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/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(NumCSECVP, "Number of compare instructions CVP'd");
44 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
45 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
46 STATISTIC(NumDSE, "Number of trivial dead stores removed");
48 //===----------------------------------------------------------------------===//
50 //===----------------------------------------------------------------------===//
53 /// \brief Struct representing the available values in the scoped hash table.
57 SimpleValue(Instruction *I) : Inst(I) {
58 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
61 bool isSentinel() const {
62 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
63 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
66 static bool canHandle(Instruction *Inst) {
67 // This can only handle non-void readnone functions.
68 if (CallInst *CI = dyn_cast<CallInst>(Inst))
69 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
70 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) ||
71 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) ||
72 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
73 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
74 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst);
80 template <> struct DenseMapInfo<SimpleValue> {
81 static inline SimpleValue getEmptyKey() {
82 return DenseMapInfo<Instruction *>::getEmptyKey();
84 static inline SimpleValue getTombstoneKey() {
85 return DenseMapInfo<Instruction *>::getTombstoneKey();
87 static unsigned getHashValue(SimpleValue Val);
88 static bool isEqual(SimpleValue LHS, SimpleValue RHS);
92 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
93 Instruction *Inst = Val.Inst;
94 // Hash in all of the operands as pointers.
95 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
96 Value *LHS = BinOp->getOperand(0);
97 Value *RHS = BinOp->getOperand(1);
98 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
101 return hash_combine(BinOp->getOpcode(), LHS, RHS);
104 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
105 Value *LHS = CI->getOperand(0);
106 Value *RHS = CI->getOperand(1);
107 CmpInst::Predicate Pred = CI->getPredicate();
108 if (Inst->getOperand(0) > Inst->getOperand(1)) {
110 Pred = CI->getSwappedPredicate();
112 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
115 if (CastInst *CI = dyn_cast<CastInst>(Inst))
116 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
118 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
119 return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
120 hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
122 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
123 return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
125 hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
127 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) ||
128 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) ||
129 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
130 isa<ShuffleVectorInst>(Inst)) &&
131 "Invalid/unknown instruction");
133 // Mix in the opcode.
136 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
139 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
140 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
142 if (LHS.isSentinel() || RHS.isSentinel())
145 if (LHSI->getOpcode() != RHSI->getOpcode())
147 if (LHSI->isIdenticalToWhenDefined(RHSI))
150 // If we're not strictly identical, we still might be a commutable instruction
151 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
152 if (!LHSBinOp->isCommutative())
155 assert(isa<BinaryOperator>(RHSI) &&
156 "same opcode, but different instruction type?");
157 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
160 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
161 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
163 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
164 assert(isa<CmpInst>(RHSI) &&
165 "same opcode, but different instruction type?");
166 CmpInst *RHSCmp = cast<CmpInst>(RHSI);
168 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
169 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
170 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
176 //===----------------------------------------------------------------------===//
178 //===----------------------------------------------------------------------===//
181 /// \brief Struct representing the available call values in the scoped hash
186 CallValue(Instruction *I) : Inst(I) {
187 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
190 bool isSentinel() const {
191 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
192 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
195 static bool canHandle(Instruction *Inst) {
196 // Don't value number anything that returns void.
197 if (Inst->getType()->isVoidTy())
200 CallInst *CI = dyn_cast<CallInst>(Inst);
201 if (!CI || !CI->onlyReadsMemory())
209 template <> struct DenseMapInfo<CallValue> {
210 static inline CallValue getEmptyKey() {
211 return DenseMapInfo<Instruction *>::getEmptyKey();
213 static inline CallValue getTombstoneKey() {
214 return DenseMapInfo<Instruction *>::getTombstoneKey();
216 static unsigned getHashValue(CallValue Val);
217 static bool isEqual(CallValue LHS, CallValue RHS);
221 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
222 Instruction *Inst = Val.Inst;
223 // Hash all of the operands as pointers and mix in the opcode.
226 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
229 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
230 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
231 if (LHS.isSentinel() || RHS.isSentinel())
233 return LHSI->isIdenticalTo(RHSI);
236 //===----------------------------------------------------------------------===//
237 // EarlyCSE implementation
238 //===----------------------------------------------------------------------===//
241 /// \brief A simple and fast domtree-based CSE pass.
