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);
343 if (CallSite CS = CallSite(I))
344 if (isMathLibCallNoop(CS, TLI))
350 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
351 /// trivially dead instruction, delete it. If that makes any of its operands
352 /// trivially dead, delete them too, recursively. Return true if any
353 /// instructions were deleted.
355 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
356 const TargetLibraryInfo *TLI) {
357 Instruction *I = dyn_cast<Instruction>(V);
358 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
361 SmallVector<Instruction*, 16> DeadInsts;
362 DeadInsts.push_back(I);
365 I = DeadInsts.pop_back_val();
367 // Null out all of the instruction's operands to see if any operand becomes
369 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
370 Value *OpV = I->getOperand(i);
371 I->setOperand(i, nullptr);
373 if (!OpV->use_empty()) continue;
375 // If the operand is an instruction that became dead as we nulled out the
376 // operand, and if it is 'trivially' dead, delete it in a future loop
378 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
379 if (isInstructionTriviallyDead(OpI, TLI))
380 DeadInsts.push_back(OpI);
383 I->eraseFromParent();
384 } while (!DeadInsts.empty());
389 /// areAllUsesEqual - Check whether the uses of a value are all the same.
390 /// This is similar to Instruction::hasOneUse() except this will also return
391 /// true when there are no uses or multiple uses that all refer to the same
393 static bool areAllUsesEqual(Instruction *I) {
394 Value::user_iterator UI = I->user_begin();
395 Value::user_iterator UE = I->user_end();
400 for (++UI; UI != UE; ++UI) {
407 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
408 /// dead PHI node, due to being a def-use chain of single-use nodes that
409 /// either forms a cycle or is terminated by a trivially dead instruction,
410 /// delete it. If that makes any of its operands trivially dead, delete them
411 /// too, recursively. Return true if a change was made.
412 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
413 const TargetLibraryInfo *TLI) {
414 SmallPtrSet<Instruction*, 4> Visited;
415 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
416 I = cast<Instruction>(*I->user_begin())) {
418 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
420 // If we find an instruction more than once, we're on a cycle that
421 // won't prove fruitful.
422 if (!Visited.insert(I).second) {
423 // Break the cycle and delete the instruction and its operands.
424 I->replaceAllUsesWith(UndefValue::get(I->getType()));
425 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
433 simplifyAndDCEInstruction(Instruction *I,
434 SmallSetVector<Instruction *, 16> &WorkList,
435 const DataLayout &DL,
436 const TargetLibraryInfo *TLI) {
437 if (isInstructionTriviallyDead(I, TLI)) {
438 // Null out all of the instruction's operands to see if any operand becomes
440 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
441 Value *OpV = I->getOperand(i);
442 I->setOperand(i, nullptr);
444 if (!OpV->use_empty() || I == OpV)
447 // If the operand is an instruction that became dead as we nulled out the
448 // operand, and if it is 'trivially' dead, delete it in a future loop
450 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
451 if (isInstructionTriviallyDead(OpI, TLI))
452 WorkList.insert(OpI);
455 I->eraseFromParent();
460 if (Value *SimpleV = SimplifyInstruction(I, DL)) {
461 // Add the users to the worklist. CAREFUL: an instruction can use itself,
462 // in the case of a phi node.
463 for (User *U : I->users()) {
465 WorkList.insert(cast<Instruction>(U));
469 // Replace the instruction with its simplified value.
470 bool Changed = false;
471 if (!I->use_empty()) {
472 I->replaceAllUsesWith(SimpleV);
475 if (isInstructionTriviallyDead(I, TLI)) {
476 I->eraseFromParent();
484 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
485 /// simplify any instructions in it and recursively delete dead instructions.
487 /// This returns true if it changed the code, note that it can delete
488 /// instructions in other blocks as well in this block.
489 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
490 const TargetLibraryInfo *TLI) {
491 bool MadeChange = false;
492 const DataLayout &DL = BB->getModule()->getDataLayout();
495 // In debug builds, ensure that the terminator of the block is never replaced
496 // or deleted by these simplifications. The idea of simplification is that it
497 // cannot introduce new instructions, and there is no way to replace the
498 // terminator of a block without introducing a new instruction.
499 AssertingVH<Instruction> TerminatorVH(&BB->back());
502 SmallSetVector<Instruction *, 16> WorkList;
503 // Iterate over the original function, only adding insts to the worklist
504 // if they actually need to be revisited. This avoids having to pre-init
505 // the worklist with the entire function's worth of instructions.
506 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end());
508 assert(!BI->isTerminator());
509 Instruction *I = &*BI;
512 // We're visiting this instruction now, so make sure it's not in the
513 // worklist from an earlier visit.
514 if (!WorkList.count(I))
515 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
518 while (!WorkList.empty()) {
519 Instruction *I = WorkList.pop_back_val();
520 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
525 //===----------------------------------------------------------------------===//
526 // Control Flow Graph Restructuring.
530 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
531 /// method is called when we're about to delete Pred as a predecessor of BB. If
532 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
534 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
535 /// nodes that collapse into identity values. For example, if we have:
536 /// x = phi(1, 0, 0, 0)
539 /// .. and delete the predecessor corresponding to the '1', this will attempt to
540 /// recursively fold the and to 0.
541 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) {
542 // This only adjusts blocks with PHI nodes.
543 if (!isa<PHINode>(BB->begin()))
546 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
547 // them down. This will leave us with single entry phi nodes and other phis
548 // that can be removed.
549 BB->removePredecessor(Pred, true);
551 WeakVH PhiIt = &BB->front();
552 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
553 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
554 Value *OldPhiIt = PhiIt;
556 if (!recursivelySimplifyInstruction(PN))
559 // If recursive simplification ended up deleting the next PHI node we would
560 // iterate to, then our iterator is invalid, restart scanning from the top
562 if (PhiIt != OldPhiIt) PhiIt = &BB->front();
567 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
568 /// predecessor is known to have one successor (DestBB!). Eliminate the edge
569 /// between them, moving the instructions in the predecessor into DestBB and
570 /// deleting the predecessor block.
572 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) {
573 // If BB has single-entry PHI nodes, fold them.
574 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
575 Value *NewVal = PN->getIncomingValue(0);
576 // Replace self referencing PHI with undef, it must be dead.
577 if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
578 PN->replaceAllUsesWith(NewVal);
579 PN->eraseFromParent();
582 BasicBlock *PredBB = DestBB->getSinglePredecessor();
583 assert(PredBB && "Block doesn't have a single predecessor!");
585 // Zap anything that took the address of DestBB. Not doing this will give the
586 // address an invalid value.
587 if (DestBB->hasAddressTaken()) {
588 BlockAddress *BA = BlockAddress::get(DestBB);
589 Constant *Replacement =
590 ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1);
591 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
593 BA->destroyConstant();
596 // Anything that branched to PredBB now branches to DestBB.
