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/LLVMContext.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Value.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/DepthFirstIterator.h"
30 #include "llvm/ADT/PostOrderIterator.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/Dominators.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/MallocHelper.h"
37 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/GetElementPtrTypeIterator.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/SSAUpdater.h"
51 STATISTIC(NumGVNInstr, "Number of instructions deleted");
52 STATISTIC(NumGVNLoad, "Number of loads deleted");
53 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
54 STATISTIC(NumGVNBlocks, "Number of blocks merged");
55 STATISTIC(NumPRELoad, "Number of loads PRE'd");
57 static cl::opt<bool> EnablePRE("enable-pre",
58 cl::init(true), cl::Hidden);
59 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
61 //===----------------------------------------------------------------------===//
63 //===----------------------------------------------------------------------===//
65 /// This class holds the mapping between values and value numbers. It is used
66 /// as an efficient mechanism to determine the expression-wise equivalence of
70 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
71 UDIV, SDIV, FDIV, UREM, SREM,
72 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
73 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
74 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
75 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
76 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
77 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
78 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
79 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
80 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
83 ExpressionOpcode opcode;
88 SmallVector<uint32_t, 4> varargs;
92 Expression(ExpressionOpcode o) : opcode(o) { }
94 bool operator==(const Expression &other) const {
95 if (opcode != other.opcode)
97 else if (opcode == EMPTY || opcode == TOMBSTONE)
99 else if (type != other.type)
101 else if (function != other.function)
103 else if (firstVN != other.firstVN)
105 else if (secondVN != other.secondVN)
107 else if (thirdVN != other.thirdVN)
110 if (varargs.size() != other.varargs.size())
113 for (size_t i = 0; i < varargs.size(); ++i)
114 if (varargs[i] != other.varargs[i])
121 bool operator!=(const Expression &other) const {
122 return !(*this == other);
128 DenseMap<Value*, uint32_t> valueNumbering;
129 DenseMap<Expression, uint32_t> expressionNumbering;
131 MemoryDependenceAnalysis* MD;
134 uint32_t nextValueNumber;
136 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
137 Expression::ExpressionOpcode getOpcode(CmpInst* C);
138 Expression::ExpressionOpcode getOpcode(CastInst* C);
139 Expression create_expression(BinaryOperator* BO);
140 Expression create_expression(CmpInst* C);
141 Expression create_expression(ShuffleVectorInst* V);
142 Expression create_expression(ExtractElementInst* C);
143 Expression create_expression(InsertElementInst* V);
144 Expression create_expression(SelectInst* V);
145 Expression create_expression(CastInst* C);
146 Expression create_expression(GetElementPtrInst* G);
147 Expression create_expression(CallInst* C);
148 Expression create_expression(Constant* C);
150 ValueTable() : nextValueNumber(1) { }
151 uint32_t lookup_or_add(Value *V);
152 uint32_t lookup(Value *V) const;
153 void add(Value *V, uint32_t num);
155 void erase(Value *v);
157 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
158 AliasAnalysis *getAliasAnalysis() const { return AA; }
159 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
160 void setDomTree(DominatorTree* D) { DT = D; }
161 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
162 void verifyRemoved(const Value *) const;
167 template <> struct DenseMapInfo<Expression> {
168 static inline Expression getEmptyKey() {
169 return Expression(Expression::EMPTY);
172 static inline Expression getTombstoneKey() {
173 return Expression(Expression::TOMBSTONE);
176 static unsigned getHashValue(const Expression e) {
177 unsigned hash = e.opcode;
179 hash = e.firstVN + hash * 37;
180 hash = e.secondVN + hash * 37;
181 hash = e.thirdVN + hash * 37;
183 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
184 (unsigned)((uintptr_t)e.type >> 9)) +
187 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
188 E = e.varargs.end(); I != E; ++I)
189 hash = *I + hash * 37;
191 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
192 (unsigned)((uintptr_t)e.function >> 9)) +
197 static bool isEqual(const Expression &LHS, const Expression &RHS) {
200 static bool isPod() { return true; }
204 //===----------------------------------------------------------------------===//
205 // ValueTable Internal Functions
206 //===----------------------------------------------------------------------===//
207 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
208 switch(BO->getOpcode()) {
209 default: // THIS SHOULD NEVER HAPPEN
210 llvm_unreachable("Binary operator with unknown opcode?");
211 case Instruction::Add: return Expression::ADD;
212 case Instruction::FAdd: return Expression::FADD;
213 case Instruction::Sub: return Expression::SUB;
214 case Instruction::FSub: return Expression::FSUB;
215 case Instruction::Mul: return Expression::MUL;
216 case Instruction::FMul: return Expression::FMUL;
217 case Instruction::UDiv: return Expression::UDIV;
218 case Instruction::SDiv: return Expression::SDIV;
219 case Instruction::FDiv: return Expression::FDIV;
220 case Instruction::URem: return Expression::UREM;
221 case Instruction::SRem: return Expression::SREM;
222 case Instruction::FRem: return Expression::FREM;
223 case Instruction::Shl: return Expression::SHL;
224 case Instruction::LShr: return Expression::LSHR;
225 case Instruction::AShr: return Expression::ASHR;
226 case Instruction::And: return Expression::AND;
227 case Instruction::Or: return Expression::OR;
228 case Instruction::Xor: return Expression::XOR;
232 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
233 if (isa<ICmpInst>(C)) {
234 switch (C->getPredicate()) {
235 default: // THIS SHOULD NEVER HAPPEN
236 llvm_unreachable("Comparison with unknown predicate?");
237 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
238 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
239 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
240 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
241 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
242 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
243 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
244 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
245 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
246 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
249 switch (C->getPredicate()) {
