• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- Local.cpp - Functions to perform local transformations -------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This family of functions perform various local transformations to the
11 // program.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DenseMapInfo.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/Hashing.h"
21 #include "llvm/ADT/None.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/ADT/TinyPtrVector.h"
29 #include "llvm/Analysis/ConstantFolding.h"
30 #include "llvm/Analysis/EHPersonalities.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/LazyValueInfo.h"
33 #include "llvm/Analysis/MemoryBuiltins.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/BinaryFormat/Dwarf.h"
37 #include "llvm/IR/Argument.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/BasicBlock.h"
40 #include "llvm/IR/CFG.h"
41 #include "llvm/IR/CallSite.h"
42 #include "llvm/IR/Constant.h"
43 #include "llvm/IR/ConstantRange.h"
44 #include "llvm/IR/Constants.h"
45 #include "llvm/IR/DIBuilder.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/DebugInfoMetadata.h"
48 #include "llvm/IR/DebugLoc.h"
49 #include "llvm/IR/DerivedTypes.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/GetElementPtrTypeIterator.h"
53 #include "llvm/IR/GlobalObject.h"
54 #include "llvm/IR/IRBuilder.h"
55 #include "llvm/IR/InstrTypes.h"
56 #include "llvm/IR/Instruction.h"
57 #include "llvm/IR/Instructions.h"
58 #include "llvm/IR/IntrinsicInst.h"
59 #include "llvm/IR/Intrinsics.h"
60 #include "llvm/IR/LLVMContext.h"
61 #include "llvm/IR/MDBuilder.h"
62 #include "llvm/IR/Metadata.h"
63 #include "llvm/IR/Module.h"
64 #include "llvm/IR/Operator.h"
65 #include "llvm/IR/PatternMatch.h"
66 #include "llvm/IR/Type.h"
67 #include "llvm/IR/Use.h"
68 #include "llvm/IR/User.h"
69 #include "llvm/IR/Value.h"
70 #include "llvm/IR/ValueHandle.h"
71 #include "llvm/Support/Casting.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/ErrorHandling.h"
74 #include "llvm/Support/KnownBits.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/ValueMapper.h"
77 #include <algorithm>
78 #include <cassert>
79 #include <climits>
80 #include <cstdint>
81 #include <iterator>
82 #include <map>
83 #include <utility>
84 
85 using namespace llvm;
86 using namespace llvm::PatternMatch;
87 
88 #define DEBUG_TYPE "local"
89 
90 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
91 
92 //===----------------------------------------------------------------------===//
93 //  Local constant propagation.
94 //
95 
96 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
97 /// constant value, convert it into an unconditional branch to the constant
98 /// destination.  This is a nontrivial operation because the successors of this
99 /// basic block must have their PHI nodes updated.
100 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
101 /// conditions and indirectbr addresses this might make dead if
102 /// DeleteDeadConditions is true.
ConstantFoldTerminator(BasicBlock * BB,bool DeleteDeadConditions,const TargetLibraryInfo * TLI,DeferredDominance * DDT)103 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
104                                   const TargetLibraryInfo *TLI,
105                                   DeferredDominance *DDT) {
106   TerminatorInst *T = BB->getTerminator();
107   IRBuilder<> Builder(T);
108 
109   // Branch - See if we are conditional jumping on constant
110   if (auto *BI = dyn_cast<BranchInst>(T)) {
111     if (BI->isUnconditional()) return false;  // Can't optimize uncond branch
112     BasicBlock *Dest1 = BI->getSuccessor(0);
113     BasicBlock *Dest2 = BI->getSuccessor(1);
114 
115     if (auto *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
116       // Are we branching on constant?
117       // YES.  Change to unconditional branch...
118       BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
119       BasicBlock *OldDest     = Cond->getZExtValue() ? Dest2 : Dest1;
120 
121       // Let the basic block know that we are letting go of it.  Based on this,
122       // it will adjust it's PHI nodes.
123       OldDest->removePredecessor(BB);
124 
125       // Replace the conditional branch with an unconditional one.
126       Builder.CreateBr(Destination);
127       BI->eraseFromParent();
128       if (DDT)
129         DDT->deleteEdge(BB, OldDest);
130       return true;
131     }
132 
133     if (Dest2 == Dest1) {       // Conditional branch to same location?
134       // This branch matches something like this:
135       //     br bool %cond, label %Dest, label %Dest
136       // and changes it into:  br label %Dest
137 
138       // Let the basic block know that we are letting go of one copy of it.
139       assert(BI->getParent() && "Terminator not inserted in block!");
140       Dest1->removePredecessor(BI->getParent());
141 
142       // Replace the conditional branch with an unconditional one.
143       Builder.CreateBr(Dest1);
144       Value *Cond = BI->getCondition();
145       BI->eraseFromParent();
146       if (DeleteDeadConditions)
147         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
148       return true;
149     }
150     return false;
151   }
152 
153   if (auto *SI = dyn_cast<SwitchInst>(T)) {
154     // If we are switching on a constant, we can convert the switch to an
155     // unconditional branch.
156     auto *CI = dyn_cast<ConstantInt>(SI->getCondition());
157     BasicBlock *DefaultDest = SI->getDefaultDest();
158     BasicBlock *TheOnlyDest = DefaultDest;
159 
160     // If the default is unreachable, ignore it when searching for TheOnlyDest.
161     if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
162         SI->getNumCases() > 0) {
163       TheOnlyDest = SI->case_begin()->getCaseSuccessor();
164     }
165 
166     // Figure out which case it goes to.
167     for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
168       // Found case matching a constant operand?
169       if (i->getCaseValue() == CI) {
170         TheOnlyDest = i->getCaseSuccessor();
171         break;
172       }
173 
174       // Check to see if this branch is going to the same place as the default
175       // dest.  If so, eliminate it as an explicit compare.
176       if (i->getCaseSuccessor() == DefaultDest) {
177         MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
178         unsigned NCases = SI->getNumCases();
179         // Fold the case metadata into the default if there will be any branches
180         // left, unless the metadata doesn't match the switch.
181         if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) {
182           // Collect branch weights into a vector.
183           SmallVector<uint32_t, 8> Weights;
184           for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
185                ++MD_i) {
186             auto *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
187             Weights.push_back(CI->getValue().getZExtValue());
188           }
189           // Merge weight of this case to the default weight.
190           unsigned idx = i->getCaseIndex();
191           Weights[0] += Weights[idx+1];
192           // Remove weight for this case.
193           std::swap(Weights[idx+1], Weights.back());
194           Weights.pop_back();
195           SI->setMetadata(LLVMContext::MD_prof,
196                           MDBuilder(BB->getContext()).
197                           createBranchWeights(Weights));
198         }
199         // Remove this entry.
200         BasicBlock *ParentBB = SI->getParent();
201         DefaultDest->removePredecessor(ParentBB);
202         i = SI->removeCase(i);
203         e = SI->case_end();
204         if (DDT)
205           DDT->deleteEdge(ParentBB, DefaultDest);
206         continue;
207       }
208 
209       // Otherwise, check to see if the switch only branches to one destination.
210       // We do this by reseting "TheOnlyDest" to null when we find two non-equal
211       // destinations.
212       if (i->getCaseSuccessor() != TheOnlyDest)
213         TheOnlyDest = nullptr;
214 
215       // Increment this iterator as we haven't removed the case.
216       ++i;
217     }
218 
219     if (CI && !TheOnlyDest) {
220       // Branching on a constant, but not any of the cases, go to the default
221       // successor.
222       TheOnlyDest = SI->getDefaultDest();
223     }
224 
225     // If we found a single destination that we can fold the switch into, do so
226     // now.
227     if (TheOnlyDest) {
228       // Insert the new branch.
229       Builder.CreateBr(TheOnlyDest);
230       BasicBlock *BB = SI->getParent();
231       std::vector <DominatorTree::UpdateType> Updates;
232       if (DDT)
233         Updates.reserve(SI->getNumSuccessors() - 1);
234 
235       // Remove entries from PHI nodes which we no longer branch to...
236       for (BasicBlock *Succ : SI->successors()) {
237         // Found case matching a constant operand?
238         if (Succ == TheOnlyDest) {
239           TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest
240         } else {
241           Succ->removePredecessor(BB);
242           if (DDT)
243             Updates.push_back({DominatorTree::Delete, BB, Succ});
244         }
245       }
246 
247       // Delete the old switch.
248       Value *Cond = SI->getCondition();
249       SI->eraseFromParent();
250       if (DeleteDeadConditions)
251         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
252       if (DDT)
253         DDT->applyUpdates(Updates);
254       return true;
255     }
256 
257     if (SI->getNumCases() == 1) {
258       // Otherwise, we can fold this switch into a conditional branch
259       // instruction if it has only one non-default destination.
260       auto FirstCase = *SI->case_begin();
261       Value *Cond = Builder.CreateICmpEQ(SI->getCondition(),
262           FirstCase.getCaseValue(), "cond");
263 
264       // Insert the new branch.
265       BranchInst *NewBr = Builder.CreateCondBr(Cond,
266                                                FirstCase.getCaseSuccessor(),
267                                                SI->getDefaultDest());
268       MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
269       if (MD && MD->getNumOperands() == 3) {
270         ConstantInt *SICase =
271             mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
272         ConstantInt *SIDef =
273             mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
274         assert(SICase && SIDef);
275         // The TrueWeight should be the weight for the single case of SI.
276         NewBr->setMetadata(LLVMContext::MD_prof,
277                         MDBuilder(BB->getContext()).
278                         createBranchWeights(SICase->getValue().getZExtValue(),
279                                             SIDef->getValue().getZExtValue()));
280       }
281 
282       // Update make.implicit metadata to the newly-created conditional branch.
283       MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit);
284       if (MakeImplicitMD)
285         NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD);
286 
287       // Delete the old switch.
288       SI->eraseFromParent();
289       return true;
290     }
291     return false;
292   }
293 
294   if (auto *IBI = dyn_cast<IndirectBrInst>(T)) {
295     // indirectbr blockaddress(@F, @BB) -> br label @BB
296     if (auto *BA =
297           dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) {
298       BasicBlock *TheOnlyDest = BA->getBasicBlock();
299       std::vector <DominatorTree::UpdateType> Updates;
300       if (DDT)
301         Updates.reserve(IBI->getNumDestinations() - 1);
302 
303       // Insert the new branch.
304       Builder.CreateBr(TheOnlyDest);
305 
306       for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
307         if (IBI->getDestination(i) == TheOnlyDest) {
308           TheOnlyDest = nullptr;
309         } else {
310           BasicBlock *ParentBB = IBI->getParent();
311           BasicBlock *DestBB = IBI->getDestination(i);
312           DestBB->removePredecessor(ParentBB);
313           if (DDT)
314             Updates.push_back({DominatorTree::Delete, ParentBB, DestBB});
315         }
316       }
317       Value *Address = IBI->getAddress();
318       IBI->eraseFromParent();
319       if (DeleteDeadConditions)
320         RecursivelyDeleteTriviallyDeadInstructions(Address, TLI);
321 
322       // If we didn't find our destination in the IBI successor list, then we
323       // have undefined behavior.  Replace the unconditional branch with an
324       // 'unreachable' instruction.
325       if (TheOnlyDest) {
326         BB->getTerminator()->eraseFromParent();
327         new UnreachableInst(BB->getContext(), BB);
328       }
329 
330       if (DDT)
331         DDT->applyUpdates(Updates);
332       return true;
333     }
334   }
335 
336   return false;
337 }
338 
339 //===----------------------------------------------------------------------===//
340 //  Local dead code elimination.
341 //
342 
343 /// isInstructionTriviallyDead - Return true if the result produced by the
344 /// instruction is not used, and the instruction has no side effects.
345 ///
isInstructionTriviallyDead(Instruction * I,const TargetLibraryInfo * TLI)346 bool llvm::isInstructionTriviallyDead(Instruction *I,
347                                       const TargetLibraryInfo *TLI) {
348   if (!I->use_empty())
349     return false;
350   return wouldInstructionBeTriviallyDead(I, TLI);
351 }
352 
wouldInstructionBeTriviallyDead(Instruction * I,const TargetLibraryInfo * TLI)353 bool llvm::wouldInstructionBeTriviallyDead(Instruction *I,
354                                            const TargetLibraryInfo *TLI) {
355   if (isa<TerminatorInst>(I))
356     return false;
357 
358   // We don't want the landingpad-like instructions removed by anything this
359   // general.
360   if (I->isEHPad())
361     return false;
362 
363   // We don't want debug info removed by anything this general, unless
364   // debug info is empty.
365   if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) {
366     if (DDI->getAddress())
367       return false;
368     return true;
369   }
370   if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) {
371     if (DVI->getValue())
372       return false;
373     return true;
374   }
375   if (DbgLabelInst *DLI = dyn_cast<DbgLabelInst>(I)) {
376     if (DLI->getLabel())
377       return false;
378     return true;
379   }
380 
381   if (!I->mayHaveSideEffects())
382     return true;
383 
384   // Special case intrinsics that "may have side effects" but can be deleted
385   // when dead.
386   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
387     // Safe to delete llvm.stacksave and launder.invariant.group if dead.
388     if (II->getIntrinsicID() == Intrinsic::stacksave ||
389         II->getIntrinsicID() == Intrinsic::launder_invariant_group)
390       return true;
391 
392     // Lifetime intrinsics are dead when their right-hand is undef.
393     if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
394         II->getIntrinsicID() == Intrinsic::lifetime_end)
395       return isa<UndefValue>(II->getArgOperand(1));
396 
397     // Assumptions are dead if their condition is trivially true.  Guards on
398     // true are operationally no-ops.  In the future we can consider more
399     // sophisticated tradeoffs for guards considering potential for check
400     // widening, but for now we keep things simple.
401     if (II->getIntrinsicID() == Intrinsic::assume ||
402         II->getIntrinsicID() == Intrinsic::experimental_guard) {
403       if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
404         return !Cond->isZero();
405 
406       return false;
407     }
408   }
409 
410   if (isAllocLikeFn(I, TLI))
411     return true;
412 
413   if (CallInst *CI = isFreeCall(I, TLI))
414     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
415       return C->isNullValue() || isa<UndefValue>(C);
416 
417   if (CallSite CS = CallSite(I))
418     if (isMathLibCallNoop(CS, TLI))
419       return true;
420 
421   return false;
422 }
423 
424 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
425 /// trivially dead instruction, delete it.  If that makes any of its operands
426 /// trivially dead, delete them too, recursively.  Return true if any
427 /// instructions were deleted.
