1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/BasicBlock.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/IntrinsicInst.h"
25 #include "llvm/LLVMContext.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Value.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/DepthFirstIterator.h"
30 #include "llvm/ADT/PostOrderIterator.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/AliasAnalysis.h"
35 #include "llvm/Analysis/ConstantFolding.h"
36 #include "llvm/Analysis/Dominators.h"
37 #include "llvm/Analysis/MemoryBuiltins.h"
38 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
39 #include "llvm/Support/CFG.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/GetElementPtrTypeIterator.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SSAUpdater.h"
53 STATISTIC(NumGVNInstr, "Number of instructions deleted");
54 STATISTIC(NumGVNLoad, "Number of loads deleted");
55 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
56 STATISTIC(NumGVNBlocks, "Number of blocks merged");
57 STATISTIC(NumPRELoad, "Number of loads PRE'd");
59 static cl::opt<bool> EnablePRE("enable-pre",
60 cl::init(true), cl::Hidden);
61 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
63 //===----------------------------------------------------------------------===//
65 //===----------------------------------------------------------------------===//
67 /// This class holds the mapping between values and value numbers. It is used
68 /// as an efficient mechanism to determine the expression-wise equivalence of
72 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
73 UDIV, SDIV, FDIV, UREM, SREM,
74 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
75 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
76 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
77 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
78 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
79 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
80 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
81 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
82 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
83 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
85 ExpressionOpcode opcode;
87 SmallVector<uint32_t, 4> varargs;
91 Expression(ExpressionOpcode o) : opcode(o) { }
93 bool operator==(const Expression &other) const {
94 if (opcode != other.opcode)
96 else if (opcode == EMPTY || opcode == TOMBSTONE)
98 else if (type != other.type)
100 else if (function != other.function)
103 if (varargs.size() != other.varargs.size())
106 for (size_t i = 0; i < varargs.size(); ++i)
107 if (varargs[i] != other.varargs[i])
114 bool operator!=(const Expression &other) const {
115 return !(*this == other);
121 DenseMap<Value*, uint32_t> valueNumbering;
122 DenseMap<Expression, uint32_t> expressionNumbering;
124 MemoryDependenceAnalysis* MD;
127 uint32_t nextValueNumber;
129 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
130 Expression::ExpressionOpcode getOpcode(CmpInst* C);
131 Expression::ExpressionOpcode getOpcode(CastInst* C);
132 Expression create_expression(BinaryOperator* BO);
133 Expression create_expression(CmpInst* C);
134 Expression create_expression(ShuffleVectorInst* V);
135 Expression create_expression(ExtractElementInst* C);
136 Expression create_expression(InsertElementInst* V);
137 Expression create_expression(SelectInst* V);
138 Expression create_expression(CastInst* C);
139 Expression create_expression(GetElementPtrInst* G);
140 Expression create_expression(CallInst* C);
141 Expression create_expression(Constant* C);
142 Expression create_expression(ExtractValueInst* C);
143 Expression create_expression(InsertValueInst* C);
145 uint32_t lookup_or_add_call(CallInst* C);
147 ValueTable() : nextValueNumber(1) { }
148 uint32_t lookup_or_add(Value *V);
149 uint32_t lookup(Value *V) const;
150 void add(Value *V, uint32_t num);
152 void erase(Value *v);
154 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
155 AliasAnalysis *getAliasAnalysis() const { return AA; }
156 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
157 void setDomTree(DominatorTree* D) { DT = D; }
158 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
159 void verifyRemoved(const Value *) const;
164 template <> struct DenseMapInfo<Expression> {
165 static inline Expression getEmptyKey() {
166 return Expression(Expression::EMPTY);
169 static inline Expression getTombstoneKey() {
170 return Expression(Expression::TOMBSTONE);
173 static unsigned getHashValue(const Expression e) {
174 unsigned hash = e.opcode;
176 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
177 (unsigned)((uintptr_t)e.type >> 9));
179 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
180 E = e.varargs.end(); I != E; ++I)
181 hash = *I + hash * 37;
183 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
184 (unsigned)((uintptr_t)e.function >> 9)) +
189 static bool isEqual(const Expression &LHS, const Expression &RHS) {
192 static bool isPod() { return true; }
196 //===----------------------------------------------------------------------===//
197 // ValueTable Internal Functions
198 //===----------------------------------------------------------------------===//
199 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
200 switch(BO->getOpcode()) {
201 default: // THIS SHOULD NEVER HAPPEN
202 llvm_unreachable("Binary operator with unknown opcode?");
203 case Instruction::Add: return Expression::ADD;
204 case Instruction::FAdd: return Expression::FADD;
205 case Instruction::Sub: return Expression::SUB;
206 case Instruction::FSub: return Expression::FSUB;
207 case Instruction::Mul: return Expression::MUL;
208 case Instruction::FMul: return Expression::FMUL;
209 case Instruction::UDiv: return Expression::UDIV;
210 case Instruction::SDiv: return Expression::SDIV;
211 case Instruction::FDiv: return Expression::FDIV;
212 case Instruction::URem: return Expression::UREM;
213 case Instruction::SRem: return Expression::SREM;
214 case Instruction::FRem: return Expression::FREM;
215 case Instruction::Shl: return Expression::SHL;
216 case Instruction::LShr: return Expression::LSHR;
217 case Instruction::AShr: return Expression::ASHR;
218 case Instruction::And: return Expression::AND;
219 case Instruction::Or: return Expression::OR;
220 case Instruction::Xor: return Expression::XOR;
224 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
225 if (isa<ICmpInst>(C)) {
226 switch (C->getPredicate()) {
227 default: // THIS SHOULD NEVER HAPPEN
228 llvm_unreachable("Comparison with unknown predicate?");
229 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
230 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
231 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
232 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
233 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
234 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
235 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
236 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
237 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
238 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
241 switch (C->getPredicate()) {
242 default: // THIS SHOULD NEVER HAPPEN
243 llvm_unreachable("Comparison with unknown predicate?");
244 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
245 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
246 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
247 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
248 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
249 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
250 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
251 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
252 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
253 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
254 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
255 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
256 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
257 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
262 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
263 switch(C->getOpcode()) {
264 default: // THIS SHOULD NEVER HAPPEN
265 llvm_unreachable("Cast operator with unknown opcode?");
266 case Instruction::Trunc: return Expression::TRUNC;
267 case Instruction::ZExt: return Expression::ZEXT;
268 case Instruction::SExt: return Expression::SEXT;
269 case Instruction::FPToUI: return Expression::FPTOUI;
270 case Instruction::FPToSI: return Expression::FPTOSI;
271 case Instruction::UIToFP: return Expression::UITOFP;
272 case Instruction::SIToFP: return Expression::SITOFP;
273 case Instruction::FPTrunc: return Expression::FPTRUNC;
