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/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/DomTreeUpdater.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"
47 #include "llvm/Transforms/Utils/Local.h"
53 #define DEBUG_TYPE "correlated-value-propagation"
55 STATISTIC(NumPhis, "Number of phis propagated");
56 STATISTIC(NumPhiCommon, "Number of phis deleted via common incoming value");
57 STATISTIC(NumSelects, "Number of selects propagated");
58 STATISTIC(NumMemAccess, "Number of memory access targets propagated");
59 STATISTIC(NumCmps, "Number of comparisons propagated");
60 STATISTIC(NumReturns, "Number of return values propagated");
61 STATISTIC(NumDeadCases, "Number of switch cases removed");
62 STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
63 STATISTIC(NumUDivs, "Number of udivs whose width was decreased");
64 STATISTIC(NumAShrs, "Number of ashr converted to lshr");
65 STATISTIC(NumSRems, "Number of srem converted to urem");
66 STATISTIC(NumOverflows, "Number of overflow checks removed");
68 static cl::opt<bool> DontProcessAdds("cvp-dont-process-adds", cl::init(true));
72 class CorrelatedValuePropagation : public FunctionPass {
76 CorrelatedValuePropagation(): FunctionPass(ID) {
77 initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
80 bool runOnFunction(Function &F) override;
82 void getAnalysisUsage(AnalysisUsage &AU) const override {
83 AU.addRequired<DominatorTreeWrapperPass>();
84 AU.addRequired<LazyValueInfoWrapperPass>();
85 AU.addPreserved<GlobalsAAWrapperPass>();
86 AU.addPreserved<DominatorTreeWrapperPass>();
90 } // end anonymous namespace
92 char CorrelatedValuePropagation::ID = 0;
94 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
95 "Value Propagation", false, false)
96 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
97 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
98 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
99 "Value Propagation", false, false)
101 // Public interface to the Value Propagation pass
102 Pass *llvm::createCorrelatedValuePropagationPass() {
103 return new CorrelatedValuePropagation();
106 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
107 if (S->getType()->isVectorTy()) return false;
108 if (isa<Constant>(S->getOperand(0))) return false;
110 Constant *C = LVI->getConstant(S->getCondition(), S->getParent(), S);
111 if (!C) return false;
113 ConstantInt *CI = dyn_cast<ConstantInt>(C);
114 if (!CI) return false;
116 Value *ReplaceWith = S->getTrueValue();
117 Value *Other = S->getFalseValue();
118 if (!CI->isOne()) std::swap(ReplaceWith, Other);
119 if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
121 S->replaceAllUsesWith(ReplaceWith);
122 S->eraseFromParent();
129 /// Try to simplify a phi with constant incoming values that match the edge
130 /// values of a non-constant value on all other edges:
132 /// %isnull = icmp eq i8* %x, null
133 /// br i1 %isnull, label %bb2, label %bb1
137 /// %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ]
140 static bool simplifyCommonValuePhi(PHINode *P, LazyValueInfo *LVI,
142 // Collect incoming constants and initialize possible common value.
143 SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants;
144 Value *CommonValue = nullptr;
145 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
146 Value *Incoming = P->getIncomingValue(i);
147 if (auto *IncomingConstant = dyn_cast<Constant>(Incoming)) {
148 IncomingConstants.push_back(std::make_pair(IncomingConstant, i));
149 } else if (!CommonValue) {
150 // The potential common value is initialized to the first non-constant.
151 CommonValue = Incoming;
152 } else if (Incoming != CommonValue) {
153 // There can be only one non-constant common value.
158 if (!CommonValue || IncomingConstants.empty())
161 // The common value must be valid in all incoming blocks.
162 BasicBlock *ToBB = P->getParent();
163 if (auto *CommonInst = dyn_cast<Instruction>(CommonValue))
164 if (!DT->dominates(CommonInst, ToBB))
167 // We have a phi with exactly 1 variable incoming value and 1 or more constant
168 // incoming values. See if all constant incoming values can be mapped back to
169 // the same incoming variable value.
170 for (auto &IncomingConstant : IncomingConstants) {
171 Constant *C = IncomingConstant.first;
172 BasicBlock *IncomingBB = P->getIncomingBlock(IncomingConstant.second);
173 if (C != LVI->getConstantOnEdge(CommonValue, IncomingBB, ToBB, P))
177 // All constant incoming values map to the same variable along the incoming
178 // edges of the phi. The phi is unnecessary.