428 bool
RecursivelyDeleteTriviallyDeadInstructions(Value * V,const TargetLibraryInfo * TLI)429 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
430                                                  const TargetLibraryInfo *TLI) {
431   Instruction *I = dyn_cast<Instruction>(V);
432   if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
433     return false;
434 
435   SmallVector<Instruction*, 16> DeadInsts;
436   DeadInsts.push_back(I);
437   RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI);
438 
439   return true;
440 }
441 
RecursivelyDeleteTriviallyDeadInstructions(SmallVectorImpl<Instruction * > & DeadInsts,const TargetLibraryInfo * TLI)442 void llvm::RecursivelyDeleteTriviallyDeadInstructions(
443     SmallVectorImpl<Instruction *> &DeadInsts, const TargetLibraryInfo *TLI) {
444   // Process the dead instruction list until empty.
445   while (!DeadInsts.empty()) {
446     Instruction &I = *DeadInsts.pop_back_val();
447     assert(I.use_empty() && "Instructions with uses are not dead.");
448     assert(isInstructionTriviallyDead(&I, TLI) &&
449            "Live instruction found in dead worklist!");
450 
451     // Don't lose the debug info while deleting the instructions.
452     salvageDebugInfo(I);
453 
454     // Null out all of the instruction's operands to see if any operand becomes
455     // dead as we go.
456     for (Use &OpU : I.operands()) {
457       Value *OpV = OpU.get();
458       OpU.set(nullptr);
459 
460       if (!OpV->use_empty())
461         continue;
462 
463       // If the operand is an instruction that became dead as we nulled out the
464       // operand, and if it is 'trivially' dead, delete it in a future loop
465       // iteration.
466       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
467         if (isInstructionTriviallyDead(OpI, TLI))
468           DeadInsts.push_back(OpI);
469     }
470 
471     I.eraseFromParent();
472   }
473 }
474 
475 /// areAllUsesEqual - Check whether the uses of a value are all the same.
476 /// This is similar to Instruction::hasOneUse() except this will also return
477 /// true when there are no uses or multiple uses that all refer to the same
478 /// value.
areAllUsesEqual(Instruction * I)479 static bool areAllUsesEqual(Instruction *I) {
480   Value::user_iterator UI = I->user_begin();
481   Value::user_iterator UE = I->user_end();
482   if (UI == UE)
483     return true;
484 
485   User *TheUse = *UI;
486   for (++UI; UI != UE; ++UI) {
487     if (*UI != TheUse)
488       return false;
489   }
490   return true;
491 }
492 
493 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
494 /// dead PHI node, due to being a def-use chain of single-use nodes that
495 /// either forms a cycle or is terminated by a trivially dead instruction,
496 /// delete it.  If that makes any of its operands trivially dead, delete them
497 /// too, recursively.  Return true if a change was made.
RecursivelyDeleteDeadPHINode(PHINode * PN,const TargetLibraryInfo * TLI)498 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
499                                         const TargetLibraryInfo *TLI) {
500   SmallPtrSet<Instruction*, 4> Visited;
501   for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
502        I = cast<Instruction>(*I->user_begin())) {
503     if (I->use_empty())
504       return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
505 
506     // If we find an instruction more than once, we're on a cycle that
507     // won't prove fruitful.
508     if (!Visited.insert(I).second) {
509       // Break the cycle and delete the instruction and its operands.
510       I->replaceAllUsesWith(UndefValue::get(I->getType()));
511       (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
512       return true;
513     }
514   }
515   return false;
516 }
517 
518 static bool
simplifyAndDCEInstruction(Instruction * I,SmallSetVector<Instruction *,16> & WorkList,const DataLayout & DL,const TargetLibraryInfo * TLI)519 simplifyAndDCEInstruction(Instruction *I,
520                           SmallSetVector<Instruction *, 16> &WorkList,
521                           const DataLayout &DL,
522                           const TargetLibraryInfo *TLI) {
523   if (isInstructionTriviallyDead(I, TLI)) {
524     salvageDebugInfo(*I);
525 
526     // Null out all of the instruction's operands to see if any operand becomes
527     // dead as we go.
528     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
529       Value *OpV = I->getOperand(i);
530       I->setOperand(i, nullptr);
531 
532       if (!OpV->use_empty() || I == OpV)
533         continue;
534 
535       // If the operand is an instruction that became dead as we nulled out the
536       // operand, and if it is 'trivially' dead, delete it in a future loop
537       // iteration.
538       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
539         if (isInstructionTriviallyDead(OpI, TLI))
540           WorkList.insert(OpI);
541     }
542 
543     I->eraseFromParent();
544 
545     return true;
546   }
547 
548   if (Value *SimpleV = SimplifyInstruction(I, DL)) {
549     // Add the users to the worklist. CAREFUL: an instruction can use itself,
550     // in the case of a phi node.
551     for (User *U : I->users()) {
552       if (U != I) {
553         WorkList.insert(cast<Instruction>(U));
554       }
555     }
556 
557     // Replace the instruction with its simplified value.
558     bool Changed = false;
559     if (!I->use_empty()) {
560       I->replaceAllUsesWith(SimpleV);
561       Changed = true;
562     }
563     if (isInstructionTriviallyDead(I, TLI)) {
564       I->eraseFromParent();
565       Changed = true;
566     }
567     return Changed;
568   }
569   return false;
570 }
571 
572 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
573 /// simplify any instructions in it and recursively delete dead instructions.
574 ///
575 /// This returns true if it changed the code, note that it can delete
576 /// instructions in other blocks as well in this block.
SimplifyInstructionsInBlock(BasicBlock * BB,const TargetLibraryInfo * TLI)577 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
578                                        const TargetLibraryInfo *TLI) {
579   bool MadeChange = false;
580   const DataLayout &DL = BB->getModule()->getDataLayout();
581 
582 #ifndef NDEBUG
583   // In debug builds, ensure that the terminator of the block is never replaced
584   // or deleted by these simplifications. The idea of simplification is that it
585   // cannot introduce new instructions, and there is no way to replace the
586   // terminator of a block without introducing a new instruction.
587   AssertingVH<Instruction> TerminatorVH(&BB->back());
588 #endif
589 
590   SmallSetVector<Instruction *, 16> WorkList;
591   // Iterate over the original function, only adding insts to the worklist
592   // if they actually need to be revisited. This avoids having to pre-init
593   // the worklist with the entire function's worth of instructions.
594   for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end());
595        BI != E;) {
596     assert(!BI->isTerminator());
597     Instruction *I = &*BI;
598     ++BI;
599 
600     // We're visiting this instruction now, so make sure it's not in the
601     // worklist from an earlier visit.
602     if (!WorkList.count(I))
603       MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
604   }
605 
606   while (!WorkList.empty()) {
607     Instruction *I = WorkList.pop_back_val();
608     MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
609   }
610   return MadeChange;
611 }
612 
613 //===----------------------------------------------------------------------===//
614 //  Control Flow Graph Restructuring.
615 //
616 
617 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
618 /// method is called when we're about to delete Pred as a predecessor of BB.  If
619 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
620 ///
621 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
622 /// nodes that collapse into identity values.  For example, if we have:
623 ///   x = phi(1, 0, 0, 0)
624 ///   y = and x, z
625 ///
626 /// .. and delete the predecessor corresponding to the '1', this will attempt to
627 /// recursively fold the and to 0.
RemovePredecessorAndSimplify(BasicBlock * BB,BasicBlock * Pred,DeferredDominance * DDT)628 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred,
629                                         DeferredDominance *DDT) {
630   // This only adjusts blocks with PHI nodes.
631   if (!isa<PHINode>(BB->begin()))
632     return;
633 
634   // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
635   // them down.  This will leave us with single entry phi nodes and other phis
636   // that can be removed.
637   BB->removePredecessor(Pred, true);
638 
639   WeakTrackingVH PhiIt = &BB->front();
640   while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
641     PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
642     Value *OldPhiIt = PhiIt;
643 
644     if (!recursivelySimplifyInstruction(PN))
645       continue;
646 
647     // If recursive simplification ended up deleting the next PHI node we would
648     // iterate to, then our iterator is invalid, restart scanning from the top
649     // of the block.
650     if (PhiIt != OldPhiIt) PhiIt = &BB->front();
651   }
652   if (DDT)
653     DDT->deleteEdge(Pred, BB);
654 }
655 
656 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
657 /// predecessor is known to have one successor (DestBB!).  Eliminate the edge
658 /// between them, moving the instructions in the predecessor into DestBB and
659 /// deleting the predecessor block.
MergeBasicBlockIntoOnlyPred(BasicBlock * DestBB,DominatorTree * DT,DeferredDominance * DDT)660 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT,
661                                        DeferredDominance *DDT) {
662   assert(!(DT && DDT) && "Cannot call with both DT and DDT.");
663 
664   // If BB has single-entry PHI nodes, fold them.
665   while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
666     Value *NewVal = PN->getIncomingValue(0);
667     // Replace self referencing PHI with undef, it must be dead.
668     if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
669     PN->replaceAllUsesWith(NewVal);
670     PN->eraseFromParent();
671   }
672 
673   BasicBlock *PredBB = DestBB->getSinglePredecessor();
674   assert(PredBB && "Block doesn't have a single predecessor!");
675 
676   bool ReplaceEntryBB = false;
677   if (PredBB == &DestBB->getParent()->getEntryBlock())
678     ReplaceEntryBB = true;
679 
680   // Deferred DT update: Collect all the edges that enter PredBB. These
681   // dominator edges will be redirected to DestBB.
682   std::vector <DominatorTree::UpdateType> Updates;
683   if (DDT && !ReplaceEntryBB) {
684     Updates.reserve(1 + (2 * pred_size(PredBB)));
685     Updates.push_back({DominatorTree::Delete, PredBB, DestBB});
686     for (auto I = pred_begin(PredBB), E = pred_end(PredBB); I != E; ++I) {
687       Updates.push_back({DominatorTree::Delete, *I, PredBB});
688       // This predecessor of PredBB may already have DestBB as a successor.
689       if (llvm::find(successors(*I), DestBB) == succ_end(*I))
690         Updates.push_back({DominatorTree::Insert, *I, DestBB});
691     }
692   }
693 
694   // Zap anything that took the address of DestBB.  Not doing this will give the
695   // address an invalid value.
696   if (DestBB->hasAddressTaken()) {
697     BlockAddress *BA = BlockAddress::get(DestBB);
698     Constant *Replacement =
699       ConstantInt::get(Type::getInt32Ty(BA->getContext()), 1);
700     BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
701                                                      BA->getType()));
702     BA->destroyConstant();
703   }
704 
705   // Anything that branched to PredBB now branches to DestBB.
706   PredBB->replaceAllUsesWith(DestBB);
707 
708   // Splice all the instructions from PredBB to DestBB.
709   PredBB->getTerminator()->eraseFromParent();
710   DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
711 
712   // If the PredBB is the entry block of the function, move DestBB up to
713   // become the entry block after we erase PredBB.
714   if (ReplaceEntryBB)
715     DestBB->moveAfter(PredBB);
716 
717   if (DT) {
718     // For some irreducible CFG we end up having forward-unreachable blocks
719     // so check if getNode returns a valid node before updating the domtree.
720     if (DomTreeNode *DTN = DT->getNode(PredBB)) {
721       BasicBlock *PredBBIDom = DTN->getIDom()->getBlock();
722       DT->changeImmediateDominator(DestBB, PredBBIDom);
723       DT->eraseNode(PredBB);
724     }
725   }
726 
727   if (DDT) {
728     DDT->deleteBB(PredBB); // Deferred deletion of BB.
729     if (ReplaceEntryBB)
730       // The entry block was removed and there is no external interface for the
731       // dominator tree to be notified of this change. In this corner-case we
732       // recalculate the entire tree.
733       DDT->recalculate(*(DestBB->getParent()));
734     else
735       DDT->applyUpdates(Updates);
736   } else {
737     PredBB->eraseFromParent(); // Nuke BB.
738   }
739 }
740 
741 /// CanMergeValues - Return true if we can choose one of these values to use
742 /// in place of the other. Note that we will always choose the non-undef
743 /// value to keep.
CanMergeValues(Value * First,Value * Second)744 static bool CanMergeValues(Value *First, Value *Second) {
745   return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
746 }
747 
748 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
749 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
750 ///
751 /// Assumption: Succ is the single successor for BB.
CanPropagatePredecessorsForPHIs(BasicBlock * BB,BasicBlock * Succ)752 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
753   assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
754 
755   LLVM_DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
756                     << Succ->getName() << "\n");
757   // Shortcut, if there is only a single predecessor it must be BB and merging
758   // is always safe
759   if (Succ->getSinglePredecessor()) return true;
760 
761   // Make a list of the predecessors of BB
762   SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
763 
764   // Look at all the phi nodes in Succ, to see if they present a conflict when
765   // merging these blocks
766   for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
767     PHINode *PN = cast<PHINode>(I);
768 
769     // If the incoming value from BB is again a PHINode in
770     // BB which has the same incoming value for *PI as PN does, we can
771     // merge the phi nodes and then the blocks can still be merged
772     PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
773     if (BBPN && BBPN->getParent() == BB) {
774       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
775         BasicBlock *IBB = PN->getIncomingBlock(PI);
776         if (BBPreds.count(IBB) &&
777             !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
778                             PN->getIncomingValue(PI))) {
779           LLVM_DEBUG(dbgs()
780                      << "Can't fold, phi node " << PN->getName() << " in "
781                      << Succ->getName() << " is conflicting with "
782                      << BBPN->getName() << " with regard to common predecessor "
783                      << IBB->getName() << "\n");
784           return false;
785         }
786       }
787     } else {
788       Value* Val = PN->getIncomingValueForBlock(BB);
789       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
790         // See if the incoming value for the common predecessor is equal to the
791         // one for BB, in which case this phi node will not prevent the merging
792         // of the block.