274 case Instruction::FPExt: return Expression::FPEXT;
275 case Instruction::PtrToInt: return Expression::PTRTOINT;
276 case Instruction::IntToPtr: return Expression::INTTOPTR;
277 case Instruction::BitCast: return Expression::BITCAST;
281 Expression ValueTable::create_expression(CallInst* C) {
284 e.type = C->getType();
285 e.function = C->getCalledFunction();
286 e.opcode = Expression::CALL;
288 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
290 e.varargs.push_back(lookup_or_add(*I));
295 Expression ValueTable::create_expression(BinaryOperator* BO) {
297 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
298 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
300 e.type = BO->getType();
301 e.opcode = getOpcode(BO);
306 Expression ValueTable::create_expression(CmpInst* C) {
309 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
310 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
312 e.type = C->getType();
313 e.opcode = getOpcode(C);
318 Expression ValueTable::create_expression(CastInst* C) {
321 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
323 e.type = C->getType();
324 e.opcode = getOpcode(C);
329 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
332 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
333 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
334 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
336 e.type = S->getType();
337 e.opcode = Expression::SHUFFLE;
342 Expression ValueTable::create_expression(ExtractElementInst* E) {
345 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
346 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
348 e.type = E->getType();
349 e.opcode = Expression::EXTRACT;
354 Expression ValueTable::create_expression(InsertElementInst* I) {
357 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
358 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
359 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
361 e.type = I->getType();
362 e.opcode = Expression::INSERT;
367 Expression ValueTable::create_expression(SelectInst* I) {
370 e.varargs.push_back(lookup_or_add(I->getCondition()));
371 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
372 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
374 e.type = I->getType();
375 e.opcode = Expression::SELECT;
380 Expression ValueTable::create_expression(GetElementPtrInst* G) {
383 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
385 e.type = G->getType();
386 e.opcode = Expression::GEP;
388 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
390 e.varargs.push_back(lookup_or_add(*I));
395 Expression ValueTable::create_expression(ExtractValueInst* E) {
398 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
399 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
401 e.varargs.push_back(*II);
403 e.type = E->getType();
404 e.opcode = Expression::EXTRACTVALUE;
409 Expression ValueTable::create_expression(InsertValueInst* E) {
412 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
413 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
414 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
416 e.varargs.push_back(*II);
418 e.type = E->getType();
419 e.opcode = Expression::INSERTVALUE;
424 //===----------------------------------------------------------------------===//
425 // ValueTable External Functions
426 //===----------------------------------------------------------------------===//
428 /// add - Insert a value into the table with a specified value number.
429 void ValueTable::add(Value *V, uint32_t num) {
430 valueNumbering.insert(std::make_pair(V, num));
433 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
434 if (AA->doesNotAccessMemory(C)) {
435 Expression exp = create_expression(C);
436 uint32_t& e = expressionNumbering[exp];
437 if (!e) e = nextValueNumber++;
438 valueNumbering[C] = e;
440 } else if (AA->onlyReadsMemory(C)) {
441 Expression exp = create_expression(C);
442 uint32_t& e = expressionNumbering[exp];
444 e = nextValueNumber++;
445 valueNumbering[C] = e;
449 e = nextValueNumber++;
450 valueNumbering[C] = e;
454 MemDepResult local_dep = MD->getDependency(C);
456 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
457 valueNumbering[C] = nextValueNumber;
458 return nextValueNumber++;
461 if (local_dep.isDef()) {
462 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
464 if (local_cdep->getNumOperands() != C->getNumOperands()) {
465 valueNumbering[C] = nextValueNumber;
466 return nextValueNumber++;
469 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
470 uint32_t c_vn = lookup_or_add(C->getOperand(i));
471 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
473 valueNumbering[C] = nextValueNumber;
474 return nextValueNumber++;
478 uint32_t v = lookup_or_add(local_cdep);
479 valueNumbering[C] = v;
484 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
485 MD->getNonLocalCallDependency(CallSite(C));
486 // FIXME: call/call dependencies for readonly calls should return def, not
487 // clobber! Move the checking logic to MemDep!
490 // Check to see if we have a single dominating call instruction that is
492 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
493 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
494 // Ignore non-local dependencies.
495 if (I->second.isNonLocal())
498 // We don't handle non-depedencies. If we already have a call, reject
499 // instruction dependencies.
500 if (I->second.isClobber() || cdep != 0) {
505 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
506 // FIXME: All duplicated with non-local case.
507 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
508 cdep = NonLocalDepCall;
517 valueNumbering[C] = nextValueNumber;
518 return nextValueNumber++;
521 if (cdep->getNumOperands() != C->getNumOperands()) {
522 valueNumbering[C] = nextValueNumber;
523 return nextValueNumber++;
525 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
526 uint32_t c_vn = lookup_or_add(C->getOperand(i));
527 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
529 valueNumbering[C] = nextValueNumber;
530 return nextValueNumber++;
534 uint32_t v = lookup_or_add(cdep);
535 valueNumbering[C] = v;
539 valueNumbering[C] = nextValueNumber;
540 return nextValueNumber++;
544 /// lookup_or_add - Returns the value number for the specified value, assigning
545 /// it a new number if it did not have one before.
546 uint32_t ValueTable::lookup_or_add(Value *V) {
547 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
548 if (VI != valueNumbering.end())
551 if (!isa<Instruction>(V)) {
552 valueNumbering[V] = nextValueNumber;
553 return nextValueNumber++;
556 Instruction* I = cast<Instruction>(V);
558 switch (I->getOpcode()) {
559 case Instruction::Call:
560 return lookup_or_add_call(cast<CallInst>(I));
561 case Instruction::Add:
562 case Instruction::FAdd:
563 case Instruction::Sub:
564 case Instruction::FSub:
565 case Instruction::Mul:
566 case Instruction::FMul:
567 case Instruction::UDiv:
568 case Instruction::SDiv:
569 case Instruction::FDiv:
570 case Instruction::URem:
571 case Instruction::SRem:
572 case Instruction::FRem:
573 case Instruction::Shl:
574 case Instruction::LShr:
575 case Instruction::AShr:
576 case Instruction::And:
577 case Instruction::Or :
578 case Instruction::Xor:
579 exp = create_expression(cast<BinaryOperator>(I));
581 case Instruction::ICmp:
582 case Instruction::FCmp:
583 exp = create_expression(cast<CmpInst>(I));
585 case Instruction::Trunc:
586 case Instruction::ZExt:
587 case Instruction::SExt:
588 case Instruction::FPToUI:
589 case Instruction::FPToSI:
590 case Instruction::UIToFP:
591 case Instruction::SIToFP:
592 case Instruction::FPTrunc:
593 case Instruction::FPExt:
594 case Instruction::PtrToInt:
595 case Instruction::IntToPtr:
596 case Instruction::BitCast:
597 exp = create_expression(cast<CastInst>(I));
599 case Instruction::Select:
600 exp = create_expression(cast<SelectInst>(I));
602 case Instruction::ExtractElement:
603 exp = create_expression(cast<ExtractElementInst>(I));
605 case Instruction::InsertElement:
606 exp = create_expression(cast<InsertElementInst>(I));
608 case Instruction::ShuffleVector:
609 exp = create_expression(cast<ShuffleVectorInst>(I));
611 case Instruction::ExtractValue:
612 exp = create_expression(cast<ExtractValueInst>(I));
614 case Instruction::InsertValue:
615 exp = create_expression(cast<InsertValueInst>(I));
617 case Instruction::GetElementPtr:
618 exp = create_expression(cast<GetElementPtrInst>(I));
621 valueNumbering[V] = nextValueNumber;
622 return nextValueNumber++;
625 uint32_t& e = expressionNumbering[exp];
626 if (!e) e = nextValueNumber++;
627 valueNumbering[V] = e;
631 /// lookup - Returns the value number of the specified value. Fails if
632 /// the value has not yet been numbered.
