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.
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 ///
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 
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 
376   if (!I->mayHaveSideEffects())
377     return true;
378 
379   // Special case intrinsics that "may have side effects" but can be deleted
380   // when dead.
381   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
382     // Safe to delete llvm.stacksave if dead.
383     if (II->getIntrinsicID() == Intrinsic::stacksave)
384       return true;
385 
386     // Lifetime intrinsics are dead when their right-hand is undef.
387     if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
388         II->getIntrinsicID() == Intrinsic::lifetime_end)
389       return isa<UndefValue>(II->getArgOperand(1));
390 
391     // Assumptions are dead if their condition is trivially true.  Guards on
392     // true are operationally no-ops.  In the future we can consider more
393     // sophisticated tradeoffs for guards considering potential for check
394     // widening, but for now we keep things simple.
395     if (II->getIntrinsicID() == Intrinsic::assume ||
396         II->getIntrinsicID() == Intrinsic::experimental_guard) {
397       if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
398         return !Cond->isZero();
399 
400       return false;
401     }
402   }
403 
404   if (isAllocLikeFn(I, TLI))
405     return true;
406 
407   if (CallInst *CI = isFreeCall(I, TLI))
408     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
409       return C->isNullValue() || isa<UndefValue>(C);
410 
411   if (CallSite CS = CallSite(I))
412     if (isMathLibCallNoop(CS, TLI))
413       return true;
414 
415   return false;
416 }
417 
418 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
419 /// trivially dead instruction, delete it.  If that makes any of its operands
420 /// trivially dead, delete them too, recursively.  Return true if any
421 /// instructions were deleted.
422 bool
423 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
424                                                  const TargetLibraryInfo *TLI) {
425   Instruction *I = dyn_cast<Instruction>(V);
426   if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
427     return false;
428 
429   SmallVector<Instruction*, 16> DeadInsts;
430   DeadInsts.push_back(I);
431 
432   do {
433     I = DeadInsts.pop_back_val();
434 
435     // Null out all of the instruction's operands to see if any operand becomes
436     // dead as we go.
437     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
438       Value *OpV = I->getOperand(i);
439       I->setOperand(i, nullptr);
440 
441       if (!OpV->use_empty()) continue;
442 
443       // If the operand is an instruction that became dead as we nulled out the
444       // operand, and if it is 'trivially' dead, delete it in a future loop
445       // iteration.
446       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
447         if (isInstructionTriviallyDead(OpI, TLI))
448           DeadInsts.push_back(OpI);
449     }
450 
451     I->eraseFromParent();
452   } while (!DeadInsts.empty());
453 
454   return true;
455 }
456 
457 /// areAllUsesEqual - Check whether the uses of a value are all the same.
458 /// This is similar to Instruction::hasOneUse() except this will also return
459 /// true when there are no uses or multiple uses that all refer to the same
460 /// value.
461 static bool areAllUsesEqual(Instruction *I) {
462   Value::user_iterator UI = I->user_begin();
463   Value::user_iterator UE = I->user_end();
464   if (UI == UE)
465     return true;
466 
467   User *TheUse = *UI;
468   for (++UI; UI != UE; ++UI) {
469     if (*UI != TheUse)
470       return false;
471   }
472   return true;
473 }
474 
475 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
476 /// dead PHI node, due to being a def-use chain of single-use nodes that
477 /// either forms a cycle or is terminated by a trivially dead instruction,
478 /// delete it.  If that makes any of its operands trivially dead, delete them
479 /// too, recursively.  Return true if a change was made.
480 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
481                                         const TargetLibraryInfo *TLI) {
482   SmallPtrSet<Instruction*, 4> Visited;
483   for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
484        I = cast<Instruction>(*I->user_begin())) {
485     if (I->use_empty())
486       return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
487 
488     // If we find an instruction more than once, we're on a cycle that
489     // won't prove fruitful.
490     if (!Visited.insert(I).second) {
491       // Break the cycle and delete the instruction and its operands.
492       I->replaceAllUsesWith(UndefValue::get(I->getType()));
493       (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
494       return true;
495     }
496   }
497   return false;
498 }
499 
500 static bool
501 simplifyAndDCEInstruction(Instruction *I,
502                           SmallSetVector<Instruction *, 16> &WorkList,
503                           const DataLayout &DL,
504                           const TargetLibraryInfo *TLI) {
505   if (isInstructionTriviallyDead(I, TLI)) {
506     // Null out all of the instruction's operands to see if any operand becomes
507     // dead as we go.
508     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
509       Value *OpV = I->getOperand(i);
510       I->setOperand(i, nullptr);
511 
512       if (!OpV->use_empty() || I == OpV)
513         continue;
514 
515       // If the operand is an instruction that became dead as we nulled out the
516       // operand, and if it is 'trivially' dead, delete it in a future loop
517       // iteration.
518       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
519         if (isInstructionTriviallyDead(OpI, TLI))
520           WorkList.insert(OpI);
521     }
522 
523     I->eraseFromParent();
524 
525     return true;
526   }
527 
528   if (Value *SimpleV = SimplifyInstruction(I, DL)) {
529     // Add the users to the worklist. CAREFUL: an instruction can use itself,
530     // in the case of a phi node.
531     for (User *U : I->users()) {
532       if (U != I) {
533         WorkList.insert(cast<Instruction>(U));
534       }
535     }
536 
537     // Replace the instruction with its simplified value.
538     bool Changed = false;
539     if (!I->use_empty()) {
540       I->replaceAllUsesWith(SimpleV);
541       Changed = true;
542     }
543     if (isInstructionTriviallyDead(I, TLI)) {
544       I->eraseFromParent();
545       Changed = true;
546     }
547     return Changed;
548   }
549   return false;
550 }
551 
552 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
553 /// simplify any instructions in it and recursively delete dead instructions.
554 ///
555 /// This returns true if it changed the code, note that it can delete
556 /// instructions in other blocks as well in this block.
557 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
558                                        const TargetLibraryInfo *TLI) {
559   bool MadeChange = false;
560   const DataLayout &DL = BB->getModule()->getDataLayout();
561 
562 #ifndef NDEBUG
563   // In debug builds, ensure that the terminator of the block is never replaced
564   // or deleted by these simplifications. The idea of simplification is that it
565   // cannot introduce new instructions, and there is no way to replace the
566   // terminator of a block without introducing a new instruction.
567   AssertingVH<Instruction> TerminatorVH(&BB->back());
568 #endif
569 
570   SmallSetVector<Instruction *, 16> WorkList;
571   // Iterate over the original function, only adding insts to the worklist
572   // if they actually need to be revisited. This avoids having to pre-init
573   // the worklist with the entire function's worth of instructions.
574   for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end());
575        BI != E;) {
576     assert(!BI->isTerminator());
577     Instruction *I = &*BI;
578     ++BI;
579 
580     // We're visiting this instruction now, so make sure it's not in the
581     // worklist from an earlier visit.
582     if (!WorkList.count(I))
583       MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
584   }
585 
586   while (!WorkList.empty()) {
587     Instruction *I = WorkList.pop_back_val();
588     MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
589   }
590   return MadeChange;
591 }
592 
593 //===----------------------------------------------------------------------===//
594 //  Control Flow Graph Restructuring.
595 //
596 
597 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
598 /// method is called when we're about to delete Pred as a predecessor of BB.  If
599 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
600 ///
601 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
602 /// nodes that collapse into identity values.  For example, if we have:
603 ///   x = phi(1, 0, 0, 0)
604 ///   y = and x, z
605 ///
606 /// .. and delete the predecessor corresponding to the '1', this will attempt to
607 /// recursively fold the and to 0.
608 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred,
609                                         DeferredDominance *DDT) {
610   // This only adjusts blocks with PHI nodes.
611   if (!isa<PHINode>(BB->begin()))
612     return;
613 
614   // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
615   // them down.  This will leave us with single entry phi nodes and other phis
616   // that can be removed.
617   BB->removePredecessor(Pred, true);
618 
619   WeakTrackingVH PhiIt = &BB->front();
620   while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
621     PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
622     Value *OldPhiIt = PhiIt;
623 
624     if (!recursivelySimplifyInstruction(PN))
625       continue;
626 
627     // If recursive simplification ended up deleting the next PHI node we would
628     // iterate to, then our iterator is invalid, restart scanning from the top
629     // of the block.
630     if (PhiIt != OldPhiIt) PhiIt = &BB->front();
631   }
632   if (DDT)
633     DDT->deleteEdge(Pred, BB);
634 }
635 
636 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
637 /// predecessor is known to have one successor (DestBB!).  Eliminate the edge
638 /// between them, moving the instructions in the predecessor into DestBB and
639 /// deleting the predecessor block.
640 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT,
641                                        DeferredDominance *DDT) {
642   assert(!(DT && DDT) && "Cannot call with both DT and DDT.");
643 
644   // If BB has single-entry PHI nodes, fold them.
645   while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
646     Value *NewVal = PN->getIncomingValue(0);
647     // Replace self referencing PHI with undef, it must be dead.
648     if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
649     PN->replaceAllUsesWith(NewVal);
650     PN->eraseFromParent();
651   }
652 
653   BasicBlock *PredBB = DestBB->getSinglePredecessor();
654   assert(PredBB && "Block doesn't have a single predecessor!");
655 
656   bool ReplaceEntryBB = false;
657   if (PredBB == &DestBB->getParent()->getEntryBlock())
658     ReplaceEntryBB = true;
659 
660   // Deferred DT update: Collect all the edges that enter PredBB. These
661   // dominator edges will be redirected to DestBB.
662   std::vector <DominatorTree::UpdateType> Updates;
663   if (DDT && !ReplaceEntryBB) {
664     Updates.reserve(1 +
665                     (2 * std::distance(pred_begin(PredBB), pred_end(PredBB))));
666     Updates.push_back({DominatorTree::Delete, PredBB, DestBB});
667     for (auto I = pred_begin(PredBB), E = pred_end(PredBB); I != E; ++I) {
668       Updates.push_back({DominatorTree::Delete, *I, PredBB});
669       // This predecessor of PredBB may already have DestBB as a successor.
