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 #include "llvm/Transforms/Scalar/GVN.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PointerIntPair.h"
24 #include "llvm/ADT/PostOrderIterator.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/SetVector.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/AliasAnalysis.h"
31 #include "llvm/Analysis/AssumptionCache.h"
32 #include "llvm/Analysis/CFG.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/InstructionSimplify.h"
35 #include "llvm/Analysis/LoopInfo.h"
36 #include "llvm/Analysis/MemoryBuiltins.h"
37 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
38 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
39 #include "llvm/Analysis/PHITransAddr.h"
40 #include "llvm/Analysis/TargetLibraryInfo.h"
41 #include "llvm/Analysis/ValueTracking.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DebugLoc.h"
50 #include "llvm/IR/DomTreeUpdater.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/PatternMatch.h"
64 #include "llvm/IR/Type.h"
65 #include "llvm/IR/Use.h"
66 #include "llvm/IR/Value.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/Casting.h"
69 #include "llvm/Support/CommandLine.h"
70 #include "llvm/Support/Compiler.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
74 #include "llvm/Transforms/Utils/Local.h"
75 #include "llvm/Transforms/Utils/SSAUpdater.h"
76 #include "llvm/Transforms/Utils/VNCoercion.h"
84 using namespace llvm::gvn;
85 using namespace llvm::VNCoercion;
86 using namespace PatternMatch;
88 #define DEBUG_TYPE "gvn"
90 STATISTIC(NumGVNInstr, "Number of instructions deleted");
91 STATISTIC(NumGVNLoad, "Number of loads deleted");
92 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
93 STATISTIC(NumGVNBlocks, "Number of blocks merged");
94 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
95 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
96 STATISTIC(NumPRELoad, "Number of loads PRE'd");
98 static cl::opt<bool> EnablePRE("enable-pre",
99 cl::init(true), cl::Hidden);
100 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
101 static cl::opt<bool> EnableMemDep("enable-gvn-memdep", cl::init(true));
103 // Maximum allowed recursion depth.
104 static cl::opt<uint32_t>
105 MaxRecurseDepth("gvn-max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
106 cl::desc("Max recurse depth in GVN (default = 1000)"));
108 static cl::opt<uint32_t> MaxNumDeps(
109 "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore,
110 cl::desc("Max number of dependences to attempt Load PRE (default = 100)"));
112 struct llvm::GVN::Expression {
115 bool commutative = false;
116 SmallVector<uint32_t, 4> varargs;
118 Expression(uint32_t o = ~2U) : opcode(o) {}
120 bool operator==(const Expression &other) const {
121 if (opcode != other.opcode)
123 if (opcode == ~0U || opcode == ~1U)
125 if (type != other.type)
127 if (varargs != other.varargs)
132 friend hash_code hash_value(const Expression &Value) {
134 Value.opcode, Value.type,
135 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
141 template <> struct DenseMapInfo<GVN::Expression> {
142 static inline GVN::Expression getEmptyKey() { return ~0U; }
143 static inline GVN::Expression getTombstoneKey() { return ~1U; }
145 static unsigned getHashValue(const GVN::Expression &e) {
146 using llvm::hash_value;
148 return static_cast<unsigned>(hash_value(e));
151 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
156 } // end namespace llvm
158 /// Represents a particular available value that we know how to materialize.
159 /// Materialization of an AvailableValue never fails. An AvailableValue is
160 /// implicitly associated with a rematerialization point which is the
161 /// location of the instruction from which it was formed.
162 struct llvm::gvn::AvailableValue {
164 SimpleVal, // A simple offsetted value that is accessed.
165 LoadVal, // A value produced by a load.
166 MemIntrin, // A memory intrinsic which is loaded from.
167 UndefVal // A UndefValue representing a value from dead block (which
168 // is not yet physically removed from the CFG).
171 /// V - The value that is live out of the block.
172 PointerIntPair<Value *, 2, ValType> Val;
174 /// Offset - The byte offset in Val that is interesting for the load query.
177 static AvailableValue get(Value *V, unsigned Offset = 0) {
179 Res.Val.setPointer(V);
180 Res.Val.setInt(SimpleVal);
185 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
187 Res.Val.setPointer(MI);
188 Res.Val.setInt(MemIntrin);
193 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
195 Res.Val.setPointer(LI);
196 Res.Val.setInt(LoadVal);
201 static AvailableValue getUndef() {
203 Res.Val.setPointer(nullptr);
204 Res.Val.setInt(UndefVal);
209 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
210 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
211 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
212 bool isUndefValue() const { return Val.getInt() == UndefVal; }
214 Value *getSimpleValue() const {
215 assert(isSimpleValue() && "Wrong accessor");
216 return Val.getPointer();
219 LoadInst *getCoercedLoadValue() const {
220 assert(isCoercedLoadValue() && "Wrong accessor");
221 return cast<LoadInst>(Val.getPointer());
224 MemIntrinsic *getMemIntrinValue() const {
225 assert(isMemIntrinValue() && "Wrong accessor");
226 return cast<MemIntrinsic>(Val.getPointer());
229 /// Emit code at the specified insertion point to adjust the value defined
230 /// here to the specified type. This handles various coercion cases.
231 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
235 /// Represents an AvailableValue which can be rematerialized at the end of
236 /// the associated BasicBlock.
237 struct llvm::gvn::AvailableValueInBlock {
238 /// BB - The basic block in question.
241 /// AV - The actual available value
244 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
245 AvailableValueInBlock Res;
247 Res.AV = std::move(AV);
251 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
252 unsigned Offset = 0) {
253 return get(BB, AvailableValue::get(V, Offset));
256 static AvailableValueInBlock getUndef(BasicBlock *BB) {
257 return get(BB, AvailableValue::getUndef());
260 /// Emit code at the end of this block to adjust the value defined here to
261 /// the specified type. This handles various coercion cases.
262 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
263 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
267 //===----------------------------------------------------------------------===//
268 // ValueTable Internal Functions
269 //===----------------------------------------------------------------------===//
271 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
273 e.type = I->getType();
274 e.opcode = I->getOpcode();
275 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
277 e.varargs.push_back(lookupOrAdd(*OI));
278 if (I->isCommutative()) {
279 // Ensure that commutative instructions that only differ by a permutation
280 // of their operands get the same value number by sorting the operand value
281 // numbers. Since all commutative instructions have two operands it is more
282 // efficient to sort by hand rather than using, say, std::sort.
283 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
284 if (e.varargs[0] > e.varargs[1])
285 std::swap(e.varargs[0], e.varargs[1]);
286 e.commutative = true;
289 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
290 // Sort the operand value numbers so x<y and y>x get the same value number.
291 CmpInst::Predicate Predicate = C->getPredicate();
292 if (e.varargs[0] > e.varargs[1]) {
293 std::swap(e.varargs[0], e.varargs[1]);
294 Predicate = CmpInst::getSwappedPredicate(Predicate);
296 e.opcode = (C->getOpcode() << 8) | Predicate;
297 e.commutative = true;
298 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
299 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
301 e.varargs.push_back(*II);
307 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
308 CmpInst::Predicate Predicate,
309 Value *LHS, Value *RHS) {
310 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
311 "Not a comparison!");
313 e.type = CmpInst::makeCmpResultType(LHS->getType());
314 e.varargs.push_back(lookupOrAdd(LHS));
315 e.varargs.push_back(lookupOrAdd(RHS));
317 // Sort the operand value numbers so x<y and y>x get the same value number.
318 if (e.varargs[0] > e.varargs[1]) {
319 std::swap(e.varargs[0], e.varargs[1]);
320 Predicate = CmpInst::getSwappedPredicate(Predicate);
322 e.opcode = (Opcode << 8) | Predicate;
323 e.commutative = true;
327 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
328 assert(EI && "Not an ExtractValueInst?");
330 e.type = EI->getType();
333 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
334 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
335 // EI might be an extract from one of our recognised intrinsics. If it
336 // is we'll synthesize a semantically equivalent expression instead on
337 // an extract value expression.
338 switch (I->getIntrinsicID()) {
339 case Intrinsic::sadd_with_overflow:
340 case Intrinsic::uadd_with_overflow:
341 e.opcode = Instruction::Add;
343 case Intrinsic::ssub_with_overflow:
344 case Intrinsic::usub_with_overflow:
345 e.opcode = Instruction::Sub;
347 case Intrinsic::smul_with_overflow:
348 case Intrinsic::umul_with_overflow:
349 e.opcode = Instruction::Mul;
356 // Intrinsic recognized. Grab its args to finish building the expression.
