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/IntrinsicInst.h"
25 #include "llvm/Value.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/PostOrderIterator.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
35 #include "llvm/Support/CFG.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 STATISTIC(NumGVNInstr, "Number of instructions deleted");
44 STATISTIC(NumGVNLoad, "Number of loads deleted");
45 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
46 STATISTIC(NumGVNBlocks, "Number of blocks merged");
47 STATISTIC(NumPRELoad, "Number of loads PRE'd");
49 static cl::opt<bool> EnablePRE("enable-pre",
50 cl::init(true), cl::Hidden);
51 cl::opt<bool> EnableLoadPRE("enable-load-pre"/*, cl::init(true)*/);
53 //===----------------------------------------------------------------------===//
55 //===----------------------------------------------------------------------===//
57 /// This class holds the mapping between values and value numbers. It is used
58 /// as an efficient mechanism to determine the expression-wise equivalence of
61 struct VISIBILITY_HIDDEN Expression {
62 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
63 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
64 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
65 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
66 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
67 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
68 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
69 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
70 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
71 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
74 ExpressionOpcode opcode;
79 SmallVector<uint32_t, 4> varargs;
83 Expression(ExpressionOpcode o) : opcode(o) { }
85 bool operator==(const Expression &other) const {
86 if (opcode != other.opcode)
88 else if (opcode == EMPTY || opcode == TOMBSTONE)
90 else if (type != other.type)
92 else if (function != other.function)
94 else if (firstVN != other.firstVN)
96 else if (secondVN != other.secondVN)
98 else if (thirdVN != other.thirdVN)
101 if (varargs.size() != other.varargs.size())
104 for (size_t i = 0; i < varargs.size(); ++i)
105 if (varargs[i] != other.varargs[i])
112 bool operator!=(const Expression &other) const {
113 return !(*this == other);
117 class VISIBILITY_HIDDEN ValueTable {
119 DenseMap<Value*, uint32_t> valueNumbering;
120 DenseMap<Expression, uint32_t> expressionNumbering;
122 MemoryDependenceAnalysis* MD;
124 uint32_t true_vn, false_vn;
126 uint32_t nextValueNumber;
128 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
129 Expression::ExpressionOpcode getOpcode(CmpInst* C);
130 Expression::ExpressionOpcode getOpcode(CastInst* C);
131 Expression create_expression(BinaryOperator* BO);
132 Expression create_expression(CmpInst* C);
133 Expression create_expression(ShuffleVectorInst* V);
134 Expression create_expression(ExtractElementInst* C);
135 Expression create_expression(InsertElementInst* V);
136 Expression create_expression(SelectInst* V);
137 Expression create_expression(CastInst* C);
138 Expression create_expression(GetElementPtrInst* G);
139 Expression create_expression(CallInst* C);
140 Expression create_expression(Constant* C);
142 ValueTable() : nextValueNumber(1) {
143 true_vn = lookup_or_add(ConstantInt::getTrue());
144 false_vn = lookup_or_add(ConstantInt::getFalse());
147 uint32_t getTrueVN() { return true_vn; }
148 uint32_t getFalseVN() { return false_vn; }
150 uint32_t lookup_or_add(Value* V);
151 uint32_t lookup(Value* V) const;
152 void add(Value* V, uint32_t num);
154 void erase(Value* v);
156 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
157 AliasAnalysis *getAliasAnalysis() const { return AA; }
158 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
159 void setDomTree(DominatorTree* D) { DT = D; }
160 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
161 void verifyRemoved(const Value *) const;
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) || isa<VICmpInst>(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) || isa<VFCmpInst>(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 /// add - Insert a value into the table with a specified value number.
415 void ValueTable::add(Value* V, uint32_t num) {
416 valueNumbering.insert(std::make_pair(V, num));
419 /// lookup_or_add - Returns the value number for the specified value, assigning
420 /// it a new number if it did not have one before.
