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/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/LLVMContext.h"
27 #include "llvm/Operator.h"
28 #include "llvm/Value.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/DepthFirstIterator.h"
31 #include "llvm/ADT/PostOrderIterator.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/ConstantFolding.h"
37 #include "llvm/Analysis/Dominators.h"
38 #include "llvm/Analysis/MemoryBuiltins.h"
39 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
40 #include "llvm/Analysis/PHITransAddr.h"
41 #include "llvm/Support/CFG.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/GetElementPtrTypeIterator.h"
46 #include "llvm/Support/IRBuilder.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Target/TargetData.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Transforms/Utils/SSAUpdater.h"
54 STATISTIC(NumGVNInstr, "Number of instructions deleted");
55 STATISTIC(NumGVNLoad, "Number of loads deleted");
56 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
57 STATISTIC(NumGVNBlocks, "Number of blocks merged");
58 STATISTIC(NumPRELoad, "Number of loads PRE'd");
60 static cl::opt<bool> EnablePRE("enable-pre",
61 cl::init(true), cl::Hidden);
62 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
64 //===----------------------------------------------------------------------===//
66 //===----------------------------------------------------------------------===//
68 /// This class holds the mapping between values and value numbers. It is used
69 /// as an efficient mechanism to determine the expression-wise equivalence of
73 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
74 UDIV, SDIV, FDIV, UREM, SREM,
75 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
76 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
77 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
78 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
79 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
80 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
81 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
82 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
83 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
84 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
86 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)
104 if (varargs.size() != other.varargs.size())
107 for (size_t i = 0; i < varargs.size(); ++i)
108 if (varargs[i] != other.varargs[i])
115 bool operator!=(const Expression &other) const {
116 return !(*this == other);
122 DenseMap<Value*, uint32_t> valueNumbering;
123 DenseMap<Expression, uint32_t> expressionNumbering;
125 MemoryDependenceAnalysis* MD;
128 uint32_t nextValueNumber;
130 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
131 Expression::ExpressionOpcode getOpcode(CmpInst* C);
132 Expression::ExpressionOpcode getOpcode(CastInst* C);
133 Expression create_expression(BinaryOperator* BO);
134 Expression create_expression(CmpInst* C);
135 Expression create_expression(ShuffleVectorInst* V);
136 Expression create_expression(ExtractElementInst* C);
137 Expression create_expression(InsertElementInst* V);
138 Expression create_expression(SelectInst* V);
139 Expression create_expression(CastInst* C);
140 Expression create_expression(GetElementPtrInst* G);
141 Expression create_expression(CallInst* C);
142 Expression create_expression(Constant* C);
143 Expression create_expression(ExtractValueInst* C);
144 Expression create_expression(InsertValueInst* C);
146 uint32_t lookup_or_add_call(CallInst* C);
148 ValueTable() : nextValueNumber(1) { }
149 uint32_t lookup_or_add(Value *V);
150 uint32_t lookup(Value *V) const;
151 void add(Value *V, uint32_t num);
153 void erase(Value *v);
155 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
156 AliasAnalysis *getAliasAnalysis() const { return AA; }
157 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
158 void setDomTree(DominatorTree* D) { DT = D; }
159 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
160 void verifyRemoved(const Value *) const;
165 template <> struct DenseMapInfo<Expression> {
166 static inline Expression getEmptyKey() {
167 return Expression(Expression::EMPTY);
170 static inline Expression getTombstoneKey() {
171 return Expression(Expression::TOMBSTONE);
174 static unsigned getHashValue(const Expression e) {
175 unsigned hash = e.opcode;
177 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
178 (unsigned)((uintptr_t)e.type >> 9));
180 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
181 E = e.varargs.end(); I != E; ++I)
182 hash = *I + hash * 37;
184 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
185 (unsigned)((uintptr_t)e.function >> 9)) +
190 static bool isEqual(const Expression &LHS, const Expression &RHS) {
196 struct isPodLike<Expression> { static const bool value = true; };
200 //===----------------------------------------------------------------------===//
201 // ValueTable Internal Functions
202 //===----------------------------------------------------------------------===//
203 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
204 switch(BO->getOpcode()) {
205 default: // THIS SHOULD NEVER HAPPEN
206 llvm_unreachable("Binary operator with unknown opcode?");
207 case Instruction::Add: return Expression::ADD;
208 case Instruction::FAdd: return Expression::FADD;
209 case Instruction::Sub: return Expression::SUB;
210 case Instruction::FSub: return Expression::FSUB;
211 case Instruction::Mul: return Expression::MUL;
212 case Instruction::FMul: return Expression::FMUL;
213 case Instruction::UDiv: return Expression::UDIV;
214 case Instruction::SDiv: return Expression::SDIV;
215 case Instruction::FDiv: return Expression::FDIV;
216 case Instruction::URem: return Expression::UREM;
217 case Instruction::SRem: return Expression::SREM;
218 case Instruction::FRem: return Expression::FREM;
219 case Instruction::Shl: return Expression::SHL;
220 case Instruction::LShr: return Expression::LSHR;
221 case Instruction::AShr: return Expression::ASHR;
222 case Instruction::And: return Expression::AND;
223 case Instruction::Or: return Expression::OR;
224 case Instruction::Xor: return Expression::XOR;
228 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
229 if (isa<ICmpInst>(C)) {
230 switch (C->getPredicate()) {
231 default: // THIS SHOULD NEVER HAPPEN
232 llvm_unreachable("Comparison with unknown predicate?");
233 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
234 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
235 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
236 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
237 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
238 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
239 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
240 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
241 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
242 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
245 switch (C->getPredicate()) {
246 default: // THIS SHOULD NEVER HAPPEN
247 llvm_unreachable("Comparison with unknown predicate?");
248 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
249 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
250 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
251 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
252 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
253 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
254 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
255 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
256 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
257 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
258 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
259 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
260 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
261 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
266 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
267 switch(C->getOpcode()) {
268 default: // THIS SHOULD NEVER HAPPEN
269 llvm_unreachable("Cast operator with unknown opcode?");
270 case Instruction::Trunc: return Expression::TRUNC;
271 case Instruction::ZExt: return Expression::ZEXT;
272 case Instruction::SExt: return Expression::SEXT;
273 case Instruction::FPToUI: return Expression::FPTOUI;
274 case Instruction::FPToSI: return Expression::FPTOSI;
275 case Instruction::UIToFP: return Expression::UITOFP;
276 case Instruction::SIToFP: return Expression::SITOFP;
277 case Instruction::FPTrunc: return Expression::FPTRUNC;
278 case Instruction::FPExt: return Expression::FPEXT;
279 case Instruction::PtrToInt: return Expression::PTRTOINT;