179 P->replaceAllUsesWith(CommonValue);
180 P->eraseFromParent();
185 static bool processPHI(PHINode *P, LazyValueInfo *LVI, DominatorTree *DT,
186 const SimplifyQuery &SQ) {
187 bool Changed = false;
189 BasicBlock *BB = P->getParent();
190 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
191 Value *Incoming = P->getIncomingValue(i);
192 if (isa<Constant>(Incoming)) continue;
194 Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
196 // Look if the incoming value is a select with a scalar condition for which
197 // LVI can tells us the value. In that case replace the incoming value with
198 // the appropriate value of the select. This often allows us to remove the
201 SelectInst *SI = dyn_cast<SelectInst>(Incoming);
204 Value *Condition = SI->getCondition();
205 if (!Condition->getType()->isVectorTy()) {
206 if (Constant *C = LVI->getConstantOnEdge(
207 Condition, P->getIncomingBlock(i), BB, P)) {
208 if (C->isOneValue()) {
209 V = SI->getTrueValue();
210 } else if (C->isZeroValue()) {
211 V = SI->getFalseValue();
213 // Once LVI learns to handle vector types, we could also add support
214 // for vector type constants that are not all zeroes or all ones.
218 // Look if the select has a constant but LVI tells us that the incoming
219 // value can never be that constant. In that case replace the incoming
220 // value with the other value of the select. This often allows us to
221 // remove the select later.
223 Constant *C = dyn_cast<Constant>(SI->getFalseValue());
226 if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
227 P->getIncomingBlock(i), BB, P) !=
228 LazyValueInfo::False)
230 V = SI->getTrueValue();
233 LLVM_DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
236 P->setIncomingValue(i, V);
240 if (Value *V = SimplifyInstruction(P, SQ)) {
241 P->replaceAllUsesWith(V);
242 P->eraseFromParent();
247 Changed = simplifyCommonValuePhi(P, LVI, DT);
255 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
256 Value *Pointer = nullptr;
257 if (LoadInst *L = dyn_cast<LoadInst>(I))
258 Pointer = L->getPointerOperand();
260 Pointer = cast<StoreInst>(I)->getPointerOperand();
262 if (isa<Constant>(Pointer)) return false;
264 Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
265 if (!C) return false;
268 I->replaceUsesOfWith(Pointer, C);
272 /// See if LazyValueInfo's ability to exploit edge conditions or range
273 /// information is sufficient to prove this comparison. Even for local
274 /// conditions, this can sometimes prove conditions instcombine can't by
275 /// exploiting range information.
276 static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
277 Value *Op0 = Cmp->getOperand(0);
278 auto *C = dyn_cast<Constant>(Cmp->getOperand(1));
282 // As a policy choice, we choose not to waste compile time on anything where
283 // the comparison is testing local values. While LVI can sometimes reason
284 // about such cases, it's not its primary purpose. We do make sure to do
285 // the block local query for uses from terminator instructions, but that's
286 // handled in the code for each terminator.
287 auto *I = dyn_cast<Instruction>(Op0);
288 if (I && I->getParent() == Cmp->getParent())
291 LazyValueInfo::Tristate Result =
292 LVI->getPredicateAt(Cmp->getPredicate(), Op0, C, Cmp);
293 if (Result == LazyValueInfo::Unknown)
297 Constant *TorF = ConstantInt::get(Type::getInt1Ty(Cmp->getContext()), Result);
298 Cmp->replaceAllUsesWith(TorF);
299 Cmp->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,
312 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
313 Value *Cond = SI->getCondition();
314 BasicBlock *BB = SI->getParent();
316 // If the condition was defined in same block as the switch then LazyValueInfo
317 // currently won't say anything useful about it, though in theory it could.
318 if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
321 // If the switch is unreachable then trying to improve it is a waste of time.
322 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
323 if (PB == PE) return false;
325 // Analyse each switch case in turn.
326 bool Changed = false;
327 DenseMap<BasicBlock*, int> SuccessorsCount;
328 for (auto *Succ : successors(BB))
329 SuccessorsCount[Succ]++;
331 for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
332 ConstantInt *Case = CI->getCaseValue();
334 // Check to see if the switch condition is equal to/not equal to the case
335 // value on every incoming edge, equal/not equal being the same each time.
336 LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
337 for (pred_iterator PI = PB; PI != PE; ++PI) {
338 // Is the switch condition equal to the case value?
339 LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
342 // Give up on this case if nothing is known.
