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/AliasAnalysis.h"
35 #include "llvm/Analysis/ConstantFolding.h"
36 #include "llvm/Analysis/Dominators.h"
37 #include "llvm/Analysis/MemoryBuiltins.h"
38 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
39 #include "llvm/Analysis/PHITransAddr.h"
40 #include "llvm/Support/CFG.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/GetElementPtrTypeIterator.h"
45 #include "llvm/Support/IRBuilder.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SSAUpdater.h"
53 STATISTIC(NumGVNInstr, "Number of instructions deleted");
54 STATISTIC(NumGVNLoad, "Number of loads deleted");
55 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
56 STATISTIC(NumGVNBlocks, "Number of blocks merged");
57 STATISTIC(NumPRELoad, "Number of loads PRE'd");
59 static cl::opt<bool> EnablePRE("enable-pre",
60 cl::init(true), cl::Hidden);
61 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
63 //===----------------------------------------------------------------------===//
65 //===----------------------------------------------------------------------===//
67 /// This class holds the mapping between values and value numbers. It is used
68 /// as an efficient mechanism to determine the expression-wise equivalence of
72 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
73 UDIV, SDIV, FDIV, UREM, SREM,
74 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
75 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
76 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
77 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
78 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
79 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
80 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
81 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
82 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
83 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
85 ExpressionOpcode opcode;
87 SmallVector<uint32_t, 4> varargs;
91 Expression(ExpressionOpcode o) : opcode(o) { }
93 bool operator==(const Expression &other) const {
94 if (opcode != other.opcode)
96 else if (opcode == EMPTY || opcode == TOMBSTONE)
98 else if (type != other.type)
100 else if (function != other.function)
103 if (varargs.size() != other.varargs.size())
106 for (size_t i = 0; i < varargs.size(); ++i)
107 if (varargs[i] != other.varargs[i])
114 bool operator!=(const Expression &other) const {
115 return !(*this == other);
121 DenseMap<Value*, uint32_t> valueNumbering;
122 DenseMap<Expression, uint32_t> expressionNumbering;
124 MemoryDependenceAnalysis* MD;
127 uint32_t nextValueNumber;
129 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
130 Expression::ExpressionOpcode getOpcode(CmpInst* C);
131 Expression::ExpressionOpcode getOpcode(CastInst* C);
132 Expression create_expression(BinaryOperator* BO);
133 Expression create_expression(CmpInst* C);
134 Expression create_expression(ShuffleVectorInst* V);
135 Expression create_expression(ExtractElementInst* C);
136 Expression create_expression(InsertElementInst* V);
137 Expression create_expression(SelectInst* V);
138 Expression create_expression(CastInst* C);
139 Expression create_expression(GetElementPtrInst* G);
140 Expression create_expression(CallInst* C);
141 Expression create_expression(Constant* C);
142 Expression create_expression(ExtractValueInst* C);
143 Expression create_expression(InsertValueInst* C);
145 uint32_t lookup_or_add_call(CallInst* C);
147 ValueTable() : nextValueNumber(1) { }
148 uint32_t lookup_or_add(Value *V);
149 uint32_t lookup(Value *V) const;
150 void add(Value *V, uint32_t num);
152 void erase(Value *v);
154 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
155 AliasAnalysis *getAliasAnalysis() const { return AA; }
156 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
157 void setDomTree(DominatorTree* D) { DT = D; }
158 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
159 void verifyRemoved(const Value *) const;
164 template <> struct DenseMapInfo<Expression> {
165 static inline Expression getEmptyKey() {
166 return Expression(Expression::EMPTY);
169 static inline Expression getTombstoneKey() {
170 return Expression(Expression::TOMBSTONE);
173 static unsigned getHashValue(const Expression e) {
174 unsigned hash = e.opcode;
176 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
177 (unsigned)((uintptr_t)e.type >> 9));
179 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
180 E = e.varargs.end(); I != E; ++I)
181 hash = *I + hash * 37;
183 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
184 (unsigned)((uintptr_t)e.function >> 9)) +
189 static bool isEqual(const Expression &LHS, const Expression &RHS) {
195 struct isPodLike<Expression> { static const bool value = true; };
199 //===----------------------------------------------------------------------===//
200 // ValueTable Internal Functions
201 //===----------------------------------------------------------------------===//
202 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
203 switch(BO->getOpcode()) {
204 default: // THIS SHOULD NEVER HAPPEN
205 llvm_unreachable("Binary operator with unknown opcode?");
206 case Instruction::Add: return Expression::ADD;
207 case Instruction::FAdd: return Expression::FADD;
208 case Instruction::Sub: return Expression::SUB;
209 case Instruction::FSub: return Expression::FSUB;
210 case Instruction::Mul: return Expression::MUL;
211 case Instruction::FMul: return Expression::FMUL;
212 case Instruction::UDiv: return Expression::UDIV;
213 case Instruction::SDiv: return Expression::SDIV;
214 case Instruction::FDiv: return Expression::FDIV;
215 case Instruction::URem: return Expression::UREM;
216 case Instruction::SRem: return Expression::SREM;
217 case Instruction::FRem: return Expression::FREM;
218 case Instruction::Shl: return Expression::SHL;
219 case Instruction::LShr: return Expression::LSHR;
220 case Instruction::AShr: return Expression::ASHR;
221 case Instruction::And: return Expression::AND;
222 case Instruction::Or: return Expression::OR;
223 case Instruction::Xor: return Expression::XOR;
227 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
228 if (isa<ICmpInst>(C)) {
229 switch (C->getPredicate()) {
230 default: // THIS SHOULD NEVER HAPPEN
231 llvm_unreachable("Comparison with unknown predicate?");
232 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
233 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
234 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
235 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
236 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
237 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
238 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
239 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
240 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
241 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
244 switch (C->getPredicate()) {
245 default: // THIS SHOULD NEVER HAPPEN
246 llvm_unreachable("Comparison with unknown predicate?");
247 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
248 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
249 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
250 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
251 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
252 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
253 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
254 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
255 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
256 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
257 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
258 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
259 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
260 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
265 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
266 switch(C->getOpcode()) {
267 default: // THIS SHOULD NEVER HAPPEN
268 llvm_unreachable("Cast operator with unknown opcode?");
269 case Instruction::Trunc: return Expression::TRUNC;
270 case Instruction::ZExt: return Expression::ZEXT;
271 case Instruction::SExt: return Expression::SEXT;
272 case Instruction::FPToUI: return Expression::FPTOUI;
273 case Instruction::FPToSI: return Expression::FPTOSI;
274 case Instruction::UIToFP: return Expression::UITOFP;
275 case Instruction::SIToFP: return Expression::SITOFP;
276 case Instruction::FPTrunc: return Expression::FPTRUNC;
277 case Instruction::FPExt: return Expression::FPEXT;
278 case Instruction::PtrToInt: return Expression::PTRTOINT;
