 
2671     DbgValReplacement DVRepl = RewriteDVRExpr(*DVR);
2672     if (!DVRepl)
2673       continue;
2674 
2675     DVR->replaceVariableLocationOp(&From, &To);
2676     DVR->setExpression(*DVRepl);
2677     LLVM_DEBUG(dbgs() << "REWRITE:  " << DVR << '\n');
2678     Changed = true;
2679   }
2680 
2681   if (!UndefOrSalvage.empty() || !UndefOrSalvageDVR.empty()) {
2682     // Try to salvage the remaining debug users.
2683     salvageDebugInfo(From);
2684     Changed = true;
2685   }
2686 
2687   return Changed;
2688 }
2689 
2690 /// Check if a bitcast between a value of type \p FromTy to type \p ToTy would
2691 /// losslessly preserve the bits and semantics of the value. This predicate is
2692 /// symmetric, i.e swapping \p FromTy and \p ToTy should give the same result.
2693 ///
2694 /// Note that Type::canLosslesslyBitCastTo is not suitable here because it
2695 /// allows semantically unequivalent bitcasts, such as <2 x i64> -> <4 x i32>,
2696 /// and also does not allow lossless pointer <-> integer conversions.
isBitCastSemanticsPreserving(const DataLayout & DL,Type * FromTy,Type * ToTy)2697 static bool isBitCastSemanticsPreserving(const DataLayout &DL, Type *FromTy,
2698                                          Type *ToTy) {
2699   // Trivially compatible types.
2700   if (FromTy == ToTy)
2701     return true;
2702 
2703   // Handle compatible pointer <-> integer conversions.
2704   if (FromTy->isIntOrPtrTy() && ToTy->isIntOrPtrTy()) {
2705     bool SameSize = DL.getTypeSizeInBits(FromTy) == DL.getTypeSizeInBits(ToTy);
2706     bool LosslessConversion = !DL.isNonIntegralPointerType(FromTy) &&
2707                               !DL.isNonIntegralPointerType(ToTy);
2708     return SameSize && LosslessConversion;
2709   }
2710 
2711   // TODO: This is not exhaustive.
2712   return false;
2713 }
2714 
replaceAllDbgUsesWith(Instruction & From,Value & To,Instruction & DomPoint,DominatorTree & DT)2715 bool llvm::replaceAllDbgUsesWith(Instruction &From, Value &To,
2716                                  Instruction &DomPoint, DominatorTree &DT) {
2717   // Exit early if From has no debug users.
2718   if (!From.isUsedByMetadata())
2719     return false;
2720 
2721   assert(&From != &To && "Can't replace something with itself");
2722 
2723   Type *FromTy = From.getType();
2724   Type *ToTy = To.getType();
2725 
2726   auto Identity = [&](DbgVariableIntrinsic &DII) -> DbgValReplacement {
2727     return DII.getExpression();
2728   };
2729   auto IdentityDVR = [&](DbgVariableRecord &DVR) -> DbgValReplacement {
2730     return DVR.getExpression();
2731   };
2732 
2733   // Handle no-op conversions.
2734   Module &M = *From.getModule();
2735   const DataLayout &DL = M.getDataLayout();
2736   if (isBitCastSemanticsPreserving(DL, FromTy, ToTy))
2737     return rewriteDebugUsers(From, To, DomPoint, DT, Identity, IdentityDVR);
2738 
2739   // Handle integer-to-integer widening and narrowing.
2740   // FIXME: Use DW_OP_convert when it's available everywhere.
2741   if (FromTy->isIntegerTy() && ToTy->isIntegerTy()) {
2742     uint64_t FromBits = FromTy->getPrimitiveSizeInBits();
2743     uint64_t ToBits = ToTy->getPrimitiveSizeInBits();
2744     assert(FromBits != ToBits && "Unexpected no-op conversion");
2745 
2746     // When the width of the result grows, assume that a debugger will only
2747     // access the low `FromBits` bits when inspecting the source variable.
2748     if (FromBits < ToBits)
2749       return rewriteDebugUsers(From, To, DomPoint, DT, Identity, IdentityDVR);
2750 
2751     // The width of the result has shrunk. Use sign/zero extension to describe
2752     // the source variable's high bits.
2753     auto SignOrZeroExt = [&](DbgVariableIntrinsic &DII) -> DbgValReplacement {
2754       DILocalVariable *Var = DII.getVariable();
2755 
2756       // Without knowing signedness, sign/zero extension isn't possible.
2757       auto Signedness = Var->getSignedness();
2758       if (!Signedness)
2759         return std::nullopt;
2760 
2761       bool Signed = *Signedness == DIBasicType::Signedness::Signed;
2762       return DIExpression::appendExt(DII.getExpression(), ToBits, FromBits,
2763                                      Signed);
2764     };
2765     // RemoveDIs: duplicate implementation working on DbgVariableRecords rather
2766     // than on dbg.value intrinsics.
2767     auto SignOrZeroExtDVR = [&](DbgVariableRecord &DVR) -> DbgValReplacement {
2768       DILocalVariable *Var = DVR.getVariable();
2769 
2770       // Without knowing signedness, sign/zero extension isn't possible.
2771       auto Signedness = Var->getSignedness();
2772       if (!Signedness)
2773         return std::nullopt;
2774 
2775       bool Signed = *Signedness == DIBasicType::Signedness::Signed;
2776       return DIExpression::appendExt(DVR.getExpression(), ToBits, FromBits,
2777                                      Signed);
2778     };
2779     return rewriteDebugUsers(From, To, DomPoint, DT, SignOrZeroExt,
2780                              SignOrZeroExtDVR);
2781   }
2782 
2783   // TODO: Floating-point conversions, vectors.
2784   return false;
2785 }
2786 
handleUnreachableTerminator(Instruction * I,SmallVectorImpl<Value * > & PoisonedValues)2787 bool llvm::handleUnreachableTerminator(
2788     Instruction *I, SmallVectorImpl<Value *> &PoisonedValues) {
2789   bool Changed = false;
2790   // RemoveDIs: erase debug-info on this instruction manually.
2791   I->dropDbgRecords();
2792   for (Use &U : I->operands()) {
2793     Value *Op = U.get();
2794     if (isa<Instruction>(Op) && !Op->getType()->isTokenTy()) {
2795       U.set(PoisonValue::get(Op->getType()));
2796       PoisonedValues.push_back(Op);
2797       Changed = true;
2798     }
2799   }
2800 
2801   return Changed;
2802 }
2803 
2804 std::pair<unsigned, unsigned>
removeAllNonTerminatorAndEHPadInstructions(BasicBlock * BB)2805 llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) {
2806   unsigned NumDeadInst = 0;
2807   unsigned NumDeadDbgInst = 0;
2808   // Delete the instructions backwards, as it has a reduced likelihood of
2809   // having to update as many def-use and use-def chains.
2810   Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2811   SmallVector<Value *> Uses;
2812   handleUnreachableTerminator(EndInst, Uses);
2813 
2814   while (EndInst != &BB->front()) {
2815     // Delete the next to last instruction.
2816     Instruction *Inst = &*--EndInst->getIterator();
2817     if (!Inst->use_empty() && !Inst->getType()->isTokenTy())
2818       Inst->replaceAllUsesWith(PoisonValue::get(Inst->getType()));
2819     if (Inst->isEHPad() || Inst->getType()->isTokenTy()) {
2820       // EHPads can't have DbgVariableRecords attached to them, but it might be
2821       // possible for things with token type.
2822       Inst->dropDbgRecords();
2823       EndInst = Inst;
2824       continue;
2825     }
2826     if (isa<DbgInfoIntrinsic>(Inst))
2827       ++NumDeadDbgInst;
2828     else
2829       ++NumDeadInst;
2830     // RemoveDIs: erasing debug-info must be done manually.
2831     Inst->dropDbgRecords();
2832     Inst->eraseFromParent();
2833   }
2834   return {NumDeadInst, NumDeadDbgInst};
2835 }
2836 
changeToUnreachable(Instruction * I,bool PreserveLCSSA,DomTreeUpdater * DTU,MemorySSAUpdater * MSSAU)2837 unsigned llvm::changeToUnreachable(Instruction *I, bool PreserveLCSSA,
2838                                    DomTreeUpdater *DTU,
2839                                    MemorySSAUpdater *MSSAU) {
2840   BasicBlock *BB = I->getParent();
2841 
2842   if (MSSAU)
2843     MSSAU->changeToUnreachable(I);
2844 
2845   SmallSet<BasicBlock *, 8> UniqueSuccessors;
2846 
2847   // Loop over all of the successors, removing BB's entry from any PHI
2848   // nodes.
2849   for (BasicBlock *Successor : successors(BB)) {
2850     Successor->removePredecessor(BB, PreserveLCSSA);
2851     if (DTU)
2852       UniqueSuccessors.insert(Successor);
2853   }
2854   auto *UI = new UnreachableInst(I->getContext(), I->getIterator());
2855   UI->setDebugLoc(I->getDebugLoc());
2856 
2857   // All instructions after this are dead.
