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/Transforms/Scalar.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/GlobalsModRef.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/LazyValueInfo.h"
20 #include "llvm/IR/CFG.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Transforms/Utils/Local.h"
32 #define DEBUG_TYPE "correlated-value-propagation"
34 STATISTIC(NumPhis, "Number of phis propagated");
35 STATISTIC(NumSelects, "Number of selects propagated");
36 STATISTIC(NumMemAccess, "Number of memory access targets propagated");
37 STATISTIC(NumCmps, "Number of comparisons propagated");
38 STATISTIC(NumReturns, "Number of return values propagated");
39 STATISTIC(NumDeadCases, "Number of switch cases removed");
40 STATISTIC(NumSDivs, "Number of sdiv converted to udiv");
41 STATISTIC(NumAShrs, "Number of ashr converted to lshr");
42 STATISTIC(NumSRems, "Number of srem converted to urem");
44 static cl::opt<bool> DontProcessAdds("cvp-dont-process-adds", cl::init(true));
47 class CorrelatedValuePropagation : public FunctionPass {
50 CorrelatedValuePropagation(): FunctionPass(ID) {
51 initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
54 bool runOnFunction(Function &F) override;
56 void getAnalysisUsage(AnalysisUsage &AU) const override {
57 AU.addRequired<LazyValueInfoWrapperPass>();
58 AU.addPreserved<GlobalsAAWrapperPass>();
63 char CorrelatedValuePropagation::ID = 0;
64 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
65 "Value Propagation", false, false)
66 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
67 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
68 "Value Propagation", false, false)
70 // Public interface to the Value Propagation pass
71 Pass *llvm::createCorrelatedValuePropagationPass() {
72 return new CorrelatedValuePropagation();
75 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
76 if (S->getType()->isVectorTy()) return false;
77 if (isa<Constant>(S->getOperand(0))) return false;
79 Constant *C = LVI->getConstant(S->getOperand(0), S->getParent(), S);
82 ConstantInt *CI = dyn_cast<ConstantInt>(C);
83 if (!CI) return false;
85 Value *ReplaceWith = S->getOperand(1);
86 Value *Other = S->getOperand(2);
87 if (!CI->isOne()) std::swap(ReplaceWith, Other);
88 if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
90 S->replaceAllUsesWith(ReplaceWith);
98 static bool processPHI(PHINode *P, LazyValueInfo *LVI) {
101 BasicBlock *BB = P->getParent();
102 for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
103 Value *Incoming = P->getIncomingValue(i);
104 if (isa<Constant>(Incoming)) continue;
106 Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
108 // Look if the incoming value is a select with a scalar condition for which
109 // LVI can tells us the value. In that case replace the incoming value with
110 // the appropriate value of the select. This often allows us to remove the
113 SelectInst *SI = dyn_cast<SelectInst>(Incoming);
116 Value *Condition = SI->getCondition();
117 if (!Condition->getType()->isVectorTy()) {
118 if (Constant *C = LVI->getConstantOnEdge(
119 Condition, P->getIncomingBlock(i), BB, P)) {
120 if (C->isOneValue()) {
121 V = SI->getTrueValue();
122 } else if (C->isZeroValue()) {
123 V = SI->getFalseValue();
125 // Once LVI learns to handle vector types, we could also add support
126 // for vector type constants that are not all zeroes or all ones.
130 // Look if the select has a constant but LVI tells us that the incoming
131 // value can never be that constant. In that case replace the incoming
132 // value with the other value of the select. This often allows us to
133 // remove the select later.
135 Constant *C = dyn_cast<Constant>(SI->getFalseValue());
138 if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
139 P->getIncomingBlock(i), BB, P) !=
140 LazyValueInfo::False)
142 V = SI->getTrueValue();
145 DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
148 P->setIncomingValue(i, V);
152 // FIXME: Provide TLI, DT, AT to SimplifyInstruction.
