1 //===-- Local.cpp - Functions to perform local transformations ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This family of functions perform various local transformations to the
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/Hashing.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/EHPersonalities.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/LazyValueInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DIBuilder.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/GlobalVariable.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/Intrinsics.h"
42 #include "llvm/IR/MDBuilder.h"
43 #include "llvm/IR/Metadata.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/IR/ValueHandle.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "local"
55 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
57 //===----------------------------------------------------------------------===//
58 // Local constant propagation.
61 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
62 /// constant value, convert it into an unconditional branch to the constant
63 /// destination. This is a nontrivial operation because the successors of this
64 /// basic block must have their PHI nodes updated.
65 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
66 /// conditions and indirectbr addresses this might make dead if
67 /// DeleteDeadConditions is true.
68 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
69 const TargetLibraryInfo *TLI) {
70 TerminatorInst *T = BB->getTerminator();
71 IRBuilder<> Builder(T);
73 // Branch - See if we are conditional jumping on constant
74 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
75 if (BI->isUnconditional()) return false; // Can't optimize uncond branch
76 BasicBlock *Dest1 = BI->getSuccessor(0);
77 BasicBlock *Dest2 = BI->getSuccessor(1);
79 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
80 // Are we branching on constant?
81 // YES. Change to unconditional branch...
82 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
83 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1;
85 //cerr << "Function: " << T->getParent()->getParent()
86 // << "\nRemoving branch from " << T->getParent()
87 // << "\n\nTo: " << OldDest << endl;
89 // Let the basic block know that we are letting go of it. Based on this,
90 // it will adjust it's PHI nodes.
91 OldDest->removePredecessor(BB);
93 // Replace the conditional branch with an unconditional one.
94 Builder.CreateBr(Destination);
95 BI->eraseFromParent();
99 if (Dest2 == Dest1) { // Conditional branch to same location?
100 // This branch matches something like this:
101 // br bool %cond, label %Dest, label %Dest
102 // and changes it into: br label %Dest
104 // Let the basic block know that we are letting go of one copy of it.
105 assert(BI->getParent() && "Terminator not inserted in block!");
106 Dest1->removePredecessor(BI->getParent());
108 // Replace the conditional branch with an unconditional one.
109 Builder.CreateBr(Dest1);
110 Value *Cond = BI->getCondition();
111 BI->eraseFromParent();
112 if (DeleteDeadConditions)
113 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
119 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) {
120 // If we are switching on a constant, we can convert the switch to an
121 // unconditional branch.
122 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition());
123 BasicBlock *DefaultDest = SI->getDefaultDest();
124 BasicBlock *TheOnlyDest = DefaultDest;
126 // If the default is unreachable, ignore it when searching for TheOnlyDest.
127 if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
128 SI->getNumCases() > 0) {
129 TheOnlyDest = SI->case_begin().getCaseSuccessor();
132 // Figure out which case it goes to.
133 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
135 // Found case matching a constant operand?
136 if (i.getCaseValue() == CI) {
137 TheOnlyDest = i.getCaseSuccessor();
141 // Check to see if this branch is going to the same place as the default
142 // dest. If so, eliminate it as an explicit compare.
143 if (i.getCaseSuccessor() == DefaultDest) {
144 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
145 unsigned NCases = SI->getNumCases();
146 // Fold the case metadata into the default if there will be any branches
147 // left, unless the metadata doesn't match the switch.
148 if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) {
149 // Collect branch weights into a vector.
150 SmallVector<uint32_t, 8> Weights;
151 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
153 auto *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
154 Weights.push_back(CI->getValue().getZExtValue());
156 // Merge weight of this case to the default weight.
157 unsigned idx = i.getCaseIndex();
158 Weights[0] += Weights[idx+1];
159 // Remove weight for this case.
160 std::swap(Weights[idx+1], Weights.back());
162 SI->setMetadata(LLVMContext::MD_prof,
163 MDBuilder(BB->getContext()).
164 createBranchWeights(Weights));
166 // Remove this entry.
167 DefaultDest->removePredecessor(SI->getParent());
173 // Otherwise, check to see if the switch only branches to one destination.
174 // We do this by reseting "TheOnlyDest" to null when we find two non-equal
176 if (i.getCaseSuccessor() != TheOnlyDest) TheOnlyDest = nullptr;
179 if (CI && !TheOnlyDest) {
180 // Branching on a constant, but not any of the cases, go to the default
182 TheOnlyDest = SI->getDefaultDest();
185 // If we found a single destination that we can fold the switch into, do so
188 // Insert the new branch.
189 Builder.CreateBr(TheOnlyDest);
190 BasicBlock *BB = SI->getParent();
192 // Remove entries from PHI nodes which we no longer branch to...
193 for (BasicBlock *Succ : SI->successors()) {
194 // Found case matching a constant operand?
195 if (Succ == TheOnlyDest)
196 TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest
198 Succ->removePredecessor(BB);
201 // Delete the old switch.
202 Value *Cond = SI->getCondition();
203 SI->eraseFromParent();
204 if (DeleteDeadConditions)
205 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
209 if (SI->getNumCases() == 1) {
210 // Otherwise, we can fold this switch into a conditional branch
211 // instruction if it has only one non-default destination.
212 SwitchInst::CaseIt FirstCase = SI->case_begin();
213 Value *Cond = Builder.CreateICmpEQ(SI->getCondition(),
214 FirstCase.getCaseValue(), "cond");
216 // Insert the new branch.
217 BranchInst *NewBr = Builder.CreateCondBr(Cond,
218 FirstCase.getCaseSuccessor(),
219 SI->getDefaultDest());
220 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
221 if (MD && MD->getNumOperands() == 3) {
222 ConstantInt *SICase =
223 mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
225 mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
226 assert(SICase && SIDef);
227 // The TrueWeight should be the weight for the single case of SI.
228 NewBr->setMetadata(LLVMContext::MD_prof,
229 MDBuilder(BB->getContext()).
230 createBranchWeights(SICase->getValue().getZExtValue(),
231 SIDef->getValue().getZExtValue()));
234 // Update make.implicit metadata to the newly-created conditional branch.
235 MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit);
237 NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD);
239 // Delete the old switch.
240 SI->eraseFromParent();
246 if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) {
247 // indirectbr blockaddress(@F, @BB) -> br label @BB
248 if (BlockAddress *BA =
249 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) {
250 BasicBlock *TheOnlyDest = BA->getBasicBlock();
251 // Insert the new branch.
252 Builder.CreateBr(TheOnlyDest);
254 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
255 if (IBI->getDestination(i) == TheOnlyDest)
256 TheOnlyDest = nullptr;
258 IBI->getDestination(i)->removePredecessor(IBI->getParent());
260 Value *Address = IBI->getAddress();
261 IBI->eraseFromParent();
262 if (DeleteDeadConditions)
263 RecursivelyDeleteTriviallyDeadInstructions(Address, TLI);
265 // If we didn't find our destination in the IBI successor list, then we
266 // have undefined behavior. Replace the unconditional branch with an
267 // 'unreachable' instruction.
269 BB->getTerminator()->eraseFromParent();
270 new UnreachableInst(BB->getContext(), BB);
281 //===----------------------------------------------------------------------===//
282 // Local dead code elimination.
285 /// isInstructionTriviallyDead - Return true if the result produced by the
286 /// instruction is not used, and the instruction has no side effects.
