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 #define DEBUG_TYPE "early-cse"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/Dominators.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/RecyclingAllocator.h"
25 #include "llvm/ADT/ScopedHashTable.h"
26 #include "llvm/ADT/Statistic.h"
29 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd");
30 STATISTIC(NumCSE, "Number of instructions CSE'd");
31 STATISTIC(NumCSELoad, "Number of load instructions CSE'd");
32 STATISTIC(NumCSECall, "Number of call instructions CSE'd");
33 STATISTIC(NumDSE, "Number of trivial dead stores removed");
35 static unsigned getHash(const void *V) {
36 return DenseMapInfo<const void*>::getHashValue(V);
39 //===----------------------------------------------------------------------===//
41 //===----------------------------------------------------------------------===//
44 /// SimpleValue - Instances of this struct represent available values in the
45 /// scoped hash table.
49 SimpleValue(Instruction *I) : Inst(I) {
50 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
53 bool isSentinel() const {
54 return Inst == DenseMapInfo<Instruction*>::getEmptyKey() ||
55 Inst == DenseMapInfo<Instruction*>::getTombstoneKey();
58 static bool canHandle(Instruction *Inst) {
59 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) ||
60 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) ||
61 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
62 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) ||
63 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst);
69 // SimpleValue is POD.
70 template<> struct isPodLike<SimpleValue> {
71 static const bool value = true;
74 template<> struct DenseMapInfo<SimpleValue> {
75 static inline SimpleValue getEmptyKey() {
76 return DenseMapInfo<Instruction*>::getEmptyKey();
78 static inline SimpleValue getTombstoneKey() {
79 return DenseMapInfo<Instruction*>::getTombstoneKey();
81 static unsigned getHashValue(SimpleValue Val);
82 static bool isEqual(SimpleValue LHS, SimpleValue RHS);
86 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) {
87 Instruction *Inst = Val.Inst;
89 // Hash in all of the operands as pointers.
91 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
92 Res ^= getHash(Inst->getOperand(i)) << i;
94 if (CastInst *CI = dyn_cast<CastInst>(Inst))
95 Res ^= getHash(CI->getType());
96 else if (CmpInst *CI = dyn_cast<CmpInst>(Inst))
97 Res ^= CI->getPredicate();
98 else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) {
99 for (ExtractValueInst::idx_iterator I = EVI->idx_begin(),
100 E = EVI->idx_end(); I != E; ++I)
102 } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) {
103 for (InsertValueInst::idx_iterator I = IVI->idx_begin(),
104 E = IVI->idx_end(); I != E; ++I)
107 // nothing extra to hash in.
108 assert((isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) ||
109 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) ||
110 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst)) &&
111 "Invalid/unknown instruction");
114 // Mix in the opcode.
115 return (Res << 1) ^ Inst->getOpcode();
118 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) {
119 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
121 if (LHS.isSentinel() || RHS.isSentinel())
124 if (LHSI->getOpcode() != RHSI->getOpcode()) return false;
125 return LHSI->isIdenticalTo(RHSI);
128 //===----------------------------------------------------------------------===//
130 //===----------------------------------------------------------------------===//
133 /// CallValue - Instances of this struct represent available call values in
134 /// the scoped hash table.
138 CallValue(Instruction *I) : Inst(I) {
139 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!");
142 bool isSentinel() const {
143 return Inst == DenseMapInfo<Instruction*>::getEmptyKey() ||
144 Inst == DenseMapInfo<Instruction*>::getTombstoneKey();
147 static bool canHandle(Instruction *Inst) {
148 // Don't value number anything that returns void.
149 if (Inst->getType()->isVoidTy())
152 CallInst *CI = dyn_cast<CallInst>(Inst);
153 if (CI == 0 || !CI->onlyReadsMemory())
162 template<> struct isPodLike<CallValue> {
163 static const bool value = true;
166 template<> struct DenseMapInfo<CallValue> {
167 static inline CallValue getEmptyKey() {
168 return DenseMapInfo<Instruction*>::getEmptyKey();
170 static inline CallValue getTombstoneKey() {
171 return DenseMapInfo<Instruction*>::getTombstoneKey();
173 static unsigned getHashValue(CallValue Val);
174 static bool isEqual(CallValue LHS, CallValue RHS);
177 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
178 Instruction *Inst = Val.Inst;
179 // Hash in all of the operands as pointers.
181 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) {
182 assert(!Inst->getOperand(i)->getType()->isMetadataTy() &&
183 "Cannot value number calls with metadata operands");
184 Res ^= getHash(Inst->getOperand(i)) << i;
187 // Mix in the opcode.
