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/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/MemoryBuiltins.h"
35 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
36 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
37 #include "llvm/Analysis/PHITransAddr.h"
38 #include "llvm/Analysis/TargetLibraryInfo.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Metadata.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
53 #include "llvm/Transforms/Utils/VNCoercion.h"
57 using namespace llvm::gvn;
58 using namespace llvm::VNCoercion;
59 using namespace PatternMatch;
61 #define DEBUG_TYPE "gvn"
63 STATISTIC(NumGVNInstr, "Number of instructions deleted");
64 STATISTIC(NumGVNLoad, "Number of loads deleted");
65 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
66 STATISTIC(NumGVNBlocks, "Number of blocks merged");
67 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
68 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
69 STATISTIC(NumPRELoad, "Number of loads PRE'd");
71 static cl::opt<bool> EnablePRE("enable-pre",
72 cl::init(true), cl::Hidden);
73 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
75 // Maximum allowed recursion depth.
76 static cl::opt<uint32_t>
77 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
78 cl::desc("Max recurse depth (default = 1000)"));
80 struct llvm::GVN::Expression {
83 SmallVector<uint32_t, 4> varargs;
85 Expression(uint32_t o = ~2U) : opcode(o) {}
87 bool operator==(const Expression &other) const {
88 if (opcode != other.opcode)
90 if (opcode == ~0U || opcode == ~1U)
92 if (type != other.type)
94 if (varargs != other.varargs)
99 friend hash_code hash_value(const Expression &Value) {
101 Value.opcode, Value.type,
102 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
107 template <> struct DenseMapInfo<GVN::Expression> {
108 static inline GVN::Expression getEmptyKey() { return ~0U; }
110 static inline GVN::Expression getTombstoneKey() { return ~1U; }
112 static unsigned getHashValue(const GVN::Expression &e) {
113 using llvm::hash_value;
114 return static_cast<unsigned>(hash_value(e));
116 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
120 } // End llvm namespace.
122 /// Represents a particular available value that we know how to materialize.
123 /// Materialization of an AvailableValue never fails. An AvailableValue is
124 /// implicitly associated with a rematerialization point which is the
125 /// location of the instruction from which it was formed.
126 struct llvm::gvn::AvailableValue {
128 SimpleVal, // A simple offsetted value that is accessed.
129 LoadVal, // A value produced by a load.
130 MemIntrin, // A memory intrinsic which is loaded from.
131 UndefVal // A UndefValue representing a value from dead block (which
132 // is not yet physically removed from the CFG).
135 /// V - The value that is live out of the block.
136 PointerIntPair<Value *, 2, ValType> Val;
138 /// Offset - The byte offset in Val that is interesting for the load query.
141 static AvailableValue get(Value *V, unsigned Offset = 0) {
143 Res.Val.setPointer(V);
144 Res.Val.setInt(SimpleVal);
149 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
151 Res.Val.setPointer(MI);
152 Res.Val.setInt(MemIntrin);
157 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
159 Res.Val.setPointer(LI);
160 Res.Val.setInt(LoadVal);
165 static AvailableValue getUndef() {
167 Res.Val.setPointer(nullptr);
168 Res.Val.setInt(UndefVal);
173 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
174 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
175 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
176 bool isUndefValue() const { return Val.getInt() == UndefVal; }
178 Value *getSimpleValue() const {
179 assert(isSimpleValue() && "Wrong accessor");
180 return Val.getPointer();
183 LoadInst *getCoercedLoadValue() const {
184 assert(isCoercedLoadValue() && "Wrong accessor");
185 return cast<LoadInst>(Val.getPointer());
188 MemIntrinsic *getMemIntrinValue() const {
189 assert(isMemIntrinValue() && "Wrong accessor");
190 return cast<MemIntrinsic>(Val.getPointer());
193 /// Emit code at the specified insertion point to adjust the value defined
194 /// here to the specified type. This handles various coercion cases.
195 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
199 /// Represents an AvailableValue which can be rematerialized at the end of
200 /// the associated BasicBlock.
201 struct llvm::gvn::AvailableValueInBlock {
202 /// BB - The basic block in question.
205 /// AV - The actual available value
208 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
209 AvailableValueInBlock Res;
211 Res.AV = std::move(AV);
215 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
216 unsigned Offset = 0) {
217 return get(BB, AvailableValue::get(V, Offset));
219 static AvailableValueInBlock getUndef(BasicBlock *BB) {
220 return get(BB, AvailableValue::getUndef());
223 /// Emit code at the end of this block to adjust the value defined here to
224 /// the specified type. This handles various coercion cases.
225 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
226 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
230 //===----------------------------------------------------------------------===//
231 // ValueTable Internal Functions
232 //===----------------------------------------------------------------------===//
234 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
236 e.type = I->getType();
237 e.opcode = I->getOpcode();
238 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
240 e.varargs.push_back(lookupOrAdd(*OI));
241 if (I->isCommutative()) {
242 // Ensure that commutative instructions that only differ by a permutation
243 // of their operands get the same value number by sorting the operand value
244 // numbers. Since all commutative instructions have two operands it is more
245 // efficient to sort by hand rather than using, say, std::sort.
246 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
247 if (e.varargs[0] > e.varargs[1])
248 std::swap(e.varargs[0], e.varargs[1]);
251 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
252 // Sort the operand value numbers so x<y and y>x get the same value number.
253 CmpInst::Predicate Predicate = C->getPredicate();
254 if (e.varargs[0] > e.varargs[1]) {
255 std::swap(e.varargs[0], e.varargs[1]);
256 Predicate = CmpInst::getSwappedPredicate(Predicate);
258 e.opcode = (C->getOpcode() << 8) | Predicate;
259 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
260 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
262 e.varargs.push_back(*II);
268 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
269 CmpInst::Predicate Predicate,
270 Value *LHS, Value *RHS) {
271 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
272 "Not a comparison!");
274 e.type = CmpInst::makeCmpResultType(LHS->getType());
275 e.varargs.push_back(lookupOrAdd(LHS));
276 e.varargs.push_back(lookupOrAdd(RHS));
278 // Sort the operand value numbers so x<y and y>x get the same value number.
279 if (e.varargs[0] > e.varargs[1]) {
280 std::swap(e.varargs[0], e.varargs[1]);
281 Predicate = CmpInst::getSwappedPredicate(Predicate);
283 e.opcode = (Opcode << 8) | Predicate;
287 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
288 assert(EI && "Not an ExtractValueInst?");
290 e.type = EI->getType();
293 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
294 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
295 // EI might be an extract from one of our recognised intrinsics. If it
296 // is we'll synthesize a semantically equivalent expression instead on
297 // an extract value expression.
298 switch (I->getIntrinsicID()) {
299 case Intrinsic::sadd_with_overflow:
300 case Intrinsic::uadd_with_overflow:
301 e.opcode = Instruction::Add;
303 case Intrinsic::ssub_with_overflow:
304 case Intrinsic::usub_with_overflow:
305 e.opcode = Instruction::Sub;
307 case Intrinsic::smul_with_overflow:
308 case Intrinsic::umul_with_overflow:
309 e.opcode = Instruction::Mul;
316 // Intrinsic recognized. Grab its args to finish building the expression.
317 assert(I->getNumArgOperands() == 2 &&
318 "Expect two args for recognised intrinsics.");
319 e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
320 e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
325 // Not a recognised intrinsic. Fall back to producing an extract value
327 e.opcode = EI->getOpcode();
328 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
330 e.varargs.push_back(lookupOrAdd(*OI));
332 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
334 e.varargs.push_back(*II);
339 //===----------------------------------------------------------------------===//
340 // ValueTable External Functions
341 //===----------------------------------------------------------------------===//
343 GVN::ValueTable::ValueTable() : nextValueNumber(1) {}
344 GVN::ValueTable::ValueTable(const ValueTable &) = default;
345 GVN::ValueTable::ValueTable(ValueTable &&) = default;
346 GVN::ValueTable::~ValueTable() = default;
348 /// add - Insert a value into the table with a specified value number.
