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, FADD, SUB, FSUB, MUL, FMUL,
63 UDIV, SDIV, FDIV, UREM, SREM,
64 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
65 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
66 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
67 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
68 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
69 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
70 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
71 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
72 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
75 ExpressionOpcode opcode;
80 SmallVector<uint32_t, 4> varargs;
84 Expression(ExpressionOpcode o) : opcode(o) { }
86 bool operator==(const Expression &other) const {
87 if (opcode != other.opcode)
89 else if (opcode == EMPTY || opcode == TOMBSTONE)
91 else if (type != other.type)
93 else if (function != other.function)
95 else if (firstVN != other.firstVN)
97 else if (secondVN != other.secondVN)
99 else if (thirdVN != other.thirdVN)
102 if (varargs.size() != other.varargs.size())
105 for (size_t i = 0; i < varargs.size(); ++i)
106 if (varargs[i] != other.varargs[i])
113 bool operator!=(const Expression &other) const {
114 return !(*this == other);
118 class VISIBILITY_HIDDEN ValueTable {
120 DenseMap<Value*, uint32_t> valueNumbering;
121 DenseMap<Expression, uint32_t> expressionNumbering;
123 MemoryDependenceAnalysis* MD;
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 uint32_t lookup_or_add(Value* V);
144 uint32_t lookup(Value* V) const;
145 void add(Value* V, uint32_t num);
147 void erase(Value* v);
149 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
150 AliasAnalysis *getAliasAnalysis() const { return AA; }
151 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
152 void setDomTree(DominatorTree* D) { DT = D; }
153 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
154 void verifyRemoved(const Value *) const;
159 template <> struct DenseMapInfo<Expression> {
160 static inline Expression getEmptyKey() {
161 return Expression(Expression::EMPTY);
164 static inline Expression getTombstoneKey() {
165 return Expression(Expression::TOMBSTONE);
168 static unsigned getHashValue(const Expression e) {
169 unsigned hash = e.opcode;
171 hash = e.firstVN + hash * 37;
172 hash = e.secondVN + hash * 37;
173 hash = e.thirdVN + hash * 37;
175 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
176 (unsigned)((uintptr_t)e.type >> 9)) +
179 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
180 E = e.varargs.end(); I != E; ++I)
181 hash = *I + hash * 37;
183 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
184 (unsigned)((uintptr_t)e.function >> 9)) +
189 static bool isEqual(const Expression &LHS, const Expression &RHS) {
192 static bool isPod() { return true; }
196 //===----------------------------------------------------------------------===//
197 // ValueTable Internal Functions
198 //===----------------------------------------------------------------------===//
199 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
200 switch(BO->getOpcode()) {
201 default: // THIS SHOULD NEVER HAPPEN
202 assert(0 && "Binary operator with unknown opcode?");
203 case Instruction::Add: return Expression::ADD;
204 case Instruction::FAdd: return Expression::FADD;
205 case Instruction::Sub: return Expression::SUB;
206 case Instruction::FSub: return Expression::FSUB;
207 case Instruction::Mul: return Expression::MUL;
208 case Instruction::FMul: return Expression::FMUL;
209 case Instruction::UDiv: return Expression::UDIV;
210 case Instruction::SDiv: return Expression::SDIV;
211 case Instruction::FDiv: return Expression::FDIV;
212 case Instruction::URem: return Expression::UREM;
213 case Instruction::SRem: return Expression::SREM;
214 case Instruction::FRem: return Expression::FREM;
215 case Instruction::Shl: return Expression::SHL;
216 case Instruction::LShr: return Expression::LSHR;
217 case Instruction::AShr: return Expression::ASHR;
218 case Instruction::And: return Expression::AND;
219 case Instruction::Or: return Expression::OR;
220 case Instruction::Xor: return Expression::XOR;
224 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
225 if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
226 switch (C->getPredicate()) {
227 default: // THIS SHOULD NEVER HAPPEN
228 assert(0 && "Comparison with unknown predicate?");
229 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
230 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
231 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
232 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
233 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
234 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
235 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
236 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
237 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
238 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
241 assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
242 switch (C->getPredicate()) {
243 default: // THIS SHOULD NEVER HAPPEN
244 assert(0 && "Comparison with unknown predicate?");
245 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
246 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
247 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
248 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
249 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
250 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
251 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
252 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
253 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
254 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
255 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
256 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
257 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
258 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