250 default: // THIS SHOULD NEVER HAPPEN
251 llvm_unreachable("Comparison with unknown predicate?");
252 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
253 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
254 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
255 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
256 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
257 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
258 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
259 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
260 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
261 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
262 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
263 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
264 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
265 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
270 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
271 switch(C->getOpcode()) {
272 default: // THIS SHOULD NEVER HAPPEN
273 llvm_unreachable("Cast operator with unknown opcode?");
274 case Instruction::Trunc: return Expression::TRUNC;
275 case Instruction::ZExt: return Expression::ZEXT;
276 case Instruction::SExt: return Expression::SEXT;
277 case Instruction::FPToUI: return Expression::FPTOUI;
278 case Instruction::FPToSI: return Expression::FPTOSI;
279 case Instruction::UIToFP: return Expression::UITOFP;
280 case Instruction::SIToFP: return Expression::SITOFP;
281 case Instruction::FPTrunc: return Expression::FPTRUNC;
282 case Instruction::FPExt: return Expression::FPEXT;
283 case Instruction::PtrToInt: return Expression::PTRTOINT;
284 case Instruction::IntToPtr: return Expression::INTTOPTR;
285 case Instruction::BitCast: return Expression::BITCAST;
289 Expression ValueTable::create_expression(CallInst* C) {
292 e.type = C->getType();
296 e.function = C->getCalledFunction();
297 e.opcode = Expression::CALL;
299 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
301 e.varargs.push_back(lookup_or_add(*I));
306 Expression ValueTable::create_expression(BinaryOperator* BO) {
309 e.firstVN = lookup_or_add(BO->getOperand(0));
310 e.secondVN = lookup_or_add(BO->getOperand(1));
313 e.type = BO->getType();
314 e.opcode = getOpcode(BO);
319 Expression ValueTable::create_expression(CmpInst* C) {
322 e.firstVN = lookup_or_add(C->getOperand(0));
323 e.secondVN = lookup_or_add(C->getOperand(1));
326 e.type = C->getType();
327 e.opcode = getOpcode(C);
332 Expression ValueTable::create_expression(CastInst* C) {
335 e.firstVN = lookup_or_add(C->getOperand(0));
339 e.type = C->getType();
340 e.opcode = getOpcode(C);
345 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
348 e.firstVN = lookup_or_add(S->getOperand(0));
349 e.secondVN = lookup_or_add(S->getOperand(1));
350 e.thirdVN = lookup_or_add(S->getOperand(2));
352 e.type = S->getType();
353 e.opcode = Expression::SHUFFLE;
358 Expression ValueTable::create_expression(ExtractElementInst* E) {
361 e.firstVN = lookup_or_add(E->getOperand(0));
362 e.secondVN = lookup_or_add(E->getOperand(1));
365 e.type = E->getType();
366 e.opcode = Expression::EXTRACT;
371 Expression ValueTable::create_expression(InsertElementInst* I) {
374 e.firstVN = lookup_or_add(I->getOperand(0));
375 e.secondVN = lookup_or_add(I->getOperand(1));
376 e.thirdVN = lookup_or_add(I->getOperand(2));
378 e.type = I->getType();
379 e.opcode = Expression::INSERT;
384 Expression ValueTable::create_expression(SelectInst* I) {
387 e.firstVN = lookup_or_add(I->getCondition());
388 e.secondVN = lookup_or_add(I->getTrueValue());
389 e.thirdVN = lookup_or_add(I->getFalseValue());
391 e.type = I->getType();
392 e.opcode = Expression::SELECT;
397 Expression ValueTable::create_expression(GetElementPtrInst* G) {
400 e.firstVN = lookup_or_add(G->getPointerOperand());
404 e.type = G->getType();
405 e.opcode = Expression::GEP;
407 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
409 e.varargs.push_back(lookup_or_add(*I));
414 //===----------------------------------------------------------------------===//
415 // ValueTable External Functions
416 //===----------------------------------------------------------------------===//
418 /// add - Insert a value into the table with a specified value number.
419 void ValueTable::add(Value *V, uint32_t num) {
420 valueNumbering.insert(std::make_pair(V, num));
423 /// lookup_or_add - Returns the value number for the specified value, assigning
424 /// it a new number if it did not have one before.
425 uint32_t ValueTable::lookup_or_add(Value *V) {
426 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
427 if (VI != valueNumbering.end())
430 if (CallInst* C = dyn_cast<CallInst>(V)) {
431 if (AA->doesNotAccessMemory(C)) {
432 Expression e = create_expression(C);
434 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
435 if (EI != expressionNumbering.end()) {
436 valueNumbering.insert(std::make_pair(V, EI->second));
439 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
440 valueNumbering.insert(std::make_pair(V, nextValueNumber));
442 return nextValueNumber++;
444 } else if (AA->onlyReadsMemory(C)) {
445 Expression e = create_expression(C);
447 if (expressionNumbering.find(e) == expressionNumbering.end()) {
448 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
449 valueNumbering.insert(std::make_pair(V, nextValueNumber));
450 return nextValueNumber++;
453 MemDepResult local_dep = MD->getDependency(C);
455 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
456 valueNumbering.insert(std::make_pair(V, nextValueNumber));
457 return nextValueNumber++;
460 if (local_dep.isDef()) {
461 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
463 if (local_cdep->getNumOperands() != C->getNumOperands()) {
464 valueNumbering.insert(std::make_pair(V, nextValueNumber));
465 return nextValueNumber++;
468 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
469 uint32_t c_vn = lookup_or_add(C->getOperand(i));
470 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
472 valueNumbering.insert(std::make_pair(V, nextValueNumber));
473 return nextValueNumber++;
477 uint32_t v = lookup_or_add(local_cdep);
478 valueNumbering.insert(std::make_pair(V, v));
483 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
484 MD->getNonLocalCallDependency(CallSite(C));
485 // FIXME: call/call dependencies for readonly calls should return def, not
486 // clobber! Move the checking logic to MemDep!
489 // Check to see if we have a single dominating call instruction that is
491 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
492 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
493 // Ignore non-local dependencies.
494 if (I->second.isNonLocal())
497 // We don't handle non-depedencies. If we already have a call, reject
498 // instruction dependencies.
499 if (I->second.isClobber() || cdep != 0) {
504 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
505 // FIXME: All duplicated with non-local case.