280 case Instruction::IntToPtr: return Expression::INTTOPTR;
281 case Instruction::BitCast: return Expression::BITCAST;
285 Expression ValueTable::create_expression(CallInst* C) {
288 e.type = C->getType();
289 e.function = C->getCalledFunction();
290 e.opcode = Expression::CALL;
292 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
294 e.varargs.push_back(lookup_or_add(*I));
299 Expression ValueTable::create_expression(BinaryOperator* BO) {
301 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
302 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
304 e.type = BO->getType();
305 e.opcode = getOpcode(BO);
310 Expression ValueTable::create_expression(CmpInst* C) {
313 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
314 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
316 e.type = C->getType();
317 e.opcode = getOpcode(C);
322 Expression ValueTable::create_expression(CastInst* C) {
325 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
327 e.type = C->getType();
328 e.opcode = getOpcode(C);
333 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
336 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
337 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
338 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
340 e.type = S->getType();
341 e.opcode = Expression::SHUFFLE;
346 Expression ValueTable::create_expression(ExtractElementInst* E) {
349 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
350 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
352 e.type = E->getType();
353 e.opcode = Expression::EXTRACT;
358 Expression ValueTable::create_expression(InsertElementInst* I) {
361 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
362 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
363 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
365 e.type = I->getType();
366 e.opcode = Expression::INSERT;
371 Expression ValueTable::create_expression(SelectInst* I) {
374 e.varargs.push_back(lookup_or_add(I->getCondition()));
375 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
376 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
378 e.type = I->getType();
379 e.opcode = Expression::SELECT;
384 Expression ValueTable::create_expression(GetElementPtrInst* G) {
387 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
389 e.type = G->getType();
390 e.opcode = Expression::GEP;
392 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
394 e.varargs.push_back(lookup_or_add(*I));
399 Expression ValueTable::create_expression(ExtractValueInst* E) {
402 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
403 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
405 e.varargs.push_back(*II);
407 e.type = E->getType();
408 e.opcode = Expression::EXTRACTVALUE;
413 Expression ValueTable::create_expression(InsertValueInst* E) {
416 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
417 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
418 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
420 e.varargs.push_back(*II);
422 e.type = E->getType();
423 e.opcode = Expression::INSERTVALUE;
428 //===----------------------------------------------------------------------===//
429 // ValueTable External Functions
430 //===----------------------------------------------------------------------===//
432 /// add - Insert a value into the table with a specified value number.
433 void ValueTable::add(Value *V, uint32_t num) {
434 valueNumbering.insert(std::make_pair(V, num));
437 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
438 if (AA->doesNotAccessMemory(C)) {
439 Expression exp = create_expression(C);
440 uint32_t& e = expressionNumbering[exp];
441 if (!e) e = nextValueNumber++;
442 valueNumbering[C] = e;
444 } else if (AA->onlyReadsMemory(C)) {
445 Expression exp = create_expression(C);
446 uint32_t& e = expressionNumbering[exp];
448 e = nextValueNumber++;
449 valueNumbering[C] = e;
453 e = nextValueNumber++;
454 valueNumbering[C] = e;
458 MemDepResult local_dep = MD->getDependency(C);
460 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
461 valueNumbering[C] = nextValueNumber;
462 return nextValueNumber++;
465 if (local_dep.isDef()) {
466 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
468 if (local_cdep->getNumOperands() != C->getNumOperands()) {
469 valueNumbering[C] = nextValueNumber;
470 return nextValueNumber++;
473 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
474 uint32_t c_vn = lookup_or_add(C->getOperand(i));
475 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
477 valueNumbering[C] = nextValueNumber;
478 return nextValueNumber++;
482 uint32_t v = lookup_or_add(local_cdep);
483 valueNumbering[C] = v;
488 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
489 MD->getNonLocalCallDependency(CallSite(C));
490 // FIXME: call/call dependencies for readonly calls should return def, not
491 // clobber! Move the checking logic to MemDep!
494 // Check to see if we have a single dominating call instruction that is
496 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
497 const NonLocalDepEntry *I = &deps[i];
498 // Ignore non-local dependencies.
499 if (I->getResult().isNonLocal())
502 // We don't handle non-depedencies. If we already have a call, reject
503 // instruction dependencies.
504 if (I->getResult().isClobber() || cdep != 0) {
509 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
510 // FIXME: All duplicated with non-local case.
511 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
512 cdep = NonLocalDepCall;
521 valueNumbering[C] = nextValueNumber;
522 return nextValueNumber++;
525 if (cdep->getNumOperands() != C->getNumOperands()) {
526 valueNumbering[C] = nextValueNumber;
527 return nextValueNumber++;
529 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
530 uint32_t c_vn = lookup_or_add(C->getOperand(i));
531 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
533 valueNumbering[C] = nextValueNumber;
534 return nextValueNumber++;
538 uint32_t v = lookup_or_add(cdep);
539 valueNumbering[C] = v;
543 valueNumbering[C] = nextValueNumber;
544 return nextValueNumber++;
548 /// lookup_or_add - Returns the value number for the specified value, assigning
549 /// it a new number if it did not have one before.
550 uint32_t ValueTable::lookup_or_add(Value *V) {
551 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
552 if (VI != valueNumbering.end())
555 if (!isa<Instruction>(V)) {
556 valueNumbering[V] = nextValueNumber;
557 return nextValueNumber++;
560 Instruction* I = cast<Instruction>(V);
562 switch (I->getOpcode()) {
563 case Instruction::Call:
564 return lookup_or_add_call(cast<CallInst>(I));
565 case Instruction::Add:
566 case Instruction::FAdd:
567 case Instruction::Sub:
568 case Instruction::FSub:
569 case Instruction::Mul:
570 case Instruction::FMul:
571 case Instruction::UDiv:
572 case Instruction::SDiv:
573 case Instruction::FDiv:
574 case Instruction::URem:
575 case Instruction::SRem:
576 case Instruction::FRem:
577 case Instruction::Shl:
578 case Instruction::LShr:
579 case Instruction::AShr:
580 case Instruction::And:
581 case Instruction::Or :
582 case Instruction::Xor:
583 exp = create_expression(cast<BinaryOperator>(I));
585 case Instruction::ICmp:
586 case Instruction::FCmp:
587 exp = create_expression(cast<CmpInst>(I));
589 case Instruction::Trunc:
590 case Instruction::ZExt:
591 case Instruction::SExt:
592 case Instruction::FPToUI:
593 case Instruction::FPToSI:
594 case Instruction::UIToFP:
595 case Instruction::SIToFP:
596 case Instruction::FPTrunc:
597 case Instruction::FPExt:
598 case Instruction::PtrToInt:
599 case Instruction::IntToPtr:
600 case Instruction::BitCast:
601 exp = create_expression(cast<CastInst>(I));
603 case Instruction::Select:
604 exp = create_expression(cast<SelectInst>(I));
606 case Instruction::ExtractElement:
607 exp = create_expression(cast<ExtractElementInst>(I));
609 case Instruction::InsertElement:
610 exp = create_expression(cast<InsertElementInst>(I));
612 case Instruction::ShuffleVector:
613 exp = create_expression(cast<ShuffleVectorInst>(I));
615 case Instruction::ExtractValue:
616 exp = create_expression(cast<ExtractValueInst>(I));
618 case Instruction::InsertValue:
619 exp = create_expression(cast<InsertValueInst>(I));
621 case Instruction::GetElementPtr:
622 exp = create_expression(cast<GetElementPtrInst>(I));
625 valueNumbering[V] = nextValueNumber;
626 return nextValueNumber++;
629 uint32_t& e = expressionNumbering[exp];
630 if (!e) e = nextValueNumber++;
631 valueNumbering[V] = e;
635 /// lookup - Returns the value number of the specified value. Fails if
636 /// the value has not yet been numbered.
