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));
63 static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false));
65 //===----------------------------------------------------------------------===//
67 //===----------------------------------------------------------------------===//
69 /// This class holds the mapping between values and value numbers. It is used
70 /// as an efficient mechanism to determine the expression-wise equivalence of
74 enum ExpressionOpcode {
75 ADD = Instruction::Add,
76 FADD = Instruction::FAdd,
77 SUB = Instruction::Sub,
78 FSUB = Instruction::FSub,
79 MUL = Instruction::Mul,
80 FMUL = Instruction::FMul,
81 UDIV = Instruction::UDiv,
82 SDIV = Instruction::SDiv,
83 FDIV = Instruction::FDiv,
84 UREM = Instruction::URem,
85 SREM = Instruction::SRem,
86 FREM = Instruction::FRem,
87 SHL = Instruction::Shl,
88 LSHR = Instruction::LShr,
89 ASHR = Instruction::AShr,
90 AND = Instruction::And,
92 XOR = Instruction::Xor,
93 TRUNC = Instruction::Trunc,
94 ZEXT = Instruction::ZExt,
95 SEXT = Instruction::SExt,
96 FPTOUI = Instruction::FPToUI,
97 FPTOSI = Instruction::FPToSI,
98 UITOFP = Instruction::UIToFP,
99 SITOFP = Instruction::SIToFP,
100 FPTRUNC = Instruction::FPTrunc,
101 FPEXT = Instruction::FPExt,
102 PTRTOINT = Instruction::PtrToInt,
103 INTTOPTR = Instruction::IntToPtr,
104 BITCAST = Instruction::BitCast,
105 ICMPEQ, ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
106 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
107 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
108 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
109 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
110 SHUFFLE, SELECT, GEP, CALL, CONSTANT,
111 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
113 ExpressionOpcode opcode;
115 SmallVector<uint32_t, 4> varargs;
119 Expression(ExpressionOpcode o) : opcode(o) { }
121 bool operator==(const Expression &other) const {
122 if (opcode != other.opcode)
124 else if (opcode == EMPTY || opcode == TOMBSTONE)
126 else if (type != other.type)
128 else if (function != other.function)
131 if (varargs.size() != other.varargs.size())
134 for (size_t i = 0; i < varargs.size(); ++i)
135 if (varargs[i] != other.varargs[i])
142 bool operator!=(const Expression &other) const {
143 return !(*this == other);
149 DenseMap<Value*, uint32_t> valueNumbering;
150 DenseMap<Expression, uint32_t> expressionNumbering;
152 MemoryDependenceAnalysis* MD;
155 uint32_t nextValueNumber;
157 Expression::ExpressionOpcode getOpcode(CmpInst* C);
158 Expression create_expression(BinaryOperator* BO);
159 Expression create_expression(CmpInst* C);
160 Expression create_expression(ShuffleVectorInst* V);
161 Expression create_expression(ExtractElementInst* C);
162 Expression create_expression(InsertElementInst* V);
163 Expression create_expression(SelectInst* V);
164 Expression create_expression(CastInst* C);
165 Expression create_expression(GetElementPtrInst* G);
166 Expression create_expression(CallInst* C);
167 Expression create_expression(Constant* C);
168 Expression create_expression(ExtractValueInst* C);
169 Expression create_expression(InsertValueInst* C);
171 uint32_t lookup_or_add_call(CallInst* C);
173 ValueTable() : nextValueNumber(1) { }
174 uint32_t lookup_or_add(Value *V);
175 uint32_t lookup(Value *V) const;
176 void add(Value *V, uint32_t num);
178 void erase(Value *v);
180 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
181 AliasAnalysis *getAliasAnalysis() const { return AA; }
182 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
183 void setDomTree(DominatorTree* D) { DT = D; }
184 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
185 void verifyRemoved(const Value *) const;
190 template <> struct DenseMapInfo<Expression> {
191 static inline Expression getEmptyKey() {
192 return Expression(Expression::EMPTY);
195 static inline Expression getTombstoneKey() {
196 return Expression(Expression::TOMBSTONE);
199 static unsigned getHashValue(const Expression e) {
200 unsigned hash = e.opcode;
202 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
203 (unsigned)((uintptr_t)e.type >> 9));
205 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
206 E = e.varargs.end(); I != E; ++I)
207 hash = *I + hash * 37;
209 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
210 (unsigned)((uintptr_t)e.function >> 9)) +
215 static bool isEqual(const Expression &LHS, const Expression &RHS) {
221 struct isPodLike<Expression> { static const bool value = true; };
225 //===----------------------------------------------------------------------===//
226 // ValueTable Internal Functions
227 //===----------------------------------------------------------------------===//
229 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
230 if (isa<ICmpInst>(C)) {
231 switch (C->getPredicate()) {
232 default: // THIS SHOULD NEVER HAPPEN
233 llvm_unreachable("Comparison with unknown predicate?");
234 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
235 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
236 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
237 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
238 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
239 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
240 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
241 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
242 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
243 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
246 switch (C->getPredicate()) {
247 default: // THIS SHOULD NEVER HAPPEN
248 llvm_unreachable("Comparison with unknown predicate?");
249 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
250 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
251 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
252 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
253 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
254 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
255 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
256 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
257 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
258 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
259 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
260 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
261 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
262 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
267 Expression ValueTable::create_expression(CallInst* C) {
270 e.type = C->getType();
271 e.function = C->getCalledFunction();
272 e.opcode = Expression::CALL;
274 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
276 e.varargs.push_back(lookup_or_add(*I));
281 Expression ValueTable::create_expression(BinaryOperator* BO) {
283 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
284 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
286 e.type = BO->getType();
287 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode());
292 Expression ValueTable::create_expression(CmpInst* C) {
295 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
296 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
298 e.type = C->getType();
299 e.opcode = getOpcode(C);
304 Expression ValueTable::create_expression(CastInst* C) {
307 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
309 e.type = C->getType();
310 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode());
315 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
318 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
319 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
320 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
322 e.type = S->getType();
323 e.opcode = Expression::SHUFFLE;
328 Expression ValueTable::create_expression(ExtractElementInst* E) {
331 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
332 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
334 e.type = E->getType();
335 e.opcode = Expression::EXTRACT;
340 Expression ValueTable::create_expression(InsertElementInst* I) {
343 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
344 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
345 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
347 e.type = I->getType();
348 e.opcode = Expression::INSERT;
353 Expression ValueTable::create_expression(SelectInst* I) {
356 e.varargs.push_back(lookup_or_add(I->getCondition()));
357 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
358 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
360 e.type = I->getType();
361 e.opcode = Expression::SELECT;
366 Expression ValueTable::create_expression(GetElementPtrInst* G) {
369 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
371 e.type = G->getType();
372 e.opcode = Expression::GEP;
374 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
376 e.varargs.push_back(lookup_or_add(*I));
381 Expression ValueTable::create_expression(ExtractValueInst* E) {
384 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
385 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
387 e.varargs.push_back(*II);
389 e.type = E->getType();
390 e.opcode = Expression::EXTRACTVALUE;
395 Expression ValueTable::create_expression(InsertValueInst* E) {
398 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
399 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
400 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
402 e.varargs.push_back(*II);
404 e.type = E->getType();
405 e.opcode = Expression::INSERTVALUE;
410 //===----------------------------------------------------------------------===//
411 // ValueTable External Functions
412 //===----------------------------------------------------------------------===//
414 /// add - Insert a value into the table with a specified value number.
