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.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/IR/DataLayout.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/RecyclingAllocator.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Transforms/Utils/Local.h"
33 using namespace llvm::PatternMatch;
35 #define DEBUG_TYPE "early-cse"
37 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
38 STATISTIC(NumCSE, "Number of instructions CSE'd");
39 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
40 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
41 STATISTIC(NumDSE, "Number of trivial dead stores removed");
43 static unsigned getHash(const void *V) {
44 return DenseMapInfo<const void*>::getHashValue(V);
47 //===----------------------------------------------------------------------===//
49 //===----------------------------------------------------------------------===//
52 /// SimpleValue - Instances of this struct represent available values in the
53 /// 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 if (isa<OverflowingBinaryOperator>(BinOp)) {
102 // Hash the overflow behavior
104 BinOp->hasNoSignedWrap() * OverflowingBinaryOperator::NoSignedWrap |
105 BinOp->hasNoUnsignedWrap() *
106 OverflowingBinaryOperator::NoUnsignedWrap;
107 return hash_combine(BinOp->getOpcode(), Overflow, LHS, RHS);
110 return hash_combine(BinOp->getOpcode(), LHS, RHS);
113 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) {
114 Value *LHS = CI->getOperand(0);
115 Value *RHS = CI->getOperand(1);
116 CmpInst::Predicate Pred = CI->getPredicate();
117 if (Inst->getOperand(0) > Inst->getOperand(1)) {
119 Pred = CI->getSwappedPredicate();
121 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS);
124 if (CastInst *CI = dyn_cast<CastInst>(Inst))
125 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0));
127 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst))
128 return hash_combine(EVI->getOpcode(), EVI->getOperand(0),
129 hash_combine_range(EVI->idx_begin(), EVI->idx_end()));
131 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst))
132 return hash_combine(IVI->getOpcode(), IVI->getOperand(0),
134 hash_combine_range(IVI->idx_begin(), IVI->idx_end()));
136 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) ||
137 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) ||
138 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) ||
139 isa<ShuffleVectorInst>(Inst)) &&
140 "Invalid/unknown instruction");
142 // Mix in the opcode.
145 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end()));
148 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
149 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
151 if (LHS.isSentinel() || RHS.isSentinel())
154 if (LHSI->getOpcode() != RHSI->getOpcode())
156 if (LHSI->isIdenticalTo(RHSI))
159 // If we're not strictly identical, we still might be a commutable instruction
160 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) {
161 if (!LHSBinOp->isCommutative())
164 assert(isa<BinaryOperator>(RHSI) &&
165 "same opcode, but different instruction type?");
166 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI);
168 // Check overflow attributes
169 if (isa<OverflowingBinaryOperator>(LHSBinOp)) {
170 assert(isa<OverflowingBinaryOperator>(RHSBinOp) &&
171 "same opcode, but different operator type?");
172 if (LHSBinOp->hasNoUnsignedWrap() != RHSBinOp->hasNoUnsignedWrap() ||
173 LHSBinOp->hasNoSignedWrap() != RHSBinOp->hasNoSignedWrap())
178 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) &&
179 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0);
181 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) {
182 assert(isa<CmpInst>(RHSI) &&
183 "same opcode, but different instruction type?");
184 CmpInst *RHSCmp = cast<CmpInst>(RHSI);
186 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) &&
187 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) &&
188 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate();
194 //===----------------------------------------------------------------------===//
196 //===----------------------------------------------------------------------===//
199 /// CallValue - Instances of this struct represent available call values in
200 /// the scoped hash table.
204 CallValue(Instruction *I) : Inst(I) {
205 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
208 bool isSentinel() const {
209 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() ||
210 Inst == DenseMapInfo<Instruction *>::getTombstoneKey();
213 static bool canHandle(Instruction *Inst) {
214 // Don't value number anything that returns void.
