1 //===- GVN.cpp - Eliminate redundant values and loads ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself, it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/BasicBlock.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Value.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Analysis/AliasAnalysis.h"
33 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
34 #include "llvm/Support/CFG.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 STATISTIC(NumGVNInstr, "Number of instructions deleted");
43 STATISTIC(NumGVNLoad, "Number of loads deleted");
44 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
45 STATISTIC(NumGVNBlocks, "Number of blocks merged");
47 static cl::opt<bool> EnablePRE("enable-pre",
48 cl::init(true), cl::Hidden);
50 //===----------------------------------------------------------------------===//
52 //===----------------------------------------------------------------------===//
54 /// This class holds the mapping between values and value numbers. It is used
55 /// as an efficient mechanism to determine the expression-wise equivalence of
58 struct VISIBILITY_HIDDEN Expression {
59 enum ExpressionOpcode { ADD, SUB, MUL, UDIV, SDIV, FDIV, UREM, SREM,
60 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
61 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
62 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
63 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
64 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
65 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
66 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
67 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
68 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
71 ExpressionOpcode opcode;
76 SmallVector<uint32_t, 4> varargs;
80 Expression(ExpressionOpcode o) : opcode(o) { }
82 bool operator==(const Expression &other) const {
83 if (opcode != other.opcode)
85 else if (opcode == EMPTY || opcode == TOMBSTONE)
87 else if (type != other.type)
89 else if (function != other.function)
91 else if (firstVN != other.firstVN)
93 else if (secondVN != other.secondVN)
95 else if (thirdVN != other.thirdVN)
98 if (varargs.size() != other.varargs.size())
101 for (size_t i = 0; i < varargs.size(); ++i)
102 if (varargs[i] != other.varargs[i])
109 bool operator!=(const Expression &other) const {
110 if (opcode != other.opcode)
112 else if (opcode == EMPTY || opcode == TOMBSTONE)
114 else if (type != other.type)
116 else if (function != other.function)
118 else if (firstVN != other.firstVN)
120 else if (secondVN != other.secondVN)
122 else if (thirdVN != other.thirdVN)
125 if (varargs.size() != other.varargs.size())
128 for (size_t i = 0; i < varargs.size(); ++i)
129 if (varargs[i] != other.varargs[i])
137 class VISIBILITY_HIDDEN ValueTable {
139 DenseMap<Value*, uint32_t> valueNumbering;
140 DenseMap<Expression, uint32_t> expressionNumbering;
142 MemoryDependenceAnalysis* MD;
145 uint32_t nextValueNumber;
147 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
148 Expression::ExpressionOpcode getOpcode(CmpInst* C);
149 Expression::ExpressionOpcode getOpcode(CastInst* C);
150 Expression create_expression(BinaryOperator* BO);
151 Expression create_expression(CmpInst* C);
152 Expression create_expression(ShuffleVectorInst* V);
153 Expression create_expression(ExtractElementInst* C);
154 Expression create_expression(InsertElementInst* V);
155 Expression create_expression(SelectInst* V);
156 Expression create_expression(CastInst* C);
157 Expression create_expression(GetElementPtrInst* G);
158 Expression create_expression(CallInst* C);
159 Expression create_expression(Constant* C);
161 ValueTable() : nextValueNumber(1) { }
162 uint32_t lookup_or_add(Value* V);
163 uint32_t lookup(Value* V) const;
164 void add(Value* V, uint32_t num);
166 void erase(Value* v);
168 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
169 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
170 void setDomTree(DominatorTree* D) { DT = D; }
171 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
176 template <> struct DenseMapInfo<Expression> {
177 static inline Expression getEmptyKey() {
178 return Expression(Expression::EMPTY);
181 static inline Expression getTombstoneKey() {
182 return Expression(Expression::TOMBSTONE);
185 static unsigned getHashValue(const Expression e) {
186 unsigned hash = e.opcode;
188 hash = e.firstVN + hash * 37;
189 hash = e.secondVN + hash * 37;
190 hash = e.thirdVN + hash * 37;
192 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
193 (unsigned)((uintptr_t)e.type >> 9)) +
196 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
197 E = e.varargs.end(); I != E; ++I)
198 hash = *I + hash * 37;
200 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
201 (unsigned)((uintptr_t)e.function >> 9)) +
206 static bool isEqual(const Expression &LHS, const Expression &RHS) {
209 static bool isPod() { return true; }
213 //===----------------------------------------------------------------------===//
214 // ValueTable Internal Functions
215 //===----------------------------------------------------------------------===//
216 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
217 switch(BO->getOpcode()) {
218 default: // THIS SHOULD NEVER HAPPEN
219 assert(0 && "Binary operator with unknown opcode?");
220 case Instruction::Add: return Expression::ADD;
221 case Instruction::Sub: return Expression::SUB;
222 case Instruction::Mul: return Expression::MUL;
223 case Instruction::UDiv: return Expression::UDIV;
224 case Instruction::SDiv: return Expression::SDIV;
225 case Instruction::FDiv: return Expression::FDIV;
