1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/BasicBlock.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/LLVMContext.h"
27 #include "llvm/Operator.h"
28 #include "llvm/Value.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/DepthFirstIterator.h"
31 #include "llvm/ADT/PostOrderIterator.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/ConstantFolding.h"
37 #include "llvm/Analysis/Dominators.h"
38 #include "llvm/Analysis/Loads.h"
39 #include "llvm/Analysis/MemoryBuiltins.h"
40 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
41 #include "llvm/Analysis/PHITransAddr.h"
42 #include "llvm/Support/CFG.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/GetElementPtrTypeIterator.h"
47 #include "llvm/Support/IRBuilder.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Target/TargetData.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
55 STATISTIC(NumGVNInstr, "Number of instructions deleted");
56 STATISTIC(NumGVNLoad, "Number of loads deleted");
57 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
58 STATISTIC(NumGVNBlocks, "Number of blocks merged");
59 STATISTIC(NumPRELoad, "Number of loads PRE'd");
61 static cl::opt<bool> EnablePRE("enable-pre",
62 cl::init(true), cl::Hidden);
63 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
64 static cl::opt<bool> EnableFullLoadPRE("enable-full-load-pre", cl::init(false));
66 //===----------------------------------------------------------------------===//
68 //===----------------------------------------------------------------------===//
70 /// This class holds the mapping between values and value numbers. It is used
71 /// as an efficient mechanism to determine the expression-wise equivalence of
75 enum ExpressionOpcode {
76 ADD = Instruction::Add,
77 FADD = Instruction::FAdd,
78 SUB = Instruction::Sub,
79 FSUB = Instruction::FSub,
80 MUL = Instruction::Mul,
81 FMUL = Instruction::FMul,
82 UDIV = Instruction::UDiv,
83 SDIV = Instruction::SDiv,
84 FDIV = Instruction::FDiv,
85 UREM = Instruction::URem,
86 SREM = Instruction::SRem,
87 FREM = Instruction::FRem,
88 SHL = Instruction::Shl,
89 LSHR = Instruction::LShr,
90 ASHR = Instruction::AShr,
91 AND = Instruction::And,
93 XOR = Instruction::Xor,
94 TRUNC = Instruction::Trunc,
95 ZEXT = Instruction::ZExt,
96 SEXT = Instruction::SExt,
97 FPTOUI = Instruction::FPToUI,
98 FPTOSI = Instruction::FPToSI,
99 UITOFP = Instruction::UIToFP,
100 SITOFP = Instruction::SIToFP,
101 FPTRUNC = Instruction::FPTrunc,
102 FPEXT = Instruction::FPExt,
103 PTRTOINT = Instruction::PtrToInt,
104 INTTOPTR = Instruction::IntToPtr,
105 BITCAST = Instruction::BitCast,
106 ICMPEQ, ICMPNE, ICMPUGT, ICMPUGE, ICMPULT, ICMPULE,
107 ICMPSGT, ICMPSGE, ICMPSLT, ICMPSLE, FCMPOEQ,
108 FCMPOGT, FCMPOGE, FCMPOLT, FCMPOLE, FCMPONE,
109 FCMPORD, FCMPUNO, FCMPUEQ, FCMPUGT, FCMPUGE,
110 FCMPULT, FCMPULE, FCMPUNE, EXTRACT, INSERT,
111 SHUFFLE, SELECT, GEP, CALL, CONSTANT,
112 INSERTVALUE, EXTRACTVALUE, EMPTY, TOMBSTONE };
114 ExpressionOpcode opcode;
116 SmallVector<uint32_t, 4> varargs;
120 Expression(ExpressionOpcode o) : opcode(o) { }
122 bool operator==(const Expression &other) const {
123 if (opcode != other.opcode)
125 else if (opcode == EMPTY || opcode == TOMBSTONE)
127 else if (type != other.type)
129 else if (function != other.function)
132 if (varargs.size() != other.varargs.size())
135 for (size_t i = 0; i < varargs.size(); ++i)
136 if (varargs[i] != other.varargs[i])
143 /*bool operator!=(const Expression &other) const {
144 return !(*this == other);
150 DenseMap<Value*, uint32_t> valueNumbering;
151 DenseMap<Expression, uint32_t> expressionNumbering;
153 MemoryDependenceAnalysis* MD;
156 uint32_t nextValueNumber;
158 Expression::ExpressionOpcode getOpcode(CmpInst* C);
159 Expression create_expression(BinaryOperator* BO);
160 Expression create_expression(CmpInst* C);
161 Expression create_expression(ShuffleVectorInst* V);
162 Expression create_expression(ExtractElementInst* C);
163 Expression create_expression(InsertElementInst* V);
164 Expression create_expression(SelectInst* V);
165 Expression create_expression(CastInst* C);
166 Expression create_expression(GetElementPtrInst* G);
167 Expression create_expression(CallInst* C);
168 Expression create_expression(ExtractValueInst* C);
169 Expression create_expression(InsertValueInst* C);
171 uint32_t lookup_or_add_call(CallInst* C);
173 ValueTable() : nextValueNumber(1) { }
174 uint32_t lookup_or_add(Value *V);
175 uint32_t lookup(Value *V) const;
176 void add(Value *V, uint32_t num);
178 void erase(Value *v);
179 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
180 AliasAnalysis *getAliasAnalysis() const { return AA; }
181 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
182 void setDomTree(DominatorTree* D) { DT = D; }
183 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
184 void verifyRemoved(const Value *) const;
189 template <> struct DenseMapInfo<Expression> {
190 static inline Expression getEmptyKey() {
191 return Expression(Expression::EMPTY);
194 static inline Expression getTombstoneKey() {
195 return Expression(Expression::TOMBSTONE);
198 static unsigned getHashValue(const Expression e) {
199 unsigned hash = e.opcode;
201 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
202 (unsigned)((uintptr_t)e.type >> 9));
204 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
205 E = e.varargs.end(); I != E; ++I)
206 hash = *I + hash * 37;
208 hash = ((unsigned)((uintptr_t)e.function >> 4) ^
209 (unsigned)((uintptr_t)e.function >> 9)) +
214 static bool isEqual(const Expression &LHS, const Expression &RHS) {
220 struct isPodLike<Expression> { static const bool value = true; };
224 //===----------------------------------------------------------------------===//
225 // ValueTable Internal Functions
226 //===----------------------------------------------------------------------===//
228 Expression::ExpressionOpcode ValueTable::getOpcode(CmpInst* C) {
229 if (isa<ICmpInst>(C)) {
230 switch (C->getPredicate()) {
231 default: // THIS SHOULD NEVER HAPPEN
232 llvm_unreachable("Comparison with unknown predicate?");
233 case ICmpInst::ICMP_EQ: return Expression::ICMPEQ;
234 case ICmpInst::ICMP_NE: return Expression::ICMPNE;
235 case ICmpInst::ICMP_UGT: return Expression::ICMPUGT;