243 /// This pass does a simple depth-first walk over the dominator tree,
244 /// eliminating trivially redundant instructions and using instsimplify to
245 /// canonicalize things as it goes. It is intended to be fast and catch obvious
246 /// cases so that instcombine and other passes are more effective. It is
247 /// expected that a later pass of GVN will catch the interesting/hard cases.
250 const TargetLibraryInfo &TLI;
251 const TargetTransformInfo &TTI;
254 typedef RecyclingAllocator<
255 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy;
256 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
257 AllocatorTy> ScopedHTType;
259 /// \brief A scoped hash table of the current values of all of our simple
260 /// scalar expressions.
262 /// As we walk down the domtree, we look to see if instructions are in this:
263 /// if so, we replace them with what we find, otherwise we insert them so
264 /// that dominated values can succeed in their lookup.
265 ScopedHTType AvailableValues;
267 /// A scoped hash table of the current values of previously encounted memory
270 /// This allows us to get efficient access to dominating loads or stores when
271 /// we have a fully redundant load. In addition to the most recent load, we
272 /// keep track of a generation count of the read, which is compared against
273 /// the current generation count. The current generation count is incremented
274 /// after every possibly writing memory operation, which ensures that we only
275 /// CSE loads with other loads that have no intervening store. Ordering
276 /// events (such as fences or atomic instructions) increment the generation
277 /// count as well; essentially, we model these as writes to all possible
278 /// locations. Note that atomic and/or volatile loads and stores can be
279 /// present the table; it is the responsibility of the consumer to inspect
280 /// the atomicity/volatility if needed.
282 Instruction *DefInst;
288 : DefInst(nullptr), Generation(0), MatchingId(-1), IsAtomic(false),
289 IsInvariant(false) {}
290 LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId,
291 bool IsAtomic, bool IsInvariant)
292 : DefInst(Inst), Generation(Generation), MatchingId(MatchingId),
293 IsAtomic(IsAtomic), IsInvariant(IsInvariant) {}
295 typedef RecyclingAllocator<BumpPtrAllocator,
296 ScopedHashTableVal<Value *, LoadValue>>
298 typedef ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>,
299 LoadMapAllocator> LoadHTType;
300 LoadHTType AvailableLoads;
302 /// \brief A scoped hash table of the current values of read-only call
305 /// It uses the same generation count as loads.
306 typedef ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>
308 CallHTType AvailableCalls;
310 /// \brief This is the current generation of the memory value.
311 unsigned CurrentGeneration;
313 /// \brief Set up the EarlyCSE runner for a particular function.
314 EarlyCSE(const TargetLibraryInfo &TLI, const TargetTransformInfo &TTI,
315 DominatorTree &DT, AssumptionCache &AC)
316 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), CurrentGeneration(0) {}
321 // Almost a POD, but needs to call the constructors for the scoped hash
322 // tables so that a new scope gets pushed on. These are RAII so that the
323 // scope gets popped when the NodeScope is destroyed.
326 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
327 CallHTType &AvailableCalls)
328 : Scope(AvailableValues), LoadScope(AvailableLoads),
329 CallScope(AvailableCalls) {}
332 NodeScope(const NodeScope &) = delete;
333 void operator=(const NodeScope &) = delete;
335 ScopedHTType::ScopeTy Scope;
336 LoadHTType::ScopeTy LoadScope;
337 CallHTType::ScopeTy CallScope;
340 // Contains all the needed information to create a stack for doing a depth
341 // first tranversal of the tree. This includes scopes for values, loads, and
342 // calls as well as the generation. There is a child iterator so that the
343 // children do not need to be store separately.
346 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
347 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n,
348 DomTreeNode::iterator child, DomTreeNode::iterator end)
349 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
350 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls),
354 unsigned currentGeneration() { return CurrentGeneration; }
355 unsigned childGeneration() { return ChildGeneration; }
356 void childGeneration(unsigned generation) { ChildGeneration = generation; }
357 DomTreeNode *node() { return Node; }
358 DomTreeNode::iterator childIter() { return ChildIter; }
359 DomTreeNode *nextChild() {
360 DomTreeNode *child = *ChildIter;
364 DomTreeNode::iterator end() { return EndIter; }
365 bool isProcessed() { return Processed; }
366 void process() { Processed = true; }
369 StackNode(const StackNode &) = delete;
370 void operator=(const StackNode &) = delete;
373 unsigned CurrentGeneration;
374 unsigned ChildGeneration;
376 DomTreeNode::iterator ChildIter;
377 DomTreeNode::iterator EndIter;
382 /// \brief Wrapper class to handle memory instructions, including loads,
383 /// stores and intrinsic loads and stores defined by the target.
