1 //===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===//
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 file implements the Correlated Value Propagation pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
15 #include "llvm/ADT/DepthFirstIterator.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/GlobalsModRef.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LazyValueInfo.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/IR/Attributes.h"
24 #include "llvm/IR/BasicBlock.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/Constant.h"
28 #include "llvm/IR/ConstantRange.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/InstrTypes.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PassManager.h"
39 #include "llvm/IR/Type.h"
40 #include "llvm/IR/Value.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Support/Casting.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Scalar.h"
52 #define DEBUG_TYPE "correlated-value-propagation"
54 STATISTIC(NumPhis, "Number of phis propagated");
55 STATISTIC(NumPhiCommon, "Number of phis deleted via common incoming value");
56 STATISTIC(NumSelects, "Number of selects propagated");
57 STATISTIC(NumMemAccess, "Number of memory access targets propagated");
58 STATISTIC(NumCmps, "Number of comparisons propagated");
59 STATISTIC(NumReturns, "Number of return values propagated");
60 STATISTIC(NumDeadCases, "Number of switch cases removed");
61 STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
62 STATISTIC(NumUDivs, "Number of udivs whose width was decreased");
63 STATISTIC(NumAShrs, "Number of ashr converted to lshr");
64 STATISTIC(NumSRems, "Number of srem converted to urem");
65 STATISTIC(NumOverflows, "Number of overflow checks removed");
67 static cl::opt<bool> DontProcessAdds("cvp-dont-process-adds", cl::init(true));
71 class CorrelatedValuePropagation : public FunctionPass {
75 CorrelatedValuePropagation(): FunctionPass(ID) {
76 initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
79 bool runOnFunction(Function &F) override;
81 void getAnalysisUsage(AnalysisUsage &AU) const override {
82 AU.addRequired<DominatorTreeWrapperPass>();
83 AU.addRequired<LazyValueInfoWrapperPass>();
84 AU.addPreserved<GlobalsAAWrapperPass>();
85 AU.addPreserved<DominatorTreeWrapperPass>();
89 } // end anonymous namespace
91 char CorrelatedValuePropagation::ID = 0;
93 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
94 "Value Propagation", false, false)
95 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
96 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
97 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
98 "Value Propagation", false, false)
100 // Public interface to the Value Propagation pass
101 Pass *llvm::createCorrelatedValuePropagationPass() {
102 return new CorrelatedValuePropagation();
105 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
106 if (S->getType()->isVectorTy()) return false;
107 if (isa<Constant>(S->getOperand(0))) return false;
109 Constant *C = LVI->getConstant(S->getCondition(), S->getParent(), S);
110 if (!C) return false;
112 ConstantInt *CI = dyn_cast<ConstantInt>(C);
113 if (!CI) return false;
115 Value *ReplaceWith = S->getTrueValue();
116 Value *Other = S->getFalseValue();
117 if (!CI->isOne()) std::swap(ReplaceWith, Other);
118 if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
120 S->replaceAllUsesWith(ReplaceWith);
121 S->eraseFromParent();
128 /// Try to simplify a phi with constant incoming values that match the edge
129 /// values of a non-constant value on all other edges:
131 /// %isnull = icmp eq i8* %x, null
132 /// br i1 %isnull, label %bb2, label %bb1
136 /// %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ]
139 static bool simplifyCommonValuePhi(PHINode *P, LazyValueInfo *LVI,
141 // Collect incoming constants and initialize possible common value.
142 SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants;
143 Value *CommonValue = nullptr;
144 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
145 Value *Incoming = P->getIncomingValue(i);
146 if (auto *IncomingConstant = dyn_cast<Constant>(Incoming)) {
147 IncomingConstants.push_back(std::make_pair(IncomingConstant, i));
148 } else if (!CommonValue) {
149 // The potential common value is initialized to the first non-constant.
150 CommonValue = Incoming;
151 } else if (Incoming != CommonValue) {
152 // There can be only one non-constant common value.
157 if (!CommonValue || IncomingConstants.empty())
160 // The common value must be valid in all incoming blocks.
161 BasicBlock *ToBB = P->getParent();
162 if (auto *CommonInst = dyn_cast<Instruction>(CommonValue))
163 if (!DT->dominates(CommonInst, ToBB))
166 // We have a phi with exactly 1 variable incoming value and 1 or more constant
167 // incoming values. See if all constant incoming values can be mapped back to
168 // the same incoming variable value.