357 assert(I->getNumArgOperands() == 2 &&
358 "Expect two args for recognised intrinsics.");
359 e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
360 e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
365 // Not a recognised intrinsic. Fall back to producing an extract value
367 e.opcode = EI->getOpcode();
368 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
370 e.varargs.push_back(lookupOrAdd(*OI));
372 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
374 e.varargs.push_back(*II);
379 //===----------------------------------------------------------------------===//
380 // ValueTable External Functions
381 //===----------------------------------------------------------------------===//
383 GVN::ValueTable::ValueTable() = default;
384 GVN::ValueTable::ValueTable(const ValueTable &) = default;
385 GVN::ValueTable::ValueTable(ValueTable &&) = default;
386 GVN::ValueTable::~ValueTable() = default;
388 /// add - Insert a value into the table with a specified value number.
389 void GVN::ValueTable::add(Value *V, uint32_t num) {
390 valueNumbering.insert(std::make_pair(V, num));
391 if (PHINode *PN = dyn_cast<PHINode>(V))
392 NumberingPhi[num] = PN;
395 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
396 if (AA->doesNotAccessMemory(C)) {
397 Expression exp = createExpr(C);
398 uint32_t e = assignExpNewValueNum(exp).first;
399 valueNumbering[C] = e;
401 } else if (MD && AA->onlyReadsMemory(C)) {
402 Expression exp = createExpr(C);
403 auto ValNum = assignExpNewValueNum(exp);
405 valueNumbering[C] = ValNum.first;
409 MemDepResult local_dep = MD->getDependency(C);
411 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
412 valueNumbering[C] = nextValueNumber;
413 return nextValueNumber++;
416 if (local_dep.isDef()) {
417 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
419 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
420 valueNumbering[C] = nextValueNumber;
421 return nextValueNumber++;
424 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
425 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
426 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
428 valueNumbering[C] = nextValueNumber;
429 return nextValueNumber++;
433 uint32_t v = lookupOrAdd(local_cdep);
434 valueNumbering[C] = v;
439 const MemoryDependenceResults::NonLocalDepInfo &deps =
440 MD->getNonLocalCallDependency(C);
441 // FIXME: Move the checking logic to MemDep!
442 CallInst* cdep = nullptr;
444 // Check to see if we have a single dominating call instruction that is
446 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
447 const NonLocalDepEntry *I = &deps[i];
448 if (I->getResult().isNonLocal())
451 // We don't handle non-definitions. If we already have a call, reject
452 // instruction dependencies.
453 if (!I->getResult().isDef() || cdep != nullptr) {
458 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
459 // FIXME: All duplicated with non-local case.
460 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
461 cdep = NonLocalDepCall;
470 valueNumbering[C] = nextValueNumber;
471 return nextValueNumber++;
474 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
475 valueNumbering[C] = nextValueNumber;
476 return nextValueNumber++;
478 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
479 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
480 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
482 valueNumbering[C] = nextValueNumber;
483 return nextValueNumber++;
487 uint32_t v = lookupOrAdd(cdep);
488 valueNumbering[C] = v;
491 valueNumbering[C] = nextValueNumber;
492 return nextValueNumber++;
496 /// Returns true if a value number exists for the specified value.
497 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
499 /// lookup_or_add - Returns the value number for the specified value, assigning
500 /// it a new number if it did not have one before.
501 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
502 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
503 if (VI != valueNumbering.end())
506 if (!isa<Instruction>(V)) {
507 valueNumbering[V] = nextValueNumber;
508 return nextValueNumber++;
511 Instruction* I = cast<Instruction>(V);
513 switch (I->getOpcode()) {
514 case Instruction::Call:
515 return lookupOrAddCall(cast<CallInst>(I));
516 case Instruction::Add:
517 case Instruction::FAdd:
518 case Instruction::Sub:
519 case Instruction::FSub:
520 case Instruction::Mul:
521 case Instruction::FMul:
522 case Instruction::UDiv:
523 case Instruction::SDiv:
524 case Instruction::FDiv:
525 case Instruction::URem:
526 case Instruction::SRem:
527 case Instruction::FRem:
528 case Instruction::Shl:
529 case Instruction::LShr:
530 case Instruction::AShr:
531 case Instruction::And:
532 case Instruction::Or:
533 case Instruction::Xor:
534 case Instruction::ICmp:
535 case Instruction::FCmp:
536 case Instruction::Trunc:
537 case Instruction::ZExt:
538 case Instruction::SExt:
539 case Instruction::FPToUI:
540 case Instruction::FPToSI:
541 case Instruction::UIToFP:
542 case Instruction::SIToFP:
543 case Instruction::FPTrunc:
544 case Instruction::FPExt:
545 case Instruction::PtrToInt:
546 case Instruction::IntToPtr:
547 case Instruction::BitCast:
548 case Instruction::Select:
549 case Instruction::ExtractElement:
550 case Instruction::InsertElement:
551 case Instruction::ShuffleVector:
552 case Instruction::InsertValue:
553 case Instruction::GetElementPtr:
556 case Instruction::ExtractValue:
557 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
559 case Instruction::PHI:
560 valueNumbering[V] = nextValueNumber;
561 NumberingPhi[nextValueNumber] = cast<PHINode>(V);
562 return nextValueNumber++;
564 valueNumbering[V] = nextValueNumber;
565 return nextValueNumber++;
568 uint32_t e = assignExpNewValueNum(exp).first;
569 valueNumbering[V] = e;
573 /// Returns the value number of the specified value. Fails if
574 /// the value has not yet been numbered.
575 uint32_t GVN::ValueTable::lookup(Value *V, bool Verify) const {
576 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
578 assert(VI != valueNumbering.end() && "Value not numbered?");
581 return (VI != valueNumbering.end()) ? VI->second : 0;
584 /// Returns the value number of the given comparison,
585 /// assigning it a new number if it did not have one before. Useful when
586 /// we deduced the result of a comparison, but don't immediately have an
587 /// instruction realizing that comparison to hand.
588 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
589 CmpInst::Predicate Predicate,
590 Value *LHS, Value *RHS) {
591 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
592 return assignExpNewValueNum(exp).first;
595 /// Remove all entries from the ValueTable.
596 void GVN::ValueTable::clear() {
597 valueNumbering.clear();
598 expressionNumbering.clear();
599 NumberingPhi.clear();
600 PhiTranslateTable.clear();
607 /// Remove a value from the value numbering.
608 void GVN::ValueTable::erase(Value *V) {
609 uint32_t Num = valueNumbering.lookup(V);
610 valueNumbering.erase(V);
611 // If V is PHINode, V <--> value number is an one-to-one mapping.
613 NumberingPhi.erase(Num);
616 /// verifyRemoved - Verify that the value is removed from all internal data
618 void GVN::ValueTable::verifyRemoved(const Value *V) const {
619 for (DenseMap<Value*, uint32_t>::const_iterator
620 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
621 assert(I->first != V && "Inst still occurs in value numbering map!");
625 //===----------------------------------------------------------------------===//
627 //===----------------------------------------------------------------------===//
629 PreservedAnalyses GVN::run(Function &F, FunctionAnalysisManager &AM) {
630 // FIXME: The order of evaluation of these 'getResult' calls is very
631 // significant! Re-ordering these variables will cause GVN when run alone to
632 // be less effective! We should fix memdep and basic-aa to not exhibit this
633 // behavior, but until then don't change the order here.
634 auto &AC = AM.getResult<AssumptionAnalysis>(F);
635 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
636 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
637 auto &AA = AM.getResult<AAManager>(F);
638 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
639 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
640 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
641 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
643 return PreservedAnalyses::all();
644 PreservedAnalyses PA;
645 PA.preserve<DominatorTreeAnalysis>();
646 PA.preserve<GlobalsAA>();
647 PA.preserve<TargetLibraryAnalysis>();
651 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
652 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
654 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
655 E = d.end(); I != E; ++I) {
656 errs() << I->first << "\n";
663 /// Return true if we can prove that the value
664 /// we're analyzing is fully available in the specified block. As we go, keep
665 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
666 /// map is actually a tri-state map with the following values:
667 /// 0) we know the block *is not* fully available.
668 /// 1) we know the block *is* fully available.
669 /// 2) we do not know whether the block is fully available or not, but we are
670 /// currently speculating that it will be.
671 /// 3) we are speculating for this block and have used that to speculate for
673 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
674 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
675 uint32_t RecurseDepth) {
676 if (RecurseDepth > MaxRecurseDepth)
679 // Optimistically assume that the block is fully available and check to see
680 // if we already know about this block in one lookup.
681 std::pair<DenseMap<BasicBlock*, char>::iterator, bool> IV =
682 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
684 // If the entry already existed for this block, return the precomputed value.
686 // If this is a speculative "available" value, mark it as being used for
687 // speculation of other blocks.
688 if (IV.first->second == 2)
689 IV.first->second = 3;
690 return IV.first->second != 0;
693 // Otherwise, see if it is fully available in all predecessors.
694 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
696 // If this block has no predecessors, it isn't live-in here.
698 goto SpeculationFailure;
700 for (; PI != PE; ++PI)
701 // If the value isn't fully available in one of our predecessors, then it
702 // isn't fully available in this block either. Undo our previous
703 // optimistic assumption and bail out.