236 case ICmpInst::ICMP_UGE: return Expression::ICMPUGE;
237 case ICmpInst::ICMP_ULT: return Expression::ICMPULT;
238 case ICmpInst::ICMP_ULE: return Expression::ICMPULE;
239 case ICmpInst::ICMP_SGT: return Expression::ICMPSGT;
240 case ICmpInst::ICMP_SGE: return Expression::ICMPSGE;
241 case ICmpInst::ICMP_SLT: return Expression::ICMPSLT;
242 case ICmpInst::ICMP_SLE: return Expression::ICMPSLE;
245 switch (C->getPredicate()) {
246 default: // THIS SHOULD NEVER HAPPEN
247 llvm_unreachable("Comparison with unknown predicate?");
248 case FCmpInst::FCMP_OEQ: return Expression::FCMPOEQ;
249 case FCmpInst::FCMP_OGT: return Expression::FCMPOGT;
250 case FCmpInst::FCMP_OGE: return Expression::FCMPOGE;
251 case FCmpInst::FCMP_OLT: return Expression::FCMPOLT;
252 case FCmpInst::FCMP_OLE: return Expression::FCMPOLE;
253 case FCmpInst::FCMP_ONE: return Expression::FCMPONE;
254 case FCmpInst::FCMP_ORD: return Expression::FCMPORD;
255 case FCmpInst::FCMP_UNO: return Expression::FCMPUNO;
256 case FCmpInst::FCMP_UEQ: return Expression::FCMPUEQ;
257 case FCmpInst::FCMP_UGT: return Expression::FCMPUGT;
258 case FCmpInst::FCMP_UGE: return Expression::FCMPUGE;
259 case FCmpInst::FCMP_ULT: return Expression::FCMPULT;
260 case FCmpInst::FCMP_ULE: return Expression::FCMPULE;
261 case FCmpInst::FCMP_UNE: return Expression::FCMPUNE;
266 Expression ValueTable::create_expression(CallInst* C) {
269 e.type = C->getType();
270 e.function = C->getCalledFunction();
271 e.opcode = Expression::CALL;
274 for (CallInst::op_iterator I = CS.arg_begin(), E = CS.arg_end();
276 e.varargs.push_back(lookup_or_add(*I));
281 Expression ValueTable::create_expression(BinaryOperator* BO) {
283 e.varargs.push_back(lookup_or_add(BO->getOperand(0)));
284 e.varargs.push_back(lookup_or_add(BO->getOperand(1)));
286 e.type = BO->getType();
287 e.opcode = static_cast<Expression::ExpressionOpcode>(BO->getOpcode());
292 Expression ValueTable::create_expression(CmpInst* C) {
295 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
296 e.varargs.push_back(lookup_or_add(C->getOperand(1)));
298 e.type = C->getType();
299 e.opcode = getOpcode(C);
304 Expression ValueTable::create_expression(CastInst* C) {
307 e.varargs.push_back(lookup_or_add(C->getOperand(0)));
309 e.type = C->getType();
310 e.opcode = static_cast<Expression::ExpressionOpcode>(C->getOpcode());
315 Expression ValueTable::create_expression(ShuffleVectorInst* S) {
318 e.varargs.push_back(lookup_or_add(S->getOperand(0)));
319 e.varargs.push_back(lookup_or_add(S->getOperand(1)));
320 e.varargs.push_back(lookup_or_add(S->getOperand(2)));
322 e.type = S->getType();
323 e.opcode = Expression::SHUFFLE;
328 Expression ValueTable::create_expression(ExtractElementInst* E) {
331 e.varargs.push_back(lookup_or_add(E->getOperand(0)));
332 e.varargs.push_back(lookup_or_add(E->getOperand(1)));
334 e.type = E->getType();
335 e.opcode = Expression::EXTRACT;
340 Expression ValueTable::create_expression(InsertElementInst* I) {
343 e.varargs.push_back(lookup_or_add(I->getOperand(0)));
344 e.varargs.push_back(lookup_or_add(I->getOperand(1)));
345 e.varargs.push_back(lookup_or_add(I->getOperand(2)));
347 e.type = I->getType();
348 e.opcode = Expression::INSERT;
353 Expression ValueTable::create_expression(SelectInst* I) {
356 e.varargs.push_back(lookup_or_add(I->getCondition()));
357 e.varargs.push_back(lookup_or_add(I->getTrueValue()));
358 e.varargs.push_back(lookup_or_add(I->getFalseValue()));
360 e.type = I->getType();
361 e.opcode = Expression::SELECT;
366 Expression ValueTable::create_expression(GetElementPtrInst* G) {
369 e.varargs.push_back(lookup_or_add(G->getPointerOperand()));
371 e.type = G->getType();
372 e.opcode = Expression::GEP;
374 for (GetElementPtrInst::op_iterator I = G->idx_begin(), E = G->idx_end();
376 e.varargs.push_back(lookup_or_add(*I));
381 Expression ValueTable::create_expression(ExtractValueInst* E) {
384 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
385 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
387 e.varargs.push_back(*II);
389 e.type = E->getType();
390 e.opcode = Expression::EXTRACTVALUE;
395 Expression ValueTable::create_expression(InsertValueInst* E) {
398 e.varargs.push_back(lookup_or_add(E->getAggregateOperand()));
399 e.varargs.push_back(lookup_or_add(E->getInsertedValueOperand()));
400 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
402 e.varargs.push_back(*II);
404 e.type = E->getType();
405 e.opcode = Expression::INSERTVALUE;
410 //===----------------------------------------------------------------------===//
411 // ValueTable External Functions
412 //===----------------------------------------------------------------------===//
414 /// add - Insert a value into the table with a specified value number.
415 void ValueTable::add(Value *V, uint32_t num) {
416 valueNumbering.insert(std::make_pair(V, num));
419 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
420 if (AA->doesNotAccessMemory(C)) {
421 Expression exp = create_expression(C);
422 uint32_t& e = expressionNumbering[exp];
423 if (!e) e = nextValueNumber++;
424 valueNumbering[C] = e;
426 } else if (AA->onlyReadsMemory(C)) {
427 Expression exp = create_expression(C);
428 uint32_t& e = expressionNumbering[exp];
430 e = nextValueNumber++;
431 valueNumbering[C] = e;
435 e = nextValueNumber++;
436 valueNumbering[C] = e;
440 MemDepResult local_dep = MD->getDependency(C);
442 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
443 valueNumbering[C] = nextValueNumber;
444 return nextValueNumber++;
447 if (local_dep.isDef()) {
448 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
450 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
451 valueNumbering[C] = nextValueNumber;
452 return nextValueNumber++;
455 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
456 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
457 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
459 valueNumbering[C] = nextValueNumber;
460 return nextValueNumber++;
464 uint32_t v = lookup_or_add(local_cdep);
465 valueNumbering[C] = v;
470 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
471 MD->getNonLocalCallDependency(CallSite(C));
472 // FIXME: call/call dependencies for readonly calls should return def, not
473 // clobber! Move the checking logic to MemDep!
