1 //===- GVN.cpp - Eliminate redundant values and loads ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the Owen Anderson and is distributed under
6 // the University of Illinois Open Source 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"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/BasicBlock.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/Value.h"
24 #include "llvm/ADT/BitVector.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/Statistic.h"
30 #include "llvm/Analysis/Dominators.h"
31 #include "llvm/Analysis/AliasAnalysis.h"
32 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Support/Compiler.h"
37 //===----------------------------------------------------------------------===//
39 //===----------------------------------------------------------------------===//
41 /// This class holds the mapping between values and value numbers. It is used
42 /// as an efficient mechanism to determine the expression-wise equivalence of
45 struct VISIBILITY_HIDDEN Expression {
46 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
47 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
48 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
49 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
50 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
51 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
52 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
53 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
54 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
55 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, EMPTY,
58 ExpressionOpcode opcode;
63 SmallVector<uint32_t, 4> varargs;
67 Expression(ExpressionOpcode o) : opcode(o) { }
69 bool operator==(const Expression &other) const {
70 if (opcode != other.opcode)
72 else if (opcode == EMPTY || opcode == TOMBSTONE)
74 else if (type != other.type)
76 else if (function != other.function)
78 else if (firstVN != other.firstVN)
80 else if (secondVN != other.secondVN)
82 else if (thirdVN != other.thirdVN)
85 if (varargs.size() != other.varargs.size())
88 for (size_t i = 0; i < varargs.size(); ++i)
89 if (varargs[i] != other.varargs[i])
96 bool operator!=(const Expression &other) const {
97 if (opcode != other.opcode)
99 else if (opcode == EMPTY || opcode == TOMBSTONE)
101 else if (type != other.type)
103 else if (function != other.function)
105 else if (firstVN != other.firstVN)
107 else if (secondVN != other.secondVN)
109 else if (thirdVN != other.thirdVN)
112 if (varargs.size() != other.varargs.size())
115 for (size_t i = 0; i < varargs.size(); ++i)
116 if (varargs[i] != other.varargs[i])
124 class VISIBILITY_HIDDEN ValueTable {
126 DenseMap<Value*, uint32_t> valueNumbering;
127 DenseMap<Expression, uint32_t> expressionNumbering;
130 uint32_t nextValueNumber;
132 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
133 Expression::ExpressionOpcode getOpcode(CmpInst* C);
134 Expression::ExpressionOpcode getOpcode(CastInst* C);
135 Expression create_expression(BinaryOperator* BO);
136 Expression create_expression(CmpInst* C);
137 Expression create_expression(ShuffleVectorInst* V);
138 Expression create_expression(ExtractElementInst* C);
139 Expression create_expression(InsertElementInst* V);
140 Expression create_expression(SelectInst* V);
141 Expression create_expression(CastInst* C);
142 Expression create_expression(GetElementPtrInst* G);
143 Expression create_expression(CallInst* C);
145 ValueTable() : nextValueNumber(1) { }
146 uint32_t lookup_or_add(Value* V);
147 uint32_t lookup(Value* V) const;
148 void add(Value* V, uint32_t num);
150 void erase(Value* v);
152 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
153 uint32_t hash_operand(Value* v);
158 template <> struct DenseMapInfo<Expression> {
159 static inline Expression getEmptyKey() {