2858   unsigned NumInstrsRemoved = 0;
2859   BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
2860   while (BBI != BBE) {
2861     if (!BBI->use_empty())
2862       BBI->replaceAllUsesWith(PoisonValue::get(BBI->getType()));
2863     BBI++->eraseFromParent();
2864     ++NumInstrsRemoved;
2865   }
2866   if (DTU) {
2867     SmallVector<DominatorTree::UpdateType, 8> Updates;
2868     Updates.reserve(UniqueSuccessors.size());
2869     for (BasicBlock *UniqueSuccessor : UniqueSuccessors)
2870       Updates.push_back({DominatorTree::Delete, BB, UniqueSuccessor});
2871     DTU->applyUpdates(Updates);
2872   }
2873   BB->flushTerminatorDbgRecords();
2874   return NumInstrsRemoved;
2875 }
2876 
createCallMatchingInvoke(InvokeInst * II)2877 CallInst *llvm::createCallMatchingInvoke(InvokeInst *II) {
2878   SmallVector<Value *, 8> Args(II->args());
2879   SmallVector<OperandBundleDef, 1> OpBundles;
2880   II->getOperandBundlesAsDefs(OpBundles);
2881   CallInst *NewCall = CallInst::Create(II->getFunctionType(),
2882                                        II->getCalledOperand(), Args, OpBundles);
2883   NewCall->setCallingConv(II->getCallingConv());
2884   NewCall->setAttributes(II->getAttributes());
2885   NewCall->setDebugLoc(II->getDebugLoc());
2886   NewCall->copyMetadata(*II);
2887 
2888   // If the invoke had profile metadata, try converting them for CallInst.
2889   uint64_t TotalWeight;
2890   if (NewCall->extractProfTotalWeight(TotalWeight)) {
2891     // Set the total weight if it fits into i32, otherwise reset.
2892     MDBuilder MDB(NewCall->getContext());
2893     auto NewWeights = uint32_t(TotalWeight) != TotalWeight
2894                           ? nullptr
2895                           : MDB.createBranchWeights({uint32_t(TotalWeight)});
2896     NewCall->setMetadata(LLVMContext::MD_prof, NewWeights);
2897   }
2898 
2899   return NewCall;
2900 }
2901 
2902 // changeToCall - Convert the specified invoke into a normal call.
changeToCall(InvokeInst * II,DomTreeUpdater * DTU)2903 CallInst *llvm::changeToCall(InvokeInst *II, DomTreeUpdater *DTU) {
2904   CallInst *NewCall = createCallMatchingInvoke(II);
2905   NewCall->takeName(II);
2906   NewCall->insertBefore(II);
2907   II->replaceAllUsesWith(NewCall);
2908 
2909   // Follow the call by a branch to the normal destination.
2910   BasicBlock *NormalDestBB = II->getNormalDest();
2911   BranchInst::Create(NormalDestBB, II->getIterator());
2912 
2913   // Update PHI nodes in the unwind destination
2914   BasicBlock *BB = II->getParent();
2915   BasicBlock *UnwindDestBB = II->getUnwindDest();
2916   UnwindDestBB->removePredecessor(BB);
2917   II->eraseFromParent();
2918   if (DTU)
2919     DTU->applyUpdates({{DominatorTree::Delete, BB, UnwindDestBB}});
2920   return NewCall;
2921 }
2922 
changeToInvokeAndSplitBasicBlock(CallInst * CI,BasicBlock * UnwindEdge,DomTreeUpdater * DTU)2923 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI,
2924                                                    BasicBlock *UnwindEdge,
2925                                                    DomTreeUpdater *DTU) {
2926   BasicBlock *BB = CI->getParent();
2927 
2928   // Convert this function call into an invoke instruction.  First, split the
2929   // basic block.
2930   BasicBlock *Split = SplitBlock(BB, CI, DTU, /*LI=*/nullptr, /*MSSAU*/ nullptr,
2931                                  CI->getName() + ".noexc");
2932 
2933   // Delete the unconditional branch inserted by SplitBlock
2934   BB->back().eraseFromParent();
2935 
2936   // Create the new invoke instruction.
2937   SmallVector<Value *, 8> InvokeArgs(CI->args());
2938   SmallVector<OperandBundleDef, 1> OpBundles;
2939 
2940   CI->getOperandBundlesAsDefs(OpBundles);
2941 
2942   // Note: we're round tripping operand bundles through memory here, and that
2943   // can potentially be avoided with a cleverer API design that we do not have
2944   // as of this time.
2945 
2946   InvokeInst *II =
2947       InvokeInst::Create(CI->getFunctionType(), CI->getCalledOperand(), Split,
2948                          UnwindEdge, InvokeArgs, OpBundles, CI->getName(), BB);
2949   II->setDebugLoc(CI->getDebugLoc());
2950   II->setCallingConv(CI->getCallingConv());
2951   II->setAttributes(CI->getAttributes());
2952   II->setMetadata(LLVMContext::MD_prof, CI->getMetadata(LLVMContext::MD_prof));
2953 
2954   if (DTU)
2955     DTU->applyUpdates({{DominatorTree::Insert, BB, UnwindEdge}});
2956 
2957   // Make sure that anything using the call now uses the invoke!  This also
2958   // updates the CallGraph if present, because it uses a WeakTrackingVH.
2959   CI->replaceAllUsesWith(II);
2960 
2961   // Delete the original call
2962   Split->front().eraseFromParent();
2963   return Split;
2964 }
2965 
markAliveBlocks(Function & F,SmallPtrSetImpl<BasicBlock * > & Reachable,DomTreeUpdater * DTU=nullptr)2966 static bool markAliveBlocks(Function &F,
2967                             SmallPtrSetImpl<BasicBlock *> &Reachable,
2968                             DomTreeUpdater *DTU = nullptr) {
2969   SmallVector<BasicBlock*, 128> Worklist;
2970   BasicBlock *BB = &F.front();
2971   Worklist.push_back(BB);
2972   Reachable.insert(BB);
2973   bool Changed = false;
2974   do {
2975     BB = Worklist.pop_back_val();
2976 
2977     // Do a quick scan of the basic block, turning any obviously unreachable
2978     // instructions into LLVM unreachable insts.  The instruction combining pass
2979     // canonicalizes unreachable insts into stores to null or undef.
2980     for (Instruction &I : *BB) {
2981       if (auto *CI = dyn_cast<CallInst>(&I)) {
2982         Value *Callee = CI->getCalledOperand();
2983         // Handle intrinsic calls.
2984         if (Function *F = dyn_cast<Function>(Callee)) {
2985           auto IntrinsicID = F->getIntrinsicID();
2986           // Assumptions that are known to be false are equivalent to
2987           // unreachable. Also, if the condition is undefined, then we make the
2988           // choice most beneficial to the optimizer, and choose that to also be
2989           // unreachable.
2990           if (IntrinsicID == Intrinsic::assume) {
2991             if (match(CI->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) {
2992               // Don't insert a call to llvm.trap right before the unreachable.
2993               changeToUnreachable(CI, false, DTU);
2994               Changed = true;
2995               break;
2996             }
2997           } else if (IntrinsicID == Intrinsic::experimental_guard) {
2998             // A call to the guard intrinsic bails out of the current
2999             // compilation unit if the predicate passed to it is false. If the
3000             // predicate is a constant false, then we know the guard will bail
3001             // out of the current compile unconditionally, so all code following
3002             // it is dead.
3003             //
3004             // Note: unlike in llvm.assume, it is not "obviously profitable" for
3005             // guards to treat `undef` as `false` since a guard on `undef` can
3006             // still be useful for widening.
3007             if (match(CI->getArgOperand(0), m_Zero()))
3008               if (!isa<UnreachableInst>(CI->getNextNode())) {
3009                 changeToUnreachable(CI->getNextNode(), false, DTU);
3010                 Changed = true;
3011                 break;
3012               }
3013           }
3014         } else if ((isa<ConstantPointerNull>(Callee) &&
3015                     !NullPointerIsDefined(CI->getFunction(),
3016                                           cast<PointerType>(Callee->getType())
3017                                               ->getAddressSpace())) ||
3018                    isa<UndefValue>(Callee)) {
3019           changeToUnreachable(CI, false, DTU);
3020           Changed = true;
3021           break;
3022         }
3023         if (CI->doesNotReturn() && !CI->isMustTailCall()) {
3024           // If we found a call to a no-return function, insert an unreachable
3025           // instruction after it.  Make sure there isn't *already* one there
3026           // though.
3027           if (!isa<UnreachableInst>(CI->getNextNonDebugInstruction())) {
3028             // Don't insert a call to llvm.trap right before the unreachable.
3029             changeToUnreachable(CI->getNextNonDebugInstruction(), false, DTU);
3030             Changed = true;
3031           }
3032           break;
3033         }
3034       } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
3035         // Store to undef and store to null are undefined and used to signal
3036         // that they should be changed to unreachable by passes that can't
3037         // modify the CFG.
3038 
3039         // Don't touch volatile stores.