349 void GVN::ValueTable::add(Value *V, uint32_t num) {
350 valueNumbering.insert(std::make_pair(V, num));
353 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
354 if (AA->doesNotAccessMemory(C)) {
355 Expression exp = createExpr(C);
356 uint32_t &e = expressionNumbering[exp];
357 if (!e) e = nextValueNumber++;
358 valueNumbering[C] = e;
360 } else if (AA->onlyReadsMemory(C)) {
361 Expression exp = createExpr(C);
362 uint32_t &e = expressionNumbering[exp];
364 e = nextValueNumber++;
365 valueNumbering[C] = e;
369 e = nextValueNumber++;
370 valueNumbering[C] = e;
374 MemDepResult local_dep = MD->getDependency(C);
376 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
377 valueNumbering[C] = nextValueNumber;
378 return nextValueNumber++;
381 if (local_dep.isDef()) {
382 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
384 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
385 valueNumbering[C] = nextValueNumber;
386 return nextValueNumber++;
389 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
390 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
391 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
393 valueNumbering[C] = nextValueNumber;
394 return nextValueNumber++;
398 uint32_t v = lookupOrAdd(local_cdep);
399 valueNumbering[C] = v;
404 const MemoryDependenceResults::NonLocalDepInfo &deps =
405 MD->getNonLocalCallDependency(CallSite(C));
406 // FIXME: Move the checking logic to MemDep!
407 CallInst* cdep = nullptr;
409 // Check to see if we have a single dominating call instruction that is
411 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
412 const NonLocalDepEntry *I = &deps[i];
413 if (I->getResult().isNonLocal())
416 // We don't handle non-definitions. If we already have a call, reject
417 // instruction dependencies.
418 if (!I->getResult().isDef() || cdep != nullptr) {
423 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
424 // FIXME: All duplicated with non-local case.
425 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
426 cdep = NonLocalDepCall;
435 valueNumbering[C] = nextValueNumber;
436 return nextValueNumber++;
439 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
440 valueNumbering[C] = nextValueNumber;
441 return nextValueNumber++;
443 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
444 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
445 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
447 valueNumbering[C] = nextValueNumber;
448 return nextValueNumber++;
452 uint32_t v = lookupOrAdd(cdep);
453 valueNumbering[C] = v;
457 valueNumbering[C] = nextValueNumber;
458 return nextValueNumber++;
462 /// Returns true if a value number exists for the specified value.
463 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
465 /// lookup_or_add - Returns the value number for the specified value, assigning
466 /// it a new number if it did not have one before.
467 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
468 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
469 if (VI != valueNumbering.end())
472 if (!isa<Instruction>(V)) {
473 valueNumbering[V] = nextValueNumber;
474 return nextValueNumber++;
477 Instruction* I = cast<Instruction>(V);
479 switch (I->getOpcode()) {
480 case Instruction::Call:
481 return lookupOrAddCall(cast<CallInst>(I));
482 case Instruction::Add:
483 case Instruction::FAdd:
484 case Instruction::Sub:
485 case Instruction::FSub:
486 case Instruction::Mul:
487 case Instruction::FMul:
488 case Instruction::UDiv:
489 case Instruction::SDiv:
490 case Instruction::FDiv:
491 case Instruction::URem:
492 case Instruction::SRem:
493 case Instruction::FRem:
494 case Instruction::Shl:
495 case Instruction::LShr:
496 case Instruction::AShr:
497 case Instruction::And:
498 case Instruction::Or:
499 case Instruction::Xor:
500 case Instruction::ICmp:
501 case Instruction::FCmp:
502 case Instruction::Trunc:
503 case Instruction::ZExt:
504 case Instruction::SExt:
505 case Instruction::FPToUI:
506 case Instruction::FPToSI:
507 case Instruction::UIToFP:
508 case Instruction::SIToFP:
509 case Instruction::FPTrunc:
510 case Instruction::FPExt:
511 case Instruction::PtrToInt:
512 case Instruction::IntToPtr:
513 case Instruction::BitCast:
514 case Instruction::Select:
515 case Instruction::ExtractElement:
516 case Instruction::InsertElement:
517 case Instruction::ShuffleVector:
518 case Instruction::InsertValue:
519 case Instruction::GetElementPtr:
522 case Instruction::ExtractValue:
523 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
526 valueNumbering[V] = nextValueNumber;
527 return nextValueNumber++;
530 uint32_t& e = expressionNumbering[exp];
531 if (!e) e = nextValueNumber++;
532 valueNumbering[V] = e;
536 /// Returns the value number of the specified value. Fails if
537 /// the value has not yet been numbered.
538 uint32_t GVN::ValueTable::lookup(Value *V) const {
539 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
540 assert(VI != valueNumbering.end() && "Value not numbered?");
544 /// Returns the value number of the given comparison,
545 /// assigning it a new number if it did not have one before. Useful when
546 /// we deduced the result of a comparison, but don't immediately have an
547 /// instruction realizing that comparison to hand.
548 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
549 CmpInst::Predicate Predicate,
550 Value *LHS, Value *RHS) {
551 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
552 uint32_t& e = expressionNumbering[exp];
553 if (!e) e = nextValueNumber++;
557 /// Remove all entries from the ValueTable.
558 void GVN::ValueTable::clear() {
559 valueNumbering.clear();
560 expressionNumbering.clear();
564 /// Remove a value from the value numbering.
565 void GVN::ValueTable::erase(Value *V) {
566 valueNumbering.erase(V);
569 /// verifyRemoved - Verify that the value is removed from all internal data
571 void GVN::ValueTable::verifyRemoved(const Value *V) const {
572 for (DenseMap<Value*, uint32_t>::const_iterator
573 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
574 assert(I->first != V && "Inst still occurs in value numbering map!");
578 //===----------------------------------------------------------------------===//
580 //===----------------------------------------------------------------------===//
582 PreservedAnalyses GVN::run(Function &F, FunctionAnalysisManager &AM) {
583 // FIXME: The order of evaluation of these 'getResult' calls is very
584 // significant! Re-ordering these variables will cause GVN when run alone to
585 // be less effective! We should fix memdep and basic-aa to not exhibit this
586 // behavior, but until then don't change the order here.
587 auto &AC = AM.getResult<AssumptionAnalysis>(F);
588 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
589 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
590 auto &AA = AM.getResult<AAManager>(F);
591 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
592 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
593 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
594 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
596 return PreservedAnalyses::all();
597 PreservedAnalyses PA;
598 PA.preserve<DominatorTreeAnalysis>();
599 PA.preserve<GlobalsAA>();
600 PA.preserve<TargetLibraryAnalysis>();
604 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
605 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) {
607 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
608 E = d.end(); I != E; ++I) {
609 errs() << I->first << "\n";
616 /// Return true if we can prove that the value
617 /// we're analyzing is fully available in the specified block. As we go, keep
618 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
619 /// map is actually a tri-state map with the following values:
620 /// 0) we know the block *is not* fully available.
621 /// 1) we know the block *is* fully available.
622 /// 2) we do not know whether the block is fully available or not, but we are
623 /// currently speculating that it will be.
624 /// 3) we are speculating for this block and have used that to speculate for
626 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
627 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
628 uint32_t RecurseDepth) {
629 if (RecurseDepth > MaxRecurseDepth)
632 // Optimistically assume that the block is fully available and check to see
633 // if we already know about this block in one lookup.
634 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
635 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
637 // If the entry already existed for this block, return the precomputed value.
639 // If this is a speculative "available" value, mark it as being used for
640 // speculation of other blocks.
641 if (IV.first->second == 2)
642 IV.first->second = 3;
643 return IV.first->second != 0;
646 // Otherwise, see if it is fully available in all predecessors.
647 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
649 // If this block has no predecessors, it isn't live-in here.