793         BasicBlock *IBB = PN->getIncomingBlock(PI);
794         if (BBPreds.count(IBB) &&
795             !CanMergeValues(Val, PN->getIncomingValue(PI))) {
796           LLVM_DEBUG(dbgs() << "Can't fold, phi node " << PN->getName()
797                             << " in " << Succ->getName()
798                             << " is conflicting with regard to common "
799                             << "predecessor " << IBB->getName() << "\n");
800           return false;
801         }
802       }
803     }
804   }
805 
806   return true;
807 }
808 
809 using PredBlockVector = SmallVector<BasicBlock *, 16>;
810 using IncomingValueMap = DenseMap<BasicBlock *, Value *>;
811 
812 /// Determines the value to use as the phi node input for a block.
813 ///
814 /// Select between \p OldVal any value that we know flows from \p BB
815 /// to a particular phi on the basis of which one (if either) is not
816 /// undef. Update IncomingValues based on the selected value.
817 ///
818 /// \param OldVal The value we are considering selecting.
819 /// \param BB The block that the value flows in from.
820 /// \param IncomingValues A map from block-to-value for other phi inputs
821 /// that we have examined.
822 ///
823 /// \returns the selected value.
selectIncomingValueForBlock(Value * OldVal,BasicBlock * BB,IncomingValueMap & IncomingValues)824 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
825                                           IncomingValueMap &IncomingValues) {
826   if (!isa<UndefValue>(OldVal)) {
827     assert((!IncomingValues.count(BB) ||
828             IncomingValues.find(BB)->second == OldVal) &&
829            "Expected OldVal to match incoming value from BB!");
830 
831     IncomingValues.insert(std::make_pair(BB, OldVal));
832     return OldVal;
833   }
834 
835   IncomingValueMap::const_iterator It = IncomingValues.find(BB);
836   if (It != IncomingValues.end()) return It->second;
837 
838   return OldVal;
839 }
840 
841 /// Create a map from block to value for the operands of a
842 /// given phi.
843 ///
844 /// Create a map from block to value for each non-undef value flowing
845 /// into \p PN.
846 ///
847 /// \param PN The phi we are collecting the map for.
848 /// \param IncomingValues [out] The map from block to value for this phi.
gatherIncomingValuesToPhi(PHINode * PN,IncomingValueMap & IncomingValues)849 static void gatherIncomingValuesToPhi(PHINode *PN,
850                                       IncomingValueMap &IncomingValues) {
851   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
852     BasicBlock *BB = PN->getIncomingBlock(i);
853     Value *V = PN->getIncomingValue(i);
854 
855     if (!isa<UndefValue>(V))
856       IncomingValues.insert(std::make_pair(BB, V));
857   }
858 }
859 
860 /// Replace the incoming undef values to a phi with the values
861 /// from a block-to-value map.
862 ///
863 /// \param PN The phi we are replacing the undefs in.
864 /// \param IncomingValues A map from block to value.
replaceUndefValuesInPhi(PHINode * PN,const IncomingValueMap & IncomingValues)865 static void replaceUndefValuesInPhi(PHINode *PN,
866                                     const IncomingValueMap &IncomingValues) {
867   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
868     Value *V = PN->getIncomingValue(i);
869 
870     if (!isa<UndefValue>(V)) continue;
871 
872     BasicBlock *BB = PN->getIncomingBlock(i);
873     IncomingValueMap::const_iterator It = IncomingValues.find(BB);
874     if (It == IncomingValues.end()) continue;
875 
876     PN->setIncomingValue(i, It->second);
877   }
878 }
879 
880 /// Replace a value flowing from a block to a phi with
881 /// potentially multiple instances of that value flowing from the
882 /// block's predecessors to the phi.
883 ///
884 /// \param BB The block with the value flowing into the phi.
885 /// \param BBPreds The predecessors of BB.
886 /// \param PN The phi that we are updating.
redirectValuesFromPredecessorsToPhi(BasicBlock * BB,const PredBlockVector & BBPreds,PHINode * PN)887 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
888                                                 const PredBlockVector &BBPreds,
889                                                 PHINode *PN) {
890   Value *OldVal = PN->removeIncomingValue(BB, false);
891   assert(OldVal && "No entry in PHI for Pred BB!");
892 
893   IncomingValueMap IncomingValues;
894 
895   // We are merging two blocks - BB, and the block containing PN - and
896   // as a result we need to redirect edges from the predecessors of BB
897   // to go to the block containing PN, and update PN
898   // accordingly. Since we allow merging blocks in the case where the
899   // predecessor and successor blocks both share some predecessors,
900   // and where some of those common predecessors might have undef
901   // values flowing into PN, we want to rewrite those values to be
902   // consistent with the non-undef values.
903 
904   gatherIncomingValuesToPhi(PN, IncomingValues);
905 
906   // If this incoming value is one of the PHI nodes in BB, the new entries
907   // in the PHI node are the entries from the old PHI.
908   if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
909     PHINode *OldValPN = cast<PHINode>(OldVal);
910     for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
911       // Note that, since we are merging phi nodes and BB and Succ might
912       // have common predecessors, we could end up with a phi node with
913       // identical incoming branches. This will be cleaned up later (and
914       // will trigger asserts if we try to clean it up now, without also
915       // simplifying the corresponding conditional branch).
916       BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
917       Value *PredVal = OldValPN->getIncomingValue(i);
918       Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
919                                                     IncomingValues);
920 
921       // And add a new incoming value for this predecessor for the
922       // newly retargeted branch.
923       PN->addIncoming(Selected, PredBB);
924     }
925   } else {
926     for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
927       // Update existing incoming values in PN for this
928       // predecessor of BB.
929       BasicBlock *PredBB = BBPreds[i];
930       Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
931                                                     IncomingValues);
932 
933       // And add a new incoming value for this predecessor for the
934       // newly retargeted branch.
935       PN->addIncoming(Selected, PredBB);
936     }
937   }
938 
939   replaceUndefValuesInPhi(PN, IncomingValues);
940 }
941 
942 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
943 /// unconditional branch, and contains no instructions other than PHI nodes,
944 /// potential side-effect free intrinsics and the branch.  If possible,
945 /// eliminate BB by rewriting all the predecessors to branch to the successor
946 /// block and return true.  If we can't transform, return false.
TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock * BB,DeferredDominance * DDT)947 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB,
948                                                    DeferredDominance *DDT) {
949   assert(BB != &BB->getParent()->getEntryBlock() &&
950          "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
951 
952   // We can't eliminate infinite loops.
953   BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
954   if (BB == Succ) return false;
955 
956   // Check to see if merging these blocks would cause conflicts for any of the
957   // phi nodes in BB or Succ. If not, we can safely merge.
958   if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
959 
960   // Check for cases where Succ has multiple predecessors and a PHI node in BB
961   // has uses which will not disappear when the PHI nodes are merged.  It is
962   // possible to handle such cases, but difficult: it requires checking whether
963   // BB dominates Succ, which is non-trivial to calculate in the case where
964   // Succ has multiple predecessors.  Also, it requires checking whether
965   // constructing the necessary self-referential PHI node doesn't introduce any
966   // conflicts; this isn't too difficult, but the previous code for doing this
967   // was incorrect.
968   //
969   // Note that if this check finds a live use, BB dominates Succ, so BB is
970   // something like a loop pre-header (or rarely, a part of an irreducible CFG);
971   // folding the branch isn't profitable in that case anyway.
972   if (!Succ->getSinglePredecessor()) {
973     BasicBlock::iterator BBI = BB->begin();
974     while (isa<PHINode>(*BBI)) {
975       for (Use &U : BBI->uses()) {
976         if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
977           if (PN->getIncomingBlock(U) != BB)
978             return false;
979         } else {
980           return false;
981         }
982       }
983       ++BBI;
984     }
985   }
986 
987   LLVM_DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
988 
989   std::vector<DominatorTree::UpdateType> Updates;
990   if (DDT) {
991     Updates.reserve(1 + (2 * pred_size(BB)));
992     Updates.push_back({DominatorTree::Delete, BB, Succ});
993     // All predecessors of BB will be moved to Succ.
994     for (auto I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
995       Updates.push_back({DominatorTree::Delete, *I, BB});
996       // This predecessor of BB may already have Succ as a successor.
997       if (llvm::find(successors(*I), Succ) == succ_end(*I))
998         Updates.push_back({DominatorTree::Insert, *I, Succ});
999     }
1000   }
1001 
1002   if (isa<PHINode>(Succ->begin())) {
1003     // If there is more than one pred of succ, and there are PHI nodes in
1004     // the successor, then we need to add incoming edges for the PHI nodes
1005     //
1006     const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
1007 
1008     // Loop over all of the PHI nodes in the successor of BB.
1009     for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
1010       PHINode *PN = cast<PHINode>(I);
1011 
1012       redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
1013     }
1014   }
1015 
1016   if (Succ->getSinglePredecessor()) {
1017     // BB is the only predecessor of Succ, so Succ will end up with exactly
1018     // the same predecessors BB had.
1019 
1020     // Copy over any phi, debug or lifetime instruction.
1021     BB->getTerminator()->eraseFromParent();
1022     Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
1023                                BB->getInstList());
1024   } else {
1025     while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
1026       // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
1027       assert(PN->use_empty() && "There shouldn't be any uses here!");
1028       PN->eraseFromParent();
1029     }
1030   }
1031 
1032   // If the unconditional branch we replaced contains llvm.loop metadata, we
1033   // add the metadata to the branch instructions in the predecessors.
1034   unsigned LoopMDKind = BB->getContext().getMDKindID("llvm.loop");
1035   Instruction *TI = BB->getTerminator();
1036   if (TI)
1037     if (MDNode *LoopMD = TI->getMetadata(LoopMDKind))
1038       for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1039         BasicBlock *Pred = *PI;
1040         Pred->getTerminator()->setMetadata(LoopMDKind, LoopMD);
1041       }
1042 
1043   // Everything that jumped to BB now goes to Succ.
1044   BB->replaceAllUsesWith(Succ);
1045   if (!Succ->hasName()) Succ->takeName(BB);
1046 
1047   if (DDT) {
1048     DDT->deleteBB(BB); // Deferred deletion of the old basic block.
1049     DDT->applyUpdates(Updates);
1050   } else {
1051     BB->eraseFromParent(); // Delete the old basic block.
1052   }
1053   return true;
1054 }
1055 
1056 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
1057 /// nodes in this block. This doesn't try to be clever about PHI nodes
1058 /// which differ only in the order of the incoming values, but instcombine
1059 /// orders them so it usually won't matter.
EliminateDuplicatePHINodes(BasicBlock * BB)1060 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
1061   // This implementation doesn't currently consider undef operands
1062   // specially. Theoretically, two phis which are identical except for
1063   // one having an undef where the other doesn't could be collapsed.
1064 
1065   struct PHIDenseMapInfo {
1066     static PHINode *getEmptyKey() {
1067       return DenseMapInfo<PHINode *>::getEmptyKey();
1068     }
1069 
1070     static PHINode *getTombstoneKey() {
1071       return DenseMapInfo<PHINode *>::getTombstoneKey();
1072     }
1073 
1074     static unsigned getHashValue(PHINode *PN) {
1075       // Compute a hash value on the operands. Instcombine will likely have
1076       // sorted them, which helps expose duplicates, but we have to check all
1077       // the operands to be safe in case instcombine hasn't run.
1078       return static_cast<unsigned>(hash_combine(
1079           hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
1080           hash_combine_range(PN->block_begin(), PN->block_end())));
1081     }
1082 
1083     static bool isEqual(PHINode *LHS, PHINode *RHS) {
1084       if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1085           RHS == getEmptyKey() || RHS == getTombstoneKey())
1086         return LHS == RHS;
1087       return LHS->isIdenticalTo(RHS);
1088     }
1089   };
1090 
1091   // Set of unique PHINodes.
1092   DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
1093 
1094   // Examine each PHI.
1095   bool Changed = false;
1096   for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
1097     auto Inserted = PHISet.insert(PN);
1098     if (!Inserted.second) {
1099       // A duplicate. Replace this PHI with its duplicate.
1100       PN->replaceAllUsesWith(*Inserted.first);
1101       PN->eraseFromParent();
1102       Changed = true;
1103 
1104       // The RAUW can change PHIs that we already visited. Start over from the
1105       // beginning.
1106       PHISet.clear();
1107       I = BB->begin();
1108     }
1109   }
1110 
1111   return Changed;
1112 }
1113 
1114 /// enforceKnownAlignment - If the specified pointer points to an object that
1115 /// we control, modify the object's alignment to PrefAlign. This isn't
1116 /// often possible though. If alignment is important, a more reliable approach
1117 /// is to simply align all global variables and allocation instructions to
1118 /// their preferred alignment from the beginning.
enforceKnownAlignment(Value * V,unsigned Align,unsigned PrefAlign,const DataLayout & DL)1119 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
1120                                       unsigned PrefAlign,
1121                                       const DataLayout &DL) {
1122   assert(PrefAlign > Align);
1123 
1124   V = V->stripPointerCasts();
1125 
1126   if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1127     // TODO: ideally, computeKnownBits ought to have used
1128     // AllocaInst::getAlignment() in its computation already, making
1129     // the below max redundant. But, as it turns out,
1130     // stripPointerCasts recurses through infinite layers of bitcasts,
1131     // while computeKnownBits is not allowed to traverse more than 6
1132     // levels.
1133     Align = std::max(AI->getAlignment(), Align);
1134     if (PrefAlign <= Align)
1135       return Align;
1136 
1137     // If the preferred alignment is greater than the natural stack alignment
1138     // then don't round up. This avoids dynamic stack realignment.
1139     if (DL.exceedsNaturalStackAlignment(PrefAlign))
1140       return Align;
1141     AI->setAlignment(PrefAlign);
1142     return PrefAlign;
1143   }
1144 
1145   if (auto *GO = dyn_cast<GlobalObject>(V)) {
1146     // TODO: as above, this shouldn't be necessary.