384 class ParseMemoryInst {
386 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
387 : IsTargetMemInst(false), Inst(Inst) {
388 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
389 if (TTI.getTgtMemIntrinsic(II, Info) && Info.NumMemRefs == 1)
390 IsTargetMemInst = true;
392 bool isLoad() const {
393 if (IsTargetMemInst) return Info.ReadMem;
394 return isa<LoadInst>(Inst);
396 bool isStore() const {
397 if (IsTargetMemInst) return Info.WriteMem;
398 return isa<StoreInst>(Inst);
400 bool isAtomic() const {
401 if (IsTargetMemInst) {
402 assert(Info.IsSimple && "need to refine IsSimple in TTI");
405 return Inst->isAtomic();
407 bool isUnordered() const {
408 if (IsTargetMemInst) {
409 assert(Info.IsSimple && "need to refine IsSimple in TTI");
412 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
413 return LI->isUnordered();
414 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
415 return SI->isUnordered();
417 // Conservative answer
418 return !Inst->isAtomic();
421 bool isVolatile() const {
422 if (IsTargetMemInst) {
423 assert(Info.IsSimple && "need to refine IsSimple in TTI");
426 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
427 return LI->isVolatile();
428 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
429 return SI->isVolatile();
431 // Conservative answer
435 bool isInvariantLoad() const {
436 if (auto *LI = dyn_cast<LoadInst>(Inst))
437 return LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr;
441 bool isMatchingMemLoc(const ParseMemoryInst &Inst) const {
442 return (getPointerOperand() == Inst.getPointerOperand() &&
443 getMatchingId() == Inst.getMatchingId());
445 bool isValid() const { return getPointerOperand() != nullptr; }
447 // For regular (non-intrinsic) loads/stores, this is set to -1. For
448 // intrinsic loads/stores, the id is retrieved from the corresponding
449 // field in the MemIntrinsicInfo structure. That field contains
450 // non-negative values only.
451 int getMatchingId() const {
452 if (IsTargetMemInst) return Info.MatchingId;
455 Value *getPointerOperand() const {
456 if (IsTargetMemInst) return Info.PtrVal;
457 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
458 return LI->getPointerOperand();
459 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
460 return SI->getPointerOperand();
464 bool mayReadFromMemory() const {
465 if (IsTargetMemInst) return Info.ReadMem;
466 return Inst->mayReadFromMemory();
468 bool mayWriteToMemory() const {
469 if (IsTargetMemInst) return Info.WriteMem;
470 return Inst->mayWriteToMemory();
474 bool IsTargetMemInst;
475 MemIntrinsicInfo Info;
479 bool processNode(DomTreeNode *Node);
481 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
482 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
484 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
485 return SI->getValueOperand();
486 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
487 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
493 bool EarlyCSE::processNode(DomTreeNode *Node) {
494 bool Changed = false;
495 BasicBlock *BB = Node->getBlock();
497 // If this block has a single predecessor, then the predecessor is the parent
498 // of the domtree node and all of the live out memory values are still current
499 // in this block. If this block has multiple predecessors, then they could
500 // have invalidated the live-out memory values of our parent value. For now,
501 // just be conservative and invalidate memory if this block has multiple
503 if (!BB->getSinglePredecessor())
506 // If this node has a single predecessor which ends in a conditional branch,
507 // we can infer the value of the branch condition given that we took this
508 // path. We need the single predecessor to ensure there's not another path
509 // which reaches this block where the condition might hold a different
510 // value. Since we're adding this to the scoped hash table (like any other
511 // def), it will have been popped if we encounter a future merge block.
512 if (BasicBlock *Pred = BB->getSinglePredecessor())
513 if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
514 if (BI->isConditional())
515 if (auto *CondInst = dyn_cast<Instruction>(BI->getCondition()))
516 if (SimpleValue::canHandle(CondInst)) {
517 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
518 auto *ConditionalConstant = (BI->getSuccessor(0) == BB) ?
