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/ValueHandle.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
51 #define DEBUG_TYPE "local"
53 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
55 //===----------------------------------------------------------------------===//
56 // Local constant propagation.
59 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
60 /// constant value, convert it into an unconditional branch to the constant
61 /// destination. This is a nontrivial operation because the successors of this
62 /// basic block must have their PHI nodes updated.
63 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
64 /// conditions and indirectbr addresses this might make dead if
65 /// DeleteDeadConditions is true.
66 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
67 const TargetLibraryInfo *TLI) {
68 TerminatorInst *T = BB->getTerminator();
69 IRBuilder<> Builder(T);
71 // Branch - See if we are conditional jumping on constant
72 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
73 if (BI->isUnconditional()) return false; // Can't optimize uncond branch
74 BasicBlock *Dest1 = BI->getSuccessor(0);
75 BasicBlock *Dest2 = BI->getSuccessor(1);
77 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
78 // Are we branching on constant?
79 // YES. Change to unconditional branch...
80 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
81 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1;
83 //cerr << "Function: " << T->getParent()->getParent()
84 // << "\nRemoving branch from " << T->getParent()
85 // << "\n\nTo: " << OldDest << endl;
87 // Let the basic block know that we are letting go of it. Based on this,
88 // it will adjust it's PHI nodes.
89 OldDest->removePredecessor(BB);
91 // Replace the conditional branch with an unconditional one.
92 Builder.CreateBr(Destination);
93 BI->eraseFromParent();
97 if (Dest2 == Dest1) { // Conditional branch to same location?
98 // This branch matches something like this:
99 // br bool %cond, label %Dest, label %Dest
100 // and changes it into: br label %Dest
102 // Let the basic block know that we are letting go of one copy of it.
103 assert(BI->getParent() && "Terminator not inserted in block!");
104 Dest1->removePredecessor(BI->getParent());
106 // Replace the conditional branch with an unconditional one.
107 Builder.CreateBr(Dest1);
108 Value *Cond = BI->getCondition();
109 BI->eraseFromParent();
110 if (DeleteDeadConditions)
111 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
117 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) {
118 // If we are switching on a constant, we can convert the switch to an
119 // unconditional branch.
120 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition());
121 BasicBlock *DefaultDest = SI->getDefaultDest();
122 BasicBlock *TheOnlyDest = DefaultDest;
124 // If the default is unreachable, ignore it when searching for TheOnlyDest.
125 if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
126 SI->getNumCases() > 0) {
127 TheOnlyDest = SI->case_begin().getCaseSuccessor();
130 // Figure out which case it goes to.
131 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
133 // Found case matching a constant operand?
134 if (i.getCaseValue() == CI) {
135 TheOnlyDest = i.getCaseSuccessor();
139 // Check to see if this branch is going to the same place as the default
140 // dest. If so, eliminate it as an explicit compare.
141 if (i.getCaseSuccessor() == DefaultDest) {
142 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
143 unsigned NCases = SI->getNumCases();
144 // Fold the case metadata into the default if there will be any branches
145 // left, unless the metadata doesn't match the switch.
146 if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) {
147 // Collect branch weights into a vector.
148 SmallVector<uint32_t, 8> Weights;
149 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
152 mdconst::dyn_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.
325 if (II->getIntrinsicID() == Intrinsic::assume) {
326 if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
327 return !Cond->isZero();
333 if (isAllocLikeFn(I, TLI)) return true;
335 if (CallInst *CI = isFreeCall(I, TLI))
336 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
337 return C->isNullValue() || isa<UndefValue>(C);
342 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
343 /// trivially dead instruction, delete it. If that makes any of its operands
344 /// trivially dead, delete them too, recursively. Return true if any
345 /// instructions were deleted.
347 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
348 const TargetLibraryInfo *TLI) {
349 Instruction *I = dyn_cast<Instruction>(V);
350 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
353 SmallVector<Instruction*, 16> DeadInsts;
354 DeadInsts.push_back(I);
357 I = DeadInsts.pop_back_val();
359 // Null out all of the instruction's operands to see if any operand becomes
361 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
362 Value *OpV = I->getOperand(i);
363 I->setOperand(i, nullptr);
365 if (!OpV->use_empty()) continue;
367 // If the operand is an instruction that became dead as we nulled out the
368 // operand, and if it is 'trivially' dead, delete it in a future loop
370 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
371 if (isInstructionTriviallyDead(OpI, TLI))
372 DeadInsts.push_back(OpI);
375 I->eraseFromParent();
376 } while (!DeadInsts.empty());
381 /// areAllUsesEqual - Check whether the uses of a value are all the same.
382 /// This is similar to Instruction::hasOneUse() except this will also return
383 /// true when there are no uses or multiple uses that all refer to the same
385 static bool areAllUsesEqual(Instruction *I) {
386 Value::user_iterator UI = I->user_begin();
387 Value::user_iterator UE = I->user_end();
392 for (++UI; UI != UE; ++UI) {
399 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
400 /// dead PHI node, due to being a def-use chain of single-use nodes that
401 /// either forms a cycle or is terminated by a trivially dead instruction,
402 /// delete it. If that makes any of its operands trivially dead, delete them
403 /// too, recursively. Return true if a change was made.