262 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
263 switch(C->getOpcode()) {
264 default: // THIS SHOULD NEVER HAPPEN
265 assert(0 && "Cast operator with unknown opcode?");
266 case Instruction::Trunc: return Expression::TRUNC;
267 case Instruction::ZExt: return Expression::ZEXT;
268 case Instruction::SExt: return Expression::SEXT;
269 case Instruction::FPToUI: return Expression::FPTOUI;
270 case Instruction::FPToSI: return Expression::FPTOSI;
271 case Instruction::UIToFP: return Expression::UITOFP;
272 case Instruction::SIToFP: return Expression::SITOFP;
273 case Instruction::FPTrunc: return Expression::FPTRUNC;
274 case Instruction::FPExt: return Expression::FPEXT;
275 case Instruction::PtrToInt: return Expression::PTRTOINT;
276 case Instruction::IntToPtr: return Expression::INTTOPTR;
277 case Instruction::BitCast: return Expression::BITCAST;
281 Expression ValueTable::create_expression(CallInst* C) {
284 e.type = C->getType();
288 e.function = C->getCalledFunction();
289 e.opcode = Expression::CALL;
291 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
293 e.varargs.push_back(lookup_or_add(*I));
298 Expression ValueTable::create_expression(BinaryOperator* BO) {
301 e.firstVN = lookup_or_add(BO->getOperand(0));
302 e.secondVN = lookup_or_add(BO->getOperand(1));
305 e.type = BO->getType();
306 e.opcode = getOpcode(BO);
311 Expression ValueTable::create_expression(CmpInst* C) {
314 e.firstVN = lookup_or_add(C->getOperand(0));
315 e.secondVN = lookup_or_add(C->getOperand(1));
318 e.type = C->getType();
319 e.opcode = getOpcode(C);
324 Expression ValueTable::create_expression(CastInst* C) {
327 e.firstVN = lookup_or_add(C->getOperand(0));
331 e.type = C->getType();
332 e.opcode = getOpcode(C);
337 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
340 e.firstVN = lookup_or_add(S->getOperand(0));
341 e.secondVN = lookup_or_add(S->getOperand(1));
342 e.thirdVN = lookup_or_add(S->getOperand(2));
344 e.type = S->getType();
345 e.opcode = Expression::SHUFFLE;
350 Expression ValueTable::create_expression(ExtractElementInst* E) {
353 e.firstVN = lookup_or_add(E->getOperand(0));
354 e.secondVN = lookup_or_add(E->getOperand(1));
357 e.type = E->getType();
358 e.opcode = Expression::EXTRACT;
363 Expression ValueTable::create_expression(InsertElementInst* I) {
366 e.firstVN = lookup_or_add(I->getOperand(0));
367 e.secondVN = lookup_or_add(I->getOperand(1));
368 e.thirdVN = lookup_or_add(I->getOperand(2));
370 e.type = I->getType();
371 e.opcode = Expression::INSERT;
376 Expression ValueTable::create_expression(SelectInst* I) {
379 e.firstVN = lookup_or_add(I->getCondition());
380 e.secondVN = lookup_or_add(I->getTrueValue());
381 e.thirdVN = lookup_or_add(I->getFalseValue());
383 e.type = I->getType();
384 e.opcode = Expression::SELECT;
389 Expression ValueTable::create_expression(GetElementPtrInst* G) {
392 e.firstVN = lookup_or_add(G->getPointerOperand());
396 e.type = G->getType();
397 e.opcode = Expression::GEP;
399 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
401 e.varargs.push_back(lookup_or_add(*I));
406 //===----------------------------------------------------------------------===//
407 // ValueTable External Functions
408 //===----------------------------------------------------------------------===//
410 /// add - Insert a value into the table with a specified value number.
411 void ValueTable::add(Value* V, uint32_t num) {
412 valueNumbering.insert(std::make_pair(V, num));
415 /// lookup_or_add - Returns the value number for the specified value, assigning
416 /// it a new number if it did not have one before.
417 uint32_t ValueTable::lookup_or_add(Value* V) {
418 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
419 if (VI != valueNumbering.end())
422 if (CallInst* C = dyn_cast<CallInst>(V)) {
423 if (AA->doesNotAccessMemory(C)) {
424 Expression e = create_expression(C);
426 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
427 if (EI != expressionNumbering.end()) {
428 valueNumbering.insert(std::make_pair(V, EI->second));
431 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
432 valueNumbering.insert(std::make_pair(V, nextValueNumber));
434 return nextValueNumber++;
436 } else if (AA->onlyReadsMemory(C)) {
437 Expression e = create_expression(C);
439 if (expressionNumbering.find(e) == expressionNumbering.end()) {
440 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
441 valueNumbering.insert(std::make_pair(V, nextValueNumber));
442 return nextValueNumber++;
445 MemDepResult local_dep = MD->getDependency(C);
447 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
448 valueNumbering.insert(std::make_pair(V, nextValueNumber));
449 return nextValueNumber++;
452 if (local_dep.isDef()) {
453 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
455 if (local_cdep->getNumOperands() != C->getNumOperands()) {
456 valueNumbering.insert(std::make_pair(V, nextValueNumber));
457 return nextValueNumber++;
460 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
461 uint32_t c_vn = lookup_or_add(C->getOperand(i));
462 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
464 valueNumbering.insert(std::make_pair(V, nextValueNumber));
465 return nextValueNumber++;
469 uint32_t v = lookup_or_add(local_cdep);
470 valueNumbering.insert(std::make_pair(V, v));
475 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
476 MD->getNonLocalCallDependency(CallSite(C));
477 // FIXME: call/call dependencies for readonly calls should return def, not
478 // clobber! Move the checking logic to MemDep!
481 // Check to see if we have a single dominating call instruction that is
483 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
484 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
485 // Ignore non-local dependencies.
486 if (I->second.isNonLocal())
489 // We don't handle non-depedencies. If we already have a call, reject
490 // instruction dependencies.
491 if (I->second.isClobber() || cdep != 0) {
496 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
497 // FIXME: All duplicated with non-local case.