506 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
507 cdep = NonLocalDepCall;
516 valueNumbering.insert(std::make_pair(V, nextValueNumber));
517 return nextValueNumber++;
520 if (cdep->getNumOperands() != C->getNumOperands()) {
521 valueNumbering.insert(std::make_pair(V, nextValueNumber));
522 return nextValueNumber++;
524 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
525 uint32_t c_vn = lookup_or_add(C->getOperand(i));
526 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
528 valueNumbering.insert(std::make_pair(V, nextValueNumber));
529 return nextValueNumber++;
533 uint32_t v = lookup_or_add(cdep);
534 valueNumbering.insert(std::make_pair(V, v));
538 valueNumbering.insert(std::make_pair(V, nextValueNumber));
539 return nextValueNumber++;
541 } else if (BinaryOperator* BO = dyn_cast<BinaryOperator>(V)) {
542 Expression e = create_expression(BO);
544 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
545 if (EI != expressionNumbering.end()) {
546 valueNumbering.insert(std::make_pair(V, EI->second));
549 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
550 valueNumbering.insert(std::make_pair(V, nextValueNumber));
552 return nextValueNumber++;
554 } else if (CmpInst* C = dyn_cast<CmpInst>(V)) {
555 Expression e = create_expression(C);
557 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
558 if (EI != expressionNumbering.end()) {
559 valueNumbering.insert(std::make_pair(V, EI->second));
562 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
563 valueNumbering.insert(std::make_pair(V, nextValueNumber));
565 return nextValueNumber++;
567 } else if (ShuffleVectorInst* U = dyn_cast<ShuffleVectorInst>(V)) {
568 Expression e = create_expression(U);
570 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
571 if (EI != expressionNumbering.end()) {
572 valueNumbering.insert(std::make_pair(V, EI->second));
575 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
576 valueNumbering.insert(std::make_pair(V, nextValueNumber));
578 return nextValueNumber++;
580 } else if (ExtractElementInst* U = dyn_cast<ExtractElementInst>(V)) {
581 Expression e = create_expression(U);
583 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
584 if (EI != expressionNumbering.end()) {
585 valueNumbering.insert(std::make_pair(V, EI->second));
588 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
589 valueNumbering.insert(std::make_pair(V, nextValueNumber));
591 return nextValueNumber++;
593 } else if (InsertElementInst* U = dyn_cast<InsertElementInst>(V)) {
594 Expression e = create_expression(U);
596 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
597 if (EI != expressionNumbering.end()) {
598 valueNumbering.insert(std::make_pair(V, EI->second));
601 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
602 valueNumbering.insert(std::make_pair(V, nextValueNumber));
604 return nextValueNumber++;
606 } else if (SelectInst* U = dyn_cast<SelectInst>(V)) {
607 Expression e = create_expression(U);
609 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
610 if (EI != expressionNumbering.end()) {
611 valueNumbering.insert(std::make_pair(V, EI->second));
614 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
615 valueNumbering.insert(std::make_pair(V, nextValueNumber));
617 return nextValueNumber++;
619 } else if (CastInst* U = dyn_cast<CastInst>(V)) {
620 Expression e = create_expression(U);
622 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
623 if (EI != expressionNumbering.end()) {
624 valueNumbering.insert(std::make_pair(V, EI->second));
627 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
628 valueNumbering.insert(std::make_pair(V, nextValueNumber));
630 return nextValueNumber++;
632 } else if (GetElementPtrInst* U = dyn_cast<GetElementPtrInst>(V)) {
633 Expression e = create_expression(U);
635 DenseMap<Expression, uint32_t>::iterator EI = expressionNumbering.find(e);
636 if (EI != expressionNumbering.end()) {
637 valueNumbering.insert(std::make_pair(V, EI->second));
640 expressionNumbering.insert(std::make_pair(e, nextValueNumber));
641 valueNumbering.insert(std::make_pair(V, nextValueNumber));
643 return nextValueNumber++;
646 valueNumbering.insert(std::make_pair(V, nextValueNumber));
647 return nextValueNumber++;
651 /// lookup - Returns the value number of the specified value. Fails if
652 /// the value has not yet been numbered.
653 uint32_t ValueTable::lookup(Value *V) const {
654 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
655 assert(VI != valueNumbering.end() && "Value not numbered?");
659 /// clear - Remove all entries from the ValueTable
660 void ValueTable::clear() {
661 valueNumbering.clear();
662 expressionNumbering.clear();
666 /// erase - Remove a value from the value numbering
667 void ValueTable::erase(Value *V) {
668 valueNumbering.erase(V);
671 /// verifyRemoved - Verify that the value is removed from all internal data
673 void ValueTable::verifyRemoved(const Value *V) const {
674 for (DenseMap<Value*, uint32_t>::iterator
675 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
676 assert(I->first != V && "Inst still occurs in value numbering map!");
680 //===----------------------------------------------------------------------===//
682 //===----------------------------------------------------------------------===//
685 struct ValueNumberScope {
686 ValueNumberScope* parent;
687 DenseMap<uint32_t, Value*> table;
689 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
695 class GVN : public FunctionPass {
696 bool runOnFunction(Function &F);
698 static char ID; // Pass identification, replacement for typeid
699 GVN() : FunctionPass(&ID) { }
702 MemoryDependenceAnalysis *MD;
706 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
708 // This transformation requires dominator postdominator info
709 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
710 AU.addRequired<DominatorTree>();
711 AU.addRequired<MemoryDependenceAnalysis>();
712 AU.addRequired<AliasAnalysis>();
714 AU.addPreserved<DominatorTree>();
715 AU.addPreserved<AliasAnalysis>();
719 // FIXME: eliminate or document these better
720 bool processLoad(LoadInst* L,
721 SmallVectorImpl<Instruction*> &toErase);
722 bool processInstruction(Instruction *I,
723 SmallVectorImpl<Instruction*> &toErase);
724 bool processNonLocalLoad(LoadInst* L,
725 SmallVectorImpl<Instruction*> &toErase);
726 bool processBlock(BasicBlock *BB);
727 void dump(DenseMap<uint32_t, Value*>& d);
728 bool iterateOnFunction(Function &F);
729 Value *CollapsePhi(PHINode* p);
730 bool performPRE(Function& F);
731 Value *lookupNumber(BasicBlock *BB, uint32_t num);
732 void cleanupGlobalSets();
733 void verifyRemoved(const Instruction *I) const;
739 // createGVNPass - The public interface to this file...
740 FunctionPass *llvm::createGVNPass() { return new GVN(); }
742 static RegisterPass<GVN> X("gvn",
743 "Global Value Numbering");
745 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
747 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
748 E = d.end(); I != E; ++I) {
749 printf("%d\n", I->first);
755 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
756 if (!isa<PHINode>(inst))
759 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
761 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
762 if (use_phi->getParent() == inst->getParent())
768 Value *GVN::CollapsePhi(PHINode *PN) {
769 Value *ConstVal = PN->hasConstantValue(DT);
770 if (!ConstVal) return 0;
772 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
776 if (DT->dominates(Inst, PN))
777 if (isSafeReplacement(PN, Inst))
782 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
783 /// we're analyzing is fully available in the specified block. As we go, keep
784 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
785 /// map is actually a tri-state map with the following values:
786 /// 0) we know the block *is not* fully available.