670       if (llvm::find(successors(*I), DestBB) == succ_end(*I))
671         Updates.push_back({DominatorTree::Insert, *I, DestBB});
672     }
673   }
674 
675   // Zap anything that took the address of DestBB.  Not doing this will give the
676   // address an invalid value.
677   if (DestBB->hasAddressTaken()) {
678     BlockAddress *BA = BlockAddress::get(DestBB);
679     Constant *Replacement =
680       ConstantInt::get(Type::getInt32Ty(BA->getContext()), 1);
681     BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
682                                                      BA->getType()));
683     BA->destroyConstant();
684   }
685 
686   // Anything that branched to PredBB now branches to DestBB.
687   PredBB->replaceAllUsesWith(DestBB);
688 
689   // Splice all the instructions from PredBB to DestBB.
690   PredBB->getTerminator()->eraseFromParent();
691   DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
692 
693   // If the PredBB is the entry block of the function, move DestBB up to
694   // become the entry block after we erase PredBB.
695   if (ReplaceEntryBB)
696     DestBB->moveAfter(PredBB);
697 
698   if (DT) {
699     // For some irreducible CFG we end up having forward-unreachable blocks
700     // so check if getNode returns a valid node before updating the domtree.
701     if (DomTreeNode *DTN = DT->getNode(PredBB)) {
702       BasicBlock *PredBBIDom = DTN->getIDom()->getBlock();
703       DT->changeImmediateDominator(DestBB, PredBBIDom);
704       DT->eraseNode(PredBB);
705     }
706   }
707 
708   if (DDT) {
709     DDT->deleteBB(PredBB); // Deferred deletion of BB.
710     if (ReplaceEntryBB)
711       // The entry block was removed and there is no external interface for the
712       // dominator tree to be notified of this change. In this corner-case we
713       // recalculate the entire tree.
714       DDT->recalculate(*(DestBB->getParent()));
715     else
716       DDT->applyUpdates(Updates);
717   } else {
718     PredBB->eraseFromParent(); // Nuke BB.
719   }
720 }
721 
722 /// CanMergeValues - Return true if we can choose one of these values to use
723 /// in place of the other. Note that we will always choose the non-undef
724 /// value to keep.
725 static bool CanMergeValues(Value *First, Value *Second) {
726   return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
727 }
728 
729 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
730 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
731 ///
732 /// Assumption: Succ is the single successor for BB.
733 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
734   assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
735 
736   DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
737         << Succ->getName() << "\n");
738   // Shortcut, if there is only a single predecessor it must be BB and merging
739   // is always safe
740   if (Succ->getSinglePredecessor()) return true;
741 
742   // Make a list of the predecessors of BB
743   SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
744 
745   // Look at all the phi nodes in Succ, to see if they present a conflict when
746   // merging these blocks
747   for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
748     PHINode *PN = cast<PHINode>(I);
749 
750     // If the incoming value from BB is again a PHINode in
751     // BB which has the same incoming value for *PI as PN does, we can
752     // merge the phi nodes and then the blocks can still be merged
753     PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
754     if (BBPN && BBPN->getParent() == BB) {
755       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
756         BasicBlock *IBB = PN->getIncomingBlock(PI);
757         if (BBPreds.count(IBB) &&
758             !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
759                             PN->getIncomingValue(PI))) {
760           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
761                 << Succ->getName() << " is conflicting with "
762                 << BBPN->getName() << " with regard to common predecessor "
763                 << IBB->getName() << "\n");
764           return false;
765         }
766       }
767     } else {
768       Value* Val = PN->getIncomingValueForBlock(BB);
769       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
770         // See if the incoming value for the common predecessor is equal to the
771         // one for BB, in which case this phi node will not prevent the merging
772         // of the block.
773         BasicBlock *IBB = PN->getIncomingBlock(PI);
774         if (BBPreds.count(IBB) &&
775             !CanMergeValues(Val, PN->getIncomingValue(PI))) {
776           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
777                 << Succ->getName() << " is conflicting with regard to common "
778                 << "predecessor " << IBB->getName() << "\n");
779           return false;
780         }
781       }
782     }
783   }
784 
785   return true;
786 }
787 
788 using PredBlockVector = SmallVector<BasicBlock *, 16>;
789 using IncomingValueMap = DenseMap<BasicBlock *, Value *>;
790 
791 /// \brief Determines the value to use as the phi node input for a block.
792 ///
793 /// Select between \p OldVal any value that we know flows from \p BB
794 /// to a particular phi on the basis of which one (if either) is not
795 /// undef. Update IncomingValues based on the selected value.
796 ///
797 /// \param OldVal The value we are considering selecting.
798 /// \param BB The block that the value flows in from.
799 /// \param IncomingValues A map from block-to-value for other phi inputs
800 /// that we have examined.
801 ///
802 /// \returns the selected value.
803 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
804                                           IncomingValueMap &IncomingValues) {
805   if (!isa<UndefValue>(OldVal)) {
806     assert((!IncomingValues.count(BB) ||
807             IncomingValues.find(BB)->second == OldVal) &&
808            "Expected OldVal to match incoming value from BB!");
809 
810     IncomingValues.insert(std::make_pair(BB, OldVal));
811     return OldVal;
812   }
813 
814   IncomingValueMap::const_iterator It = IncomingValues.find(BB);
815   if (It != IncomingValues.end()) return It->second;
816 
817   return OldVal;
818 }
819 
820 /// \brief Create a map from block to value for the operands of a
821 /// given phi.
822 ///
823 /// Create a map from block to value for each non-undef value flowing
824 /// into \p PN.
825 ///
826 /// \param PN The phi we are collecting the map for.
827 /// \param IncomingValues [out] The map from block to value for this phi.
828 static void gatherIncomingValuesToPhi(PHINode *PN,
829                                       IncomingValueMap &IncomingValues) {
830   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
831     BasicBlock *BB = PN->getIncomingBlock(i);
832     Value *V = PN->getIncomingValue(i);
833 
834     if (!isa<UndefValue>(V))
835       IncomingValues.insert(std::make_pair(BB, V));
836   }
837 }
838 
839 /// \brief Replace the incoming undef values to a phi with the values
840 /// from a block-to-value map.
841 ///
842 /// \param PN The phi we are replacing the undefs in.
843 /// \param IncomingValues A map from block to value.
844 static void replaceUndefValuesInPhi(PHINode *PN,
845                                     const IncomingValueMap &IncomingValues) {
846   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
847     Value *V = PN->getIncomingValue(i);
848 
849     if (!isa<UndefValue>(V)) continue;
850 
851     BasicBlock *BB = PN->getIncomingBlock(i);
852     IncomingValueMap::const_iterator It = IncomingValues.find(BB);
853     if (It == IncomingValues.end()) continue;
854 
855     PN->setIncomingValue(i, It->second);
856   }
857 }
858 
859 /// \brief Replace a value flowing from a block to a phi with
860 /// potentially multiple instances of that value flowing from the
861 /// block's predecessors to the phi.
862 ///
863 /// \param BB The block with the value flowing into the phi.
864 /// \param BBPreds The predecessors of BB.
865 /// \param PN The phi that we are updating.
866 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
867                                                 const PredBlockVector &BBPreds,
868                                                 PHINode *PN) {
869   Value *OldVal = PN->removeIncomingValue(BB, false);
870   assert(OldVal && "No entry in PHI for Pred BB!");
871 
872   IncomingValueMap IncomingValues;
873 
874   // We are merging two blocks - BB, and the block containing PN - and
875   // as a result we need to redirect edges from the predecessors of BB
876   // to go to the block containing PN, and update PN
877   // accordingly. Since we allow merging blocks in the case where the
878   // predecessor and successor blocks both share some predecessors,
879   // and where some of those common predecessors might have undef
880   // values flowing into PN, we want to rewrite those values to be
881   // consistent with the non-undef values.
882 
883   gatherIncomingValuesToPhi(PN, IncomingValues);
884 
885   // If this incoming value is one of the PHI nodes in BB, the new entries
886   // in the PHI node are the entries from the old PHI.
887   if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
888     PHINode *OldValPN = cast<PHINode>(OldVal);
889     for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
890       // Note that, since we are merging phi nodes and BB and Succ might
891       // have common predecessors, we could end up with a phi node with
892       // identical incoming branches. This will be cleaned up later (and
893       // will trigger asserts if we try to clean it up now, without also
894       // simplifying the corresponding conditional branch).
895       BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
896       Value *PredVal = OldValPN->getIncomingValue(i);
897       Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
898                                                     IncomingValues);
899 
900       // And add a new incoming value for this predecessor for the
901       // newly retargeted branch.
902       PN->addIncoming(Selected, PredBB);
903     }
904   } else {
905     for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
906       // Update existing incoming values in PN for this
907       // predecessor of BB.
908       BasicBlock *PredBB = BBPreds[i];
909       Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
910                                                     IncomingValues);
911 
912       // And add a new incoming value for this predecessor for the
913       // newly retargeted branch.
914       PN->addIncoming(Selected, PredBB);
915     }
916   }
917 
918   replaceUndefValuesInPhi(PN, IncomingValues);
919 }
920 
921 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
922 /// unconditional branch, and contains no instructions other than PHI nodes,
923 /// potential side-effect free intrinsics and the branch.  If possible,
924 /// eliminate BB by rewriting all the predecessors to branch to the successor
925 /// block and return true.  If we can't transform, return false.
926 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB,
927                                                    DeferredDominance *DDT) {
928   assert(BB != &BB->getParent()->getEntryBlock() &&
929          "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
930 
931   // We can't eliminate infinite loops.
932   BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
933   if (BB == Succ) return false;
934 
935   // Check to see if merging these blocks would cause conflicts for any of the
936   // phi nodes in BB or Succ. If not, we can safely merge.
937   if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
938 
939   // Check for cases where Succ has multiple predecessors and a PHI node in BB
940   // has uses which will not disappear when the PHI nodes are merged.  It is
941   // possible to handle such cases, but difficult: it requires checking whether
942   // BB dominates Succ, which is non-trivial to calculate in the case where
943   // Succ has multiple predecessors.  Also, it requires checking whether
944   // constructing the necessary self-referential PHI node doesn't introduce any
945   // conflicts; this isn't too difficult, but the previous code for doing this
946   // was incorrect.