153 const DataLayout &DL = BB->getModule()->getDataLayout();
154 if (Value *V = SimplifyInstruction(P, DL)) {
155 P->replaceAllUsesWith(V);
156 P->eraseFromParent();
166 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
167 Value *Pointer = nullptr;
168 if (LoadInst *L = dyn_cast<LoadInst>(I))
169 Pointer = L->getPointerOperand();
171 Pointer = cast<StoreInst>(I)->getPointerOperand();
173 if (isa<Constant>(Pointer)) return false;
175 Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
176 if (!C) return false;
179 I->replaceUsesOfWith(Pointer, C);
183 /// See if LazyValueInfo's ability to exploit edge conditions or range
184 /// information is sufficient to prove this comparison. Even for local
185 /// conditions, this can sometimes prove conditions instcombine can't by
186 /// exploiting range information.
187 static bool processCmp(CmpInst *C, LazyValueInfo *LVI) {
188 Value *Op0 = C->getOperand(0);
189 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
190 if (!Op1) return false;
192 // As a policy choice, we choose not to waste compile time on anything where
193 // the comparison is testing local values. While LVI can sometimes reason
194 // about such cases, it's not its primary purpose. We do make sure to do
195 // the block local query for uses from terminator instructions, but that's
196 // handled in the code for each terminator.
197 auto *I = dyn_cast<Instruction>(Op0);
198 if (I && I->getParent() == C->getParent())
201 LazyValueInfo::Tristate Result =
202 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, C);
203 if (Result == LazyValueInfo::Unknown) return false;
206 if (Result == LazyValueInfo::True)
207 C->replaceAllUsesWith(ConstantInt::getTrue(C->getContext()));
209 C->replaceAllUsesWith(ConstantInt::getFalse(C->getContext()));
210 C->eraseFromParent();
215 /// Simplify a switch instruction by removing cases which can never fire. If the
216 /// uselessness of a case could be determined locally then constant propagation
217 /// would already have figured it out. Instead, walk the predecessors and
218 /// statically evaluate cases based on information available on that edge. Cases
219 /// that cannot fire no matter what the incoming edge can safely be removed. If
220 /// a case fires on every incoming edge then the entire switch can be removed
221 /// and replaced with a branch to the case destination.
222 static bool processSwitch(SwitchInst *SI, LazyValueInfo *LVI) {
223 Value *Cond = SI->getCondition();
224 BasicBlock *BB = SI->getParent();
226 // If the condition was defined in same block as the switch then LazyValueInfo
227 // currently won't say anything useful about it, though in theory it could.
228 if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
231 // If the switch is unreachable then trying to improve it is a waste of time.
232 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
233 if (PB == PE) return false;
235 // Analyse each switch case in turn. This is done in reverse order so that
236 // removing a case doesn't cause trouble for the iteration.
237 bool Changed = false;
238 for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
239 ConstantInt *Case = CI->getCaseValue();
241 // Check to see if the switch condition is equal to/not equal to the case
242 // value on every incoming edge, equal/not equal being the same each time.
243 LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
244 for (pred_iterator PI = PB; PI != PE; ++PI) {
245 // Is the switch condition equal to the case value?
246 LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
249 // Give up on this case if nothing is known.
250 if (Value == LazyValueInfo::Unknown) {
251 State = LazyValueInfo::Unknown;
255 // If this was the first edge to be visited, record that all other edges
256 // need to give the same result.
262 // If this case is known to fire for some edges and known not to fire for
263 // others then there is nothing we can do - give up.
264 if (Value != State) {
265 State = LazyValueInfo::Unknown;
270 if (State == LazyValueInfo::False) {
271 // This case never fires - remove it.
272 CI->getCaseSuccessor()->removePredecessor(BB);
273 CI = SI->removeCase(CI);
276 // The condition can be modified by removePredecessor's PHI simplification
278 Cond = SI->getCondition();
284 if (State == LazyValueInfo::True) {
285 // This case always fires. Arrange for the switch to be turned into an
286 // unconditional branch by replacing the switch condition with the case
288 SI->setCondition(Case);
289 NumDeadCases += SI->getNumCases();
294 // Increment the case iterator sense we didn't delete it.
299 // If the switch has been simplified to the point where it can be replaced
300 // by a branch then do so now.
301 ConstantFoldTerminator(BB);
306 /// Infer nonnull attributes for the arguments at the specified callsite.
307 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
308 SmallVector<unsigned, 4> Indices;
311 for (Value *V : CS.args()) {
312 PointerType *Type = dyn_cast<PointerType>(V->getType());
313 // Try to mark pointer typed parameters as non-null. We skip the
314 // relatively expensive analysis for constants which are obviously either
315 // null or non-null to start with.