288 bool llvm::isInstructionTriviallyDead(Instruction *I,
289 const TargetLibraryInfo *TLI) {
290 if (!I->use_empty() || isa<TerminatorInst>(I)) return false;
292 // We don't want the landingpad-like instructions removed by anything this
297 // We don't want debug info removed by anything this general, unless
298 // debug info is empty.
299 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) {
300 if (DDI->getAddress())
304 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) {
310 if (!I->mayHaveSideEffects()) return true;
312 // Special case intrinsics that "may have side effects" but can be deleted
314 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
315 // Safe to delete llvm.stacksave if dead.
316 if (II->getIntrinsicID() == Intrinsic::stacksave)
319 // Lifetime intrinsics are dead when their right-hand is undef.
320 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
321 II->getIntrinsicID() == Intrinsic::lifetime_end)
322 return isa<UndefValue>(II->getArgOperand(1));
324 // Assumptions are dead if their condition is trivially true. Guards on
325 // true are operationally no-ops. In the future we can consider more
326 // sophisticated tradeoffs for guards considering potential for check
327 // widening, but for now we keep things simple.
328 if (II->getIntrinsicID() == Intrinsic::assume ||
329 II->getIntrinsicID() == Intrinsic::experimental_guard) {
330 if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
331 return !Cond->isZero();
337 if (isAllocLikeFn(I, TLI)) return true;
339 if (CallInst *CI = isFreeCall(I, TLI))
340 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
341 return C->isNullValue() || isa<UndefValue>(C);
346 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
347 /// trivially dead instruction, delete it. If that makes any of its operands
348 /// trivially dead, delete them too, recursively. Return true if any
349 /// instructions were deleted.
351 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
352 const TargetLibraryInfo *TLI) {
353 Instruction *I = dyn_cast<Instruction>(V);
354 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
357 SmallVector<Instruction*, 16> DeadInsts;
358 DeadInsts.push_back(I);
361 I = DeadInsts.pop_back_val();
363 // Null out all of the instruction's operands to see if any operand becomes
365 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
366 Value *OpV = I->getOperand(i);
367 I->setOperand(i, nullptr);
369 if (!OpV->use_empty()) continue;
371 // If the operand is an instruction that became dead as we nulled out the
372 // operand, and if it is 'trivially' dead, delete it in a future loop
374 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
375 if (isInstructionTriviallyDead(OpI, TLI))
376 DeadInsts.push_back(OpI);
379 I->eraseFromParent();
380 } while (!DeadInsts.empty());
385 /// areAllUsesEqual - Check whether the uses of a value are all the same.
386 /// This is similar to Instruction::hasOneUse() except this will also return
387 /// true when there are no uses or multiple uses that all refer to the same
389 static bool areAllUsesEqual(Instruction *I) {
390 Value::user_iterator UI = I->user_begin();
391 Value::user_iterator UE = I->user_end();
396 for (++UI; UI != UE; ++UI) {
403 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
404 /// dead PHI node, due to being a def-use chain of single-use nodes that
405 /// either forms a cycle or is terminated by a trivially dead instruction,
406 /// delete it. If that makes any of its operands trivially dead, delete them
407 /// too, recursively. Return true if a change was made.
408 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
409 const TargetLibraryInfo *TLI) {
410 SmallPtrSet<Instruction*, 4> Visited;
411 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
412 I = cast<Instruction>(*I->user_begin())) {
414 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
416 // If we find an instruction more than once, we're on a cycle that
417 // won't prove fruitful.
418 if (!Visited.insert(I).second) {
419 // Break the cycle and delete the instruction and its operands.
420 I->replaceAllUsesWith(UndefValue::get(I->getType()));
421 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
429 simplifyAndDCEInstruction(Instruction *I,
430 SmallSetVector<Instruction *, 16> &WorkList,
431 const DataLayout &DL,
432 const TargetLibraryInfo *TLI) {
433 if (isInstructionTriviallyDead(I, TLI)) {
434 // Null out all of the instruction's operands to see if any operand becomes
436 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
437 Value *OpV = I->getOperand(i);
438 I->setOperand(i, nullptr);
440 if (!OpV->use_empty() || I == OpV)
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
446 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
447 if (isInstructionTriviallyDead(OpI, TLI))
448 WorkList.insert(OpI);
451 I->eraseFromParent();
456 if (Value *SimpleV = SimplifyInstruction(I, DL)) {
457 // Add the users to the worklist. CAREFUL: an instruction can use itself,
458 // in the case of a phi node.
459 for (User *U : I->users()) {
461 WorkList.insert(cast<Instruction>(U));
465 // Replace the instruction with its simplified value.
466 bool Changed = false;
467 if (!I->use_empty()) {
468 I->replaceAllUsesWith(SimpleV);
471 if (isInstructionTriviallyDead(I, TLI)) {
472 I->eraseFromParent();
480 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
481 /// simplify any instructions in it and recursively delete dead instructions.
483 /// This returns true if it changed the code, note that it can delete
484 /// instructions in other blocks as well in this block.
485 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
486 const TargetLibraryInfo *TLI) {
487 bool MadeChange = false;
488 const DataLayout &DL = BB->getModule()->getDataLayout();
491 // In debug builds, ensure that the terminator of the block is never replaced
492 // or deleted by these simplifications. The idea of simplification is that it
493 // cannot introduce new instructions, and there is no way to replace the
494 // terminator of a block without introducing a new instruction.
495 AssertingVH<Instruction> TerminatorVH(&BB->back());
498 SmallSetVector<Instruction *, 16> WorkList;
499 // Iterate over the original function, only adding insts to the worklist
500 // if they actually need to be revisited. This avoids having to pre-init
501 // the worklist with the entire function's worth of instructions.
502 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end());
504 assert(!BI->isTerminator());
505 Instruction *I = &*BI;
508 // We're visiting this instruction now, so make sure it's not in the
509 // worklist from an earlier visit.
510 if (!WorkList.count(I))
511 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
514 while (!WorkList.empty()) {
515 Instruction *I = WorkList.pop_back_val();
516 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
521 //===----------------------------------------------------------------------===//
522 // Control Flow Graph Restructuring.
526 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
527 /// method is called when we're about to delete Pred as a predecessor of BB. If
528 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
530 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
531 /// nodes that collapse into identity values. For example, if we have:
532 /// x = phi(1, 0, 0, 0)
535 /// .. and delete the predecessor corresponding to the '1', this will attempt to
536 /// recursively fold the and to 0.
537 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) {
538 // This only adjusts blocks with PHI nodes.
539 if (!isa<PHINode>(BB->begin()))
542 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
543 // them down. This will leave us with single entry phi nodes and other phis
544 // that can be removed.
545 BB->removePredecessor(Pred, true);
547 WeakVH PhiIt = &BB->front();
548 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
549 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
550 Value *OldPhiIt = PhiIt;
552 if (!recursivelySimplifyInstruction(PN))
555 // If recursive simplification ended up deleting the next PHI node we would
556 // iterate to, then our iterator is invalid, restart scanning from the top
558 if (PhiIt != OldPhiIt) PhiIt = &BB->front();
563 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
564 /// predecessor is known to have one successor (DestBB!). Eliminate the edge
565 /// between them, moving the instructions in the predecessor into DestBB and
566 /// deleting the predecessor block.
568 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) {
569 // If BB has single-entry PHI nodes, fold them.
