1 //===- GVN.cpp - Eliminate redundant values and loads ------------===//
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 global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "gvn"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/BasicBlock.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/ParameterAttributes.h"
24 #include "llvm/Value.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/DepthFirstIterator.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/SparseBitVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Analysis/AliasAnalysis.h"
33 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
34 #include "llvm/Support/CFG.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Target/TargetData.h"
43 STATISTIC(NumGVNInstr, "Number of instructions deleted");
44 STATISTIC(NumGVNLoad, "Number of loads deleted");
46 //===----------------------------------------------------------------------===//
48 //===----------------------------------------------------------------------===//
50 /// This class holds the mapping between values and value numbers. It is used
51 /// as an efficient mechanism to determine the expression-wise equivalence of
54 struct VISIBILITY_HIDDEN Expression {
55 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
56 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
57 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
58 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
59 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
60 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
61 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
62 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
63 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
64 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, EMPTY,
67 ExpressionOpcode opcode;
72 SmallVector<uint32_t, 4> varargs;
76 Expression(ExpressionOpcode o) : opcode(o) { }
78 bool operator==(const Expression &other) const {
79 if (opcode != other.opcode)
81 else if (opcode == EMPTY || opcode == TOMBSTONE)
83 else if (type != other.type)
85 else if (function != other.function)
87 else if (firstVN != other.firstVN)
89 else if (secondVN != other.secondVN)
91 else if (thirdVN != other.thirdVN)
94 if (varargs.size() != other.varargs.size())
97 for (size_t i = 0; i < varargs.size(); ++i)
98 if (varargs[i] != other.varargs[i])
105 bool operator!=(const Expression &other) const {
106 if (opcode != other.opcode)
108 else if (opcode == EMPTY || opcode == TOMBSTONE)
110 else if (type != other.type)
112 else if (function != other.function)
114 else if (firstVN != other.firstVN)
116 else if (secondVN != other.secondVN)
118 else if (thirdVN != other.thirdVN)
121 if (varargs.size() != other.varargs.size())
124 for (size_t i = 0; i < varargs.size(); ++i)
125 if (varargs[i] != other.varargs[i])
133 class VISIBILITY_HIDDEN ValueTable {
135 DenseMap<Value*, uint32_t> valueNumbering;
136 DenseMap<Expression, uint32_t> expressionNumbering;
139 uint32_t nextValueNumber;
141 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
142 Expression::ExpressionOpcode getOpcode(CmpInst* C);
143 Expression::ExpressionOpcode getOpcode(CastInst* C);
144 Expression create_expression(BinaryOperator* BO);
145 Expression create_expression(CmpInst* C);
146 Expression create_expression(ShuffleVectorInst* V);
147 Expression create_expression(ExtractElementInst* C);
148 Expression create_expression(InsertElementInst* V);
149 Expression create_expression(SelectInst* V);
150 Expression create_expression(CastInst* C);
151 Expression create_expression(GetElementPtrInst* G);
152 Expression create_expression(CallInst* C);
154 ValueTable() : nextValueNumber(1) { }
155 uint32_t lookup_or_add(Value* V);
156 uint32_t lookup(Value* V) const;
157 void add(Value* V, uint32_t num);
159 void erase(Value* v);
161 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
166 template <> struct DenseMapInfo<Expression> {
167 static inline Expression getEmptyKey() {
168 return Expression(Expression::EMPTY);
171 static inline Expression getTombstoneKey() {
172 return Expression(Expression::TOMBSTONE);
175 static unsigned getHashValue(const Expression e) {
176 unsigned hash = e.opcode;
178 hash = e.firstVN + hash * 37;
179 hash = e.secondVN + hash * 37;
180 hash = e.thirdVN + hash * 37;
182 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
183 (unsigned)((uintptr_t)e.type >> 9)) +
186 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
187 E = e.varargs.end(); I != E; ++I)
188 hash = *I + hash * 37;
190 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
191 (unsigned)((uintptr_t)e.function >> 9)) +
196 static bool isEqual(const Expression &LHS, const Expression &RHS) {
199 static bool isPod() { return true; }
203 //===----------------------------------------------------------------------===//
204 // ValueTable Internal Functions
205 //===----------------------------------------------------------------------===//
206 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
207 switch(BO->getOpcode()) {
208 default: // THIS SHOULD NEVER HAPPEN
209 assert(0 && "Binary operator with unknown opcode?");
210 case Instruction::Add: return Expression::ADD;
211 case Instruction::Sub: return Expression::SUB;
212 case Instruction::Mul: return Expression::MUL;
213 case Instruction::UDiv: return Expression::UDIV;
214 case Instruction::SDiv: return Expression::SDIV;
215 case Instruction::FDiv: return Expression::FDIV;
216 case Instruction::URem: return Expression::UREM;
217 case Instruction::SRem: return Expression::SREM;