633 uint32_t ValueTable::lookup(Value *V) const {
634 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
635 assert(VI != valueNumbering.end() && "Value not numbered?");
639 /// clear - Remove all entries from the ValueTable
640 void ValueTable::clear() {
641 valueNumbering.clear();
642 expressionNumbering.clear();
646 /// erase - Remove a value from the value numbering
647 void ValueTable::erase(Value *V) {
648 valueNumbering.erase(V);
651 /// verifyRemoved - Verify that the value is removed from all internal data
653 void ValueTable::verifyRemoved(const Value *V) const {
654 for (DenseMap<Value*, uint32_t>::const_iterator
655 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
656 assert(I->first != V && "Inst still occurs in value numbering map!");
660 //===----------------------------------------------------------------------===//
662 //===----------------------------------------------------------------------===//
665 struct ValueNumberScope {
666 ValueNumberScope* parent;
667 DenseMap<uint32_t, Value*> table;
669 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
675 class GVN : public FunctionPass {
676 bool runOnFunction(Function &F);
678 static char ID; // Pass identification, replacement for typeid
679 explicit GVN(bool nopre = false, bool noloads = false)
680 : FunctionPass(&ID), NoPRE(nopre), NoLoads(noloads), MD(0) { }
685 MemoryDependenceAnalysis *MD;
689 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
691 // This transformation requires dominator postdominator info
692 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
693 AU.addRequired<DominatorTree>();
695 AU.addRequired<MemoryDependenceAnalysis>();
696 AU.addRequired<AliasAnalysis>();
698 AU.addPreserved<DominatorTree>();
699 AU.addPreserved<AliasAnalysis>();
703 // FIXME: eliminate or document these better
704 bool processLoad(LoadInst* L,
705 SmallVectorImpl<Instruction*> &toErase);
706 bool processInstruction(Instruction *I,
707 SmallVectorImpl<Instruction*> &toErase);
708 bool processNonLocalLoad(LoadInst* L,
709 SmallVectorImpl<Instruction*> &toErase);
710 bool processBlock(BasicBlock *BB);
711 void dump(DenseMap<uint32_t, Value*>& d);
712 bool iterateOnFunction(Function &F);
713 Value *CollapsePhi(PHINode* p);
714 bool performPRE(Function& F);
715 Value *lookupNumber(BasicBlock *BB, uint32_t num);
716 void cleanupGlobalSets();
717 void verifyRemoved(const Instruction *I) const;
723 // createGVNPass - The public interface to this file...
724 FunctionPass *llvm::createGVNPass(bool NoPRE, bool NoLoads) {
725 return new GVN(NoPRE, NoLoads);
728 static RegisterPass<GVN> X("gvn",
729 "Global Value Numbering");
731 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
733 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
734 E = d.end(); I != E; ++I) {
735 printf("%d\n", I->first);
741 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
742 if (!isa<PHINode>(inst))
745 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
747 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
748 if (use_phi->getParent() == inst->getParent())
754 Value *GVN::CollapsePhi(PHINode *PN) {
755 Value *ConstVal = PN->hasConstantValue(DT);
756 if (!ConstVal) return 0;
758 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
762 if (DT->dominates(Inst, PN))
763 if (isSafeReplacement(PN, Inst))
768 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
769 /// we're analyzing is fully available in the specified block. As we go, keep
770 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
771 /// map is actually a tri-state map with the following values:
772 /// 0) we know the block *is not* fully available.
773 /// 1) we know the block *is* fully available.
774 /// 2) we do not know whether the block is fully available or not, but we are
775 /// currently speculating that it will be.
776 /// 3) we are speculating for this block and have used that to speculate for
778 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
779 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
780 // Optimistically assume that the block is fully available and check to see
781 // if we already know about this block in one lookup.
782 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
783 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
785 // If the entry already existed for this block, return the precomputed value.
787 // If this is a speculative "available" value, mark it as being used for
788 // speculation of other blocks.
789 if (IV.first->second == 2)
790 IV.first->second = 3;
791 return IV.first->second != 0;
794 // Otherwise, see if it is fully available in all predecessors.
795 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
797 // If this block has no predecessors, it isn't live-in here.
799 goto SpeculationFailure;
801 for (; PI != PE; ++PI)
802 // If the value isn't fully available in one of our predecessors, then it
803 // isn't fully available in this block either. Undo our previous
804 // optimistic assumption and bail out.
805 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
806 goto SpeculationFailure;
810 // SpeculationFailure - If we get here, we found out that this is not, after
811 // all, a fully-available block. We have a problem if we speculated on this and
812 // used the speculation to mark other blocks as available.
814 char &BBVal = FullyAvailableBlocks[BB];
816 // If we didn't speculate on this, just return with it set to false.
822 // If we did speculate on this value, we could have blocks set to 1 that are
823 // incorrect. Walk the (transitive) successors of this block and mark them as
825 SmallVector<BasicBlock*, 32> BBWorklist;
826 BBWorklist.push_back(BB);
828 while (!BBWorklist.empty()) {
829 BasicBlock *Entry = BBWorklist.pop_back_val();
830 // Note that this sets blocks to 0 (unavailable) if they happen to not
831 // already be in FullyAvailableBlocks. This is safe.
832 char &EntryVal = FullyAvailableBlocks[Entry];
833 if (EntryVal == 0) continue; // Already unavailable.
835 // Mark as unavailable.
838 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
839 BBWorklist.push_back(*I);
846 /// CanCoerceMustAliasedValueToLoad - Return true if
847 /// CoerceAvailableValueToLoadType will succeed.
848 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
850 const TargetData &TD) {
851 // If the loaded or stored value is an first class array or struct, don't try
852 // to transform them. We need to be able to bitcast to integer.
853 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
854 isa<StructType>(StoredVal->getType()) ||
855 isa<ArrayType>(StoredVal->getType()))
858 // The store has to be at least as big as the load.
859 if (TD.getTypeSizeInBits(StoredVal->getType()) <
860 TD.getTypeSizeInBits(LoadTy))
867 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
868 /// then a load from a must-aliased pointer of a different type, try to coerce
869 /// the stored value. LoadedTy is the type of the load we want to replace and
870 /// InsertPt is the place to insert new instructions.
872 /// If we can't do it, return null.
873 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
874 const Type *LoadedTy,
875 Instruction *InsertPt,
876 const TargetData &TD) {
877 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
880 const Type *StoredValTy = StoredVal->getType();
882 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
883 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
885 // If the store and reload are the same size, we can always reuse it.