597 PredBB->replaceAllUsesWith(DestBB);
599 // Splice all the instructions from PredBB to DestBB.
600 PredBB->getTerminator()->eraseFromParent();
601 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
603 // If the PredBB is the entry block of the function, move DestBB up to
604 // become the entry block after we erase PredBB.
605 if (PredBB == &DestBB->getParent()->getEntryBlock())
606 DestBB->moveAfter(PredBB);
609 BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock();
610 DT->changeImmediateDominator(DestBB, PredBBIDom);
611 DT->eraseNode(PredBB);
614 PredBB->eraseFromParent();
617 /// CanMergeValues - Return true if we can choose one of these values to use
618 /// in place of the other. Note that we will always choose the non-undef
620 static bool CanMergeValues(Value *First, Value *Second) {
621 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
624 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
625 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
627 /// Assumption: Succ is the single successor for BB.
629 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
630 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
632 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
633 << Succ->getName() << "\n");
634 // Shortcut, if there is only a single predecessor it must be BB and merging
636 if (Succ->getSinglePredecessor()) return true;
638 // Make a list of the predecessors of BB
639 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
641 // Look at all the phi nodes in Succ, to see if they present a conflict when
642 // merging these blocks
643 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
644 PHINode *PN = cast<PHINode>(I);
646 // If the incoming value from BB is again a PHINode in
647 // BB which has the same incoming value for *PI as PN does, we can
648 // merge the phi nodes and then the blocks can still be merged
649 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
650 if (BBPN && BBPN->getParent() == BB) {
651 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
652 BasicBlock *IBB = PN->getIncomingBlock(PI);
653 if (BBPreds.count(IBB) &&
654 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
655 PN->getIncomingValue(PI))) {
656 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
657 << Succ->getName() << " is conflicting with "
658 << BBPN->getName() << " with regard to common predecessor "
659 << IBB->getName() << "\n");
664 Value* Val = PN->getIncomingValueForBlock(BB);
665 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
666 // See if the incoming value for the common predecessor is equal to the
667 // one for BB, in which case this phi node will not prevent the merging
669 BasicBlock *IBB = PN->getIncomingBlock(PI);
670 if (BBPreds.count(IBB) &&
671 !CanMergeValues(Val, PN->getIncomingValue(PI))) {
672 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
673 << Succ->getName() << " is conflicting with regard to common "
674 << "predecessor " << IBB->getName() << "\n");
684 typedef SmallVector<BasicBlock *, 16> PredBlockVector;
685 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap;
687 /// \brief Determines the value to use as the phi node input for a block.
689 /// Select between \p OldVal any value that we know flows from \p BB
690 /// to a particular phi on the basis of which one (if either) is not
691 /// undef. Update IncomingValues based on the selected value.
693 /// \param OldVal The value we are considering selecting.
694 /// \param BB The block that the value flows in from.
695 /// \param IncomingValues A map from block-to-value for other phi inputs
696 /// that we have examined.
698 /// \returns the selected value.
699 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
700 IncomingValueMap &IncomingValues) {
701 if (!isa<UndefValue>(OldVal)) {
702 assert((!IncomingValues.count(BB) ||
703 IncomingValues.find(BB)->second == OldVal) &&
704 "Expected OldVal to match incoming value from BB!");
706 IncomingValues.insert(std::make_pair(BB, OldVal));
710 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
711 if (It != IncomingValues.end()) return It->second;
716 /// \brief Create a map from block to value for the operands of a
719 /// Create a map from block to value for each non-undef value flowing
722 /// \param PN The phi we are collecting the map for.
723 /// \param IncomingValues [out] The map from block to value for this phi.
724 static void gatherIncomingValuesToPhi(PHINode *PN,
725 IncomingValueMap &IncomingValues) {
726 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
727 BasicBlock *BB = PN->getIncomingBlock(i);
728 Value *V = PN->getIncomingValue(i);
730 if (!isa<UndefValue>(V))
731 IncomingValues.insert(std::make_pair(BB, V));
735 /// \brief Replace the incoming undef values to a phi with the values
736 /// from a block-to-value map.
738 /// \param PN The phi we are replacing the undefs in.
739 /// \param IncomingValues A map from block to value.
740 static void replaceUndefValuesInPhi(PHINode *PN,
741 const IncomingValueMap &IncomingValues) {
742 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
743 Value *V = PN->getIncomingValue(i);
745 if (!isa<UndefValue>(V)) continue;
747 BasicBlock *BB = PN->getIncomingBlock(i);
748 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
749 if (It == IncomingValues.end()) continue;
751 PN->setIncomingValue(i, It->second);
755 /// \brief Replace a value flowing from a block to a phi with
756 /// potentially multiple instances of that value flowing from the
757 /// block's predecessors to the phi.
759 /// \param BB The block with the value flowing into the phi.
760 /// \param BBPreds The predecessors of BB.
761 /// \param PN The phi that we are updating.
762 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
763 const PredBlockVector &BBPreds,
765 Value *OldVal = PN->removeIncomingValue(BB, false);
766 assert(OldVal && "No entry in PHI for Pred BB!");
768 IncomingValueMap IncomingValues;
770 // We are merging two blocks - BB, and the block containing PN - and
771 // as a result we need to redirect edges from the predecessors of BB
772 // to go to the block containing PN, and update PN
773 // accordingly. Since we allow merging blocks in the case where the
774 // predecessor and successor blocks both share some predecessors,
775 // and where some of those common predecessors might have undef
776 // values flowing into PN, we want to rewrite those values to be
777 // consistent with the non-undef values.
779 gatherIncomingValuesToPhi(PN, IncomingValues);
781 // If this incoming value is one of the PHI nodes in BB, the new entries
782 // in the PHI node are the entries from the old PHI.
783 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
784 PHINode *OldValPN = cast<PHINode>(OldVal);
785 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
786 // Note that, since we are merging phi nodes and BB and Succ might
787 // have common predecessors, we could end up with a phi node with
788 // identical incoming branches. This will be cleaned up later (and
789 // will trigger asserts if we try to clean it up now, without also
790 // simplifying the corresponding conditional branch).
791 BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
792 Value *PredVal = OldValPN->getIncomingValue(i);
793 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
796 // And add a new incoming value for this predecessor for the
797 // newly retargeted branch.
798 PN->addIncoming(Selected, PredBB);
801 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
802 // Update existing incoming values in PN for this
803 // predecessor of BB.
804 BasicBlock *PredBB = BBPreds[i];
805 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
808 // And add a new incoming value for this predecessor for the
809 // newly retargeted branch.