1147     Align = std::max(GO->getAlignment(), Align);
1148     if (PrefAlign <= Align)
1149       return Align;
1150 
1151     // If there is a large requested alignment and we can, bump up the alignment
1152     // of the global.  If the memory we set aside for the global may not be the
1153     // memory used by the final program then it is impossible for us to reliably
1154     // enforce the preferred alignment.
1155     if (!GO->canIncreaseAlignment())
1156       return Align;
1157 
1158     GO->setAlignment(PrefAlign);
1159     return PrefAlign;
1160   }
1161 
1162   return Align;
1163 }
1164 
getOrEnforceKnownAlignment(Value * V,unsigned PrefAlign,const DataLayout & DL,const Instruction * CxtI,AssumptionCache * AC,const DominatorTree * DT)1165 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
1166                                           const DataLayout &DL,
1167                                           const Instruction *CxtI,
1168                                           AssumptionCache *AC,
1169                                           const DominatorTree *DT) {
1170   assert(V->getType()->isPointerTy() &&
1171          "getOrEnforceKnownAlignment expects a pointer!");
1172 
1173   KnownBits Known = computeKnownBits(V, DL, 0, AC, CxtI, DT);
1174   unsigned TrailZ = Known.countMinTrailingZeros();
1175 
1176   // Avoid trouble with ridiculously large TrailZ values, such as
1177   // those computed from a null pointer.
1178   TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1179 
1180   unsigned Align = 1u << std::min(Known.getBitWidth() - 1, TrailZ);
1181 
1182   // LLVM doesn't support alignments larger than this currently.
1183   Align = std::min(Align, +Value::MaximumAlignment);
1184 
1185   if (PrefAlign > Align)
1186     Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1187 
1188   // We don't need to make any adjustment.
1189   return Align;
1190 }
1191 
1192 ///===---------------------------------------------------------------------===//
1193 ///  Dbg Intrinsic utilities
1194 ///
1195 
1196 /// See if there is a dbg.value intrinsic for DIVar before I.
LdStHasDebugValue(DILocalVariable * DIVar,DIExpression * DIExpr,Instruction * I)1197 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr,
1198                               Instruction *I) {
1199   // Since we can't guarantee that the original dbg.declare instrinsic
1200   // is removed by LowerDbgDeclare(), we need to make sure that we are
1201   // not inserting the same dbg.value intrinsic over and over.
1202   BasicBlock::InstListType::iterator PrevI(I);
1203   if (PrevI != I->getParent()->getInstList().begin()) {
1204     --PrevI;
1205     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1206       if (DVI->getValue() == I->getOperand(0) &&
1207           DVI->getVariable() == DIVar &&
1208           DVI->getExpression() == DIExpr)
1209         return true;
1210   }
1211   return false;
1212 }
1213 
1214 /// See if there is a dbg.value intrinsic for DIVar for the PHI node.
PhiHasDebugValue(DILocalVariable * DIVar,DIExpression * DIExpr,PHINode * APN)1215 static bool PhiHasDebugValue(DILocalVariable *DIVar,
1216                              DIExpression *DIExpr,
1217                              PHINode *APN) {
1218   // Since we can't guarantee that the original dbg.declare instrinsic
1219   // is removed by LowerDbgDeclare(), we need to make sure that we are
1220   // not inserting the same dbg.value intrinsic over and over.
1221   SmallVector<DbgValueInst *, 1> DbgValues;
1222   findDbgValues(DbgValues, APN);
1223   for (auto *DVI : DbgValues) {
1224     assert(DVI->getValue() == APN);
1225     if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr))
1226       return true;
1227   }
1228   return false;
1229 }
1230 
1231 /// Check if the alloc size of \p ValTy is large enough to cover the variable
1232 /// (or fragment of the variable) described by \p DII.
1233 ///
1234 /// This is primarily intended as a helper for the different
1235 /// ConvertDebugDeclareToDebugValue functions. The dbg.declare/dbg.addr that is
1236 /// converted describes an alloca'd variable, so we need to use the
1237 /// alloc size of the value when doing the comparison. E.g. an i1 value will be
1238 /// identified as covering an n-bit fragment, if the store size of i1 is at
1239 /// least n bits.
valueCoversEntireFragment(Type * ValTy,DbgInfoIntrinsic * DII)1240 static bool valueCoversEntireFragment(Type *ValTy, DbgInfoIntrinsic *DII) {
1241   const DataLayout &DL = DII->getModule()->getDataLayout();
1242   uint64_t ValueSize = DL.getTypeAllocSizeInBits(ValTy);
1243   if (auto FragmentSize = DII->getFragmentSizeInBits())
1244     return ValueSize >= *FragmentSize;
1245   // We can't always calculate the size of the DI variable (e.g. if it is a
1246   // VLA). Try to use the size of the alloca that the dbg intrinsic describes
1247   // intead.
1248   if (DII->isAddressOfVariable())
1249     if (auto *AI = dyn_cast_or_null<AllocaInst>(DII->getVariableLocation()))
1250       if (auto FragmentSize = AI->getAllocationSizeInBits(DL))
1251         return ValueSize >= *FragmentSize;
1252   // Could not determine size of variable. Conservatively return false.
1253   return false;
1254 }
1255 
1256 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1257 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic.
ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic * DII,StoreInst * SI,DIBuilder & Builder)1258 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1259                                            StoreInst *SI, DIBuilder &Builder) {
1260   assert(DII->isAddressOfVariable());
1261   auto *DIVar = DII->getVariable();
1262   assert(DIVar && "Missing variable");
1263   auto *DIExpr = DII->getExpression();
1264   Value *DV = SI->getOperand(0);
1265 
1266   if (!valueCoversEntireFragment(SI->getValueOperand()->getType(), DII)) {
1267     // FIXME: If storing to a part of the variable described by the dbg.declare,
1268     // then we want to insert a dbg.value for the corresponding fragment.
1269     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: "
1270                       << *DII << '\n');
1271     // For now, when there is a store to parts of the variable (but we do not
1272     // know which part) we insert an dbg.value instrinsic to indicate that we
1273     // know nothing about the variable's content.
1274     DV = UndefValue::get(DV->getType());
1275     if (!LdStHasDebugValue(DIVar, DIExpr, SI))
1276       Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, DII->getDebugLoc(),
1277                                       SI);
1278     return;
1279   }
1280 
1281   // If an argument is zero extended then use argument directly. The ZExt
1282   // may be zapped by an optimization pass in future.
1283   Argument *ExtendedArg = nullptr;
1284   if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1285     ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1286   if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1287     ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1288   if (ExtendedArg) {
1289     // If this DII was already describing only a fragment of a variable, ensure
1290     // that fragment is appropriately narrowed here.
1291     // But if a fragment wasn't used, describe the value as the original
1292     // argument (rather than the zext or sext) so that it remains described even
1293     // if the sext/zext is optimized away. This widens the variable description,
1294     // leaving it up to the consumer to know how the smaller value may be
1295     // represented in a larger register.
1296     if (auto Fragment = DIExpr->getFragmentInfo()) {
1297       unsigned FragmentOffset = Fragment->OffsetInBits;
1298       SmallVector<uint64_t, 3> Ops(DIExpr->elements_begin(),
1299                                    DIExpr->elements_end() - 3);
1300       Ops.push_back(dwarf::DW_OP_LLVM_fragment);
1301       Ops.push_back(FragmentOffset);
1302       const DataLayout &DL = DII->getModule()->getDataLayout();
1303       Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType()));
1304       DIExpr = Builder.createExpression(Ops);
1305     }
1306     DV = ExtendedArg;
1307   }
1308   if (!LdStHasDebugValue(DIVar, DIExpr, SI))
1309     Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, DII->getDebugLoc(),
1310                                     SI);
1311 }
1312 
1313 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1314 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic.
ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic * DII,LoadInst * LI,DIBuilder & Builder)1315 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1316                                            LoadInst *LI, DIBuilder &Builder) {
1317   auto *DIVar = DII->getVariable();
1318   auto *DIExpr = DII->getExpression();
1319   assert(DIVar && "Missing variable");
1320 
1321   if (LdStHasDebugValue(DIVar, DIExpr, LI))
1322     return;
1323 
1324   if (!valueCoversEntireFragment(LI->getType(), DII)) {
1325     // FIXME: If only referring to a part of the variable described by the
1326     // dbg.declare, then we want to insert a dbg.value for the corresponding
1327     // fragment.
1328     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: "
1329                       << *DII << '\n');
1330     return;
1331   }
1332 
1333   // We are now tracking the loaded value instead of the address. In the
1334   // future if multi-location support is added to the IR, it might be
1335   // preferable to keep tracking both the loaded value and the original
1336   // address in case the alloca can not be elided.
1337   Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1338       LI, DIVar, DIExpr, DII->getDebugLoc(), (Instruction *)nullptr);
1339   DbgValue->insertAfter(LI);
1340 }
1341 
1342 /// Inserts a llvm.dbg.value intrinsic after a phi that has an associated
1343 /// llvm.dbg.declare or llvm.dbg.addr intrinsic.
ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic * DII,PHINode * APN,DIBuilder & Builder)1344 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1345                                            PHINode *APN, DIBuilder &Builder) {
1346   auto *DIVar = DII->getVariable();
1347   auto *DIExpr = DII->getExpression();
1348   assert(DIVar && "Missing variable");
1349 
1350   if (PhiHasDebugValue(DIVar, DIExpr, APN))
1351     return;
1352 
1353   if (!valueCoversEntireFragment(APN->getType(), DII)) {
1354     // FIXME: If only referring to a part of the variable described by the
1355     // dbg.declare, then we want to insert a dbg.value for the corresponding
1356     // fragment.
1357     LLVM_DEBUG(dbgs() << "Failed to convert dbg.declare to dbg.value: "
1358                       << *DII << '\n');
1359     return;
1360   }
1361 
1362   BasicBlock *BB = APN->getParent();
1363   auto InsertionPt = BB->getFirstInsertionPt();
1364 
1365   // The block may be a catchswitch block, which does not have a valid
1366   // insertion point.
1367   // FIXME: Insert dbg.value markers in the successors when appropriate.
1368   if (InsertionPt != BB->end())
1369     Builder.insertDbgValueIntrinsic(APN, DIVar, DIExpr, DII->getDebugLoc(),
1370                                     &*InsertionPt);
1371 }
1372 
1373 /// Determine whether this alloca is either a VLA or an array.
isArray(AllocaInst * AI)1374 static bool isArray(AllocaInst *AI) {
1375   return AI->isArrayAllocation() ||
1376     AI->getType()->getElementType()->isArrayTy();
1377 }
1378 
1379 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1380 /// of llvm.dbg.value intrinsics.
LowerDbgDeclare(Function & F)1381 bool llvm::LowerDbgDeclare(Function &F) {
1382   DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1383   SmallVector<DbgDeclareInst *, 4> Dbgs;
1384   for (auto &FI : F)
1385     for (Instruction &BI : FI)
1386       if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1387         Dbgs.push_back(DDI);
1388 
1389   if (Dbgs.empty())
1390     return false;
1391 
1392   for (auto &I : Dbgs) {
1393     DbgDeclareInst *DDI = I;
1394     AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1395     // If this is an alloca for a scalar variable, insert a dbg.value
1396     // at each load and store to the alloca and erase the dbg.declare.
1397     // The dbg.values allow tracking a variable even if it is not
1398     // stored on the stack, while the dbg.declare can only describe
1399     // the stack slot (and at a lexical-scope granularity). Later
1400     // passes will attempt to elide the stack slot.
1401     if (!AI || isArray(AI))
1402       continue;
1403 
1404     // A volatile load/store means that the alloca can't be elided anyway.
1405     if (llvm::any_of(AI->users(), [](User *U) -> bool {
1406           if (LoadInst *LI = dyn_cast<LoadInst>(U))
1407             return LI->isVolatile();
1408           if (StoreInst *SI = dyn_cast<StoreInst>(U))
1409             return SI->isVolatile();
1410           return false;
1411         }))
1412       continue;
1413 
1414     for (auto &AIUse : AI->uses()) {
1415       User *U = AIUse.getUser();
1416       if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1417         if (AIUse.getOperandNo() == 1)
1418           ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1419       } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1420         ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1421       } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1422         // This is a call by-value or some other instruction that takes a
1423         // pointer to the variable. Insert a *value* intrinsic that describes
1424         // the variable by dereferencing the alloca.
1425         auto *DerefExpr =
1426             DIExpression::append(DDI->getExpression(), dwarf::DW_OP_deref);
1427         DIB.insertDbgValueIntrinsic(AI, DDI->getVariable(), DerefExpr,
1428                                     DDI->getDebugLoc(), CI);
1429       }
1430     }
1431     DDI->eraseFromParent();
1432   }
1433   return true;
1434 }
1435 
1436 /// Propagate dbg.value intrinsics through the newly inserted PHIs.
insertDebugValuesForPHIs(BasicBlock * BB,SmallVectorImpl<PHINode * > & InsertedPHIs)1437 void llvm::insertDebugValuesForPHIs(BasicBlock *BB,
1438                                     SmallVectorImpl<PHINode *> &InsertedPHIs) {
1439   assert(BB && "No BasicBlock to clone dbg.value(s) from.");
1440   if (InsertedPHIs.size() == 0)
1441     return;
1442 
1443   // Map existing PHI nodes to their dbg.values.
1444   ValueToValueMapTy DbgValueMap;
1445   for (auto &I : *BB) {
1446     if (auto DbgII = dyn_cast<DbgInfoIntrinsic>(&I)) {
1447       if (auto *Loc = dyn_cast_or_null<PHINode>(DbgII->getVariableLocation()))
1448         DbgValueMap.insert({Loc, DbgII});
1449     }
1450   }
1451   if (DbgValueMap.size() == 0)
1452     return;
1453 
1454   // Then iterate through the new PHIs and look to see if they use one of the
1455   // previously mapped PHIs. If so, insert a new dbg.value intrinsic that will
1456   // propagate the info through the new PHI.
1457   LLVMContext &C = BB->getContext();
1458   for (auto PHI : InsertedPHIs) {
1459     BasicBlock *Parent = PHI->getParent();
1460     // Avoid inserting an intrinsic into an EH block.