886 if (StoreSize == LoadSize) {
887 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
888 // Pointer to Pointer -> use bitcast.
889 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
892 // Convert source pointers to integers, which can be bitcast.
893 if (isa<PointerType>(StoredValTy)) {
894 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
895 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
898 const Type *TypeToCastTo = LoadedTy;
899 if (isa<PointerType>(TypeToCastTo))
900 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
902 if (StoredValTy != TypeToCastTo)
903 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
905 // Cast to pointer if the load needs a pointer type.
906 if (isa<PointerType>(LoadedTy))
907 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
912 // If the loaded value is smaller than the available value, then we can
913 // extract out a piece from it. If the available value is too small, then we
914 // can't do anything.
915 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
917 // Convert source pointers to integers, which can be manipulated.
918 if (isa<PointerType>(StoredValTy)) {
919 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
920 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
923 // Convert vectors and fp to integer, which can be manipulated.
924 if (!isa<IntegerType>(StoredValTy)) {
925 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
926 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
929 // If this is a big-endian system, we need to shift the value down to the low
930 // bits so that a truncate will work.
931 if (TD.isBigEndian()) {
932 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
933 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
936 // Truncate the integer to the right size now.
937 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
938 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
940 if (LoadedTy == NewIntTy)
943 // If the result is a pointer, inttoptr.
944 if (isa<PointerType>(LoadedTy))
945 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
947 // Otherwise, bitcast.
948 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
951 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
952 /// be expressed as a base pointer plus a constant offset. Return the base and
953 /// offset to the caller.
954 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
955 const TargetData &TD) {
956 Operator *PtrOp = dyn_cast<Operator>(Ptr);
957 if (PtrOp == 0) return Ptr;
959 // Just look through bitcasts.
960 if (PtrOp->getOpcode() == Instruction::BitCast)
961 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
963 // If this is a GEP with constant indices, we can look through it.
964 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
965 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
967 gep_type_iterator GTI = gep_type_begin(GEP);
968 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
970 ConstantInt *OpC = cast<ConstantInt>(*I);
971 if (OpC->isZero()) continue;
973 // Handle a struct and array indices which add their offset to the pointer.
974 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
975 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
977 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
978 Offset += OpC->getSExtValue()*Size;
982 // Re-sign extend from the pointer size if needed to get overflow edge cases
984 unsigned PtrSize = TD.getPointerSizeInBits();
986 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
988 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
992 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
993 /// memdep query of a load that ends up being a clobbering memory write (store,
994 /// memset, memcpy, memmove). This means that the write *may* provide bits used
995 /// by the load but we can't be sure because the pointers don't mustalias.
997 /// Check this case to see if there is anything more we can do before we give
998 /// up. This returns -1 if we have to give up, or a byte number in the stored
999 /// value of the piece that feeds the load.
1000 static int AnalyzeLoadFromClobberingWrite(LoadInst *L, Value *WritePtr,
1001 uint64_t WriteSizeInBits,
1002 const TargetData &TD) {
1003 // If the loaded or stored value is an first class array or struct, don't try
1004 // to transform them. We need to be able to bitcast to integer.
1005 if (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()))
1008 int64_t StoreOffset = 0, LoadOffset = 0;
1009 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1011 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
1012 if (StoreBase != LoadBase)
1015 // If the load and store are to the exact same address, they should have been
1016 // a must alias. AA must have gotten confused.
1017 // FIXME: Study to see if/when this happens.
1018 if (LoadOffset == StoreOffset) {
1020 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1021 << "Base = " << *StoreBase << "\n"
1022 << "Store Ptr = " << *WritePtr << "\n"
1023 << "Store Offs = " << StoreOffset << "\n"
1024 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1025 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1026 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1032 // If the load and store don't overlap at all, the store doesn't provide
1033 // anything to the load. In this case, they really don't alias at all, AA
1034 // must have gotten confused.
1035 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1036 // remove this check, as it is duplicated with what we have below.
1037 uint64_t LoadSize = TD.getTypeSizeInBits(L->getType());
1039 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1041 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1045 bool isAAFailure = false;
1046 if (StoreOffset < LoadOffset) {
1047 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1049 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1053 errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1054 << "Base = " << *StoreBase << "\n"
1055 << "Store Ptr = " << *WritePtr << "\n"
1056 << "Store Offs = " << StoreOffset << "\n"
1057 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1058 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1059 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1065 // If the Load isn't completely contained within the stored bits, we don't
1066 // have all the bits to feed it. We could do something crazy in the future
1067 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1069 if (StoreOffset > LoadOffset ||
1070 StoreOffset+StoreSize < LoadOffset+LoadSize)
1073 // Okay, we can do this transformation. Return the number of bytes into the
1074 // store that the load is.
1075 return LoadOffset-StoreOffset;
1078 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1079 /// memdep query of a load that ends up being a clobbering store.
1080 static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI,
1081 const TargetData &TD) {
1082 // Cannot handle reading from store of first-class aggregate yet.
1083 if (isa<StructType>(DepSI->getOperand(0)->getType()) ||
1084 isa<ArrayType>(DepSI->getOperand(0)->getType()))
1087 Value *StorePtr = DepSI->getPointerOperand();
1088 uint64_t StoreSize = TD.getTypeSizeInBits(StorePtr->getType());
1089 return AnalyzeLoadFromClobberingWrite(L, StorePtr, StoreSize, TD);
1092 static int AnalyzeLoadFromClobberingMemInst(LoadInst *L, MemIntrinsic *MI,
1093 const TargetData &TD) {
1094 // If the mem operation is a non-constant size, we can't handle it.
1095 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1096 if (SizeCst == 0) return -1;
1097 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1099 // If this is memset, we just need to see if the offset is valid in the size
1101 if (MI->getIntrinsicID() == Intrinsic::memset)
1102 return AnalyzeLoadFromClobberingWrite(L, MI->getDest(), MemSizeInBits, TD);
1104 // If we have a memcpy/memmove, the only case we can handle is if this is a
1105 // copy from constant memory. In that case, we can read directly from the
1107 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1109 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1110 if (Src == 0) return -1;
1112 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1113 if (GV == 0 || !GV->isConstant()) return -1;
1115 // See if the access is within the bounds of the transfer.
1117 AnalyzeLoadFromClobberingWrite(L, MI->getDest(), MemSizeInBits, TD);
1121 // Otherwise, see if we can constant fold a load from the constant with the
1122 // offset applied as appropriate.
1123 Src = ConstantExpr::getBitCast(Src,
1124 llvm::Type::getInt8PtrTy(Src->getContext()));
1125 Constant *OffsetCst =
1126 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1127 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1128 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(L->getType()));
1129 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1135 /// GetStoreValueForLoad - This function is called when we have a
1136 /// memdep query of a load that ends up being a clobbering store. This means
1137 /// that the store *may* provide bits used by the load but we can't be sure
1138 /// because the pointers don't mustalias. Check this case to see if there is
1139 /// anything more we can do before we give up.