3040         if (SI->isVolatile()) continue;
3041 
3042         Value *Ptr = SI->getOperand(1);
3043 
3044         if (isa<UndefValue>(Ptr) ||
3045             (isa<ConstantPointerNull>(Ptr) &&
3046              !NullPointerIsDefined(SI->getFunction(),
3047                                    SI->getPointerAddressSpace()))) {
3048           changeToUnreachable(SI, false, DTU);
3049           Changed = true;
3050           break;
3051         }
3052       }
3053     }
3054 
3055     Instruction *Terminator = BB->getTerminator();
3056     if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
3057       // Turn invokes that call 'nounwind' functions into ordinary calls.
3058       Value *Callee = II->getCalledOperand();
3059       if ((isa<ConstantPointerNull>(Callee) &&
3060            !NullPointerIsDefined(BB->getParent())) ||
3061           isa<UndefValue>(Callee)) {
3062         changeToUnreachable(II, false, DTU);
3063         Changed = true;
3064       } else {
3065         if (II->doesNotReturn() &&
3066             !isa<UnreachableInst>(II->getNormalDest()->front())) {
3067           // If we found an invoke of a no-return function,
3068           // create a new empty basic block with an `unreachable` terminator,
3069           // and set it as the normal destination for the invoke,
3070           // unless that is already the case.
3071           // Note that the original normal destination could have other uses.
3072           BasicBlock *OrigNormalDest = II->getNormalDest();
3073           OrigNormalDest->removePredecessor(II->getParent());
3074           LLVMContext &Ctx = II->getContext();
3075           BasicBlock *UnreachableNormalDest = BasicBlock::Create(
3076               Ctx, OrigNormalDest->getName() + ".unreachable",
3077               II->getFunction(), OrigNormalDest);
3078           new UnreachableInst(Ctx, UnreachableNormalDest);
3079           II->setNormalDest(UnreachableNormalDest);
3080           if (DTU)
3081             DTU->applyUpdates(
3082                 {{DominatorTree::Delete, BB, OrigNormalDest},
3083                  {DominatorTree::Insert, BB, UnreachableNormalDest}});
3084           Changed = true;
3085         }
3086         if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
3087           if (II->use_empty() && !II->mayHaveSideEffects()) {
3088             // jump to the normal destination branch.
3089             BasicBlock *NormalDestBB = II->getNormalDest();
3090             BasicBlock *UnwindDestBB = II->getUnwindDest();
3091             BranchInst::Create(NormalDestBB, II->getIterator());
3092             UnwindDestBB->removePredecessor(II->getParent());
3093             II->eraseFromParent();
3094             if (DTU)
3095               DTU->applyUpdates({{DominatorTree::Delete, BB, UnwindDestBB}});
3096           } else
3097             changeToCall(II, DTU);
3098           Changed = true;
3099         }
3100       }
3101     } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
3102       // Remove catchpads which cannot be reached.
3103       struct CatchPadDenseMapInfo {
3104         static CatchPadInst *getEmptyKey() {
3105           return DenseMapInfo<CatchPadInst *>::getEmptyKey();
3106         }
3107 
3108         static CatchPadInst *getTombstoneKey() {
3109           return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
3110         }
3111 
3112         static unsigned getHashValue(CatchPadInst *CatchPad) {
3113           return static_cast<unsigned>(hash_combine_range(
3114               CatchPad->value_op_begin(), CatchPad->value_op_end()));
3115         }
3116 
3117         static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
3118           if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
3119               RHS == getEmptyKey() || RHS == getTombstoneKey())
3120             return LHS == RHS;
3121           return LHS->isIdenticalTo(RHS);
3122         }
3123       };
3124 
3125       SmallDenseMap<BasicBlock *, int, 8> NumPerSuccessorCases;
3126       // Set of unique CatchPads.
3127       SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
3128                     CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
3129           HandlerSet;
3130       detail::DenseSetEmpty Empty;
3131       for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
3132                                              E = CatchSwitch->handler_end();
3133            I != E; ++I) {
3134         BasicBlock *HandlerBB = *I;
3135         if (DTU)
3136           ++NumPerSuccessorCases[HandlerBB];
3137         auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
3138         if (!HandlerSet.insert({CatchPad, Empty}).second) {
3139           if (DTU)
3140             --NumPerSuccessorCases[HandlerBB];
3141           CatchSwitch->removeHandler(I);
3142           --I;
3143           --E;
3144           Changed = true;
3145         }
3146       }
3147       if (DTU) {
3148         std::vector<DominatorTree::UpdateType> Updates;
3149         for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
3150           if (I.second == 0)
3151             Updates.push_back({DominatorTree::Delete, BB, I.first});
3152         DTU->applyUpdates(Updates);
3153       }
3154     }
3155 
3156     Changed |= ConstantFoldTerminator(BB, true, nullptr, DTU);
3157     for (BasicBlock *Successor : successors(BB))
3158       if (Reachable.insert(Successor).second)
3159         Worklist.push_back(Successor);
3160   } while (!Worklist.empty());
3161   return Changed;
3162 }
3163 
removeUnwindEdge(BasicBlock * BB,DomTreeUpdater * DTU)3164 Instruction *llvm::removeUnwindEdge(BasicBlock *BB, DomTreeUpdater *DTU) {
3165   Instruction *TI = BB->getTerminator();
3166 
3167   if (auto *II = dyn_cast<InvokeInst>(TI))
3168     return changeToCall(II, DTU);
3169 
3170   Instruction *NewTI;
3171   BasicBlock *UnwindDest;
3172 
3173   if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
3174     NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI->getIterator());
3175     UnwindDest = CRI->getUnwindDest();
3176   } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
3177     auto *NewCatchSwitch = CatchSwitchInst::Create(
3178         CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
3179         CatchSwitch->getName(), CatchSwitch->getIterator());
3180     for (BasicBlock *PadBB : CatchSwitch->handlers())
3181       NewCatchSwitch->addHandler(PadBB);
3182 
3183     NewTI = NewCatchSwitch;
3184     UnwindDest = CatchSwitch->getUnwindDest();
3185   } else {
3186     llvm_unreachable("Could not find unwind successor");
3187   }
3188 
3189   NewTI->takeName(TI);
3190   NewTI->setDebugLoc(TI->getDebugLoc());
3191   UnwindDest->removePredecessor(BB);
3192   TI->replaceAllUsesWith(NewTI);
3193   TI->eraseFromParent();
3194   if (DTU)
3195     DTU->applyUpdates({{DominatorTree::Delete, BB, UnwindDest}});
3196   return NewTI;
3197 }
3198 
3199 /// removeUnreachableBlocks - Remove blocks that are not reachable, even
3200 /// if they are in a dead cycle.  Return true if a change was made, false
3201 /// otherwise.
removeUnreachableBlocks(Function & F,DomTreeUpdater * DTU,MemorySSAUpdater * MSSAU)3202 bool llvm::removeUnreachableBlocks(Function &F, DomTreeUpdater *DTU,
3203                                    MemorySSAUpdater *MSSAU) {
3204   SmallPtrSet<BasicBlock *, 16> Reachable;
3205   bool Changed = markAliveBlocks(F, Reachable, DTU);
3206 
3207   // If there are unreachable blocks in the CFG...
3208   if (Reachable.size() == F.size())
3209     return Changed;
3210 
3211   assert(Reachable.size() < F.size());
3212 
3213   // Are there any blocks left to actually delete?
3214   SmallSetVector<BasicBlock *, 8> BlocksToRemove;
3215   for (BasicBlock &BB : F) {
3216     // Skip reachable basic blocks
3217     if (Reachable.count(&BB))
3218       continue;
3219     // Skip already-deleted blocks
3220     if (DTU && DTU->isBBPendingDeletion(&BB))
3221       continue;
3222     BlocksToRemove.insert(&BB);
3223   }
3224 
3225   if (BlocksToRemove.empty())
3226     return Changed;
3227 
3228   Changed = true;
3229   NumRemoved += BlocksToRemove.size();
3230 
3231   if (MSSAU)
3232     MSSAU->removeBlocks(BlocksToRemove);
3233 
3234   DeleteDeadBlocks(BlocksToRemove.takeVector(), DTU);
3235 
3236   return Changed;
3237 }
3238 
combineMetadata(Instruction * K,const Instruction * J,ArrayRef<unsigned> KnownIDs,bool DoesKMove)3239 void llvm::combineMetadata(Instruction *K, const Instruction *J,
3240                            ArrayRef<unsigned> KnownIDs, bool DoesKMove) {
3241   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
3242   K->dropUnknownNonDebugMetadata(KnownIDs);
3243   K->getAllMetadataOtherThanDebugLoc(Metadata);
3244   for (const auto &MD : Metadata) {
3245     unsigned Kind = MD.first;
3246     MDNode *JMD = J->getMetadata(Kind);
3247     MDNode *KMD = MD.second;
3248 
3249     switch (Kind) {
3250       default:
3251         K->setMetadata(Kind, nullptr); // Remove unknown metadata
3252         break;
3253       case LLVMContext::MD_dbg:
3254         llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
3255       case LLVMContext::MD_DIAssignID:
3256         K->mergeDIAssignID(J);
3257         break;
3258       case LLVMContext::MD_tbaa:
3259         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
3260         break;
3261       case LLVMContext::MD_alias_scope:
3262         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
3263         break;
3264       case LLVMContext::MD_noalias:
3265       case LLVMContext::MD_mem_parallel_loop_access:
3266         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
3267         break;
3268       case LLVMContext::MD_access_group:
3269         K->setMetadata(LLVMContext::MD_access_group,
3270                        intersectAccessGroups(K, J));
3271         break;
3272       case LLVMContext::MD_range:
3273         if (DoesKMove || !K->hasMetadata(LLVMContext::MD_noundef))
3274           K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
3275         break;
3276       case LLVMContext::MD_fpmath:
3277         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
3278         break;
3279       case LLVMContext::MD_invariant_load:
3280         // If K moves, only set the !invariant.load if it is present in both
3281         // instructions.