415 void ValueTable::add(Value *V, uint32_t num) {
416 valueNumbering.insert(std::make_pair(V, num));
419 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
420 if (AA->doesNotAccessMemory(C)) {
421 Expression exp = create_expression(C);
422 uint32_t& e = expressionNumbering[exp];
423 if (!e) e = nextValueNumber++;
424 valueNumbering[C] = e;
426 } else if (AA->onlyReadsMemory(C)) {
427 Expression exp = create_expression(C);
428 uint32_t& e = expressionNumbering[exp];
430 e = nextValueNumber++;
431 valueNumbering[C] = e;
435 e = nextValueNumber++;
436 valueNumbering[C] = e;
440 MemDepResult local_dep = MD->getDependency(C);
442 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
443 valueNumbering[C] = nextValueNumber;
444 return nextValueNumber++;
447 if (local_dep.isDef()) {
448 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
450 if (local_cdep->getNumOperands() != C->getNumOperands()) {
451 valueNumbering[C] = nextValueNumber;
452 return nextValueNumber++;
455 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
456 uint32_t c_vn = lookup_or_add(C->getOperand(i));
457 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
459 valueNumbering[C] = nextValueNumber;
460 return nextValueNumber++;
464 uint32_t v = lookup_or_add(local_cdep);
465 valueNumbering[C] = v;
470 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
471 MD->getNonLocalCallDependency(CallSite(C));
472 // FIXME: call/call dependencies for readonly calls should return def, not
473 // clobber! Move the checking logic to MemDep!
476 // Check to see if we have a single dominating call instruction that is
478 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
479 const NonLocalDepEntry *I = &deps[i];
480 // Ignore non-local dependencies.
481 if (I->getResult().isNonLocal())
484 // We don't handle non-depedencies. If we already have a call, reject
485 // instruction dependencies.
486 if (I->getResult().isClobber() || cdep != 0) {
491 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
492 // FIXME: All duplicated with non-local case.
493 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
494 cdep = NonLocalDepCall;
503 valueNumbering[C] = nextValueNumber;
504 return nextValueNumber++;
507 if (cdep->getNumOperands() != C->getNumOperands()) {
508 valueNumbering[C] = nextValueNumber;
509 return nextValueNumber++;
511 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
512 uint32_t c_vn = lookup_or_add(C->getOperand(i));
513 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
515 valueNumbering[C] = nextValueNumber;
516 return nextValueNumber++;
520 uint32_t v = lookup_or_add(cdep);
521 valueNumbering[C] = v;
525 valueNumbering[C] = nextValueNumber;
526 return nextValueNumber++;
530 /// lookup_or_add - Returns the value number for the specified value, assigning
531 /// it a new number if it did not have one before.
532 uint32_t ValueTable::lookup_or_add(Value *V) {
533 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
534 if (VI != valueNumbering.end())
537 if (!isa<Instruction>(V)) {
538 valueNumbering[V] = nextValueNumber;
539 return nextValueNumber++;
542 Instruction* I = cast<Instruction>(V);
544 switch (I->getOpcode()) {
545 case Instruction::Call:
546 return lookup_or_add_call(cast<CallInst>(I));
547 case Instruction::Add:
548 case Instruction::FAdd:
549 case Instruction::Sub:
550 case Instruction::FSub:
551 case Instruction::Mul:
552 case Instruction::FMul:
553 case Instruction::UDiv:
554 case Instruction::SDiv:
555 case Instruction::FDiv:
556 case Instruction::URem:
557 case Instruction::SRem:
558 case Instruction::FRem:
559 case Instruction::Shl:
560 case Instruction::LShr:
561 case Instruction::AShr:
562 case Instruction::And:
563 case Instruction::Or :
564 case Instruction::Xor:
565 exp = create_expression(cast<BinaryOperator>(I));
567 case Instruction::ICmp:
568 case Instruction::FCmp:
569 exp = create_expression(cast<CmpInst>(I));
571 case Instruction::Trunc:
572 case Instruction::ZExt:
573 case Instruction::SExt:
574 case Instruction::FPToUI:
575 case Instruction::FPToSI:
576 case Instruction::UIToFP:
577 case Instruction::SIToFP:
578 case Instruction::FPTrunc:
579 case Instruction::FPExt:
580 case Instruction::PtrToInt:
581 case Instruction::IntToPtr:
582 case Instruction::BitCast:
583 exp = create_expression(cast<CastInst>(I));
585 case Instruction::Select:
586 exp = create_expression(cast<SelectInst>(I));
588 case Instruction::ExtractElement:
589 exp = create_expression(cast<ExtractElementInst>(I));
591 case Instruction::InsertElement:
592 exp = create_expression(cast<InsertElementInst>(I));
594 case Instruction::ShuffleVector:
595 exp = create_expression(cast<ShuffleVectorInst>(I));
597 case Instruction::ExtractValue:
598 exp = create_expression(cast<ExtractValueInst>(I));
600 case Instruction::InsertValue:
601 exp = create_expression(cast<InsertValueInst>(I));
603 case Instruction::GetElementPtr:
604 exp = create_expression(cast<GetElementPtrInst>(I));
607 valueNumbering[V] = nextValueNumber;
608 return nextValueNumber++;
611 uint32_t& e = expressionNumbering[exp];
612 if (!e) e = nextValueNumber++;
613 valueNumbering[V] = e;
617 /// lookup - Returns the value number of the specified value. Fails if
618 /// the value has not yet been numbered.
