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/InstructionSimplify.h"
39 #include "llvm/Analysis/Loads.h"
40 #include "llvm/Analysis/MemoryBuiltins.h"
41 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
42 #include "llvm/Analysis/PHITransAddr.h"
43 #include "llvm/Support/Allocator.h"
44 #include "llvm/Support/CFG.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/IRBuilder.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Target/TargetData.h"
52 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
53 #include "llvm/Transforms/Utils/Local.h"
54 #include "llvm/Transforms/Utils/SSAUpdater.h"
58 STATISTIC(NumGVNInstr, "Number of instructions deleted");
59 STATISTIC(NumGVNLoad, "Number of loads deleted");
60 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
61 STATISTIC(NumGVNBlocks, "Number of blocks merged");
62 STATISTIC(NumPRELoad, "Number of loads PRE'd");
64 static cl::opt<bool> EnablePRE("enable-pre",
65 cl::init(true), cl::Hidden);
66 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
68 //===----------------------------------------------------------------------===//
70 //===----------------------------------------------------------------------===//
72 /// This class holds the mapping between values and value numbers. It is used
73 /// as an efficient mechanism to determine the expression-wise equivalence of
77 enum ExpressionOpcode {
78 ADD = Instruction::Add,
79 FADD = Instruction::FAdd,
80 SUB = Instruction::Sub,
81 FSUB = Instruction::FSub,
82 MUL = Instruction::Mul,
83 FMUL = Instruction::FMul,
84 UDIV = Instruction::UDiv,
85 SDIV = Instruction::SDiv,
86 FDIV = Instruction::FDiv,
87 UREM = Instruction::URem,
88 SREM = Instruction::SRem,
89 FREM = Instruction::FRem,
90 SHL = Instruction::Shl,
91 LSHR = Instruction::LShr,
92 ASHR = Instruction::AShr,
93 AND = Instruction::And,
95 XOR = Instruction::Xor,
96 TRUNC = Instruction::Trunc,
97 ZEXT = Instruction::ZExt,
98 SEXT = Instruction::SExt,
99 FPTOUI = Instruction::FPToUI,
100 FPTOSI = Instruction::FPToSI,
101 UITOFP = Instruction::UIToFP,
102 SITOFP = Instruction::SIToFP,
103 FPTRUNC = Instruction::FPTrunc,
104 FPEXT = Instruction::FPExt,
105 PTRTOINT = Instruction::PtrToInt,
106 INTTOPTR = Instruction::IntToPtr,
107 BITCAST = Instruction::BitCast,
108 ICMPEQ, ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
109 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
110 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
111 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
112 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
113 SHUFFLE, SELECT, GEP, CALL, CONSTANT,
114 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
116 ExpressionOpcode opcode;
118 SmallVector<uint32_t, 4> varargs;
122 Expression(ExpressionOpcode o) : opcode(o) { }
124 bool operator==(const Expression &other) const {
125 if (opcode != other.opcode)
127 else if (opcode == EMPTY || opcode == TOMBSTONE)
129 else if (type != other.type)
131 else if (function != other.function)
134 if (varargs.size() != other.varargs.size())
137 for (size_t i = 0; i < varargs.size(); ++i)
138 if (varargs[i] != other.varargs[i])
145 /*bool operator!=(const Expression &other) const {
146 return !(*this == other);
152 DenseMap<Value*, uint32_t> valueNumbering;
153 DenseMap<Expression, uint32_t> expressionNumbering;
155 MemoryDependenceAnalysis* MD;
158 uint32_t nextValueNumber;
160 Expression::ExpressionOpcode getOpcode(CmpInst* C);
161 Expression create_expression(BinaryOperator* BO);
162 Expression create_expression(CmpInst* C);
163 Expression create_expression(ShuffleVectorInst* V);
164 Expression create_expression(ExtractElementInst* C);
165 Expression create_expression(InsertElementInst* V);
166 Expression create_expression(SelectInst* V);
167 Expression create_expression(CastInst* C);
168 Expression create_expression(GetElementPtrInst* G);
169 Expression create_expression(CallInst* C);
170 Expression create_expression(ExtractValueInst* C);
171 Expression create_expression(InsertValueInst* C);
173 uint32_t lookup_or_add_call(CallInst* C);
175 ValueTable() : nextValueNumber(1) { }
176 uint32_t lookup_or_add(Value *V);
177 uint32_t lookup(Value *V) const;
178 void add(Value *V, uint32_t num);
180 void erase(Value *v);
181 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
182 AliasAnalysis *getAliasAnalysis() const { return AA; }
183 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
184 void setDomTree(DominatorTree* D) { DT = D; }
185 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
186 void verifyRemoved(const Value *) const;
191 template <> struct DenseMapInfo<Expression> {
192 static inline Expression getEmptyKey() {
193 return Expression(Expression::EMPTY);
196 static inline Expression getTombstoneKey() {
197 return Expression(Expression::TOMBSTONE);
200 static unsigned getHashValue(const Expression e) {
201 unsigned hash = e.opcode;
203 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
204 (unsigned)((uintptr_t)e.type >> 9));
206 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
207 E = e.varargs.end(); I != E; ++I)
208 hash = *I + hash * 37;
210 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
211 (unsigned)((uintptr_t)e.function >> 9)) +
216 static bool isEqual(const Expression &LHS, const Expression &RHS) {
222 struct isPodLike<Expression> { static const bool value = true; };
226 //===----------------------------------------------------------------------===//
227 // ValueTable Internal Functions
228 //===----------------------------------------------------------------------===//
230 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
231 if (isa<ICmpInst>(C)) {
232 switch (C->getPredicate()) {
233 default: // THIS SHOULD NEVER HAPPEN
234 llvm_unreachable("Comparison with unknown predicate?");
235 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
236 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
237 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
238 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
239 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
240 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
241 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
242 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
243 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
244 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
247 switch (C->getPredicate()) {
248 default: // THIS SHOULD NEVER HAPPEN
249 llvm_unreachable("Comparison with unknown predicate?");
250 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
251 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
252 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
253 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
254 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
255 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
256 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
257 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
258 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
259 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
260 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
261 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
262 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
263 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
268 Expression ValueTable::create_expression(CallInst* C) {
271 e.type = C->getType();
272 e.function = C->getCalledFunction();
273 e.opcode = Expression::CALL;
276 for (CallInst::op_iterator I = CS.arg_begin(), E = CS.arg_end();
278 e.varargs.push_back(lookup_or_add(*I));
283 Expression ValueTable::create_expression(BinaryOperator* BO) {
285 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
286 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
288 e.type = BO->getType();
289 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode());
294 Expression ValueTable::create_expression(CmpInst* C) {
297 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
298 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
300 e.type = C->getType();
301 e.opcode = getOpcode(C);
306 Expression ValueTable::create_expression(CastInst* C) {
309 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
311 e.type = C->getType();
312 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode());
317 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
320 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
321 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
322 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
324 e.type = S->getType();
325 e.opcode = Expression::SHUFFLE;
330 Expression ValueTable::create_expression(ExtractElementInst* E) {
333 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
334 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
336 e.type = E->getType();
337 e.opcode = Expression::EXTRACT;
342 Expression ValueTable::create_expression(InsertElementInst* I) {
345 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
346 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
347 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
349 e.type = I->getType();
350 e.opcode = Expression::INSERT;
355 Expression ValueTable::create_expression(SelectInst* I) {
358 e.varargs.push_back(lookup_or_add(I->getCondition()));
359 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
360 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
362 e.type = I->getType();
363 e.opcode = Expression::SELECT;
368 Expression ValueTable::create_expression(GetElementPtrInst* G) {
371 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
373 e.type = G->getType();
374 e.opcode = Expression::GEP;
376 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
378 e.varargs.push_back(lookup_or_add(*I));
383 Expression ValueTable::create_expression(ExtractValueInst* E) {
386 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
387 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
389 e.varargs.push_back(*II);
391 e.type = E->getType();
392 e.opcode = Expression::EXTRACTVALUE;
397 Expression ValueTable::create_expression(InsertValueInst* E) {
400 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
401 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
402 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
404 e.varargs.push_back(*II);
406 e.type = E->getType();
407 e.opcode = Expression::INSERTVALUE;
412 //===----------------------------------------------------------------------===//
413 // ValueTable External Functions
414 //===----------------------------------------------------------------------===//
416 /// add - Insert a value into the table with a specified value number.