3282         if (DoesKMove)
3283           K->setMetadata(Kind, JMD);
3284         break;
3285       case LLVMContext::MD_nonnull:
3286         if (DoesKMove || !K->hasMetadata(LLVMContext::MD_noundef))
3287           K->setMetadata(Kind, JMD);
3288         break;
3289       case LLVMContext::MD_invariant_group:
3290         // Preserve !invariant.group in K.
3291         break;
3292       case LLVMContext::MD_mmra:
3293         // Combine MMRAs
3294         break;
3295       case LLVMContext::MD_align:
3296         if (DoesKMove || !K->hasMetadata(LLVMContext::MD_noundef))
3297           K->setMetadata(
3298               Kind, MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
3299         break;
3300       case LLVMContext::MD_dereferenceable:
3301       case LLVMContext::MD_dereferenceable_or_null:
3302         if (DoesKMove)
3303           K->setMetadata(Kind,
3304             MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
3305         break;
3306       case LLVMContext::MD_preserve_access_index:
3307         // Preserve !preserve.access.index in K.
3308         break;
3309       case LLVMContext::MD_noundef:
3310         // If K does move, keep noundef if it is present in both instructions.
3311         if (DoesKMove)
3312           K->setMetadata(Kind, JMD);
3313         break;
3314       case LLVMContext::MD_nontemporal:
3315         // Preserve !nontemporal if it is present on both instructions.
3316         K->setMetadata(Kind, JMD);
3317         break;
3318       case LLVMContext::MD_prof:
3319         if (DoesKMove)
3320           K->setMetadata(Kind, MDNode::getMergedProfMetadata(KMD, JMD, K, J));
3321         break;
3322     }
3323   }
3324   // Set !invariant.group from J if J has it. If both instructions have it
3325   // then we will just pick it from J - even when they are different.
3326   // Also make sure that K is load or store - f.e. combining bitcast with load
3327   // could produce bitcast with invariant.group metadata, which is invalid.
3328   // FIXME: we should try to preserve both invariant.group md if they are
3329   // different, but right now instruction can only have one invariant.group.
3330   if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
3331     if (isa<LoadInst>(K) || isa<StoreInst>(K))
3332       K->setMetadata(LLVMContext::MD_invariant_group, JMD);
3333 
3334   // Merge MMRAs.
3335   // This is handled separately because we also want to handle cases where K
3336   // doesn't have tags but J does.
3337   auto JMMRA = J->getMetadata(LLVMContext::MD_mmra);
3338   auto KMMRA = K->getMetadata(LLVMContext::MD_mmra);
3339   if (JMMRA || KMMRA) {
3340     K->setMetadata(LLVMContext::MD_mmra,
3341                    MMRAMetadata::combine(K->getContext(), JMMRA, KMMRA));
3342   }
3343 }
3344 
combineMetadataForCSE(Instruction * K,const Instruction * J,bool KDominatesJ)3345 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J,
3346                                  bool KDominatesJ) {
3347   unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
3348                          LLVMContext::MD_alias_scope,
3349                          LLVMContext::MD_noalias,
3350                          LLVMContext::MD_range,
3351                          LLVMContext::MD_fpmath,
3352                          LLVMContext::MD_invariant_load,
3353                          LLVMContext::MD_nonnull,
3354                          LLVMContext::MD_invariant_group,
3355                          LLVMContext::MD_align,
3356                          LLVMContext::MD_dereferenceable,
3357                          LLVMContext::MD_dereferenceable_or_null,
3358                          LLVMContext::MD_access_group,
3359                          LLVMContext::MD_preserve_access_index,
3360                          LLVMContext::MD_prof,
3361                          LLVMContext::MD_nontemporal,
3362                          LLVMContext::MD_noundef,
3363                          LLVMContext::MD_mmra};
3364   combineMetadata(K, J, KnownIDs, KDominatesJ);
3365 }
3366 
copyMetadataForLoad(LoadInst & Dest,const LoadInst & Source)3367 void llvm::copyMetadataForLoad(LoadInst &Dest, const LoadInst &Source) {
3368   SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
3369   Source.getAllMetadata(MD);
3370   MDBuilder MDB(Dest.getContext());
3371   Type *NewType = Dest.getType();
3372   const DataLayout &DL = Source.getDataLayout();
3373   for (const auto &MDPair : MD) {
3374     unsigned ID = MDPair.first;
3375     MDNode *N = MDPair.second;
3376     // Note, essentially every kind of metadata should be preserved here! This
3377     // routine is supposed to clone a load instruction changing *only its type*.
3378     // The only metadata it makes sense to drop is metadata which is invalidated
3379     // when the pointer type changes. This should essentially never be the case
3380     // in LLVM, but we explicitly switch over only known metadata to be
3381     // conservatively correct. If you are adding metadata to LLVM which pertains
3382     // to loads, you almost certainly want to add it here.
3383     switch (ID) {
3384     case LLVMContext::MD_dbg:
3385     case LLVMContext::MD_tbaa:
3386     case LLVMContext::MD_prof:
3387     case LLVMContext::MD_fpmath:
3388     case LLVMContext::MD_tbaa_struct:
3389     case LLVMContext::MD_invariant_load:
3390     case LLVMContext::MD_alias_scope:
3391     case LLVMContext::MD_noalias:
3392     case LLVMContext::MD_nontemporal:
3393     case LLVMContext::MD_mem_parallel_loop_access:
3394     case LLVMContext::MD_access_group:
3395     case LLVMContext::MD_noundef:
3396       // All of these directly apply.
3397       Dest.setMetadata(ID, N);
3398       break;
3399 
3400     case LLVMContext::MD_nonnull:
3401       copyNonnullMetadata(Source, N, Dest);
3402       break;
3403 
3404     case LLVMContext::MD_align:
3405     case LLVMContext::MD_dereferenceable:
3406     case LLVMContext::MD_dereferenceable_or_null:
3407       // These only directly apply if the new type is also a pointer.
3408       if (NewType->isPointerTy())
3409         Dest.setMetadata(ID, N);
3410       break;
3411 
3412     case LLVMContext::MD_range:
3413       copyRangeMetadata(DL, Source, N, Dest);
3414       break;
3415     }
3416   }
3417 }
3418 
patchReplacementInstruction(Instruction * I,Value * Repl)3419 void llvm::patchReplacementInstruction(Instruction *I, Value *Repl) {
3420   auto *ReplInst = dyn_cast<Instruction>(Repl);
3421   if (!ReplInst)
3422     return;
3423 
3424   // Patch the replacement so that it is not more restrictive than the value
3425   // being replaced.
3426   WithOverflowInst *UnusedWO;
3427   // When replacing the result of a llvm.*.with.overflow intrinsic with a
3428   // overflowing binary operator, nuw/nsw flags may no longer hold.
3429   if (isa<OverflowingBinaryOperator>(ReplInst) &&
3430       match(I, m_ExtractValue<0>(m_WithOverflowInst(UnusedWO))))
3431     ReplInst->dropPoisonGeneratingFlags();
3432   // Note that if 'I' is a load being replaced by some operation,
3433   // for example, by an arithmetic operation, then andIRFlags()
3434   // would just erase all math flags from the original arithmetic
3435   // operation, which is clearly not wanted and not needed.
3436   else if (!isa<LoadInst>(I))
3437     ReplInst->andIRFlags(I);
3438 
3439   // FIXME: If both the original and replacement value are part of the
3440   // same control-flow region (meaning that the execution of one
3441   // guarantees the execution of the other), then we can combine the
3442   // noalias scopes here and do better than the general conservative
3443   // answer used in combineMetadata().
3444 
3445   // In general, GVN unifies expressions over different control-flow
3446   // regions, and so we need a conservative combination of the noalias
3447   // scopes.
