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/IR/Attributes.h"
42 #include "llvm/IR/BasicBlock.h"
43 #include "llvm/IR/CallSite.h"
44 #include "llvm/IR/Constant.h"
45 #include "llvm/IR/Constants.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/IR/Dominators.h"
49 #include "llvm/IR/Function.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PassManager.h"
60 #include "llvm/IR/PatternMatch.h"
61 #include "llvm/IR/Type.h"
62 #include "llvm/IR/Use.h"
63 #include "llvm/IR/Value.h"
64 #include "llvm/Pass.h"
65 #include "llvm/Support/Casting.h"
66 #include "llvm/Support/CommandLine.h"
67 #include "llvm/Support/Compiler.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/raw_ostream.h"
70 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
71 #include "llvm/Transforms/Utils/Local.h"
72 #include "llvm/Transforms/Utils/SSAUpdater.h"
73 #include "llvm/Transforms/Utils/VNCoercion.h"
81 using namespace llvm::gvn;
82 using namespace llvm::VNCoercion;
83 using namespace PatternMatch;
85 #define DEBUG_TYPE "gvn"
87 STATISTIC(NumGVNInstr, "Number of instructions deleted");
88 STATISTIC(NumGVNLoad, "Number of loads deleted");
89 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
90 STATISTIC(NumGVNBlocks, "Number of blocks merged");
91 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
92 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
93 STATISTIC(NumPRELoad, "Number of loads PRE'd");
95 static cl::opt<bool> EnablePRE("enable-pre",
96 cl::init(true), cl::Hidden);
97 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
99 // Maximum allowed recursion depth.
100 static cl::opt<uint32_t>
101 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
102 cl::desc("Max recurse depth (default = 1000)"));
104 struct llvm::GVN::Expression {
107 bool commutative = false;
108 SmallVector<uint32_t, 4> varargs;
110 Expression(uint32_t o = ~2U) : opcode(o) {}
112 bool operator==(const Expression &other) const {
113 if (opcode != other.opcode)
115 if (opcode == ~0U || opcode == ~1U)
117 if (type != other.type)
119 if (varargs != other.varargs)
124 friend hash_code hash_value(const Expression &Value) {
126 Value.opcode, Value.type,
127 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
133 template <> struct DenseMapInfo<GVN::Expression> {
134 static inline GVN::Expression getEmptyKey() { return ~0U; }
135 static inline GVN::Expression getTombstoneKey() { return ~1U; }
137 static unsigned getHashValue(const GVN::Expression &e) {
138 using llvm::hash_value;
140 return static_cast<unsigned>(hash_value(e));
143 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
148 } // end namespace llvm
150 /// Represents a particular available value that we know how to materialize.
151 /// Materialization of an AvailableValue never fails. An AvailableValue is
152 /// implicitly associated with a rematerialization point which is the
153 /// location of the instruction from which it was formed.
154 struct llvm::gvn::AvailableValue {
156 SimpleVal, // A simple offsetted value that is accessed.
157 LoadVal, // A value produced by a load.
158 MemIntrin, // A memory intrinsic which is loaded from.
159 UndefVal // A UndefValue representing a value from dead block (which
160 // is not yet physically removed from the CFG).
163 /// V - The value that is live out of the block.
164 PointerIntPair<Value *, 2, ValType> Val;
166 /// Offset - The byte offset in Val that is interesting for the load query.
169 static AvailableValue get(Value *V, unsigned Offset = 0) {
171 Res.Val.setPointer(V);
172 Res.Val.setInt(SimpleVal);
177 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
179 Res.Val.setPointer(MI);
180 Res.Val.setInt(MemIntrin);
185 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
187 Res.Val.setPointer(LI);
188 Res.Val.setInt(LoadVal);
193 static AvailableValue getUndef() {
195 Res.Val.setPointer(nullptr);
196 Res.Val.setInt(UndefVal);
201 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
202 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
203 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
204 bool isUndefValue() const { return Val.getInt() == UndefVal; }
206 Value *getSimpleValue() const {
207 assert(isSimpleValue() && "Wrong accessor");
208 return Val.getPointer();
211 LoadInst *getCoercedLoadValue() const {
212 assert(isCoercedLoadValue() && "Wrong accessor");
213 return cast<LoadInst>(Val.getPointer());
216 MemIntrinsic *getMemIntrinValue() const {
217 assert(isMemIntrinValue() && "Wrong accessor");
218 return cast<MemIntrinsic>(Val.getPointer());
221 /// Emit code at the specified insertion point to adjust the value defined
222 /// here to the specified type. This handles various coercion cases.
223 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
227 /// Represents an AvailableValue which can be rematerialized at the end of
228 /// the associated BasicBlock.
229 struct llvm::gvn::AvailableValueInBlock {
230 /// BB - The basic block in question.
233 /// AV - The actual available value
236 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
237 AvailableValueInBlock Res;
239 Res.AV = std::move(AV);
243 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
244 unsigned Offset = 0) {
245 return get(BB, AvailableValue::get(V, Offset));
248 static AvailableValueInBlock getUndef(BasicBlock *BB) {
249 return get(BB, AvailableValue::getUndef());
252 /// Emit code at the end of this block to adjust the value defined here to
253 /// the specified type. This handles various coercion cases.
254 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
255 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
259 //===----------------------------------------------------------------------===//
260 // ValueTable Internal Functions
261 //===----------------------------------------------------------------------===//
263 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
265 e.type = I->getType();
266 e.opcode = I->getOpcode();
267 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
269 e.varargs.push_back(lookupOrAdd(*OI));
270 if (I->isCommutative()) {
271 // Ensure that commutative instructions that only differ by a permutation
272 // of their operands get the same value number by sorting the operand value
273 // numbers. Since all commutative instructions have two operands it is more
274 // efficient to sort by hand rather than using, say, std::sort.
275 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
276 if (e.varargs[0] > e.varargs[1])
277 std::swap(e.varargs[0], e.varargs[1]);
278 e.commutative = true;
281 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
282 // Sort the operand value numbers so x<y and y>x get the same value number.
283 CmpInst::Predicate Predicate = C->getPredicate();
284 if (e.varargs[0] > e.varargs[1]) {
285 std::swap(e.varargs[0], e.varargs[1]);
286 Predicate = CmpInst::getSwappedPredicate(Predicate);
288 e.opcode = (C->getOpcode() << 8) | Predicate;
289 e.commutative = true;
290 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
291 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
293 e.varargs.push_back(*II);
299 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
300 CmpInst::Predicate Predicate,
301 Value *LHS, Value *RHS) {
302 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
303 "Not a comparison!");
305 e.type = CmpInst::makeCmpResultType(LHS->getType());
306 e.varargs.push_back(lookupOrAdd(LHS));
307 e.varargs.push_back(lookupOrAdd(RHS));
309 // Sort the operand value numbers so x<y and y>x get the same value number.
310 if (e.varargs[0] > e.varargs[1]) {
311 std::swap(e.varargs[0], e.varargs[1]);
312 Predicate = CmpInst::getSwappedPredicate(Predicate);
314 e.opcode = (Opcode << 8) | Predicate;
315 e.commutative = true;
319 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
320 assert(EI && "Not an ExtractValueInst?");
322 e.type = EI->getType();
325 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
326 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
327 // EI might be an extract from one of our recognised intrinsics. If it
328 // is we'll synthesize a semantically equivalent expression instead on
329 // an extract value expression.
330 switch (I->getIntrinsicID()) {
331 case Intrinsic::sadd_with_overflow:
332 case Intrinsic::uadd_with_overflow:
333 e.opcode = Instruction::Add;
335 case Intrinsic::ssub_with_overflow:
336 case Intrinsic::usub_with_overflow:
337 e.opcode = Instruction::Sub;
339 case Intrinsic::smul_with_overflow:
340 case Intrinsic::umul_with_overflow:
341 e.opcode = Instruction::Mul;
348 // Intrinsic recognized. Grab its args to finish building the expression.
349 assert(I->getNumArgOperands() == 2 &&
350 "Expect two args for recognised intrinsics.");
351 e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
352 e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
357 // Not a recognised intrinsic. Fall back to producing an extract value
359 e.opcode = EI->getOpcode();
360 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
362 e.varargs.push_back(lookupOrAdd(*OI));
364 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
366 e.varargs.push_back(*II);
371 //===----------------------------------------------------------------------===//
372 // ValueTable External Functions
373 //===----------------------------------------------------------------------===//
375 GVN::ValueTable::ValueTable() = default;
376 GVN::ValueTable::ValueTable(const ValueTable &) = default;
377 GVN::ValueTable::ValueTable(ValueTable &&) = default;
378 GVN::ValueTable::~ValueTable() = default;
380 /// add - Insert a value into the table with a specified value number.