417 void ValueTable::add(Value *V, uint32_t num) {
418 valueNumbering.insert(std::make_pair(V, num));
421 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
422 if (AA->doesNotAccessMemory(C)) {
423 Expression exp = create_expression(C);
424 uint32_t& e = expressionNumbering[exp];
425 if (!e) e = nextValueNumber++;
426 valueNumbering[C] = e;
428 } else if (AA->onlyReadsMemory(C)) {
429 Expression exp = create_expression(C);
430 uint32_t& e = expressionNumbering[exp];
432 e = nextValueNumber++;
433 valueNumbering[C] = e;
437 e = nextValueNumber++;
438 valueNumbering[C] = e;
442 MemDepResult local_dep = MD->getDependency(C);
444 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
445 valueNumbering[C] = nextValueNumber;
446 return nextValueNumber++;
449 if (local_dep.isDef()) {
450 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
452 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
453 valueNumbering[C] = nextValueNumber;
454 return nextValueNumber++;
457 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
458 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
459 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
461 valueNumbering[C] = nextValueNumber;
462 return nextValueNumber++;
466 uint32_t v = lookup_or_add(local_cdep);
467 valueNumbering[C] = v;
472 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
473 MD->getNonLocalCallDependency(CallSite(C));
474 // FIXME: call/call dependencies for readonly calls should return def, not
475 // clobber! Move the checking logic to MemDep!
478 // Check to see if we have a single dominating call instruction that is
480 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
481 const NonLocalDepEntry *I = &deps[i];
482 // Ignore non-local dependencies.
483 if (I->getResult().isNonLocal())
486 // We don't handle non-depedencies. If we already have a call, reject
487 // instruction dependencies.
488 if (I->getResult().isClobber() || cdep != 0) {
493 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
494 // FIXME: All duplicated with non-local case.
495 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
496 cdep = NonLocalDepCall;
505 valueNumbering[C] = nextValueNumber;
506 return nextValueNumber++;
509 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
510 valueNumbering[C] = nextValueNumber;
511 return nextValueNumber++;
513 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
514 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
515 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
517 valueNumbering[C] = nextValueNumber;
518 return nextValueNumber++;
522 uint32_t v = lookup_or_add(cdep);
523 valueNumbering[C] = v;
527 valueNumbering[C] = nextValueNumber;
528 return nextValueNumber++;
532 /// lookup_or_add - Returns the value number for the specified value, assigning
533 /// it a new number if it did not have one before.
534 uint32_t ValueTable::lookup_or_add(Value *V) {
535 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
536 if (VI != valueNumbering.end())
539 if (!isa<Instruction>(V)) {
540 valueNumbering[V] = nextValueNumber;
541 return nextValueNumber++;
544 Instruction* I = cast<Instruction>(V);
546 switch (I->getOpcode()) {
547 case Instruction::Call:
548 return lookup_or_add_call(cast<CallInst>(I));
549 case Instruction::Add:
550 case Instruction::FAdd:
551 case Instruction::Sub:
552 case Instruction::FSub:
553 case Instruction::Mul:
554 case Instruction::FMul:
555 case Instruction::UDiv:
556 case Instruction::SDiv:
557 case Instruction::FDiv:
558 case Instruction::URem:
559 case Instruction::SRem:
560 case Instruction::FRem:
561 case Instruction::Shl:
562 case Instruction::LShr:
563 case Instruction::AShr:
564 case Instruction::And:
565 case Instruction::Or :
566 case Instruction::Xor:
567 exp = create_expression(cast<BinaryOperator>(I));
569 case Instruction::ICmp:
570 case Instruction::FCmp:
571 exp = create_expression(cast<CmpInst>(I));
573 case Instruction::Trunc:
574 case Instruction::ZExt:
575 case Instruction::SExt:
576 case Instruction::FPToUI:
577 case Instruction::FPToSI:
578 case Instruction::UIToFP:
579 case Instruction::SIToFP:
580 case Instruction::FPTrunc:
581 case Instruction::FPExt:
582 case Instruction::PtrToInt:
583 case Instruction::IntToPtr:
584 case Instruction::BitCast:
585 exp = create_expression(cast<CastInst>(I));
587 case Instruction::Select:
588 exp = create_expression(cast<SelectInst>(I));
590 case Instruction::ExtractElement:
591 exp = create_expression(cast<ExtractElementInst>(I));
593 case Instruction::InsertElement:
594 exp = create_expression(cast<InsertElementInst>(I));
596 case Instruction::ShuffleVector:
597 exp = create_expression(cast<ShuffleVectorInst>(I));
599 case Instruction::ExtractValue:
600 exp = create_expression(cast<ExtractValueInst>(I));
602 case Instruction::InsertValue:
603 exp = create_expression(cast<InsertValueInst>(I));
605 case Instruction::GetElementPtr:
606 exp = create_expression(cast<GetElementPtrInst>(I));
609 valueNumbering[V] = nextValueNumber;
610 return nextValueNumber++;
613 uint32_t& e = expressionNumbering[exp];
614 if (!e) e = nextValueNumber++;
615 valueNumbering[V] = e;
619 /// lookup - Returns the value number of the specified value. Fails if
620 /// the value has not yet been numbered.