404 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
405 const TargetLibraryInfo *TLI) {
406 SmallPtrSet<Instruction*, 4> Visited;
407 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
408 I = cast<Instruction>(*I->user_begin())) {
410 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
412 // If we find an instruction more than once, we're on a cycle that
413 // won't prove fruitful.
414 if (!Visited.insert(I).second) {
415 // Break the cycle and delete the instruction and its operands.
416 I->replaceAllUsesWith(UndefValue::get(I->getType()));
417 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
425 simplifyAndDCEInstruction(Instruction *I,
426 SmallSetVector<Instruction *, 16> &WorkList,
427 const DataLayout &DL,
428 const TargetLibraryInfo *TLI) {
429 if (isInstructionTriviallyDead(I, TLI)) {
430 // Null out all of the instruction's operands to see if any operand becomes
432 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
433 Value *OpV = I->getOperand(i);
434 I->setOperand(i, nullptr);
436 if (!OpV->use_empty() || I == OpV)
439 // If the operand is an instruction that became dead as we nulled out the
440 // operand, and if it is 'trivially' dead, delete it in a future loop
442 if (Instruction *OpI = dyn_cast<Instruction>(OpV))
443 if (isInstructionTriviallyDead(OpI, TLI))
444 WorkList.insert(OpI);
447 I->eraseFromParent();
452 if (Value *SimpleV = SimplifyInstruction(I, DL)) {
453 // Add the users to the worklist. CAREFUL: an instruction can use itself,
454 // in the case of a phi node.
455 for (User *U : I->users())
457 WorkList.insert(cast<Instruction>(U));
459 // Replace the instruction with its simplified value.
460 I->replaceAllUsesWith(SimpleV);
461 I->eraseFromParent();
467 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
468 /// simplify any instructions in it and recursively delete dead instructions.
470 /// This returns true if it changed the code, note that it can delete
471 /// instructions in other blocks as well in this block.
472 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
473 const TargetLibraryInfo *TLI) {
474 bool MadeChange = false;
475 const DataLayout &DL = BB->getModule()->getDataLayout();
478 // In debug builds, ensure that the terminator of the block is never replaced
479 // or deleted by these simplifications. The idea of simplification is that it
480 // cannot introduce new instructions, and there is no way to replace the
481 // terminator of a block without introducing a new instruction.
482 AssertingVH<Instruction> TerminatorVH(&BB->back());
485 SmallSetVector<Instruction *, 16> WorkList;
486 // Iterate over the original function, only adding insts to the worklist
487 // if they actually need to be revisited. This avoids having to pre-init
488 // the worklist with the entire function's worth of instructions.
489 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end()); BI != E;) {
490 assert(!BI->isTerminator());
491 Instruction *I = &*BI;
494 // We're visiting this instruction now, so make sure it's not in the
495 // worklist from an earlier visit.
496 if (!WorkList.count(I))
497 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
500 while (!WorkList.empty()) {
501 Instruction *I = WorkList.pop_back_val();
502 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
507 //===----------------------------------------------------------------------===//
508 // Control Flow Graph Restructuring.
512 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
513 /// method is called when we're about to delete Pred as a predecessor of BB. If
514 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
516 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
517 /// nodes that collapse into identity values. For example, if we have:
518 /// x = phi(1, 0, 0, 0)
521 /// .. and delete the predecessor corresponding to the '1', this will attempt to
522 /// recursively fold the and to 0.
523 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) {
524 // This only adjusts blocks with PHI nodes.
525 if (!isa<PHINode>(BB->begin()))
528 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
529 // them down. This will leave us with single entry phi nodes and other phis
530 // that can be removed.
531 BB->removePredecessor(Pred, true);
533 WeakVH PhiIt = &BB->front();
534 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
535 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
536 Value *OldPhiIt = PhiIt;
538 if (!recursivelySimplifyInstruction(PN))
541 // If recursive simplification ended up deleting the next PHI node we would
542 // iterate to, then our iterator is invalid, restart scanning from the top
544 if (PhiIt != OldPhiIt) PhiIt = &BB->front();
549 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
550 /// predecessor is known to have one successor (DestBB!). Eliminate the edge
551 /// between them, moving the instructions in the predecessor into DestBB and
552 /// deleting the predecessor block.
554 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) {
555 // If BB has single-entry PHI nodes, fold them.
556 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
557 Value *NewVal = PN->getIncomingValue(0);
558 // Replace self referencing PHI with undef, it must be dead.
559 if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
560 PN->replaceAllUsesWith(NewVal);
561 PN->eraseFromParent();
564 BasicBlock *PredBB = DestBB->getSinglePredecessor();
565 assert(PredBB && "Block doesn't have a single predecessor!");
567 // Zap anything that took the address of DestBB. Not doing this will give the
568 // address an invalid value.
569 if (DestBB->hasAddressTaken()) {
570 BlockAddress *BA = BlockAddress::get(DestBB);
571 Constant *Replacement =
572 ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1);
573 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
575 BA->destroyConstant();
578 // Anything that branched to PredBB now branches to DestBB.
579 PredBB->replaceAllUsesWith(DestBB);
581 // Splice all the instructions from PredBB to DestBB.
582 PredBB->getTerminator()->eraseFromParent();
583 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
585 // If the PredBB is the entry block of the function, move DestBB up to
586 // become the entry block after we erase PredBB.