160 return Expression(Expression::EMPTY);
163 static inline Expression getTombstoneKey() {
164 return Expression(Expression::TOMBSTONE);
167 static unsigned getHashValue(const Expression e) {
168 unsigned hash = e.opcode;
170 hash = e.firstVN + hash * 37;
171 hash = e.secondVN + hash * 37;
172 hash = e.thirdVN + hash * 37;
174 hash = (unsigned)((uintptr_t)e.type >> 4) ^
175 (unsigned)((uintptr_t)e.type >> 9) +
178 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
179 E = e.varargs.end(); I != E; ++I)
180 hash = *I + hash * 37;
182 hash = (unsigned)((uintptr_t)e.function >> 4) ^
183 (unsigned)((uintptr_t)e.function >> 9) +
188 static bool isEqual(const Expression &LHS, const Expression &RHS) {
191 static bool isPod() { return true; }
195 //===----------------------------------------------------------------------===//
196 // ValueTable Internal Functions
197 //===----------------------------------------------------------------------===//
198 Expression::ExpressionOpcode
199 ValueTable::getOpcode(BinaryOperator* BO) {
200 switch(BO->getOpcode()) {
201 case Instruction::Add:
202 return Expression::ADD;
203 case Instruction::Sub:
204 return Expression::SUB;
205 case Instruction::Mul:
206 return Expression::MUL;
207 case Instruction::UDiv:
208 return Expression::UDIV;
209 case Instruction::SDiv:
210 return Expression::SDIV;
211 case Instruction::FDiv:
212 return Expression::FDIV;
213 case Instruction::URem:
214 return Expression::UREM;
215 case Instruction::SRem:
216 return Expression::SREM;
217 case Instruction::FRem:
218 return Expression::FREM;
219 case Instruction::Shl:
220 return Expression::SHL;
221 case Instruction::LShr:
222 return Expression::LSHR;
223 case Instruction::AShr:
224 return Expression::ASHR;
225 case Instruction::And:
226 return Expression::AND;
227 case Instruction::Or:
228 return Expression::OR;
229 case Instruction::Xor:
230 return Expression::XOR;
232 // THIS SHOULD NEVER HAPPEN
234 assert(0 && "Binary operator with unknown opcode?");
235 return Expression::ADD;
239 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
240 if (C->getOpcode() == Instruction::ICmp) {
241 switch (C->getPredicate()) {
242 case ICmpInst::ICMP_EQ:
243 return Expression::ICMPEQ;
244 case ICmpInst::ICMP_NE:
245 return Expression::ICMPNE;
246 case ICmpInst::ICMP_UGT:
247 return Expression::ICMPUGT;
248 case ICmpInst::ICMP_UGE:
249 return Expression::ICMPUGE;
250 case ICmpInst::ICMP_ULT:
251 return Expression::ICMPULT;
252 case ICmpInst::ICMP_ULE:
253 return Expression::ICMPULE;
254 case ICmpInst::ICMP_SGT:
255 return Expression::ICMPSGT;
256 case ICmpInst::ICMP_SGE:
257 return Expression::ICMPSGE;
258 case ICmpInst::ICMP_SLT:
259 return Expression::ICMPSLT;
260 case ICmpInst::ICMP_SLE:
261 return Expression::ICMPSLE;
263 // THIS SHOULD NEVER HAPPEN
265 assert(0 && "Comparison with unknown predicate?");
266 return Expression::ICMPEQ;
269 switch (C->getPredicate()) {
270 case FCmpInst::FCMP_OEQ:
271 return Expression::FCMPOEQ;
272 case FCmpInst::FCMP_OGT:
273 return Expression::FCMPOGT;
274 case FCmpInst::FCMP_OGE:
275 return Expression::FCMPOGE;
276 case FCmpInst::FCMP_OLT:
277 return Expression::FCMPOLT;
278 case FCmpInst::FCMP_OLE:
279 return Expression::FCMPOLE;
280 case FCmpInst::FCMP_ONE:
281 return Expression::FCMPONE;
282 case FCmpInst::FCMP_ORD:
283 return Expression::FCMPORD;
284 case FCmpInst::FCMP_UNO:
285 return Expression::FCMPUNO;
286 case FCmpInst::FCMP_UEQ:
287 return Expression::FCMPUEQ;
288 case FCmpInst::FCMP_UGT:
289 return Expression::FCMPUGT;
290 case FCmpInst::FCMP_UGE:
291 return Expression::FCMPUGE;
292 case FCmpInst::FCMP_ULT:
293 return Expression::FCMPULT;
294 case FCmpInst::FCMP_ULE:
295 return Expression::FCMPULE;
296 case FCmpInst::FCMP_UNE:
297 return Expression::FCMPUNE;
299 // THIS SHOULD NEVER HAPPEN
