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.
267 const TargetLibraryInfo &TLI;
268 const TargetTransformInfo &TTI;
271 typedef RecyclingAllocator<
272 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy;
273 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
274 AllocatorTy> ScopedHTType;
276 /// \brief A scoped hash table of the current values of all of our simple
277 /// scalar expressions.
279 /// As we walk down the domtree, we look to see if instructions are in this:
280 /// if so, we replace them with what we find, otherwise we insert them so
281 /// that dominated values can succeed in their lookup.
282 ScopedHTType AvailableValues;
284 /// \brief A scoped hash table of the current values of loads.
286 /// This allows us to get efficient access to dominating loads when we have
287 /// a fully redundant load. In addition to the most recent load, we keep
288 /// track of a generation count of the read, which is compared against the
289 /// current generation count. The current generation count is incremented
290 /// after every possibly writing memory operation, which ensures that we only
291 /// CSE loads with other loads that have no intervening store.
296 LoadValue() : Data(nullptr), Generation(0), MatchingId(-1) {}
297 LoadValue(Value *Data, unsigned Generation, unsigned MatchingId)
298 : Data(Data), Generation(Generation), MatchingId(MatchingId) {}
300 typedef RecyclingAllocator<BumpPtrAllocator,
301 ScopedHashTableVal<Value *, LoadValue>>
303 typedef ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>,
304 LoadMapAllocator> LoadHTType;
305 LoadHTType AvailableLoads;
307 /// \brief A scoped hash table of the current values of read-only call
310 /// It uses the same generation count as loads.
311 typedef ScopedHashTable<CallValue, std::pair<Value *, unsigned>> CallHTType;
312 CallHTType AvailableCalls;
314 /// \brief This is the current generation of the memory value.
315 unsigned CurrentGeneration;
317 /// \brief Set up the EarlyCSE runner for a particular function.
318 EarlyCSE(const TargetLibraryInfo &TLI, const TargetTransformInfo &TTI,
319 DominatorTree &DT, AssumptionCache &AC)
320 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), CurrentGeneration(0) {}
325 // Almost a POD, but needs to call the constructors for the scoped hash
326 // tables so that a new scope gets pushed on. These are RAII so that the
327 // scope gets popped when the NodeScope is destroyed.
330 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
331 CallHTType &AvailableCalls)
332 : Scope(AvailableValues), LoadScope(AvailableLoads),
333 CallScope(AvailableCalls) {}
336 NodeScope(const NodeScope &) = delete;
337 void operator=(const NodeScope &) = delete;
339 ScopedHTType::ScopeTy Scope;
340 LoadHTType::ScopeTy LoadScope;
341 CallHTType::ScopeTy CallScope;
344 // Contains all the needed information to create a stack for doing a depth
345 // first tranversal of the tree. This includes scopes for values, loads, and
346 // calls as well as the generation. There is a child iterator so that the
347 // children do not need to be store spearately.
350 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
351 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n,
352 DomTreeNode::iterator child, DomTreeNode::iterator end)
353 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
354 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls),
358 unsigned currentGeneration() { return CurrentGeneration; }
359 unsigned childGeneration() { return ChildGeneration; }
360 void childGeneration(unsigned generation) { ChildGeneration = generation; }
361 DomTreeNode *node() { return Node; }
362 DomTreeNode::iterator childIter() { return ChildIter; }
363 DomTreeNode *nextChild() {
364 DomTreeNode *child = *ChildIter;
368 DomTreeNode::iterator end() { return EndIter; }
369 bool isProcessed() { return Processed; }
370 void process() { Processed = true; }
373 StackNode(const StackNode &) = delete;
374 void operator=(const StackNode &) = delete;
377 unsigned CurrentGeneration;
378 unsigned ChildGeneration;
380 DomTreeNode::iterator ChildIter;
381 DomTreeNode::iterator EndIter;
386 /// \brief Wrapper class to handle memory instructions, including loads,
387 /// stores and intrinsic loads and stores defined by the target.