226 case Instruction::URem: return Expression::UREM;
227 case Instruction::SRem: return Expression::SREM;
228 case Instruction::FRem: return Expression::FREM;
229 case Instruction::Shl: return Expression::SHL;
230 case Instruction::LShr: return Expression::LSHR;
231 case Instruction::AShr: return Expression::ASHR;
232 case Instruction::And: return Expression::AND;
233 case Instruction::Or: return Expression::OR;
234 case Instruction::Xor: return Expression::XOR;
238 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
239 if (isa<ICmpInst>(C) || isa<VICmpInst>(C)) {
240 switch (C->getPredicate()) {
241 default: // THIS SHOULD NEVER HAPPEN
242 assert(0 && "Comparison with unknown predicate?");
243 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
244 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
245 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
246 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
247 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
248 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
249 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
250 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
251 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
252 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
255 assert((isa<FCmpInst>(C) || isa<VFCmpInst>(C)) && "Unknown compare");
256 switch (C->getPredicate()) {
257 default: // THIS SHOULD NEVER HAPPEN
258 assert(0 && "Comparison with unknown predicate?");
259 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
260 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
261 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
262 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
263 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
264 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
265 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
266 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
267 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
268 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
269 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
270 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
271 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
272 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
276 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
277 switch(C->getOpcode()) {
278 default: // THIS SHOULD NEVER HAPPEN
279 assert(0 && "Cast operator with unknown opcode?");
280 case Instruction::Trunc: return Expression::TRUNC;
281 case Instruction::ZExt: return Expression::ZEXT;
282 case Instruction::SExt: return Expression::SEXT;
283 case Instruction::FPToUI: return Expression::FPTOUI;
284 case Instruction::FPToSI: return Expression::FPTOSI;
285 case Instruction::UIToFP: return Expression::UITOFP;
286 case Instruction::SIToFP: return Expression::SITOFP;
287 case Instruction::FPTrunc: return Expression::FPTRUNC;
288 case Instruction::FPExt: return Expression::FPEXT;
289 case Instruction::PtrToInt: return Expression::PTRTOINT;
290 case Instruction::IntToPtr: return Expression::INTTOPTR;
291 case Instruction::BitCast: return Expression::BITCAST;
295 Expression ValueTable::create_expression(CallInst* C) {
298 e.type = C->getType();
302 e.function = C->getCalledFunction();
303 e.opcode = Expression::CALL;
305 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
307 e.varargs.push_back(lookup_or_add(*I));
312 Expression ValueTable::create_expression(BinaryOperator* BO) {
315 e.firstVN = lookup_or_add(BO->getOperand(0));
316 e.secondVN = lookup_or_add(BO->getOperand(1));
319 e.type = BO->getType();
320 e.opcode = getOpcode(BO);
325 Expression ValueTable::create_expression(CmpInst* C) {
328 e.firstVN = lookup_or_add(C->getOperand(0));
329 e.secondVN = lookup_or_add(C->getOperand(1));
332 e.type = C->getType();
333 e.opcode = getOpcode(C);
338 Expression ValueTable::create_expression(CastInst* C) {
341 e.firstVN = lookup_or_add(C->getOperand(0));
345 e.type = C->getType();
346 e.opcode = getOpcode(C);
351 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
354 e.firstVN = lookup_or_add(S->getOperand(0));
355 e.secondVN = lookup_or_add(S->getOperand(1));
356 e.thirdVN = lookup_or_add(S->getOperand(2));
358 e.type = S->getType();
359 e.opcode = Expression::SHUFFLE;
364 Expression ValueTable::create_expression(ExtractElementInst* E) {
367 e.firstVN = lookup_or_add(E->getOperand(0));
368 e.secondVN = lookup_or_add(E->getOperand(1));
371 e.type = E->getType();
372 e.opcode = Expression::EXTRACT;
377 Expression ValueTable::create_expression(InsertElementInst* I) {
380 e.firstVN = lookup_or_add(I->getOperand(0));
381 e.secondVN = lookup_or_add(I->getOperand(1));
382 e.thirdVN = lookup_or_add(I->getOperand(2));
384 e.type = I->getType();
385 e.opcode = Expression::INSERT;
390 Expression ValueTable::create_expression(SelectInst* I) {
393 e.firstVN = lookup_or_add(I->getCondition());
394 e.secondVN = lookup_or_add(I->getTrueValue());
395 e.thirdVN = lookup_or_add(I->getFalseValue());
397 e.type = I->getType();
398 e.opcode = Expression::SELECT;
403 Expression ValueTable::create_expression(GetElementPtrInst* G) {
406 e.firstVN = lookup_or_add(G->getPointerOperand());
410 e.type = G->getType();
411 e.opcode = Expression::GEP;
413 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
415 e.varargs.push_back(lookup_or_add(*I));
420 //===----------------------------------------------------------------------===//
421 // ValueTable External Functions
422 //===----------------------------------------------------------------------===//
424 /// add - Insert a value into the table with a specified value number.