169 for (auto &IncomingConstant : IncomingConstants) {
170 Constant *C = IncomingConstant.first;
171 BasicBlock *IncomingBB = P->getIncomingBlock(IncomingConstant.second);
172 if (C != LVI->getConstantOnEdge(CommonValue, IncomingBB, ToBB, P))
176 // All constant incoming values map to the same variable along the incoming
177 // edges of the phi. The phi is unnecessary.
178 P->replaceAllUsesWith(CommonValue);
179 P->eraseFromParent();
184 static bool processPHI(PHINode *P, LazyValueInfo *LVI, DominatorTree *DT,
185 const SimplifyQuery &SQ) {
186 bool Changed = false;
188 BasicBlock *BB = P->getParent();
189 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
190 Value *Incoming = P->getIncomingValue(i);
191 if (isa<Constant>(Incoming)) continue;
193 Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
195 // Look if the incoming value is a select with a scalar condition for which
196 // LVI can tells us the value. In that case replace the incoming value with
197 // the appropriate value of the select. This often allows us to remove the
200 SelectInst *SI = dyn_cast<SelectInst>(Incoming);
203 Value *Condition = SI->getCondition();
204 if (!Condition->getType()->isVectorTy()) {
205 if (Constant *C = LVI->getConstantOnEdge(
206 Condition, P->getIncomingBlock(i), BB, P)) {
207 if (C->isOneValue()) {
208 V = SI->getTrueValue();
209 } else if (C->isZeroValue()) {
210 V = SI->getFalseValue();
212 // Once LVI learns to handle vector types, we could also add support
213 // for vector type constants that are not all zeroes or all ones.
217 // Look if the select has a constant but LVI tells us that the incoming
218 // value can never be that constant. In that case replace the incoming
219 // value with the other value of the select. This often allows us to
220 // remove the select later.
222 Constant *C = dyn_cast<Constant>(SI->getFalseValue());
225 if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
226 P->getIncomingBlock(i), BB, P) !=
227 LazyValueInfo::False)
229 V = SI->getTrueValue();
232 LLVM_DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
235 P->setIncomingValue(i, V);
239 if (Value *V = SimplifyInstruction(P, SQ)) {
240 P->replaceAllUsesWith(V);
241 P->eraseFromParent();
246 Changed = simplifyCommonValuePhi(P, LVI, DT);
254 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
255 Value *Pointer = nullptr;
256 if (LoadInst *L = dyn_cast<LoadInst>(I))
257 Pointer = L->getPointerOperand();
259 Pointer = cast<StoreInst>(I)->getPointerOperand();
261 if (isa<Constant>(Pointer)) return false;
263 Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
264 if (!C) return false;
267 I->replaceUsesOfWith(Pointer, C);
271 /// See if LazyValueInfo's ability to exploit edge conditions or range
272 /// information is sufficient to prove this comparison. Even for local
273 /// conditions, this can sometimes prove conditions instcombine can't by
274 /// exploiting range information.
275 static bool processCmp(CmpInst *C, LazyValueInfo *LVI) {
276 Value *Op0 = C->getOperand(0);
277 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
278 if (!Op1) return false;
280 // As a policy choice, we choose not to waste compile time on anything where
281 // the comparison is testing local values. While LVI can sometimes reason
282 // about such cases, it's not its primary purpose. We do make sure to do
283 // the block local query for uses from terminator instructions, but that's
284 // handled in the code for each terminator.
285 auto *I = dyn_cast<Instruction>(Op0);
286 if (I && I->getParent() == C->getParent())
289 LazyValueInfo::Tristate Result =
290 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, C);
291 if (Result == LazyValueInfo::Unknown) return false;
294 if (Result == LazyValueInfo::True)
295 C->replaceAllUsesWith(ConstantInt::getTrue(C->getContext()));
297 C->replaceAllUsesWith(ConstantInt::getFalse(C->getContext()));
298 C->eraseFromParent();
303 /// Simplify a switch instruction by removing cases which can never fire. If the
304 /// uselessness of a case could be determined locally then constant propagation
305 /// would already have figured it out. Instead, walk the predecessors and
306 /// statically evaluate cases based on information available on that edge. Cases
307 /// that cannot fire no matter what the incoming edge can safely be removed. If
308 /// a case fires on every incoming edge then the entire switch can be removed
309 /// and replaced with a branch to the case destination.
310 static bool processSwitch(SwitchInst *SI, LazyValueInfo *LVI, DominatorTree *DT) {
311 Value *Cond = SI->getCondition();
312 BasicBlock *BB = SI->getParent();
314 // If the condition was defined in same block as the switch then LazyValueInfo
315 // currently won't say anything useful about it, though in theory it could.
316 if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
319 // If the switch is unreachable then trying to improve it is a waste of time.
