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/AssumptionCache.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/Analysis/MemorySSA.h"
23 #include "llvm/Analysis/MemorySSAUpdater.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/PatternMatch.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/RecyclingAllocator.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Transforms/Utils/Local.h"
39 using namespace llvm::PatternMatch;
41 #define DEBUG_TYPE "early-cse"
43 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
44 STATISTIC(NumCSE, "Number of instructions CSE'd");
45 STATISTIC(NumCSECVP, "Number of compare instructions CVP'd");
46 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
47 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
48 STATISTIC(NumDSE, "Number of trivial dead stores removed");
50 //===----------------------------------------------------------------------===//
52 //===----------------------------------------------------------------------===//
55 /// \brief Struct representing the available values in the scoped hash table.
59 SimpleValue(Instruction *I) : Inst(I) {
60 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
63 bool isSentinel() const {
64 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
65 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
68 static bool canHandle(Instruction *Inst) {
69 // This can only handle non-void readnone functions.
70 if (CallInst *CI = dyn_cast<CallInst>(Inst))
71 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
72 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) ||
73 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) ||
74 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
75 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
76 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst);
82 template <> struct DenseMapInfo<SimpleValue> {
83 static inline SimpleValue getEmptyKey() {
84 return DenseMapInfo<Instruction *>::getEmptyKey();
86 static inline SimpleValue getTombstoneKey() {
87 return DenseMapInfo<Instruction *>::getTombstoneKey();
89 static unsigned getHashValue(SimpleValue Val);
90 static bool isEqual(SimpleValue LHS, SimpleValue RHS);
94 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
95 Instruction *Inst = Val.Inst;
96 // Hash in all of the operands as pointers.
97 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
98 Value *LHS = BinOp->getOperand(0);
99 Value *RHS = BinOp->getOperand(1);
100 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
103 return hash_combine(BinOp->getOpcode(), LHS, RHS);
106 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
107 Value *LHS = CI->getOperand(0);
108 Value *RHS = CI->getOperand(1);
109 CmpInst::Predicate Pred = CI->getPredicate();
110 if (Inst->getOperand(0) > Inst->getOperand(1)) {
112 Pred = CI->getSwappedPredicate();
114 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
117 if (CastInst *CI = dyn_cast<CastInst>(Inst))
118 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
120 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
121 return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
122 hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
124 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
125 return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
127 hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
129 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) ||
130 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) ||
131 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
132 isa<ShuffleVectorInst>(Inst)) &&
133 "Invalid/unknown instruction");
135 // Mix in the opcode.
138 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
141 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
142 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
144 if (LHS.isSentinel() || RHS.isSentinel())
147 if (LHSI->getOpcode() != RHSI->getOpcode())
149 if (LHSI->isIdenticalToWhenDefined(RHSI))
152 // If we're not strictly identical, we still might be a commutable instruction
153 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
154 if (!LHSBinOp->isCommutative())
157 assert(isa<BinaryOperator>(RHSI) &&
158 "same opcode, but different instruction type?");
159 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
162 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
163 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
165 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
166 assert(isa<CmpInst>(RHSI) &&
167 "same opcode, but different instruction type?");
168 CmpInst *RHSCmp = cast<CmpInst>(RHSI);
170 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
171 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
172 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
178 //===----------------------------------------------------------------------===//
180 //===----------------------------------------------------------------------===//
183 /// \brief Struct representing the available call values in the scoped hash
188 CallValue(Instruction *I) : Inst(I) {
189 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
192 bool isSentinel() const {
193 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
194 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
197 static bool canHandle(Instruction *Inst) {
198 // Don't value number anything that returns void.
199 if (Inst->getType()->isVoidTy())
202 CallInst *CI = dyn_cast<CallInst>(Inst);
203 if (!CI || !CI->onlyReadsMemory())
211 template <> struct DenseMapInfo<CallValue> {
212 static inline CallValue getEmptyKey() {
213 return DenseMapInfo<Instruction *>::getEmptyKey();
215 static inline CallValue getTombstoneKey() {
216 return DenseMapInfo<Instruction *>::getTombstoneKey();
218 static unsigned getHashValue(CallValue Val);
219 static bool isEqual(CallValue LHS, CallValue RHS);
223 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
224 Instruction *Inst = Val.Inst;
225 // Hash all of the operands as pointers and mix in the opcode.