619 uint32_t ValueTable::lookup(Value *V) const {
620 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
621 assert(VI != valueNumbering.end() && "Value not numbered?");
625 /// clear - Remove all entries from the ValueTable
626 void ValueTable::clear() {
627 valueNumbering.clear();
628 expressionNumbering.clear();
632 /// erase - Remove a value from the value numbering
633 void ValueTable::erase(Value *V) {
634 valueNumbering.erase(V);
637 /// verifyRemoved - Verify that the value is removed from all internal data
639 void ValueTable::verifyRemoved(const Value *V) const {
640 for (DenseMap<Value*, uint32_t>::const_iterator
641 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
642 assert(I->first != V && "Inst still occurs in value numbering map!");
646 //===----------------------------------------------------------------------===//
648 //===----------------------------------------------------------------------===//
651 struct ValueNumberScope {
652 ValueNumberScope* parent;
653 DenseMap<uint32_t, Value*> table;
655 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
661 class GVN : public FunctionPass {
662 bool runOnFunction(Function &F);
664 static char ID; // Pass identification, replacement for typeid
665 explicit GVN(bool nopre = false, bool noloads = false)
666 : FunctionPass(&ID), NoPRE(nopre), NoLoads(noloads), MD(0) { }
671 MemoryDependenceAnalysis *MD;
675 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
677 // List of critical edges to be split between iterations.
678 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
680 // This transformation requires dominator postdominator info
681 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
682 AU.addRequired<DominatorTree>();
684 AU.addRequired<MemoryDependenceAnalysis>();
685 AU.addRequired<AliasAnalysis>();
687 AU.addPreserved<DominatorTree>();
688 AU.addPreserved<AliasAnalysis>();
692 // FIXME: eliminate or document these better
693 bool processLoad(LoadInst* L,
694 SmallVectorImpl<Instruction*> &toErase);
695 bool processInstruction(Instruction *I,
696 SmallVectorImpl<Instruction*> &toErase);
697 bool processNonLocalLoad(LoadInst* L,
698 SmallVectorImpl<Instruction*> &toErase);
699 bool processBlock(BasicBlock *BB);
700 void dump(DenseMap<uint32_t, Value*>& d);
701 bool iterateOnFunction(Function &F);
702 Value *CollapsePhi(PHINode* p);
703 bool performPRE(Function& F);
704 Value *lookupNumber(BasicBlock *BB, uint32_t num);
705 void cleanupGlobalSets();
706 void verifyRemoved(const Instruction *I) const;
707 bool splitCriticalEdges();
713 // createGVNPass - The public interface to this file...
714 FunctionPass *llvm::createGVNPass(bool NoPRE, bool NoLoads) {
715 return new GVN(NoPRE, NoLoads);
718 static RegisterPass<GVN> X("gvn",
719 "Global Value Numbering");
721 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
723 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
724 E = d.end(); I != E; ++I) {
725 errs() << I->first << "\n";
731 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
732 if (!isa<PHINode>(inst))
735 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
737 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
738 if (use_phi->getParent() == inst->getParent())
744 Value *GVN::CollapsePhi(PHINode *PN) {
745 Value *ConstVal = PN->hasConstantValue(DT);
746 if (!ConstVal) return 0;
748 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
752 if (DT->dominates(Inst, PN))
753 if (isSafeReplacement(PN, Inst))
758 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
759 /// we're analyzing is fully available in the specified block. As we go, keep
760 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
761 /// map is actually a tri-state map with the following values:
762 /// 0) we know the block *is not* fully available.
763 /// 1) we know the block *is* fully available.
764 /// 2) we do not know whether the block is fully available or not, but we are
765 /// currently speculating that it will be.
766 /// 3) we are speculating for this block and have used that to speculate for
768 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
769 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
770 // Optimistically assume that the block is fully available and check to see
771 // if we already know about this block in one lookup.
772 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
773 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
775 // If the entry already existed for this block, return the precomputed value.
777 // If this is a speculative "available" value, mark it as being used for
778 // speculation of other blocks.
779 if (IV.first->second == 2)
780 IV.first->second = 3;
781 return IV.first->second != 0;
784 // Otherwise, see if it is fully available in all predecessors.
785 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
787 // If this block has no predecessors, it isn't live-in here.
789 goto SpeculationFailure;
791 for (; PI != PE; ++PI)
792 // If the value isn't fully available in one of our predecessors, then it
793 // isn't fully available in this block either. Undo our previous
794 // optimistic assumption and bail out.
795 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
796 goto SpeculationFailure;
800 // SpeculationFailure - If we get here, we found out that this is not, after
801 // all, a fully-available block. We have a problem if we speculated on this and
802 // used the speculation to mark other blocks as available.
804 char &BBVal = FullyAvailableBlocks[BB];
806 // If we didn't speculate on this, just return with it set to false.
812 // If we did speculate on this value, we could have blocks set to 1 that are
813 // incorrect. Walk the (transitive) successors of this block and mark them as
815 SmallVector<BasicBlock*, 32> BBWorklist;
816 BBWorklist.push_back(BB);
819 BasicBlock *Entry = BBWorklist.pop_back_val();
820 // Note that this sets blocks to 0 (unavailable) if they happen to not
821 // already be in FullyAvailableBlocks. This is safe.
822 char &EntryVal = FullyAvailableBlocks[Entry];
823 if (EntryVal == 0) continue; // Already unavailable.
825 // Mark as unavailable.
828 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
829 BBWorklist.push_back(*I);
830 } while (!BBWorklist.empty());
836 /// CanCoerceMustAliasedValueToLoad - Return true if
837 /// CoerceAvailableValueToLoadType will succeed.
838 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
840 const TargetData &TD) {
841 // If the loaded or stored value is an first class array or struct, don't try
842 // to transform them. We need to be able to bitcast to integer.
843 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
844 StoredVal->getType()->isStructTy() ||
845 StoredVal->getType()->isArrayTy())
848 // The store has to be at least as big as the load.