421 uint32_t ValueTable::lookup_or_add(Value* V) {
422 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
423 if (VI != valueNumbering.end())
426 if (CallInst* C = dyn_cast<CallInst>(V)) {
427 if (AA->doesNotAccessMemory(C)) {
428 Expression e = create_expression(C);
430 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
431 if (EI != expressionNumbering.end()) {
432 valueNumbering.insert(std::make_pair(V, EI->second));
435 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
436 valueNumbering.insert(std::make_pair(V, nextValueNumber));
438 return nextValueNumber++;
440 } else if (AA->onlyReadsMemory(C)) {
441 Expression e = create_expression(C);
443 if (expressionNumbering.find(e) == expressionNumbering.end()) {
444 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
445 valueNumbering.insert(std::make_pair(V, nextValueNumber));
446 return nextValueNumber++;
449 MemDepResult local_dep = MD->getDependency(C);
451 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
452 valueNumbering.insert(std::make_pair(V, nextValueNumber));
453 return nextValueNumber++;
456 if (local_dep.isDef()) {
457 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
459 if (local_cdep->getNumOperands() != C->getNumOperands()) {
460 valueNumbering.insert(std::make_pair(V, nextValueNumber));
461 return nextValueNumber++;
464 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
465 uint32_t c_vn = lookup_or_add(C->getOperand(i));
466 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
468 valueNumbering.insert(std::make_pair(V, nextValueNumber));
469 return nextValueNumber++;
473 uint32_t v = lookup_or_add(local_cdep);
474 valueNumbering.insert(std::make_pair(V, v));
479 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
480 MD->getNonLocalCallDependency(CallSite(C));
481 // FIXME: call/call dependencies for readonly calls should return def, not
482 // clobber! Move the checking logic to MemDep!
485 // Check to see if we have a single dominating call instruction that is
487 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
488 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
489 // Ignore non-local dependencies.
490 if (I->second.isNonLocal())
493 // We don't handle non-depedencies. If we already have a call, reject
494 // instruction dependencies.
495 if (I->second.isClobber() || cdep != 0) {
500 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
501 // FIXME: All duplicated with non-local case.
502 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
503 cdep = NonLocalDepCall;
512 valueNumbering.insert(std::make_pair(V, nextValueNumber));
513 return nextValueNumber++;
516 if (cdep->getNumOperands() != C->getNumOperands()) {
517 valueNumbering.insert(std::make_pair(V, nextValueNumber));
518 return nextValueNumber++;
520 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
521 uint32_t c_vn = lookup_or_add(C->getOperand(i));
522 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
524 valueNumbering.insert(std::make_pair(V, nextValueNumber));
525 return nextValueNumber++;
529 uint32_t v = lookup_or_add(cdep);
530 valueNumbering.insert(std::make_pair(V, v));
534 valueNumbering.insert(std::make_pair(V, nextValueNumber));
535 return nextValueNumber++;
537 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
538 Expression e = create_expression(BO);
540 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
541 if (EI != expressionNumbering.end()) {
542 valueNumbering.insert(std::make_pair(V, EI->second));
545 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
546 valueNumbering.insert(std::make_pair(V, nextValueNumber));
548 return nextValueNumber++;
550 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
551 Expression e = create_expression(C);
553 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
554 if (EI != expressionNumbering.end()) {
555 valueNumbering.insert(std::make_pair(V, EI->second));
558 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
559 valueNumbering.insert(std::make_pair(V, nextValueNumber));
561 return nextValueNumber++;
563 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
564 Expression e = create_expression(U);
566 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
567 if (EI != expressionNumbering.end()) {
568 valueNumbering.insert(std::make_pair(V, EI->second));
571 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
572 valueNumbering.insert(std::make_pair(V, nextValueNumber));
574 return nextValueNumber++;
576 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
577 Expression e = create_expression(U);
579 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
580 if (EI != expressionNumbering.end()) {
581 valueNumbering.insert(std::make_pair(V, EI->second));
584 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
585 valueNumbering.insert(std::make_pair(V, nextValueNumber));
587 return nextValueNumber++;
589 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
590 Expression e = create_expression(U);
592 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
593 if (EI != expressionNumbering.end()) {
594 valueNumbering.insert(std::make_pair(V, EI->second));
597 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
598 valueNumbering.insert(std::make_pair(V, nextValueNumber));
600 return nextValueNumber++;
602 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
603 Expression e = create_expression(U);
605 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
606 if (EI != expressionNumbering.end()) {
607 valueNumbering.insert(std::make_pair(V, EI->second));
610 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
611 valueNumbering.insert(std::make_pair(V, nextValueNumber));
613 return nextValueNumber++;
615 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
616 Expression e = create_expression(U);
618 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
619 if (EI != expressionNumbering.end()) {
620 valueNumbering.insert(std::make_pair(V, EI->second));
623 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
624 valueNumbering.insert(std::make_pair(V, nextValueNumber));
626 return nextValueNumber++;
628 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
629 Expression e = create_expression(U);
631 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
632 if (EI != expressionNumbering.end()) {
633 valueNumbering.insert(std::make_pair(V, EI->second));
636 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
637 valueNumbering.insert(std::make_pair(V, nextValueNumber));
639 return nextValueNumber++;
642 valueNumbering.insert(std::make_pair(V, nextValueNumber));
643 return nextValueNumber++;
647 /// lookup - Returns the value number of the specified value. Fails if
648 /// the value has not yet been numbered.