519 ConstantInt::getTrue(BB->getContext()) :
520 ConstantInt::getFalse(BB->getContext());
521 AvailableValues.insert(CondInst, ConditionalConstant);
522 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
523 << CondInst->getName() << "' as " << *ConditionalConstant
524 << " in " << BB->getName() << "\n");
525 // Replace all dominated uses with the known value.
527 replaceDominatedUsesWith(CondInst, ConditionalConstant, DT,
528 BasicBlockEdge(Pred, BB))) {
530 NumCSECVP = NumCSECVP + Count;
534 /// LastStore - Keep track of the last non-volatile store that we saw... for
535 /// as long as there in no instruction that reads memory. If we see a store
536 /// to the same location, we delete the dead store. This zaps trivial dead
537 /// stores which can occur in bitfield code among other things.
538 Instruction *LastStore = nullptr;
540 const DataLayout &DL = BB->getModule()->getDataLayout();
542 // See if any instructions in the block can be eliminated. If so, do it. If
543 // not, add them to AvailableValues.
544 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
545 Instruction *Inst = &*I++;
547 // Dead instructions should just be removed.
548 if (isInstructionTriviallyDead(Inst, &TLI)) {
549 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
550 Inst->eraseFromParent();
556 // Skip assume intrinsics, they don't really have side effects (although
557 // they're marked as such to ensure preservation of control dependencies),
558 // and this pass will not disturb any of the assumption's control
560 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
561 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
565 // Skip invariant.start intrinsics since they only read memory, and we can
566 // forward values across it. Also, we dont need to consume the last store
567 // since the semantics of invariant.start allow us to perform DSE of the
568 // last store, if there was a store following invariant.start. Consider:
571 // invariant.start(p)
573 // We can DSE the store to 30, since the store 40 to invariant location p
574 // causes undefined behaviour.
575 if (match(Inst, m_Intrinsic<Intrinsic::invariant_start>()))
578 if (match(Inst, m_Intrinsic<Intrinsic::experimental_guard>())) {
580 dyn_cast<Instruction>(cast<CallInst>(Inst)->getArgOperand(0))) {
581 // The condition we're on guarding here is true for all dominated
583 if (SimpleValue::canHandle(CondI))
584 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
587 // Guard intrinsics read all memory, but don't write any memory.
588 // Accordingly, don't update the generation but consume the last store (to
589 // avoid an incorrect DSE).
594 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
595 // its simpler value.
596 if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) {
597 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
599 if (!Inst->use_empty()) {
600 Inst->replaceAllUsesWith(V);
603 if (isInstructionTriviallyDead(Inst, &TLI)) {
604 Inst->eraseFromParent();
614 // If this is a simple instruction that we can value number, process it.
615 if (SimpleValue::canHandle(Inst)) {
616 // See if the instruction has an available value. If so, use it.
617 if (Value *V = AvailableValues.lookup(Inst)) {
618 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
619 if (auto *I = dyn_cast<Instruction>(V))
621 Inst->replaceAllUsesWith(V);
622 Inst->eraseFromParent();
628 // Otherwise, just remember that this value is available.
629 AvailableValues.insert(Inst, Inst);
633 ParseMemoryInst MemInst(Inst, TTI);
634 // If this is a non-volatile load, process it.
635 if (MemInst.isValid() && MemInst.isLoad()) {
636 // (conservatively) we can't peak past the ordering implied by this
637 // operation, but we can add this load to our set of available values
638 if (MemInst.isVolatile() || !MemInst.isUnordered()) {
643 // If we have an available version of this load, and if it is the right
644 // generation or the load is known to be from an invariant location,
645 // replace this instruction.
647 // A dominating invariant load implies that the location loaded from is
648 // unchanging beginning at the point of the invariant load, so the load
649 // we're CSE'ing _away_ does not need to be invariant, only the available
650 // load we're CSE'ing _to_ does.
651 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
652 if (InVal.DefInst != nullptr &&
653 (InVal.Generation == CurrentGeneration || InVal.IsInvariant) &&
654 InVal.MatchingId == MemInst.getMatchingId() &&
655 // We don't yet handle removing loads with ordering of any kind.