381 void GVN::ValueTable::add(Value *V, uint32_t num) {
382 valueNumbering.insert(std::make_pair(V, num));
383 if (PHINode *PN = dyn_cast<PHINode>(V))
384 NumberingPhi[num] = PN;
387 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
388 if (AA->doesNotAccessMemory(C)) {
389 Expression exp = createExpr(C);
390 uint32_t e = assignExpNewValueNum(exp).first;
391 valueNumbering[C] = e;
393 } else if (AA->onlyReadsMemory(C)) {
394 Expression exp = createExpr(C);
395 auto ValNum = assignExpNewValueNum(exp);
397 valueNumbering[C] = ValNum.first;
401 uint32_t e = assignExpNewValueNum(exp).first;
402 valueNumbering[C] = e;
406 MemDepResult local_dep = MD->getDependency(C);
408 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
409 valueNumbering[C] = nextValueNumber;
410 return nextValueNumber++;
413 if (local_dep.isDef()) {
414 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
416 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
417 valueNumbering[C] = nextValueNumber;
418 return nextValueNumber++;
421 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
422 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
423 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
425 valueNumbering[C] = nextValueNumber;
426 return nextValueNumber++;
430 uint32_t v = lookupOrAdd(local_cdep);
431 valueNumbering[C] = v;
436 const MemoryDependenceResults::NonLocalDepInfo &deps =
437 MD->getNonLocalCallDependency(CallSite(C));
438 // FIXME: Move the checking logic to MemDep!
439 CallInst* cdep = nullptr;
441 // Check to see if we have a single dominating call instruction that is
443 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
444 const NonLocalDepEntry *I = &deps[i];
445 if (I->getResult().isNonLocal())
448 // We don't handle non-definitions. If we already have a call, reject
449 // instruction dependencies.
450 if (!I->getResult().isDef() || cdep != nullptr) {
455 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
456 // FIXME: All duplicated with non-local case.
457 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
458 cdep = NonLocalDepCall;
467 valueNumbering[C] = nextValueNumber;
468 return nextValueNumber++;
471 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
472 valueNumbering[C] = nextValueNumber;
473 return nextValueNumber++;
475 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
476 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
477 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
479 valueNumbering[C] = nextValueNumber;
480 return nextValueNumber++;
484 uint32_t v = lookupOrAdd(cdep);
485 valueNumbering[C] = v;
488 valueNumbering[C] = nextValueNumber;
489 return nextValueNumber++;
493 /// Returns true if a value number exists for the specified value.
494 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
496 /// lookup_or_add - Returns the value number for the specified value, assigning
497 /// it a new number if it did not have one before.
498 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
499 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
500 if (VI != valueNumbering.end())
503 if (!isa<Instruction>(V)) {
504 valueNumbering[V] = nextValueNumber;
505 return nextValueNumber++;
508 Instruction* I = cast<Instruction>(V);
510 switch (I->getOpcode()) {
511 case Instruction::Call:
512 return lookupOrAddCall(cast<CallInst>(I));
513 case Instruction::Add:
514 case Instruction::FAdd:
515 case Instruction::Sub:
516 case Instruction::FSub:
517 case Instruction::Mul:
518 case Instruction::FMul:
519 case Instruction::UDiv:
520 case Instruction::SDiv:
521 case Instruction::FDiv:
522 case Instruction::URem:
523 case Instruction::SRem:
524 case Instruction::FRem:
525 case Instruction::Shl:
526 case Instruction::LShr:
527 case Instruction::AShr:
528 case Instruction::And:
529 case Instruction::Or:
530 case Instruction::Xor:
531 case Instruction::ICmp:
532 case Instruction::FCmp:
533 case Instruction::Trunc:
534 case Instruction::ZExt:
535 case Instruction::SExt:
536 case Instruction::FPToUI:
537 case Instruction::FPToSI:
538 case Instruction::UIToFP:
539 case Instruction::SIToFP:
540 case Instruction::FPTrunc:
541 case Instruction::FPExt:
542 case Instruction::PtrToInt:
543 case Instruction::IntToPtr:
544 case Instruction::BitCast:
545 case Instruction::Select:
546 case Instruction::ExtractElement:
547 case Instruction::InsertElement:
548 case Instruction::ShuffleVector:
549 case Instruction::InsertValue:
550 case Instruction::GetElementPtr:
553 case Instruction::ExtractValue:
554 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
556 case Instruction::PHI:
557 valueNumbering[V] = nextValueNumber;
558 NumberingPhi[nextValueNumber] = cast<PHINode>(V);
559 return nextValueNumber++;
561 valueNumbering[V] = nextValueNumber;
562 return nextValueNumber++;
565 uint32_t e = assignExpNewValueNum(exp).first;
566 valueNumbering[V] = e;
570 /// Returns the value number of the specified value. Fails if
571 /// the value has not yet been numbered.
572 uint32_t GVN::ValueTable::lookup(Value *V, bool Verify) const {
573 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
575 assert(VI != valueNumbering.end() && "Value not numbered?");
578 return (VI != valueNumbering.end()) ? VI->second : 0;
581 /// Returns the value number of the given comparison,
582 /// assigning it a new number if it did not have one before. Useful when
583 /// we deduced the result of a comparison, but don't immediately have an
584 /// instruction realizing that comparison to hand.
585 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
586 CmpInst::Predicate Predicate,
587 Value *LHS, Value *RHS) {
588 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
589 return assignExpNewValueNum(exp).first;
592 /// Remove all entries from the ValueTable.
593 void GVN::ValueTable::clear() {
594 valueNumbering.clear();
595 expressionNumbering.clear();
596 NumberingPhi.clear();
597 PhiTranslateTable.clear();
604 /// Remove a value from the value numbering.
605 void GVN::ValueTable::erase(Value *V) {
606 uint32_t Num = valueNumbering.lookup(V);
607 valueNumbering.erase(V);
608 // If V is PHINode, V <--> value number is an one-to-one mapping.
610 NumberingPhi.erase(Num);
613 /// verifyRemoved - Verify that the value is removed from all internal data
615 void GVN::ValueTable::verifyRemoved(const Value *V) const {
616 for (DenseMap<Value*, uint32_t>::const_iterator
617 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
618 assert(I->first != V && "Inst still occurs in value numbering map!");
622 //===----------------------------------------------------------------------===//
624 //===----------------------------------------------------------------------===//
626 PreservedAnalyses GVN::run(Function &F, FunctionAnalysisManager &AM) {
627 // FIXME: The order of evaluation of these 'getResult' calls is very
628 // significant! Re-ordering these variables will cause GVN when run alone to
629 // be less effective! We should fix memdep and basic-aa to not exhibit this
630 // behavior, but until then don't change the order here.
631 auto &AC = AM.getResult<AssumptionAnalysis>(F);
632 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
633 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
634 auto &AA = AM.getResult<AAManager>(F);
635 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
636 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
637 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
638 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep, LI, &ORE);
640 return PreservedAnalyses::all();
641 PreservedAnalyses PA;
642 PA.preserve<DominatorTreeAnalysis>();
643 PA.preserve<GlobalsAA>();
644 PA.preserve<TargetLibraryAnalysis>();
648 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
649 LLVM_DUMP_METHOD void GVN::dump(DenseMap<uint32_t, Value*>& d) const {
651 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
652 E = d.end(); I != E; ++I) {
653 errs() << I->first << "\n";
660 /// Return true if we can prove that the value
661 /// we're analyzing is fully available in the specified block. As we go, keep
662 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
663 /// map is actually a tri-state map with the following values:
664 /// 0) we know the block *is not* fully available.
665 /// 1) we know the block *is* fully available.
666 /// 2) we do not know whether the block is fully available or not, but we are
667 /// currently speculating that it will be.
668 /// 3) we are speculating for this block and have used that to speculate for
670 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
671 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
672 uint32_t RecurseDepth) {
673 if (RecurseDepth > MaxRecurseDepth)
676 // Optimistically assume that the block is fully available and check to see
677 // if we already know about this block in one lookup.
678 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
679 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
681 // If the entry already existed for this block, return the precomputed value.
683 // If this is a speculative "available" value, mark it as being used for
684 // speculation of other blocks.
685 if (IV.first->second == 2)
686 IV.first->second = 3;
687 return IV.first->second != 0;
690 // Otherwise, see if it is fully available in all predecessors.
691 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
693 // If this block has no predecessors, it isn't live-in here.
695 goto SpeculationFailure;
697 for (; PI != PE; ++PI)
698 // If the value isn't fully available in one of our predecessors, then it
699 // isn't fully available in this block either. Undo our previous
700 // optimistic assumption and bail out.