188 return (Res << 1) ^ Inst->getOpcode();
191 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
192 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
193 if (LHS.isSentinel() || RHS.isSentinel())
195 return LHSI->isIdenticalTo(RHSI);
199 //===----------------------------------------------------------------------===//
201 //===----------------------------------------------------------------------===//
205 /// EarlyCSE - This pass does a simple depth-first walk over the dominator
206 /// tree, eliminating trivially redundant instructions and using instsimplify
207 /// to canonicalize things as it goes. It is intended to be fast and catch
208 /// obvious cases so that instcombine and other passes are more effective. It
209 /// is expected that a later pass of GVN will catch the interesting/hard
211 class EarlyCSE : public FunctionPass {
213 const TargetData *TD;
215 typedef RecyclingAllocator<BumpPtrAllocator,
216 ScopedHashTableVal<SimpleValue, Value*> > AllocatorTy;
217 typedef ScopedHashTable<SimpleValue, Value*, DenseMapInfo<SimpleValue>,
218 AllocatorTy> ScopedHTType;
220 /// AvailableValues - This scoped hash table contains the current values of
221 /// all of our simple scalar expressions. As we walk down the domtree, we
222 /// look to see if instructions are in this: if so, we replace them with what
223 /// we find, otherwise we insert them so that dominated values can succeed in
225 ScopedHTType *AvailableValues;
227 /// AvailableLoads - This scoped hash table contains the current values
228 /// of loads. This allows us to get efficient access to dominating loads when
229 /// we have a fully redundant load. In addition to the most recent load, we
230 /// keep track of a generation count of the read, which is compared against
231 /// the current generation count. The current generation count is
232 /// incremented after every possibly writing memory operation, which ensures
233 /// that we only CSE loads with other loads that have no intervening store.
234 typedef RecyclingAllocator<BumpPtrAllocator,
235 ScopedHashTableVal<Value*, std::pair<Value*, unsigned> > > LoadMapAllocator;
236 typedef ScopedHashTable<Value*, std::pair<Value*, unsigned>,
237 DenseMapInfo<Value*>, LoadMapAllocator> LoadHTType;
238 LoadHTType *AvailableLoads;
240 /// AvailableCalls - This scoped hash table contains the current values
241 /// of read-only call values. It uses the same generation count as loads.
242 typedef ScopedHashTable<CallValue, std::pair<Value*, unsigned> > CallHTType;
243 CallHTType *AvailableCalls;
245 /// CurrentGeneration - This is the current generation of the memory value.
246 unsigned CurrentGeneration;
249 explicit EarlyCSE() : FunctionPass(ID) {
250 initializeEarlyCSEPass(*PassRegistry::getPassRegistry());
253 bool runOnFunction(Function &F);
257 bool processNode(DomTreeNode *Node);
259 // This transformation requires dominator postdominator info
260 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
261 AU.addRequired<DominatorTree>();
262 AU.setPreservesCFG();
267 char EarlyCSE::ID = 0;
269 // createEarlyCSEPass - The public interface to this file.
270 FunctionPass *llvm::createEarlyCSEPass() {
271 return new EarlyCSE();
274 INITIALIZE_PASS_BEGIN(EarlyCSE, "early-cse", "Early CSE", false, false)
275 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
276 INITIALIZE_PASS_END(EarlyCSE, "early-cse", "Early CSE", false, false)
278 bool EarlyCSE::processNode(DomTreeNode *Node) {
279 // Define a scope in the scoped hash table. When we are done processing this
280 // domtree node and recurse back up to our parent domtree node, this will pop
281 // off all the values we install.
282 ScopedHTType::ScopeTy Scope(*AvailableValues);
284 // Define a scope for the load values so that anything we add will get
285 // popped when we recurse back up to our parent domtree node.
286 LoadHTType::ScopeTy LoadScope(*AvailableLoads);
288 // Define a scope for the call values so that anything we add will get
289 // popped when we recurse back up to our parent domtree node.
290 CallHTType::ScopeTy CallScope(*AvailableCalls);
292 BasicBlock *BB = Node->getBlock();
294 // If this block has a single predecessor, then the predecessor is the parent
295 // of the domtree node and all of the live out memory values are still current
296 // in this block. If this block has multiple predecessors, then they could
297 // have invalidated the live-out memory values of our parent value. For now,
298 // just be conservative and invalidate memory if this block has multiple
300 if (BB->getSinglePredecessor() == 0)
303 /// LastStore - Keep track of the last non-volatile store that we saw... for
304 /// as long as there in no instruction that reads memory. If we see a store
305 /// to the same location, we delete the dead store. This zaps trivial dead
306 /// stores which can occur in bitfield code among other things.