704 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
705 goto SpeculationFailure;
709 // If we get here, we found out that this is not, after
710 // all, a fully-available block. We have a problem if we speculated on this and
711 // used the speculation to mark other blocks as available.
713 char &BBVal = FullyAvailableBlocks[BB];
715 // If we didn't speculate on this, just return with it set to false.
721 // If we did speculate on this value, we could have blocks set to 1 that are
722 // incorrect. Walk the (transitive) successors of this block and mark them as
724 SmallVector<BasicBlock*, 32> BBWorklist;
725 BBWorklist.push_back(BB);
728 BasicBlock *Entry = BBWorklist.pop_back_val();
729 // Note that this sets blocks to 0 (unavailable) if they happen to not
730 // already be in FullyAvailableBlocks. This is safe.
731 char &EntryVal = FullyAvailableBlocks[Entry];
732 if (EntryVal == 0) continue; // Already unavailable.
734 // Mark as unavailable.
737 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
738 } while (!BBWorklist.empty());
743 /// Given a set of loads specified by ValuesPerBlock,
744 /// construct SSA form, allowing us to eliminate LI. This returns the value
745 /// that should be used at LI's definition site.
746 static Value *ConstructSSAForLoadSet(LoadInst *LI,
747 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
749 // Check for the fully redundant, dominating load case. In this case, we can
750 // just use the dominating value directly.
751 if (ValuesPerBlock.size() == 1 &&
752 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
754 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
755 "Dead BB dominate this block");
756 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
759 // Otherwise, we have to construct SSA form.
760 SmallVector<PHINode*, 8> NewPHIs;
761 SSAUpdater SSAUpdate(&NewPHIs);
762 SSAUpdate.Initialize(LI->getType(), LI->getName());
764 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
765 BasicBlock *BB = AV.BB;
767 if (SSAUpdate.HasValueForBlock(BB))
770 // If the value is the load that we will be eliminating, and the block it's
771 // available in is the block that the load is in, then don't add it as
772 // SSAUpdater will resolve the value to the relevant phi which may let it
773 // avoid phi construction entirely if there's actually only one value.
774 if (BB == LI->getParent() &&
775 ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == LI) ||
776 (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == LI)))
779 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
782 // Perform PHI construction.
783 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
786 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
787 Instruction *InsertPt,
790 Type *LoadTy = LI->getType();
791 const DataLayout &DL = LI->getModule()->getDataLayout();
792 if (isSimpleValue()) {
793 Res = getSimpleValue();
794 if (Res->getType() != LoadTy) {
795 Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
797 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset
798 << " " << *getSimpleValue() << '\n'
802 } else if (isCoercedLoadValue()) {
803 LoadInst *Load = getCoercedLoadValue();
804 if (Load->getType() == LoadTy && Offset == 0) {
807 Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
808 // We would like to use gvn.markInstructionForDeletion here, but we can't
809 // because the load is already memoized into the leader map table that GVN
810 // tracks. It is potentially possible to remove the load from the table,
811 // but then there all of the operations based on it would need to be
812 // rehashed. Just leave the dead load around.
813 gvn.getMemDep().removeInstruction(Load);
814 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset
815 << " " << *getCoercedLoadValue() << '\n'
819 } else if (isMemIntrinValue()) {
820 Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
822 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
823 << " " << *getMemIntrinValue() << '\n'
827 assert(isUndefValue() && "Should be UndefVal");
828 LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
829 return UndefValue::get(LoadTy);
831 assert(Res && "failed to materialize?");
835 static bool isLifetimeStart(const Instruction *Inst) {
836 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
837 return II->getIntrinsicID() == Intrinsic::lifetime_start;
841 /// Try to locate the three instruction involved in a missed
842 /// load-elimination case that is due to an intervening store.
843 static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo,
845 OptimizationRemarkEmitter *ORE) {
848 User *OtherAccess = nullptr;
850 OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
851 R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
854 for (auto *U : LI->getPointerOperand()->users())
855 if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
856 DT->dominates(cast<Instruction>(U), LI)) {
857 // FIXME: for now give up if there are multiple memory accesses that
858 // dominate the load. We need further analysis to decide which one is
859 // that we're forwarding from.
861 OtherAccess = nullptr;
867 R << " in favor of " << NV("OtherAccess", OtherAccess);
869 R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
874 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
875 Value *Address, AvailableValue &Res) {
876 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
877 "expected a local dependence");
878 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
880 const DataLayout &DL = LI->getModule()->getDataLayout();
882 if (DepInfo.isClobber()) {
883 // If the dependence is to a store that writes to a superset of the bits
884 // read by the load, we can extract the bits we need for the load from the
886 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
887 // Can't forward from non-atomic to atomic without violating memory model.
888 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
890 analyzeLoadFromClobberingStore(LI->getType(), Address, DepSI, DL);
892 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
898 // Check to see if we have something like this:
901 // if we have this, replace the later with an extraction from the former.
902 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
903 // If this is a clobber and L is the first instruction in its block, then
904 // we have the first instruction in the entry block.
905 // Can't forward from non-atomic to atomic without violating memory model.
906 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
908 analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
911 Res = AvailableValue::getLoad(DepLI, Offset);
917 // If the clobbering value is a memset/memcpy/memmove, see if we can
918 // forward a value on from it.
919 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
920 if (Address && !LI->isAtomic()) {
921 int Offset = analyzeLoadFromClobberingMemInst(LI->getType(), Address,
924 Res = AvailableValue::getMI(DepMI, Offset);
929 // Nothing known about this clobber, have to be conservative
931 // fast print dep, using operator<< on instruction is too slow.
932 dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
933 Instruction *I = DepInfo.getInst();
934 dbgs() << " is clobbered by " << *I << '\n';);
935 if (ORE->allowExtraAnalysis(DEBUG_TYPE))
936 reportMayClobberedLoad(LI, DepInfo, DT, ORE);
940 assert(DepInfo.isDef() && "follows from above");
942 Instruction *DepInst = DepInfo.getInst();
944 // Loading the allocation -> undef.
945 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
946 // Loading immediately after lifetime begin -> undef.
947 isLifetimeStart(DepInst)) {
948 Res = AvailableValue::get(UndefValue::get(LI->getType()));
952 // Loading from calloc (which zero initializes memory) -> zero
953 if (isCallocLikeFn(DepInst, TLI)) {
954 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
958 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
959 // Reject loads and stores that are to the same address but are of
960 // different types if we have to. If the stored value is larger or equal to
961 // the loaded value, we can reuse it.
962 if (S->getValueOperand()->getType() != LI->getType() &&
963 !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
967 // Can't forward from non-atomic to atomic without violating memory model.
968 if (S->isAtomic() < LI->isAtomic())
971 Res = AvailableValue::get(S->getValueOperand());
975 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
976 // If the types mismatch and we can't handle it, reject reuse of the load.
977 // If the stored value is larger or equal to the loaded value, we can reuse
979 if (LD->getType() != LI->getType() &&
980 !canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
983 // Can't forward from non-atomic to atomic without violating memory model.
984 if (LD->isAtomic() < LI->isAtomic())
987 Res = AvailableValue::getLoad(LD);
991 // Unknown def - must be conservative
993 // fast print dep, using operator<< on instruction is too slow.
994 dbgs() << "GVN: load "; LI->printAsOperand(dbgs());
995 dbgs() << " has unknown def " << *DepInst << '\n';);
999 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1000 AvailValInBlkVect &ValuesPerBlock,
1001 UnavailBlkVect &UnavailableBlocks) {
1002 // Filter out useless results (non-locals, etc). Keep track of the blocks
1003 // where we have a value available in repl, also keep track of whether we see
1004 // dependencies that produce an unknown value for the load (such as a call
1005 // that could potentially clobber the load).
1006 unsigned NumDeps = Deps.size();
1007 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1008 BasicBlock *DepBB = Deps[i].getBB();
1009 MemDepResult DepInfo = Deps[i].getResult();
1011 if (DeadBlocks.count(DepBB)) {
1012 // Dead dependent mem-op disguise as a load evaluating the same value
1013 // as the load in question.
1014 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1018 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1019 UnavailableBlocks.push_back(DepBB);
1023 // The address being loaded in this non-local block may not be the same as
1024 // the pointer operand of the load if PHI translation occurs. Make sure
1025 // to consider the right address.
1026 Value *Address = Deps[i].getAddress();
1029 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1030 // subtlety: because we know this was a non-local dependency, we know
1031 // it's safe to materialize anywhere between the instruction within
1032 // DepInfo and the end of it's block.
1033 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1036 UnavailableBlocks.push_back(DepBB);
1040 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1041 "post condition violation");
1044 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1045 UnavailBlkVect &UnavailableBlocks) {
1046 // Okay, we have *some* definitions of the value. This means that the value
1047 // is available in some of our (transitive) predecessors. Lets think about
1048 // doing PRE of this load. This will involve inserting a new load into the
1049 // predecessor when it's not available. We could do this in general, but
1050 // prefer to not increase code size. As such, we only do this when we know
1051 // that we only have to insert *one* load (which means we're basically moving
1052 // the load, not inserting a new one).