649 uint32_t ValueTable::lookup(Value* V) const {
650 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
651 assert(VI != valueNumbering.end() && "Value not numbered?");
655 /// clear - Remove all entries from the ValueTable
656 void ValueTable::clear() {
657 valueNumbering.clear();
658 expressionNumbering.clear();
662 /// erase - Remove a value from the value numbering
663 void ValueTable::erase(Value* V) {
664 valueNumbering.erase(V);
667 /// verifyRemoved - Verify that the value is removed from all internal data
669 void ValueTable::verifyRemoved(const Value *V) const {
670 for (DenseMap<Value*, uint32_t>::iterator
671 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
672 assert(I->first != V && "Inst still occurs in value numbering map!");
676 //===----------------------------------------------------------------------===//
678 //===----------------------------------------------------------------------===//
681 struct VISIBILITY_HIDDEN ValueNumberScope {
682 ValueNumberScope* parent;
683 DenseMap<uint32_t, Value*> table;
685 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
691 class VISIBILITY_HIDDEN GVN : public FunctionPass {
692 bool runOnFunction(Function &F);
694 static char ID; // Pass identification, replacement for typeid
695 GVN() : FunctionPass(&ID) { }
698 MemoryDependenceAnalysis *MD;
702 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
704 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
708 // This transformation requires dominator postdominator info
709 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
710 AU.addRequired<DominatorTree>();
711 AU.addRequired<MemoryDependenceAnalysis>();
712 AU.addRequired<AliasAnalysis>();
714 AU.addPreserved<DominatorTree>();
715 AU.addPreserved<AliasAnalysis>();
719 // FIXME: eliminate or document these better
720 bool processLoad(LoadInst* L,
721 SmallVectorImpl<Instruction*> &toErase);
722 bool processInstruction(Instruction* I,
723 SmallVectorImpl<Instruction*> &toErase);
724 bool processNonLocalLoad(LoadInst* L,
725 SmallVectorImpl<Instruction*> &toErase);
726 bool processBlock(BasicBlock* BB);
727 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig,
728 DenseMap<BasicBlock*, Value*> &Phis,
729 bool top_level = false);
730 void dump(DenseMap<uint32_t, Value*>& d);
731 bool iterateOnFunction(Function &F);
732 Value* CollapsePhi(PHINode* p);
733 bool isSafeReplacement(PHINode* p, Instruction* inst);
734 bool performPRE(Function& F);
735 Value* lookupNumber(BasicBlock* BB, uint32_t num);
736 bool mergeBlockIntoPredecessor(BasicBlock* BB);
737 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno);
738 void cleanupGlobalSets();
739 void verifyRemoved(const Instruction *I) const;
745 // createGVNPass - The public interface to this file...
746 FunctionPass *llvm::createGVNPass() { return new GVN(); }
748 static RegisterPass<GVN> X("gvn",
749 "Global Value Numbering");
751 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
753 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
754 E = d.end(); I != E; ++I) {
755 printf("%d\n", I->first);
761 Value* GVN::CollapsePhi(PHINode* p) {
762 Value* constVal = p->hasConstantValue();
763 if (!constVal) return 0;
765 Instruction* inst = dyn_cast<Instruction>(constVal);
769 if (DT->dominates(inst, p))
770 if (isSafeReplacement(p, inst))
775 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
776 if (!isa<PHINode>(inst))
779 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
781 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
782 if (use_phi->getParent() == inst->getParent())
788 /// GetValueForBlock - Get the value to use within the specified basic block.
789 /// available values are in Phis.
790 Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig,
791 DenseMap<BasicBlock*, Value*> &Phis,
794 // If we have already computed this value, return the previously computed val.
795 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
796 if (V != Phis.end() && !top_level) return V->second;
798 // If the block is unreachable, just return undef, since this path
799 // can't actually occur at runtime.
800 if (!DT->isReachableFromEntry(BB))
801 return Phis[BB] = UndefValue::get(orig->getType());
803 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
804 Value *ret = GetValueForBlock(Pred, orig, Phis);
809 // Get the number of predecessors of this block so we can reserve space later.
810 // If there is already a PHI in it, use the #preds from it, otherwise count.
811 // Getting it from the PHI is constant time.
813 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin()))
814 NumPreds = ExistingPN->getNumIncomingValues();
816 NumPreds = std::distance(pred_begin(BB), pred_end(BB));
818 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
819 // now, then get values to fill in the incoming values for the PHI.
820 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
822 PN->reserveOperandSpace(NumPreds);
824 Phis.insert(std::make_pair(BB, PN));
826 // Fill in the incoming values for the block.