215 if (Inst->getType()->isVoidTy())
218 CallInst *CI = dyn_cast<CallInst>(Inst);
219 if (!CI || !CI->onlyReadsMemory())
227 template <> struct DenseMapInfo<CallValue> {
228 static inline CallValue getEmptyKey() {
229 return DenseMapInfo<Instruction *>::getEmptyKey();
231 static inline CallValue getTombstoneKey() {
232 return DenseMapInfo<Instruction *>::getTombstoneKey();
234 static unsigned getHashValue(CallValue Val);
235 static bool isEqual(CallValue LHS, CallValue RHS);
239 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
240 Instruction *Inst = Val.Inst;
241 // Hash in all of the operands as pointers.
243 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) {
244 assert(!Inst->getOperand(i)->getType()->isMetadataTy() &&
245 "Cannot value number calls with metadata operands");
246 Res ^= getHash(Inst->getOperand(i)) << (i & 0xF);
249 // Mix in the opcode.
250 return (Res << 1) ^ Inst->getOpcode();
253 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
254 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
255 if (LHS.isSentinel() || RHS.isSentinel())
257 return LHSI->isIdenticalTo(RHSI);
260 //===----------------------------------------------------------------------===//
262 //===----------------------------------------------------------------------===//
266 /// EarlyCSE - This pass does a simple depth-first walk over the dominator
267 /// tree, eliminating trivially redundant instructions and using instsimplify
268 /// to canonicalize things as it goes. It is intended to be fast and catch
269 /// obvious cases so that instcombine and other passes are more effective. It
270 /// is expected that a later pass of GVN will catch the interesting/hard
272 class EarlyCSE : public FunctionPass {
274 const DataLayout *DL;
275 const TargetLibraryInfo *TLI;
278 typedef RecyclingAllocator<
279 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy;
280 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>,
281 AllocatorTy> ScopedHTType;
283 /// AvailableValues - This scoped hash table contains the current values of
284 /// all of our simple scalar expressions. As we walk down the domtree, we
285 /// look to see if instructions are in this: if so, we replace them with what
286 /// we find, otherwise we insert them so that dominated values can succeed in
288 ScopedHTType *AvailableValues;
290 /// AvailableLoads - This scoped hash table contains the current values
291 /// of loads. This allows us to get efficient access to dominating loads when
292 /// we have a fully redundant load. In addition to the most recent load, we
293 /// keep track of a generation count of the read, which is compared against
294 /// the current generation count. The current generation count is
295 /// incremented after every possibly writing memory operation, which ensures
296 /// that we only CSE loads with other loads that have no intervening store.
297 typedef RecyclingAllocator<
299 ScopedHashTableVal<Value *, std::pair<Value *, unsigned>>>
301 typedef ScopedHashTable<Value *, std::pair<Value *, unsigned>,
302 DenseMapInfo<Value *>, LoadMapAllocator> LoadHTType;
303 LoadHTType *AvailableLoads;
305 /// AvailableCalls - This scoped hash table contains the current values
306 /// of read-only call values. It uses the same generation count as loads.
307 typedef ScopedHashTable<CallValue, std::pair<Value *, unsigned>> CallHTType;
308 CallHTType *AvailableCalls;
310 /// CurrentGeneration - This is the current generation of the memory value.
311 unsigned CurrentGeneration;
314 explicit EarlyCSE() : FunctionPass(ID) {
315 initializeEarlyCSEPass(*PassRegistry::getPassRegistry());
318 bool runOnFunction(Function &F) override;
321 // NodeScope - almost a POD, but needs to call the constructors for the
322 // scoped hash tables so that a new scope gets pushed on. These are RAII so
323 // that the 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 &) LLVM_DELETED_FUNCTION;
333 void operator=(const NodeScope &) LLVM_DELETED_FUNCTION;
335 ScopedHTType::ScopeTy Scope;
336 LoadHTType::ScopeTy LoadScope;
337 CallHTType::ScopeTy CallScope;
340 // StackNode - contains all the needed information to create a stack for
341 // doing a depth first tranversal of the tree. This includes scopes for
342 // values, loads, and calls as well as the generation. There is a child
343 // iterator so that the children do not need to be store spearately.