849 if (TD.getTypeSizeInBits(StoredVal->getType()) <
850 TD.getTypeSizeInBits(LoadTy))
857 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
858 /// then a load from a must-aliased pointer of a different type, try to coerce
859 /// the stored value. LoadedTy is the type of the load we want to replace and
860 /// InsertPt is the place to insert new instructions.
862 /// If we can't do it, return null.
863 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
864 const Type *LoadedTy,
865 Instruction *InsertPt,
866 const TargetData &TD) {
867 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
870 const Type *StoredValTy = StoredVal->getType();
872 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
873 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
875 // If the store and reload are the same size, we can always reuse it.
876 if (StoreSize == LoadSize) {
877 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) {
878 // Pointer to Pointer -> use bitcast.
879 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
882 // Convert source pointers to integers, which can be bitcast.
883 if (StoredValTy->isPointerTy()) {
884 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
885 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
888 const Type *TypeToCastTo = LoadedTy;
889 if (TypeToCastTo->isPointerTy())
890 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
892 if (StoredValTy != TypeToCastTo)
893 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
895 // Cast to pointer if the load needs a pointer type.
896 if (LoadedTy->isPointerTy())
897 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
902 // If the loaded value is smaller than the available value, then we can
903 // extract out a piece from it. If the available value is too small, then we
904 // can't do anything.
905 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
907 // Convert source pointers to integers, which can be manipulated.
908 if (StoredValTy->isPointerTy()) {
909 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
910 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
913 // Convert vectors and fp to integer, which can be manipulated.
914 if (!StoredValTy->isIntegerTy()) {
915 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
916 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
919 // If this is a big-endian system, we need to shift the value down to the low
920 // bits so that a truncate will work.
921 if (TD.isBigEndian()) {
922 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
923 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
926 // Truncate the integer to the right size now.
927 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
928 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
930 if (LoadedTy == NewIntTy)
933 // If the result is a pointer, inttoptr.
934 if (LoadedTy->isPointerTy())
935 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
937 // Otherwise, bitcast.
938 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
941 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
942 /// be expressed as a base pointer plus a constant offset. Return the base and
943 /// offset to the caller.
944 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
945 const TargetData &TD) {
946 Operator *PtrOp = dyn_cast<Operator>(Ptr);
947 if (PtrOp == 0) return Ptr;
949 // Just look through bitcasts.
950 if (PtrOp->getOpcode() == Instruction::BitCast)
951 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
953 // If this is a GEP with constant indices, we can look through it.
954 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
955 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
957 gep_type_iterator GTI = gep_type_begin(GEP);
958 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
960 ConstantInt *OpC = cast<ConstantInt>(*I);
961 if (OpC->isZero()) continue;
963 // Handle a struct and array indices which add their offset to the pointer.
964 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
965 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
967 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
968 Offset += OpC->getSExtValue()*Size;
972 // Re-sign extend from the pointer size if needed to get overflow edge cases
974 unsigned PtrSize = TD.getPointerSizeInBits();
976 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
978 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
982 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
983 /// memdep query of a load that ends up being a clobbering memory write (store,
984 /// memset, memcpy, memmove). This means that the write *may* provide bits used
985 /// by the load but we can't be sure because the pointers don't mustalias.
987 /// Check this case to see if there is anything more we can do before we give
988 /// up. This returns -1 if we have to give up, or a byte number in the stored
989 /// value of the piece that feeds the load.
990 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
992 uint64_t WriteSizeInBits,
993 const TargetData &TD) {
994 // If the loaded or stored value is an first class array or struct, don't try
995 // to transform them. We need to be able to bitcast to integer.
996 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
999 int64_t StoreOffset = 0, LoadOffset = 0;
1000 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1002 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
1003 if (StoreBase != LoadBase)
1006 // If the load and store are to the exact same address, they should have been
1007 // a must alias. AA must have gotten confused.
1008 // FIXME: Study to see if/when this happens.
1009 if (LoadOffset == StoreOffset) {
1011 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1012 << "Base = " << *StoreBase << "\n"
1013 << "Store Ptr = " << *WritePtr << "\n"
1014 << "Store Offs = " << StoreOffset << "\n"
1015 << "Load Ptr = " << *LoadPtr << "\n";
1021 // If the load and store don't overlap at all, the store doesn't provide
1022 // anything to the load. In this case, they really don't alias at all, AA
1023 // must have gotten confused.
1024 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1025 // remove this check, as it is duplicated with what we have below.
1026 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
1028 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1030 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1034 bool isAAFailure = false;
1035 if (StoreOffset < LoadOffset) {
1036 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1038 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1042 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1043 << "Base = " << *StoreBase << "\n"
1044 << "Store Ptr = " << *WritePtr << "\n"
1045 << "Store Offs = " << StoreOffset << "\n"
1046 << "Load Ptr = " << *LoadPtr << "\n";
1052 // If the Load isn't completely contained within the stored bits, we don't
1053 // have all the bits to feed it. We could do something crazy in the future
1054 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1056 if (StoreOffset > LoadOffset ||
1057 StoreOffset+StoreSize < LoadOffset+LoadSize)
1060 // Okay, we can do this transformation. Return the number of bytes into the
1061 // store that the load is.
1062 return LoadOffset-StoreOffset;
1065 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1066 /// memdep query of a load that ends up being a clobbering store.
1067 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
1069 const TargetData &TD) {
1070 // Cannot handle reading from store of first-class aggregate yet.
1071 if (DepSI->getOperand(0)->getType()->isStructTy() ||
1072 DepSI->getOperand(0)->getType()->isArrayTy())
1075 Value *StorePtr = DepSI->getPointerOperand();
1076 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1077 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1078 StorePtr, StoreSize, TD);
1081 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
1083 const TargetData &TD) {
1084 // If the mem operation is a non-constant size, we can't handle it.
1085 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1086 if (SizeCst == 0) return -1;
1087 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1089 // If this is memset, we just need to see if the offset is valid in the size
1091 if (MI->getIntrinsicID() == Intrinsic::memset)
1092 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1095 // If we have a memcpy/memmove, the only case we can handle is if this is a
1096 // copy from constant memory. In that case, we can read directly from the
1098 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1100 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1101 if (Src == 0) return -1;
1103 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1104 if (GV == 0 || !GV->isConstant()) return -1;
1106 // See if the access is within the bounds of the transfer.
