1 //===- EarlyCSE.cpp - Simple and fast CSE pass ----------------------------===//
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 pass performs a simple dominator tree walk that eliminates trivially
11 // redundant instructions.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Scalar/EarlyCSE.h"
16 #include "llvm/ADT/Hashing.h"
17 #include "llvm/ADT/ScopedHashTable.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/GlobalsModRef.h"
20 #include "llvm/Analysis/AssumptionCache.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/PatternMatch.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/RecyclingAllocator.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Scalar.h"
34 #include "llvm/Transforms/Utils/Local.h"
37 using namespace llvm::PatternMatch;
39 #define DEBUG_TYPE "early-cse"
41 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
42 STATISTIC(NumCSE, "Number of instructions CSE'd");
43 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
44 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
45 STATISTIC(NumDSE, "Number of trivial dead stores removed");
47 //===----------------------------------------------------------------------===//
49 //===----------------------------------------------------------------------===//
52 /// \brief Struct representing the available values in the scoped hash table.
56 SimpleValue(Instruction *I) : Inst(I) {
57 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
60 bool isSentinel() const {
61 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
62 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
65 static bool canHandle(Instruction *Inst) {
66 // This can only handle non-void readnone functions.
67 if (CallInst *CI = dyn_cast<CallInst>(Inst))
68 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
69 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) ||
70 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) ||
71 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
72 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
73 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst);
79 template <> struct DenseMapInfo<SimpleValue> {
80 static inline SimpleValue getEmptyKey() {
81 return DenseMapInfo<Instruction *>::getEmptyKey();
83 static inline SimpleValue getTombstoneKey() {
84 return DenseMapInfo<Instruction *>::getTombstoneKey();
86 static unsigned getHashValue(SimpleValue Val);
87 static bool isEqual(SimpleValue LHS, SimpleValue RHS);
91 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
92 Instruction *Inst = Val.Inst;
93 // Hash in all of the operands as pointers.
94 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
95 Value *LHS = BinOp->getOperand(0);
96 Value *RHS = BinOp->getOperand(1);
97 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
100 if (isa<OverflowingBinaryOperator>(BinOp)) {
101 // Hash the overflow behavior
103 BinOp->hasNoSignedWrap() * OverflowingBinaryOperator::NoSignedWrap |
104 BinOp->hasNoUnsignedWrap() *
105 OverflowingBinaryOperator::NoUnsignedWrap;
106 return hash_combine(BinOp->getOpcode(), Overflow, LHS, RHS);
109 return hash_combine(BinOp->getOpcode(), LHS, RHS);
112 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
113 Value *LHS = CI->getOperand(0);
114 Value *RHS = CI->getOperand(1);
115 CmpInst::Predicate Pred = CI->getPredicate();
116 if (Inst->getOperand(0) > Inst->getOperand(1)) {
118 Pred = CI->getSwappedPredicate();
120 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
123 if (CastInst *CI = dyn_cast<CastInst>(Inst))
124 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
126 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
127 return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
128 hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
130 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
131 return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
133 hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
135 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) ||
136 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) ||
137 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
138 isa<ShuffleVectorInst>(Inst)) &&
139 "Invalid/unknown instruction");
141 // Mix in the opcode.
144 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
147 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
148 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
150 if (LHS.isSentinel() || RHS.isSentinel())
153 if (LHSI->getOpcode() != RHSI->getOpcode())
155 if (LHSI->isIdenticalTo(RHSI))
158 // If we're not strictly identical, we still might be a commutable instruction
159 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
160 if (!LHSBinOp->isCommutative())
163 assert(isa<BinaryOperator>(RHSI) &&
164 "same opcode, but different instruction type?");
165 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
167 // Check overflow attributes
168 if (isa<OverflowingBinaryOperator>(LHSBinOp)) {
169 assert(isa<OverflowingBinaryOperator>(RHSBinOp) &&
170 "same opcode, but different operator type?");
171 if (LHSBinOp->hasNoUnsignedWrap() != RHSBinOp->hasNoUnsignedWrap() ||
172 LHSBinOp->hasNoSignedWrap() != RHSBinOp->hasNoSignedWrap())
177 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
178 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
180 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
181 assert(isa<CmpInst>(RHSI) &&
182 "same opcode, but different instruction type?");
183 CmpInst *RHSCmp = cast<CmpInst>(RHSI);
185 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
186 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
187 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
193 //===----------------------------------------------------------------------===//
195 //===----------------------------------------------------------------------===//
198 /// \brief Struct representing the available call values in the scoped hash
203 CallValue(Instruction *I) : Inst(I) {
204 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
207 bool isSentinel() const {
208 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
209 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
212 static bool canHandle(Instruction *Inst) {
213 // Don't value number anything that returns void.