476 // Check to see if we have a single dominating call instruction that is
478 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
479 const NonLocalDepEntry *I = &deps[i];
480 // Ignore non-local dependencies.
481 if (I->getResult().isNonLocal())
484 // We don't handle non-depedencies. If we already have a call, reject
485 // instruction dependencies.
486 if (I->getResult().isClobber() || cdep != 0) {
491 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
492 // FIXME: All duplicated with non-local case.
493 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
494 cdep = NonLocalDepCall;
503 valueNumbering[C] = nextValueNumber;
504 return nextValueNumber++;
507 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
508 valueNumbering[C] = nextValueNumber;
509 return nextValueNumber++;
511 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
512 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
513 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
515 valueNumbering[C] = nextValueNumber;
516 return nextValueNumber++;
520 uint32_t v = lookup_or_add(cdep);
521 valueNumbering[C] = v;
525 valueNumbering[C] = nextValueNumber;
526 return nextValueNumber++;
530 /// lookup_or_add - Returns the value number for the specified value, assigning
531 /// it a new number if it did not have one before.
532 uint32_t ValueTable::lookup_or_add(Value *V) {
533 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
534 if (VI != valueNumbering.end())
537 if (!isa<Instruction>(V)) {
538 valueNumbering[V] = nextValueNumber;
539 return nextValueNumber++;
542 Instruction* I = cast<Instruction>(V);
544 switch (I->getOpcode()) {
545 case Instruction::Call:
546 return lookup_or_add_call(cast<CallInst>(I));
547 case Instruction::Add:
548 case Instruction::FAdd:
549 case Instruction::Sub:
550 case Instruction::FSub:
551 case Instruction::Mul:
552 case Instruction::FMul:
553 case Instruction::UDiv:
554 case Instruction::SDiv:
555 case Instruction::FDiv:
556 case Instruction::URem:
557 case Instruction::SRem:
558 case Instruction::FRem:
559 case Instruction::Shl:
560 case Instruction::LShr:
561 case Instruction::AShr:
562 case Instruction::And:
563 case Instruction::Or :
564 case Instruction::Xor:
565 exp = create_expression(cast<BinaryOperator>(I));
567 case Instruction::ICmp:
568 case Instruction::FCmp:
569 exp = create_expression(cast<CmpInst>(I));
571 case Instruction::Trunc:
572 case Instruction::ZExt:
573 case Instruction::SExt:
574 case Instruction::FPToUI:
575 case Instruction::FPToSI:
576 case Instruction::UIToFP:
577 case Instruction::SIToFP:
578 case Instruction::FPTrunc:
579 case Instruction::FPExt:
580 case Instruction::PtrToInt:
581 case Instruction::IntToPtr:
582 case Instruction::BitCast:
583 exp = create_expression(cast<CastInst>(I));
585 case Instruction::Select:
586 exp = create_expression(cast<SelectInst>(I));
588 case Instruction::ExtractElement:
589 exp = create_expression(cast<ExtractElementInst>(I));
591 case Instruction::InsertElement:
592 exp = create_expression(cast<InsertElementInst>(I));
594 case Instruction::ShuffleVector:
595 exp = create_expression(cast<ShuffleVectorInst>(I));
597 case Instruction::ExtractValue:
598 exp = create_expression(cast<ExtractValueInst>(I));
600 case Instruction::InsertValue:
601 exp = create_expression(cast<InsertValueInst>(I));
603 case Instruction::GetElementPtr:
604 exp = create_expression(cast<GetElementPtrInst>(I));
607 valueNumbering[V] = nextValueNumber;
608 return nextValueNumber++;
611 uint32_t& e = expressionNumbering[exp];
612 if (!e) e = nextValueNumber++;
613 valueNumbering[V] = e;
617 /// lookup - Returns the value number of the specified value. Fails if
618 /// the value has not yet been numbered.
619 uint32_t ValueTable::lookup(Value *V) const {
620 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
621 assert(VI != valueNumbering.end() && "Value not numbered?");
625 /// clear - Remove all entries from the ValueTable
626 void ValueTable::clear() {
627 valueNumbering.clear();
628 expressionNumbering.clear();
632 /// erase - Remove a value from the value numbering
633 void ValueTable::erase(Value *V) {
634 valueNumbering.erase(V);
637 /// verifyRemoved - Verify that the value is removed from all internal data
639 void ValueTable::verifyRemoved(const Value *V) const {
640 for (DenseMap<Value*, uint32_t>::const_iterator
641 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
642 assert(I->first != V && "Inst still occurs in value numbering map!");
646 //===----------------------------------------------------------------------===//
648 //===----------------------------------------------------------------------===//
651 struct ValueNumberScope {
652 ValueNumberScope* parent;
653 DenseMap<uint32_t, Value*> table;
655 ValueNumberScope(ValueNumberScope* p) : parent(p) { }
661 class GVN : public FunctionPass {
662 bool runOnFunction(Function &F);
664 static char ID; // Pass identification, replacement for typeid
665 explicit GVN(bool noloads = false)
666 : FunctionPass(ID), NoLoads(noloads), MD(0) { }
670 MemoryDependenceAnalysis *MD;
674 DenseMap<BasicBlock*, ValueNumberScope*> localAvail;
676 // List of critical edges to be split between iterations.
677 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
679 // This transformation requires dominator postdominator info
680 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
681 AU.addRequired<DominatorTree>();
683 AU.addRequired<MemoryDependenceAnalysis>();
684 AU.addRequired<AliasAnalysis>();
686 AU.addPreserved<DominatorTree>();
687 AU.addPreserved<AliasAnalysis>();
691 // FIXME: eliminate or document these better
692 bool processLoad(LoadInst* L,
693 SmallVectorImpl<Instruction*> &toErase);
694 bool processInstruction(Instruction *I,
695 SmallVectorImpl<Instruction*> &toErase);
696 bool processNonLocalLoad(LoadInst* L,
697 SmallVectorImpl<Instruction*> &toErase);
698 bool processBlock(BasicBlock *BB);
699 void dump(DenseMap<uint32_t, Value*>& d);
700 bool iterateOnFunction(Function &F);
701 Value *CollapsePhi(PHINode* p);
702 bool performPRE(Function& F);
703 Value *lookupNumber(BasicBlock *BB, uint32_t num);
704 void cleanupGlobalSets();
705 void verifyRemoved(const Instruction *I) const;
706 bool splitCriticalEdges();
712 // createGVNPass - The public interface to this file...