425 void ValueTable::add(Value* V, uint32_t num) {
426 valueNumbering.insert(std::make_pair(V, num));
429 /// lookup_or_add - Returns the value number for the specified value, assigning
430 /// it a new number if it did not have one before.
431 uint32_t ValueTable::lookup_or_add(Value* V) {
432 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
433 if (VI != valueNumbering.end())
436 if (CallInst* C = dyn_cast<CallInst>(V)) {
437 if (AA->doesNotAccessMemory(C)) {
438 Expression e = create_expression(C);
440 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
441 if (EI != expressionNumbering.end()) {
442 valueNumbering.insert(std::make_pair(V, EI->second));
445 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
446 valueNumbering.insert(std::make_pair(V, nextValueNumber));
448 return nextValueNumber++;
450 } else if (AA->onlyReadsMemory(C)) {
451 Expression e = create_expression(C);
453 if (expressionNumbering.find(e) == expressionNumbering.end()) {
454 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
455 valueNumbering.insert(std::make_pair(V, nextValueNumber));
456 return nextValueNumber++;
459 MemDepResult local_dep = MD->getDependency(C);
461 if (local_dep.isNone()) {
462 valueNumbering.insert(std::make_pair(V, nextValueNumber));
463 return nextValueNumber++;
464 } else if (Instruction *LocalDepInst = local_dep.getInst()) {
465 // FIXME: INDENT PROPERLY!
466 if (!isa<CallInst>(LocalDepInst)) {
467 valueNumbering.insert(std::make_pair(V, nextValueNumber));
468 return nextValueNumber++;
471 CallInst* local_cdep = cast<CallInst>(LocalDepInst);
473 // FIXME: INDENT PROPERLY.
474 if (local_cdep->getCalledFunction() != C->getCalledFunction() ||
475 local_cdep->getNumOperands() != C->getNumOperands()) {
476 valueNumbering.insert(std::make_pair(V, nextValueNumber));
477 return nextValueNumber++;
478 } else if (!C->getCalledFunction()) {
479 valueNumbering.insert(std::make_pair(V, nextValueNumber));
480 return nextValueNumber++;
482 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
483 uint32_t c_vn = lookup_or_add(C->getOperand(i));
484 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
486 valueNumbering.insert(std::make_pair(V, nextValueNumber));
487 return nextValueNumber++;
491 uint32_t v = lookup_or_add(local_cdep);
492 valueNumbering.insert(std::make_pair(V, v));
498 SmallVector<std::pair<BasicBlock*, MemDepResult>, 32> deps;
499 MD->getNonLocalDependency(C, deps);
502 for (SmallVector<std::pair<BasicBlock*, MemDepResult>, 32>
503 ::iterator I = deps.begin(), E = deps.end(); I != E; ++I) {
504 if (I->second.isNone()) {
505 valueNumbering.insert(std::make_pair(V, nextValueNumber));
507 return nextValueNumber++;
508 } else if (Instruction *NonLocalDepInst = I->second.getInst()) {
509 // FIXME: INDENT PROPERLY
510 // FIXME: All duplicated with non-local case.
511 if (cdep == 0 && DT->properlyDominates(I->first, C->getParent())) {
512 if (CallInst* CD = dyn_cast<CallInst>(NonLocalDepInst))
515 valueNumbering.insert(std::make_pair(V, nextValueNumber));
516 return nextValueNumber++;
519 valueNumbering.insert(std::make_pair(V, nextValueNumber));
520 return nextValueNumber++;
526 valueNumbering.insert(std::make_pair(V, nextValueNumber));
527 return nextValueNumber++;
530 // FIXME: THIS ISN'T SAFE: CONSIDER:
535 // This doesn't guarantee all-paths availability!