947   //
948   // Note that if this check finds a live use, BB dominates Succ, so BB is
949   // something like a loop pre-header (or rarely, a part of an irreducible CFG);
950   // folding the branch isn't profitable in that case anyway.
951   if (!Succ->getSinglePredecessor()) {
952     BasicBlock::iterator BBI = BB->begin();
953     while (isa<PHINode>(*BBI)) {
954       for (Use &U : BBI->uses()) {
955         if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
956           if (PN->getIncomingBlock(U) != BB)
957             return false;
958         } else {
959           return false;
960         }
961       }
962       ++BBI;
963     }
964   }
965 
966   DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
967 
968   std::vector<DominatorTree::UpdateType> Updates;
969   if (DDT) {
970     Updates.reserve(1 + (2 * std::distance(pred_begin(BB), pred_end(BB))));
971     Updates.push_back({DominatorTree::Delete, BB, Succ});
972     // All predecessors of BB will be moved to Succ.
973     for (auto I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
974       Updates.push_back({DominatorTree::Delete, *I, BB});
975       // This predecessor of BB may already have Succ as a successor.
976       if (llvm::find(successors(*I), Succ) == succ_end(*I))
977         Updates.push_back({DominatorTree::Insert, *I, Succ});
978     }
979   }
980 
981   if (isa<PHINode>(Succ->begin())) {
982     // If there is more than one pred of succ, and there are PHI nodes in
983     // the successor, then we need to add incoming edges for the PHI nodes
984     //
985     const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
986 
987     // Loop over all of the PHI nodes in the successor of BB.
988     for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
989       PHINode *PN = cast<PHINode>(I);
990 
991       redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
992     }
993   }
994 
995   if (Succ->getSinglePredecessor()) {
996     // BB is the only predecessor of Succ, so Succ will end up with exactly
997     // the same predecessors BB had.
998 
999     // Copy over any phi, debug or lifetime instruction.
1000     BB->getTerminator()->eraseFromParent();
1001     Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
1002                                BB->getInstList());
1003   } else {
1004     while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
1005       // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
1006       assert(PN->use_empty() && "There shouldn't be any uses here!");
1007       PN->eraseFromParent();
1008     }
1009   }
1010 
1011   // If the unconditional branch we replaced contains llvm.loop metadata, we
1012   // add the metadata to the branch instructions in the predecessors.
1013   unsigned LoopMDKind = BB->getContext().getMDKindID("llvm.loop");
1014   Instruction *TI = BB->getTerminator();
1015   if (TI)
1016     if (MDNode *LoopMD = TI->getMetadata(LoopMDKind))
1017       for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1018         BasicBlock *Pred = *PI;
1019         Pred->getTerminator()->setMetadata(LoopMDKind, LoopMD);
1020       }
1021 
1022   // Everything that jumped to BB now goes to Succ.
1023   BB->replaceAllUsesWith(Succ);
1024   if (!Succ->hasName()) Succ->takeName(BB);
1025 
1026   if (DDT) {
1027     DDT->deleteBB(BB); // Deferred deletion of the old basic block.
1028     DDT->applyUpdates(Updates);
1029   } else {
1030     BB->eraseFromParent(); // Delete the old basic block.
1031   }
1032   return true;
1033 }
1034 
1035 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
1036 /// nodes in this block. This doesn't try to be clever about PHI nodes
1037 /// which differ only in the order of the incoming values, but instcombine
1038 /// orders them so it usually won't matter.
1039 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
1040   // This implementation doesn't currently consider undef operands
1041   // specially. Theoretically, two phis which are identical except for
1042   // one having an undef where the other doesn't could be collapsed.
1043 
1044   struct PHIDenseMapInfo {
1045     static PHINode *getEmptyKey() {
1046       return DenseMapInfo<PHINode *>::getEmptyKey();
1047     }
1048 
1049     static PHINode *getTombstoneKey() {
1050       return DenseMapInfo<PHINode *>::getTombstoneKey();
1051     }
1052 
1053     static unsigned getHashValue(PHINode *PN) {
1054       // Compute a hash value on the operands. Instcombine will likely have
1055       // sorted them, which helps expose duplicates, but we have to check all
1056       // the operands to be safe in case instcombine hasn't run.
1057       return static_cast<unsigned>(hash_combine(
1058           hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
1059           hash_combine_range(PN->block_begin(), PN->block_end())));
1060     }
1061 
1062     static bool isEqual(PHINode *LHS, PHINode *RHS) {
1063       if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1064           RHS == getEmptyKey() || RHS == getTombstoneKey())
1065         return LHS == RHS;
1066       return LHS->isIdenticalTo(RHS);
1067     }
1068   };
1069 
1070   // Set of unique PHINodes.
1071   DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
1072 
1073   // Examine each PHI.
1074   bool Changed = false;
1075   for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
1076     auto Inserted = PHISet.insert(PN);
1077     if (!Inserted.second) {
1078       // A duplicate. Replace this PHI with its duplicate.
1079       PN->replaceAllUsesWith(*Inserted.first);
1080       PN->eraseFromParent();
1081       Changed = true;
1082 
1083       // The RAUW can change PHIs that we already visited. Start over from the
1084       // beginning.
1085       PHISet.clear();
1086       I = BB->begin();
1087     }
1088   }
1089 
1090   return Changed;
1091 }
1092 
1093 /// enforceKnownAlignment - If the specified pointer points to an object that
1094 /// we control, modify the object's alignment to PrefAlign. This isn't
1095 /// often possible though. If alignment is important, a more reliable approach
1096 /// is to simply align all global variables and allocation instructions to
1097 /// their preferred alignment from the beginning.
1098 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
1099                                       unsigned PrefAlign,
1100                                       const DataLayout &DL) {
1101   assert(PrefAlign > Align);
1102 
1103   V = V->stripPointerCasts();
1104 
1105   if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1106     // TODO: ideally, computeKnownBits ought to have used
1107     // AllocaInst::getAlignment() in its computation already, making
1108     // the below max redundant. But, as it turns out,
1109     // stripPointerCasts recurses through infinite layers of bitcasts,
1110     // while computeKnownBits is not allowed to traverse more than 6
1111     // levels.
1112     Align = std::max(AI->getAlignment(), Align);
1113     if (PrefAlign <= Align)
1114       return Align;
1115 
1116     // If the preferred alignment is greater than the natural stack alignment
1117     // then don't round up. This avoids dynamic stack realignment.
1118     if (DL.exceedsNaturalStackAlignment(PrefAlign))
1119       return Align;
1120     AI->setAlignment(PrefAlign);
1121     return PrefAlign;
1122   }
1123 
1124   if (auto *GO = dyn_cast<GlobalObject>(V)) {
1125     // TODO: as above, this shouldn't be necessary.
1126     Align = std::max(GO->getAlignment(), Align);
1127     if (PrefAlign <= Align)
1128       return Align;
1129 
1130     // If there is a large requested alignment and we can, bump up the alignment
1131     // of the global.  If the memory we set aside for the global may not be the
1132     // memory used by the final program then it is impossible for us to reliably
1133     // enforce the preferred alignment.
1134     if (!GO->canIncreaseAlignment())
1135       return Align;
1136 
1137     GO->setAlignment(PrefAlign);
1138     return PrefAlign;
1139   }
1140 
1141   return Align;
1142 }
1143 
1144 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
1145                                           const DataLayout &DL,
1146                                           const Instruction *CxtI,
1147                                           AssumptionCache *AC,
1148                                           const DominatorTree *DT) {
1149   assert(V->getType()->isPointerTy() &&
1150          "getOrEnforceKnownAlignment expects a pointer!");
1151 
1152   KnownBits Known = computeKnownBits(V, DL, 0, AC, CxtI, DT);
1153   unsigned TrailZ = Known.countMinTrailingZeros();
1154 
1155   // Avoid trouble with ridiculously large TrailZ values, such as
1156   // those computed from a null pointer.
1157   TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1158 
1159   unsigned Align = 1u << std::min(Known.getBitWidth() - 1, TrailZ);
1160 
1161   // LLVM doesn't support alignments larger than this currently.
1162   Align = std::min(Align, +Value::MaximumAlignment);
1163 
1164   if (PrefAlign > Align)
1165     Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1166 
1167   // We don't need to make any adjustment.
1168   return Align;
1169 }
1170 
1171 ///===---------------------------------------------------------------------===//
1172 ///  Dbg Intrinsic utilities
1173 ///
1174 
1175 /// See if there is a dbg.value intrinsic for DIVar before I.
1176 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr,
1177                               Instruction *I) {
1178   // Since we can't guarantee that the original dbg.declare instrinsic
1179   // is removed by LowerDbgDeclare(), we need to make sure that we are
1180   // not inserting the same dbg.value intrinsic over and over.
1181   BasicBlock::InstListType::iterator PrevI(I);
1182   if (PrevI != I->getParent()->getInstList().begin()) {
1183     --PrevI;
1184     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1185       if (DVI->getValue() == I->getOperand(0) &&
1186           DVI->getVariable() == DIVar &&
1187           DVI->getExpression() == DIExpr)
1188         return true;
1189   }
1190   return false;
1191 }
1192 
1193 /// See if there is a dbg.value intrinsic for DIVar for the PHI node.
1194 static bool PhiHasDebugValue(DILocalVariable *DIVar,
1195                              DIExpression *DIExpr,
1196                              PHINode *APN) {
1197   // Since we can't guarantee that the original dbg.declare instrinsic
1198   // is removed by LowerDbgDeclare(), we need to make sure that we are
1199   // not inserting the same dbg.value intrinsic over and over.
1200   SmallVector<DbgValueInst *, 1> DbgValues;
1201   findDbgValues(DbgValues, APN);
1202   for (auto *DVI : DbgValues) {
1203     assert(DVI->getValue() == APN);
1204     if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr))
1205       return true;
1206   }
1207   return false;
1208 }
1209 
1210 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1211 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic.