713 FunctionPass *llvm::createGVNPass(bool NoLoads) {
714 return new GVN(NoLoads);
717 INITIALIZE_PASS(GVN, "gvn", "Global Value Numbering", false, false);
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 errs() << I->first << "\n";
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);
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);
828 } while (!BBWorklist.empty());
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 (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
842 StoredVal->getType()->isStructTy() ||
843 StoredVal->getType()->isArrayTy())
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.getTypeStoreSizeInBits(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 (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) {
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 (StoredValTy->isPointerTy()) {
882 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
883 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
886 const Type *TypeToCastTo = LoadedTy;
887 if (TypeToCastTo->isPointerTy())
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 (LoadedTy->isPointerTy())
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 (StoredValTy->isPointerTy()) {
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 (!StoredValTy->isIntegerTy()) {
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 (LoadedTy->isPointerTy())
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 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
981 /// memdep query of a load that ends up being a clobbering memory write (store,
982 /// memset, memcpy, memmove). This means that the write *may* provide bits used
983 /// by the load but we can't be sure because the pointers don't mustalias.
985 /// Check this case to see if there is anything more we can do before we give
986 /// up. This returns -1 if we have to give up, or a byte number in the stored
987 /// value of the piece that feeds the load.
988 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
990 uint64_t WriteSizeInBits,
991 const TargetData &TD) {
992 // If the loaded or stored value is an first class array or struct, don't try
993 // to transform them. We need to be able to bitcast to integer.
994 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
997 int64_t StoreOffset = 0, LoadOffset = 0;
998 Value *StoreBase = GetBaseWithConstantOffset(WritePtr, StoreOffset, TD);
1000 GetBaseWithConstantOffset(LoadPtr, 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. One case is forwarding a memset
1007 // to a load from the base of the memset.
1009 if (LoadOffset == StoreOffset) {
1010 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
1011 << "Base = " << *StoreBase << "\n"
1012 << "Store Ptr = " << *WritePtr << "\n"
1013 << "Store Offs = " << StoreOffset << "\n"
1014 << "Load Ptr = " << *LoadPtr << "\n";
1019 // If the load and store don't overlap at all, the store doesn't provide
1020 // anything to the load. In this case, they really don't alias at all, AA
1021 // must have gotten confused.
1022 // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
1023 // remove this check, as it is duplicated with what we have below.
1024 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
1026 if ((WriteSizeInBits & 7) | (LoadSize & 7))
1028 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
1032 bool isAAFailure = false;
1033 if (StoreOffset < LoadOffset)
1034 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1036 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1040 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1041 << "Base = " << *StoreBase << "\n"
1042 << "Store Ptr = " << *WritePtr << "\n"
1043 << "Store Offs = " << StoreOffset << "\n"
1044 << "Load Ptr = " << *LoadPtr << "\n";
1050 // If the Load isn't completely contained within the stored bits, we don't
1051 // have all the bits to feed it. We could do something crazy in the future
1052 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1054 if (StoreOffset > LoadOffset ||
1055 StoreOffset+StoreSize < LoadOffset+LoadSize)
1058 // Okay, we can do this transformation. Return the number of bytes into the
1059 // store that the load is.
1060 return LoadOffset-StoreOffset;
1063 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1064 /// memdep query of a load that ends up being a clobbering store.
1065 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
1067 const TargetData &TD) {
1068 // Cannot handle reading from store of first-class aggregate yet.
1069 if (DepSI->getOperand(0)->getType()->isStructTy() ||
1070 DepSI->getOperand(0)->getType()->isArrayTy())
1073 Value *StorePtr = DepSI->getPointerOperand();
1074 uint64_t StoreSize = TD.getTypeSizeInBits(DepSI->getOperand(0)->getType());
1075 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1076 StorePtr, StoreSize, TD);
1079 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
1081 const TargetData &TD) {
1082 // If the mem operation is a non-constant size, we can't handle it.
1083 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1084 if (SizeCst == 0) return -1;
1085 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1087 // If this is memset, we just need to see if the offset is valid in the size
1089 if (MI->getIntrinsicID() == Intrinsic::memset)
1090 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1093 // If we have a memcpy/memmove, the only case we can handle is if this is a
1094 // copy from constant memory. In that case, we can read directly from the
1096 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1098 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1099 if (Src == 0) return -1;
1101 GlobalVariable *GV = dyn_cast<GlobalVariable>(Src->getUnderlyingObject());
1102 if (GV == 0 || !GV->isConstant()) return -1;
1104 // See if the access is within the bounds of the transfer.
1105 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1106 MI->getDest(), MemSizeInBits, TD);
1110 // Otherwise, see if we can constant fold a load from the constant with the
1111 // offset applied as appropriate.
1112 Src = ConstantExpr::getBitCast(Src,
1113 llvm::Type::getInt8PtrTy(Src->getContext()));
1114 Constant *OffsetCst =
1115 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1116 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1117 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1118 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1124 /// GetStoreValueForLoad - This function is called when we have a
1125 /// memdep query of a load that ends up being a clobbering store. This means
1126 /// that the store *may* provide bits used by the load but we can't be sure
1127 /// because the pointers don't mustalias. Check this case to see if there is
1128 /// anything more we can do before we give up.
1129 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1131 Instruction *InsertPt, const TargetData &TD){
1132 LLVMContext &Ctx = SrcVal->getType()->getContext();
1134 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1135 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1137 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1139 // Compute which bits of the stored value are being used by the load. Convert
1140 // to an integer type to start with.
1141 if (SrcVal->getType()->isPointerTy())
1142 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
1143 if (!SrcVal->getType()->isIntegerTy())
1144 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
1147 // Shift the bits to the least significant depending on endianness.
1149 if (TD.isLittleEndian())
1150 ShiftAmt = Offset*8;
1152 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1155 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
1157 if (LoadSize != StoreSize)
1158 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
1161 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1164 /// GetMemInstValueForLoad - This function is called when we have a
1165 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1166 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1167 const Type *LoadTy, Instruction *InsertPt,
1168 const TargetData &TD){
1169 LLVMContext &Ctx = LoadTy->getContext();
1170 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1172 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1174 // We know that this method is only called when the mem transfer fully
1175 // provides the bits for the load.
1176 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1177 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1178 // independently of what the offset is.
1179 Value *Val = MSI->getValue();
1181 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1183 Value *OneElt = Val;
1185 // Splat the value out to the right number of bits.
1186 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1187 // If we can double the number of bytes set, do it.
1188 if (NumBytesSet*2 <= LoadSize) {
1189 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1190 Val = Builder.CreateOr(Val, ShVal);
1195 // Otherwise insert one byte at a time.
1196 Value *ShVal = Builder.CreateShl(Val, 1*8);
1197 Val = Builder.CreateOr(OneElt, ShVal);
1201 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1204 // Otherwise, this is a memcpy/memmove from a constant global.
1205 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1206 Constant *Src = cast<Constant>(MTI->getSource());
1208 // Otherwise, see if we can constant fold a load from the constant with the
1209 // offset applied as appropriate.
1210 Src = ConstantExpr::getBitCast(Src,
1211 llvm::Type::getInt8PtrTy(Src->getContext()));
1212 Constant *OffsetCst =
1213 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1214 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1215 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1216 return ConstantFoldLoadFromConstPtr(Src, &TD);
1221 struct AvailableValueInBlock {
1222 /// BB - The basic block in question.