536 if (cdep->getCalledFunction() != C->getCalledFunction() ||
537 cdep->getNumOperands() != C->getNumOperands()) {
538 valueNumbering.insert(std::make_pair(V, nextValueNumber));
539 return nextValueNumber++;
540 } else if (!C->getCalledFunction()) {
541 valueNumbering.insert(std::make_pair(V, nextValueNumber));
542 return nextValueNumber++;
544 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
545 uint32_t c_vn = lookup_or_add(C->getOperand(i));
546 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
548 valueNumbering.insert(std::make_pair(V, nextValueNumber));
549 return nextValueNumber++;
553 uint32_t v = lookup_or_add(cdep);
554 valueNumbering.insert(std::make_pair(V, v));
559 valueNumbering.insert(std::make_pair(V, nextValueNumber));
560 return nextValueNumber++;
562 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
563 Expression e = create_expression(BO);
565 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
566 if (EI != expressionNumbering.end()) {
567 valueNumbering.insert(std::make_pair(V, EI->second));
570 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
571 valueNumbering.insert(std::make_pair(V, nextValueNumber));
573 return nextValueNumber++;
575 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
576 Expression e = create_expression(C);
578 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
579 if (EI != expressionNumbering.end()) {
580 valueNumbering.insert(std::make_pair(V, EI->second));
583 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
584 valueNumbering.insert(std::make_pair(V, nextValueNumber));
586 return nextValueNumber++;
588 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
589 Expression e = create_expression(U);
591 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
592 if (EI != expressionNumbering.end()) {
593 valueNumbering.insert(std::make_pair(V, EI->second));
596 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
597 valueNumbering.insert(std::make_pair(V, nextValueNumber));
599 return nextValueNumber++;
601 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
602 Expression e = create_expression(U);
604 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
605 if (EI != expressionNumbering.end()) {
606 valueNumbering.insert(std::make_pair(V, EI->second));
609 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
610 valueNumbering.insert(std::make_pair(V, nextValueNumber));
612 return nextValueNumber++;
614 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
615 Expression e = create_expression(U);
617 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
618 if (EI != expressionNumbering.end()) {
619 valueNumbering.insert(std::make_pair(V, EI->second));
622 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
623 valueNumbering.insert(std::make_pair(V, nextValueNumber));
625 return nextValueNumber++;
627 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
628 Expression e = create_expression(U);
630 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
631 if (EI != expressionNumbering.end()) {
632 valueNumbering.insert(std::make_pair(V, EI->second));
635 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
636 valueNumbering.insert(std::make_pair(V, nextValueNumber));
638 return nextValueNumber++;
640 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
641 Expression e = create_expression(U);
643 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
644 if (EI != expressionNumbering.end()) {
645 valueNumbering.insert(std::make_pair(V, EI->second));
648 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
649 valueNumbering.insert(std::make_pair(V, nextValueNumber));
651 return nextValueNumber++;
653 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
654 Expression e = create_expression(U);
656 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
657 if (EI != expressionNumbering.end()) {
658 valueNumbering.insert(std::make_pair(V, EI->second));
661 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
662 valueNumbering.insert(std::make_pair(V, nextValueNumber));
664 return nextValueNumber++;
667 valueNumbering.insert(std::make_pair(V, nextValueNumber));
668 return nextValueNumber++;
672 /// lookup - Returns the value number of the specified value. Fails if
673 /// the value has not yet been numbered.
