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 // Note that this pass does the value numbering itself, it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/BasicBlock.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Value.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.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/Transforms/Utils/BasicBlockUtils.h"
42 STATISTIC(NumGVNInstr, "Number of instructions deleted");
43 STATISTIC(NumGVNLoad, "Number of loads deleted");
44 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
45 STATISTIC(NumGVNBlocks, "Number of blocks merged");
46 STATISTIC(NumPRELoad, "Number of loads PRE'd");
48 static cl::opt<bool> EnablePRE("enable-pre",
49 cl::init(true), cl::Hidden);
50 cl::opt<bool> EnableLoadPRE("enable-load-pre");
52 //===----------------------------------------------------------------------===//
54 //===----------------------------------------------------------------------===//
56 /// This class holds the mapping between values and value numbers. It is used
57 /// as an efficient mechanism to determine the expression-wise equivalence of
60 struct VISIBILITY_HIDDEN Expression {
61 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
62 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
63 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
64 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
65 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
66 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
67 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
68 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
69 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
70 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
73 ExpressionOpcode opcode;
78 SmallVector<uint32_t, 4> varargs;
82 Expression(ExpressionOpcode o) : opcode(o) { }
84 bool operator==(const Expression &other) const {
85 if (opcode != other.opcode)
87 else if (opcode == EMPTY || opcode == TOMBSTONE)
89 else if (type != other.type)
91 else if (function != other.function)
93 else if (firstVN != other.firstVN)
95 else if (secondVN != other.secondVN)
97 else if (thirdVN != other.thirdVN)
100 if (varargs.size() != other.varargs.size())
103 for (size_t i = 0; i < varargs.size(); ++i)
104 if (varargs[i] != other.varargs[i])
111 bool operator!=(const Expression &other) const {
112 if (opcode != other.opcode)
114 else if (opcode == EMPTY || opcode == TOMBSTONE)
116 else if (type != other.type)
118 else if (function != other.function)
120 else if (firstVN != other.firstVN)
122 else if (secondVN != other.secondVN)
124 else if (thirdVN != other.thirdVN)
127 if (varargs.size() != other.varargs.size())
130 for (size_t i = 0; i < varargs.size(); ++i)
131 if (varargs[i] != other.varargs[i])
139 class VISIBILITY_HIDDEN ValueTable {
141 DenseMap<Value*, uint32_t> valueNumbering;
142 DenseMap<Expression, uint32_t> expressionNumbering;
144 MemoryDependenceAnalysis* MD;
147 uint32_t nextValueNumber;
149 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
150 Expression::ExpressionOpcode getOpcode(CmpInst* C);
151 Expression::ExpressionOpcode getOpcode(CastInst* C);
152 Expression create_expression(BinaryOperator* BO);
153 Expression create_expression(CmpInst* C);
154 Expression create_expression(ShuffleVectorInst* V);
155 Expression create_expression(ExtractElementInst* C);
156 Expression create_expression(InsertElementInst* V);
157 Expression create_expression(SelectInst* V);
158 Expression create_expression(CastInst* C);
159 Expression create_expression(GetElementPtrInst* G);
160 Expression create_expression(CallInst* C);
161 Expression create_expression(Constant* C);
163 ValueTable() : nextValueNumber(1) { }
164 uint32_t lookup_or_add(Value* V);
165 uint32_t lookup(Value* V) const;
166 void add(Value* V, uint32_t num);
168 void erase(Value* v);
170 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
171 AliasAnalysis *getAliasAnalysis() const { return AA; }
172 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
173 void setDomTree(DominatorTree* D) { DT = D; }
174 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
179 template <> struct DenseMapInfo<Expression> {
180 static inline Expression getEmptyKey() {
181 return Expression(Expression::EMPTY);
184 static inline Expression getTombstoneKey() {
185 return Expression(Expression::TOMBSTONE);
188 static unsigned getHashValue(const Expression e) {
189 unsigned hash = e.opcode;
191 hash = e.firstVN + hash * 37;
192 hash = e.secondVN + hash * 37;
193 hash = e.thirdVN + hash * 37;
195 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
196 (unsigned)((uintptr_t)e.type >> 9)) +
199 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
200 E = e.varargs.end(); I != E; ++I)
201 hash = *I + hash * 37;
203 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
204 (unsigned)((uintptr_t)e.function >> 9)) +
209 static bool isEqual(const Expression &LHS, const Expression &RHS) {
212 static bool isPod() { return true; }
216 //===----------------------------------------------------------------------===//
217 // ValueTable Internal Functions
218 //===----------------------------------------------------------------------===//
219 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
220 switch(BO->getOpcode()) {
221 default: // THIS SHOULD NEVER HAPPEN
222 assert(0 && "Binary operator with unknown opcode?");
223 case Instruction::Add: return Expression::ADD;
224 case Instruction::Sub: return Expression::SUB;
225 case Instruction::Mul: return Expression::MUL;
226 case Instruction::UDiv: return Expression::UDIV;
227 case Instruction::SDiv: return Expression::SDIV;
228 case Instruction::FDiv: return Expression::FDIV;