1107 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1108 MI->getDest(), MemSizeInBits, TD);
1112 // Otherwise, see if we can constant fold a load from the constant with the
1113 // offset applied as appropriate.
1114 Src = ConstantExpr::getBitCast(Src,
1115 llvm::Type::getInt8PtrTy(Src->getContext()));
1116 Constant *OffsetCst =
1117 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1118 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1119 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1120 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1126 /// GetStoreValueForLoad - This function is called when we have a
1127 /// memdep query of a load that ends up being a clobbering store. This means
1128 /// that the store *may* provide bits used by the load but we can't be sure
1129 /// because the pointers don't mustalias. Check this case to see if there is
1130 /// anything more we can do before we give up.
1131 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1133 Instruction *InsertPt, const TargetData &TD){
1134 LLVMContext &Ctx = SrcVal->getType()->getContext();
1136 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1137 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1139 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1141 // Compute which bits of the stored value are being used by the load. Convert
1142 // to an integer type to start with.
1143 if (SrcVal->getType()->isPointerTy())
1144 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
1145 if (!SrcVal->getType()->isIntegerTy())
1146 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1149 // Shift the bits to the least significant depending on endianness.
1151 if (TD.isLittleEndian())
1152 ShiftAmt = Offset*8;
1154 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1157 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
1159 if (LoadSize != StoreSize)
1160 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1163 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1166 /// GetMemInstValueForLoad - This function is called when we have a
1167 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1168 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1169 const Type *LoadTy, Instruction *InsertPt,
1170 const TargetData &TD){
1171 LLVMContext &Ctx = LoadTy->getContext();
1172 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1174 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1176 // We know that this method is only called when the mem transfer fully
1177 // provides the bits for the load.
1178 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1179 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1180 // independently of what the offset is.
1181 Value *Val = MSI->getValue();
1183 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1185 Value *OneElt = Val;
1187 // Splat the value out to the right number of bits.
1188 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1189 // If we can double the number of bytes set, do it.
1190 if (NumBytesSet*2 <= LoadSize) {
1191 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1192 Val = Builder.CreateOr(Val, ShVal);
1197 // Otherwise insert one byte at a time.
1198 Value *ShVal = Builder.CreateShl(Val, 1*8);
1199 Val = Builder.CreateOr(OneElt, ShVal);
1203 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1206 // Otherwise, this is a memcpy/memmove from a constant global.
1207 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1208 Constant *Src = cast<Constant>(MTI->getSource());
1210 // Otherwise, see if we can constant fold a load from the constant with the
1211 // offset applied as appropriate.
1212 Src = ConstantExpr::getBitCast(Src,
1213 llvm::Type::getInt8PtrTy(Src->getContext()));
1214 Constant *OffsetCst =
1215 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1216 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1217 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1218 return ConstantFoldLoadFromConstPtr(Src, &TD);
1223 struct AvailableValueInBlock {
1224 /// BB - The basic block in question.
1227 SimpleVal, // A simple offsetted value that is accessed.
1228 MemIntrin // A memory intrinsic which is loaded from.
1231 /// V - The value that is live out of the block.
1232 PointerIntPair<Value *, 1, ValType> Val;
1234 /// Offset - The byte offset in Val that is interesting for the load query.
1237 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1238 unsigned Offset = 0) {
1239 AvailableValueInBlock Res;
1241 Res.Val.setPointer(V);
1242 Res.Val.setInt(SimpleVal);
1243 Res.Offset = Offset;
1247 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1248 unsigned Offset = 0) {
1249 AvailableValueInBlock Res;
1251 Res.Val.setPointer(MI);
1252 Res.Val.setInt(MemIntrin);
1253 Res.Offset = Offset;
1257 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1258 Value *getSimpleValue() const {
1259 assert(isSimpleValue() && "Wrong accessor");
1260 return Val.getPointer();
1263 MemIntrinsic *getMemIntrinValue() const {
1264 assert(!isSimpleValue() && "Wrong accessor");
1265 return cast<MemIntrinsic>(Val.getPointer());
1268 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1269 /// defined here to the specified type. This handles various coercion cases.
1270 Value *MaterializeAdjustedValue(const Type *LoadTy,
1271 const TargetData *TD) const {
1273 if (isSimpleValue()) {
1274 Res = getSimpleValue();
1275 if (Res->getType() != LoadTy) {
1276 assert(TD && "Need target data to handle type mismatch case");
1277 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1280 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1281 << *getSimpleValue() << '\n'
1282 << *Res << '\n' << "\n\n\n");
1285 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1286 LoadTy, BB->getTerminator(), *TD);
1287 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1288 << " " << *getMemIntrinValue() << '\n'
1289 << *Res << '\n' << "\n\n\n");
1295 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1296 /// construct SSA form, allowing us to eliminate LI. This returns the value
1297 /// that should be used at LI's definition site.
1298 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1299 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1300 const TargetData *TD,
1301 const DominatorTree &DT,
1302 AliasAnalysis *AA) {
1303 // Check for the fully redundant, dominating load case. In this case, we can
1304 // just use the dominating value directly.
1305 if (ValuesPerBlock.size() == 1 &&
1306 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1307 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1309 // Otherwise, we have to construct SSA form.
1310 SmallVector<PHINode*, 8> NewPHIs;
1311 SSAUpdater SSAUpdate(&NewPHIs);
1312 SSAUpdate.Initialize(LI);
1314 const Type *LoadTy = LI->getType();
1316 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1317 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1318 BasicBlock *BB = AV.BB;
1320 if (SSAUpdate.HasValueForBlock(BB))
1323 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1326 // Perform PHI construction.
1327 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1329 // If new PHI nodes were created, notify alias analysis.
1330 if (V->getType()->isPointerTy())
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<NonLocalDepResult, 64> Deps;
1349 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1351 //DEBUG(dbgs() << "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 dbgs() << "GVN: non-local load ";
1365 WriteAsOperand(dbgs(), LI);
1366 dbgs() << " 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(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1491 // Perform PHI construction.
1492 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1493 VN.getAliasAnalysis());
1494 LI->replaceAllUsesWith(V);
1496 if (isa<PHINode>(V))
1498 if (V->getType()->isPointerTy())
1499 MD->invalidateCachedPointerInfo(V);
1501 toErase.push_back(LI);
1506 if (!EnablePRE || !EnableLoadPRE)
1509 // Okay, we have *some* definitions of the value. This means that the value
1510 // is available in some of our (transitive) predecessors. Lets think about
1511 // doing PRE of this load. This will involve inserting a new load into the
1512 // predecessor when it's not available. We could do this in general, but
1513 // prefer to not increase code size. As such, we only do this when we know
1514 // that we only have to insert *one* load (which means we're basically moving
1515 // the load, not inserting a new one).