1461     if (Parent->getFirstNonPHI()->isEHPad())
1462       continue;
1463     auto PhiMAV = MetadataAsValue::get(C, ValueAsMetadata::get(PHI));
1464     for (auto VI : PHI->operand_values()) {
1465       auto V = DbgValueMap.find(VI);
1466       if (V != DbgValueMap.end()) {
1467         auto *DbgII = cast<DbgInfoIntrinsic>(V->second);
1468         Instruction *NewDbgII = DbgII->clone();
1469         NewDbgII->setOperand(0, PhiMAV);
1470         auto InsertionPt = Parent->getFirstInsertionPt();
1471         assert(InsertionPt != Parent->end() && "Ill-formed basic block");
1472         NewDbgII->insertBefore(&*InsertionPt);
1473       }
1474     }
1475   }
1476 }
1477 
1478 /// Finds all intrinsics declaring local variables as living in the memory that
1479 /// 'V' points to. This may include a mix of dbg.declare and
1480 /// dbg.addr intrinsics.
FindDbgAddrUses(Value * V)1481 TinyPtrVector<DbgInfoIntrinsic *> llvm::FindDbgAddrUses(Value *V) {
1482   // This function is hot. Check whether the value has any metadata to avoid a
1483   // DenseMap lookup.
1484   if (!V->isUsedByMetadata())
1485     return {};
1486   auto *L = LocalAsMetadata::getIfExists(V);
1487   if (!L)
1488     return {};
1489   auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L);
1490   if (!MDV)
1491     return {};
1492 
1493   TinyPtrVector<DbgInfoIntrinsic *> Declares;
1494   for (User *U : MDV->users()) {
1495     if (auto *DII = dyn_cast<DbgInfoIntrinsic>(U))
1496       if (DII->isAddressOfVariable())
1497         Declares.push_back(DII);
1498   }
1499 
1500   return Declares;
1501 }
1502 
findDbgValues(SmallVectorImpl<DbgValueInst * > & DbgValues,Value * V)1503 void llvm::findDbgValues(SmallVectorImpl<DbgValueInst *> &DbgValues, Value *V) {
1504   // This function is hot. Check whether the value has any metadata to avoid a
1505   // DenseMap lookup.
1506   if (!V->isUsedByMetadata())
1507     return;
1508   if (auto *L = LocalAsMetadata::getIfExists(V))
1509     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1510       for (User *U : MDV->users())
1511         if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1512           DbgValues.push_back(DVI);
1513 }
1514 
findDbgUsers(SmallVectorImpl<DbgInfoIntrinsic * > & DbgUsers,Value * V)1515 void llvm::findDbgUsers(SmallVectorImpl<DbgInfoIntrinsic *> &DbgUsers,
1516                         Value *V) {
1517   // This function is hot. Check whether the value has any metadata to avoid a
1518   // DenseMap lookup.
1519   if (!V->isUsedByMetadata())
1520     return;
1521   if (auto *L = LocalAsMetadata::getIfExists(V))
1522     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1523       for (User *U : MDV->users())
1524         if (DbgInfoIntrinsic *DII = dyn_cast<DbgInfoIntrinsic>(U))
1525           DbgUsers.push_back(DII);
1526 }
1527 
replaceDbgDeclare(Value * Address,Value * NewAddress,Instruction * InsertBefore,DIBuilder & Builder,bool DerefBefore,int Offset,bool DerefAfter)1528 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1529                              Instruction *InsertBefore, DIBuilder &Builder,
1530                              bool DerefBefore, int Offset, bool DerefAfter) {
1531   auto DbgAddrs = FindDbgAddrUses(Address);
1532   for (DbgInfoIntrinsic *DII : DbgAddrs) {
1533     DebugLoc Loc = DII->getDebugLoc();
1534     auto *DIVar = DII->getVariable();
1535     auto *DIExpr = DII->getExpression();
1536     assert(DIVar && "Missing variable");
1537     DIExpr = DIExpression::prepend(DIExpr, DerefBefore, Offset, DerefAfter);
1538     // Insert llvm.dbg.declare immediately before InsertBefore, and remove old
1539     // llvm.dbg.declare.
1540     Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1541     if (DII == InsertBefore)
1542       InsertBefore = InsertBefore->getNextNode();
1543     DII->eraseFromParent();
1544   }
1545   return !DbgAddrs.empty();
1546 }
1547 
replaceDbgDeclareForAlloca(AllocaInst * AI,Value * NewAllocaAddress,DIBuilder & Builder,bool DerefBefore,int Offset,bool DerefAfter)1548 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1549                                       DIBuilder &Builder, bool DerefBefore,
1550                                       int Offset, bool DerefAfter) {
1551   return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1552                            DerefBefore, Offset, DerefAfter);
1553 }
1554 
replaceOneDbgValueForAlloca(DbgValueInst * DVI,Value * NewAddress,DIBuilder & Builder,int Offset)1555 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress,
1556                                         DIBuilder &Builder, int Offset) {
1557   DebugLoc Loc = DVI->getDebugLoc();
1558   auto *DIVar = DVI->getVariable();
1559   auto *DIExpr = DVI->getExpression();
1560   assert(DIVar && "Missing variable");
1561 
1562   // This is an alloca-based llvm.dbg.value. The first thing it should do with
1563   // the alloca pointer is dereference it. Otherwise we don't know how to handle
1564   // it and give up.
1565   if (!DIExpr || DIExpr->getNumElements() < 1 ||
1566       DIExpr->getElement(0) != dwarf::DW_OP_deref)
1567     return;
1568 
1569   // Insert the offset immediately after the first deref.
1570   // We could just change the offset argument of dbg.value, but it's unsigned...
1571   if (Offset) {
1572     SmallVector<uint64_t, 4> Ops;
1573     Ops.push_back(dwarf::DW_OP_deref);
1574     DIExpression::appendOffset(Ops, Offset);
1575     Ops.append(DIExpr->elements_begin() + 1, DIExpr->elements_end());
1576     DIExpr = Builder.createExpression(Ops);
1577   }
1578 
1579   Builder.insertDbgValueIntrinsic(NewAddress, DIVar, DIExpr, Loc, DVI);
1580   DVI->eraseFromParent();
1581 }
1582 
replaceDbgValueForAlloca(AllocaInst * AI,Value * NewAllocaAddress,DIBuilder & Builder,int Offset)1583 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1584                                     DIBuilder &Builder, int Offset) {
1585   if (auto *L = LocalAsMetadata::getIfExists(AI))
1586     if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
1587       for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) {
1588         Use &U = *UI++;
1589         if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser()))
1590           replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset);
1591       }
1592 }
1593 
1594 /// Wrap \p V in a ValueAsMetadata instance.
wrapValueInMetadata(LLVMContext & C,Value * V)1595 static MetadataAsValue *wrapValueInMetadata(LLVMContext &C, Value *V) {
1596   return MetadataAsValue::get(C, ValueAsMetadata::get(V));
1597 }
1598 
salvageDebugInfo(Instruction & I)1599 bool llvm::salvageDebugInfo(Instruction &I) {
1600   SmallVector<DbgInfoIntrinsic *, 1> DbgUsers;
1601   findDbgUsers(DbgUsers, &I);
1602   if (DbgUsers.empty())
1603     return false;
1604 
1605   auto &M = *I.getModule();
1606   auto &DL = M.getDataLayout();
1607   auto &Ctx = I.getContext();
1608   auto wrapMD = [&](Value *V) { return wrapValueInMetadata(Ctx, V); };
1609 
1610   auto doSalvage = [&](DbgInfoIntrinsic *DII, SmallVectorImpl<uint64_t> &Ops) {
1611     auto *DIExpr = DII->getExpression();
1612     if (!Ops.empty()) {
1613       // Do not add DW_OP_stack_value for DbgDeclare and DbgAddr, because they
1614       // are implicitly pointing out the value as a DWARF memory location
1615       // description.
1616       bool WithStackValue = isa<DbgValueInst>(DII);
1617       DIExpr = DIExpression::prependOpcodes(DIExpr, Ops, WithStackValue);
1618     }
1619     DII->setOperand(0, wrapMD(I.getOperand(0)));
1620     DII->setOperand(2, MetadataAsValue::get(Ctx, DIExpr));
1621     LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n');
1622   };
1623 
1624   auto applyOffset = [&](DbgInfoIntrinsic *DII, uint64_t Offset) {
1625     SmallVector<uint64_t, 8> Ops;
1626     DIExpression::appendOffset(Ops, Offset);
1627     doSalvage(DII, Ops);
1628   };
1629 
1630   auto applyOps = [&](DbgInfoIntrinsic *DII,
1631                       std::initializer_list<uint64_t> Opcodes) {
1632     SmallVector<uint64_t, 8> Ops(Opcodes);
1633     doSalvage(DII, Ops);
1634   };
1635 
1636   if (auto *CI = dyn_cast<CastInst>(&I)) {
1637     if (!CI->isNoopCast(DL))
1638       return false;
1639 
1640     // No-op casts are irrelevant for debug info.
1641     MetadataAsValue *CastSrc = wrapMD(I.getOperand(0));
1642     for (auto *DII : DbgUsers) {
1643       DII->setOperand(0, CastSrc);
1644       LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n');
1645     }
1646     return true;
1647   } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1648     unsigned BitWidth =
1649         M.getDataLayout().getIndexSizeInBits(GEP->getPointerAddressSpace());
1650     // Rewrite a constant GEP into a DIExpression.  Since we are performing
1651     // arithmetic to compute the variable's *value* in the DIExpression, we
1652     // need to mark the expression with a DW_OP_stack_value.
1653     APInt Offset(BitWidth, 0);
1654     if (GEP->accumulateConstantOffset(M.getDataLayout(), Offset))
1655       for (auto *DII : DbgUsers)
1656         applyOffset(DII, Offset.getSExtValue());
1657     return true;
1658   } else if (auto *BI = dyn_cast<BinaryOperator>(&I)) {
1659     // Rewrite binary operations with constant integer operands.
1660     auto *ConstInt = dyn_cast<ConstantInt>(I.getOperand(1));
1661     if (!ConstInt || ConstInt->getBitWidth() > 64)
1662       return false;
1663 
1664     uint64_t Val = ConstInt->getSExtValue();
1665     for (auto *DII : DbgUsers) {
1666       switch (BI->getOpcode()) {
1667       case Instruction::Add:
1668         applyOffset(DII, Val);
1669         break;
1670       case Instruction::Sub:
1671         applyOffset(DII, -int64_t(Val));
1672         break;
1673       case Instruction::Mul:
1674         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mul});
1675         break;
1676       case Instruction::SDiv:
1677         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_div});
1678         break;
1679       case Instruction::SRem:
1680         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mod});
1681         break;
1682       case Instruction::Or:
1683         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_or});
1684         break;
1685       case Instruction::And:
1686         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_and});
1687         break;
1688       case Instruction::Xor:
1689         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_xor});
1690         break;
1691       case Instruction::Shl:
1692         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shl});
1693         break;
1694       case Instruction::LShr:
1695         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shr});
1696         break;
1697       case Instruction::AShr:
1698         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shra});
1699         break;
1700       default:
1701         // TODO: Salvage constants from each kind of binop we know about.
1702         return false;
1703       }
1704     }
1705     return true;
1706   } else if (isa<LoadInst>(&I)) {
1707     MetadataAsValue *AddrMD = wrapMD(I.getOperand(0));
1708     for (auto *DII : DbgUsers) {
1709       // Rewrite the load into DW_OP_deref.
1710       auto *DIExpr = DII->getExpression();
1711       DIExpr = DIExpression::prepend(DIExpr, DIExpression::WithDeref);
1712       DII->setOperand(0, AddrMD);
1713       DII->setOperand(2, MetadataAsValue::get(Ctx, DIExpr));
1714       LLVM_DEBUG(dbgs() << "SALVAGE:  " << *DII << '\n');
1715     }
1716     return true;
1717   }
1718   return false;
1719 }
1720 
1721 /// A replacement for a dbg.value expression.
1722 using DbgValReplacement = Optional<DIExpression *>;
1723 
1724 /// Point debug users of \p From to \p To using exprs given by \p RewriteExpr,
1725 /// possibly moving/deleting users to prevent use-before-def. Returns true if
1726 /// changes are made.
rewriteDebugUsers(Instruction & From,Value & To,Instruction & DomPoint,DominatorTree & DT,function_ref<DbgValReplacement (DbgInfoIntrinsic & DII)> RewriteExpr)1727 static bool rewriteDebugUsers(
1728     Instruction &From, Value &To, Instruction &DomPoint, DominatorTree &DT,
1729     function_ref<DbgValReplacement(DbgInfoIntrinsic &DII)> RewriteExpr) {
1730   // Find debug users of From.
1731   SmallVector<DbgInfoIntrinsic *, 1> Users;
1732   findDbgUsers(Users, &From);
1733   if (Users.empty())
1734     return false;
1735 
1736   // Prevent use-before-def of To.
1737   bool Changed = false;
1738   SmallPtrSet<DbgInfoIntrinsic *, 1> DeleteOrSalvage;
1739   if (isa<Instruction>(&To)) {
1740     bool DomPointAfterFrom = From.getNextNonDebugInstruction() == &DomPoint;
1741 
1742     for (auto *DII : Users) {
1743       // It's common to see a debug user between From and DomPoint. Move it
1744       // after DomPoint to preserve the variable update without any reordering.
1745       if (DomPointAfterFrom && DII->getNextNonDebugInstruction() == &DomPoint) {
1746         LLVM_DEBUG(dbgs() << "MOVE:  " << *DII << '\n');
1747         DII->moveAfter(&DomPoint);
1748         Changed = true;
1749 
1750       // Users which otherwise aren't dominated by the replacement value must
1751       // be salvaged or deleted.
1752       } else if (!DT.dominates(&DomPoint, DII)) {
1753         DeleteOrSalvage.insert(DII);
1754       }
1755     }
1756   }
1757 
1758   // Update debug users without use-before-def risk.