279 case Instruction::IntToPtr: return Expression::INTTOPTR;
280 case Instruction::BitCast: return Expression::BITCAST;
284 Expression ValueTable::create_expression(CallInst* C) {
287 e.type = C->getType();
288 e.function = C->getCalledFunction();
289 e.opcode = Expression::CALL;
291 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
293 e.varargs.push_back(lookup_or_add(*I));
298 Expression ValueTable::create_expression(BinaryOperator* BO) {
300 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
301 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
303 e.type = BO->getType();
304 e.opcode = getOpcode(BO);
309 Expression ValueTable::create_expression(CmpInst* C) {
312 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
313 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
315 e.type = C->getType();
316 e.opcode = getOpcode(C);
321 Expression ValueTable::create_expression(CastInst* C) {
324 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
326 e.type = C->getType();
327 e.opcode = getOpcode(C);
332 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
335 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
336 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
337 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
339 e.type = S->getType();
340 e.opcode = Expression::SHUFFLE;
345 Expression ValueTable::create_expression(ExtractElementInst* E) {
348 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
349 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
351 e.type = E->getType();
352 e.opcode = Expression::EXTRACT;
357 Expression ValueTable::create_expression(InsertElementInst* I) {
360 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
361 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
362 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
364 e.type = I->getType();
365 e.opcode = Expression::INSERT;
370 Expression ValueTable::create_expression(SelectInst* I) {
373 e.varargs.push_back(lookup_or_add(I->getCondition()));
374 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
375 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
377 e.type = I->getType();
378 e.opcode = Expression::SELECT;
383 Expression ValueTable::create_expression(GetElementPtrInst* G) {
386 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
388 e.type = G->getType();
389 e.opcode = Expression::GEP;
391 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
393 e.varargs.push_back(lookup_or_add(*I));
398 Expression ValueTable::create_expression(ExtractValueInst* E) {
401 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
402 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
404 e.varargs.push_back(*II);
406 e.type = E->getType();
407 e.opcode = Expression::EXTRACTVALUE;
412 Expression ValueTable::create_expression(InsertValueInst* E) {
415 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
416 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
417 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
419 e.varargs.push_back(*II);
421 e.type = E->getType();
422 e.opcode = Expression::INSERTVALUE;
427 //===----------------------------------------------------------------------===//
428 // ValueTable External Functions
429 //===----------------------------------------------------------------------===//
431 /// add - Insert a value into the table with a specified value number.
432 void ValueTable::add(Value *V, uint32_t num) {
433 valueNumbering.insert(std::make_pair(V, num));
436 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
437 if (AA->doesNotAccessMemory(C)) {
438 Expression exp = create_expression(C);
439 uint32_t& e = expressionNumbering[exp];
440 if (!e) e = nextValueNumber++;
441 valueNumbering[C] = e;
443 } else if (AA->onlyReadsMemory(C)) {
444 Expression exp = create_expression(C);
445 uint32_t& e = expressionNumbering[exp];
447 e = nextValueNumber++;
448 valueNumbering[C] = e;
452 e = nextValueNumber++;
453 valueNumbering[C] = e;
457 MemDepResult local_dep = MD->getDependency(C);
459 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
460 valueNumbering[C] = nextValueNumber;
461 return nextValueNumber++;
464 if (local_dep.isDef()) {
465 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
467 if (local_cdep->getNumOperands() != C->getNumOperands()) {
468 valueNumbering[C] = nextValueNumber;
469 return nextValueNumber++;
472 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
473 uint32_t c_vn = lookup_or_add(C->getOperand(i));
474 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
476 valueNumbering[C] = nextValueNumber;
477 return nextValueNumber++;
481 uint32_t v = lookup_or_add(local_cdep);
482 valueNumbering[C] = v;
487 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
488 MD->getNonLocalCallDependency(CallSite(C));
489 // FIXME: call/call dependencies for readonly calls should return def, not
490 // clobber! Move the checking logic to MemDep!
493 // Check to see if we have a single dominating call instruction that is
495 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
496 const NonLocalDepEntry *I = &deps[i];
497 // Ignore non-local dependencies.
498 if (I->getResult().isNonLocal())
501 // We don't handle non-depedencies. If we already have a call, reject
502 // instruction dependencies.
503 if (I->getResult().isClobber() || cdep != 0) {
508 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
509 // FIXME: All duplicated with non-local case.
510 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
511 cdep = NonLocalDepCall;
520 valueNumbering[C] = nextValueNumber;
521 return nextValueNumber++;
524 if (cdep->getNumOperands() != C->getNumOperands()) {
525 valueNumbering[C] = nextValueNumber;
526 return nextValueNumber++;
528 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
529 uint32_t c_vn = lookup_or_add(C->getOperand(i));
530 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
532 valueNumbering[C] = nextValueNumber;
533 return nextValueNumber++;
537 uint32_t v = lookup_or_add(cdep);
538 valueNumbering[C] = v;
542 valueNumbering[C] = nextValueNumber;
543 return nextValueNumber++;
547 /// lookup_or_add - Returns the value number for the specified value, assigning
548 /// it a new number if it did not have one before.
549 uint32_t ValueTable::lookup_or_add(Value *V) {
550 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
551 if (VI != valueNumbering.end())
554 if (!isa<Instruction>(V)) {
555 valueNumbering[V] = nextValueNumber;
556 return nextValueNumber++;
559 Instruction* I = cast<Instruction>(V);
561 switch (I->getOpcode()) {
562 case Instruction::Call:
563 return lookup_or_add_call(cast<CallInst>(I));
564 case Instruction::Add:
565 case Instruction::FAdd:
566 case Instruction::Sub:
567 case Instruction::FSub:
568 case Instruction::Mul:
569 case Instruction::FMul:
570 case Instruction::UDiv:
571 case Instruction::SDiv:
572 case Instruction::FDiv:
573 case Instruction::URem:
574 case Instruction::SRem:
575 case Instruction::FRem:
576 case Instruction::Shl:
577 case Instruction::LShr:
578 case Instruction::AShr:
579 case Instruction::And:
580 case Instruction::Or :
581 case Instruction::Xor:
582 exp = create_expression(cast<BinaryOperator>(I));
584 case Instruction::ICmp:
585 case Instruction::FCmp:
586 exp = create_expression(cast<CmpInst>(I));
588 case Instruction::Trunc:
589 case Instruction::ZExt:
590 case Instruction::SExt:
591 case Instruction::FPToUI:
592 case Instruction::FPToSI:
593 case Instruction::UIToFP:
594 case Instruction::SIToFP:
595 case Instruction::FPTrunc:
596 case Instruction::FPExt:
597 case Instruction::PtrToInt:
598 case Instruction::IntToPtr:
599 case Instruction::BitCast:
600 exp = create_expression(cast<CastInst>(I));
602 case Instruction::Select:
603 exp = create_expression(cast<SelectInst>(I));
605 case Instruction::ExtractElement:
606 exp = create_expression(cast<ExtractElementInst>(I));
608 case Instruction::InsertElement:
609 exp = create_expression(cast<InsertElementInst>(I));
611 case Instruction::ShuffleVector:
612 exp = create_expression(cast<ShuffleVectorInst>(I));
614 case Instruction::ExtractValue:
615 exp = create_expression(cast<ExtractValueInst>(I));
617 case Instruction::InsertValue:
618 exp = create_expression(cast<InsertValueInst>(I));
620 case Instruction::GetElementPtr:
621 exp = create_expression(cast<GetElementPtrInst>(I));
624 valueNumbering[V] = nextValueNumber;
625 return nextValueNumber++;
628 uint32_t& e = expressionNumbering[exp];
629 if (!e) e = nextValueNumber++;
630 valueNumbering[V] = e;
634 /// lookup - Returns the value number of the specified value. Fails if
635 /// the value has not yet been numbered.