229 case Instruction::URem: return Expression::UREM;
230 case Instruction::SRem: return Expression::SREM;
231 case Instruction::FRem: return Expression::FREM;
232 case Instruction::Shl: return Expression::SHL;
233 case Instruction::LShr: return Expression::LSHR;
234 case Instruction::AShr: return Expression::ASHR;
235 case Instruction::And: return Expression::AND;
236 case Instruction::Or: return Expression::OR;
237 case Instruction::Xor: return Expression::XOR;
241 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
242 if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
243 switch (C->getPredicate()) {
244 default: // THIS SHOULD NEVER HAPPEN
245 assert(0 && "Comparison with unknown predicate?");
246 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
247 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
248 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
249 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
250 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
251 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
252 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
253 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
254 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
255 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
258 assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
259 switch (C->getPredicate()) {
260 default: // THIS SHOULD NEVER HAPPEN
261 assert(0 && "Comparison with unknown predicate?");
262 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
263 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
264 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
265 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
266 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
267 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
268 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
269 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
270 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
271 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
272 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
273 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
274 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
275 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
279 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
280 switch(C->getOpcode()) {
281 default: // THIS SHOULD NEVER HAPPEN
282 assert(0 && "Cast operator with unknown opcode?");
283 case Instruction::Trunc: return Expression::TRUNC;
284 case Instruction::ZExt: return Expression::ZEXT;
285 case Instruction::SExt: return Expression::SEXT;
286 case Instruction::FPToUI: return Expression::FPTOUI;
287 case Instruction::FPToSI: return Expression::FPTOSI;
288 case Instruction::UIToFP: return Expression::UITOFP;
289 case Instruction::SIToFP: return Expression::SITOFP;
290 case Instruction::FPTrunc: return Expression::FPTRUNC;
291 case Instruction::FPExt: return Expression::FPEXT;
292 case Instruction::PtrToInt: return Expression::PTRTOINT;
293 case Instruction::IntToPtr: return Expression::INTTOPTR;
294 case Instruction::BitCast: return Expression::BITCAST;
298 Expression ValueTable::create_expression(CallInst* C) {
301 e.type = C->getType();
305 e.function = C->getCalledFunction();
306 e.opcode = Expression::CALL;
308 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
310 e.varargs.push_back(lookup_or_add(*I));
315 Expression ValueTable::create_expression(BinaryOperator* BO) {
318 e.firstVN = lookup_or_add(BO->getOperand(0));
319 e.secondVN = lookup_or_add(BO->getOperand(1));
322 e.type = BO->getType();
323 e.opcode = getOpcode(BO);
328 Expression ValueTable::create_expression(CmpInst* C) {
331 e.firstVN = lookup_or_add(C->getOperand(0));
332 e.secondVN = lookup_or_add(C->getOperand(1));
335 e.type = C->getType();
336 e.opcode = getOpcode(C);
341 Expression ValueTable::create_expression(CastInst* C) {
344 e.firstVN = lookup_or_add(C->getOperand(0));
348 e.type = C->getType();
349 e.opcode = getOpcode(C);
354 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
357 e.firstVN = lookup_or_add(S->getOperand(0));
358 e.secondVN = lookup_or_add(S->getOperand(1));
359 e.thirdVN = lookup_or_add(S->getOperand(2));
361 e.type = S->getType();
362 e.opcode = Expression::SHUFFLE;
367 Expression ValueTable::create_expression(ExtractElementInst* E) {
370 e.firstVN = lookup_or_add(E->getOperand(0));
371 e.secondVN = lookup_or_add(E->getOperand(1));
374 e.type = E->getType();
375 e.opcode = Expression::EXTRACT;
380 Expression ValueTable::create_expression(InsertElementInst* I) {
383 e.firstVN = lookup_or_add(I->getOperand(0));
384 e.secondVN = lookup_or_add(I->getOperand(1));
385 e.thirdVN = lookup_or_add(I->getOperand(2));
387 e.type = I->getType();
388 e.opcode = Expression::INSERT;
393 Expression ValueTable::create_expression(SelectInst* I) {
396 e.firstVN = lookup_or_add(I->getCondition());
397 e.secondVN = lookup_or_add(I->getTrueValue());
398 e.thirdVN = lookup_or_add(I->getFalseValue());
400 e.type = I->getType();
401 e.opcode = Expression::SELECT;
406 Expression ValueTable::create_expression(GetElementPtrInst* G) {
409 e.firstVN = lookup_or_add(G->getPointerOperand());
413 e.type = G->getType();
414 e.opcode = Expression::GEP;
416 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
418 e.varargs.push_back(lookup_or_add(*I));
423 //===----------------------------------------------------------------------===//
424 // ValueTable External Functions
425 //===----------------------------------------------------------------------===//
427 /// add - Insert a value into the table with a specified value number.