1140 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1142 Instruction *InsertPt, const TargetData &TD){
1143 LLVMContext &Ctx = SrcVal->getType()->getContext();
1145 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1146 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1149 // Compute which bits of the stored value are being used by the load. Convert
1150 // to an integer type to start with.
1151 if (isa<PointerType>(SrcVal->getType()))
1152 SrcVal = new PtrToIntInst(SrcVal, TD.getIntPtrType(Ctx), "tmp", InsertPt);
1153 if (!isa<IntegerType>(SrcVal->getType()))
1154 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1157 // Shift the bits to the least significant depending on endianness.
1159 if (TD.isLittleEndian())
1160 ShiftAmt = Offset*8;
1162 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1165 SrcVal = BinaryOperator::CreateLShr(SrcVal,
1166 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
1168 if (LoadSize != StoreSize)
1169 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1172 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1175 /// GetMemInstValueForLoad - This function is called when we have a
1176 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1177 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1178 const Type *LoadTy, Instruction *InsertPt,
1179 const TargetData &TD){
1180 LLVMContext &Ctx = LoadTy->getContext();
1181 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1183 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1185 // We know that this method is only called when the mem transfer fully
1186 // provides the bits for the load.
1187 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1188 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1189 // independently of what the offset is.
1190 Value *Val = MSI->getValue();
1192 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1194 Value *OneElt = Val;
1196 // Splat the value out to the right number of bits.
1197 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1198 // If we can double the number of bytes set, do it.
1199 if (NumBytesSet*2 <= LoadSize) {
1200 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1201 Val = Builder.CreateOr(Val, ShVal);
1206 // Otherwise insert one byte at a time.
1207 Value *ShVal = Builder.CreateShl(Val, 1*8);
1208 Val = Builder.CreateOr(OneElt, ShVal);
1212 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1215 // Otherwise, this is a memcpy/memmove from a constant global.
1216 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1217 Constant *Src = cast<Constant>(MTI->getSource());
1219 // Otherwise, see if we can constant fold a load from the constant with the
1220 // offset applied as appropriate.
1221 Src = ConstantExpr::getBitCast(Src,
1222 llvm::Type::getInt8PtrTy(Src->getContext()));
1223 Constant *OffsetCst =
1224 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1225 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1226 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1227 return ConstantFoldLoadFromConstPtr(Src, &TD);
1232 struct AvailableValueInBlock {
1233 /// BB - The basic block in question.
1236 SimpleVal, // A simple offsetted value that is accessed.
1237 MemIntrin // A memory intrinsic which is loaded from.
1240 /// V - The value that is live out of the block.
1241 PointerIntPair<Value *, 1, ValType> Val;
1243 /// Offset - The byte offset in Val that is interesting for the load query.
1246 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1247 unsigned Offset = 0) {
1248 AvailableValueInBlock Res;
1250 Res.Val.setPointer(V);
1251 Res.Val.setInt(SimpleVal);
1252 Res.Offset = Offset;
1256 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1257 unsigned Offset = 0) {
1258 AvailableValueInBlock Res;
1260 Res.Val.setPointer(MI);
1261 Res.Val.setInt(MemIntrin);
1262 Res.Offset = Offset;
1266 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1267 Value *getSimpleValue() const {
1268 assert(isSimpleValue() && "Wrong accessor");
1269 return Val.getPointer();
1272 MemIntrinsic *getMemIntrinValue() const {
1273 assert(!isSimpleValue() && "Wrong accessor");
1274 return cast<MemIntrinsic>(Val.getPointer());
1278 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1279 /// construct SSA form, allowing us to eliminate LI. This returns the value
1280 /// that should be used at LI's definition site.
1281 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1282 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1283 const TargetData *TD,
1284 AliasAnalysis *AA) {
1285 SmallVector<PHINode*, 8> NewPHIs;
1286 SSAUpdater SSAUpdate(&NewPHIs);
1287 SSAUpdate.Initialize(LI);
1289 const Type *LoadTy = LI->getType();
1291 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1292 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1293 BasicBlock *BB = AV.BB;
1295 if (SSAUpdate.HasValueForBlock(BB))
1298 unsigned Offset = AV.Offset;
1300 Value *AvailableVal;
1301 if (AV.isSimpleValue()) {
1302 AvailableVal = AV.getSimpleValue();
1303 if (AvailableVal->getType() != LoadTy) {
1304 assert(TD && "Need target data to handle type mismatch case");
1305 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy,
1306 BB->getTerminator(), *TD);
1308 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1309 << *AV.getSimpleValue() << '\n'
1310 << *AvailableVal << '\n' << "\n\n\n");
1313 AvailableVal = GetMemInstValueForLoad(AV.getMemIntrinValue(), Offset,
1314 LoadTy, BB->getTerminator(), *TD);
1315 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1316 << " " << *AV.getMemIntrinValue() << '\n'
1317 << *AvailableVal << '\n' << "\n\n\n");
1319 SSAUpdate.AddAvailableValue(BB, AvailableVal);
1322 // Perform PHI construction.
1323 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1325 // If new PHI nodes were created, notify alias analysis.
1326 if (isa<PointerType>(V->getType()))
1327 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1328 AA->copyValue(LI, NewPHIs[i]);
1333 static bool isLifetimeStart(Instruction *Inst) {
1334 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1335 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1339 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1340 /// non-local by performing PHI construction.
1341 bool GVN::processNonLocalLoad(LoadInst *LI,
1342 SmallVectorImpl<Instruction*> &toErase) {
1343 // Find the non-local dependencies of the load.
1344 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
1345 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1347 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
1348 // << Deps.size() << *LI << '\n');
1350 // If we had to process more than one hundred blocks to find the
1351 // dependencies, this load isn't worth worrying about. Optimizing
1352 // it will be too expensive.
1353 if (Deps.size() > 100)
1356 // If we had a phi translation failure, we'll have a single entry which is a
1357 // clobber in the current block. Reject this early.
1358 if (Deps.size() == 1 && Deps[0].second.isClobber()) {
1360 errs() << "GVN: non-local load ";
1361 WriteAsOperand(errs(), LI);
1362 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n';
1367 // Filter out useless results (non-locals, etc). Keep track of the blocks
1368 // where we have a value available in repl, also keep track of whether we see
1369 // dependencies that produce an unknown value for the load (such as a call
1370 // that could potentially clobber the load).
1371 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1372 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1374 const TargetData *TD = 0;
1376 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1377 BasicBlock *DepBB = Deps[i].first;
1378 MemDepResult DepInfo = Deps[i].second;
1380 if (DepInfo.isClobber()) {
1381 // If the dependence is to a store that writes to a superset of the bits
1382 // read by the load, we can extract the bits we need for the load from the
1384 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1386 TD = getAnalysisIfAvailable<TargetData>();
1388 int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD);
1390 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1391 DepSI->getOperand(0),
1398 // If the clobbering value is a memset/memcpy/memmove, see if we can
1399 // forward a value on from it.