656 !MemInst.isVolatile() && MemInst.isUnordered() &&
657 // We can't replace an atomic load with one which isn't also atomic.
658 InVal.IsAtomic >= MemInst.isAtomic()) {
659 Value *Op = getOrCreateResult(InVal.DefInst, Inst->getType());
661 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
662 << " to: " << *InVal.DefInst << '\n');
663 if (!Inst->use_empty())
664 Inst->replaceAllUsesWith(Op);
665 Inst->eraseFromParent();
672 // Otherwise, remember that we have this instruction.
673 AvailableLoads.insert(
674 MemInst.getPointerOperand(),
675 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
676 MemInst.isAtomic(), MemInst.isInvariantLoad()));
681 // If this instruction may read from memory, forget LastStore.
682 // Load/store intrinsics will indicate both a read and a write to
683 // memory. The target may override this (e.g. so that a store intrinsic
684 // does not read from memory, and thus will be treated the same as a
685 // regular store for commoning purposes).
686 if (Inst->mayReadFromMemory() &&
687 !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
690 // If this is a read-only call, process it.
691 if (CallValue::canHandle(Inst)) {
692 // If we have an available version of this call, and if it is the right
693 // generation, replace this instruction.
694 std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(Inst);
695 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
696 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
697 << " to: " << *InVal.first << '\n');
698 if (!Inst->use_empty())
699 Inst->replaceAllUsesWith(InVal.first);
700 Inst->eraseFromParent();
706 // Otherwise, remember that we have this instruction.
707 AvailableCalls.insert(
708 Inst, std::pair<Instruction *, unsigned>(Inst, CurrentGeneration));
712 // A release fence requires that all stores complete before it, but does
713 // not prevent the reordering of following loads 'before' the fence. As a
714 // result, we don't need to consider it as writing to memory and don't need
715 // to advance the generation. We do need to prevent DSE across the fence,
716 // but that's handled above.
717 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
718 if (FI->getOrdering() == AtomicOrdering::Release) {
719 assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above");
723 // write back DSE - If we write back the same value we just loaded from
724 // the same location and haven't passed any intervening writes or ordering
725 // operations, we can remove the write. The primary benefit is in allowing
726 // the available load table to remain valid and value forward past where
727 // the store originally was.
728 if (MemInst.isValid() && MemInst.isStore()) {
729 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand());
731 InVal.DefInst == getOrCreateResult(Inst, InVal.DefInst->getType()) &&
732 InVal.Generation == CurrentGeneration &&
733 InVal.MatchingId == MemInst.getMatchingId() &&
734 // We don't yet handle removing stores with ordering of any kind.
735 !MemInst.isVolatile() && MemInst.isUnordered()) {
736 assert((!LastStore ||
737 ParseMemoryInst(LastStore, TTI).getPointerOperand() ==
738 MemInst.getPointerOperand()) &&
739 "can't have an intervening store!");
740 DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << *Inst << '\n');
741 Inst->eraseFromParent();
744 // We can avoid incrementing the generation count since we were able
745 // to eliminate this store.
750 // Okay, this isn't something we can CSE at all. Check to see if it is
751 // something that could modify memory. If so, our available memory values
752 // cannot be used so bump the generation count.
753 if (Inst->mayWriteToMemory()) {
756 if (MemInst.isValid() && MemInst.isStore()) {
757 // We do a trivial form of DSE if there are two stores to the same
758 // location with no intervening loads. Delete the earlier store.
759 // At the moment, we don't remove ordered stores, but do remove
760 // unordered atomic stores. There's no special requirement (for
761 // unordered atomics) about removing atomic stores only in favor of
762 // other atomic stores since we we're going to execute the non-atomic
763 // one anyway and the atomic one might never have become visible.
