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/Dominators.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/MemoryBuiltins.h"
37 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/GetElementPtrTypeIterator.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/SSAUpdater.h"
51 STATISTIC(NumGVNInstr, "Number of instructions deleted");
52 STATISTIC(NumGVNLoad, "Number of loads deleted");
53 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
54 STATISTIC(NumGVNBlocks, "Number of blocks merged");
55 STATISTIC(NumPRELoad, "Number of loads PRE'd");
57 static cl::opt<bool> EnablePRE("enable-pre",
58 cl::init(true), cl::Hidden);
59 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
61 //===----------------------------------------------------------------------===//
63 //===----------------------------------------------------------------------===//
65 /// This class holds the mapping between values and value numbers. It is used
66 /// as an efficient mechanism to determine the expression-wise equivalence of
70 enum ExpressionOpcode { ADD, FADD, SUB, FSUB, MUL, FMUL,
71 UDIV, SDIV, FDIV, UREM, SREM,
72 FREM, SHL, LSHR, ASHR, AND, OR, XOR, ICMPEQ,
73 ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
74 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
75 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
76 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
77 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
78 SHUFFLE, SELECT, TRUNC, ZEXT, SEXT, FPTOUI,
79 FPTOSI, UITOFP, SITOFP, FPTRUNC, FPEXT,
80 PTRTOINT, INTTOPTR, BITCAST, GEP, CALL, CONSTANT,
81 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
83 ExpressionOpcode opcode;
85 SmallVector<uint32_t, 4> varargs;
89 Expression(ExpressionOpcode o) : opcode(o) { }
91 bool operator==(const Expression &other) const {
92 if (opcode != other.opcode)
94 else if (opcode == EMPTY || opcode == TOMBSTONE)
96 else if (type != other.type)
98 else if (function != other.function)
101 if (varargs.size() != other.varargs.size())
104 for (size_t i = 0; i < varargs.size(); ++i)
105 if (varargs[i] != other.varargs[i])
112 bool operator!=(const Expression &other) const {
113 return !(*this == other);
119 DenseMap<Value*, uint32_t> valueNumbering;
120 DenseMap<Expression, uint32_t> expressionNumbering;
122 MemoryDependenceAnalysis* MD;
125 uint32_t nextValueNumber;
127 Expression::ExpressionOpcode getOpcode(BinaryOperator* BO);
128 Expression::ExpressionOpcode getOpcode(CmpInst* C);
129 Expression::ExpressionOpcode getOpcode(CastInst* C);
130 Expression create_expression(BinaryOperator* BO);
131 Expression create_expression(CmpInst* C);
132 Expression create_expression(ShuffleVectorInst* V);
133 Expression create_expression(ExtractElementInst* C);
134 Expression create_expression(InsertElementInst* V);
135 Expression create_expression(SelectInst* V);
136 Expression create_expression(CastInst* C);
137 Expression create_expression(GetElementPtrInst* G);
138 Expression create_expression(CallInst* C);
139 Expression create_expression(Constant* C);
140 Expression create_expression(ExtractValueInst* C);
141 Expression create_expression(InsertValueInst* C);
143 uint32_t lookup_or_add_call(CallInst* C);
145 ValueTable() : nextValueNumber(1) { }
146 uint32_t lookup_or_add(Value *V);
147 uint32_t lookup(Value *V) const;
148 void add(Value *V, uint32_t num);
150 void erase(Value *v);
152 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
153 AliasAnalysis *getAliasAnalysis() const { return AA; }
154 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
155 void setDomTree(DominatorTree* D) { DT = D; }
156 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
157 void verifyRemoved(const Value *) const;
162 template <> struct DenseMapInfo<Expression> {
163 static inline Expression getEmptyKey() {
164 return Expression(Expression::EMPTY);
167 static inline Expression getTombstoneKey() {
168 return Expression(Expression::TOMBSTONE);
171 static unsigned getHashValue(const Expression e) {
172 unsigned hash = e.opcode;
174 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
175 (unsigned)((uintptr_t)e.type >> 9));
177 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
178 E = e.varargs.end(); I != E; ++I)
179 hash = *I + hash * 37;
181 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
182 (unsigned)((uintptr_t)e.function >> 9)) +
187 static bool isEqual(const Expression &LHS, const Expression &RHS) {
190 static bool isPod() { return true; }
194 //===----------------------------------------------------------------------===//
195 // ValueTable Internal Functions
196 //===----------------------------------------------------------------------===//
197 Expression::ExpressionOpcode ValueTable::getOpcode(BinaryOperator* BO) {
198 switch(BO->getOpcode()) {
199 default: // THIS SHOULD NEVER HAPPEN
200 llvm_unreachable("Binary operator with unknown opcode?");
201 case Instruction::Add: return Expression::ADD;
202 case Instruction::FAdd: return Expression::FADD;
203 case Instruction::Sub: return Expression::SUB;
204 case Instruction::FSub: return Expression::FSUB;
205 case Instruction::Mul: return Expression::MUL;
206 case Instruction::FMul: return Expression::FMUL;
207 case Instruction::UDiv: return Expression::UDIV;
208 case Instruction::SDiv: return Expression::SDIV;
209 case Instruction::FDiv: return Expression::FDIV;
210 case Instruction::URem: return Expression::UREM;
211 case Instruction::SRem: return Expression::SREM;
212 case Instruction::FRem: return Expression::FREM;
213 case Instruction::Shl: return Expression::SHL;
214 case Instruction::LShr: return Expression::LSHR;
215 case Instruction::AShr: return Expression::ASHR;
216 case Instruction::And: return Expression::AND;
217 case Instruction::Or: return Expression::OR;
218 case Instruction::Xor: return Expression::XOR;
222 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
223 if (isa<ICmpInst>(C)) {
224 switch (C->getPredicate()) {
225 default: // THIS SHOULD NEVER HAPPEN
226 llvm_unreachable("Comparison with unknown predicate?");
227 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
228 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
229 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
230 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
231 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
232 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
233 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
234 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
235 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
236 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
239 switch (C->getPredicate()) {
240 default: // THIS SHOULD NEVER HAPPEN
241 llvm_unreachable("Comparison with unknown predicate?");
242 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
243 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
244 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
245 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
246 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
247 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
248 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
249 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
250 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
