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/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/RecyclingAllocator.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Transforms/Scalar.h"
33 #include "llvm/Transforms/Utils/Local.h"
36 using namespace llvm::PatternMatch;
38 #define DEBUG_TYPE "early-cse"
40 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
41 STATISTIC(NumCSE, "Number of instructions CSE'd");
42 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
43 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
44 STATISTIC(NumDSE, "Number of trivial dead stores removed");
46 //===----------------------------------------------------------------------===//
48 //===----------------------------------------------------------------------===//
51 /// \brief Struct representing the available values in the scoped hash table.
55 SimpleValue(Instruction *I) : Inst(I) {
56 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
59 bool isSentinel() const {
60 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
61 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
64 static bool canHandle(Instruction *Inst) {
65 // This can only handle non-void readnone functions.
66 if (CallInst *CI = dyn_cast<CallInst>(Inst))
67 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy();
68 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) ||
69 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) ||
70 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
71 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
72 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst);
78 template <> struct DenseMapInfo<SimpleValue> {
79 static inline SimpleValue getEmptyKey() {
80 return DenseMapInfo<Instruction *>::getEmptyKey();
82 static inline SimpleValue getTombstoneKey() {
83 return DenseMapInfo<Instruction *>::getTombstoneKey();
85 static unsigned getHashValue(SimpleValue Val);
86 static bool isEqual(SimpleValue LHS, SimpleValue RHS);
90 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
91 Instruction *Inst = Val.Inst;
92 // Hash in all of the operands as pointers.
93 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) {
94 Value *LHS = BinOp->getOperand(0);
95 Value *RHS = BinOp->getOperand(1);
96 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1))
99 if (isa<OverflowingBinaryOperator>(BinOp)) {
100 // Hash the overflow behavior
102 BinOp->hasNoSignedWrap() * OverflowingBinaryOperator::NoSignedWrap |
103 BinOp->hasNoUnsignedWrap() *
104 OverflowingBinaryOperator::NoUnsignedWrap;
105 return hash_combine(BinOp->getOpcode(), Overflow, LHS, RHS);
108 return hash_combine(BinOp->getOpcode(), LHS, RHS);
111 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
112 Value *LHS = CI->getOperand(0);
113 Value *RHS = CI->getOperand(1);
114 CmpInst::Predicate Pred = CI->getPredicate();
115 if (Inst->getOperand(0) > Inst->getOperand(1)) {
117 Pred = CI->getSwappedPredicate();
119 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
122 if (CastInst *CI = dyn_cast<CastInst>(Inst))
123 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
125 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
126 return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
127 hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
129 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
130 return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
132 hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
134 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) ||
135 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) ||
136 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
137 isa<ShuffleVectorInst>(Inst)) &&
138 "Invalid/unknown instruction");
140 // Mix in the opcode.
143 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
146 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
147 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
149 if (LHS.isSentinel() || RHS.isSentinel())
152 if (LHSI->getOpcode() != RHSI->getOpcode())
154 if (LHSI->isIdenticalTo(RHSI))
157 // If we're not strictly identical, we still might be a commutable instruction
158 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
159 if (!LHSBinOp->isCommutative())
162 assert(isa<BinaryOperator>(RHSI) &&
163 "same opcode, but different instruction type?");
164 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
166 // Check overflow attributes
167 if (isa<OverflowingBinaryOperator>(LHSBinOp)) {
168 assert(isa<OverflowingBinaryOperator>(RHSBinOp) &&
169 "same opcode, but different operator type?");
170 if (LHSBinOp->hasNoUnsignedWrap() != RHSBinOp->hasNoUnsignedWrap() ||
171 LHSBinOp->hasNoSignedWrap() != RHSBinOp->hasNoSignedWrap())
176 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
177 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
179 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
180 assert(isa<CmpInst>(RHSI) &&
181 "same opcode, but different instruction type?");
182 CmpInst *RHSCmp = cast<CmpInst>(RHSI);
184 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
185 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
186 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
192 //===----------------------------------------------------------------------===//
194 //===----------------------------------------------------------------------===//
197 /// \brief Struct representing the available call values in the scoped hash
202 CallValue(Instruction *I) : Inst(I) {
203 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
206 bool isSentinel() const {
207 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
208 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
211 static bool canHandle(Instruction *Inst) {
212 // Don't value number anything that returns void.