1225 SimpleVal, // A simple offsetted value that is accessed.
1226 MemIntrin // A memory intrinsic which is loaded from.
1229 /// V - The value that is live out of the block.
1230 PointerIntPair<Value *, 1, ValType> Val;
1232 /// Offset - The byte offset in Val that is interesting for the load query.
1235 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1236 unsigned Offset = 0) {
1237 AvailableValueInBlock Res;
1239 Res.Val.setPointer(V);
1240 Res.Val.setInt(SimpleVal);
1241 Res.Offset = Offset;
1245 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1246 unsigned Offset = 0) {
1247 AvailableValueInBlock Res;
1249 Res.Val.setPointer(MI);
1250 Res.Val.setInt(MemIntrin);
1251 Res.Offset = Offset;
1255 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1256 Value *getSimpleValue() const {
1257 assert(isSimpleValue() && "Wrong accessor");
1258 return Val.getPointer();
1261 MemIntrinsic *getMemIntrinValue() const {
1262 assert(!isSimpleValue() && "Wrong accessor");
1263 return cast<MemIntrinsic>(Val.getPointer());
1266 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1267 /// defined here to the specified type. This handles various coercion cases.
1268 Value *MaterializeAdjustedValue(const Type *LoadTy,
1269 const TargetData *TD) const {
1271 if (isSimpleValue()) {
1272 Res = getSimpleValue();
1273 if (Res->getType() != LoadTy) {
1274 assert(TD && "Need target data to handle type mismatch case");
1275 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1278 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1279 << *getSimpleValue() << '\n'
1280 << *Res << '\n' << "\n\n\n");
1283 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1284 LoadTy, BB->getTerminator(), *TD);
1285 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1286 << " " << *getMemIntrinValue() << '\n'
1287 << *Res << '\n' << "\n\n\n");
1295 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1296 /// construct SSA form, allowing us to eliminate LI. This returns the value
1297 /// that should be used at LI's definition site.
1298 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1299 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1300 const TargetData *TD,
1301 const DominatorTree &DT,
1302 AliasAnalysis *AA) {
1303 // Check for the fully redundant, dominating load case. In this case, we can
1304 // just use the dominating value directly.
1305 if (ValuesPerBlock.size() == 1 &&
1306 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1307 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1309 // Otherwise, we have to construct SSA form.
1310 SmallVector<PHINode*, 8> NewPHIs;
1311 SSAUpdater SSAUpdate(&NewPHIs);
1312 SSAUpdate.Initialize(LI->getType(), LI->getName());
1314 const Type *LoadTy = LI->getType();
1316 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1317 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1318 BasicBlock *BB = AV.BB;
1320 if (SSAUpdate.HasValueForBlock(BB))
1323 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1326 // Perform PHI construction.
1327 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1329 // If new PHI nodes were created, notify alias analysis.
1330 if (V->getType()->isPointerTy())
1331 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1332 AA->copyValue(LI, NewPHIs[i]);
1337 static bool isLifetimeStart(const Instruction *Inst) {
1338 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1339 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1343 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1344 /// non-local by performing PHI construction.
1345 bool GVN::processNonLocalLoad(LoadInst *LI,
1346 SmallVectorImpl<Instruction*> &toErase) {
1347 // Find the non-local dependencies of the load.
1348 SmallVector<NonLocalDepResult, 64> Deps;
1349 MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
1351 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1352 // << Deps.size() << *LI << '\n');
1354 // If we had to process more than one hundred blocks to find the
1355 // dependencies, this load isn't worth worrying about. Optimizing
1356 // it will be too expensive.
1357 if (Deps.size() > 100)
1360 // If we had a phi translation failure, we'll have a single entry which is a
1361 // clobber in the current block. Reject this early.
1362 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
1364 dbgs() << "GVN: non-local load ";
1365 WriteAsOperand(dbgs(), LI);
1366 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1371 // Filter out useless results (non-locals, etc). Keep track of the blocks
1372 // where we have a value available in repl, also keep track of whether we see
1373 // dependencies that produce an unknown value for the load (such as a call
1374 // that could potentially clobber the load).
1375 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1376 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1378 const TargetData *TD = 0;
1380 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1381 BasicBlock *DepBB = Deps[i].getBB();
1382 MemDepResult DepInfo = Deps[i].getResult();
1384 if (DepInfo.isClobber()) {
1385 // The address being loaded in this non-local block may not be the same as
1386 // the pointer operand of the load if PHI translation occurs. Make sure
1387 // to consider the right address.
1388 Value *Address = Deps[i].getAddress();
1390 // If the dependence is to a store that writes to a superset of the bits
1391 // read by the load, we can extract the bits we need for the load from the
1393 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1395 TD = getAnalysisIfAvailable<TargetData>();
1396 if (TD && Address) {
1397 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1400 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1401 DepSI->getOperand(0),
1408 // If the clobbering value is a memset/memcpy/memmove, see if we can
1409 // forward a value on from it.
1410 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1412 TD = getAnalysisIfAvailable<TargetData>();
1413 if (TD && Address) {
1414 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1417 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1424 UnavailableBlocks.push_back(DepBB);
1428 Instruction *DepInst = DepInfo.getInst();
1430 // Loading the allocation -> undef.
1431 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1432 // Loading immediately after lifetime begin -> undef.
1433 isLifetimeStart(DepInst)) {
1434 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1435 UndefValue::get(LI->getType())));
1439 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1440 // Reject loads and stores that are to the same address but are of
1441 // different types if we have to.
1442 if (S->getOperand(0)->getType() != LI->getType()) {
1444 TD = getAnalysisIfAvailable<TargetData>();
1446 // If the stored value is larger or equal to the loaded value, we can
1448 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getOperand(0),
1449 LI->getType(), *TD)) {
1450 UnavailableBlocks.push_back(DepBB);
1455 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1460 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1461 // If the types mismatch and we can't handle it, reject reuse of the load.
1462 if (LD->getType() != LI->getType()) {
1464 TD = getAnalysisIfAvailable<TargetData>();
1466 // If the stored value is larger or equal to the loaded value, we can
1468 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1469 UnavailableBlocks.push_back(DepBB);
1473 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1477 UnavailableBlocks.push_back(DepBB);
1481 // If we have no predecessors that produce a known value for this load, exit
1483 if (ValuesPerBlock.empty()) return false;
1485 // If all of the instructions we depend on produce a known value for this
1486 // load, then it is fully redundant and we can use PHI insertion to compute
1487 // its value. Insert PHIs and remove the fully redundant value now.
1488 if (UnavailableBlocks.empty()) {
1489 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1491 // Perform PHI construction.