827 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
828 Value* val = GetValueForBlock(*PI, orig, Phis);
829 PN->addIncoming(val, *PI);
832 VN.getAliasAnalysis()->copyValue(orig, PN);
834 // Attempt to collapse PHI nodes that are trivially redundant
835 Value* v = CollapsePhi(PN);
837 // Cache our phi construction results
838 if (LoadInst* L = dyn_cast<LoadInst>(orig))
839 phiMap[L->getPointerOperand()].insert(PN);
841 phiMap[orig].insert(PN);
846 PN->replaceAllUsesWith(v);
847 if (isa<PointerType>(v->getType()))
848 MD->invalidateCachedPointerInfo(v);
850 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
851 E = Phis.end(); I != E; ++I)
855 DEBUG(cerr << "GVN removed: " << *PN);
856 MD->removeInstruction(PN);
857 PN->eraseFromParent();
858 DEBUG(verifyRemoved(PN));
864 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
865 /// we're analyzing is fully available in the specified block. As we go, keep
866 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
867 /// map is actually a tri-state map with the following values:
868 /// 0) we know the block *is not* fully available.
869 /// 1) we know the block *is* fully available.
870 /// 2) we do not know whether the block is fully available or not, but we are
871 /// currently speculating that it will be.
872 /// 3) we are speculating for this block and have used that to speculate for
874 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
875 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
876 // Optimistically assume that the block is fully available and check to see
877 // if we already know about this block in one lookup.
878 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
879 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
881 // If the entry already existed for this block, return the precomputed value.
883 // If this is a speculative "available" value, mark it as being used for
884 // speculation of other blocks.
885 if (IV.first->second == 2)
886 IV.first->second = 3;
887 return IV.first->second != 0;
890 // Otherwise, see if it is fully available in all predecessors.
891 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
893 // If this block has no predecessors, it isn't live-in here.
895 goto SpeculationFailure;
897 for (; PI != PE; ++PI)
898 // If the value isn't fully available in one of our predecessors, then it
899 // isn't fully available in this block either. Undo our previous
900 // optimistic assumption and bail out.
901 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
902 goto SpeculationFailure;
906 // SpeculationFailure - If we get here, we found out that this is not, after
907 // all, a fully-available block. We have a problem if we speculated on this and
908 // used the speculation to mark other blocks as available.
910 char &BBVal = FullyAvailableBlocks[BB];
912 // If we didn't speculate on this, just return with it set to false.
918 // If we did speculate on this value, we could have blocks set to 1 that are
919 // incorrect. Walk the (transitive) successors of this block and mark them as
921 SmallVector<BasicBlock*, 32> BBWorklist;
922 BBWorklist.push_back(BB);
924 while (!BBWorklist.empty()) {
925 BasicBlock *Entry = BBWorklist.pop_back_val();
926 // Note that this sets blocks to 0 (unavailable) if they happen to not
927 // already be in FullyAvailableBlocks. This is safe.
928 char &EntryVal = FullyAvailableBlocks[Entry];
929 if (EntryVal == 0) continue; // Already unavailable.
931 // Mark as unavailable.
934 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
935 BBWorklist.push_back(*I);
941 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
942 /// non-local by performing PHI construction.
943 bool GVN::processNonLocalLoad(LoadInst *LI,
944 SmallVectorImpl<Instruction*> &toErase) {
945 // Find the non-local dependencies of the load.
946 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
947 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
949 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI);
951 // If we had to process more than one hundred blocks to find the
952 // dependencies, this load isn't worth worrying about. Optimizing
953 // it will be too expensive.
954 if (Deps.size() > 100)
957 // If we had a phi translation failure, we'll have a single entry which is a
958 // clobber in the current block. Reject this early.
959 if (Deps.size() == 1 && Deps[0].second.isClobber())
962 // Filter out useless results (non-locals, etc). Keep track of the blocks
963 // where we have a value available in repl, also keep track of whether we see
964 // dependencies that produce an unknown value for the load (such as a call
965 // that could potentially clobber the load).
966 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
967 SmallVector<BasicBlock*, 16> UnavailableBlocks;
969 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
970 BasicBlock *DepBB = Deps[i].first;
971 MemDepResult DepInfo = Deps[i].second;
973 if (DepInfo.isClobber()) {
974 UnavailableBlocks.push_back(DepBB);
978 Instruction *DepInst = DepInfo.getInst();
980 // Loading the allocation -> undef.