251 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
252 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
253 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
254 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
255 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
260 Expression::ExpressionOpcode ValueTable::getOpcode(CastInst* C) {
261 switch(C->getOpcode()) {
262 default: // THIS SHOULD NEVER HAPPEN
263 llvm_unreachable("Cast operator with unknown opcode?");
264 case Instruction::Trunc: return Expression::TRUNC;
265 case Instruction::ZExt: return Expression::ZEXT;
266 case Instruction::SExt: return Expression::SEXT;
267 case Instruction::FPToUI: return Expression::FPTOUI;
268 case Instruction::FPToSI: return Expression::FPTOSI;
269 case Instruction::UIToFP: return Expression::UITOFP;
270 case Instruction::SIToFP: return Expression::SITOFP;
271 case Instruction::FPTrunc: return Expression::FPTRUNC;
272 case Instruction::FPExt: return Expression::FPEXT;
273 case Instruction::PtrToInt: return Expression::PTRTOINT;
274 case Instruction::IntToPtr: return Expression::INTTOPTR;
275 case Instruction::BitCast: return Expression::BITCAST;
279 Expression ValueTable::create_expression(CallInst* C) {
282 e.type = C->getType();
283 e.function = C->getCalledFunction();
284 e.opcode = Expression::CALL;
286 for (CallInst::op_iterator I = C->op_begin()+1, E = C->op_end();
288 e.varargs.push_back(lookup_or_add(*I));
293 Expression ValueTable::create_expression(BinaryOperator* BO) {
295 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
296 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
298 e.type = BO->getType();
299 e.opcode = getOpcode(BO);
304 Expression ValueTable::create_expression(CmpInst* C) {
307 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
308 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
310 e.type = C->getType();
311 e.opcode = getOpcode(C);
316 Expression ValueTable::create_expression(CastInst* C) {
319 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
321 e.type = C->getType();
322 e.opcode = getOpcode(C);
327 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
330 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
331 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
332 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
334 e.type = S->getType();
335 e.opcode = Expression::SHUFFLE;
340 Expression ValueTable::create_expression(ExtractElementInst* E) {
343 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
344 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
346 e.type = E->getType();
347 e.opcode = Expression::EXTRACT;
352 Expression ValueTable::create_expression(InsertElementInst* I) {
355 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
356 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
357 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
359 e.type = I->getType();
360 e.opcode = Expression::INSERT;
365 Expression ValueTable::create_expression(SelectInst* I) {
368 e.varargs.push_back(lookup_or_add(I->getCondition()));
369 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
370 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
372 e.type = I->getType();
373 e.opcode = Expression::SELECT;
378 Expression ValueTable::create_expression(GetElementPtrInst* G) {
381 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
383 e.type = G->getType();
384 e.opcode = Expression::GEP;
386 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
388 e.varargs.push_back(lookup_or_add(*I));
393 Expression ValueTable::create_expression(ExtractValueInst* E) {
396 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
397 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
399 e.varargs.push_back(*II);
401 e.type = E->getType();
402 e.opcode = Expression::EXTRACTVALUE;
407 Expression ValueTable::create_expression(InsertValueInst* E) {
410 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
411 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
412 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
414 e.varargs.push_back(*II);
416 e.type = E->getType();
417 e.opcode = Expression::INSERTVALUE;
422 //===----------------------------------------------------------------------===//
423 // ValueTable External Functions
424 //===----------------------------------------------------------------------===//
426 /// add - Insert a value into the table with a specified value number.
427 void ValueTable::add(Value *V, uint32_t num) {
428 valueNumbering.insert(std::make_pair(V, num));
431 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
432 if (AA->doesNotAccessMemory(C)) {
433 Expression exp = create_expression(C);
434 uint32_t& e = expressionNumbering[exp];
435 if (!e) e = nextValueNumber++;
436 valueNumbering[C] = e;
438 } else if (AA->onlyReadsMemory(C)) {
439 Expression exp = create_expression(C);
440 uint32_t& e = expressionNumbering[exp];
442 e = nextValueNumber++;
443 valueNumbering[C] = e;
447 MemDepResult local_dep = MD->getDependency(C);
449 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
450 valueNumbering[C] = nextValueNumber;
451 return nextValueNumber++;
454 if (local_dep.isDef()) {
455 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
457 if (local_cdep->getNumOperands() != C->getNumOperands()) {
458 valueNumbering[C] = nextValueNumber;
459 return nextValueNumber++;
462 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
463 uint32_t c_vn = lookup_or_add(C->getOperand(i));
464 uint32_t cd_vn = lookup_or_add(local_cdep->getOperand(i));
466 valueNumbering[C] = nextValueNumber;
467 return nextValueNumber++;
471 uint32_t v = lookup_or_add(local_cdep);
472 valueNumbering[C] = v;
477 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
478 MD->getNonLocalCallDependency(CallSite(C));
479 // FIXME: call/call dependencies for readonly calls should return def, not
480 // clobber! Move the checking logic to MemDep!
483 // Check to see if we have a single dominating call instruction that is
485 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
486 const MemoryDependenceAnalysis::NonLocalDepEntry *I = &deps[i];
487 // Ignore non-local dependencies.
488 if (I->second.isNonLocal())
491 // We don't handle non-depedencies. If we already have a call, reject
492 // instruction dependencies.
493 if (I->second.isClobber() || cdep != 0) {
498 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->second.getInst());
499 // FIXME: All duplicated with non-local case.
500 if (NonLocalDepCall && DT->properlyDominates(I->first, C->getParent())){
501 cdep = NonLocalDepCall;
510 valueNumbering[C] = nextValueNumber;
511 return nextValueNumber++;
514 if (cdep->getNumOperands() != C->getNumOperands()) {
515 valueNumbering[C] = nextValueNumber;
516 return nextValueNumber++;
518 for (unsigned i = 1; i < C->getNumOperands(); ++i) {
519 uint32_t c_vn = lookup_or_add(C->getOperand(i));
520 uint32_t cd_vn = lookup_or_add(cdep->getOperand(i));
522 valueNumbering[C] = nextValueNumber;
523 return nextValueNumber++;
527 uint32_t v = lookup_or_add(cdep);
528 valueNumbering[C] = v;
532 valueNumbering[C] = nextValueNumber;
533 return nextValueNumber++;
537 /// lookup_or_add - Returns the value number for the specified value, assigning
538 /// it a new number if it did not have one before.