1400 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1402 TD = getAnalysisIfAvailable<TargetData>();
1404 int Offset = AnalyzeLoadFromClobberingMemInst(LI, DepMI, *TD);
1406 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1413 UnavailableBlocks.push_back(DepBB);
1417 Instruction *DepInst = DepInfo.getInst();
1419 // Loading the allocation -> undef.
1420 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1421 // Loading immediately after lifetime begin -> undef.
1422 isLifetimeStart(DepInst)) {
1423 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1424 UndefValue::get(LI->getType())));
1428 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1429 // Reject loads and stores that are to the same address but are of
1430 // different types if we have to.
1431 if (S->getOperand(0)->getType() != LI->getType()) {
1433 TD = getAnalysisIfAvailable<TargetData>();
1435 // If the stored value is larger or equal to the loaded value, we can
1437 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1438 LI->getType(), *TD)) {
1439 UnavailableBlocks.push_back(DepBB);
1444 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1449 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1450 // If the types mismatch and we can't handle it, reject reuse of the load.
1451 if (LD->getType() != LI->getType()) {
1453 TD = getAnalysisIfAvailable<TargetData>();
1455 // If the stored value is larger or equal to the loaded value, we can
1457 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1458 UnavailableBlocks.push_back(DepBB);
1462 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1466 UnavailableBlocks.push_back(DepBB);
1470 // If we have no predecessors that produce a known value for this load, exit
1472 if (ValuesPerBlock.empty()) return false;
1474 // If all of the instructions we depend on produce a known value for this
1475 // load, then it is fully redundant and we can use PHI insertion to compute
1476 // its value. Insert PHIs and remove the fully redundant value now.
1477 if (UnavailableBlocks.empty()) {
1478 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1480 // Perform PHI construction.
1481 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1482 VN.getAliasAnalysis());
1483 LI->replaceAllUsesWith(V);
1485 if (isa<PHINode>(V))
1487 if (isa<PointerType>(V->getType()))
1488 MD->invalidateCachedPointerInfo(V);
1489 toErase.push_back(LI);
1494 if (!EnablePRE || !EnableLoadPRE)
1497 // Okay, we have *some* definitions of the value. This means that the value
1498 // is available in some of our (transitive) predecessors. Lets think about
1499 // doing PRE of this load. This will involve inserting a new load into the
1500 // predecessor when it's not available. We could do this in general, but
1501 // prefer to not increase code size. As such, we only do this when we know
1502 // that we only have to insert *one* load (which means we're basically moving
1503 // the load, not inserting a new one).
1505 SmallPtrSet<BasicBlock *, 4> Blockers;
1506 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1507 Blockers.insert(UnavailableBlocks[i]);
1509 // Lets find first basic block with more than one predecessor. Walk backwards
1510 // through predecessors if needed.
1511 BasicBlock *LoadBB = LI->getParent();
1512 BasicBlock *TmpBB = LoadBB;
1514 bool isSinglePred = false;
1515 bool allSingleSucc = true;
1516 while (TmpBB->getSinglePredecessor()) {
1517 isSinglePred = true;
1518 TmpBB = TmpBB->getSinglePredecessor();
1519 if (!TmpBB) // If haven't found any, bail now.
1521 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1523 if (Blockers.count(TmpBB))
1525 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1526 allSingleSucc = false;
1532 // If we have a repl set with LI itself in it, this means we have a loop where
1533 // at least one of the values is LI. Since this means that we won't be able
1534 // to eliminate LI even if we insert uses in the other predecessors, we will
1535 // end up increasing code size. Reject this by scanning for LI.
1536 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1537 if (ValuesPerBlock[i].isSimpleValue() &&
1538 ValuesPerBlock[i].getSimpleValue() == LI)
1541 // FIXME: It is extremely unclear what this loop is doing, other than
1542 // artificially restricting loadpre.
1545 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1546 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1547 if (AV.isSimpleValue())
1548 // "Hot" Instruction is in some loop (because it dominates its dep.
1550 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1551 if (DT->dominates(LI, I)) {
1557 // We are interested only in "hot" instructions. We don't want to do any
1558 // mis-optimizations here.
1563 // Okay, we have some hope :). Check to see if the loaded value is fully
1564 // available in all but one predecessor.
1565 // FIXME: If we could restructure the CFG, we could make a common pred with
1566 // all the preds that don't have an available LI and insert a new load into
1568 BasicBlock *UnavailablePred = 0;
1570 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1571 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1572 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1573 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1574 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1576 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1578 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1581 // If this load is not available in multiple predecessors, reject it.
1582 if (UnavailablePred && UnavailablePred != *PI)
1584 UnavailablePred = *PI;
1587 assert(UnavailablePred != 0 &&
1588 "Fully available value should be eliminated above!");
1590 // We don't currently handle critical edges :(
1591 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1592 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1593 << UnavailablePred->getName() << "': " << *LI << '\n');
1597 // Do PHI translation to get its value in the predecessor if necessary. The
1598 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1600 // FIXME: This may insert a computation, but we don't tell scalar GVN
1601 // optimization stuff about it. How do we do this?
1602 SmallVector<Instruction*, 8> NewInsts;
1605 // If all preds have a single successor, then we know it is safe to insert the
1606 // load on the pred (?!?), so we can insert code to materialize the pointer if
1607 // it is not available.
1608 if (allSingleSucc) {
1609 LoadPtr = MD->InsertPHITranslatedPointer(LI->getOperand(0), LoadBB,
1610 UnavailablePred, TD, *DT,NewInsts);
1612 LoadPtr = MD->GetAvailablePHITranslatedValue(LI->getOperand(0), LoadBB,
1613 UnavailablePred, TD, *DT);
1616 // Assign value numbers to these new instructions.
1617 for (SmallVector<Instruction*, 8>::iterator NI = NewInsts.begin(),
1618 NE = NewInsts.end(); NI != NE; ++NI) {
1619 // FIXME: We really _ought_ to insert these value numbers into their
1620 // parent's availability map. However, in doing so, we risk getting into
1621 // ordering issues. If a block hasn't been processed yet, we would be
1622 // marking a value as AVAIL-IN, which isn't what we intend.
1623 VN.lookup_or_add(*NI);
1626 // If we couldn't find or insert a computation of this phi translated value,
1629 DEBUG(errs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1630 << *LI->getOperand(0) << "\n");
1634 // Make sure it is valid to move this load here. We have to watch out for:
1635 // @1 = getelementptr (i8* p, ...
1636 // test p and branch if == 0
1638 // It is valid to have the getelementptr before the test, even if p can be 0,
1639 // as getelementptr only does address arithmetic.
1640 // If we are not pushing the value through any multiple-successor blocks
1641 // we do not have this case. Otherwise, check that the load is safe to
1642 // put anywhere; this can be improved, but should be conservatively safe.
1643 if (!allSingleSucc &&
1644 // FIXME: REEVALUTE THIS.
1645 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator())) {
1646 assert(NewInsts.empty() && "Should not have inserted instructions");
1650 // Okay, we can eliminate this load by inserting a reload in the predecessor
1651 // and using PHI construction to get the value in the other predecessors, do
1653 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1654 DEBUG(if (!NewInsts.empty())
1655 errs() << "INSERTED " << NewInsts.size() << " INSTS: "
1656 << *NewInsts.back() << '\n');
1658 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1660 UnavailablePred->getTerminator());
1662 // Add the newly created load.