301 assert(0 && "Comparison with unknown predicate?");
302 return Expression::FCMPOEQ;
307 Expression::ExpressionOpcode
308 ValueTable::getOpcode(CastInst* C) {
309 switch(C->getOpcode()) {
310 case Instruction::Trunc:
311 return Expression::TRUNC;
312 case Instruction::ZExt:
313 return Expression::ZEXT;
314 case Instruction::SExt:
315 return Expression::SEXT;
316 case Instruction::FPToUI:
317 return Expression::FPTOUI;
318 case Instruction::FPToSI:
319 return Expression::FPTOSI;
320 case Instruction::UIToFP:
321 return Expression::UITOFP;
322 case Instruction::SIToFP:
323 return Expression::SITOFP;
324 case Instruction::FPTrunc:
325 return Expression::FPTRUNC;
326 case Instruction::FPExt:
327 return Expression::FPEXT;
328 case Instruction::PtrToInt:
329 return Expression::PTRTOINT;
330 case Instruction::IntToPtr:
331 return Expression::INTTOPTR;
332 case Instruction::BitCast:
333 return Expression::BITCAST;
335 // THIS SHOULD NEVER HAPPEN
337 assert(0 && "Cast operator with unknown opcode?");
338 return Expression::BITCAST;
342 uint32_t ValueTable::hash_operand(Value* v) {
343 if (CallInst* CI = dyn_cast<CallInst>(v))
344 if (CI->getCalledFunction() &&
345 !AA->doesNotAccessMemory(CI->getCalledFunction()))
346 return nextValueNumber++;
348 return lookup_or_add(v);
351 Expression ValueTable::create_expression(CallInst* C) {
354 e.type = C->getType();
358 e.function = C->getCalledFunction();
359 e.opcode = Expression::CALL;
361 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
363 e.varargs.push_back(hash_operand(*I));
368 Expression ValueTable::create_expression(BinaryOperator* BO) {
371 e.firstVN = hash_operand(BO->getOperand(0));
372 e.secondVN = hash_operand(BO->getOperand(1));
375 e.type = BO->getType();
376 e.opcode = getOpcode(BO);
381 Expression ValueTable::create_expression(CmpInst* C) {
384 e.firstVN = hash_operand(C->getOperand(0));
385 e.secondVN = hash_operand(C->getOperand(1));
388 e.type = C->getType();
389 e.opcode = getOpcode(C);
394 Expression ValueTable::create_expression(CastInst* C) {
397 e.firstVN = hash_operand(C->getOperand(0));
401 e.type = C->getType();
402 e.opcode = getOpcode(C);
407 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
410 e.firstVN = hash_operand(S->getOperand(0));
411 e.secondVN = hash_operand(S->getOperand(1));
412 e.thirdVN = hash_operand(S->getOperand(2));
414 e.type = S->getType();
415 e.opcode = Expression::SHUFFLE;
420 Expression ValueTable::create_expression(ExtractElementInst* E) {
423 e.firstVN = hash_operand(E->getOperand(0));
424 e.secondVN = hash_operand(E->getOperand(1));
427 e.type = E->getType();
428 e.opcode = Expression::EXTRACT;
433 Expression ValueTable::create_expression(InsertElementInst* I) {
436 e.firstVN = hash_operand(I->getOperand(0));
437 e.secondVN = hash_operand(I->getOperand(1));
438 e.thirdVN = hash_operand(I->getOperand(2));
440 e.type = I->getType();
441 e.opcode = Expression::INSERT;
446 Expression ValueTable::create_expression(SelectInst* I) {
449 e.firstVN = hash_operand(I->getCondition());
450 e.secondVN = hash_operand(I->getTrueValue());
451 e.thirdVN = hash_operand(I->getFalseValue());
453 e.type = I->getType();
454 e.opcode = Expression::SELECT;
459 Expression ValueTable::create_expression(GetElementPtrInst* G) {
462 e.firstVN = hash_operand(G->getPointerOperand());
466 e.type = G->getType();
467 e.opcode = Expression::GEP;
469 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
471 e.varargs.push_back(hash_operand(*I));
476 //===----------------------------------------------------------------------===//
477 // ValueTable External Functions
478 //===----------------------------------------------------------------------===//
480 /// lookup_or_add - Returns the value number for the specified value, assigning
481 /// it a new number if it did not have one before.