636 uint32_t ValueTable::lookup(Value *V) const {
637 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
638 assert(VI != valueNumbering.end() && "Value not numbered?");
642 /// clear - Remove all entries from the ValueTable
643 void ValueTable::clear() {
644 valueNumbering.clear();
645 expressionNumbering.clear();
649 /// erase - Remove a value from the value numbering
650 void ValueTable::erase(Value *V) {
651 valueNumbering.erase(V);
654 /// verifyRemoved - Verify that the value is removed from all internal data
656 void ValueTable::verifyRemoved(const Value *V) const {
657 for (DenseMap<Value*, uint32_t>::const_iterator
658 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
659 assert(I->first != V && "Inst still occurs in value numbering map!");
663 //===----------------------------------------------------------------------===//
665 //===----------------------------------------------------------------------===//
668 struct ValueNumberScope {
669 ValueNumberScope* parent;
670 DenseMap<uint32_t, Value*> table;
672 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
678 class GVN : public FunctionPass {
679 bool runOnFunction(Function &F);
681 static char ID; // Pass identification, replacement for typeid
682 explicit GVN(bool nopre = false, bool noloads = false)
683 : FunctionPass(&ID), NoPRE(nopre), NoLoads(noloads), MD(0) { }
688 MemoryDependenceAnalysis *MD;
692 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
694 // This transformation requires dominator postdominator info
695 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
696 AU.addRequired<DominatorTree>();
698 AU.addRequired<MemoryDependenceAnalysis>();
699 AU.addRequired<AliasAnalysis>();
701 AU.addPreserved<DominatorTree>();
702 AU.addPreserved<AliasAnalysis>();
706 // FIXME: eliminate or document these better
707 bool processLoad(LoadInst* L,
708 SmallVectorImpl<Instruction*> &toErase);
709 bool processInstruction(Instruction *I,
710 SmallVectorImpl<Instruction*> &toErase);
711 bool processNonLocalLoad(LoadInst* L,
712 SmallVectorImpl<Instruction*> &toErase);
713 bool processBlock(BasicBlock *BB);
714 void dump(DenseMap<uint32_t, Value*>& d);
715 bool iterateOnFunction(Function &F);
716 Value *CollapsePhi(PHINode* p);
717 bool performPRE(Function& F);
718 Value *lookupNumber(BasicBlock *BB, uint32_t num);
719 void cleanupGlobalSets();
720 void verifyRemoved(const Instruction *I) const;
726 // createGVNPass - The public interface to this file...
727 FunctionPass *llvm::createGVNPass(bool NoPRE, bool NoLoads) {
728 return new GVN(NoPRE, NoLoads);
731 static RegisterPass<GVN> X("gvn",
732 "Global Value Numbering");
734 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
736 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
737 E = d.end(); I != E; ++I) {
738 errs() << I->first << "\n";
744 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
745 if (!isa<PHINode>(inst))
748 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
750 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
751 if (use_phi->getParent() == inst->getParent())
757 Value *GVN::CollapsePhi(PHINode *PN) {
758 Value *ConstVal = PN->hasConstantValue(DT);
759 if (!ConstVal) return 0;
761 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
765 if (DT->dominates(Inst, PN))
766 if (isSafeReplacement(PN, Inst))
771 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
772 /// we're analyzing is fully available in the specified block. As we go, keep
773 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
774 /// map is actually a tri-state map with the following values:
775 /// 0) we know the block *is not* fully available.
776 /// 1) we know the block *is* fully available.
777 /// 2) we do not know whether the block is fully available or not, but we are
778 /// currently speculating that it will be.
779 /// 3) we are speculating for this block and have used that to speculate for
781 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
782 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
783 // Optimistically assume that the block is fully available and check to see
784 // if we already know about this block in one lookup.
785 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
786 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
788 // If the entry already existed for this block, return the precomputed value.
790 // If this is a speculative "available" value, mark it as being used for
791 // speculation of other blocks.
792 if (IV.first->second == 2)
793 IV.first->second = 3;
794 return IV.first->second != 0;
797 // Otherwise, see if it is fully available in all predecessors.
798 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
800 // If this block has no predecessors, it isn't live-in here.
802 goto SpeculationFailure;
804 for (; PI != PE; ++PI)
805 // If the value isn't fully available in one of our predecessors, then it
806 // isn't fully available in this block either. Undo our previous
807 // optimistic assumption and bail out.
808 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
809 goto SpeculationFailure;
813 // SpeculationFailure - If we get here, we found out that this is not, after
814 // all, a fully-available block. We have a problem if we speculated on this and
815 // used the speculation to mark other blocks as available.
817 char &BBVal = FullyAvailableBlocks[BB];
819 // If we didn't speculate on this, just return with it set to false.
825 // If we did speculate on this value, we could have blocks set to 1 that are
826 // incorrect. Walk the (transitive) successors of this block and mark them as
828 SmallVector<BasicBlock*, 32> BBWorklist;
829 BBWorklist.push_back(BB);
831 while (!BBWorklist.empty()) {
832 BasicBlock *Entry = BBWorklist.pop_back_val();
833 // Note that this sets blocks to 0 (unavailable) if they happen to not
834 // already be in FullyAvailableBlocks. This is safe.
835 char &EntryVal = FullyAvailableBlocks[Entry];
836 if (EntryVal == 0) continue; // Already unavailable.
838 // Mark as unavailable.
841 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
842 BBWorklist.push_back(*I);
849 /// CanCoerceMustAliasedValueToLoad - Return true if
850 /// CoerceAvailableValueToLoadType will succeed.
851 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
853 const TargetData &TD) {
854 // If the loaded or stored value is an first class array or struct, don't try
855 // to transform them. We need to be able to bitcast to integer.
856 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
857 isa<StructType>(StoredVal->getType()) ||
858 isa<ArrayType>(StoredVal->getType()))
861 // The store has to be at least as big as the load.
862 if (TD.getTypeSizeInBits(StoredVal->getType()) <
863 TD.getTypeSizeInBits(LoadTy))
870 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
871 /// then a load from a must-aliased pointer of a different type, try to coerce
872 /// the stored value. LoadedTy is the type of the load we want to replace and
873 /// InsertPt is the place to insert new instructions.