674 uint32_t ValueTable::lookup(Value* V) const {
675 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
676 assert(VI != valueNumbering.end() && "Value not numbered?");
680 /// clear - Remove all entries from the ValueTable
681 void ValueTable::clear() {
682 valueNumbering.clear();
683 expressionNumbering.clear();
687 /// erase - Remove a value from the value numbering
688 void ValueTable::erase(Value* V) {
689 valueNumbering.erase(V);
692 //===----------------------------------------------------------------------===//
694 //===----------------------------------------------------------------------===//
697 struct VISIBILITY_HIDDEN ValueNumberScope {
698 ValueNumberScope* parent;
699 DenseMap<uint32_t, Value*> table;
701 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
707 class VISIBILITY_HIDDEN GVN : public FunctionPass {
708 bool runOnFunction(Function &F);
710 static char ID; // Pass identification, replacement for typeid
711 GVN() : FunctionPass(&ID) { }
715 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
717 typedef DenseMap<Value*, SmallPtrSet<Instruction*, 4> > PhiMapType;
721 // This transformation requires dominator postdominator info
722 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
723 AU.addRequired<DominatorTree>();
724 AU.addRequired<MemoryDependenceAnalysis>();
725 AU.addRequired<AliasAnalysis>();
727 AU.addPreserved<DominatorTree>();
728 AU.addPreserved<AliasAnalysis>();
732 // FIXME: eliminate or document these better
733 bool processLoad(LoadInst* L,
734 DenseMap<Value*, LoadInst*> &lastLoad,
735 SmallVectorImpl<Instruction*> &toErase);
736 bool processInstruction(Instruction* I,
737 DenseMap<Value*, LoadInst*>& lastSeenLoad,
738 SmallVectorImpl<Instruction*> &toErase);
739 bool processNonLocalLoad(LoadInst* L,
740 SmallVectorImpl<Instruction*> &toErase);
741 bool processBlock(DomTreeNode* DTN);
742 Value *GetValueForBlock(BasicBlock *BB, LoadInst* orig,
743 DenseMap<BasicBlock*, Value*> &Phis,
744 bool top_level = false);
745 void dump(DenseMap<uint32_t, Value*>& d);
746 bool iterateOnFunction(Function &F);
747 Value* CollapsePhi(PHINode* p);
748 bool isSafeReplacement(PHINode* p, Instruction* inst);
749 bool performPRE(Function& F);
750 Value* lookupNumber(BasicBlock* BB, uint32_t num);
751 bool mergeBlockIntoPredecessor(BasicBlock* BB);
752 void cleanupGlobalSets();
758 // createGVNPass - The public interface to this file...
759 FunctionPass *llvm::createGVNPass() { return new GVN(); }
761 static RegisterPass<GVN> X("gvn",
762 "Global Value Numbering");
764 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
766 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
767 E = d.end(); I != E; ++I) {
768 printf("%d\n", I->first);
774 Value* GVN::CollapsePhi(PHINode* p) {
775 DominatorTree &DT = getAnalysis<DominatorTree>();
776 Value* constVal = p->hasConstantValue();
778 if (!constVal) return 0;
780 Instruction* inst = dyn_cast<Instruction>(constVal);
784 if (DT.dominates(inst, p))
785 if (isSafeReplacement(p, inst))
790 bool GVN::isSafeReplacement(PHINode* p, Instruction* inst) {
791 if (!isa<PHINode>(inst))
794 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
796 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
797 if (use_phi->getParent() == inst->getParent())
803 /// GetValueForBlock - Get the value to use within the specified basic block.
804 /// available values are in Phis.
805 Value *GVN::GetValueForBlock(BasicBlock *BB, LoadInst* orig,
806 DenseMap<BasicBlock*, Value*> &Phis,
809 // If we have already computed this value, return the previously computed val.
810 DenseMap<BasicBlock*, Value*>::iterator V = Phis.find(BB);
811 if (V != Phis.end() && !top_level) return V->second;
813 // If the block is unreachable, just return undef, since this path
814 // can't actually occur at runtime.
815 if (!getAnalysis<DominatorTree>().isReachableFromEntry(BB))
816 return Phis[BB] = UndefValue::get(orig->getType());
818 BasicBlock* singlePred = BB->getSinglePredecessor();
820 Value *ret = GetValueForBlock(singlePred, orig, Phis);
825 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
826 // now, then get values to fill in the incoming values for the PHI.
827 PHINode *PN = PHINode::Create(orig->getType(), orig->getName()+".rle",
829 PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
831 if (Phis.count(BB) == 0)
832 Phis.insert(std::make_pair(BB, PN));
834 // Fill in the incoming values for the block.
835 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
836 Value* val = GetValueForBlock(*PI, orig, Phis);
837 PN->addIncoming(val, *PI);
840 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
841 AA.copyValue(orig, PN);
843 // Attempt to collapse PHI nodes that are trivially redundant
844 Value* v = CollapsePhi(PN);
846 // Cache our phi construction results
847 phiMap[orig->getPointerOperand()].insert(PN);
851 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
853 MD.removeInstruction(PN);
854 PN->replaceAllUsesWith(v);
856 for (DenseMap<BasicBlock*, Value*>::iterator I = Phis.begin(),
857 E = Phis.end(); I != E; ++I)
861 PN->eraseFromParent();
867 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
868 /// non-local by performing PHI construction.