218 case Instruction::FRem: return Expression::FREM;
219 case Instruction::Shl: return Expression::SHL;
220 case Instruction::LShr: return Expression::LSHR;
221 case Instruction::AShr: return Expression::ASHR;
222 case Instruction::And: return Expression::AND;
223 case Instruction::Or: return Expression::OR;
224 case Instruction::Xor: return Expression::XOR;
228 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
229 if (isa<ICmpInst>(C)) {
230 switch (C->getPredicate()) {
231 default: // THIS SHOULD NEVER HAPPEN
232 assert(0 && "Comparison with unknown predicate?");
233 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
234 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
235 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
236 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
237 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
238 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
239 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
240 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
241 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
242 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
245 assert(isa<FCmpInst>(C) && "Unknown compare");
246 switch (C->getPredicate()) {
247 default: // THIS SHOULD NEVER HAPPEN
248 assert(0 && "Comparison with unknown predicate?");
249 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
250 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
251 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
252 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
253 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
254 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
255 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
256 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
257 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
258 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
259 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
260 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
261 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
262 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
266 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
267 switch(C->getOpcode()) {
268 default: // THIS SHOULD NEVER HAPPEN
269 assert(0 && "Cast operator with unknown opcode?");
270 case Instruction::Trunc: return Expression::TRUNC;
271 case Instruction::ZExt: return Expression::ZEXT;
272 case Instruction::SExt: return Expression::SEXT;
273 case Instruction::FPToUI: return Expression::FPTOUI;
274 case Instruction::FPToSI: return Expression::FPTOSI;
275 case Instruction::UIToFP: return Expression::UITOFP;
276 case Instruction::SIToFP: return Expression::SITOFP;
277 case Instruction::FPTrunc: return Expression::FPTRUNC;
278 case Instruction::FPExt: return Expression::FPEXT;
279 case Instruction::PtrToInt: return Expression::PTRTOINT;
280 case Instruction::IntToPtr: return Expression::INTTOPTR;
281 case Instruction::BitCast: return Expression::BITCAST;
285 Expression ValueTable::create_expression(CallInst* C) {
288 e.type = C->getType();
292 e.function = C->getCalledFunction();
293 e.opcode = Expression::CALL;
295 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
297 e.varargs.push_back(lookup_or_add(*I));
302 Expression ValueTable::create_expression(BinaryOperator* BO) {
305 e.firstVN = lookup_or_add(BO->getOperand(0));
306 e.secondVN = lookup_or_add(BO->getOperand(1));
309 e.type = BO->getType();
310 e.opcode = getOpcode(BO);
315 Expression ValueTable::create_expression(CmpInst* C) {
318 e.firstVN = lookup_or_add(C->getOperand(0));
319 e.secondVN = lookup_or_add(C->getOperand(1));
322 e.type = C->getType();
323 e.opcode = getOpcode(C);
328 Expression ValueTable::create_expression(CastInst* C) {
331 e.firstVN = lookup_or_add(C->getOperand(0));
335 e.type = C->getType();
336 e.opcode = getOpcode(C);
341 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
344 e.firstVN = lookup_or_add(S->getOperand(0));
345 e.secondVN = lookup_or_add(S->getOperand(1));
346 e.thirdVN = lookup_or_add(S->getOperand(2));
348 e.type = S->getType();
349 e.opcode = Expression::SHUFFLE;
354 Expression ValueTable::create_expression(ExtractElementInst* E) {
357 e.firstVN = lookup_or_add(E->getOperand(0));
358 e.secondVN = lookup_or_add(E->getOperand(1));
361 e.type = E->getType();
362 e.opcode = Expression::EXTRACT;
367 Expression ValueTable::create_expression(InsertElementInst* I) {
370 e.firstVN = lookup_or_add(I->getOperand(0));
371 e.secondVN = lookup_or_add(I->getOperand(1));
372 e.thirdVN = lookup_or_add(I->getOperand(2));
374 e.type = I->getType();
375 e.opcode = Expression::INSERT;
380 Expression ValueTable::create_expression(SelectInst* I) {
383 e.firstVN = lookup_or_add(I->getCondition());
384 e.secondVN = lookup_or_add(I->getTrueValue());
385 e.thirdVN = lookup_or_add(I->getFalseValue());
387 e.type = I->getType();
388 e.opcode = Expression::SELECT;
393 Expression ValueTable::create_expression(GetElementPtrInst* G) {
396 e.firstVN = lookup_or_add(G->getPointerOperand());
400 e.type = G->getType();
401 e.opcode = Expression::GEP;
403 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
405 e.varargs.push_back(lookup_or_add(*I));
410 //===----------------------------------------------------------------------===//
411 // ValueTable External Functions
412 //===----------------------------------------------------------------------===//
414 /// lookup_or_add - Returns the value number for the specified value, assigning
415 /// it a new number if it did not have one before.