621 uint32_t ValueTable::lookup(Value *V) const {
622 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
623 assert(VI != valueNumbering.end() && "Value not numbered?");
627 /// clear - Remove all entries from the ValueTable
628 void ValueTable::clear() {
629 valueNumbering.clear();
630 expressionNumbering.clear();
634 /// erase - Remove a value from the value numbering
635 void ValueTable::erase(Value *V) {
636 valueNumbering.erase(V);
639 /// verifyRemoved - Verify that the value is removed from all internal data
641 void ValueTable::verifyRemoved(const Value *V) const {
642 for (DenseMap<Value*, uint32_t>::const_iterator
643 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
644 assert(I->first != V && "Inst still occurs in value numbering map!");
648 //===----------------------------------------------------------------------===//
650 //===----------------------------------------------------------------------===//
653 struct ValueNumberScope {
654 ValueNumberScope* parent;
655 DenseMap<uint32_t, Value*> table;
657 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
663 class GVN : public FunctionPass {
664 bool runOnFunction(Function &F);
666 static char ID; // Pass identification, replacement for typeid
667 explicit GVN(bool noloads = false)
668 : FunctionPass(ID), NoLoads(noloads), MD(0) {
669 initializeGVNPass(*PassRegistry::getPassRegistry());
674 MemoryDependenceAnalysis *MD;
676 const TargetData* TD;
680 /// NumberTable - A mapping from value numers to lists of Value*'s that
681 /// have that value number. Use lookupNumber to query it.
682 DenseMap<uint32_t, std::pair<Value*, void*> > NumberTable;
683 BumpPtrAllocator TableAllocator;
685 /// insert_table - Push a new Value to the NumberTable onto the list for
686 /// its value number.
687 void insert_table(uint32_t N, Value *V) {
688 std::pair<Value*, void*>& Curr = NumberTable[N];
694 std::pair<Value*, void*>* Node =
695 TableAllocator.Allocate<std::pair<Value*, void*> >();
697 Node->second = Curr.second;
701 /// erase_table - Scan the list of values corresponding to a given value
702 /// number, and remove the given value if encountered.
703 void erase_table(uint32_t N, Value *V) {
704 std::pair<Value*, void*>* Prev = 0;
705 std::pair<Value*, void*>* Curr = &NumberTable[N];
707 while (Curr->first != V) {
709 Curr = static_cast<std::pair<Value*, void*>*>(Curr->second);
713 Prev->second = Curr->second;
718 std::pair<Value*, void*>* Next =
719 static_cast<std::pair<Value*, void*>*>(Curr->second);
720 Curr->first = Next->first;
721 Curr->second = Next->second;
726 // List of critical edges to be split between iterations.
727 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
729 // This transformation requires dominator postdominator info
730 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
731 AU.addRequired<DominatorTree>();
733 AU.addRequired<MemoryDependenceAnalysis>();
734 AU.addRequired<AliasAnalysis>();
736 AU.addPreserved<DominatorTree>();
737 AU.addPreserved<AliasAnalysis>();
741 // FIXME: eliminate or document these better
742 bool processLoad(LoadInst* L,
743 SmallVectorImpl<Instruction*> &toErase);
744 bool processInstruction(Instruction *I,
745 SmallVectorImpl<Instruction*> &toErase);
746 bool processNonLocalLoad(LoadInst* L,
747 SmallVectorImpl<Instruction*> &toErase);
748 bool processBlock(BasicBlock *BB);
749 void dump(DenseMap<uint32_t, Value*>& d);
750 bool iterateOnFunction(Function &F);
751 bool performPRE(Function& F);
752 Value *lookupNumber(BasicBlock *BB, uint32_t num);
753 void cleanupGlobalSets();
754 void verifyRemoved(const Instruction *I) const;
755 bool splitCriticalEdges();
761 // createGVNPass - The public interface to this file...
762 FunctionPass *llvm::createGVNPass(bool NoLoads) {
763 return new GVN(NoLoads);
766 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
767 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
768 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
769 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
770 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
772 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
774 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
775 E = d.end(); I != E; ++I) {
776 errs() << I->first << "\n";
782 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
783 /// we're analyzing is fully available in the specified block. As we go, keep
784 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
785 /// map is actually a tri-state map with the following values:
786 /// 0) we know the block *is not* fully available.
787 /// 1) we know the block *is* fully available.
788 /// 2) we do not know whether the block is fully available or not, but we are
789 /// currently speculating that it will be.
790 /// 3) we are speculating for this block and have used that to speculate for
792 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
793 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
794 // Optimistically assume that the block is fully available and check to see
795 // if we already know about this block in one lookup.
796 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
797 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
799 // If the entry already existed for this block, return the precomputed value.
801 // If this is a speculative "available" value, mark it as being used for
802 // speculation of other blocks.
803 if (IV.first->second == 2)
804 IV.first->second = 3;
805 return IV.first->second != 0;
808 // Otherwise, see if it is fully available in all predecessors.
809 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
811 // If this block has no predecessors, it isn't live-in here.
813 goto SpeculationFailure;
815 for (; PI != PE; ++PI)
816 // If the value isn't fully available in one of our predecessors, then it
817 // isn't fully available in this block either. Undo our previous
818 // optimistic assumption and bail out.
819 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
820 goto SpeculationFailure;
824 // SpeculationFailure - If we get here, we found out that this is not, after
825 // all, a fully-available block. We have a problem if we speculated on this and
826 // used the speculation to mark other blocks as available.
828 char &BBVal = FullyAvailableBlocks[BB];
830 // If we didn't speculate on this, just return with it set to false.
836 // If we did speculate on this value, we could have blocks set to 1 that are
837 // incorrect. Walk the (transitive) successors of this block and mark them as
839 SmallVector<BasicBlock*, 32> BBWorklist;
840 BBWorklist.push_back(BB);
843 BasicBlock *Entry = BBWorklist.pop_back_val();
844 // Note that this sets blocks to 0 (unavailable) if they happen to not
845 // already be in FullyAvailableBlocks. This is safe.
846 char &EntryVal = FullyAvailableBlocks[Entry];
847 if (EntryVal == 0) continue; // Already unavailable.
849 // Mark as unavailable.
852 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
853 BBWorklist.push_back(*I);
854 } while (!BBWorklist.empty());
860 /// CanCoerceMustAliasedValueToLoad - Return true if
861 /// CoerceAvailableValueToLoadType will succeed.
862 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
864 const TargetData &TD) {
865 // If the loaded or stored value is an first class array or struct, don't try
866 // to transform them. We need to be able to bitcast to integer.
867 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
868 StoredVal->getType()->isStructTy() ||
869 StoredVal->getType()->isArrayTy())
872 // The store has to be at least as big as the load.
873 if (TD.getTypeSizeInBits(StoredVal->getType()) <
874 TD.getTypeSizeInBits(LoadTy))
881 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
882 /// then a load from a must-aliased pointer of a different type, try to coerce
883 /// the stored value. LoadedTy is the type of the load we want to replace and
884 /// InsertPt is the place to insert new instructions.
886 /// If we can't do it, return null.
887 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
888 const Type *LoadedTy,
889 Instruction *InsertPt,
890 const TargetData &TD) {
891 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
894 const Type *StoredValTy = StoredVal->getType();
896 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy);
897 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
899 // If the store and reload are the same size, we can always reuse it.
900 if (StoreSize == LoadSize) {
901 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) {
902 // Pointer to Pointer -> use bitcast.
903 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
906 // Convert source pointers to integers, which can be bitcast.