701 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
702 goto SpeculationFailure;
706 // If we get here, we found out that this is not, after
707 // all, a fully-available block. We have a problem if we speculated on this and
708 // used the speculation to mark other blocks as available.
710 char &BBVal = FullyAvailableBlocks[BB];
712 // If we didn't speculate on this, just return with it set to false.
718 // If we did speculate on this value, we could have blocks set to 1 that are
719 // incorrect. Walk the (transitive) successors of this block and mark them as
721 SmallVector<BasicBlock*, 32> BBWorklist;
722 BBWorklist.push_back(BB);
725 BasicBlock *Entry = BBWorklist.pop_back_val();
726 // Note that this sets blocks to 0 (unavailable) if they happen to not
727 // already be in FullyAvailableBlocks. This is safe.
728 char &EntryVal = FullyAvailableBlocks[Entry];
729 if (EntryVal == 0) continue; // Already unavailable.
731 // Mark as unavailable.
734 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
735 } while (!BBWorklist.empty());
740 /// Given a set of loads specified by ValuesPerBlock,
741 /// construct SSA form, allowing us to eliminate LI. This returns the value
742 /// that should be used at LI's definition site.
743 static Value *ConstructSSAForLoadSet(LoadInst *LI,
744 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
746 // Check for the fully redundant, dominating load case. In this case, we can
747 // just use the dominating value directly.
748 if (ValuesPerBlock.size() == 1 &&
749 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
751 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
752 "Dead BB dominate this block");
753 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
756 // Otherwise, we have to construct SSA form.
757 SmallVector<PHINode*, 8> NewPHIs;
758 SSAUpdater SSAUpdate(&NewPHIs);
759 SSAUpdate.Initialize(LI->getType(), LI->getName());
761 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
762 BasicBlock *BB = AV.BB;
764 if (SSAUpdate.HasValueForBlock(BB))
767 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
770 // Perform PHI construction.
771 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
774 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
775 Instruction *InsertPt,
778 Type *LoadTy = LI->getType();
779 const DataLayout &DL = LI->getModule()->getDataLayout();
780 if (isSimpleValue()) {
781 Res = getSimpleValue();
782 if (Res->getType() != LoadTy) {
783 Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
785 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
786 << *getSimpleValue() << '\n'
787 << *Res << '\n' << "\n\n\n");
789 } else if (isCoercedLoadValue()) {
790 LoadInst *Load = getCoercedLoadValue();
791 if (Load->getType() == LoadTy && Offset == 0) {
794 Res = getLoadValueForLoad(Load, Offset, LoadTy, InsertPt, DL);
795 // We would like to use gvn.markInstructionForDeletion here, but we can't
796 // because the load is already memoized into the leader map table that GVN
797 // tracks. It is potentially possible to remove the load from the table,
798 // but then there all of the operations based on it would need to be
799 // rehashed. Just leave the dead load around.
800 gvn.getMemDep().removeInstruction(Load);
801 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
802 << *getCoercedLoadValue() << '\n'
806 } else if (isMemIntrinValue()) {
807 Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
809 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
810 << " " << *getMemIntrinValue() << '\n'
811 << *Res << '\n' << "\n\n\n");
813 assert(isUndefValue() && "Should be UndefVal");
814 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
815 return UndefValue::get(LoadTy);
817 assert(Res && "failed to materialize?");
821 static bool isLifetimeStart(const Instruction *Inst) {
822 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
823 return II->getIntrinsicID() == Intrinsic::lifetime_start;
827 /// \brief Try to locate the three instruction involved in a missed
828 /// load-elimination case that is due to an intervening store.
829 static void reportMayClobberedLoad(LoadInst *LI, MemDepResult DepInfo,
831 OptimizationRemarkEmitter *ORE) {
834 User *OtherAccess = nullptr;
836 OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered", LI);
837 R << "load of type " << NV("Type", LI->getType()) << " not eliminated"
840 for (auto *U : LI->getPointerOperand()->users())
841 if (U != LI && (isa<LoadInst>(U) || isa<StoreInst>(U)) &&
842 DT->dominates(cast<Instruction>(U), LI)) {
843 // FIXME: for now give up if there are multiple memory accesses that
844 // dominate the load. We need further analysis to decide which one is
845 // that we're forwarding from.
847 OtherAccess = nullptr;
853 R << " in favor of " << NV("OtherAccess", OtherAccess);
855 R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst());
860 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
861 Value *Address, AvailableValue &Res) {
862 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
863 "expected a local dependence");
864 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
866 const DataLayout &DL = LI->getModule()->getDataLayout();
868 if (DepInfo.isClobber()) {
869 // If the dependence is to a store that writes to a superset of the bits
870 // read by the load, we can extract the bits we need for the load from the
872 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
873 // Can't forward from non-atomic to atomic without violating memory model.
874 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
876 analyzeLoadFromClobberingStore(LI->getType(), Address, DepSI, DL);
878 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
884 // Check to see if we have something like this:
887 // if we have this, replace the later with an extraction from the former.
888 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
889 // If this is a clobber and L is the first instruction in its block, then
890 // we have the first instruction in the entry block.
891 // Can't forward from non-atomic to atomic without violating memory model.
892 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
894 analyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
897 Res = AvailableValue::getLoad(DepLI, Offset);
903 // If the clobbering value is a memset/memcpy/memmove, see if we can
904 // forward a value on from it.
905 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
906 if (Address && !LI->isAtomic()) {
907 int Offset = analyzeLoadFromClobberingMemInst(LI->getType(), Address,
910 Res = AvailableValue::getMI(DepMI, Offset);
915 // Nothing known about this clobber, have to be conservative
917 // fast print dep, using operator<< on instruction is too slow.
918 dbgs() << "GVN: load ";
919 LI->printAsOperand(dbgs());
920 Instruction *I = DepInfo.getInst();
921 dbgs() << " is clobbered by " << *I << '\n';
923 if (ORE->allowExtraAnalysis(DEBUG_TYPE))
924 reportMayClobberedLoad(LI, DepInfo, DT, ORE);
928 assert(DepInfo.isDef() && "follows from above");
930 Instruction *DepInst = DepInfo.getInst();
932 // Loading the allocation -> undef.
933 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
934 // Loading immediately after lifetime begin -> undef.
935 isLifetimeStart(DepInst)) {
936 Res = AvailableValue::get(UndefValue::get(LI->getType()));
940 // Loading from calloc (which zero initializes memory) -> zero
941 if (isCallocLikeFn(DepInst, TLI)) {
942 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
946 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
947 // Reject loads and stores that are to the same address but are of
948 // different types if we have to. If the stored value is larger or equal to
949 // the loaded value, we can reuse it.
950 if (S->getValueOperand()->getType() != LI->getType() &&
951 !canCoerceMustAliasedValueToLoad(S->getValueOperand(),
955 // Can't forward from non-atomic to atomic without violating memory model.
956 if (S->isAtomic() < LI->isAtomic())
959 Res = AvailableValue::get(S->getValueOperand());
963 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
964 // If the types mismatch and we can't handle it, reject reuse of the load.
965 // If the stored value is larger or equal to the loaded value, we can reuse
967 if (LD->getType() != LI->getType() &&
968 !canCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
971 // Can't forward from non-atomic to atomic without violating memory model.
972 if (LD->isAtomic() < LI->isAtomic())
975 Res = AvailableValue::getLoad(LD);
979 // Unknown def - must be conservative
981 // fast print dep, using operator<< on instruction is too slow.
982 dbgs() << "GVN: load ";
983 LI->printAsOperand(dbgs());
984 dbgs() << " has unknown def " << *DepInst << '\n';
989 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
990 AvailValInBlkVect &ValuesPerBlock,
991 UnavailBlkVect &UnavailableBlocks) {
992 // Filter out useless results (non-locals, etc). Keep track of the blocks
993 // where we have a value available in repl, also keep track of whether we see
994 // dependencies that produce an unknown value for the load (such as a call
995 // that could potentially clobber the load).
996 unsigned NumDeps = Deps.size();
997 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
998 BasicBlock *DepBB = Deps[i].getBB();
999 MemDepResult DepInfo = Deps[i].getResult();
1001 if (DeadBlocks.count(DepBB)) {
1002 // Dead dependent mem-op disguise as a load evaluating the same value
1003 // as the load in question.
1004 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1008 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1009 UnavailableBlocks.push_back(DepBB);
1013 // The address being loaded in this non-local block may not be the same as
1014 // the pointer operand of the load if PHI translation occurs. Make sure
1015 // to consider the right address.
1016 Value *Address = Deps[i].getAddress();
1019 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1020 // subtlety: because we know this was a non-local dependency, we know
1021 // it's safe to materialize anywhere between the instruction within
1022 // DepInfo and the end of it's block.