498 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
499 cdep = NonLocalDepCall;
508 valueNumbering.insert(std::make_pair(V, nextValueNumber));
509 return nextValueNumber++;
512 if (cdep->getNumOperands() != C->getNumOperands()) {
513 valueNumbering.insert(std::make_pair(V, nextValueNumber));
514 return nextValueNumber++;
516 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
517 uint32_t c_vn = lookup_or_add(C->getOperand(i));
518 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
520 valueNumbering.insert(std::make_pair(V, nextValueNumber));
521 return nextValueNumber++;
525 uint32_t v = lookup_or_add(cdep);
526 valueNumbering.insert(std::make_pair(V, v));
530 valueNumbering.insert(std::make_pair(V, nextValueNumber));
531 return nextValueNumber++;
533 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
534 Expression e = create_expression(BO);
536 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
537 if (EI != expressionNumbering.end()) {
538 valueNumbering.insert(std::make_pair(V, EI->second));
541 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
542 valueNumbering.insert(std::make_pair(V, nextValueNumber));
544 return nextValueNumber++;
546 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
547 Expression e = create_expression(C);
549 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
550 if (EI != expressionNumbering.end()) {
551 valueNumbering.insert(std::make_pair(V, EI->second));
554 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
555 valueNumbering.insert(std::make_pair(V, nextValueNumber));
557 return nextValueNumber++;
559 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
560 Expression e = create_expression(U);
562 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
563 if (EI != expressionNumbering.end()) {
564 valueNumbering.insert(std::make_pair(V, EI->second));
567 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
568 valueNumbering.insert(std::make_pair(V, nextValueNumber));
570 return nextValueNumber++;
572 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
573 Expression e = create_expression(U);
575 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
576 if (EI != expressionNumbering.end()) {
577 valueNumbering.insert(std::make_pair(V, EI->second));
580 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
581 valueNumbering.insert(std::make_pair(V, nextValueNumber));
583 return nextValueNumber++;
585 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
586 Expression e = create_expression(U);
588 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
589 if (EI != expressionNumbering.end()) {
590 valueNumbering.insert(std::make_pair(V, EI->second));
593 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
594 valueNumbering.insert(std::make_pair(V, nextValueNumber));
596 return nextValueNumber++;
598 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
599 Expression e = create_expression(U);
601 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
602 if (EI != expressionNumbering.end()) {
603 valueNumbering.insert(std::make_pair(V, EI->second));
606 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
607 valueNumbering.insert(std::make_pair(V, nextValueNumber));
609 return nextValueNumber++;
611 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
612 Expression e = create_expression(U);
614 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
615 if (EI != expressionNumbering.end()) {
616 valueNumbering.insert(std::make_pair(V, EI->second));
619 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
620 valueNumbering.insert(std::make_pair(V, nextValueNumber));
622 return nextValueNumber++;
624 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
625 Expression e = create_expression(U);
627 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
628 if (EI != expressionNumbering.end()) {
629 valueNumbering.insert(std::make_pair(V, EI->second));
632 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
633 valueNumbering.insert(std::make_pair(V, nextValueNumber));
635 return nextValueNumber++;
638 valueNumbering.insert(std::make_pair(V, nextValueNumber));
639 return nextValueNumber++;
643 /// lookup - Returns the value number of the specified value. Fails if
644 /// the value has not yet been numbered.
645 uint32_t ValueTable::lookup(Value* V) const {
646 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
647 assert(VI != valueNumbering.end() && "Value not numbered?");
651 /// clear - Remove all entries from the ValueTable
652 void ValueTable::clear() {
653 valueNumbering.clear();
654 expressionNumbering.clear();
658 /// erase - Remove a value from the value numbering
659 void ValueTable::erase(Value* V) {
660 valueNumbering.erase(V);
663 /// verifyRemoved - Verify that the value is removed from all internal data
665 void ValueTable::verifyRemoved(const Value *V) const {
666 for (DenseMap<Value*, uint32_t>::iterator
667 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
668 assert(I->first != V && "Inst still occurs in value numbering map!");
672 //===----------------------------------------------------------------------===//
674 //===----------------------------------------------------------------------===//
677 struct VISIBILITY_HIDDEN ValueNumberScope {
678 ValueNumberScope* parent;
679 DenseMap<uint32_t, Value*> table;
681 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
687 class VISIBILITY_HIDDEN GVN : public FunctionPass {
688 bool runOnFunction(Function &F);
690 static char ID; // Pass identification, replacement for typeid
691 GVN() : FunctionPass(&ID) { }
694 MemoryDependenceAnalysis *MD;
698 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
700 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
704 // This transformation requires dominator postdominator info
705 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
706 AU.addRequired<DominatorTree>();
707 AU.addRequired<MemoryDependenceAnalysis>();
708 AU.addRequired<AliasAnalysis>();
710 AU.addPreserved<DominatorTree>();
711 AU.addPreserved<AliasAnalysis>();
715 // FIXME: eliminate or document these better
716 bool processLoad(LoadInst* L,
717 SmallVectorImpl<Instruction*> &toErase);
718 bool processInstruction(Instruction* I,
719 SmallVectorImpl<Instruction*> &toErase);
720 bool processNonLocalLoad(LoadInst* L,
721 SmallVectorImpl<Instruction*> &toErase);
722 bool processBlock(BasicBlock* BB);
723 Value *GetValueForBlock(BasicBlock *BB, Instruction* orig,
724 DenseMap<BasicBlock*, Value*> &Phis,
725 bool top_level = false);
726 void dump(DenseMap<uint32_t, Value*>& d);
727 bool iterateOnFunction(Function &F);
728 Value* CollapsePhi(PHINode* p);
729 bool isSafeReplacement(PHINode* p, Instruction* inst);
730 bool performPRE(Function& F);
731 Value* lookupNumber(BasicBlock* BB, uint32_t num);
732 bool mergeBlockIntoPredecessor(BasicBlock* BB);
733 Value* AttemptRedundancyElimination(Instruction* orig, unsigned valno);
734 void cleanupGlobalSets();
735 void verifyRemoved(const Instruction *I) const;
741 // createGVNPass - The public interface to this file...