637 uint32_t ValueTable::lookup(Value *V) const {
638 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
639 assert(VI != valueNumbering.end() && "Value not numbered?");
643 /// clear - Remove all entries from the ValueTable
644 void ValueTable::clear() {
645 valueNumbering.clear();
646 expressionNumbering.clear();
650 /// erase - Remove a value from the value numbering
651 void ValueTable::erase(Value *V) {
652 valueNumbering.erase(V);
655 /// verifyRemoved - Verify that the value is removed from all internal data
657 void ValueTable::verifyRemoved(const Value *V) const {
658 for (DenseMap<Value*, uint32_t>::const_iterator
659 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
660 assert(I->first != V && "Inst still occurs in value numbering map!");
664 //===----------------------------------------------------------------------===//
666 //===----------------------------------------------------------------------===//
669 struct ValueNumberScope {
670 ValueNumberScope* parent;
671 DenseMap<uint32_t, Value*> table;
673 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
679 class GVN : public FunctionPass {
680 bool runOnFunction(Function &F);
682 static char ID; // Pass identification, replacement for typeid
683 explicit GVN(bool nopre = false, bool noloads = false)
684 : FunctionPass(&ID), NoPRE(nopre), NoLoads(noloads), MD(0) { }
689 MemoryDependenceAnalysis *MD;
693 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
695 // This transformation requires dominator postdominator info
696 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
697 AU.addRequired<DominatorTree>();
699 AU.addRequired<MemoryDependenceAnalysis>();
700 AU.addRequired<AliasAnalysis>();
702 AU.addPreserved<DominatorTree>();
703 AU.addPreserved<AliasAnalysis>();
707 // FIXME: eliminate or document these better
708 bool processLoad(LoadInst* L,
709 SmallVectorImpl<Instruction*> &toErase);
710 bool processInstruction(Instruction *I,
711 SmallVectorImpl<Instruction*> &toErase);
712 bool processNonLocalLoad(LoadInst* L,
713 SmallVectorImpl<Instruction*> &toErase);
714 bool processBlock(BasicBlock *BB);
715 void dump(DenseMap<uint32_t, Value*>& d);
716 bool iterateOnFunction(Function &F);
717 Value *CollapsePhi(PHINode* p);
718 bool performPRE(Function& F);
719 Value *lookupNumber(BasicBlock *BB, uint32_t num);
720 void cleanupGlobalSets();
721 void verifyRemoved(const Instruction *I) const;
727 // createGVNPass - The public interface to this file...
728 FunctionPass *llvm::createGVNPass(bool NoPRE, bool NoLoads) {
729 return new GVN(NoPRE, NoLoads);
732 static RegisterPass<GVN> X("gvn",
733 "Global Value Numbering");
735 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
737 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
738 E = d.end(); I != E; ++I) {
739 errs() << I->first << "\n";
745 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
746 if (!isa<PHINode>(inst))
749 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
751 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
752 if (use_phi->getParent() == inst->getParent())
758 Value *GVN::CollapsePhi(PHINode *PN) {
759 Value *ConstVal = PN->hasConstantValue(DT);
760 if (!ConstVal) return 0;
762 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
766 if (DT->dominates(Inst, PN))
767 if (isSafeReplacement(PN, Inst))
772 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
773 /// we're analyzing is fully available in the specified block. As we go, keep
774 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
775 /// map is actually a tri-state map with the following values:
776 /// 0) we know the block *is not* fully available.
777 /// 1) we know the block *is* fully available.
778 /// 2) we do not know whether the block is fully available or not, but we are
779 /// currently speculating that it will be.
780 /// 3) we are speculating for this block and have used that to speculate for
782 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
783 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
784 // Optimistically assume that the block is fully available and check to see
785 // if we already know about this block in one lookup.
786 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
787 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
789 // If the entry already existed for this block, return the precomputed value.
791 // If this is a speculative "available" value, mark it as being used for
792 // speculation of other blocks.
793 if (IV.first->second == 2)
794 IV.first->second = 3;
795 return IV.first->second != 0;
798 // Otherwise, see if it is fully available in all predecessors.
799 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
801 // If this block has no predecessors, it isn't live-in here.
803 goto SpeculationFailure;
805 for (; PI != PE; ++PI)
806 // If the value isn't fully available in one of our predecessors, then it
807 // isn't fully available in this block either. Undo our previous
808 // optimistic assumption and bail out.
809 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
810 goto SpeculationFailure;
814 // SpeculationFailure - If we get here, we found out that this is not, after
815 // all, a fully-available block. We have a problem if we speculated on this and
816 // used the speculation to mark other blocks as available.
818 char &BBVal = FullyAvailableBlocks[BB];
820 // If we didn't speculate on this, just return with it set to false.
826 // If we did speculate on this value, we could have blocks set to 1 that are
827 // incorrect. Walk the (transitive) successors of this block and mark them as
829 SmallVector<BasicBlock*, 32> BBWorklist;
830 BBWorklist.push_back(BB);
832 while (!BBWorklist.empty()) {
833 BasicBlock *Entry = BBWorklist.pop_back_val();
834 // Note that this sets blocks to 0 (unavailable) if they happen to not
835 // already be in FullyAvailableBlocks. This is safe.
836 char &EntryVal = FullyAvailableBlocks[Entry];
837 if (EntryVal == 0) continue; // Already unavailable.
839 // Mark as unavailable.
842 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
843 BBWorklist.push_back(*I);
850 /// CanCoerceMustAliasedValueToLoad - Return true if
851 /// CoerceAvailableValueToLoadType will succeed.
852 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
854 const TargetData &TD) {
855 // If the loaded or stored value is an first class array or struct, don't try
856 // to transform them. We need to be able to bitcast to integer.
857 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
858 isa<StructType>(StoredVal->getType()) ||
859 isa<ArrayType>(StoredVal->getType()))
862 // The store has to be at least as big as the load.
863 if (TD.getTypeSizeInBits(StoredVal->getType()) <
864 TD.getTypeSizeInBits(LoadTy))
871 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
872 /// then a load from a must-aliased pointer of a different type, try to coerce
873 /// the stored value. LoadedTy is the type of the load we want to replace and
874 /// InsertPt is the place to insert new instructions.
876 /// If we can't do it, return null.