651 goto SpeculationFailure;
653 for (; PI != PE; ++PI)
654 // If the value isn't fully available in one of our predecessors, then it
655 // isn't fully available in this block either. Undo our previous
656 // optimistic assumption and bail out.
657 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
658 goto SpeculationFailure;
662 // If we get here, we found out that this is not, after
663 // all, a fully-available block. We have a problem if we speculated on this and
664 // used the speculation to mark other blocks as available.
666 char &BBVal = FullyAvailableBlocks[BB];
668 // If we didn't speculate on this, just return with it set to false.
674 // If we did speculate on this value, we could have blocks set to 1 that are
675 // incorrect. Walk the (transitive) successors of this block and mark them as
677 SmallVector<BasicBlock*, 32> BBWorklist;
678 BBWorklist.push_back(BB);
681 BasicBlock *Entry = BBWorklist.pop_back_val();
682 // Note that this sets blocks to 0 (unavailable) if they happen to not
683 // already be in FullyAvailableBlocks. This is safe.
684 char &EntryVal = FullyAvailableBlocks[Entry];
685 if (EntryVal == 0) continue; // Already unavailable.
687 // Mark as unavailable.
690 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
691 } while (!BBWorklist.empty());
699 /// Given a set of loads specified by ValuesPerBlock,
700 /// construct SSA form, allowing us to eliminate LI. This returns the value
701 /// that should be used at LI's definition site.
702 static Value *ConstructSSAForLoadSet(LoadInst *LI,
703 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
705 // Check for the fully redundant, dominating load case. In this case, we can
706 // just use the dominating value directly.
707 if (ValuesPerBlock.size() == 1 &&
708 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
710 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
711 "Dead BB dominate this block");
712 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
715 // Otherwise, we have to construct SSA form.
716 SmallVector<PHINode*, 8> NewPHIs;
717 SSAUpdater SSAUpdate(&NewPHIs);
718 SSAUpdate.Initialize(LI->getType(), LI->getName());
720 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
721 BasicBlock *BB = AV.BB;
723 if (SSAUpdate.HasValueForBlock(BB))
726 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
729 // Perform PHI construction.
730 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
733 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
734 Instruction *InsertPt,
737 Type *LoadTy = LI->getType();
738 const DataLayout &DL = LI->getModule()->getDataLayout();
739 if (isSimpleValue()) {
740 Res = getSimpleValue();
741 if (Res->getType() != LoadTy) {
742 Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
744 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
745 << *getSimpleValue() << '\n'
746 << *Res << '\n' << "\n\n\n");
748 } else if (isCoercedLoadValue()) {
749 LoadInst *Load = getCoercedLoadValue();
750 if (Load->getType() == LoadTy && Offset == 0) {
753 Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
754 // We would like to use gvn.markInstructionForDeletion here, but we can't
755 // because the load is already memoized into the leader map table that GVN
756 // tracks. It is potentially possible to remove the load from the table,
757 // but then there all of the operations based on it would need to be
758 // rehashed. Just leave the dead load around.
759 gvn.getMemDep().removeInstruction(Load);
760 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
761 << *getCoercedLoadValue() << '\n'
765 } else if (isMemIntrinValue()) {
766 Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
768 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
769 << " " << *getMemIntrinValue() << '\n'
770 << *Res << '\n' << "\n\n\n");
772 assert(isUndefValue() && "Should be UndefVal");
773 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
774 return UndefValue::get(LoadTy);
776 assert(Res && "failed to materialize?");
780 static bool isLifetimeStart(const Instruction *Inst) {
781 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
782 return II->getIntrinsicID() == Intrinsic::lifetime_start;
786 /// \brief Try to locate the three instruction involved in a missed
787 /// load-elimination case that is due to an intervening store.
788 static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo,
790 OptimizationRemarkEmitter *ORE) {
792 User *OtherAccess = nullptr;
794 OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
795 R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
798 for (auto *U : LI->getPointerOperand()->users())
799 if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
800 DT->dominates(cast<Instruction>(U), LI)) {
801 // FIXME: for now give up if there are multiple memory accesses that
802 // dominate the load. We need further analysis to decide which one is
803 // that we're forwarding from.
805 OtherAccess = nullptr;
811 R << " in favor of " << NV("OtherAccess", OtherAccess);
813 R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
818 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
819 Value *Address, AvailableValue &Res) {
821 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
822 "expected a local dependence");
823 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
825 const DataLayout &DL = LI->getModule()->getDataLayout();
827 if (DepInfo.isClobber()) {
828 // If the dependence is to a store that writes to a superset of the bits
829 // read by the load, we can extract the bits we need for the load from the
831 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
832 // Can't forward from non-atomic to atomic without violating memory model.
833 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
835 analyzeLoadFromClobberingStore(LI->getType(), Address, DepSI, DL);
837 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
843 // Check to see if we have something like this:
846 // if we have this, replace the later with an extraction from the former.
847 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
848 // If this is a clobber and L is the first instruction in its block, then
849 // we have the first instruction in the entry block.
850 // Can't forward from non-atomic to atomic without violating memory model.
851 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
853 analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
856 Res = AvailableValue::getLoad(DepLI, Offset);
862 // If the clobbering value is a memset/memcpy/memmove, see if we can
863 // forward a value on from it.
864 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
865 if (Address && !LI->isAtomic()) {
866 int Offset = analyzeLoadFromClobberingMemInst(LI->getType(), Address,
869 Res = AvailableValue::getMI(DepMI, Offset);
874 // Nothing known about this clobber, have to be conservative
876 // fast print dep, using operator<< on instruction is too slow.
877 dbgs() << "GVN: load ";
878 LI->printAsOperand(dbgs());
879 Instruction *I = DepInfo.getInst();
880 dbgs() << " is clobbered by " << *I << '\n';
883 if (ORE->allowExtraAnalysis())
884 reportMayClobberedLoad(LI, DepInfo, DT, ORE);
888 assert(DepInfo.isDef() && "follows from above");
890 Instruction *DepInst = DepInfo.getInst();
892 // Loading the allocation -> undef.
893 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
894 // Loading immediately after lifetime begin -> undef.
895 isLifetimeStart(DepInst)) {
896 Res = AvailableValue::get(UndefValue::get(LI->getType()));
900 // Loading from calloc (which zero initializes memory) -> zero
901 if (isCallocLikeFn(DepInst, TLI)) {
902 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
906 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
907 // Reject loads and stores that are to the same address but are of
908 // different types if we have to. If the stored value is larger or equal to
909 // the loaded value, we can reuse it.
910 if (S->getValueOperand()->getType() != LI->getType() &&
911 !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
915 // Can't forward from non-atomic to atomic without violating memory model.
916 if (S->isAtomic() < LI->isAtomic())
919 Res = AvailableValue::get(S->getValueOperand());
923 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
924 // If the types mismatch and we can't handle it, reject reuse of the load.
925 // If the stored value is larger or equal to the loaded value, we can reuse
927 if (LD->getType() != LI->getType() &&
928 !canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
931 // Can't forward from non-atomic to atomic without violating memory model.
932 if (LD->isAtomic() < LI->isAtomic())
935 Res = AvailableValue::getLoad(LD);
939 // Unknown def - must be conservative
941 // fast print dep, using operator<< on instruction is too slow.
942 dbgs() << "GVN: load ";
943 LI->printAsOperand(dbgs());
944 dbgs() << " has unknown def " << *DepInst << '\n';
949 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
950 AvailValInBlkVect &ValuesPerBlock,
951 UnavailBlkVect &UnavailableBlocks) {
953 // Filter out useless results (non-locals, etc). Keep track of the blocks
954 // where we have a value available in repl, also keep track of whether we see
955 // dependencies that produce an unknown value for the load (such as a call
956 // that could potentially clobber the load).
