1 //===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the Correlated Value Propagation pass.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
14 #include "llvm/ADT/DepthFirstIterator.h"
15 #include "llvm/ADT/Optional.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/DomTreeUpdater.h"
19 #include "llvm/Analysis/GlobalsModRef.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LazyValueInfo.h"
22 #include "llvm/IR/Attributes.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/CFG.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/Constant.h"
27 #include "llvm/IR/ConstantRange.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InstrTypes.h"
33 #include "llvm/IR/Instruction.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/PassManager.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/IR/Value.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Casting.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Scalar.h"
46 #include "llvm/Transforms/Utils/Local.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");
66 STATISTIC(NumSaturating,
67 "Number of saturating arithmetics converted to normal arithmetics");
69 static cl::opt<bool> DontAddNoWrapFlags("cvp-dont-add-nowrap-flags", cl::init(false));
73 class CorrelatedValuePropagation : public FunctionPass {
77 CorrelatedValuePropagation(): FunctionPass(ID) {
78 initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
81 bool runOnFunction(Function &F) override;
83 void getAnalysisUsage(AnalysisUsage &AU) const override {
84 AU.addRequired<DominatorTreeWrapperPass>();
85 AU.addRequired<LazyValueInfoWrapperPass>();
86 AU.addPreserved<GlobalsAAWrapperPass>();
87 AU.addPreserved<DominatorTreeWrapperPass>();
91 } // end anonymous namespace
93 char CorrelatedValuePropagation::ID = 0;
95 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
96 "Value Propagation", false, false)
97 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
98 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
99 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
100 "Value Propagation", false, false)
102 // Public interface to the Value Propagation pass
103 Pass *llvm::createCorrelatedValuePropagationPass() {
104 return new CorrelatedValuePropagation();
107 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
108 if (S->getType()->isVectorTy()) return false;
109 if (isa<Constant>(S->getOperand(0))) return false;
111 Constant *C = LVI->getConstant(S->getCondition(), S->getParent(), S);
112 if (!C) return false;
114 ConstantInt *CI = dyn_cast<ConstantInt>(C);
115 if (!CI) return false;
117 Value *ReplaceWith = S->getTrueValue();
118 Value *Other = S->getFalseValue();
119 if (!CI->isOne()) std::swap(ReplaceWith, Other);
120 if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
122 S->replaceAllUsesWith(ReplaceWith);
123 S->eraseFromParent();
130 /// Try to simplify a phi with constant incoming values that match the edge
131 /// values of a non-constant value on all other edges:
133 /// %isnull = icmp eq i8* %x, null
134 /// br i1 %isnull, label %bb2, label %bb1
138 /// %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ]
141 static bool simplifyCommonValuePhi(PHINode *P, LazyValueInfo *LVI,
143 // Collect incoming constants and initialize possible common value.
144 SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants;
145 Value *CommonValue = nullptr;
146 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
147 Value *Incoming = P->getIncomingValue(i);
148 if (auto *IncomingConstant = dyn_cast<Constant>(Incoming)) {
149 IncomingConstants.push_back(std::make_pair(IncomingConstant, i));
150 } else if (!CommonValue) {
151 // The potential common value is initialized to the first non-constant.
152 CommonValue = Incoming;
153 } else if (Incoming != CommonValue) {
154 // There can be only one non-constant common value.
159 if (!CommonValue || IncomingConstants.empty())
162 // The common value must be valid in all incoming blocks.
163 BasicBlock *ToBB = P->getParent();
164 if (auto *CommonInst = dyn_cast<Instruction>(CommonValue))
165 if (!DT->dominates(CommonInst, ToBB))
168 // We have a phi with exactly 1 variable incoming value and 1 or more constant
169 // incoming values. See if all constant incoming values can be mapped back to
170 // the same incoming variable value.
171 for (auto &IncomingConstant : IncomingConstants) {
172 Constant *C = IncomingConstant.first;
173 BasicBlock *IncomingBB = P->getIncomingBlock(IncomingConstant.second);
174 if (C != LVI->getConstantOnEdge(CommonValue, IncomingBB, ToBB, P))
178 // All constant incoming values map to the same variable along the incoming
179 // edges of the phi. The phi is unnecessary.