228 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
231 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
232 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
233 if (LHS.isSentinel() || RHS.isSentinel())
235 return LHSI->isIdenticalTo(RHSI);
238 //===----------------------------------------------------------------------===//
239 // EarlyCSE implementation
240 //===----------------------------------------------------------------------===//
243 /// \brief A simple and fast domtree-based CSE pass.
245 /// This pass does a simple depth-first walk over the dominator tree,
246 /// eliminating trivially redundant instructions and using instsimplify to
247 /// canonicalize things as it goes. It is intended to be fast and catch obvious
248 /// cases so that instcombine and other passes are more effective. It is
249 /// expected that a later pass of GVN will catch the interesting/hard cases.
252 const TargetLibraryInfo &TLI;
253 const TargetTransformInfo &TTI;
256 const SimplifyQuery SQ;
258 std::unique_ptr<MemorySSAUpdater> MSSAUpdater;
259 typedef RecyclingAllocator<
260 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy;
261 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
262 AllocatorTy> ScopedHTType;
264 /// \brief A scoped hash table of the current values of all of our simple
265 /// scalar expressions.
267 /// As we walk down the domtree, we look to see if instructions are in this:
268 /// if so, we replace them with what we find, otherwise we insert them so
269 /// that dominated values can succeed in their lookup.
270 ScopedHTType AvailableValues;
272 /// A scoped hash table of the current values of previously encounted memory
275 /// This allows us to get efficient access to dominating loads or stores when
276 /// we have a fully redundant load. In addition to the most recent load, we
277 /// keep track of a generation count of the read, which is compared against
278 /// the current generation count. The current generation count is incremented
279 /// after every possibly writing memory operation, which ensures that we only
280 /// CSE loads with other loads that have no intervening store. Ordering
281 /// events (such as fences or atomic instructions) increment the generation
282 /// count as well; essentially, we model these as writes to all possible
283 /// locations. Note that atomic and/or volatile loads and stores can be
284 /// present the table; it is the responsibility of the consumer to inspect
285 /// the atomicity/volatility if needed.
287 Instruction *DefInst;
293 : DefInst(nullptr), Generation(0), MatchingId(-1), IsAtomic(false),
294 IsInvariant(false) {}
295 LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId,
296 bool IsAtomic, bool IsInvariant)
297 : DefInst(Inst), Generation(Generation), MatchingId(MatchingId),
298 IsAtomic(IsAtomic), IsInvariant(IsInvariant) {}
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<Instruction *, unsigned>>
313 CallHTType AvailableCalls;
315 /// \brief This is the current generation of the memory value.
316 unsigned CurrentGeneration;
318 /// \brief Set up the EarlyCSE runner for a particular function.
319 EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI,
320 const TargetTransformInfo &TTI, DominatorTree &DT,
321 AssumptionCache &AC, MemorySSA *MSSA)
322 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA),
323 MSSAUpdater(make_unique<MemorySSAUpdater>(MSSA)), CurrentGeneration(0) {
329 // Almost a POD, but needs to call the constructors for the scoped hash
330 // tables so that a new scope gets pushed on. These are RAII so that the
331 // scope gets popped when the NodeScope is destroyed.
334 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
335 CallHTType &AvailableCalls)
336 : Scope(AvailableValues), LoadScope(AvailableLoads),
337 CallScope(AvailableCalls) {}
340 NodeScope(const NodeScope &) = delete;
341 void operator=(const NodeScope &) = delete;
343 ScopedHTType::ScopeTy Scope;
344 LoadHTType::ScopeTy LoadScope;
345 CallHTType::ScopeTy CallScope;
348 // Contains all the needed information to create a stack for doing a depth
349 // first traversal of the tree. This includes scopes for values, loads, and
350 // calls as well as the generation. There is a child iterator so that the
351 // children do not need to be store separately.