213 if (Inst->getType()->isVoidTy())
216 CallInst *CI = dyn_cast<CallInst>(Inst);
217 if (!CI || !CI->onlyReadsMemory())
225 template <> struct DenseMapInfo<CallValue> {
226 static inline CallValue getEmptyKey() {
227 return DenseMapInfo<Instruction *>::getEmptyKey();
229 static inline CallValue getTombstoneKey() {
230 return DenseMapInfo<Instruction *>::getTombstoneKey();
232 static unsigned getHashValue(CallValue Val);
233 static bool isEqual(CallValue LHS, CallValue RHS);
237 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
238 Instruction *Inst = Val.Inst;
239 // Hash all of the operands as pointers and mix in the opcode.
242 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
245 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
246 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
247 if (LHS.isSentinel() || RHS.isSentinel())
249 return LHSI->isIdenticalTo(RHSI);
252 //===----------------------------------------------------------------------===//
253 // EarlyCSE implementation
254 //===----------------------------------------------------------------------===//
257 /// \brief A simple and fast domtree-based CSE pass.
259 /// This pass does a simple depth-first walk over the dominator tree,
260 /// eliminating trivially redundant instructions and using instsimplify to
261 /// canonicalize things as it goes. It is intended to be fast and catch obvious
262 /// cases so that instcombine and other passes are more effective. It is
263 /// 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.
292 typedef RecyclingAllocator<
294 ScopedHashTableVal<Value *, std::pair<Value *, unsigned>>>
296 typedef ScopedHashTable<Value *, std::pair<Value *, unsigned>,
297 DenseMapInfo<Value *>, LoadMapAllocator> LoadHTType;
298 LoadHTType AvailableLoads;
300 /// \brief A scoped hash table of the current values of read-only call
303 /// It uses the same generation count as loads.
304 typedef ScopedHashTable<CallValue, std::pair<Value *, unsigned>> CallHTType;
305 CallHTType AvailableCalls;
307 /// \brief This is the current generation of the memory value.
308 unsigned CurrentGeneration;
310 /// \brief Set up the EarlyCSE runner for a particular function.
311 EarlyCSE(Function &F, const TargetLibraryInfo &TLI,
312 const TargetTransformInfo &TTI, DominatorTree &DT,
314 : F(F), TLI(TLI), TTI(TTI), DT(DT), AC(AC), CurrentGeneration(0) {}
319 // Almost a POD, but needs to call the constructors for the scoped hash
320 // tables so that a new scope gets pushed on. These are RAII so that the
321 // scope gets popped when the NodeScope is destroyed.
324 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
325 CallHTType &AvailableCalls)
326 : Scope(AvailableValues), LoadScope(AvailableLoads),
327 CallScope(AvailableCalls) {}
330 NodeScope(const NodeScope &) = delete;
331 void operator=(const NodeScope &) = delete;
333 ScopedHTType::ScopeTy Scope;
334 LoadHTType::ScopeTy LoadScope;
335 CallHTType::ScopeTy CallScope;
338 // Contains all the needed information to create a stack for doing a depth
339 // first tranversal of the tree. This includes scopes for values, loads, and
340 // calls as well as the generation. There is a child iterator so that the
341 // children do not need to be store spearately.
344 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads,
345 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n,
346 DomTreeNode::iterator child, DomTreeNode::iterator end)
347 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child),
348 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls),
352 unsigned currentGeneration() { return CurrentGeneration; }
353 unsigned childGeneration() { return ChildGeneration; }
354 void childGeneration(unsigned generation) { ChildGeneration = generation; }
355 DomTreeNode *node() { return Node; }
356 DomTreeNode::iterator childIter() { return ChildIter; }
357 DomTreeNode *nextChild() {
358 DomTreeNode *child = *ChildIter;
362 DomTreeNode::iterator end() { return EndIter; }
363 bool isProcessed() { return Processed; }
364 void process() { Processed = true; }
367 StackNode(const StackNode &) = delete;
368 void operator=(const StackNode &) = delete;
371 unsigned CurrentGeneration;
372 unsigned ChildGeneration;
374 DomTreeNode::iterator ChildIter;
375 DomTreeNode::iterator EndIter;
380 /// \brief Wrapper class to handle memory instructions, including loads,
381 /// stores and intrinsic loads and stores defined by the target.