214 if (Inst->getType()->isVoidTy())
217 CallInst *CI = dyn_cast<CallInst>(Inst);
218 if (!CI || !CI->onlyReadsMemory())
226 template <> struct DenseMapInfo<CallValue> {
227 static inline CallValue getEmptyKey() {
228 return DenseMapInfo<Instruction *>::getEmptyKey();
230 static inline CallValue getTombstoneKey() {
231 return DenseMapInfo<Instruction *>::getTombstoneKey();
233 static unsigned getHashValue(CallValue Val);
234 static bool isEqual(CallValue LHS, CallValue RHS);
238 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
239 Instruction *Inst = Val.Inst;
240 // Hash all of the operands as pointers and mix in the opcode.
243 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
246 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
247 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
248 if (LHS.isSentinel() || RHS.isSentinel())
250 return LHSI->isIdenticalTo(RHSI);
253 //===----------------------------------------------------------------------===//
254 // EarlyCSE implementation
255 //===----------------------------------------------------------------------===//
258 /// \brief A simple and fast domtree-based CSE pass.
260 /// This pass does a simple depth-first walk over the dominator tree,
261 /// eliminating trivially redundant instructions and using instsimplify to
262 /// canonicalize things as it goes. It is intended to be fast and catch obvious
263 /// cases so that instcombine and other passes are more effective. It is
264 /// expected that a later pass of GVN will catch the interesting/hard cases.
268 const TargetLibraryInfo &TLI;
269 const TargetTransformInfo &TTI;
272 typedef RecyclingAllocator<
273 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy;
274 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
275 AllocatorTy> ScopedHTType;
277 /// \brief A scoped hash table of the current values of all of our simple
278 /// scalar expressions.
280 /// As we walk down the domtree, we look to see if instructions are in this:
281 /// if so, we replace them with what we find, otherwise we insert them so
282 /// that dominated values can succeed in their lookup.
283 ScopedHTType AvailableValues;
285 /// \brief A scoped hash table of the current values of loads.
287 /// This allows us to get efficient access to dominating loads when we have
288 /// a fully redundant load. In addition to the most recent load, we keep
289 /// track of a generation count of the read, which is compared against the
290 /// current generation count. The current generation count is incremented
291 /// after every possibly writing memory operation, which ensures that we only
292 /// CSE loads with other loads that have no intervening store.
293 typedef RecyclingAllocator<
295 ScopedHashTableVal<Value *, std::pair<Value *, unsigned>>>
297 typedef ScopedHashTable<Value *, std::pair<Value *, unsigned>,
298 DenseMapInfo<Value *>, LoadMapAllocator> LoadHTType;
299 LoadHTType AvailableLoads;
301 /// \brief A scoped hash table of the current values of read-only call
304 /// It uses the same generation count as loads.
305 typedef ScopedHashTable<CallValue, std::pair<Value *, unsigned>> CallHTType;
306 CallHTType AvailableCalls;
308 /// \brief This is the current generation of the memory value.
309 unsigned CurrentGeneration;
311 /// \brief Set up the EarlyCSE runner for a particular function.
312 EarlyCSE(Function &F, const TargetLibraryInfo &TLI,
313 const TargetTransformInfo &TTI, DominatorTree &DT,
315 : F(F), TLI(TLI), TTI(TTI), DT(DT), AC(AC), CurrentGeneration(0) {}
320 // Almost a POD, but needs to call the constructors for the scoped hash
321 // tables so that a new scope gets pushed on. These are RAII so that the
322 // scope gets popped when the NodeScope is destroyed.
325 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
326 CallHTType &AvailableCalls)
327 : Scope(AvailableValues), LoadScope(AvailableLoads),
328 CallScope(AvailableCalls) {}
331 NodeScope(const NodeScope &) = delete;
332 void operator=(const NodeScope &) = delete;
334 ScopedHTType::ScopeTy Scope;
335 LoadHTType::ScopeTy LoadScope;
336 CallHTType::ScopeTy CallScope;
339 // Contains all the needed information to create a stack for doing a depth
340 // first tranversal of the tree. This includes scopes for values, loads, and
341 // calls as well as the generation. There is a child iterator so that the
342 // children do not need to be store spearately.