742 FunctionPass *llvm::createGVNPass() { return new GVN(); }
744 static RegisterPass<GVN> X("gvn",
745 "Global Value Numbering");
747 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
749 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
750 E = d.end(); I != E; ++I) {
751 printf("%d\n", I->first);
757 Value* GVN::CollapsePhi(PHINode* p) {
758 Value* constVal = p->hasConstantValue();
759 if (!constVal) return 0;
761 Instruction* inst = dyn_cast<Instruction>(constVal);
765 if (DT->dominates(inst, p))
766 if (isSafeReplacement(p, inst))
771 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
772 if (!isa<PHINode>(inst))
775 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
777 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
778 if (use_phi->getParent() == inst->getParent())
784 /// GetValueForBlock - Get the value to use within the specified basic block.
785 /// available values are in Phis.
786 Value *GVN::GetValueForBlock(BasicBlock *BB, Instruction* orig,
787 DenseMap<BasicBlock*, Value*> &Phis,
790 // If we have already computed this value, return the previously computed val.
791 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
792 if (V != Phis.end() && !top_level) return V->second;
794 // If the block is unreachable, just return undef, since this path
795 // can't actually occur at runtime.
796 if (!DT->isReachableFromEntry(BB))
797 return Phis[BB] = UndefValue::get(orig->getType());
799 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
800 Value *ret = GetValueForBlock(Pred, orig, Phis);
805 // Get the number of predecessors of this block so we can reserve space later.
806 // If there is already a PHI in it, use the #preds from it, otherwise count.
807 // Getting it from the PHI is constant time.
809 if (PHINode *ExistingPN = dyn_cast<PHINode>(BB->begin()))
810 NumPreds = ExistingPN->getNumIncomingValues();
812 NumPreds = std::distance(pred_begin(BB), pred_end(BB));
814 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
815 // now, then get values to fill in the incoming values for the PHI.
816 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
818 PN->reserveOperandSpace(NumPreds);
820 Phis.insert(std::make_pair(BB, PN));
822 // Fill in the incoming values for the block.
823 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
824 Value* val = GetValueForBlock(*PI, orig, Phis);
825 PN->addIncoming(val, *PI);
828 VN.getAliasAnalysis()->copyValue(orig, PN);
830 // Attempt to collapse PHI nodes that are trivially redundant
831 Value* v = CollapsePhi(PN);
833 // Cache our phi construction results
834 if (LoadInst* L = dyn_cast<LoadInst>(orig))
835 phiMap[L->getPointerOperand()].insert(PN);
837 phiMap[orig].insert(PN);
842 PN->replaceAllUsesWith(v);
843 if (isa<PointerType>(v->getType()))
844 MD->invalidateCachedPointerInfo(v);
846 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
847 E = Phis.end(); I != E; ++I)
851 DEBUG(cerr << "GVN removed: " << *PN);
852 MD->removeInstruction(PN);
853 PN->eraseFromParent();
854 DEBUG(verifyRemoved(PN));
860 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
861 /// we're analyzing is fully available in the specified block. As we go, keep
862 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
863 /// map is actually a tri-state map with the following values:
864 /// 0) we know the block *is not* fully available.
865 /// 1) we know the block *is* fully available.
866 /// 2) we do not know whether the block is fully available or not, but we are
867 /// currently speculating that it will be.
868 /// 3) we are speculating for this block and have used that to speculate for
870 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
871 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
872 // Optimistically assume that the block is fully available and check to see
873 // if we already know about this block in one lookup.
874 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
875 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
877 // If the entry already existed for this block, return the precomputed value.
879 // If this is a speculative "available" value, mark it as being used for
880 // speculation of other blocks.
881 if (IV.first->second == 2)
882 IV.first->second = 3;
883 return IV.first->second != 0;
886 // Otherwise, see if it is fully available in all predecessors.
887 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
889 // If this block has no predecessors, it isn't live-in here.
891 goto SpeculationFailure;
893 for (; PI != PE; ++PI)
894 // If the value isn't fully available in one of our predecessors, then it
895 // isn't fully available in this block either. Undo our previous
896 // optimistic assumption and bail out.
897 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
898 goto SpeculationFailure;
902 // SpeculationFailure - If we get here, we found out that this is not, after
903 // all, a fully-available block. We have a problem if we speculated on this and
904 // used the speculation to mark other blocks as available.
906 char &BBVal = FullyAvailableBlocks[BB];
908 // If we didn't speculate on this, just return with it set to false.
914 // If we did speculate on this value, we could have blocks set to 1 that are
915 // incorrect. Walk the (transitive) successors of this block and mark them as
917 SmallVector<BasicBlock*, 32> BBWorklist;
918 BBWorklist.push_back(BB);
920 while (!BBWorklist.empty()) {
921 BasicBlock *Entry = BBWorklist.pop_back_val();
922 // Note that this sets blocks to 0 (unavailable) if they happen to not
923 // already be in FullyAvailableBlocks. This is safe.
924 char &EntryVal = FullyAvailableBlocks[Entry];
925 if (EntryVal == 0) continue; // Already unavailable.
927 // Mark as unavailable.
930 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
931 BBWorklist.push_back(*I);
937 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
938 /// non-local by performing PHI construction.
939 bool GVN::processNonLocalLoad(LoadInst *LI,
940 SmallVectorImpl<Instruction*> &toErase) {
941 // Find the non-local dependencies of the load.