388 class ParseMemoryInst {
390 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
391 : IsTargetMemInst(false), Inst(Inst) {
392 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
393 if (TTI.getTgtMemIntrinsic(II, Info) && Info.NumMemRefs == 1)
394 IsTargetMemInst = true;
396 bool isLoad() const {
397 if (IsTargetMemInst) return Info.ReadMem;
398 return isa<LoadInst>(Inst);
400 bool isStore() const {
401 if (IsTargetMemInst) return Info.WriteMem;
402 return isa<StoreInst>(Inst);
404 bool isSimple() const {
405 if (IsTargetMemInst) return Info.IsSimple;
406 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
407 return LI->isSimple();
408 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
409 return SI->isSimple();
411 return Inst->isAtomic();
413 bool isMatchingMemLoc(const ParseMemoryInst &Inst) const {
414 return (getPointerOperand() == Inst.getPointerOperand() &&
415 getMatchingId() == Inst.getMatchingId());
417 bool isValid() const { return getPointerOperand() != nullptr; }
419 // For regular (non-intrinsic) loads/stores, this is set to -1. For
420 // intrinsic loads/stores, the id is retrieved from the corresponding
421 // field in the MemIntrinsicInfo structure. That field contains
422 // non-negative values only.
423 int getMatchingId() const {
424 if (IsTargetMemInst) return Info.MatchingId;
427 Value *getPointerOperand() const {
428 if (IsTargetMemInst) return Info.PtrVal;
429 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
430 return LI->getPointerOperand();
431 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
432 return SI->getPointerOperand();
436 bool mayReadFromMemory() const {
437 if (IsTargetMemInst) return Info.ReadMem;
438 return Inst->mayReadFromMemory();
440 bool mayWriteToMemory() const {
441 if (IsTargetMemInst) return Info.WriteMem;
442 return Inst->mayWriteToMemory();
446 bool IsTargetMemInst;
447 MemIntrinsicInfo Info;
451 bool processNode(DomTreeNode *Node);
453 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
454 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
456 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
457 return SI->getValueOperand();
458 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
459 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
465 bool EarlyCSE::processNode(DomTreeNode *Node) {
466 BasicBlock *BB = Node->getBlock();
468 // If this block has a single predecessor, then the predecessor is the parent
469 // of the domtree node and all of the live out memory values are still current
470 // in this block. If this block has multiple predecessors, then they could
471 // have invalidated the live-out memory values of our parent value. For now,
472 // just be conservative and invalidate memory if this block has multiple
474 if (!BB->getSinglePredecessor())
477 // If this node has a single predecessor which ends in a conditional branch,
478 // we can infer the value of the branch condition given that we took this
479 // path. We need the single predeccesor to ensure there's not another path
480 // which reaches this block where the condition might hold a different
481 // value. Since we're adding this to the scoped hash table (like any other
482 // def), it will have been popped if we encounter a future merge block.
483 if (BasicBlock *Pred = BB->getSinglePredecessor())
484 if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
485 if (BI->isConditional())
486 if (auto *CondInst = dyn_cast<Instruction>(BI->getCondition()))
487 if (SimpleValue::canHandle(CondInst)) {
488 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
489 auto *ConditionalConstant = (BI->getSuccessor(0) == BB) ?
490 ConstantInt::getTrue(BB->getContext()) :
491 ConstantInt::getFalse(BB->getContext());
492 AvailableValues.insert(CondInst, ConditionalConstant);
493 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
494 << CondInst->getName() << "' as " << *ConditionalConstant
495 << " in " << BB->getName() << "\n");
496 // Replace all dominated uses with the known value
497 replaceDominatedUsesWith(CondInst, ConditionalConstant, DT,
498 BasicBlockEdge(Pred, BB));
501 /// LastStore - Keep track of the last non-volatile store that we saw... for
502 /// as long as there in no instruction that reads memory. If we see a store
503 /// to the same location, we delete the dead store. This zaps trivial dead
504 /// stores which can occur in bitfield code among other things.
505 Instruction *LastStore = nullptr;
507 bool Changed = false;
508 const DataLayout &DL = BB->getModule()->getDataLayout();
510 // See if any instructions in the block can be eliminated. If so, do it. If
511 // not, add them to AvailableValues.