343 if (Value == LazyValueInfo::Unknown) {
344 State = LazyValueInfo::Unknown;
348 // If this was the first edge to be visited, record that all other edges
349 // need to give the same result.
355 // If this case is known to fire for some edges and known not to fire for
356 // others then there is nothing we can do - give up.
357 if (Value != State) {
358 State = LazyValueInfo::Unknown;
363 if (State == LazyValueInfo::False) {
364 // This case never fires - remove it.
365 BasicBlock *Succ = CI->getCaseSuccessor();
366 Succ->removePredecessor(BB);
367 CI = SI->removeCase(CI);
370 // The condition can be modified by removePredecessor's PHI simplification
372 Cond = SI->getCondition();
376 if (--SuccessorsCount[Succ] == 0)
377 DTU.deleteEdge(BB, Succ);
380 if (State == LazyValueInfo::True) {
381 // This case always fires. Arrange for the switch to be turned into an
382 // unconditional branch by replacing the switch condition with the case
384 SI->setCondition(Case);
385 NumDeadCases += SI->getNumCases();
390 // Increment the case iterator since we didn't delete it.
395 // If the switch has been simplified to the point where it can be replaced
396 // by a branch then do so now.
397 ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false,
398 /*TLI = */ nullptr, &DTU);
402 // See if we can prove that the given overflow intrinsic will not overflow.
403 static bool willNotOverflow(IntrinsicInst *II, LazyValueInfo *LVI) {
404 using OBO = OverflowingBinaryOperator;
405 auto NoWrap = [&] (Instruction::BinaryOps BinOp, unsigned NoWrapKind) {
406 Value *RHS = II->getOperand(1);
407 ConstantRange RRange = LVI->getConstantRange(RHS, II->getParent(), II);
408 ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
409 BinOp, RRange, NoWrapKind);
410 // As an optimization, do not compute LRange if we do not need it.
411 if (NWRegion.isEmptySet())
413 Value *LHS = II->getOperand(0);
414 ConstantRange LRange = LVI->getConstantRange(LHS, II->getParent(), II);
415 return NWRegion.contains(LRange);
417 switch (II->getIntrinsicID()) {
420 case Intrinsic::uadd_with_overflow:
421 return NoWrap(Instruction::Add, OBO::NoUnsignedWrap);
422 case Intrinsic::sadd_with_overflow:
423 return NoWrap(Instruction::Add, OBO::NoSignedWrap);
424 case Intrinsic::usub_with_overflow:
425 return NoWrap(Instruction::Sub, OBO::NoUnsignedWrap);
426 case Intrinsic::ssub_with_overflow:
427 return NoWrap(Instruction::Sub, OBO::NoSignedWrap);
432 static void processOverflowIntrinsic(IntrinsicInst *II) {
434 Value *NewOp = nullptr;
435 switch (II->getIntrinsicID()) {
437 llvm_unreachable("Unexpected instruction.");
438 case Intrinsic::uadd_with_overflow:
439 case Intrinsic::sadd_with_overflow:
440 NewOp = B.CreateAdd(II->getOperand(0), II->getOperand(1), II->getName());
442 case Intrinsic::usub_with_overflow:
443 case Intrinsic::ssub_with_overflow:
444 NewOp = B.CreateSub(II->getOperand(0), II->getOperand(1), II->getName());
448 Value *NewI = B.CreateInsertValue(UndefValue::get(II->getType()), NewOp, 0);
449 NewI = B.CreateInsertValue(NewI, ConstantInt::getFalse(II->getContext()), 1);
450 II->replaceAllUsesWith(NewI);
451 II->eraseFromParent();
454 /// Infer nonnull attributes for the arguments at the specified callsite.
455 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
456 SmallVector<unsigned, 4> ArgNos;
459 if (auto *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
460 if (willNotOverflow(II, LVI)) {
461 processOverflowIntrinsic(II);
466 for (Value *V : CS.args()) {
467 PointerType *Type = dyn_cast<PointerType>(V->getType());
468 // Try to mark pointer typed parameters as non-null. We skip the
469 // relatively expensive analysis for constants which are obviously either
470 // null or non-null to start with.