570 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
571 Value *NewVal = PN->getIncomingValue(0);
572 // Replace self referencing PHI with undef, it must be dead.
573 if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
574 PN->replaceAllUsesWith(NewVal);
575 PN->eraseFromParent();
578 BasicBlock *PredBB = DestBB->getSinglePredecessor();
579 assert(PredBB && "Block doesn't have a single predecessor!");
581 // Zap anything that took the address of DestBB. Not doing this will give the
582 // address an invalid value.
583 if (DestBB->hasAddressTaken()) {
584 BlockAddress *BA = BlockAddress::get(DestBB);
585 Constant *Replacement =
586 ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1);
587 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
589 BA->destroyConstant();
592 // Anything that branched to PredBB now branches to DestBB.
593 PredBB->replaceAllUsesWith(DestBB);
595 // Splice all the instructions from PredBB to DestBB.
596 PredBB->getTerminator()->eraseFromParent();
597 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
599 // If the PredBB is the entry block of the function, move DestBB up to
600 // become the entry block after we erase PredBB.
601 if (PredBB == &DestBB->getParent()->getEntryBlock())
602 DestBB->moveAfter(PredBB);
605 BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock();
606 DT->changeImmediateDominator(DestBB, PredBBIDom);
607 DT->eraseNode(PredBB);
610 PredBB->eraseFromParent();
613 /// CanMergeValues - Return true if we can choose one of these values to use
614 /// in place of the other. Note that we will always choose the non-undef
616 static bool CanMergeValues(Value *First, Value *Second) {
617 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
620 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
621 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
623 /// Assumption: Succ is the single successor for BB.
625 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
626 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
628 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
629 << Succ->getName() << "\n");
630 // Shortcut, if there is only a single predecessor it must be BB and merging
632 if (Succ->getSinglePredecessor()) return true;
634 // Make a list of the predecessors of BB
635 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
637 // Look at all the phi nodes in Succ, to see if they present a conflict when
638 // merging these blocks
639 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
640 PHINode *PN = cast<PHINode>(I);
642 // If the incoming value from BB is again a PHINode in
643 // BB which has the same incoming value for *PI as PN does, we can
644 // merge the phi nodes and then the blocks can still be merged
645 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
646 if (BBPN && BBPN->getParent() == BB) {
647 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
648 BasicBlock *IBB = PN->getIncomingBlock(PI);
649 if (BBPreds.count(IBB) &&
650 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
651 PN->getIncomingValue(PI))) {
652 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
653 << Succ->getName() << " is conflicting with "
654 << BBPN->getName() << " with regard to common predecessor "
655 << IBB->getName() << "\n");
660 Value* Val = PN->getIncomingValueForBlock(BB);
661 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
662 // See if the incoming value for the common predecessor is equal to the
663 // one for BB, in which case this phi node will not prevent the merging
665 BasicBlock *IBB = PN->getIncomingBlock(PI);
666 if (BBPreds.count(IBB) &&
667 !CanMergeValues(Val, PN->getIncomingValue(PI))) {
668 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
669 << Succ->getName() << " is conflicting with regard to common "
670 << "predecessor " << IBB->getName() << "\n");
680 typedef SmallVector<BasicBlock *, 16> PredBlockVector;
681 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap;
683 /// \brief Determines the value to use as the phi node input for a block.
685 /// Select between \p OldVal any value that we know flows from \p BB
686 /// to a particular phi on the basis of which one (if either) is not
687 /// undef. Update IncomingValues based on the selected value.
689 /// \param OldVal The value we are considering selecting.
690 /// \param BB The block that the value flows in from.
691 /// \param IncomingValues A map from block-to-value for other phi inputs
692 /// that we have examined.
694 /// \returns the selected value.
695 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
696 IncomingValueMap &IncomingValues) {
697 if (!isa<UndefValue>(OldVal)) {
698 assert((!IncomingValues.count(BB) ||
699 IncomingValues.find(BB)->second == OldVal) &&
700 "Expected OldVal to match incoming value from BB!");
702 IncomingValues.insert(std::make_pair(BB, OldVal));
706 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
707 if (It != IncomingValues.end()) return It->second;
712 /// \brief Create a map from block to value for the operands of a
715 /// Create a map from block to value for each non-undef value flowing
718 /// \param PN The phi we are collecting the map for.
719 /// \param IncomingValues [out] The map from block to value for this phi.
720 static void gatherIncomingValuesToPhi(PHINode *PN,
721 IncomingValueMap &IncomingValues) {
722 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
723 BasicBlock *BB = PN->getIncomingBlock(i);
724 Value *V = PN->getIncomingValue(i);
726 if (!isa<UndefValue>(V))
727 IncomingValues.insert(std::make_pair(BB, V));
731 /// \brief Replace the incoming undef values to a phi with the values
732 /// from a block-to-value map.
734 /// \param PN The phi we are replacing the undefs in.
735 /// \param IncomingValues A map from block to value.
736 static void replaceUndefValuesInPhi(PHINode *PN,
737 const IncomingValueMap &IncomingValues) {
738 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
739 Value *V = PN->getIncomingValue(i);
741 if (!isa<UndefValue>(V)) continue;
743 BasicBlock *BB = PN->getIncomingBlock(i);
744 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
745 if (It == IncomingValues.end()) continue;
747 PN->setIncomingValue(i, It->second);
751 /// \brief Replace a value flowing from a block to a phi with
752 /// potentially multiple instances of that value flowing from the
753 /// block's predecessors to the phi.
755 /// \param BB The block with the value flowing into the phi.
756 /// \param BBPreds The predecessors of BB.
757 /// \param PN The phi that we are updating.
758 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
759 const PredBlockVector &BBPreds,
761 Value *OldVal = PN->removeIncomingValue(BB, false);
762 assert(OldVal && "No entry in PHI for Pred BB!");
764 IncomingValueMap IncomingValues;
766 // We are merging two blocks - BB, and the block containing PN - and
767 // as a result we need to redirect edges from the predecessors of BB
768 // to go to the block containing PN, and update PN
769 // accordingly. Since we allow merging blocks in the case where the
770 // predecessor and successor blocks both share some predecessors,
771 // and where some of those common predecessors might have undef
772 // values flowing into PN, we want to rewrite those values to be
773 // consistent with the non-undef values.
775 gatherIncomingValuesToPhi(PN, IncomingValues);
777 // If this incoming value is one of the PHI nodes in BB, the new entries
778 // in the PHI node are the entries from the old PHI.
779 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
780 PHINode *OldValPN = cast<PHINode>(OldVal);
781 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
782 // Note that, since we are merging phi nodes and BB and Succ might
783 // have common predecessors, we could end up with a phi node with
784 // identical incoming branches. This will be cleaned up later (and
785 // will trigger asserts if we try to clean it up now, without also
786 // simplifying the corresponding conditional branch).
787 BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
788 Value *PredVal = OldValPN->getIncomingValue(i);
789 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
792 // And add a new incoming value for this predecessor for the
793 // newly retargeted branch.
794 PN->addIncoming(Selected, PredBB);
797 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
798 // Update existing incoming values in PN for this
799 // predecessor of BB.
800 BasicBlock *PredBB = BBPreds[i];
801 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
804 // And add a new incoming value for this predecessor for the
805 // newly retargeted branch.