1023 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1026 UnavailableBlocks.push_back(DepBB);
1030 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1031 "post condition violation");
1034 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1035 UnavailBlkVect &UnavailableBlocks) {
1036 // Okay, we have *some* definitions of the value. This means that the value
1037 // is available in some of our (transitive) predecessors. Lets think about
1038 // doing PRE of this load. This will involve inserting a new load into the
1039 // predecessor when it's not available. We could do this in general, but
1040 // prefer to not increase code size. As such, we only do this when we know
1041 // that we only have to insert *one* load (which means we're basically moving
1042 // the load, not inserting a new one).
1044 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1045 UnavailableBlocks.end());
1047 // Let's find the first basic block with more than one predecessor. Walk
1048 // backwards through predecessors if needed.
1049 BasicBlock *LoadBB = LI->getParent();
1050 BasicBlock *TmpBB = LoadBB;
1052 while (TmpBB->getSinglePredecessor()) {
1053 TmpBB = TmpBB->getSinglePredecessor();
1054 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1056 if (Blockers.count(TmpBB))
1059 // If any of these blocks has more than one successor (i.e. if the edge we
1060 // just traversed was critical), then there are other paths through this
1061 // block along which the load may not be anticipated. Hoisting the load
1062 // above this block would be adding the load to execution paths along
1063 // which it was not previously executed.
1064 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1071 // Check to see how many predecessors have the loaded value fully
1073 MapVector<BasicBlock *, Value *> PredLoads;
1074 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1075 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1076 FullyAvailableBlocks[AV.BB] = true;
1077 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1078 FullyAvailableBlocks[UnavailableBB] = false;
1080 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1081 for (BasicBlock *Pred : predecessors(LoadBB)) {
1082 // If any predecessor block is an EH pad that does not allow non-PHI
1083 // instructions before the terminator, we can't PRE the load.
1084 if (Pred->getTerminator()->isEHPad()) {
1086 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1087 << Pred->getName() << "': " << *LI << '\n');
1091 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1095 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1096 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1097 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1098 << Pred->getName() << "': " << *LI << '\n');
1102 if (LoadBB->isEHPad()) {
1104 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1105 << Pred->getName() << "': " << *LI << '\n');
1109 CriticalEdgePred.push_back(Pred);
1111 // Only add the predecessors that will not be split for now.
1112 PredLoads[Pred] = nullptr;
1116 // Decide whether PRE is profitable for this load.
1117 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1118 assert(NumUnavailablePreds != 0 &&
1119 "Fully available value should already be eliminated!");
1121 // If this load is unavailable in multiple predecessors, reject it.
1122 // FIXME: If we could restructure the CFG, we could make a common pred with
1123 // all the preds that don't have an available LI and insert a new load into
1125 if (NumUnavailablePreds != 1)
1128 // Split critical edges, and update the unavailable predecessors accordingly.
1129 for (BasicBlock *OrigPred : CriticalEdgePred) {
1130 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1131 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1132 PredLoads[NewPred] = nullptr;
1133 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1134 << LoadBB->getName() << '\n');
1137 // Check if the load can safely be moved to all the unavailable predecessors.
1138 bool CanDoPRE = true;
1139 const DataLayout &DL = LI->getModule()->getDataLayout();
1140 SmallVector<Instruction*, 8> NewInsts;
1141 for (auto &PredLoad : PredLoads) {
1142 BasicBlock *UnavailablePred = PredLoad.first;
1144 // Do PHI translation to get its value in the predecessor if necessary. The
1145 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1147 // If all preds have a single successor, then we know it is safe to insert
1148 // the load on the pred (?!?), so we can insert code to materialize the
1149 // pointer if it is not available.
1150 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1151 Value *LoadPtr = nullptr;
1152 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1155 // If we couldn't find or insert a computation of this phi translated value,
1158 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1159 << *LI->getPointerOperand() << "\n");
1164 PredLoad.second = LoadPtr;
1168 while (!NewInsts.empty()) {
1169 Instruction *I = NewInsts.pop_back_val();
1170 if (MD) MD->removeInstruction(I);
1171 I->eraseFromParent();
1173 // HINT: Don't revert the edge-splitting as following transformation may
1174 // also need to split these critical edges.
1175 return !CriticalEdgePred.empty();
1178 // Okay, we can eliminate this load by inserting a reload in the predecessor
1179 // and using PHI construction to get the value in the other predecessors, do
1181 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1182 DEBUG(if (!NewInsts.empty())
1183 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1184 << *NewInsts.back() << '\n');
1186 // Assign value numbers to the new instructions.
1187 for (Instruction *I : NewInsts) {
1188 // Instructions that have been inserted in predecessor(s) to materialize
1189 // the load address do not retain their original debug locations. Doing
1190 // so could lead to confusing (but correct) source attributions.
1191 // FIXME: How do we retain source locations without causing poor debugging
1193 I->setDebugLoc(DebugLoc());
1195 // FIXME: We really _ought_ to insert these value numbers into their
1196 // parent's availability map. However, in doing so, we risk getting into
1197 // ordering issues. If a block hasn't been processed yet, we would be
1198 // marking a value as AVAIL-IN, which isn't what we intend.
1202 for (const auto &PredLoad : PredLoads) {
1203 BasicBlock *UnavailablePred = PredLoad.first;
1204 Value *LoadPtr = PredLoad.second;
1206 auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1207 LI->isVolatile(), LI->getAlignment(),
1208 LI->getOrdering(), LI->getSyncScopeID(),
1209 UnavailablePred->getTerminator());
1210 NewLoad->setDebugLoc(LI->getDebugLoc());
1212 // Transfer the old load's AA tags to the new load.
1214 LI->getAAMetadata(Tags);
1216 NewLoad->setAAMetadata(Tags);
1218 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1219 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1220 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1221 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1222 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1223 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1225 // We do not propagate the old load's debug location, because the new
1226 // load now lives in a different BB, and we want to avoid a jumpy line
1228 // FIXME: How do we retain source locations without causing poor debugging
1231 // Add the newly created load.
1232 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1234 MD->invalidateCachedPointerInfo(LoadPtr);
1235 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1238 // Perform PHI construction.
1239 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1240 LI->replaceAllUsesWith(V);
1241 if (isa<PHINode>(V))
1243 if (Instruction *I = dyn_cast<Instruction>(V))
1244 I->setDebugLoc(LI->getDebugLoc());
1245 if (V->getType()->isPtrOrPtrVectorTy())
1246 MD->invalidateCachedPointerInfo(V);
1247 markInstructionForDeletion(LI);
1248 ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadPRE", LI)
1249 << "load eliminated by PRE");
1254 static void reportLoadElim(LoadInst *LI, Value *AvailableValue,
1255 OptimizationRemarkEmitter *ORE) {
1256 using namespace ore;
1258 ORE->emit(OptimizationRemark(DEBUG_TYPE, "LoadElim", LI)
1259 << "load of type " << NV("Type", LI->getType()) << " eliminated"
1260 << setExtraArgs() << " in favor of "
1261 << NV("InfavorOfValue", AvailableValue));
1264 /// Attempt to eliminate a load whose dependencies are
1265 /// non-local by performing PHI construction.
1266 bool GVN::processNonLocalLoad(LoadInst *LI) {
1267 // non-local speculations are not allowed under asan.
1268 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1271 // Step 1: Find the non-local dependencies of the load.
1273 MD->getNonLocalPointerDependency(LI, Deps);
1275 // If we had to process more than one hundred blocks to find the
1276 // dependencies, this load isn't worth worrying about. Optimizing
1277 // it will be too expensive.
1278 unsigned NumDeps = Deps.size();
1282 // If we had a phi translation failure, we'll have a single entry which is a
1283 // clobber in the current block. Reject this early.
1285 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1287 dbgs() << "GVN: non-local load ";
1288 LI->printAsOperand(dbgs());
1289 dbgs() << " has unknown dependencies\n";
1294 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1295 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1296 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1297 OE = GEP->idx_end();
1299 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1300 performScalarPRE(I);
1303 // Step 2: Analyze the availability of the load
1304 AvailValInBlkVect ValuesPerBlock;
1305 UnavailBlkVect UnavailableBlocks;
1306 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1308 // If we have no predecessors that produce a known value for this load, exit
1310 if (ValuesPerBlock.empty())
1313 // Step 3: Eliminate fully redundancy.
1315 // If all of the instructions we depend on produce a known value for this
1316 // load, then it is fully redundant and we can use PHI insertion to compute
1317 // its value. Insert PHIs and remove the fully redundant value now.
1318 if (UnavailableBlocks.empty()) {
1319 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1321 // Perform PHI construction.