1517 SmallPtrSet<BasicBlock *, 4> Blockers;
1518 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1519 Blockers.insert(UnavailableBlocks[i]);
1521 // Lets find first basic block with more than one predecessor. Walk backwards
1522 // through predecessors if needed.
1523 BasicBlock *LoadBB = LI->getParent();
1524 BasicBlock *TmpBB = LoadBB;
1526 bool isSinglePred = false;
1527 bool allSingleSucc = true;
1528 while (TmpBB->getSinglePredecessor()) {
1529 isSinglePred = true;
1530 TmpBB = TmpBB->getSinglePredecessor();
1531 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1533 if (Blockers.count(TmpBB))
1535 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1536 allSingleSucc = false;
1542 // If we have a repl set with LI itself in it, this means we have a loop where
1543 // at least one of the values is LI. Since this means that we won't be able
1544 // to eliminate LI even if we insert uses in the other predecessors, we will
1545 // end up increasing code size. Reject this by scanning for LI.
1546 if (!EnableFullLoadPRE) {
1547 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1548 if (ValuesPerBlock[i].isSimpleValue() &&
1549 ValuesPerBlock[i].getSimpleValue() == LI)
1553 // FIXME: It is extremely unclear what this loop is doing, other than
1554 // artificially restricting loadpre.
1557 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1558 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1559 if (AV.isSimpleValue())
1560 // "Hot" Instruction is in some loop (because it dominates its dep.
1562 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1563 if (DT->dominates(LI, I)) {
1569 // We are interested only in "hot" instructions. We don't want to do any
1570 // mis-optimizations here.
1575 // Check to see how many predecessors have the loaded value fully
1577 DenseMap<BasicBlock*, Value*> PredLoads;
1578 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1579 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1580 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1581 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1582 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1584 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1586 BasicBlock *Pred = *PI;
1587 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1590 PredLoads[Pred] = 0;
1592 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1593 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1594 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1595 << Pred->getName() << "': " << *LI << '\n');
1598 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1599 toSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1604 // Decide whether PRE is profitable for this load.
1605 unsigned NumUnavailablePreds = PredLoads.size();
1606 assert(NumUnavailablePreds != 0 &&
1607 "Fully available value should be eliminated above!");
1608 if (!EnableFullLoadPRE) {
1609 // If this load is unavailable in multiple predecessors, reject it.
1610 // FIXME: If we could restructure the CFG, we could make a common pred with
1611 // all the preds that don't have an available LI and insert a new load into
1613 if (NumUnavailablePreds != 1)
1617 // Check if the load can safely be moved to all the unavailable predecessors.
1618 bool CanDoPRE = true;
1619 SmallVector<Instruction*, 8> NewInsts;
1620 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1621 E = PredLoads.end(); I != E; ++I) {
1622 BasicBlock *UnavailablePred = I->first;
1624 // Do PHI translation to get its value in the predecessor if necessary. The
1625 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1627 // If all preds have a single successor, then we know it is safe to insert
1628 // the load on the pred (?!?), so we can insert code to materialize the
1629 // pointer if 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, DT);
1637 LoadPtr = Address.getAddr();
1640 // If we couldn't find or insert a computation of this phi translated value,
1643 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1644 << *LI->getOperand(0) << "\n");
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,
1661 UnavailablePred->getTerminator(),
1662 LI->getAlignment(), TD)) {
1667 I->second = LoadPtr;
1671 while (!NewInsts.empty())
1672 NewInsts.pop_back_val()->eraseFromParent();
1676 // Okay, we can eliminate this load by inserting a reload in the predecessor
1677 // and using PHI construction to get the value in the other predecessors, do
1679 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1680 DEBUG(if (!NewInsts.empty())
1681 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1682 << *NewInsts.back() << '\n');
1684 // Assign value numbers to the new instructions.
1685 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1686 // FIXME: We really _ought_ to insert these value numbers into their
1687 // parent's availability map. However, in doing so, we risk getting into
1688 // ordering issues. If a block hasn't been processed yet, we would be
1689 // marking a value as AVAIL-IN, which isn't what we intend.
1690 VN.lookup_or_add(NewInsts[i]);
1693 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1694 E = PredLoads.end(); I != E; ++I) {
1695 BasicBlock *UnavailablePred = I->first;
1696 Value *LoadPtr = I->second;
1698 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1700 UnavailablePred->getTerminator());
1702 // Add the newly created load.
1703 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1705 MD->invalidateCachedPointerInfo(LoadPtr);
1706 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1709 // Perform PHI construction.
1710 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1711 VN.getAliasAnalysis());
1712 LI->replaceAllUsesWith(V);
1713 if (isa<PHINode>(V))
1715 if (V->getType()->isPointerTy())
1716 MD->invalidateCachedPointerInfo(V);
1718 toErase.push_back(LI);
1723 /// processLoad - Attempt to eliminate a load, first by eliminating it
1724 /// locally, and then attempting non-local elimination if that fails.
1725 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1729 if (L->isVolatile())
1732 // ... to a pointer that has been loaded from before...
1733 MemDepResult Dep = MD->getDependency(L);
1735 // If the value isn't available, don't do anything!
1736 if (Dep.isClobber()) {
1737 // Check to see if we have something like this:
1738 // store i32 123, i32* %P
1739 // %A = bitcast i32* %P to i8*
1740 // %B = gep i8* %A, i32 1
1743 // We could do that by recognizing if the clobber instructions are obviously
1744 // a common base + constant offset, and if the previous store (or memset)
1745 // completely covers this load. This sort of thing can happen in bitfield
1747 Value *AvailVal = 0;
1748 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1749 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1750 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1751 L->getPointerOperand(),
1754 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1755 L->getType(), L, *TD);
1758 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1759 // a value on from it.
1760 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1761 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1762 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1763 L->getPointerOperand(),
1766 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1771 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1772 << *AvailVal << '\n' << *L << "\n\n\n");
1774 // Replace the load!