428 void ValueTable::add(Value* V, uint32_t num) {
429 valueNumbering.insert(std::make_pair(V, num));
432 /// lookup_or_add - Returns the value number for the specified value, assigning
433 /// it a new number if it did not have one before.
434 uint32_t ValueTable::lookup_or_add(Value* V) {
435 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
436 if (VI != valueNumbering.end())
439 if (CallInst* C = dyn_cast<CallInst>(V)) {
440 if (AA->doesNotAccessMemory(C)) {
441 Expression e = create_expression(C);
443 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
444 if (EI != expressionNumbering.end()) {
445 valueNumbering.insert(std::make_pair(V, EI->second));
448 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
449 valueNumbering.insert(std::make_pair(V, nextValueNumber));
451 return nextValueNumber++;
453 } else if (AA->onlyReadsMemory(C)) {
454 Expression e = create_expression(C);
456 if (expressionNumbering.find(e) == expressionNumbering.end()) {
457 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
458 valueNumbering.insert(std::make_pair(V, nextValueNumber));
459 return nextValueNumber++;
462 MemDepResult local_dep = MD->getDependency(C);
464 if (local_dep.isNone()) {
465 valueNumbering.insert(std::make_pair(V, nextValueNumber));
466 return nextValueNumber++;
469 if (Instruction *LocalDepInst = local_dep.getInst()) {
470 if (!isa<CallInst>(LocalDepInst)) {
471 valueNumbering.insert(std::make_pair(V, nextValueNumber));
472 return nextValueNumber++;
475 CallInst* local_cdep = cast<CallInst>(LocalDepInst);
477 if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
478 local_cdep->getNumOperands() != C->getNumOperands()) {
479 valueNumbering.insert(std::make_pair(V, nextValueNumber));
480 return nextValueNumber++;
483 if (!C->getCalledFunction()) {
484 valueNumbering.insert(std::make_pair(V, nextValueNumber));
485 return nextValueNumber++;
488 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
489 uint32_t c_vn = lookup_or_add(C->getOperand(i));
490 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
492 valueNumbering.insert(std::make_pair(V, nextValueNumber));
493 return nextValueNumber++;
497 uint32_t v = lookup_or_add(local_cdep);
498 valueNumbering.insert(std::make_pair(V, v));
503 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
504 MD->getNonLocalDependency(C);
507 // Check to see if we have a single dominating call instruction that is
509 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
510 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
511 // Ignore non-local dependencies.
512 if (I->second.isNonLocal())
515 // We don't handle non-depedencies. If we already have a call, reject
516 // instruction dependencies.
517 if (I->second.isNone() || cdep != 0) {
522 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
523 // FIXME: All duplicated with non-local case.
524 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
525 cdep = NonLocalDepCall;
534 valueNumbering.insert(std::make_pair(V, nextValueNumber));
535 return nextValueNumber++;
538 if (cdep->getCalledFunction() != C->getCalledFunction() ||
539 cdep->getNumOperands() != C->getNumOperands()) {
540 valueNumbering.insert(std::make_pair(V, nextValueNumber));
541 return nextValueNumber++;
543 if (!C->getCalledFunction()) {
544 valueNumbering.insert(std::make_pair(V, nextValueNumber));
545 return nextValueNumber++;
547 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
548 uint32_t c_vn = lookup_or_add(C->getOperand(i));
549 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
551 valueNumbering.insert(std::make_pair(V, nextValueNumber));
552 return nextValueNumber++;
556 uint32_t v = lookup_or_add(cdep);
557 valueNumbering.insert(std::make_pair(V, v));
561 valueNumbering.insert(std::make_pair(V, nextValueNumber));
562 return nextValueNumber++;
564 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
565 Expression e = create_expression(BO);
567 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
568 if (EI != expressionNumbering.end()) {
569 valueNumbering.insert(std::make_pair(V, EI->second));
572 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
573 valueNumbering.insert(std::make_pair(V, nextValueNumber));
575 return nextValueNumber++;
577 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
578 Expression e = create_expression(C);
580 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
581 if (EI != expressionNumbering.end()) {
582 valueNumbering.insert(std::make_pair(V, EI->second));
585 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
586 valueNumbering.insert(std::make_pair(V, nextValueNumber));
588 return nextValueNumber++;
590 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
591 Expression e = create_expression(U);
593 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
594 if (EI != expressionNumbering.end()) {
595 valueNumbering.insert(std::make_pair(V, EI->second));
598 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
599 valueNumbering.insert(std::make_pair(V, nextValueNumber));
601 return nextValueNumber++;
603 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
604 Expression e = create_expression(U);
606 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
607 if (EI != expressionNumbering.end()) {
608 valueNumbering.insert(std::make_pair(V, EI->second));