482 uint32_t ValueTable::lookup_or_add(Value* V) {
483 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
484 if (VI != valueNumbering.end())
487 if (CallInst* C = dyn_cast<CallInst>(V)) {
488 if (C->getCalledFunction() &&
489 (AA->doesNotAccessMemory(C->getCalledFunction()) ||
490 AA->onlyReadsMemory(C->getCalledFunction()))) {
491 Expression e = create_expression(C);
493 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
494 if (EI != expressionNumbering.end()) {
495 valueNumbering.insert(std::make_pair(V, EI->second));
498 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
499 valueNumbering.insert(std::make_pair(V, nextValueNumber));
501 return nextValueNumber++;
504 valueNumbering.insert(std::make_pair(V, nextValueNumber));
505 return nextValueNumber++;
507 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
508 Expression e = create_expression(BO);
510 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
511 if (EI != expressionNumbering.end()) {
512 valueNumbering.insert(std::make_pair(V, EI->second));
515 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
516 valueNumbering.insert(std::make_pair(V, nextValueNumber));
518 return nextValueNumber++;
520 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
521 Expression e = create_expression(C);
523 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
524 if (EI != expressionNumbering.end()) {
525 valueNumbering.insert(std::make_pair(V, EI->second));
528 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
529 valueNumbering.insert(std::make_pair(V, nextValueNumber));
531 return nextValueNumber++;
533 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
534 Expression e = create_expression(U);
536 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
537 if (EI != expressionNumbering.end()) {
538 valueNumbering.insert(std::make_pair(V, EI->second));
541 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
542 valueNumbering.insert(std::make_pair(V, nextValueNumber));
544 return nextValueNumber++;
546 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
547 Expression e = create_expression(U);
549 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
550 if (EI != expressionNumbering.end()) {
551 valueNumbering.insert(std::make_pair(V, EI->second));
554 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
555 valueNumbering.insert(std::make_pair(V, nextValueNumber));
557 return nextValueNumber++;
559 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
560 Expression e = create_expression(U);
562 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
563 if (EI != expressionNumbering.end()) {
564 valueNumbering.insert(std::make_pair(V, EI->second));
567 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
568 valueNumbering.insert(std::make_pair(V, nextValueNumber));
570 return nextValueNumber++;
572 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
573 Expression e = create_expression(U);
575 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
576 if (EI != expressionNumbering.end()) {
577 valueNumbering.insert(std::make_pair(V, EI->second));
580 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
581 valueNumbering.insert(std::make_pair(V, nextValueNumber));
583 return nextValueNumber++;
585 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
586 Expression e = create_expression(U);
588 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
589 if (EI != expressionNumbering.end()) {
590 valueNumbering.insert(std::make_pair(V, EI->second));
593 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
594 valueNumbering.insert(std::make_pair(V, nextValueNumber));
596 return nextValueNumber++;
598 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
599 Expression e = create_expression(U);
601 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
602 if (EI != expressionNumbering.end()) {
603 valueNumbering.insert(std::make_pair(V, EI->second));
606 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
607 valueNumbering.insert(std::make_pair(V, nextValueNumber));
609 return nextValueNumber++;
612 valueNumbering.insert(std::make_pair(V, nextValueNumber));
613 return nextValueNumber++;
617 /// lookup - Returns the value number of the specified value. Fails if
618 /// the value has not yet been numbered.