539 uint32_t ValueTable::lookup_or_add(Value *V) {
540 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
541 if (VI != valueNumbering.end())
544 if (!isa<Instruction>(V)) {
545 valueNumbering[V] = nextValueNumber;
546 return nextValueNumber++;
549 Instruction* I = cast<Instruction>(V);
551 switch (I->getOpcode()) {
552 case Instruction::Call:
553 return lookup_or_add_call(cast<CallInst>(I));
554 case Instruction::Add:
555 case Instruction::FAdd:
556 case Instruction::Sub:
557 case Instruction::FSub:
558 case Instruction::Mul:
559 case Instruction::FMul:
560 case Instruction::UDiv:
561 case Instruction::SDiv:
562 case Instruction::FDiv:
563 case Instruction::URem:
564 case Instruction::SRem:
565 case Instruction::FRem:
566 case Instruction::Shl:
567 case Instruction::LShr:
568 case Instruction::AShr:
569 case Instruction::And:
570 case Instruction::Or :
571 case Instruction::Xor:
572 exp = create_expression(cast<BinaryOperator>(I));
574 case Instruction::ICmp:
575 case Instruction::FCmp:
576 exp = create_expression(cast<CmpInst>(I));
578 case Instruction::Trunc:
579 case Instruction::ZExt:
580 case Instruction::SExt:
581 case Instruction::FPToUI:
582 case Instruction::FPToSI:
583 case Instruction::UIToFP:
584 case Instruction::SIToFP:
585 case Instruction::FPTrunc:
586 case Instruction::FPExt:
587 case Instruction::PtrToInt:
588 case Instruction::IntToPtr:
589 case Instruction::BitCast:
590 exp = create_expression(cast<CastInst>(I));
592 case Instruction::Select:
593 exp = create_expression(cast<SelectInst>(I));
595 case Instruction::ExtractElement:
596 exp = create_expression(cast<ExtractElementInst>(I));
598 case Instruction::InsertElement:
599 exp = create_expression(cast<InsertElementInst>(I));
601 case Instruction::ShuffleVector:
602 exp = create_expression(cast<ShuffleVectorInst>(I));
604 case Instruction::ExtractValue:
605 exp = create_expression(cast<ExtractValueInst>(I));
607 case Instruction::InsertValue:
608 exp = create_expression(cast<InsertValueInst>(I));
610 case Instruction::GetElementPtr:
611 exp = create_expression(cast<GetElementPtrInst>(I));
614 valueNumbering[V] = nextValueNumber;
615 return nextValueNumber++;
618 uint32_t& e = expressionNumbering[exp];
619 if (!e) e = nextValueNumber++;
620 valueNumbering[V] = e;
624 /// lookup - Returns the value number of the specified value. Fails if
625 /// the value has not yet been numbered.
626 uint32_t ValueTable::lookup(Value *V) const {
627 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
628 assert(VI != valueNumbering.end() && "Value not numbered?");
632 /// clear - Remove all entries from the ValueTable
633 void ValueTable::clear() {
634 valueNumbering.clear();
635 expressionNumbering.clear();
639 /// erase - Remove a value from the value numbering
640 void ValueTable::erase(Value *V) {
641 valueNumbering.erase(V);
644 /// verifyRemoved - Verify that the value is removed from all internal data
646 void ValueTable::verifyRemoved(const Value *V) const {
647 for (DenseMap<Value*, uint32_t>::iterator
648 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
649 assert(I->first != V && "Inst still occurs in value numbering map!");
653 //===----------------------------------------------------------------------===//
655 //===----------------------------------------------------------------------===//
658 struct ValueNumberScope {
659 ValueNumberScope* parent;
660 DenseMap<uint32_t, Value*> table;
662 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
668 class GVN : public FunctionPass {
669 bool runOnFunction(Function &F);
671 static char ID; // Pass identification, replacement for typeid
672 GVN(bool nopre = false) : FunctionPass(&ID), NoPRE(nopre) { }
676 MemoryDependenceAnalysis *MD;
680 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
682 // This transformation requires dominator postdominator info
683 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
684 AU.addRequired<DominatorTree>();
685 AU.addRequired<MemoryDependenceAnalysis>();
686 AU.addRequired<AliasAnalysis>();
688 AU.addPreserved<DominatorTree>();
689 AU.addPreserved<AliasAnalysis>();
693 // FIXME: eliminate or document these better
694 bool processLoad(LoadInst* L,
695 SmallVectorImpl<Instruction*> &toErase);
696 bool processInstruction(Instruction *I,
697 SmallVectorImpl<Instruction*> &toErase);
698 bool processNonLocalLoad(LoadInst* L,
699 SmallVectorImpl<Instruction*> &toErase);
700 bool processBlock(BasicBlock *BB);
701 void dump(DenseMap<uint32_t, Value*>& d);
702 bool iterateOnFunction(Function &F);
703 Value *CollapsePhi(PHINode* p);
704 bool performPRE(Function& F);
705 Value *lookupNumber(BasicBlock *BB, uint32_t num);
706 void cleanupGlobalSets();
707 void verifyRemoved(const Instruction *I) const;
713 // createGVNPass - The public interface to this file...
714 FunctionPass *llvm::createGVNPass(bool NoPRE) { return new GVN(NoPRE); }
716 static RegisterPass<GVN> X("gvn",
717 "Global Value Numbering");
719 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
721 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
722 E = d.end(); I != E; ++I) {
723 printf("%d\n", I->first);
729 static bool isSafeReplacement(PHINode* p, Instruction *inst) {
730 if (!isa<PHINode>(inst))
733 for (Instruction::use_iterator UI = p->use_begin(), E = p->use_end();
735 if (PHINode* use_phi = dyn_cast<PHINode>(UI))
736 if (use_phi->getParent() == inst->getParent())
742 Value *GVN::CollapsePhi(PHINode *PN) {
743 Value *ConstVal = PN->hasConstantValue(DT);
744 if (!ConstVal) return 0;
746 Instruction *Inst = dyn_cast<Instruction>(ConstVal);
750 if (DT->dominates(Inst, PN))
751 if (isSafeReplacement(PN, Inst))
756 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
757 /// we're analyzing is fully available in the specified block. As we go, keep
758 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
759 /// map is actually a tri-state map with the following values:
760 /// 0) we know the block *is not* fully available.
761 /// 1) we know the block *is* fully available.
762 /// 2) we do not know whether the block is fully available or not, but we are
763 /// currently speculating that it will be.
764 /// 3) we are speculating for this block and have used that to speculate for
766 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
767 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
768 // Optimistically assume that the block is fully available and check to see
769 // if we already know about this block in one lookup.
770 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
771 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
773 // If the entry already existed for this block, return the precomputed value.
775 // If this is a speculative "available" value, mark it as being used for
776 // speculation of other blocks.