3448   combineMetadataForCSE(ReplInst, I, false);
3449 }
3450 
3451 template <typename RootType, typename ShouldReplaceFn>
replaceDominatedUsesWith(Value * From,Value * To,const RootType & Root,const ShouldReplaceFn & ShouldReplace)3452 static unsigned replaceDominatedUsesWith(Value *From, Value *To,
3453                                          const RootType &Root,
3454                                          const ShouldReplaceFn &ShouldReplace) {
3455   assert(From->getType() == To->getType());
3456 
3457   unsigned Count = 0;
3458   for (Use &U : llvm::make_early_inc_range(From->uses())) {
3459     if (!ShouldReplace(Root, U))
3460       continue;
3461     LLVM_DEBUG(dbgs() << "Replace dominated use of '";
3462                From->printAsOperand(dbgs());
3463                dbgs() << "' with " << *To << " in " << *U.getUser() << "\n");
3464     U.set(To);
3465     ++Count;
3466   }
3467   return Count;
3468 }
3469 
replaceNonLocalUsesWith(Instruction * From,Value * To)3470 unsigned llvm::replaceNonLocalUsesWith(Instruction *From, Value *To) {
3471    assert(From->getType() == To->getType());
3472    auto *BB = From->getParent();
3473    unsigned Count = 0;
3474 
3475    for (Use &U : llvm::make_early_inc_range(From->uses())) {
3476     auto *I = cast<Instruction>(U.getUser());
3477     if (I->getParent() == BB)
3478       continue;
3479     U.set(To);
3480     ++Count;
3481   }
3482   return Count;
3483 }
3484 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlockEdge & Root)3485 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
3486                                         DominatorTree &DT,
3487                                         const BasicBlockEdge &Root) {
3488   auto Dominates = [&DT](const BasicBlockEdge &Root, const Use &U) {
3489     return DT.dominates(Root, U);
3490   };
3491   return ::replaceDominatedUsesWith(From, To, Root, Dominates);
3492 }
3493 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlock * BB)3494 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
3495                                         DominatorTree &DT,
3496                                         const BasicBlock *BB) {
3497   auto Dominates = [&DT](const BasicBlock *BB, const Use &U) {
3498     return DT.dominates(BB, U);
3499   };
3500   return ::replaceDominatedUsesWith(From, To, BB, Dominates);
3501 }
3502 
replaceDominatedUsesWithIf(Value * From,Value * To,DominatorTree & DT,const BasicBlockEdge & Root,function_ref<bool (const Use & U,const Value * To)> ShouldReplace)3503 unsigned llvm::replaceDominatedUsesWithIf(
3504     Value *From, Value *To, DominatorTree &DT, const BasicBlockEdge &Root,
3505     function_ref<bool(const Use &U, const Value *To)> ShouldReplace) {
3506   auto DominatesAndShouldReplace =
3507       [&DT, &ShouldReplace, To](const BasicBlockEdge &Root, const Use &U) {
3508         return DT.dominates(Root, U) && ShouldReplace(U, To);
3509       };
3510   return ::replaceDominatedUsesWith(From, To, Root, DominatesAndShouldReplace);
3511 }
3512 
replaceDominatedUsesWithIf(Value * From,Value * To,DominatorTree & DT,const BasicBlock * BB,function_ref<bool (const Use & U,const Value * To)> ShouldReplace)3513 unsigned llvm::replaceDominatedUsesWithIf(
3514     Value *From, Value *To, DominatorTree &DT, const BasicBlock *BB,
3515     function_ref<bool(const Use &U, const Value *To)> ShouldReplace) {
3516   auto DominatesAndShouldReplace = [&DT, &ShouldReplace,
3517                                     To](const BasicBlock *BB, const Use &U) {
3518     return DT.dominates(BB, U) && ShouldReplace(U, To);
3519   };
3520   return ::replaceDominatedUsesWith(From, To, BB, DominatesAndShouldReplace);
3521 }
3522 
callsGCLeafFunction(const CallBase * Call,const TargetLibraryInfo & TLI)3523 bool llvm::callsGCLeafFunction(const CallBase *Call,
3524                                const TargetLibraryInfo &TLI) {
3525   // Check if the function is specifically marked as a gc leaf function.
3526   if (Call->hasFnAttr("gc-leaf-function"))
3527     return true;
3528   if (const Function *F = Call->getCalledFunction()) {
3529     if (F->hasFnAttribute("gc-leaf-function"))
3530       return true;
3531 
3532     if (auto IID = F->getIntrinsicID()) {
3533       // Most LLVM intrinsics do not take safepoints.
3534       return IID != Intrinsic::experimental_gc_statepoint &&
3535              IID != Intrinsic::experimental_deoptimize &&
3536              IID != Intrinsic::memcpy_element_unordered_atomic &&
3537              IID != Intrinsic::memmove_element_unordered_atomic;
3538     }
3539   }
3540 
3541   // Lib calls can be materialized by some passes, and won't be
3542   // marked as 'gc-leaf-function.' All available Libcalls are
3543   // GC-leaf.
3544   LibFunc LF;
3545   if (TLI.getLibFunc(*Call, LF)) {
3546     return TLI.has(LF);
3547   }
3548 
3549   return false;
3550 }
3551 
copyNonnullMetadata(const LoadInst & OldLI,MDNode * N,LoadInst & NewLI)3552 void llvm::copyNonnullMetadata(const LoadInst &OldLI, MDNode *N,
3553                                LoadInst &NewLI) {
3554   auto *NewTy = NewLI.getType();
3555 
3556   // This only directly applies if the new type is also a pointer.
3557   if (NewTy->isPointerTy()) {
3558     NewLI.setMetadata(LLVMContext::MD_nonnull, N);
3559     return;
3560   }
3561 
3562   // The only other translation we can do is to integral loads with !range
3563   // metadata.
3564   if (!NewTy->isIntegerTy())
3565     return;
3566 
3567   MDBuilder MDB(NewLI.getContext());
3568   const Value *Ptr = OldLI.getPointerOperand();
3569   auto *ITy = cast<IntegerType>(NewTy);
3570   auto *NullInt = ConstantExpr::getPtrToInt(
3571       ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
3572   auto *NonNullInt = ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
3573   NewLI.setMetadata(LLVMContext::MD_range,
3574                     MDB.createRange(NonNullInt, NullInt));
3575 }
3576 
copyRangeMetadata(const DataLayout & DL,const LoadInst & OldLI,MDNode * N,LoadInst & NewLI)3577 void llvm::copyRangeMetadata(const DataLayout &DL, const LoadInst &OldLI,
3578                              MDNode *N, LoadInst &NewLI) {
3579   auto *NewTy = NewLI.getType();
3580   // Simply copy the metadata if the type did not change.
3581   if (NewTy == OldLI.getType()) {
3582     NewLI.setMetadata(LLVMContext::MD_range, N);
3583     return;
3584   }
3585 
3586   // Give up unless it is converted to a pointer where there is a single very
3587   // valuable mapping we can do reliably.
3588   // FIXME: It would be nice to propagate this in more ways, but the type
3589   // conversions make it hard.
3590   if (!NewTy->isPointerTy())
3591     return;
3592 
3593   unsigned BitWidth = DL.getPointerTypeSizeInBits(NewTy);
3594   if (BitWidth == OldLI.getType()->getScalarSizeInBits() &&
3595       !getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) {
3596     MDNode *NN = MDNode::get(OldLI.getContext(), std::nullopt);
3597     NewLI.setMetadata(LLVMContext::MD_nonnull, NN);
3598   }
3599 }
3600 
dropDebugUsers(Instruction & I)3601 void llvm::dropDebugUsers(Instruction &I) {
3602   SmallVector<DbgVariableIntrinsic *, 1> DbgUsers;
3603   SmallVector<DbgVariableRecord *, 1> DPUsers;
3604   findDbgUsers(DbgUsers, &I, &DPUsers);
3605   for (auto *DII : DbgUsers)
3606     DII->eraseFromParent();
3607   for (auto *DVR : DPUsers)
3608     DVR->eraseFromParent();
3609 }
3610 
hoistAllInstructionsInto(BasicBlock * DomBlock,Instruction * InsertPt,BasicBlock * BB)3611 void llvm::hoistAllInstructionsInto(BasicBlock *DomBlock, Instruction *InsertPt,
3612                                     BasicBlock *BB) {
3613   // Since we are moving the instructions out of its basic block, we do not
3614   // retain their original debug locations (DILocations) and debug intrinsic
3615   // instructions.
3616   //
3617   // Doing so would degrade the debugging experience and adversely affect the
3618   // accuracy of profiling information.
3619   //
3620   // Currently, when hoisting the instructions, we take the following actions:
3621   // - Remove their debug intrinsic instructions.
3622   // - Set their debug locations to the values from the insertion point.
3623   //
3624   // As per PR39141 (comment #8), the more fundamental reason why the dbg.values
3625   // need to be deleted, is because there will not be any instructions with a
3626   // DILocation in either branch left after performing the transformation. We
3627   // can only insert a dbg.value after the two branches are joined again.
3628   //
3629   // See PR38762, PR39243 for more details.
3630   //
3631   // TODO: Extend llvm.dbg.value to take more than one SSA Value (PR39141) to
3632   // encode predicated DIExpressions that yield different results on different
3633   // code paths.
3634 
3635   for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE;) {
3636     Instruction *I = &*II;
3637     I->dropUBImplyingAttrsAndMetadata();
3638     if (I->isUsedByMetadata())
3639       dropDebugUsers(*I);
3640     // RemoveDIs: drop debug-info too as the following code does.
3641     I->dropDbgRecords();
3642     if (I->isDebugOrPseudoInst()) {
3643       // Remove DbgInfo and pseudo probe Intrinsics.