1322 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1323 LI->replaceAllUsesWith(V);
1325 if (isa<PHINode>(V))
1327 if (Instruction *I = dyn_cast<Instruction>(V))
1328 // If instruction I has debug info, then we should not update it.
1329 // Also, if I has a null DebugLoc, then it is still potentially incorrect
1330 // to propagate LI's DebugLoc because LI may not post-dominate I.
1331 if (LI->getDebugLoc() && LI->getParent() == I->getParent())
1332 I->setDebugLoc(LI->getDebugLoc());
1333 if (V->getType()->isPtrOrPtrVectorTy())
1334 MD->invalidateCachedPointerInfo(V);
1335 markInstructionForDeletion(LI);
1337 reportLoadElim(LI, V, ORE);
1341 // Step 4: Eliminate partial redundancy.
1342 if (!EnablePRE || !EnableLoadPRE)
1345 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1348 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1349 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1350 "This function can only be called with llvm.assume intrinsic");
1351 Value *V = IntrinsicI->getArgOperand(0);
1353 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1354 if (Cond->isZero()) {
1355 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1356 // Insert a new store to null instruction before the load to indicate that
1357 // this code is not reachable. FIXME: We could insert unreachable
1358 // instruction directly because we can modify the CFG.
1359 new StoreInst(UndefValue::get(Int8Ty),
1360 Constant::getNullValue(Int8Ty->getPointerTo()),
1363 markInstructionForDeletion(IntrinsicI);
1367 Constant *True = ConstantInt::getTrue(V->getContext());
1368 bool Changed = false;
1370 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1371 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1373 // This property is only true in dominated successors, propagateEquality
1374 // will check dominance for us.
1375 Changed |= propagateEquality(V, True, Edge, false);
1378 // We can replace assume value with true, which covers cases like this:
1379 // call void @llvm.assume(i1 %cmp)
1380 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1381 ReplaceWithConstMap[V] = True;
1383 // If one of *cmp *eq operand is const, adding it to map will cover this:
1384 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1385 // call void @llvm.assume(i1 %cmp)
1386 // ret float %0 ; will change it to ret float 3.000000e+00
1387 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1388 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1389 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1390 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1391 CmpI->getFastMathFlags().noNaNs())) {
1392 Value *CmpLHS = CmpI->getOperand(0);
1393 Value *CmpRHS = CmpI->getOperand(1);
1394 if (isa<Constant>(CmpLHS))
1395 std::swap(CmpLHS, CmpRHS);
1396 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1398 // If only one operand is constant.
1399 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1400 ReplaceWithConstMap[CmpLHS] = RHSConst;
1406 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1407 auto *ReplInst = dyn_cast<Instruction>(Repl);
1411 // Patch the replacement so that it is not more restrictive than the value
1413 // Note that if 'I' is a load being replaced by some operation,
1414 // for example, by an arithmetic operation, then andIRFlags()
1415 // would just erase all math flags from the original arithmetic
1416 // operation, which is clearly not wanted and not needed.
1417 if (!isa<LoadInst>(I))
1418 ReplInst->andIRFlags(I);
1420 // FIXME: If both the original and replacement value are part of the
1421 // same control-flow region (meaning that the execution of one
1422 // guarantees the execution of the other), then we can combine the
1423 // noalias scopes here and do better than the general conservative
1424 // answer used in combineMetadata().
1426 // In general, GVN unifies expressions over different control-flow
1427 // regions, and so we need a conservative combination of the noalias
1429 static const unsigned KnownIDs[] = {
1430 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1431 LLVMContext::MD_noalias, LLVMContext::MD_range,
1432 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1433 LLVMContext::MD_invariant_group};
1434 combineMetadata(ReplInst, I, KnownIDs);
1437 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1438 patchReplacementInstruction(I, Repl);
1439 I->replaceAllUsesWith(Repl);
1442 /// Attempt to eliminate a load, first by eliminating it
1443 /// locally, and then attempting non-local elimination if that fails.
1444 bool GVN::processLoad(LoadInst *L) {
1448 // This code hasn't been audited for ordered or volatile memory access
1449 if (!L->isUnordered())
1452 if (L->use_empty()) {
1453 markInstructionForDeletion(L);
1457 // ... to a pointer that has been loaded from before...
1458 MemDepResult Dep = MD->getDependency(L);
1460 // If it is defined in another block, try harder.
1461 if (Dep.isNonLocal())
1462 return processNonLocalLoad(L);
1464 // Only handle the local case below
1465 if (!Dep.isDef() && !Dep.isClobber()) {
1466 // This might be a NonFuncLocal or an Unknown
1468 // fast print dep, using operator<< on instruction is too slow.
1469 dbgs() << "GVN: load ";
1470 L->printAsOperand(dbgs());
1471 dbgs() << " has unknown dependence\n";
1477 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1478 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1480 // Replace the load!
1481 patchAndReplaceAllUsesWith(L, AvailableValue);
1482 markInstructionForDeletion(L);
1484 reportLoadElim(L, AvailableValue, ORE);
1485 // Tell MDA to rexamine the reused pointer since we might have more
1486 // information after forwarding it.
1487 if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy())
1488 MD->invalidateCachedPointerInfo(AvailableValue);
1495 /// Return a pair the first field showing the value number of \p Exp and the
1496 /// second field showing whether it is a value number newly created.
1497 std::pair<uint32_t, bool>
1498 GVN::ValueTable::assignExpNewValueNum(Expression &Exp) {
1499 uint32_t &e = expressionNumbering[Exp];
1500 bool CreateNewValNum = !e;
1501 if (CreateNewValNum) {
1502 Expressions.push_back(Exp);
1503 if (ExprIdx.size() < nextValueNumber + 1)
1504 ExprIdx.resize(nextValueNumber * 2);
1505 e = nextValueNumber;
1506 ExprIdx[nextValueNumber++] = nextExprNumber++;
1508 return {e, CreateNewValNum};
1511 /// Return whether all the values related with the same \p num are
1512 /// defined in \p BB.
1513 bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB,
1515 LeaderTableEntry *Vals = &Gvn.LeaderTable[Num];
1516 while (Vals && Vals->BB == BB)
1521 /// Wrap phiTranslateImpl to provide caching functionality.
1522 uint32_t GVN::ValueTable::phiTranslate(const BasicBlock *Pred,
1523 const BasicBlock *PhiBlock, uint32_t Num,
1525 auto FindRes = PhiTranslateTable.find({Num, Pred});
1526 if (FindRes != PhiTranslateTable.end())
1527 return FindRes->second;
1528 uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn);
1529 PhiTranslateTable.insert({{Num, Pred}, NewNum});
1533 /// Translate value number \p Num using phis, so that it has the values of
1535 uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred,
1536 const BasicBlock *PhiBlock,
1537 uint32_t Num, GVN &Gvn) {
1538 if (PHINode *PN = NumberingPhi[Num]) {
1539 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
1540 if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred)
1541 if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false))
1547 // If there is any value related with Num is defined in a BB other than
1548 // PhiBlock, it cannot depend on a phi in PhiBlock without going through
1549 // a backedge. We can do an early exit in that case to save compile time.
1550 if (!areAllValsInBB(Num, PhiBlock, Gvn))
1553 if (Num >= ExprIdx.size() || ExprIdx[Num] == 0)
1555 Expression Exp = Expressions[ExprIdx[Num]];
1557 for (unsigned i = 0; i < Exp.varargs.size(); i++) {
1558 // For InsertValue and ExtractValue, some varargs are index numbers
1559 // instead of value numbers. Those index numbers should not be
1561 if ((i > 1 && Exp.opcode == Instruction::InsertValue) ||
1562 (i > 0 && Exp.opcode == Instruction::ExtractValue))
1564 Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn);
1567 if (Exp.commutative) {
1568 assert(Exp.varargs.size() == 2 && "Unsupported commutative expression!");
1569 if (Exp.varargs[0] > Exp.varargs[1]) {
1570 std::swap(Exp.varargs[0], Exp.varargs[1]);
1571 uint32_t Opcode = Exp.opcode >> 8;
1572 if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
1573 Exp.opcode = (Opcode << 8) |
1574 CmpInst::getSwappedPredicate(
1575 static_cast<CmpInst::Predicate>(Exp.opcode & 255));
1579 if (uint32_t NewNum = expressionNumbering[Exp])
1584 /// Erase stale entry from phiTranslate cache so phiTranslate can be computed
1586 void GVN::ValueTable::eraseTranslateCacheEntry(uint32_t Num,
1587 const BasicBlock &CurrBlock) {
1588 for (const BasicBlock *Pred : predecessors(&CurrBlock)) {
1589 auto FindRes = PhiTranslateTable.find({Num, Pred});
1590 if (FindRes != PhiTranslateTable.end())
1591 PhiTranslateTable.erase(FindRes);
1595 // In order to find a leader for a given value number at a
1596 // specific basic block, we first obtain the list of all Values for that number,
1597 // and then scan the list to find one whose block dominates the block in
1598 // question. This is fast because dominator tree queries consist of only
1599 // a few comparisons of DFS numbers.