981 if (isa<AllocationInst>(DepInst)) {
982 ValuesPerBlock.push_back(std::make_pair(DepBB,
983 UndefValue::get(LI->getType())));
987 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
988 // Reject loads and stores that are to the same address but are of
990 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
991 // of bitfield access, it would be interesting to optimize for it at some
993 if (S->getOperand(0)->getType() != LI->getType()) {
994 UnavailableBlocks.push_back(DepBB);
998 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
1000 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
1001 if (LD->getType() != LI->getType()) {
1002 UnavailableBlocks.push_back(DepBB);
1005 ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
1007 UnavailableBlocks.push_back(DepBB);
1012 // If we have no predecessors that produce a known value for this load, exit
1014 if (ValuesPerBlock.empty()) return false;
1016 // If all of the instructions we depend on produce a known value for this
1017 // load, then it is fully redundant and we can use PHI insertion to compute
1018 // its value. Insert PHIs and remove the fully redundant value now.
1019 if (UnavailableBlocks.empty()) {
1020 // Use cached PHI construction information from previous runs
1021 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
1022 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated?
1023 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
1025 if ((*I)->getParent() == LI->getParent()) {
1026 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI);
1027 LI->replaceAllUsesWith(*I);
1028 if (isa<PointerType>((*I)->getType()))
1029 MD->invalidateCachedPointerInfo(*I);
1030 toErase.push_back(LI);
1035 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
1038 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI);
1040 DenseMap<BasicBlock*, Value*> BlockReplValues;
1041 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1042 // Perform PHI construction.
1043 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1044 LI->replaceAllUsesWith(v);
1046 if (isa<PHINode>(v))
1048 if (isa<PointerType>(v->getType()))
1049 MD->invalidateCachedPointerInfo(v);
1050 toErase.push_back(LI);
1055 if (!EnablePRE || !EnableLoadPRE)
1058 // Okay, we have *some* definitions of the value. This means that the value
1059 // is available in some of our (transitive) predecessors. Lets think about
1060 // doing PRE of this load. This will involve inserting a new load into the
1061 // predecessor when it's not available. We could do this in general, but
1062 // prefer to not increase code size. As such, we only do this when we know
1063 // that we only have to insert *one* load (which means we're basically moving
1064 // the load, not inserting a new one).
1066 // Everything we do here is based on local predecessors of LI's block. If it
1067 // only has one predecessor, bail now.
1068 BasicBlock *LoadBB = LI->getParent();
1069 if (LoadBB->getSinglePredecessor())
1072 // If we have a repl set with LI itself in it, this means we have a loop where
1073 // at least one of the values is LI. Since this means that we won't be able
1074 // to eliminate LI even if we insert uses in the other predecessors, we will
1075 // end up increasing code size. Reject this by scanning for LI.
1076 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1077 if (ValuesPerBlock[i].second == LI)
1080 // Okay, we have some hope :). Check to see if the loaded value is fully
1081 // available in all but one predecessor.
1082 // FIXME: If we could restructure the CFG, we could make a common pred with
1083 // all the preds that don't have an available LI and insert a new load into
1085 BasicBlock *UnavailablePred = 0;
1087 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1088 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1089 FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
1090 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1091 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1093 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1095 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1098 // If this load is not available in multiple predecessors, reject it.
1099 if (UnavailablePred && UnavailablePred != *PI)
1101 UnavailablePred = *PI;
1104 assert(UnavailablePred != 0 &&
1105 "Fully available value should be eliminated above!");
1107 // If the loaded pointer is PHI node defined in this block, do PHI translation
1108 // to get its value in the predecessor.
1109 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
1111 // Make sure the value is live in the predecessor. If it was defined by a
1112 // non-PHI instruction in this block, we don't know how to recompute it above.
1113 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
1114 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
1115 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
1116 << *LPInst << *LI << "\n");
1120 // We don't currently handle critical edges :(
1121 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1122 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1123 << UnavailablePred->getName() << "': " << *LI);
1127 // Okay, we can eliminate this load by inserting a reload in the predecessor
1128 // and using PHI construction to get the value in the other predecessors, do
1130 DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI);
1132 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1134 UnavailablePred->getTerminator());
1136 DenseMap<BasicBlock*, Value*> BlockReplValues;
1137 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1138 BlockReplValues[UnavailablePred] = NewLoad;
1140 // Perform PHI construction.
1141 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1142 LI->replaceAllUsesWith(v);
1143 if (isa<PHINode>(v))
1145 if (isa<PointerType>(v->getType()))
1146 MD->invalidateCachedPointerInfo(v);
1147 toErase.push_back(LI);
1152 /// processLoad - Attempt to eliminate a load, first by eliminating it
1153 /// locally, and then attempting non-local elimination if that fails.
1154 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1155 if (L->isVolatile())
1158 Value* pointer = L->getPointerOperand();
1160 // ... to a pointer that has been loaded from before...