957 unsigned NumDeps = Deps.size();
958 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
959 BasicBlock *DepBB = Deps[i].getBB();
960 MemDepResult DepInfo = Deps[i].getResult();
962 if (DeadBlocks.count(DepBB)) {
963 // Dead dependent mem-op disguise as a load evaluating the same value
964 // as the load in question.
965 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
969 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
970 UnavailableBlocks.push_back(DepBB);
974 // The address being loaded in this non-local block may not be the same as
975 // the pointer operand of the load if PHI translation occurs. Make sure
976 // to consider the right address.
977 Value *Address = Deps[i].getAddress();
980 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
981 // subtlety: because we know this was a non-local dependency, we know
982 // it's safe to materialize anywhere between the instruction within
983 // DepInfo and the end of it's block.
984 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
987 UnavailableBlocks.push_back(DepBB);
991 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
992 "post condition violation");
995 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
996 UnavailBlkVect &UnavailableBlocks) {
997 // Okay, we have *some* definitions of the value. This means that the value
998 // is available in some of our (transitive) predecessors. Lets think about
999 // doing PRE of this load. This will involve inserting a new load into the
1000 // predecessor when it's not available. We could do this in general, but
1001 // prefer to not increase code size. As such, we only do this when we know
1002 // that we only have to insert *one* load (which means we're basically moving
1003 // the load, not inserting a new one).
1005 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1006 UnavailableBlocks.end());
1008 // Let's find the first basic block with more than one predecessor. Walk
1009 // backwards through predecessors if needed.
1010 BasicBlock *LoadBB = LI->getParent();
1011 BasicBlock *TmpBB = LoadBB;
1013 while (TmpBB->getSinglePredecessor()) {
1014 TmpBB = TmpBB->getSinglePredecessor();
1015 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1017 if (Blockers.count(TmpBB))
1020 // If any of these blocks has more than one successor (i.e. if the edge we
1021 // just traversed was critical), then there are other paths through this
1022 // block along which the load may not be anticipated. Hoisting the load
1023 // above this block would be adding the load to execution paths along
1024 // which it was not previously executed.
1025 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1032 // Check to see how many predecessors have the loaded value fully
1034 MapVector<BasicBlock *, Value *> PredLoads;
1035 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1036 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1037 FullyAvailableBlocks[AV.BB] = true;
1038 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1039 FullyAvailableBlocks[UnavailableBB] = false;
1041 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1042 for (BasicBlock *Pred : predecessors(LoadBB)) {
1043 // If any predecessor block is an EH pad that does not allow non-PHI
1044 // instructions before the terminator, we can't PRE the load.
1045 if (Pred->getTerminator()->isEHPad()) {
1047 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1048 << Pred->getName() << "': " << *LI << '\n');
1052 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1056 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1057 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1058 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1059 << Pred->getName() << "': " << *LI << '\n');
1063 if (LoadBB->isEHPad()) {
1065 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1066 << Pred->getName() << "': " << *LI << '\n');
1070 CriticalEdgePred.push_back(Pred);
1072 // Only add the predecessors that will not be split for now.
1073 PredLoads[Pred] = nullptr;
1077 // Decide whether PRE is profitable for this load.
1078 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1079 assert(NumUnavailablePreds != 0 &&
1080 "Fully available value should already be eliminated!");
1082 // If this load is unavailable in multiple predecessors, reject it.
1083 // FIXME: If we could restructure the CFG, we could make a common pred with
1084 // all the preds that don't have an available LI and insert a new load into
1086 if (NumUnavailablePreds != 1)
1089 // Split critical edges, and update the unavailable predecessors accordingly.
1090 for (BasicBlock *OrigPred : CriticalEdgePred) {
1091 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1092 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1093 PredLoads[NewPred] = nullptr;
1094 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1095 << LoadBB->getName() << '\n');
1098 // Check if the load can safely be moved to all the unavailable predecessors.
1099 bool CanDoPRE = true;
1100 const DataLayout &DL = LI->getModule()->getDataLayout();
1101 SmallVector<Instruction*, 8> NewInsts;
1102 for (auto &PredLoad : PredLoads) {
1103 BasicBlock *UnavailablePred = PredLoad.first;
1105 // Do PHI translation to get its value in the predecessor if necessary. The
1106 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1108 // If all preds have a single successor, then we know it is safe to insert
1109 // the load on the pred (?!?), so we can insert code to materialize the
1110 // pointer if it is not available.
1111 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1112 Value *LoadPtr = nullptr;
1113 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1116 // If we couldn't find or insert a computation of this phi translated value,
1119 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1120 << *LI->getPointerOperand() << "\n");
1125 PredLoad.second = LoadPtr;
1129 while (!NewInsts.empty()) {
1130 Instruction *I = NewInsts.pop_back_val();
1131 if (MD) MD->removeInstruction(I);
1132 I->eraseFromParent();
1134 // HINT: Don't revert the edge-splitting as following transformation may
1135 // also need to split these critical edges.
1136 return !CriticalEdgePred.empty();
1139 // Okay, we can eliminate this load by inserting a reload in the predecessor
1140 // and using PHI construction to get the value in the other predecessors, do
1142 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1143 DEBUG(if (!NewInsts.empty())
1144 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1145 << *NewInsts.back() << '\n');
1147 // Assign value numbers to the new instructions.
1148 for (Instruction *I : NewInsts) {
1149 // Instructions that have been inserted in predecessor(s) to materialize
1150 // the load address do not retain their original debug locations. Doing
1151 // so could lead to confusing (but correct) source attributions.
1152 // FIXME: How do we retain source locations without causing poor debugging
1154 I->setDebugLoc(DebugLoc());
1156 // FIXME: We really _ought_ to insert these value numbers into their
1157 // parent's availability map. However, in doing so, we risk getting into
1158 // ordering issues. If a block hasn't been processed yet, we would be
1159 // marking a value as AVAIL-IN, which isn't what we intend.
1163 for (const auto &PredLoad : PredLoads) {
1164 BasicBlock *UnavailablePred = PredLoad.first;
1165 Value *LoadPtr = PredLoad.second;
1167 auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1168 LI->isVolatile(), LI->getAlignment(),
1169 LI->getOrdering(), LI->getSynchScope(),
1170 UnavailablePred->getTerminator());
1172 // Transfer the old load's AA tags to the new load.
1174 LI->getAAMetadata(Tags);
1176 NewLoad->setAAMetadata(Tags);
1178 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1179 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1180 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1181 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1182 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1183 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1185 // We do not propagate the old load's debug location, because the new
1186 // load now lives in a different BB, and we want to avoid a jumpy line
1188 // FIXME: How do we retain source locations without causing poor debugging
1191 // Add the newly created load.
1192 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1194 MD->invalidateCachedPointerInfo(LoadPtr);
1195 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1198 // Perform PHI construction.
1199 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1200 LI->replaceAllUsesWith(V);
1201 if (isa<PHINode>(V))
1203 if (Instruction *I = dyn_cast<Instruction>(V))
1204 I->setDebugLoc(LI->getDebugLoc());
1205 if (V->getType()->getScalarType()->isPointerTy())
1206 MD->invalidateCachedPointerInfo(V);
1207 markInstructionForDeletion(LI);
1208 ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1209 << "load eliminated by PRE");
1214 static void reportLoadElim(LoadInst *LI, Value *AvailableValue,
1215 OptimizationRemarkEmitter *ORE) {
1216 using namespace ore;
1217 ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1218 << "load of type " << NV("Type", LI->getType()) << " eliminated"
1219 << setExtraArgs() << " in favor of "
1220 << NV("InfavorOfValue", AvailableValue));
1223 /// Attempt to eliminate a load whose dependencies are
1224 /// non-local by performing PHI construction.
1225 bool GVN::processNonLocalLoad(LoadInst *LI) {
1226 // non-local speculations are not allowed under asan.
1227 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1230 // Step 1: Find the non-local dependencies of the load.