1212 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1213                                            StoreInst *SI, DIBuilder &Builder) {
1214   assert(DII->isAddressOfVariable());
1215   auto *DIVar = DII->getVariable();
1216   assert(DIVar && "Missing variable");
1217   auto *DIExpr = DII->getExpression();
1218   Value *DV = SI->getOperand(0);
1219 
1220   // If an argument is zero extended then use argument directly. The ZExt
1221   // may be zapped by an optimization pass in future.
1222   Argument *ExtendedArg = nullptr;
1223   if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1224     ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1225   if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1226     ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1227   if (ExtendedArg) {
1228     // If this DII was already describing only a fragment of a variable, ensure
1229     // that fragment is appropriately narrowed here.
1230     // But if a fragment wasn't used, describe the value as the original
1231     // argument (rather than the zext or sext) so that it remains described even
1232     // if the sext/zext is optimized away. This widens the variable description,
1233     // leaving it up to the consumer to know how the smaller value may be
1234     // represented in a larger register.
1235     if (auto Fragment = DIExpr->getFragmentInfo()) {
1236       unsigned FragmentOffset = Fragment->OffsetInBits;
1237       SmallVector<uint64_t, 3> Ops(DIExpr->elements_begin(),
1238                                    DIExpr->elements_end() - 3);
1239       Ops.push_back(dwarf::DW_OP_LLVM_fragment);
1240       Ops.push_back(FragmentOffset);
1241       const DataLayout &DL = DII->getModule()->getDataLayout();
1242       Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType()));
1243       DIExpr = Builder.createExpression(Ops);
1244     }
1245     DV = ExtendedArg;
1246   }
1247   if (!LdStHasDebugValue(DIVar, DIExpr, SI))
1248     Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, DII->getDebugLoc(),
1249                                     SI);
1250 }
1251 
1252 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1253 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic.
1254 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1255                                            LoadInst *LI, DIBuilder &Builder) {
1256   auto *DIVar = DII->getVariable();
1257   auto *DIExpr = DII->getExpression();
1258   assert(DIVar && "Missing variable");
1259 
1260   if (LdStHasDebugValue(DIVar, DIExpr, LI))
1261     return;
1262 
1263   // We are now tracking the loaded value instead of the address. In the
1264   // future if multi-location support is added to the IR, it might be
1265   // preferable to keep tracking both the loaded value and the original
1266   // address in case the alloca can not be elided.
1267   Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1268       LI, DIVar, DIExpr, DII->getDebugLoc(), (Instruction *)nullptr);
1269   DbgValue->insertAfter(LI);
1270 }
1271 
1272 /// Inserts a llvm.dbg.value intrinsic after a phi that has an associated
1273 /// llvm.dbg.declare or llvm.dbg.addr intrinsic.
1274 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII,
1275                                            PHINode *APN, DIBuilder &Builder) {
1276   auto *DIVar = DII->getVariable();
1277   auto *DIExpr = DII->getExpression();
1278   assert(DIVar && "Missing variable");
1279 
1280   if (PhiHasDebugValue(DIVar, DIExpr, APN))
1281     return;
1282 
1283   BasicBlock *BB = APN->getParent();
1284   auto InsertionPt = BB->getFirstInsertionPt();
1285 
1286   // The block may be a catchswitch block, which does not have a valid
1287   // insertion point.
1288   // FIXME: Insert dbg.value markers in the successors when appropriate.
1289   if (InsertionPt != BB->end())
1290     Builder.insertDbgValueIntrinsic(APN, DIVar, DIExpr, DII->getDebugLoc(),
1291                                     &*InsertionPt);
1292 }
1293 
1294 /// Determine whether this alloca is either a VLA or an array.
1295 static bool isArray(AllocaInst *AI) {
1296   return AI->isArrayAllocation() ||
1297     AI->getType()->getElementType()->isArrayTy();
1298 }
1299 
1300 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1301 /// of llvm.dbg.value intrinsics.
1302 bool llvm::LowerDbgDeclare(Function &F) {
1303   DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1304   SmallVector<DbgDeclareInst *, 4> Dbgs;
1305   for (auto &FI : F)
1306     for (Instruction &BI : FI)
1307       if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1308         Dbgs.push_back(DDI);
1309 
1310   if (Dbgs.empty())
1311     return false;
1312 
1313   for (auto &I : Dbgs) {
1314     DbgDeclareInst *DDI = I;
1315     AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1316     // If this is an alloca for a scalar variable, insert a dbg.value
1317     // at each load and store to the alloca and erase the dbg.declare.
1318     // The dbg.values allow tracking a variable even if it is not
1319     // stored on the stack, while the dbg.declare can only describe
1320     // the stack slot (and at a lexical-scope granularity). Later
1321     // passes will attempt to elide the stack slot.
1322     if (AI && !isArray(AI)) {
1323       for (auto &AIUse : AI->uses()) {
1324         User *U = AIUse.getUser();
1325         if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1326           if (AIUse.getOperandNo() == 1)
1327             ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1328         } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1329           ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1330         } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1331           // This is a call by-value or some other instruction that
1332           // takes a pointer to the variable. Insert a *value*
1333           // intrinsic that describes the alloca.
1334           DIB.insertDbgValueIntrinsic(AI, DDI->getVariable(),
1335                                       DDI->getExpression(), DDI->getDebugLoc(),
1336                                       CI);
1337         }
1338       }
1339       DDI->eraseFromParent();
1340     }
1341   }
1342   return true;
1343 }
1344 
1345 /// Propagate dbg.value intrinsics through the newly inserted PHIs.
1346 void llvm::insertDebugValuesForPHIs(BasicBlock *BB,
1347                                     SmallVectorImpl<PHINode *> &InsertedPHIs) {
1348   assert(BB && "No BasicBlock to clone dbg.value(s) from.");
1349   if (InsertedPHIs.size() == 0)
1350     return;
1351 
1352   // Map existing PHI nodes to their dbg.values.
1353   ValueToValueMapTy DbgValueMap;
1354   for (auto &I : *BB) {
1355     if (auto DbgII = dyn_cast<DbgInfoIntrinsic>(&I)) {
1356       if (auto *Loc = dyn_cast_or_null<PHINode>(DbgII->getVariableLocation()))
1357         DbgValueMap.insert({Loc, DbgII});
1358     }
1359   }
1360   if (DbgValueMap.size() == 0)
1361     return;
1362 
1363   // Then iterate through the new PHIs and look to see if they use one of the
1364   // previously mapped PHIs. If so, insert a new dbg.value intrinsic that will
1365   // propagate the info through the new PHI.
1366   LLVMContext &C = BB->getContext();
1367   for (auto PHI : InsertedPHIs) {
1368     for (auto VI : PHI->operand_values()) {
1369       auto V = DbgValueMap.find(VI);
1370       if (V != DbgValueMap.end()) {
1371         auto *DbgII = cast<DbgInfoIntrinsic>(V->second);
1372         Instruction *NewDbgII = DbgII->clone();
1373         auto PhiMAV = MetadataAsValue::get(C, ValueAsMetadata::get(PHI));
1374         NewDbgII->setOperand(0, PhiMAV);
1375         BasicBlock *Parent = PHI->getParent();
1376         auto InsertionPt = Parent->getFirstInsertionPt();
1377         assert(InsertionPt != Parent->end() && "Ill-formed basic block");
1378         NewDbgII->insertBefore(&*InsertionPt);
1379       }
1380     }
1381   }
1382 }
1383 
1384 /// Finds all intrinsics declaring local variables as living in the memory that
1385 /// 'V' points to. This may include a mix of dbg.declare and
1386 /// dbg.addr intrinsics.
1387 TinyPtrVector<DbgInfoIntrinsic *> llvm::FindDbgAddrUses(Value *V) {
1388   auto *L = LocalAsMetadata::getIfExists(V);
1389   if (!L)
1390     return {};
1391   auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L);
1392   if (!MDV)
1393     return {};
1394 
1395   TinyPtrVector<DbgInfoIntrinsic *> Declares;
1396   for (User *U : MDV->users()) {
1397     if (auto *DII = dyn_cast<DbgInfoIntrinsic>(U))
1398       if (DII->isAddressOfVariable())
1399         Declares.push_back(DII);
1400   }
1401 
1402   return Declares;
1403 }
1404 
1405 void llvm::findDbgValues(SmallVectorImpl<DbgValueInst *> &DbgValues, Value *V) {
1406   if (auto *L = LocalAsMetadata::getIfExists(V))
1407     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1408       for (User *U : MDV->users())
1409         if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1410           DbgValues.push_back(DVI);
1411 }
1412 
1413 void llvm::findDbgUsers(SmallVectorImpl<DbgInfoIntrinsic *> &DbgUsers,
1414                         Value *V) {
1415   if (auto *L = LocalAsMetadata::getIfExists(V))
1416     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1417       for (User *U : MDV->users())
1418         if (DbgInfoIntrinsic *DII = dyn_cast<DbgInfoIntrinsic>(U))
1419           DbgUsers.push_back(DII);
1420 }
1421 
1422 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1423                              Instruction *InsertBefore, DIBuilder &Builder,
1424                              bool DerefBefore, int Offset, bool DerefAfter) {
1425   auto DbgAddrs = FindDbgAddrUses(Address);
1426   for (DbgInfoIntrinsic *DII : DbgAddrs) {
1427     DebugLoc Loc = DII->getDebugLoc();
1428     auto *DIVar = DII->getVariable();
1429     auto *DIExpr = DII->getExpression();
1430     assert(DIVar && "Missing variable");
1431     DIExpr = DIExpression::prepend(DIExpr, DerefBefore, Offset, DerefAfter);
1432     // Insert llvm.dbg.declare immediately after InsertBefore, and remove old
1433     // llvm.dbg.declare.
1434     Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1435     if (DII == InsertBefore)
1436       InsertBefore = &*std::next(InsertBefore->getIterator());
1437     DII->eraseFromParent();
1438   }
1439   return !DbgAddrs.empty();
1440 }
1441 
1442 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1443                                       DIBuilder &Builder, bool DerefBefore,
1444                                       int Offset, bool DerefAfter) {
1445   return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1446                            DerefBefore, Offset, DerefAfter);
1447 }
1448 
1449 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress,
1450                                         DIBuilder &Builder, int Offset) {
1451   DebugLoc Loc = DVI->getDebugLoc();
1452   auto *DIVar = DVI->getVariable();
1453   auto *DIExpr = DVI->getExpression();
1454   assert(DIVar && "Missing variable");
1455 
1456   // This is an alloca-based llvm.dbg.value. The first thing it should do with
1457   // the alloca pointer is dereference it. Otherwise we don't know how to handle
1458   // it and give up.