1161 MemDepResult dep = MD->getDependency(L);
1163 // If the value isn't available, don't do anything!
1164 if (dep.isClobber())
1167 // If it is defined in another block, try harder.
1168 if (dep.isNonLocal())
1169 return processNonLocalLoad(L, toErase);
1171 Instruction *DepInst = dep.getInst();
1172 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1173 // Only forward substitute stores to loads of the same type.
1174 // FIXME: Could do better!
1175 if (DepSI->getPointerOperand()->getType() != pointer->getType())
1179 L->replaceAllUsesWith(DepSI->getOperand(0));
1180 if (isa<PointerType>(DepSI->getOperand(0)->getType()))
1181 MD->invalidateCachedPointerInfo(DepSI->getOperand(0));
1182 toErase.push_back(L);
1187 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1188 // Only forward substitute stores to loads of the same type.
1189 // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian.
1190 if (DepLI->getType() != L->getType())
1194 L->replaceAllUsesWith(DepLI);
1195 if (isa<PointerType>(DepLI->getType()))
1196 MD->invalidateCachedPointerInfo(DepLI);
1197 toErase.push_back(L);
1202 // If this load really doesn't depend on anything, then we must be loading an
1203 // undef value. This can happen when loading for a fresh allocation with no
1204 // intervening stores, for example.
1205 if (isa<AllocationInst>(DepInst)) {
1206 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1207 toErase.push_back(L);
1215 Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
1216 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1217 if (I == localAvail.end())
1220 ValueNumberScope* locals = I->second;
1223 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
1224 if (I != locals->table.end())
1227 locals = locals->parent;
1233 /// AttemptRedundancyElimination - If the "fast path" of redundancy elimination
1234 /// by inheritance from the dominator fails, see if we can perform phi
1235 /// construction to eliminate the redundancy.
1236 Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) {
1237 BasicBlock* BaseBlock = orig->getParent();
1239 SmallPtrSet<BasicBlock*, 4> Visited;
1240 SmallVector<BasicBlock*, 8> Stack;
1241 Stack.push_back(BaseBlock);
1243 DenseMap<BasicBlock*, Value*> Results;
1245 // Walk backwards through our predecessors, looking for instances of the
1246 // value number we're looking for. Instances are recorded in the Results
1247 // map, which is then used to perform phi construction.
1248 while (!Stack.empty()) {
1249 BasicBlock* Current = Stack.back();
1252 // If we've walked all the way to a proper dominator, then give up. Cases
1253 // where the instance is in the dominator will have been caught by the fast
1254 // path, and any cases that require phi construction further than this are
1255 // probably not worth it anyways. Note that this is a SIGNIFICANT compile
1256 // time improvement.
1257 if (DT->properlyDominates(Current, orig->getParent())) return 0;
1259 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
1260 localAvail.find(Current);
1261 if (LA == localAvail.end()) return 0;
1262 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);
1264 if (V != LA->second->table.end()) {
1265 // Found an instance, record it.
1266 Results.insert(std::make_pair(Current, V->second));
1270 // If we reach the beginning of the function, then give up.
1271 if (pred_begin(Current) == pred_end(Current))
1274 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current);
1276 if (Visited.insert(*PI))
1277 Stack.push_back(*PI);
1280 // If we didn't find instances, give up. Otherwise, perform phi construction.
1281 if (Results.size() == 0)
1284 return GetValueForBlock(BaseBlock, orig, Results, true);
1287 /// processInstruction - When calculating availability, handle an instruction
1288 /// by inserting it into the appropriate sets
1289 bool GVN::processInstruction(Instruction *I,
1290 SmallVectorImpl<Instruction*> &toErase) {
1291 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1292 bool changed = processLoad(L, toErase);
1295 unsigned num = VN.lookup_or_add(L);
1296 localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
1302 uint32_t nextNum = VN.getNextUnusedValueNumber();
1303 unsigned num = VN.lookup_or_add(I);
1305 if (BranchInst* BI = dyn_cast<BranchInst>(I)) {
1306 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1308 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1311 Value* branchCond = BI->getCondition();
1312 uint32_t condVN = VN.lookup_or_add(branchCond);
1314 BasicBlock* trueSucc = BI->getSuccessor(0);
1315 BasicBlock* falseSucc = BI->getSuccessor(1);
1317 localAvail[trueSucc]->table.insert(std::make_pair(condVN,
1318 ConstantInt::getTrue()));
1319 localAvail[falseSucc]->table.insert(std::make_pair(condVN,
1320 ConstantInt::getFalse()));
1323 // Allocations are always uniquely numbered, so we can save time and memory
1324 // by fast failing them.