810 PN->addIncoming(Selected, PredBB);
814 replaceUndefValuesInPhi(PN, IncomingValues);
817 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
818 /// unconditional branch, and contains no instructions other than PHI nodes,
819 /// potential side-effect free intrinsics and the branch. If possible,
820 /// eliminate BB by rewriting all the predecessors to branch to the successor
821 /// block and return true. If we can't transform, return false.
822 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) {
823 assert(BB != &BB->getParent()->getEntryBlock() &&
824 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
826 // We can't eliminate infinite loops.
827 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
828 if (BB == Succ) return false;
830 // Check to see if merging these blocks would cause conflicts for any of the
831 // phi nodes in BB or Succ. If not, we can safely merge.
832 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
834 // Check for cases where Succ has multiple predecessors and a PHI node in BB
835 // has uses which will not disappear when the PHI nodes are merged. It is
836 // possible to handle such cases, but difficult: it requires checking whether
837 // BB dominates Succ, which is non-trivial to calculate in the case where
838 // Succ has multiple predecessors. Also, it requires checking whether
839 // constructing the necessary self-referential PHI node doesn't introduce any
840 // conflicts; this isn't too difficult, but the previous code for doing this
843 // Note that if this check finds a live use, BB dominates Succ, so BB is
844 // something like a loop pre-header (or rarely, a part of an irreducible CFG);
845 // folding the branch isn't profitable in that case anyway.
846 if (!Succ->getSinglePredecessor()) {
847 BasicBlock::iterator BBI = BB->begin();
848 while (isa<PHINode>(*BBI)) {
849 for (Use &U : BBI->uses()) {
850 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
851 if (PN->getIncomingBlock(U) != BB)
861 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
863 if (isa<PHINode>(Succ->begin())) {
864 // If there is more than one pred of succ, and there are PHI nodes in
865 // the successor, then we need to add incoming edges for the PHI nodes
867 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
869 // Loop over all of the PHI nodes in the successor of BB.
870 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
871 PHINode *PN = cast<PHINode>(I);
873 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
877 if (Succ->getSinglePredecessor()) {
878 // BB is the only predecessor of Succ, so Succ will end up with exactly
879 // the same predecessors BB had.
881 // Copy over any phi, debug or lifetime instruction.
882 BB->getTerminator()->eraseFromParent();
883 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
886 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
887 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
888 assert(PN->use_empty() && "There shouldn't be any uses here!");
889 PN->eraseFromParent();
893 // If the unconditional branch we replaced contains llvm.loop metadata, we
894 // add the metadata to the branch instructions in the predecessors.
895 unsigned LoopMDKind = BB->getContext().getMDKindID("llvm.loop");
896 Instruction *TI = BB->getTerminator();
898 if (MDNode *LoopMD = TI->getMetadata(LoopMDKind))
899 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
900 BasicBlock *Pred = *PI;
901 Pred->getTerminator()->setMetadata(LoopMDKind, LoopMD);
904 // Everything that jumped to BB now goes to Succ.
905 BB->replaceAllUsesWith(Succ);
906 if (!Succ->hasName()) Succ->takeName(BB);
907 BB->eraseFromParent(); // Delete the old basic block.
911 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
912 /// nodes in this block. This doesn't try to be clever about PHI nodes
913 /// which differ only in the order of the incoming values, but instcombine
914 /// orders them so it usually won't matter.
916 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
917 // This implementation doesn't currently consider undef operands
918 // specially. Theoretically, two phis which are identical except for
919 // one having an undef where the other doesn't could be collapsed.
921 struct PHIDenseMapInfo {
922 static PHINode *getEmptyKey() {
923 return DenseMapInfo<PHINode *>::getEmptyKey();
925 static PHINode *getTombstoneKey() {
926 return DenseMapInfo<PHINode *>::getTombstoneKey();
928 static unsigned getHashValue(PHINode *PN) {
929 // Compute a hash value on the operands. Instcombine will likely have
930 // sorted them, which helps expose duplicates, but we have to check all
931 // the operands to be safe in case instcombine hasn't run.
932 return static_cast<unsigned>(hash_combine(
933 hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
934 hash_combine_range(PN->block_begin(), PN->block_end())));
936 static bool isEqual(PHINode *LHS, PHINode *RHS) {
937 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
938 RHS == getEmptyKey() || RHS == getTombstoneKey())
940 return LHS->isIdenticalTo(RHS);
944 // Set of unique PHINodes.
945 DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
948 bool Changed = false;
949 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
950 auto Inserted = PHISet.insert(PN);
951 if (!Inserted.second) {
952 // A duplicate. Replace this PHI with its duplicate.
953 PN->replaceAllUsesWith(*Inserted.first);
954 PN->eraseFromParent();
957 // The RAUW can change PHIs that we already visited. Start over from the
967 /// enforceKnownAlignment - If the specified pointer points to an object that
968 /// we control, modify the object's alignment to PrefAlign. This isn't
969 /// often possible though. If alignment is important, a more reliable approach
970 /// is to simply align all global variables and allocation instructions to
971 /// their preferred alignment from the beginning.
973 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
975 const DataLayout &DL) {
976 assert(PrefAlign > Align);
978 V = V->stripPointerCasts();
980 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
981 // TODO: ideally, computeKnownBits ought to have used
982 // AllocaInst::getAlignment() in its computation already, making
983 // the below max redundant. But, as it turns out,
984 // stripPointerCasts recurses through infinite layers of bitcasts,
985 // while computeKnownBits is not allowed to traverse more than 6
987 Align = std::max(AI->getAlignment(), Align);
988 if (PrefAlign <= Align)
991 // If the preferred alignment is greater than the natural stack alignment
992 // then don't round up. This avoids dynamic stack realignment.
993 if (DL.exceedsNaturalStackAlignment(PrefAlign))
995 AI->setAlignment(PrefAlign);
999 if (auto *GO = dyn_cast<GlobalObject>(V)) {
1000 // TODO: as above, this shouldn't be necessary.
1001 Align = std::max(GO->getAlignment(), Align);
1002 if (PrefAlign <= Align)
1005 // If there is a large requested alignment and we can, bump up the alignment
1006 // of the global. If the memory we set aside for the global may not be the
1007 // memory used by the final program then it is impossible for us to reliably
1008 // enforce the preferred alignment.
1009 if (!GO->canIncreaseAlignment())
1012 GO->setAlignment(PrefAlign);
1019 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
1020 const DataLayout &DL,
1021 const Instruction *CxtI,
1022 AssumptionCache *AC,
1023 const DominatorTree *DT) {
1024 assert(V->getType()->isPointerTy() &&
1025 "getOrEnforceKnownAlignment expects a pointer!");
1026 unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType());
1028 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1029 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT);
1030 unsigned TrailZ = KnownZero.countTrailingOnes();
1032 // Avoid trouble with ridiculously large TrailZ values, such as
1033 // those computed from a null pointer.