611 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
612 valueNumbering.insert(std::make_pair(V, nextValueNumber));
614 return nextValueNumber++;
616 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
617 Expression e = create_expression(U);
619 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
620 if (EI != expressionNumbering.end()) {
621 valueNumbering.insert(std::make_pair(V, EI->second));
624 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
625 valueNumbering.insert(std::make_pair(V, nextValueNumber));
627 return nextValueNumber++;
629 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
630 Expression e = create_expression(U);
632 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
633 if (EI != expressionNumbering.end()) {
634 valueNumbering.insert(std::make_pair(V, EI->second));
637 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
638 valueNumbering.insert(std::make_pair(V, nextValueNumber));
640 return nextValueNumber++;
642 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
643 Expression e = create_expression(U);
645 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
646 if (EI != expressionNumbering.end()) {
647 valueNumbering.insert(std::make_pair(V, EI->second));
650 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
651 valueNumbering.insert(std::make_pair(V, nextValueNumber));
653 return nextValueNumber++;
655 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
656 Expression e = create_expression(U);
658 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
659 if (EI != expressionNumbering.end()) {
660 valueNumbering.insert(std::make_pair(V, EI->second));
663 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
664 valueNumbering.insert(std::make_pair(V, nextValueNumber));
666 return nextValueNumber++;
669 valueNumbering.insert(std::make_pair(V, nextValueNumber));
670 return nextValueNumber++;
674 /// lookup - Returns the value number of the specified value. Fails if
675 /// the value has not yet been numbered.
676 uint32_t ValueTable::lookup(Value* V) const {
677 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
678 assert(VI != valueNumbering.end() && "Value not numbered?");
682 /// clear - Remove all entries from the ValueTable
683 void ValueTable::clear() {
684 valueNumbering.clear();
685 expressionNumbering.clear();
689 /// erase - Remove a value from the value numbering
690 void ValueTable::erase(Value* V) {
691 valueNumbering.erase(V);
694 //===----------------------------------------------------------------------===//
696 //===----------------------------------------------------------------------===//
699 struct VISIBILITY_HIDDEN ValueNumberScope {
700 ValueNumberScope* parent;
701 DenseMap<uint32_t, Value*> table;
703 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
709 class VISIBILITY_HIDDEN GVN : public FunctionPass {
710 bool runOnFunction(Function &F);
712 static char ID; // Pass identification, replacement for typeid
713 GVN() : FunctionPass(&ID) { }
716 MemoryDependenceAnalysis *MD;
720 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
722 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
726 // This transformation requires dominator postdominator info
727 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
728 AU.addRequired<DominatorTree>();
729 AU.addRequired<MemoryDependenceAnalysis>();
730 AU.addRequired<AliasAnalysis>();
732 AU.addPreserved<DominatorTree>();
733 AU.addPreserved<AliasAnalysis>();
737 // FIXME: eliminate or document these better
738 bool processLoad(LoadInst* L,
739 DenseMap<Value*, LoadInst*> &lastLoad,
740 SmallVectorImpl<Instruction*> &toErase);
741 bool processInstruction(Instruction* I,
742 DenseMap<Value*, LoadInst*>& lastSeenLoad,
743 SmallVectorImpl<Instruction*> &toErase);
744 bool processNonLocalLoad(LoadInst* L,
745 SmallVectorImpl<Instruction*> &toErase);
746 bool processBlock(DomTreeNode* DTN);
747 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
748 DenseMap<BasicBlock*, Value*> &Phis,
749 bool top_level = false);
750 void dump(DenseMap<uint32_t, Value*>& d);
751 bool iterateOnFunction(Function &F);
752 Value* CollapsePhi(PHINode* p);
753 bool isSafeReplacement(PHINode* p, Instruction* inst);
754 bool performPRE(Function& F);
755 Value* lookupNumber(BasicBlock* BB, uint32_t num);
756 bool mergeBlockIntoPredecessor(BasicBlock* BB);
757 void cleanupGlobalSets();
763 // createGVNPass - The public interface to this file...
764 FunctionPass *llvm::createGVNPass() { return new GVN(); }
766 static RegisterPass<GVN> X("gvn",
767 "Global Value Numbering");
769 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
771 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
772 E = d.end(); I != E; ++I) {
773 printf("%d\n", I->first);
779 Value* GVN::CollapsePhi(PHINode* p) {
780 Value* constVal = p->hasConstantValue();
781 if (!constVal) return 0;
783 Instruction* inst = dyn_cast<Instruction>(constVal);
787 if (DT->dominates(inst, p))
788 if (isSafeReplacement(p, inst))
793 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
794 if (!isa<PHINode>(inst))
797 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
799 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
800 if (use_phi->getParent() == inst->getParent())
806 /// GetValueForBlock - Get the value to use within the specified basic block.