1459   if (!DIExpr || DIExpr->getNumElements() < 1 ||
1460       DIExpr->getElement(0) != dwarf::DW_OP_deref)
1461     return;
1462 
1463   // Insert the offset immediately after the first deref.
1464   // We could just change the offset argument of dbg.value, but it's unsigned...
1465   if (Offset) {
1466     SmallVector<uint64_t, 4> Ops;
1467     Ops.push_back(dwarf::DW_OP_deref);
1468     DIExpression::appendOffset(Ops, Offset);
1469     Ops.append(DIExpr->elements_begin() + 1, DIExpr->elements_end());
1470     DIExpr = Builder.createExpression(Ops);
1471   }
1472 
1473   Builder.insertDbgValueIntrinsic(NewAddress, DIVar, DIExpr, Loc, DVI);
1474   DVI->eraseFromParent();
1475 }
1476 
1477 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1478                                     DIBuilder &Builder, int Offset) {
1479   if (auto *L = LocalAsMetadata::getIfExists(AI))
1480     if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
1481       for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) {
1482         Use &U = *UI++;
1483         if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser()))
1484           replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset);
1485       }
1486 }
1487 
1488 void llvm::salvageDebugInfo(Instruction &I) {
1489   SmallVector<DbgInfoIntrinsic *, 1> DbgUsers;
1490   findDbgUsers(DbgUsers, &I);
1491   if (DbgUsers.empty())
1492     return;
1493 
1494   auto &M = *I.getModule();
1495 
1496   auto wrapMD = [&](Value *V) {
1497     return MetadataAsValue::get(I.getContext(), ValueAsMetadata::get(V));
1498   };
1499 
1500   auto doSalvage = [&](DbgInfoIntrinsic *DII, SmallVectorImpl<uint64_t> &Ops) {
1501     auto *DIExpr = DII->getExpression();
1502     DIExpr = DIExpression::doPrepend(DIExpr, Ops,
1503                                      DIExpression::WithStackValue);
1504     DII->setOperand(0, wrapMD(I.getOperand(0)));
1505     DII->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr));
1506     DEBUG(dbgs() << "SALVAGE: " << *DII << '\n');
1507   };
1508 
1509   auto applyOffset = [&](DbgInfoIntrinsic *DII, uint64_t Offset) {
1510     SmallVector<uint64_t, 8> Ops;
1511     DIExpression::appendOffset(Ops, Offset);
1512     doSalvage(DII, Ops);
1513   };
1514 
1515   auto applyOps = [&](DbgInfoIntrinsic *DII,
1516                       std::initializer_list<uint64_t> Opcodes) {
1517     SmallVector<uint64_t, 8> Ops(Opcodes);
1518     doSalvage(DII, Ops);
1519   };
1520 
1521   if (auto *CI = dyn_cast<CastInst>(&I)) {
1522     if (!CI->isNoopCast(M.getDataLayout()))
1523       return;
1524 
1525     // No-op casts are irrelevant for debug info.
1526     MetadataAsValue *CastSrc = wrapMD(I.getOperand(0));
1527     for (auto *DII : DbgUsers) {
1528       DII->setOperand(0, CastSrc);
1529       DEBUG(dbgs() << "SALVAGE: " << *DII << '\n');
1530     }
1531   } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1532     unsigned BitWidth =
1533         M.getDataLayout().getIndexSizeInBits(GEP->getPointerAddressSpace());
1534     // Rewrite a constant GEP into a DIExpression.  Since we are performing
1535     // arithmetic to compute the variable's *value* in the DIExpression, we
1536     // need to mark the expression with a DW_OP_stack_value.
1537     APInt Offset(BitWidth, 0);
1538     if (GEP->accumulateConstantOffset(M.getDataLayout(), Offset))
1539       for (auto *DII : DbgUsers)
1540         applyOffset(DII, Offset.getSExtValue());
1541   } else if (auto *BI = dyn_cast<BinaryOperator>(&I)) {
1542     auto *ConstInt = dyn_cast<ConstantInt>(I.getOperand(1));
1543     if (!ConstInt || ConstInt->getBitWidth() > 64)
1544       return;
1545 
1546     uint64_t Val = ConstInt->getSExtValue();
1547     for (auto *DII : DbgUsers) {
1548       switch (BI->getOpcode()) {
1549       case Instruction::Add:
1550         applyOffset(DII, Val);
1551         break;
1552       case Instruction::Sub:
1553         applyOffset(DII, -int64_t(Val));
1554         break;
1555       case Instruction::Mul:
1556         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mul});
1557         break;
1558       case Instruction::SDiv:
1559         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_div});
1560         break;
1561       case Instruction::SRem:
1562         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mod});
1563         break;
1564       case Instruction::Or:
1565         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_or});
1566         break;
1567       case Instruction::And:
1568         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_and});
1569         break;
1570       case Instruction::Xor:
1571         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_xor});
1572         break;
1573       case Instruction::Shl:
1574         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shl});
1575         break;
1576       case Instruction::LShr:
1577         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shr});
1578         break;
1579       case Instruction::AShr:
1580         applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shra});
1581         break;
1582       default:
1583         // TODO: Salvage constants from each kind of binop we know about.
1584         continue;
1585       }
1586     }
1587   } else if (isa<LoadInst>(&I)) {
1588     MetadataAsValue *AddrMD = wrapMD(I.getOperand(0));
1589     for (auto *DII : DbgUsers) {
1590       // Rewrite the load into DW_OP_deref.
1591       auto *DIExpr = DII->getExpression();
1592       DIExpr = DIExpression::prepend(DIExpr, DIExpression::WithDeref);
1593       DII->setOperand(0, AddrMD);
1594       DII->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr));
1595       DEBUG(dbgs() << "SALVAGE:  " << *DII << '\n');
1596     }
1597   }
1598 }
1599 
1600 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) {
1601   unsigned NumDeadInst = 0;
1602   // Delete the instructions backwards, as it has a reduced likelihood of
1603   // having to update as many def-use and use-def chains.
1604   Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1605   while (EndInst != &BB->front()) {
1606     // Delete the next to last instruction.
1607     Instruction *Inst = &*--EndInst->getIterator();
1608     if (!Inst->use_empty() && !Inst->getType()->isTokenTy())
1609       Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1610     if (Inst->isEHPad() || Inst->getType()->isTokenTy()) {
1611       EndInst = Inst;
1612       continue;
1613     }
1614     if (!isa<DbgInfoIntrinsic>(Inst))
1615       ++NumDeadInst;
1616     Inst->eraseFromParent();
1617   }
1618   return NumDeadInst;
1619 }
1620 
1621 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap,
1622                                    bool PreserveLCSSA, DeferredDominance *DDT) {
1623   BasicBlock *BB = I->getParent();
1624   std::vector <DominatorTree::UpdateType> Updates;
1625 
1626   // Loop over all of the successors, removing BB's entry from any PHI
1627   // nodes.
1628   if (DDT)
1629     Updates.reserve(BB->getTerminator()->getNumSuccessors());
1630   for (BasicBlock *Successor : successors(BB)) {
1631     Successor->removePredecessor(BB, PreserveLCSSA);
1632     if (DDT)
1633       Updates.push_back({DominatorTree::Delete, BB, Successor});
1634   }
1635   // Insert a call to llvm.trap right before this.  This turns the undefined
1636   // behavior into a hard fail instead of falling through into random code.
1637   if (UseLLVMTrap) {
1638     Function *TrapFn =
1639       Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1640     CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1641     CallTrap->setDebugLoc(I->getDebugLoc());
1642   }
1643   new UnreachableInst(I->getContext(), I);
1644 
1645   // All instructions after this are dead.
1646   unsigned NumInstrsRemoved = 0;
1647   BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1648   while (BBI != BBE) {
1649     if (!BBI->use_empty())
1650       BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1651     BB->getInstList().erase(BBI++);
1652     ++NumInstrsRemoved;
1653   }
1654   if (DDT)
1655     DDT->applyUpdates(Updates);
1656   return NumInstrsRemoved;
1657 }
1658 
1659 /// changeToCall - Convert the specified invoke into a normal call.
1660 static void changeToCall(InvokeInst *II, DeferredDominance *DDT = nullptr) {
1661   SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1662   SmallVector<OperandBundleDef, 1> OpBundles;
1663   II->getOperandBundlesAsDefs(OpBundles);
1664   CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1665                                        "", II);
1666   NewCall->takeName(II);
1667   NewCall->setCallingConv(II->getCallingConv());
1668   NewCall->setAttributes(II->getAttributes());
1669   NewCall->setDebugLoc(II->getDebugLoc());
1670   II->replaceAllUsesWith(NewCall);
1671 
1672   // Follow the call by a branch to the normal destination.
1673   BasicBlock *NormalDestBB = II->getNormalDest();
1674   BranchInst::Create(NormalDestBB, II);
1675 
1676   // Update PHI nodes in the unwind destination
1677   BasicBlock *BB = II->getParent();
1678   BasicBlock *UnwindDestBB = II->getUnwindDest();
1679   UnwindDestBB->removePredecessor(BB);
1680   II->eraseFromParent();
1681   if (DDT)
1682     DDT->deleteEdge(BB, UnwindDestBB);
1683 }
1684 
1685 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI,
1686                                                    BasicBlock *UnwindEdge) {
1687   BasicBlock *BB = CI->getParent();
1688 
1689   // Convert this function call into an invoke instruction.  First, split the
1690   // basic block.
1691   BasicBlock *Split =
1692       BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc");
1693 
1694   // Delete the unconditional branch inserted by splitBasicBlock
1695   BB->getInstList().pop_back();
1696 
1697   // Create the new invoke instruction.