1600 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1601 LeaderTableEntry Vals = LeaderTable[num];
1602 if (!Vals.Val) return nullptr;
1604 Value *Val = nullptr;
1605 if (DT->dominates(Vals.BB, BB)) {
1607 if (isa<Constant>(Val)) return Val;
1610 LeaderTableEntry* Next = Vals.Next;
1612 if (DT->dominates(Next->BB, BB)) {
1613 if (isa<Constant>(Next->Val)) return Next->Val;
1614 if (!Val) Val = Next->Val;
1623 /// There is an edge from 'Src' to 'Dst'. Return
1624 /// true if every path from the entry block to 'Dst' passes via this edge. In
1625 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1626 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1627 DominatorTree *DT) {
1628 // While in theory it is interesting to consider the case in which Dst has
1629 // more than one predecessor, because Dst might be part of a loop which is
1630 // only reachable from Src, in practice it is pointless since at the time
1631 // GVN runs all such loops have preheaders, which means that Dst will have
1632 // been changed to have only one predecessor, namely Src.
1633 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1634 assert((!Pred || Pred == E.getStart()) &&
1635 "No edge between these basic blocks!");
1636 return Pred != nullptr;
1639 void GVN::assignBlockRPONumber(Function &F) {
1640 uint32_t NextBlockNumber = 1;
1641 ReversePostOrderTraversal<Function *> RPOT(&F);
1642 for (BasicBlock *BB : RPOT)
1643 BlockRPONumber[BB] = NextBlockNumber++;
1646 // Tries to replace instruction with const, using information from
1647 // ReplaceWithConstMap.
1648 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1649 bool Changed = false;
1650 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1651 Value *Operand = Instr->getOperand(OpNum);
1652 auto it = ReplaceWithConstMap.find(Operand);
1653 if (it != ReplaceWithConstMap.end()) {
1654 assert(!isa<Constant>(Operand) &&
1655 "Replacing constants with constants is invalid");
1656 DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1657 << " in instruction " << *Instr << '\n');
1658 Instr->setOperand(OpNum, it->second);
1665 /// The given values are known to be equal in every block
1666 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1667 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1668 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1669 /// value starting from the end of Root.Start.
1670 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1671 bool DominatesByEdge) {
1672 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1673 Worklist.push_back(std::make_pair(LHS, RHS));
1674 bool Changed = false;
1675 // For speed, compute a conservative fast approximation to
1676 // DT->dominates(Root, Root.getEnd());
1677 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1679 while (!Worklist.empty()) {
1680 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1681 LHS = Item.first; RHS = Item.second;
1685 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1687 // Don't try to propagate equalities between constants.
1688 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1691 // Prefer a constant on the right-hand side, or an Argument if no constants.
1692 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1693 std::swap(LHS, RHS);
1694 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1696 // If there is no obvious reason to prefer the left-hand side over the
1697 // right-hand side, ensure the longest lived term is on the right-hand side,
1698 // so the shortest lived term will be replaced by the longest lived.
1699 // This tends to expose more simplifications.
1700 uint32_t LVN = VN.lookupOrAdd(LHS);
1701 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1702 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1703 // Move the 'oldest' value to the right-hand side, using the value number
1704 // as a proxy for age.
1705 uint32_t RVN = VN.lookupOrAdd(RHS);
1707 std::swap(LHS, RHS);
1712 // If value numbering later sees that an instruction in the scope is equal
1713 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1714 // the invariant that instructions only occur in the leader table for their
1715 // own value number (this is used by removeFromLeaderTable), do not do this
1716 // if RHS is an instruction (if an instruction in the scope is morphed into
1717 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1718 // using the leader table is about compiling faster, not optimizing better).
1719 // The leader table only tracks basic blocks, not edges. Only add to if we
1720 // have the simple case where the edge dominates the end.
1721 if (RootDominatesEnd && !isa<Instruction>(RHS))
1722 addToLeaderTable(LVN, RHS, Root.getEnd());
1724 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1725 // LHS always has at least one use that is not dominated by Root, this will
1726 // never do anything if LHS has only one use.
1727 if (!LHS->hasOneUse()) {
1728 unsigned NumReplacements =
1730 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1731 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1733 Changed |= NumReplacements > 0;
1734 NumGVNEqProp += NumReplacements;
1737 // Now try to deduce additional equalities from this one. For example, if
1738 // the known equality was "(A != B)" == "false" then it follows that A and B
1739 // are equal in the scope. Only boolean equalities with an explicit true or
1740 // false RHS are currently supported.
1741 if (!RHS->getType()->isIntegerTy(1))
1742 // Not a boolean equality - bail out.
1744 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1746 // RHS neither 'true' nor 'false' - bail out.
1748 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1749 bool isKnownTrue = CI->isMinusOne();
1750 bool isKnownFalse = !isKnownTrue;
1752 // If "A && B" is known true then both A and B are known true. If "A || B"
1753 // is known false then both A and B are known false.
1755 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1756 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1757 Worklist.push_back(std::make_pair(A, RHS));
1758 Worklist.push_back(std::make_pair(B, RHS));
1762 // If we are propagating an equality like "(A == B)" == "true" then also
1763 // propagate the equality A == B. When propagating a comparison such as
1764 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1765 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1766 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1768 // If "A == B" is known true, or "A != B" is known false, then replace
1769 // A with B everywhere in the scope.
1770 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1771 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1772 Worklist.push_back(std::make_pair(Op0, Op1));
1774 // Handle the floating point versions of equality comparisons too.
1775 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1776 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1778 // Floating point -0.0 and 0.0 compare equal, so we can only
1779 // propagate values if we know that we have a constant and that
1780 // its value is non-zero.
1782 // FIXME: We should do this optimization if 'no signed zeros' is
1783 // applicable via an instruction-level fast-math-flag or some other
1784 // indicator that relaxed FP semantics are being used.
1786 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
1787 Worklist.push_back(std::make_pair(Op0, Op1));
1790 // If "A >= B" is known true, replace "A < B" with false everywhere.
1791 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
1792 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
1793 // Since we don't have the instruction "A < B" immediately to hand, work
1794 // out the value number that it would have and use that to find an
1795 // appropriate instruction (if any).
1796 uint32_t NextNum = VN.getNextUnusedValueNumber();
1797 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
1798 // If the number we were assigned was brand new then there is no point in
1799 // looking for an instruction realizing it: there cannot be one!
1800 if (Num < NextNum) {
1801 Value *NotCmp = findLeader(Root.getEnd(), Num);
1802 if (NotCmp && isa<Instruction>(NotCmp)) {
1803 unsigned NumReplacements =
1805 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
1806 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
1808 Changed |= NumReplacements > 0;
1809 NumGVNEqProp += NumReplacements;
1812 // Ensure that any instruction in scope that gets the "A < B" value number
1813 // is replaced with false.
1814 // The leader table only tracks basic blocks, not edges. Only add to if we
1815 // have the simple case where the edge dominates the end.
1816 if (RootDominatesEnd)
1817 addToLeaderTable(Num, NotVal, Root.getEnd());
1826 /// When calculating availability, handle an instruction
1827 /// by inserting it into the appropriate sets
1828 bool GVN::processInstruction(Instruction *I) {
1829 // Ignore dbg info intrinsics.
1830 if (isa<DbgInfoIntrinsic>(I))
1833 // If the instruction can be easily simplified then do so now in preference
1834 // to value numbering it. Value numbering often exposes redundancies, for
1835 // example if it determines that %y is equal to %x then the instruction
1836 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1837 const DataLayout &DL = I->getModule()->getDataLayout();
1838 if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) {
1839 bool Changed = false;
1840 if (!I->use_empty()) {
1841 I->replaceAllUsesWith(V);
1844 if (isInstructionTriviallyDead(I, TLI)) {
1845 markInstructionForDeletion(I);
1849 if (MD && V->getType()->isPtrOrPtrVectorTy())
1850 MD->invalidateCachedPointerInfo(V);
1856 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
1857 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
1858 return processAssumeIntrinsic(IntrinsicI);
1860 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1861 if (processLoad(LI))
1864 unsigned Num = VN.lookupOrAdd(LI);
1865 addToLeaderTable(Num, LI, LI->getParent());
1869 // For conditional branches, we can perform simple conditional propagation on
1870 // the condition value itself.