875 /// If we can't do it, return null.
876 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
877 const Type *LoadedTy,
878 Instruction *InsertPt,
879 const TargetData &TD) {
880 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
883 const Type *StoredValTy = StoredVal->getType();
885 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
886 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
888 // If the store and reload are the same size, we can always reuse it.
889 if (StoreSize == LoadSize) {
890 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
891 // Pointer to Pointer -> use bitcast.
892 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
895 // Convert source pointers to integers, which can be bitcast.
896 if (isa<PointerType>(StoredValTy)) {
897 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
898 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
901 const Type *TypeToCastTo = LoadedTy;
902 if (isa<PointerType>(TypeToCastTo))
903 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
905 if (StoredValTy != TypeToCastTo)
906 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
908 // Cast to pointer if the load needs a pointer type.
909 if (isa<PointerType>(LoadedTy))
910 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
915 // If the loaded value is smaller than the available value, then we can
916 // extract out a piece from it. If the available value is too small, then we
917 // can't do anything.
918 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
920 // Convert source pointers to integers, which can be manipulated.
921 if (isa<PointerType>(StoredValTy)) {
922 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
923 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
926 // Convert vectors and fp to integer, which can be manipulated.
927 if (!isa<IntegerType>(StoredValTy)) {
928 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
929 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
932 // If this is a big-endian system, we need to shift the value down to the low
933 // bits so that a truncate will work.
934 if (TD.isBigEndian()) {
935 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
936 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
939 // Truncate the integer to the right size now.
940 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
941 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
943 if (LoadedTy == NewIntTy)
946 // If the result is a pointer, inttoptr.
947 if (isa<PointerType>(LoadedTy))
948 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
950 // Otherwise, bitcast.
951 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
954 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
955 /// be expressed as a base pointer plus a constant offset. Return the base and
956 /// offset to the caller.
957 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
958 const TargetData &TD) {
959 Operator *PtrOp = dyn_cast<Operator>(Ptr);
960 if (PtrOp == 0) return Ptr;
962 // Just look through bitcasts.
963 if (PtrOp->getOpcode() == Instruction::BitCast)
964 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
966 // If this is a GEP with constant indices, we can look through it.
967 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
968 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
970 gep_type_iterator GTI = gep_type_begin(GEP);
971 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
973 ConstantInt *OpC = cast<ConstantInt>(*I);
974 if (OpC->isZero()) continue;
976 // Handle a struct and array indices which add their offset to the pointer.
977 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
978 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
980 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
981 Offset += OpC->getSExtValue()*Size;
985 // Re-sign extend from the pointer size if needed to get overflow edge cases
987 unsigned PtrSize = TD.getPointerSizeInBits();
989 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
991 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
995 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
996 /// memdep query of a load that ends up being a clobbering memory write (store,
997 /// memset, memcpy, memmove). This means that the write *may* provide bits used
998 /// by the load but we can't be sure because the pointers don't mustalias.
1000 /// Check this case to see if there is anything more we can do before we give
1001 /// up. This returns -1 if we have to give up, or a byte number in the stored
1002 /// value of the piece that feeds the load.
1003 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
1005 uint64_t WriteSizeInBits,
1006 const TargetData &TD) {
1007 // If the loaded or stored value is an first class array or struct, don't try
1008 // to transform them. We need to be able to bitcast to integer.
1009 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy))
1012 int64_t StoreOffset = 0, LoadOffset = 0;
1013 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1015 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
1016 if (StoreBase != LoadBase)
1019 // If the load and store are to the exact same address, they should have been
1020 // a must alias. AA must have gotten confused.
1021 // FIXME: Study to see if/when this happens.
1022 if (LoadOffset == StoreOffset) {
1024 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1025 << "Base = " << *StoreBase << "\n"
1026 << "Store Ptr = " << *WritePtr << "\n"
1027 << "Store Offs = " << StoreOffset << "\n"
1028 << "Load Ptr = " << *LoadPtr << "\n";
1034 // If the load and store don't overlap at all, the store doesn't provide
1035 // anything to the load. In this case, they really don't alias at all, AA
1036 // must have gotten confused.
1037 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1038 // remove this check, as it is duplicated with what we have below.
1039 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
1041 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1043 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1047 bool isAAFailure = false;
1048 if (StoreOffset < LoadOffset) {
1049 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1051 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1055 errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1056 << "Base = " << *StoreBase << "\n"
1057 << "Store Ptr = " << *WritePtr << "\n"
1058 << "Store Offs = " << StoreOffset << "\n"
1059 << "Load Ptr = " << *LoadPtr << "\n";
1065 // If the Load isn't completely contained within the stored bits, we don't
1066 // have all the bits to feed it. We could do something crazy in the future
1067 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1069 if (StoreOffset > LoadOffset ||
1070 StoreOffset+StoreSize < LoadOffset+LoadSize)
1073 // Okay, we can do this transformation. Return the number of bytes into the
1074 // store that the load is.
1075 return LoadOffset-StoreOffset;
1078 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1079 /// memdep query of a load that ends up being a clobbering store.
1080 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
1082 const TargetData &TD) {
1083 // Cannot handle reading from store of first-class aggregate yet.
1084 if (isa<StructType>(DepSI->getOperand(0)->getType()) ||
1085 isa<ArrayType>(DepSI->getOperand(0)->getType()))
1088 Value *StorePtr = DepSI->getPointerOperand();
1089 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1090 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1091 StorePtr, StoreSize, TD);
1094 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
1096 const TargetData &TD) {
1097 // If the mem operation is a non-constant size, we can't handle it.
1098 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1099 if (SizeCst == 0) return -1;
1100 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1102 // If this is memset, we just need to see if the offset is valid in the size
1104 if (MI->getIntrinsicID() == Intrinsic::memset)
1105 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1108 // If we have a memcpy/memmove, the only case we can handle is if this is a
1109 // copy from constant memory. In that case, we can read directly from the
1111 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1113 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1114 if (Src == 0) return -1;
1116 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1117 if (GV == 0 || !GV->isConstant()) return -1;
1119 // See if the access is within the bounds of the transfer.
1120 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1121 MI->getDest(), MemSizeInBits, TD);
1125 // Otherwise, see if we can constant fold a load from the constant with the
1126 // offset applied as appropriate.
1127 Src = ConstantExpr::getBitCast(Src,
1128 llvm::Type::getInt8PtrTy(Src->getContext()));
1129 Constant *OffsetCst =
1130 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1131 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1132 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1133 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1139 /// GetStoreValueForLoad - This function is called when we have a
1140 /// memdep query of a load that ends up being a clobbering store. This means
1141 /// that the store *may* provide bits used by the load but we can't be sure
1142 /// because the pointers don't mustalias. Check this case to see if there is
1143 /// anything more we can do before we give up.
1144 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1146 Instruction *InsertPt, const TargetData &TD){
1147 LLVMContext &Ctx = SrcVal->getType()->getContext();
1149 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1150 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1152 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1154 // Compute which bits of the stored value are being used by the load. Convert
1155 // to an integer type to start with.