382 class ParseMemoryInst {
384 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI)
385 : Load(false), Store(false), Vol(false), MayReadFromMemory(false),
386 MayWriteToMemory(false), MatchingId(-1), Ptr(nullptr) {
387 MayReadFromMemory = Inst->mayReadFromMemory();
388 MayWriteToMemory = Inst->mayWriteToMemory();
389 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
390 MemIntrinsicInfo Info;
391 if (!TTI.getTgtMemIntrinsic(II, Info))
393 if (Info.NumMemRefs == 1) {
394 Store = Info.WriteMem;
396 MatchingId = Info.MatchingId;
397 MayReadFromMemory = Info.ReadMem;
398 MayWriteToMemory = Info.WriteMem;
402 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
404 Vol = !LI->isSimple();
405 Ptr = LI->getPointerOperand();
406 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
408 Vol = !SI->isSimple();
409 Ptr = SI->getPointerOperand();
412 bool isLoad() { return Load; }
413 bool isStore() { return Store; }
414 bool isVolatile() { return Vol; }
415 bool isMatchingMemLoc(const ParseMemoryInst &Inst) {
416 return Ptr == Inst.Ptr && MatchingId == Inst.MatchingId;
418 bool isValid() { return Ptr != nullptr; }
419 int getMatchingId() { return MatchingId; }
420 Value *getPtr() { return Ptr; }
421 bool mayReadFromMemory() { return MayReadFromMemory; }
422 bool mayWriteToMemory() { return MayWriteToMemory; }
428 bool MayReadFromMemory;
429 bool MayWriteToMemory;
430 // For regular (non-intrinsic) loads/stores, this is set to -1. For
431 // intrinsic loads/stores, the id is retrieved from the corresponding
432 // field in the MemIntrinsicInfo structure. That field contains
433 // non-negative values only.
438 bool processNode(DomTreeNode *Node);
440 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const {
441 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
443 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
444 return SI->getValueOperand();
445 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported");
446 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst),
452 bool EarlyCSE::processNode(DomTreeNode *Node) {
453 BasicBlock *BB = Node->getBlock();
455 // If this block has a single predecessor, then the predecessor is the parent
456 // of the domtree node and all of the live out memory values are still current
457 // in this block. If this block has multiple predecessors, then they could
458 // have invalidated the live-out memory values of our parent value. For now,
459 // just be conservative and invalidate memory if this block has multiple
461 if (!BB->getSinglePredecessor())
464 // If this node has a single predecessor which ends in a conditional branch,
465 // we can infer the value of the branch condition given that we took this
466 // path. We need the single predeccesor to ensure there's not another path
467 // which reaches this block where the condition might hold a different
468 // value. Since we're adding this to the scoped hash table (like any other
469 // def), it will have been popped if we encounter a future merge block.
470 if (BasicBlock *Pred = BB->getSinglePredecessor())
471 if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
472 if (BI->isConditional())
473 if (auto *CondInst = dyn_cast<Instruction>(BI->getCondition()))
474 if (SimpleValue::canHandle(CondInst)) {
475 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB);
476 auto *ConditionalConstant = (BI->getSuccessor(0) == BB) ?
477 ConstantInt::getTrue(BB->getContext()) :
478 ConstantInt::getFalse(BB->getContext());
479 AvailableValues.insert(CondInst, ConditionalConstant);
480 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '"
481 << CondInst->getName() << "' as " << *ConditionalConstant
482 << " in " << BB->getName() << "\n");
483 // Replace all dominated uses with the known value
484 replaceDominatedUsesWith(CondInst, ConditionalConstant, DT,
485 BasicBlockEdge(Pred, BB));
488 /// LastStore - Keep track of the last non-volatile store that we saw... for
489 /// as long as there in no instruction that reads memory. If we see a store
490 /// to the same location, we delete the dead store. This zaps trivial dead
491 /// stores which can occur in bitfield code among other things.