1232 MD->getNonLocalPointerDependency(LI, Deps);
1234 // If we had to process more than one hundred blocks to find the
1235 // dependencies, this load isn't worth worrying about. Optimizing
1236 // it will be too expensive.
1237 unsigned NumDeps = Deps.size();
1241 // If we had a phi translation failure, we'll have a single entry which is a
1242 // clobber in the current block. Reject this early.
1244 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1246 dbgs() << "GVN: non-local load ";
1247 LI->printAsOperand(dbgs());
1248 dbgs() << " has unknown dependencies\n";
1253 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1254 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1255 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1256 OE = GEP->idx_end();
1258 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1259 performScalarPRE(I);
1262 // Step 2: Analyze the availability of the load
1263 AvailValInBlkVect ValuesPerBlock;
1264 UnavailBlkVect UnavailableBlocks;
1265 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1267 // If we have no predecessors that produce a known value for this load, exit
1269 if (ValuesPerBlock.empty())
1272 // Step 3: Eliminate fully redundancy.
1274 // If all of the instructions we depend on produce a known value for this
1275 // load, then it is fully redundant and we can use PHI insertion to compute
1276 // its value. Insert PHIs and remove the fully redundant value now.
1277 if (UnavailableBlocks.empty()) {
1278 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1280 // Perform PHI construction.
1281 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1282 LI->replaceAllUsesWith(V);
1284 if (isa<PHINode>(V))
1286 if (Instruction *I = dyn_cast<Instruction>(V))
1287 // If instruction I has debug info, then we should not update it.
1288 // Also, if I has a null DebugLoc, then it is still potentially incorrect
1289 // to propagate LI's DebugLoc because LI may not post-dominate I.
1290 if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1291 I->setDebugLoc(LI->getDebugLoc());
1292 if (V->getType()->getScalarType()->isPointerTy())
1293 MD->invalidateCachedPointerInfo(V);
1294 markInstructionForDeletion(LI);
1296 reportLoadElim(LI, V, ORE);
1300 // Step 4: Eliminate partial redundancy.
1301 if (!EnablePRE || !EnableLoadPRE)
1304 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1307 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1308 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1309 "This function can only be called with llvm.assume intrinsic");
1310 Value *V = IntrinsicI->getArgOperand(0);
1312 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1313 if (Cond->isZero()) {
1314 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1315 // Insert a new store to null instruction before the load to indicate that
1316 // this code is not reachable. FIXME: We could insert unreachable
1317 // instruction directly because we can modify the CFG.
1318 new StoreInst(UndefValue::get(Int8Ty),
1319 Constant::getNullValue(Int8Ty->getPointerTo()),
1322 markInstructionForDeletion(IntrinsicI);
1326 Constant *True = ConstantInt::getTrue(V->getContext());
1327 bool Changed = false;
1329 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1330 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1332 // This property is only true in dominated successors, propagateEquality
1333 // will check dominance for us.
1334 Changed |= propagateEquality(V, True, Edge, false);
1337 // We can replace assume value with true, which covers cases like this:
1338 // call void @llvm.assume(i1 %cmp)
1339 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1340 ReplaceWithConstMap[V] = True;
1342 // If one of *cmp *eq operand is const, adding it to map will cover this:
1343 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1344 // call void @llvm.assume(i1 %cmp)
1345 // ret float %0 ; will change it to ret float 3.000000e+00
1346 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1347 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1348 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1349 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1350 CmpI->getFastMathFlags().noNaNs())) {
1351 Value *CmpLHS = CmpI->getOperand(0);
1352 Value *CmpRHS = CmpI->getOperand(1);
1353 if (isa<Constant>(CmpLHS))
1354 std::swap(CmpLHS, CmpRHS);
1355 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1357 // If only one operand is constant.
1358 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1359 ReplaceWithConstMap[CmpLHS] = RHSConst;
1365 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1366 auto *ReplInst = dyn_cast<Instruction>(Repl);
1370 // Patch the replacement so that it is not more restrictive than the value
1372 // Note that if 'I' is a load being replaced by some operation,
1373 // for example, by an arithmetic operation, then andIRFlags()
1374 // would just erase all math flags from the original arithmetic
1375 // operation, which is clearly not wanted and not needed.
1376 if (!isa<LoadInst>(I))
1377 ReplInst->andIRFlags(I);
1379 // FIXME: If both the original and replacement value are part of the
1380 // same control-flow region (meaning that the execution of one
1381 // guarantees the execution of the other), then we can combine the
1382 // noalias scopes here and do better than the general conservative
1383 // answer used in combineMetadata().
1385 // In general, GVN unifies expressions over different control-flow
1386 // regions, and so we need a conservative combination of the noalias
1388 static const unsigned KnownIDs[] = {
1389 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1390 LLVMContext::MD_noalias, LLVMContext::MD_range,
1391 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1392 LLVMContext::MD_invariant_group};
1393 combineMetadata(ReplInst, I, KnownIDs);
1396 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1397 patchReplacementInstruction(I, Repl);
1398 I->replaceAllUsesWith(Repl);
1401 /// Attempt to eliminate a load, first by eliminating it
1402 /// locally, and then attempting non-local elimination if that fails.
1403 bool GVN::processLoad(LoadInst *L) {
1407 // This code hasn't been audited for ordered or volatile memory access
1408 if (!L->isUnordered())
1411 if (L->use_empty()) {
1412 markInstructionForDeletion(L);
1416 // ... to a pointer that has been loaded from before...
1417 MemDepResult Dep = MD->getDependency(L);
1419 // If it is defined in another block, try harder.
1420 if (Dep.isNonLocal())
1421 return processNonLocalLoad(L);
1423 // Only handle the local case below
1424 if (!Dep.isDef() && !Dep.isClobber()) {
1425 // This might be a NonFuncLocal or an Unknown
1427 // fast print dep, using operator<< on instruction is too slow.
1428 dbgs() << "GVN: load ";
1429 L->printAsOperand(dbgs());
1430 dbgs() << " has unknown dependence\n";
1436 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1437 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1439 // Replace the load!
1440 patchAndReplaceAllUsesWith(L, AvailableValue);
1441 markInstructionForDeletion(L);
1443 reportLoadElim(L, AvailableValue, ORE);
1444 // Tell MDA to rexamine the reused pointer since we might have more
1445 // information after forwarding it.
1446 if (MD && AvailableValue->getType()->getScalarType()->isPointerTy())
1447 MD->invalidateCachedPointerInfo(AvailableValue);
1454 // In order to find a leader for a given value number at a
1455 // specific basic block, we first obtain the list of all Values for that number,
1456 // and then scan the list to find one whose block dominates the block in
1457 // question. This is fast because dominator tree queries consist of only
1458 // a few comparisons of DFS numbers.
1459 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1460 LeaderTableEntry Vals = LeaderTable[num];
1461 if (!Vals.Val) return nullptr;
1463 Value *Val = nullptr;
1464 if (DT->dominates(Vals.BB, BB)) {
1466 if (isa<Constant>(Val)) return Val;
1469 LeaderTableEntry* Next = Vals.Next;
1471 if (DT->dominates(Next->BB, BB)) {
1472 if (isa<Constant>(Next->Val)) return Next->Val;
1473 if (!Val) Val = Next->Val;
1482 /// There is an edge from 'Src' to 'Dst'. Return
1483 /// true if every path from the entry block to 'Dst' passes via this edge. In
1484 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1485 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1486 DominatorTree *DT) {
1487 // While in theory it is interesting to consider the case in which Dst has
1488 // more than one predecessor, because Dst might be part of a loop which is
1489 // only reachable from Src, in practice it is pointless since at the time
1490 // GVN runs all such loops have preheaders, which means that Dst will have
1491 // been changed to have only one predecessor, namely Src.
1492 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1493 assert((!Pred || Pred == E.getStart()) &&
1494 "No edge between these basic blocks!");
1495 return Pred != nullptr;
1498 // Tries to replace instruction with const, using information from
1499 // ReplaceWithConstMap.