180 P->replaceAllUsesWith(CommonValue);
181 P->eraseFromParent();
186 static bool processPHI(PHINode *P, LazyValueInfo *LVI, DominatorTree *DT,
187 const SimplifyQuery &SQ) {
188 bool Changed = false;
190 BasicBlock *BB = P->getParent();
191 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
192 Value *Incoming = P->getIncomingValue(i);
193 if (isa<Constant>(Incoming)) continue;
195 Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
197 // Look if the incoming value is a select with a scalar condition for which
198 // LVI can tells us the value. In that case replace the incoming value with
199 // the appropriate value of the select. This often allows us to remove the
202 SelectInst *SI = dyn_cast<SelectInst>(Incoming);
205 Value *Condition = SI->getCondition();
206 if (!Condition->getType()->isVectorTy()) {
207 if (Constant *C = LVI->getConstantOnEdge(
208 Condition, P->getIncomingBlock(i), BB, P)) {
209 if (C->isOneValue()) {
210 V = SI->getTrueValue();
211 } else if (C->isZeroValue()) {
212 V = SI->getFalseValue();
214 // Once LVI learns to handle vector types, we could also add support
215 // for vector type constants that are not all zeroes or all ones.
219 // Look if the select has a constant but LVI tells us that the incoming
220 // value can never be that constant. In that case replace the incoming
221 // value with the other value of the select. This often allows us to
222 // remove the select later.
224 Constant *C = dyn_cast<Constant>(SI->getFalseValue());
227 if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
228 P->getIncomingBlock(i), BB, P) !=
229 LazyValueInfo::False)
231 V = SI->getTrueValue();
234 LLVM_DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
237 P->setIncomingValue(i, V);
241 if (Value *V = SimplifyInstruction(P, SQ)) {
242 P->replaceAllUsesWith(V);
243 P->eraseFromParent();
248 Changed = simplifyCommonValuePhi(P, LVI, DT);
256 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
257 Value *Pointer = nullptr;
258 if (LoadInst *L = dyn_cast<LoadInst>(I))
259 Pointer = L->getPointerOperand();
261 Pointer = cast<StoreInst>(I)->getPointerOperand();
263 if (isa<Constant>(Pointer)) return false;
265 Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
266 if (!C) return false;
269 I->replaceUsesOfWith(Pointer, C);
273 /// See if LazyValueInfo's ability to exploit edge conditions or range
274 /// information is sufficient to prove this comparison. Even for local
275 /// conditions, this can sometimes prove conditions instcombine can't by
276 /// exploiting range information.
277 static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
278 Value *Op0 = Cmp->getOperand(0);
279 auto *C = dyn_cast<Constant>(Cmp->getOperand(1));
283 // As a policy choice, we choose not to waste compile time on anything where
284 // the comparison is testing local values. While LVI can sometimes reason
285 // about such cases, it's not its primary purpose. We do make sure to do
286 // the block local query for uses from terminator instructions, but that's
287 // handled in the code for each terminator.
288 auto *I = dyn_cast<Instruction>(Op0);
289 if (I && I->getParent() == Cmp->getParent())
292 LazyValueInfo::Tristate Result =
293 LVI->getPredicateAt(Cmp->getPredicate(), Op0, C, Cmp);
294 if (Result == LazyValueInfo::Unknown)
298 Constant *TorF = ConstantInt::get(Type::getInt1Ty(Cmp->getContext()), Result);
299 Cmp->replaceAllUsesWith(TorF);
300 Cmp->eraseFromParent();
304 /// Simplify a switch instruction by removing cases which can never fire. If the
305 /// uselessness of a case could be determined locally then constant propagation
306 /// would already have figured it out. Instead, walk the predecessors and
307 /// statically evaluate cases based on information available on that edge. Cases
308 /// that cannot fire no matter what the incoming edge can safely be removed. If
309 /// a case fires on every incoming edge then the entire switch can be removed
310 /// and replaced with a branch to the case destination.
311 static bool processSwitch(SwitchInst *I, LazyValueInfo *LVI,
313 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
314 Value *Cond = I->getCondition();
315 BasicBlock *BB = I->getParent();
317 // If the condition was defined in same block as the switch then LazyValueInfo
318 // currently won't say anything useful about it, though in theory it could.
319 if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
322 // If the switch is unreachable then trying to improve it is a waste of time.
323 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
324 if (PB == PE) return false;
326 // Analyse each switch case in turn.
327 bool Changed = false;
328 DenseMap<BasicBlock*, int> SuccessorsCount;
329 for (auto *Succ : successors(BB))
330 SuccessorsCount[Succ]++;
332 { // Scope for SwitchInstProfUpdateWrapper. It must not live during
333 // ConstantFoldTerminator() as the underlying SwitchInst can be changed.
334 SwitchInstProfUpdateWrapper SI(*I);
336 for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
337 ConstantInt *Case = CI->getCaseValue();
339 // Check to see if the switch condition is equal to/not equal to the case
340 // value on every incoming edge, equal/not equal being the same each time.