416 uint32_t ValueTable::lookup_or_add(Value* V) {
417 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
418 if (VI != valueNumbering.end())
421 if (CallInst* C = dyn_cast<CallInst>(V)) {
422 if (AA->doesNotAccessMemory(C)) {
423 Expression e = create_expression(C);
425 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
426 if (EI != expressionNumbering.end()) {
427 valueNumbering.insert(std::make_pair(V, EI->second));
430 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
431 valueNumbering.insert(std::make_pair(V, nextValueNumber));
433 return nextValueNumber++;
436 valueNumbering.insert(std::make_pair(V, nextValueNumber));
437 return nextValueNumber++;
439 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
440 Expression e = create_expression(BO);
442 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
443 if (EI != expressionNumbering.end()) {
444 valueNumbering.insert(std::make_pair(V, EI->second));
447 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
448 valueNumbering.insert(std::make_pair(V, nextValueNumber));
450 return nextValueNumber++;
452 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
453 Expression e = create_expression(C);
455 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
456 if (EI != expressionNumbering.end()) {
457 valueNumbering.insert(std::make_pair(V, EI->second));
460 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
461 valueNumbering.insert(std::make_pair(V, nextValueNumber));
463 return nextValueNumber++;
465 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
466 Expression e = create_expression(U);
468 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
469 if (EI != expressionNumbering.end()) {
470 valueNumbering.insert(std::make_pair(V, EI->second));
473 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
474 valueNumbering.insert(std::make_pair(V, nextValueNumber));
476 return nextValueNumber++;
478 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
479 Expression e = create_expression(U);
481 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
482 if (EI != expressionNumbering.end()) {
483 valueNumbering.insert(std::make_pair(V, EI->second));
486 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
487 valueNumbering.insert(std::make_pair(V, nextValueNumber));
489 return nextValueNumber++;
491 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
492 Expression e = create_expression(U);
494 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
495 if (EI != expressionNumbering.end()) {
496 valueNumbering.insert(std::make_pair(V, EI->second));
499 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
500 valueNumbering.insert(std::make_pair(V, nextValueNumber));
502 return nextValueNumber++;
504 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
505 Expression e = create_expression(U);
507 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
508 if (EI != expressionNumbering.end()) {
509 valueNumbering.insert(std::make_pair(V, EI->second));
512 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
513 valueNumbering.insert(std::make_pair(V, nextValueNumber));
515 return nextValueNumber++;
517 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
518 Expression e = create_expression(U);
520 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
521 if (EI != expressionNumbering.end()) {
522 valueNumbering.insert(std::make_pair(V, EI->second));
525 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
526 valueNumbering.insert(std::make_pair(V, nextValueNumber));
528 return nextValueNumber++;
530 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
531 Expression e = create_expression(U);
533 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
534 if (EI != expressionNumbering.end()) {
535 valueNumbering.insert(std::make_pair(V, EI->second));
538 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
539 valueNumbering.insert(std::make_pair(V, nextValueNumber));
541 return nextValueNumber++;
544 valueNumbering.insert(std::make_pair(V, nextValueNumber));
545 return nextValueNumber++;
549 /// lookup - Returns the value number of the specified value. Fails if
550 /// the value has not yet been numbered.