1054 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1055 UnavailableBlocks.end());
1057 // Let's find the first basic block with more than one predecessor. Walk
1058 // backwards through predecessors if needed.
1059 BasicBlock *LoadBB = LI->getParent();
1060 BasicBlock *TmpBB = LoadBB;
1061 bool IsSafeToSpeculativelyExecute = isSafeToSpeculativelyExecute(LI);
1063 // Check that there is no implicit control flow instructions above our load in
1064 // its block. If there is an instruction that doesn't always pass the
1065 // execution to the following instruction, then moving through it may become
1066 // invalid. For example:
1071 // guard(0 <= index && index < LEN);
1074 // It is illegal to move the array access to any point above the guard,
1075 // because if the index is out of bounds we should deoptimize rather than
1076 // access the array.
1077 // Check that there is no guard in this block above our instruction.
1078 if (!IsSafeToSpeculativelyExecute && ICF->isDominatedByICFIFromSameBlock(LI))
1080 while (TmpBB->getSinglePredecessor()) {
1081 TmpBB = TmpBB->getSinglePredecessor();
1082 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1084 if (Blockers.count(TmpBB))
1087 // If any of these blocks has more than one successor (i.e. if the edge we
1088 // just traversed was critical), then there are other paths through this
1089 // block along which the load may not be anticipated. Hoisting the load
1090 // above this block would be adding the load to execution paths along
1091 // which it was not previously executed.
1092 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1095 // Check that there is no implicit control flow in a block above.
1096 if (!IsSafeToSpeculativelyExecute && ICF->hasICF(TmpBB))
1103 // Check to see how many predecessors have the loaded value fully
1105 MapVector<BasicBlock *, Value *> PredLoads;
1106 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1107 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1108 FullyAvailableBlocks[AV.BB] = true;
1109 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1110 FullyAvailableBlocks[UnavailableBB] = false;
1112 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1113 for (BasicBlock *Pred : predecessors(LoadBB)) {
1114 // If any predecessor block is an EH pad that does not allow non-PHI
1115 // instructions before the terminator, we can't PRE the load.
1116 if (Pred->getTerminator()->isEHPad()) {
1118 dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1119 << Pred->getName() << "': " << *LI << '\n');
1123 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1127 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1128 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1130 dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1131 << Pred->getName() << "': " << *LI << '\n');
1135 if (LoadBB->isEHPad()) {
1137 dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1138 << Pred->getName() << "': " << *LI << '\n');
1142 CriticalEdgePred.push_back(Pred);
1144 // Only add the predecessors that will not be split for now.
1145 PredLoads[Pred] = nullptr;
1149 // Decide whether PRE is profitable for this load.
1150 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1151 assert(NumUnavailablePreds != 0 &&
1152 "Fully available value should already be eliminated!");
1154 // If this load is unavailable in multiple predecessors, reject it.
1155 // FIXME: If we could restructure the CFG, we could make a common pred with
1156 // all the preds that don't have an available LI and insert a new load into
1158 if (NumUnavailablePreds != 1)
1161 // Split critical edges, and update the unavailable predecessors accordingly.
1162 for (BasicBlock *OrigPred : CriticalEdgePred) {
1163 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1164 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1165 PredLoads[NewPred] = nullptr;
1166 LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1167 << LoadBB->getName() << '\n');
1170 // Check if the load can safely be moved to all the unavailable predecessors.
1171 bool CanDoPRE = true;
1172 const DataLayout &DL = LI->getModule()->getDataLayout();
1173 SmallVector<Instruction*, 8> NewInsts;
1174 for (auto &PredLoad : PredLoads) {
1175 BasicBlock *UnavailablePred = PredLoad.first;
1177 // Do PHI translation to get its value in the predecessor if necessary. The
1178 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1180 // If all preds have a single successor, then we know it is safe to insert
1181 // the load on the pred (?!?), so we can insert code to materialize the
1182 // pointer if it is not available.
1183 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1184 Value *LoadPtr = nullptr;
1185 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1188 // If we couldn't find or insert a computation of this phi translated value,
1191 LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1192 << *LI->getPointerOperand() << "\n");
1197 PredLoad.second = LoadPtr;
1201 while (!NewInsts.empty()) {
1202 Instruction *I = NewInsts.pop_back_val();
1203 markInstructionForDeletion(I);
1205 // HINT: Don't revert the edge-splitting as following transformation may
1206 // also need to split these critical edges.
1207 return !CriticalEdgePred.empty();
1210 // Okay, we can eliminate this load by inserting a reload in the predecessor
1211 // and using PHI construction to get the value in the other predecessors, do
1213 LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1214 LLVM_DEBUG(if (!NewInsts.empty()) dbgs()
1215 << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts.back()
1218 // Assign value numbers to the new instructions.
1219 for (Instruction *I : NewInsts) {
1220 // Instructions that have been inserted in predecessor(s) to materialize
1221 // the load address do not retain their original debug locations. Doing
1222 // so could lead to confusing (but correct) source attributions.
1223 // FIXME: How do we retain source locations without causing poor debugging
1225 I->setDebugLoc(DebugLoc());
1227 // FIXME: We really _ought_ to insert these value numbers into their
1228 // parent's availability map. However, in doing so, we risk getting into
1229 // ordering issues. If a block hasn't been processed yet, we would be
1230 // marking a value as AVAIL-IN, which isn't what we intend.
1234 for (const auto &PredLoad : PredLoads) {
1235 BasicBlock *UnavailablePred = PredLoad.first;
1236 Value *LoadPtr = PredLoad.second;
1238 auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1239 LI->isVolatile(), LI->getAlignment(),
1240 LI->getOrdering(), LI->getSyncScopeID(),
1241 UnavailablePred->getTerminator());
1242 NewLoad->setDebugLoc(LI->getDebugLoc());
1244 // Transfer the old load's AA tags to the new load.
1246 LI->getAAMetadata(Tags);
1248 NewLoad->setAAMetadata(Tags);
1250 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1251 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1252 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1253 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1254 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1255 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1257 // We do not propagate the old load's debug location, because the new
1258 // load now lives in a different BB, and we want to avoid a jumpy line
1260 // FIXME: How do we retain source locations without causing poor debugging
1263 // Add the newly created load.
1264 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1266 MD->invalidateCachedPointerInfo(LoadPtr);
1267 LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1270 // Perform PHI construction.
1271 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1272 LI->replaceAllUsesWith(V);
1273 if (isa<PHINode>(V))
1275 if (Instruction *I = dyn_cast<Instruction>(V))
1276 I->setDebugLoc(LI->getDebugLoc());
1277 if (V->getType()->isPtrOrPtrVectorTy())
1278 MD->invalidateCachedPointerInfo(V);
1279 markInstructionForDeletion(LI);
1281 return OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1282 << "load eliminated by PRE";
1288 static void reportLoadElim(LoadInst *LI, Value *AvailableValue,
1289 OptimizationRemarkEmitter *ORE) {
1290 using namespace ore;
1293 return OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1294 << "load of type " << NV("Type", LI->getType()) << " eliminated"
1295 << setExtraArgs() << " in favor of "
1296 << NV("InfavorOfValue", AvailableValue);
1300 /// Attempt to eliminate a load whose dependencies are
1301 /// non-local by performing PHI construction.
1302 bool GVN::processNonLocalLoad(LoadInst *LI) {
1303 // non-local speculations are not allowed under asan.
1304 if (LI->getParent()->getParent()->hasFnAttribute(
1305 Attribute::SanitizeAddress) ||
1306 LI->getParent()->getParent()->hasFnAttribute(
1307 Attribute::SanitizeHWAddress))
1310 // Step 1: Find the non-local dependencies of the load.
1312 MD->getNonLocalPointerDependency(LI, Deps);
1314 // If we had to process more than one hundred blocks to find the
1315 // dependencies, this load isn't worth worrying about. Optimizing
1316 // it will be too expensive.
1317 unsigned NumDeps = Deps.size();
1318 if (NumDeps > MaxNumDeps)
1321 // If we had a phi translation failure, we'll have a single entry which is a
1322 // clobber in the current block. Reject this early.
1324 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1325 LLVM_DEBUG(dbgs() << "GVN: non-local load "; LI->printAsOperand(dbgs());
1326 dbgs() << " has unknown dependencies\n";);
1330 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1331 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1332 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1333 OE = GEP->idx_end();
1335 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1336 performScalarPRE(I);
1339 // Step 2: Analyze the availability of the load
1340 AvailValInBlkVect ValuesPerBlock;
1341 UnavailBlkVect UnavailableBlocks;
1342 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1344 // If we have no predecessors that produce a known value for this load, exit
1346 if (ValuesPerBlock.empty())
1349 // Step 3: Eliminate fully redundancy.