1034 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1036 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
1038 // LLVM doesn't support alignments larger than this currently.
1039 Align = std::min(Align, +Value::MaximumAlignment);
1041 if (PrefAlign > Align)
1042 Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1044 // We don't need to make any adjustment.
1048 ///===---------------------------------------------------------------------===//
1049 /// Dbg Intrinsic utilities
1052 /// See if there is a dbg.value intrinsic for DIVar before I.
1053 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr,
1055 // Since we can't guarantee that the original dbg.declare instrinsic
1056 // is removed by LowerDbgDeclare(), we need to make sure that we are
1057 // not inserting the same dbg.value intrinsic over and over.
1058 llvm::BasicBlock::InstListType::iterator PrevI(I);
1059 if (PrevI != I->getParent()->getInstList().begin()) {
1061 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1062 if (DVI->getValue() == I->getOperand(0) &&
1063 DVI->getOffset() == 0 &&
1064 DVI->getVariable() == DIVar &&
1065 DVI->getExpression() == DIExpr)
1071 /// See if there is a dbg.value intrinsic for DIVar for the PHI node.
1072 static bool PhiHasDebugValue(DILocalVariable *DIVar,
1073 DIExpression *DIExpr,
1075 // Since we can't guarantee that the original dbg.declare instrinsic
1076 // is removed by LowerDbgDeclare(), we need to make sure that we are
1077 // not inserting the same dbg.value intrinsic over and over.
1078 DbgValueList DbgValues;
1079 FindAllocaDbgValues(DbgValues, APN);
1080 for (auto DVI : DbgValues) {
1081 assert (DVI->getValue() == APN);
1082 assert (DVI->getOffset() == 0);
1083 if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr))
1089 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1090 /// that has an associated llvm.dbg.decl intrinsic.
1091 void llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1092 StoreInst *SI, DIBuilder &Builder) {
1093 auto *DIVar = DDI->getVariable();
1094 auto *DIExpr = DDI->getExpression();
1095 assert(DIVar && "Missing variable");
1097 // If an argument is zero extended then use argument directly. The ZExt
1098 // may be zapped by an optimization pass in future.
1099 Argument *ExtendedArg = nullptr;
1100 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1101 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1102 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1103 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1105 // We're now only describing a subset of the variable. The fragment we're
1106 // describing will always be smaller than the variable size, because
1107 // VariableSize == Size of Alloca described by DDI. Since SI stores
1108 // to the alloca described by DDI, if it's first operand is an extend,
1109 // we're guaranteed that before extension, the value was narrower than
1110 // the size of the alloca, hence the size of the described variable.
1111 SmallVector<uint64_t, 3> Ops;
1112 unsigned FragmentOffset = 0;
1113 // If this already is a bit fragment, we drop the bit fragment from the
1114 // expression and record the offset.
1115 auto Fragment = DIExpr->getFragmentInfo();
1117 Ops.append(DIExpr->elements_begin(), DIExpr->elements_end()-3);
1118 FragmentOffset = Fragment->OffsetInBits;
1120 Ops.append(DIExpr->elements_begin(), DIExpr->elements_end());
1122 Ops.push_back(dwarf::DW_OP_LLVM_fragment);
1123 Ops.push_back(FragmentOffset);
1124 const DataLayout &DL = DDI->getModule()->getDataLayout();
1125 Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType()));
1126 auto NewDIExpr = Builder.createExpression(Ops);
1127 if (!LdStHasDebugValue(DIVar, NewDIExpr, SI))
1128 Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, NewDIExpr,
1129 DDI->getDebugLoc(), SI);
1130 } else if (!LdStHasDebugValue(DIVar, DIExpr, SI))
1131 Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr,
1132 DDI->getDebugLoc(), SI);
1135 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1136 /// that has an associated llvm.dbg.decl intrinsic.
1137 void llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1138 LoadInst *LI, DIBuilder &Builder) {
1139 auto *DIVar = DDI->getVariable();
1140 auto *DIExpr = DDI->getExpression();
1141 assert(DIVar && "Missing variable");
1143 if (LdStHasDebugValue(DIVar, DIExpr, LI))
1146 // We are now tracking the loaded value instead of the address. In the
1147 // future if multi-location support is added to the IR, it might be
1148 // preferable to keep tracking both the loaded value and the original
1149 // address in case the alloca can not be elided.
1150 Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1151 LI, 0, DIVar, DIExpr, DDI->getDebugLoc(), (Instruction *)nullptr);
1152 DbgValue->insertAfter(LI);
1155 /// Inserts a llvm.dbg.value intrinsic after a phi
1156 /// that has an associated llvm.dbg.decl intrinsic.
1157 void llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1158 PHINode *APN, DIBuilder &Builder) {
1159 auto *DIVar = DDI->getVariable();
1160 auto *DIExpr = DDI->getExpression();
1161 assert(DIVar && "Missing variable");
1163 if (PhiHasDebugValue(DIVar, DIExpr, APN))
1166 BasicBlock *BB = APN->getParent();
1167 auto InsertionPt = BB->getFirstInsertionPt();
1169 // The block may be a catchswitch block, which does not have a valid
1171 // FIXME: Insert dbg.value markers in the successors when appropriate.
1172 if (InsertionPt != BB->end())
1173 Builder.insertDbgValueIntrinsic(APN, 0, DIVar, DIExpr, DDI->getDebugLoc(),
1177 /// Determine whether this alloca is either a VLA or an array.
1178 static bool isArray(AllocaInst *AI) {
1179 return AI->isArrayAllocation() ||
1180 AI->getType()->getElementType()->isArrayTy();
1183 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1184 /// of llvm.dbg.value intrinsics.
1185 bool llvm::LowerDbgDeclare(Function &F) {
1186 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1187 SmallVector<DbgDeclareInst *, 4> Dbgs;
1189 for (Instruction &BI : FI)
1190 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1191 Dbgs.push_back(DDI);
1196 for (auto &I : Dbgs) {
1197 DbgDeclareInst *DDI = I;
1198 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1199 // If this is an alloca for a scalar variable, insert a dbg.value
1200 // at each load and store to the alloca and erase the dbg.declare.
1201 // The dbg.values allow tracking a variable even if it is not
1202 // stored on the stack, while the dbg.declare can only describe
1203 // the stack slot (and at a lexical-scope granularity). Later
1204 // passes will attempt to elide the stack slot.
1205 if (AI && !isArray(AI)) {
1206 for (auto &AIUse : AI->uses()) {
1207 User *U = AIUse.getUser();
1208 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1209 if (AIUse.getOperandNo() == 1)
1210 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1211 } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1212 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1213 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1214 // This is a call by-value or some other instruction that
1215 // takes a pointer to the variable. Insert a *value*
1216 // intrinsic that describes the alloca.