1759   for (auto *DII : Users) {
1760     if (DeleteOrSalvage.count(DII))
1761       continue;
1762 
1763     LLVMContext &Ctx = DII->getContext();
1764     DbgValReplacement DVR = RewriteExpr(*DII);
1765     if (!DVR)
1766       continue;
1767 
1768     DII->setOperand(0, wrapValueInMetadata(Ctx, &To));
1769     DII->setOperand(2, MetadataAsValue::get(Ctx, *DVR));
1770     LLVM_DEBUG(dbgs() << "REWRITE:  " << *DII << '\n');
1771     Changed = true;
1772   }
1773 
1774   if (!DeleteOrSalvage.empty()) {
1775     // Try to salvage the remaining debug users.
1776     Changed |= salvageDebugInfo(From);
1777 
1778     // Delete the debug users which weren't salvaged.
1779     for (auto *DII : DeleteOrSalvage) {
1780       if (DII->getVariableLocation() == &From) {
1781         LLVM_DEBUG(dbgs() << "Erased UseBeforeDef:  " << *DII << '\n');
1782         DII->eraseFromParent();
1783         Changed = true;
1784       }
1785     }
1786   }
1787 
1788   return Changed;
1789 }
1790 
1791 /// Check if a bitcast between a value of type \p FromTy to type \p ToTy would
1792 /// losslessly preserve the bits and semantics of the value. This predicate is
1793 /// symmetric, i.e swapping \p FromTy and \p ToTy should give the same result.
1794 ///
1795 /// Note that Type::canLosslesslyBitCastTo is not suitable here because it
1796 /// allows semantically unequivalent bitcasts, such as <2 x i64> -> <4 x i32>,
1797 /// and also does not allow lossless pointer <-> integer conversions.
isBitCastSemanticsPreserving(const DataLayout & DL,Type * FromTy,Type * ToTy)1798 static bool isBitCastSemanticsPreserving(const DataLayout &DL, Type *FromTy,
1799                                          Type *ToTy) {
1800   // Trivially compatible types.
1801   if (FromTy == ToTy)
1802     return true;
1803 
1804   // Handle compatible pointer <-> integer conversions.
1805   if (FromTy->isIntOrPtrTy() && ToTy->isIntOrPtrTy()) {
1806     bool SameSize = DL.getTypeSizeInBits(FromTy) == DL.getTypeSizeInBits(ToTy);
1807     bool LosslessConversion = !DL.isNonIntegralPointerType(FromTy) &&
1808                               !DL.isNonIntegralPointerType(ToTy);
1809     return SameSize && LosslessConversion;
1810   }
1811 
1812   // TODO: This is not exhaustive.
1813   return false;
1814 }
1815 
replaceAllDbgUsesWith(Instruction & From,Value & To,Instruction & DomPoint,DominatorTree & DT)1816 bool llvm::replaceAllDbgUsesWith(Instruction &From, Value &To,
1817                                  Instruction &DomPoint, DominatorTree &DT) {
1818   // Exit early if From has no debug users.
1819   if (!From.isUsedByMetadata())
1820     return false;
1821 
1822   assert(&From != &To && "Can't replace something with itself");
1823 
1824   Type *FromTy = From.getType();
1825   Type *ToTy = To.getType();
1826 
1827   auto Identity = [&](DbgInfoIntrinsic &DII) -> DbgValReplacement {
1828     return DII.getExpression();
1829   };
1830 
1831   // Handle no-op conversions.
1832   Module &M = *From.getModule();
1833   const DataLayout &DL = M.getDataLayout();
1834   if (isBitCastSemanticsPreserving(DL, FromTy, ToTy))
1835     return rewriteDebugUsers(From, To, DomPoint, DT, Identity);
1836 
1837   // Handle integer-to-integer widening and narrowing.
1838   // FIXME: Use DW_OP_convert when it's available everywhere.
1839   if (FromTy->isIntegerTy() && ToTy->isIntegerTy()) {
1840     uint64_t FromBits = FromTy->getPrimitiveSizeInBits();
1841     uint64_t ToBits = ToTy->getPrimitiveSizeInBits();
1842     assert(FromBits != ToBits && "Unexpected no-op conversion");
1843 
1844     // When the width of the result grows, assume that a debugger will only
1845     // access the low `FromBits` bits when inspecting the source variable.
1846     if (FromBits < ToBits)
1847       return rewriteDebugUsers(From, To, DomPoint, DT, Identity);
1848 
1849     // The width of the result has shrunk. Use sign/zero extension to describe
1850     // the source variable's high bits.
1851     auto SignOrZeroExt = [&](DbgInfoIntrinsic &DII) -> DbgValReplacement {
1852       DILocalVariable *Var = DII.getVariable();
1853 
1854       // Without knowing signedness, sign/zero extension isn't possible.
1855       auto Signedness = Var->getSignedness();
1856       if (!Signedness)
1857         return None;
1858 
1859       bool Signed = *Signedness == DIBasicType::Signedness::Signed;
1860 
1861       if (!Signed) {
1862         // In the unsigned case, assume that a debugger will initialize the
1863         // high bits to 0 and do a no-op conversion.
1864         return Identity(DII);
1865       } else {
1866         // In the signed case, the high bits are given by sign extension, i.e:
1867         //   (To >> (ToBits - 1)) * ((2 ^ FromBits) - 1)
1868         // Calculate the high bits and OR them together with the low bits.
1869         SmallVector<uint64_t, 8> Ops({dwarf::DW_OP_dup, dwarf::DW_OP_constu,
1870                                       (ToBits - 1), dwarf::DW_OP_shr,
1871                                       dwarf::DW_OP_lit0, dwarf::DW_OP_not,
1872                                       dwarf::DW_OP_mul, dwarf::DW_OP_or});
1873         return DIExpression::appendToStack(DII.getExpression(), Ops);
1874       }
1875     };
1876     return rewriteDebugUsers(From, To, DomPoint, DT, SignOrZeroExt);
1877   }
1878 
1879   // TODO: Floating-point conversions, vectors.
1880   return false;
1881 }
1882 
removeAllNonTerminatorAndEHPadInstructions(BasicBlock * BB)1883 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) {
1884   unsigned NumDeadInst = 0;
1885   // Delete the instructions backwards, as it has a reduced likelihood of
1886   // having to update as many def-use and use-def chains.
1887   Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1888   while (EndInst != &BB->front()) {
1889     // Delete the next to last instruction.
1890     Instruction *Inst = &*--EndInst->getIterator();
1891     if (!Inst->use_empty() && !Inst->getType()->isTokenTy())
1892       Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1893     if (Inst->isEHPad() || Inst->getType()->isTokenTy()) {
1894       EndInst = Inst;
1895       continue;
1896     }
1897     if (!isa<DbgInfoIntrinsic>(Inst))
1898       ++NumDeadInst;
1899     Inst->eraseFromParent();
1900   }
1901   return NumDeadInst;
1902 }
1903 
changeToUnreachable(Instruction * I,bool UseLLVMTrap,bool PreserveLCSSA,DeferredDominance * DDT)1904 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap,
1905                                    bool PreserveLCSSA, DeferredDominance *DDT) {
1906   BasicBlock *BB = I->getParent();
1907   std::vector <DominatorTree::UpdateType> Updates;
1908 
1909   // Loop over all of the successors, removing BB's entry from any PHI
1910   // nodes.
1911   if (DDT)
1912     Updates.reserve(BB->getTerminator()->getNumSuccessors());
1913   for (BasicBlock *Successor : successors(BB)) {
1914     Successor->removePredecessor(BB, PreserveLCSSA);
1915     if (DDT)
1916       Updates.push_back({DominatorTree::Delete, BB, Successor});
1917   }
1918   // Insert a call to llvm.trap right before this.  This turns the undefined
1919   // behavior into a hard fail instead of falling through into random code.
1920   if (UseLLVMTrap) {
1921     Function *TrapFn =
1922       Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1923     CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1924     CallTrap->setDebugLoc(I->getDebugLoc());
1925   }
1926   new UnreachableInst(I->getContext(), I);
1927 
1928   // All instructions after this are dead.
1929   unsigned NumInstrsRemoved = 0;
1930   BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1931   while (BBI != BBE) {
1932     if (!BBI->use_empty())
1933       BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1934     BB->getInstList().erase(BBI++);
1935     ++NumInstrsRemoved;
1936   }
1937   if (DDT)
1938     DDT->applyUpdates(Updates);
1939   return NumInstrsRemoved;
1940 }
1941 
1942 /// changeToCall - Convert the specified invoke into a normal call.
changeToCall(InvokeInst * II,DeferredDominance * DDT=nullptr)1943 static void changeToCall(InvokeInst *II, DeferredDominance *DDT = nullptr) {
1944   SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1945   SmallVector<OperandBundleDef, 1> OpBundles;
1946   II->getOperandBundlesAsDefs(OpBundles);
1947   CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1948                                        "", II);
1949   NewCall->takeName(II);
1950   NewCall->setCallingConv(II->getCallingConv());
1951   NewCall->setAttributes(II->getAttributes());
1952   NewCall->setDebugLoc(II->getDebugLoc());
1953   II->replaceAllUsesWith(NewCall);
1954 
1955   // Follow the call by a branch to the normal destination.
1956   BasicBlock *NormalDestBB = II->getNormalDest();
1957   BranchInst::Create(NormalDestBB, II);
1958 
1959   // Update PHI nodes in the unwind destination
1960   BasicBlock *BB = II->getParent();
1961   BasicBlock *UnwindDestBB = II->getUnwindDest();
1962   UnwindDestBB->removePredecessor(BB);
1963   II->eraseFromParent();
1964   if (DDT)
1965     DDT->deleteEdge(BB, UnwindDestBB);
1966 }
1967 
changeToInvokeAndSplitBasicBlock(CallInst * CI,BasicBlock * UnwindEdge)1968 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI,
1969                                                    BasicBlock *UnwindEdge) {
1970   BasicBlock *BB = CI->getParent();
1971 
1972   // Convert this function call into an invoke instruction.  First, split the
1973   // basic block.
1974   BasicBlock *Split =
1975       BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc");
1976 
1977   // Delete the unconditional branch inserted by splitBasicBlock
1978   BB->getInstList().pop_back();
1979 
1980   // Create the new invoke instruction.
1981   SmallVector<Value *, 8> InvokeArgs(CI->arg_begin(), CI->arg_end());
1982   SmallVector<OperandBundleDef, 1> OpBundles;
1983 
1984   CI->getOperandBundlesAsDefs(OpBundles);
1985 
1986   // Note: we're round tripping operand bundles through memory here, and that
1987   // can potentially be avoided with a cleverer API design that we do not have
1988   // as of this time.
1989 
1990   InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge,
1991                                       InvokeArgs, OpBundles, CI->getName(), BB);
1992   II->setDebugLoc(CI->getDebugLoc());
1993   II->setCallingConv(CI->getCallingConv());
1994   II->setAttributes(CI->getAttributes());
1995 
1996   // Make sure that anything using the call now uses the invoke!  This also
1997   // updates the CallGraph if present, because it uses a WeakTrackingVH.
1998   CI->replaceAllUsesWith(II);
1999 
2000   // Delete the original call
2001   Split->getInstList().pop_front();
2002   return Split;
2003 }
2004 
markAliveBlocks(Function & F,SmallPtrSetImpl<BasicBlock * > & Reachable,DeferredDominance * DDT=nullptr)2005 static bool markAliveBlocks(Function &F,
2006                             SmallPtrSetImpl<BasicBlock*> &Reachable,
2007                             DeferredDominance *DDT = nullptr) {
2008   SmallVector<BasicBlock*, 128> Worklist;
2009   BasicBlock *BB = &F.front();
2010   Worklist.push_back(BB);
2011   Reachable.insert(BB);
2012   bool Changed = false;
2013   do {
2014     BB = Worklist.pop_back_val();
2015 
2016     // Do a quick scan of the basic block, turning any obviously unreachable
2017     // instructions into LLVM unreachable insts.  The instruction combining pass
2018     // canonicalizes unreachable insts into stores to null or undef.
2019     for (Instruction &I : *BB) {
2020       if (auto *CI = dyn_cast<CallInst>(&I)) {
2021         Value *Callee = CI->getCalledValue();
2022         // Handle intrinsic calls.
2023         if (Function *F = dyn_cast<Function>(Callee)) {
2024           auto IntrinsicID = F->getIntrinsicID();
2025           // Assumptions that are known to be false are equivalent to
2026           // unreachable. Also, if the condition is undefined, then we make the
2027           // choice most beneficial to the optimizer, and choose that to also be
2028           // unreachable.
2029           if (IntrinsicID == Intrinsic::assume) {
2030             if (match(CI->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) {
2031               // Don't insert a call to llvm.trap right before the unreachable.
2032               changeToUnreachable(CI, false, false, DDT);
2033               Changed = true;
2034               break;
2035             }
2036           } else if (IntrinsicID == Intrinsic::experimental_guard) {
2037             // A call to the guard intrinsic bails out of the current
2038             // compilation unit if the predicate passed to it is false. If the
2039             // predicate is a constant false, then we know the guard will bail
2040             // out of the current compile unconditionally, so all code following
2041             // it is dead.
2042             //
2043             // Note: unlike in llvm.assume, it is not "obviously profitable" for
2044             // guards to treat `undef` as `false` since a guard on `undef` can
2045             // still be useful for widening.
2046             if (match(CI->getArgOperand(0), m_Zero()))
2047               if (!isa<UnreachableInst>(CI->getNextNode())) {
2048                 changeToUnreachable(CI->getNextNode(), /*UseLLVMTrap=*/false,
2049                                     false, DDT);
2050                 Changed = true;
2051                 break;
2052               }
2053           }
2054         } else if ((isa<ConstantPointerNull>(Callee) &&
2055                     !NullPointerIsDefined(CI->getFunction())) ||
2056                    isa<UndefValue>(Callee)) {
2057           changeToUnreachable(CI, /*UseLLVMTrap=*/false, false, DDT);
2058           Changed = true;
2059           break;
2060         }
2061         if (CI->doesNotReturn()) {
2062           // If we found a call to a no-return function, insert an unreachable
2063           // instruction after it.  Make sure there isn't *already* one there
2064           // though.
2065           if (!isa<UnreachableInst>(CI->getNextNode())) {
2066             // Don't insert a call to llvm.trap right before the unreachable.