1351 // If all of the instructions we depend on produce a known value for this
1352 // load, then it is fully redundant and we can use PHI insertion to compute
1353 // its value. Insert PHIs and remove the fully redundant value now.
1354 if (UnavailableBlocks.empty()) {
1355 LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1357 // Perform PHI construction.
1358 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1359 LI->replaceAllUsesWith(V);
1361 if (isa<PHINode>(V))
1363 if (Instruction *I = dyn_cast<Instruction>(V))
1364 // If instruction I has debug info, then we should not update it.
1365 // Also, if I has a null DebugLoc, then it is still potentially incorrect
1366 // to propagate LI's DebugLoc because LI may not post-dominate I.
1367 if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1368 I->setDebugLoc(LI->getDebugLoc());
1369 if (V->getType()->isPtrOrPtrVectorTy())
1370 MD->invalidateCachedPointerInfo(V);
1371 markInstructionForDeletion(LI);
1373 reportLoadElim(LI, V, ORE);
1377 // Step 4: Eliminate partial redundancy.
1378 if (!EnablePRE || !EnableLoadPRE)
1381 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1384 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1385 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1386 "This function can only be called with llvm.assume intrinsic");
1387 Value *V = IntrinsicI->getArgOperand(0);
1389 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1390 if (Cond->isZero()) {
1391 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1392 // Insert a new store to null instruction before the load to indicate that
1393 // this code is not reachable. FIXME: We could insert unreachable
1394 // instruction directly because we can modify the CFG.
1395 new StoreInst(UndefValue::get(Int8Ty),
1396 Constant::getNullValue(Int8Ty->getPointerTo()),
1399 markInstructionForDeletion(IntrinsicI);
1401 } else if (isa<Constant>(V)) {
1402 // If it's not false, and constant, it must evaluate to true. This means our
1403 // assume is assume(true), and thus, pointless, and we don't want to do
1404 // anything more here.
1408 Constant *True = ConstantInt::getTrue(V->getContext());
1409 bool Changed = false;
1411 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1412 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1414 // This property is only true in dominated successors, propagateEquality
1415 // will check dominance for us.
1416 Changed |= propagateEquality(V, True, Edge, false);
1419 // We can replace assume value with true, which covers cases like this:
1420 // call void @llvm.assume(i1 %cmp)
1421 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1422 ReplaceWithConstMap[V] = True;
1424 // If one of *cmp *eq operand is const, adding it to map will cover this:
1425 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1426 // call void @llvm.assume(i1 %cmp)
1427 // ret float %0 ; will change it to ret float 3.000000e+00
1428 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1429 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1430 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1431 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1432 CmpI->getFastMathFlags().noNaNs())) {
1433 Value *CmpLHS = CmpI->getOperand(0);
1434 Value *CmpRHS = CmpI->getOperand(1);
1435 if (isa<Constant>(CmpLHS))
1436 std::swap(CmpLHS, CmpRHS);
1437 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1439 // If only one operand is constant.
1440 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1441 ReplaceWithConstMap[CmpLHS] = RHSConst;
1447 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1448 patchReplacementInstruction(I, Repl);
1449 I->replaceAllUsesWith(Repl);
1452 /// Attempt to eliminate a load, first by eliminating it
1453 /// locally, and then attempting non-local elimination if that fails.
1454 bool GVN::processLoad(LoadInst *L) {
1458 // This code hasn't been audited for ordered or volatile memory access
1459 if (!L->isUnordered())
1462 if (L->use_empty()) {
1463 markInstructionForDeletion(L);
1467 // ... to a pointer that has been loaded from before...
1468 MemDepResult Dep = MD->getDependency(L);
1470 // If it is defined in another block, try harder.
1471 if (Dep.isNonLocal())
1472 return processNonLocalLoad(L);
1474 // Only handle the local case below
1475 if (!Dep.isDef() && !Dep.isClobber()) {
1476 // This might be a NonFuncLocal or an Unknown
1478 // fast print dep, using operator<< on instruction is too slow.
1479 dbgs() << "GVN: load "; L->printAsOperand(dbgs());
1480 dbgs() << " has unknown dependence\n";);
1485 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1486 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1488 // Replace the load!
1489 patchAndReplaceAllUsesWith(L, AvailableValue);
1490 markInstructionForDeletion(L);
1492 reportLoadElim(L, AvailableValue, ORE);
1493 // Tell MDA to rexamine the reused pointer since we might have more
1494 // information after forwarding it.
1495 if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1496 MD->invalidateCachedPointerInfo(AvailableValue);
1503 /// Return a pair the first field showing the value number of \p Exp and the
1504 /// second field showing whether it is a value number newly created.
1505 std::pair<uint32_t, bool>
1506 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1507 uint32_t &e = expressionNumbering[Exp];
1508 bool CreateNewValNum = !e;
1509 if (CreateNewValNum) {
1510 Expressions.push_back(Exp);
1511 if (ExprIdx.size() < nextValueNumber + 1)
1512 ExprIdx.resize(nextValueNumber * 2);
1513 e = nextValueNumber;
1514 ExprIdx[nextValueNumber++] = nextExprNumber++;
1516 return {e, CreateNewValNum};
1519 /// Return whether all the values related with the same \p num are
1520 /// defined in \p BB.
1521 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1523 LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1524 while (Vals && Vals->BB == BB)
1529 /// Wrap phiTranslateImpl to provide caching functionality.
1530 uint32_t GVN::ValueTable::phiTranslate(const BasicBlock *Pred,
1531 const BasicBlock *PhiBlock, uint32_t Num,
1533 auto FindRes = PhiTranslateTable.find({Num, Pred});
1534 if (FindRes != PhiTranslateTable.end())
1535 return FindRes->second;
1536 uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1537 PhiTranslateTable.insert({{Num, Pred}, NewNum});
1541 /// Translate value number \p Num using phis, so that it has the values of
1543 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1544 const BasicBlock *PhiBlock,
1545 uint32_t Num, GVN &Gvn) {
1546 if (PHINode *PN = NumberingPhi[Num]) {
1547 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1548 if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1549 if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1555 // If there is any value related with Num is defined in a BB other than
1556 // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1557 // a backedge. We can do an early exit in that case to save compile time.
1558 if (!areAllValsInBB(Num, PhiBlock, Gvn))
1561 if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1563 Expression Exp = Expressions[ExprIdx[Num]];
1565 for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1566 // For InsertValue and ExtractValue, some varargs are index numbers
1567 // instead of value numbers. Those index numbers should not be
1569 if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1570 (i > 0 && Exp.opcode == Instruction::ExtractValue))
1572 Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1575 if (Exp.commutative) {
1576 assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1577 if (Exp.varargs[0] > Exp.varargs[1]) {
1578 std::swap(Exp.varargs[0], Exp.varargs[1]);
1579 uint32_t Opcode = Exp.opcode >> 8;
1580 if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1581 Exp.opcode = (Opcode << 8) |
1582 CmpInst::getSwappedPredicate(
1583 static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1587 if (uint32_t NewNum = expressionNumbering[Exp])
1592 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1594 void GVN::ValueTable::eraseTranslateCacheEntry(uint32_t Num,
1595 const BasicBlock &CurrBlock) {
1596 for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1597 auto FindRes = PhiTranslateTable.find({Num, Pred});
1598 if (FindRes != PhiTranslateTable.end())
1599 PhiTranslateTable.erase(FindRes);
1603 // In order to find a leader for a given value number at a
1604 // specific basic block, we first obtain the list of all Values for that number,
1605 // and then scan the list to find one whose block dominates the block in
1606 // question. This is fast because dominator tree queries consist of only
1607 // a few comparisons of DFS numbers.
1608 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1609 LeaderTableEntry Vals = LeaderTable[num];
1610 if (!Vals.Val) return nullptr;
1612 Value *Val = nullptr;
1613 if (DT->dominates(Vals.BB, BB)) {
1615 if (isa<Constant>(Val)) return Val;
1618 LeaderTableEntry* Next = Vals.Next;
1620 if (DT->dominates(Next->BB, BB)) {
1621 if (isa<Constant>(Next->Val)) return Next->Val;
1622 if (!Val) Val = Next->Val;
1631 /// There is an edge from 'Src' to 'Dst'. Return
1632 /// true if every path from the entry block to 'Dst' passes via this edge. In
1633 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1634 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1635 DominatorTree *DT) {
1636 // While in theory it is interesting to consider the case in which Dst has
1637 // more than one predecessor, because Dst might be part of a loop which is
1638 // only reachable from Src, in practice it is pointless since at the time
1639 // GVN runs all such loops have preheaders, which means that Dst will have
1640 // been changed to have only one predecessor, namely Src.