787 /// 1) we know the block *is* fully available.
788 /// 2) we do not know whether the block is fully available or not, but we are
789 /// currently speculating that it will be.
790 /// 3) we are speculating for this block and have used that to speculate for
792 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
793 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
794 // Optimistically assume that the block is fully available and check to see
795 // if we already know about this block in one lookup.
796 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
797 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
799 // If the entry already existed for this block, return the precomputed value.
801 // If this is a speculative "available" value, mark it as being used for
802 // speculation of other blocks.
803 if (IV.first->second == 2)
804 IV.first->second = 3;
805 return IV.first->second != 0;
808 // Otherwise, see if it is fully available in all predecessors.
809 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
811 // If this block has no predecessors, it isn't live-in here.
813 goto SpeculationFailure;
815 for (; PI != PE; ++PI)
816 // If the value isn't fully available in one of our predecessors, then it
817 // isn't fully available in this block either. Undo our previous
818 // optimistic assumption and bail out.
819 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
820 goto SpeculationFailure;
824 // SpeculationFailure - If we get here, we found out that this is not, after
825 // all, a fully-available block. We have a problem if we speculated on this and
826 // used the speculation to mark other blocks as available.
828 char &BBVal = FullyAvailableBlocks[BB];
830 // If we didn't speculate on this, just return with it set to false.
836 // If we did speculate on this value, we could have blocks set to 1 that are
837 // incorrect. Walk the (transitive) successors of this block and mark them as
839 SmallVector<BasicBlock*, 32> BBWorklist;
840 BBWorklist.push_back(BB);
842 while (!BBWorklist.empty()) {
843 BasicBlock *Entry = BBWorklist.pop_back_val();
844 // Note that this sets blocks to 0 (unavailable) if they happen to not
845 // already be in FullyAvailableBlocks. This is safe.
846 char &EntryVal = FullyAvailableBlocks[Entry];
847 if (EntryVal == 0) continue; // Already unavailable.
849 // Mark as unavailable.
852 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
853 BBWorklist.push_back(*I);
860 /// CanCoerceMustAliasedValueToLoad - Return true if
861 /// CoerceAvailableValueToLoadType will succeed.
862 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
864 const TargetData &TD) {
865 // If the loaded or stored value is an first class array or struct, don't try
866 // to transform them. We need to be able to bitcast to integer.
867 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
868 isa<StructType>(StoredVal->getType()) ||
869 isa<ArrayType>(StoredVal->getType()))
872 // The store has to be at least as big as the load.
873 if (TD.getTypeSizeInBits(StoredVal->getType()) <
874 TD.getTypeSizeInBits(LoadTy))
881 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
882 /// then a load from a must-aliased pointer of a different type, try to coerce
883 /// the stored value. LoadedTy is the type of the load we want to replace and
884 /// InsertPt is the place to insert new instructions.
886 /// If we can't do it, return null.
887 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
888 const Type *LoadedTy,
889 Instruction *InsertPt,
890 const TargetData &TD) {
891 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
894 const Type *StoredValTy = StoredVal->getType();
896 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
897 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
899 // If the store and reload are the same size, we can always reuse it.
900 if (StoreSize == LoadSize) {
901 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
902 // Pointer to Pointer -> use bitcast.
903 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
906 // Convert source pointers to integers, which can be bitcast.
907 if (isa<PointerType>(StoredValTy)) {
908 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
909 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
912 const Type *TypeToCastTo = LoadedTy;
913 if (isa<PointerType>(TypeToCastTo))
914 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
916 if (StoredValTy != TypeToCastTo)
917 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
919 // Cast to pointer if the load needs a pointer type.
920 if (isa<PointerType>(LoadedTy))
921 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
926 // If the loaded value is smaller than the available value, then we can
927 // extract out a piece from it. If the available value is too small, then we
928 // can't do anything.
929 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
931 // Convert source pointers to integers, which can be manipulated.
932 if (isa<PointerType>(StoredValTy)) {
933 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
934 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
937 // Convert vectors and fp to integer, which can be manipulated.
938 if (!isa<IntegerType>(StoredValTy)) {
939 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
940 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
943 // If this is a big-endian system, we need to shift the value down to the low
944 // bits so that a truncate will work.
945 if (TD.isBigEndian()) {
946 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
947 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
950 // Truncate the integer to the right size now.
951 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
952 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
954 if (LoadedTy == NewIntTy)
957 // If the result is a pointer, inttoptr.
958 if (isa<PointerType>(LoadedTy))
959 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
961 // Otherwise, bitcast.
962 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
965 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
966 /// be expressed as a base pointer plus a constant offset. Return the base and
967 /// offset to the caller.
968 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
969 const TargetData &TD) {
970 Operator *PtrOp = dyn_cast<Operator>(Ptr);
971 if (PtrOp == 0) return Ptr;
973 // Just look through bitcasts.
974 if (PtrOp->getOpcode() == Instruction::BitCast)
975 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
977 // If this is a GEP with constant indices, we can look through it.
978 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
979 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
981 gep_type_iterator GTI = gep_type_begin(GEP);
982 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
984 ConstantInt *OpC = cast<ConstantInt>(*I);
985 if (OpC->isZero()) continue;
987 // Handle a struct and array indices which add their offset to the pointer.
988 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
989 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
991 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
992 Offset += OpC->getSExtValue()*Size;
996 // Re-sign extend from the pointer size if needed to get overflow edge cases
998 unsigned PtrSize = TD.getPointerSizeInBits();
1000 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1002 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1006 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1007 /// memdep query of a load that ends up being a clobbering store. This means
1008 /// that the store *may* provide bits used by the load but we can't be sure
1009 /// because the pointers don't mustalias. Check this case to see if there is
1010 /// anything more we can do before we give up. This returns -1 if we have to
1011 /// give up, or a byte number in the stored value of the piece that feeds the
1013 static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI,
1014 const TargetData &TD) {
1015 // If the loaded or stored value is an first class array or struct, don't try
1016 // to transform them. We need to be able to bitcast to integer.