1325 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1326 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1330 // Collapse PHI nodes
1331 if (PHINode* p = dyn_cast<PHINode>(I)) {
1332 Value* constVal = CollapsePhi(p);
1335 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1337 PI->second.erase(p);
1339 p->replaceAllUsesWith(constVal);
1340 if (isa<PointerType>(constVal->getType()))
1341 MD->invalidateCachedPointerInfo(constVal);
1344 toErase.push_back(p);
1346 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1349 // If the number we were assigned was a brand new VN, then we don't
1350 // need to do a lookup to see if the number already exists
1351 // somewhere in the domtree: it can't!
1352 } else if (num == nextNum) {
1353 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1355 // Perform fast-path value-number based elimination of values inherited from
1357 } else if (Value* repl = lookupNumber(I->getParent(), num)) {
1360 I->replaceAllUsesWith(repl);
1361 if (isa<PointerType>(repl->getType()))
1362 MD->invalidateCachedPointerInfo(repl);
1363 toErase.push_back(I);
1367 // Perform slow-pathvalue-number based elimination with phi construction.
1368 } else if (Value* repl = AttemptRedundancyElimination(I, num)) {
1371 I->replaceAllUsesWith(repl);
1372 if (isa<PointerType>(repl->getType()))
1373 MD->invalidateCachedPointerInfo(repl);
1374 toErase.push_back(I);
1378 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1384 /// runOnFunction - This is the main transformation entry point for a function.
1385 bool GVN::runOnFunction(Function& F) {
1386 MD = &getAnalysis<MemoryDependenceAnalysis>();
1387 DT = &getAnalysis<DominatorTree>();
1388 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1392 bool changed = false;
1393 bool shouldContinue = true;
1395 // Merge unconditional branches, allowing PRE to catch more
1396 // optimization opportunities.
1397 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1398 BasicBlock* BB = FI;
1400 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1401 if (removedBlock) NumGVNBlocks++;
1403 changed |= removedBlock;
1406 unsigned Iteration = 0;
1408 while (shouldContinue) {
1409 DEBUG(cerr << "GVN iteration: " << Iteration << "\n");
1410 shouldContinue = iterateOnFunction(F);
1411 changed |= shouldContinue;
1416 bool PREChanged = true;
1417 while (PREChanged) {
1418 PREChanged = performPRE(F);
1419 changed |= PREChanged;
1422 // FIXME: Should perform GVN again after PRE does something. PRE can move
1423 // computations into blocks where they become fully redundant. Note that
1424 // we can't do this until PRE's critical edge splitting updates memdep.
1425 // Actually, when this happens, we should just fully integrate PRE into GVN.
1427 cleanupGlobalSets();
1433 bool GVN::processBlock(BasicBlock* BB) {
1434 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1435 // incrementing BI before processing an instruction).
1436 SmallVector<Instruction*, 8> toErase;
1437 bool changed_function = false;
1439 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1441 changed_function |= processInstruction(BI, toErase);
1442 if (toErase.empty()) {
1447 // If we need some instructions deleted, do it now.
1448 NumGVNInstr += toErase.size();
1450 // Avoid iterator invalidation.
1451 bool AtStart = BI == BB->begin();
1455 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1456 E = toErase.end(); I != E; ++I) {
1457 DEBUG(cerr << "GVN removed: " << **I);
1458 MD->removeInstruction(*I);
1459 (*I)->eraseFromParent();
1460 DEBUG(verifyRemoved(*I));
1470 return changed_function;
1473 /// performPRE - Perform a purely local form of PRE that looks for diamond
1474 /// control flow patterns and attempts to perform simple PRE at the join point.
1475 bool GVN::performPRE(Function& F) {
1476 bool Changed = false;
1477 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1478 DenseMap<BasicBlock*, Value*> predMap;
1479 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1480 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1481 BasicBlock* CurrentBlock = *DI;
1483 // Nothing to PRE in the entry block.
1484 if (CurrentBlock == &F.getEntryBlock()) continue;
1486 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1487 BE = CurrentBlock->end(); BI != BE; ) {
1488 Instruction *CurInst = BI++;
1490 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
1491 isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) ||
1492 CurInst->mayReadFromMemory() || CurInst->mayWriteToMemory() ||
1493 isa<DbgInfoIntrinsic>(CurInst))
1496 uint32_t valno = VN.lookup(CurInst);
1498 // Look for the predecessors for PRE opportunities. We're
1499 // only trying to solve the basic diamond case, where
1500 // a value is computed in the successor and one predecessor,
1501 // but not the other. We also explicitly disallow cases
1502 // where the successor is its own predecessor, because they're
1503 // more complicated to get right.