341 LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
342 for (pred_iterator PI = PB; PI != PE; ++PI) {
343 // Is the switch condition equal to the case value?
344 LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
347 // Give up on this case if nothing is known.
348 if (Value == LazyValueInfo::Unknown) {
349 State = LazyValueInfo::Unknown;
353 // If this was the first edge to be visited, record that all other edges
354 // need to give the same result.
360 // If this case is known to fire for some edges and known not to fire for
361 // others then there is nothing we can do - give up.
362 if (Value != State) {
363 State = LazyValueInfo::Unknown;
368 if (State == LazyValueInfo::False) {
369 // This case never fires - remove it.
370 BasicBlock *Succ = CI->getCaseSuccessor();
371 Succ->removePredecessor(BB);
372 CI = SI.removeCase(CI);
375 // The condition can be modified by removePredecessor's PHI simplification
377 Cond = SI->getCondition();
381 if (--SuccessorsCount[Succ] == 0)
382 DTU.applyUpdatesPermissive({{DominatorTree::Delete, BB, Succ}});
385 if (State == LazyValueInfo::True) {
386 // This case always fires. Arrange for the switch to be turned into an
387 // unconditional branch by replacing the switch condition with the case
389 SI->setCondition(Case);
390 NumDeadCases += SI->getNumCases();
395 // Increment the case iterator since we didn't delete it.
401 // If the switch has been simplified to the point where it can be replaced
402 // by a branch then do so now.
403 ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false,
404 /*TLI = */ nullptr, &DTU);
408 // See if we can prove that the given binary op intrinsic will not overflow.
409 static bool willNotOverflow(BinaryOpIntrinsic *BO, LazyValueInfo *LVI) {
410 ConstantRange LRange = LVI->getConstantRange(
411 BO->getLHS(), BO->getParent(), BO);
412 ConstantRange RRange = LVI->getConstantRange(
413 BO->getRHS(), BO->getParent(), BO);
414 ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
415 BO->getBinaryOp(), RRange, BO->getNoWrapKind());
416 return NWRegion.contains(LRange);
419 static void processOverflowIntrinsic(WithOverflowInst *WO) {
421 Value *NewOp = B.CreateBinOp(
422 WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), WO->getName());
423 // Constant-folding could have happened.
424 if (auto *Inst = dyn_cast<Instruction>(NewOp)) {
426 Inst->setHasNoSignedWrap();
428 Inst->setHasNoUnsignedWrap();
431 Value *NewI = B.CreateInsertValue(UndefValue::get(WO->getType()), NewOp, 0);
432 NewI = B.CreateInsertValue(NewI, ConstantInt::getFalse(WO->getContext()), 1);
433 WO->replaceAllUsesWith(NewI);
434 WO->eraseFromParent();
438 static void processSaturatingInst(SaturatingInst *SI) {
439 BinaryOperator *BinOp = BinaryOperator::Create(
440 SI->getBinaryOp(), SI->getLHS(), SI->getRHS(), SI->getName(), SI);
441 BinOp->setDebugLoc(SI->getDebugLoc());
443 BinOp->setHasNoSignedWrap();
445 BinOp->setHasNoUnsignedWrap();
447 SI->replaceAllUsesWith(BinOp);
448 SI->eraseFromParent();
452 /// Infer nonnull attributes for the arguments at the specified callsite.
453 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
454 SmallVector<unsigned, 4> ArgNos;
457 if (auto *WO = dyn_cast<WithOverflowInst>(CS.getInstruction())) {
458 if (WO->getLHS()->getType()->isIntegerTy() && willNotOverflow(WO, LVI)) {
459 processOverflowIntrinsic(WO);
464 if (auto *SI = dyn_cast<SaturatingInst>(CS.getInstruction())) {
465 if (SI->getType()->isIntegerTy() && willNotOverflow(SI, LVI)) {
466 processSaturatingInst(SI);
471 // Deopt bundle operands are intended to capture state with minimal
472 // perturbance of the code otherwise. If we can find a constant value for
473 // any such operand and remove a use of the original value, that's
474 // desireable since it may allow further optimization of that value (e.g. via
475 // single use rules in instcombine). Since deopt uses tend to,
476 // idiomatically, appear along rare conditional paths, it's reasonable likely
477 // we may have a conditional fact with which LVI can fold.