907 if (StoredValTy->isPointerTy()) {
908 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
909 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
912 const Type *TypeToCastTo = LoadedTy;
913 if (TypeToCastTo->isPointerTy())
914 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
916 if (StoredValTy != TypeToCastTo)
917 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
919 // Cast to pointer if the load needs a pointer type.
920 if (LoadedTy->isPointerTy())
921 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
926 // If the loaded value is smaller than the available value, then we can
927 // extract out a piece from it. If the available value is too small, then we
928 // can't do anything.
929 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
931 // Convert source pointers to integers, which can be manipulated.
932 if (StoredValTy->isPointerTy()) {
933 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
934 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
937 // Convert vectors and fp to integer, which can be manipulated.
938 if (!StoredValTy->isIntegerTy()) {
939 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
940 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
943 // If this is a big-endian system, we need to shift the value down to the low
944 // bits so that a truncate will work.
945 if (TD.isBigEndian()) {
946 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
947 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
950 // Truncate the integer to the right size now.
951 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
952 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
954 if (LoadedTy == NewIntTy)
957 // If the result is a pointer, inttoptr.
958 if (LoadedTy->isPointerTy())
959 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
961 // Otherwise, bitcast.
962 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
965 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
966 /// be expressed as a base pointer plus a constant offset. Return the base and
967 /// offset to the caller.
968 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
969 const TargetData &TD) {
970 Operator *PtrOp = dyn_cast<Operator>(Ptr);
971 if (PtrOp == 0) return Ptr;
973 // Just look through bitcasts.
974 if (PtrOp->getOpcode() == Instruction::BitCast)
975 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
977 // If this is a GEP with constant indices, we can look through it.
978 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
979 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
981 gep_type_iterator GTI = gep_type_begin(GEP);
982 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
984 ConstantInt *OpC = cast<ConstantInt>(*I);
985 if (OpC->isZero()) continue;
987 // Handle a struct and array indices which add their offset to the pointer.
988 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
989 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
991 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
992 Offset += OpC->getSExtValue()*Size;
996 // Re-sign extend from the pointer size if needed to get overflow edge cases
998 unsigned PtrSize = TD.getPointerSizeInBits();
1000 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
1002 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
1006 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
1007 /// memdep query of a load that ends up being a clobbering memory write (store,
1008 /// memset, memcpy, memmove). This means that the write *may* provide bits used
1009 /// by the load but we can't be sure because the pointers don't mustalias.
1011 /// Check this case to see if there is anything more we can do before we give
1012 /// up. This returns -1 if we have to give up, or a byte number in the stored
1013 /// value of the piece that feeds the load.
1014 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
1016 uint64_t WriteSizeInBits,
1017 const TargetData &TD) {
1018 // If the loaded or stored value is an first class array or struct, don't try
1019 // to transform them. We need to be able to bitcast to integer.
1020 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
1023 int64_t StoreOffset = 0, LoadOffset = 0;
1024 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1026 GetBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
1027 if (StoreBase != LoadBase)
1030 // If the load and store are to the exact same address, they should have been
1031 // a must alias. AA must have gotten confused.
1032 // FIXME: Study to see if/when this happens. One case is forwarding a memset
1033 // to a load from the base of the memset.
1035 if (LoadOffset == StoreOffset) {
1036 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1037 << "Base = " << *StoreBase << "\n"
1038 << "Store Ptr = " << *WritePtr << "\n"
1039 << "Store Offs = " << StoreOffset << "\n"
1040 << "Load Ptr = " << *LoadPtr << "\n";
1045 // If the load and store don't overlap at all, the store doesn't provide
1046 // anything to the load. In this case, they really don't alias at all, AA
1047 // must have gotten confused.
1048 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
1050 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1052 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1056 bool isAAFailure = false;
1057 if (StoreOffset < LoadOffset)
1058 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1060 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1064 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1065 << "Base = " << *StoreBase << "\n"
1066 << "Store Ptr = " << *WritePtr << "\n"
1067 << "Store Offs = " << StoreOffset << "\n"
1068 << "Load Ptr = " << *LoadPtr << "\n";
1074 // If the Load isn't completely contained within the stored bits, we don't
1075 // have all the bits to feed it. We could do something crazy in the future
1076 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1078 if (StoreOffset > LoadOffset ||
1079 StoreOffset+StoreSize < LoadOffset+LoadSize)
1082 // Okay, we can do this transformation. Return the number of bytes into the
1083 // store that the load is.
1084 return LoadOffset-StoreOffset;
1087 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1088 /// memdep query of a load that ends up being a clobbering store.
1089 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
1091 const TargetData &TD) {
1092 // Cannot handle reading from store of first-class aggregate yet.
1093 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1094 DepSI->getValueOperand()->getType()->isArrayTy())
1097 Value *StorePtr = DepSI->getPointerOperand();
1098 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1099 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1100 StorePtr, StoreSize, TD);
1103 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
1105 const TargetData &TD) {
1106 // If the mem operation is a non-constant size, we can't handle it.
1107 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1108 if (SizeCst == 0) return -1;
1109 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1111 // If this is memset, we just need to see if the offset is valid in the size
1113 if (MI->getIntrinsicID() == Intrinsic::memset)
1114 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1117 // If we have a memcpy/memmove, the only case we can handle is if this is a
1118 // copy from constant memory. In that case, we can read directly from the
1120 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1122 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1123 if (Src == 0) return -1;
1125 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1126 if (GV == 0 || !GV->isConstant()) return -1;
1128 // See if the access is within the bounds of the transfer.
1129 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1130 MI->getDest(), MemSizeInBits, TD);
1134 // Otherwise, see if we can constant fold a load from the constant with the
1135 // offset applied as appropriate.
1136 Src = ConstantExpr::getBitCast(Src,
1137 llvm::Type::getInt8PtrTy(Src->getContext()));
1138 Constant *OffsetCst =
1139 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1140 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1141 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1142 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1148 /// GetStoreValueForLoad - This function is called when we have a
1149 /// memdep query of a load that ends up being a clobbering store. This means
1150 /// that the store *may* provide bits used by the load but we can't be sure
1151 /// because the pointers don't mustalias. Check this case to see if there is
1152 /// anything more we can do before we give up.
1153 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1155 Instruction *InsertPt, const TargetData &TD){
1156 LLVMContext &Ctx = SrcVal->getType()->getContext();
1158 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1159 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1161 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1163 // Compute which bits of the stored value are being used by the load. Convert
1164 // to an integer type to start with.
1165 if (SrcVal->getType()->isPointerTy())
1166 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
1167 if (!SrcVal->getType()->isIntegerTy())
1168 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1171 // Shift the bits to the least significant depending on endianness.
1173 if (TD.isLittleEndian())
1174 ShiftAmt = Offset*8;
1176 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1179 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
1181 if (LoadSize != StoreSize)
1182 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1185 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1188 /// GetMemInstValueForLoad - This function is called when we have a
1189 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1190 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1191 const Type *LoadTy, Instruction *InsertPt,
1192 const TargetData &TD){
1193 LLVMContext &Ctx = LoadTy->getContext();
1194 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1196 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1198 // We know that this method is only called when the mem transfer fully
1199 // provides the bits for the load.