1500 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1501 bool Changed = false;
1502 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1503 Value *Operand = Instr->getOperand(OpNum);
1504 auto it = ReplaceWithConstMap.find(Operand);
1505 if (it != ReplaceWithConstMap.end()) {
1506 assert(!isa<Constant>(Operand) &&
1507 "Replacing constants with constants is invalid");
1508 DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1509 << " in instruction " << *Instr << '\n');
1510 Instr->setOperand(OpNum, it->second);
1517 /// The given values are known to be equal in every block
1518 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1519 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1520 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1521 /// value starting from the end of Root.Start.
1522 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1523 bool DominatesByEdge) {
1524 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1525 Worklist.push_back(std::make_pair(LHS, RHS));
1526 bool Changed = false;
1527 // For speed, compute a conservative fast approximation to
1528 // DT->dominates(Root, Root.getEnd());
1529 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1531 while (!Worklist.empty()) {
1532 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1533 LHS = Item.first; RHS = Item.second;
1537 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1539 // Don't try to propagate equalities between constants.
1540 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1543 // Prefer a constant on the right-hand side, or an Argument if no constants.
1544 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1545 std::swap(LHS, RHS);
1546 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1548 // If there is no obvious reason to prefer the left-hand side over the
1549 // right-hand side, ensure the longest lived term is on the right-hand side,
1550 // so the shortest lived term will be replaced by the longest lived.
1551 // This tends to expose more simplifications.
1552 uint32_t LVN = VN.lookupOrAdd(LHS);
1553 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1554 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1555 // Move the 'oldest' value to the right-hand side, using the value number
1556 // as a proxy for age.
1557 uint32_t RVN = VN.lookupOrAdd(RHS);
1559 std::swap(LHS, RHS);
1564 // If value numbering later sees that an instruction in the scope is equal
1565 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1566 // the invariant that instructions only occur in the leader table for their
1567 // own value number (this is used by removeFromLeaderTable), do not do this
1568 // if RHS is an instruction (if an instruction in the scope is morphed into
1569 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1570 // using the leader table is about compiling faster, not optimizing better).
1571 // The leader table only tracks basic blocks, not edges. Only add to if we
1572 // have the simple case where the edge dominates the end.
1573 if (RootDominatesEnd && !isa<Instruction>(RHS))
1574 addToLeaderTable(LVN, RHS, Root.getEnd());
1576 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1577 // LHS always has at least one use that is not dominated by Root, this will
1578 // never do anything if LHS has only one use.
1579 if (!LHS->hasOneUse()) {
1580 unsigned NumReplacements =
1582 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1583 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1585 Changed |= NumReplacements > 0;
1586 NumGVNEqProp += NumReplacements;
1589 // Now try to deduce additional equalities from this one. For example, if
1590 // the known equality was "(A != B)" == "false" then it follows that A and B
1591 // are equal in the scope. Only boolean equalities with an explicit true or
1592 // false RHS are currently supported.
1593 if (!RHS->getType()->isIntegerTy(1))
1594 // Not a boolean equality - bail out.
1596 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1598 // RHS neither 'true' nor 'false' - bail out.
1600 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1601 bool isKnownTrue = CI->isAllOnesValue();
1602 bool isKnownFalse = !isKnownTrue;
1604 // If "A && B" is known true then both A and B are known true. If "A || B"
1605 // is known false then both A and B are known false.
1607 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1608 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1609 Worklist.push_back(std::make_pair(A, RHS));
1610 Worklist.push_back(std::make_pair(B, RHS));
1614 // If we are propagating an equality like "(A == B)" == "true" then also
1615 // propagate the equality A == B. When propagating a comparison such as
1616 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1617 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1618 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1620 // If "A == B" is known true, or "A != B" is known false, then replace
1621 // A with B everywhere in the scope.
1622 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1623 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1624 Worklist.push_back(std::make_pair(Op0, Op1));
1626 // Handle the floating point versions of equality comparisons too.
1627 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1628 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1630 // Floating point -0.0 and 0.0 compare equal, so we can only
1631 // propagate values if we know that we have a constant and that
1632 // its value is non-zero.
1634 // FIXME: We should do this optimization if 'no signed zeros' is
1635 // applicable via an instruction-level fast-math-flag or some other
1636 // indicator that relaxed FP semantics are being used.
1638 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1639 Worklist.push_back(std::make_pair(Op0, Op1));
1642 // If "A >= B" is known true, replace "A < B" with false everywhere.
1643 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1644 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1645 // Since we don't have the instruction "A < B" immediately to hand, work
1646 // out the value number that it would have and use that to find an
1647 // appropriate instruction (if any).
1648 uint32_t NextNum = VN.getNextUnusedValueNumber();
1649 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1650 // If the number we were assigned was brand new then there is no point in
1651 // looking for an instruction realizing it: there cannot be one!
1652 if (Num < NextNum) {
1653 Value *NotCmp = findLeader(Root.getEnd(), Num);
1654 if (NotCmp && isa<Instruction>(NotCmp)) {
1655 unsigned NumReplacements =
1657 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1658 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1660 Changed |= NumReplacements > 0;
1661 NumGVNEqProp += NumReplacements;
1664 // Ensure that any instruction in scope that gets the "A < B" value number
1665 // is replaced with false.
1666 // The leader table only tracks basic blocks, not edges. Only add to if we
1667 // have the simple case where the edge dominates the end.
1668 if (RootDominatesEnd)
1669 addToLeaderTable(Num, NotVal, Root.getEnd());
1678 /// When calculating availability, handle an instruction
1679 /// by inserting it into the appropriate sets
1680 bool GVN::processInstruction(Instruction *I) {
1681 // Ignore dbg info intrinsics.
1682 if (isa<DbgInfoIntrinsic>(I))
1685 // If the instruction can be easily simplified then do so now in preference
1686 // to value numbering it. Value numbering often exposes redundancies, for
1687 // example if it determines that %y is equal to %x then the instruction
1688 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1689 const DataLayout &DL = I->getModule()->getDataLayout();
1690 if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1691 bool Changed = false;
1692 if (!I->use_empty()) {
1693 I->replaceAllUsesWith(V);
1696 if (isInstructionTriviallyDead(I, TLI)) {
1697 markInstructionForDeletion(I);
1701 if (MD && V->getType()->getScalarType()->isPointerTy())
1702 MD->invalidateCachedPointerInfo(V);
1708 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1709 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1710 return processAssumeIntrinsic(IntrinsicI);
1712 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1713 if (processLoad(LI))
1716 unsigned Num = VN.lookupOrAdd(LI);
1717 addToLeaderTable(Num, LI, LI->getParent());
1721 // For conditional branches, we can perform simple conditional propagation on
1722 // the condition value itself.
1723 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1724 if (!BI->isConditional())
1727 if (isa<Constant>(BI->getCondition()))
1728 return processFoldableCondBr(BI);
1730 Value *BranchCond = BI->getCondition();
1731 BasicBlock *TrueSucc = BI->getSuccessor(0);
1732 BasicBlock *FalseSucc = BI->getSuccessor(1);
1733 // Avoid multiple edges early.
1734 if (TrueSucc == FalseSucc)
1737 BasicBlock *Parent = BI->getParent();
1738 bool Changed = false;
1740 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1741 BasicBlockEdge TrueE(Parent, TrueSucc);
1742 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1744 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1745 BasicBlockEdge FalseE(Parent, FalseSucc);
1746 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1751 // For switches, propagate the case values into the case destinations.
1752 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1753 Value *SwitchCond = SI->getCondition();
1754 BasicBlock *Parent = SI->getParent();
1755 bool Changed = false;
1757 // Remember how many outgoing edges there are to every successor.
1758 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
1759 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1760 ++SwitchEdges[SI->getSuccessor(i)];
1762 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1764 BasicBlock *Dst = i->getCaseSuccessor();
1765 // If there is only a single edge, propagate the case value into it.