3644       II = I->eraseFromParent();
3645       continue;
3646     }
3647     I->setDebugLoc(InsertPt->getDebugLoc());
3648     ++II;
3649   }
3650   DomBlock->splice(InsertPt->getIterator(), BB, BB->begin(),
3651                    BB->getTerminator()->getIterator());
3652 }
3653 
getExpressionForConstant(DIBuilder & DIB,const Constant & C,Type & Ty)3654 DIExpression *llvm::getExpressionForConstant(DIBuilder &DIB, const Constant &C,
3655                                              Type &Ty) {
3656   // Create integer constant expression.
3657   auto createIntegerExpression = [&DIB](const Constant &CV) -> DIExpression * {
3658     const APInt &API = cast<ConstantInt>(&CV)->getValue();
3659     std::optional<int64_t> InitIntOpt = API.trySExtValue();
3660     return InitIntOpt ? DIB.createConstantValueExpression(
3661                             static_cast<uint64_t>(*InitIntOpt))
3662                       : nullptr;
3663   };
3664 
3665   if (isa<ConstantInt>(C))
3666     return createIntegerExpression(C);
3667 
3668   auto *FP = dyn_cast<ConstantFP>(&C);
3669   if (FP && Ty.isFloatingPointTy() && Ty.getScalarSizeInBits() <= 64) {
3670     const APFloat &APF = FP->getValueAPF();
3671     APInt const &API = APF.bitcastToAPInt();
3672     if (auto Temp = API.getZExtValue())
3673       return DIB.createConstantValueExpression(static_cast<uint64_t>(Temp));
3674     return DIB.createConstantValueExpression(*API.getRawData());
3675   }
3676 
3677   if (!Ty.isPointerTy())
3678     return nullptr;
3679 
3680   if (isa<ConstantPointerNull>(C))
3681     return DIB.createConstantValueExpression(0);
3682 
3683   if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(&C))
3684     if (CE->getOpcode() == Instruction::IntToPtr) {
3685       const Value *V = CE->getOperand(0);
3686       if (auto CI = dyn_cast_or_null<ConstantInt>(V))
3687         return createIntegerExpression(*CI);
3688     }
3689   return nullptr;
3690 }
3691 
remapDebugVariable(ValueToValueMapTy & Mapping,Instruction * Inst)3692 void llvm::remapDebugVariable(ValueToValueMapTy &Mapping, Instruction *Inst) {
3693   auto RemapDebugOperands = [&Mapping](auto *DV, auto Set) {
3694     for (auto *Op : Set) {
3695       auto I = Mapping.find(Op);
3696       if (I != Mapping.end())
3697         DV->replaceVariableLocationOp(Op, I->second, /*AllowEmpty=*/true);
3698     }
3699   };
3700   auto RemapAssignAddress = [&Mapping](auto *DA) {
3701     auto I = Mapping.find(DA->getAddress());
3702     if (I != Mapping.end())
3703       DA->setAddress(I->second);
3704   };
3705   if (auto DVI = dyn_cast<DbgVariableIntrinsic>(Inst))
3706     RemapDebugOperands(DVI, DVI->location_ops());
3707   if (auto DAI = dyn_cast<DbgAssignIntrinsic>(Inst))
3708     RemapAssignAddress(DAI);
3709   for (DbgVariableRecord &DVR : filterDbgVars(Inst->getDbgRecordRange())) {
3710     RemapDebugOperands(&DVR, DVR.location_ops());
3711     if (DVR.isDbgAssign())
3712       RemapAssignAddress(&DVR);
3713   }
3714 }
3715 
3716 namespace {
3717 
3718 /// A potential constituent of a bitreverse or bswap expression. See
3719 /// collectBitParts for a fuller explanation.
3720 struct BitPart {
BitPart__anon34d338391211::BitPart3721   BitPart(Value *P, unsigned BW) : Provider(P) {
3722     Provenance.resize(BW);
3723   }
3724 
3725   /// The Value that this is a bitreverse/bswap of.
3726   Value *Provider;
3727 
3728   /// The "provenance" of each bit. Provenance[A] = B means that bit A
3729   /// in Provider becomes bit B in the result of this expression.
3730   SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128.
3731 
3732   enum { Unset = -1 };
3733 };
3734 
3735 } // end anonymous namespace
3736 
3737 /// Analyze the specified subexpression and see if it is capable of providing
3738 /// pieces of a bswap or bitreverse. The subexpression provides a potential
3739 /// piece of a bswap or bitreverse if it can be proved that each non-zero bit in
3740 /// the output of the expression came from a corresponding bit in some other
3741 /// value. This function is recursive, and the end result is a mapping of
3742 /// bitnumber to bitnumber. It is the caller's responsibility to validate that
3743 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse.
3744 ///
3745 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know
3746 /// that the expression deposits the low byte of %X into the high byte of the
3747 /// result and that all other bits are zero. This expression is accepted and a
3748 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to
3749 /// [0-7].
3750 ///
3751 /// For vector types, all analysis is performed at the per-element level. No
3752 /// cross-element analysis is supported (shuffle/insertion/reduction), and all
3753 /// constant masks must be splatted across all elements.
3754 ///
3755 /// To avoid revisiting values, the BitPart results are memoized into the
3756 /// provided map. To avoid unnecessary copying of BitParts, BitParts are
3757 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to
3758 /// store BitParts objects, not pointers. As we need the concept of a nullptr
3759 /// BitParts (Value has been analyzed and the analysis failed), we an Optional
3760 /// type instead to provide the same functionality.
3761 ///
3762 /// Because we pass around references into \c BPS, we must use a container that
3763 /// does not invalidate internal references (std::map instead of DenseMap).
3764 static const std::optional<BitPart> &
collectBitParts(Value * V,bool MatchBSwaps,bool MatchBitReversals,std::map<Value *,std::optional<BitPart>> & BPS,int Depth,bool & FoundRoot)3765 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
3766                 std::map<Value *, std::optional<BitPart>> &BPS, int Depth,
3767                 bool &FoundRoot) {
3768   auto I = BPS.find(V);
3769   if (I != BPS.end())
3770     return I->second;
3771 
3772   auto &Result = BPS[V] = std::nullopt;
3773   auto BitWidth = V->getType()->getScalarSizeInBits();
3774 
3775   // Can't do integer/elements > 128 bits.
3776   if (BitWidth > 128)
3777     return Result;
3778 
3779   // Prevent stack overflow by limiting the recursion depth
3780   if (Depth == BitPartRecursionMaxDepth) {
3781     LLVM_DEBUG(dbgs() << "collectBitParts max recursion depth reached.\n");
3782     return Result;
3783   }
3784 
3785   if (auto *I = dyn_cast<Instruction>(V)) {
3786     Value *X, *Y;
3787     const APInt *C;
3788 
3789     // If this is an or instruction, it may be an inner node of the bswap.
3790     if (match(V, m_Or(m_Value(X), m_Value(Y)))) {
3791       // Check we have both sources and they are from the same provider.
3792       const auto &A = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3793                                       Depth + 1, FoundRoot);
3794       if (!A || !A->Provider)
3795         return Result;
3796 
3797       const auto &B = collectBitParts(Y, MatchBSwaps, MatchBitReversals, BPS,
3798                                       Depth + 1, FoundRoot);
3799       if (!B || A->Provider != B->Provider)
3800         return Result;
3801 
3802       // Try and merge the two together.
3803       Result = BitPart(A->Provider, BitWidth);
3804       for (unsigned BitIdx = 0; BitIdx < BitWidth; ++BitIdx) {
3805         if (A->Provenance[BitIdx] != BitPart::Unset &&
3806             B->Provenance[BitIdx] != BitPart::Unset &&
3807             A->Provenance[BitIdx] != B->Provenance[BitIdx])
3808           return Result = std::nullopt;
3809 
3810         if (A->Provenance[BitIdx] == BitPart::Unset)
3811           Result->Provenance[BitIdx] = B->Provenance[BitIdx];
3812         else
3813           Result->Provenance[BitIdx] = A->Provenance[BitIdx];
3814       }
3815 
3816       return Result;
3817     }
3818 
3819     // If this is a logical shift by a constant, recurse then shift the result.
3820     if (match(V, m_LogicalShift(m_Value(X), m_APInt(C)))) {
3821       const APInt &BitShift = *C;
3822 
3823       // Ensure the shift amount is defined.
3824       if (BitShift.uge(BitWidth))
3825         return Result;
3826 
3827       // For bswap-only, limit shift amounts to whole bytes, for an early exit.
3828       if (!MatchBitReversals && (BitShift.getZExtValue() % 8) != 0)
3829         return Result;
3830 
3831       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3832                                         Depth + 1, FoundRoot);
3833       if (!Res)
3834         return Result;
3835       Result = Res;
3836 
3837       // Perform the "shift" on BitProvenance.
3838       auto &P = Result->Provenance;
3839       if (I->getOpcode() == Instruction::Shl) {
3840         P.erase(std::prev(P.end(), BitShift.getZExtValue()), P.end());
3841         P.insert(P.begin(), BitShift.getZExtValue(), BitPart::Unset);
3842       } else {
3843         P.erase(P.begin(), std::next(P.begin(), BitShift.getZExtValue()));
3844         P.insert(P.end(), BitShift.getZExtValue(), BitPart::Unset);
3845       }
3846 
3847       return Result;
3848     }
3849 
3850     // If this is a logical 'and' with a mask that clears bits, recurse then
3851     // unset the appropriate bits.