777 if (IV.first->second == 2)
778 IV.first->second = 3;
779 return IV.first->second != 0;
782 // Otherwise, see if it is fully available in all predecessors.
783 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
785 // If this block has no predecessors, it isn't live-in here.
787 goto SpeculationFailure;
789 for (; PI != PE; ++PI)
790 // If the value isn't fully available in one of our predecessors, then it
791 // isn't fully available in this block either. Undo our previous
792 // optimistic assumption and bail out.
793 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
794 goto SpeculationFailure;
798 // SpeculationFailure - If we get here, we found out that this is not, after
799 // all, a fully-available block. We have a problem if we speculated on this and
800 // used the speculation to mark other blocks as available.
802 char &BBVal = FullyAvailableBlocks[BB];
804 // If we didn't speculate on this, just return with it set to false.
810 // If we did speculate on this value, we could have blocks set to 1 that are
811 // incorrect. Walk the (transitive) successors of this block and mark them as
813 SmallVector<BasicBlock*, 32> BBWorklist;
814 BBWorklist.push_back(BB);
816 while (!BBWorklist.empty()) {
817 BasicBlock *Entry = BBWorklist.pop_back_val();
818 // Note that this sets blocks to 0 (unavailable) if they happen to not
819 // already be in FullyAvailableBlocks. This is safe.
820 char &EntryVal = FullyAvailableBlocks[Entry];
821 if (EntryVal == 0) continue; // Already unavailable.
823 // Mark as unavailable.
826 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
827 BBWorklist.push_back(*I);
834 /// CanCoerceMustAliasedValueToLoad - Return true if
835 /// CoerceAvailableValueToLoadType will succeed.
836 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
838 const TargetData &TD) {
839 // If the loaded or stored value is an first class array or struct, don't try
840 // to transform them. We need to be able to bitcast to integer.
841 if (isa<StructType>(LoadTy) || isa<ArrayType>(LoadTy) ||
842 isa<StructType>(StoredVal->getType()) ||
843 isa<ArrayType>(StoredVal->getType()))
846 // The store has to be at least as big as the load.
847 if (TD.getTypeSizeInBits(StoredVal->getType()) <
848 TD.getTypeSizeInBits(LoadTy))
855 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
856 /// then a load from a must-aliased pointer of a different type, try to coerce
857 /// the stored value. LoadedTy is the type of the load we want to replace and
858 /// InsertPt is the place to insert new instructions.
860 /// If we can't do it, return null.
861 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
862 const Type *LoadedTy,
863 Instruction *InsertPt,
864 const TargetData &TD) {
865 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
868 const Type *StoredValTy = StoredVal->getType();
870 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
871 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
873 // If the store and reload are the same size, we can always reuse it.
874 if (StoreSize == LoadSize) {
875 if (isa<PointerType>(StoredValTy) && isa<PointerType>(LoadedTy)) {
876 // Pointer to Pointer -> use bitcast.
877 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
880 // Convert source pointers to integers, which can be bitcast.
881 if (isa<PointerType>(StoredValTy)) {
882 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
883 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
886 const Type *TypeToCastTo = LoadedTy;
887 if (isa<PointerType>(TypeToCastTo))
888 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
890 if (StoredValTy != TypeToCastTo)
891 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
893 // Cast to pointer if the load needs a pointer type.
894 if (isa<PointerType>(LoadedTy))
895 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
900 // If the loaded value is smaller than the available value, then we can
901 // extract out a piece from it. If the available value is too small, then we
902 // can't do anything.
903 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
905 // Convert source pointers to integers, which can be manipulated.
906 if (isa<PointerType>(StoredValTy)) {
907 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
908 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
911 // Convert vectors and fp to integer, which can be manipulated.
912 if (!isa<IntegerType>(StoredValTy)) {
913 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
914 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
917 // If this is a big-endian system, we need to shift the value down to the low
918 // bits so that a truncate will work.
919 if (TD.isBigEndian()) {
920 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
921 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
924 // Truncate the integer to the right size now.
925 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
926 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
928 if (LoadedTy == NewIntTy)
931 // If the result is a pointer, inttoptr.
932 if (isa<PointerType>(LoadedTy))
933 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
935 // Otherwise, bitcast.
936 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
939 /// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
940 /// be expressed as a base pointer plus a constant offset. Return the base and
941 /// offset to the caller.
942 static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
943 const TargetData &TD) {
944 Operator *PtrOp = dyn_cast<Operator>(Ptr);
945 if (PtrOp == 0) return Ptr;
947 // Just look through bitcasts.
948 if (PtrOp->getOpcode() == Instruction::BitCast)
949 return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
951 // If this is a GEP with constant indices, we can look through it.
952 GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
953 if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
955 gep_type_iterator GTI = gep_type_begin(GEP);
956 for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
958 ConstantInt *OpC = cast<ConstantInt>(*I);
959 if (OpC->isZero()) continue;
961 // Handle a struct and array indices which add their offset to the pointer.
962 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
963 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
965 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
966 Offset += OpC->getSExtValue()*Size;
970 // Re-sign extend from the pointer size if needed to get overflow edge cases
972 unsigned PtrSize = TD.getPointerSizeInBits();
974 Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
976 return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
980 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
981 /// memdep query of a load that ends up being a clobbering store. This means
982 /// that the store *may* provide bits used by the load but we can't be sure
983 /// because the pointers don't mustalias. Check this case to see if there is
984 /// anything more we can do before we give up. This returns -1 if we have to
985 /// give up, or a byte number in the stored value of the piece that feeds the
987 static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI,
988 const TargetData &TD) {
989 // If the loaded or stored value is an first class array or struct, don't try
990 // to transform them. We need to be able to bitcast to integer.
991 if (isa<StructType>(L->getType()) || isa<ArrayType>(L->getType()) ||
992 isa<StructType>(DepSI->getOperand(0)->getType()) ||
993 isa<ArrayType>(DepSI->getOperand(0)->getType()))
996 int64_t StoreOffset = 0, LoadOffset = 0;
998 GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
1000 GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
1001 if (StoreBase != LoadBase)
1004 // If the load and store are to the exact same address, they should have been
1005 // a must alias. AA must have gotten confused.
1006 // FIXME: Study to see if/when this happens.
1007 if (LoadOffset == StoreOffset) {
1009 errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1010 << "Base = " << *StoreBase << "\n"
1011 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
1012 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
1013 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1014 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1015 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1021 // If the load and store don't overlap at all, the store doesn't provide
1022 // anything to the load. In this case, they really don't alias at all, AA
1023 // must have gotten confused.