587 if (PredBB == &DestBB->getParent()->getEntryBlock())
588 DestBB->moveAfter(PredBB);
591 BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock();
592 DT->changeImmediateDominator(DestBB, PredBBIDom);
593 DT->eraseNode(PredBB);
596 PredBB->eraseFromParent();
599 /// CanMergeValues - Return true if we can choose one of these values to use
600 /// in place of the other. Note that we will always choose the non-undef
602 static bool CanMergeValues(Value *First, Value *Second) {
603 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
606 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
607 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
609 /// Assumption: Succ is the single successor for BB.
611 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
612 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
614 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
615 << Succ->getName() << "\n");
616 // Shortcut, if there is only a single predecessor it must be BB and merging
618 if (Succ->getSinglePredecessor()) return true;
620 // Make a list of the predecessors of BB
621 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
623 // Look at all the phi nodes in Succ, to see if they present a conflict when
624 // merging these blocks
625 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
626 PHINode *PN = cast<PHINode>(I);
628 // If the incoming value from BB is again a PHINode in
629 // BB which has the same incoming value for *PI as PN does, we can
630 // merge the phi nodes and then the blocks can still be merged
631 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
632 if (BBPN && BBPN->getParent() == BB) {
633 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
634 BasicBlock *IBB = PN->getIncomingBlock(PI);
635 if (BBPreds.count(IBB) &&
636 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
637 PN->getIncomingValue(PI))) {
638 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
639 << Succ->getName() << " is conflicting with "
640 << BBPN->getName() << " with regard to common predecessor "
641 << IBB->getName() << "\n");
646 Value* Val = PN->getIncomingValueForBlock(BB);
647 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
648 // See if the incoming value for the common predecessor is equal to the
649 // one for BB, in which case this phi node will not prevent the merging
651 BasicBlock *IBB = PN->getIncomingBlock(PI);
652 if (BBPreds.count(IBB) &&
653 !CanMergeValues(Val, PN->getIncomingValue(PI))) {
654 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
655 << Succ->getName() << " is conflicting with regard to common "
656 << "predecessor " << IBB->getName() << "\n");
666 typedef SmallVector<BasicBlock *, 16> PredBlockVector;
667 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap;
669 /// \brief Determines the value to use as the phi node input for a block.
671 /// Select between \p OldVal any value that we know flows from \p BB
672 /// to a particular phi on the basis of which one (if either) is not
673 /// undef. Update IncomingValues based on the selected value.
675 /// \param OldVal The value we are considering selecting.
676 /// \param BB The block that the value flows in from.
677 /// \param IncomingValues A map from block-to-value for other phi inputs
678 /// that we have examined.
680 /// \returns the selected value.
681 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
682 IncomingValueMap &IncomingValues) {
683 if (!isa<UndefValue>(OldVal)) {
684 assert((!IncomingValues.count(BB) ||
685 IncomingValues.find(BB)->second == OldVal) &&
686 "Expected OldVal to match incoming value from BB!");
688 IncomingValues.insert(std::make_pair(BB, OldVal));
692 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
693 if (It != IncomingValues.end()) return It->second;
698 /// \brief Create a map from block to value for the operands of a
701 /// Create a map from block to value for each non-undef value flowing
704 /// \param PN The phi we are collecting the map for.
705 /// \param IncomingValues [out] The map from block to value for this phi.
706 static void gatherIncomingValuesToPhi(PHINode *PN,
707 IncomingValueMap &IncomingValues) {
708 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
709 BasicBlock *BB = PN->getIncomingBlock(i);
710 Value *V = PN->getIncomingValue(i);
712 if (!isa<UndefValue>(V))
713 IncomingValues.insert(std::make_pair(BB, V));
717 /// \brief Replace the incoming undef values to a phi with the values
718 /// from a block-to-value map.
720 /// \param PN The phi we are replacing the undefs in.
721 /// \param IncomingValues A map from block to value.
722 static void replaceUndefValuesInPhi(PHINode *PN,
723 const IncomingValueMap &IncomingValues) {
724 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
725 Value *V = PN->getIncomingValue(i);
727 if (!isa<UndefValue>(V)) continue;
729 BasicBlock *BB = PN->getIncomingBlock(i);
730 IncomingValueMap::const_iterator It = IncomingValues.find(BB);
731 if (It == IncomingValues.end()) continue;
733 PN->setIncomingValue(i, It->second);
737 /// \brief Replace a value flowing from a block to a phi with
738 /// potentially multiple instances of that value flowing from the
739 /// block's predecessors to the phi.
741 /// \param BB The block with the value flowing into the phi.
742 /// \param BBPreds The predecessors of BB.
743 /// \param PN The phi that we are updating.
744 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
745 const PredBlockVector &BBPreds,
747 Value *OldVal = PN->removeIncomingValue(BB, false);
748 assert(OldVal && "No entry in PHI for Pred BB!");
750 IncomingValueMap IncomingValues;
752 // We are merging two blocks - BB, and the block containing PN - and
753 // as a result we need to redirect edges from the predecessors of BB
754 // to go to the block containing PN, and update PN
755 // accordingly. Since we allow merging blocks in the case where the
756 // predecessor and successor blocks both share some predecessors,
757 // and where some of those common predecessors might have undef
758 // values flowing into PN, we want to rewrite those values to be
759 // consistent with the non-undef values.