1156 if (isa<PointerType>(SrcVal->getType()))
1157 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
1158 if (!isa<IntegerType>(SrcVal->getType()))
1159 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1162 // Shift the bits to the least significant depending on endianness.
1164 if (TD.isLittleEndian())
1165 ShiftAmt = Offset*8;
1167 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1170 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
1172 if (LoadSize != StoreSize)
1173 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1176 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1179 /// GetMemInstValueForLoad - This function is called when we have a
1180 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1181 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1182 const Type *LoadTy, Instruction *InsertPt,
1183 const TargetData &TD){
1184 LLVMContext &Ctx = LoadTy->getContext();
1185 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1187 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1189 // We know that this method is only called when the mem transfer fully
1190 // provides the bits for the load.
1191 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1192 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1193 // independently of what the offset is.
1194 Value *Val = MSI->getValue();
1196 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1198 Value *OneElt = Val;
1200 // Splat the value out to the right number of bits.
1201 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1202 // If we can double the number of bytes set, do it.
1203 if (NumBytesSet*2 <= LoadSize) {
1204 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1205 Val = Builder.CreateOr(Val, ShVal);
1210 // Otherwise insert one byte at a time.
1211 Value *ShVal = Builder.CreateShl(Val, 1*8);
1212 Val = Builder.CreateOr(OneElt, ShVal);
1216 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1219 // Otherwise, this is a memcpy/memmove from a constant global.
1220 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1221 Constant *Src = cast<Constant>(MTI->getSource());
1223 // Otherwise, see if we can constant fold a load from the constant with the
1224 // offset applied as appropriate.
1225 Src = ConstantExpr::getBitCast(Src,
1226 llvm::Type::getInt8PtrTy(Src->getContext()));
1227 Constant *OffsetCst =
1228 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1229 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1230 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1231 return ConstantFoldLoadFromConstPtr(Src, &TD);
1236 struct AvailableValueInBlock {
1237 /// BB - The basic block in question.
1240 SimpleVal, // A simple offsetted value that is accessed.
1241 MemIntrin // A memory intrinsic which is loaded from.
1244 /// V - The value that is live out of the block.
1245 PointerIntPair<Value *, 1, ValType> Val;
1247 /// Offset - The byte offset in Val that is interesting for the load query.
1250 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1251 unsigned Offset = 0) {
1252 AvailableValueInBlock Res;
1254 Res.Val.setPointer(V);
1255 Res.Val.setInt(SimpleVal);
1256 Res.Offset = Offset;
1260 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1261 unsigned Offset = 0) {
1262 AvailableValueInBlock Res;
1264 Res.Val.setPointer(MI);
1265 Res.Val.setInt(MemIntrin);
1266 Res.Offset = Offset;
1270 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1271 Value *getSimpleValue() const {
1272 assert(isSimpleValue() && "Wrong accessor");
1273 return Val.getPointer();
1276 MemIntrinsic *getMemIntrinValue() const {
1277 assert(!isSimpleValue() && "Wrong accessor");
1278 return cast<MemIntrinsic>(Val.getPointer());
1282 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1283 /// construct SSA form, allowing us to eliminate LI. This returns the value
1284 /// that should be used at LI's definition site.
1285 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1286 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1287 const TargetData *TD,
1288 AliasAnalysis *AA) {
1289 SmallVector<PHINode*, 8> NewPHIs;
1290 SSAUpdater SSAUpdate(&NewPHIs);
1291 SSAUpdate.Initialize(LI);
1293 const Type *LoadTy = LI->getType();
1295 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1296 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1297 BasicBlock *BB = AV.BB;
1299 if (SSAUpdate.HasValueForBlock(BB))
1302 unsigned Offset = AV.Offset;
1304 Value *AvailableVal;
1305 if (AV.isSimpleValue()) {
1306 AvailableVal = AV.getSimpleValue();
1307 if (AvailableVal->getType() != LoadTy) {
1308 assert(TD && "Need target data to handle type mismatch case");
1309 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy,
1310 BB->getTerminator(), *TD);
1312 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1313 << *AV.getSimpleValue() << '\n'
1314 << *AvailableVal << '\n' << "\n\n\n");
1317 AvailableVal = GetMemInstValueForLoad(AV.getMemIntrinValue(), Offset,
1318 LoadTy, BB->getTerminator(), *TD);
1319 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1320 << " " << *AV.getMemIntrinValue() << '\n'
1321 << *AvailableVal << '\n' << "\n\n\n");
1323 SSAUpdate.AddAvailableValue(BB, AvailableVal);
1326 // Perform PHI construction.
1327 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1329 // If new PHI nodes were created, notify alias analysis.
1330 if (isa<PointerType>(V->getType()))
1331 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1332 AA->copyValue(LI, NewPHIs[i]);
1337 static bool isLifetimeStart(Instruction *Inst) {
1338 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1339 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1343 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1344 /// non-local by performing PHI construction.
1345 bool GVN::processNonLocalLoad(LoadInst *LI,
1346 SmallVectorImpl<Instruction*> &toErase) {
1347 // Find the non-local dependencies of the load.
1348 SmallVector<NonLocalDepEntry, 64> Deps;
1349 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1351 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
1352 // << Deps.size() << *LI << '\n');
1354 // If we had to process more than one hundred blocks to find the
1355 // dependencies, this load isn't worth worrying about. Optimizing
1356 // it will be too expensive.
1357 if (Deps.size() > 100)
1360 // If we had a phi translation failure, we'll have a single entry which is a
1361 // clobber in the current block. Reject this early.
1362 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
1364 errs() << "GVN: non-local load ";
1365 WriteAsOperand(errs(), LI);
1366 errs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1371 // Filter out useless results (non-locals, etc). Keep track of the blocks
1372 // where we have a value available in repl, also keep track of whether we see
1373 // dependencies that produce an unknown value for the load (such as a call
1374 // that could potentially clobber the load).
1375 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1376 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1378 const TargetData *TD = 0;
1380 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1381 BasicBlock *DepBB = Deps[i].getBB();
1382 MemDepResult DepInfo = Deps[i].getResult();
1384 if (DepInfo.isClobber()) {
1385 // The address being loaded in this non-local block may not be the same as
1386 // the pointer operand of the load if PHI translation occurs. Make sure
1387 // to consider the right address.
1388 Value *Address = Deps[i].getAddress();
1390 // If the dependence is to a store that writes to a superset of the bits
1391 // read by the load, we can extract the bits we need for the load from the
1393 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1395 TD = getAnalysisIfAvailable<TargetData>();
1396 if (TD && Address) {
1397 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1400 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1401 DepSI->getOperand(0),
1408 // If the clobbering value is a memset/memcpy/memmove, see if we can
1409 // forward a value on from it.
1410 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1412 TD = getAnalysisIfAvailable<TargetData>();
1413 if (TD && Address) {
1414 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1417 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1424 UnavailableBlocks.push_back(DepBB);
1428 Instruction *DepInst = DepInfo.getInst();
1430 // Loading the allocation -> undef.