1017 if (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()) ||
1018 isa<StructType>(DepSI->getOperand(0)->getType()) ||
1019 isa<ArrayType>(DepSI->getOperand(0)->getType()))
1022 int64_t StoreOffset = 0, LoadOffset = 0;
1024 GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
1026 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
1027 if (StoreBase != LoadBase)
1030 // If the load and store are to the exact same address, they should have been
1031 // a must alias. AA must have gotten confused.
1032 // FIXME: Study to see if/when this happens.
1033 if (LoadOffset == StoreOffset) {
1035 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1036 << "Base = " << *StoreBase << "\n"
1037 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
1038 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
1039 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1040 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1041 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1047 // If the load and store don't overlap at all, the store doesn't provide
1048 // anything to the load. In this case, they really don't alias at all, AA
1049 // must have gotten confused.
1050 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1051 // remove this check, as it is duplicated with what we have below.
1052 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1053 uint64_t LoadSize = TD.getTypeSizeInBits(L->getType());
1055 if ((StoreSize & 7) | (LoadSize & 7))
1057 StoreSize >>= 3; // Convert to bytes.
1061 bool isAAFailure = false;
1062 if (StoreOffset < LoadOffset) {
1063 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1065 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1069 errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1070 << "Base = " << *StoreBase << "\n"
1071 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
1072 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
1073 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1074 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1075 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1081 // If the Load isn't completely contained within the stored bits, we don't
1082 // have all the bits to feed it. We could do something crazy in the future
1083 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1085 if (StoreOffset > LoadOffset ||
1086 StoreOffset+StoreSize < LoadOffset+LoadSize)
1089 // Okay, we can do this transformation. Return the number of bytes into the
1090 // store that the load is.
1091 return LoadOffset-StoreOffset;
1095 /// GetStoreValueForLoad - This function is called when we have a
1096 /// memdep query of a load that ends up being a clobbering store. This means
1097 /// that the store *may* provide bits used by the load but we can't be sure
1098 /// because the pointers don't mustalias. Check this case to see if there is
1099 /// anything more we can do before we give up.
1100 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1102 Instruction *InsertPt, const TargetData &TD){
1103 LLVMContext &Ctx = SrcVal->getType()->getContext();
1105 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1106 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1109 // Compute which bits of the stored value are being used by the load. Convert
1110 // to an integer type to start with.
1111 if (isa<PointerType>(SrcVal->getType()))
1112 SrcVal = new PtrToIntInst(SrcVal, TD.getIntPtrType(Ctx), "tmp", InsertPt);
1113 if (!isa<IntegerType>(SrcVal->getType()))
1114 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1117 // Shift the bits to the least significant depending on endianness.
1119 if (TD.isLittleEndian()) {
1120 ShiftAmt = Offset*8;
1122 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1126 SrcVal = BinaryOperator::CreateLShr(SrcVal,
1127 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
1129 if (LoadSize != StoreSize)
1130 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1133 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1136 struct AvailableValueInBlock {
1137 /// BB - The basic block in question.
1139 /// V - The value that is live out of the block.
1141 /// Offset - The byte offset in V that is interesting for the load query.
1144 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1145 unsigned Offset = 0) {
1146 AvailableValueInBlock Res;
1149 Res.Offset = Offset;
1154 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1155 /// construct SSA form, allowing us to eliminate LI. This returns the value
1156 /// that should be used at LI's definition site.
1157 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1158 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1159 const TargetData *TD,
1160 AliasAnalysis *AA) {
1161 SmallVector<PHINode*, 8> NewPHIs;
1162 SSAUpdater SSAUpdate(&NewPHIs);
1163 SSAUpdate.Initialize(LI);
1165 const Type *LoadTy = LI->getType();
1167 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1168 BasicBlock *BB = ValuesPerBlock[i].BB;
1169 Value *AvailableVal = ValuesPerBlock[i].V;
1170 unsigned Offset = ValuesPerBlock[i].Offset;
1172 if (SSAUpdate.HasValueForBlock(BB))
1175 if (AvailableVal->getType() != LoadTy) {
1176 assert(TD && "Need target data to handle type mismatch case");
1177 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy,
1178 BB->getTerminator(), *TD);
1181 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
1182 << *ValuesPerBlock[i].V << '\n'
1183 << *AvailableVal << '\n' << "\n\n\n");
1187 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
1188 << *ValuesPerBlock[i].V << '\n'
1189 << *AvailableVal << '\n' << "\n\n\n");
1192 SSAUpdate.AddAvailableValue(BB, AvailableVal);
1195 // Perform PHI construction.
1196 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1198 // If new PHI nodes were created, notify alias analysis.
1199 if (isa<PointerType>(V->getType()))
1200 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1201 AA->copyValue(LI, NewPHIs[i]);
1206 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1207 /// non-local by performing PHI construction.
1208 bool GVN::processNonLocalLoad(LoadInst *LI,
1209 SmallVectorImpl<Instruction*> &toErase) {
1210 // Find the non-local dependencies of the load.
1211 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
1212 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1214 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
1215 // << Deps.size() << *LI << '\n');
1217 // If we had to process more than one hundred blocks to find the
1218 // dependencies, this load isn't worth worrying about. Optimizing
1219 // it will be too expensive.
1220 if (Deps.size() > 100)
1223 // If we had a phi translation failure, we'll have a single entry which is a
1224 // clobber in the current block. Reject this early.
1225 if (Deps.size() == 1 && Deps[0].second.isClobber()) {
1227 errs() << "GVN: non-local load ";
1228 WriteAsOperand(errs(), LI);
1229 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n';
1234 // Filter out useless results (non-locals, etc). Keep track of the blocks
1235 // where we have a value available in repl, also keep track of whether we see
1236 // dependencies that produce an unknown value for the load (such as a call
1237 // that could potentially clobber the load).
1238 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1239 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1241 const TargetData *TD = 0;
1243 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1244 BasicBlock *DepBB = Deps[i].first;
1245 MemDepResult DepInfo = Deps[i].second;
1247 if (DepInfo.isClobber()) {
1248 // If the dependence is to a store that writes to a superset of the bits
1249 // read by the load, we can extract the bits we need for the load from the
1251 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1253 TD = getAnalysisIfAvailable<TargetData>();
1255 int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD);
1257 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1258 DepSI->getOperand(0),
1265 // FIXME: Handle memset/memcpy.
1266 UnavailableBlocks.push_back(DepBB);
1270 Instruction *DepInst = DepInfo.getInst();
1272 // Loading the allocation -> undef.