942 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
943 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
945 //DEBUG(cerr << "INVESTIGATING NONLOCAL LOAD: " << Deps.size() << *LI);
947 // If we had to process more than one hundred blocks to find the
948 // dependencies, this load isn't worth worrying about. Optimizing
949 // it will be too expensive.
950 if (Deps.size() > 100)
953 // If we had a phi translation failure, we'll have a single entry which is a
954 // clobber in the current block. Reject this early.
955 if (Deps.size() == 1 && Deps[0].second.isClobber())
958 // Filter out useless results (non-locals, etc). Keep track of the blocks
959 // where we have a value available in repl, also keep track of whether we see
960 // dependencies that produce an unknown value for the load (such as a call
961 // that could potentially clobber the load).
962 SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
963 SmallVector<BasicBlock*, 16> UnavailableBlocks;
965 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
966 BasicBlock *DepBB = Deps[i].first;
967 MemDepResult DepInfo = Deps[i].second;
969 if (DepInfo.isClobber()) {
970 UnavailableBlocks.push_back(DepBB);
974 Instruction *DepInst = DepInfo.getInst();
976 // Loading the allocation -> undef.
977 if (isa<AllocationInst>(DepInst)) {
978 ValuesPerBlock.push_back(std::make_pair(DepBB,
979 UndefValue::get(LI->getType())));
983 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
984 // Reject loads and stores that are to the same address but are of
986 // NOTE: 403.gcc does have this case (e.g. in readonly_fields_p) because
987 // of bitfield access, it would be interesting to optimize for it at some
989 if (S->getOperand(0)->getType() != LI->getType()) {
990 UnavailableBlocks.push_back(DepBB);
994 ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));
996 } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
997 if (LD->getType() != LI->getType()) {
998 UnavailableBlocks.push_back(DepBB);
1001 ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
1003 UnavailableBlocks.push_back(DepBB);
1008 // If we have no predecessors that produce a known value for this load, exit
1010 if (ValuesPerBlock.empty()) return false;
1012 // If all of the instructions we depend on produce a known value for this
1013 // load, then it is fully redundant and we can use PHI insertion to compute
1014 // its value. Insert PHIs and remove the fully redundant value now.
1015 if (UnavailableBlocks.empty()) {
1016 // Use cached PHI construction information from previous runs
1017 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
1018 // FIXME: What does phiMap do? Are we positive it isn't getting invalidated?
1019 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
1021 if ((*I)->getParent() == LI->getParent()) {
1022 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD #1: " << *LI);
1023 LI->replaceAllUsesWith(*I);
1024 if (isa<PointerType>((*I)->getType()))
1025 MD->invalidateCachedPointerInfo(*I);
1026 toErase.push_back(LI);
1031 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
1034 DEBUG(cerr << "GVN REMOVING NONLOCAL LOAD: " << *LI);
1036 DenseMap<BasicBlock*, Value*> BlockReplValues;
1037 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1038 // Perform PHI construction.
1039 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1040 LI->replaceAllUsesWith(v);
1042 if (isa<PHINode>(v))
1044 if (isa<PointerType>(v->getType()))
1045 MD->invalidateCachedPointerInfo(v);
1046 toErase.push_back(LI);
1051 if (!EnablePRE || !EnableLoadPRE)
1054 // Okay, we have *some* definitions of the value. This means that the value
1055 // is available in some of our (transitive) predecessors. Lets think about
1056 // doing PRE of this load. This will involve inserting a new load into the
1057 // predecessor when it's not available. We could do this in general, but
1058 // prefer to not increase code size. As such, we only do this when we know
1059 // that we only have to insert *one* load (which means we're basically moving
1060 // the load, not inserting a new one).
1062 SmallPtrSet<BasicBlock *, 4> Blockers;
1063 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1064 Blockers.insert(UnavailableBlocks[i]);
1066 // Lets find first basic block with more than one predecessor. Walk backwards
1067 // through predecessors if needed.
1068 BasicBlock *LoadBB = LI->getParent();
1069 BasicBlock *TmpBB = LoadBB;
1071 bool isSinglePred = false;
1072 while (TmpBB->getSinglePredecessor()) {
1073 isSinglePred = true;
1074 TmpBB = TmpBB->getSinglePredecessor();
1075 if (!TmpBB) // If haven't found any, bail now.
1077 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1079 if (Blockers.count(TmpBB))
1086 // If we have a repl set with LI itself in it, this means we have a loop where
1087 // at least one of the values is LI. Since this means that we won't be able
1088 // to eliminate LI even if we insert uses in the other predecessors, we will
1089 // end up increasing code size. Reject this by scanning for LI.
1090 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1091 if (ValuesPerBlock[i].second == LI)
1096 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1097 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second))
1098 // "Hot" Instruction is in some loop (because it dominates its dep.
1100 if (DT->dominates(LI, I)) {
1105 // We are interested only in "hot" instructions. We don't want to do any
1106 // mis-optimizations here.
1111 // Okay, we have some hope :). Check to see if the loaded value is fully
1112 // available in all but one predecessor.
1113 // FIXME: If we could restructure the CFG, we could make a common pred with
1114 // all the preds that don't have an available LI and insert a new load into
1116 BasicBlock *UnavailablePred = 0;
1118 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1119 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1120 FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
1121 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1122 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1124 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1126 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1129 // If this load is not available in multiple predecessors, reject it.