1698   SmallVector<Value *, 8> InvokeArgs(CI->arg_begin(), CI->arg_end());
1699   SmallVector<OperandBundleDef, 1> OpBundles;
1700 
1701   CI->getOperandBundlesAsDefs(OpBundles);
1702 
1703   // Note: we're round tripping operand bundles through memory here, and that
1704   // can potentially be avoided with a cleverer API design that we do not have
1705   // as of this time.
1706 
1707   InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge,
1708                                       InvokeArgs, OpBundles, CI->getName(), BB);
1709   II->setDebugLoc(CI->getDebugLoc());
1710   II->setCallingConv(CI->getCallingConv());
1711   II->setAttributes(CI->getAttributes());
1712 
1713   // Make sure that anything using the call now uses the invoke!  This also
1714   // updates the CallGraph if present, because it uses a WeakTrackingVH.
1715   CI->replaceAllUsesWith(II);
1716 
1717   // Delete the original call
1718   Split->getInstList().pop_front();
1719   return Split;
1720 }
1721 
1722 static bool markAliveBlocks(Function &F,
1723                             SmallPtrSetImpl<BasicBlock*> &Reachable,
1724                             DeferredDominance *DDT = nullptr) {
1725   SmallVector<BasicBlock*, 128> Worklist;
1726   BasicBlock *BB = &F.front();
1727   Worklist.push_back(BB);
1728   Reachable.insert(BB);
1729   bool Changed = false;
1730   do {
1731     BB = Worklist.pop_back_val();
1732 
1733     // Do a quick scan of the basic block, turning any obviously unreachable
1734     // instructions into LLVM unreachable insts.  The instruction combining pass
1735     // canonicalizes unreachable insts into stores to null or undef.
1736     for (Instruction &I : *BB) {
1737       // Assumptions that are known to be false are equivalent to unreachable.
1738       // Also, if the condition is undefined, then we make the choice most
1739       // beneficial to the optimizer, and choose that to also be unreachable.
1740       if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1741         if (II->getIntrinsicID() == Intrinsic::assume) {
1742           if (match(II->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) {
1743             // Don't insert a call to llvm.trap right before the unreachable.
1744             changeToUnreachable(II, false, false, DDT);
1745             Changed = true;
1746             break;
1747           }
1748         }
1749 
1750         if (II->getIntrinsicID() == Intrinsic::experimental_guard) {
1751           // A call to the guard intrinsic bails out of the current compilation
1752           // unit if the predicate passed to it is false.  If the predicate is a
1753           // constant false, then we know the guard will bail out of the current
1754           // compile unconditionally, so all code following it is dead.
1755           //
1756           // Note: unlike in llvm.assume, it is not "obviously profitable" for
1757           // guards to treat `undef` as `false` since a guard on `undef` can
1758           // still be useful for widening.
1759           if (match(II->getArgOperand(0), m_Zero()))
1760             if (!isa<UnreachableInst>(II->getNextNode())) {
1761               changeToUnreachable(II->getNextNode(), /*UseLLVMTrap=*/false,
1762                                   false, DDT);
1763               Changed = true;
1764               break;
1765             }
1766         }
1767       }
1768 
1769       if (auto *CI = dyn_cast<CallInst>(&I)) {
1770         Value *Callee = CI->getCalledValue();
1771         if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1772           changeToUnreachable(CI, /*UseLLVMTrap=*/false, false, DDT);
1773           Changed = true;
1774           break;
1775         }
1776         if (CI->doesNotReturn()) {
1777           // If we found a call to a no-return function, insert an unreachable
1778           // instruction after it.  Make sure there isn't *already* one there
1779           // though.
1780           if (!isa<UnreachableInst>(CI->getNextNode())) {
1781             // Don't insert a call to llvm.trap right before the unreachable.
1782             changeToUnreachable(CI->getNextNode(), false, false, DDT);
1783             Changed = true;
1784           }
1785           break;
1786         }
1787       }
1788 
1789       // Store to undef and store to null are undefined and used to signal that
1790       // they should be changed to unreachable by passes that can't modify the
1791       // CFG.
1792       if (auto *SI = dyn_cast<StoreInst>(&I)) {
1793         // Don't touch volatile stores.
1794         if (SI->isVolatile()) continue;
1795 
1796         Value *Ptr = SI->getOperand(1);
1797 
1798         if (isa<UndefValue>(Ptr) ||
1799             (isa<ConstantPointerNull>(Ptr) &&
1800              SI->getPointerAddressSpace() == 0)) {
1801           changeToUnreachable(SI, true, false, DDT);
1802           Changed = true;
1803           break;
1804         }
1805       }
1806     }
1807 
1808     TerminatorInst *Terminator = BB->getTerminator();
1809     if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
1810       // Turn invokes that call 'nounwind' functions into ordinary calls.
1811       Value *Callee = II->getCalledValue();
1812       if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1813         changeToUnreachable(II, true, false, DDT);
1814         Changed = true;
1815       } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
1816         if (II->use_empty() && II->onlyReadsMemory()) {
1817           // jump to the normal destination branch.
1818           BasicBlock *NormalDestBB = II->getNormalDest();
1819           BasicBlock *UnwindDestBB = II->getUnwindDest();
1820           BranchInst::Create(NormalDestBB, II);
1821           UnwindDestBB->removePredecessor(II->getParent());
1822           II->eraseFromParent();
1823           if (DDT)
1824             DDT->deleteEdge(BB, UnwindDestBB);
1825         } else
1826           changeToCall(II, DDT);
1827         Changed = true;
1828       }
1829     } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
1830       // Remove catchpads which cannot be reached.
1831       struct CatchPadDenseMapInfo {
1832         static CatchPadInst *getEmptyKey() {
1833           return DenseMapInfo<CatchPadInst *>::getEmptyKey();
1834         }
1835 
1836         static CatchPadInst *getTombstoneKey() {
1837           return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
1838         }
1839 
1840         static unsigned getHashValue(CatchPadInst *CatchPad) {
1841           return static_cast<unsigned>(hash_combine_range(
1842               CatchPad->value_op_begin(), CatchPad->value_op_end()));
1843         }
1844 
1845         static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
1846           if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1847               RHS == getEmptyKey() || RHS == getTombstoneKey())
1848             return LHS == RHS;
1849           return LHS->isIdenticalTo(RHS);
1850         }
1851       };
1852 
1853       // Set of unique CatchPads.
1854       SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
1855                     CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
1856           HandlerSet;
1857       detail::DenseSetEmpty Empty;
1858       for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
1859                                              E = CatchSwitch->handler_end();
1860            I != E; ++I) {
1861         BasicBlock *HandlerBB = *I;
1862         auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
1863         if (!HandlerSet.insert({CatchPad, Empty}).second) {
1864           CatchSwitch->removeHandler(I);
1865           --I;
1866           --E;
1867           Changed = true;
1868         }
1869       }
1870     }
1871 
1872     Changed |= ConstantFoldTerminator(BB, true, nullptr, DDT);
1873     for (BasicBlock *Successor : successors(BB))
1874       if (Reachable.insert(Successor).second)
1875         Worklist.push_back(Successor);
1876   } while (!Worklist.empty());
1877   return Changed;
1878 }
1879 
1880 void llvm::removeUnwindEdge(BasicBlock *BB, DeferredDominance *DDT) {
1881   TerminatorInst *TI = BB->getTerminator();
1882 
1883   if (auto *II = dyn_cast<InvokeInst>(TI)) {
1884     changeToCall(II, DDT);
1885     return;
1886   }
1887 
1888   TerminatorInst *NewTI;
1889   BasicBlock *UnwindDest;
1890 
1891   if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
1892     NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
1893     UnwindDest = CRI->getUnwindDest();
1894   } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
1895     auto *NewCatchSwitch = CatchSwitchInst::Create(
1896         CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
1897         CatchSwitch->getName(), CatchSwitch);
1898     for (BasicBlock *PadBB : CatchSwitch->handlers())
1899       NewCatchSwitch->addHandler(PadBB);
1900 
1901     NewTI = NewCatchSwitch;
1902     UnwindDest = CatchSwitch->getUnwindDest();
1903   } else {
1904     llvm_unreachable("Could not find unwind successor");
1905   }
1906 
1907   NewTI->takeName(TI);
1908   NewTI->setDebugLoc(TI->getDebugLoc());
1909   UnwindDest->removePredecessor(BB);
1910   TI->replaceAllUsesWith(NewTI);
1911   TI->eraseFromParent();
1912   if (DDT)
1913     DDT->deleteEdge(BB, UnwindDest);
1914 }
1915 
1916 /// removeUnreachableBlocks - Remove blocks that are not reachable, even
1917 /// if they are in a dead cycle.  Return true if a change was made, false
1918 /// otherwise. If `LVI` is passed, this function preserves LazyValueInfo
1919 /// after modifying the CFG.
1920 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI,
1921                                    DeferredDominance *DDT) {
1922   SmallPtrSet<BasicBlock*, 16> Reachable;
1923   bool Changed = markAliveBlocks(F, Reachable, DDT);
1924 
1925   // If there are unreachable blocks in the CFG...
1926   if (Reachable.size() == F.size())
1927     return Changed;
1928 
1929   assert(Reachable.size() < F.size());
1930   NumRemoved += F.size()-Reachable.size();
1931 
1932   // Loop over all of the basic blocks that are not reachable, dropping all of
1933   // their internal references. Update DDT and LVI if available.