1217 SmallVector<uint64_t, 1> NewDIExpr;
1218 auto *DIExpr = DDI->getExpression();
1219 NewDIExpr.push_back(dwarf::DW_OP_deref);
1220 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1221 DIB.insertDbgValueIntrinsic(AI, 0, DDI->getVariable(),
1222 DIB.createExpression(NewDIExpr),
1223 DDI->getDebugLoc(), CI);
1226 DDI->eraseFromParent();
1232 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
1233 /// alloca 'V', if any.
1234 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) {
1235 if (auto *L = LocalAsMetadata::getIfExists(V))
1236 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1237 for (User *U : MDV->users())
1238 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1244 /// FindAllocaDbgValues - Finds the llvm.dbg.value intrinsics describing the
1245 /// alloca 'V', if any.
1246 void llvm::FindAllocaDbgValues(DbgValueList &DbgValues, Value *V) {
1247 if (auto *L = LocalAsMetadata::getIfExists(V))
1248 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1249 for (User *U : MDV->users())
1250 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1251 DbgValues.push_back(DVI);
1254 static void DIExprAddDeref(SmallVectorImpl<uint64_t> &Expr) {
1255 Expr.push_back(dwarf::DW_OP_deref);
1258 static void DIExprAddOffset(SmallVectorImpl<uint64_t> &Expr, int Offset) {
1260 Expr.push_back(dwarf::DW_OP_plus);
1261 Expr.push_back(Offset);
1262 } else if (Offset < 0) {
1263 Expr.push_back(dwarf::DW_OP_minus);
1264 Expr.push_back(-Offset);
1268 static DIExpression *BuildReplacementDIExpr(DIBuilder &Builder,
1269 DIExpression *DIExpr, bool Deref,
1271 if (!Deref && !Offset)
1273 // Create a copy of the original DIDescriptor for user variable, prepending
1274 // "deref" operation to a list of address elements, as new llvm.dbg.declare
1275 // will take a value storing address of the memory for variable, not
1277 SmallVector<uint64_t, 4> NewDIExpr;
1279 DIExprAddDeref(NewDIExpr);
1280 DIExprAddOffset(NewDIExpr, Offset);
1282 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1283 return Builder.createExpression(NewDIExpr);
1286 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1287 Instruction *InsertBefore, DIBuilder &Builder,
1288 bool Deref, int Offset) {
1289 DbgDeclareInst *DDI = FindAllocaDbgDeclare(Address);
1292 DebugLoc Loc = DDI->getDebugLoc();
1293 auto *DIVar = DDI->getVariable();
1294 auto *DIExpr = DDI->getExpression();
1295 assert(DIVar && "Missing variable");
1297 DIExpr = BuildReplacementDIExpr(Builder, DIExpr, Deref, Offset);
1299 // Insert llvm.dbg.declare immediately after the original alloca, and remove
1300 // old llvm.dbg.declare.
1301 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1302 DDI->eraseFromParent();
1306 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1307 DIBuilder &Builder, bool Deref, int Offset) {
1308 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1312 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress,
1313 DIBuilder &Builder, int Offset) {
1314 DebugLoc Loc = DVI->getDebugLoc();
1315 auto *DIVar = DVI->getVariable();
1316 auto *DIExpr = DVI->getExpression();
1317 assert(DIVar && "Missing variable");
1319 // This is an alloca-based llvm.dbg.value. The first thing it should do with
1320 // the alloca pointer is dereference it. Otherwise we don't know how to handle
1322 if (!DIExpr || DIExpr->getNumElements() < 1 ||
1323 DIExpr->getElement(0) != dwarf::DW_OP_deref)
1326 // Insert the offset immediately after the first deref.
1327 // We could just change the offset argument of dbg.value, but it's unsigned...
1329 SmallVector<uint64_t, 4> NewDIExpr;
1330 DIExprAddDeref(NewDIExpr);
1331 DIExprAddOffset(NewDIExpr, Offset);
1332 NewDIExpr.append(DIExpr->elements_begin() + 1, DIExpr->elements_end());
1333 DIExpr = Builder.createExpression(NewDIExpr);
1336 Builder.insertDbgValueIntrinsic(NewAddress, DVI->getOffset(), DIVar, DIExpr,
1338 DVI->eraseFromParent();
1341 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1342 DIBuilder &Builder, int Offset) {
1343 if (auto *L = LocalAsMetadata::getIfExists(AI))
1344 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
1345 for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) {
1347 if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser()))
1348 replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset);
1352 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) {
1353 unsigned NumDeadInst = 0;
1354 // Delete the instructions backwards, as it has a reduced likelihood of
1355 // having to update as many def-use and use-def chains.
1356 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
1357 while (EndInst != &BB->front()) {
1358 // Delete the next to last instruction.
1359 Instruction *Inst = &*--EndInst->getIterator();
1360 if (!Inst->use_empty() && !Inst->getType()->isTokenTy())
1361 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
1362 if (Inst->isEHPad() || Inst->getType()->isTokenTy()) {
1366 if (!isa<DbgInfoIntrinsic>(Inst))
1368 Inst->eraseFromParent();
1373 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap,
1374 bool PreserveLCSSA) {
1375 BasicBlock *BB = I->getParent();
1376 // Loop over all of the successors, removing BB's entry from any PHI
1378 for (BasicBlock *Successor : successors(BB))
1379 Successor->removePredecessor(BB, PreserveLCSSA);
1381 // Insert a call to llvm.trap right before this. This turns the undefined
1382 // behavior into a hard fail instead of falling through into random code.
1385 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1386 CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1387 CallTrap->setDebugLoc(I->getDebugLoc());
1389 new UnreachableInst(I->getContext(), I);
1391 // All instructions after this are dead.
1392 unsigned NumInstrsRemoved = 0;
1393 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1394 while (BBI != BBE) {
1395 if (!BBI->use_empty())
1396 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1397 BB->getInstList().erase(BBI++);
1400 return NumInstrsRemoved;
1403 /// changeToCall - Convert the specified invoke into a normal call.
1404 static void changeToCall(InvokeInst *II) {
1405 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1406 SmallVector<OperandBundleDef, 1> OpBundles;
1407 II->getOperandBundlesAsDefs(OpBundles);
1408 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1410 NewCall->takeName(II);
1411 NewCall->setCallingConv(II->getCallingConv());
1412 NewCall->setAttributes(II->getAttributes());
1413 NewCall->setDebugLoc(II->getDebugLoc());
1414 II->replaceAllUsesWith(NewCall);
1416 // Follow the call by a branch to the normal destination.