877 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
878 const Type *LoadedTy,
879 Instruction *InsertPt,
880 const TargetData &TD) {
881 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
884 const Type *StoredValTy = StoredVal->getType();
886 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
887 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
889 // If the store and reload are the same size, we can always reuse it.
890 if (StoreSize == LoadSize) {
891 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
892 // Pointer to Pointer -> use bitcast.
893 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
896 // Convert source pointers to integers, which can be bitcast.
897 if (isa<PointerType>(StoredValTy)) {
898 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
899 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
902 const Type *TypeToCastTo = LoadedTy;
903 if (isa<PointerType>(TypeToCastTo))
904 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
906 if (StoredValTy != TypeToCastTo)
907 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
909 // Cast to pointer if the load needs a pointer type.
910 if (isa<PointerType>(LoadedTy))
911 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
916 // If the loaded value is smaller than the available value, then we can
917 // extract out a piece from it. If the available value is too small, then we
918 // can't do anything.
919 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
921 // Convert source pointers to integers, which can be manipulated.
922 if (isa<PointerType>(StoredValTy)) {
923 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
924 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
927 // Convert vectors and fp to integer, which can be manipulated.
928 if (!isa<IntegerType>(StoredValTy)) {
929 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
930 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
933 // If this is a big-endian system, we need to shift the value down to the low
934 // bits so that a truncate will work.
935 if (TD.isBigEndian()) {
936 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
937 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
940 // Truncate the integer to the right size now.
941 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
942 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
944 if (LoadedTy == NewIntTy)
947 // If the result is a pointer, inttoptr.
948 if (isa<PointerType>(LoadedTy))
949 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
951 // Otherwise, bitcast.
952 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
955 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
956 /// be expressed as a base pointer plus a constant offset. Return the base and
957 /// offset to the caller.
958 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
959 const TargetData &TD) {
960 Operator *PtrOp = dyn_cast<Operator>(Ptr);
961 if (PtrOp == 0) return Ptr;
963 // Just look through bitcasts.
964 if (PtrOp->getOpcode() == Instruction::BitCast)
965 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
967 // If this is a GEP with constant indices, we can look through it.
968 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
969 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
971 gep_type_iterator GTI = gep_type_begin(GEP);
972 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
974 ConstantInt *OpC = cast<ConstantInt>(*I);
975 if (OpC->isZero()) continue;
977 // Handle a struct and array indices which add their offset to the pointer.
978 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
979 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
981 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
982 Offset += OpC->getSExtValue()*Size;
986 // Re-sign extend from the pointer size if needed to get overflow edge cases
988 unsigned PtrSize = TD.getPointerSizeInBits();
990 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
992 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
996 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
997 /// memdep query of a load that ends up being a clobbering memory write (store,
998 /// memset, memcpy, memmove). This means that the write *may* provide bits used
999 /// by the load but we can't be sure because the pointers don't mustalias.
1001 /// Check this case to see if there is anything more we can do before we give
1002 /// up. This returns -1 if we have to give up, or a byte number in the stored
1003 /// value of the piece that feeds the load.
1004 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
1006 uint64_t WriteSizeInBits,
1007 const TargetData &TD) {
1008 // If the loaded or stored value is an first class array or struct, don't try
1009 // to transform them. We need to be able to bitcast to integer.
1010 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy))
1013 int64_t StoreOffset = 0, LoadOffset = 0;
1014 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1016 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
1017 if (StoreBase != LoadBase)
1020 // If the load and store are to the exact same address, they should have been
1021 // a must alias. AA must have gotten confused.
1022 // FIXME: Study to see if/when this happens.
1023 if (LoadOffset == StoreOffset) {
1025 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1026 << "Base = " << *StoreBase << "\n"
1027 << "Store Ptr = " << *WritePtr << "\n"
1028 << "Store Offs = " << StoreOffset << "\n"
1029 << "Load Ptr = " << *LoadPtr << "\n";
1035 // If the load and store don't overlap at all, the store doesn't provide
1036 // anything to the load. In this case, they really don't alias at all, AA
1037 // must have gotten confused.
1038 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1039 // remove this check, as it is duplicated with what we have below.
1040 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
1042 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1044 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1048 bool isAAFailure = false;
1049 if (StoreOffset < LoadOffset) {
1050 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1052 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1056 errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1057 << "Base = " << *StoreBase << "\n"
1058 << "Store Ptr = " << *WritePtr << "\n"
1059 << "Store Offs = " << StoreOffset << "\n"
1060 << "Load Ptr = " << *LoadPtr << "\n";
1066 // If the Load isn't completely contained within the stored bits, we don't
1067 // have all the bits to feed it. We could do something crazy in the future
1068 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1070 if (StoreOffset > LoadOffset ||
1071 StoreOffset+StoreSize < LoadOffset+LoadSize)
1074 // Okay, we can do this transformation. Return the number of bytes into the
1075 // store that the load is.
1076 return LoadOffset-StoreOffset;
1079 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1080 /// memdep query of a load that ends up being a clobbering store.
1081 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
1083 const TargetData &TD) {
1084 // Cannot handle reading from store of first-class aggregate yet.
1085 if (isa<StructType>(DepSI->getOperand(0)->getType()) ||
1086 isa<ArrayType>(DepSI->getOperand(0)->getType()))
1089 Value *StorePtr = DepSI->getPointerOperand();
1090 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1091 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1092 StorePtr, StoreSize, TD);
1095 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
1097 const TargetData &TD) {
1098 // If the mem operation is a non-constant size, we can't handle it.
1099 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1100 if (SizeCst == 0) return -1;
1101 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1103 // If this is memset, we just need to see if the offset is valid in the size
1105 if (MI->getIntrinsicID() == Intrinsic::memset)
1106 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1109 // If we have a memcpy/memmove, the only case we can handle is if this is a
1110 // copy from constant memory. In that case, we can read directly from the
1112 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1114 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1115 if (Src == 0) return -1;
1117 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1118 if (GV == 0 || !GV->isConstant()) return -1;
1120 // See if the access is within the bounds of the transfer.
1121 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1122 MI->getDest(), MemSizeInBits, TD);
1126 // Otherwise, see if we can constant fold a load from the constant with the
1127 // offset applied as appropriate.
1128 Src = ConstantExpr::getBitCast(Src,
1129 llvm::Type::getInt8PtrTy(Src->getContext()));
1130 Constant *OffsetCst =
1131 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1132 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1133 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1134 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1140 /// GetStoreValueForLoad - This function is called when we have a
1141 /// memdep query of a load that ends up being a clobbering store. This means
1142 /// that the store *may* provide bits used by the load but we can't be sure
1143 /// because the pointers don't mustalias. Check this case to see if there is
1144 /// anything more we can do before we give up.
1145 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1147 Instruction *InsertPt, const TargetData &TD){
1148 LLVMContext &Ctx = SrcVal->getType()->getContext();
1150 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1151 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1153 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1155 // Compute which bits of the stored value are being used by the load. Convert
1156 // to an integer type to start with.