1130 if (UnavailablePred && UnavailablePred != *PI)
1132 UnavailablePred = *PI;
1135 assert(UnavailablePred != 0 &&
1136 "Fully available value should be eliminated above!");
1138 // If the loaded pointer is PHI node defined in this block, do PHI translation
1139 // to get its value in the predecessor.
1140 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
1142 // Make sure the value is live in the predecessor. If it was defined by a
1143 // non-PHI instruction in this block, we don't know how to recompute it above.
1144 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
1145 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
1146 DEBUG(cerr << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
1147 << *LPInst << *LI << "\n");
1151 // We don't currently handle critical edges :(
1152 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1153 DEBUG(cerr << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1154 << UnavailablePred->getName() << "': " << *LI);
1158 // Okay, we can eliminate this load by inserting a reload in the predecessor
1159 // and using PHI construction to get the value in the other predecessors, do
1161 DEBUG(cerr << "GVN REMOVING PRE LOAD: " << *LI);
1163 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1165 UnavailablePred->getTerminator());
1167 SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
1168 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
1170 ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
1172 DenseMap<BasicBlock*, Value*> BlockReplValues;
1173 BlockReplValues.insert(ValuesPerBlock.begin(), ValuesPerBlock.end());
1174 BlockReplValues[UnavailablePred] = NewLoad;
1176 // Perform PHI construction.
1177 Value* v = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
1178 LI->replaceAllUsesWith(v);
1179 if (isa<PHINode>(v))
1181 if (isa<PointerType>(v->getType()))
1182 MD->invalidateCachedPointerInfo(v);
1183 toErase.push_back(LI);
1188 /// processLoad - Attempt to eliminate a load, first by eliminating it
1189 /// locally, and then attempting non-local elimination if that fails.
1190 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1191 if (L->isVolatile())
1194 Value* pointer = L->getPointerOperand();
1196 // ... to a pointer that has been loaded from before...
1197 MemDepResult dep = MD->getDependency(L);
1199 // If the value isn't available, don't do anything!
1200 if (dep.isClobber()) {
1202 // fast print dep, using operator<< on instruction would be too slow
1203 DOUT << "GVN: load ";
1204 WriteAsOperand(*DOUT.stream(), L);
1205 Instruction *I = dep.getInst();
1206 DOUT << " is clobbered by " << *I;
1211 // If it is defined in another block, try harder.
1212 if (dep.isNonLocal())
1213 return processNonLocalLoad(L, toErase);
1215 Instruction *DepInst = dep.getInst();
1216 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1217 // Only forward substitute stores to loads of the same type.
1218 // FIXME: Could do better!
1219 if (DepSI->getPointerOperand()->getType() != pointer->getType())
1223 L->replaceAllUsesWith(DepSI->getOperand(0));
1224 if (isa<PointerType>(DepSI->getOperand(0)->getType()))
1225 MD->invalidateCachedPointerInfo(DepSI->getOperand(0));
1226 toErase.push_back(L);
1231 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1232 // Only forward substitute stores to loads of the same type.
1233 // FIXME: Could do better! load i32 -> load i8 -> truncate on little endian.
1234 if (DepLI->getType() != L->getType())
1238 L->replaceAllUsesWith(DepLI);
1239 if (isa<PointerType>(DepLI->getType()))
1240 MD->invalidateCachedPointerInfo(DepLI);
1241 toErase.push_back(L);
1246 // If this load really doesn't depend on anything, then we must be loading an
1247 // undef value. This can happen when loading for a fresh allocation with no
1248 // intervening stores, for example.
1249 if (isa<AllocationInst>(DepInst)) {
1250 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1251 toErase.push_back(L);
1259 Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
1260 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1261 if (I == localAvail.end())
1264 ValueNumberScope* locals = I->second;
1267 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
1268 if (I != locals->table.end())
1271 locals = locals->parent;
1277 /// AttemptRedundancyElimination - If the "fast path" of redundancy elimination
1278 /// by inheritance from the dominator fails, see if we can perform phi
1279 /// construction to eliminate the redundancy.
1280 Value* GVN::AttemptRedundancyElimination(Instruction* orig, unsigned valno) {
1281 BasicBlock* BaseBlock = orig->getParent();
1283 SmallPtrSet<BasicBlock*, 4> Visited;
1284 SmallVector<BasicBlock*, 8> Stack;
1285 Stack.push_back(BaseBlock);
1287 DenseMap<BasicBlock*, Value*> Results;
1289 // Walk backwards through our predecessors, looking for instances of the
1290 // value number we're looking for. Instances are recorded in the Results
1291 // map, which is then used to perform phi construction.
1292 while (!Stack.empty()) {
1293 BasicBlock* Current = Stack.back();
1296 // If we've walked all the way to a proper dominator, then give up. Cases
1297 // where the instance is in the dominator will have been caught by the fast
1298 // path, and any cases that require phi construction further than this are
1299 // probably not worth it anyways. Note that this is a SIGNIFICANT compile
1300 // time improvement.
1301 if (DT->properlyDominates(Current, orig->getParent())) return 0;
1303 DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
1304 localAvail.find(Current);
1305 if (LA == localAvail.end()) return 0;
1306 DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);
1308 if (V != LA->second->table.end()) {
1309 // Found an instance, record it.
1310 Results.insert(std::make_pair(Current, V->second));
1314 // If we reach the beginning of the function, then give up.
1315 if (pred_begin(Current) == pred_end(Current))
1318 for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current);
1320 if (Visited.insert(*PI))
1321 Stack.push_back(*PI);
1324 // If we didn't find instances, give up. Otherwise, perform phi construction.