619 uint32_t ValueTable::lookup(Value* V) const {
620 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
621 if (VI != valueNumbering.end())
624 assert(0 && "Value not numbered?");
629 /// clear - Remove all entries from the ValueTable
630 void ValueTable::clear() {
631 valueNumbering.clear();
632 expressionNumbering.clear();
636 /// erase - Remove a value from the value numbering
637 void ValueTable::erase(Value* V) {
638 valueNumbering.erase(V);
641 //===----------------------------------------------------------------------===//
642 // ValueNumberedSet Class
643 //===----------------------------------------------------------------------===//
645 class ValueNumberedSet {
647 SmallPtrSet<Value*, 8> contents;
650 ValueNumberedSet() { numbers.resize(1); }
651 ValueNumberedSet(const ValueNumberedSet& other) {
652 numbers = other.numbers;
653 contents = other.contents;
656 typedef SmallPtrSet<Value*, 8>::iterator iterator;
658 iterator begin() { return contents.begin(); }
659 iterator end() { return contents.end(); }
661 bool insert(Value* v) { return contents.insert(v); }
662 void insert(iterator I, iterator E) { contents.insert(I, E); }
663 void erase(Value* v) { contents.erase(v); }
664 unsigned count(Value* v) { return contents.count(v); }
665 size_t size() { return contents.size(); }
667 void set(unsigned i) {
668 if (i >= numbers.size())
674 void operator=(const ValueNumberedSet& other) {
675 contents = other.contents;
676 numbers = other.numbers;
679 void reset(unsigned i) {
680 if (i < numbers.size())
684 bool test(unsigned i) {
685 if (i >= numbers.size())
688 return numbers.test(i);
698 //===----------------------------------------------------------------------===//
700 //===----------------------------------------------------------------------===//
704 class VISIBILITY_HIDDEN GVN : public FunctionPass {
705 bool runOnFunction(Function &F);
707 static char ID; // Pass identification, replacement for typeid
708 GVN() : FunctionPass((intptr_t)&ID) { }
713 DenseMap<BasicBlock*, ValueNumberedSet> availableOut;
715 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
719 // This transformation requires dominator postdominator info
720 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
721 AU.setPreservesCFG();
722 AU.addRequired<DominatorTree>();
723 AU.addRequired<MemoryDependenceAnalysis>();
724 AU.addRequired<AliasAnalysis>();
725 AU.addPreserved<AliasAnalysis>();
726 AU.addPreserved<MemoryDependenceAnalysis>();
730 // FIXME: eliminate or document these better
731 Value* find_leader(ValueNumberedSet& vals, uint32_t v) ;
732 void val_insert(ValueNumberedSet& s, Value* v);
733 bool processLoad(LoadInst* L,
734 DenseMap<Value*, LoadInst*>& lastLoad,
735 SmallVector<Instruction*, 4>& toErase);
736 bool processInstruction(Instruction* I,
737 ValueNumberedSet& currAvail,
738 DenseMap<Value*, LoadInst*>& lastSeenLoad,
739 SmallVector<Instruction*, 4>& toErase);
740 bool processNonLocalLoad(LoadInst* L,
741 SmallVector<Instruction*, 4>& toErase);
742 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
743 DenseMap<BasicBlock*, Value*> &Phis,
744 bool top_level = false);
745 void dump(DenseMap<BasicBlock*, Value*>& d);
746 bool iterateOnFunction(Function &F);
747 Value* CollapsePhi(PHINode* p);
748 bool isSafeReplacement(PHINode* p, Instruction* inst);
755 // createGVNPass - The public interface to this file...
756 FunctionPass *llvm::createGVNPass() { return new GVN(); }
758 static RegisterPass<GVN> X("gvn",
759 "Global Value Numbering");
761 STATISTIC(NumGVNInstr, "Number of instructions deleted");
762 STATISTIC(NumGVNLoad, "Number of loads deleted");
764 /// find_leader - Given a set and a value number, return the first
765 /// element of the set with that value number, or 0 if no such element
767 Value* GVN::find_leader(ValueNumberedSet& vals, uint32_t v) {
771 for (ValueNumberedSet::iterator I = vals.begin(), E = vals.end();
773 if (v == VN.lookup(*I))
776 assert(0 && "No leader found, but present bit is set?");
780 /// val_insert - Insert a value into a set only if there is not a value
781 /// with the same value number already in the set
782 void GVN::val_insert(ValueNumberedSet& s, Value* v) {
783 uint32_t num = VN.lookup(v);
788 void GVN::dump(DenseMap<BasicBlock*, Value*>& d) {
790 for (DenseMap<BasicBlock*, Value*>::iterator I = d.begin(),
791 E = d.end(); I != E; ++I) {
792 if (I->second == MemoryDependenceAnalysis::None)
800 Value* GVN::CollapsePhi(PHINode* p) {
801 DominatorTree &DT = getAnalysis<DominatorTree>();
802 Value* constVal = p->hasConstantValue();
805 if (Instruction* inst = dyn_cast<Instruction>(constVal)) {
806 if (DT.dominates(inst, p))
807 if (isSafeReplacement(p, inst))
817 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
818 if (!isa<PHINode>(inst))
821 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
823 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
824 if (use_phi->getParent() == inst->getParent())
830 /// GetValueForBlock - Get the value to use within the specified basic block.