1157 if (isa<PointerType>(SrcVal->getType()))
1158 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
1159 if (!isa<IntegerType>(SrcVal->getType()))
1160 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1163 // Shift the bits to the least significant depending on endianness.
1165 if (TD.isLittleEndian())
1166 ShiftAmt = Offset*8;
1168 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1171 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
1173 if (LoadSize != StoreSize)
1174 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1177 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1180 /// GetMemInstValueForLoad - This function is called when we have a
1181 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1182 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1183 const Type *LoadTy, Instruction *InsertPt,
1184 const TargetData &TD){
1185 LLVMContext &Ctx = LoadTy->getContext();
1186 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1188 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1190 // We know that this method is only called when the mem transfer fully
1191 // provides the bits for the load.
1192 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1193 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1194 // independently of what the offset is.
1195 Value *Val = MSI->getValue();
1197 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1199 Value *OneElt = Val;
1201 // Splat the value out to the right number of bits.
1202 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1203 // If we can double the number of bytes set, do it.
1204 if (NumBytesSet*2 <= LoadSize) {
1205 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1206 Val = Builder.CreateOr(Val, ShVal);
1211 // Otherwise insert one byte at a time.
1212 Value *ShVal = Builder.CreateShl(Val, 1*8);
1213 Val = Builder.CreateOr(OneElt, ShVal);
1217 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1220 // Otherwise, this is a memcpy/memmove from a constant global.
1221 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1222 Constant *Src = cast<Constant>(MTI->getSource());
1224 // Otherwise, see if we can constant fold a load from the constant with the
1225 // offset applied as appropriate.
1226 Src = ConstantExpr::getBitCast(Src,
1227 llvm::Type::getInt8PtrTy(Src->getContext()));
1228 Constant *OffsetCst =
1229 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1230 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1231 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1232 return ConstantFoldLoadFromConstPtr(Src, &TD);
1237 struct AvailableValueInBlock {
1238 /// BB - The basic block in question.
1241 SimpleVal, // A simple offsetted value that is accessed.
1242 MemIntrin // A memory intrinsic which is loaded from.
1245 /// V - The value that is live out of the block.
1246 PointerIntPair<Value *, 1, ValType> Val;
1248 /// Offset - The byte offset in Val that is interesting for the load query.
1251 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1252 unsigned Offset = 0) {
1253 AvailableValueInBlock Res;
1255 Res.Val.setPointer(V);
1256 Res.Val.setInt(SimpleVal);
1257 Res.Offset = Offset;
1261 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1262 unsigned Offset = 0) {
1263 AvailableValueInBlock Res;
1265 Res.Val.setPointer(MI);
1266 Res.Val.setInt(MemIntrin);
1267 Res.Offset = Offset;
1271 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1272 Value *getSimpleValue() const {
1273 assert(isSimpleValue() && "Wrong accessor");
1274 return Val.getPointer();
1277 MemIntrinsic *getMemIntrinValue() const {
1278 assert(!isSimpleValue() && "Wrong accessor");
1279 return cast<MemIntrinsic>(Val.getPointer());
1282 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1283 /// defined here to the specified type. This handles various coercion cases.
1284 Value *MaterializeAdjustedValue(const Type *LoadTy,
1285 const TargetData *TD) const {
1287 if (isSimpleValue()) {
1288 Res = getSimpleValue();
1289 if (Res->getType() != LoadTy) {
1290 assert(TD && "Need target data to handle type mismatch case");
1291 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1294 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1295 << *getSimpleValue() << '\n'
1296 << *Res << '\n' << "\n\n\n");
1299 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1300 LoadTy, BB->getTerminator(), *TD);
1301 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1302 << " " << *getMemIntrinValue() << '\n'
1303 << *Res << '\n' << "\n\n\n");
1309 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1310 /// construct SSA form, allowing us to eliminate LI. This returns the value
1311 /// that should be used at LI's definition site.
1312 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1313 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1314 const TargetData *TD,
1315 const DominatorTree &DT,
1316 AliasAnalysis *AA) {
1317 // Check for the fully redundant, dominating load case. In this case, we can
1318 // just use the dominating value directly.
1319 if (ValuesPerBlock.size() == 1 &&
1320 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1321 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1323 // Otherwise, we have to construct SSA form.
1324 SmallVector<PHINode*, 8> NewPHIs;
1325 SSAUpdater SSAUpdate(&NewPHIs);
1326 SSAUpdate.Initialize(LI);
1328 const Type *LoadTy = LI->getType();
1330 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1331 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1332 BasicBlock *BB = AV.BB;
1334 if (SSAUpdate.HasValueForBlock(BB))
1337 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1340 // Perform PHI construction.
1341 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1343 // If new PHI nodes were created, notify alias analysis.
1344 if (isa<PointerType>(V->getType()))
1345 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1346 AA->copyValue(LI, NewPHIs[i]);
1351 static bool isLifetimeStart(Instruction *Inst) {
1352 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1353 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1357 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1358 /// non-local by performing PHI construction.
1359 bool GVN::processNonLocalLoad(LoadInst *LI,
1360 SmallVectorImpl<Instruction*> &toErase) {
1361 // Find the non-local dependencies of the load.
1362 SmallVector<NonLocalDepResult, 64> Deps;
1363 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1365 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
1366 // << Deps.size() << *LI << '\n');
1368 // If we had to process more than one hundred blocks to find the
1369 // dependencies, this load isn't worth worrying about. Optimizing
1370 // it will be too expensive.
1371 if (Deps.size() > 100)
1374 // If we had a phi translation failure, we'll have a single entry which is a
1375 // clobber in the current block. Reject this early.
1376 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
1378 errs() << "GVN: non-local load ";
1379 WriteAsOperand(errs(), LI);
1380 errs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1385 // Filter out useless results (non-locals, etc). Keep track of the blocks
1386 // where we have a value available in repl, also keep track of whether we see
1387 // dependencies that produce an unknown value for the load (such as a call
1388 // that could potentially clobber the load).
1389 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1390 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1392 const TargetData *TD = 0;
1394 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1395 BasicBlock *DepBB = Deps[i].getBB();
1396 MemDepResult DepInfo = Deps[i].getResult();
1398 if (DepInfo.isClobber()) {
1399 // The address being loaded in this non-local block may not be the same as
1400 // the pointer operand of the load if PHI translation occurs. Make sure
1401 // to consider the right address.
1402 Value *Address = Deps[i].getAddress();
1404 // If the dependence is to a store that writes to a superset of the bits
1405 // read by the load, we can extract the bits we need for the load from the
1407 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1409 TD = getAnalysisIfAvailable<TargetData>();
1410 if (TD && Address) {
1411 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1414 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1415 DepSI->getOperand(0),
1422 // If the clobbering value is a memset/memcpy/memmove, see if we can
1423 // forward a value on from it.
1424 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1426 TD = getAnalysisIfAvailable<TargetData>();
1427 if (TD && Address) {
1428 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1431 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1438 UnavailableBlocks.push_back(DepBB);
1442 Instruction *DepInst = DepInfo.getInst();
1444 // Loading the allocation -> undef.