869 bool GVN::processNonLocalLoad(LoadInst* L,
870 SmallVectorImpl<Instruction*> &toErase) {
871 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
873 // Find the non-local dependencies of the load
874 SmallVector<std::pair<BasicBlock*, MemDepResult>, 32> deps;
875 MD.getNonLocalDependency(L, deps);
877 // If we had to process more than one hundred blocks to find the
878 // dependencies, this load isn't worth worrying about. Optimizing
879 // it will be too expensive.
880 if (deps.size() > 100)
883 BasicBlock *EntryBlock = &L->getParent()->getParent()->getEntryBlock();
885 DenseMap<BasicBlock*, Value*> repl;
887 // Filter out useless results (non-locals, etc)
888 for (SmallVector<std::pair<BasicBlock*, MemDepResult>, 32>::iterator
889 I = deps.begin(), E = deps.end(); I != E; ++I) {
890 if (I->second.isNone()) {
891 repl[I->first] = UndefValue::get(L->getType());
895 if (I->second.isNonLocal()) {
896 // If this is a non-local dependency in the entry block, then we depend on
897 // the value live-in at the start of the function. We could insert a load
898 // in the entry block to get this, but for now we'll just bail out.
899 // FIXME: Consider emitting a load in the entry block to catch this case!
900 if (I->first == EntryBlock)
905 if (StoreInst* S = dyn_cast<StoreInst>(I->second.getInst())) {
906 if (S->getPointerOperand() != L->getPointerOperand())
908 repl[I->first] = S->getOperand(0);
909 } else if (LoadInst* LD = dyn_cast<LoadInst>(I->second.getInst())) {
910 if (LD->getPointerOperand() != L->getPointerOperand())
918 // Use cached PHI construction information from previous runs
919 SmallPtrSet<Instruction*, 4>& p = phiMap[L->getPointerOperand()];
920 for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
922 if ((*I)->getParent() == L->getParent()) {
923 MD.removeInstruction(L);
924 L->replaceAllUsesWith(*I);
925 toErase.push_back(L);
930 repl.insert(std::make_pair((*I)->getParent(), *I));
933 // Perform PHI construction
934 SmallPtrSet<BasicBlock*, 4> visited;
935 Value* v = GetValueForBlock(L->getParent(), L, repl, true);
937 MD.removeInstruction(L);
938 L->replaceAllUsesWith(v);
939 toErase.push_back(L);
945 /// processLoad - Attempt to eliminate a load, first by eliminating it
946 /// locally, and then attempting non-local elimination if that fails.
947 bool GVN::processLoad(LoadInst *L, DenseMap<Value*, LoadInst*> &lastLoad,
948 SmallVectorImpl<Instruction*> &toErase) {
949 if (L->isVolatile()) {
950 lastLoad[L->getPointerOperand()] = L;
954 Value* pointer = L->getPointerOperand();
955 LoadInst*& last = lastLoad[pointer];
957 // ... to a pointer that has been loaded from before...
958 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
959 bool removedNonLocal = false;
960 MemDepResult dep = MD.getDependency(L);
961 if (dep.isNonLocal() &&
962 L->getParent() != &L->getParent()->getParent()->getEntryBlock()) {
963 removedNonLocal = processNonLocalLoad(L, toErase);
965 if (!removedNonLocal)
968 return removedNonLocal;
972 bool deletedLoad = false;
974 // Walk up the dependency chain until we either find
975 // a dependency we can use, or we can't walk any further
976 while (Instruction *DepInst = dep.getInst()) {
977 // ... that depends on a store ...
978 if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
979 if (S->getPointerOperand() == pointer) {
981 MD.removeInstruction(L);
983 L->replaceAllUsesWith(S->getOperand(0));
984 toErase.push_back(L);
989 // Whether we removed it or not, we can't
992 } else if (!isa<LoadInst>(DepInst)) {
993 // Only want to handle loads below.
996 // If we don't depend on a store, and we haven't
997 // been loaded before, bail.
999 } else if (DepInst == last) {
1001 MD.removeInstruction(L);
1003 L->replaceAllUsesWith(last);
1004 toErase.push_back(L);
1009 dep = MD.getDependencyFrom(L, DepInst, DepInst->getParent());
1013 // If this load really doesn't depend on anything, then we must be loading an
1014 // undef value. This can happen when loading for a fresh allocation with no
1015 // intervening stores, for example.