320 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
321 if (PB == PE) return false;
323 // Analyse each switch case in turn.
324 bool Changed = false;
325 DenseMap<BasicBlock*, int> SuccessorsCount;
326 for (auto *Succ : successors(BB))
327 SuccessorsCount[Succ]++;
329 for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
330 ConstantInt *Case = CI->getCaseValue();
332 // Check to see if the switch condition is equal to/not equal to the case
333 // value on every incoming edge, equal/not equal being the same each time.
334 LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
335 for (pred_iterator PI = PB; PI != PE; ++PI) {
336 // Is the switch condition equal to the case value?
337 LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
340 // Give up on this case if nothing is known.
341 if (Value == LazyValueInfo::Unknown) {
342 State = LazyValueInfo::Unknown;
346 // If this was the first edge to be visited, record that all other edges
347 // need to give the same result.
353 // If this case is known to fire for some edges and known not to fire for
354 // others then there is nothing we can do - give up.
355 if (Value != State) {
356 State = LazyValueInfo::Unknown;
361 if (State == LazyValueInfo::False) {
362 // This case never fires - remove it.
363 BasicBlock *Succ = CI->getCaseSuccessor();
364 Succ->removePredecessor(BB);
365 CI = SI->removeCase(CI);
368 // The condition can be modified by removePredecessor's PHI simplification
370 Cond = SI->getCondition();
374 if (--SuccessorsCount[Succ] == 0)
375 DT->deleteEdge(BB, Succ);
378 if (State == LazyValueInfo::True) {
379 // This case always fires. Arrange for the switch to be turned into an
380 // unconditional branch by replacing the switch condition with the case
382 SI->setCondition(Case);
383 NumDeadCases += SI->getNumCases();
388 // Increment the case iterator since we didn't delete it.
393 // If the switch has been simplified to the point where it can be replaced
394 // by a branch then do so now.
395 DeferredDominance DDT(*DT);
396 ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false,
397 /*TLI = */ nullptr, &DDT);
404 // See if we can prove that the given overflow intrinsic will not overflow.
405 static bool willNotOverflow(IntrinsicInst *II, LazyValueInfo *LVI) {
406 using OBO = OverflowingBinaryOperator;
407 auto NoWrap = [&] (Instruction::BinaryOps BinOp, unsigned NoWrapKind) {
408 Value *RHS = II->getOperand(1);
409 ConstantRange RRange = LVI->getConstantRange(RHS, II->getParent(), II);
410 ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
411 BinOp, RRange, NoWrapKind);
412 // As an optimization, do not compute LRange if we do not need it.
413 if (NWRegion.isEmptySet())
415 Value *LHS = II->getOperand(0);
416 ConstantRange LRange = LVI->getConstantRange(LHS, II->getParent(), II);
417 return NWRegion.contains(LRange);
419 switch (II->getIntrinsicID()) {
422 case Intrinsic::uadd_with_overflow:
423 return NoWrap(Instruction::Add, OBO::NoUnsignedWrap);
424 case Intrinsic::sadd_with_overflow:
425 return NoWrap(Instruction::Add, OBO::NoSignedWrap);
426 case Intrinsic::usub_with_overflow:
427 return NoWrap(Instruction::Sub, OBO::NoUnsignedWrap);
428 case Intrinsic::ssub_with_overflow:
429 return NoWrap(Instruction::Sub, OBO::NoSignedWrap);
434 static void processOverflowIntrinsic(IntrinsicInst *II) {
435 Value *NewOp = nullptr;
436 switch (II->getIntrinsicID()) {
438 llvm_unreachable("Unexpected instruction.");
439 case Intrinsic::uadd_with_overflow:
440 case Intrinsic::sadd_with_overflow:
441 NewOp = BinaryOperator::CreateAdd(II->getOperand(0), II->getOperand(1),
444 case Intrinsic::usub_with_overflow:
445 case Intrinsic::ssub_with_overflow:
446 NewOp = BinaryOperator::CreateSub(II->getOperand(0), II->getOperand(1),
452 Value *NewI = B.CreateInsertValue(UndefValue::get(II->getType()), NewOp, 0);
453 NewI = B.CreateInsertValue(NewI, ConstantInt::getFalse(II->getContext()), 1);
454 II->replaceAllUsesWith(NewI);
455 II->eraseFromParent();
458 /// Infer nonnull attributes for the arguments at the specified callsite.