1641 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1642 assert((!Pred || Pred == E.getStart()) &&
1643 "No edge between these basic blocks!");
1644 return Pred != nullptr;
1647 void GVN::assignBlockRPONumber(Function &F) {
1648 BlockRPONumber.clear();
1649 uint32_t NextBlockNumber = 1;
1650 ReversePostOrderTraversal<Function *> RPOT(&F);
1651 for (BasicBlock *BB : RPOT)
1652 BlockRPONumber[BB] = NextBlockNumber++;
1653 InvalidBlockRPONumbers = false;
1656 // Tries to replace instruction with const, using information from
1657 // ReplaceWithConstMap.
1658 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1659 bool Changed = false;
1660 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1661 Value *Operand = Instr->getOperand(OpNum);
1662 auto it = ReplaceWithConstMap.find(Operand);
1663 if (it != ReplaceWithConstMap.end()) {
1664 assert(!isa<Constant>(Operand) &&
1665 "Replacing constants with constants is invalid");
1666 LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "
1667 << *it->second << " in instruction " << *Instr << '\n');
1668 Instr->setOperand(OpNum, it->second);
1675 /// The given values are known to be equal in every block
1676 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1677 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1678 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1679 /// value starting from the end of Root.Start.
1680 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1681 bool DominatesByEdge) {
1682 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1683 Worklist.push_back(std::make_pair(LHS, RHS));
1684 bool Changed = false;
1685 // For speed, compute a conservative fast approximation to
1686 // DT->dominates(Root, Root.getEnd());
1687 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1689 while (!Worklist.empty()) {
1690 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1691 LHS = Item.first; RHS = Item.second;
1695 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1697 // Don't try to propagate equalities between constants.
1698 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1701 // Prefer a constant on the right-hand side, or an Argument if no constants.
1702 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1703 std::swap(LHS, RHS);
1704 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1706 // If there is no obvious reason to prefer the left-hand side over the
1707 // right-hand side, ensure the longest lived term is on the right-hand side,
1708 // so the shortest lived term will be replaced by the longest lived.
1709 // This tends to expose more simplifications.
1710 uint32_t LVN = VN.lookupOrAdd(LHS);
1711 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1712 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1713 // Move the 'oldest' value to the right-hand side, using the value number
1714 // as a proxy for age.
1715 uint32_t RVN = VN.lookupOrAdd(RHS);
1717 std::swap(LHS, RHS);
1722 // If value numbering later sees that an instruction in the scope is equal
1723 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1724 // the invariant that instructions only occur in the leader table for their
1725 // own value number (this is used by removeFromLeaderTable), do not do this
1726 // if RHS is an instruction (if an instruction in the scope is morphed into
1727 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1728 // using the leader table is about compiling faster, not optimizing better).
1729 // The leader table only tracks basic blocks, not edges. Only add to if we
1730 // have the simple case where the edge dominates the end.
1731 if (RootDominatesEnd && !isa<Instruction>(RHS))
1732 addToLeaderTable(LVN, RHS, Root.getEnd());
1734 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1735 // LHS always has at least one use that is not dominated by Root, this will
1736 // never do anything if LHS has only one use.
1737 if (!LHS->hasOneUse()) {
1738 unsigned NumReplacements =
1740 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1741 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1743 Changed |= NumReplacements > 0;
1744 NumGVNEqProp += NumReplacements;
1745 // Cached information for anything that uses LHS will be invalid.
1747 MD->invalidateCachedPointerInfo(LHS);
1750 // Now try to deduce additional equalities from this one. For example, if
1751 // the known equality was "(A != B)" == "false" then it follows that A and B
1752 // are equal in the scope. Only boolean equalities with an explicit true or
1753 // false RHS are currently supported.
1754 if (!RHS->getType()->isIntegerTy(1))
1755 // Not a boolean equality - bail out.
1757 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1759 // RHS neither 'true' nor 'false' - bail out.
1761 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1762 bool isKnownTrue = CI->isMinusOne();
1763 bool isKnownFalse = !isKnownTrue;
1765 // If "A && B" is known true then both A and B are known true. If "A || B"
1766 // is known false then both A and B are known false.
1768 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1769 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1770 Worklist.push_back(std::make_pair(A, RHS));
1771 Worklist.push_back(std::make_pair(B, RHS));
1775 // If we are propagating an equality like "(A == B)" == "true" then also
1776 // propagate the equality A == B. When propagating a comparison such as
1777 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1778 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1779 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1781 // If "A == B" is known true, or "A != B" is known false, then replace
1782 // A with B everywhere in the scope.
1783 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1784 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1785 Worklist.push_back(std::make_pair(Op0, Op1));
1787 // Handle the floating point versions of equality comparisons too.
1788 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1789 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1791 // Floating point -0.0 and 0.0 compare equal, so we can only
1792 // propagate values if we know that we have a constant and that
1793 // its value is non-zero.
1795 // FIXME: We should do this optimization if 'no signed zeros' is
1796 // applicable via an instruction-level fast-math-flag or some other
1797 // indicator that relaxed FP semantics are being used.
1799 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1800 Worklist.push_back(std::make_pair(Op0, Op1));
1803 // If "A >= B" is known true, replace "A < B" with false everywhere.
1804 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1805 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1806 // Since we don't have the instruction "A < B" immediately to hand, work
1807 // out the value number that it would have and use that to find an
1808 // appropriate instruction (if any).
1809 uint32_t NextNum = VN.getNextUnusedValueNumber();
1810 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1811 // If the number we were assigned was brand new then there is no point in
1812 // looking for an instruction realizing it: there cannot be one!
1813 if (Num < NextNum) {
1814 Value *NotCmp = findLeader(Root.getEnd(), Num);
1815 if (NotCmp && isa<Instruction>(NotCmp)) {
1816 unsigned NumReplacements =
1818 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1819 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1821 Changed |= NumReplacements > 0;
1822 NumGVNEqProp += NumReplacements;
1823 // Cached information for anything that uses NotCmp will be invalid.
1825 MD->invalidateCachedPointerInfo(NotCmp);
1828 // Ensure that any instruction in scope that gets the "A < B" value number
1829 // is replaced with false.
1830 // The leader table only tracks basic blocks, not edges. Only add to if we
1831 // have the simple case where the edge dominates the end.
1832 if (RootDominatesEnd)
1833 addToLeaderTable(Num, NotVal, Root.getEnd());
1842 /// When calculating availability, handle an instruction
1843 /// by inserting it into the appropriate sets
1844 bool GVN::processInstruction(Instruction *I) {
1845 // Ignore dbg info intrinsics.
1846 if (isa<DbgInfoIntrinsic>(I))
1849 // If the instruction can be easily simplified then do so now in preference
1850 // to value numbering it. Value numbering often exposes redundancies, for
1851 // example if it determines that %y is equal to %x then the instruction
1852 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1853 const DataLayout &DL = I->getModule()->getDataLayout();
1854 if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1855 bool Changed = false;
1856 if (!I->use_empty()) {
1857 I->replaceAllUsesWith(V);
1860 if (isInstructionTriviallyDead(I, TLI)) {
1861 markInstructionForDeletion(I);
1865 if (MD && V->getType()->isPtrOrPtrVectorTy())
1866 MD->invalidateCachedPointerInfo(V);
1872 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1873 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1874 return processAssumeIntrinsic(IntrinsicI);
1876 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1877 if (processLoad(LI))
1880 unsigned Num = VN.lookupOrAdd(LI);
1881 addToLeaderTable(Num, LI, LI->getParent());
1885 // For conditional branches, we can perform simple conditional propagation on
1886 // the condition value itself.
1887 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1888 if (!BI->isConditional())
1891 if (isa<Constant>(BI->getCondition()))
1892 return processFoldableCondBr(BI);
1894 Value *BranchCond = BI->getCondition();
1895 BasicBlock *TrueSucc = BI->getSuccessor(0);
1896 BasicBlock *FalseSucc = BI->getSuccessor(1);
1897 // Avoid multiple edges early.
1898 if (TrueSucc == FalseSucc)
1901 BasicBlock *Parent = BI->getParent();
1902 bool Changed = false;
1904 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1905 BasicBlockEdge TrueE(Parent, TrueSucc);
1906 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1908 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1909 BasicBlockEdge FalseE(Parent, FalseSucc);
1910 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1915 // For switches, propagate the case values into the case destinations.
1916 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1917 Value *SwitchCond = SI->getCondition();
1918 BasicBlock *Parent = SI->getParent();
1919 bool Changed = false;
1921 // Remember how many outgoing edges there are to every successor.
1922 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
1923 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1924 ++SwitchEdges[SI->getSuccessor(i)];
1926 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1928 BasicBlock *Dst = i->getCaseSuccessor();
1929 // If there is only a single edge, propagate the case value into it.
1930 if (SwitchEdges.lookup(Dst) == 1) {
1931 BasicBlockEdge E(Parent, Dst);
1932 Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1938 // Instructions with void type don't return a value, so there's
1939 // no point in trying to find redundancies in them.