1273 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) {
1274 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1275 UndefValue::get(LI->getType())));
1279 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1280 // Reject loads and stores that are to the same address but are of
1281 // different types if we have to.
1282 if (S->getOperand(0)->getType() != LI->getType()) {
1284 TD = getAnalysisIfAvailable<TargetData>();
1286 // If the stored value is larger or equal to the loaded value, we can
1288 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1289 LI->getType(), *TD)) {
1290 UnavailableBlocks.push_back(DepBB);
1295 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1300 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1301 // If the types mismatch and we can't handle it, reject reuse of the load.
1302 if (LD->getType() != LI->getType()) {
1304 TD = getAnalysisIfAvailable<TargetData>();
1306 // If the stored value is larger or equal to the loaded value, we can
1308 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1309 UnavailableBlocks.push_back(DepBB);
1313 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1317 UnavailableBlocks.push_back(DepBB);
1321 // If we have no predecessors that produce a known value for this load, exit
1323 if (ValuesPerBlock.empty()) return false;
1325 // If all of the instructions we depend on produce a known value for this
1326 // load, then it is fully redundant and we can use PHI insertion to compute
1327 // its value. Insert PHIs and remove the fully redundant value now.
1328 if (UnavailableBlocks.empty()) {
1329 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1331 // Perform PHI construction.
1332 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1333 VN.getAliasAnalysis());
1334 LI->replaceAllUsesWith(V);
1336 if (isa<PHINode>(V))
1338 if (isa<PointerType>(V->getType()))
1339 MD->invalidateCachedPointerInfo(V);
1340 toErase.push_back(LI);
1345 if (!EnablePRE || !EnableLoadPRE)
1348 // Okay, we have *some* definitions of the value. This means that the value
1349 // is available in some of our (transitive) predecessors. Lets think about
1350 // doing PRE of this load. This will involve inserting a new load into the
1351 // predecessor when it's not available. We could do this in general, but
1352 // prefer to not increase code size. As such, we only do this when we know
1353 // that we only have to insert *one* load (which means we're basically moving
1354 // the load, not inserting a new one).
1356 SmallPtrSet<BasicBlock *, 4> Blockers;
1357 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1358 Blockers.insert(UnavailableBlocks[i]);
1360 // Lets find first basic block with more than one predecessor. Walk backwards
1361 // through predecessors if needed.
1362 BasicBlock *LoadBB = LI->getParent();
1363 BasicBlock *TmpBB = LoadBB;
1365 bool isSinglePred = false;
1366 bool allSingleSucc = true;
1367 while (TmpBB->getSinglePredecessor()) {
1368 isSinglePred = true;
1369 TmpBB = TmpBB->getSinglePredecessor();
1370 if (!TmpBB) // If haven't found any, bail now.
1372 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1374 if (Blockers.count(TmpBB))
1376 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1377 allSingleSucc = false;
1383 // If we have a repl set with LI itself in it, this means we have a loop where
1384 // at least one of the values is LI. Since this means that we won't be able
1385 // to eliminate LI even if we insert uses in the other predecessors, we will
1386 // end up increasing code size. Reject this by scanning for LI.
1387 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1388 if (ValuesPerBlock[i].V == LI)
1393 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1394 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].V))
1395 // "Hot" Instruction is in some loop (because it dominates its dep.
1397 if (DT->dominates(LI, I)) {
1402 // We are interested only in "hot" instructions. We don't want to do any
1403 // mis-optimizations here.
1408 // Okay, we have some hope :). Check to see if the loaded value is fully
1409 // available in all but one predecessor.
1410 // FIXME: If we could restructure the CFG, we could make a common pred with
1411 // all the preds that don't have an available LI and insert a new load into
1413 BasicBlock *UnavailablePred = 0;
1415 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1416 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1417 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1418 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1419 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1421 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1423 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1426 // If this load is not available in multiple predecessors, reject it.
1427 if (UnavailablePred && UnavailablePred != *PI)
1429 UnavailablePred = *PI;
1432 assert(UnavailablePred != 0 &&
1433 "Fully available value should be eliminated above!");
1435 // If the loaded pointer is PHI node defined in this block, do PHI translation
1436 // to get its value in the predecessor.
1437 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
1439 // Make sure the value is live in the predecessor. If it was defined by a
1440 // non-PHI instruction in this block, we don't know how to recompute it above.
1441 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
1442 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
1443 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
1444 << *LPInst << '\n' << *LI << "\n");
1448 // We don't currently handle critical edges :(
1449 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1450 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1451 << UnavailablePred->getName() << "': " << *LI << '\n');
1455 // Make sure it is valid to move this load here. We have to watch out for:
1456 // @1 = getelementptr (i8* p, ...
1457 // test p and branch if == 0
1459 // It is valid to have the getelementptr before the test, even if p can be 0,
1460 // as getelementptr only does address arithmetic.
1461 // If we are not pushing the value through any multiple-successor blocks
1462 // we do not have this case. Otherwise, check that the load is safe to
1463 // put anywhere; this can be improved, but should be conservatively safe.
1464 if (!allSingleSucc &&
1465 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator()))
1468 // Okay, we can eliminate this load by inserting a reload in the predecessor
1469 // and using PHI construction to get the value in the other predecessors, do
1471 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1473 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1475 UnavailablePred->getTerminator());
1477 // Add the newly created load.
1478 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));
1480 // Perform PHI construction.
1481 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1482 VN.getAliasAnalysis());
1483 LI->replaceAllUsesWith(V);
1484 if (isa<PHINode>(V))
1486 if (isa<PointerType>(V->getType()))
1487 MD->invalidateCachedPointerInfo(V);
1488 toErase.push_back(LI);
1493 /// processLoad - Attempt to eliminate a load, first by eliminating it
1494 /// locally, and then attempting non-local elimination if that fails.
1495 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1496 if (L->isVolatile())
1499 // ... to a pointer that has been loaded from before...
1500 MemDepResult Dep = MD->getDependency(L);
1502 // If the value isn't available, don't do anything!
1503 if (Dep.isClobber()) {
1504 // FIXME: We should handle memset/memcpy/memmove as dependent instructions
1505 // to forward the value if available.