806 PN->addIncoming(Selected, PredBB);
810 replaceUndefValuesInPhi(PN, IncomingValues);
813 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
814 /// unconditional branch, and contains no instructions other than PHI nodes,
815 /// potential side-effect free intrinsics and the branch. If possible,
816 /// eliminate BB by rewriting all the predecessors to branch to the successor
817 /// block and return true. If we can't transform, return false.
818 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) {
819 assert(BB != &BB->getParent()->getEntryBlock() &&
820 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
822 // We can't eliminate infinite loops.
823 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
824 if (BB == Succ) return false;
826 // Check to see if merging these blocks would cause conflicts for any of the
827 // phi nodes in BB or Succ. If not, we can safely merge.
828 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
830 // Check for cases where Succ has multiple predecessors and a PHI node in BB
831 // has uses which will not disappear when the PHI nodes are merged. It is
832 // possible to handle such cases, but difficult: it requires checking whether
833 // BB dominates Succ, which is non-trivial to calculate in the case where
834 // Succ has multiple predecessors. Also, it requires checking whether
835 // constructing the necessary self-referential PHI node doesn't introduce any
836 // conflicts; this isn't too difficult, but the previous code for doing this
839 // Note that if this check finds a live use, BB dominates Succ, so BB is
840 // something like a loop pre-header (or rarely, a part of an irreducible CFG);
841 // folding the branch isn't profitable in that case anyway.
842 if (!Succ->getSinglePredecessor()) {
843 BasicBlock::iterator BBI = BB->begin();
844 while (isa<PHINode>(*BBI)) {
845 for (Use &U : BBI->uses()) {
846 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
847 if (PN->getIncomingBlock(U) != BB)
857 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
859 if (isa<PHINode>(Succ->begin())) {
860 // If there is more than one pred of succ, and there are PHI nodes in
861 // the successor, then we need to add incoming edges for the PHI nodes
863 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
865 // Loop over all of the PHI nodes in the successor of BB.
866 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
867 PHINode *PN = cast<PHINode>(I);
869 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
873 if (Succ->getSinglePredecessor()) {
874 // BB is the only predecessor of Succ, so Succ will end up with exactly
875 // the same predecessors BB had.
877 // Copy over any phi, debug or lifetime instruction.
878 BB->getTerminator()->eraseFromParent();
879 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
882 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
883 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
884 assert(PN->use_empty() && "There shouldn't be any uses here!");
885 PN->eraseFromParent();
889 // Everything that jumped to BB now goes to Succ.
890 BB->replaceAllUsesWith(Succ);
891 if (!Succ->hasName()) Succ->takeName(BB);
892 BB->eraseFromParent(); // Delete the old basic block.
896 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
897 /// nodes in this block. This doesn't try to be clever about PHI nodes
898 /// which differ only in the order of the incoming values, but instcombine
899 /// orders them so it usually won't matter.
901 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
902 // This implementation doesn't currently consider undef operands
903 // specially. Theoretically, two phis which are identical except for
904 // one having an undef where the other doesn't could be collapsed.
906 struct PHIDenseMapInfo {
907 static PHINode *getEmptyKey() {
908 return DenseMapInfo<PHINode *>::getEmptyKey();
910 static PHINode *getTombstoneKey() {
911 return DenseMapInfo<PHINode *>::getTombstoneKey();
913 static unsigned getHashValue(PHINode *PN) {
914 // Compute a hash value on the operands. Instcombine will likely have
915 // sorted them, which helps expose duplicates, but we have to check all
916 // the operands to be safe in case instcombine hasn't run.
917 return static_cast<unsigned>(hash_combine(
918 hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
919 hash_combine_range(PN->block_begin(), PN->block_end())));
921 static bool isEqual(PHINode *LHS, PHINode *RHS) {
922 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
923 RHS == getEmptyKey() || RHS == getTombstoneKey())
925 return LHS->isIdenticalTo(RHS);
929 // Set of unique PHINodes.
930 DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
933 bool Changed = false;
934 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
935 auto Inserted = PHISet.insert(PN);
936 if (!Inserted.second) {
937 // A duplicate. Replace this PHI with its duplicate.
938 PN->replaceAllUsesWith(*Inserted.first);
939 PN->eraseFromParent();
942 // The RAUW can change PHIs that we already visited. Start over from the
952 /// enforceKnownAlignment - If the specified pointer points to an object that
953 /// we control, modify the object's alignment to PrefAlign. This isn't
954 /// often possible though. If alignment is important, a more reliable approach
955 /// is to simply align all global variables and allocation instructions to
956 /// their preferred alignment from the beginning.
958 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
960 const DataLayout &DL) {
961 assert(PrefAlign > Align);
963 V = V->stripPointerCasts();
965 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
966 // TODO: ideally, computeKnownBits ought to have used
967 // AllocaInst::getAlignment() in its computation already, making
968 // the below max redundant. But, as it turns out,
969 // stripPointerCasts recurses through infinite layers of bitcasts,
970 // while computeKnownBits is not allowed to traverse more than 6
972 Align = std::max(AI->getAlignment(), Align);
973 if (PrefAlign <= Align)
976 // If the preferred alignment is greater than the natural stack alignment
977 // then don't round up. This avoids dynamic stack realignment.
978 if (DL.exceedsNaturalStackAlignment(PrefAlign))
980 AI->setAlignment(PrefAlign);
984 if (auto *GO = dyn_cast<GlobalObject>(V)) {
985 // TODO: as above, this shouldn't be necessary.
986 Align = std::max(GO->getAlignment(), Align);
987 if (PrefAlign <= Align)
990 // If there is a large requested alignment and we can, bump up the alignment
991 // of the global. If the memory we set aside for the global may not be the
992 // memory used by the final program then it is impossible for us to reliably
993 // enforce the preferred alignment.
994 if (!GO->canIncreaseAlignment())
997 GO->setAlignment(PrefAlign);
1004 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that
1005 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
1006 /// and it is more than the alignment of the ultimate object, see if we can
1007 /// increase the alignment of the ultimate object, making this check succeed.
1008 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
1009 const DataLayout &DL,
1010 const Instruction *CxtI,
1011 AssumptionCache *AC,
1012 const DominatorTree *DT) {
1013 assert(V->getType()->isPointerTy() &&
1014 "getOrEnforceKnownAlignment expects a pointer!");
1015 unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType());
1017 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1018 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT);
1019 unsigned TrailZ = KnownZero.countTrailingOnes();
1021 // Avoid trouble with ridiculously large TrailZ values, such as
1022 // those computed from a null pointer.
1023 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1025 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
1027 // LLVM doesn't support alignments larger than this currently.
1028 Align = std::min(Align, +Value::MaximumAlignment);
1030 if (PrefAlign > Align)
1031 Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1033 // We don't need to make any adjustment.
1037 ///===---------------------------------------------------------------------===//
1038 /// Dbg Intrinsic utilities
1041 /// See if there is a dbg.value intrinsic for DIVar before I.
1042 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr,
1044 // Since we can't guarantee that the original dbg.declare instrinsic
1045 // is removed by LowerDbgDeclare(), we need to make sure that we are
1046 // not inserting the same dbg.value intrinsic over and over.
1047 llvm::BasicBlock::InstListType::iterator PrevI(I);
1048 if (PrevI != I->getParent()->getInstList().begin()) {
1050 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1051 if (DVI->getValue() == I->getOperand(0) &&
1052 DVI->getOffset() == 0 &&
1053 DVI->getVariable() == DIVar &&
1054 DVI->getExpression() == DIExpr)
1060 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1061 /// that has an associated llvm.dbg.decl intrinsic.