1663 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));
1665 // Perform PHI construction.
1666 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1667 VN.getAliasAnalysis());
1668 LI->replaceAllUsesWith(V);
1669 if (isa<PHINode>(V))
1671 if (isa<PointerType>(V->getType()))
1672 MD->invalidateCachedPointerInfo(V);
1673 toErase.push_back(LI);
1678 /// processLoad - Attempt to eliminate a load, first by eliminating it
1679 /// locally, and then attempting non-local elimination if that fails.
1680 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1684 if (L->isVolatile())
1687 // ... to a pointer that has been loaded from before...
1688 MemDepResult Dep = MD->getDependency(L);
1690 // If the value isn't available, don't do anything!
1691 if (Dep.isClobber()) {
1692 // Check to see if we have something like this:
1693 // store i32 123, i32* %P
1694 // %A = bitcast i32* %P to i8*
1695 // %B = gep i8* %A, i32 1
1698 // We could do that by recognizing if the clobber instructions are obviously
1699 // a common base + constant offset, and if the previous store (or memset)
1700 // completely covers this load. This sort of thing can happen in bitfield
1702 Value *AvailVal = 0;
1703 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1704 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1705 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD);
1707 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1708 L->getType(), L, *TD);
1711 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1712 // a value on from it.
1713 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1714 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1715 int Offset = AnalyzeLoadFromClobberingMemInst(L, DepMI, *TD);
1717 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1722 DEBUG(errs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1723 << *AvailVal << '\n' << *L << "\n\n\n");
1725 // Replace the load!
1726 L->replaceAllUsesWith(AvailVal);
1727 if (isa<PointerType>(AvailVal->getType()))
1728 MD->invalidateCachedPointerInfo(AvailVal);
1729 toErase.push_back(L);
1735 // fast print dep, using operator<< on instruction would be too slow
1736 errs() << "GVN: load ";
1737 WriteAsOperand(errs(), L);
1738 Instruction *I = Dep.getInst();
1739 errs() << " is clobbered by " << *I << '\n';
1744 // If it is defined in another block, try harder.
1745 if (Dep.isNonLocal())
1746 return processNonLocalLoad(L, toErase);
1748 Instruction *DepInst = Dep.getInst();
1749 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1750 Value *StoredVal = DepSI->getOperand(0);
1752 // The store and load are to a must-aliased pointer, but they may not
1753 // actually have the same type. See if we know how to reuse the stored
1754 // value (depending on its type).
1755 const TargetData *TD = 0;
1756 if (StoredVal->getType() != L->getType()) {
1757 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1758 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1763 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1764 << '\n' << *L << "\n\n\n");
1771 L->replaceAllUsesWith(StoredVal);
1772 if (isa<PointerType>(StoredVal->getType()))
1773 MD->invalidateCachedPointerInfo(StoredVal);
1774 toErase.push_back(L);
1779 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1780 Value *AvailableVal = DepLI;
1782 // The loads are of a must-aliased pointer, but they may not actually have
1783 // the same type. See if we know how to reuse the previously loaded value
1784 // (depending on its type).
1785 const TargetData *TD = 0;
1786 if (DepLI->getType() != L->getType()) {
1787 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1788 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1789 if (AvailableVal == 0)
1792 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1793 << "\n" << *L << "\n\n\n");
1800 L->replaceAllUsesWith(AvailableVal);
1801 if (isa<PointerType>(DepLI->getType()))
1802 MD->invalidateCachedPointerInfo(DepLI);
1803 toErase.push_back(L);
1808 // If this load really doesn't depend on anything, then we must be loading an
1809 // undef value. This can happen when loading for a fresh allocation with no
1810 // intervening stores, for example.
1811 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1812 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1813 toErase.push_back(L);
1818 // If this load occurs either right after a lifetime begin,
1819 // then the loaded value is undefined.
1820 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1821 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1822 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1823 toErase.push_back(L);
1832 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1833 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1834 if (I == localAvail.end())
1837 ValueNumberScope *Locals = I->second;
1839 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1840 if (I != Locals->table.end())
1842 Locals = Locals->parent;
1849 /// processInstruction - When calculating availability, handle an instruction
1850 /// by inserting it into the appropriate sets
1851 bool GVN::processInstruction(Instruction *I,
1852 SmallVectorImpl<Instruction*> &toErase) {
1853 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1854 bool Changed = processLoad(LI, toErase);
1857 unsigned Num = VN.lookup_or_add(LI);
1858 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1864 uint32_t NextNum = VN.getNextUnusedValueNumber();
1865 unsigned Num = VN.lookup_or_add(I);
1867 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1868 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1870 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1873 Value *BranchCond = BI->getCondition();
1874 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1876 BasicBlock *TrueSucc = BI->getSuccessor(0);
1877 BasicBlock *FalseSucc = BI->getSuccessor(1);
1879 if (TrueSucc->getSinglePredecessor())
1880 localAvail[TrueSucc]->table[CondVN] =
1881 ConstantInt::getTrue(TrueSucc->getContext());
1882 if (FalseSucc->getSinglePredecessor())
1883 localAvail[FalseSucc]->table[CondVN] =
1884 ConstantInt::getFalse(TrueSucc->getContext());
1888 // Allocations are always uniquely numbered, so we can save time and memory
1889 // by fast failing them.
1890 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1891 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1895 // Collapse PHI nodes
1896 if (PHINode* p = dyn_cast<PHINode>(I)) {
1897 Value *constVal = CollapsePhi(p);
1900 p->replaceAllUsesWith(constVal);
1901 if (MD && isa<PointerType>(constVal->getType()))
1902 MD->invalidateCachedPointerInfo(constVal);
1905 toErase.push_back(p);
1907 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1910 // If the number we were assigned was a brand new VN, then we don't
1911 // need to do a lookup to see if the number already exists
1912 // somewhere in the domtree: it can't!
1913 } else if (Num == NextNum) {
1914 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1916 // Perform fast-path value-number based elimination of values inherited from
1918 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1921 I->replaceAllUsesWith(repl);
1922 if (MD && isa<PointerType>(repl->getType()))
1923 MD->invalidateCachedPointerInfo(repl);
1924 toErase.push_back(I);
1928 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1934 /// runOnFunction - This is the main transformation entry point for a function.
1935 bool GVN::runOnFunction(Function& F) {
1937 MD = &getAnalysis<MemoryDependenceAnalysis>();
1938 DT = &getAnalysis<DominatorTree>();
1939 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1943 bool Changed = false;
1944 bool ShouldContinue = true;
1946 // Merge unconditional branches, allowing PRE to catch more
1947 // optimization opportunities.