354 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
355 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n,
356 DomTreeNode::iterator child, DomTreeNode::iterator end)
357 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
358 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls),
362 unsigned currentGeneration() { return CurrentGeneration; }
363 unsigned childGeneration() { return ChildGeneration; }
364 void childGeneration(unsigned generation) { ChildGeneration = generation; }
365 DomTreeNode *node() { return Node; }
366 DomTreeNode::iterator childIter() { return ChildIter; }
367 DomTreeNode *nextChild() {
368 DomTreeNode *child = *ChildIter;
372 DomTreeNode::iterator end() { return EndIter; }
373 bool isProcessed() { return Processed; }
374 void process() { Processed = true; }
377 StackNode(const StackNode &) = delete;
378 void operator=(const StackNode &) = delete;
381 unsigned CurrentGeneration;
382 unsigned ChildGeneration;
384 DomTreeNode::iterator ChildIter;
385 DomTreeNode::iterator EndIter;
390 /// \brief Wrapper class to handle memory instructions, including loads,
391 /// stores and intrinsic loads and stores defined by the target.
392 class ParseMemoryInst {
394 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
395 : IsTargetMemInst(false), Inst(Inst) {
396 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
397 if (TTI.getTgtMemIntrinsic(II, Info))
398 IsTargetMemInst = true;
400 bool isLoad() const {
401 if (IsTargetMemInst) return Info.ReadMem;
402 return isa<LoadInst>(Inst);
404 bool isStore() const {
405 if (IsTargetMemInst) return Info.WriteMem;
406 return isa<StoreInst>(Inst);
408 bool isAtomic() const {
410 return Info.Ordering != AtomicOrdering::NotAtomic;
411 return Inst->isAtomic();
413 bool isUnordered() const {
415 return Info.isUnordered();
417 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
418 return LI->isUnordered();
419 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
420 return SI->isUnordered();
422 // Conservative answer
423 return !Inst->isAtomic();
426 bool isVolatile() const {
428 return Info.IsVolatile;
430 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
431 return LI->isVolatile();
432 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
433 return SI->isVolatile();
435 // Conservative answer
439 bool isInvariantLoad() const {
440 if (auto *LI = dyn_cast<LoadInst>(Inst))
441 return LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr;
445 bool isMatchingMemLoc(const ParseMemoryInst &Inst) const {
446 return (getPointerOperand() == Inst.getPointerOperand() &&
447 getMatchingId() == Inst.getMatchingId());
449 bool isValid() const { return getPointerOperand() != nullptr; }
451 // For regular (non-intrinsic) loads/stores, this is set to -1. For
452 // intrinsic loads/stores, the id is retrieved from the corresponding
453 // field in the MemIntrinsicInfo structure. That field contains
454 // non-negative values only.
455 int getMatchingId() const {
456 if (IsTargetMemInst) return Info.MatchingId;
459 Value *getPointerOperand() const {
460 if (IsTargetMemInst) return Info.PtrVal;
461 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
462 return LI->getPointerOperand();
463 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
464 return SI->getPointerOperand();
468 bool mayReadFromMemory() const {
469 if (IsTargetMemInst) return Info.ReadMem;
470 return Inst->mayReadFromMemory();
472 bool mayWriteToMemory() const {
473 if (IsTargetMemInst) return Info.WriteMem;
474 return Inst->mayWriteToMemory();
478 bool IsTargetMemInst;
479 MemIntrinsicInfo Info;
483 bool processNode(DomTreeNode *Node);
485 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
486 if (auto *LI = dyn_cast<LoadInst>(Inst))
488 if (auto *SI = dyn_cast<StoreInst>(Inst))
489 return SI->getValueOperand();
490 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
491 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
495 bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration,
496 Instruction *EarlierInst, Instruction *LaterInst);
498 void removeMSSA(Instruction *Inst) {
501 // Removing a store here can leave MemorySSA in an unoptimized state by
502 // creating MemoryPhis that have identical arguments and by creating
503 // MemoryUses whose defining access is not an actual clobber. We handle the
504 // phi case eagerly here. The non-optimized MemoryUse case is lazily
505 // updated by MemorySSA getClobberingMemoryAccess.
506 if (MemoryAccess *MA = MSSA->getMemoryAccess(Inst)) {
507 // Optimize MemoryPhi nodes that may become redundant by having all the
508 // same input values once MA is removed.
509 SmallVector<MemoryPhi *, 4> PhisToCheck;
510 SmallVector<MemoryAccess *, 8> WorkQueue;
511 WorkQueue.push_back(MA);
512 // Process MemoryPhi nodes in FIFO order using a ever-growing vector since
513 // we shouldn't be processing that many phis and this will avoid an
514 // allocation in almost all cases.