1506 //if (isa<MemIntrinsic>(Dep.getInst()))
1507 //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n";
1509 // Check to see if we have something like this:
1510 // store i32 123, i32* %P
1511 // %A = bitcast i32* %P to i8*
1512 // %B = gep i8* %A, i32 1
1515 // We could do that by recognizing if the clobber instructions are obviously
1516 // a common base + constant offset, and if the previous store (or memset)
1517 // completely covers this load. This sort of thing can happen in bitfield
1519 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1520 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1521 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD);
1523 Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1524 L->getType(), L, *TD);
1525 DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n'
1526 << *AvailVal << '\n' << *L << "\n\n\n");
1528 // Replace the load!
1529 L->replaceAllUsesWith(AvailVal);
1530 if (isa<PointerType>(AvailVal->getType()))
1531 MD->invalidateCachedPointerInfo(AvailVal);
1532 toErase.push_back(L);
1539 // fast print dep, using operator<< on instruction would be too slow
1540 errs() << "GVN: load ";
1541 WriteAsOperand(errs(), L);
1542 Instruction *I = Dep.getInst();
1543 errs() << " is clobbered by " << *I << '\n';
1548 // If it is defined in another block, try harder.
1549 if (Dep.isNonLocal())
1550 return processNonLocalLoad(L, toErase);
1552 Instruction *DepInst = Dep.getInst();
1553 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1554 Value *StoredVal = DepSI->getOperand(0);
1556 // The store and load are to a must-aliased pointer, but they may not
1557 // actually have the same type. See if we know how to reuse the stored
1558 // value (depending on its type).
1559 const TargetData *TD = 0;
1560 if (StoredVal->getType() != L->getType() &&
1561 (TD = getAnalysisIfAvailable<TargetData>())) {
1562 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1567 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1568 << '\n' << *L << "\n\n\n");
1572 L->replaceAllUsesWith(StoredVal);
1573 if (isa<PointerType>(StoredVal->getType()))
1574 MD->invalidateCachedPointerInfo(StoredVal);
1575 toErase.push_back(L);
1580 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1581 Value *AvailableVal = DepLI;
1583 // The loads are of a must-aliased pointer, but they may not actually have
1584 // the same type. See if we know how to reuse the previously loaded value
1585 // (depending on its type).
1586 const TargetData *TD = 0;
1587 if (DepLI->getType() != L->getType() &&
1588 (TD = getAnalysisIfAvailable<TargetData>())) {
1589 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1590 if (AvailableVal == 0)
1593 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1594 << "\n" << *L << "\n\n\n");
1598 L->replaceAllUsesWith(AvailableVal);
1599 if (isa<PointerType>(DepLI->getType()))
1600 MD->invalidateCachedPointerInfo(DepLI);
1601 toErase.push_back(L);
1606 // If this load really doesn't depend on anything, then we must be loading an
1607 // undef value. This can happen when loading for a fresh allocation with no
1608 // intervening stores, for example.
1609 if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) {
1610 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1611 toErase.push_back(L);
1619 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1620 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1621 if (I == localAvail.end())
1624 ValueNumberScope *Locals = I->second;
1626 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1627 if (I != Locals->table.end())
1629 Locals = Locals->parent;
1636 /// processInstruction - When calculating availability, handle an instruction
1637 /// by inserting it into the appropriate sets
1638 bool GVN::processInstruction(Instruction *I,
1639 SmallVectorImpl<Instruction*> &toErase) {
1640 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1641 bool Changed = processLoad(LI, toErase);
1644 unsigned Num = VN.lookup_or_add(LI);
1645 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1651 uint32_t NextNum = VN.getNextUnusedValueNumber();
1652 unsigned Num = VN.lookup_or_add(I);
1654 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1655 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1657 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1660 Value *BranchCond = BI->getCondition();
1661 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1663 BasicBlock *TrueSucc = BI->getSuccessor(0);
1664 BasicBlock *FalseSucc = BI->getSuccessor(1);
1666 if (TrueSucc->getSinglePredecessor())
1667 localAvail[TrueSucc]->table[CondVN] =
1668 ConstantInt::getTrue(TrueSucc->getContext());
1669 if (FalseSucc->getSinglePredecessor())
1670 localAvail[FalseSucc]->table[CondVN] =
1671 ConstantInt::getFalse(TrueSucc->getContext());
1675 // Allocations are always uniquely numbered, so we can save time and memory
1676 // by fast failing them.
1677 } else if (isa<AllocationInst>(I) || isa<TerminatorInst>(I)) {
1678 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1682 // Collapse PHI nodes
1683 if (PHINode* p = dyn_cast<PHINode>(I)) {
1684 Value *constVal = CollapsePhi(p);
1687 p->replaceAllUsesWith(constVal);
1688 if (isa<PointerType>(constVal->getType()))
1689 MD->invalidateCachedPointerInfo(constVal);
1692 toErase.push_back(p);
1694 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1697 // If the number we were assigned was a brand new VN, then we don't
1698 // need to do a lookup to see if the number already exists
1699 // somewhere in the domtree: it can't!
1700 } else if (Num == NextNum) {
1701 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1703 // Perform fast-path value-number based elimination of values inherited from
1705 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1708 I->replaceAllUsesWith(repl);
1709 if (isa<PointerType>(repl->getType()))
1710 MD->invalidateCachedPointerInfo(repl);
1711 toErase.push_back(I);
1715 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1721 /// runOnFunction - This is the main transformation entry point for a function.
1722 bool GVN::runOnFunction(Function& F) {
1723 MD = &getAnalysis<MemoryDependenceAnalysis>();
1724 DT = &getAnalysis<DominatorTree>();
1725 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1729 bool Changed = false;
1730 bool ShouldContinue = true;
1732 // Merge unconditional branches, allowing PRE to catch more
1733 // optimization opportunities.
1734 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1735 BasicBlock *BB = FI;
1737 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1738 if (removedBlock) NumGVNBlocks++;
1740 Changed |= removedBlock;
1743 unsigned Iteration = 0;
1745 while (ShouldContinue) {
1746 DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
1747 ShouldContinue = iterateOnFunction(F);
1748 Changed |= ShouldContinue;
1753 bool PREChanged = true;
1754 while (PREChanged) {
1755 PREChanged = performPRE(F);
1756 Changed |= PREChanged;
1759 // FIXME: Should perform GVN again after PRE does something. PRE can move
1760 // computations into blocks where they become fully redundant. Note that
1761 // we can't do this until PRE's critical edge splitting updates memdep.
1762 // Actually, when this happens, we should just fully integrate PRE into GVN.