1200 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1201 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1202 // independently of what the offset is.
1203 Value *Val = MSI->getValue();
1205 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1207 Value *OneElt = Val;
1209 // Splat the value out to the right number of bits.
1210 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1211 // If we can double the number of bytes set, do it.
1212 if (NumBytesSet*2 <= LoadSize) {
1213 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1214 Val = Builder.CreateOr(Val, ShVal);
1219 // Otherwise insert one byte at a time.
1220 Value *ShVal = Builder.CreateShl(Val, 1*8);
1221 Val = Builder.CreateOr(OneElt, ShVal);
1225 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1228 // Otherwise, this is a memcpy/memmove from a constant global.
1229 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1230 Constant *Src = cast<Constant>(MTI->getSource());
1232 // Otherwise, see if we can constant fold a load from the constant with the
1233 // offset applied as appropriate.
1234 Src = ConstantExpr::getBitCast(Src,
1235 llvm::Type::getInt8PtrTy(Src->getContext()));
1236 Constant *OffsetCst =
1237 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1238 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1239 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1240 return ConstantFoldLoadFromConstPtr(Src, &TD);
1245 struct AvailableValueInBlock {
1246 /// BB - The basic block in question.
1249 SimpleVal, // A simple offsetted value that is accessed.
1250 MemIntrin // A memory intrinsic which is loaded from.
1253 /// V - The value that is live out of the block.
1254 PointerIntPair<Value *, 1, ValType> Val;
1256 /// Offset - The byte offset in Val that is interesting for the load query.
1259 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1260 unsigned Offset = 0) {
1261 AvailableValueInBlock Res;
1263 Res.Val.setPointer(V);
1264 Res.Val.setInt(SimpleVal);
1265 Res.Offset = Offset;
1269 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1270 unsigned Offset = 0) {
1271 AvailableValueInBlock Res;
1273 Res.Val.setPointer(MI);
1274 Res.Val.setInt(MemIntrin);
1275 Res.Offset = Offset;
1279 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1280 Value *getSimpleValue() const {
1281 assert(isSimpleValue() && "Wrong accessor");
1282 return Val.getPointer();
1285 MemIntrinsic *getMemIntrinValue() const {
1286 assert(!isSimpleValue() && "Wrong accessor");
1287 return cast<MemIntrinsic>(Val.getPointer());
1290 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1291 /// defined here to the specified type. This handles various coercion cases.
1292 Value *MaterializeAdjustedValue(const Type *LoadTy,
1293 const TargetData *TD) const {
1295 if (isSimpleValue()) {
1296 Res = getSimpleValue();
1297 if (Res->getType() != LoadTy) {
1298 assert(TD && "Need target data to handle type mismatch case");
1299 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1302 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1303 << *getSimpleValue() << '\n'
1304 << *Res << '\n' << "\n\n\n");
1307 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1308 LoadTy, BB->getTerminator(), *TD);
1309 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1310 << " " << *getMemIntrinValue() << '\n'
1311 << *Res << '\n' << "\n\n\n");
1319 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1320 /// construct SSA form, allowing us to eliminate LI. This returns the value
1321 /// that should be used at LI's definition site.
1322 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1323 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1324 const TargetData *TD,
1325 const DominatorTree &DT,
1326 AliasAnalysis *AA) {
1327 // Check for the fully redundant, dominating load case. In this case, we can
1328 // just use the dominating value directly.
1329 if (ValuesPerBlock.size() == 1 &&
1330 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1331 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1333 // Otherwise, we have to construct SSA form.
1334 SmallVector<PHINode*, 8> NewPHIs;
1335 SSAUpdater SSAUpdate(&NewPHIs);
1336 SSAUpdate.Initialize(LI->getType(), LI->getName());
1338 const Type *LoadTy = LI->getType();
1340 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1341 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1342 BasicBlock *BB = AV.BB;
1344 if (SSAUpdate.HasValueForBlock(BB))
1347 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1350 // Perform PHI construction.
1351 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1353 // If new PHI nodes were created, notify alias analysis.
1354 if (V->getType()->isPointerTy())
1355 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1356 AA->copyValue(LI, NewPHIs[i]);
1361 static bool isLifetimeStart(const Instruction *Inst) {
1362 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1363 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1367 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1368 /// non-local by performing PHI construction.
1369 bool GVN::processNonLocalLoad(LoadInst *LI,
1370 SmallVectorImpl<Instruction*> &toErase) {
1371 // Find the non-local dependencies of the load.
1372 SmallVector<NonLocalDepResult, 64> Deps;
1373 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1374 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1375 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1376 // << Deps.size() << *LI << '\n');
1378 // If we had to process more than one hundred blocks to find the
1379 // dependencies, this load isn't worth worrying about. Optimizing
1380 // it will be too expensive.
1381 if (Deps.size() > 100)
1384 // If we had a phi translation failure, we'll have a single entry which is a
1385 // clobber in the current block. Reject this early.
1386 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
1388 dbgs() << "GVN: non-local load ";
1389 WriteAsOperand(dbgs(), LI);
1390 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1395 // Filter out useless results (non-locals, etc). Keep track of the blocks
1396 // where we have a value available in repl, also keep track of whether we see
1397 // dependencies that produce an unknown value for the load (such as a call
1398 // that could potentially clobber the load).
1399 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1400 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1402 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1403 BasicBlock *DepBB = Deps[i].getBB();
1404 MemDepResult DepInfo = Deps[i].getResult();
1406 if (DepInfo.isClobber()) {
1407 // The address being loaded in this non-local block may not be the same as
1408 // the pointer operand of the load if PHI translation occurs. Make sure
1409 // to consider the right address.
1410 Value *Address = Deps[i].getAddress();
1412 // If the dependence is to a store that writes to a superset of the bits
1413 // read by the load, we can extract the bits we need for the load from the
1415 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1416 if (TD && Address) {
1417 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1420 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1421 DepSI->getValueOperand(),
1428 // If the clobbering value is a memset/memcpy/memmove, see if we can
1429 // forward a value on from it.
1430 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1431 if (TD && Address) {
1432 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1435 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1442 UnavailableBlocks.push_back(DepBB);
1446 Instruction *DepInst = DepInfo.getInst();
1448 // Loading the allocation -> undef.
1449 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1450 // Loading immediately after lifetime begin -> undef.
1451 isLifetimeStart(DepInst)) {
1452 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1453 UndefValue::get(LI->getType())));
1457 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1458 // Reject loads and stores that are to the same address but are of
1459 // different types if we have to.
1460 if (S->getValueOperand()->getType() != LI->getType()) {
1461 // If the stored value is larger or equal to the loaded value, we can
1463 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1464 LI->getType(), *TD)) {
1465 UnavailableBlocks.push_back(DepBB);
1470 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1471 S->getValueOperand()));
1475 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1476 // If the types mismatch and we can't handle it, reject reuse of the load.