1492 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1493 VN.getAliasAnalysis());
1494 LI->replaceAllUsesWith(V);
1496 if (isa<PHINode>(V))
1498 if (V->getType()->isPointerTy())
1499 MD->invalidateCachedPointerInfo(V);
1501 toErase.push_back(LI);
1506 if (!EnablePRE || !EnableLoadPRE)
1509 // Okay, we have *some* definitions of the value. This means that the value
1510 // is available in some of our (transitive) predecessors. Lets think about
1511 // doing PRE of this load. This will involve inserting a new load into the
1512 // predecessor when it's not available. We could do this in general, but
1513 // prefer to not increase code size. As such, we only do this when we know
1514 // that we only have to insert *one* load (which means we're basically moving
1515 // the load, not inserting a new one).
1517 SmallPtrSet<BasicBlock *, 4> Blockers;
1518 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1519 Blockers.insert(UnavailableBlocks[i]);
1521 // Lets find first basic block with more than one predecessor. Walk backwards
1522 // through predecessors if needed.
1523 BasicBlock *LoadBB = LI->getParent();
1524 BasicBlock *TmpBB = LoadBB;
1526 bool isSinglePred = false;
1527 bool allSingleSucc = true;
1528 while (TmpBB->getSinglePredecessor()) {
1529 isSinglePred = true;
1530 TmpBB = TmpBB->getSinglePredecessor();
1531 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1533 if (Blockers.count(TmpBB))
1535 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1536 allSingleSucc = false;
1542 // If we have a repl set with LI itself in it, this means we have a loop where
1543 // at least one of the values is LI. Since this means that we won't be able
1544 // to eliminate LI even if we insert uses in the other predecessors, we will
1545 // end up increasing code size. Reject this by scanning for LI.
1546 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1547 if (ValuesPerBlock[i].isSimpleValue() &&
1548 ValuesPerBlock[i].getSimpleValue() == LI) {
1549 // Skip cases where LI is the only definition, even for EnableFullLoadPRE.
1550 if (!EnableFullLoadPRE || e == 1)
1555 // FIXME: It is extremely unclear what this loop is doing, other than
1556 // artificially restricting loadpre.
1559 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1560 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1561 if (AV.isSimpleValue())
1562 // "Hot" Instruction is in some loop (because it dominates its dep.
1564 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1565 if (DT->dominates(LI, I)) {
1571 // We are interested only in "hot" instructions. We don't want to do any
1572 // mis-optimizations here.
1577 // Check to see how many predecessors have the loaded value fully
1579 DenseMap<BasicBlock*, Value*> PredLoads;
1580 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1581 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1582 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1583 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1584 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1586 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1587 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1589 BasicBlock *Pred = *PI;
1590 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1593 PredLoads[Pred] = 0;
1595 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1596 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1597 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1598 << Pred->getName() << "': " << *LI << '\n');
1601 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1602 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1605 if (!NeedToSplit.empty()) {
1606 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1610 // Decide whether PRE is profitable for this load.
1611 unsigned NumUnavailablePreds = PredLoads.size();
1612 assert(NumUnavailablePreds != 0 &&
1613 "Fully available value should be eliminated above!");
1614 if (!EnableFullLoadPRE) {
1615 // If this load is unavailable in multiple predecessors, reject it.
1616 // FIXME: If we could restructure the CFG, we could make a common pred with
1617 // all the preds that don't have an available LI and insert a new load into
1619 if (NumUnavailablePreds != 1)
1623 // Check if the load can safely be moved to all the unavailable predecessors.
1624 bool CanDoPRE = true;
1625 SmallVector<Instruction*, 8> NewInsts;
1626 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1627 E = PredLoads.end(); I != E; ++I) {
1628 BasicBlock *UnavailablePred = I->first;
1630 // Do PHI translation to get its value in the predecessor if necessary. The
1631 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1633 // If all preds have a single successor, then we know it is safe to insert
1634 // the load on the pred (?!?), so we can insert code to materialize the
1635 // pointer if it is not available.
1636 PHITransAddr Address(LI->getOperand(0), TD);
1638 if (allSingleSucc) {
1639 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1642 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1643 LoadPtr = Address.getAddr();
1646 // If we couldn't find or insert a computation of this phi translated value,
1649 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1650 << *LI->getOperand(0) << "\n");
1655 // Make sure it is valid to move this load here. We have to watch out for:
1656 // @1 = getelementptr (i8* p, ...
1657 // test p and branch if == 0
1659 // It is valid to have the getelementptr before the test, even if p can be 0,
1660 // as getelementptr only does address arithmetic.
1661 // If we are not pushing the value through any multiple-successor blocks
1662 // we do not have this case. Otherwise, check that the load is safe to
1663 // put anywhere; this can be improved, but should be conservatively safe.
1664 if (!allSingleSucc &&
1665 // FIXME: REEVALUTE THIS.
1666 !isSafeToLoadUnconditionally(LoadPtr,
1667 UnavailablePred->getTerminator(),
1668 LI->getAlignment(), TD)) {
1673 I->second = LoadPtr;
1677 while (!NewInsts.empty())
1678 NewInsts.pop_back_val()->eraseFromParent();
1682 // Okay, we can eliminate this load by inserting a reload in the predecessor
1683 // and using PHI construction to get the value in the other predecessors, do
1685 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1686 DEBUG(if (!NewInsts.empty())
1687 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1688 << *NewInsts.back() << '\n');
1690 // Assign value numbers to the new instructions.
1691 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1692 // FIXME: We really _ought_ to insert these value numbers into their
1693 // parent's availability map. However, in doing so, we risk getting into
1694 // ordering issues. If a block hasn't been processed yet, we would be
1695 // marking a value as AVAIL-IN, which isn't what we intend.
1696 VN.lookup_or_add(NewInsts[i]);
1699 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1700 E = PredLoads.end(); I != E; ++I) {
1701 BasicBlock *UnavailablePred = I->first;
1702 Value *LoadPtr = I->second;
1704 Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1706 UnavailablePred->getTerminator());
1708 // Add the newly created load.
1709 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1711 MD->invalidateCachedPointerInfo(LoadPtr);
1712 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1715 // Perform PHI construction.
1716 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1717 VN.getAliasAnalysis());
1718 LI->replaceAllUsesWith(V);
1719 if (isa<PHINode>(V))
1721 if (V->getType()->isPointerTy())
1722 MD->invalidateCachedPointerInfo(V);
1724 toErase.push_back(LI);
1729 /// processLoad - Attempt to eliminate a load, first by eliminating it
1730 /// locally, and then attempting non-local elimination if that fails.
1731 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1735 if (L->isVolatile())
1738 // ... to a pointer that has been loaded from before...
1739 MemDepResult Dep = MD->getDependency(L);
1741 // If the value isn't available, don't do anything!
1742 if (Dep.isClobber()) {
1743 // Check to see if we have something like this:
1744 // store i32 123, i32* %P
1745 // %A = bitcast i32* %P to i8*
1746 // %B = gep i8* %A, i32 1
1749 // We could do that by recognizing if the clobber instructions are obviously
1750 // a common base + constant offset, and if the previous store (or memset)
1751 // completely covers this load. This sort of thing can happen in bitfield
1753 Value *AvailVal = 0;
1754 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1755 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1756 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1757 L->getPointerOperand(),
1760 AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
1761 L->getType(), L, *TD);
1764 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1765 // a value on from it.