1766 if (SwitchEdges.lookup(Dst) == 1) {
1767 BasicBlockEdge E(Parent, Dst);
1768 Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1774 // Instructions with void type don't return a value, so there's
1775 // no point in trying to find redundancies in them.
1776 if (I->getType()->isVoidTy())
1779 uint32_t NextNum = VN.getNextUnusedValueNumber();
1780 unsigned Num = VN.lookupOrAdd(I);
1782 // Allocations are always uniquely numbered, so we can save time and memory
1783 // by fast failing them.
1784 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1785 addToLeaderTable(Num, I, I->getParent());
1789 // If the number we were assigned was a brand new VN, then we don't
1790 // need to do a lookup to see if the number already exists
1791 // somewhere in the domtree: it can't!
1792 if (Num >= NextNum) {
1793 addToLeaderTable(Num, I, I->getParent());
1797 // Perform fast-path value-number based elimination of values inherited from
1799 Value *Repl = findLeader(I->getParent(), Num);
1801 // Failure, just remember this instance for future use.
1802 addToLeaderTable(Num, I, I->getParent());
1804 } else if (Repl == I) {
1805 // If I was the result of a shortcut PRE, it might already be in the table
1806 // and the best replacement for itself. Nothing to do.
1811 patchAndReplaceAllUsesWith(I, Repl);
1812 if (MD && Repl->getType()->getScalarType()->isPointerTy())
1813 MD->invalidateCachedPointerInfo(Repl);
1814 markInstructionForDeletion(I);
1818 /// runOnFunction - This is the main transformation entry point for a function.
1819 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1820 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1821 MemoryDependenceResults *RunMD, LoopInfo *LI,
1822 OptimizationRemarkEmitter *RunORE) {
1827 VN.setAliasAnalysis(&RunAA);
1832 bool Changed = false;
1833 bool ShouldContinue = true;
1835 // Merge unconditional branches, allowing PRE to catch more
1836 // optimization opportunities.
1837 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1838 BasicBlock *BB = &*FI++;
1840 bool removedBlock = MergeBlockIntoPredecessor(BB, DT, LI, MD);
1844 Changed |= removedBlock;
1847 unsigned Iteration = 0;
1848 while (ShouldContinue) {
1849 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1850 ShouldContinue = iterateOnFunction(F);
1851 Changed |= ShouldContinue;
1856 // Fabricate val-num for dead-code in order to suppress assertion in
1858 assignValNumForDeadCode();
1859 bool PREChanged = true;
1860 while (PREChanged) {
1861 PREChanged = performPRE(F);
1862 Changed |= PREChanged;
1866 // FIXME: Should perform GVN again after PRE does something. PRE can move
1867 // computations into blocks where they become fully redundant. Note that
1868 // we can't do this until PRE's critical edge splitting updates memdep.
1869 // Actually, when this happens, we should just fully integrate PRE into GVN.
1871 cleanupGlobalSets();
1872 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
1879 bool GVN::processBlock(BasicBlock *BB) {
1880 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
1881 // (and incrementing BI before processing an instruction).
1882 assert(InstrsToErase.empty() &&
1883 "We expect InstrsToErase to be empty across iterations");
1884 if (DeadBlocks.count(BB))
1887 // Clearing map before every BB because it can be used only for single BB.
1888 ReplaceWithConstMap.clear();
1889 bool ChangedFunction = false;
1891 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1893 if (!ReplaceWithConstMap.empty())
1894 ChangedFunction |= replaceOperandsWithConsts(&*BI);
1895 ChangedFunction |= processInstruction(&*BI);
1897 if (InstrsToErase.empty()) {
1902 // If we need some instructions deleted, do it now.
1903 NumGVNInstr += InstrsToErase.size();
1905 // Avoid iterator invalidation.
1906 bool AtStart = BI == BB->begin();
1910 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
1911 E = InstrsToErase.end(); I != E; ++I) {
1912 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
1913 if (MD) MD->removeInstruction(*I);
1914 DEBUG(verifyRemoved(*I));
1915 (*I)->eraseFromParent();
1917 InstrsToErase.clear();
1925 return ChangedFunction;
1928 // Instantiate an expression in a predecessor that lacked it.
1929 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
1930 unsigned int ValNo) {
1931 // Because we are going top-down through the block, all value numbers
1932 // will be available in the predecessor by the time we need them. Any
1933 // that weren't originally present will have been instantiated earlier
1935 bool success = true;
1936 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
1937 Value *Op = Instr->getOperand(i);
1938 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1940 // This could be a newly inserted instruction, in which case, we won't
1941 // find a value number, and should give up before we hurt ourselves.
1942 // FIXME: Rewrite the infrastructure to let it easier to value number
1943 // and process newly inserted instructions.
1944 if (!VN.exists(Op)) {
1948 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
1949 Instr->setOperand(i, V);
1956 // Fail out if we encounter an operand that is not available in
1957 // the PRE predecessor. This is typically because of loads which
1958 // are not value numbered precisely.
1962 Instr->insertBefore(Pred->getTerminator());
1963 Instr->setName(Instr->getName() + ".pre");
1964 Instr->setDebugLoc(Instr->getDebugLoc());
1965 VN.add(Instr, ValNo);
1967 // Update the availability map to include the new instruction.
1968 addToLeaderTable(ValNo, Instr, Pred);
1972 bool GVN::performScalarPRE(Instruction *CurInst) {
1973 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
1974 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
1975 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1976 isa<DbgInfoIntrinsic>(CurInst))
1979 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
1980 // sinking the compare again, and it would force the code generator to
1981 // move the i1 from processor flags or predicate registers into a general
1982 // purpose register.
1983 if (isa<CmpInst>(CurInst))
1986 // We don't currently value number ANY inline asm calls.
1987 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
1988 if (CallI->isInlineAsm())
1991 uint32_t ValNo = VN.lookup(CurInst);
1993 // Look for the predecessors for PRE opportunities. We're
1994 // only trying to solve the basic diamond case, where
1995 // a value is computed in the successor and one predecessor,
1996 // but not the other. We also explicitly disallow cases
1997 // where the successor is its own predecessor, because they're
1998 // more complicated to get right.
1999 unsigned NumWith = 0;
2000 unsigned NumWithout = 0;
2001 BasicBlock *PREPred = nullptr;
2002 BasicBlock *CurrentBlock = CurInst->getParent();
2004 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2005 for (BasicBlock *P : predecessors(CurrentBlock)) {
2006 // We're not interested in PRE where the block is its
2007 // own predecessor, or in blocks with predecessors
2008 // that are not reachable.
2009 if (P == CurrentBlock) {
2012 } else if (!DT->isReachableFromEntry(P)) {
2017 Value *predV = findLeader(P, ValNo);
2019 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2022 } else if (predV == CurInst) {
2023 /* CurInst dominates this predecessor. */
2027 predMap.push_back(std::make_pair(predV, P));
2032 // Don't do PRE when it might increase code size, i.e. when
2033 // we would need to insert instructions in more than one pred.
2034 if (NumWithout > 1 || NumWith == 0)
2037 // We may have a case where all predecessors have the instruction,
2038 // and we just need to insert a phi node. Otherwise, perform
2040 Instruction *PREInstr = nullptr;
2042 if (NumWithout != 0) {
2043 // Don't do PRE across indirect branch.
2044 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2047 // We can't do PRE safely on a critical edge, so instead we schedule
2048 // the edge to be split and perform the PRE the next time we iterate
2050 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2051 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2052 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2055 // We need to insert somewhere, so let's give it a shot
2056 PREInstr = CurInst->clone();
2057 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2058 // If we failed insertion, make sure we remove the instruction.
2059 DEBUG(verifyRemoved(PREInstr));
2065 // Either we should have filled in the PRE instruction, or we should
2066 // not have needed insertions.
2067 assert (PREInstr != nullptr || NumWithout == 0);
2071 // Create a PHI to make the value available in this block.