1775 L->replaceAllUsesWith(AvailVal);
1776 if (AvailVal->getType()->isPointerTy())
1777 MD->invalidateCachedPointerInfo(AvailVal);
1779 toErase.push_back(L);
1785 // fast print dep, using operator<< on instruction would be too slow
1786 dbgs() << "GVN: load ";
1787 WriteAsOperand(dbgs(), L);
1788 Instruction *I = Dep.getInst();
1789 dbgs() << " is clobbered by " << *I << '\n';
1794 // If it is defined in another block, try harder.
1795 if (Dep.isNonLocal())
1796 return processNonLocalLoad(L, toErase);
1798 Instruction *DepInst = Dep.getInst();
1799 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1800 Value *StoredVal = DepSI->getOperand(0);
1802 // The store and load are to a must-aliased pointer, but they may not
1803 // actually have the same type. See if we know how to reuse the stored
1804 // value (depending on its type).
1805 const TargetData *TD = 0;
1806 if (StoredVal->getType() != L->getType()) {
1807 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1808 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1813 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1814 << '\n' << *L << "\n\n\n");
1821 L->replaceAllUsesWith(StoredVal);
1822 if (StoredVal->getType()->isPointerTy())
1823 MD->invalidateCachedPointerInfo(StoredVal);
1825 toErase.push_back(L);
1830 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1831 Value *AvailableVal = DepLI;
1833 // The loads are of a must-aliased pointer, but they may not actually have
1834 // the same type. See if we know how to reuse the previously loaded value
1835 // (depending on its type).
1836 const TargetData *TD = 0;
1837 if (DepLI->getType() != L->getType()) {
1838 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1839 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1840 if (AvailableVal == 0)
1843 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1844 << "\n" << *L << "\n\n\n");
1851 L->replaceAllUsesWith(AvailableVal);
1852 if (DepLI->getType()->isPointerTy())
1853 MD->invalidateCachedPointerInfo(DepLI);
1855 toErase.push_back(L);
1860 // If this load really doesn't depend on anything, then we must be loading an
1861 // undef value. This can happen when loading for a fresh allocation with no
1862 // intervening stores, for example.
1863 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1864 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1866 toErase.push_back(L);
1871 // If this load occurs either right after a lifetime begin,
1872 // then the loaded value is undefined.
1873 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1874 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1875 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1877 toErase.push_back(L);
1886 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1887 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1888 if (I == localAvail.end())
1891 ValueNumberScope *Locals = I->second;
1893 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1894 if (I != Locals->table.end())
1896 Locals = Locals->parent;
1903 /// processInstruction - When calculating availability, handle an instruction
1904 /// by inserting it into the appropriate sets
1905 bool GVN::processInstruction(Instruction *I,
1906 SmallVectorImpl<Instruction*> &toErase) {
1907 // Ignore dbg info intrinsics.
1908 if (isa<DbgInfoIntrinsic>(I))
1911 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1912 bool Changed = processLoad(LI, toErase);
1915 unsigned Num = VN.lookup_or_add(LI);
1916 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1922 uint32_t NextNum = VN.getNextUnusedValueNumber();
1923 unsigned Num = VN.lookup_or_add(I);
1925 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1926 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1928 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1931 Value *BranchCond = BI->getCondition();
1932 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1934 BasicBlock *TrueSucc = BI->getSuccessor(0);
1935 BasicBlock *FalseSucc = BI->getSuccessor(1);
1937 if (TrueSucc->getSinglePredecessor())
1938 localAvail[TrueSucc]->table[CondVN] =
1939 ConstantInt::getTrue(TrueSucc->getContext());
1940 if (FalseSucc->getSinglePredecessor())
1941 localAvail[FalseSucc]->table[CondVN] =
1942 ConstantInt::getFalse(TrueSucc->getContext());
1946 // Allocations are always uniquely numbered, so we can save time and memory
1947 // by fast failing them.
1948 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1949 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1953 // Collapse PHI nodes
1954 if (PHINode* p = dyn_cast<PHINode>(I)) {
1955 Value *constVal = CollapsePhi(p);
1958 p->replaceAllUsesWith(constVal);
1959 if (MD && constVal->getType()->isPointerTy())
1960 MD->invalidateCachedPointerInfo(constVal);
1963 toErase.push_back(p);
1965 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1968 // If the number we were assigned was a brand new VN, then we don't
1969 // need to do a lookup to see if the number already exists
1970 // somewhere in the domtree: it can't!
1971 } else if (Num == NextNum) {
1972 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1974 // Perform fast-path value-number based elimination of values inherited from
1976 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1979 I->replaceAllUsesWith(repl);
1980 if (MD && repl->getType()->isPointerTy())
1981 MD->invalidateCachedPointerInfo(repl);
1982 toErase.push_back(I);
1986 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1992 /// runOnFunction - This is the main transformation entry point for a function.
1993 bool GVN::runOnFunction(Function& F) {
1995 MD = &getAnalysis<MemoryDependenceAnalysis>();
1996 DT = &getAnalysis<DominatorTree>();
1997 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2001 bool Changed = false;
2002 bool ShouldContinue = true;
2004 // Merge unconditional branches, allowing PRE to catch more
2005 // optimization opportunities.
2006 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2007 BasicBlock *BB = FI;
2009 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2010 if (removedBlock) NumGVNBlocks++;
2012 Changed |= removedBlock;
2015 unsigned Iteration = 0;
2017 while (ShouldContinue) {
2018 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2019 ShouldContinue = iterateOnFunction(F);
2020 if (splitCriticalEdges())
2021 ShouldContinue = true;
2022 Changed |= ShouldContinue;
2027 bool PREChanged = true;
2028 while (PREChanged) {
2029 PREChanged = performPRE(F);
2030 Changed |= PREChanged;
2033 // FIXME: Should perform GVN again after PRE does something. PRE can move
2034 // computations into blocks where they become fully redundant. Note that
2035 // we can't do this until PRE's critical edge splitting updates memdep.
2036 // Actually, when this happens, we should just fully integrate PRE into GVN.
2038 cleanupGlobalSets();
2044 bool GVN::processBlock(BasicBlock *BB) {
2045 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
2046 // incrementing BI before processing an instruction).
2047 SmallVector<Instruction*, 8> toErase;
2048 bool ChangedFunction = false;
2050 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2052 ChangedFunction |= processInstruction(BI, toErase);
2053 if (toErase.empty()) {
2058 // If we need some instructions deleted, do it now.