551 uint32_t ValueTable::lookup(Value* V) const {
552 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
553 assert(VI != valueNumbering.end() && "Value not numbered?");
557 /// clear - Remove all entries from the ValueTable
558 void ValueTable::clear() {
559 valueNumbering.clear();
560 expressionNumbering.clear();
564 /// erase - Remove a value from the value numbering
565 void ValueTable::erase(Value* V) {
566 valueNumbering.erase(V);
569 //===----------------------------------------------------------------------===//
570 // ValueNumberedSet Class
571 //===----------------------------------------------------------------------===//
573 class VISIBILITY_HIDDEN ValueNumberedSet {
575 SmallPtrSet<Value*, 8> contents;
576 SparseBitVector<64> numbers;
578 ValueNumberedSet() { }
579 ValueNumberedSet(const ValueNumberedSet& other) {
580 numbers = other.numbers;
581 contents = other.contents;
584 typedef SmallPtrSet<Value*, 8>::iterator iterator;
586 iterator begin() { return contents.begin(); }
587 iterator end() { return contents.end(); }
589 bool insert(Value* v) { return contents.insert(v); }
590 void insert(iterator I, iterator E) { contents.insert(I, E); }
591 void erase(Value* v) { contents.erase(v); }
592 unsigned count(Value* v) { return contents.count(v); }
593 size_t size() { return contents.size(); }
595 void set(unsigned i) {
599 void operator=(const ValueNumberedSet& other) {
600 contents = other.contents;
601 numbers = other.numbers;
604 void reset(unsigned i) {
608 bool test(unsigned i) {
609 return numbers.test(i);
614 //===----------------------------------------------------------------------===//
616 //===----------------------------------------------------------------------===//
620 class VISIBILITY_HIDDEN GVN : public FunctionPass {
621 bool runOnFunction(Function &F);
623 static char ID; // Pass identification, replacement for typeid
624 GVN() : FunctionPass((intptr_t)&ID) { }
629 DenseMap<BasicBlock*, ValueNumberedSet> availableOut;
631 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
635 // This transformation requires dominator postdominator info
636 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
637 AU.setPreservesCFG();
638 AU.addRequired<DominatorTree>();
639 AU.addRequired<MemoryDependenceAnalysis>();
640 AU.addRequired<AliasAnalysis>();
641 AU.addRequired<TargetData>();
642 AU.addPreserved<AliasAnalysis>();
643 AU.addPreserved<MemoryDependenceAnalysis>();
644 AU.addPreserved<TargetData>();
648 // FIXME: eliminate or document these better
649 Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
650 void val_insert(ValueNumberedSet& s, Value* v);
651 bool processLoad(LoadInst* L,
652 DenseMap<Value*, LoadInst*> &lastLoad,
653 SmallVectorImpl<Instruction*> &toErase);
654 bool processInstruction(Instruction* I,
655 ValueNumberedSet& currAvail,
656 DenseMap<Value*, LoadInst*>& lastSeenLoad,
657 SmallVectorImpl<Instruction*> &toErase);
658 bool processNonLocalLoad(LoadInst* L,
659 SmallVectorImpl<Instruction*> &toErase);
660 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
661 DenseMap<BasicBlock*, Value*> &Phis,
662 bool top_level = false);
663 void dump(DenseMap<BasicBlock*, Value*>& d);
664 bool iterateOnFunction(Function &F);
665 Value* CollapsePhi(PHINode* p);
666 bool isSafeReplacement(PHINode* p, Instruction* inst);
672 // createGVNPass - The public interface to this file...
673 FunctionPass *llvm::createGVNPass() { return new GVN(); }
675 static RegisterPass<GVN> X("gvn",
676 "Global Value Numbering");
678 /// find_leader - Given a set and a value number, return the first
679 /// element of the set with that value number, or 0 if no such element
681 Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) {
685 for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end();
687 if (v == VN.lookup(*I))
690 assert(0 && "No leader found, but present bit is set?");
694 /// val_insert - Insert a value into a set only if there is not a value
695 /// with the same value number already in the set
696 void GVN::val_insert(ValueNumberedSet& s, Value* v) {
697 uint32_t num = VN.lookup(v);
702 void GVN::dump(DenseMap<BasicBlock*, Value*>& d) {
704 for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(),
705 E = d.end(); I != E; ++I) {
706 if (I->second == MemoryDependenceAnalysis::None)
714 Value* GVN::CollapsePhi(PHINode* p) {
715 DominatorTree &DT = getAnalysis<DominatorTree>();
716 Value* constVal = p->hasConstantValue();
718 if (!constVal) return 0;
720 Instruction* inst = dyn_cast<Instruction>(constVal);
724 if (DT.dominates(inst, p))
725 if (isSafeReplacement(p, inst))
730 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
731 if (!isa<PHINode>(inst))
734 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
736 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
737 if (use_phi->getParent() == inst->getParent())
743 /// GetValueForBlock - Get the value to use within the specified basic block.