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 &) LLVM_DELETED_FUNCTION;
370 void operator=(const StackNode &) LLVM_DELETED_FUNCTION;
373 unsigned CurrentGeneration;
374 unsigned ChildGeneration;
376 DomTreeNode::iterator ChildIter;
377 DomTreeNode::iterator EndIter;
382 bool processNode(DomTreeNode *Node);
384 // This transformation requires dominator postdominator info
385 void getAnalysisUsage(AnalysisUsage &AU) const override {
386 AU.addRequired<AssumptionCacheTracker>();
387 AU.addRequired<DominatorTreeWrapperPass>();
388 AU.addRequired<TargetLibraryInfoWrapperPass>();
389 AU.setPreservesCFG();
394 char EarlyCSE::ID = 0;
396 // createEarlyCSEPass - The public interface to this file.
397 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSE(); }
399 INITIALIZE_PASS_BEGIN(EarlyCSE, "early-cse", "Early CSE", false, false)
400 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
401 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
402 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
403 INITIALIZE_PASS_END(EarlyCSE, "early-cse", "Early CSE", false, false)
405 bool EarlyCSE::processNode(DomTreeNode *Node) {
406 BasicBlock *BB = Node->getBlock();
408 // If this block has a single predecessor, then the predecessor is the parent
409 // of the domtree node and all of the live out memory values are still current
410 // in this block. If this block has multiple predecessors, then they could
411 // have invalidated the live-out memory values of our parent value. For now,
412 // just be conservative and invalidate memory if this block has multiple
414 if (!BB->getSinglePredecessor())
417 /// LastStore - Keep track of the last non-volatile store that we saw... for
418 /// as long as there in no instruction that reads memory. If we see a store
419 /// to the same location, we delete the dead store. This zaps trivial dead
420 /// stores which can occur in bitfield code among other things.
421 StoreInst *LastStore = nullptr;
423 bool Changed = false;
425 // See if any instructions in the block can be eliminated. If so, do it. If
426 // not, add them to AvailableValues.
427 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
428 Instruction *Inst = I++;
430 // Dead instructions should just be removed.
431 if (isInstructionTriviallyDead(Inst, TLI)) {
432 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
433 Inst->eraseFromParent();
439 // Skip assume intrinsics, they don't really have side effects (although
440 // they're marked as such to ensure preservation of control dependencies),
441 // and this pass will not disturb any of the assumption's control
443 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
444 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
448 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
449 // its simpler value.
450 if (Value *V = SimplifyInstruction(Inst, DL, TLI, DT, AC)) {
451 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
452 Inst->replaceAllUsesWith(V);
453 Inst->eraseFromParent();
459 // If this is a simple instruction that we can value number, process it.
460 if (SimpleValue::canHandle(Inst)) {
461 // See if the instruction has an available value. If so, use it.
462 if (Value *V = AvailableValues->lookup(Inst)) {
463 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
464 Inst->replaceAllUsesWith(V);
465 Inst->eraseFromParent();
471 // Otherwise, just remember that this value is available.
472 AvailableValues->insert(Inst, Inst);
476 // If this is a non-volatile load, process it.
477 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
478 // Ignore volatile loads.
479 if (!LI->isSimple()) {
484 // If we have an available version of this load, and if it is the right
485 // generation, replace this instruction.
486 std::pair<Value *, unsigned> InVal =
487 AvailableLoads->lookup(Inst->getOperand(0));
488 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
489 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst
490 << " to: " << *InVal.first << '\n');
491 if (!Inst->use_empty())
492 Inst->replaceAllUsesWith(InVal.first);
493 Inst->eraseFromParent();
499 // Otherwise, remember that we have this instruction.
500 AvailableLoads->insert(Inst->getOperand(0), std::pair<Value *, unsigned>(
501 Inst, CurrentGeneration));
506 // If this instruction may read from memory, forget LastStore.
507 if (Inst->mayReadFromMemory())
510 // If this is a read-only call, process it.
