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 if (CallInst *CI = dyn_cast<CallInst>(Inst))
149 return CI->onlyReadsMemory();
157 template<> struct isPodLike<CallValue> {
158 static const bool value = true;
161 template<> struct DenseMapInfo<CallValue> {
162 static inline CallValue getEmptyKey() {
163 return DenseMapInfo<Instruction*>::getEmptyKey();
165 static inline CallValue getTombstoneKey() {
166 return DenseMapInfo<Instruction*>::getTombstoneKey();
168 static unsigned getHashValue(CallValue Val);
169 static bool isEqual(CallValue LHS, CallValue RHS);
172 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) {
173 Instruction *Inst = Val.Inst;
174 // Hash in all of the operands as pointers.
176 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
177 Res ^= getHash(Inst->getOperand(i)) << i;
178 // Mix in the opcode.
179 return (Res << 1) ^ Inst->getOpcode();
182 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) {
183 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst;
184 if (LHS.isSentinel() || RHS.isSentinel())
186 return LHSI->isIdenticalTo(RHSI);
190 //===----------------------------------------------------------------------===//
192 //===----------------------------------------------------------------------===//
196 /// EarlyCSE - This pass does a simple depth-first walk over the dominator
197 /// tree, eliminating trivially redundant instructions and using instsimplify
198 /// to canonicalize things as it goes. It is intended to be fast and catch
199 /// obvious cases so that instcombine and other passes are more effective. It
200 /// is expected that a later pass of GVN will catch the interesting/hard
202 class EarlyCSE : public FunctionPass {
204 const TargetData *TD;
206 typedef RecyclingAllocator<BumpPtrAllocator,
207 ScopedHashTableVal<SimpleValue, Value*> > AllocatorTy;
208 typedef ScopedHashTable<SimpleValue, Value*, DenseMapInfo<SimpleValue>,
209 AllocatorTy> ScopedHTType;
211 /// AvailableValues - This scoped hash table contains the current values of
212 /// all of our simple scalar expressions. As we walk down the domtree, we
213 /// look to see if instructions are in this: if so, we replace them with what
214 /// we find, otherwise we insert them so that dominated values can succeed in
216 ScopedHTType *AvailableValues;
218 /// AvailableLoads - This scoped hash table contains the current values
219 /// of loads. This allows us to get efficient access to dominating loads when
220 /// we have a fully redundant load. In addition to the most recent load, we
221 /// keep track of a generation count of the read, which is compared against
222 /// the current generation count. The current generation count is
223 /// incremented after every possibly writing memory operation, which ensures
224 /// that we only CSE loads with other loads that have no intervening store.
225 typedef RecyclingAllocator<BumpPtrAllocator,
226 ScopedHashTableVal<Value*, std::pair<Value*, unsigned> > > LoadMapAllocator;
227 typedef ScopedHashTable<Value*, std::pair<Value*, unsigned>,
228 DenseMapInfo<Value*>, LoadMapAllocator> LoadHTType;
229 LoadHTType *AvailableLoads;
231 /// AvailableCalls - This scoped hash table contains the current values
232 /// of read-only call values. It uses the same generation count as loads.
233 typedef ScopedHashTable<CallValue, std::pair<Value*, unsigned> > CallHTType;
234 CallHTType *AvailableCalls;
236 /// CurrentGeneration - This is the current generation of the memory value.
237 unsigned CurrentGeneration;
240 explicit EarlyCSE() : FunctionPass(ID) {
241 initializeEarlyCSEPass(*PassRegistry::getPassRegistry());
244 bool runOnFunction(Function &F);
248 bool processNode(DomTreeNode *Node);
250 // This transformation requires dominator postdominator info
251 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
252 AU.addRequired<DominatorTree>();
253 AU.setPreservesCFG();
258 char EarlyCSE::ID = 0;
260 // createEarlyCSEPass - The public interface to this file.
261 FunctionPass *llvm::createEarlyCSEPass() {
262 return new EarlyCSE();
265 INITIALIZE_PASS_BEGIN(EarlyCSE, "early-cse", "Early CSE", false, false)
266 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
267 INITIALIZE_PASS_END(EarlyCSE, "early-cse", "Early CSE", false, false)
269 bool EarlyCSE::processNode(DomTreeNode *Node) {
270 // Define a scope in the scoped hash table. When we are done processing this
271 // domtree node and recurse back up to our parent domtree node, this will pop
272 // off all the values we install.
273 ScopedHTType::ScopeTy Scope(*AvailableValues);
275 // Define a scope for the load values so that anything we add will get
276 // popped when we recurse back up to our parent domtree node.
277 LoadHTType::ScopeTy LoadScope(*AvailableLoads);
279 // Define a scope for the call values so that anything we add will get
280 // popped when we recurse back up to our parent domtree node.