459 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
460 SmallVector<unsigned, 4> ArgNos;
463 if (auto *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
464 if (willNotOverflow(II, LVI)) {
465 processOverflowIntrinsic(II);
470 for (Value *V : CS.args()) {
471 PointerType *Type = dyn_cast<PointerType>(V->getType());
472 // Try to mark pointer typed parameters as non-null. We skip the
473 // relatively expensive analysis for constants which are obviously either
474 // null or non-null to start with.
475 if (Type && !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
477 LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
478 ConstantPointerNull::get(Type),
479 CS.getInstruction()) == LazyValueInfo::False)
480 ArgNos.push_back(ArgNo);
484 assert(ArgNo == CS.arg_size() && "sanity check");
489 AttributeList AS = CS.getAttributes();
490 LLVMContext &Ctx = CS.getInstruction()->getContext();
491 AS = AS.addParamAttribute(Ctx, ArgNos,
492 Attribute::get(Ctx, Attribute::NonNull));
493 CS.setAttributes(AS);
498 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
499 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
500 for (Value *O : SDI->operands()) {
501 auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
502 if (Result != LazyValueInfo::True)
508 /// Try to shrink a udiv/urem's width down to the smallest power of two that's
509 /// sufficient to contain its operands.
510 static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
511 assert(Instr->getOpcode() == Instruction::UDiv ||
512 Instr->getOpcode() == Instruction::URem);
513 if (Instr->getType()->isVectorTy())
516 // Find the smallest power of two bitwidth that's sufficient to hold Instr's
518 auto OrigWidth = Instr->getType()->getIntegerBitWidth();
519 ConstantRange OperandRange(OrigWidth, /*isFullset=*/false);
520 for (Value *Operand : Instr->operands()) {
521 OperandRange = OperandRange.unionWith(
522 LVI->getConstantRange(Operand, Instr->getParent()));
524 // Don't shrink below 8 bits wide.
525 unsigned NewWidth = std::max<unsigned>(
526 PowerOf2Ceil(OperandRange.getUnsignedMax().getActiveBits()), 8);
527 // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
529 if (NewWidth >= OrigWidth)
533 auto *TruncTy = Type::getIntNTy(Instr->getContext(), NewWidth);
534 auto *LHS = CastInst::Create(Instruction::Trunc, Instr->getOperand(0), TruncTy,
535 Instr->getName() + ".lhs.trunc", Instr);
536 auto *RHS = CastInst::Create(Instruction::Trunc, Instr->getOperand(1), TruncTy,
537 Instr->getName() + ".rhs.trunc", Instr);
539 BinaryOperator::Create(Instr->getOpcode(), LHS, RHS, Instr->getName(), Instr);
540 auto *Zext = CastInst::Create(Instruction::ZExt, BO, Instr->getType(),
541 Instr->getName() + ".zext", Instr);
542 if (BO->getOpcode() == Instruction::UDiv)
543 BO->setIsExact(Instr->isExact());
545 Instr->replaceAllUsesWith(Zext);
546 Instr->eraseFromParent();
550 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
551 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
555 auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
556 SDI->getName(), SDI);
557 SDI->replaceAllUsesWith(BO);
558 SDI->eraseFromParent();
560 // Try to process our new urem.
561 processUDivOrURem(BO, LVI);
566 /// See if LazyValueInfo's ability to exploit edge conditions or range
567 /// information is sufficient to prove the both operands of this SDiv are
568 /// positive. If this is the case, replace the SDiv with a UDiv. Even for local
569 /// conditions, this can sometimes prove conditions instcombine can't by
570 /// exploiting range information.
571 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
572 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
576 auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
577 SDI->getName(), SDI);
578 BO->setIsExact(SDI->isExact());
579 SDI->replaceAllUsesWith(BO);
580 SDI->eraseFromParent();
582 // Try to simplify our new udiv.
583 processUDivOrURem(BO, LVI);
588 static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
589 if (SDI->getType()->isVectorTy())
592 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
593 if (LVI->getPredicateAt(ICmpInst::ICMP_SGE, SDI->getOperand(0), Zero, SDI) !=
598 auto *BO = BinaryOperator::CreateLShr(SDI->getOperand(0), SDI->getOperand(1),
599 SDI->getName(), SDI);
600 BO->setIsExact(SDI->isExact());
601 SDI->replaceAllUsesWith(BO);
602 SDI->eraseFromParent();
607 static bool processAdd(BinaryOperator *AddOp, LazyValueInfo *LVI) {
608 using OBO = OverflowingBinaryOperator;
613 if (AddOp->getType()->isVectorTy())
616 bool NSW = AddOp->hasNoSignedWrap();
617 bool NUW = AddOp->hasNoUnsignedWrap();
621 BasicBlock *BB = AddOp->getParent();
623 Value *LHS = AddOp->getOperand(0);
624 Value *RHS = AddOp->getOperand(1);
626 ConstantRange LRange = LVI->getConstantRange(LHS, BB, AddOp);
628 // Initialize RRange only if we need it. If we know that guaranteed no wrap
629 // range for the given LHS range is empty don't spend time calculating the
630 // range for the RHS.