807 /// available values are in Phis.
808 Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
809 DenseMap<BasicBlock*, Value*> &Phis,
812 // If we have already computed this value, return the previously computed val.
813 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
814 if (V != Phis.end() && !top_level) return V->second;
816 // If the block is unreachable, just return undef, since this path
817 // can't actually occur at runtime.
818 if (!DT->isReachableFromEntry(BB))
819 return Phis[BB] = UndefValue::get(orig->getType());
821 BasicBlock* singlePred = BB->getSinglePredecessor();
823 Value *ret = GetValueForBlock(singlePred, orig, Phis);
828 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
829 // now, then get values to fill in the incoming values for the PHI.
830 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
832 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
834 if (Phis.count(BB) == 0)
835 Phis.insert(std::make_pair(BB, PN));
837 // Fill in the incoming values for the block.
838 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
839 Value* val = GetValueForBlock(*PI, orig, Phis);
840 PN->addIncoming(val, *PI);
843 VN.getAliasAnalysis()->copyValue(orig, PN);
845 // Attempt to collapse PHI nodes that are trivially redundant
846 Value* v = CollapsePhi(PN);
848 // Cache our phi construction results
849 phiMap[orig->getPointerOperand()].insert(PN);
853 PN->replaceAllUsesWith(v);
855 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
856 E = Phis.end(); I != E; ++I)
860 DEBUG(cerr << "GVN removed: " << *PN);
861 MD->removeInstruction(PN);
862 PN->eraseFromParent();
868 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
869 /// we're analyzing is fully available in the specified block. As we go, keep
870 /// track of which blocks we know it is fully alive or not in
871 /// FullyAvailableBlocks.
872 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
873 DenseMap<BasicBlock*, bool> &FullyAvailableBlocks) {
874 // Optimistically assume that the block is fully available and check to see
875 // if we already know about this block in one lookup.
876 std::pair<DenseMap<BasicBlock*, bool>::iterator, bool> IV =
877 FullyAvailableBlocks.insert(std::make_pair(BB, true));
879 // If the entry already existed for this block, return the precomputed value.
881 return IV.first->second;
883 // Otherwise, see if it is fully available in all predecessors.
884 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
886 // If this block has no predecessors, it isn't live-in here.
888 return FullyAvailableBlocks[BB] = false;
890 for (; PI != PE; ++PI)
891 // If the value isn't fully available in one of our predecessors, then it
892 // isn't fully available in this block either. Undo our previous
893 // optimistic assumption and bail out.
894 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
895 return FullyAvailableBlocks[BB] = false;
900 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
901 /// non-local by performing PHI construction.
902 bool GVN::processNonLocalLoad(LoadInst *LI,
903 SmallVectorImpl<Instruction*> &toErase) {
904 // Find the non-local dependencies of the load.
905 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
906 MD->getNonLocalDependency(LI);
907 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << deps.size() << *LI);
909 // If we had to process more than one hundred blocks to find the
910 // dependencies, this load isn't worth worrying about. Optimizing
911 // it will be too expensive.
912 if (deps.size() > 100)
915 BasicBlock *EntryBlock = &LI->getParent()->getParent()->getEntryBlock();
917 // Filter out useless results (non-locals, etc). Keep track of the blocks
918 // where we have a value available in repl, also keep track of whether we see
919 // dependencies that produce an unknown value for the load (such as a call
920 // that could potentially clobber the load).
921 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
922 SmallVector<BasicBlock*, 16> UnavailableBlocks;
924 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
925 BasicBlock *DepBB = deps[i].first;
926 MemDepResult DepInfo = deps[i].second;
928 if (DepInfo.isNonLocal()) {
929 // If this is a non-local dependency in the entry block, then we depend on
930 // the value live-in at the start of the function. We could insert a load
931 // in the entry block to get this, but for now we'll just bail out.