345 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
346 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n,
347 DomTreeNode::iterator child, DomTreeNode::iterator end)
348 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
349 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls),
353 unsigned currentGeneration() { return CurrentGeneration; }
354 unsigned childGeneration() { return ChildGeneration; }
355 void childGeneration(unsigned generation) { ChildGeneration = generation; }
356 DomTreeNode *node() { return Node; }
357 DomTreeNode::iterator childIter() { return ChildIter; }
358 DomTreeNode *nextChild() {
359 DomTreeNode *child = *ChildIter;
363 DomTreeNode::iterator end() { return EndIter; }
364 bool isProcessed() { return Processed; }
365 void process() { Processed = true; }
368 StackNode(const StackNode &) = delete;
369 void operator=(const StackNode &) = delete;
372 unsigned CurrentGeneration;
373 unsigned ChildGeneration;
375 DomTreeNode::iterator ChildIter;
376 DomTreeNode::iterator EndIter;
381 /// \brief Wrapper class to handle memory instructions, including loads,
382 /// stores and intrinsic loads and stores defined by the target.
383 class ParseMemoryInst {
385 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
386 : Load(false), Store(false), Vol(false), MayReadFromMemory(false),
387 MayWriteToMemory(false), MatchingId(-1), Ptr(nullptr) {
388 MayReadFromMemory = Inst->mayReadFromMemory();
389 MayWriteToMemory = Inst->mayWriteToMemory();
390 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
391 MemIntrinsicInfo Info;
392 if (!TTI.getTgtMemIntrinsic(II, Info))
394 if (Info.NumMemRefs == 1) {
395 Store = Info.WriteMem;
397 MatchingId = Info.MatchingId;
398 MayReadFromMemory = Info.ReadMem;
399 MayWriteToMemory = Info.WriteMem;
403 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
405 Vol = !LI->isSimple();
406 Ptr = LI->getPointerOperand();
407 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
409 Vol = !SI->isSimple();
410 Ptr = SI->getPointerOperand();
413 bool isLoad() { return Load; }
414 bool isStore() { return Store; }
415 bool isVolatile() { return Vol; }
416 bool isMatchingMemLoc(const ParseMemoryInst &Inst) {
417 return Ptr == Inst.Ptr && MatchingId == Inst.MatchingId;
419 bool isValid() { return Ptr != nullptr; }
420 int getMatchingId() { return MatchingId; }
421 Value *getPtr() { return Ptr; }
422 bool mayReadFromMemory() { return MayReadFromMemory; }
423 bool mayWriteToMemory() { return MayWriteToMemory; }
429 bool MayReadFromMemory;
430 bool MayWriteToMemory;
431 // For regular (non-intrinsic) loads/stores, this is set to -1. For
432 // intrinsic loads/stores, the id is retrieved from the corresponding
433 // field in the MemIntrinsicInfo structure. That field contains
434 // non-negative values only.
439 bool processNode(DomTreeNode *Node);
441 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
442 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
444 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
445 return SI->getValueOperand();
446 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
447 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
453 bool EarlyCSE::processNode(DomTreeNode *Node) {
454 BasicBlock *BB = Node->getBlock();
456 // If this block has a single predecessor, then the predecessor is the parent
457 // of the domtree node and all of the live out memory values are still current
458 // in this block. If this block has multiple predecessors, then they could
459 // have invalidated the live-out memory values of our parent value. For now,
460 // just be conservative and invalidate memory if this block has multiple
462 if (!BB->getSinglePredecessor())
465 // If this node has a single predecessor which ends in a conditional branch,
466 // we can infer the value of the branch condition given that we took this
467 // path. We need the single predeccesor to ensure there's not another path
468 // which reaches this block where the condition might hold a different
469 // value. Since we're adding this to the scoped hash table (like any other
470 // def), it will have been popped if we encounter a future merge block.
471 if (BasicBlock *Pred = BB->getSinglePredecessor())
472 if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
473 if (BI->isConditional())
474 if (auto *CondInst = dyn_cast<Instruction>(BI->getCondition()))
475 if (SimpleValue::canHandle(CondInst)) {
476 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
477 auto *ConditionalConstant = (BI->getSuccessor(0) == BB) ?