511 if (CallValue::canHandle(Inst)) {
512 // If we have an available version of this call, and if it is the right
513 // generation, replace this instruction.
514 std::pair<Value *, unsigned> InVal = AvailableCalls->lookup(Inst);
515 if (InVal.first != nullptr && InVal.second == CurrentGeneration) {
516 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst
517 << " to: " << *InVal.first << '\n');
518 if (!Inst->use_empty())
519 Inst->replaceAllUsesWith(InVal.first);
520 Inst->eraseFromParent();
526 // Otherwise, remember that we have this instruction.
527 AvailableCalls->insert(
528 Inst, std::pair<Value *, unsigned>(Inst, CurrentGeneration));
532 // Okay, this isn't something we can CSE at all. Check to see if it is
533 // something that could modify memory. If so, our available memory values
534 // cannot be used so bump the generation count.
535 if (Inst->mayWriteToMemory()) {
538 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
539 // We do a trivial form of DSE if there are two stores to the same
540 // location with no intervening loads. Delete the earlier store.
542 LastStore->getPointerOperand() == SI->getPointerOperand()) {
543 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore
544 << " due to: " << *Inst << '\n');
545 LastStore->eraseFromParent();
549 // fallthrough - we can exploit information about this store
552 // Okay, we just invalidated anything we knew about loaded values. Try
553 // to salvage *something* by remembering that the stored value is a live
554 // version of the pointer. It is safe to forward from volatile stores
555 // to non-volatile loads, so we don't have to check for volatility of
557 AvailableLoads->insert(SI->getPointerOperand(),
558 std::pair<Value *, unsigned>(
559 SI->getValueOperand(), CurrentGeneration));
561 // Remember that this was the last store we saw for DSE.
571 bool EarlyCSE::runOnFunction(Function &F) {
572 if (skipOptnoneFunction(F))
575 // Note, deque is being used here because there is significant performance
576 // gains over vector when the container becomes very large due to the
577 // specific access patterns. For more information see the mailing list
578 // discussion on this:
579 // http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html
580 std::deque<StackNode *> nodesToProcess;
582 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
583 DL = DLP ? &DLP->getDataLayout() : nullptr;
584 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
585 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
586 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
588 // Tables that the pass uses when walking the domtree.
589 ScopedHTType AVTable;
590 AvailableValues = &AVTable;
591 LoadHTType LoadTable;
592 AvailableLoads = &LoadTable;
593 CallHTType CallTable;
594 AvailableCalls = &CallTable;
596 CurrentGeneration = 0;
597 bool Changed = false;
599 // Process the root node.
600 nodesToProcess.push_back(new StackNode(
601 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration,
602 DT->getRootNode(), DT->getRootNode()->begin(), DT->getRootNode()->end()));
604 // Save the current generation.
605 unsigned LiveOutGeneration = CurrentGeneration;
607 // Process the stack.
608 while (!nodesToProcess.empty()) {
609 // Grab the first item off the stack. Set the current generation, remove
610 // the node from the stack, and process it.
611 StackNode *NodeToProcess = nodesToProcess.back();
613 // Initialize class members.
614 CurrentGeneration = NodeToProcess->currentGeneration();
616 // Check if the node needs to be processed.
617 if (!NodeToProcess->isProcessed()) {
619 Changed |= processNode(NodeToProcess->node());
620 NodeToProcess->childGeneration(CurrentGeneration);
621 NodeToProcess->process();
622 } else if (NodeToProcess->childIter() != NodeToProcess->end()) {
623 // Push the next child onto the stack.
624 DomTreeNode *child = NodeToProcess->nextChild();
625 nodesToProcess.push_back(
626 new StackNode(AvailableValues, AvailableLoads, AvailableCalls,
627 NodeToProcess->childGeneration(), child, child->begin(),
630 // It has been processed, and there are no more children to process,
631 // so delete it and pop it off the stack.
632 delete NodeToProcess;
633 nodesToProcess.pop_back();
635 } // while (!nodes...)
637 // Reset the current generation.
638 CurrentGeneration = LiveOutGeneration;