1445 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1446 // Loading immediately after lifetime begin -> undef.
1447 isLifetimeStart(DepInst)) {
1448 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1449 UndefValue::get(LI->getType())));
1453 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1454 // Reject loads and stores that are to the same address but are of
1455 // different types if we have to.
1456 if (S->getOperand(0)->getType() != LI->getType()) {
1458 TD = getAnalysisIfAvailable<TargetData>();
1460 // If the stored value is larger or equal to the loaded value, we can
1462 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1463 LI->getType(), *TD)) {
1464 UnavailableBlocks.push_back(DepBB);
1469 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1474 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1475 // If the types mismatch and we can't handle it, reject reuse of the load.
1476 if (LD->getType() != LI->getType()) {
1478 TD = getAnalysisIfAvailable<TargetData>();
1480 // If the stored value is larger or equal to the loaded value, we can
1482 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1483 UnavailableBlocks.push_back(DepBB);
1487 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1491 UnavailableBlocks.push_back(DepBB);
1495 // If we have no predecessors that produce a known value for this load, exit
1497 if (ValuesPerBlock.empty()) return false;
1499 // If all of the instructions we depend on produce a known value for this
1500 // load, then it is fully redundant and we can use PHI insertion to compute
1501 // its value. Insert PHIs and remove the fully redundant value now.
1502 if (UnavailableBlocks.empty()) {
1503 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1505 // Perform PHI construction.
1506 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1507 VN.getAliasAnalysis());
1508 LI->replaceAllUsesWith(V);
1510 if (isa<PHINode>(V))
1512 if (isa<PointerType>(V->getType()))
1513 MD->invalidateCachedPointerInfo(V);
1514 toErase.push_back(LI);
1519 if (!EnablePRE || !EnableLoadPRE)
1522 // Okay, we have *some* definitions of the value. This means that the value
1523 // is available in some of our (transitive) predecessors. Lets think about
1524 // doing PRE of this load. This will involve inserting a new load into the
1525 // predecessor when it's not available. We could do this in general, but
1526 // prefer to not increase code size. As such, we only do this when we know
1527 // that we only have to insert *one* load (which means we're basically moving
1528 // the load, not inserting a new one).
1530 SmallPtrSet<BasicBlock *, 4> Blockers;
1531 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1532 Blockers.insert(UnavailableBlocks[i]);
1534 // Lets find first basic block with more than one predecessor. Walk backwards
1535 // through predecessors if needed.
1536 BasicBlock *LoadBB = LI->getParent();
1537 BasicBlock *TmpBB = LoadBB;
1539 bool isSinglePred = false;
1540 bool allSingleSucc = true;
1541 while (TmpBB->getSinglePredecessor()) {
1542 isSinglePred = true;
1543 TmpBB = TmpBB->getSinglePredecessor();
1544 if (!TmpBB) // If haven't found any, bail now.
1546 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1548 if (Blockers.count(TmpBB))
1550 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1551 allSingleSucc = false;
1557 // If we have a repl set with LI itself in it, this means we have a loop where
1558 // at least one of the values is LI. Since this means that we won't be able
1559 // to eliminate LI even if we insert uses in the other predecessors, we will
1560 // end up increasing code size. Reject this by scanning for LI.
1561 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1562 if (ValuesPerBlock[i].isSimpleValue() &&
1563 ValuesPerBlock[i].getSimpleValue() == LI)
1566 // FIXME: It is extremely unclear what this loop is doing, other than
1567 // artificially restricting loadpre.
1570 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1571 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1572 if (AV.isSimpleValue())
1573 // "Hot" Instruction is in some loop (because it dominates its dep.
1575 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1576 if (DT->dominates(LI, I)) {
1582 // We are interested only in "hot" instructions. We don't want to do any
1583 // mis-optimizations here.
1588 // Okay, we have some hope :). Check to see if the loaded value is fully
1589 // available in all but one predecessor.
1590 // FIXME: If we could restructure the CFG, we could make a common pred with
1591 // all the preds that don't have an available LI and insert a new load into
1593 BasicBlock *UnavailablePred = 0;
1595 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1596 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1597 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1598 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1599 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1601 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1603 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1606 // If this load is not available in multiple predecessors, reject it.
1607 if (UnavailablePred && UnavailablePred != *PI)
1609 UnavailablePred = *PI;
1612 assert(UnavailablePred != 0 &&
1613 "Fully available value should be eliminated above!");
1615 // We don't currently handle critical edges :(
1616 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1617 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1618 << UnavailablePred->getName() << "': " << *LI << '\n');
1622 // Do PHI translation to get its value in the predecessor if necessary. The
1623 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1625 SmallVector<Instruction*, 8> NewInsts;
1627 // If all preds have a single successor, then we know it is safe to insert the
1628 // load on the pred (?!?), so we can insert code to materialize the pointer if
1629 // it is not available.
1630 PHITransAddr Address(LI->getOperand(0), TD);
1632 if (allSingleSucc) {
1633 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1636 Address.PHITranslateValue(LoadBB, UnavailablePred);
1637 LoadPtr = Address.getAddr();
1639 // Make sure the value is live in the predecessor.
1640 if (Instruction *Inst = dyn_cast_or_null<Instruction>(LoadPtr))
1641 if (!DT->dominates(Inst->getParent(), UnavailablePred))
1645 // If we couldn't find or insert a computation of this phi translated value,
1648 assert(NewInsts.empty() && "Shouldn't insert insts on failure");
1649 DEBUG(errs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1650 << *LI->getOperand(0) << "\n");
1654 // Assign value numbers to these new instructions.
1655 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1656 // FIXME: We really _ought_ to insert these value numbers into their
1657 // parent's availability map. However, in doing so, we risk getting into
1658 // ordering issues. If a block hasn't been processed yet, we would be
1659 // marking a value as AVAIL-IN, which isn't what we intend.
1660 VN.lookup_or_add(NewInsts[i]);
1663 // Make sure it is valid to move this load here. We have to watch out for:
1664 // @1 = getelementptr (i8* p, ...
1665 // test p and branch if == 0
1667 // It is valid to have the getelementptr before the test, even if p can be 0,
1668 // as getelementptr only does address arithmetic.
1669 // If we are not pushing the value through any multiple-successor blocks
1670 // we do not have this case. Otherwise, check that the load is safe to
1671 // put anywhere; this can be improved, but should be conservatively safe.
1672 if (!allSingleSucc &&
1673 // FIXME: REEVALUTE THIS.
1674 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) {
1675 assert(NewInsts.empty() && "Should not have inserted instructions");
1679 // Okay, we can eliminate this load by inserting a reload in the predecessor
1680 // and using PHI construction to get the value in the other predecessors, do
1682 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1683 DEBUG(if (!NewInsts.empty())
1684 errs() << "INSERTED " << NewInsts.size() << " INSTS: "
1685 << *NewInsts.back() << '\n');
1687 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1689 UnavailablePred->getTerminator());
1691 // Add the newly created load.