1871 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1872 if (!BI->isConditional())
1875 if (isa<Constant>(BI->getCondition()))
1876 return processFoldableCondBr(BI);
1878 Value *BranchCond = BI->getCondition();
1879 BasicBlock *TrueSucc = BI->getSuccessor(0);
1880 BasicBlock *FalseSucc = BI->getSuccessor(1);
1881 // Avoid multiple edges early.
1882 if (TrueSucc == FalseSucc)
1885 BasicBlock *Parent = BI->getParent();
1886 bool Changed = false;
1888 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
1889 BasicBlockEdge TrueE(Parent, TrueSucc);
1890 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
1892 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
1893 BasicBlockEdge FalseE(Parent, FalseSucc);
1894 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
1899 // For switches, propagate the case values into the case destinations.
1900 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
1901 Value *SwitchCond = SI->getCondition();
1902 BasicBlock *Parent = SI->getParent();
1903 bool Changed = false;
1905 // Remember how many outgoing edges there are to every successor.
1906 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
1907 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
1908 ++SwitchEdges[SI->getSuccessor(i)];
1910 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
1912 BasicBlock *Dst = i->getCaseSuccessor();
1913 // If there is only a single edge, propagate the case value into it.
1914 if (SwitchEdges.lookup(Dst) == 1) {
1915 BasicBlockEdge E(Parent, Dst);
1916 Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true);
1922 // Instructions with void type don't return a value, so there's
1923 // no point in trying to find redundancies in them.
1924 if (I->getType()->isVoidTy())
1927 uint32_t NextNum = VN.getNextUnusedValueNumber();
1928 unsigned Num = VN.lookupOrAdd(I);
1930 // Allocations are always uniquely numbered, so we can save time and memory
1931 // by fast failing them.
1932 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1933 addToLeaderTable(Num, I, I->getParent());
1937 // If the number we were assigned was a brand new VN, then we don't
1938 // need to do a lookup to see if the number already exists
1939 // somewhere in the domtree: it can't!
1940 if (Num >= NextNum) {
1941 addToLeaderTable(Num, I, I->getParent());
1945 // Perform fast-path value-number based elimination of values inherited from
1947 Value *Repl = findLeader(I->getParent(), Num);
1949 // Failure, just remember this instance for future use.
1950 addToLeaderTable(Num, I, I->getParent());
1952 } else if (Repl == I) {
1953 // If I was the result of a shortcut PRE, it might already be in the table
1954 // and the best replacement for itself. Nothing to do.
1959 patchAndReplaceAllUsesWith(I, Repl);
1960 if (MD && Repl->getType()->isPtrOrPtrVectorTy())
1961 MD->invalidateCachedPointerInfo(Repl);
1962 markInstructionForDeletion(I);
1966 /// runOnFunction - This is the main transformation entry point for a function.
1967 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
1968 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
1969 MemoryDependenceResults *RunMD, LoopInfo *LI,
1970 OptimizationRemarkEmitter *RunORE) {
1975 VN.setAliasAnalysis(&RunAA);
1980 bool Changed = false;
1981 bool ShouldContinue = true;
1983 // Merge unconditional branches, allowing PRE to catch more
1984 // optimization opportunities.
1985 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1986 BasicBlock *BB = &*FI++;
1988 bool removedBlock = MergeBlockIntoPredecessor(BB, DT, LI, MD);
1992 Changed |= removedBlock;
1995 unsigned Iteration = 0;
1996 while (ShouldContinue) {
1997 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1998 ShouldContinue = iterateOnFunction(F);
1999 Changed |= ShouldContinue;
2004 // Fabricate val-num for dead-code in order to suppress assertion in
2006 assignValNumForDeadCode();
2007 assignBlockRPONumber(F);
2008 bool PREChanged = true;
2009 while (PREChanged) {
2010 PREChanged = performPRE(F);
2011 Changed |= PREChanged;
2015 // FIXME: Should perform GVN again after PRE does something. PRE can move
2016 // computations into blocks where they become fully redundant. Note that
2017 // we can't do this until PRE's critical edge splitting updates memdep.
2018 // Actually, when this happens, we should just fully integrate PRE into GVN.
2020 cleanupGlobalSets();
2021 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2028 bool GVN::processBlock(BasicBlock *BB) {
2029 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2030 // (and incrementing BI before processing an instruction).
2031 assert(InstrsToErase.empty() &&
2032 "We expect InstrsToErase to be empty across iterations");
2033 if (DeadBlocks.count(BB))
2036 // Clearing map before every BB because it can be used only for single BB.
2037 ReplaceWithConstMap.clear();
2038 bool ChangedFunction = false;
2040 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2042 if (!ReplaceWithConstMap.empty())
2043 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2044 ChangedFunction |= processInstruction(&*BI);
2046 if (InstrsToErase.empty()) {
2051 // If we need some instructions deleted, do it now.
2052 NumGVNInstr += InstrsToErase.size();
2054 // Avoid iterator invalidation.
2055 bool AtStart = BI == BB->begin();
2059 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2060 E = InstrsToErase.end(); I != E; ++I) {
2061 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2062 if (MD) MD->removeInstruction(*I);
2063 DEBUG(verifyRemoved(*I));
2064 (*I)->eraseFromParent();
2066 InstrsToErase.clear();
2074 return ChangedFunction;
2077 // Instantiate an expression in a predecessor that lacked it.
2078 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2079 BasicBlock *Curr, unsigned int ValNo) {
2080 // Because we are going top-down through the block, all value numbers
2081 // will be available in the predecessor by the time we need them. Any
2082 // that weren't originally present will have been instantiated earlier
2084 bool success = true;
2085 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2086 Value *Op = Instr->getOperand(i);
2087 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2089 // This could be a newly inserted instruction, in which case, we won't
2090 // find a value number, and should give up before we hurt ourselves.
2091 // FIXME: Rewrite the infrastructure to let it easier to value number
2092 // and process newly inserted instructions.
2093 if (!VN.exists(Op)) {
2098 VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this);
2099 if (Value *V = findLeader(Pred, TValNo)) {
2100 Instr->setOperand(i, V);
2107 // Fail out if we encounter an operand that is not available in
2108 // the PRE predecessor. This is typically because of loads which
2109 // are not value numbered precisely.
2113 Instr->insertBefore(Pred->getTerminator());
2114 Instr->setName(Instr->getName() + ".pre");
2115 Instr->setDebugLoc(Instr->getDebugLoc());
2117 unsigned Num = VN.lookupOrAdd(Instr);
2120 // Update the availability map to include the new instruction.
2121 addToLeaderTable(Num, Instr, Pred);
2125 bool GVN::performScalarPRE(Instruction *CurInst) {
2126 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2127 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2128 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2129 isa<DbgInfoIntrinsic>(CurInst))
2132 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2133 // sinking the compare again, and it would force the code generator to
2134 // move the i1 from processor flags or predicate registers into a general
2135 // purpose register.
2136 if (isa<CmpInst>(CurInst))
2139 // We don't currently value number ANY inline asm calls.
2140 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2141 if (CallI->isInlineAsm())
2144 uint32_t ValNo = VN.lookup(CurInst);
2146 // Look for the predecessors for PRE opportunities. We're
2147 // only trying to solve the basic diamond case, where
2148 // a value is computed in the successor and one predecessor,
2149 // but not the other. We also explicitly disallow cases
2150 // where the successor is its own predecessor, because they're
2151 // more complicated to get right.
2152 unsigned NumWith = 0;
2153 unsigned NumWithout = 0;
2154 BasicBlock *PREPred = nullptr;
2155 BasicBlock *CurrentBlock = CurInst->getParent();
2157 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2158 for (BasicBlock *P : predecessors(CurrentBlock)) {
2159 // We're not interested in PRE where blocks with predecessors that are
2161 if (!DT->isReachableFromEntry(P)) {
2165 // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and
2166 // when CurInst has operand defined in CurrentBlock (so it may be defined
2167 // by phi in the loop header).
2168 if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] &&
2169 llvm::any_of(CurInst->operands(), [&](const Use &U) {
2170 if (auto *Inst = dyn_cast<Instruction>(U.get()))
2171 return Inst->getParent() == CurrentBlock;
2178 uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this);
2179 Value *predV = findLeader(P, TValNo);
2181 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2184 } else if (predV == CurInst) {
2185 /* CurInst dominates this predecessor. */
2189 predMap.push_back(std::make_pair(predV, P));
2194 // Don't do PRE when it might increase code size, i.e. when
2195 // we would need to insert instructions in more than one pred.
2196 if (NumWithout > 1 || NumWith == 0)
2199 // We may have a case where all predecessors have the instruction,
2200 // and we just need to insert a phi node. Otherwise, perform
2202 Instruction *PREInstr = nullptr;
2204 if (NumWithout != 0) {
2205 // Don't do PRE across indirect branch.