1017 // If this load depends directly on an allocation, there isn't
1018 // anything stored there; therefore, we can optimize this load
1020 MD.removeInstruction(L);
1021 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1022 toErase.push_back(L);
1033 Value* GVN::lookupNumber(BasicBlock* BB, uint32_t num) {
1034 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1035 if (I == localAvail.end())
1038 ValueNumberScope* locals = I->second;
1041 DenseMap<uint32_t, Value*>::iterator I = locals->table.find(num);
1042 if (I != locals->table.end())
1045 locals = locals->parent;
1051 /// processInstruction - When calculating availability, handle an instruction
1052 /// by inserting it into the appropriate sets
1053 bool GVN::processInstruction(Instruction *I,
1054 DenseMap<Value*, LoadInst*> &lastSeenLoad,
1055 SmallVectorImpl<Instruction*> &toErase) {
1056 if (LoadInst* L = dyn_cast<LoadInst>(I)) {
1057 bool changed = processLoad(L, lastSeenLoad, toErase);
1060 unsigned num = VN.lookup_or_add(L);
1061 localAvail[I->getParent()]->table.insert(std::make_pair(num, L));
1067 uint32_t nextNum = VN.getNextUnusedValueNumber();
1068 unsigned num = VN.lookup_or_add(I);
1070 // Allocations are always uniquely numbered, so we can save time and memory
1071 // by fast failing them.
1072 if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1073 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1077 // Collapse PHI nodes
1078 if (PHINode* p = dyn_cast<PHINode>(I)) {
1079 Value* constVal = CollapsePhi(p);
1082 for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
1084 if (PI->second.count(p))
1085 PI->second.erase(p);
1087 p->replaceAllUsesWith(constVal);
1088 toErase.push_back(p);
1090 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1093 // If the number we were assigned was a brand new VN, then we don't
1094 // need to do a lookup to see if the number already exists
1095 // somewhere in the domtree: it can't!
1096 } else if (num == nextNum) {
1097 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1099 // Perform value-number based elimination
1100 } else if (Value* repl = lookupNumber(I->getParent(), num)) {
1102 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
1103 MD.removeInstruction(I);
1106 I->replaceAllUsesWith(repl);
1107 toErase.push_back(I);
1110 localAvail[I->getParent()]->table.insert(std::make_pair(num, I));
1116 // GVN::runOnFunction - This is the main transformation entry point for a
1119 bool GVN::runOnFunction(Function& F) {
1120 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1121 VN.setMemDep(&getAnalysis<MemoryDependenceAnalysis>());
1122 VN.setDomTree(&getAnalysis<DominatorTree>());
1124 bool changed = false;
1125 bool shouldContinue = true;
1127 // Merge unconditional branches, allowing PRE to catch more
1128 // optimization opportunities.
1129 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1130 BasicBlock* BB = FI;
1132 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1133 if (removedBlock) NumGVNBlocks++;
1135 changed |= removedBlock;
1138 while (shouldContinue) {
1139 shouldContinue = iterateOnFunction(F);
1140 changed |= shouldContinue;
1144 bool PREChanged = true;
1145 while (PREChanged) {
1146 PREChanged = performPRE(F);
1147 changed |= PREChanged;
1151 cleanupGlobalSets();
1157 bool GVN::processBlock(DomTreeNode* DTN) {
1158 BasicBlock* BB = DTN->getBlock();
1160 SmallVector<Instruction*, 8> toErase;
1161 DenseMap<Value*, LoadInst*> lastSeenLoad;
1162 bool changed_function = false;
1166 new ValueNumberScope(localAvail[DTN->getIDom()->getBlock()]);
1168 localAvail[BB] = new ValueNumberScope(0);
1170 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1172 changed_function |= processInstruction(BI, lastSeenLoad, toErase);
1173 if (toErase.empty()) {
1178 // If we need some instructions deleted, do it now.
1179 NumGVNInstr += toErase.size();
1181 // Avoid iterator invalidation.
1182 bool AtStart = BI == BB->begin();
1186 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1187 E = toErase.end(); I != E; ++I)
1188 (*I)->eraseFromParent();
1198 return changed_function;
1201 /// performPRE - Perform a purely local form of PRE that looks for diamond
1202 /// control flow patterns and attempts to perform simple PRE at the join point.
1203 bool GVN::performPRE(Function& F) {
1204 bool changed = false;
1205 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1206 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1207 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1208 BasicBlock* CurrentBlock = *DI;
1210 // Nothing to PRE in the entry block.