932 if (DepBB == EntryBlock)
933 UnavailableBlocks.push_back(DepBB);
937 if (DepInfo.isNone()) {
938 ValuesPerBlock.push_back(std::make_pair(DepBB,
939 UndefValue::get(LI->getType())));
943 if (StoreInst* S = dyn_cast<StoreInst>(DepInfo.getInst())) {
944 // Reject loads and stores that are to the same address but are of
946 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
947 // of bitfield access, it would be interesting to optimize for it at some
949 if (S->getOperand(0)->getType() != LI->getType()) {
950 UnavailableBlocks.push_back(DepBB);
954 if (S->getPointerOperand() != LI->getPointerOperand() &&
955 VN.getAliasAnalysis()->alias(S->getPointerOperand(), 1,
956 LI->getPointerOperand(), 1)
957 != AliasAnalysis::MustAlias) {
958 UnavailableBlocks.push_back(DepBB);
961 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
963 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInfo.getInst())) {
964 if (LD->getType() != LI->getType()) {
965 UnavailableBlocks.push_back(DepBB);
969 if (LD->getPointerOperand() != LI->getPointerOperand() &&
970 VN.getAliasAnalysis()->alias(LD->getPointerOperand(), 1,
971 LI->getPointerOperand(), 1)
972 != AliasAnalysis::MustAlias) {
973 UnavailableBlocks.push_back(DepBB);
976 ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
978 UnavailableBlocks.push_back(DepBB);
983 // If we have no predecessors that produce a known value for this load, exit
985 if (ValuesPerBlock.empty()) return false;
987 // If all of the instructions we depend on produce a known value for this
988 // load, then it is fully redundant and we can use PHI insertion to compute
989 // its value. Insert PHIs and remove the fully redundant value now.
990 if (UnavailableBlocks.empty()) {
991 // Use cached PHI construction information from previous runs
992 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
993 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
995 if ((*I)->getParent() == LI->getParent()) {
996 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI);
997 LI->replaceAllUsesWith(*I);
998 toErase.push_back(LI);
1003 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
1006 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI);
1008 DenseMap<BasicBlock*, Value*> BlockReplValues;
1009 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1010 // Perform PHI construction.
1011 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1012 LI->replaceAllUsesWith(v);
1013 toErase.push_back(LI);
1018 if (!EnablePRE || !EnableLoadPRE)
1021 // Okay, we have *some* definitions of the value. This means that the value
1022 // is available in some of our (transitive) predecessors. Lets think about
1023 // doing PRE of this load. This will involve inserting a new load into the
1024 // predecessor when it's not available. We could do this in general, but
1025 // prefer to not increase code size. As such, we only do this when we know
1026 // that we only have to insert *one* load (which means we're basically moving
1027 // the load, not inserting a new one).
1029 // Everything we do here is based on local predecessors of LI's block. If it
1030 // only has one predecessor, bail now.
1031 BasicBlock *LoadBB = LI->getParent();
1032 if (LoadBB->getSinglePredecessor())
1035 // If we have a repl set with LI itself in it, this means we have a loop where
1036 // at least one of the values is LI. Since this means that we won't be able
1037 // to eliminate LI even if we insert uses in the other predecessors, we will
1038 // end up increasing code size. Reject this by scanning for LI.
1039 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1040 if (ValuesPerBlock[i].second == LI)
1043 // Okay, we have some hope :). Check to see if the loaded value is fully
1044 // available in all but one predecessor.
1045 // FIXME: If we could restructure the CFG, we could make a common pred with
1046 // all the preds that don't have an available LI and insert a new load into
1048 BasicBlock *UnavailablePred = 0;
1050 DenseMap<BasicBlock*, bool> FullyAvailableBlocks;
1051 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1052 FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
1053 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1054 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1056 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1058 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1061 // If this load is not available in multiple predecessors, reject it.
1062 if (UnavailablePred && UnavailablePred != *PI)
1064 UnavailablePred = *PI;
1067 assert(UnavailablePred != 0 &&
1068 "Fully available value should be eliminated above!");
1070 // If the loaded pointer is PHI node defined in this block, do PHI translation
1071 // to get its value in the predecessor.
1072 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
1074 // Make sure the value is live in the predecessor. If it was defined by a
1075 // non-PHI instruction in this block, we don't know how to recompute it above.
1076 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
1077 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
1078 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
1079 << *LPInst << *LI << "\n");
1083 // We don't currently handle critical edges :(
1084 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1085 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1086 << UnavailablePred->getName() << "': " << *LI);
1090 // Okay, we can eliminate this load by inserting a reload in the predecessor
1091 // and using PHI construction to get the value in the other predecessors, do
1093 DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI);
1095 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1097 UnavailablePred->getTerminator());
1099 DenseMap<BasicBlock*, Value*> BlockReplValues;
1100 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1101 BlockReplValues[UnavailablePred] = NewLoad;
1103 // Perform PHI construction.
1104 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1105 LI->replaceAllUsesWith(v);
1106 toErase.push_back(LI);
1111 /// processLoad - Attempt to eliminate a load, first by eliminating it
1112 /// locally, and then attempting non-local elimination if that fails.
1113 bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
1114 SmallVectorImpl<Instruction*> &toErase) {
1115 if (L->isVolatile()) {
1116 lastLoad[L->getPointerOperand()] = L;
1120 Value* pointer = L->getPointerOperand();
1121 LoadInst*& last = lastLoad[pointer];
1123 // ... to a pointer that has been loaded from before...