765 ParseMemoryInst LastStoreMemInst(LastStore, TTI);
766 assert(LastStoreMemInst.isUnordered() &&
767 !LastStoreMemInst.isVolatile() &&
768 "Violated invariant");
769 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
770 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
771 << " due to: " << *Inst << '\n');
772 LastStore->eraseFromParent();
777 // fallthrough - we can exploit information about this store
780 // Okay, we just invalidated anything we knew about loaded values. Try
781 // to salvage *something* by remembering that the stored value is a live
782 // version of the pointer. It is safe to forward from volatile stores
783 // to non-volatile loads, so we don't have to check for volatility of
785 AvailableLoads.insert(
786 MemInst.getPointerOperand(),
787 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(),
788 MemInst.isAtomic(), /*IsInvariant=*/false));
790 // Remember that this was the last unordered store we saw for DSE. We
791 // don't yet handle DSE on ordered or volatile stores since we don't
792 // have a good way to model the ordering requirement for following
793 // passes once the store is removed. We could insert a fence, but
794 // since fences are slightly stronger than stores in their ordering,
795 // it's not clear this is a profitable transform. Another option would
796 // be to merge the ordering with that of the post dominating store.
797 if (MemInst.isUnordered() && !MemInst.isVolatile())
808 bool EarlyCSE::run() {
809 // Note, deque is being used here because there is significant performance
810 // gains over vector when the container becomes very large due to the
811 // specific access patterns. For more information see the mailing list
812 // discussion on this:
813 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
814 std::deque<StackNode *> nodesToProcess;
816 bool Changed = false;
818 // Process the root node.
819 nodesToProcess.push_back(new StackNode(
820 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
821 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end()));
823 // Save the current generation.
824 unsigned LiveOutGeneration = CurrentGeneration;
826 // Process the stack.
827 while (!nodesToProcess.empty()) {
828 // Grab the first item off the stack. Set the current generation, remove
829 // the node from the stack, and process it.
830 StackNode *NodeToProcess = nodesToProcess.back();
832 // Initialize class members.
833 CurrentGeneration = NodeToProcess->currentGeneration();
835 // Check if the node needs to be processed.
836 if (!NodeToProcess->isProcessed()) {
838 Changed |= processNode(NodeToProcess->node());
839 NodeToProcess->childGeneration(CurrentGeneration);
840 NodeToProcess->process();
841 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
842 // Push the next child onto the stack.
843 DomTreeNode *child = NodeToProcess->nextChild();
844 nodesToProcess.push_back(
845 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
846 NodeToProcess->childGeneration(), child, child->begin(),
849 // It has been processed, and there are no more children to process,
850 // so delete it and pop it off the stack.
851 delete NodeToProcess;
852 nodesToProcess.pop_back();
854 } // while (!nodes...)
856 // Reset the current generation.
857 CurrentGeneration = LiveOutGeneration;
862 PreservedAnalyses EarlyCSEPass::run(Function &F,
863 FunctionAnalysisManager &AM) {
864 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
865 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
866 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
867 auto &AC = AM.getResult<AssumptionAnalysis>(F);
869 EarlyCSE CSE(TLI, TTI, DT, AC);
872 return PreservedAnalyses::all();
874 // CSE preserves the dominator tree because it doesn't mutate the CFG.
875 // FIXME: Bundle this with other CFG-preservation.
876 PreservedAnalyses PA;
877 PA.preserve<DominatorTreeAnalysis>();
878 PA.preserve<GlobalsAA>();
883 /// \brief A simple and fast domtree-based CSE pass.
885 /// This pass does a simple depth-first walk over the dominator tree,
886 /// eliminating trivially redundant instructions and using instsimplify to
887 /// canonicalize things as it goes. It is intended to be fast and catch obvious
888 /// cases so that instcombine and other passes are more effective. It is
889 /// expected that a later pass of GVN will catch the interesting/hard cases.
890 class EarlyCSELegacyPass : public FunctionPass {
894 EarlyCSELegacyPass() : FunctionPass(ID) {
895 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
898 bool runOnFunction(Function &F) override {
902 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
903 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
904 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
905 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
907 EarlyCSE CSE(TLI, TTI, DT, AC);
912 void getAnalysisUsage(AnalysisUsage &AU) const override {
913 AU.addRequired<AssumptionCacheTracker>();
914 AU.addRequired<DominatorTreeWrapperPass>();
915 AU.addRequired<TargetLibraryInfoWrapperPass>();
916 AU.addRequired<TargetTransformInfoWrapperPass>();
917 AU.addPreserved<GlobalsAAWrapperPass>();
918 AU.setPreservesCFG();
923 char EarlyCSELegacyPass::ID = 0;
925 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSELegacyPass(); }
927 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
929 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
930 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
931 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
932 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
933 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)