515 for (unsigned I = 0; I < WorkQueue.size(); ++I) {
516 MemoryAccess *WI = WorkQueue[I];
518 for (auto *U : WI->users())
519 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(U))
520 PhisToCheck.push_back(MP);
522 MSSAUpdater->removeMemoryAccess(WI);
524 for (MemoryPhi *MP : PhisToCheck) {
525 MemoryAccess *FirstIn = MP->getIncomingValue(0);
526 if (all_of(MP->incoming_values(),
527 [=](Use &In) { return In == FirstIn; }))
528 WorkQueue.push_back(MP);
537 /// Determine if the memory referenced by LaterInst is from the same heap
538 /// version as EarlierInst.
539 /// This is currently called in two scenarios:
551 /// in both cases we want to verify that there are no possible writes to the
552 /// memory referenced by p between the earlier and later instruction.
553 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration,
554 unsigned LaterGeneration,
555 Instruction *EarlierInst,
556 Instruction *LaterInst) {
557 // Check the simple memory generation tracking first.
558 if (EarlierGeneration == LaterGeneration)
564 // Since we know LaterDef dominates LaterInst and EarlierInst dominates
565 // LaterInst, if LaterDef dominates EarlierInst then it can't occur between
566 // EarlierInst and LaterInst and neither can any other write that potentially
567 // clobbers LaterInst.
568 MemoryAccess *LaterDef =
569 MSSA->getWalker()->getClobberingMemoryAccess(LaterInst);
570 return MSSA->dominates(LaterDef, MSSA->getMemoryAccess(EarlierInst));
573 bool EarlyCSE::processNode(DomTreeNode *Node) {
574 bool Changed = false;
575 BasicBlock *BB = Node->getBlock();
577 // If this block has a single predecessor, then the predecessor is the parent
578 // of the domtree node and all of the live out memory values are still current
579 // in this block. If this block has multiple predecessors, then they could
580 // have invalidated the live-out memory values of our parent value. For now,
581 // just be conservative and invalidate memory if this block has multiple
583 if (!BB->getSinglePredecessor())
586 // If this node has a single predecessor which ends in a conditional branch,
587 // we can infer the value of the branch condition given that we took this
588 // path. We need the single predecessor to ensure there's not another path
589 // which reaches this block where the condition might hold a different
590 // value. Since we're adding this to the scoped hash table (like any other
591 // def), it will have been popped if we encounter a future merge block.
592 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
593 auto *BI = dyn_cast<BranchInst>(Pred->getTerminator());
594 if (BI && BI->isConditional()) {
595 auto *CondInst = dyn_cast<Instruction>(BI->getCondition());
596 if (CondInst && SimpleValue::canHandle(CondInst)) {
597 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
598 auto *TorF = (BI->getSuccessor(0) == BB)
599 ? ConstantInt::getTrue(BB->getContext())
600 : ConstantInt::getFalse(BB->getContext());
601 AvailableValues.insert(CondInst, TorF);
602 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
603 << CondInst->getName() << "' as " << *TorF << " in "
604 << BB->getName() << "\n");
605 // Replace all dominated uses with the known value.
606 if (unsigned Count = replaceDominatedUsesWith(
607 CondInst, TorF, DT, BasicBlockEdge(Pred, BB))) {
609 NumCSECVP = NumCSECVP + Count;
615 /// LastStore - Keep track of the last non-volatile store that we saw... for
616 /// as long as there in no instruction that reads memory. If we see a store
617 /// to the same location, we delete the dead store. This zaps trivial dead
618 /// stores which can occur in bitfield code among other things.
619 Instruction *LastStore = nullptr;
621 // See if any instructions in the block can be eliminated. If so, do it. If
622 // not, add them to AvailableValues.
623 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
624 Instruction *Inst = &*I++;
626 // Dead instructions should just be removed.
627 if (isInstructionTriviallyDead(Inst, &TLI)) {
628 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
630 Inst->eraseFromParent();
636 // Skip assume intrinsics, they don't really have side effects (although
637 // they're marked as such to ensure preservation of control dependencies),
638 // and this pass will not bother with its removal. However, we should mark
639 // its condition as true for all dominated blocks.