1124 bool removedNonLocal = false;
1125 MemDepResult dep = MD->getDependency(L);
1126 if (dep.isNonLocal() &&
1127 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
1128 removedNonLocal = processNonLocalLoad(L, toErase);
1130 if (!removedNonLocal)
1133 return removedNonLocal;
1137 bool deletedLoad = false;
1139 // Walk up the dependency chain until we either find
1140 // a dependency we can use, or we can't walk any further
1141 while (Instruction *DepInst = dep.getInst()) {
1142 // ... that depends on a store ...
1143 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
1144 if (S->getPointerOperand() == pointer) {
1146 L->replaceAllUsesWith(S->getOperand(0));
1147 toErase.push_back(L);
1152 // Whether we removed it or not, we can't
1155 } else if (!isa<LoadInst>(DepInst)) {
1156 // Only want to handle loads below.
1159 // If we don't depend on a store, and we haven't
1160 // been loaded before, bail.
1162 } else if (DepInst == last) {
1164 L->replaceAllUsesWith(last);
1165 toErase.push_back(L);
1170 dep = MD->getDependencyFrom(L, DepInst, DepInst->getParent());
1174 // If this load really doesn't depend on anything, then we must be loading an
1175 // undef value. This can happen when loading for a fresh allocation with no
1176 // intervening stores, for example.
1178 // If this load depends directly on an allocation, there isn't
1179 // anything stored there; therefore, we can optimize this load
1181 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1182 toErase.push_back(L);
1193 Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
1194 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1195 if (I == localAvail.end())
1198 ValueNumberScope* locals = I->second;
1201 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
1202 if (I != locals->table.end())
1205 locals = locals->parent;
1211 /// processInstruction - When calculating availability, handle an instruction
1212 /// by inserting it into the appropriate sets
1213 bool GVN::processInstruction(Instruction *I,
1214 DenseMap<Value*, LoadInst*> &lastSeenLoad,
1215 SmallVectorImpl<Instruction*> &toErase) {
1216 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1217 bool changed = processLoad(L, lastSeenLoad, toErase);
1220 unsigned num = VN.lookup_or_add(L);
1221 localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
1227 uint32_t nextNum = VN.getNextUnusedValueNumber();
1228 unsigned num = VN.lookup_or_add(I);
1230 // Allocations are always uniquely numbered, so we can save time and memory
1231 // by fast failing them.
1232 if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1233 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1237 // Collapse PHI nodes
1238 if (PHINode* p = dyn_cast<PHINode>(I)) {
1239 Value* constVal = CollapsePhi(p);
1242 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1244 if (PI->second.count(p))
1245 PI->second.erase(p);
1247 p->replaceAllUsesWith(constVal);
1248 toErase.push_back(p);
1250 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1253 // If the number we were assigned was a brand new VN, then we don't
1254 // need to do a lookup to see if the number already exists
1255 // somewhere in the domtree: it can't!
1256 } else if (num == nextNum) {
1257 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1259 // Perform value-number based elimination
1260 } else if (Value* repl = lookupNumber(I->getParent(), num)) {
1263 I->replaceAllUsesWith(repl);
1264 toErase.push_back(I);
1267 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1273 // GVN::runOnFunction - This is the main transformation entry point for a
1276 bool GVN::runOnFunction(Function& F) {
1277 MD = &getAnalysis<MemoryDependenceAnalysis>();
1278 DT = &getAnalysis<DominatorTree>();
1279 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1283 bool changed = false;
1284 bool shouldContinue = true;
1286 // Merge unconditional branches, allowing PRE to catch more
1287 // optimization opportunities.
1288 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1289 BasicBlock* BB = FI;
1291 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1292 if (removedBlock) NumGVNBlocks++;
1294 changed |= removedBlock;
1297 while (shouldContinue) {
1298 shouldContinue = iterateOnFunction(F);
1299 changed |= shouldContinue;
1303 bool PREChanged = true;
1304 while (PREChanged) {
1305 PREChanged = performPRE(F);
1306 changed |= PREChanged;
1310 cleanupGlobalSets();
1316 bool GVN::processBlock(DomTreeNode* DTN) {
1317 BasicBlock* BB = DTN->getBlock();
1318 SmallVector<Instruction*, 8> toErase;
1319 DenseMap<Value*, LoadInst*> lastSeenLoad;
1320 bool changed_function = false;
1324 new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
1326 localAvail[BB] = new ValueNumberScope(0);
1328 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1330 changed_function |= processInstruction(BI, lastSeenLoad, toErase);
1331 if (toErase.empty()) {
1336 // If we need some instructions deleted, do it now.
1337 NumGVNInstr += toErase.size();
1339 // Avoid iterator invalidation.