1692 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));
1694 // Perform PHI construction.
1695 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1696 VN.getAliasAnalysis());
1697 LI->replaceAllUsesWith(V);
1698 if (isa<PHINode>(V))
1700 if (isa<PointerType>(V->getType()))
1701 MD->invalidateCachedPointerInfo(V);
1702 toErase.push_back(LI);
1707 /// processLoad - Attempt to eliminate a load, first by eliminating it
1708 /// locally, and then attempting non-local elimination if that fails.
1709 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1713 if (L->isVolatile())
1716 // ... to a pointer that has been loaded from before...
1717 MemDepResult Dep = MD->getDependency(L);
1719 // If the value isn't available, don't do anything!
1720 if (Dep.isClobber()) {
1721 // Check to see if we have something like this:
1722 // store i32 123, i32* %P
1723 // %A = bitcast i32* %P to i8*
1724 // %B = gep i8* %A, i32 1
1727 // We could do that by recognizing if the clobber instructions are obviously
1728 // a common base + constant offset, and if the previous store (or memset)
1729 // completely covers this load. This sort of thing can happen in bitfield
1731 Value *AvailVal = 0;
1732 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1733 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1734 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1735 L->getPointerOperand(),
1738 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1739 L->getType(), L, *TD);
1742 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1743 // a value on from it.
1744 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1745 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1746 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1747 L->getPointerOperand(),
1750 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1755 DEBUG(errs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1756 << *AvailVal << '\n' << *L << "\n\n\n");
1758 // Replace the load!
1759 L->replaceAllUsesWith(AvailVal);
1760 if (isa<PointerType>(AvailVal->getType()))
1761 MD->invalidateCachedPointerInfo(AvailVal);
1762 toErase.push_back(L);
1768 // fast print dep, using operator<< on instruction would be too slow
1769 errs() << "GVN: load ";
1770 WriteAsOperand(errs(), L);
1771 Instruction *I = Dep.getInst();
1772 errs() << " is clobbered by " << *I << '\n';
1777 // If it is defined in another block, try harder.
1778 if (Dep.isNonLocal())
1779 return processNonLocalLoad(L, toErase);
1781 Instruction *DepInst = Dep.getInst();
1782 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1783 Value *StoredVal = DepSI->getOperand(0);
1785 // The store and load are to a must-aliased pointer, but they may not
1786 // actually have the same type. See if we know how to reuse the stored
1787 // value (depending on its type).
1788 const TargetData *TD = 0;
1789 if (StoredVal->getType() != L->getType()) {
1790 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1791 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1796 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1797 << '\n' << *L << "\n\n\n");
1804 L->replaceAllUsesWith(StoredVal);
1805 if (isa<PointerType>(StoredVal->getType()))
1806 MD->invalidateCachedPointerInfo(StoredVal);
1807 toErase.push_back(L);
1812 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1813 Value *AvailableVal = DepLI;
1815 // The loads are of a must-aliased pointer, but they may not actually have
1816 // the same type. See if we know how to reuse the previously loaded value
1817 // (depending on its type).
1818 const TargetData *TD = 0;
1819 if (DepLI->getType() != L->getType()) {
1820 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1821 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1822 if (AvailableVal == 0)
1825 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1826 << "\n" << *L << "\n\n\n");
1833 L->replaceAllUsesWith(AvailableVal);
1834 if (isa<PointerType>(DepLI->getType()))
1835 MD->invalidateCachedPointerInfo(DepLI);
1836 toErase.push_back(L);
1841 // If this load really doesn't depend on anything, then we must be loading an
1842 // undef value. This can happen when loading for a fresh allocation with no
1843 // intervening stores, for example.
1844 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1845 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1846 toErase.push_back(L);
1851 // If this load occurs either right after a lifetime begin,
1852 // then the loaded value is undefined.
1853 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1854 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1855 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1856 toErase.push_back(L);
1865 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1866 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1867 if (I == localAvail.end())
1870 ValueNumberScope *Locals = I->second;
1872 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1873 if (I != Locals->table.end())
1875 Locals = Locals->parent;
1882 /// processInstruction - When calculating availability, handle an instruction
1883 /// by inserting it into the appropriate sets
1884 bool GVN::processInstruction(Instruction *I,
1885 SmallVectorImpl<Instruction*> &toErase) {
1886 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1887 bool Changed = processLoad(LI, toErase);
1890 unsigned Num = VN.lookup_or_add(LI);
1891 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1897 uint32_t NextNum = VN.getNextUnusedValueNumber();
1898 unsigned Num = VN.lookup_or_add(I);
1900 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1901 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1903 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1906 Value *BranchCond = BI->getCondition();
1907 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1909 BasicBlock *TrueSucc = BI->getSuccessor(0);
1910 BasicBlock *FalseSucc = BI->getSuccessor(1);
1912 if (TrueSucc->getSinglePredecessor())
1913 localAvail[TrueSucc]->table[CondVN] =
1914 ConstantInt::getTrue(TrueSucc->getContext());
1915 if (FalseSucc->getSinglePredecessor())
1916 localAvail[FalseSucc]->table[CondVN] =
1917 ConstantInt::getFalse(TrueSucc->getContext());
1921 // Allocations are always uniquely numbered, so we can save time and memory
1922 // by fast failing them.
1923 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1924 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1928 // Collapse PHI nodes
1929 if (PHINode* p = dyn_cast<PHINode>(I)) {
1930 Value *constVal = CollapsePhi(p);
1933 p->replaceAllUsesWith(constVal);
1934 if (MD && isa<PointerType>(constVal->getType()))
1935 MD->invalidateCachedPointerInfo(constVal);
1938 toErase.push_back(p);
1940 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1943 // If the number we were assigned was a brand new VN, then we don't
1944 // need to do a lookup to see if the number already exists
1945 // somewhere in the domtree: it can't!
1946 } else if (Num == NextNum) {
1947 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1949 // Perform fast-path value-number based elimination of values inherited from
1951 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1954 I->replaceAllUsesWith(repl);
1955 if (MD && isa<PointerType>(repl->getType()))
1956 MD->invalidateCachedPointerInfo(repl);
1957 toErase.push_back(I);
1961 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1967 /// runOnFunction - This is the main transformation entry point for a function.
1968 bool GVN::runOnFunction(Function& F) {
1970 MD = &getAnalysis<MemoryDependenceAnalysis>();
1971 DT = &getAnalysis<DominatorTree>();
1972 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1976 bool Changed = false;
1977 bool ShouldContinue = true;
1979 // Merge unconditional branches, allowing PRE to catch more
1980 // optimization opportunities.