1417 BranchInst::Create(II->getNormalDest(), II);
1419 // Update PHI nodes in the unwind destination
1420 II->getUnwindDest()->removePredecessor(II->getParent());
1421 II->eraseFromParent();
1424 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI,
1425 BasicBlock *UnwindEdge) {
1426 BasicBlock *BB = CI->getParent();
1428 // Convert this function call into an invoke instruction. First, split the
1431 BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc");
1433 // Delete the unconditional branch inserted by splitBasicBlock
1434 BB->getInstList().pop_back();
1436 // Create the new invoke instruction.
1437 SmallVector<Value *, 8> InvokeArgs(CI->arg_begin(), CI->arg_end());
1438 SmallVector<OperandBundleDef, 1> OpBundles;
1440 CI->getOperandBundlesAsDefs(OpBundles);
1442 // Note: we're round tripping operand bundles through memory here, and that
1443 // can potentially be avoided with a cleverer API design that we do not have
1446 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge,
1447 InvokeArgs, OpBundles, CI->getName(), BB);
1448 II->setDebugLoc(CI->getDebugLoc());
1449 II->setCallingConv(CI->getCallingConv());
1450 II->setAttributes(CI->getAttributes());
1452 // Make sure that anything using the call now uses the invoke! This also
1453 // updates the CallGraph if present, because it uses a WeakVH.
1454 CI->replaceAllUsesWith(II);
1456 // Delete the original call
1457 Split->getInstList().pop_front();
1461 static bool markAliveBlocks(Function &F,
1462 SmallPtrSetImpl<BasicBlock*> &Reachable) {
1464 SmallVector<BasicBlock*, 128> Worklist;
1465 BasicBlock *BB = &F.front();
1466 Worklist.push_back(BB);
1467 Reachable.insert(BB);
1468 bool Changed = false;
1470 BB = Worklist.pop_back_val();
1472 // Do a quick scan of the basic block, turning any obviously unreachable
1473 // instructions into LLVM unreachable insts. The instruction combining pass
1474 // canonicalizes unreachable insts into stores to null or undef.
1475 for (Instruction &I : *BB) {
1476 // Assumptions that are known to be false are equivalent to unreachable.
1477 // Also, if the condition is undefined, then we make the choice most
1478 // beneficial to the optimizer, and choose that to also be unreachable.
1479 if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1480 if (II->getIntrinsicID() == Intrinsic::assume) {
1481 if (match(II->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) {
1482 // Don't insert a call to llvm.trap right before the unreachable.
1483 changeToUnreachable(II, false);
1489 if (II->getIntrinsicID() == Intrinsic::experimental_guard) {
1490 // A call to the guard intrinsic bails out of the current compilation
1491 // unit if the predicate passed to it is false. If the predicate is a
1492 // constant false, then we know the guard will bail out of the current
1493 // compile unconditionally, so all code following it is dead.
1495 // Note: unlike in llvm.assume, it is not "obviously profitable" for
1496 // guards to treat `undef` as `false` since a guard on `undef` can
1497 // still be useful for widening.
1498 if (match(II->getArgOperand(0), m_Zero()))
1499 if (!isa<UnreachableInst>(II->getNextNode())) {
1500 changeToUnreachable(II->getNextNode(), /*UseLLVMTrap=*/ false);
1507 if (auto *CI = dyn_cast<CallInst>(&I)) {
1508 Value *Callee = CI->getCalledValue();
1509 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1510 changeToUnreachable(CI, /*UseLLVMTrap=*/false);
1514 if (CI->doesNotReturn()) {
1515 // If we found a call to a no-return function, insert an unreachable
1516 // instruction after it. Make sure there isn't *already* one there
1518 if (!isa<UnreachableInst>(CI->getNextNode())) {
1519 // Don't insert a call to llvm.trap right before the unreachable.
1520 changeToUnreachable(CI->getNextNode(), false);
1527 // Store to undef and store to null are undefined and used to signal that
1528 // they should be changed to unreachable by passes that can't modify the
1530 if (auto *SI = dyn_cast<StoreInst>(&I)) {
1531 // Don't touch volatile stores.
1532 if (SI->isVolatile()) continue;
1534 Value *Ptr = SI->getOperand(1);
1536 if (isa<UndefValue>(Ptr) ||
1537 (isa<ConstantPointerNull>(Ptr) &&
1538 SI->getPointerAddressSpace() == 0)) {
1539 changeToUnreachable(SI, true);
1546 TerminatorInst *Terminator = BB->getTerminator();
1547 if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
1548 // Turn invokes that call 'nounwind' functions into ordinary calls.
1549 Value *Callee = II->getCalledValue();
1550 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1551 changeToUnreachable(II, true);
1553 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
1554 if (II->use_empty() && II->onlyReadsMemory()) {
1555 // jump to the normal destination branch.
1556 BranchInst::Create(II->getNormalDest(), II);
1557 II->getUnwindDest()->removePredecessor(II->getParent());
1558 II->eraseFromParent();
1563 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
1564 // Remove catchpads which cannot be reached.
1565 struct CatchPadDenseMapInfo {
1566 static CatchPadInst *getEmptyKey() {
1567 return DenseMapInfo<CatchPadInst *>::getEmptyKey();
1569 static CatchPadInst *getTombstoneKey() {
1570 return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
1572 static unsigned getHashValue(CatchPadInst *CatchPad) {
1573 return static_cast<unsigned>(hash_combine_range(
1574 CatchPad->value_op_begin(), CatchPad->value_op_end()));
1576 static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
1577 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1578 RHS == getEmptyKey() || RHS == getTombstoneKey())
1580 return LHS->isIdenticalTo(RHS);
1584 // Set of unique CatchPads.
1585 SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
1586 CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
1588 detail::DenseSetEmpty Empty;
1589 for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
1590 E = CatchSwitch->handler_end();
1592 BasicBlock *HandlerBB = *I;
1593 auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
1594 if (!HandlerSet.insert({CatchPad, Empty}).second) {
1595 CatchSwitch->removeHandler(I);
1603 Changed |= ConstantFoldTerminator(BB, true);
1604 for (BasicBlock *Successor : successors(BB))
1605 if (Reachable.insert(Successor).second)
1606 Worklist.push_back(Successor);
1607 } while (!Worklist.empty());
1611 void llvm::removeUnwindEdge(BasicBlock *BB) {
1612 TerminatorInst *TI = BB->getTerminator();
1614 if (auto *II = dyn_cast<InvokeInst>(TI)) {
1619 TerminatorInst *NewTI;
1620 BasicBlock *UnwindDest;
1622 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
1623 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
1624 UnwindDest = CRI->getUnwindDest();
1625 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
1626 auto *NewCatchSwitch = CatchSwitchInst::Create(
1627 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
1628 CatchSwitch->getName(), CatchSwitch);
1629 for (BasicBlock *PadBB : CatchSwitch->handlers())
1630 NewCatchSwitch->addHandler(PadBB);
1632 NewTI = NewCatchSwitch;
1633 UnwindDest = CatchSwitch->getUnwindDest();
1635 llvm_unreachable("Could not find unwind successor");
1638 NewTI->takeName(TI);
1639 NewTI->setDebugLoc(TI->getDebugLoc());
1640 UnwindDest->removePredecessor(BB);
1641 TI->replaceAllUsesWith(NewTI);
1642 TI->eraseFromParent();
1645 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even
1646 /// if they are in a dead cycle. Return true if a change was made, false
1648 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI) {
1649 SmallPtrSet<BasicBlock*, 16> Reachable;
1650 bool Changed = markAliveBlocks(F, Reachable);
1652 // If there are unreachable blocks in the CFG...