1766 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1767 if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
1768 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1769 L->getPointerOperand(),
1772 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1777 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1778 << *AvailVal << '\n' << *L << "\n\n\n");
1780 // Replace the load!
1781 L->replaceAllUsesWith(AvailVal);
1782 if (AvailVal->getType()->isPointerTy())
1783 MD->invalidateCachedPointerInfo(AvailVal);
1785 toErase.push_back(L);
1791 // fast print dep, using operator<< on instruction would be too slow
1792 dbgs() << "GVN: load ";
1793 WriteAsOperand(dbgs(), L);
1794 Instruction *I = Dep.getInst();
1795 dbgs() << " is clobbered by " << *I << '\n';
1800 // If it is defined in another block, try harder.
1801 if (Dep.isNonLocal())
1802 return processNonLocalLoad(L, toErase);
1804 Instruction *DepInst = Dep.getInst();
1805 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1806 Value *StoredVal = DepSI->getOperand(0);
1808 // The store and load are to a must-aliased pointer, but they may not
1809 // actually have the same type. See if we know how to reuse the stored
1810 // value (depending on its type).
1811 const TargetData *TD = 0;
1812 if (StoredVal->getType() != L->getType()) {
1813 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1814 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1819 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1820 << '\n' << *L << "\n\n\n");
1827 L->replaceAllUsesWith(StoredVal);
1828 if (StoredVal->getType()->isPointerTy())
1829 MD->invalidateCachedPointerInfo(StoredVal);
1831 toErase.push_back(L);
1836 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1837 Value *AvailableVal = DepLI;
1839 // The loads are of a must-aliased pointer, but they may not actually have
1840 // the same type. See if we know how to reuse the previously loaded value
1841 // (depending on its type).
1842 const TargetData *TD = 0;
1843 if (DepLI->getType() != L->getType()) {
1844 if ((TD = getAnalysisIfAvailable<TargetData>())) {
1845 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1846 if (AvailableVal == 0)
1849 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1850 << "\n" << *L << "\n\n\n");
1857 L->replaceAllUsesWith(AvailableVal);
1858 if (DepLI->getType()->isPointerTy())
1859 MD->invalidateCachedPointerInfo(DepLI);
1861 toErase.push_back(L);
1866 // If this load really doesn't depend on anything, then we must be loading an
1867 // undef value. This can happen when loading for a fresh allocation with no
1868 // intervening stores, for example.
1869 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1870 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1872 toErase.push_back(L);
1877 // If this load occurs either right after a lifetime begin,
1878 // then the loaded value is undefined.
1879 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1880 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1881 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1883 toErase.push_back(L);
1892 Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
1893 DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
1894 if (I == localAvail.end())
1897 ValueNumberScope *Locals = I->second;
1899 DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
1900 if (I != Locals->table.end())
1902 Locals = Locals->parent;
1909 /// processInstruction - When calculating availability, handle an instruction
1910 /// by inserting it into the appropriate sets
1911 bool GVN::processInstruction(Instruction *I,
1912 SmallVectorImpl<Instruction*> &toErase) {
1913 // Ignore dbg info intrinsics.
1914 if (isa<DbgInfoIntrinsic>(I))
1917 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1918 bool Changed = processLoad(LI, toErase);
1921 unsigned Num = VN.lookup_or_add(LI);
1922 localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
1928 uint32_t NextNum = VN.getNextUnusedValueNumber();
1929 unsigned Num = VN.lookup_or_add(I);
1931 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1932 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1934 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1937 Value *BranchCond = BI->getCondition();
1938 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1940 BasicBlock *TrueSucc = BI->getSuccessor(0);
1941 BasicBlock *FalseSucc = BI->getSuccessor(1);
1943 if (TrueSucc->getSinglePredecessor())
1944 localAvail[TrueSucc]->table[CondVN] =
1945 ConstantInt::getTrue(TrueSucc->getContext());
1946 if (FalseSucc->getSinglePredecessor())
1947 localAvail[FalseSucc]->table[CondVN] =
1948 ConstantInt::getFalse(TrueSucc->getContext());
1952 // Allocations are always uniquely numbered, so we can save time and memory
1953 // by fast failing them.
1954 } else if (isa<AllocaInst>(I) || isa<TerminatorInst>(I)) {
1955 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1959 // Collapse PHI nodes
1960 if (PHINode* p = dyn_cast<PHINode>(I)) {
1961 Value *constVal = CollapsePhi(p);
1964 p->replaceAllUsesWith(constVal);
1965 if (MD && constVal->getType()->isPointerTy())
1966 MD->invalidateCachedPointerInfo(constVal);
1969 toErase.push_back(p);
1971 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1974 // If the number we were assigned was a brand new VN, then we don't
1975 // need to do a lookup to see if the number already exists
1976 // somewhere in the domtree: it can't!
1977 } else if (Num == NextNum) {
1978 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1980 // Perform fast-path value-number based elimination of values inherited from
1982 } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
1985 I->replaceAllUsesWith(repl);
1986 if (MD && repl->getType()->isPointerTy())
1987 MD->invalidateCachedPointerInfo(repl);
1988 toErase.push_back(I);
1992 localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
1998 /// runOnFunction - This is the main transformation entry point for a function.
1999 bool GVN::runOnFunction(Function& F) {
2001 MD = &getAnalysis<MemoryDependenceAnalysis>();
2002 DT = &getAnalysis<DominatorTree>();
2003 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2007 bool Changed = false;
2008 bool ShouldContinue = true;
2010 // Merge unconditional branches, allowing PRE to catch more
2011 // optimization opportunities.
2012 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2013 BasicBlock *BB = FI;
2015 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2016 if (removedBlock) ++NumGVNBlocks;
2018 Changed |= removedBlock;
2021 unsigned Iteration = 0;
2023 while (ShouldContinue) {
2024 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2025 ShouldContinue = iterateOnFunction(F);
2026 if (splitCriticalEdges())
2027 ShouldContinue = true;
2028 Changed |= ShouldContinue;
2033 bool PREChanged = true;
2034 while (PREChanged) {
2035 PREChanged = performPRE(F);
2036 Changed |= PREChanged;
2039 // FIXME: Should perform GVN again after PRE does something. PRE can move
2040 // computations into blocks where they become fully redundant. Note that
2041 // we can't do this until PRE's critical edge splitting updates memdep.
2042 // Actually, when this happens, we should just fully integrate PRE into GVN.
2044 cleanupGlobalSets();
2050 bool GVN::processBlock(BasicBlock *BB) {
2051 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
2052 // incrementing BI before processing an instruction).
2053 SmallVector<Instruction*, 8> toErase;
2054 bool ChangedFunction = false;
2056 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2058 ChangedFunction |= processInstruction(BI, toErase);
2059 if (toErase.empty()) {
2064 // If we need some instructions deleted, do it now.