492 Instruction *LastStore = nullptr;
494 bool Changed = false;
495 const DataLayout &DL = BB->getModule()->getDataLayout();
497 // See if any instructions in the block can be eliminated. If so, do it. If
498 // not, add them to AvailableValues.
499 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
500 Instruction *Inst = I++;
502 // Dead instructions should just be removed.
503 if (isInstructionTriviallyDead(Inst, &TLI)) {
504 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
505 Inst->eraseFromParent();
511 // Skip assume intrinsics, they don't really have side effects (although
512 // they're marked as such to ensure preservation of control dependencies),
513 // and this pass will not disturb any of the assumption's control
515 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
516 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
520 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
521 // its simpler value.
522 if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) {
523 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
524 Inst->replaceAllUsesWith(V);
525 Inst->eraseFromParent();
531 // If this is a simple instruction that we can value number, process it.
532 if (SimpleValue::canHandle(Inst)) {
533 // See if the instruction has an available value. If so, use it.
534 if (Value *V = AvailableValues.lookup(Inst)) {
535 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
536 Inst->replaceAllUsesWith(V);
537 Inst->eraseFromParent();
543 // Otherwise, just remember that this value is available.
544 AvailableValues.insert(Inst, Inst);
548 ParseMemoryInst MemInst(Inst, TTI);
549 // If this is a non-volatile load, process it.
550 if (MemInst.isValid() && MemInst.isLoad()) {
551 // Ignore volatile loads.
552 if (MemInst.isVolatile()) {
554 // Don't CSE across synchronization boundaries.
555 if (Inst->mayWriteToMemory())
560 // If we have an available version of this load, and if it is the right
561 // generation, replace this instruction.
562 std::pair<Value *, unsigned> InVal =
563 AvailableLoads.lookup(MemInst.getPtr());
564 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
565 Value *Op = getOrCreateResult(InVal.first, Inst->getType());
567 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
568 << " to: " << *InVal.first << '\n');
569 if (!Inst->use_empty())
570 Inst->replaceAllUsesWith(Op);
571 Inst->eraseFromParent();
578 // Otherwise, remember that we have this instruction.
579 AvailableLoads.insert(MemInst.getPtr(), std::pair<Value *, unsigned>(
580 Inst, CurrentGeneration));
585 // If this instruction may read from memory, forget LastStore.
586 // Load/store intrinsics will indicate both a read and a write to
587 // memory. The target may override this (e.g. so that a store intrinsic
588 // does not read from memory, and thus will be treated the same as a
589 // regular store for commoning purposes).
590 if (Inst->mayReadFromMemory() &&
591 !(MemInst.isValid() && !MemInst.mayReadFromMemory()))
594 // If this is a read-only call, process it.
595 if (CallValue::canHandle(Inst)) {
596 // If we have an available version of this call, and if it is the right
597 // generation, replace this instruction.
598 std::pair<Value *, unsigned> InVal = AvailableCalls.lookup(Inst);
599 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
600 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
601 << " to: " << *InVal.first << '\n');
602 if (!Inst->use_empty())
603 Inst->replaceAllUsesWith(InVal.first);
604 Inst->eraseFromParent();
610 // Otherwise, remember that we have this instruction.
611 AvailableCalls.insert(
612 Inst, std::pair<Value *, unsigned>(Inst, CurrentGeneration));
616 // Okay, this isn't something we can CSE at all. Check to see if it is
617 // something that could modify memory. If so, our available memory values
618 // cannot be used so bump the generation count.
619 if (Inst->mayWriteToMemory()) {
622 if (MemInst.isValid() && MemInst.isStore()) {
623 // We do a trivial form of DSE if there are two stores to the same
624 // location with no intervening loads. Delete the earlier store.