640 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
642 dyn_cast<Instruction>(cast<CallInst>(Inst)->getArgOperand(0));
643 if (CondI && SimpleValue::canHandle(CondI)) {
644 DEBUG(dbgs() << "EarlyCSE considering assumption: " << *Inst << '\n');
645 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
647 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
651 // Skip invariant.start intrinsics since they only read memory, and we can
652 // forward values across it. Also, we dont need to consume the last store
653 // since the semantics of invariant.start allow us to perform DSE of the
654 // last store, if there was a store following invariant.start. Consider:
657 // invariant.start(p)
659 // We can DSE the store to 30, since the store 40 to invariant location p
660 // causes undefined behaviour.
661 if (match(Inst, m_Intrinsic<Intrinsic::invariant_start>()))
664 if (match(Inst, m_Intrinsic<Intrinsic::experimental_guard>())) {
666 dyn_cast<Instruction>(cast<CallInst>(Inst)->getArgOperand(0))) {
667 if (SimpleValue::canHandle(CondI)) {
668 // Do we already know the actual value of this condition?
669 if (auto *KnownCond = AvailableValues.lookup(CondI)) {
670 // Is the condition known to be true?
671 if (isa<ConstantInt>(KnownCond) &&
672 cast<ConstantInt>(KnownCond)->isOneValue()) {
673 DEBUG(dbgs() << "EarlyCSE removing guard: " << *Inst << '\n');
675 Inst->eraseFromParent();
679 // Use the known value if it wasn't true.
680 cast<CallInst>(Inst)->setArgOperand(0, KnownCond);
682 // The condition we're on guarding here is true for all dominated
684 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
688 // Guard intrinsics read all memory, but don't write any memory.
689 // Accordingly, don't update the generation but consume the last store (to
690 // avoid an incorrect DSE).
695 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
696 // its simpler value.
697 if (Value *V = SimplifyInstruction(Inst, SQ)) {
698 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
700 if (!Inst->use_empty()) {
701 Inst->replaceAllUsesWith(V);
704 if (isInstructionTriviallyDead(Inst, &TLI)) {
706 Inst->eraseFromParent();
716 // If this is a simple instruction that we can value number, process it.
717 if (SimpleValue::canHandle(Inst)) {
718 // See if the instruction has an available value. If so, use it.
719 if (Value *V = AvailableValues.lookup(Inst)) {
720 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
721 if (auto *I = dyn_cast<Instruction>(V))
723 Inst->replaceAllUsesWith(V);
725 Inst->eraseFromParent();
731 // Otherwise, just remember that this value is available.
732 AvailableValues.insert(Inst, Inst);
736 ParseMemoryInst MemInst(Inst, TTI);
737 // If this is a non-volatile load, process it.
738 if (MemInst.isValid() && MemInst.isLoad()) {
739 // (conservatively) we can't peak past the ordering implied by this
740 // operation, but we can add this load to our set of available values
741 if (MemInst.isVolatile() || !MemInst.isUnordered()) {
746 // If we have an available version of this load, and if it is the right
747 // generation or the load is known to be from an invariant location,
748 // replace this instruction.
750 // If either the dominating load or the current load are invariant, then
751 // we can assume the current load loads the same value as the dominating
753 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
754 if (InVal.DefInst != nullptr &&
755 InVal.MatchingId == MemInst.getMatchingId() &&
756 // We don't yet handle removing loads with ordering of any kind.
757 !MemInst.isVolatile() && MemInst.isUnordered() &&
758 // We can't replace an atomic load with one which isn't also atomic.
759 InVal.IsAtomic >= MemInst.isAtomic() &&
760 (InVal.IsInvariant || MemInst.isInvariantLoad() ||
761 isSameMemGeneration(InVal.Generation, CurrentGeneration,
762 InVal.DefInst, Inst))) {
763 Value *Op = getOrCreateResult(InVal.DefInst, Inst->getType());
765 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
766 << " to: " << *InVal.DefInst << '\n');
767 if (!Inst->use_empty())
768 Inst->replaceAllUsesWith(Op);
770 Inst->eraseFromParent();
777 // Otherwise, remember that we have this instruction.
778 AvailableLoads.insert(
779 MemInst.getPointerOperand(),
780 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
781 MemInst.isAtomic(), MemInst.isInvariantLoad()));
786 // If this instruction may read from memory or throw (and potentially read
787 // from memory in the exception handler), forget LastStore. Load/store
788 // intrinsics will indicate both a read and a write to memory. The target
789 // may override this (e.g. so that a store intrinsic does not read from
790 // memory, and thus will be treated the same as a regular store for
791 // commoning purposes).