1764 cleanupGlobalSets();
1770 bool GVN::processBlock(BasicBlock *BB) {
1771 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1772 // incrementing BI before processing an instruction).
1773 SmallVector<Instruction*, 8> toErase;
1774 bool ChangedFunction = false;
1776 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1778 ChangedFunction |= processInstruction(BI, toErase);
1779 if (toErase.empty()) {
1784 // If we need some instructions deleted, do it now.
1785 NumGVNInstr += toErase.size();
1787 // Avoid iterator invalidation.
1788 bool AtStart = BI == BB->begin();
1792 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1793 E = toErase.end(); I != E; ++I) {
1794 DEBUG(errs() << "GVN removed: " << **I << '\n');
1795 MD->removeInstruction(*I);
1796 (*I)->eraseFromParent();
1797 DEBUG(verifyRemoved(*I));
1807 return ChangedFunction;
1810 /// performPRE - Perform a purely local form of PRE that looks for diamond
1811 /// control flow patterns and attempts to perform simple PRE at the join point.
1812 bool GVN::performPRE(Function& F) {
1813 bool Changed = false;
1814 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1815 DenseMap<BasicBlock*, Value*> predMap;
1816 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1817 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1818 BasicBlock *CurrentBlock = *DI;
1820 // Nothing to PRE in the entry block.
1821 if (CurrentBlock == &F.getEntryBlock()) continue;
1823 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1824 BE = CurrentBlock->end(); BI != BE; ) {
1825 Instruction *CurInst = BI++;
1827 if (isa<AllocationInst>(CurInst) ||
1828 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
1829 CurInst->getType()->isVoidTy() ||
1830 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1831 isa<DbgInfoIntrinsic>(CurInst))
1834 uint32_t ValNo = VN.lookup(CurInst);
1836 // Look for the predecessors for PRE opportunities. We're
1837 // only trying to solve the basic diamond case, where
1838 // a value is computed in the successor and one predecessor,
1839 // but not the other. We also explicitly disallow cases
1840 // where the successor is its own predecessor, because they're
1841 // more complicated to get right.
1842 unsigned NumWith = 0;
1843 unsigned NumWithout = 0;
1844 BasicBlock *PREPred = 0;
1847 for (pred_iterator PI = pred_begin(CurrentBlock),
1848 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1849 // We're not interested in PRE where the block is its
1850 // own predecessor, on in blocks with predecessors
1851 // that are not reachable.
1852 if (*PI == CurrentBlock) {
1855 } else if (!localAvail.count(*PI)) {
1860 DenseMap<uint32_t, Value*>::iterator predV =
1861 localAvail[*PI]->table.find(ValNo);
1862 if (predV == localAvail[*PI]->table.end()) {
1865 } else if (predV->second == CurInst) {
1868 predMap[*PI] = predV->second;
1873 // Don't do PRE when it might increase code size, i.e. when
1874 // we would need to insert instructions in more than one pred.
1875 if (NumWithout != 1 || NumWith == 0)
1878 // We can't do PRE safely on a critical edge, so instead we schedule
1879 // the edge to be split and perform the PRE the next time we iterate
1881 unsigned SuccNum = 0;
1882 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1884 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1889 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
1890 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
1894 // Instantiate the expression the in predecessor that lacked it.
1895 // Because we are going top-down through the block, all value numbers
1896 // will be available in the predecessor by the time we need them. Any
1897 // that weren't original present will have been instantiated earlier
1899 Instruction *PREInstr = CurInst->clone();
1900 bool success = true;
1901 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1902 Value *Op = PREInstr->getOperand(i);
1903 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1906 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
1907 PREInstr->setOperand(i, V);
1914 // Fail out if we encounter an operand that is not available in
1915 // the PRE predecessor. This is typically because of loads which
1916 // are not value numbered precisely.
1919 DEBUG(verifyRemoved(PREInstr));
1923 PREInstr->insertBefore(PREPred->getTerminator());
1924 PREInstr->setName(CurInst->getName() + ".pre");
1925 predMap[PREPred] = PREInstr;
1926 VN.add(PREInstr, ValNo);
1929 // Update the availability map to include the new instruction.
1930 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
1932 // Create a PHI to make the value available in this block.
1933 PHINode* Phi = PHINode::Create(CurInst->getType(),
1934 CurInst->getName() + ".pre-phi",
1935 CurrentBlock->begin());
1936 for (pred_iterator PI = pred_begin(CurrentBlock),
1937 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1938 Phi->addIncoming(predMap[*PI], *PI);
1941 localAvail[CurrentBlock]->table[ValNo] = Phi;
1943 CurInst->replaceAllUsesWith(Phi);
1944 if (isa<PointerType>(Phi->getType()))
1945 MD->invalidateCachedPointerInfo(Phi);
1948 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
1949 MD->removeInstruction(CurInst);
1950 CurInst->eraseFromParent();
1951 DEBUG(verifyRemoved(CurInst));
1956 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1957 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1958 SplitCriticalEdge(I->first, I->second, this);
1960 return Changed || toSplit.size();
1963 /// iterateOnFunction - Executes one iteration of GVN
1964 bool GVN::iterateOnFunction(Function &F) {
1965 cleanupGlobalSets();
1967 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1968 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
1970 localAvail[DI->getBlock()] =
1971 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
1973 localAvail[DI->getBlock()] = new ValueNumberScope(0);
1976 // Top-down walk of the dominator tree
1977 bool Changed = false;
1979 // Needed for value numbering with phi construction to work.
1980 ReversePostOrderTraversal<Function*> RPOT(&F);
1981 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
1982 RE = RPOT.end(); RI != RE; ++RI)
1983 Changed |= processBlock(*RI);
1985 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1986 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
1987 Changed |= processBlock(DI->getBlock());
1993 void GVN::cleanupGlobalSets() {
1996 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1997 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2002 /// verifyRemoved - Verify that the specified instruction does not occur in our
2003 /// internal data structures.
2004 void GVN::verifyRemoved(const Instruction *Inst) const {
2005 VN.verifyRemoved(Inst);
2007 // Walk through the value number scope to make sure the instruction isn't
2008 // ferreted away in it.
2009 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2010 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2011 const ValueNumberScope *VNS = I->second;
2014 for (DenseMap<uint32_t, Value*>::iterator
2015 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2016 assert(II->second != Inst && "Inst still in value numbering scope!");