2206 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2209 // We can't do PRE safely on a critical edge, so instead we schedule
2210 // the edge to be split and perform the PRE the next time we iterate
2212 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2213 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2214 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2217 // We need to insert somewhere, so let's give it a shot
2218 PREInstr = CurInst->clone();
2219 if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) {
2220 // If we failed insertion, make sure we remove the instruction.
2221 DEBUG(verifyRemoved(PREInstr));
2222 PREInstr->deleteValue();
2227 // Either we should have filled in the PRE instruction, or we should
2228 // not have needed insertions.
2229 assert(PREInstr != nullptr || NumWithout == 0);
2233 // Create a PHI to make the value available in this block.
2235 PHINode::Create(CurInst->getType(), predMap.size(),
2236 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2237 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2238 if (Value *V = predMap[i].first)
2239 Phi->addIncoming(V, predMap[i].second);
2241 Phi->addIncoming(PREInstr, PREPred);
2245 // After creating a new PHI for ValNo, the phi translate result for ValNo will
2246 // be changed, so erase the related stale entries in phi translate cache.
2247 VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock);
2248 addToLeaderTable(ValNo, Phi, CurrentBlock);
2249 Phi->setDebugLoc(CurInst->getDebugLoc());
2250 CurInst->replaceAllUsesWith(Phi);
2251 if (MD && Phi->getType()->isPtrOrPtrVectorTy())
2252 MD->invalidateCachedPointerInfo(Phi);
2254 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2256 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2258 MD->removeInstruction(CurInst);
2259 DEBUG(verifyRemoved(CurInst));
2260 CurInst->eraseFromParent();
2266 /// Perform a purely local form of PRE that looks for diamond
2267 /// control flow patterns and attempts to perform simple PRE at the join point.
2268 bool GVN::performPRE(Function &F) {
2269 bool Changed = false;
2270 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2271 // Nothing to PRE in the entry block.
2272 if (CurrentBlock == &F.getEntryBlock())
2275 // Don't perform PRE on an EH pad.
2276 if (CurrentBlock->isEHPad())
2279 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2280 BE = CurrentBlock->end();
2282 Instruction *CurInst = &*BI++;
2283 Changed |= performScalarPRE(CurInst);
2287 if (splitCriticalEdges())
2293 /// Split the critical edge connecting the given two blocks, and return
2294 /// the block inserted to the critical edge.
2295 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2297 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2299 MD->invalidateCachedPredecessors();
2303 /// Split critical edges found during the previous
2304 /// iteration that may enable further optimization.
2305 bool GVN::splitCriticalEdges() {
2306 if (toSplit.empty())
2309 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2310 SplitCriticalEdge(Edge.first, Edge.second,
2311 CriticalEdgeSplittingOptions(DT));
2312 } while (!toSplit.empty());
2313 if (MD) MD->invalidateCachedPredecessors();
2317 /// Executes one iteration of GVN
2318 bool GVN::iterateOnFunction(Function &F) {
2319 cleanupGlobalSets();
2321 // Top-down walk of the dominator tree
2322 bool Changed = false;
2323 // Needed for value numbering with phi construction to work.
2324 // RPOT walks the graph in its constructor and will not be invalidated during
2326 ReversePostOrderTraversal<Function *> RPOT(&F);
2327 for (BasicBlock *BB : RPOT)
2328 Changed |= processBlock(BB);
2333 void GVN::cleanupGlobalSets() {
2335 LeaderTable.clear();
2336 BlockRPONumber.clear();
2337 TableAllocator.Reset();
2340 /// Verify that the specified instruction does not occur in our
2341 /// internal data structures.
2342 void GVN::verifyRemoved(const Instruction *Inst) const {
2343 VN.verifyRemoved(Inst);
2345 // Walk through the value number scope to make sure the instruction isn't
2346 // ferreted away in it.
2347 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2348 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2349 const LeaderTableEntry *Node = &I->second;
2350 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2352 while (Node->Next) {
2354 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2359 /// BB is declared dead, which implied other blocks become dead as well. This
2360 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2361 /// live successors, update their phi nodes by replacing the operands
2362 /// corresponding to dead blocks with UndefVal.
2363 void GVN::addDeadBlock(BasicBlock *BB) {
2364 SmallVector<BasicBlock *, 4> NewDead;
2365 SmallSetVector<BasicBlock *, 4> DF;
2367 NewDead.push_back(BB);
2368 while (!NewDead.empty()) {
2369 BasicBlock *D = NewDead.pop_back_val();
2370 if (DeadBlocks.count(D))
2373 // All blocks dominated by D are dead.
2374 SmallVector<BasicBlock *, 8> Dom;
2375 DT->getDescendants(D, Dom);
2376 DeadBlocks.insert(Dom.begin(), Dom.end());
2378 // Figure out the dominance-frontier(D).
2379 for (BasicBlock *B : Dom) {
2380 for (BasicBlock *S : successors(B)) {
2381 if (DeadBlocks.count(S))
2384 bool AllPredDead = true;
2385 for (BasicBlock *P : predecessors(S))
2386 if (!DeadBlocks.count(P)) {
2387 AllPredDead = false;
2392 // S could be proved dead later on. That is why we don't update phi
2393 // operands at this moment.
2396 // While S is not dominated by D, it is dead by now. This could take
2397 // place if S already have a dead predecessor before D is declared
2399 NewDead.push_back(S);
2405 // For the dead blocks' live successors, update their phi nodes by replacing
2406 // the operands corresponding to dead blocks with UndefVal.
2407 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2410 if (DeadBlocks.count(B))
2413 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2414 for (BasicBlock *P : Preds) {
2415 if (!DeadBlocks.count(P))
2418 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2419 if (BasicBlock *S = splitCriticalEdges(P, B))
2420 DeadBlocks.insert(P = S);
2423 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2424 PHINode &Phi = cast<PHINode>(*II);
2425 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2426 UndefValue::get(Phi.getType()));
2432 // If the given branch is recognized as a foldable branch (i.e. conditional
2433 // branch with constant condition), it will perform following analyses and
2435 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2436 // R be the target of the dead out-coming edge.
2437 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2438 // edge. The result of this step will be {X| X is dominated by R}
2439 // 2) Identify those blocks which haves at least one dead predecessor. The
2440 // result of this step will be dominance-frontier(R).
2441 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2442 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2444 // Return true iff *NEW* dead code are found.
2445 bool GVN::processFoldableCondBr(BranchInst *BI) {
2446 if (!BI || BI->isUnconditional())
2449 // If a branch has two identical successors, we cannot declare either dead.
2450 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2453 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2457 BasicBlock *DeadRoot =
2458 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2459 if (DeadBlocks.count(DeadRoot))
2462 if (!DeadRoot->getSinglePredecessor())
2463 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2465 addDeadBlock(DeadRoot);
2469 // performPRE() will trigger assert if it comes across an instruction without
2470 // associated val-num. As it normally has far more live instructions than dead
2471 // instructions, it makes more sense just to "fabricate" a val-number for the
2472 // dead code than checking if instruction involved is dead or not.
2473 void GVN::assignValNumForDeadCode() {
2474 for (BasicBlock *BB : DeadBlocks) {
2475 for (Instruction &Inst : *BB) {
2476 unsigned ValNum = VN.lookupOrAdd(&Inst);
2477 addToLeaderTable(ValNum, &Inst, BB);
2482 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2484 static char ID; // Pass identification, replacement for typeid
2486 explicit GVNLegacyPass(bool NoLoads = false)
2487 : FunctionPass(ID), NoLoads(NoLoads) {
2488 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2491 bool runOnFunction(Function &F) override {
2492 if (skipFunction(F))
2495 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
2497 return Impl.runImpl(
2498 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2499 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2500 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2501 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2503 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep(),
2504 LIWP ? &LIWP->getLoopInfo() : nullptr,
2505 &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE());
2508 void getAnalysisUsage(AnalysisUsage &AU) const override {
2509 AU.addRequired<AssumptionCacheTracker>();
2510 AU.addRequired<DominatorTreeWrapperPass>();
2511 AU.addRequired<TargetLibraryInfoWrapperPass>();
2513 AU.addRequired<MemoryDependenceWrapperPass>();
2514 AU.addRequired<AAResultsWrapperPass>();
2516 AU.addPreserved<DominatorTreeWrapperPass>();
2517 AU.addPreserved<GlobalsAAWrapperPass>();
2518 AU.addPreserved<TargetLibraryInfoWrapperPass>();
2519 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2527 char GVNLegacyPass::ID = 0;
2529 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2530 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2531 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2532 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2533 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2534 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2535 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2536 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2537 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2539 // The public interface to this file...
2540 FunctionPass *llvm::createGVNPass(bool NoLoads) {
2541 return new GVNLegacyPass(NoLoads);