512 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
513 Instruction *Inst = &*I++;
515 // Dead instructions should just be removed.
516 if (isInstructionTriviallyDead(Inst, &TLI)) {
517 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
518 Inst->eraseFromParent();
524 // Skip assume intrinsics, they don't really have side effects (although
525 // they're marked as such to ensure preservation of control dependencies),
526 // and this pass will not disturb any of the assumption's control
528 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
529 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
533 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
534 // its simpler value.
535 if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) {
536 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
537 Inst->replaceAllUsesWith(V);
538 Inst->eraseFromParent();
544 // If this is a simple instruction that we can value number, process it.
545 if (SimpleValue::canHandle(Inst)) {
546 // See if the instruction has an available value. If so, use it.
547 if (Value *V = AvailableValues.lookup(Inst)) {
548 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
549 Inst->replaceAllUsesWith(V);
550 Inst->eraseFromParent();
556 // Otherwise, just remember that this value is available.
557 AvailableValues.insert(Inst, Inst);
561 ParseMemoryInst MemInst(Inst, TTI);
562 // If this is a non-volatile load, process it.
563 if (MemInst.isValid() && MemInst.isLoad()) {
564 // Ignore volatile or ordered loads.
565 if (!MemInst.isSimple()) {
567 // Don't CSE across synchronization boundaries.
568 if (Inst->mayWriteToMemory())
573 // If we have an available version of this load, and if it is the right
574 // generation, replace this instruction.
575 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
576 if (InVal.Data != nullptr && InVal.Generation == CurrentGeneration &&
577 InVal.MatchingId == MemInst.getMatchingId()) {
578 Value *Op = getOrCreateResult(InVal.Data, Inst->getType());
580 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
581 << " to: " << *InVal.Data << '\n');
582 if (!Inst->use_empty())
583 Inst->replaceAllUsesWith(Op);
584 Inst->eraseFromParent();
591 // Otherwise, remember that we have this instruction.
592 AvailableLoads.insert(
593 MemInst.getPointerOperand(),
594 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId()));
599 // If this instruction may read from memory, forget LastStore.
600 // Load/store intrinsics will indicate both a read and a write to
601 // memory. The target may override this (e.g. so that a store intrinsic
602 // does not read from memory, and thus will be treated the same as a
603 // regular store for commoning purposes).
604 if (Inst->mayReadFromMemory() &&
605 !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
608 // If this is a read-only call, process it.
609 if (CallValue::canHandle(Inst)) {
610 // If we have an available version of this call, and if it is the right
611 // generation, replace this instruction.
612 std::pair<Value *, unsigned> InVal = AvailableCalls.lookup(Inst);
613 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
614 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
615 << " to: " << *InVal.first << '\n');
616 if (!Inst->use_empty())
617 Inst->replaceAllUsesWith(InVal.first);
618 Inst->eraseFromParent();
624 // Otherwise, remember that we have this instruction.
625 AvailableCalls.insert(
626 Inst, std::pair<Value *, unsigned>(Inst, CurrentGeneration));
630 // A release fence requires that all stores complete before it, but does
631 // not prevent the reordering of following loads 'before' the fence. As a
632 // result, we don't need to consider it as writing to memory and don't need
633 // to advance the generation. We do need to prevent DSE across the fence,
634 // but that's handled above.
635 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
636 if (FI->getOrdering() == Release) {
637 assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above");
641 // Okay, this isn't something we can CSE at all. Check to see if it is
642 // something that could modify memory. If so, our available memory values
643 // cannot be used so bump the generation count.
644 if (Inst->mayWriteToMemory()) {
647 if (MemInst.isValid() && MemInst.isStore()) {
648 // We do a trivial form of DSE if there are two stores to the same
649 // location with no intervening loads. Delete the earlier store.