3852     if (match(V, m_And(m_Value(X), m_APInt(C)))) {
3853       const APInt &AndMask = *C;
3854 
3855       // Check that the mask allows a multiple of 8 bits for a bswap, for an
3856       // early exit.
3857       unsigned NumMaskedBits = AndMask.popcount();
3858       if (!MatchBitReversals && (NumMaskedBits % 8) != 0)
3859         return Result;
3860 
3861       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3862                                         Depth + 1, FoundRoot);
3863       if (!Res)
3864         return Result;
3865       Result = Res;
3866 
3867       for (unsigned BitIdx = 0; BitIdx < BitWidth; ++BitIdx)
3868         // If the AndMask is zero for this bit, clear the bit.
3869         if (AndMask[BitIdx] == 0)
3870           Result->Provenance[BitIdx] = BitPart::Unset;
3871       return Result;
3872     }
3873 
3874     // If this is a zext instruction zero extend the result.
3875     if (match(V, m_ZExt(m_Value(X)))) {
3876       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3877                                         Depth + 1, FoundRoot);
3878       if (!Res)
3879         return Result;
3880 
3881       Result = BitPart(Res->Provider, BitWidth);
3882       auto NarrowBitWidth = X->getType()->getScalarSizeInBits();
3883       for (unsigned BitIdx = 0; BitIdx < NarrowBitWidth; ++BitIdx)
3884         Result->Provenance[BitIdx] = Res->Provenance[BitIdx];
3885       for (unsigned BitIdx = NarrowBitWidth; BitIdx < BitWidth; ++BitIdx)
3886         Result->Provenance[BitIdx] = BitPart::Unset;
3887       return Result;
3888     }
3889 
3890     // If this is a truncate instruction, extract the lower bits.
3891     if (match(V, m_Trunc(m_Value(X)))) {
3892       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3893                                         Depth + 1, FoundRoot);
3894       if (!Res)
3895         return Result;
3896 
3897       Result = BitPart(Res->Provider, BitWidth);
3898       for (unsigned BitIdx = 0; BitIdx < BitWidth; ++BitIdx)
3899         Result->Provenance[BitIdx] = Res->Provenance[BitIdx];
3900       return Result;
3901     }
3902 
3903     // BITREVERSE - most likely due to us previous matching a partial
3904     // bitreverse.
3905     if (match(V, m_BitReverse(m_Value(X)))) {
3906       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3907                                         Depth + 1, FoundRoot);
3908       if (!Res)
3909         return Result;
3910 
3911       Result = BitPart(Res->Provider, BitWidth);
3912       for (unsigned BitIdx = 0; BitIdx < BitWidth; ++BitIdx)
3913         Result->Provenance[(BitWidth - 1) - BitIdx] = Res->Provenance[BitIdx];
3914       return Result;
3915     }
3916 
3917     // BSWAP - most likely due to us previous matching a partial bswap.
3918     if (match(V, m_BSwap(m_Value(X)))) {
3919       const auto &Res = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3920                                         Depth + 1, FoundRoot);
3921       if (!Res)
3922         return Result;
3923 
3924       unsigned ByteWidth = BitWidth / 8;
3925       Result = BitPart(Res->Provider, BitWidth);
3926       for (unsigned ByteIdx = 0; ByteIdx < ByteWidth; ++ByteIdx) {
3927         unsigned ByteBitOfs = ByteIdx * 8;
3928         for (unsigned BitIdx = 0; BitIdx < 8; ++BitIdx)
3929           Result->Provenance[(BitWidth - 8 - ByteBitOfs) + BitIdx] =
3930               Res->Provenance[ByteBitOfs + BitIdx];
3931       }
3932       return Result;
3933     }
3934 
3935     // Funnel 'double' shifts take 3 operands, 2 inputs and the shift
3936     // amount (modulo).
3937     // fshl(X,Y,Z): (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
3938     // fshr(X,Y,Z): (X << (BW - (Z % BW))) | (Y >> (Z % BW))
3939     if (match(V, m_FShl(m_Value(X), m_Value(Y), m_APInt(C))) ||
3940         match(V, m_FShr(m_Value(X), m_Value(Y), m_APInt(C)))) {
3941       // We can treat fshr as a fshl by flipping the modulo amount.
3942       unsigned ModAmt = C->urem(BitWidth);
3943       if (cast<IntrinsicInst>(I)->getIntrinsicID() == Intrinsic::fshr)
3944         ModAmt = BitWidth - ModAmt;
3945 
3946       // For bswap-only, limit shift amounts to whole bytes, for an early exit.
3947       if (!MatchBitReversals && (ModAmt % 8) != 0)
3948         return Result;
3949 
3950       // Check we have both sources and they are from the same provider.
3951       const auto &LHS = collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS,
3952                                         Depth + 1, FoundRoot);
3953       if (!LHS || !LHS->Provider)
3954         return Result;
3955 
3956       const auto &RHS = collectBitParts(Y, MatchBSwaps, MatchBitReversals, BPS,
3957                                         Depth + 1, FoundRoot);
3958       if (!RHS || LHS->Provider != RHS->Provider)
3959         return Result;
3960 
3961       unsigned StartBitRHS = BitWidth - ModAmt;
3962       Result = BitPart(LHS->Provider, BitWidth);
3963       for (unsigned BitIdx = 0; BitIdx < StartBitRHS; ++BitIdx)
3964         Result->Provenance[BitIdx + ModAmt] = LHS->Provenance[BitIdx];
3965       for (unsigned BitIdx = 0; BitIdx < ModAmt; ++BitIdx)
3966         Result->Provenance[BitIdx] = RHS->Provenance[BitIdx + StartBitRHS];
3967       return Result;
3968     }
3969   }
3970 
3971   // If we've already found a root input value then we're never going to merge
3972   // these back together.
3973   if (FoundRoot)
3974     return Result;
3975 
3976   // Okay, we got to something that isn't a shift, 'or', 'and', etc. This must
3977   // be the root input value to the bswap/bitreverse.
3978   FoundRoot = true;
3979   Result = BitPart(V, BitWidth);
3980   for (unsigned BitIdx = 0; BitIdx < BitWidth; ++BitIdx)
3981     Result->Provenance[BitIdx] = BitIdx;
3982   return Result;
3983 }
3984 
bitTransformIsCorrectForBSwap(unsigned From,unsigned To,unsigned BitWidth)3985 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To,
3986                                           unsigned BitWidth) {
3987   if (From % 8 != To % 8)
3988     return false;
3989   // Convert from bit indices to byte indices and check for a byte reversal.
3990   From >>= 3;
3991   To >>= 3;
3992   BitWidth >>= 3;
3993   return From == BitWidth - To - 1;
3994 }
3995 
bitTransformIsCorrectForBitReverse(unsigned From,unsigned To,unsigned BitWidth)3996 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To,
3997                                                unsigned BitWidth) {
3998   return From == BitWidth - To - 1;
3999 }
4000 
recognizeBSwapOrBitReverseIdiom(Instruction * I,bool MatchBSwaps,bool MatchBitReversals,SmallVectorImpl<Instruction * > & InsertedInsts)4001 bool llvm::recognizeBSwapOrBitReverseIdiom(
4002     Instruction *I, bool MatchBSwaps, bool MatchBitReversals,
4003     SmallVectorImpl<Instruction *> &InsertedInsts) {
4004   if (!match(I, m_Or(m_Value(), m_Value())) &&
4005       !match(I, m_FShl(m_Value(), m_Value(), m_Value())) &&
4006       !match(I, m_FShr(m_Value(), m_Value(), m_Value())) &&
4007       !match(I, m_BSwap(m_Value())))
4008     return false;
4009   if (!MatchBSwaps && !MatchBitReversals)
4010     return false;
4011   Type *ITy = I->getType();
4012   if (!ITy->isIntOrIntVectorTy() || ITy->getScalarSizeInBits() > 128)
4013     return false;  // Can't do integer/elements > 128 bits.
4014 
4015   // Try to find all the pieces corresponding to the bswap.
4016   bool FoundRoot = false;
4017   std::map<Value *, std::optional<BitPart>> BPS;
4018   const auto &Res =
4019       collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS, 0, FoundRoot);
4020   if (!Res)
4021     return false;
4022   ArrayRef<int8_t> BitProvenance = Res->Provenance;
4023   assert(all_of(BitProvenance,
4024                 [](int8_t I) { return I == BitPart::Unset || 0 <= I; }) &&
4025          "Illegal bit provenance index");
4026 
4027   // If the upper bits are zero, then attempt to perform as a truncated op.
4028   Type *DemandedTy = ITy;
4029   if (BitProvenance.back() == BitPart::Unset) {
4030     while (!BitProvenance.empty() && BitProvenance.back() == BitPart::Unset)
4031       BitProvenance = BitProvenance.drop_back();
4032     if (BitProvenance.empty())
4033       return false; // TODO - handle null value?
4034     DemandedTy = Type::getIntNTy(I->getContext(), BitProvenance.size());
4035     if (auto *IVecTy = dyn_cast<VectorType>(ITy))
4036       DemandedTy = VectorType::get(DemandedTy, IVecTy);
4037   }
4038 
4039   // Check BitProvenance hasn't found a source larger than the result type.