281 CallHTType::ScopeTy CallScope(*AvailableCalls);
283 BasicBlock *BB = Node->getBlock();
285 // If this block has a single predecessor, then the predecessor is the parent
286 // of the domtree node and all of the live out memory values are still current
287 // in this block. If this block has multiple predecessors, then they could
288 // have invalidated the live-out memory values of our parent value. For now,
289 // just be conservative and invalidate memory if this block has multiple
291 if (BB->getSinglePredecessor() == 0)
294 /// LastStore - Keep track of the last non-volatile store that we saw... for
295 /// as long as there in no instruction that reads memory. If we see a store
296 /// to the same location, we delete the dead store. This zaps trivial dead
297 /// stores which can occur in bitfield code among other things.
298 StoreInst *LastStore = 0;
300 bool Changed = false;
302 // See if any instructions in the block can be eliminated. If so, do it. If
303 // not, add them to AvailableValues.
304 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
305 Instruction *Inst = I++;
307 // Dead instructions should just be removed.
308 if (isInstructionTriviallyDead(Inst)) {
309 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n');
310 Inst->eraseFromParent();
316 // If the instruction can be simplified (e.g. X+0 = X) then replace it with
317 // its simpler value.
318 if (Value *V = SimplifyInstruction(Inst, TD, DT)) {
319 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n');
320 Inst->replaceAllUsesWith(V);
321 Inst->eraseFromParent();
327 // If this is a simple instruction that we can value number, process it.
328 if (SimpleValue::canHandle(Inst)) {
329 // See if the instruction has an available value. If so, use it.
330 if (Value *V = AvailableValues->lookup(Inst)) {
331 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n');
332 Inst->replaceAllUsesWith(V);
333 Inst->eraseFromParent();
339 // Otherwise, just remember that this value is available.
340 AvailableValues->insert(Inst, Inst);
344 // If this is a non-volatile load, process it.
345 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
346 // Ignore volatile loads.
347 if (LI->isVolatile()) {
352 // If we have an available version of this load, and if it is the right
353 // generation, replace this instruction.
354 std::pair<Value*, unsigned> InVal =
355 AvailableLoads->lookup(Inst->getOperand(0));
356 if (InVal.first != 0 && InVal.second == CurrentGeneration) {
357 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst << " to: "
358 << *InVal.first << '\n');
359 if (!Inst->use_empty()) Inst->replaceAllUsesWith(InVal.first);
360 Inst->eraseFromParent();
366 // Otherwise, remember that we have this instruction.
367 AvailableLoads->insert(Inst->getOperand(0),
368 std::pair<Value*, unsigned>(Inst, CurrentGeneration));
373 // If this instruction may read from memory, forget LastStore.
374 if (Inst->mayReadFromMemory())
377 // If this is a read-only call, process it.
378 if (CallValue::canHandle(Inst)) {
379 // If we have an available version of this call, and if it is the right
380 // generation, replace this instruction.
381 std::pair<Value*, unsigned> InVal = AvailableCalls->lookup(Inst);
382 if (InVal.first != 0 && InVal.second == CurrentGeneration) {
383 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst << " to: "
384 << *InVal.first << '\n');
385 if (!Inst->use_empty()) Inst->replaceAllUsesWith(InVal.first);
386 Inst->eraseFromParent();
392 // Otherwise, remember that we have this instruction.
393 AvailableCalls->insert(Inst,
394 std::pair<Value*, unsigned>(Inst, CurrentGeneration));
398 // Okay, this isn't something we can CSE at all. Check to see if it is
399 // something that could modify memory. If so, our available memory values
400 // cannot be used so bump the generation count.
401 if (Inst->mayWriteToMemory()) {
404 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
405 // We do a trivial form of DSE if there are two stores to the same
406 // location with no intervening loads. Delete the earlier store.
408 LastStore->getPointerOperand() == SI->getPointerOperand()) {
409 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore << " due to: "
411 LastStore->eraseFromParent();
418 // Okay, we just invalidated anything we knew about loaded values. Try
419 // to salvage *something* by remembering that the stored value is a live
420 // version of the pointer. It is safe to forward from volatile stores
421 // to non-volatile loads, so we don't have to check for volatility of
423 AvailableLoads->insert(SI->getPointerOperand(),
424 std::pair<Value*, unsigned>(SI->getValueOperand(), CurrentGeneration));
426 // Remember that this was the last store we saw for DSE.
427 if (!SI->isVolatile())
433 unsigned LiveOutGeneration = CurrentGeneration;
434 for (DomTreeNode::iterator I = Node->begin(), E = Node->end(); I != E; ++I) {
435 Changed |= processNode(*I);
436 // Pop any generation changes off the stack from the recursive walk.
437 CurrentGeneration = LiveOutGeneration;
443 bool EarlyCSE::runOnFunction(Function &F) {
444 TD = getAnalysisIfAvailable<TargetData>();
445 DT = &getAnalysis<DominatorTree>();
447 // Tables that the pass uses when walking the domtree.
448 ScopedHTType AVTable;
449 AvailableValues = &AVTable;
450 LoadHTType LoadTable;
451 AvailableLoads = &LoadTable;
452 CallHTType CallTable;
453 AvailableCalls = &CallTable;
455 CurrentGeneration = 0;
456 return processNode(DT->getRootNode());