761 gatherIncomingValuesToPhi(PN, IncomingValues);
763 // If this incoming value is one of the PHI nodes in BB, the new entries
764 // in the PHI node are the entries from the old PHI.
765 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
766 PHINode *OldValPN = cast<PHINode>(OldVal);
767 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
768 // Note that, since we are merging phi nodes and BB and Succ might
769 // have common predecessors, we could end up with a phi node with
770 // identical incoming branches. This will be cleaned up later (and
771 // will trigger asserts if we try to clean it up now, without also
772 // simplifying the corresponding conditional branch).
773 BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
774 Value *PredVal = OldValPN->getIncomingValue(i);
775 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
778 // And add a new incoming value for this predecessor for the
779 // newly retargeted branch.
780 PN->addIncoming(Selected, PredBB);
783 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
784 // Update existing incoming values in PN for this
785 // predecessor of BB.
786 BasicBlock *PredBB = BBPreds[i];
787 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
790 // And add a new incoming value for this predecessor for the
791 // newly retargeted branch.
792 PN->addIncoming(Selected, PredBB);
796 replaceUndefValuesInPhi(PN, IncomingValues);
799 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
800 /// unconditional branch, and contains no instructions other than PHI nodes,
801 /// potential side-effect free intrinsics and the branch. If possible,
802 /// eliminate BB by rewriting all the predecessors to branch to the successor
803 /// block and return true. If we can't transform, return false.
804 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) {
805 assert(BB != &BB->getParent()->getEntryBlock() &&
806 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
808 // We can't eliminate infinite loops.
809 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
810 if (BB == Succ) return false;
812 // Check to see if merging these blocks would cause conflicts for any of the
813 // phi nodes in BB or Succ. If not, we can safely merge.
814 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
816 // Check for cases where Succ has multiple predecessors and a PHI node in BB
817 // has uses which will not disappear when the PHI nodes are merged. It is
818 // possible to handle such cases, but difficult: it requires checking whether
819 // BB dominates Succ, which is non-trivial to calculate in the case where
820 // Succ has multiple predecessors. Also, it requires checking whether
821 // constructing the necessary self-referential PHI node doesn't introduce any
822 // conflicts; this isn't too difficult, but the previous code for doing this
825 // Note that if this check finds a live use, BB dominates Succ, so BB is
826 // something like a loop pre-header (or rarely, a part of an irreducible CFG);
827 // folding the branch isn't profitable in that case anyway.
828 if (!Succ->getSinglePredecessor()) {
829 BasicBlock::iterator BBI = BB->begin();
830 while (isa<PHINode>(*BBI)) {
831 for (Use &U : BBI->uses()) {
832 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
833 if (PN->getIncomingBlock(U) != BB)
843 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
845 if (isa<PHINode>(Succ->begin())) {
846 // If there is more than one pred of succ, and there are PHI nodes in
847 // the successor, then we need to add incoming edges for the PHI nodes
849 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
851 // Loop over all of the PHI nodes in the successor of BB.
852 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
853 PHINode *PN = cast<PHINode>(I);
855 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
859 if (Succ->getSinglePredecessor()) {
860 // BB is the only predecessor of Succ, so Succ will end up with exactly
861 // the same predecessors BB had.
863 // Copy over any phi, debug or lifetime instruction.
864 BB->getTerminator()->eraseFromParent();
865 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
868 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
869 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
870 assert(PN->use_empty() && "There shouldn't be any uses here!");
871 PN->eraseFromParent();
875 // Everything that jumped to BB now goes to Succ.
876 BB->replaceAllUsesWith(Succ);
877 if (!Succ->hasName()) Succ->takeName(BB);
878 BB->eraseFromParent(); // Delete the old basic block.
882 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
883 /// nodes in this block. This doesn't try to be clever about PHI nodes
884 /// which differ only in the order of the incoming values, but instcombine
885 /// orders them so it usually won't matter.
887 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
888 // This implementation doesn't currently consider undef operands
889 // specially. Theoretically, two phis which are identical except for
890 // one having an undef where the other doesn't could be collapsed.
892 struct PHIDenseMapInfo {
893 static PHINode *getEmptyKey() {
894 return DenseMapInfo<PHINode *>::getEmptyKey();
896 static PHINode *getTombstoneKey() {
897 return DenseMapInfo<PHINode *>::getTombstoneKey();
899 static unsigned getHashValue(PHINode *PN) {
900 // Compute a hash value on the operands. Instcombine will likely have
901 // sorted them, which helps expose duplicates, but we have to check all
902 // the operands to be safe in case instcombine hasn't run.
903 return static_cast<unsigned>(hash_combine(
904 hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
905 hash_combine_range(PN->block_begin(), PN->block_end())));
907 static bool isEqual(PHINode *LHS, PHINode *RHS) {
908 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
909 RHS == getEmptyKey() || RHS == getTombstoneKey())
911 return LHS->isIdenticalTo(RHS);
915 // Set of unique PHINodes.
916 DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
919 bool Changed = false;
920 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
921 auto Inserted = PHISet.insert(PN);
922 if (!Inserted.second) {
923 // A duplicate. Replace this PHI with its duplicate.