1940 if (I->getType()->isVoidTy())
1943 uint32_t NextNum = VN.getNextUnusedValueNumber();
1944 unsigned Num = VN.lookupOrAdd(I);
1946 // Allocations are always uniquely numbered, so we can save time and memory
1947 // by fast failing them.
1948 if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) {
1949 addToLeaderTable(Num, I, I->getParent());
1953 // If the number we were assigned was a brand new VN, then we don't
1954 // need to do a lookup to see if the number already exists
1955 // somewhere in the domtree: it can't!
1956 if (Num >= NextNum) {
1957 addToLeaderTable(Num, I, I->getParent());
1961 // Perform fast-path value-number based elimination of values inherited from
1963 Value *Repl = findLeader(I->getParent(), Num);
1965 // Failure, just remember this instance for future use.
1966 addToLeaderTable(Num, I, I->getParent());
1968 } else if (Repl == I) {
1969 // If I was the result of a shortcut PRE, it might already be in the table
1970 // and the best replacement for itself. Nothing to do.
1975 patchAndReplaceAllUsesWith(I, Repl);
1976 if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1977 MD->invalidateCachedPointerInfo(Repl);
1978 markInstructionForDeletion(I);
1982 /// runOnFunction - This is the main transformation entry point for a function.
1983 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1984 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1985 MemoryDependenceResults *RunMD, LoopInfo *LI,
1986 OptimizationRemarkEmitter *RunORE) {
1991 VN.setAliasAnalysis(&RunAA);
1993 ImplicitControlFlowTracking ImplicitCFT(DT);
1997 InvalidBlockRPONumbers = true;
1999 bool Changed = false;
2000 bool ShouldContinue = true;
2002 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
2003 // Merge unconditional branches, allowing PRE to catch more
2004 // optimization opportunities.
2005 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2006 BasicBlock *BB = &*FI++;
2008 bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, nullptr, MD);
2012 Changed |= removedBlock;
2015 unsigned Iteration = 0;
2016 while (ShouldContinue) {
2017 LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2018 ShouldContinue = iterateOnFunction(F);
2019 Changed |= ShouldContinue;
2024 // Fabricate val-num for dead-code in order to suppress assertion in
2026 assignValNumForDeadCode();
2027 bool PREChanged = true;
2028 while (PREChanged) {
2029 PREChanged = performPRE(F);
2030 Changed |= PREChanged;
2034 // FIXME: Should perform GVN again after PRE does something. PRE can move
2035 // computations into blocks where they become fully redundant. Note that
2036 // we can't do this until PRE's critical edge splitting updates memdep.
2037 // Actually, when this happens, we should just fully integrate PRE into GVN.
2039 cleanupGlobalSets();
2040 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2047 bool GVN::processBlock(BasicBlock *BB) {
2048 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2049 // (and incrementing BI before processing an instruction).
2050 assert(InstrsToErase.empty() &&
2051 "We expect InstrsToErase to be empty across iterations");
2052 if (DeadBlocks.count(BB))
2055 // Clearing map before every BB because it can be used only for single BB.
2056 ReplaceWithConstMap.clear();
2057 bool ChangedFunction = false;
2059 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2061 if (!ReplaceWithConstMap.empty())
2062 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2063 ChangedFunction |= processInstruction(&*BI);
2065 if (InstrsToErase.empty()) {
2070 // If we need some instructions deleted, do it now.
2071 NumGVNInstr += InstrsToErase.size();
2073 // Avoid iterator invalidation.
2074 bool AtStart = BI == BB->begin();
2078 for (auto *I : InstrsToErase) {
2079 assert(I->getParent() == BB && "Removing instruction from wrong block?");
2080 LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n');
2081 salvageDebugInfo(*I);
2082 if (MD) MD->removeInstruction(I);
2083 LLVM_DEBUG(verifyRemoved(I));
2084 ICF->removeInstruction(I);
2085 I->eraseFromParent();
2087 InstrsToErase.clear();
2095 return ChangedFunction;
2098 // Instantiate an expression in a predecessor that lacked it.
2099 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2100 BasicBlock *Curr, unsigned int ValNo) {
2101 // Because we are going top-down through the block, all value numbers
2102 // will be available in the predecessor by the time we need them. Any
2103 // that weren't originally present will have been instantiated earlier
2105 bool success = true;
2106 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2107 Value *Op = Instr->getOperand(i);
2108 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2110 // This could be a newly inserted instruction, in which case, we won't
2111 // find a value number, and should give up before we hurt ourselves.
2112 // FIXME: Rewrite the infrastructure to let it easier to value number
2113 // and process newly inserted instructions.
2114 if (!VN.exists(Op)) {
2119 VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2120 if (Value *V = findLeader(Pred, TValNo)) {
2121 Instr->setOperand(i, V);
2128 // Fail out if we encounter an operand that is not available in
2129 // the PRE predecessor. This is typically because of loads which
2130 // are not value numbered precisely.
2134 Instr->insertBefore(Pred->getTerminator());
2135 Instr->setName(Instr->getName() + ".pre");
2136 Instr->setDebugLoc(Instr->getDebugLoc());
2138 unsigned Num = VN.lookupOrAdd(Instr);
2141 // Update the availability map to include the new instruction.
2142 addToLeaderTable(Num, Instr, Pred);
2146 bool GVN::performScalarPRE(Instruction *CurInst) {
2147 if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() ||
2148 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2149 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2150 isa<DbgInfoIntrinsic>(CurInst))
2153 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2154 // sinking the compare again, and it would force the code generator to
2155 // move the i1 from processor flags or predicate registers into a general
2156 // purpose register.
2157 if (isa<CmpInst>(CurInst))
2160 // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from
2161 // sinking the addressing mode computation back to its uses. Extending the
2162 // GEP's live range increases the register pressure, and therefore it can
2163 // introduce unnecessary spills.
2165 // This doesn't prevent Load PRE. PHI translation will make the GEP available
2166 // to the load by moving it to the predecessor block if necessary.
2167 if (isa<GetElementPtrInst>(CurInst))
2170 // We don't currently value number ANY inline asm calls.
2171 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2172 if (CallI->isInlineAsm())
2175 uint32_t ValNo = VN.lookup(CurInst);
2177 // Look for the predecessors for PRE opportunities. We're
2178 // only trying to solve the basic diamond case, where
2179 // a value is computed in the successor and one predecessor,
2180 // but not the other. We also explicitly disallow cases
2181 // where the successor is its own predecessor, because they're
2182 // more complicated to get right.
2183 unsigned NumWith = 0;
2184 unsigned NumWithout = 0;
2185 BasicBlock *PREPred = nullptr;
2186 BasicBlock *CurrentBlock = CurInst->getParent();
2188 // Update the RPO numbers for this function.
2189 if (InvalidBlockRPONumbers)
2190 assignBlockRPONumber(*CurrentBlock->getParent());
2192 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2193 for (BasicBlock *P : predecessors(CurrentBlock)) {
2194 // We're not interested in PRE where blocks with predecessors that are
2196 if (!DT->isReachableFromEntry(P)) {
2200 // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2201 // when CurInst has operand defined in CurrentBlock (so it may be defined
2202 // by phi in the loop header).
2203 assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&
2204 "Invalid BlockRPONumber map.");
2205 if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2206 llvm::any_of(CurInst->operands(), [&](const Use &U) {
2207 if (auto *Inst = dyn_cast<Instruction>(U.get()))
2208 return Inst->getParent() == CurrentBlock;
2215 uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2216 Value *predV = findLeader(P, TValNo);
2218 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2221 } else if (predV == CurInst) {
2222 /* CurInst dominates this predecessor. */
2226 predMap.push_back(std::make_pair(predV, P));
2231 // Don't do PRE when it might increase code size, i.e. when
2232 // we would need to insert instructions in more than one pred.
2233 if (NumWithout > 1 || NumWith == 0)
2236 // We may have a case where all predecessors have the instruction,
2237 // and we just need to insert a phi node. Otherwise, perform
2239 Instruction *PREInstr = nullptr;
2241 if (NumWithout != 0) {
2242 if (!isSafeToSpeculativelyExecute(CurInst)) {
2243 // It is only valid to insert a new instruction if the current instruction
2244 // is always executed. An instruction with implicit control flow could
2245 // prevent us from doing it. If we cannot speculate the execution, then
2246 // PRE should be prohibited.
2247 if (ICF->isDominatedByICFIFromSameBlock(CurInst))
2251 // Don't do PRE across indirect branch.
2252 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2255 // We can't do PRE safely on a critical edge, so instead we schedule
2256 // the edge to be split and perform the PRE the next time we iterate
2258 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2259 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2260 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2263 // We need to insert somewhere, so let's give it a shot
2264 PREInstr = CurInst->clone();
2265 if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2266 // If we failed insertion, make sure we remove the instruction.