1477 if (LD->getType() != LI->getType()) {
1478 // If the stored value is larger or equal to the loaded value, we can
1480 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1481 UnavailableBlocks.push_back(DepBB);
1485 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1489 UnavailableBlocks.push_back(DepBB);
1493 // If we have no predecessors that produce a known value for this load, exit
1495 if (ValuesPerBlock.empty()) return false;
1497 // If all of the instructions we depend on produce a known value for this
1498 // load, then it is fully redundant and we can use PHI insertion to compute
1499 // its value. Insert PHIs and remove the fully redundant value now.
1500 if (UnavailableBlocks.empty()) {
1501 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1503 // Perform PHI construction.
1504 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1505 VN.getAliasAnalysis());
1506 LI->replaceAllUsesWith(V);
1508 if (isa<PHINode>(V))
1510 if (V->getType()->isPointerTy())
1511 MD->invalidateCachedPointerInfo(V);
1513 toErase.push_back(LI);
1518 if (!EnablePRE || !EnableLoadPRE)
1521 // Okay, we have *some* definitions of the value. This means that the value
1522 // is available in some of our (transitive) predecessors. Lets think about
1523 // doing PRE of this load. This will involve inserting a new load into the
1524 // predecessor when it's not available. We could do this in general, but
1525 // prefer to not increase code size. As such, we only do this when we know
1526 // that we only have to insert *one* load (which means we're basically moving
1527 // the load, not inserting a new one).
1529 SmallPtrSet<BasicBlock *, 4> Blockers;
1530 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1531 Blockers.insert(UnavailableBlocks[i]);
1533 // Lets find first basic block with more than one predecessor. Walk backwards
1534 // through predecessors if needed.
1535 BasicBlock *LoadBB = LI->getParent();
1536 BasicBlock *TmpBB = LoadBB;
1538 bool isSinglePred = false;
1539 bool allSingleSucc = true;
1540 while (TmpBB->getSinglePredecessor()) {
1541 isSinglePred = true;
1542 TmpBB = TmpBB->getSinglePredecessor();
1543 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1545 if (Blockers.count(TmpBB))
1548 // If any of these blocks has more than one successor (i.e. if the edge we
1549 // just traversed was critical), then there are other paths through this
1550 // block along which the load may not be anticipated. Hoisting the load
1551 // above this block would be adding the load to execution paths along
1552 // which it was not previously executed.
1553 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1560 // FIXME: It is extremely unclear what this loop is doing, other than
1561 // artificially restricting loadpre.
1564 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1565 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1566 if (AV.isSimpleValue())
1567 // "Hot" Instruction is in some loop (because it dominates its dep.
1569 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1570 if (DT->dominates(LI, I)) {
1576 // We are interested only in "hot" instructions. We don't want to do any
1577 // mis-optimizations here.
1582 // Check to see how many predecessors have the loaded value fully
1584 DenseMap<BasicBlock*, Value*> PredLoads;
1585 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1586 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1587 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1588 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1589 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1591 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1592 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1594 BasicBlock *Pred = *PI;
1595 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1598 PredLoads[Pred] = 0;
1600 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1601 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1602 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1603 << Pred->getName() << "': " << *LI << '\n');
1606 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1607 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1610 if (!NeedToSplit.empty()) {
1611 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1615 // Decide whether PRE is profitable for this load.
1616 unsigned NumUnavailablePreds = PredLoads.size();
1617 assert(NumUnavailablePreds != 0 &&
1618 "Fully available value should be eliminated above!");
1620 // If this load is unavailable in multiple predecessors, reject it.
1621 // FIXME: If we could restructure the CFG, we could make a common pred with
1622 // all the preds that don't have an available LI and insert a new load into
1624 if (NumUnavailablePreds != 1)
1627 // Check if the load can safely be moved to all the unavailable predecessors.
1628 bool CanDoPRE = true;
1629 SmallVector<Instruction*, 8> NewInsts;
1630 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1631 E = PredLoads.end(); I != E; ++I) {
1632 BasicBlock *UnavailablePred = I->first;
1634 // Do PHI translation to get its value in the predecessor if necessary. The
1635 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1637 // If all preds have a single successor, then we know it is safe to insert
1638 // the load on the pred (?!?), so we can insert code to materialize the
1639 // pointer if it is not available.
1640 PHITransAddr Address(LI->getPointerOperand(), TD);
1642 if (allSingleSucc) {
1643 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1646 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1647 LoadPtr = Address.getAddr();
1650 // If we couldn't find or insert a computation of this phi translated value,
1653 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1654 << *LI->getPointerOperand() << "\n");
1659 // Make sure it is valid to move this load here. We have to watch out for:
1660 // @1 = getelementptr (i8* p, ...
1661 // test p and branch if == 0
1663 // It is valid to have the getelementptr before the test, even if p can be 0,
1664 // as getelementptr only does address arithmetic.
1665 // If we are not pushing the value through any multiple-successor blocks
1666 // we do not have this case. Otherwise, check that the load is safe to
1667 // put anywhere; this can be improved, but should be conservatively safe.
1668 if (!allSingleSucc &&
1669 // FIXME: REEVALUTE THIS.
1670 !isSafeToLoadUnconditionally(LoadPtr,
1671 UnavailablePred->getTerminator(),
1672 LI->getAlignment(), TD)) {
1677 I->second = LoadPtr;
1681 while (!NewInsts.empty())
1682 NewInsts.pop_back_val()->eraseFromParent();
1686 // Okay, we can eliminate this load by inserting a reload in the predecessor
1687 // and using PHI construction to get the value in the other predecessors, do
1689 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1690 DEBUG(if (!NewInsts.empty())
1691 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1692 << *NewInsts.back() << '\n');
1694 // Assign value numbers to the new instructions.
1695 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1696 // FIXME: We really _ought_ to insert these value numbers into their
1697 // parent's availability map. However, in doing so, we risk getting into
1698 // ordering issues. If a block hasn't been processed yet, we would be
1699 // marking a value as AVAIL-IN, which isn't what we intend.
1700 VN.lookup_or_add(NewInsts[i]);
1703 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1704 E = PredLoads.end(); I != E; ++I) {
1705 BasicBlock *UnavailablePred = I->first;
1706 Value *LoadPtr = I->second;
1708 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1710 UnavailablePred->getTerminator());
1712 // Add the newly created load.
1713 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1715 MD->invalidateCachedPointerInfo(LoadPtr);
1716 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1719 // Perform PHI construction.
1720 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1721 VN.getAliasAnalysis());
1722 LI->replaceAllUsesWith(V);
1723 if (isa<PHINode>(V))
1725 if (V->getType()->isPointerTy())
1726 MD->invalidateCachedPointerInfo(V);
1728 toErase.push_back(LI);
1733 /// processLoad - Attempt to eliminate a load, first by eliminating it
1734 /// locally, and then attempting non-local elimination if that fails.
1735 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1739 if (L->isVolatile())
1742 // ... to a pointer that has been loaded from before...
1743 MemDepResult Dep = MD->getDependency(L);
1745 // If the value isn't available, don't do anything!