1325 if (Results.size() == 0)
1328 return GetValueForBlock(BaseBlock, orig, Results, true);
1331 /// processInstruction - When calculating availability, handle an instruction
1332 /// by inserting it into the appropriate sets
1333 bool GVN::processInstruction(Instruction *I,
1334 SmallVectorImpl<Instruction*> &toErase) {
1335 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1336 bool changed = processLoad(L, toErase);
1339 unsigned num = VN.lookup_or_add(L);
1340 localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
1346 uint32_t nextNum = VN.getNextUnusedValueNumber();
1347 unsigned num = VN.lookup_or_add(I);
1349 if (BranchInst* BI = dyn_cast<BranchInst>(I)) {
1350 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1352 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1355 Value* branchCond = BI->getCondition();
1356 uint32_t condVN = VN.lookup_or_add(branchCond);
1358 BasicBlock* trueSucc = BI->getSuccessor(0);
1359 BasicBlock* falseSucc = BI->getSuccessor(1);
1361 if (trueSucc->getSinglePredecessor())
1362 localAvail[trueSucc]->table[condVN] = ConstantInt::getTrue();
1363 if (falseSucc->getSinglePredecessor())
1364 localAvail[falseSucc]->table[condVN] = ConstantInt::getFalse();
1368 // Allocations are always uniquely numbered, so we can save time and memory
1369 // by fast failing them.
1370 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1371 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1375 // Collapse PHI nodes
1376 if (PHINode* p = dyn_cast<PHINode>(I)) {
1377 Value* constVal = CollapsePhi(p);
1380 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1382 PI->second.erase(p);
1384 p->replaceAllUsesWith(constVal);
1385 if (isa<PointerType>(constVal->getType()))
1386 MD->invalidateCachedPointerInfo(constVal);
1389 toErase.push_back(p);
1391 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1394 // If the number we were assigned was a brand new VN, then we don't
1395 // need to do a lookup to see if the number already exists
1396 // somewhere in the domtree: it can't!
1397 } else if (num == nextNum) {
1398 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1400 // Perform fast-path value-number based elimination of values inherited from
1402 } else if (Value* repl = lookupNumber(I->getParent(), num)) {
1405 I->replaceAllUsesWith(repl);
1406 if (isa<PointerType>(repl->getType()))
1407 MD->invalidateCachedPointerInfo(repl);
1408 toErase.push_back(I);
1412 // Perform slow-pathvalue-number based elimination with phi construction.
1413 } else if (Value* repl = AttemptRedundancyElimination(I, num)) {
1416 I->replaceAllUsesWith(repl);
1417 if (isa<PointerType>(repl->getType()))
1418 MD->invalidateCachedPointerInfo(repl);
1419 toErase.push_back(I);
1423 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1429 /// runOnFunction - This is the main transformation entry point for a function.
1430 bool GVN::runOnFunction(Function& F) {
1431 MD = &getAnalysis<MemoryDependenceAnalysis>();
1432 DT = &getAnalysis<DominatorTree>();
1433 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1437 bool changed = false;
1438 bool shouldContinue = true;
1440 // Merge unconditional branches, allowing PRE to catch more
1441 // optimization opportunities.
1442 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1443 BasicBlock* BB = FI;
1445 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1446 if (removedBlock) NumGVNBlocks++;
1448 changed |= removedBlock;
1451 unsigned Iteration = 0;
1453 while (shouldContinue) {
1454 DEBUG(cerr << "GVN iteration: " << Iteration << "\n");
1455 shouldContinue = iterateOnFunction(F);
1456 changed |= shouldContinue;
1461 bool PREChanged = true;
1462 while (PREChanged) {
1463 PREChanged = performPRE(F);
1464 changed |= PREChanged;
1467 // FIXME: Should perform GVN again after PRE does something. PRE can move
1468 // computations into blocks where they become fully redundant. Note that
1469 // we can't do this until PRE's critical edge splitting updates memdep.
1470 // Actually, when this happens, we should just fully integrate PRE into GVN.
1472 cleanupGlobalSets();
1478 bool GVN::processBlock(BasicBlock* BB) {
1479 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1480 // incrementing BI before processing an instruction).
1481 SmallVector<Instruction*, 8> toErase;
1482 bool changed_function = false;
1484 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1486 changed_function |= processInstruction(BI, toErase);
1487 if (toErase.empty()) {
1492 // If we need some instructions deleted, do it now.
1493 NumGVNInstr += toErase.size();
1495 // Avoid iterator invalidation.
1496 bool AtStart = BI == BB->begin();
1500 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1501 E = toErase.end(); I != E; ++I) {
1502 DEBUG(cerr << "GVN removed: " << **I);
1503 MD->removeInstruction(*I);
1504 (*I)->eraseFromParent();
1505 DEBUG(verifyRemoved(*I));
1515 return changed_function;
1518 /// performPRE - Perform a purely local form of PRE that looks for diamond
1519 /// control flow patterns and attempts to perform simple PRE at the join point.
1520 bool GVN::performPRE(Function& F) {
1521 bool Changed = false;
1522 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1523 DenseMap<BasicBlock*, Value*> predMap;
1524 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1525 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1526 BasicBlock* CurrentBlock = *DI;
1528 // Nothing to PRE in the entry block.