1024 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1025 // remove this check, as it is duplicated with what we have below.
1026 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1027 uint64_t LoadSize = TD.getTypeSizeInBits(L->getType());
1029 if ((StoreSize & 7) | (LoadSize & 7))
1031 StoreSize >>= 3; // Convert to bytes.
1035 bool isAAFailure = false;
1036 if (StoreOffset < LoadOffset) {
1037 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1039 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1043 errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1044 << "Base = " << *StoreBase << "\n"
1045 << "Store Ptr = " << *DepSI->getPointerOperand() << "\n"
1046 << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
1047 << "Load Ptr = " << *L->getPointerOperand() << "\n"
1048 << "Load Offs = " << LoadOffset << " - " << *L << "\n\n";
1049 errs() << "'" << L->getParent()->getParent()->getName() << "'"
1055 // If the Load isn't completely contained within the stored bits, we don't
1056 // have all the bits to feed it. We could do something crazy in the future
1057 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1059 if (StoreOffset > LoadOffset ||
1060 StoreOffset+StoreSize < LoadOffset+LoadSize)
1063 // Okay, we can do this transformation. Return the number of bytes into the
1064 // store that the load is.
1065 return LoadOffset-StoreOffset;
1069 /// GetStoreValueForLoad - This function is called when we have a
1070 /// memdep query of a load that ends up being a clobbering store. This means
1071 /// that the store *may* provide bits used by the load but we can't be sure
1072 /// because the pointers don't mustalias. Check this case to see if there is
1073 /// anything more we can do before we give up.
1074 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1076 Instruction *InsertPt, const TargetData &TD){
1077 LLVMContext &Ctx = SrcVal->getType()->getContext();
1079 uint64_t StoreSize = TD.getTypeSizeInBits(SrcVal->getType())/8;
1080 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1083 // Compute which bits of the stored value are being used by the load. Convert
1084 // to an integer type to start with.
1085 if (isa<PointerType>(SrcVal->getType()))
1086 SrcVal = new PtrToIntInst(SrcVal, TD.getIntPtrType(Ctx), "tmp", InsertPt);
1087 if (!isa<IntegerType>(SrcVal->getType()))
1088 SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1091 // Shift the bits to the least significant depending on endianness.
1093 if (TD.isLittleEndian()) {
1094 ShiftAmt = Offset*8;
1096 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1100 SrcVal = BinaryOperator::CreateLShr(SrcVal,
1101 ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);
1103 if (LoadSize != StoreSize)
1104 SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1107 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1110 struct AvailableValueInBlock {
1111 /// BB - The basic block in question.
1113 /// V - The value that is live out of the block.
1115 /// Offset - The byte offset in V that is interesting for the load query.
1118 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1119 unsigned Offset = 0) {
1120 AvailableValueInBlock Res;
1123 Res.Offset = Offset;
1128 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1129 /// construct SSA form, allowing us to eliminate LI. This returns the value
1130 /// that should be used at LI's definition site.
1131 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1132 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1133 const TargetData *TD,
1134 AliasAnalysis *AA) {
1135 SmallVector<PHINode*, 8> NewPHIs;
1136 SSAUpdater SSAUpdate(&NewPHIs);
1137 SSAUpdate.Initialize(LI);
1139 const Type *LoadTy = LI->getType();
1141 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1142 BasicBlock *BB = ValuesPerBlock[i].BB;
1143 Value *AvailableVal = ValuesPerBlock[i].V;
1144 unsigned Offset = ValuesPerBlock[i].Offset;
1146 if (SSAUpdate.HasValueForBlock(BB))
1149 if (AvailableVal->getType() != LoadTy) {
1150 assert(TD && "Need target data to handle type mismatch case");
1151 AvailableVal = GetStoreValueForLoad(AvailableVal, Offset, LoadTy,
1152 BB->getTerminator(), *TD);
1155 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
1156 << *ValuesPerBlock[i].V << '\n'
1157 << *AvailableVal << '\n' << "\n\n\n");
1161 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
1162 << *ValuesPerBlock[i].V << '\n'
1163 << *AvailableVal << '\n' << "\n\n\n");
1166 SSAUpdate.AddAvailableValue(BB, AvailableVal);
1169 // Perform PHI construction.
1170 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1172 // If new PHI nodes were created, notify alias analysis.
1173 if (isa<PointerType>(V->getType()))
1174 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1175 AA->copyValue(LI, NewPHIs[i]);
1180 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1181 /// non-local by performing PHI construction.
1182 bool GVN::processNonLocalLoad(LoadInst *LI,
1183 SmallVectorImpl<Instruction*> &toErase) {
1184 // Find the non-local dependencies of the load.
1185 SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
1186 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1188 //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
1189 // << Deps.size() << *LI << '\n');
1191 // If we had to process more than one hundred blocks to find the
1192 // dependencies, this load isn't worth worrying about. Optimizing
1193 // it will be too expensive.
1194 if (Deps.size() > 100)
1197 // If we had a phi translation failure, we'll have a single entry which is a
1198 // clobber in the current block. Reject this early.
1199 if (Deps.size() == 1 && Deps[0].second.isClobber()) {
1201 errs() << "GVN: non-local load ";
1202 WriteAsOperand(errs(), LI);
1203 errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n';
1208 // Filter out useless results (non-locals, etc). Keep track of the blocks
1209 // where we have a value available in repl, also keep track of whether we see
1210 // dependencies that produce an unknown value for the load (such as a call
1211 // that could potentially clobber the load).
1212 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1213 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1215 const TargetData *TD = 0;
1217 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1218 BasicBlock *DepBB = Deps[i].first;
1219 MemDepResult DepInfo = Deps[i].second;
1221 if (DepInfo.isClobber()) {
1222 // If the dependence is to a store that writes to a superset of the bits
1223 // read by the load, we can extract the bits we need for the load from the
1225 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1227 TD = getAnalysisIfAvailable<TargetData>();
1229 int Offset = AnalyzeLoadFromClobberingStore(LI, DepSI, *TD);
1231 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1232 DepSI->getOperand(0),
1239 // FIXME: Handle memset/memcpy.
1240 UnavailableBlocks.push_back(DepBB);
1244 Instruction *DepInst = DepInfo.getInst();
1246 // Loading the allocation -> undef.
1247 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1248 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1249 UndefValue::get(LI->getType())));
1253 // Loading immediately after lifetime begin or end -> undef.