307 StoreInst *LastStore = 0;
309 bool Changed = false;
311 // See if any instructions in the block can be eliminated. If so, do it. If
312 // not, add them to AvailableValues.
313 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
314 Instruction *Inst = I++;
316 // Dead instructions should just be removed.
317 if (isInstructionTriviallyDead(Inst)) {
318 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
319 Inst->eraseFromParent();
325 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
326 // its simpler value.
327 if (Value *V = SimplifyInstruction(Inst, TD, DT)) {
328 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
329 Inst->replaceAllUsesWith(V);
330 Inst->eraseFromParent();
336 // If this is a simple instruction that we can value number, process it.
337 if (SimpleValue::canHandle(Inst)) {
338 // See if the instruction has an available value. If so, use it.
339 if (Value *V = AvailableValues->lookup(Inst)) {
340 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
341 Inst->replaceAllUsesWith(V);
342 Inst->eraseFromParent();
348 // Otherwise, just remember that this value is available.
349 AvailableValues->insert(Inst, Inst);
353 // If this is a non-volatile load, process it.
354 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
355 // Ignore volatile loads.
356 if (LI->isVolatile()) {
361 // If we have an available version of this load, and if it is the right
362 // generation, replace this instruction.
363 std::pair<Value*, unsigned> InVal =
364 AvailableLoads->lookup(Inst->getOperand(0));
365 if (InVal.first != 0 && InVal.second == CurrentGeneration) {
366 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst << " to: "
367 << *InVal.first << '\n');
368 if (!Inst->use_empty()) Inst->replaceAllUsesWith(InVal.first);
369 Inst->eraseFromParent();
375 // Otherwise, remember that we have this instruction.
376 AvailableLoads->insert(Inst->getOperand(0),
377 std::pair<Value*, unsigned>(Inst, CurrentGeneration));
382 // If this instruction may read from memory, forget LastStore.
383 if (Inst->mayReadFromMemory())
386 // If this is a read-only call, process it.
387 if (CallValue::canHandle(Inst)) {
388 // If we have an available version of this call, and if it is the right
389 // generation, replace this instruction.
390 std::pair<Value*, unsigned> InVal = AvailableCalls->lookup(Inst);
391 if (InVal.first != 0 && InVal.second == CurrentGeneration) {
392 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst << " to: "
393 << *InVal.first << '\n');
394 if (!Inst->use_empty()) Inst->replaceAllUsesWith(InVal.first);
395 Inst->eraseFromParent();
401 // Otherwise, remember that we have this instruction.
402 AvailableCalls->insert(Inst,
403 std::pair<Value*, unsigned>(Inst, CurrentGeneration));
407 // Okay, this isn't something we can CSE at all. Check to see if it is
408 // something that could modify memory. If so, our available memory values
409 // cannot be used so bump the generation count.
410 if (Inst->mayWriteToMemory()) {
413 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
414 // We do a trivial form of DSE if there are two stores to the same
415 // location with no intervening loads. Delete the earlier store.
417 LastStore->getPointerOperand() == SI->getPointerOperand()) {
418 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore << " due to: "
420 LastStore->eraseFromParent();
427 // Okay, we just invalidated anything we knew about loaded values. Try
428 // to salvage *something* by remembering that the stored value is a live
429 // version of the pointer. It is safe to forward from volatile stores
430 // to non-volatile loads, so we don't have to check for volatility of
432 AvailableLoads->insert(SI->getPointerOperand(),
433 std::pair<Value*, unsigned>(SI->getValueOperand(), CurrentGeneration));
435 // Remember that this was the last store we saw for DSE.
436 if (!SI->isVolatile())
442 unsigned LiveOutGeneration = CurrentGeneration;
443 for (DomTreeNode::iterator I = Node->begin(), E = Node->end(); I != E; ++I) {
444 Changed |= processNode(*I);
445 // Pop any generation changes off the stack from the recursive walk.
446 CurrentGeneration = LiveOutGeneration;
452 bool EarlyCSE::runOnFunction(Function &F) {
453 TD = getAnalysisIfAvailable<TargetData>();
454 DT = &getAnalysis<DominatorTree>();
456 // Tables that the pass uses when walking the domtree.
457 ScopedHTType AVTable;
458 AvailableValues = &AVTable;
459 LoadHTType LoadTable;
460 AvailableLoads = &LoadTable;
461 CallHTType CallTable;
462 AvailableCalls = &CallTable;
464 CurrentGeneration = 0;
465 return processNode(DT->getRootNode());