1504 unsigned numWith = 0;
1505 unsigned numWithout = 0;
1506 BasicBlock* PREPred = 0;
1509 for (pred_iterator PI = pred_begin(CurrentBlock),
1510 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1511 // We're not interested in PRE where the block is its
1512 // own predecessor, on in blocks with predecessors
1513 // that are not reachable.
1514 if (*PI == CurrentBlock) {
1517 } else if (!localAvail.count(*PI)) {
1522 DenseMap<uint32_t, Value*>::iterator predV =
1523 localAvail[*PI]->table.find(valno);
1524 if (predV == localAvail[*PI]->table.end()) {
1527 } else if (predV->second == CurInst) {
1530 predMap[*PI] = predV->second;
1535 // Don't do PRE when it might increase code size, i.e. when
1536 // we would need to insert instructions in more than one pred.
1537 if (numWithout != 1 || numWith == 0)
1540 // We can't do PRE safely on a critical edge, so instead we schedule
1541 // the edge to be split and perform the PRE the next time we iterate
1543 unsigned succNum = 0;
1544 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1546 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1551 if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
1552 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
1556 // Instantiate the expression the in predecessor that lacked it.
1557 // Because we are going top-down through the block, all value numbers
1558 // will be available in the predecessor by the time we need them. Any
1559 // that weren't original present will have been instantiated earlier
1561 Instruction* PREInstr = CurInst->clone();
1562 bool success = true;
1563 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1564 Value *Op = PREInstr->getOperand(i);
1565 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1568 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
1569 PREInstr->setOperand(i, V);
1576 // Fail out if we encounter an operand that is not available in
1577 // the PRE predecessor. This is typically because of loads which
1578 // are not value numbered precisely.
1581 DEBUG(verifyRemoved(PREInstr));
1585 PREInstr->insertBefore(PREPred->getTerminator());
1586 PREInstr->setName(CurInst->getName() + ".pre");
1587 predMap[PREPred] = PREInstr;
1588 VN.add(PREInstr, valno);
1591 // Update the availability map to include the new instruction.
1592 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
1594 // Create a PHI to make the value available in this block.
1595 PHINode* Phi = PHINode::Create(CurInst->getType(),
1596 CurInst->getName() + ".pre-phi",
1597 CurrentBlock->begin());
1598 for (pred_iterator PI = pred_begin(CurrentBlock),
1599 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1600 Phi->addIncoming(predMap[*PI], *PI);
1603 localAvail[CurrentBlock]->table[valno] = Phi;
1605 CurInst->replaceAllUsesWith(Phi);
1606 if (isa<PointerType>(Phi->getType()))
1607 MD->invalidateCachedPointerInfo(Phi);
1610 DEBUG(cerr << "GVN PRE removed: " << *CurInst);
1611 MD->removeInstruction(CurInst);
1612 CurInst->eraseFromParent();
1613 DEBUG(verifyRemoved(CurInst));
1618 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1619 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1620 SplitCriticalEdge(I->first, I->second, this);
1622 return Changed || toSplit.size();
1625 /// iterateOnFunction - Executes one iteration of GVN
1626 bool GVN::iterateOnFunction(Function &F) {
1627 cleanupGlobalSets();
1629 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1630 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
1632 localAvail[DI->getBlock()] =
1633 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
1635 localAvail[DI->getBlock()] = new ValueNumberScope(0);
1638 // Top-down walk of the dominator tree
1639 bool changed = false;
1641 // Needed for value numbering with phi construction to work.
1642 ReversePostOrderTraversal<Function*> RPOT(&F);
1643 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
1644 RE = RPOT.end(); RI != RE; ++RI)
1645 changed |= processBlock(*RI);
1647 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1648 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
1649 changed |= processBlock(DI->getBlock());
1655 void GVN::cleanupGlobalSets() {
1659 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1660 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
1665 /// verifyRemoved - Verify that the specified instruction does not occur in our
1666 /// internal data structures.
1667 void GVN::verifyRemoved(const Instruction *Inst) const {
1668 VN.verifyRemoved(Inst);
1670 // Walk through the PHI map to make sure the instruction isn't hiding in there
1672 for (PhiMapType::iterator
1673 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) {
1674 assert(I->first != Inst && "Inst is still a key in PHI map!");
1676 for (SmallPtrSet<Instruction*, 4>::iterator
1677 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) {
1678 assert(*II != Inst && "Inst is still a value in PHI map!");
1682 // Walk through the value number scope to make sure the instruction isn't
1683 // ferreted away in it.
1684 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1685 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
1686 const ValueNumberScope *VNS = I->second;
1689 for (DenseMap<uint32_t, Value*>::iterator
1690 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
1691 assert(II->second != Inst && "Inst still in value numbering scope!");