924 PN->replaceAllUsesWith(*Inserted.first);
925 PN->eraseFromParent();
928 // The RAUW can change PHIs that we already visited. Start over from the
938 /// enforceKnownAlignment - If the specified pointer points to an object that
939 /// we control, modify the object's alignment to PrefAlign. This isn't
940 /// often possible though. If alignment is important, a more reliable approach
941 /// is to simply align all global variables and allocation instructions to
942 /// their preferred alignment from the beginning.
944 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
946 const DataLayout &DL) {
947 assert(PrefAlign > Align);
949 V = V->stripPointerCasts();
951 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
952 // TODO: ideally, computeKnownBits ought to have used
953 // AllocaInst::getAlignment() in its computation already, making
954 // the below max redundant. But, as it turns out,
955 // stripPointerCasts recurses through infinite layers of bitcasts,
956 // while computeKnownBits is not allowed to traverse more than 6
958 Align = std::max(AI->getAlignment(), Align);
959 if (PrefAlign <= Align)
962 // If the preferred alignment is greater than the natural stack alignment
963 // then don't round up. This avoids dynamic stack realignment.
964 if (DL.exceedsNaturalStackAlignment(PrefAlign))
966 AI->setAlignment(PrefAlign);
970 if (auto *GO = dyn_cast<GlobalObject>(V)) {
971 // TODO: as above, this shouldn't be necessary.
972 Align = std::max(GO->getAlignment(), Align);
973 if (PrefAlign <= Align)
976 // If there is a large requested alignment and we can, bump up the alignment
977 // of the global. If the memory we set aside for the global may not be the
978 // memory used by the final program then it is impossible for us to reliably
979 // enforce the preferred alignment.
980 if (!GO->canIncreaseAlignment())
983 GO->setAlignment(PrefAlign);
990 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that
991 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
992 /// and it is more than the alignment of the ultimate object, see if we can
993 /// increase the alignment of the ultimate object, making this check succeed.
994 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
995 const DataLayout &DL,
996 const Instruction *CxtI,
998 const DominatorTree *DT) {
999 assert(V->getType()->isPointerTy() &&
1000 "getOrEnforceKnownAlignment expects a pointer!");
1001 unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType());
1003 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1004 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT);
1005 unsigned TrailZ = KnownZero.countTrailingOnes();
1007 // Avoid trouble with ridiculously large TrailZ values, such as
1008 // those computed from a null pointer.
1009 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1011 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
1013 // LLVM doesn't support alignments larger than this currently.
1014 Align = std::min(Align, +Value::MaximumAlignment);
1016 if (PrefAlign > Align)
1017 Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1019 // We don't need to make any adjustment.
1023 ///===---------------------------------------------------------------------===//
1024 /// Dbg Intrinsic utilities
1027 /// See if there is a dbg.value intrinsic for DIVar before I.
1028 static bool LdStHasDebugValue(const DILocalVariable *DIVar, Instruction *I) {
1029 // Since we can't guarantee that the original dbg.declare instrinsic
1030 // is removed by LowerDbgDeclare(), we need to make sure that we are
1031 // not inserting the same dbg.value intrinsic over and over.
1032 llvm::BasicBlock::InstListType::iterator PrevI(I);
1033 if (PrevI != I->getParent()->getInstList().begin()) {
1035 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1036 if (DVI->getValue() == I->getOperand(0) &&
1037 DVI->getOffset() == 0 &&
1038 DVI->getVariable() == DIVar)
1044 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1045 /// that has an associated llvm.dbg.decl intrinsic.
1046 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1047 StoreInst *SI, DIBuilder &Builder) {
1048 auto *DIVar = DDI->getVariable();
1049 auto *DIExpr = DDI->getExpression();
1050 assert(DIVar && "Missing variable");
1052 if (LdStHasDebugValue(DIVar, SI))
1055 // If an argument is zero extended then use argument directly. The ZExt
1056 // may be zapped by an optimization pass in future.
1057 Argument *ExtendedArg = nullptr;
1058 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1059 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1060 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1061 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1063 // We're now only describing a subset of the variable. The piece we're
1064 // describing will always be smaller than the variable size, because
1065 // VariableSize == Size of Alloca described by DDI. Since SI stores
1066 // to the alloca described by DDI, if it's first operand is an extend,
1067 // we're guaranteed that before extension, the value was narrower than
1068 // the size of the alloca, hence the size of the described variable.
1069 SmallVector<uint64_t, 3> NewDIExpr;
1070 unsigned PieceOffset = 0;
1071 // If this already is a bit piece, we drop the bit piece from the expression
1072 // and record the offset.
1073 if (DIExpr->isBitPiece()) {
1074 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end()-3);
1075 PieceOffset = DIExpr->getBitPieceOffset();
1077 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1079 NewDIExpr.push_back(dwarf::DW_OP_bit_piece);
1080 NewDIExpr.push_back(PieceOffset); //Offset
1081 const DataLayout &DL = DDI->getModule()->getDataLayout();
1082 NewDIExpr.push_back(DL.getTypeSizeInBits(ExtendedArg->getType())); // Size
1083 Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar,
1084 Builder.createExpression(NewDIExpr),
1085 DDI->getDebugLoc(), SI);
1088 Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr,
1089 DDI->getDebugLoc(), SI);
1093 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1094 /// that has an associated llvm.dbg.decl intrinsic.