1431 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1432 // Loading immediately after lifetime begin -> undef.
1433 isLifetimeStart(DepInst)) {
1434 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1435 UndefValue::get(LI->getType())));
1439 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1440 // Reject loads and stores that are to the same address but are of
1441 // different types if we have to.
1442 if (S->getOperand(0)->getType() != LI->getType()) {
1444 TD = getAnalysisIfAvailable<TargetData>();
1446 // If the stored value is larger or equal to the loaded value, we can
1448 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1449 LI->getType(), *TD)) {
1450 UnavailableBlocks.push_back(DepBB);
1455 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1460 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1461 // If the types mismatch and we can't handle it, reject reuse of the load.
1462 if (LD->getType() != LI->getType()) {
1464 TD = getAnalysisIfAvailable<TargetData>();
1466 // If the stored value is larger or equal to the loaded value, we can
1468 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1469 UnavailableBlocks.push_back(DepBB);
1473 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1477 UnavailableBlocks.push_back(DepBB);
1481 // If we have no predecessors that produce a known value for this load, exit
1483 if (ValuesPerBlock.empty()) return false;
1485 // If all of the instructions we depend on produce a known value for this
1486 // load, then it is fully redundant and we can use PHI insertion to compute
1487 // its value. Insert PHIs and remove the fully redundant value now.
1488 if (UnavailableBlocks.empty()) {
1489 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1491 // Perform PHI construction.
1492 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1493 VN.getAliasAnalysis());
1494 LI->replaceAllUsesWith(V);
1496 if (isa<PHINode>(V))
1498 if (isa<PointerType>(V->getType()))
1499 MD->invalidateCachedPointerInfo(V);
1500 toErase.push_back(LI);
1505 if (!EnablePRE || !EnableLoadPRE)
1508 // Okay, we have *some* definitions of the value. This means that the value
1509 // is available in some of our (transitive) predecessors. Lets think about
1510 // doing PRE of this load. This will involve inserting a new load into the
1511 // predecessor when it's not available. We could do this in general, but
1512 // prefer to not increase code size. As such, we only do this when we know
1513 // that we only have to insert *one* load (which means we're basically moving
1514 // the load, not inserting a new one).
1516 SmallPtrSet<BasicBlock *, 4> Blockers;
1517 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1518 Blockers.insert(UnavailableBlocks[i]);
1520 // Lets find first basic block with more than one predecessor. Walk backwards
1521 // through predecessors if needed.
1522 BasicBlock *LoadBB = LI->getParent();
1523 BasicBlock *TmpBB = LoadBB;
1525 bool isSinglePred = false;
1526 bool allSingleSucc = true;
1527 while (TmpBB->getSinglePredecessor()) {
1528 isSinglePred = true;
1529 TmpBB = TmpBB->getSinglePredecessor();
1530 if (!TmpBB) // If haven't found any, bail now.
1532 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1534 if (Blockers.count(TmpBB))
1536 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1537 allSingleSucc = false;
1543 // If we have a repl set with LI itself in it, this means we have a loop where
1544 // at least one of the values is LI. Since this means that we won't be able
1545 // to eliminate LI even if we insert uses in the other predecessors, we will
1546 // end up increasing code size. Reject this by scanning for LI.
1547 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1548 if (ValuesPerBlock[i].isSimpleValue() &&
1549 ValuesPerBlock[i].getSimpleValue() == LI)
1552 // FIXME: It is extremely unclear what this loop is doing, other than
1553 // artificially restricting loadpre.
1556 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1557 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1558 if (AV.isSimpleValue())
1559 // "Hot" Instruction is in some loop (because it dominates its dep.
1561 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1562 if (DT->dominates(LI, I)) {
1568 // We are interested only in "hot" instructions. We don't want to do any
1569 // mis-optimizations here.
1574 // Okay, we have some hope :). Check to see if the loaded value is fully
1575 // available in all but one predecessor.
1576 // FIXME: If we could restructure the CFG, we could make a common pred with
1577 // all the preds that don't have an available LI and insert a new load into
1579 BasicBlock *UnavailablePred = 0;
1581 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1582 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1583 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1584 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1585 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1587 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1589 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1592 // If this load is not available in multiple predecessors, reject it.
1593 if (UnavailablePred && UnavailablePred != *PI)
1595 UnavailablePred = *PI;
1598 assert(UnavailablePred != 0 &&
1599 "Fully available value should be eliminated above!");
1601 // We don't currently handle critical edges :(
1602 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1603 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1604 << UnavailablePred->getName() << "': " << *LI << '\n');
1608 // Do PHI translation to get its value in the predecessor if necessary. The
1609 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1611 SmallVector<Instruction*, 8> NewInsts;
1613 // If all preds have a single successor, then we know it is safe to insert the
1614 // load on the pred (?!?), so we can insert code to materialize the pointer if
1615 // it is not available.
1616 PHITransAddr Address(LI->getOperand(0), TD);
1618 if (allSingleSucc) {
1619 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1622 Address.PHITranslateValue(LoadBB, UnavailablePred);
1623 LoadPtr = Address.getAddr();
1625 // Make sure the value is live in the predecessor.
1626 if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr))
1627 if (!DT->dominates(Inst->getParent(), UnavailablePred))
1631 // If we couldn't find or insert a computation of this phi translated value,
1634 assert(NewInsts.empty() && "Shouldn't insert insts on failure");
1635 DEBUG(errs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1636 << *LI->getOperand(0) << "\n");
1640 // Assign value numbers to these new instructions.
1641 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1642 // FIXME: We really _ought_ to insert these value numbers into their
1643 // parent's availability map. However, in doing so, we risk getting into
1644 // ordering issues. If a block hasn't been processed yet, we would be
1645 // marking a value as AVAIL-IN, which isn't what we intend.
1646 VN.lookup_or_add(NewInsts[i]);
1649 // Make sure it is valid to move this load here. We have to watch out for:
1650 // @1 = getelementptr (i8* p, ...
1651 // test p and branch if == 0
1653 // It is valid to have the getelementptr before the test, even if p can be 0,
1654 // as getelementptr only does address arithmetic.
1655 // If we are not pushing the value through any multiple-successor blocks
1656 // we do not have this case. Otherwise, check that the load is safe to
1657 // put anywhere; this can be improved, but should be conservatively safe.
1658 if (!allSingleSucc &&
1659 // FIXME: REEVALUTE THIS.
1660 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) {
1661 assert(NewInsts.empty() && "Should not have inserted instructions");
1665 // Okay, we can eliminate this load by inserting a reload in the predecessor
1666 // and using PHI construction to get the value in the other predecessors, do
1668 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1669 DEBUG(if (!NewInsts.empty())
1670 errs() << "INSERTED " << NewInsts.size() << " INSTS: "
1671 << *NewInsts.back() << '\n');
1673 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1675 UnavailablePred->getTerminator());
1677 // Add the newly created load.