1062 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1063 StoreInst *SI, DIBuilder &Builder) {
1064 auto *DIVar = DDI->getVariable();
1065 auto *DIExpr = DDI->getExpression();
1066 assert(DIVar && "Missing variable");
1068 // If an argument is zero extended then use argument directly. The ZExt
1069 // may be zapped by an optimization pass in future.
1070 Argument *ExtendedArg = nullptr;
1071 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1072 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1073 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1074 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1076 // We're now only describing a subset of the variable. The piece we're
1077 // describing will always be smaller than the variable size, because
1078 // VariableSize == Size of Alloca described by DDI. Since SI stores
1079 // to the alloca described by DDI, if it's first operand is an extend,
1080 // we're guaranteed that before extension, the value was narrower than
1081 // the size of the alloca, hence the size of the described variable.
1082 SmallVector<uint64_t, 3> Ops;
1083 unsigned PieceOffset = 0;
1084 // If this already is a bit piece, we drop the bit piece from the expression
1085 // and record the offset.
1086 if (DIExpr->isBitPiece()) {
1087 Ops.append(DIExpr->elements_begin(), DIExpr->elements_end()-3);
1088 PieceOffset = DIExpr->getBitPieceOffset();
1090 Ops.append(DIExpr->elements_begin(), DIExpr->elements_end());
1092 Ops.push_back(dwarf::DW_OP_bit_piece);
1093 Ops.push_back(PieceOffset); // Offset
1094 const DataLayout &DL = DDI->getModule()->getDataLayout();
1095 Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType())); // Size
1096 auto NewDIExpr = Builder.createExpression(Ops);
1097 if (!LdStHasDebugValue(DIVar, NewDIExpr, SI))
1098 Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, NewDIExpr,
1099 DDI->getDebugLoc(), SI);
1100 } else if (!LdStHasDebugValue(DIVar, DIExpr, SI))
1101 Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr,
1102 DDI->getDebugLoc(), SI);
1106 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1107 /// that has an associated llvm.dbg.decl intrinsic.
1108 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1109 LoadInst *LI, DIBuilder &Builder) {
1110 auto *DIVar = DDI->getVariable();
1111 auto *DIExpr = DDI->getExpression();
1112 assert(DIVar && "Missing variable");
1114 if (LdStHasDebugValue(DIVar, DIExpr, LI))
1117 // We are now tracking the loaded value instead of the address. In the
1118 // future if multi-location support is added to the IR, it might be
1119 // preferable to keep tracking both the loaded value and the original
1120 // address in case the alloca can not be elided.
1121 Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1122 LI, 0, DIVar, DIExpr, DDI->getDebugLoc(), (Instruction *)nullptr);
1123 DbgValue->insertAfter(LI);
1127 /// Determine whether this alloca is either a VLA or an array.
1128 static bool isArray(AllocaInst *AI) {
1129 return AI->isArrayAllocation() ||
1130 AI->getType()->getElementType()->isArrayTy();
1133 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1134 /// of llvm.dbg.value intrinsics.
1135 bool llvm::LowerDbgDeclare(Function &F) {
1136 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1137 SmallVector<DbgDeclareInst *, 4> Dbgs;
1139 for (Instruction &BI : FI)
1140 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1141 Dbgs.push_back(DDI);
1146 for (auto &I : Dbgs) {
1147 DbgDeclareInst *DDI = I;
1148 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1149 // If this is an alloca for a scalar variable, insert a dbg.value
1150 // at each load and store to the alloca and erase the dbg.declare.
1151 // The dbg.values allow tracking a variable even if it is not
1152 // stored on the stack, while the dbg.declare can only describe
1153 // the stack slot (and at a lexical-scope granularity). Later
1154 // passes will attempt to elide the stack slot.
1155 if (AI && !isArray(AI)) {
1156 for (auto &AIUse : AI->uses()) {
1157 User *U = AIUse.getUser();
1158 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1159 if (AIUse.getOperandNo() == 1)
1160 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1161 } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1162 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1163 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1164 // This is a call by-value or some other instruction that
1165 // takes a pointer to the variable. Insert a *value*
1166 // intrinsic that describes the alloca.
1167 SmallVector<uint64_t, 1> NewDIExpr;
1168 auto *DIExpr = DDI->getExpression();
1169 NewDIExpr.push_back(dwarf::DW_OP_deref);
1170 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1171 DIB.insertDbgValueIntrinsic(AI, 0, DDI->getVariable(),
1172 DIB.createExpression(NewDIExpr),
1173 DDI->getDebugLoc(), CI);
1176 DDI->eraseFromParent();
1182 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
1183 /// alloca 'V', if any.
1184 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) {
1185 if (auto *L = LocalAsMetadata::getIfExists(V))
1186 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1187 for (User *U : MDV->users())
1188 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1194 static void DIExprAddDeref(SmallVectorImpl<uint64_t> &Expr) {
1195 Expr.push_back(dwarf::DW_OP_deref);
1198 static void DIExprAddOffset(SmallVectorImpl<uint64_t> &Expr, int Offset) {
1200 Expr.push_back(dwarf::DW_OP_plus);
1201 Expr.push_back(Offset);
1202 } else if (Offset < 0) {
1203 Expr.push_back(dwarf::DW_OP_minus);
1204 Expr.push_back(-Offset);
1208 static DIExpression *BuildReplacementDIExpr(DIBuilder &Builder,
1209 DIExpression *DIExpr, bool Deref,
1211 if (!Deref && !Offset)
1213 // Create a copy of the original DIDescriptor for user variable, prepending
1214 // "deref" operation to a list of address elements, as new llvm.dbg.declare
1215 // will take a value storing address of the memory for variable, not
1217 SmallVector<uint64_t, 4> NewDIExpr;
1219 DIExprAddDeref(NewDIExpr);
1220 DIExprAddOffset(NewDIExpr, Offset);
1222 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1223 return Builder.createExpression(NewDIExpr);
1226 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1227 Instruction *InsertBefore, DIBuilder &Builder,
1228 bool Deref, int Offset) {
1229 DbgDeclareInst *DDI = FindAllocaDbgDeclare(Address);
1232 DebugLoc Loc = DDI->getDebugLoc();
1233 auto *DIVar = DDI->getVariable();
1234 auto *DIExpr = DDI->getExpression();
1235 assert(DIVar && "Missing variable");
1237 DIExpr = BuildReplacementDIExpr(Builder, DIExpr, Deref, Offset);
1239 // Insert llvm.dbg.declare immediately after the original alloca, and remove
1240 // old llvm.dbg.declare.
1241 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1242 DDI->eraseFromParent();
1246 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1247 DIBuilder &Builder, bool Deref, int Offset) {
1248 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1252 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress,
1253 DIBuilder &Builder, int Offset) {
1254 DebugLoc Loc = DVI->getDebugLoc();
1255 auto *DIVar = DVI->getVariable();
1256 auto *DIExpr = DVI->getExpression();
1257 assert(DIVar && "Missing variable");
1259 // This is an alloca-based llvm.dbg.value. The first thing it should do with
1260 // the alloca pointer is dereference it. Otherwise we don't know how to handle
1262 if (!DIExpr || DIExpr->getNumElements() < 1 ||
1263 DIExpr->getElement(0) != dwarf::DW_OP_deref)
1266 // Insert the offset immediately after the first deref.