792 if ((Inst->mayReadFromMemory() || Inst->mayThrow()) &&
793 !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
796 // If this is a read-only call, process it.
797 if (CallValue::canHandle(Inst)) {
798 // If we have an available version of this call, and if it is the right
799 // generation, replace this instruction.
800 std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(Inst);
801 if (InVal.first != nullptr &&
802 isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first,
804 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
805 << " to: " << *InVal.first << '\n');
806 if (!Inst->use_empty())
807 Inst->replaceAllUsesWith(InVal.first);
809 Inst->eraseFromParent();
815 // Otherwise, remember that we have this instruction.
816 AvailableCalls.insert(
817 Inst, std::pair<Instruction *, unsigned>(Inst, CurrentGeneration));
821 // A release fence requires that all stores complete before it, but does
822 // not prevent the reordering of following loads 'before' the fence. As a
823 // result, we don't need to consider it as writing to memory and don't need
824 // to advance the generation. We do need to prevent DSE across the fence,
825 // but that's handled above.
826 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
827 if (FI->getOrdering() == AtomicOrdering::Release) {
828 assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above");
832 // write back DSE - If we write back the same value we just loaded from
833 // the same location and haven't passed any intervening writes or ordering
834 // operations, we can remove the write. The primary benefit is in allowing
835 // the available load table to remain valid and value forward past where
836 // the store originally was.
837 if (MemInst.isValid() && MemInst.isStore()) {
838 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
840 InVal.DefInst == getOrCreateResult(Inst, InVal.DefInst->getType()) &&
841 InVal.MatchingId == MemInst.getMatchingId() &&
842 // We don't yet handle removing stores with ordering of any kind.
843 !MemInst.isVolatile() && MemInst.isUnordered() &&
844 isSameMemGeneration(InVal.Generation, CurrentGeneration,
845 InVal.DefInst, Inst)) {
846 // It is okay to have a LastStore to a different pointer here if MemorySSA
847 // tells us that the load and store are from the same memory generation.
848 // In that case, LastStore should keep its present value since we're
849 // removing the current store.
850 assert((!LastStore ||
851 ParseMemoryInst(LastStore, TTI).getPointerOperand() ==
852 MemInst.getPointerOperand() ||
854 "can't have an intervening store if not using MemorySSA!");
855 DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << *Inst << '\n');
857 Inst->eraseFromParent();
860 // We can avoid incrementing the generation count since we were able
861 // to eliminate this store.
866 // Okay, this isn't something we can CSE at all. Check to see if it is
867 // something that could modify memory. If so, our available memory values
868 // cannot be used so bump the generation count.
869 if (Inst->mayWriteToMemory()) {
872 if (MemInst.isValid() && MemInst.isStore()) {
873 // We do a trivial form of DSE if there are two stores to the same
874 // location with no intervening loads. Delete the earlier store.
875 // At the moment, we don't remove ordered stores, but do remove
876 // unordered atomic stores. There's no special requirement (for
877 // unordered atomics) about removing atomic stores only in favor of
878 // other atomic stores since we we're going to execute the non-atomic
879 // one anyway and the atomic one might never have become visible.
881 ParseMemoryInst LastStoreMemInst(LastStore, TTI);
882 assert(LastStoreMemInst.isUnordered() &&
883 !LastStoreMemInst.isVolatile() &&
884 "Violated invariant");
885 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
886 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
887 << " due to: " << *Inst << '\n');
888 removeMSSA(LastStore);
889 LastStore->eraseFromParent();
894 // fallthrough - we can exploit information about this store
897 // Okay, we just invalidated anything we knew about loaded values. Try
898 // to salvage *something* by remembering that the stored value is a live
899 // version of the pointer. It is safe to forward from volatile stores
900 // to non-volatile loads, so we don't have to check for volatility of
902 AvailableLoads.insert(
903 MemInst.getPointerOperand(),
904 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
905 MemInst.isAtomic(), /*IsInvariant=*/false));
907 // Remember that this was the last unordered store we saw for DSE. We
908 // don't yet handle DSE on ordered or volatile stores since we don't
909 // have a good way to model the ordering requirement for following
910 // passes once the store is removed. We could insert a fence, but
911 // since fences are slightly stronger than stores in their ordering,
912 // it's not clear this is a profitable transform. Another option would
913 // be to merge the ordering with that of the post dominating store.