626 ParseMemoryInst LastStoreMemInst(LastStore, TTI);
627 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) {
628 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
629 << " due to: " << *Inst << '\n');
630 LastStore->eraseFromParent();
635 // fallthrough - we can exploit information about this store
638 // Okay, we just invalidated anything we knew about loaded values. Try
639 // to salvage *something* by remembering that the stored value is a live
640 // version of the pointer. It is safe to forward from volatile stores
641 // to non-volatile loads, so we don't have to check for volatility of
643 AvailableLoads.insert(MemInst.getPtr(), std::pair<Value *, unsigned>(
644 Inst, CurrentGeneration));
646 // Remember that this was the last store we saw for DSE.
647 if (!MemInst.isVolatile())
656 bool EarlyCSE::run() {
657 // Note, deque is being used here because there is significant performance
658 // gains over vector when the container becomes very large due to the
659 // specific access patterns. For more information see the mailing list
660 // discussion on this:
661 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
662 std::deque<StackNode *> nodesToProcess;
664 bool Changed = false;
666 // Process the root node.
667 nodesToProcess.push_back(new StackNode(
668 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
669 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end()));
671 // Save the current generation.
672 unsigned LiveOutGeneration = CurrentGeneration;
674 // Process the stack.
675 while (!nodesToProcess.empty()) {
676 // Grab the first item off the stack. Set the current generation, remove
677 // the node from the stack, and process it.
678 StackNode *NodeToProcess = nodesToProcess.back();
680 // Initialize class members.
681 CurrentGeneration = NodeToProcess->currentGeneration();
683 // Check if the node needs to be processed.
684 if (!NodeToProcess->isProcessed()) {
686 Changed |= processNode(NodeToProcess->node());
687 NodeToProcess->childGeneration(CurrentGeneration);
688 NodeToProcess->process();
689 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
690 // Push the next child onto the stack.
691 DomTreeNode *child = NodeToProcess->nextChild();
692 nodesToProcess.push_back(
693 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
694 NodeToProcess->childGeneration(), child, child->begin(),
697 // It has been processed, and there are no more children to process,
698 // so delete it and pop it off the stack.
699 delete NodeToProcess;
700 nodesToProcess.pop_back();
702 } // while (!nodes...)
704 // Reset the current generation.
705 CurrentGeneration = LiveOutGeneration;
710 PreservedAnalyses EarlyCSEPass::run(Function &F,
711 AnalysisManager<Function> *AM) {
712 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F);
713 auto &TTI = AM->getResult<TargetIRAnalysis>(F);
714 auto &DT = AM->getResult<DominatorTreeAnalysis>(F);
715 auto &AC = AM->getResult<AssumptionAnalysis>(F);
717 EarlyCSE CSE(F, TLI, TTI, DT, AC);
720 return PreservedAnalyses::all();
722 // CSE preserves the dominator tree because it doesn't mutate the CFG.
723 // FIXME: Bundle this with other CFG-preservation.
724 PreservedAnalyses PA;
725 PA.preserve<DominatorTreeAnalysis>();
730 /// \brief A simple and fast domtree-based CSE pass.
732 /// This pass does a simple depth-first walk over the dominator tree,
733 /// eliminating trivially redundant instructions and using instsimplify to
734 /// canonicalize things as it goes. It is intended to be fast and catch obvious
735 /// cases so that instcombine and other passes are more effective. It is
736 /// expected that a later pass of GVN will catch the interesting/hard cases.
737 class EarlyCSELegacyPass : public FunctionPass {
741 EarlyCSELegacyPass() : FunctionPass(ID) {
742 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry());
745 bool runOnFunction(Function &F) override {
746 if (skipOptnoneFunction(F))
749 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
750 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
751 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
752 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
754 EarlyCSE CSE(F, TLI, TTI, DT, AC);
759 void getAnalysisUsage(AnalysisUsage &AU) const override {
760 AU.addRequired<AssumptionCacheTracker>();
761 AU.addRequired<DominatorTreeWrapperPass>();
762 AU.addRequired<TargetLibraryInfoWrapperPass>();
763 AU.addRequired<TargetTransformInfoWrapperPass>();
764 AU.setPreservesCFG();
769 char EarlyCSELegacyPass::ID = 0;
771 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSELegacyPass(); }
773 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false,
775 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
776 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
777 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
778 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
779 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false)