478 if (auto DeoptBundle = CS.getOperandBundle(LLVMContext::OB_deopt)) {
479 bool Progress = false;
480 for (const Use &ConstU : DeoptBundle->Inputs) {
481 Use &U = const_cast<Use&>(ConstU);
483 if (V->getType()->isVectorTy()) continue;
484 if (isa<Constant>(V)) continue;
486 Constant *C = LVI->getConstant(V, CS.getParent(), CS.getInstruction());
495 for (Value *V : CS.args()) {
496 PointerType *Type = dyn_cast<PointerType>(V->getType());
497 // Try to mark pointer typed parameters as non-null. We skip the
498 // relatively expensive analysis for constants which are obviously either
499 // null or non-null to start with.
500 if (Type && !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
502 LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
503 ConstantPointerNull::get(Type),
504 CS.getInstruction()) == LazyValueInfo::False)
505 ArgNos.push_back(ArgNo);
509 assert(ArgNo == CS.arg_size() && "sanity check");
514 AttributeList AS = CS.getAttributes();
515 LLVMContext &Ctx = CS.getInstruction()->getContext();
516 AS = AS.addParamAttribute(Ctx, ArgNos,
517 Attribute::get(Ctx, Attribute::NonNull));
518 CS.setAttributes(AS);
523 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
524 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
525 for (Value *O : SDI->operands()) {
526 auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
527 if (Result != LazyValueInfo::True)
533 /// Try to shrink a udiv/urem's width down to the smallest power of two that's
534 /// sufficient to contain its operands.
535 static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
536 assert(Instr->getOpcode() == Instruction::UDiv ||
537 Instr->getOpcode() == Instruction::URem);
538 if (Instr->getType()->isVectorTy())
541 // Find the smallest power of two bitwidth that's sufficient to hold Instr's
543 auto OrigWidth = Instr->getType()->getIntegerBitWidth();
544 ConstantRange OperandRange(OrigWidth, /*isFullSet=*/false);
545 for (Value *Operand : Instr->operands()) {
546 OperandRange = OperandRange.unionWith(
547 LVI->getConstantRange(Operand, Instr->getParent()));
549 // Don't shrink below 8 bits wide.
550 unsigned NewWidth = std::max<unsigned>(
551 PowerOf2Ceil(OperandRange.getUnsignedMax().getActiveBits()), 8);
552 // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
554 if (NewWidth >= OrigWidth)
558 IRBuilder<> B{Instr};
559 auto *TruncTy = Type::getIntNTy(Instr->getContext(), NewWidth);
560 auto *LHS = B.CreateTruncOrBitCast(Instr->getOperand(0), TruncTy,
561 Instr->getName() + ".lhs.trunc");
562 auto *RHS = B.CreateTruncOrBitCast(Instr->getOperand(1), TruncTy,
563 Instr->getName() + ".rhs.trunc");
564 auto *BO = B.CreateBinOp(Instr->getOpcode(), LHS, RHS, Instr->getName());
565 auto *Zext = B.CreateZExt(BO, Instr->getType(), Instr->getName() + ".zext");
566 if (auto *BinOp = dyn_cast<BinaryOperator>(BO))
567 if (BinOp->getOpcode() == Instruction::UDiv)
568 BinOp->setIsExact(Instr->isExact());
570 Instr->replaceAllUsesWith(Zext);
571 Instr->eraseFromParent();
575 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
576 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
580 auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
581 SDI->getName(), SDI);
582 BO->setDebugLoc(SDI->getDebugLoc());
583 SDI->replaceAllUsesWith(BO);
584 SDI->eraseFromParent();
586 // Try to process our new urem.
587 processUDivOrURem(BO, LVI);
592 /// See if LazyValueInfo's ability to exploit edge conditions or range
593 /// information is sufficient to prove the both operands of this SDiv are
594 /// positive. If this is the case, replace the SDiv with a UDiv. Even for local
595 /// conditions, this can sometimes prove conditions instcombine can't by
596 /// exploiting range information.
597 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
598 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
602 auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
603 SDI->getName(), SDI);
604 BO->setDebugLoc(SDI->getDebugLoc());
605 BO->setIsExact(SDI->isExact());
606 SDI->replaceAllUsesWith(BO);
607 SDI->eraseFromParent();
609 // Try to simplify our new udiv.