1981 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1982 BasicBlock *BB = FI;
1984 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1985 if (removedBlock) NumGVNBlocks++;
1987 Changed |= removedBlock;
1990 unsigned Iteration = 0;
1992 while (ShouldContinue) {
1993 DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
1994 ShouldContinue = iterateOnFunction(F);
1995 Changed |= ShouldContinue;
2000 bool PREChanged = true;
2001 while (PREChanged) {
2002 PREChanged = performPRE(F);
2003 Changed |= PREChanged;
2006 // FIXME: Should perform GVN again after PRE does something. PRE can move
2007 // computations into blocks where they become fully redundant. Note that
2008 // we can't do this until PRE's critical edge splitting updates memdep.
2009 // Actually, when this happens, we should just fully integrate PRE into GVN.
2011 cleanupGlobalSets();
2017 bool GVN::processBlock(BasicBlock *BB) {
2018 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
2019 // incrementing BI before processing an instruction).
2020 SmallVector<Instruction*, 8> toErase;
2021 bool ChangedFunction = false;
2023 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2025 ChangedFunction |= processInstruction(BI, toErase);
2026 if (toErase.empty()) {
2031 // If we need some instructions deleted, do it now.
2032 NumGVNInstr += toErase.size();
2034 // Avoid iterator invalidation.
2035 bool AtStart = BI == BB->begin();
2039 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2040 E = toErase.end(); I != E; ++I) {
2041 DEBUG(errs() << "GVN removed: " << **I << '\n');
2042 if (MD) MD->removeInstruction(*I);
2043 (*I)->eraseFromParent();
2044 DEBUG(verifyRemoved(*I));
2054 return ChangedFunction;
2057 /// performPRE - Perform a purely local form of PRE that looks for diamond
2058 /// control flow patterns and attempts to perform simple PRE at the join point.
2059 bool GVN::performPRE(Function &F) {
2060 bool Changed = false;
2061 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
2062 DenseMap<BasicBlock*, Value*> predMap;
2063 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2064 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2065 BasicBlock *CurrentBlock = *DI;
2067 // Nothing to PRE in the entry block.
2068 if (CurrentBlock == &F.getEntryBlock()) continue;
2070 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2071 BE = CurrentBlock->end(); BI != BE; ) {
2072 Instruction *CurInst = BI++;
2074 if (isa<AllocaInst>(CurInst) ||
2075 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2076 CurInst->getType()->isVoidTy() ||
2077 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2078 isa<DbgInfoIntrinsic>(CurInst))
2081 uint32_t ValNo = VN.lookup(CurInst);
2083 // Look for the predecessors for PRE opportunities. We're
2084 // only trying to solve the basic diamond case, where
2085 // a value is computed in the successor and one predecessor,
2086 // but not the other. We also explicitly disallow cases
2087 // where the successor is its own predecessor, because they're
2088 // more complicated to get right.
2089 unsigned NumWith = 0;
2090 unsigned NumWithout = 0;
2091 BasicBlock *PREPred = 0;
2094 for (pred_iterator PI = pred_begin(CurrentBlock),
2095 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2096 // We're not interested in PRE where the block is its
2097 // own predecessor, on in blocks with predecessors
2098 // that are not reachable.
2099 if (*PI == CurrentBlock) {
2102 } else if (!localAvail.count(*PI)) {
2107 DenseMap<uint32_t, Value*>::iterator predV =
2108 localAvail[*PI]->table.find(ValNo);
2109 if (predV == localAvail[*PI]->table.end()) {
2112 } else if (predV->second == CurInst) {
2115 predMap[*PI] = predV->second;
2120 // Don't do PRE when it might increase code size, i.e. when
2121 // we would need to insert instructions in more than one pred.
2122 if (NumWithout != 1 || NumWith == 0)
2125 // Don't do PRE across indirect branch.
2126 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2129 // We can't do PRE safely on a critical edge, so instead we schedule
2130 // the edge to be split and perform the PRE the next time we iterate
2132 unsigned SuccNum = 0;
2133 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
2135 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
2140 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2141 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2145 // Instantiate the expression the in predecessor that lacked it.
2146 // Because we are going top-down through the block, all value numbers
2147 // will be available in the predecessor by the time we need them. Any
2148 // that weren't original present will have been instantiated earlier
2150 Instruction *PREInstr = CurInst->clone();
2151 bool success = true;
2152 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2153 Value *Op = PREInstr->getOperand(i);
2154 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2157 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2158 PREInstr->setOperand(i, V);
2165 // Fail out if we encounter an operand that is not available in
2166 // the PRE predecessor. This is typically because of loads which
2167 // are not value numbered precisely.
2170 DEBUG(verifyRemoved(PREInstr));
2174 PREInstr->insertBefore(PREPred->getTerminator());
2175 PREInstr->setName(CurInst->getName() + ".pre");
2176 predMap[PREPred] = PREInstr;
2177 VN.add(PREInstr, ValNo);
2180 // Update the availability map to include the new instruction.
2181 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
2183 // Create a PHI to make the value available in this block.
2184 PHINode* Phi = PHINode::Create(CurInst->getType(),
2185 CurInst->getName() + ".pre-phi",
2186 CurrentBlock->begin());
2187 for (pred_iterator PI = pred_begin(CurrentBlock),
2188 PE = pred_end(CurrentBlock); PI != PE; ++PI)
2189 Phi->addIncoming(predMap[*PI], *PI);
2192 localAvail[CurrentBlock]->table[ValNo] = Phi;
2194 CurInst->replaceAllUsesWith(Phi);
2195 if (MD && isa<PointerType>(Phi->getType()))
2196 MD->invalidateCachedPointerInfo(Phi);
2199 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
2200 if (MD) MD->removeInstruction(CurInst);
2201 CurInst->eraseFromParent();
2202 DEBUG(verifyRemoved(CurInst));
2207 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
2208 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
2209 SplitCriticalEdge(I->first, I->second, this);
2211 return Changed || toSplit.size();
2214 /// iterateOnFunction - Executes one iteration of GVN
2215 bool GVN::iterateOnFunction(Function &F) {
2216 cleanupGlobalSets();
2218 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2219 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
2221 localAvail[DI->getBlock()] =
2222 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
2224 localAvail[DI->getBlock()] = new ValueNumberScope(0);
2227 // Top-down walk of the dominator tree
2228 bool Changed = false;
2230 // Needed for value numbering with phi construction to work.
2231 ReversePostOrderTraversal<Function*> RPOT(&F);
2232 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2233 RE = RPOT.end(); RI != RE; ++RI)
2234 Changed |= processBlock(*RI);
2236 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2237 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2238 Changed |= processBlock(DI->getBlock());
2244 void GVN::cleanupGlobalSets() {
2247 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2248 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2253 /// verifyRemoved - Verify that the specified instruction does not occur in our
2254 /// internal data structures.
2255 void GVN::verifyRemoved(const Instruction *Inst) const {
2256 VN.verifyRemoved(Inst);
2258 // Walk through the value number scope to make sure the instruction isn't
2259 // ferreted away in it.
2260 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
2261 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2262 const ValueNumberScope *VNS = I->second;
2265 for (DenseMap<uint32_t, Value*>::const_iterator
2266 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2267 assert(II->second != Inst && "Inst still in value numbering scope!");