631 Optional<ConstantRange> RRange;
632 auto LazyRRange = [&] () {
634 RRange = LVI->getConstantRange(RHS, BB, AddOp);
635 return RRange.getValue();
638 bool Changed = false;
640 ConstantRange NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
641 BinaryOperator::Add, LRange, OBO::NoUnsignedWrap);
642 if (!NUWRange.isEmptySet()) {
643 bool NewNUW = NUWRange.contains(LazyRRange());
644 AddOp->setHasNoUnsignedWrap(NewNUW);
649 ConstantRange NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
650 BinaryOperator::Add, LRange, OBO::NoSignedWrap);
651 if (!NSWRange.isEmptySet()) {
652 bool NewNSW = NSWRange.contains(LazyRRange());
653 AddOp->setHasNoSignedWrap(NewNSW);
661 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
662 if (Constant *C = LVI->getConstant(V, At->getParent(), At))
665 // TODO: The following really should be sunk inside LVI's core algorithm, or
666 // at least the outer shims around such.
667 auto *C = dyn_cast<CmpInst>(V);
668 if (!C) return nullptr;
670 Value *Op0 = C->getOperand(0);
671 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
672 if (!Op1) return nullptr;
674 LazyValueInfo::Tristate Result =
675 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
676 if (Result == LazyValueInfo::Unknown)
679 return (Result == LazyValueInfo::True) ?
680 ConstantInt::getTrue(C->getContext()) :
681 ConstantInt::getFalse(C->getContext());
684 static bool runImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT,
685 const SimplifyQuery &SQ) {
686 bool FnChanged = false;
687 // Visiting in a pre-order depth-first traversal causes us to simplify early
688 // blocks before querying later blocks (which require us to analyze early
689 // blocks). Eagerly simplifying shallow blocks means there is strictly less
690 // work to do for deep blocks. This also means we don't visit unreachable
692 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
693 bool BBChanged = false;
694 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
695 Instruction *II = &*BI++;
696 switch (II->getOpcode()) {
697 case Instruction::Select:
698 BBChanged |= processSelect(cast<SelectInst>(II), LVI);
700 case Instruction::PHI:
701 BBChanged |= processPHI(cast<PHINode>(II), LVI, DT, SQ);
703 case Instruction::ICmp:
704 case Instruction::FCmp:
705 BBChanged |= processCmp(cast<CmpInst>(II), LVI);
707 case Instruction::Load:
708 case Instruction::Store:
709 BBChanged |= processMemAccess(II, LVI);
711 case Instruction::Call:
712 case Instruction::Invoke:
713 BBChanged |= processCallSite(CallSite(II), LVI);
715 case Instruction::SRem:
716 BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
718 case Instruction::SDiv:
719 BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
721 case Instruction::UDiv:
722 case Instruction::URem:
723 BBChanged |= processUDivOrURem(cast<BinaryOperator>(II), LVI);
725 case Instruction::AShr:
726 BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
728 case Instruction::Add:
729 BBChanged |= processAdd(cast<BinaryOperator>(II), LVI);
734 Instruction *Term = BB->getTerminator();
735 switch (Term->getOpcode()) {
736 case Instruction::Switch:
737 BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI, DT);
739 case Instruction::Ret: {
740 auto *RI = cast<ReturnInst>(Term);
741 // Try to determine the return value if we can. This is mainly here to
742 // simplify the writing of unit tests, but also helps to enable IPO by
743 // constant folding the return values of callees.
744 auto *RetVal = RI->getReturnValue();
745 if (!RetVal) break; // handle "ret void"
746 if (isa<Constant>(RetVal)) break; // nothing to do
747 if (auto *C = getConstantAt(RetVal, RI, LVI)) {
749 RI->replaceUsesOfWith(RetVal, C);
755 FnChanged |= BBChanged;
761 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
765 LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
766 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
768 return runImpl(F, LVI, DT, getBestSimplifyQuery(*this, F));
772 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
773 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
774 DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
776 bool Changed = runImpl(F, LVI, DT, getBestSimplifyQuery(AM, F));
779 return PreservedAnalyses::all();
780 PreservedAnalyses PA;
781 PA.preserve<GlobalsAA>();
782 PA.preserve<DominatorTreeAnalysis>();