471 if (Type && !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
473 LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
474 ConstantPointerNull::get(Type),
475 CS.getInstruction()) == LazyValueInfo::False)
476 ArgNos.push_back(ArgNo);
480 assert(ArgNo == CS.arg_size() && "sanity check");
485 AttributeList AS = CS.getAttributes();
486 LLVMContext &Ctx = CS.getInstruction()->getContext();
487 AS = AS.addParamAttribute(Ctx, ArgNos,
488 Attribute::get(Ctx, Attribute::NonNull));
489 CS.setAttributes(AS);
494 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
495 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
496 for (Value *O : SDI->operands()) {
497 auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
498 if (Result != LazyValueInfo::True)
504 /// Try to shrink a udiv/urem's width down to the smallest power of two that's
505 /// sufficient to contain its operands.
506 static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
507 assert(Instr->getOpcode() == Instruction::UDiv ||
508 Instr->getOpcode() == Instruction::URem);
509 if (Instr->getType()->isVectorTy())
512 // Find the smallest power of two bitwidth that's sufficient to hold Instr's
514 auto OrigWidth = Instr->getType()->getIntegerBitWidth();
515 ConstantRange OperandRange(OrigWidth, /*isFullset=*/false);
516 for (Value *Operand : Instr->operands()) {
517 OperandRange = OperandRange.unionWith(
518 LVI->getConstantRange(Operand, Instr->getParent()));
520 // Don't shrink below 8 bits wide.
521 unsigned NewWidth = std::max<unsigned>(
522 PowerOf2Ceil(OperandRange.getUnsignedMax().getActiveBits()), 8);
523 // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
525 if (NewWidth >= OrigWidth)
529 IRBuilder<> B{Instr};
530 auto *TruncTy = Type::getIntNTy(Instr->getContext(), NewWidth);
531 auto *LHS = B.CreateTruncOrBitCast(Instr->getOperand(0), TruncTy,
532 Instr->getName() + ".lhs.trunc");
533 auto *RHS = B.CreateTruncOrBitCast(Instr->getOperand(1), TruncTy,
534 Instr->getName() + ".rhs.trunc");
535 auto *BO = B.CreateBinOp(Instr->getOpcode(), LHS, RHS, Instr->getName());
536 auto *Zext = B.CreateZExt(BO, Instr->getType(), Instr->getName() + ".zext");
537 if (auto *BinOp = dyn_cast<BinaryOperator>(BO))
538 if (BinOp->getOpcode() == Instruction::UDiv)
539 BinOp->setIsExact(Instr->isExact());
541 Instr->replaceAllUsesWith(Zext);
542 Instr->eraseFromParent();
546 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
547 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
551 auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
552 SDI->getName(), SDI);
553 BO->setDebugLoc(SDI->getDebugLoc());
554 SDI->replaceAllUsesWith(BO);
555 SDI->eraseFromParent();
557 // Try to process our new urem.
558 processUDivOrURem(BO, LVI);
563 /// See if LazyValueInfo's ability to exploit edge conditions or range
564 /// information is sufficient to prove the both operands of this SDiv are
565 /// positive. If this is the case, replace the SDiv with a UDiv. Even for local
566 /// conditions, this can sometimes prove conditions instcombine can't by
567 /// exploiting range information.
568 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
569 if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
573 auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
574 SDI->getName(), SDI);
575 BO->setDebugLoc(SDI->getDebugLoc());
576 BO->setIsExact(SDI->isExact());
577 SDI->replaceAllUsesWith(BO);
578 SDI->eraseFromParent();
580 // Try to simplify our new udiv.
581 processUDivOrURem(BO, LVI);
586 static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
587 if (SDI->getType()->isVectorTy())
590 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
591 if (LVI->getPredicateAt(ICmpInst::ICMP_SGE, SDI->getOperand(0), Zero, SDI) !=
596 auto *BO = BinaryOperator::CreateLShr(SDI->getOperand(0), SDI->getOperand(1),
597 SDI->getName(), SDI);
598 BO->setDebugLoc(SDI->getDebugLoc());
599 BO->setIsExact(SDI->isExact());
600 SDI->replaceAllUsesWith(BO);
601 SDI->eraseFromParent();
606 static bool processAdd(BinaryOperator *AddOp, LazyValueInfo *LVI) {
607 using OBO = OverflowingBinaryOperator;
612 if (AddOp->getType()->isVectorTy())
615 bool NSW = AddOp->hasNoSignedWrap();
616 bool NUW = AddOp->hasNoUnsignedWrap();
620 BasicBlock *BB = AddOp->getParent();
622 Value *LHS = AddOp->getOperand(0);
623 Value *RHS = AddOp->getOperand(1);
625 ConstantRange LRange = LVI->getConstantRange(LHS, BB, AddOp);
627 // Initialize RRange only if we need it. If we know that guaranteed no wrap
628 // range for the given LHS range is empty don't spend time calculating the
629 // range for the RHS.