1678 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));
1680 // Perform PHI construction.
1681 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1682 VN.getAliasAnalysis());
1683 LI->replaceAllUsesWith(V);
1684 if (isa<PHINode>(V))
1686 if (isa<PointerType>(V->getType()))
1687 MD->invalidateCachedPointerInfo(V);
1688 toErase.push_back(LI);
1693 /// processLoad - Attempt to eliminate a load, first by eliminating it
1694 /// locally, and then attempting non-local elimination if that fails.
1695 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1699 if (L->isVolatile())
1702 // ... to a pointer that has been loaded from before...
1703 MemDepResult Dep = MD->getDependency(L);
1705 // If the value isn't available, don't do anything!
1706 if (Dep.isClobber()) {
1707 // Check to see if we have something like this:
1708 // store i32 123, i32* %P
1709 // %A = bitcast i32* %P to i8*
1710 // %B = gep i8* %A, i32 1
1713 // We could do that by recognizing if the clobber instructions are obviously
1714 // a common base + constant offset, and if the previous store (or memset)
1715 // completely covers this load. This sort of thing can happen in bitfield
1717 Value *AvailVal = 0;
1718 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1719 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1720 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1721 L->getPointerOperand(),
1724 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1725 L->getType(), L, *TD);
1728 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1729 // a value on from it.
1730 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1731 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1732 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1733 L->getPointerOperand(),
1736 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1741 DEBUG(errs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1742 << *AvailVal << '\n' << *L << "\n\n\n");
1744 // Replace the load!
1745 L->replaceAllUsesWith(AvailVal);
1746 if (isa<PointerType>(AvailVal->getType()))
1747 MD->invalidateCachedPointerInfo(AvailVal);
1748 toErase.push_back(L);
1754 // fast print dep, using operator<< on instruction would be too slow
1755 errs() << "GVN: load ";
1756 WriteAsOperand(errs(), L);
1757 Instruction *I = Dep.getInst();
1758 errs() << " is clobbered by " << *I << '\n';
1763 // If it is defined in another block, try harder.
1764 if (Dep.isNonLocal())
1765 return processNonLocalLoad(L, toErase);
1767 Instruction *DepInst = Dep.getInst();
1768 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1769 Value *StoredVal = DepSI->getOperand(0);
1771 // The store and load are to a must-aliased pointer, but they may not
1772 // actually have the same type. See if we know how to reuse the stored
1773 // value (depending on its type).
1774 const TargetData *TD = 0;
1775 if (StoredVal->getType() != L->getType()) {
1776 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1777 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1782 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1783 << '\n' << *L << "\n\n\n");
1790 L->replaceAllUsesWith(StoredVal);
1791 if (isa<PointerType>(StoredVal->getType()))
1792 MD->invalidateCachedPointerInfo(StoredVal);
1793 toErase.push_back(L);
1798 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1799 Value *AvailableVal = DepLI;
1801 // The loads are of a must-aliased pointer, but they may not actually have
1802 // the same type. See if we know how to reuse the previously loaded value
1803 // (depending on its type).
1804 const TargetData *TD = 0;
1805 if (DepLI->getType() != L->getType()) {
1806 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1807 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1808 if (AvailableVal == 0)
1811 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1812 << "\n" << *L << "\n\n\n");
1819 L->replaceAllUsesWith(AvailableVal);
1820 if (isa<PointerType>(DepLI->getType()))
1821 MD->invalidateCachedPointerInfo(DepLI);
1822 toErase.push_back(L);
1827 // If this load really doesn't depend on anything, then we must be loading an
1828 // undef value. This can happen when loading for a fresh allocation with no
1829 // intervening stores, for example.
1830 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1831 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1832 toErase.push_back(L);
1837 // If this load occurs either right after a lifetime begin,
1838 // then the loaded value is undefined.
1839 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1840 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1841 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1842 toErase.push_back(L);
1851 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1852 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1853 if (I == localAvail.end())
1856 ValueNumberScope *Locals = I->second;
1858 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1859 if (I != Locals->table.end())
1861 Locals = Locals->parent;
1868 /// processInstruction - When calculating availability, handle an instruction
1869 /// by inserting it into the appropriate sets
1870 bool GVN::processInstruction(Instruction *I,
1871 SmallVectorImpl<Instruction*> &toErase) {
1872 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1873 bool Changed = processLoad(LI, toErase);
1876 unsigned Num = VN.lookup_or_add(LI);
1877 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1883 uint32_t NextNum = VN.getNextUnusedValueNumber();
1884 unsigned Num = VN.lookup_or_add(I);
1886 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1887 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1889 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1892 Value *BranchCond = BI->getCondition();
1893 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1895 BasicBlock *TrueSucc = BI->getSuccessor(0);
1896 BasicBlock *FalseSucc = BI->getSuccessor(1);
1898 if (TrueSucc->getSinglePredecessor())
1899 localAvail[TrueSucc]->table[CondVN] =
1900 ConstantInt::getTrue(TrueSucc->getContext());
1901 if (FalseSucc->getSinglePredecessor())
1902 localAvail[FalseSucc]->table[CondVN] =
1903 ConstantInt::getFalse(TrueSucc->getContext());
1907 // Allocations are always uniquely numbered, so we can save time and memory
1908 // by fast failing them.
1909 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1910 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1914 // Collapse PHI nodes
1915 if (PHINode* p = dyn_cast<PHINode>(I)) {
1916 Value *constVal = CollapsePhi(p);
1919 p->replaceAllUsesWith(constVal);
1920 if (MD && isa<PointerType>(constVal->getType()))
1921 MD->invalidateCachedPointerInfo(constVal);
1924 toErase.push_back(p);
1926 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1929 // If the number we were assigned was a brand new VN, then we don't
1930 // need to do a lookup to see if the number already exists
1931 // somewhere in the domtree: it can't!
1932 } else if (Num == NextNum) {
1933 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1935 // Perform fast-path value-number based elimination of values inherited from
1937 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1940 I->replaceAllUsesWith(repl);
1941 if (MD && isa<PointerType>(repl->getType()))
1942 MD->invalidateCachedPointerInfo(repl);
1943 toErase.push_back(I);
1947 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1953 /// runOnFunction - This is the main transformation entry point for a function.
1954 bool GVN::runOnFunction(Function& F) {
1956 MD = &getAnalysis<MemoryDependenceAnalysis>();
1957 DT = &getAnalysis<DominatorTree>();
1958 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1962 bool Changed = false;
1963 bool ShouldContinue = true;
1965 // Merge unconditional branches, allowing PRE to catch more
1966 // optimization opportunities.