1746 if (Dep.isClobber()) {
1747 // Check to see if we have something like this:
1748 // store i32 123, i32* %P
1749 // %A = bitcast i32* %P to i8*
1750 // %B = gep i8* %A, i32 1
1753 // We could do that by recognizing if the clobber instructions are obviously
1754 // a common base + constant offset, and if the previous store (or memset)
1755 // completely covers this load. This sort of thing can happen in bitfield
1757 Value *AvailVal = 0;
1758 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1760 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1761 L->getPointerOperand(),
1764 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1765 L->getType(), L, *TD);
1768 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1769 // a value on from it.
1770 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1772 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1773 L->getPointerOperand(),
1776 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1781 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1782 << *AvailVal << '\n' << *L << "\n\n\n");
1784 // Replace the load!
1785 L->replaceAllUsesWith(AvailVal);
1786 if (AvailVal->getType()->isPointerTy())
1787 MD->invalidateCachedPointerInfo(AvailVal);
1789 toErase.push_back(L);
1795 // fast print dep, using operator<< on instruction would be too slow
1796 dbgs() << "GVN: load ";
1797 WriteAsOperand(dbgs(), L);
1798 Instruction *I = Dep.getInst();
1799 dbgs() << " is clobbered by " << *I << '\n';
1804 // If it is defined in another block, try harder.
1805 if (Dep.isNonLocal())
1806 return processNonLocalLoad(L, toErase);
1808 Instruction *DepInst = Dep.getInst();
1809 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1810 Value *StoredVal = DepSI->getValueOperand();
1812 // The store and load are to a must-aliased pointer, but they may not
1813 // actually have the same type. See if we know how to reuse the stored
1814 // value (depending on its type).
1815 if (StoredVal->getType() != L->getType()) {
1817 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1822 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1823 << '\n' << *L << "\n\n\n");
1830 L->replaceAllUsesWith(StoredVal);
1831 if (StoredVal->getType()->isPointerTy())
1832 MD->invalidateCachedPointerInfo(StoredVal);
1834 toErase.push_back(L);
1839 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1840 Value *AvailableVal = DepLI;
1842 // The loads are of a must-aliased pointer, but they may not actually have
1843 // the same type. See if we know how to reuse the previously loaded value
1844 // (depending on its type).
1845 if (DepLI->getType() != L->getType()) {
1847 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1848 if (AvailableVal == 0)
1851 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1852 << "\n" << *L << "\n\n\n");
1859 L->replaceAllUsesWith(AvailableVal);
1860 if (DepLI->getType()->isPointerTy())
1861 MD->invalidateCachedPointerInfo(DepLI);
1863 toErase.push_back(L);
1868 // If this load really doesn't depend on anything, then we must be loading an
1869 // undef value. This can happen when loading for a fresh allocation with no
1870 // intervening stores, for example.
1871 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1872 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1874 toErase.push_back(L);
1879 // If this load occurs either right after a lifetime begin,
1880 // then the loaded value is undefined.
1881 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1882 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1883 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1885 toErase.push_back(L);
1894 // lookupNumber - In order to find a leader for a given value number at a
1895 // specific basic block, we first obtain the list of all Values for that number,
1896 // and then scan the list to find one whose block dominates the block in
1897 // question. This is fast because dominator tree queries consist of only
1898 // a few comparisons of DFS numbers.
1899 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1900 std::pair<Value*, void*> Vals = NumberTable[num];
1901 if (!Vals.first) return 0;
1902 Instruction *Inst = dyn_cast<Instruction>(Vals.first);
1903 if (!Inst) return Vals.first;
1904 BasicBlock *Parent = Inst->getParent();
1905 if (DT->dominates(Parent, BB))
1908 std::pair<Value*, void*>* Next =
1909 static_cast<std::pair<Value*, void*>*>(Vals.second);
1911 Instruction *CurrInst = dyn_cast<Instruction>(Next->first);
1912 if (!CurrInst) return Next->first;
1914 BasicBlock *Parent = CurrInst->getParent();
1915 if (DT->dominates(Parent, BB))
1918 Next = static_cast<std::pair<Value*, void*>*>(Next->second);
1925 /// processInstruction - When calculating availability, handle an instruction
1926 /// by inserting it into the appropriate sets
1927 bool GVN::processInstruction(Instruction *I,
1928 SmallVectorImpl<Instruction*> &toErase) {
1929 // Ignore dbg info intrinsics.
1930 if (isa<DbgInfoIntrinsic>(I))
1933 // If the instruction can be easily simplified then do so now in preference
1934 // to value numbering it. Value numbering often exposes redundancies, for
1935 // example if it determines that %y is equal to %x then the instruction
1936 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1937 if (Value *V = SimplifyInstruction(I, TD, DT)) {
1938 I->replaceAllUsesWith(V);
1939 if (MD && V->getType()->isPointerTy())
1940 MD->invalidateCachedPointerInfo(V);
1942 toErase.push_back(I);
1946 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1947 bool Changed = processLoad(LI, toErase);
1950 unsigned Num = VN.lookup_or_add(LI);
1951 insert_table(Num, LI);
1957 uint32_t NextNum = VN.getNextUnusedValueNumber();
1958 unsigned Num = VN.lookup_or_add(I);
1960 // Allocations are always uniquely numbered, so we can save time and memory
1961 // by fast failing them.
1962 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1963 insert_table(Num, I);
1967 if (isa<PHINode>(I)) {
1968 insert_table(Num, I);
1970 // If the number we were assigned was a brand new VN, then we don't
1971 // need to do a lookup to see if the number already exists
1972 // somewhere in the domtree: it can't!
1973 } else if (Num == NextNum) {
1974 insert_table(Num, I);
1976 // Perform fast-path value-number based elimination of values inherited from
1978 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1981 I->replaceAllUsesWith(repl);
1982 if (MD && repl->getType()->isPointerTy())
1983 MD->invalidateCachedPointerInfo(repl);
1984 toErase.push_back(I);
1988 insert_table(Num, I);
1994 /// runOnFunction - This is the main transformation entry point for a function.
1995 bool GVN::runOnFunction(Function& F) {
1997 MD = &getAnalysis<MemoryDependenceAnalysis>();
1998 DT = &getAnalysis<DominatorTree>();
1999 TD = getAnalysisIfAvailable<TargetData>();
2000 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2004 bool Changed = false;
2005 bool ShouldContinue = true;
2007 // Merge unconditional branches, allowing PRE to catch more
2008 // optimization opportunities.
2009 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2010 BasicBlock *BB = FI;
2012 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2013 if (removedBlock) ++NumGVNBlocks;
2015 Changed |= removedBlock;
2018 unsigned Iteration = 0;
2020 while (ShouldContinue) {
2021 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2022 ShouldContinue = iterateOnFunction(F);
2023 if (splitCriticalEdges())
2024 ShouldContinue = true;
2025 Changed |= ShouldContinue;
2030 bool PREChanged = true;
2031 while (PREChanged) {
2032 PREChanged = performPRE(F);
2033 Changed |= PREChanged;
2036 // FIXME: Should perform GVN again after PRE does something. PRE can move
2037 // computations into blocks where they become fully redundant. Note that
2038 // we can't do this until PRE's critical edge splitting updates memdep.