651 ParseMemoryInst LastStoreMemInst(LastStore, TTI);
652 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
653 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
654 << " due to: " << *Inst << '\n');
655 LastStore->eraseFromParent();
660 // fallthrough - we can exploit information about this store
663 // Okay, we just invalidated anything we knew about loaded values. Try
664 // to salvage *something* by remembering that the stored value is a live
665 // version of the pointer. It is safe to forward from volatile stores
666 // to non-volatile loads, so we don't have to check for volatility of
668 AvailableLoads.insert(
669 MemInst.getPointerOperand(),
670 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId()));
672 // Remember that this was the last normal store we saw for DSE.
673 if (MemInst.isSimple())
682 bool EarlyCSE::run() {
683 // Note, deque is being used here because there is significant performance
684 // gains over vector when the container becomes very large due to the
685 // specific access patterns. For more information see the mailing list
686 // discussion on this:
687 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
688 std::deque<StackNode *> nodesToProcess;
690 bool Changed = false;
692 // Process the root node.
693 nodesToProcess.push_back(new StackNode(
694 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
695 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end()));
697 // Save the current generation.
698 unsigned LiveOutGeneration = CurrentGeneration;
700 // Process the stack.
701 while (!nodesToProcess.empty()) {
702 // Grab the first item off the stack. Set the current generation, remove
703 // the node from the stack, and process it.
704 StackNode *NodeToProcess = nodesToProcess.back();
706 // Initialize class members.
707 CurrentGeneration = NodeToProcess->currentGeneration();
709 // Check if the node needs to be processed.
710 if (!NodeToProcess->isProcessed()) {
712 Changed |= processNode(NodeToProcess->node());
713 NodeToProcess->childGeneration(CurrentGeneration);
714 NodeToProcess->process();
715 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
716 // Push the next child onto the stack.
717 DomTreeNode *child = NodeToProcess->nextChild();
718 nodesToProcess.push_back(
719 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
720 NodeToProcess->childGeneration(), child, child->begin(),
723 // It has been processed, and there are no more children to process,
724 // so delete it and pop it off the stack.
725 delete NodeToProcess;
726 nodesToProcess.pop_back();
728 } // while (!nodes...)
730 // Reset the current generation.
731 CurrentGeneration = LiveOutGeneration;
736 PreservedAnalyses EarlyCSEPass::run(Function &F,
737 AnalysisManager<Function> *AM) {
738 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
739 auto &TTI = AM->getResult<TargetIRAnalysis>(F);
740 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
741 auto &AC = AM->getResult<AssumptionAnalysis>(F);
743 EarlyCSE CSE(TLI, TTI, DT, AC);
746 return PreservedAnalyses::all();
748 // CSE preserves the dominator tree because it doesn't mutate the CFG.
749 // FIXME: Bundle this with other CFG-preservation.
750 PreservedAnalyses PA;
751 PA.preserve<DominatorTreeAnalysis>();
756 /// \brief A simple and fast domtree-based CSE pass.
758 /// This pass does a simple depth-first walk over the dominator tree,
759 /// eliminating trivially redundant instructions and using instsimplify to
760 /// canonicalize things as it goes. It is intended to be fast and catch obvious
761 /// cases so that instcombine and other passes are more effective. It is
762 /// expected that a later pass of GVN will catch the interesting/hard cases.
763 class EarlyCSELegacyPass : public FunctionPass {
767 EarlyCSELegacyPass() : FunctionPass(ID) {
768 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
771 bool runOnFunction(Function &F) override {
772 if (skipOptnoneFunction(F))
775 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
776 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
777 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
778 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
780 EarlyCSE CSE(TLI, TTI, DT, AC);
785 void getAnalysisUsage(AnalysisUsage &AU) const override {
786 AU.addRequired<AssumptionCacheTracker>();
787 AU.addRequired<DominatorTreeWrapperPass>();
788 AU.addRequired<TargetLibraryInfoWrapperPass>();
789 AU.addRequired<TargetTransformInfoWrapperPass>();
790 AU.addPreserved<GlobalsAAWrapperPass>();
791 AU.setPreservesCFG();
796 char EarlyCSELegacyPass::ID = 0;
798 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSELegacyPass(); }
800 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
802 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
803 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
804 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
805 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
806 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)