1254 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1255 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1256 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1257 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1258 UndefValue::get(LI->getType())));
1262 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1263 // Reject loads and stores that are to the same address but are of
1264 // different types if we have to.
1265 if (S->getOperand(0)->getType() != LI->getType()) {
1267 TD = getAnalysisIfAvailable<TargetData>();
1269 // If the stored value is larger or equal to the loaded value, we can
1271 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1272 LI->getType(), *TD)) {
1273 UnavailableBlocks.push_back(DepBB);
1278 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1283 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1284 // If the types mismatch and we can't handle it, reject reuse of the load.
1285 if (LD->getType() != LI->getType()) {
1287 TD = getAnalysisIfAvailable<TargetData>();
1289 // If the stored value is larger or equal to the loaded value, we can
1291 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1292 UnavailableBlocks.push_back(DepBB);
1296 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1300 UnavailableBlocks.push_back(DepBB);
1304 // If we have no predecessors that produce a known value for this load, exit
1306 if (ValuesPerBlock.empty()) return false;
1308 // If all of the instructions we depend on produce a known value for this
1309 // load, then it is fully redundant and we can use PHI insertion to compute
1310 // its value. Insert PHIs and remove the fully redundant value now.
1311 if (UnavailableBlocks.empty()) {
1312 DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1314 // Perform PHI construction.
1315 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1316 VN.getAliasAnalysis());
1317 LI->replaceAllUsesWith(V);
1319 if (isa<PHINode>(V))
1321 if (isa<PointerType>(V->getType()))
1322 MD->invalidateCachedPointerInfo(V);
1323 toErase.push_back(LI);
1328 if (!EnablePRE || !EnableLoadPRE)
1331 // Okay, we have *some* definitions of the value. This means that the value
1332 // is available in some of our (transitive) predecessors. Lets think about
1333 // doing PRE of this load. This will involve inserting a new load into the
1334 // predecessor when it's not available. We could do this in general, but
1335 // prefer to not increase code size. As such, we only do this when we know
1336 // that we only have to insert *one* load (which means we're basically moving
1337 // the load, not inserting a new one).
1339 SmallPtrSet<BasicBlock *, 4> Blockers;
1340 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1341 Blockers.insert(UnavailableBlocks[i]);
1343 // Lets find first basic block with more than one predecessor. Walk backwards
1344 // through predecessors if needed.
1345 BasicBlock *LoadBB = LI->getParent();
1346 BasicBlock *TmpBB = LoadBB;
1348 bool isSinglePred = false;
1349 bool allSingleSucc = true;
1350 while (TmpBB->getSinglePredecessor()) {
1351 isSinglePred = true;
1352 TmpBB = TmpBB->getSinglePredecessor();
1353 if (!TmpBB) // If haven't found any, bail now.
1355 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1357 if (Blockers.count(TmpBB))
1359 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1360 allSingleSucc = false;
1366 // If we have a repl set with LI itself in it, this means we have a loop where
1367 // at least one of the values is LI. Since this means that we won't be able
1368 // to eliminate LI even if we insert uses in the other predecessors, we will
1369 // end up increasing code size. Reject this by scanning for LI.
1370 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1371 if (ValuesPerBlock[i].V == LI)
1376 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1377 if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].V))
1378 // "Hot" Instruction is in some loop (because it dominates its dep.
1380 if (DT->dominates(LI, I)) {
1385 // We are interested only in "hot" instructions. We don't want to do any
1386 // mis-optimizations here.
1391 // Okay, we have some hope :). Check to see if the loaded value is fully
1392 // available in all but one predecessor.
1393 // FIXME: If we could restructure the CFG, we could make a common pred with
1394 // all the preds that don't have an available LI and insert a new load into
1396 BasicBlock *UnavailablePred = 0;
1398 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1399 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1400 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1401 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1402 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1404 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1406 if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
1409 // If this load is not available in multiple predecessors, reject it.
1410 if (UnavailablePred && UnavailablePred != *PI)
1412 UnavailablePred = *PI;
1415 assert(UnavailablePred != 0 &&
1416 "Fully available value should be eliminated above!");
1418 // If the loaded pointer is PHI node defined in this block, do PHI translation
1419 // to get its value in the predecessor.
1420 Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);
1422 // Make sure the value is live in the predecessor. If it was defined by a
1423 // non-PHI instruction in this block, we don't know how to recompute it above.
1424 if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
1425 if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
1426 DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
1427 << *LPInst << '\n' << *LI << "\n");
1431 // We don't currently handle critical edges :(
1432 if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
1433 DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
1434 << UnavailablePred->getName() << "': " << *LI << '\n');
1438 // Make sure it is valid to move this load here. We have to watch out for:
1439 // @1 = getelementptr (i8* p, ...
1440 // test p and branch if == 0
1442 // It is valid to have the getelementptr before the test, even if p can be 0,
1443 // as getelementptr only does address arithmetic.
1444 // If we are not pushing the value through any multiple-successor blocks
1445 // we do not have this case. Otherwise, check that the load is safe to
1446 // put anywhere; this can be improved, but should be conservatively safe.
1447 if (!allSingleSucc &&
1448 !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator()))
1451 // Okay, we can eliminate this load by inserting a reload in the predecessor
1452 // and using PHI construction to get the value in the other predecessors, do
1454 DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1456 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1458 UnavailablePred->getTerminator());
1460 // Add the newly created load.
1461 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,NewLoad));
1463 // Perform PHI construction.
1464 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD,
1465 VN.getAliasAnalysis());
1466 LI->replaceAllUsesWith(V);
1467 if (isa<PHINode>(V))
1469 if (isa<PointerType>(V->getType()))
1470 MD->invalidateCachedPointerInfo(V);
1471 toErase.push_back(LI);
1476 /// processLoad - Attempt to eliminate a load, first by eliminating it
1477 /// locally, and then attempting non-local elimination if that fails.
1478 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1479 if (L->isVolatile())
1482 // ... to a pointer that has been loaded from before...
1483 MemDepResult Dep = MD->getDependency(L);
1485 // If the value isn't available, don't do anything!
1486 if (Dep.isClobber()) {
1487 // FIXME: We should handle memset/memcpy/memmove as dependent instructions
1488 // to forward the value if available.