914 if (MemInst.isUnordered() && !MemInst.isVolatile())
925 bool EarlyCSE::run() {
926 // Note, deque is being used here because there is significant performance
927 // gains over vector when the container becomes very large due to the
928 // specific access patterns. For more information see the mailing list
929 // discussion on this:
930 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
931 std::deque<StackNode *> nodesToProcess;
933 bool Changed = false;
935 // Process the root node.
936 nodesToProcess.push_back(new StackNode(
937 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
938 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end()));
940 // Save the current generation.
941 unsigned LiveOutGeneration = CurrentGeneration;
943 // Process the stack.
944 while (!nodesToProcess.empty()) {
945 // Grab the first item off the stack. Set the current generation, remove
946 // the node from the stack, and process it.
947 StackNode *NodeToProcess = nodesToProcess.back();
949 // Initialize class members.
950 CurrentGeneration = NodeToProcess->currentGeneration();
952 // Check if the node needs to be processed.
953 if (!NodeToProcess->isProcessed()) {
955 Changed |= processNode(NodeToProcess->node());
956 NodeToProcess->childGeneration(CurrentGeneration);
957 NodeToProcess->process();
958 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
959 // Push the next child onto the stack.
960 DomTreeNode *child = NodeToProcess->nextChild();
961 nodesToProcess.push_back(
962 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
963 NodeToProcess->childGeneration(), child, child->begin(),
966 // It has been processed, and there are no more children to process,
967 // so delete it and pop it off the stack.
968 delete NodeToProcess;
969 nodesToProcess.pop_back();
971 } // while (!nodes...)
973 // Reset the current generation.
974 CurrentGeneration = LiveOutGeneration;
979 PreservedAnalyses EarlyCSEPass::run(Function &F,
980 FunctionAnalysisManager &AM) {
981 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
982 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
983 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
984 auto &AC = AM.getResult<AssumptionAnalysis>(F);
986 UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr;
988 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
991 return PreservedAnalyses::all();
993 PreservedAnalyses PA;
994 PA.preserveSet<CFGAnalyses>();
995 PA.preserve<GlobalsAA>();
997 PA.preserve<MemorySSAAnalysis>();
1002 /// \brief A simple and fast domtree-based CSE pass.
1004 /// This pass does a simple depth-first walk over the dominator tree,
1005 /// eliminating trivially redundant instructions and using instsimplify to
1006 /// canonicalize things as it goes. It is intended to be fast and catch obvious
1007 /// cases so that instcombine and other passes are more effective. It is
1008 /// expected that a later pass of GVN will catch the interesting/hard cases.
1009 template<bool UseMemorySSA>
1010 class EarlyCSELegacyCommonPass : public FunctionPass {
1014 EarlyCSELegacyCommonPass() : FunctionPass(ID) {
1016 initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry());
1018 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
1021 bool runOnFunction(Function &F) override {
1022 if (skipFunction(F))
1025 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1026 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1027 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1028 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1030 UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr;
1032 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA);
1037 void getAnalysisUsage(AnalysisUsage &AU) const override {
1038 AU.addRequired<AssumptionCacheTracker>();
1039 AU.addRequired<DominatorTreeWrapperPass>();
1040 AU.addRequired<TargetLibraryInfoWrapperPass>();
1041 AU.addRequired<TargetTransformInfoWrapperPass>();
1043 AU.addRequired<MemorySSAWrapperPass>();
1044 AU.addPreserved<MemorySSAWrapperPass>();
1046 AU.addPreserved<GlobalsAAWrapperPass>();
1047 AU.setPreservesCFG();
1052 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>;
1055 char EarlyCSELegacyPass::ID = 0;
1057 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
1059 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1060 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1061 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1062 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1063 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)
1065 using EarlyCSEMemSSALegacyPass =
1066 EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>;
1069 char EarlyCSEMemSSALegacyPass::ID = 0;
1071 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) {
1073 return new EarlyCSEMemSSALegacyPass();
1075 return new EarlyCSELegacyPass();
1078 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1079 "Early CSE w/ MemorySSA", false, false)
1080 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1081 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1082 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1083 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1084 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1085 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa",
1086 "Early CSE w/ MemorySSA", false, false)