1653 if (Reachable.size() == F.size())
1656 assert(Reachable.size() < F.size());
1657 NumRemoved += F.size()-Reachable.size();
1659 // Loop over all of the basic blocks that are not reachable, dropping all of
1660 // their internal references...
1661 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) {
1662 if (Reachable.count(&*BB))
1665 for (BasicBlock *Successor : successors(&*BB))
1666 if (Reachable.count(Successor))
1667 Successor->removePredecessor(&*BB);
1669 LVI->eraseBlock(&*BB);
1670 BB->dropAllReferences();
1673 for (Function::iterator I = ++F.begin(); I != F.end();)
1674 if (!Reachable.count(&*I))
1675 I = F.getBasicBlockList().erase(I);
1682 void llvm::combineMetadata(Instruction *K, const Instruction *J,
1683 ArrayRef<unsigned> KnownIDs) {
1684 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1685 K->dropUnknownNonDebugMetadata(KnownIDs);
1686 K->getAllMetadataOtherThanDebugLoc(Metadata);
1687 for (const auto &MD : Metadata) {
1688 unsigned Kind = MD.first;
1689 MDNode *JMD = J->getMetadata(Kind);
1690 MDNode *KMD = MD.second;
1694 K->setMetadata(Kind, nullptr); // Remove unknown metadata
1696 case LLVMContext::MD_dbg:
1697 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1698 case LLVMContext::MD_tbaa:
1699 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1701 case LLVMContext::MD_alias_scope:
1702 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1704 case LLVMContext::MD_noalias:
1705 case LLVMContext::MD_mem_parallel_loop_access:
1706 K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1708 case LLVMContext::MD_range:
1709 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1711 case LLVMContext::MD_fpmath:
1712 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1714 case LLVMContext::MD_invariant_load:
1715 // Only set the !invariant.load if it is present in both instructions.
1716 K->setMetadata(Kind, JMD);
1718 case LLVMContext::MD_nonnull:
1719 // Only set the !nonnull if it is present in both instructions.
1720 K->setMetadata(Kind, JMD);
1722 case LLVMContext::MD_invariant_group:
1723 // Preserve !invariant.group in K.
1725 case LLVMContext::MD_align:
1726 K->setMetadata(Kind,
1727 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1729 case LLVMContext::MD_dereferenceable:
1730 case LLVMContext::MD_dereferenceable_or_null:
1731 K->setMetadata(Kind,
1732 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1736 // Set !invariant.group from J if J has it. If both instructions have it
1737 // then we will just pick it from J - even when they are different.
1738 // Also make sure that K is load or store - f.e. combining bitcast with load
1739 // could produce bitcast with invariant.group metadata, which is invalid.
1740 // FIXME: we should try to preserve both invariant.group md if they are
1741 // different, but right now instruction can only have one invariant.group.
1742 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
1743 if (isa<LoadInst>(K) || isa<StoreInst>(K))
1744 K->setMetadata(LLVMContext::MD_invariant_group, JMD);
1747 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J) {
1748 unsigned KnownIDs[] = {
1749 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1750 LLVMContext::MD_noalias, LLVMContext::MD_range,
1751 LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull,
1752 LLVMContext::MD_invariant_group, LLVMContext::MD_align,
1753 LLVMContext::MD_dereferenceable,
1754 LLVMContext::MD_dereferenceable_or_null};
1755 combineMetadata(K, J, KnownIDs);
1758 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1760 const BasicBlockEdge &Root) {
1761 assert(From->getType() == To->getType());
1764 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1767 if (DT.dominates(Root, U)) {
1769 DEBUG(dbgs() << "Replace dominated use of '"
1770 << From->getName() << "' as "
1771 << *To << " in " << *U << "\n");
1778 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1780 const BasicBlock *BB) {
1781 assert(From->getType() == To->getType());
1784 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1787 auto *I = cast<Instruction>(U.getUser());
1788 if (DT.properlyDominates(BB, I->getParent())) {
1790 DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as "
1791 << *To << " in " << *U << "\n");
1798 bool llvm::callsGCLeafFunction(ImmutableCallSite CS) {
1799 // Check if the function is specifically marked as a gc leaf function.
1800 if (CS.hasFnAttr("gc-leaf-function"))
1802 if (const Function *F = CS.getCalledFunction()) {
1803 if (F->hasFnAttribute("gc-leaf-function"))
1806 if (auto IID = F->getIntrinsicID())
1807 // Most LLVM intrinsics do not take safepoints.
1808 return IID != Intrinsic::experimental_gc_statepoint &&
1809 IID != Intrinsic::experimental_deoptimize;
1816 /// A potential constituent of a bitreverse or bswap expression. See
1817 /// collectBitParts for a fuller explanation.
1819 BitPart(Value *P, unsigned BW) : Provider(P) {
1820 Provenance.resize(BW);
1823 /// The Value that this is a bitreverse/bswap of.
1825 /// The "provenance" of each bit. Provenance[A] = B means that bit A
1826 /// in Provider becomes bit B in the result of this expression.
1827 SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128.
1829 enum { Unset = -1 };
1831 } // end anonymous namespace
1833 /// Analyze the specified subexpression and see if it is capable of providing
1834 /// pieces of a bswap or bitreverse. The subexpression provides a potential
1835 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in
1836 /// the output of the expression came from a corresponding bit in some other
1837 /// value. This function is recursive, and the end result is a mapping of
1838 /// bitnumber to bitnumber. It is the caller's responsibility to validate that
1839 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse.
1841 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know
1842 /// that the expression deposits the low byte of %X into the high byte of the
1843 /// result and that all other bits are zero. This expression is accepted and a
1844 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to
1847 /// To avoid revisiting values, the BitPart results are memoized into the
1848 /// provided map. To avoid unnecessary copying of BitParts, BitParts are
1849 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to
1850 /// store BitParts objects, not pointers. As we need the concept of a nullptr
1851 /// BitParts (Value has been analyzed and the analysis failed), we an Optional
1852 /// type instead to provide the same functionality.
1854 /// Because we pass around references into \c BPS, we must use a container that
1855 /// does not invalidate internal references (std::map instead of DenseMap).