2067             changeToUnreachable(CI->getNextNode(), false, false, DDT);
2068             Changed = true;
2069           }
2070           break;
2071         }
2072       } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
2073         // Store to undef and store to null are undefined and used to signal
2074         // that they should be changed to unreachable by passes that can't
2075         // modify the CFG.
2076 
2077         // Don't touch volatile stores.
2078         if (SI->isVolatile()) continue;
2079 
2080         Value *Ptr = SI->getOperand(1);
2081 
2082         if (isa<UndefValue>(Ptr) ||
2083             (isa<ConstantPointerNull>(Ptr) &&
2084              !NullPointerIsDefined(SI->getFunction(),
2085                                    SI->getPointerAddressSpace()))) {
2086           changeToUnreachable(SI, true, false, DDT);
2087           Changed = true;
2088           break;
2089         }
2090       }
2091     }
2092 
2093     TerminatorInst *Terminator = BB->getTerminator();
2094     if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
2095       // Turn invokes that call 'nounwind' functions into ordinary calls.
2096       Value *Callee = II->getCalledValue();
2097       if ((isa<ConstantPointerNull>(Callee) &&
2098            !NullPointerIsDefined(BB->getParent())) ||
2099           isa<UndefValue>(Callee)) {
2100         changeToUnreachable(II, true, false, DDT);
2101         Changed = true;
2102       } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
2103         if (II->use_empty() && II->onlyReadsMemory()) {
2104           // jump to the normal destination branch.
2105           BasicBlock *NormalDestBB = II->getNormalDest();
2106           BasicBlock *UnwindDestBB = II->getUnwindDest();
2107           BranchInst::Create(NormalDestBB, II);
2108           UnwindDestBB->removePredecessor(II->getParent());
2109           II->eraseFromParent();
2110           if (DDT)
2111             DDT->deleteEdge(BB, UnwindDestBB);
2112         } else
2113           changeToCall(II, DDT);
2114         Changed = true;
2115       }
2116     } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
2117       // Remove catchpads which cannot be reached.
2118       struct CatchPadDenseMapInfo {
2119         static CatchPadInst *getEmptyKey() {
2120           return DenseMapInfo<CatchPadInst *>::getEmptyKey();
2121         }
2122 
2123         static CatchPadInst *getTombstoneKey() {
2124           return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
2125         }
2126 
2127         static unsigned getHashValue(CatchPadInst *CatchPad) {
2128           return static_cast<unsigned>(hash_combine_range(
2129               CatchPad->value_op_begin(), CatchPad->value_op_end()));
2130         }
2131 
2132         static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
2133           if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
2134               RHS == getEmptyKey() || RHS == getTombstoneKey())
2135             return LHS == RHS;
2136           return LHS->isIdenticalTo(RHS);
2137         }
2138       };
2139 
2140       // Set of unique CatchPads.
2141       SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
2142                     CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
2143           HandlerSet;
2144       detail::DenseSetEmpty Empty;
2145       for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
2146                                              E = CatchSwitch->handler_end();
2147            I != E; ++I) {
2148         BasicBlock *HandlerBB = *I;
2149         auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
2150         if (!HandlerSet.insert({CatchPad, Empty}).second) {
2151           CatchSwitch->removeHandler(I);
2152           --I;
2153           --E;
2154           Changed = true;
2155         }
2156       }
2157     }
2158 
2159     Changed |= ConstantFoldTerminator(BB, true, nullptr, DDT);
2160     for (BasicBlock *Successor : successors(BB))
2161       if (Reachable.insert(Successor).second)
2162         Worklist.push_back(Successor);
2163   } while (!Worklist.empty());
2164   return Changed;
2165 }
2166 
removeUnwindEdge(BasicBlock * BB,DeferredDominance * DDT)2167 void llvm::removeUnwindEdge(BasicBlock *BB, DeferredDominance *DDT) {
2168   TerminatorInst *TI = BB->getTerminator();
2169 
2170   if (auto *II = dyn_cast<InvokeInst>(TI)) {
2171     changeToCall(II, DDT);
2172     return;
2173   }
2174 
2175   TerminatorInst *NewTI;
2176   BasicBlock *UnwindDest;
2177 
2178   if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
2179     NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
2180     UnwindDest = CRI->getUnwindDest();
2181   } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
2182     auto *NewCatchSwitch = CatchSwitchInst::Create(
2183         CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
2184         CatchSwitch->getName(), CatchSwitch);
2185     for (BasicBlock *PadBB : CatchSwitch->handlers())
2186       NewCatchSwitch->addHandler(PadBB);
2187 
2188     NewTI = NewCatchSwitch;
2189     UnwindDest = CatchSwitch->getUnwindDest();
2190   } else {
2191     llvm_unreachable("Could not find unwind successor");
2192   }
2193 
2194   NewTI->takeName(TI);
2195   NewTI->setDebugLoc(TI->getDebugLoc());
2196   UnwindDest->removePredecessor(BB);
2197   TI->replaceAllUsesWith(NewTI);
2198   TI->eraseFromParent();
2199   if (DDT)
2200     DDT->deleteEdge(BB, UnwindDest);
2201 }
2202 
2203 /// removeUnreachableBlocks - Remove blocks that are not reachable, even
2204 /// if they are in a dead cycle.  Return true if a change was made, false
2205 /// otherwise. If `LVI` is passed, this function preserves LazyValueInfo
2206 /// after modifying the CFG.
removeUnreachableBlocks(Function & F,LazyValueInfo * LVI,DeferredDominance * DDT)2207 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI,
2208                                    DeferredDominance *DDT) {
2209   SmallPtrSet<BasicBlock*, 16> Reachable;
2210   bool Changed = markAliveBlocks(F, Reachable, DDT);
2211 
2212   // If there are unreachable blocks in the CFG...
2213   if (Reachable.size() == F.size())
2214     return Changed;
2215 
2216   assert(Reachable.size() < F.size());
2217   NumRemoved += F.size()-Reachable.size();
2218 
2219   // Loop over all of the basic blocks that are not reachable, dropping all of
2220   // their internal references. Update DDT and LVI if available.
2221   std::vector <DominatorTree::UpdateType> Updates;
2222   for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ++I) {
2223     auto *BB = &*I;
2224     if (Reachable.count(BB))
2225       continue;
2226     for (BasicBlock *Successor : successors(BB)) {
2227       if (Reachable.count(Successor))
2228         Successor->removePredecessor(BB);
2229       if (DDT)
2230         Updates.push_back({DominatorTree::Delete, BB, Successor});
2231     }
2232     if (LVI)
2233       LVI->eraseBlock(BB);
2234     BB->dropAllReferences();
2235   }
2236 
2237   for (Function::iterator I = ++F.begin(); I != F.end();) {
2238     auto *BB = &*I;
2239     if (Reachable.count(BB)) {
2240       ++I;
2241       continue;
2242     }
2243     if (DDT) {
2244       DDT->deleteBB(BB); // deferred deletion of BB.
2245       ++I;
2246     } else {
2247       I = F.getBasicBlockList().erase(I);
2248     }
2249   }
2250 
2251   if (DDT)
2252     DDT->applyUpdates(Updates);
2253   return true;
2254 }
2255 
combineMetadata(Instruction * K,const Instruction * J,ArrayRef<unsigned> KnownIDs)2256 void llvm::combineMetadata(Instruction *K, const Instruction *J,
2257                            ArrayRef<unsigned> KnownIDs) {
2258   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
2259   K->dropUnknownNonDebugMetadata(KnownIDs);
2260   K->getAllMetadataOtherThanDebugLoc(Metadata);
2261   for (const auto &MD : Metadata) {
2262     unsigned Kind = MD.first;
2263     MDNode *JMD = J->getMetadata(Kind);
2264     MDNode *KMD = MD.second;
2265 
2266     switch (Kind) {
2267       default:
2268         K->setMetadata(Kind, nullptr); // Remove unknown metadata
2269         break;
2270       case LLVMContext::MD_dbg:
2271         llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
2272       case LLVMContext::MD_tbaa:
2273         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2274         break;
2275       case LLVMContext::MD_alias_scope:
2276         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
2277         break;
2278       case LLVMContext::MD_noalias:
2279       case LLVMContext::MD_mem_parallel_loop_access:
2280         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
2281         break;
2282       case LLVMContext::MD_range:
2283         K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
2284         break;
2285       case LLVMContext::MD_fpmath:
2286         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2287         break;
2288       case LLVMContext::MD_invariant_load:
2289         // Only set the !invariant.load if it is present in both instructions.
2290         K->setMetadata(Kind, JMD);
2291         break;
2292       case LLVMContext::MD_nonnull:
2293         // Only set the !nonnull if it is present in both instructions.
2294         K->setMetadata(Kind, JMD);
2295         break;
2296       case LLVMContext::MD_invariant_group:
2297         // Preserve !invariant.group in K.
2298         break;
2299       case LLVMContext::MD_align:
2300         K->setMetadata(Kind,
2301           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
2302         break;
2303       case LLVMContext::MD_dereferenceable:
2304       case LLVMContext::MD_dereferenceable_or_null:
2305         K->setMetadata(Kind,
2306           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
2307         break;
2308     }
2309   }
2310   // Set !invariant.group from J if J has it. If both instructions have it
2311   // then we will just pick it from J - even when they are different.
2312   // Also make sure that K is load or store - f.e. combining bitcast with load
2313   // could produce bitcast with invariant.group metadata, which is invalid.
2314   // FIXME: we should try to preserve both invariant.group md if they are
2315   // different, but right now instruction can only have one invariant.group.
2316   if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
2317     if (isa<LoadInst>(K) || isa<StoreInst>(K))
2318       K->setMetadata(LLVMContext::MD_invariant_group, JMD);
2319 }
2320 
combineMetadataForCSE(Instruction * K,const Instruction * J)2321 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J) {
2322   unsigned KnownIDs[] = {
2323       LLVMContext::MD_tbaa,            LLVMContext::MD_alias_scope,
2324       LLVMContext::MD_noalias,         LLVMContext::MD_range,
2325       LLVMContext::MD_invariant_load,  LLVMContext::MD_nonnull,
2326       LLVMContext::MD_invariant_group, LLVMContext::MD_align,
2327       LLVMContext::MD_dereferenceable,
2328       LLVMContext::MD_dereferenceable_or_null};
2329   combineMetadata(K, J, KnownIDs);
2330 }
2331 
2332 template <typename RootType, typename DominatesFn>
replaceDominatedUsesWith(Value * From,Value * To,const RootType & Root,const DominatesFn & Dominates)2333 static unsigned replaceDominatedUsesWith(Value *From, Value *To,
2334                                          const RootType &Root,
2335                                          const DominatesFn &Dominates) {
2336   assert(From->getType() == To->getType());
2337 
2338   unsigned Count = 0;
2339   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2340        UI != UE;) {
2341     Use &U = *UI++;
2342     if (!Dominates(Root, U))
2343       continue;
2344     U.set(To);
2345     LLVM_DEBUG(dbgs() << "Replace dominated use of '" << From->getName()
2346                       << "' as " << *To << " in " << *U << "\n");
2347     ++Count;
2348   }
2349   return Count;
2350 }
2351 
replaceNonLocalUsesWith(Instruction * From,Value * To)2352 unsigned llvm::replaceNonLocalUsesWith(Instruction *From, Value *To) {
2353    assert(From->getType() == To->getType());
2354    auto *BB = From->getParent();
2355    unsigned Count = 0;
2356 
2357   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2358        UI != UE;) {
2359     Use &U = *UI++;
2360     auto *I = cast<Instruction>(U.getUser());
2361     if (I->getParent() == BB)
2362       continue;
2363     U.set(To);
2364     ++Count;
2365   }
2366   return Count;
2367 }
2368 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlockEdge & Root)2369 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
2370                                         DominatorTree &DT,
2371                                         const BasicBlockEdge &Root) {
2372   auto Dominates = [&DT](const BasicBlockEdge &Root, const Use &U) {
2373     return DT.dominates(Root, U);
2374   };
2375   return ::replaceDominatedUsesWith(From, To, Root, Dominates);
2376 }
2377 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlock * BB)2378 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
2379                                         DominatorTree &DT,
2380                                         const BasicBlock *BB) {
2381   auto ProperlyDominates = [&DT](const BasicBlock *BB, const Use &U) {
2382     auto *I = cast<Instruction>(U.getUser())->getParent();
2383     return DT.properlyDominates(BB, I);
2384   };
2385   return ::replaceDominatedUsesWith(From, To, BB, ProperlyDominates);
2386 }
2387 
callsGCLeafFunction(ImmutableCallSite CS,const TargetLibraryInfo & TLI)2388 bool llvm::callsGCLeafFunction(ImmutableCallSite CS,
2389                                const TargetLibraryInfo &TLI) {
2390   // Check if the function is specifically marked as a gc leaf function.
2391   if (CS.hasFnAttr("gc-leaf-function"))
2392     return true;
2393   if (const Function *F = CS.getCalledFunction()) {
2394     if (F->hasFnAttribute("gc-leaf-function"))
2395       return true;
2396 
2397     if (auto IID = F->getIntrinsicID())
2398       // Most LLVM intrinsics do not take safepoints.
2399       return IID != Intrinsic::experimental_gc_statepoint &&
2400              IID != Intrinsic::experimental_deoptimize;
2401   }
2402 
2403   // Lib calls can be materialized by some passes, and won't be
2404   // marked as 'gc-leaf-function.' All available Libcalls are
2405   // GC-leaf.
2406   LibFunc LF;
2407   if (TLI.getLibFunc(CS, LF)) {
2408     return TLI.has(LF);
2409   }
2410 
2411   return false;
2412 }
2413 
copyNonnullMetadata(const LoadInst & OldLI,MDNode * N,LoadInst & NewLI)2414 void llvm::copyNonnullMetadata(const LoadInst &OldLI, MDNode *N,
2415                                LoadInst &NewLI) {
2416   auto *NewTy = NewLI.getType();
2417 
2418   // This only directly applies if the new type is also a pointer.
2419   if (NewTy->isPointerTy()) {
2420     NewLI.setMetadata(LLVMContext::MD_nonnull, N);
2421     return;
2422   }
2423 
2424   // The only other translation we can do is to integral loads with !range
2425   // metadata.