744 /// available values are in Phis.
745 Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
746 DenseMap<BasicBlock*, Value*> &Phis,
749 // If we have already computed this value, return the previously computed val.
750 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
751 if (V != Phis.end() && !top_level) return V->second;
753 BasicBlock* singlePred = BB->getSinglePredecessor();
755 Value *ret = GetValueForBlock(singlePred, orig, Phis);
760 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
761 // now, then get values to fill in the incoming values for the PHI.
762 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
764 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
766 if (Phis.count(BB) == 0)
767 Phis.insert(std::make_pair(BB, PN));
769 // Fill in the incoming values for the block.
770 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
771 Value* val = GetValueForBlock(*PI, orig, Phis);
772 PN->addIncoming(val, *PI);
775 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
776 AA.copyValue(orig, PN);
778 // Attempt to collapse PHI nodes that are trivially redundant
779 Value* v = CollapsePhi(PN);
781 // Cache our phi construction results
782 phiMap[orig->getPointerOperand()].insert(PN);
786 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
788 MD.removeInstruction(PN);
789 PN->replaceAllUsesWith(v);
791 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
792 E = Phis.end(); I != E; ++I)
796 PN->eraseFromParent();
802 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
803 /// non-local by performing PHI construction.
804 bool GVN::processNonLocalLoad(LoadInst* L,
805 SmallVectorImpl<Instruction*> &toErase) {
806 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
808 // Find the non-local dependencies of the load
809 DenseMap<BasicBlock*, Value*> deps;
810 MD.getNonLocalDependency(L, deps);
812 DenseMap<BasicBlock*, Value*> repl;
814 // Filter out useless results (non-locals, etc)
815 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
817 if (I->second == MemoryDependenceAnalysis::None)
820 if (I->second == MemoryDependenceAnalysis::NonLocal)
823 if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
824 if (S->getPointerOperand() != L->getPointerOperand())
826 repl[I->first] = S->getOperand(0);
827 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
828 if (LD->getPointerOperand() != L->getPointerOperand())
836 // Use cached PHI construction information from previous runs
837 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
838 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
840 if ((*I)->getParent() == L->getParent()) {
841 MD.removeInstruction(L);
842 L->replaceAllUsesWith(*I);
843 toErase.push_back(L);
848 repl.insert(std::make_pair((*I)->getParent(), *I));
851 // Perform PHI construction
852 SmallPtrSet<BasicBlock*, 4> visited;
853 Value* v = GetValueForBlock(L->getParent(), L, repl, true);
855 MD.removeInstruction(L);
856 L->replaceAllUsesWith(v);
857 toErase.push_back(L);
863 /// processLoad - Attempt to eliminate a load, first by eliminating it
864 /// locally, and then attempting non-local elimination if that fails.
865 bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
866 SmallVectorImpl<Instruction*> &toErase) {
867 if (L->isVolatile()) {
868 lastLoad[L->getPointerOperand()] = L;
872 Value* pointer = L->getPointerOperand();
873 LoadInst*& last = lastLoad[pointer];
875 // ... to a pointer that has been loaded from before...
876 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
877 bool removedNonLocal = false;
878 Instruction* dep = MD.getDependency(L);
879 if (dep == MemoryDependenceAnalysis::NonLocal &&
880 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
881 removedNonLocal = processNonLocalLoad(L, toErase);
883 if (!removedNonLocal)
886 return removedNonLocal;
890 bool deletedLoad = false;
892 // Walk up the dependency chain until we either find
893 // a dependency we can use, or we can't walk any further
894 while (dep != MemoryDependenceAnalysis::None &&
895 dep != MemoryDependenceAnalysis::NonLocal &&
896 (isa<LoadInst>(dep) || isa<StoreInst>(dep))) {
897 // ... that depends on a store ...
898 if (StoreInst* S = dyn_cast<StoreInst>(dep)) {
899 if (S->getPointerOperand() == pointer) {
901 MD.removeInstruction(L);
903 L->replaceAllUsesWith(S->getOperand(0));
904 toErase.push_back(L);
909 // Whether we removed it or not, we can't
913 // If we don't depend on a store, and we haven't
914 // been loaded before, bail.