630 Optional<ConstantRange> RRange;
631 auto LazyRRange = [&] () {
633 RRange = LVI->getConstantRange(RHS, BB, AddOp);
634 return RRange.getValue();
637 bool Changed = false;
639 ConstantRange NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
640 BinaryOperator::Add, LRange, OBO::NoUnsignedWrap);
641 if (!NUWRange.isEmptySet()) {
642 bool NewNUW = NUWRange.contains(LazyRRange());
643 AddOp->setHasNoUnsignedWrap(NewNUW);
648 ConstantRange NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
649 BinaryOperator::Add, LRange, OBO::NoSignedWrap);
650 if (!NSWRange.isEmptySet()) {
651 bool NewNSW = NSWRange.contains(LazyRRange());
652 AddOp->setHasNoSignedWrap(NewNSW);
660 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
661 if (Constant *C = LVI->getConstant(V, At->getParent(), At))
664 // TODO: The following really should be sunk inside LVI's core algorithm, or
665 // at least the outer shims around such.
666 auto *C = dyn_cast<CmpInst>(V);
667 if (!C) return nullptr;
669 Value *Op0 = C->getOperand(0);
670 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
671 if (!Op1) return nullptr;
673 LazyValueInfo::Tristate Result =
674 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
675 if (Result == LazyValueInfo::Unknown)
678 return (Result == LazyValueInfo::True) ?
679 ConstantInt::getTrue(C->getContext()) :
680 ConstantInt::getFalse(C->getContext());
683 static bool runImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT,
684 const SimplifyQuery &SQ) {
685 bool FnChanged = false;
686 // Visiting in a pre-order depth-first traversal causes us to simplify early
687 // blocks before querying later blocks (which require us to analyze early
688 // blocks). Eagerly simplifying shallow blocks means there is strictly less
689 // work to do for deep blocks. This also means we don't visit unreachable
691 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
692 bool BBChanged = false;
693 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
694 Instruction *II = &*BI++;
695 switch (II->getOpcode()) {
696 case Instruction::Select:
697 BBChanged |= processSelect(cast<SelectInst>(II), LVI);
699 case Instruction::PHI:
700 BBChanged |= processPHI(cast<PHINode>(II), LVI, DT, SQ);
702 case Instruction::ICmp:
703 case Instruction::FCmp:
704 BBChanged |= processCmp(cast<CmpInst>(II), LVI);
706 case Instruction::Load:
707 case Instruction::Store:
708 BBChanged |= processMemAccess(II, LVI);
710 case Instruction::Call:
711 case Instruction::Invoke:
712 BBChanged |= processCallSite(CallSite(II), LVI);
714 case Instruction::SRem:
715 BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
717 case Instruction::SDiv:
718 BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
720 case Instruction::UDiv:
721 case Instruction::URem:
722 BBChanged |= processUDivOrURem(cast<BinaryOperator>(II), LVI);
724 case Instruction::AShr:
725 BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
727 case Instruction::Add:
728 BBChanged |= processAdd(cast<BinaryOperator>(II), LVI);
733 Instruction *Term = BB->getTerminator();
734 switch (Term->getOpcode()) {
735 case Instruction::Switch:
736 BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI, DT);
738 case Instruction::Ret: {
739 auto *RI = cast<ReturnInst>(Term);
740 // Try to determine the return value if we can. This is mainly here to
741 // simplify the writing of unit tests, but also helps to enable IPO by
742 // constant folding the return values of callees.
743 auto *RetVal = RI->getReturnValue();
744 if (!RetVal) break; // handle "ret void"
745 if (isa<Constant>(RetVal)) break; // nothing to do
746 if (auto *C = getConstantAt(RetVal, RI, LVI)) {
748 RI->replaceUsesOfWith(RetVal, C);
754 FnChanged |= BBChanged;
760 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
764 LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
765 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
767 return runImpl(F, LVI, DT, getBestSimplifyQuery(*this, F));
771 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
772 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
773 DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
775 bool Changed = runImpl(F, LVI, DT, getBestSimplifyQuery(AM, F));
778 return PreservedAnalyses::all();
779 PreservedAnalyses PA;
780 PA.preserve<GlobalsAA>();
781 PA.preserve<DominatorTreeAnalysis>();