1267 // We could just change the offset argument of dbg.value, but it's unsigned...
1269 SmallVector<uint64_t, 4> NewDIExpr;
1270 DIExprAddDeref(NewDIExpr);
1271 DIExprAddOffset(NewDIExpr, Offset);
1272 NewDIExpr.append(DIExpr->elements_begin() + 1, DIExpr->elements_end());
1273 DIExpr = Builder.createExpression(NewDIExpr);
1276 Builder.insertDbgValueIntrinsic(NewAddress, DVI->getOffset(), DIVar, DIExpr,
1278 DVI->eraseFromParent();
1281 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1282 DIBuilder &Builder, int Offset) {
1283 if (auto *L = LocalAsMetadata::getIfExists(AI))
1284 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
1285 for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) {
1287 if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser()))
1288 replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset);
1292 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) {
1293 unsigned NumDeadInst = 0;
1294 // Delete the instructions backwards, as it has a reduced likelihood of
1295 // having to update as many def-use and use-def chains.
1296 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1297 while (EndInst != &BB->front()) {
1298 // Delete the next to last instruction.
1299 Instruction *Inst = &*--EndInst->getIterator();
1300 if (!Inst->use_empty() && !Inst->getType()->isTokenTy())
1301 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1302 if (Inst->isEHPad() || Inst->getType()->isTokenTy()) {
1306 if (!isa<DbgInfoIntrinsic>(Inst))
1308 Inst->eraseFromParent();
1313 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap) {
1314 BasicBlock *BB = I->getParent();
1315 // Loop over all of the successors, removing BB's entry from any PHI
1317 for (BasicBlock *Successor : successors(BB))
1318 Successor->removePredecessor(BB);
1320 // Insert a call to llvm.trap right before this. This turns the undefined
1321 // behavior into a hard fail instead of falling through into random code.
1324 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1325 CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1326 CallTrap->setDebugLoc(I->getDebugLoc());
1328 new UnreachableInst(I->getContext(), I);
1330 // All instructions after this are dead.
1331 unsigned NumInstrsRemoved = 0;
1332 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1333 while (BBI != BBE) {
1334 if (!BBI->use_empty())
1335 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1336 BB->getInstList().erase(BBI++);
1339 return NumInstrsRemoved;
1342 /// changeToCall - Convert the specified invoke into a normal call.
1343 static void changeToCall(InvokeInst *II) {
1344 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1345 SmallVector<OperandBundleDef, 1> OpBundles;
1346 II->getOperandBundlesAsDefs(OpBundles);
1347 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1349 NewCall->takeName(II);
1350 NewCall->setCallingConv(II->getCallingConv());
1351 NewCall->setAttributes(II->getAttributes());
1352 NewCall->setDebugLoc(II->getDebugLoc());
1353 II->replaceAllUsesWith(NewCall);
1355 // Follow the call by a branch to the normal destination.
1356 BranchInst::Create(II->getNormalDest(), II);
1358 // Update PHI nodes in the unwind destination
1359 II->getUnwindDest()->removePredecessor(II->getParent());
1360 II->eraseFromParent();
1363 static bool markAliveBlocks(Function &F,
1364 SmallPtrSetImpl<BasicBlock*> &Reachable) {
1366 SmallVector<BasicBlock*, 128> Worklist;
1367 BasicBlock *BB = &F.front();
1368 Worklist.push_back(BB);
1369 Reachable.insert(BB);
1370 bool Changed = false;
1372 BB = Worklist.pop_back_val();
1374 // Do a quick scan of the basic block, turning any obviously unreachable
1375 // instructions into LLVM unreachable insts. The instruction combining pass
1376 // canonicalizes unreachable insts into stores to null or undef.
1377 for (Instruction &I : *BB) {
1378 // Assumptions that are known to be false are equivalent to unreachable.
1379 // Also, if the condition is undefined, then we make the choice most
1380 // beneficial to the optimizer, and choose that to also be unreachable.
1381 if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1382 if (II->getIntrinsicID() == Intrinsic::assume) {
1383 if (match(II->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) {
1384 // Don't insert a call to llvm.trap right before the unreachable.
1385 changeToUnreachable(II, false);
1391 if (II->getIntrinsicID() == Intrinsic::experimental_guard) {
1392 // A call to the guard intrinsic bails out of the current compilation
1393 // unit if the predicate passed to it is false. If the predicate is a
1394 // constant false, then we know the guard will bail out of the current
1395 // compile unconditionally, so all code following it is dead.
1397 // Note: unlike in llvm.assume, it is not "obviously profitable" for
1398 // guards to treat `undef` as `false` since a guard on `undef` can
1399 // still be useful for widening.
1400 if (match(II->getArgOperand(0), m_Zero()))
1401 if (!isa<UnreachableInst>(II->getNextNode())) {
1402 changeToUnreachable(II->getNextNode(), /*UseLLVMTrap=*/ false);
1409 if (auto *CI = dyn_cast<CallInst>(&I)) {
1410 Value *Callee = CI->getCalledValue();
1411 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1412 changeToUnreachable(CI, /*UseLLVMTrap=*/false);
1416 if (CI->doesNotReturn()) {
1417 // If we found a call to a no-return function, insert an unreachable
1418 // instruction after it. Make sure there isn't *already* one there
1420 if (!isa<UnreachableInst>(CI->getNextNode())) {
1421 // Don't insert a call to llvm.trap right before the unreachable.
1422 changeToUnreachable(CI->getNextNode(), false);
1429 // Store to undef and store to null are undefined and used to signal that
1430 // they should be changed to unreachable by passes that can't modify the
1432 if (auto *SI = dyn_cast<StoreInst>(&I)) {
1433 // Don't touch volatile stores.
1434 if (SI->isVolatile()) continue;
1436 Value *Ptr = SI->getOperand(1);
1438 if (isa<UndefValue>(Ptr) ||
1439 (isa<ConstantPointerNull>(Ptr) &&
1440 SI->getPointerAddressSpace() == 0)) {
1441 changeToUnreachable(SI, true);
1448 TerminatorInst *Terminator = BB->getTerminator();
1449 if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
1450 // Turn invokes that call 'nounwind' functions into ordinary calls.
1451 Value *Callee = II->getCalledValue();
1452 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1453 changeToUnreachable(II, true);
1455 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
1456 if (II->use_empty() && II->onlyReadsMemory()) {
1457 // jump to the normal destination branch.
1458 BranchInst::Create(II->getNormalDest(), II);
1459 II->getUnwindDest()->removePredecessor(II->getParent());
1460 II->eraseFromParent();
1465 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
1466 // Remove catchpads which cannot be reached.
1467 struct CatchPadDenseMapInfo {
1468 static CatchPadInst *getEmptyKey() {
1469 return DenseMapInfo<CatchPadInst *>::getEmptyKey();
1471 static CatchPadInst *getTombstoneKey() {
1472 return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
1474 static unsigned getHashValue(CatchPadInst *CatchPad) {
1475 return static_cast<unsigned>(hash_combine_range(
1476 CatchPad->value_op_begin(), CatchPad->value_op_end()));
1478 static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
1479 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1480 RHS == getEmptyKey() || RHS == getTombstoneKey())
1482 return LHS->isIdenticalTo(RHS);
1486 // Set of unique CatchPads.