1967 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1968 BasicBlock *BB = FI;
1970 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1971 if (removedBlock) NumGVNBlocks++;
1973 Changed |= removedBlock;
1976 unsigned Iteration = 0;
1978 while (ShouldContinue) {
1979 DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
1980 ShouldContinue = iterateOnFunction(F);
1981 Changed |= ShouldContinue;
1986 bool PREChanged = true;
1987 while (PREChanged) {
1988 PREChanged = performPRE(F);
1989 Changed |= PREChanged;
1992 // FIXME: Should perform GVN again after PRE does something. PRE can move
1993 // computations into blocks where they become fully redundant. Note that
1994 // we can't do this until PRE's critical edge splitting updates memdep.
1995 // Actually, when this happens, we should just fully integrate PRE into GVN.
1997 cleanupGlobalSets();
2003 bool GVN::processBlock(BasicBlock *BB) {
2004 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
2005 // incrementing BI before processing an instruction).
2006 SmallVector<Instruction*, 8> toErase;
2007 bool ChangedFunction = false;
2009 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2011 ChangedFunction |= processInstruction(BI, toErase);
2012 if (toErase.empty()) {
2017 // If we need some instructions deleted, do it now.
2018 NumGVNInstr += toErase.size();
2020 // Avoid iterator invalidation.
2021 bool AtStart = BI == BB->begin();
2025 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2026 E = toErase.end(); I != E; ++I) {
2027 DEBUG(errs() << "GVN removed: " << **I << '\n');
2028 if (MD) MD->removeInstruction(*I);
2029 (*I)->eraseFromParent();
2030 DEBUG(verifyRemoved(*I));
2040 return ChangedFunction;
2043 /// performPRE - Perform a purely local form of PRE that looks for diamond
2044 /// control flow patterns and attempts to perform simple PRE at the join point.
2045 bool GVN::performPRE(Function &F) {
2046 bool Changed = false;
2047 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
2048 DenseMap<BasicBlock*, Value*> predMap;
2049 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2050 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2051 BasicBlock *CurrentBlock = *DI;
2053 // Nothing to PRE in the entry block.
2054 if (CurrentBlock == &F.getEntryBlock()) continue;
2056 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2057 BE = CurrentBlock->end(); BI != BE; ) {
2058 Instruction *CurInst = BI++;
2060 if (isa<AllocaInst>(CurInst) ||
2061 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2062 CurInst->getType()->isVoidTy() ||
2063 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2064 isa<DbgInfoIntrinsic>(CurInst))
2067 uint32_t ValNo = VN.lookup(CurInst);
2069 // Look for the predecessors for PRE opportunities. We're
2070 // only trying to solve the basic diamond case, where
2071 // a value is computed in the successor and one predecessor,
2072 // but not the other. We also explicitly disallow cases
2073 // where the successor is its own predecessor, because they're
2074 // more complicated to get right.
2075 unsigned NumWith = 0;
2076 unsigned NumWithout = 0;
2077 BasicBlock *PREPred = 0;
2080 for (pred_iterator PI = pred_begin(CurrentBlock),
2081 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2082 // We're not interested in PRE where the block is its
2083 // own predecessor, on in blocks with predecessors
2084 // that are not reachable.
2085 if (*PI == CurrentBlock) {
2088 } else if (!localAvail.count(*PI)) {
2093 DenseMap<uint32_t, Value*>::iterator predV =
2094 localAvail[*PI]->table.find(ValNo);
2095 if (predV == localAvail[*PI]->table.end()) {
2098 } else if (predV->second == CurInst) {
2101 predMap[*PI] = predV->second;
2106 // Don't do PRE when it might increase code size, i.e. when
2107 // we would need to insert instructions in more than one pred.
2108 if (NumWithout != 1 || NumWith == 0)
2111 // Don't do PRE across indirect branch.
2112 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2115 // We can't do PRE safely on a critical edge, so instead we schedule
2116 // the edge to be split and perform the PRE the next time we iterate
2118 unsigned SuccNum = 0;
2119 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
2121 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
2126 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2127 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2131 // Instantiate the expression the in predecessor that lacked it.
2132 // Because we are going top-down through the block, all value numbers
2133 // will be available in the predecessor by the time we need them. Any
2134 // that weren't original present will have been instantiated earlier
2136 Instruction *PREInstr = CurInst->clone();
2137 bool success = true;
2138 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2139 Value *Op = PREInstr->getOperand(i);
2140 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2143 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2144 PREInstr->setOperand(i, V);
2151 // Fail out if we encounter an operand that is not available in
2152 // the PRE predecessor. This is typically because of loads which
2153 // are not value numbered precisely.
2156 DEBUG(verifyRemoved(PREInstr));
2160 PREInstr->insertBefore(PREPred->getTerminator());
2161 PREInstr->setName(CurInst->getName() + ".pre");
2162 predMap[PREPred] = PREInstr;
2163 VN.add(PREInstr, ValNo);
2166 // Update the availability map to include the new instruction.
2167 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
2169 // Create a PHI to make the value available in this block.
2170 PHINode* Phi = PHINode::Create(CurInst->getType(),
2171 CurInst->getName() + ".pre-phi",
2172 CurrentBlock->begin());
2173 for (pred_iterator PI = pred_begin(CurrentBlock),
2174 PE = pred_end(CurrentBlock); PI != PE; ++PI)
2175 Phi->addIncoming(predMap[*PI], *PI);
2178 localAvail[CurrentBlock]->table[ValNo] = Phi;
2180 CurInst->replaceAllUsesWith(Phi);
2181 if (MD && isa<PointerType>(Phi->getType()))
2182 MD->invalidateCachedPointerInfo(Phi);
2185 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
2186 if (MD) MD->removeInstruction(CurInst);
2187 CurInst->eraseFromParent();
2188 DEBUG(verifyRemoved(CurInst));
2193 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
2194 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
2195 SplitCriticalEdge(I->first, I->second, this);
2197 return Changed || toSplit.size();
2200 /// iterateOnFunction - Executes one iteration of GVN
2201 bool GVN::iterateOnFunction(Function &F) {
2202 cleanupGlobalSets();
2204 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2205 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
2207 localAvail[DI->getBlock()] =
2208 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
2210 localAvail[DI->getBlock()] = new ValueNumberScope(0);
2213 // Top-down walk of the dominator tree
2214 bool Changed = false;
2216 // Needed for value numbering with phi construction to work.
2217 ReversePostOrderTraversal<Function*> RPOT(&F);
2218 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2219 RE = RPOT.end(); RI != RE; ++RI)
2220 Changed |= processBlock(*RI);
2222 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2223 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2224 Changed |= processBlock(DI->getBlock());
2230 void GVN::cleanupGlobalSets() {
2233 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2234 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2239 /// verifyRemoved - Verify that the specified instruction does not occur in our
2240 /// internal data structures.
2241 void GVN::verifyRemoved(const Instruction *Inst) const {
2242 VN.verifyRemoved(Inst);
2244 // Walk through the value number scope to make sure the instruction isn't
2245 // ferreted away in it.
2246 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
2247 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2248 const ValueNumberScope *VNS = I->second;
2251 for (DenseMap<uint32_t, Value*>::const_iterator
2252 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2253 assert(II->second != Inst && "Inst still in value numbering scope!");