2039 // Actually, when this happens, we should just fully integrate PRE into GVN.
2041 cleanupGlobalSets();
2047 bool GVN::processBlock(BasicBlock *BB) {
2048 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
2049 // incrementing BI before processing an instruction).
2050 SmallVector<Instruction*, 8> toErase;
2051 bool ChangedFunction = false;
2053 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2055 ChangedFunction |= processInstruction(BI, toErase);
2056 if (toErase.empty()) {
2061 // If we need some instructions deleted, do it now.
2062 NumGVNInstr += toErase.size();
2064 // Avoid iterator invalidation.
2065 bool AtStart = BI == BB->begin();
2069 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2070 E = toErase.end(); I != E; ++I) {
2071 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2072 if (MD) MD->removeInstruction(*I);
2073 (*I)->eraseFromParent();
2074 DEBUG(verifyRemoved(*I));
2084 return ChangedFunction;
2087 /// performPRE - Perform a purely local form of PRE that looks for diamond
2088 /// control flow patterns and attempts to perform simple PRE at the join point.
2089 bool GVN::performPRE(Function &F) {
2090 bool Changed = false;
2091 DenseMap<BasicBlock*, Value*> predMap;
2092 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2093 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2094 BasicBlock *CurrentBlock = *DI;
2096 // Nothing to PRE in the entry block.
2097 if (CurrentBlock == &F.getEntryBlock()) continue;
2099 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2100 BE = CurrentBlock->end(); BI != BE; ) {
2101 Instruction *CurInst = BI++;
2103 if (isa<AllocaInst>(CurInst) ||
2104 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2105 CurInst->getType()->isVoidTy() ||
2106 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2107 isa<DbgInfoIntrinsic>(CurInst))
2110 // We don't currently value number ANY inline asm calls.
2111 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2112 if (CallI->isInlineAsm())
2115 uint32_t ValNo = VN.lookup(CurInst);
2117 // Look for the predecessors for PRE opportunities. We're
2118 // only trying to solve the basic diamond case, where
2119 // a value is computed in the successor and one predecessor,
2120 // but not the other. We also explicitly disallow cases
2121 // where the successor is its own predecessor, because they're
2122 // more complicated to get right.
2123 unsigned NumWith = 0;
2124 unsigned NumWithout = 0;
2125 BasicBlock *PREPred = 0;
2128 for (pred_iterator PI = pred_begin(CurrentBlock),
2129 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2130 BasicBlock *P = *PI;
2131 // We're not interested in PRE where the block is its
2132 // own predecessor, or in blocks with predecessors
2133 // that are not reachable.
2134 if (P == CurrentBlock) {
2137 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2142 Value* predV = lookupNumber(P, ValNo);
2146 } else if (predV == CurInst) {
2154 // Don't do PRE when it might increase code size, i.e. when
2155 // we would need to insert instructions in more than one pred.
2156 if (NumWithout != 1 || NumWith == 0)
2159 // Don't do PRE across indirect branch.
2160 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2163 // We can't do PRE safely on a critical edge, so instead we schedule
2164 // the edge to be split and perform the PRE the next time we iterate
2166 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2167 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2168 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2172 // Instantiate the expression in the predecessor that lacked it.
2173 // Because we are going top-down through the block, all value numbers
2174 // will be available in the predecessor by the time we need them. Any
2175 // that weren't originally present will have been instantiated earlier
2177 Instruction *PREInstr = CurInst->clone();
2178 bool success = true;
2179 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2180 Value *Op = PREInstr->getOperand(i);
2181 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2184 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2185 PREInstr->setOperand(i, V);
2192 // Fail out if we encounter an operand that is not available in
2193 // the PRE predecessor. This is typically because of loads which
2194 // are not value numbered precisely.
2197 DEBUG(verifyRemoved(PREInstr));
2201 PREInstr->insertBefore(PREPred->getTerminator());
2202 PREInstr->setName(CurInst->getName() + ".pre");
2203 predMap[PREPred] = PREInstr;
2204 VN.add(PREInstr, ValNo);
2207 // Update the availability map to include the new instruction.
2208 insert_table(ValNo, PREInstr);
2210 // Create a PHI to make the value available in this block.
2211 PHINode* Phi = PHINode::Create(CurInst->getType(),
2212 CurInst->getName() + ".pre-phi",
2213 CurrentBlock->begin());
2214 for (pred_iterator PI = pred_begin(CurrentBlock),
2215 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2216 BasicBlock *P = *PI;
2217 Phi->addIncoming(predMap[P], P);
2221 insert_table(ValNo, Phi);
2223 CurInst->replaceAllUsesWith(Phi);
2224 if (MD && Phi->getType()->isPointerTy())
2225 MD->invalidateCachedPointerInfo(Phi);
2227 erase_table(ValNo, CurInst);
2229 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2230 if (MD) MD->removeInstruction(CurInst);
2231 CurInst->eraseFromParent();
2232 DEBUG(verifyRemoved(CurInst));
2237 if (splitCriticalEdges())
2243 /// splitCriticalEdges - Split critical edges found during the previous
2244 /// iteration that may enable further optimization.
2245 bool GVN::splitCriticalEdges() {
2246 if (toSplit.empty())
2249 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2250 SplitCriticalEdge(Edge.first, Edge.second, this);
2251 } while (!toSplit.empty());
2252 if (MD) MD->invalidateCachedPredecessors();
2256 /// iterateOnFunction - Executes one iteration of GVN
2257 bool GVN::iterateOnFunction(Function &F) {
2258 cleanupGlobalSets();
2260 // Top-down walk of the dominator tree
2261 bool Changed = false;
2263 // Needed for value numbering with phi construction to work.
2264 ReversePostOrderTraversal<Function*> RPOT(&F);
2265 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2266 RE = RPOT.end(); RI != RE; ++RI)
2267 Changed |= processBlock(*RI);
2269 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2270 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2271 Changed |= processBlock(DI->getBlock());
2277 void GVN::cleanupGlobalSets() {
2279 NumberTable.clear();
2280 TableAllocator.Reset();
2283 /// verifyRemoved - Verify that the specified instruction does not occur in our
2284 /// internal data structures.
2285 void GVN::verifyRemoved(const Instruction *Inst) const {
2286 VN.verifyRemoved(Inst);
2288 // Walk through the value number scope to make sure the instruction isn't
2289 // ferreted away in it.
2290 for (DenseMap<uint32_t, std::pair<Value*, void*> >::const_iterator
2291 I = NumberTable.begin(), E = NumberTable.end(); I != E; ++I) {
2292 std::pair<Value*, void*> const * Node = &I->second;
2293 assert(Node->first != Inst && "Inst still in value numbering scope!");
2295 while (Node->second) {
2296 Node = static_cast<std::pair<Value*, void*>*>(Node->second);
2297 assert(Node->first != Inst && "Inst still in value numbering scope!");