1095 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1096 LoadInst *LI, DIBuilder &Builder) {
1097 auto *DIVar = DDI->getVariable();
1098 auto *DIExpr = DDI->getExpression();
1099 assert(DIVar && "Missing variable");
1101 if (LdStHasDebugValue(DIVar, LI))
1104 // We are now tracking the loaded value instead of the address. In the
1105 // future if multi-location support is added to the IR, it might be
1106 // preferable to keep tracking both the loaded value and the original
1107 // address in case the alloca can not be elided.
1108 Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1109 LI, 0, DIVar, DIExpr, DDI->getDebugLoc(), (Instruction *)nullptr);
1110 DbgValue->insertAfter(LI);
1114 /// Determine whether this alloca is either a VLA or an array.
1115 static bool isArray(AllocaInst *AI) {
1116 return AI->isArrayAllocation() ||
1117 AI->getType()->getElementType()->isArrayTy();
1120 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1121 /// of llvm.dbg.value intrinsics.
1122 bool llvm::LowerDbgDeclare(Function &F) {
1123 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1124 SmallVector<DbgDeclareInst *, 4> Dbgs;
1126 for (Instruction &BI : FI)
1127 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1128 Dbgs.push_back(DDI);
1133 for (auto &I : Dbgs) {
1134 DbgDeclareInst *DDI = I;
1135 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1136 // If this is an alloca for a scalar variable, insert a dbg.value
1137 // at each load and store to the alloca and erase the dbg.declare.
1138 // The dbg.values allow tracking a variable even if it is not
1139 // stored on the stack, while the dbg.declare can only describe
1140 // the stack slot (and at a lexical-scope granularity). Later
1141 // passes will attempt to elide the stack slot.
1142 if (AI && !isArray(AI)) {
1143 for (User *U : AI->users())
1144 if (StoreInst *SI = dyn_cast<StoreInst>(U))
1145 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1146 else if (LoadInst *LI = dyn_cast<LoadInst>(U))
1147 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1148 else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1149 // This is a call by-value or some other instruction that
1150 // takes a pointer to the variable. Insert a *value*
1151 // intrinsic that describes the alloca.
1152 SmallVector<uint64_t, 1> NewDIExpr;
1153 auto *DIExpr = DDI->getExpression();
1154 NewDIExpr.push_back(dwarf::DW_OP_deref);
1155 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1156 DIB.insertDbgValueIntrinsic(AI, 0, DDI->getVariable(),
1157 DIB.createExpression(NewDIExpr),
1158 DDI->getDebugLoc(), CI);
1160 DDI->eraseFromParent();
1166 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
1167 /// alloca 'V', if any.
1168 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) {
1169 if (auto *L = LocalAsMetadata::getIfExists(V))
1170 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1171 for (User *U : MDV->users())
1172 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1178 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1179 Instruction *InsertBefore, DIBuilder &Builder,
1180 bool Deref, int Offset) {
1181 DbgDeclareInst *DDI = FindAllocaDbgDeclare(Address);
1184 DebugLoc Loc = DDI->getDebugLoc();
1185 auto *DIVar = DDI->getVariable();
1186 auto *DIExpr = DDI->getExpression();
1187 assert(DIVar && "Missing variable");
1189 if (Deref || Offset) {
1190 // Create a copy of the original DIDescriptor for user variable, prepending
1191 // "deref" operation to a list of address elements, as new llvm.dbg.declare
1192 // will take a value storing address of the memory for variable, not
1194 SmallVector<uint64_t, 4> NewDIExpr;
1196 NewDIExpr.push_back(dwarf::DW_OP_deref);
1198 NewDIExpr.push_back(dwarf::DW_OP_plus);
1199 NewDIExpr.push_back(Offset);
1200 } else if (Offset < 0) {
1201 NewDIExpr.push_back(dwarf::DW_OP_minus);
1202 NewDIExpr.push_back(-Offset);
1205 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1206 DIExpr = Builder.createExpression(NewDIExpr);
1209 // Insert llvm.dbg.declare immediately after the original alloca, and remove
1210 // old llvm.dbg.declare.
1211 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1212 DDI->eraseFromParent();
1216 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1217 DIBuilder &Builder, bool Deref, int Offset) {
1218 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1222 void llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap) {
1223 BasicBlock *BB = I->getParent();
1224 // Loop over all of the successors, removing BB's entry from any PHI
1226 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1227 (*SI)->removePredecessor(BB);
1229 // Insert a call to llvm.trap right before this. This turns the undefined
1230 // behavior into a hard fail instead of falling through into random code.
1233 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1234 CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1235 CallTrap->setDebugLoc(I->getDebugLoc());
1237 new UnreachableInst(I->getContext(), I);
1239 // All instructions after this are dead.
1240 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1241 while (BBI != BBE) {
1242 if (!BBI->use_empty())
1243 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1244 BB->getInstList().erase(BBI++);
1248 /// changeToCall - Convert the specified invoke into a normal call.