1489 //if (isa<MemIntrinsic>(Dep.getInst()))
1490 //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n";
1492 // Check to see if we have something like this:
1493 // store i32 123, i32* %P
1494 // %A = bitcast i32* %P to i8*
1495 // %B = gep i8* %A, i32 1
1498 // We could do that by recognizing if the clobber instructions are obviously
1499 // a common base + constant offset, and if the previous store (or memset)
1500 // completely covers this load. This sort of thing can happen in bitfield
1502 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1503 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1504 int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, *TD);
1506 Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1507 L->getType(), L, *TD);
1508 DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n'
1509 << *AvailVal << '\n' << *L << "\n\n\n");
1511 // Replace the load!
1512 L->replaceAllUsesWith(AvailVal);
1513 if (isa<PointerType>(AvailVal->getType()))
1514 MD->invalidateCachedPointerInfo(AvailVal);
1515 toErase.push_back(L);
1522 // fast print dep, using operator<< on instruction would be too slow
1523 errs() << "GVN: load ";
1524 WriteAsOperand(errs(), L);
1525 Instruction *I = Dep.getInst();
1526 errs() << " is clobbered by " << *I << '\n';
1531 // If it is defined in another block, try harder.
1532 if (Dep.isNonLocal())
1533 return processNonLocalLoad(L, toErase);
1535 Instruction *DepInst = Dep.getInst();
1536 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1537 Value *StoredVal = DepSI->getOperand(0);
1539 // The store and load are to a must-aliased pointer, but they may not
1540 // actually have the same type. See if we know how to reuse the stored
1541 // value (depending on its type).
1542 const TargetData *TD = 0;
1543 if (StoredVal->getType() != L->getType()) {
1544 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1545 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1550 DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1551 << '\n' << *L << "\n\n\n");
1558 L->replaceAllUsesWith(StoredVal);
1559 if (isa<PointerType>(StoredVal->getType()))
1560 MD->invalidateCachedPointerInfo(StoredVal);
1561 toErase.push_back(L);
1566 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1567 Value *AvailableVal = DepLI;
1569 // The loads are of a must-aliased pointer, but they may not actually have
1570 // the same type. See if we know how to reuse the previously loaded value
1571 // (depending on its type).
1572 const TargetData *TD = 0;
1573 if (DepLI->getType() != L->getType()) {
1574 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1575 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1576 if (AvailableVal == 0)
1579 DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1580 << "\n" << *L << "\n\n\n");
1587 L->replaceAllUsesWith(AvailableVal);
1588 if (isa<PointerType>(DepLI->getType()))
1589 MD->invalidateCachedPointerInfo(DepLI);
1590 toErase.push_back(L);
1595 // If this load really doesn't depend on anything, then we must be loading an
1596 // undef value. This can happen when loading for a fresh allocation with no
1597 // intervening stores, for example.
1598 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1599 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1600 toErase.push_back(L);
1605 // If this load occurs either right after a lifetime begin or a lifetime end,
1606 // then the loaded value is undefined.
1607 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1608 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1609 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1610 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1611 toErase.push_back(L);
1620 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1621 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1622 if (I == localAvail.end())
1625 ValueNumberScope *Locals = I->second;
1627 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1628 if (I != Locals->table.end())
1630 Locals = Locals->parent;
1637 /// processInstruction - When calculating availability, handle an instruction
1638 /// by inserting it into the appropriate sets
1639 bool GVN::processInstruction(Instruction *I,
1640 SmallVectorImpl<Instruction*> &toErase) {
1641 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1642 bool Changed = processLoad(LI, toErase);
1645 unsigned Num = VN.lookup_or_add(LI);
1646 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1652 uint32_t NextNum = VN.getNextUnusedValueNumber();
1653 unsigned Num = VN.lookup_or_add(I);
1655 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1656 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1658 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1661 Value *BranchCond = BI->getCondition();
1662 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1664 BasicBlock *TrueSucc = BI->getSuccessor(0);
1665 BasicBlock *FalseSucc = BI->getSuccessor(1);
1667 if (TrueSucc->getSinglePredecessor())
1668 localAvail[TrueSucc]->table[CondVN] =
1669 ConstantInt::getTrue(TrueSucc->getContext());
1670 if (FalseSucc->getSinglePredecessor())
1671 localAvail[FalseSucc]->table[CondVN] =
1672 ConstantInt::getFalse(TrueSucc->getContext());
1676 // Allocations are always uniquely numbered, so we can save time and memory
1677 // by fast failing them.
1678 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1679 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1683 // Collapse PHI nodes
1684 if (PHINode* p = dyn_cast<PHINode>(I)) {
1685 Value *constVal = CollapsePhi(p);
1688 p->replaceAllUsesWith(constVal);
1689 if (isa<PointerType>(constVal->getType()))
1690 MD->invalidateCachedPointerInfo(constVal);
1693 toErase.push_back(p);
1695 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1698 // If the number we were assigned was a brand new VN, then we don't
1699 // need to do a lookup to see if the number already exists
1700 // somewhere in the domtree: it can't!
1701 } else if (Num == NextNum) {
1702 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1704 // Perform fast-path value-number based elimination of values inherited from
1706 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1709 I->replaceAllUsesWith(repl);
1710 if (isa<PointerType>(repl->getType()))
1711 MD->invalidateCachedPointerInfo(repl);
1712 toErase.push_back(I);
1716 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1722 /// runOnFunction - This is the main transformation entry point for a function.
1723 bool GVN::runOnFunction(Function& F) {
1724 MD = &getAnalysis<MemoryDependenceAnalysis>();
1725 DT = &getAnalysis<DominatorTree>();
1726 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1730 bool Changed = false;
1731 bool ShouldContinue = true;
1733 // Merge unconditional branches, allowing PRE to catch more
1734 // optimization opportunities.
1735 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1736 BasicBlock *BB = FI;
1738 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1739 if (removedBlock) NumGVNBlocks++;
1741 Changed |= removedBlock;
1744 unsigned Iteration = 0;
1746 while (ShouldContinue) {
1747 DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
1748 ShouldContinue = iterateOnFunction(F);
1749 Changed |= ShouldContinue;
1754 bool PREChanged = true;
1755 while (PREChanged) {
1756 PREChanged = performPRE(F);
1757 Changed |= PREChanged;
1760 // FIXME: Should perform GVN again after PRE does something. PRE can move
1761 // computations into blocks where they become fully redundant. Note that
1762 // we can't do this until PRE's critical edge splitting updates memdep.
1763 // Actually, when this happens, we should just fully integrate PRE into GVN.
1765 cleanupGlobalSets();
1771 bool GVN::processBlock(BasicBlock *BB) {
1772 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1773 // incrementing BI before processing an instruction).