4040   unsigned DemandedBW = DemandedTy->getScalarSizeInBits();
4041   if (DemandedBW > ITy->getScalarSizeInBits())
4042     return false;
4043 
4044   // Now, is the bit permutation correct for a bswap or a bitreverse? We can
4045   // only byteswap values with an even number of bytes.
4046   APInt DemandedMask = APInt::getAllOnes(DemandedBW);
4047   bool OKForBSwap = MatchBSwaps && (DemandedBW % 16) == 0;
4048   bool OKForBitReverse = MatchBitReversals;
4049   for (unsigned BitIdx = 0;
4050        (BitIdx < DemandedBW) && (OKForBSwap || OKForBitReverse); ++BitIdx) {
4051     if (BitProvenance[BitIdx] == BitPart::Unset) {
4052       DemandedMask.clearBit(BitIdx);
4053       continue;
4054     }
4055     OKForBSwap &= bitTransformIsCorrectForBSwap(BitProvenance[BitIdx], BitIdx,
4056                                                 DemandedBW);
4057     OKForBitReverse &= bitTransformIsCorrectForBitReverse(BitProvenance[BitIdx],
4058                                                           BitIdx, DemandedBW);
4059   }
4060 
4061   Intrinsic::ID Intrin;
4062   if (OKForBSwap)
4063     Intrin = Intrinsic::bswap;
4064   else if (OKForBitReverse)
4065     Intrin = Intrinsic::bitreverse;
4066   else
4067     return false;
4068 
4069   Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy);
4070   Value *Provider = Res->Provider;
4071 
4072   // We may need to truncate the provider.
4073   if (DemandedTy != Provider->getType()) {
4074     auto *Trunc =
4075         CastInst::CreateIntegerCast(Provider, DemandedTy, false, "trunc", I->getIterator());
4076     InsertedInsts.push_back(Trunc);
4077     Provider = Trunc;
4078   }
4079 
4080   Instruction *Result = CallInst::Create(F, Provider, "rev", I->getIterator());
4081   InsertedInsts.push_back(Result);
4082 
4083   if (!DemandedMask.isAllOnes()) {
4084     auto *Mask = ConstantInt::get(DemandedTy, DemandedMask);
4085     Result = BinaryOperator::Create(Instruction::And, Result, Mask, "mask", I->getIterator());
4086     InsertedInsts.push_back(Result);
4087   }
4088 
4089   // We may need to zeroextend back to the result type.
4090   if (ITy != Result->getType()) {
4091     auto *ExtInst = CastInst::CreateIntegerCast(Result, ITy, false, "zext", I->getIterator());
4092     InsertedInsts.push_back(ExtInst);
4093   }
4094 
4095   return true;
4096 }
4097 
4098 // CodeGen has special handling for some string functions that may replace
4099 // them with target-specific intrinsics.  Since that'd skip our interceptors
4100 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses,
4101 // we mark affected calls as NoBuiltin, which will disable optimization
4102 // in CodeGen.
maybeMarkSanitizerLibraryCallNoBuiltin(CallInst * CI,const TargetLibraryInfo * TLI)4103 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin(
4104     CallInst *CI, const TargetLibraryInfo *TLI) {
4105   Function *F = CI->getCalledFunction();
4106   LibFunc Func;
4107   if (F && !F->hasLocalLinkage() && F->hasName() &&
4108       TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) &&
4109       !F->doesNotAccessMemory())
4110     CI->addFnAttr(Attribute::NoBuiltin);
4111 }
4112 
canReplaceOperandWithVariable(const Instruction * I,unsigned OpIdx)4113 bool llvm::canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx) {
4114   // We can't have a PHI with a metadata type.
4115   if (I->getOperand(OpIdx)->getType()->isMetadataTy())
4116     return false;
4117 
4118   // Early exit.
4119   if (!isa<Constant>(I->getOperand(OpIdx)))
4120     return true;
4121 
4122   switch (I->getOpcode()) {
4123   default:
4124     return true;
4125   case Instruction::Call:
4126   case Instruction::Invoke: {
4127     const auto &CB = cast<CallBase>(*I);
4128 
4129     // Can't handle inline asm. Skip it.
4130     if (CB.isInlineAsm())
4131       return false;
4132 
4133     // Constant bundle operands may need to retain their constant-ness for
4134     // correctness.
4135     if (CB.isBundleOperand(OpIdx))
4136       return false;
4137 
4138     if (OpIdx < CB.arg_size()) {
4139       // Some variadic intrinsics require constants in the variadic arguments,
4140       // which currently aren't markable as immarg.
4141       if (isa<IntrinsicInst>(CB) &&
4142           OpIdx >= CB.getFunctionType()->getNumParams()) {
4143         // This is known to be OK for stackmap.
4144         return CB.getIntrinsicID() == Intrinsic::experimental_stackmap;
4145       }
4146 
4147       // gcroot is a special case, since it requires a constant argument which
4148       // isn't also required to be a simple ConstantInt.
4149       if (CB.getIntrinsicID() == Intrinsic::gcroot)
4150         return false;
4151 
4152       // Some intrinsic operands are required to be immediates.
4153       return !CB.paramHasAttr(OpIdx, Attribute::ImmArg);
4154     }
4155 
4156     // It is never allowed to replace the call argument to an intrinsic, but it
4157     // may be possible for a call.
4158     return !isa<IntrinsicInst>(CB);
4159   }
4160   case Instruction::ShuffleVector:
4161     // Shufflevector masks are constant.
4162     return OpIdx != 2;
4163   case Instruction::Switch:
4164   case Instruction::ExtractValue:
4165     // All operands apart from the first are constant.
4166     return OpIdx == 0;
4167   case Instruction::InsertValue:
4168     // All operands apart from the first and the second are constant.
4169     return OpIdx < 2;
4170   case Instruction::Alloca:
4171     // Static allocas (constant size in the entry block) are handled by
4172     // prologue/epilogue insertion so they're free anyway. We definitely don't
4173     // want to make them non-constant.
4174     return !cast<AllocaInst>(I)->isStaticAlloca();
4175   case Instruction::GetElementPtr:
4176     if (OpIdx == 0)
4177       return true;
4178     gep_type_iterator It = gep_type_begin(I);
4179     for (auto E = std::next(It, OpIdx); It != E; ++It)
4180       if (It.isStruct())
4181         return false;
4182     return true;
4183   }
4184 }
4185 
invertCondition(Value * Condition)4186 Value *llvm::invertCondition(Value *Condition) {
4187   // First: Check if it's a constant
4188   if (Constant *C = dyn_cast<Constant>(Condition))
4189     return ConstantExpr::getNot(C);
4190 
4191   // Second: If the condition is already inverted, return the original value
4192   Value *NotCondition;
4193   if (match(Condition, m_Not(m_Value(NotCondition))))
4194     return NotCondition;
4195 
4196   BasicBlock *Parent = nullptr;
4197   Instruction *Inst = dyn_cast<Instruction>(Condition);
4198   if (Inst)
4199     Parent = Inst->getParent();
4200   else if (Argument *Arg = dyn_cast<Argument>(Condition))
4201     Parent = &Arg->getParent()->getEntryBlock();
4202   assert(Parent && "Unsupported condition to invert");
4203 
4204   // Third: Check all the users for an invert
4205   for (User *U : Condition->users())
4206     if (Instruction *I = dyn_cast<Instruction>(U))
4207       if (I->getParent() == Parent && match(I, m_Not(m_Specific(Condition))))
4208         return I;
4209 
4210   // Last option: Create a new instruction
4211   auto *Inverted =
4212       BinaryOperator::CreateNot(Condition, Condition->getName() + ".inv");
4213   if (Inst && !isa<PHINode>(Inst))
4214     Inverted->insertAfter(Inst);
4215   else
4216     Inverted->insertBefore(&*Parent->getFirstInsertionPt());
4217   return Inverted;
4218 }
4219 
inferAttributesFromOthers(Function & F)4220 bool llvm::inferAttributesFromOthers(Function &F) {
4221   // Note: We explicitly check for attributes rather than using cover functions
4222   // because some of the cover functions include the logic being implemented.
4223 
4224   bool Changed = false;
4225   // readnone + not convergent implies nosync
4226   if (!F.hasFnAttribute(Attribute::NoSync) &&
4227       F.doesNotAccessMemory() && !F.isConvergent()) {
4228     F.setNoSync();
4229     Changed = true;
4230   }
4231 
4232   // readonly implies nofree
4233   if (!F.hasFnAttribute(Attribute::NoFree) && F.onlyReadsMemory()) {
4234     F.setDoesNotFreeMemory();
4235     Changed = true;
4236   }
4237 
4238   // willreturn implies mustprogress
4239   if (!F.hasFnAttribute(Attribute::MustProgress) && F.willReturn()) {
4240     F.setMustProgress();
4241     Changed = true;
4242   }
4243 
4244   // TODO: There are a bunch of cases of restrictive memory effects we
4245   // can infer by inspecting arguments of argmemonly-ish functions.
4246 
4247   return Changed;
4248 }
4249