1857 static const Optional<BitPart> &
1858 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
1859 std::map<Value *, Optional<BitPart>> &BPS) {
1860 auto I = BPS.find(V);
1864 auto &Result = BPS[V] = None;
1865 auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1867 if (Instruction *I = dyn_cast<Instruction>(V)) {
1868 // If this is an or instruction, it may be an inner node of the bswap.
1869 if (I->getOpcode() == Instruction::Or) {
1870 auto &A = collectBitParts(I->getOperand(0), MatchBSwaps,
1871 MatchBitReversals, BPS);
1872 auto &B = collectBitParts(I->getOperand(1), MatchBSwaps,
1873 MatchBitReversals, BPS);
1877 // Try and merge the two together.
1878 if (!A->Provider || A->Provider != B->Provider)
1881 Result = BitPart(A->Provider, BitWidth);
1882 for (unsigned i = 0; i < A->Provenance.size(); ++i) {
1883 if (A->Provenance[i] != BitPart::Unset &&
1884 B->Provenance[i] != BitPart::Unset &&
1885 A->Provenance[i] != B->Provenance[i])
1886 return Result = None;
1888 if (A->Provenance[i] == BitPart::Unset)
1889 Result->Provenance[i] = B->Provenance[i];
1891 Result->Provenance[i] = A->Provenance[i];
1897 // If this is a logical shift by a constant, recurse then shift the result.
1898 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1900 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1901 // Ensure the shift amount is defined.
1902 if (BitShift > BitWidth)
1905 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1906 MatchBitReversals, BPS);
1911 // Perform the "shift" on BitProvenance.
1912 auto &P = Result->Provenance;
1913 if (I->getOpcode() == Instruction::Shl) {
1914 P.erase(std::prev(P.end(), BitShift), P.end());
1915 P.insert(P.begin(), BitShift, BitPart::Unset);
1917 P.erase(P.begin(), std::next(P.begin(), BitShift));
1918 P.insert(P.end(), BitShift, BitPart::Unset);
1924 // If this is a logical 'and' with a mask that clears bits, recurse then
1925 // unset the appropriate bits.
1926 if (I->getOpcode() == Instruction::And &&
1927 isa<ConstantInt>(I->getOperand(1))) {
1928 APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1);
1929 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1931 // Check that the mask allows a multiple of 8 bits for a bswap, for an
1933 unsigned NumMaskedBits = AndMask.countPopulation();
1934 if (!MatchBitReversals && NumMaskedBits % 8 != 0)
1937 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1938 MatchBitReversals, BPS);
1943 for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1)
1944 // If the AndMask is zero for this bit, clear the bit.
1945 if ((AndMask & Bit) == 0)
1946 Result->Provenance[i] = BitPart::Unset;
1950 // If this is a zext instruction zero extend the result.
1951 if (I->getOpcode() == Instruction::ZExt) {
1952 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps,
1953 MatchBitReversals, BPS);
1957 Result = BitPart(Res->Provider, BitWidth);
1958 auto NarrowBitWidth =
1959 cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth();
1960 for (unsigned i = 0; i < NarrowBitWidth; ++i)
1961 Result->Provenance[i] = Res->Provenance[i];
1962 for (unsigned i = NarrowBitWidth; i < BitWidth; ++i)
1963 Result->Provenance[i] = BitPart::Unset;
1968 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1969 // the input value to the bswap/bitreverse.
1970 Result = BitPart(V, BitWidth);
1971 for (unsigned i = 0; i < BitWidth; ++i)
1972 Result->Provenance[i] = i;
1976 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To,
1977 unsigned BitWidth) {
1978 if (From % 8 != To % 8)
1980 // Convert from bit indices to byte indices and check for a byte reversal.
1984 return From == BitWidth - To - 1;
1987 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To,
1988 unsigned BitWidth) {
1989 return From == BitWidth - To - 1;
1992 /// Given an OR instruction, check to see if this is a bitreverse
1993 /// idiom. If so, insert the new intrinsic and return true.
1994 bool llvm::recognizeBSwapOrBitReverseIdiom(
1995 Instruction *I, bool MatchBSwaps, bool MatchBitReversals,
1996 SmallVectorImpl<Instruction *> &InsertedInsts) {
1997 if (Operator::getOpcode(I) != Instruction::Or)
1999 if (!MatchBSwaps && !MatchBitReversals)
2001 IntegerType *ITy = dyn_cast<IntegerType>(I->getType());
2002 if (!ITy || ITy->getBitWidth() > 128)
2003 return false; // Can't do vectors or integers > 128 bits.
2004 unsigned BW = ITy->getBitWidth();
2006 unsigned DemandedBW = BW;
2007 IntegerType *DemandedTy = ITy;
2008 if (I->hasOneUse()) {
2009 if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) {
2010 DemandedTy = cast<IntegerType>(Trunc->getType());
2011 DemandedBW = DemandedTy->getBitWidth();
2015 // Try to find all the pieces corresponding to the bswap.
2016 std::map<Value *, Optional<BitPart>> BPS;
2017 auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS);
2020 auto &BitProvenance = Res->Provenance;
2022 // Now, is the bit permutation correct for a bswap or a bitreverse? We can
2023 // only byteswap values with an even number of bytes.
2024 bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true;
2025 for (unsigned i = 0; i < DemandedBW; ++i) {
2027 bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW);
2029 bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW);
2032 Intrinsic::ID Intrin;
2033 if (OKForBSwap && MatchBSwaps)
2034 Intrin = Intrinsic::bswap;
2035 else if (OKForBitReverse && MatchBitReversals)
2036 Intrin = Intrinsic::bitreverse;
2040 if (ITy != DemandedTy) {
2041 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy);
2042 Value *Provider = Res->Provider;
2043 IntegerType *ProviderTy = cast<IntegerType>(Provider->getType());
2044 // We may need to truncate the provider.
2045 if (DemandedTy != ProviderTy) {
2046 auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy,
2048 InsertedInsts.push_back(Trunc);
2051 auto *CI = CallInst::Create(F, Provider, "rev", I);
2052 InsertedInsts.push_back(CI);
2053 auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I);
2054 InsertedInsts.push_back(ExtInst);
2058 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy);
2059 InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I));
2063 // CodeGen has special handling for some string functions that may replace
2064 // them with target-specific intrinsics. Since that'd skip our interceptors
2065 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses,
2066 // we mark affected calls as NoBuiltin, which will disable optimization
2068 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin(
2069 CallInst *CI, const TargetLibraryInfo *TLI) {
2070 Function *F = CI->getCalledFunction();
2072 if (F && !F->hasLocalLinkage() && F->hasName() &&
2073 TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) &&
2074 !F->doesNotAccessMemory())
2075 CI->addAttribute(AttributeSet::FunctionIndex, Attribute::NoBuiltin);