2059 NumGVNInstr += toErase.size();
2061 // Avoid iterator invalidation.
2062 bool AtStart = BI == BB->begin();
2066 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2067 E = toErase.end(); I != E; ++I) {
2068 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2069 if (MD) MD->removeInstruction(*I);
2070 (*I)->eraseFromParent();
2071 DEBUG(verifyRemoved(*I));
2081 return ChangedFunction;
2084 /// performPRE - Perform a purely local form of PRE that looks for diamond
2085 /// control flow patterns and attempts to perform simple PRE at the join point.
2086 bool GVN::performPRE(Function &F) {
2087 bool Changed = false;
2088 DenseMap<BasicBlock*, Value*> predMap;
2089 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2090 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2091 BasicBlock *CurrentBlock = *DI;
2093 // Nothing to PRE in the entry block.
2094 if (CurrentBlock == &F.getEntryBlock()) continue;
2096 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2097 BE = CurrentBlock->end(); BI != BE; ) {
2098 Instruction *CurInst = BI++;
2100 if (isa<AllocaInst>(CurInst) ||
2101 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2102 CurInst->getType()->isVoidTy() ||
2103 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2104 isa<DbgInfoIntrinsic>(CurInst))
2107 uint32_t ValNo = VN.lookup(CurInst);
2109 // Look for the predecessors for PRE opportunities. We're
2110 // only trying to solve the basic diamond case, where
2111 // a value is computed in the successor and one predecessor,
2112 // but not the other. We also explicitly disallow cases
2113 // where the successor is its own predecessor, because they're
2114 // more complicated to get right.
2115 unsigned NumWith = 0;
2116 unsigned NumWithout = 0;
2117 BasicBlock *PREPred = 0;
2120 for (pred_iterator PI = pred_begin(CurrentBlock),
2121 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2122 // We're not interested in PRE where the block is its
2123 // own predecessor, or in blocks with predecessors
2124 // that are not reachable.
2125 if (*PI == CurrentBlock) {
2128 } else if (!localAvail.count(*PI)) {
2133 DenseMap<uint32_t, Value*>::iterator predV =
2134 localAvail[*PI]->table.find(ValNo);
2135 if (predV == localAvail[*PI]->table.end()) {
2138 } else if (predV->second == CurInst) {
2141 predMap[*PI] = predV->second;
2146 // Don't do PRE when it might increase code size, i.e. when
2147 // we would need to insert instructions in more than one pred.
2148 if (NumWithout != 1 || NumWith == 0)
2151 // Don't do PRE across indirect branch.
2152 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2155 // We can't do PRE safely on a critical edge, so instead we schedule
2156 // the edge to be split and perform the PRE the next time we iterate
2158 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2159 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2160 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2164 // Instantiate the expression in the predecessor that lacked it.
2165 // Because we are going top-down through the block, all value numbers
2166 // will be available in the predecessor by the time we need them. Any
2167 // that weren't originally present will have been instantiated earlier
2169 Instruction *PREInstr = CurInst->clone();
2170 bool success = true;
2171 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2172 Value *Op = PREInstr->getOperand(i);
2173 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2176 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2177 PREInstr->setOperand(i, V);
2184 // Fail out if we encounter an operand that is not available in
2185 // the PRE predecessor. This is typically because of loads which
2186 // are not value numbered precisely.
2189 DEBUG(verifyRemoved(PREInstr));
2193 PREInstr->insertBefore(PREPred->getTerminator());
2194 PREInstr->setName(CurInst->getName() + ".pre");
2195 predMap[PREPred] = PREInstr;
2196 VN.add(PREInstr, ValNo);
2199 // Update the availability map to include the new instruction.
2200 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
2202 // Create a PHI to make the value available in this block.
2203 PHINode* Phi = PHINode::Create(CurInst->getType(),
2204 CurInst->getName() + ".pre-phi",
2205 CurrentBlock->begin());
2206 for (pred_iterator PI = pred_begin(CurrentBlock),
2207 PE = pred_end(CurrentBlock); PI != PE; ++PI)
2208 Phi->addIncoming(predMap[*PI], *PI);
2211 localAvail[CurrentBlock]->table[ValNo] = Phi;
2213 CurInst->replaceAllUsesWith(Phi);
2214 if (MD && Phi->getType()->isPointerTy())
2215 MD->invalidateCachedPointerInfo(Phi);
2218 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2219 if (MD) MD->removeInstruction(CurInst);
2220 CurInst->eraseFromParent();
2221 DEBUG(verifyRemoved(CurInst));
2226 if (splitCriticalEdges())
2232 /// splitCriticalEdges - Split critical edges found during the previous
2233 /// iteration that may enable further optimization.
2234 bool GVN::splitCriticalEdges() {
2235 if (toSplit.empty())
2238 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2239 SplitCriticalEdge(Edge.first, Edge.second, this);
2240 } while (!toSplit.empty());
2241 MD->invalidateCachedPredecessors();
2245 /// iterateOnFunction - Executes one iteration of GVN
2246 bool GVN::iterateOnFunction(Function &F) {
2247 cleanupGlobalSets();
2249 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2250 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
2252 localAvail[DI->getBlock()] =
2253 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
2255 localAvail[DI->getBlock()] = new ValueNumberScope(0);
2258 // Top-down walk of the dominator tree
2259 bool Changed = false;
2261 // Needed for value numbering with phi construction to work.
2262 ReversePostOrderTraversal<Function*> RPOT(&F);
2263 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2264 RE = RPOT.end(); RI != RE; ++RI)
2265 Changed |= processBlock(*RI);
2267 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2268 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2269 Changed |= processBlock(DI->getBlock());
2275 void GVN::cleanupGlobalSets() {
2278 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2279 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2284 /// verifyRemoved - Verify that the specified instruction does not occur in our
2285 /// internal data structures.
2286 void GVN::verifyRemoved(const Instruction *Inst) const {
2287 VN.verifyRemoved(Inst);
2289 // Walk through the value number scope to make sure the instruction isn't
2290 // ferreted away in it.
2291 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
2292 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2293 const ValueNumberScope *VNS = I->second;
2296 for (DenseMap<uint32_t, Value*>::const_iterator
2297 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2298 assert(II->second != Inst && "Inst still in value numbering scope!");