1249 static void changeToCall(InvokeInst *II) {
1250 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1251 SmallVector<OperandBundleDef, 1> OpBundles;
1252 II->getOperandBundlesAsDefs(OpBundles);
1253 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1255 NewCall->takeName(II);
1256 NewCall->setCallingConv(II->getCallingConv());
1257 NewCall->setAttributes(II->getAttributes());
1258 NewCall->setDebugLoc(II->getDebugLoc());
1259 II->replaceAllUsesWith(NewCall);
1261 // Follow the call by a branch to the normal destination.
1262 BranchInst::Create(II->getNormalDest(), II);
1264 // Update PHI nodes in the unwind destination
1265 II->getUnwindDest()->removePredecessor(II->getParent());
1266 II->eraseFromParent();
1269 static bool markAliveBlocks(Function &F,
1270 SmallPtrSetImpl<BasicBlock*> &Reachable) {
1272 SmallVector<BasicBlock*, 128> Worklist;
1273 BasicBlock *BB = &F.front();
1274 Worklist.push_back(BB);
1275 Reachable.insert(BB);
1276 bool Changed = false;
1278 BB = Worklist.pop_back_val();
1280 // Do a quick scan of the basic block, turning any obviously unreachable
1281 // instructions into LLVM unreachable insts. The instruction combining pass
1282 // canonicalizes unreachable insts into stores to null or undef.
1283 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){
1284 // Assumptions that are known to be false are equivalent to unreachable.
1285 // Also, if the condition is undefined, then we make the choice most
1286 // beneficial to the optimizer, and choose that to also be unreachable.
1287 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
1288 if (II->getIntrinsicID() == Intrinsic::assume) {
1289 bool MakeUnreachable = false;
1290 if (isa<UndefValue>(II->getArgOperand(0)))
1291 MakeUnreachable = true;
1292 else if (ConstantInt *Cond =
1293 dyn_cast<ConstantInt>(II->getArgOperand(0)))
1294 MakeUnreachable = Cond->isZero();
1296 if (MakeUnreachable) {
1297 // Don't insert a call to llvm.trap right before the unreachable.
1298 changeToUnreachable(&*BBI, false);
1304 if (CallInst *CI = dyn_cast<CallInst>(BBI)) {
1305 if (CI->doesNotReturn()) {
1306 // If we found a call to a no-return function, insert an unreachable
1307 // instruction after it. Make sure there isn't *already* one there
1310 if (!isa<UnreachableInst>(BBI)) {
1311 // Don't insert a call to llvm.trap right before the unreachable.
1312 changeToUnreachable(&*BBI, false);
1319 // Store to undef and store to null are undefined and used to signal that
1320 // they should be changed to unreachable by passes that can't modify the
1322 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
1323 // Don't touch volatile stores.
1324 if (SI->isVolatile()) continue;
1326 Value *Ptr = SI->getOperand(1);
1328 if (isa<UndefValue>(Ptr) ||
1329 (isa<ConstantPointerNull>(Ptr) &&
1330 SI->getPointerAddressSpace() == 0)) {
1331 changeToUnreachable(SI, true);
1338 TerminatorInst *Terminator = BB->getTerminator();
1339 if (auto *II = dyn_cast<InvokeInst>(Terminator)) {
1340 // Turn invokes that call 'nounwind' functions into ordinary calls.
1341 Value *Callee = II->getCalledValue();
1342 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1343 changeToUnreachable(II, true);
1345 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
1346 if (II->use_empty() && II->onlyReadsMemory()) {
1347 // jump to the normal destination branch.
1348 BranchInst::Create(II->getNormalDest(), II);
1349 II->getUnwindDest()->removePredecessor(II->getParent());
1350 II->eraseFromParent();
1355 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) {
1356 // Remove catchpads which cannot be reached.
1357 struct CatchPadDenseMapInfo {
1358 static CatchPadInst *getEmptyKey() {
1359 return DenseMapInfo<CatchPadInst *>::getEmptyKey();
1361 static CatchPadInst *getTombstoneKey() {
1362 return DenseMapInfo<CatchPadInst *>::getTombstoneKey();
1364 static unsigned getHashValue(CatchPadInst *CatchPad) {
1365 return static_cast<unsigned>(hash_combine_range(
1366 CatchPad->value_op_begin(), CatchPad->value_op_end()));
1368 static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) {
1369 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
1370 RHS == getEmptyKey() || RHS == getTombstoneKey())
1372 return LHS->isIdenticalTo(RHS);
1376 // Set of unique CatchPads.