1340 bool AtStart = BI == BB->begin();
1344 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1345 E = toErase.end(); I != E; ++I) {
1346 DEBUG(cerr << "GVN removed: " << **I);
1347 MD->removeInstruction(*I);
1348 (*I)->eraseFromParent();
1359 return changed_function;
1362 /// performPRE - Perform a purely local form of PRE that looks for diamond
1363 /// control flow patterns and attempts to perform simple PRE at the join point.
1364 bool GVN::performPRE(Function& F) {
1365 bool Changed = false;
1366 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1367 DenseMap<BasicBlock*, Value*> predMap;
1368 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1369 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1370 BasicBlock* CurrentBlock = *DI;
1372 // Nothing to PRE in the entry block.
1373 if (CurrentBlock == &F.getEntryBlock()) continue;
1375 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1376 BE = CurrentBlock->end(); BI != BE; ) {
1377 Instruction *CurInst = BI++;
1379 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
1380 isa<PHINode>(CurInst) || CurInst->mayReadFromMemory() ||
1381 CurInst->mayWriteToMemory())
1384 uint32_t valno = VN.lookup(CurInst);
1386 // Look for the predecessors for PRE opportunities. We're
1387 // only trying to solve the basic diamond case, where
1388 // a value is computed in the successor and one predecessor,
1389 // but not the other. We also explicitly disallow cases
1390 // where the successor is its own predecessor, because they're
1391 // more complicated to get right.
1392 unsigned numWith = 0;
1393 unsigned numWithout = 0;
1394 BasicBlock* PREPred = 0;
1397 for (pred_iterator PI = pred_begin(CurrentBlock),
1398 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1399 // We're not interested in PRE where the block is its
1400 // own predecessor, on in blocks with predecessors
1401 // that are not reachable.
1402 if (*PI == CurrentBlock) {
1405 } else if (!localAvail.count(*PI)) {
1410 DenseMap<uint32_t, Value*>::iterator predV =
1411 localAvail[*PI]->table.find(valno);
1412 if (predV == localAvail[*PI]->table.end()) {
1415 } else if (predV->second == CurInst) {
1418 predMap[*PI] = predV->second;
1423 // Don't do PRE when it might increase code size, i.e. when
1424 // we would need to insert instructions in more than one pred.
1425 if (numWithout != 1 || numWith == 0)
1428 // We can't do PRE safely on a critical edge, so instead we schedule
1429 // the edge to be split and perform the PRE the next time we iterate
1431 unsigned succNum = 0;
1432 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1434 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1439 if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
1440 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
1445 // Instantiate the expression the in predecessor that lacked it.
1446 // Because we are going top-down through the block, all value numbers
1447 // will be available in the predecessor by the time we need them. Any
1448 // that weren't original present will have been instantiated earlier
1450 Instruction* PREInstr = CurInst->clone();
1451 bool success = true;
1452 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1453 Value *Op = PREInstr->getOperand(i);
1454 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1457 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
1458 PREInstr->setOperand(i, V);
1465 // Fail out if we encounter an operand that is not available in
1466 // the PRE predecessor. This is typically because of loads which
1467 // are not value numbered precisely.
1473 PREInstr->insertBefore(PREPred->getTerminator());
1474 PREInstr->setName(CurInst->getName() + ".pre");
1475 predMap[PREPred] = PREInstr;
1476 VN.add(PREInstr, valno);
1479 // Update the availability map to include the new instruction.
1480 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
1482 // Create a PHI to make the value available in this block.
1483 PHINode* Phi = PHINode::Create(CurInst->getType(),
1484 CurInst->getName() + ".pre-phi",
1485 CurrentBlock->begin());
1486 for (pred_iterator PI = pred_begin(CurrentBlock),
1487 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1488 Phi->addIncoming(predMap[*PI], *PI);
1491 localAvail[CurrentBlock]->table[valno] = Phi;
1493 CurInst->replaceAllUsesWith(Phi);
1496 DEBUG(cerr << "GVN PRE removed: " << *CurInst);
1497 MD->removeInstruction(CurInst);
1498 CurInst->eraseFromParent();
1503 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1504 I = toSplit.begin(), E = toSplit.end(); I != E; ++I) {
1505 SplitCriticalEdge(I->first, I->second, this);
1506 BasicBlock* NewBlock = I->first->getSuccessor(I->second);
1507 localAvail[NewBlock] =
1508 new ValueNumberScope(localAvail[I->first->getParent()]);
1514 // iterateOnFunction - Executes one iteration of GVN
1515 bool GVN::iterateOnFunction(Function &F) {
1516 cleanupGlobalSets();
1518 // Top-down walk of the dominator tree
1519 bool changed = false;
1520 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1521 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
1522 changed |= processBlock(*DI);
1527 void GVN::cleanupGlobalSets() {
1531 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1532 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)