831 /// available values are in Phis.
832 Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
833 DenseMap<BasicBlock*, Value*> &Phis,
836 // If we have already computed this value, return the previously computed val.
837 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
838 if (V != Phis.end() && !top_level) return V->second;
840 BasicBlock* singlePred = BB->getSinglePredecessor();
842 Value *ret = GetValueForBlock(singlePred, orig, Phis);
846 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
847 // now, then get values to fill in the incoming values for the PHI.
848 PHINode *PN = new PHINode(orig->getType(), orig->getName()+".rle",
850 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
852 if (Phis.count(BB) == 0)
853 Phis.insert(std::make_pair(BB, PN));
855 // Fill in the incoming values for the block.
856 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
857 Value* val = GetValueForBlock(*PI, orig, Phis);
859 PN->addIncoming(val, *PI);
862 // Attempt to collapse PHI nodes that are trivially redundant
863 Value* v = CollapsePhi(PN);
865 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
867 MD.removeInstruction(PN);
868 PN->replaceAllUsesWith(v);
870 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
871 E = Phis.end(); I != E; ++I)
875 PN->eraseFromParent();
882 // Cache our phi construction results
883 phiMap[orig->getPointerOperand()].insert(PN);
887 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
888 /// non-local by performing PHI construction.
889 bool GVN::processNonLocalLoad(LoadInst* L,
890 SmallVector<Instruction*, 4>& toErase) {
891 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
893 // Find the non-local dependencies of the load
894 DenseMap<BasicBlock*, Value*> deps;
895 MD.getNonLocalDependency(L, deps);
897 DenseMap<BasicBlock*, Value*> repl;
899 // Filter out useless results (non-locals, etc)
900 for (DenseMap<BasicBlock*, Value*>::iterator I = deps.begin(), E = deps.end();
902 if (I->second == MemoryDependenceAnalysis::None) {
904 } else if (I->second == MemoryDependenceAnalysis::NonLocal) {
906 } else if (StoreInst* S = dyn_cast<StoreInst>(I->second)) {
907 if (S->getPointerOperand() == L->getPointerOperand())
908 repl[I->first] = S->getOperand(0);
911 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second)) {
912 if (LD->getPointerOperand() == L->getPointerOperand())
920 // Use cached PHI construction information from previous runs
921 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
922 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
924 if ((*I)->getParent() == L->getParent()) {
925 MD.removeInstruction(L);
926 L->replaceAllUsesWith(*I);
927 toErase.push_back(L);
932 repl.insert(std::make_pair((*I)->getParent(), *I));
936 // Perform PHI construction
937 SmallPtrSet<BasicBlock*, 4> visited;
938 Value* v = GetValueForBlock(L->getParent(), L, repl, true);
940 MD.removeInstruction(L);
941 L->replaceAllUsesWith(v);
942 toErase.push_back(L);
948 /// processLoad - Attempt to eliminate a load, first by eliminating it
949 /// locally, and then attempting non-local elimination if that fails.
950 bool GVN::processLoad(LoadInst* L,
951 DenseMap<Value*, LoadInst*>& lastLoad,
952 SmallVector<Instruction*, 4>& toErase) {
953 if (L->isVolatile()) {
954 lastLoad[L->getPointerOperand()] = L;
958 Value* pointer = L->getPointerOperand();
959 LoadInst*& last = lastLoad[pointer];
961 // ... to a pointer that has been loaded from before...
962 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
963 bool removedNonLocal = false;
964 Instruction* dep = MD.getDependency(L);
965 if (dep == MemoryDependenceAnalysis::NonLocal &&
966 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
967 removedNonLocal = processNonLocalLoad(L, toErase);
969 if (!removedNonLocal)
972 return removedNonLocal;
976 bool deletedLoad = false;
978 // Walk up the dependency chain until we either find
979 // a dependency we can use, or we can't walk any further
980 while (dep != MemoryDependenceAnalysis::None &&
981 dep != MemoryDependenceAnalysis::NonLocal &&
982 (isa<LoadInst>(dep) || isa<StoreInst>(dep))) {
983 // ... that depends on a store ...