1934   std::vector <DominatorTree::UpdateType> Updates;
1935   for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ++I) {
1936     auto *BB = &*I;
1937     if (Reachable.count(BB))
1938       continue;
1939     for (BasicBlock *Successor : successors(BB)) {
1940       if (Reachable.count(Successor))
1941         Successor->removePredecessor(BB);
1942       if (DDT)
1943         Updates.push_back({DominatorTree::Delete, BB, Successor});
1944     }
1945     if (LVI)
1946       LVI->eraseBlock(BB);
1947     BB->dropAllReferences();
1948   }
1949 
1950   for (Function::iterator I = ++F.begin(); I != F.end();) {
1951     auto *BB = &*I;
1952     if (Reachable.count(BB)) {
1953       ++I;
1954       continue;
1955     }
1956     if (DDT) {
1957       DDT->deleteBB(BB); // deferred deletion of BB.
1958       ++I;
1959     } else {
1960       I = F.getBasicBlockList().erase(I);
1961     }
1962   }
1963 
1964   if (DDT)
1965     DDT->applyUpdates(Updates);
1966   return true;
1967 }
1968 
1969 void llvm::combineMetadata(Instruction *K, const Instruction *J,
1970                            ArrayRef<unsigned> KnownIDs) {
1971   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1972   K->dropUnknownNonDebugMetadata(KnownIDs);
1973   K->getAllMetadataOtherThanDebugLoc(Metadata);
1974   for (const auto &MD : Metadata) {
1975     unsigned Kind = MD.first;
1976     MDNode *JMD = J->getMetadata(Kind);
1977     MDNode *KMD = MD.second;
1978 
1979     switch (Kind) {
1980       default:
1981         K->setMetadata(Kind, nullptr); // Remove unknown metadata
1982         break;
1983       case LLVMContext::MD_dbg:
1984         llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1985       case LLVMContext::MD_tbaa:
1986         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1987         break;
1988       case LLVMContext::MD_alias_scope:
1989         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1990         break;
1991       case LLVMContext::MD_noalias:
1992       case LLVMContext::MD_mem_parallel_loop_access:
1993         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1994         break;
1995       case LLVMContext::MD_range:
1996         K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1997         break;
1998       case LLVMContext::MD_fpmath:
1999         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2000         break;
2001       case LLVMContext::MD_invariant_load:
2002         // Only set the !invariant.load if it is present in both instructions.
2003         K->setMetadata(Kind, JMD);
2004         break;
2005       case LLVMContext::MD_nonnull:
2006         // Only set the !nonnull if it is present in both instructions.
2007         K->setMetadata(Kind, JMD);
2008         break;
2009       case LLVMContext::MD_invariant_group:
2010         // Preserve !invariant.group in K.
2011         break;
2012       case LLVMContext::MD_align:
2013         K->setMetadata(Kind,
2014           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
2015         break;
2016       case LLVMContext::MD_dereferenceable:
2017       case LLVMContext::MD_dereferenceable_or_null:
2018         K->setMetadata(Kind,
2019           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
2020         break;
2021     }
2022   }
2023   // Set !invariant.group from J if J has it. If both instructions have it
2024   // then we will just pick it from J - even when they are different.
2025   // Also make sure that K is load or store - f.e. combining bitcast with load
2026   // could produce bitcast with invariant.group metadata, which is invalid.
2027   // FIXME: we should try to preserve both invariant.group md if they are
2028   // different, but right now instruction can only have one invariant.group.
2029   if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
2030     if (isa<LoadInst>(K) || isa<StoreInst>(K))
2031       K->setMetadata(LLVMContext::MD_invariant_group, JMD);
2032 }
2033 
2034 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J) {
2035   unsigned KnownIDs[] = {
2036       LLVMContext::MD_tbaa,            LLVMContext::MD_alias_scope,
2037       LLVMContext::MD_noalias,         LLVMContext::MD_range,
2038       LLVMContext::MD_invariant_load,  LLVMContext::MD_nonnull,
2039       LLVMContext::MD_invariant_group, LLVMContext::MD_align,
2040       LLVMContext::MD_dereferenceable,
2041       LLVMContext::MD_dereferenceable_or_null};
2042   combineMetadata(K, J, KnownIDs);
2043 }
2044 
2045 template <typename RootType, typename DominatesFn>
2046 static unsigned replaceDominatedUsesWith(Value *From, Value *To,
2047                                          const RootType &Root,
2048                                          const DominatesFn &Dominates) {
2049   assert(From->getType() == To->getType());
2050 
2051   unsigned Count = 0;
2052   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2053        UI != UE;) {
2054     Use &U = *UI++;
2055     if (!Dominates(Root, U))
2056       continue;
2057     U.set(To);
2058     DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as "
2059                  << *To << " in " << *U << "\n");
2060     ++Count;
2061   }
2062   return Count;
2063 }
2064 
2065 unsigned llvm::replaceNonLocalUsesWith(Instruction *From, Value *To) {
2066    assert(From->getType() == To->getType());
2067    auto *BB = From->getParent();
2068    unsigned Count = 0;
2069 
2070   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2071        UI != UE;) {
2072     Use &U = *UI++;
2073     auto *I = cast<Instruction>(U.getUser());
2074     if (I->getParent() == BB)
2075       continue;
2076     U.set(To);
2077     ++Count;
2078   }
2079   return Count;
2080 }
2081 
2082 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
2083                                         DominatorTree &DT,
2084                                         const BasicBlockEdge &Root) {
2085   auto Dominates = [&DT](const BasicBlockEdge &Root, const Use &U) {
2086     return DT.dominates(Root, U);
2087   };
2088   return ::replaceDominatedUsesWith(From, To, Root, Dominates);
2089 }
2090 
2091 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
2092                                         DominatorTree &DT,
2093                                         const BasicBlock *BB) {
2094   auto ProperlyDominates = [&DT](const BasicBlock *BB, const Use &U) {
2095     auto *I = cast<Instruction>(U.getUser())->getParent();
2096     return DT.properlyDominates(BB, I);
2097   };
2098   return ::replaceDominatedUsesWith(From, To, BB, ProperlyDominates);
2099 }
2100 
2101 bool llvm::callsGCLeafFunction(ImmutableCallSite CS,
2102                                const TargetLibraryInfo &TLI) {
2103   // Check if the function is specifically marked as a gc leaf function.
2104   if (CS.hasFnAttr("gc-leaf-function"))
2105     return true;
2106   if (const Function *F = CS.getCalledFunction()) {
2107     if (F->hasFnAttribute("gc-leaf-function"))
2108       return true;
2109 
2110     if (auto IID = F->getIntrinsicID())
2111       // Most LLVM intrinsics do not take safepoints.
2112       return IID != Intrinsic::experimental_gc_statepoint &&
2113              IID != Intrinsic::experimental_deoptimize;
2114   }
2115 
2116   // Lib calls can be materialized by some passes, and won't be
2117   // marked as 'gc-leaf-function.' All available Libcalls are
2118   // GC-leaf.
2119   LibFunc LF;
2120   if (TLI.getLibFunc(CS, LF)) {
2121     return TLI.has(LF);
2122   }
2123 
2124   return false;
2125 }
2126 
2127 void llvm::copyNonnullMetadata(const LoadInst &OldLI, MDNode *N,
2128                                LoadInst &NewLI) {
2129   auto *NewTy = NewLI.getType();
2130 
2131   // This only directly applies if the new type is also a pointer.
2132   if (NewTy->isPointerTy()) {
2133     NewLI.setMetadata(LLVMContext::MD_nonnull, N);
2134     return;
2135   }
2136 
2137   // The only other translation we can do is to integral loads with !range
2138   // metadata.
2139   if (!NewTy->isIntegerTy())
2140     return;
2141 
2142   MDBuilder MDB(NewLI.getContext());
2143   const Value *Ptr = OldLI.getPointerOperand();
2144   auto *ITy = cast<IntegerType>(NewTy);
2145   auto *NullInt = ConstantExpr::getPtrToInt(
2146       ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
2147   auto *NonNullInt = ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
2148   NewLI.setMetadata(LLVMContext::MD_range,
2149                     MDB.createRange(NonNullInt, NullInt));
2150 }
2151 
2152 void llvm::copyRangeMetadata(const DataLayout &DL, const LoadInst &OldLI,
2153                              MDNode *N, LoadInst &NewLI) {
2154   auto *NewTy = NewLI.getType();
2155 
2156   // Give up unless it is converted to a pointer where there is a single very
2157   // valuable mapping we can do reliably.
2158   // FIXME: It would be nice to propagate this in more ways, but the type
2159   // conversions make it hard.
2160   if (!NewTy->isPointerTy())
2161     return;
2162 
2163   unsigned BitWidth = DL.getIndexTypeSizeInBits(NewTy);
2164   if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) {
2165     MDNode *NN = MDNode::get(OldLI.getContext(), None);
2166     NewLI.setMetadata(LLVMContext::MD_nonnull, NN);
2167   }
2168 }
2169 
2170 namespace {
2171 
2172 /// A potential constituent of a bitreverse or bswap expression. See
2173 /// collectBitParts for a fuller explanation.
2174 struct BitPart {
2175   BitPart(Value *P, unsigned BW) : Provider(P) {
2176     Provenance.resize(BW);
2177   }
2178 
2179   /// The Value that this is a bitreverse/bswap of.
2180   Value *Provider;
2181 
2182   /// The "provenance" of each bit. Provenance[A] = B means that bit A
2183   /// in Provider becomes bit B in the result of this expression.
2184   SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128.
2185 
2186   enum { Unset = -1 };
2187 };
2188 
2189 } // end anonymous namespace
2190 
2191 /// Analyze the specified subexpression and see if it is capable of providing
2192 /// pieces of a bswap or bitreverse. The subexpression provides a potential
2193 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in
2194 /// the output of the expression came from a corresponding bit in some other
2195 /// value. This function is recursive, and the end result is a mapping of
2196 /// bitnumber to bitnumber. It is the caller's responsibility to validate that
2197 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse.
2198 ///
2199 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know
2200 /// that the expression deposits the low byte of %X into the high byte of the
2201 /// result and that all other bits are zero. This expression is accepted and a
2202 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to
2203 /// [0-7].
2204 ///
2205 /// To avoid revisiting values, the BitPart results are memoized into the
2206 /// provided map. To avoid unnecessary copying of BitParts, BitParts are
2207 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to
2208 /// store BitParts objects, not pointers. As we need the concept of a nullptr
2209 /// BitParts (Value has been analyzed and the analysis failed), we an Optional
2210 /// type instead to provide the same functionality.
2211 ///
2212 /// Because we pass around references into \c BPS, we must use a container that
2213 /// does not invalidate internal references (std::map instead of DenseMap).