1377 SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4,
1378 CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>>
1380 detail::DenseSetEmpty Empty;
1381 for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(),
1382 E = CatchSwitch->handler_end();
1384 BasicBlock *HandlerBB = *I;
1385 auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI());
1386 if (!HandlerSet.insert({CatchPad, Empty}).second) {
1387 CatchSwitch->removeHandler(I);
1395 Changed |= ConstantFoldTerminator(BB, true);
1396 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1397 if (Reachable.insert(*SI).second)
1398 Worklist.push_back(*SI);
1399 } while (!Worklist.empty());
1403 void llvm::removeUnwindEdge(BasicBlock *BB) {
1404 TerminatorInst *TI = BB->getTerminator();
1406 if (auto *II = dyn_cast<InvokeInst>(TI)) {
1411 TerminatorInst *NewTI;
1412 BasicBlock *UnwindDest;
1414 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
1415 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
1416 UnwindDest = CRI->getUnwindDest();
1417 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
1418 auto *NewCatchSwitch = CatchSwitchInst::Create(
1419 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
1420 CatchSwitch->getName(), CatchSwitch);
1421 for (BasicBlock *PadBB : CatchSwitch->handlers())
1422 NewCatchSwitch->addHandler(PadBB);
1424 NewTI = NewCatchSwitch;
1425 UnwindDest = CatchSwitch->getUnwindDest();
1427 llvm_unreachable("Could not find unwind successor");
1430 NewTI->takeName(TI);
1431 NewTI->setDebugLoc(TI->getDebugLoc());
1432 UnwindDest->removePredecessor(BB);
1433 TI->replaceAllUsesWith(NewTI);
1434 TI->eraseFromParent();
1437 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even
1438 /// if they are in a dead cycle. Return true if a change was made, false
1440 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI) {
1441 SmallPtrSet<BasicBlock*, 128> Reachable;
1442 bool Changed = markAliveBlocks(F, Reachable);
1444 // If there are unreachable blocks in the CFG...
1445 if (Reachable.size() == F.size())
1448 assert(Reachable.size() < F.size());
1449 NumRemoved += F.size()-Reachable.size();
1451 // Loop over all of the basic blocks that are not reachable, dropping all of
1452 // their internal references...
1453 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) {
1454 if (Reachable.count(&*BB))
1457 for (succ_iterator SI = succ_begin(&*BB), SE = succ_end(&*BB); SI != SE;
1459 if (Reachable.count(*SI))
1460 (*SI)->removePredecessor(&*BB);
1462 LVI->eraseBlock(&*BB);
1463 BB->dropAllReferences();
1466 for (Function::iterator I = ++F.begin(); I != F.end();)
1467 if (!Reachable.count(&*I))
1468 I = F.getBasicBlockList().erase(I);
1475 void llvm::combineMetadata(Instruction *K, const Instruction *J,
1476 ArrayRef<unsigned> KnownIDs) {
1477 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1478 K->dropUnknownNonDebugMetadata(KnownIDs);
1479 K->getAllMetadataOtherThanDebugLoc(Metadata);
1480 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
1481 unsigned Kind = Metadata[i].first;
1482 MDNode *JMD = J->getMetadata(Kind);
1483 MDNode *KMD = Metadata[i].second;
1487 K->setMetadata(Kind, nullptr); // Remove unknown metadata
1489 case LLVMContext::MD_dbg:
1490 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1491 case LLVMContext::MD_tbaa:
1492 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1494 case LLVMContext::MD_alias_scope:
1495 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1497 case LLVMContext::MD_noalias:
1498 K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1500 case LLVMContext::MD_range:
1501 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1503 case LLVMContext::MD_fpmath:
1504 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1506 case LLVMContext::MD_invariant_load:
1507 // Only set the !invariant.load if it is present in both instructions.
1508 K->setMetadata(Kind, JMD);
1510 case LLVMContext::MD_nonnull:
1511 // Only set the !nonnull if it is present in both instructions.
1512 K->setMetadata(Kind, JMD);
1514 case LLVMContext::MD_invariant_group:
1515 // Preserve !invariant.group in K.
1517 case LLVMContext::MD_align:
1518 K->setMetadata(Kind,
1519 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1521 case LLVMContext::MD_dereferenceable:
1522 case LLVMContext::MD_dereferenceable_or_null:
1523 K->setMetadata(Kind,
1524 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1528 // Set !invariant.group from J if J has it. If both instructions have it
1529 // then we will just pick it from J - even when they are different.
1530 // Also make sure that K is load or store - f.e. combining bitcast with load
1531 // could produce bitcast with invariant.group metadata, which is invalid.
1532 // FIXME: we should try to preserve both invariant.group md if they are
1533 // different, but right now instruction can only have one invariant.group.
1534 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
1535 if (isa<LoadInst>(K) || isa<StoreInst>(K))
1536 K->setMetadata(LLVMContext::MD_invariant_group, JMD);
1539 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1541 const BasicBlockEdge &Root) {
1542 assert(From->getType() == To->getType());
1545 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1548 if (DT.dominates(Root, U)) {
1550 DEBUG(dbgs() << "Replace dominated use of '"
1551 << From->getName() << "' as "
1552 << *To << " in " << *U << "\n");
1559 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1561 const BasicBlock *BB) {
1562 assert(From->getType() == To->getType());
1565 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1568 auto *I = cast<Instruction>(U.getUser());
1569 if (DT.dominates(BB, I->getParent())) {
1571 DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as "
1572 << *To << " in " << *U << "\n");
1579 bool llvm::callsGCLeafFunction(ImmutableCallSite CS) {
1580 if (isa<IntrinsicInst>(CS.getInstruction()))
1581 // Most LLVM intrinsics are things which can never take a safepoint.
1582 // As a result, we don't need to have the stack parsable at the
1583 // callsite. This is a highly useful optimization since intrinsic
1584 // calls are fairly prevalent, particularly in debug builds.
1587 // Check if the function is specifically marked as a gc leaf function.
1588 if (CS.hasFnAttr("gc-leaf-function"))
1590 if (const Function *F = CS.getCalledFunction())
1591 return F->hasFnAttribute("gc-leaf-function");