2426   if (!NewTy->isIntegerTy())
2427     return;
2428 
2429   MDBuilder MDB(NewLI.getContext());
2430   const Value *Ptr = OldLI.getPointerOperand();
2431   auto *ITy = cast<IntegerType>(NewTy);
2432   auto *NullInt = ConstantExpr::getPtrToInt(
2433       ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
2434   auto *NonNullInt = ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
2435   NewLI.setMetadata(LLVMContext::MD_range,
2436                     MDB.createRange(NonNullInt, NullInt));
2437 }
2438 
copyRangeMetadata(const DataLayout & DL,const LoadInst & OldLI,MDNode * N,LoadInst & NewLI)2439 void llvm::copyRangeMetadata(const DataLayout &DL, const LoadInst &OldLI,
2440                              MDNode *N, LoadInst &NewLI) {
2441   auto *NewTy = NewLI.getType();
2442 
2443   // Give up unless it is converted to a pointer where there is a single very
2444   // valuable mapping we can do reliably.
2445   // FIXME: It would be nice to propagate this in more ways, but the type
2446   // conversions make it hard.
2447   if (!NewTy->isPointerTy())
2448     return;
2449 
2450   unsigned BitWidth = DL.getIndexTypeSizeInBits(NewTy);
2451   if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) {
2452     MDNode *NN = MDNode::get(OldLI.getContext(), None);
2453     NewLI.setMetadata(LLVMContext::MD_nonnull, NN);
2454   }
2455 }
2456 
2457 namespace {
2458 
2459 /// A potential constituent of a bitreverse or bswap expression. See
2460 /// collectBitParts for a fuller explanation.
2461 struct BitPart {
BitPart__anond699c33e0a11::BitPart2462   BitPart(Value *P, unsigned BW) : Provider(P) {
2463     Provenance.resize(BW);
2464   }
2465 
2466   /// The Value that this is a bitreverse/bswap of.
2467   Value *Provider;
2468 
2469   /// The "provenance" of each bit. Provenance[A] = B means that bit A
2470   /// in Provider becomes bit B in the result of this expression.
2471   SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128.
2472 
2473   enum { Unset = -1 };
2474 };
2475 
2476 } // end anonymous namespace
2477 
2478 /// Analyze the specified subexpression and see if it is capable of providing
2479 /// pieces of a bswap or bitreverse. The subexpression provides a potential
2480 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in
2481 /// the output of the expression came from a corresponding bit in some other
2482 /// value. This function is recursive, and the end result is a mapping of
2483 /// bitnumber to bitnumber. It is the caller's responsibility to validate that
2484 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse.
2485 ///
2486 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know
2487 /// that the expression deposits the low byte of %X into the high byte of the
2488 /// result and that all other bits are zero. This expression is accepted and a
2489 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to
2490 /// [0-7].
2491 ///
2492 /// To avoid revisiting values, the BitPart results are memoized into the
2493 /// provided map. To avoid unnecessary copying of BitParts, BitParts are
2494 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to
2495 /// store BitParts objects, not pointers. As we need the concept of a nullptr
2496 /// BitParts (Value has been analyzed and the analysis failed), we an Optional
2497 /// type instead to provide the same functionality.
2498 ///
2499 /// Because we pass around references into \c BPS, we must use a container that
2500 /// does not invalidate internal references (std::map instead of DenseMap).
2501 static const Optional<BitPart> &
collectBitParts(Value * V,bool MatchBSwaps,bool MatchBitReversals,std::map<Value *,Optional<BitPart>> & BPS)2502 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
2503                 std::map<Value *, Optional<BitPart>> &BPS) {
2504   auto I = BPS.find(V);
2505   if (I != BPS.end())
2506     return I->second;
2507 
2508   auto &Result = BPS[V] = None;
2509   auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2510 
2511   if (Instruction *I = dyn_cast<Instruction>(V)) {
2512     // If this is an or instruction, it may be an inner node of the bswap.
2513     if (I->getOpcode() == Instruction::Or) {
2514       auto &A = collectBitParts(I->getOperand(0), MatchBSwaps,
2515                                 MatchBitReversals, BPS);
2516       auto &B = collectBitParts(I->getOperand(1), MatchBSwaps,
2517                                 MatchBitReversals, BPS);
2518       if (!A || !B)
2519         return Result;
2520 
2521       // Try and merge the two together.
2522       if (!A->Provider || A->Provider != B->Provider)
2523         return Result;
2524 
2525       Result = BitPart(A->Provider, BitWidth);
2526       for (unsigned i = 0; i < A->Provenance.size(); ++i) {
2527         if (A->Provenance[i] != BitPart::Unset &&
2528             B->Provenance[i] != BitPart::Unset &&
2529             A->Provenance[i] != B->Provenance[i])
2530           return Result = None;
2531 
2532         if (A->Provenance[i] == BitPart::Unset)
2533           Result->Provenance[i] = B->Provenance[i];
2534         else
2535           Result->Provenance[i] = A->Provenance[i];
2536       }
2537 
2538       return Result;
2539     }
2540 
2541     // If this is a logical shift by a constant, recurse then shift the result.
2542     if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
2543       unsigned BitShift =
2544           cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
2545       // Ensure the shift amount is defined.
2546       if (BitShift > BitWidth)
2547         return Result;
2548 
2549       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2550                                   MatchBitReversals, BPS);
2551       if (!Res)
2552         return Result;
2553       Result = Res;
2554 
2555       // Perform the "shift" on BitProvenance.
2556       auto &P = Result->Provenance;
2557       if (I->getOpcode() == Instruction::Shl) {
2558         P.erase(std::prev(P.end(), BitShift), P.end());
2559         P.insert(P.begin(), BitShift, BitPart::Unset);
2560       } else {
2561         P.erase(P.begin(), std::next(P.begin(), BitShift));
2562         P.insert(P.end(), BitShift, BitPart::Unset);
2563       }
2564 
2565       return Result;
2566     }
2567 
2568     // If this is a logical 'and' with a mask that clears bits, recurse then
2569     // unset the appropriate bits.
2570     if (I->getOpcode() == Instruction::And &&
2571         isa<ConstantInt>(I->getOperand(1))) {
2572       APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1);
2573       const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
2574 
2575       // Check that the mask allows a multiple of 8 bits for a bswap, for an
2576       // early exit.
2577       unsigned NumMaskedBits = AndMask.countPopulation();
2578       if (!MatchBitReversals && NumMaskedBits % 8 != 0)
2579         return Result;
2580 
2581       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2582                                   MatchBitReversals, BPS);
2583       if (!Res)
2584         return Result;
2585       Result = Res;
2586 
2587       for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1)
2588         // If the AndMask is zero for this bit, clear the bit.
2589         if ((AndMask & Bit) == 0)
2590           Result->Provenance[i] = BitPart::Unset;
2591       return Result;
2592     }
2593 
2594     // If this is a zext instruction zero extend the result.
2595     if (I->getOpcode() == Instruction::ZExt) {
2596       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2597                                   MatchBitReversals, BPS);
2598       if (!Res)
2599         return Result;
2600 
2601       Result = BitPart(Res->Provider, BitWidth);
2602       auto NarrowBitWidth =
2603           cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth();
2604       for (unsigned i = 0; i < NarrowBitWidth; ++i)
2605         Result->Provenance[i] = Res->Provenance[i];
2606       for (unsigned i = NarrowBitWidth; i < BitWidth; ++i)
2607         Result->Provenance[i] = BitPart::Unset;
2608       return Result;
2609     }
2610   }
2611 
2612   // Okay, we got to something that isn't a shift, 'or' or 'and'.  This must be
2613   // the input value to the bswap/bitreverse.
2614   Result = BitPart(V, BitWidth);
2615   for (unsigned i = 0; i < BitWidth; ++i)
2616     Result->Provenance[i] = i;
2617   return Result;
2618 }
2619 
bitTransformIsCorrectForBSwap(unsigned From,unsigned To,unsigned BitWidth)2620 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To,
2621                                           unsigned BitWidth) {
2622   if (From % 8 != To % 8)
2623     return false;
2624   // Convert from bit indices to byte indices and check for a byte reversal.
2625   From >>= 3;
2626   To >>= 3;
2627   BitWidth >>= 3;
2628   return From == BitWidth - To - 1;
2629 }
2630 
bitTransformIsCorrectForBitReverse(unsigned From,unsigned To,unsigned BitWidth)2631 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To,
2632                                                unsigned BitWidth) {
2633   return From == BitWidth - To - 1;
2634 }
2635 
recognizeBSwapOrBitReverseIdiom(Instruction * I,bool MatchBSwaps,bool MatchBitReversals,SmallVectorImpl<Instruction * > & InsertedInsts)2636 bool llvm::recognizeBSwapOrBitReverseIdiom(
2637     Instruction *I, bool MatchBSwaps, bool MatchBitReversals,
2638     SmallVectorImpl<Instruction *> &InsertedInsts) {
2639   if (Operator::getOpcode(I) != Instruction::Or)
2640     return false;
2641   if (!MatchBSwaps && !MatchBitReversals)
2642     return false;
2643   IntegerType *ITy = dyn_cast<IntegerType>(I->getType());
2644   if (!ITy || ITy->getBitWidth() > 128)
2645     return false;   // Can't do vectors or integers > 128 bits.
2646   unsigned BW = ITy->getBitWidth();
2647 
2648   unsigned DemandedBW = BW;
2649   IntegerType *DemandedTy = ITy;
2650   if (I->hasOneUse()) {
2651     if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) {
2652       DemandedTy = cast<IntegerType>(Trunc->getType());
2653       DemandedBW = DemandedTy->getBitWidth();
2654     }
2655   }
2656 
2657   // Try to find all the pieces corresponding to the bswap.
2658   std::map<Value *, Optional<BitPart>> BPS;
2659   auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS);
2660   if (!Res)
2661     return false;
2662   auto &BitProvenance = Res->Provenance;
2663 
2664   // Now, is the bit permutation correct for a bswap or a bitreverse? We can
2665   // only byteswap values with an even number of bytes.
2666   bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true;
2667   for (unsigned i = 0; i < DemandedBW; ++i) {
2668     OKForBSwap &=
2669         bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW);
2670     OKForBitReverse &=
2671         bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW);
2672   }
2673 
2674   Intrinsic::ID Intrin;
2675   if (OKForBSwap && MatchBSwaps)
2676     Intrin = Intrinsic::bswap;
2677   else if (OKForBitReverse && MatchBitReversals)
2678     Intrin = Intrinsic::bitreverse;
2679   else
2680     return false;
2681 
2682   if (ITy != DemandedTy) {
2683     Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy);
2684     Value *Provider = Res->Provider;
2685     IntegerType *ProviderTy = cast<IntegerType>(Provider->getType());
2686     // We may need to truncate the provider.
2687     if (DemandedTy != ProviderTy) {
2688       auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy,
2689                                      "trunc", I);
2690       InsertedInsts.push_back(Trunc);
2691       Provider = Trunc;
2692     }
2693     auto *CI = CallInst::Create(F, Provider, "rev", I);
2694     InsertedInsts.push_back(CI);
2695     auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I);
2696     InsertedInsts.push_back(ExtInst);
2697     return true;
2698   }
2699 
2700   Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy);
2701   InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I));
2702   return true;
2703 }
2704 
2705 // CodeGen has special handling for some string functions that may replace
2706 // them with target-specific intrinsics.  Since that'd skip our interceptors
2707 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses,
2708 // we mark affected calls as NoBuiltin, which will disable optimization
2709 // in CodeGen.
maybeMarkSanitizerLibraryCallNoBuiltin(CallInst * CI,const TargetLibraryInfo * TLI)2710 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin(
2711     CallInst *CI, const TargetLibraryInfo *TLI) {
2712   Function *F = CI->getCalledFunction();
2713   LibFunc Func;
2714   if (F && !F->hasLocalLinkage() && F->hasName() &&
2715       TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) &&
2716       !F->doesNotAccessMemory())
2717     CI->addAttribute(AttributeList::FunctionIndex, Attribute::NoBuiltin);
2718 }
2719 
canReplaceOperandWithVariable(const Instruction * I,unsigned OpIdx)2720 bool llvm::canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx) {
2721   // We can't have a PHI with a metadata type.
2722   if (I->getOperand(OpIdx)->getType()->isMetadataTy())
2723     return false;
2724 
2725   // Early exit.
2726   if (!isa<Constant>(I->getOperand(OpIdx)))
2727     return true;
2728 
2729   switch (I->getOpcode()) {
2730   default:
2731     return true;
2732   case Instruction::Call:
2733   case Instruction::Invoke:
2734     // Can't handle inline asm. Skip it.
2735     if (isa<InlineAsm>(ImmutableCallSite(I).getCalledValue()))
2736       return false;
2737     // Many arithmetic intrinsics have no issue taking a
2738     // variable, however it's hard to distingish these from
2739     // specials such as @llvm.frameaddress that require a constant.
2740     if (isa<IntrinsicInst>(I))
2741       return false;
2742 
2743     // Constant bundle operands may need to retain their constant-ness for
2744     // correctness.
2745     if (ImmutableCallSite(I).isBundleOperand(OpIdx))
2746       return false;
2747     return true;
2748   case Instruction::ShuffleVector:
2749     // Shufflevector masks are constant.
2750     return OpIdx != 2;
2751   case Instruction::Switch:
2752   case Instruction::ExtractValue:
2753     // All operands apart from the first are constant.
2754     return OpIdx == 0;
2755   case Instruction::InsertValue:
2756     // All operands apart from the first and the second are constant.
2757     return OpIdx < 2;
2758   case Instruction::Alloca:
2759     // Static allocas (constant size in the entry block) are handled by
2760     // prologue/epilogue insertion so they're free anyway. We definitely don't
2761     // want to make them non-constant.
2762     return !cast<AllocaInst>(I)->isStaticAlloca();
2763   case Instruction::GetElementPtr:
2764     if (OpIdx == 0)
2765       return true;
2766     gep_type_iterator It = gep_type_begin(I);
2767     for (auto E = std::next(It, OpIdx); It != E; ++It)
2768       if (It.isStruct())
2769         return false;
2770     return true;
2771   }
2772 }
2773