1774 SmallVector<Instruction*, 8> toErase;
1775 bool ChangedFunction = false;
1777 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1779 ChangedFunction |= processInstruction(BI, toErase);
1780 if (toErase.empty()) {
1785 // If we need some instructions deleted, do it now.
1786 NumGVNInstr += toErase.size();
1788 // Avoid iterator invalidation.
1789 bool AtStart = BI == BB->begin();
1793 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1794 E = toErase.end(); I != E; ++I) {
1795 DEBUG(errs() << "GVN removed: " << **I << '\n');
1796 MD->removeInstruction(*I);
1797 (*I)->eraseFromParent();
1798 DEBUG(verifyRemoved(*I));
1808 return ChangedFunction;
1811 /// performPRE - Perform a purely local form of PRE that looks for diamond
1812 /// control flow patterns and attempts to perform simple PRE at the join point.
1813 bool GVN::performPRE(Function& F) {
1814 bool Changed = false;
1815 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
1816 DenseMap<BasicBlock*, Value*> predMap;
1817 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1818 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1819 BasicBlock *CurrentBlock = *DI;
1821 // Nothing to PRE in the entry block.
1822 if (CurrentBlock == &F.getEntryBlock()) continue;
1824 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1825 BE = CurrentBlock->end(); BI != BE; ) {
1826 Instruction *CurInst = BI++;
1828 if (isa<AllocaInst>(CurInst) ||
1829 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
1830 CurInst->getType()->isVoidTy() ||
1831 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1832 isa<DbgInfoIntrinsic>(CurInst))
1835 uint32_t ValNo = VN.lookup(CurInst);
1837 // Look for the predecessors for PRE opportunities. We're
1838 // only trying to solve the basic diamond case, where
1839 // a value is computed in the successor and one predecessor,
1840 // but not the other. We also explicitly disallow cases
1841 // where the successor is its own predecessor, because they're
1842 // more complicated to get right.
1843 unsigned NumWith = 0;
1844 unsigned NumWithout = 0;
1845 BasicBlock *PREPred = 0;
1848 for (pred_iterator PI = pred_begin(CurrentBlock),
1849 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1850 // We're not interested in PRE where the block is its
1851 // own predecessor, on in blocks with predecessors
1852 // that are not reachable.
1853 if (*PI == CurrentBlock) {
1856 } else if (!localAvail.count(*PI)) {
1861 DenseMap<uint32_t, Value*>::iterator predV =
1862 localAvail[*PI]->table.find(ValNo);
1863 if (predV == localAvail[*PI]->table.end()) {
1866 } else if (predV->second == CurInst) {
1869 predMap[*PI] = predV->second;
1874 // Don't do PRE when it might increase code size, i.e. when
1875 // we would need to insert instructions in more than one pred.
1876 if (NumWithout != 1 || NumWith == 0)
1879 // We can't do PRE safely on a critical edge, so instead we schedule
1880 // the edge to be split and perform the PRE the next time we iterate
1882 unsigned SuccNum = 0;
1883 for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
1885 if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
1890 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
1891 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
1895 // Instantiate the expression the in predecessor that lacked it.
1896 // Because we are going top-down through the block, all value numbers
1897 // will be available in the predecessor by the time we need them. Any
1898 // that weren't original present will have been instantiated earlier
1900 Instruction *PREInstr = CurInst->clone();
1901 bool success = true;
1902 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1903 Value *Op = PREInstr->getOperand(i);
1904 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1907 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
1908 PREInstr->setOperand(i, V);
1915 // Fail out if we encounter an operand that is not available in
1916 // the PRE predecessor. This is typically because of loads which
1917 // are not value numbered precisely.
1920 DEBUG(verifyRemoved(PREInstr));
1924 PREInstr->insertBefore(PREPred->getTerminator());
1925 PREInstr->setName(CurInst->getName() + ".pre");
1926 predMap[PREPred] = PREInstr;
1927 VN.add(PREInstr, ValNo);
1930 // Update the availability map to include the new instruction.
1931 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
1933 // Create a PHI to make the value available in this block.
1934 PHINode* Phi = PHINode::Create(CurInst->getType(),
1935 CurInst->getName() + ".pre-phi",
1936 CurrentBlock->begin());
1937 for (pred_iterator PI = pred_begin(CurrentBlock),
1938 PE = pred_end(CurrentBlock); PI != PE; ++PI)
1939 Phi->addIncoming(predMap[*PI], *PI);
1942 localAvail[CurrentBlock]->table[ValNo] = Phi;
1944 CurInst->replaceAllUsesWith(Phi);
1945 if (isa<PointerType>(Phi->getType()))
1946 MD->invalidateCachedPointerInfo(Phi);
1949 DEBUG(errs() << "GVN PRE removed: " << *CurInst << '\n');
1950 MD->removeInstruction(CurInst);
1951 CurInst->eraseFromParent();
1952 DEBUG(verifyRemoved(CurInst));
1957 for (SmallVector<std::pair<TerminatorInst*, unsigned>, 4>::iterator
1958 I = toSplit.begin(), E = toSplit.end(); I != E; ++I)
1959 SplitCriticalEdge(I->first, I->second, this);
1961 return Changed || toSplit.size();
1964 /// iterateOnFunction - Executes one iteration of GVN
1965 bool GVN::iterateOnFunction(Function &F) {
1966 cleanupGlobalSets();
1968 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1969 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
1971 localAvail[DI->getBlock()] =
1972 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
1974 localAvail[DI->getBlock()] = new ValueNumberScope(0);
1977 // Top-down walk of the dominator tree
1978 bool Changed = false;
1980 // Needed for value numbering with phi construction to work.
1981 ReversePostOrderTraversal<Function*> RPOT(&F);
1982 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
1983 RE = RPOT.end(); RI != RE; ++RI)
1984 Changed |= processBlock(*RI);
1986 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
1987 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
1988 Changed |= processBlock(DI->getBlock());
1994 void GVN::cleanupGlobalSets() {
1997 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
1998 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2003 /// verifyRemoved - Verify that the specified instruction does not occur in our
2004 /// internal data structures.
2005 void GVN::verifyRemoved(const Instruction *Inst) const {
2006 VN.verifyRemoved(Inst);
2008 // Walk through the value number scope to make sure the instruction isn't
2009 // ferreted away in it.
2010 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2011 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2012 const ValueNumberScope *VNS = I->second;
2015 for (DenseMap<uint32_t, Value*>::iterator
2016 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2017 assert(II->second != Inst && "Inst still in value numbering scope!");