610 processUDivOrURem(BO, LVI);
615 static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
616 if (SDI->getType()->isVectorTy())
619 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
620 if (LVI->getPredicateAt(ICmpInst::ICMP_SGE, SDI->getOperand(0), Zero, SDI) !=
625 auto *BO = BinaryOperator::CreateLShr(SDI->getOperand(0), SDI->getOperand(1),
626 SDI->getName(), SDI);
627 BO->setDebugLoc(SDI->getDebugLoc());
628 BO->setIsExact(SDI->isExact());
629 SDI->replaceAllUsesWith(BO);
630 SDI->eraseFromParent();
635 static bool processBinOp(BinaryOperator *BinOp, LazyValueInfo *LVI) {
636 using OBO = OverflowingBinaryOperator;
638 if (DontAddNoWrapFlags)
641 if (BinOp->getType()->isVectorTy())
644 bool NSW = BinOp->hasNoSignedWrap();
645 bool NUW = BinOp->hasNoUnsignedWrap();
649 BasicBlock *BB = BinOp->getParent();
651 Value *LHS = BinOp->getOperand(0);
652 Value *RHS = BinOp->getOperand(1);
654 ConstantRange LRange = LVI->getConstantRange(LHS, BB, BinOp);
655 ConstantRange RRange = LVI->getConstantRange(RHS, BB, BinOp);
657 bool Changed = false;
659 ConstantRange NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
660 BinOp->getOpcode(), RRange, OBO::NoUnsignedWrap);
661 bool NewNUW = NUWRange.contains(LRange);
662 BinOp->setHasNoUnsignedWrap(NewNUW);
666 ConstantRange NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
667 BinOp->getOpcode(), RRange, OBO::NoSignedWrap);
668 bool NewNSW = NSWRange.contains(LRange);
669 BinOp->setHasNoSignedWrap(NewNSW);
676 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
677 if (Constant *C = LVI->getConstant(V, At->getParent(), At))
680 // TODO: The following really should be sunk inside LVI's core algorithm, or
681 // at least the outer shims around such.
682 auto *C = dyn_cast<CmpInst>(V);
683 if (!C) return nullptr;
685 Value *Op0 = C->getOperand(0);
686 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
687 if (!Op1) return nullptr;
689 LazyValueInfo::Tristate Result =
690 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
691 if (Result == LazyValueInfo::Unknown)
694 return (Result == LazyValueInfo::True) ?
695 ConstantInt::getTrue(C->getContext()) :
696 ConstantInt::getFalse(C->getContext());
699 static bool runImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT,
700 const SimplifyQuery &SQ) {
701 bool FnChanged = false;
702 // Visiting in a pre-order depth-first traversal causes us to simplify early
703 // blocks before querying later blocks (which require us to analyze early
704 // blocks). Eagerly simplifying shallow blocks means there is strictly less
705 // work to do for deep blocks. This also means we don't visit unreachable
707 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
708 bool BBChanged = false;
709 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
710 Instruction *II = &*BI++;
711 switch (II->getOpcode()) {
712 case Instruction::Select:
713 BBChanged |= processSelect(cast<SelectInst>(II), LVI);
715 case Instruction::PHI:
716 BBChanged |= processPHI(cast<PHINode>(II), LVI, DT, SQ);
718 case Instruction::ICmp:
719 case Instruction::FCmp:
720 BBChanged |= processCmp(cast<CmpInst>(II), LVI);
722 case Instruction::Load:
723 case Instruction::Store:
724 BBChanged |= processMemAccess(II, LVI);
726 case Instruction::Call:
727 case Instruction::Invoke:
728 BBChanged |= processCallSite(CallSite(II), LVI);
730 case Instruction::SRem:
731 BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
733 case Instruction::SDiv:
734 BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
736 case Instruction::UDiv:
737 case Instruction::URem:
738 BBChanged |= processUDivOrURem(cast<BinaryOperator>(II), LVI);
740 case Instruction::AShr:
741 BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
743 case Instruction::Add:
744 case Instruction::Sub:
745 BBChanged |= processBinOp(cast<BinaryOperator>(II), LVI);
750 Instruction *Term = BB->getTerminator();
751 switch (Term->getOpcode()) {
752 case Instruction::Switch:
753 BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI, DT);
755 case Instruction::Ret: {
756 auto *RI = cast<ReturnInst>(Term);
757 // Try to determine the return value if we can. This is mainly here to
758 // simplify the writing of unit tests, but also helps to enable IPO by
759 // constant folding the return values of callees.
760 auto *RetVal = RI->getReturnValue();
761 if (!RetVal) break; // handle "ret void"
762 if (isa<Constant>(RetVal)) break; // nothing to do
763 if (auto *C = getConstantAt(RetVal, RI, LVI)) {
765 RI->replaceUsesOfWith(RetVal, C);
771 FnChanged |= BBChanged;
777 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
781 LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
782 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
784 return runImpl(F, LVI, DT, getBestSimplifyQuery(*this, F));
788 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
789 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
790 DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
792 bool Changed = runImpl(F, LVI, DT, getBestSimplifyQuery(AM, F));
795 return PreservedAnalyses::all();
796 PreservedAnalyses PA;
797 PA.preserve<GlobalsAA>();
798 PA.preserve<DominatorTreeAnalysis>();