1948 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1949 BasicBlock *BB = FI;
1951 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1952 if (removedBlock) NumGVNBlocks++;
1954 Changed |= removedBlock;
1957 unsigned Iteration = 0;
1959 while (ShouldContinue) {
1960 DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
1961 ShouldContinue = iterateOnFunction(F);
1962 Changed |= ShouldContinue;
1967 bool PREChanged = true;
1968 while (PREChanged) {
1969 PREChanged = performPRE(F);
1970 Changed |= PREChanged;
1973 // FIXME: Should perform GVN again after PRE does something. PRE can move
1974 // computations into blocks where they become fully redundant. Note that
1975 // we can't do this until PRE's critical edge splitting updates memdep.
1976 // Actually, when this happens, we should just fully integrate PRE into GVN.
1978 cleanupGlobalSets();
1984 bool GVN::processBlock(BasicBlock *BB) {
1985 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1986 // incrementing BI before processing an instruction).
1987 SmallVector<Instruction*, 8> toErase;
1988 bool ChangedFunction = false;
1990 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1992 ChangedFunction |= processInstruction(BI, toErase);
1993 if (toErase.empty()) {
1998 // If we need some instructions deleted, do it now.
1999 NumGVNInstr += toErase.size();
2001 // Avoid iterator invalidation.
2002 bool AtStart = BI == BB->begin();
2006 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2007 E = toErase.end(); I != E; ++I) {
2008 DEBUG(errs() << "GVN removed: " << **I << '\n');
2009 if (MD) MD->removeInstruction(*I);
2010 (*I)->eraseFromParent();
2011 DEBUG(verifyRemoved(*I));
2021 return ChangedFunction;
2024 /// performPRE - Perform a purely local form of PRE that looks for diamond
2025 /// control flow patterns and attempts to perform simple PRE at the join point.
2026 bool GVN::performPRE(Function &F) {
2027 bool Changed = false;
2028 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
2029 DenseMap<BasicBlock*, Value*> predMap;
2030 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2031 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2032 BasicBlock *CurrentBlock = *DI;
2034 // Nothing to PRE in the entry block.
2035 if (CurrentBlock == &F.getEntryBlock()) continue;
2037 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2038 BE = CurrentBlock->end(); BI != BE; ) {
2039 Instruction *CurInst = BI++;
2041 if (isa<AllocaInst>(CurInst) ||
2042 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2043 CurInst->getType()->isVoidTy() ||
2044 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2045 isa<DbgInfoIntrinsic>(CurInst))
2048 uint32_t ValNo = VN.lookup(CurInst);
2050 // Look for the predecessors for PRE opportunities. We're
2051 // only trying to solve the basic diamond case, where
2052 // a value is computed in the successor and one predecessor,
2053 // but not the other. We also explicitly disallow cases
2054 // where the successor is its own predecessor, because they're
2055 // more complicated to get right.
2056 unsigned NumWith = 0;
2057 unsigned NumWithout = 0;
2058 BasicBlock *PREPred = 0;
2061 for (pred_iterator PI = pred_begin(CurrentBlock),
2062 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2063 // We're not interested in PRE where the block is its
2064 // own predecessor, on in blocks with predecessors
2065 // that are not reachable.
2066 if (*PI == CurrentBlock) {
2069 } else if (!localAvail.count(*PI)) {
2074 DenseMap<uint32_t, Value*>::iterator predV =
2075 localAvail[*PI]->table.find(ValNo);
2076 if (predV == localAvail[*PI]->table.end()) {
2079 } else if (predV->second == CurInst) {
2082 predMap[*PI] = predV->second;
2087 // Don't do PRE when it might increase code size, i.e. when
2088 // we would need to insert instructions in more than one pred.
2089 if (NumWithout != 1 || NumWith == 0)
2092 // Don't do PRE across indirect branch.
2093 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2096 // We can't do PRE safely on a critical edge, so instead we schedule
2097 // the edge to be split and perform the PRE the next time we iterate
2099 unsigned SuccNum = 0;
2100 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
2102 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
2107 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2108 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2112 // Instantiate the expression the in predecessor that lacked it.
2113 // Because we are going top-down through the block, all value numbers
2114 // will be available in the predecessor by the time we need them. Any
2115 // that weren't original present will have been instantiated earlier
2117 Instruction *PREInstr = CurInst->clone();
2118 bool success = true;
2119 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2120 Value *Op = PREInstr->getOperand(i);
2121 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2124 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2125 PREInstr->setOperand(i, V);
2132 // Fail out if we encounter an operand that is not available in
2133 // the PRE predecessor. This is typically because of loads which
2134 // are not value numbered precisely.
2137 DEBUG(verifyRemoved(PREInstr));
2141 PREInstr->insertBefore(PREPred->getTerminator());
2142 PREInstr->setName(CurInst->getName() + ".pre");
2143 predMap[PREPred] = PREInstr;
2144 VN.add(PREInstr, ValNo);
2147 // Update the availability map to include the new instruction.
2148 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
2150 // Create a PHI to make the value available in this block.
2151 PHINode* Phi = PHINode::Create(CurInst->getType(),
2152 CurInst->getName() + ".pre-phi",
2153 CurrentBlock->begin());
2154 for (pred_iterator PI = pred_begin(CurrentBlock),
2155 PE = pred_end(CurrentBlock); PI != PE; ++PI)
2156 Phi->addIncoming(predMap[*PI], *PI);
2159 localAvail[CurrentBlock]->table[ValNo] = Phi;
2161 CurInst->replaceAllUsesWith(Phi);
2162 if (MD && isa<PointerType>(Phi->getType()))
2163 MD->invalidateCachedPointerInfo(Phi);
2166 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
2167 if (MD) MD->removeInstruction(CurInst);
2168 CurInst->eraseFromParent();
2169 DEBUG(verifyRemoved(CurInst));
2174 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
2175 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
2176 SplitCriticalEdge(I->first, I->second, this);
2178 return Changed || toSplit.size();
2181 /// iterateOnFunction - Executes one iteration of GVN
2182 bool GVN::iterateOnFunction(Function &F) {
2183 cleanupGlobalSets();
2185 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2186 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
2188 localAvail[DI->getBlock()] =
2189 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
2191 localAvail[DI->getBlock()] = new ValueNumberScope(0);
2194 // Top-down walk of the dominator tree
2195 bool Changed = false;
2197 // Needed for value numbering with phi construction to work.
2198 ReversePostOrderTraversal<Function*> RPOT(&F);
2199 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2200 RE = RPOT.end(); RI != RE; ++RI)
2201 Changed |= processBlock(*RI);
2203 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2204 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2205 Changed |= processBlock(DI->getBlock());
2211 void GVN::cleanupGlobalSets() {
2214 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2215 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2220 /// verifyRemoved - Verify that the specified instruction does not occur in our
2221 /// internal data structures.
2222 void GVN::verifyRemoved(const Instruction *Inst) const {
2223 VN.verifyRemoved(Inst);
2225 // Walk through the value number scope to make sure the instruction isn't
2226 // ferreted away in it.
2227 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
2228 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2229 const ValueNumberScope *VNS = I->second;
2232 for (DenseMap<uint32_t, Value*>::const_iterator
2233 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2234 assert(II->second != Inst && "Inst still in value numbering scope!");