1211 if (CurrentBlock == &F.getEntryBlock()) continue;
1213 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1214 BE = CurrentBlock->end(); BI != BE; ) {
1215 if (isa<AllocationInst>(BI) || isa<TerminatorInst>(BI) ||
1216 isa<PHINode>(BI) || BI->mayReadFromMemory() ||
1217 BI->mayWriteToMemory()) {
1222 uint32_t valno = VN.lookup(BI);
1224 // Look for the predecessors for PRE opportunities. We're
1225 // only trying to solve the basic diamond case, where
1226 // a value is computed in the successor and one predecessor,
1227 // but not the other. We also explicitly disallow cases
1228 // where the successor is its own predecessor, because they're
1229 // more complicated to get right.
1230 unsigned numWith = 0;
1231 unsigned numWithout = 0;
1232 BasicBlock* PREPred = 0;
1233 DenseMap<BasicBlock*, Value*> predMap;
1234 for (pred_iterator PI = pred_begin(CurrentBlock),
1235 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1236 // We're not interested in PRE where the block is its
1237 // own predecessor, on in blocks with predecessors
1238 // that are not reachable.
1239 if (*PI == CurrentBlock) {
1242 } else if (!localAvail.count(*PI)) {
1247 DenseMap<uint32_t, Value*>::iterator predV =
1248 localAvail[*PI]->table.find(valno);
1249 if (predV == localAvail[*PI]->table.end()) {
1252 } else if (predV->second == BI) {
1255 predMap[*PI] = predV->second;
1260 // Don't do PRE when it might increase code size, i.e. when
1261 // we would need to insert instructions in more than one pred.
1262 if (numWithout != 1 || numWith == 0) {
1267 // We can't do PRE safely on a critical edge, so instead we schedule
1268 // the edge to be split and perform the PRE the next time we iterate
1270 unsigned succNum = 0;
1271 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1273 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1278 if (isCriticalEdge(PREPred->getTerminator(), succNum)) {
1279 toSplit.push_back(std::make_pair(PREPred->getTerminator(), succNum));
1285 // Instantiate the expression the in predecessor that lacked it.
1286 // Because we are going top-down through the block, all value numbers
1287 // will be available in the predecessor by the time we need them. Any
1288 // that weren't original present will have been instantiated earlier
1290 Instruction* PREInstr = BI->clone();
1291 bool success = true;
1292 for (unsigned i = 0; i < BI->getNumOperands(); ++i) {
1293 Value* op = BI->getOperand(i);
1294 if (isa<Argument>(op) || isa<Constant>(op) || isa<GlobalValue>(op))
1295 PREInstr->setOperand(i, op);
1297 Value* V = lookupNumber(PREPred, VN.lookup(op));
1302 PREInstr->setOperand(i, V);
1306 // Fail out if we encounter an operand that is not available in
1307 // the PRE predecessor. This is typically because of loads which
1308 // are not value numbered precisely.
1315 PREInstr->insertBefore(PREPred->getTerminator());
1316 PREInstr->setName(BI->getName() + ".pre");
1317 predMap[PREPred] = PREInstr;
1318 VN.add(PREInstr, valno);
1321 // Update the availability map to include the new instruction.
1322 localAvail[PREPred]->table.insert(std::make_pair(valno, PREInstr));
1324 // Create a PHI to make the value available in this block.
1325 PHINode* Phi = PHINode::Create(BI->getType(),
1326 BI->getName() + ".pre-phi",
1327 CurrentBlock->begin());
1328 for (pred_iterator PI = pred_begin(CurrentBlock),
1329 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1330 Phi->addIncoming(predMap[*PI], *PI);
1333 localAvail[CurrentBlock]->table[valno] = Phi;
1335 BI->replaceAllUsesWith(Phi);
1338 Instruction* erase = BI;
1340 erase->eraseFromParent();
1346 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1347 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1348 SplitCriticalEdge(I->first, I->second, this);
1350 return changed || toSplit.size();
1353 // iterateOnFunction - Executes one iteration of GVN
1354 bool GVN::iterateOnFunction(Function &F) {
1355 DominatorTree &DT = getAnalysis<DominatorTree>();
1357 cleanupGlobalSets();
1359 // Top-down walk of the dominator tree
1360 bool changed = false;
1361 for (df_iterator<DomTreeNode*> DI = df_begin(DT.getRootNode()),
1362 DE = df_end(DT.getRootNode()); DI != DE; ++DI)
1363 changed |= processBlock(*DI);
1368 void GVN::cleanupGlobalSets() {
1372 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1373 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)