916 } else if (dep == last) {
918 MD.removeInstruction(L);
920 L->replaceAllUsesWith(last);
921 toErase.push_back(L);
927 dep = MD.getDependency(L, dep);
931 if (dep != MemoryDependenceAnalysis::None &&
932 dep != MemoryDependenceAnalysis::NonLocal &&
933 isa<AllocationInst>(dep)) {
934 // Check that this load is actually from the
935 // allocation we found
936 Value* v = L->getOperand(0);
938 if (BitCastInst *BC = dyn_cast<BitCastInst>(v))
939 v = BC->getOperand(0);
940 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v))
941 v = GEP->getOperand(0);
946 // If this load depends directly on an allocation, there isn't
947 // anything stored there; therefore, we can optimize this load
949 MD.removeInstruction(L);
951 L->replaceAllUsesWith(UndefValue::get(L->getType()));
952 toErase.push_back(L);
964 /// processInstruction - When calculating availability, handle an instruction
965 /// by inserting it into the appropriate sets
966 bool GVN::processInstruction(Instruction *I, ValueNumberedSet &currAvail,
967 DenseMap<Value*, LoadInst*> &lastSeenLoad,
968 SmallVectorImpl<Instruction*> &toErase) {
969 if (LoadInst* L = dyn_cast<LoadInst>(I))
970 return processLoad(L, lastSeenLoad, toErase);
972 // Allocations are always uniquely numbered, so we can save time and memory
973 // by fast failing them.
974 if (isa<AllocationInst>(I))
977 unsigned num = VN.lookup_or_add(I);
979 // Collapse PHI nodes
980 if (PHINode* p = dyn_cast<PHINode>(I)) {
981 Value* constVal = CollapsePhi(p);
984 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
986 if (PI->second.count(p))
989 p->replaceAllUsesWith(constVal);
990 toErase.push_back(p);
992 // Perform value-number based elimination
993 } else if (currAvail.test(num)) {
994 Value* repl = find_leader(currAvail, num);
996 if (CallInst* CI = dyn_cast<CallInst>(I)) {
997 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
998 if (!AA.doesNotAccessMemory(CI)) {
999 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1000 if (cast<Instruction>(repl)->getParent() != CI->getParent() ||
1001 MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) {
1002 // There must be an intervening may-alias store, so nothing from
1003 // this point on will be able to be replaced with the preceding call
1004 currAvail.erase(repl);
1005 currAvail.insert(I);
1013 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1014 MD.removeInstruction(I);
1017 I->replaceAllUsesWith(repl);
1018 toErase.push_back(I);
1020 } else if (!I->isTerminator()) {
1022 currAvail.insert(I);
1028 // GVN::runOnFunction - This is the main transformation entry point for a
1031 bool GVN::runOnFunction(Function& F) {
1032 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1034 bool changed = false;
1035 bool shouldContinue = true;
1037 while (shouldContinue) {
1038 shouldContinue = iterateOnFunction(F);
1039 changed |= shouldContinue;
1046 // GVN::iterateOnFunction - Executes one iteration of GVN
1047 bool GVN::iterateOnFunction(Function &F) {
1048 // Clean out global sets from any previous functions
1050 availableOut.clear();
1053 bool changed_function = false;
1055 DominatorTree &DT = getAnalysis<DominatorTree>();
1057 SmallVector<Instruction*, 8> toErase;
1058 DenseMap<Value*, LoadInst*> lastSeenLoad;
1059 DenseMap<DomTreeNode*, size_t> numChildrenVisited;
1061 // Top-down walk of the dominator tree
1062 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1063 E = df_end(DT.getRootNode()); DI != E; ++DI) {
1065 // Get the set to update for this block
1066 ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
1067 lastSeenLoad.clear();
1069 BasicBlock* BB = DI->getBlock();
1071 // A block inherits AVAIL_OUT from its dominator
1072 if (DI->getIDom() != 0) {
1073 currAvail = availableOut[DI->getIDom()->getBlock()];
1075 numChildrenVisited[DI->getIDom()]++;
1077 if (numChildrenVisited[DI->getIDom()] == DI->getIDom()->getNumChildren()) {
1078 availableOut.erase(DI->getIDom()->getBlock());
1079 numChildrenVisited.erase(DI->getIDom());
1083 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1085 changed_function |= processInstruction(BI, currAvail,
1086 lastSeenLoad, toErase);
1087 if (toErase.empty()) {
1092 // If we need some instructions deleted, do it now.
1093 NumGVNInstr += toErase.size();
1095 // Avoid iterator invalidation.
1096 bool AtStart = BI == BB->begin();
1100 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1101 E = toErase.end(); I != E; ++I)
1102 (*I)->eraseFromParent();
1113 return changed_function;