2073 PHINode::Create(CurInst->getType(), predMap.size(),
2074 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2075 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2076 if (Value *V = predMap[i].first)
2077 Phi->addIncoming(V, predMap[i].second);
2079 Phi->addIncoming(PREInstr, PREPred);
2083 addToLeaderTable(ValNo, Phi, CurrentBlock);
2084 Phi->setDebugLoc(CurInst->getDebugLoc());
2085 CurInst->replaceAllUsesWith(Phi);
2086 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2087 MD->invalidateCachedPointerInfo(Phi);
2089 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2091 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2093 MD->removeInstruction(CurInst);
2094 DEBUG(verifyRemoved(CurInst));
2095 CurInst->eraseFromParent();
2101 /// Perform a purely local form of PRE that looks for diamond
2102 /// control flow patterns and attempts to perform simple PRE at the join point.
2103 bool GVN::performPRE(Function &F) {
2104 bool Changed = false;
2105 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2106 // Nothing to PRE in the entry block.
2107 if (CurrentBlock == &F.getEntryBlock())
2110 // Don't perform PRE on an EH pad.
2111 if (CurrentBlock->isEHPad())
2114 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2115 BE = CurrentBlock->end();
2117 Instruction *CurInst = &*BI++;
2118 Changed |= performScalarPRE(CurInst);
2122 if (splitCriticalEdges())
2128 /// Split the critical edge connecting the given two blocks, and return
2129 /// the block inserted to the critical edge.
2130 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2132 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2134 MD->invalidateCachedPredecessors();
2138 /// Split critical edges found during the previous
2139 /// iteration that may enable further optimization.
2140 bool GVN::splitCriticalEdges() {
2141 if (toSplit.empty())
2144 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2145 SplitCriticalEdge(Edge.first, Edge.second,
2146 CriticalEdgeSplittingOptions(DT));
2147 } while (!toSplit.empty());
2148 if (MD) MD->invalidateCachedPredecessors();
2152 /// Executes one iteration of GVN
2153 bool GVN::iterateOnFunction(Function &F) {
2154 cleanupGlobalSets();
2156 // Top-down walk of the dominator tree
2157 bool Changed = false;
2158 // Needed for value numbering with phi construction to work.
2159 // RPOT walks the graph in its constructor and will not be invalidated during
2161 ReversePostOrderTraversal<Function *> RPOT(&F);
2162 for (BasicBlock *BB : RPOT)
2163 Changed |= processBlock(BB);
2168 void GVN::cleanupGlobalSets() {
2170 LeaderTable.clear();
2171 TableAllocator.Reset();
2174 /// Verify that the specified instruction does not occur in our
2175 /// internal data structures.
2176 void GVN::verifyRemoved(const Instruction *Inst) const {
2177 VN.verifyRemoved(Inst);
2179 // Walk through the value number scope to make sure the instruction isn't
2180 // ferreted away in it.
2181 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2182 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2183 const LeaderTableEntry *Node = &I->second;
2184 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2186 while (Node->Next) {
2188 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2193 /// BB is declared dead, which implied other blocks become dead as well. This
2194 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2195 /// live successors, update their phi nodes by replacing the operands
2196 /// corresponding to dead blocks with UndefVal.
2197 void GVN::addDeadBlock(BasicBlock *BB) {
2198 SmallVector<BasicBlock *, 4> NewDead;
2199 SmallSetVector<BasicBlock *, 4> DF;
2201 NewDead.push_back(BB);
2202 while (!NewDead.empty()) {
2203 BasicBlock *D = NewDead.pop_back_val();
2204 if (DeadBlocks.count(D))
2207 // All blocks dominated by D are dead.
2208 SmallVector<BasicBlock *, 8> Dom;
2209 DT->getDescendants(D, Dom);
2210 DeadBlocks.insert(Dom.begin(), Dom.end());
2212 // Figure out the dominance-frontier(D).
2213 for (BasicBlock *B : Dom) {
2214 for (BasicBlock *S : successors(B)) {
2215 if (DeadBlocks.count(S))
2218 bool AllPredDead = true;
2219 for (BasicBlock *P : predecessors(S))
2220 if (!DeadBlocks.count(P)) {
2221 AllPredDead = false;
2226 // S could be proved dead later on. That is why we don't update phi
2227 // operands at this moment.
2230 // While S is not dominated by D, it is dead by now. This could take
2231 // place if S already have a dead predecessor before D is declared
2233 NewDead.push_back(S);
2239 // For the dead blocks' live successors, update their phi nodes by replacing
2240 // the operands corresponding to dead blocks with UndefVal.
2241 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2244 if (DeadBlocks.count(B))
2247 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2248 for (BasicBlock *P : Preds) {
2249 if (!DeadBlocks.count(P))
2252 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2253 if (BasicBlock *S = splitCriticalEdges(P, B))
2254 DeadBlocks.insert(P = S);
2257 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2258 PHINode &Phi = cast<PHINode>(*II);
2259 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2260 UndefValue::get(Phi.getType()));
2266 // If the given branch is recognized as a foldable branch (i.e. conditional
2267 // branch with constant condition), it will perform following analyses and
2269 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2270 // R be the target of the dead out-coming edge.
2271 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2272 // edge. The result of this step will be {X| X is dominated by R}
2273 // 2) Identify those blocks which haves at least one dead predecessor. The
2274 // result of this step will be dominance-frontier(R).
2275 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2276 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2278 // Return true iff *NEW* dead code are found.
2279 bool GVN::processFoldableCondBr(BranchInst *BI) {
2280 if (!BI || BI->isUnconditional())
2283 // If a branch has two identical successors, we cannot declare either dead.
2284 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2287 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2291 BasicBlock *DeadRoot =
2292 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2293 if (DeadBlocks.count(DeadRoot))
2296 if (!DeadRoot->getSinglePredecessor())
2297 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2299 addDeadBlock(DeadRoot);
2303 // performPRE() will trigger assert if it comes across an instruction without
2304 // associated val-num. As it normally has far more live instructions than dead
2305 // instructions, it makes more sense just to "fabricate" a val-number for the
2306 // dead code than checking if instruction involved is dead or not.
2307 void GVN::assignValNumForDeadCode() {
2308 for (BasicBlock *BB : DeadBlocks) {
2309 for (Instruction &Inst : *BB) {
2310 unsigned ValNum = VN.lookupOrAdd(&Inst);
2311 addToLeaderTable(ValNum, &Inst, BB);
2316 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2318 static char ID; // Pass identification, replacement for typeid
2319 explicit GVNLegacyPass(bool NoLoads = false)
2320 : FunctionPass(ID), NoLoads(NoLoads) {
2321 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2324 bool runOnFunction(Function &F) override {
2325 if (skipFunction(F))
2328 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2330 return Impl.runImpl(
2331 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2332 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2333 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2334 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2336 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2337 LIWP ? &LIWP->getLoopInfo() : nullptr,
2338 &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2341 void getAnalysisUsage(AnalysisUsage &AU) const override {
2342 AU.addRequired<AssumptionCacheTracker>();
2343 AU.addRequired<DominatorTreeWrapperPass>();
2344 AU.addRequired<TargetLibraryInfoWrapperPass>();
2346 AU.addRequired<MemoryDependenceWrapperPass>();
2347 AU.addRequired<AAResultsWrapperPass>();
2349 AU.addPreserved<DominatorTreeWrapperPass>();
2350 AU.addPreserved<GlobalsAAWrapperPass>();
2351 AU.addPreserved<TargetLibraryInfoWrapperPass>();
2352 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2360 char GVNLegacyPass::ID = 0;
2362 // The public interface to this file...
2363 FunctionPass *llvm::createGVNPass(bool NoLoads) {
2364 return new GVNLegacyPass(NoLoads);
2367 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2368 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2369 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2370 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2371 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2372 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2373 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2374 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2375 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)