2065 NumGVNInstr += toErase.size();
2067 // Avoid iterator invalidation.
2068 bool AtStart = BI == BB->begin();
2072 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
2073 E = toErase.end(); I != E; ++I) {
2074 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2075 if (MD) MD->removeInstruction(*I);
2076 (*I)->eraseFromParent();
2077 DEBUG(verifyRemoved(*I));
2087 return ChangedFunction;
2090 /// performPRE - Perform a purely local form of PRE that looks for diamond
2091 /// control flow patterns and attempts to perform simple PRE at the join point.
2092 bool GVN::performPRE(Function &F) {
2093 bool Changed = false;
2094 DenseMap<BasicBlock*, Value*> predMap;
2095 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2096 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2097 BasicBlock *CurrentBlock = *DI;
2099 // Nothing to PRE in the entry block.
2100 if (CurrentBlock == &F.getEntryBlock()) continue;
2102 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2103 BE = CurrentBlock->end(); BI != BE; ) {
2104 Instruction *CurInst = BI++;
2106 if (isa<AllocaInst>(CurInst) ||
2107 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2108 CurInst->getType()->isVoidTy() ||
2109 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2110 isa<DbgInfoIntrinsic>(CurInst))
2113 // We don't currently value number ANY inline asm calls.
2114 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2115 if (CallI->isInlineAsm())
2118 uint32_t ValNo = VN.lookup(CurInst);
2120 // Look for the predecessors for PRE opportunities. We're
2121 // only trying to solve the basic diamond case, where
2122 // a value is computed in the successor and one predecessor,
2123 // but not the other. We also explicitly disallow cases
2124 // where the successor is its own predecessor, because they're
2125 // more complicated to get right.
2126 unsigned NumWith = 0;
2127 unsigned NumWithout = 0;
2128 BasicBlock *PREPred = 0;
2131 for (pred_iterator PI = pred_begin(CurrentBlock),
2132 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2133 BasicBlock *P = *PI;
2134 // We're not interested in PRE where the block is its
2135 // own predecessor, or in blocks with predecessors
2136 // that are not reachable.
2137 if (P == CurrentBlock) {
2140 } else if (!localAvail.count(P)) {
2145 DenseMap<uint32_t, Value*>::iterator predV =
2146 localAvail[P]->table.find(ValNo);
2147 if (predV == localAvail[P]->table.end()) {
2150 } else if (predV->second == CurInst) {
2153 predMap[P] = predV->second;
2158 // Don't do PRE when it might increase code size, i.e. when
2159 // we would need to insert instructions in more than one pred.
2160 if (NumWithout != 1 || NumWith == 0)
2163 // Don't do PRE across indirect branch.
2164 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2167 // We can't do PRE safely on a critical edge, so instead we schedule
2168 // the edge to be split and perform the PRE the next time we iterate
2170 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2171 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2172 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2176 // Instantiate the expression in the predecessor that lacked it.
2177 // Because we are going top-down through the block, all value numbers
2178 // will be available in the predecessor by the time we need them. Any
2179 // that weren't originally present will have been instantiated earlier
2181 Instruction *PREInstr = CurInst->clone();
2182 bool success = true;
2183 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2184 Value *Op = PREInstr->getOperand(i);
2185 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2188 if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
2189 PREInstr->setOperand(i, V);
2196 // Fail out if we encounter an operand that is not available in
2197 // the PRE predecessor. This is typically because of loads which
2198 // are not value numbered precisely.
2201 DEBUG(verifyRemoved(PREInstr));
2205 PREInstr->insertBefore(PREPred->getTerminator());
2206 PREInstr->setName(CurInst->getName() + ".pre");
2207 predMap[PREPred] = PREInstr;
2208 VN.add(PREInstr, ValNo);
2211 // Update the availability map to include the new instruction.
2212 localAvail[PREPred]->table.insert(std::make_pair(ValNo, PREInstr));
2214 // Create a PHI to make the value available in this block.
2215 PHINode* Phi = PHINode::Create(CurInst->getType(),
2216 CurInst->getName() + ".pre-phi",
2217 CurrentBlock->begin());
2218 for (pred_iterator PI = pred_begin(CurrentBlock),
2219 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2220 BasicBlock *P = *PI;
2221 Phi->addIncoming(predMap[P], P);
2225 localAvail[CurrentBlock]->table[ValNo] = Phi;
2227 CurInst->replaceAllUsesWith(Phi);
2228 if (MD && Phi->getType()->isPointerTy())
2229 MD->invalidateCachedPointerInfo(Phi);
2232 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2233 if (MD) MD->removeInstruction(CurInst);
2234 CurInst->eraseFromParent();
2235 DEBUG(verifyRemoved(CurInst));
2240 if (splitCriticalEdges())
2246 /// splitCriticalEdges - Split critical edges found during the previous
2247 /// iteration that may enable further optimization.
2248 bool GVN::splitCriticalEdges() {
2249 if (toSplit.empty())
2252 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2253 SplitCriticalEdge(Edge.first, Edge.second, this);
2254 } while (!toSplit.empty());
2255 if (MD) MD->invalidateCachedPredecessors();
2259 /// iterateOnFunction - Executes one iteration of GVN
2260 bool GVN::iterateOnFunction(Function &F) {
2261 cleanupGlobalSets();
2263 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2264 DE = df_end(DT->getRootNode()); DI != DE; ++DI) {
2266 localAvail[DI->getBlock()] =
2267 new ValueNumberScope(localAvail[DI->getIDom()->getBlock()]);
2269 localAvail[DI->getBlock()] = new ValueNumberScope(0);
2272 // Top-down walk of the dominator tree
2273 bool Changed = false;
2275 // Needed for value numbering with phi construction to work.
2276 ReversePostOrderTraversal<Function*> RPOT(&F);
2277 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2278 RE = RPOT.end(); RI != RE; ++RI)
2279 Changed |= processBlock(*RI);
2281 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2282 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2283 Changed |= processBlock(DI->getBlock());
2289 void GVN::cleanupGlobalSets() {
2292 for (DenseMap<BasicBlock*, ValueNumberScope*>::iterator
2293 I = localAvail.begin(), E = localAvail.end(); I != E; ++I)
2298 /// verifyRemoved - Verify that the specified instruction does not occur in our
2299 /// internal data structures.
2300 void GVN::verifyRemoved(const Instruction *Inst) const {
2301 VN.verifyRemoved(Inst);
2303 // Walk through the value number scope to make sure the instruction isn't
2304 // ferreted away in it.
2305 for (DenseMap<BasicBlock*, ValueNumberScope*>::const_iterator
2306 I = localAvail.begin(), E = localAvail.end(); I != E; ++I) {
2307 const ValueNumberScope *VNS = I->second;
2310 for (DenseMap<uint32_t, Value*>::const_iterator
2311 II = VNS->table.begin(), IE = VNS->table.end(); II != IE; ++II) {
2312 assert(II->second != Inst && "Inst still in value numbering scope!");