984 if (StoreInst* S = dyn_cast<StoreInst>(dep)) {
985 if (S->getPointerOperand() == pointer) {
987 MD.removeInstruction(L);
989 L->replaceAllUsesWith(S->getOperand(0));
990 toErase.push_back(L);
995 // Whether we removed it or not, we can't
999 // If we don't depend on a store, and we haven't
1000 // been loaded before, bail.
1002 } else if (dep == last) {
1004 MD.removeInstruction(L);
1006 L->replaceAllUsesWith(last);
1007 toErase.push_back(L);
1013 dep = MD.getDependency(L, dep);
1023 /// processInstruction - When calculating availability, handle an instruction
1024 /// by inserting it into the appropriate sets
1025 bool GVN::processInstruction(Instruction* I,
1026 ValueNumberedSet& currAvail,
1027 DenseMap<Value*, LoadInst*>& lastSeenLoad,
1028 SmallVector<Instruction*, 4>& toErase) {
1029 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1030 return processLoad(L, lastSeenLoad, toErase);
1033 unsigned num = VN.lookup_or_add(I);
1035 // Collapse PHI nodes
1036 if (PHINode* p = dyn_cast<PHINode>(I)) {
1037 Value* constVal = CollapsePhi(p);
1040 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1042 if (PI->second.count(p))
1043 PI->second.erase(p);
1045 p->replaceAllUsesWith(constVal);
1046 toErase.push_back(p);
1048 // Perform value-number based elimination
1049 } else if (currAvail.test(num)) {
1050 Value* repl = find_leader(currAvail, num);
1052 if (CallInst* CI = dyn_cast<CallInst>(I)) {
1053 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
1054 if (CI->getCalledFunction() &&
1055 !AA.doesNotAccessMemory(CI->getCalledFunction())) {
1056 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1057 if (cast<Instruction>(repl)->getParent() != CI->getParent() ||
1058 MD.getDependency(CI) != MD.getDependency(cast<CallInst>(repl))) {
1059 // There must be an intervening may-alias store, so nothing from
1060 // this point on will be able to be replaced with the preceding call
1061 currAvail.erase(repl);
1062 currAvail.insert(I);
1070 I->replaceAllUsesWith(repl);
1071 toErase.push_back(I);
1073 } else if (!I->isTerminator()) {
1075 currAvail.insert(I);
1081 // GVN::runOnFunction - This is the main transformation entry point for a
1084 bool GVN::runOnFunction(Function& F) {
1085 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1087 bool changed = false;
1088 bool shouldContinue = true;
1090 while (shouldContinue) {
1091 shouldContinue = iterateOnFunction(F);
1092 changed |= shouldContinue;
1099 // GVN::iterateOnFunction - Executes one iteration of GVN
1100 bool GVN::iterateOnFunction(Function &F) {
1101 // Clean out global sets from any previous functions
1103 availableOut.clear();
1106 bool changed_function = false;
1108 DominatorTree &DT = getAnalysis<DominatorTree>();
1110 SmallVector<Instruction*, 4> toErase;
1112 // Top-down walk of the dominator tree
1113 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1114 E = df_end(DT.getRootNode()); DI != E; ++DI) {
1116 // Get the set to update for this block
1117 ValueNumberedSet& currAvail = availableOut[DI->getBlock()];
1118 DenseMap<Value*, LoadInst*> lastSeenLoad;
1120 BasicBlock* BB = DI->getBlock();
1122 // A block inherits AVAIL_OUT from its dominator
1123 if (DI->getIDom() != 0)
1124 currAvail = availableOut[DI->getIDom()->getBlock()];
1126 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1128 changed_function |= processInstruction(BI, currAvail,
1129 lastSeenLoad, toErase);
1131 NumGVNInstr += toErase.size();
1133 // Avoid iterator invalidation
1136 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1137 E = toErase.end(); I != E; ++I)
1138 (*I)->eraseFromParent();
1144 return changed_function;