478 ConstantInt::getTrue(BB->getContext()) :
479 ConstantInt::getFalse(BB->getContext());
480 AvailableValues.insert(CondInst, ConditionalConstant);
481 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
482 << CondInst->getName() << "' as " << *ConditionalConstant
483 << " in " << BB->getName() << "\n");
484 // Replace all dominated uses with the known value
485 replaceDominatedUsesWith(CondInst, ConditionalConstant, DT,
486 BasicBlockEdge(Pred, BB));
489 /// LastStore - Keep track of the last non-volatile store that we saw... for
490 /// as long as there in no instruction that reads memory. If we see a store
491 /// to the same location, we delete the dead store. This zaps trivial dead
492 /// stores which can occur in bitfield code among other things.
493 Instruction *LastStore = nullptr;
495 bool Changed = false;
496 const DataLayout &DL = BB->getModule()->getDataLayout();
498 // See if any instructions in the block can be eliminated. If so, do it. If
499 // not, add them to AvailableValues.
500 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
501 Instruction *Inst = I++;
503 // Dead instructions should just be removed.
504 if (isInstructionTriviallyDead(Inst, &TLI)) {
505 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
506 Inst->eraseFromParent();
512 // Skip assume intrinsics, they don't really have side effects (although
513 // they're marked as such to ensure preservation of control dependencies),
514 // and this pass will not disturb any of the assumption's control
516 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
517 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
521 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
522 // its simpler value.
523 if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) {
524 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
525 Inst->replaceAllUsesWith(V);
526 Inst->eraseFromParent();
532 // If this is a simple instruction that we can value number, process it.
533 if (SimpleValue::canHandle(Inst)) {
534 // See if the instruction has an available value. If so, use it.
535 if (Value *V = AvailableValues.lookup(Inst)) {
536 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
537 Inst->replaceAllUsesWith(V);
538 Inst->eraseFromParent();
544 // Otherwise, just remember that this value is available.
545 AvailableValues.insert(Inst, Inst);
549 ParseMemoryInst MemInst(Inst, TTI);
550 // If this is a non-volatile load, process it.
551 if (MemInst.isValid() && MemInst.isLoad()) {
552 // Ignore volatile loads.
553 if (MemInst.isVolatile()) {
555 // Don't CSE across synchronization boundaries.
556 if (Inst->mayWriteToMemory())
561 // If we have an available version of this load, and if it is the right
562 // generation, replace this instruction.
563 std::pair<Value *, unsigned> InVal =
564 AvailableLoads.lookup(MemInst.getPtr());
565 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
566 Value *Op = getOrCreateResult(InVal.first, Inst->getType());
568 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
569 << " to: " << *InVal.first << '\n');
570 if (!Inst->use_empty())
571 Inst->replaceAllUsesWith(Op);
572 Inst->eraseFromParent();
579 // Otherwise, remember that we have this instruction.
580 AvailableLoads.insert(MemInst.getPtr(), std::pair<Value *, unsigned>(
581 Inst, CurrentGeneration));
586 // If this instruction may read from memory, forget LastStore.
587 // Load/store intrinsics will indicate both a read and a write to
588 // memory. The target may override this (e.g. so that a store intrinsic
589 // does not read from memory, and thus will be treated the same as a
590 // regular store for commoning purposes).
591 if (Inst->mayReadFromMemory() &&
592 !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
595 // If this is a read-only call, process it.
596 if (CallValue::canHandle(Inst)) {
597 // If we have an available version of this call, and if it is the right
598 // generation, replace this instruction.
599 std::pair<Value *, unsigned> InVal = AvailableCalls.lookup(Inst);
600 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
601 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
602 << " to: " << *InVal.first << '\n');
603 if (!Inst->use_empty())
604 Inst->replaceAllUsesWith(InVal.first);
605 Inst->eraseFromParent();
611 // Otherwise, remember that we have this instruction.
612 AvailableCalls.insert(
613 Inst, std::pair<Value *, unsigned>(Inst, CurrentGeneration));
617 // A release fence requires that all stores complete before it, but does
618 // not prevent the reordering of following loads 'before' the fence. As a
619 // result, we don't need to consider it as writing to memory and don't need
620 // to advance the generation. We do need to prevent DSE across the fence,
621 // but that's handled above.
622 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
623 if (FI->getOrdering() == Release) {
624 assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above");
628 // Okay, this isn't something we can CSE at all. Check to see if it is
629 // something that could modify memory. If so, our available memory values
630 // cannot be used so bump the generation count.