1529 if (CurrentBlock == &F.getEntryBlock()) continue;
1531 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1532 BE = CurrentBlock->end(); BI != BE; ) {
1533 Instruction *CurInst = BI++;
1535 if (isa<AllocationInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
1536 isa<PHINode>(CurInst) || (CurInst->getType() == Type::VoidTy) ||
1537 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1538 isa<DbgInfoIntrinsic>(CurInst))
1541 uint32_t valno = VN.lookup(CurInst);
1543 // Look for the predecessors for PRE opportunities. We're
1544 // only trying to solve the basic diamond case, where
1545 // a value is computed in the successor and one predecessor,
1546 // but not the other. We also explicitly disallow cases
1547 // where the successor is its own predecessor, because they're
1548 // more complicated to get right.
1549 unsigned numWith = 0;
1550 unsigned numWithout = 0;
1551 BasicBlock* PREPred = 0;
1554 for (pred_iterator PI = pred_begin(CurrentBlock),
1555 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1556 // We're not interested in PRE where the block is its
1557 // own predecessor, on in blocks with predecessors
1558 // that are not reachable.
1559 if (*PI == CurrentBlock) {
1562 } else if (!localAvail.count(*PI)) {
1567 DenseMap<uint32_t, Value*>::iterator predV =
1568 localAvail[*PI]->table.find(valno);
1569 if (predV == localAvail[*PI]->table.end()) {
1572 } else if (predV->second == CurInst) {
1575 predMap[*PI] = predV->second;
1580 // Don't do PRE when it might increase code size, i.e. when
1581 // we would need to insert instructions in more than one pred.
1582 if (numWithout != 1 || numWith == 0)
1585 // We can't do PRE safely on a critical edge, so instead we schedule
1586 // the edge to be split and perform the PRE the next time we iterate
1588 unsigned succNum = 0;
1589 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1591 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1596 if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
1597 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
1601 // Instantiate the expression the in predecessor that lacked it.
1602 // Because we are going top-down through the block, all value numbers
1603 // will be available in the predecessor by the time we need them. Any
1604 // that weren't original present will have been instantiated earlier
1606 Instruction* PREInstr = CurInst->clone();
1607 bool success = true;
1608 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1609 Value *Op = PREInstr->getOperand(i);
1610 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1613 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
1614 PREInstr->setOperand(i, V);
1621 // Fail out if we encounter an operand that is not available in
1622 // the PRE predecessor. This is typically because of loads which
1623 // are not value numbered precisely.
1626 DEBUG(verifyRemoved(PREInstr));
1630 PREInstr->insertBefore(PREPred->getTerminator());
1631 PREInstr->setName(CurInst->getName() + ".pre");
1632 predMap[PREPred] = PREInstr;
1633 VN.add(PREInstr, valno);
1636 // Update the availability map to include the new instruction.
1637 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
1639 // Create a PHI to make the value available in this block.
1640 PHINode* Phi = PHINode::Create(CurInst->getType(),
1641 CurInst->getName() + ".pre-phi",
1642 CurrentBlock->begin());
1643 for (pred_iterator PI = pred_begin(CurrentBlock),
1644 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1645 Phi->addIncoming(predMap[*PI], *PI);
1648 localAvail[CurrentBlock]->table[valno] = Phi;
1650 CurInst->replaceAllUsesWith(Phi);
1651 if (isa<PointerType>(Phi->getType()))
1652 MD->invalidateCachedPointerInfo(Phi);
1655 DEBUG(cerr << "GVN PRE removed: " << *CurInst);
1656 MD->removeInstruction(CurInst);
1657 CurInst->eraseFromParent();
1658 DEBUG(verifyRemoved(CurInst));
1663 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1664 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1665 SplitCriticalEdge(I->first, I->second, this);
1667 return Changed || toSplit.size();
1670 /// iterateOnFunction - Executes one iteration of GVN
1671 bool GVN::iterateOnFunction(Function &F) {
1672 cleanupGlobalSets();
1674 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1675 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
1677 localAvail[DI->getBlock()] =
1678 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
1680 localAvail[DI->getBlock()] = new ValueNumberScope(0);
1683 // Top-down walk of the dominator tree
1684 bool changed = false;
1686 // Needed for value numbering with phi construction to work.
1687 ReversePostOrderTraversal<Function*> RPOT(&F);
1688 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
1689 RE = RPOT.end(); RI != RE; ++RI)
1690 changed |= processBlock(*RI);
1692 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1693 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
1694 changed |= processBlock(DI->getBlock());
1700 void GVN::cleanupGlobalSets() {
1704 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1705 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
1710 /// verifyRemoved - Verify that the specified instruction does not occur in our
1711 /// internal data structures.
1712 void GVN::verifyRemoved(const Instruction *Inst) const {
1713 VN.verifyRemoved(Inst);
1715 // Walk through the PHI map to make sure the instruction isn't hiding in there
1717 for (PhiMapType::iterator
1718 I = phiMap.begin(), E = phiMap.end(); I != E; ++I) {
1719 assert(I->first != Inst && "Inst is still a key in PHI map!");
1721 for (SmallPtrSet<Instruction*, 4>::iterator
1722 II = I->second.begin(), IE = I->second.end(); II != IE; ++II) {
1723 assert(*II != Inst && "Inst is still a value in PHI map!");
1727 // Walk through the value number scope to make sure the instruction isn't
1728 // ferreted away in it.
1729 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1730 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
1731 const ValueNumberScope *VNS = I->second;
1734 for (DenseMap<uint32_t, Value*>::iterator
1735 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
1736 assert(II->second != Inst && "Inst still in value numbering scope!");