1487 SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
1488 CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
1490 detail::DenseSetEmpty Empty;
1491 for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
1492 E = CatchSwitch->handler_end();
1494 BasicBlock *HandlerBB = *I;
1495 auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
1496 if (!HandlerSet.insert({CatchPad, Empty}).second) {
1497 CatchSwitch->removeHandler(I);
1505 Changed |= ConstantFoldTerminator(BB, true);
1506 for (BasicBlock *Successor : successors(BB))
1507 if (Reachable.insert(Successor).second)
1508 Worklist.push_back(Successor);
1509 } while (!Worklist.empty());
1513 void llvm::removeUnwindEdge(BasicBlock *BB) {
1514 TerminatorInst *TI = BB->getTerminator();
1516 if (auto *II = dyn_cast<InvokeInst>(TI)) {
1521 TerminatorInst *NewTI;
1522 BasicBlock *UnwindDest;
1524 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
1525 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
1526 UnwindDest = CRI->getUnwindDest();
1527 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
1528 auto *NewCatchSwitch = CatchSwitchInst::Create(
1529 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
1530 CatchSwitch->getName(), CatchSwitch);
1531 for (BasicBlock *PadBB : CatchSwitch->handlers())
1532 NewCatchSwitch->addHandler(PadBB);
1534 NewTI = NewCatchSwitch;
1535 UnwindDest = CatchSwitch->getUnwindDest();
1537 llvm_unreachable("Could not find unwind successor");
1540 NewTI->takeName(TI);
1541 NewTI->setDebugLoc(TI->getDebugLoc());
1542 UnwindDest->removePredecessor(BB);
1543 TI->replaceAllUsesWith(NewTI);
1544 TI->eraseFromParent();
1547 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even
1548 /// if they are in a dead cycle. Return true if a change was made, false
1550 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI) {
1551 SmallPtrSet<BasicBlock*, 16> Reachable;
1552 bool Changed = markAliveBlocks(F, Reachable);
1554 // If there are unreachable blocks in the CFG...
1555 if (Reachable.size() == F.size())
1558 assert(Reachable.size() < F.size());
1559 NumRemoved += F.size()-Reachable.size();
1561 // Loop over all of the basic blocks that are not reachable, dropping all of
1562 // their internal references...
1563 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) {
1564 if (Reachable.count(&*BB))
1567 for (BasicBlock *Successor : successors(&*BB))
1568 if (Reachable.count(Successor))
1569 Successor->removePredecessor(&*BB);
1571 LVI->eraseBlock(&*BB);
1572 BB->dropAllReferences();
1575 for (Function::iterator I = ++F.begin(); I != F.end();)
1576 if (!Reachable.count(&*I))
1577 I = F.getBasicBlockList().erase(I);
1584 void llvm::combineMetadata(Instruction *K, const Instruction *J,
1585 ArrayRef<unsigned> KnownIDs) {
1586 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1587 K->dropUnknownNonDebugMetadata(KnownIDs);
1588 K->getAllMetadataOtherThanDebugLoc(Metadata);
1589 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
1590 unsigned Kind = Metadata[i].first;
1591 MDNode *JMD = J->getMetadata(Kind);
1592 MDNode *KMD = Metadata[i].second;
1596 K->setMetadata(Kind, nullptr); // Remove unknown metadata
1598 case LLVMContext::MD_dbg:
1599 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1600 case LLVMContext::MD_tbaa:
1601 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1603 case LLVMContext::MD_alias_scope:
1604 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1606 case LLVMContext::MD_noalias:
1607 case LLVMContext::MD_mem_parallel_loop_access:
1608 K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1610 case LLVMContext::MD_range:
1611 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1613 case LLVMContext::MD_fpmath:
1614 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1616 case LLVMContext::MD_invariant_load:
1617 // Only set the !invariant.load if it is present in both instructions.
1618 K->setMetadata(Kind, JMD);
1620 case LLVMContext::MD_nonnull:
1621 // Only set the !nonnull if it is present in both instructions.
1622 K->setMetadata(Kind, JMD);
1624 case LLVMContext::MD_invariant_group:
1625 // Preserve !invariant.group in K.
1627 case LLVMContext::MD_align:
1628 K->setMetadata(Kind,
1629 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1631 case LLVMContext::MD_dereferenceable:
1632 case LLVMContext::MD_dereferenceable_or_null:
1633 K->setMetadata(Kind,
1634 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1638 // Set !invariant.group from J if J has it. If both instructions have it
1639 // then we will just pick it from J - even when they are different.
1640 // Also make sure that K is load or store - f.e. combining bitcast with load
1641 // could produce bitcast with invariant.group metadata, which is invalid.
1642 // FIXME: we should try to preserve both invariant.group md if they are
1643 // different, but right now instruction can only have one invariant.group.
1644 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
1645 if (isa<LoadInst>(K) || isa<StoreInst>(K))
1646 K->setMetadata(LLVMContext::MD_invariant_group, JMD);
1649 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1651 const BasicBlockEdge &Root) {
1652 assert(From->getType() == To->getType());
1655 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1658 if (DT.dominates(Root, U)) {
1660 DEBUG(dbgs() << "Replace dominated use of '"
1661 << From->getName() << "' as "
1662 << *To << " in " << *U << "\n");
1669 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1671 const BasicBlock *BB) {
1672 assert(From->getType() == To->getType());
1675 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1678 auto *I = cast<Instruction>(U.getUser());
1679 if (DT.properlyDominates(BB, I->getParent())) {
1681 DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as "
1682 << *To << " in " << *U << "\n");
1689 bool llvm::callsGCLeafFunction(ImmutableCallSite CS) {
1690 // Check if the function is specifically marked as a gc leaf function.
1691 if (CS.hasFnAttr("gc-leaf-function"))
1693 if (const Function *F = CS.getCalledFunction()) {
1694 if (F->hasFnAttribute("gc-leaf-function"))
1697 if (auto IID = F->getIntrinsicID())
1698 // Most LLVM intrinsics do not take safepoints.
1699 return IID != Intrinsic::experimental_gc_statepoint &&
1700 IID != Intrinsic::experimental_deoptimize;
1706 /// A potential constituent of a bitreverse or bswap expression. See
1707 /// collectBitParts for a fuller explanation.
1709 BitPart(Value *P, unsigned BW) : Provider(P) {
1710 Provenance.resize(BW);
1713 /// The Value that this is a bitreverse/bswap of.
1715 /// The "provenance" of each bit. Provenance[A] = B means that bit A
1716 /// in Provider becomes bit B in the result of this expression.
1717 SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128.
1719 enum { Unset = -1 };
1722 /// Analyze the specified subexpression and see if it is capable of providing
1723 /// pieces of a bswap or bitreverse. The subexpression provides a potential
1724 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in
1725 /// the output of the expression came from a corresponding bit in some other
1726 /// value. This function is recursive, and the end result is a mapping of
1727 /// bitnumber to bitnumber. It is the caller's responsibility to validate that
1728 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse.
1730 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know
1731 /// that the expression deposits the low byte of %X into the high byte of the
1732 /// result and that all other bits are zero. This expression is accepted and a
1733 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to
1736 /// To avoid revisiting values, the BitPart results are memoized into the
1737 /// provided map. To avoid unnecessary copying of BitParts, BitParts are
1738 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to
1739 /// store BitParts objects, not pointers. As we need the concept of a nullptr
1740 /// BitParts (Value has been analyzed and the analysis failed), we an Optional
1741 /// type instead to provide the same functionality.