2214 static const Optional<BitPart> &
2215 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
2216                 std::map<Value *, Optional<BitPart>> &BPS) {
2217   auto I = BPS.find(V);
2218   if (I != BPS.end())
2219     return I->second;
2220 
2221   auto &Result = BPS[V] = None;
2222   auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2223 
2224   if (Instruction *I = dyn_cast<Instruction>(V)) {
2225     // If this is an or instruction, it may be an inner node of the bswap.
2226     if (I->getOpcode() == Instruction::Or) {
2227       auto &A = collectBitParts(I->getOperand(0), MatchBSwaps,
2228                                 MatchBitReversals, BPS);
2229       auto &B = collectBitParts(I->getOperand(1), MatchBSwaps,
2230                                 MatchBitReversals, BPS);
2231       if (!A || !B)
2232         return Result;
2233 
2234       // Try and merge the two together.
2235       if (!A->Provider || A->Provider != B->Provider)
2236         return Result;
2237 
2238       Result = BitPart(A->Provider, BitWidth);
2239       for (unsigned i = 0; i < A->Provenance.size(); ++i) {
2240         if (A->Provenance[i] != BitPart::Unset &&
2241             B->Provenance[i] != BitPart::Unset &&
2242             A->Provenance[i] != B->Provenance[i])
2243           return Result = None;
2244 
2245         if (A->Provenance[i] == BitPart::Unset)
2246           Result->Provenance[i] = B->Provenance[i];
2247         else
2248           Result->Provenance[i] = A->Provenance[i];
2249       }
2250 
2251       return Result;
2252     }
2253 
2254     // If this is a logical shift by a constant, recurse then shift the result.
2255     if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
2256       unsigned BitShift =
2257           cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
2258       // Ensure the shift amount is defined.
2259       if (BitShift > BitWidth)
2260         return Result;
2261 
2262       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2263                                   MatchBitReversals, BPS);
2264       if (!Res)
2265         return Result;
2266       Result = Res;
2267 
2268       // Perform the "shift" on BitProvenance.
2269       auto &P = Result->Provenance;
2270       if (I->getOpcode() == Instruction::Shl) {
2271         P.erase(std::prev(P.end(), BitShift), P.end());
2272         P.insert(P.begin(), BitShift, BitPart::Unset);
2273       } else {
2274         P.erase(P.begin(), std::next(P.begin(), BitShift));
2275         P.insert(P.end(), BitShift, BitPart::Unset);
2276       }
2277 
2278       return Result;
2279     }
2280 
2281     // If this is a logical 'and' with a mask that clears bits, recurse then
2282     // unset the appropriate bits.
2283     if (I->getOpcode() == Instruction::And &&
2284         isa<ConstantInt>(I->getOperand(1))) {
2285       APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1);
2286       const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
2287 
2288       // Check that the mask allows a multiple of 8 bits for a bswap, for an
2289       // early exit.
2290       unsigned NumMaskedBits = AndMask.countPopulation();
2291       if (!MatchBitReversals && NumMaskedBits % 8 != 0)
2292         return Result;
2293 
2294       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2295                                   MatchBitReversals, BPS);
2296       if (!Res)
2297         return Result;
2298       Result = Res;
2299 
2300       for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1)
2301         // If the AndMask is zero for this bit, clear the bit.
2302         if ((AndMask & Bit) == 0)
2303           Result->Provenance[i] = BitPart::Unset;
2304       return Result;
2305     }
2306 
2307     // If this is a zext instruction zero extend the result.
2308     if (I->getOpcode() == Instruction::ZExt) {
2309       auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
2310                                   MatchBitReversals, BPS);
2311       if (!Res)
2312         return Result;
2313 
2314       Result = BitPart(Res->Provider, BitWidth);
2315       auto NarrowBitWidth =
2316           cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth();
2317       for (unsigned i = 0; i < NarrowBitWidth; ++i)
2318         Result->Provenance[i] = Res->Provenance[i];
2319       for (unsigned i = NarrowBitWidth; i < BitWidth; ++i)
2320         Result->Provenance[i] = BitPart::Unset;
2321       return Result;
2322     }
2323   }
2324 
2325   // Okay, we got to something that isn't a shift, 'or' or 'and'.  This must be
2326   // the input value to the bswap/bitreverse.
2327   Result = BitPart(V, BitWidth);
2328   for (unsigned i = 0; i < BitWidth; ++i)
2329     Result->Provenance[i] = i;
2330   return Result;
2331 }
2332 
2333 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To,
2334                                           unsigned BitWidth) {
2335   if (From % 8 != To % 8)
2336     return false;
2337   // Convert from bit indices to byte indices and check for a byte reversal.
2338   From >>= 3;
2339   To >>= 3;
2340   BitWidth >>= 3;
2341   return From == BitWidth - To - 1;
2342 }
2343 
2344 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To,
2345                                                unsigned BitWidth) {
2346   return From == BitWidth - To - 1;
2347 }
2348 
2349 bool llvm::recognizeBSwapOrBitReverseIdiom(
2350     Instruction *I, bool MatchBSwaps, bool MatchBitReversals,
2351     SmallVectorImpl<Instruction *> &InsertedInsts) {
2352   if (Operator::getOpcode(I) != Instruction::Or)
2353     return false;
2354   if (!MatchBSwaps && !MatchBitReversals)
2355     return false;
2356   IntegerType *ITy = dyn_cast<IntegerType>(I->getType());
2357   if (!ITy || ITy->getBitWidth() > 128)
2358     return false;   // Can't do vectors or integers > 128 bits.
2359   unsigned BW = ITy->getBitWidth();
2360 
2361   unsigned DemandedBW = BW;
2362   IntegerType *DemandedTy = ITy;
2363   if (I->hasOneUse()) {
2364     if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) {
2365       DemandedTy = cast<IntegerType>(Trunc->getType());
2366       DemandedBW = DemandedTy->getBitWidth();
2367     }
2368   }
2369 
2370   // Try to find all the pieces corresponding to the bswap.
2371   std::map<Value *, Optional<BitPart>> BPS;
2372   auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS);
2373   if (!Res)
2374     return false;
2375   auto &BitProvenance = Res->Provenance;
2376 
2377   // Now, is the bit permutation correct for a bswap or a bitreverse? We can
2378   // only byteswap values with an even number of bytes.
2379   bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true;
2380   for (unsigned i = 0; i < DemandedBW; ++i) {
2381     OKForBSwap &=
2382         bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW);
2383     OKForBitReverse &=
2384         bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW);
2385   }
2386 
2387   Intrinsic::ID Intrin;
2388   if (OKForBSwap && MatchBSwaps)
2389     Intrin = Intrinsic::bswap;
2390   else if (OKForBitReverse && MatchBitReversals)
2391     Intrin = Intrinsic::bitreverse;
2392   else
2393     return false;
2394 
2395   if (ITy != DemandedTy) {
2396     Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy);
2397     Value *Provider = Res->Provider;
2398     IntegerType *ProviderTy = cast<IntegerType>(Provider->getType());
2399     // We may need to truncate the provider.
2400     if (DemandedTy != ProviderTy) {
2401       auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy,
2402                                      "trunc", I);
2403       InsertedInsts.push_back(Trunc);
2404       Provider = Trunc;
2405     }
2406     auto *CI = CallInst::Create(F, Provider, "rev", I);
2407     InsertedInsts.push_back(CI);
2408     auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I);
2409     InsertedInsts.push_back(ExtInst);
2410     return true;
2411   }
2412 
2413   Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy);
2414   InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I));
2415   return true;
2416 }
2417 
2418 // CodeGen has special handling for some string functions that may replace
2419 // them with target-specific intrinsics.  Since that'd skip our interceptors
2420 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses,
2421 // we mark affected calls as NoBuiltin, which will disable optimization
2422 // in CodeGen.
2423 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin(
2424     CallInst *CI, const TargetLibraryInfo *TLI) {
2425   Function *F = CI->getCalledFunction();
2426   LibFunc Func;
2427   if (F && !F->hasLocalLinkage() && F->hasName() &&
2428       TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) &&
2429       !F->doesNotAccessMemory())
2430     CI->addAttribute(AttributeList::FunctionIndex, Attribute::NoBuiltin);
2431 }
2432 
2433 bool llvm::canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx) {
2434   // We can't have a PHI with a metadata type.
2435   if (I->getOperand(OpIdx)->getType()->isMetadataTy())
2436     return false;
2437 
2438   // Early exit.
2439   if (!isa<Constant>(I->getOperand(OpIdx)))
2440     return true;
2441 
2442   switch (I->getOpcode()) {
2443   default:
2444     return true;
2445   case Instruction::Call:
2446   case Instruction::Invoke:
2447     // Can't handle inline asm. Skip it.
2448     if (isa<InlineAsm>(ImmutableCallSite(I).getCalledValue()))
2449       return false;
2450     // Many arithmetic intrinsics have no issue taking a
2451     // variable, however it's hard to distingish these from
2452     // specials such as @llvm.frameaddress that require a constant.
2453     if (isa<IntrinsicInst>(I))
2454       return false;
2455 
2456     // Constant bundle operands may need to retain their constant-ness for
2457     // correctness.
2458     if (ImmutableCallSite(I).isBundleOperand(OpIdx))
2459       return false;
2460     return true;
2461   case Instruction::ShuffleVector:
2462     // Shufflevector masks are constant.
2463     return OpIdx != 2;
2464   case Instruction::Switch:
2465   case Instruction::ExtractValue:
2466     // All operands apart from the first are constant.
2467     return OpIdx == 0;
2468   case Instruction::InsertValue:
2469     // All operands apart from the first and the second are constant.
2470     return OpIdx < 2;
2471   case Instruction::Alloca:
2472     // Static allocas (constant size in the entry block) are handled by
2473     // prologue/epilogue insertion so they're free anyway. We definitely don't
2474     // want to make them non-constant.
2475     return !dyn_cast<AllocaInst>(I)->isStaticAlloca();
2476   case Instruction::GetElementPtr:
2477     if (OpIdx == 0)
2478       return true;
2479     gep_type_iterator It = gep_type_begin(I);
2480     for (auto E = std::next(It, OpIdx); It != E; ++It)
2481       if (It.isStruct())
2482         return false;
2483     return true;
2484   }
2485 }
2486