631 if (Inst->mayWriteToMemory()) {
634 if (MemInst.isValid() && MemInst.isStore()) {
635 // We do a trivial form of DSE if there are two stores to the same
636 // location with no intervening loads. Delete the earlier store.
638 ParseMemoryInst LastStoreMemInst(LastStore, TTI);
639 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
640 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
641 << " due to: " << *Inst << '\n');
642 LastStore->eraseFromParent();
647 // fallthrough - we can exploit information about this store
650 // Okay, we just invalidated anything we knew about loaded values. Try
651 // to salvage *something* by remembering that the stored value is a live
652 // version of the pointer. It is safe to forward from volatile stores
653 // to non-volatile loads, so we don't have to check for volatility of
655 AvailableLoads.insert(MemInst.getPtr(), std::pair<Value *, unsigned>(
656 Inst, CurrentGeneration));
658 // Remember that this was the last store we saw for DSE.
659 if (!MemInst.isVolatile())
668 bool EarlyCSE::run() {
669 // Note, deque is being used here because there is significant performance
670 // gains over vector when the container becomes very large due to the
671 // specific access patterns. For more information see the mailing list
672 // discussion on this:
673 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
674 std::deque<StackNode *> nodesToProcess;
676 bool Changed = false;
678 // Process the root node.
679 nodesToProcess.push_back(new StackNode(
680 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
681 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end()));
683 // Save the current generation.
684 unsigned LiveOutGeneration = CurrentGeneration;
686 // Process the stack.
687 while (!nodesToProcess.empty()) {
688 // Grab the first item off the stack. Set the current generation, remove
689 // the node from the stack, and process it.
690 StackNode *NodeToProcess = nodesToProcess.back();
692 // Initialize class members.
693 CurrentGeneration = NodeToProcess->currentGeneration();
695 // Check if the node needs to be processed.
696 if (!NodeToProcess->isProcessed()) {
698 Changed |= processNode(NodeToProcess->node());
699 NodeToProcess->childGeneration(CurrentGeneration);
700 NodeToProcess->process();
701 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
702 // Push the next child onto the stack.
703 DomTreeNode *child = NodeToProcess->nextChild();
704 nodesToProcess.push_back(
705 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
706 NodeToProcess->childGeneration(), child, child->begin(),
709 // It has been processed, and there are no more children to process,
710 // so delete it and pop it off the stack.
711 delete NodeToProcess;
712 nodesToProcess.pop_back();
714 } // while (!nodes...)
716 // Reset the current generation.
717 CurrentGeneration = LiveOutGeneration;
722 PreservedAnalyses EarlyCSEPass::run(Function &F,
723 AnalysisManager<Function> *AM) {
724 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
725 auto &TTI = AM->getResult<TargetIRAnalysis>(F);
726 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
727 auto &AC = AM->getResult<AssumptionAnalysis>(F);
729 EarlyCSE CSE(F, TLI, TTI, DT, AC);
732 return PreservedAnalyses::all();
734 // CSE preserves the dominator tree because it doesn't mutate the CFG.
735 // FIXME: Bundle this with other CFG-preservation.
736 PreservedAnalyses PA;
737 PA.preserve<DominatorTreeAnalysis>();
742 /// \brief A simple and fast domtree-based CSE pass.
744 /// This pass does a simple depth-first walk over the dominator tree,
745 /// eliminating trivially redundant instructions and using instsimplify to
746 /// canonicalize things as it goes. It is intended to be fast and catch obvious
747 /// cases so that instcombine and other passes are more effective. It is
748 /// expected that a later pass of GVN will catch the interesting/hard cases.
749 class EarlyCSELegacyPass : public FunctionPass {
753 EarlyCSELegacyPass() : FunctionPass(ID) {
754 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
757 bool runOnFunction(Function &F) override {
758 if (skipOptnoneFunction(F))
761 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
762 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
763 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
764 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
766 EarlyCSE CSE(F, TLI, TTI, DT, AC);
771 void getAnalysisUsage(AnalysisUsage &AU) const override {
772 AU.addRequired<AssumptionCacheTracker>();
773 AU.addRequired<DominatorTreeWrapperPass>();
774 AU.addRequired<TargetLibraryInfoWrapperPass>();
775 AU.addRequired<TargetTransformInfoWrapperPass>();
776 AU.addPreserved<GlobalsAAWrapperPass>();
777 AU.setPreservesCFG();
782 char EarlyCSELegacyPass::ID = 0;
784 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSELegacyPass(); }
786 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
788 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
789 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
790 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
791 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
792 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)