1743 /// Because we pass around references into \c BPS, we must use a container that
1744 /// does not invalidate internal references (std::map instead of DenseMap).
1746 static const Optional<BitPart> &
1747 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
1748 std::map<Value *, Optional<BitPart>> &BPS) {
1749 auto I = BPS.find(V);
1753 auto &Result = BPS[V] = None;
1754 auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1756 if (Instruction *I = dyn_cast<Instruction>(V)) {
1757 // If this is an or instruction, it may be an inner node of the bswap.
1758 if (I->getOpcode() == Instruction::Or) {
1759 auto &A = collectBitParts(I->getOperand(0), MatchBSwaps,
1760 MatchBitReversals, BPS);
1761 auto &B = collectBitParts(I->getOperand(1), MatchBSwaps,
1762 MatchBitReversals, BPS);
1766 // Try and merge the two together.
1767 if (!A->Provider || A->Provider != B->Provider)
1770 Result = BitPart(A->Provider, BitWidth);
1771 for (unsigned i = 0; i < A->Provenance.size(); ++i) {
1772 if (A->Provenance[i] != BitPart::Unset &&
1773 B->Provenance[i] != BitPart::Unset &&
1774 A->Provenance[i] != B->Provenance[i])
1775 return Result = None;
1777 if (A->Provenance[i] == BitPart::Unset)
1778 Result->Provenance[i] = B->Provenance[i];
1780 Result->Provenance[i] = A->Provenance[i];
1786 // If this is a logical shift by a constant, recurse then shift the result.
1787 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1789 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1790 // Ensure the shift amount is defined.
1791 if (BitShift > BitWidth)
1794 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1795 MatchBitReversals, BPS);
1800 // Perform the "shift" on BitProvenance.
1801 auto &P = Result->Provenance;
1802 if (I->getOpcode() == Instruction::Shl) {
1803 P.erase(std::prev(P.end(), BitShift), P.end());
1804 P.insert(P.begin(), BitShift, BitPart::Unset);
1806 P.erase(P.begin(), std::next(P.begin(), BitShift));
1807 P.insert(P.end(), BitShift, BitPart::Unset);
1813 // If this is a logical 'and' with a mask that clears bits, recurse then
1814 // unset the appropriate bits.
1815 if (I->getOpcode() == Instruction::And &&
1816 isa<ConstantInt>(I->getOperand(1))) {
1817 APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1);
1818 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1820 // Check that the mask allows a multiple of 8 bits for a bswap, for an
1822 unsigned NumMaskedBits = AndMask.countPopulation();
1823 if (!MatchBitReversals && NumMaskedBits % 8 != 0)
1826 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1827 MatchBitReversals, BPS);
1832 for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1)
1833 // If the AndMask is zero for this bit, clear the bit.
1834 if ((AndMask & Bit) == 0)
1835 Result->Provenance[i] = BitPart::Unset;
1839 // If this is a zext instruction zero extend the result.
1840 if (I->getOpcode() == Instruction::ZExt) {
1841 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1842 MatchBitReversals, BPS);
1846 Result = BitPart(Res->Provider, BitWidth);
1847 auto NarrowBitWidth =
1848 cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth();
1849 for (unsigned i = 0; i < NarrowBitWidth; ++i)
1850 Result->Provenance[i] = Res->Provenance[i];
1851 for (unsigned i = NarrowBitWidth; i < BitWidth; ++i)
1852 Result->Provenance[i] = BitPart::Unset;
1857 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1858 // the input value to the bswap/bitreverse.
1859 Result = BitPart(V, BitWidth);
1860 for (unsigned i = 0; i < BitWidth; ++i)
1861 Result->Provenance[i] = i;
1865 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To,
1866 unsigned BitWidth) {
1867 if (From % 8 != To % 8)
1869 // Convert from bit indices to byte indices and check for a byte reversal.
1873 return From == BitWidth - To - 1;
1876 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To,
1877 unsigned BitWidth) {
1878 return From == BitWidth - To - 1;
1881 /// Given an OR instruction, check to see if this is a bitreverse
1882 /// idiom. If so, insert the new intrinsic and return true.
1883 bool llvm::recognizeBSwapOrBitReverseIdiom(
1884 Instruction *I, bool MatchBSwaps, bool MatchBitReversals,
1885 SmallVectorImpl<Instruction *> &InsertedInsts) {
1886 if (Operator::getOpcode(I) != Instruction::Or)
1888 if (!MatchBSwaps && !MatchBitReversals)
1890 IntegerType *ITy = dyn_cast<IntegerType>(I->getType());
1891 if (!ITy || ITy->getBitWidth() > 128)
1892 return false; // Can't do vectors or integers > 128 bits.
1893 unsigned BW = ITy->getBitWidth();
1895 unsigned DemandedBW = BW;
1896 IntegerType *DemandedTy = ITy;
1897 if (I->hasOneUse()) {
1898 if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) {
1899 DemandedTy = cast<IntegerType>(Trunc->getType());
1900 DemandedBW = DemandedTy->getBitWidth();
1904 // Try to find all the pieces corresponding to the bswap.
1905 std::map<Value *, Optional<BitPart>> BPS;
1906 auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS);
1909 auto &BitProvenance = Res->Provenance;
1911 // Now, is the bit permutation correct for a bswap or a bitreverse? We can
1912 // only byteswap values with an even number of bytes.
1913 bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true;
1914 for (unsigned i = 0; i < DemandedBW; ++i) {
1916 bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW);
1918 bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW);
1921 Intrinsic::ID Intrin;
1922 if (OKForBSwap && MatchBSwaps)
1923 Intrin = Intrinsic::bswap;
1924 else if (OKForBitReverse && MatchBitReversals)
1925 Intrin = Intrinsic::bitreverse;
1929 if (ITy != DemandedTy) {
1930 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy);
1931 Value *Provider = Res->Provider;
1932 IntegerType *ProviderTy = cast<IntegerType>(Provider->getType());
1933 // We may need to truncate the provider.
1934 if (DemandedTy != ProviderTy) {
1935 auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy,
1937 InsertedInsts.push_back(Trunc);
1940 auto *CI = CallInst::Create(F, Provider, "rev", I);
1941 InsertedInsts.push_back(CI);
1942 auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I);
1943 InsertedInsts.push_back(ExtInst);
1947 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy);
1948 InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I));
1952 // CodeGen has special handling for some string functions that may replace
1953 // them with target-specific intrinsics. Since that'd skip our interceptors
1954 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses,
1955 // we mark affected calls as NoBuiltin, which will disable optimization
1957 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin(CallInst *CI,
1958 const TargetLibraryInfo *TLI) {
1959 Function *F = CI->getCalledFunction();
1961 if (!F || F->hasLocalLinkage() || !F->hasName() ||
1962 !TLI->getLibFunc(F->getName(), Func))
1966 case LibFunc::memcmp:
1967 case LibFunc::memchr:
1968 case LibFunc::strcpy:
1969 case LibFunc::stpcpy:
1970 case LibFunc::strcmp:
1971 case LibFunc::strlen:
1972 case LibFunc::strnlen:
1973 CI->addAttribute(AttributeSet::FunctionIndex, Attribute::NoBuiltin);