2267 LLVM_DEBUG(verifyRemoved(PREInstr));
2268 PREInstr->deleteValue();
2273 // Either we should have filled in the PRE instruction, or we should
2274 // not have needed insertions.
2275 assert(PREInstr != nullptr || NumWithout == 0);
2279 // Create a PHI to make the value available in this block.
2281 PHINode::Create(CurInst->getType(), predMap.size(),
2282 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2283 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2284 if (Value *V = predMap[i].first) {
2285 // If we use an existing value in this phi, we have to patch the original
2286 // value because the phi will be used to replace a later value.
2287 patchReplacementInstruction(CurInst, V);
2288 Phi->addIncoming(V, predMap[i].second);
2290 Phi->addIncoming(PREInstr, PREPred);
2294 // After creating a new PHI for ValNo, the phi translate result for ValNo will
2295 // be changed, so erase the related stale entries in phi translate cache.
2296 VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2297 addToLeaderTable(ValNo, Phi, CurrentBlock);
2298 Phi->setDebugLoc(CurInst->getDebugLoc());
2299 CurInst->replaceAllUsesWith(Phi);
2300 if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2301 MD->invalidateCachedPointerInfo(Phi);
2303 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2305 LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2307 MD->removeInstruction(CurInst);
2308 LLVM_DEBUG(verifyRemoved(CurInst));
2309 // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes
2310 // some assertion failures.
2311 ICF->removeInstruction(CurInst);
2312 CurInst->eraseFromParent();
2318 /// Perform a purely local form of PRE that looks for diamond
2319 /// control flow patterns and attempts to perform simple PRE at the join point.
2320 bool GVN::performPRE(Function &F) {
2321 bool Changed = false;
2322 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2323 // Nothing to PRE in the entry block.
2324 if (CurrentBlock == &F.getEntryBlock())
2327 // Don't perform PRE on an EH pad.
2328 if (CurrentBlock->isEHPad())
2331 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2332 BE = CurrentBlock->end();
2334 Instruction *CurInst = &*BI++;
2335 Changed |= performScalarPRE(CurInst);
2339 if (splitCriticalEdges())
2345 /// Split the critical edge connecting the given two blocks, and return
2346 /// the block inserted to the critical edge.
2347 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2349 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2351 MD->invalidateCachedPredecessors();
2352 InvalidBlockRPONumbers = true;
2356 /// Split critical edges found during the previous
2357 /// iteration that may enable further optimization.
2358 bool GVN::splitCriticalEdges() {
2359 if (toSplit.empty())
2362 std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val();
2363 SplitCriticalEdge(Edge.first, Edge.second,
2364 CriticalEdgeSplittingOptions(DT));
2365 } while (!toSplit.empty());
2366 if (MD) MD->invalidateCachedPredecessors();
2367 InvalidBlockRPONumbers = true;
2371 /// Executes one iteration of GVN
2372 bool GVN::iterateOnFunction(Function &F) {
2373 cleanupGlobalSets();
2375 // Top-down walk of the dominator tree
2376 bool Changed = false;
2377 // Needed for value numbering with phi construction to work.
2378 // RPOT walks the graph in its constructor and will not be invalidated during
2380 ReversePostOrderTraversal<Function *> RPOT(&F);
2382 for (BasicBlock *BB : RPOT)
2383 Changed |= processBlock(BB);
2388 void GVN::cleanupGlobalSets() {
2390 LeaderTable.clear();
2391 BlockRPONumber.clear();
2392 TableAllocator.Reset();
2394 InvalidBlockRPONumbers = true;
2397 /// Verify that the specified instruction does not occur in our
2398 /// internal data structures.
2399 void GVN::verifyRemoved(const Instruction *Inst) const {
2400 VN.verifyRemoved(Inst);
2402 // Walk through the value number scope to make sure the instruction isn't
2403 // ferreted away in it.
2404 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2405 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2406 const LeaderTableEntry *Node = &I->second;
2407 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2409 while (Node->Next) {
2411 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2416 /// BB is declared dead, which implied other blocks become dead as well. This
2417 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2418 /// live successors, update their phi nodes by replacing the operands
2419 /// corresponding to dead blocks with UndefVal.
2420 void GVN::addDeadBlock(BasicBlock *BB) {
2421 SmallVector<BasicBlock *, 4> NewDead;
2422 SmallSetVector<BasicBlock *, 4> DF;
2424 NewDead.push_back(BB);
2425 while (!NewDead.empty()) {
2426 BasicBlock *D = NewDead.pop_back_val();
2427 if (DeadBlocks.count(D))
2430 // All blocks dominated by D are dead.
2431 SmallVector<BasicBlock *, 8> Dom;
2432 DT->getDescendants(D, Dom);
2433 DeadBlocks.insert(Dom.begin(), Dom.end());
2435 // Figure out the dominance-frontier(D).
2436 for (BasicBlock *B : Dom) {
2437 for (BasicBlock *S : successors(B)) {
2438 if (DeadBlocks.count(S))
2441 bool AllPredDead = true;
2442 for (BasicBlock *P : predecessors(S))
2443 if (!DeadBlocks.count(P)) {
2444 AllPredDead = false;
2449 // S could be proved dead later on. That is why we don't update phi
2450 // operands at this moment.
2453 // While S is not dominated by D, it is dead by now. This could take
2454 // place if S already have a dead predecessor before D is declared
2456 NewDead.push_back(S);
2462 // For the dead blocks' live successors, update their phi nodes by replacing
2463 // the operands corresponding to dead blocks with UndefVal.
2464 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2467 if (DeadBlocks.count(B))
2470 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2471 for (BasicBlock *P : Preds) {
2472 if (!DeadBlocks.count(P))
2475 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2476 if (BasicBlock *S = splitCriticalEdges(P, B))
2477 DeadBlocks.insert(P = S);
2480 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2481 PHINode &Phi = cast<PHINode>(*II);
2482 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2483 UndefValue::get(Phi.getType()));
2485 MD->invalidateCachedPointerInfo(&Phi);
2491 // If the given branch is recognized as a foldable branch (i.e. conditional
2492 // branch with constant condition), it will perform following analyses and
2494 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2495 // R be the target of the dead out-coming edge.
2496 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2497 // edge. The result of this step will be {X| X is dominated by R}
2498 // 2) Identify those blocks which haves at least one dead predecessor. The
2499 // result of this step will be dominance-frontier(R).
2500 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2501 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2503 // Return true iff *NEW* dead code are found.
2504 bool GVN::processFoldableCondBr(BranchInst *BI) {
2505 if (!BI || BI->isUnconditional())
2508 // If a branch has two identical successors, we cannot declare either dead.
2509 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2512 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2516 BasicBlock *DeadRoot =
2517 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2518 if (DeadBlocks.count(DeadRoot))
2521 if (!DeadRoot->getSinglePredecessor())
2522 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2524 addDeadBlock(DeadRoot);
2528 // performPRE() will trigger assert if it comes across an instruction without
2529 // associated val-num. As it normally has far more live instructions than dead
2530 // instructions, it makes more sense just to "fabricate" a val-number for the
2531 // dead code than checking if instruction involved is dead or not.
2532 void GVN::assignValNumForDeadCode() {
2533 for (BasicBlock *BB : DeadBlocks) {
2534 for (Instruction &Inst : *BB) {
2535 unsigned ValNum = VN.lookupOrAdd(&Inst);
2536 addToLeaderTable(ValNum, &Inst, BB);
2541 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2543 static char ID; // Pass identification, replacement for typeid
2545 explicit GVNLegacyPass(bool NoMemDepAnalysis = !EnableMemDep)
2546 : FunctionPass(ID), NoMemDepAnalysis(NoMemDepAnalysis) {
2547 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2550 bool runOnFunction(Function &F) override {
2551 if (skipFunction(F))
2554 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2556 return Impl.runImpl(
2557 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2558 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2559 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2560 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2561 NoMemDepAnalysis ? nullptr
2562 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2563 LIWP ? &LIWP->getLoopInfo() : nullptr,
2564 &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2567 void getAnalysisUsage(AnalysisUsage &AU) const override {
2568 AU.addRequired<AssumptionCacheTracker>();
2569 AU.addRequired<DominatorTreeWrapperPass>();
2570 AU.addRequired<TargetLibraryInfoWrapperPass>();
2571 if (!NoMemDepAnalysis)
2572 AU.addRequired<MemoryDependenceWrapperPass>();
2573 AU.addRequired<AAResultsWrapperPass>();
2575 AU.addPreserved<DominatorTreeWrapperPass>();
2576 AU.addPreserved<GlobalsAAWrapperPass>();
2577 AU.addPreserved<TargetLibraryInfoWrapperPass>();
2578 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2582 bool NoMemDepAnalysis;
2586 char GVNLegacyPass::ID = 0;
2588 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2589 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2590 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2591 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2592 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2593 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2594 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2595 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2596 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2598 // The public interface to this file...
2599 FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) {
2600 return new GVNLegacyPass(NoMemDepAnalysis);