316 if (Type && !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
318 LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
319 ConstantPointerNull::get(Type),
320 CS.getInstruction()) == LazyValueInfo::False)
321 Indices.push_back(ArgNo + 1);
325 assert(ArgNo == CS.arg_size() && "sanity check");
330 AttributeList AS = CS.getAttributes();
331 LLVMContext &Ctx = CS.getInstruction()->getContext();
332 AS = AS.addAttribute(Ctx, Indices, Attribute::get(Ctx, Attribute::NonNull));
333 CS.setAttributes(AS);
338 // Helper function to rewrite srem and sdiv. As a policy choice, we choose not
339 // to waste compile time on anything where the operands are local defs. While
340 // LVI can sometimes reason about such cases, it's not its primary purpose.
341 static bool hasLocalDefs(BinaryOperator *SDI) {
342 for (Value *O : SDI->operands()) {
343 auto *I = dyn_cast<Instruction>(O);
344 if (I && I->getParent() == SDI->getParent())
350 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
351 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
352 for (Value *O : SDI->operands()) {
353 auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
354 if (Result != LazyValueInfo::True)
360 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
361 if (SDI->getType()->isVectorTy() || hasLocalDefs(SDI) ||
362 !hasPositiveOperands(SDI, LVI))
366 auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
367 SDI->getName(), SDI);
368 SDI->replaceAllUsesWith(BO);
369 SDI->eraseFromParent();
373 /// See if LazyValueInfo's ability to exploit edge conditions or range
374 /// information is sufficient to prove the both operands of this SDiv are
375 /// positive. If this is the case, replace the SDiv with a UDiv. Even for local
376 /// conditions, this can sometimes prove conditions instcombine can't by
377 /// exploiting range information.
378 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
379 if (SDI->getType()->isVectorTy() || hasLocalDefs(SDI) ||
380 !hasPositiveOperands(SDI, LVI))
384 auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
385 SDI->getName(), SDI);
386 BO->setIsExact(SDI->isExact());
387 SDI->replaceAllUsesWith(BO);
388 SDI->eraseFromParent();
393 static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
394 if (SDI->getType()->isVectorTy() || hasLocalDefs(SDI))
397 Constant *Zero = ConstantInt::get(SDI->getType(), 0);
398 if (LVI->getPredicateAt(ICmpInst::ICMP_SGE, SDI->getOperand(0), Zero, SDI) !=
403 auto *BO = BinaryOperator::CreateLShr(SDI->getOperand(0), SDI->getOperand(1),
404 SDI->getName(), SDI);
405 BO->setIsExact(SDI->isExact());
406 SDI->replaceAllUsesWith(BO);
407 SDI->eraseFromParent();
412 static bool processAdd(BinaryOperator *AddOp, LazyValueInfo *LVI) {
413 typedef OverflowingBinaryOperator OBO;
418 if (AddOp->getType()->isVectorTy() || hasLocalDefs(AddOp))
421 bool NSW = AddOp->hasNoSignedWrap();
422 bool NUW = AddOp->hasNoUnsignedWrap();
426 BasicBlock *BB = AddOp->getParent();
428 Value *LHS = AddOp->getOperand(0);
429 Value *RHS = AddOp->getOperand(1);
431 ConstantRange LRange = LVI->getConstantRange(LHS, BB, AddOp);
433 // Initialize RRange only if we need it. If we know that guaranteed no wrap
434 // range for the given LHS range is empty don't spend time calculating the
435 // range for the RHS.
436 Optional<ConstantRange> RRange;
437 auto LazyRRange = [&] () {
439 RRange = LVI->getConstantRange(RHS, BB, AddOp);
440 return RRange.getValue();
443 bool Changed = false;
445 ConstantRange NUWRange =
446 LRange.makeGuaranteedNoWrapRegion(BinaryOperator::Add, LRange,
447 OBO::NoUnsignedWrap);
448 if (!NUWRange.isEmptySet()) {
449 bool NewNUW = NUWRange.contains(LazyRRange());
450 AddOp->setHasNoUnsignedWrap(NewNUW);
455 ConstantRange NSWRange =
456 LRange.makeGuaranteedNoWrapRegion(BinaryOperator::Add, LRange,
458 if (!NSWRange.isEmptySet()) {
459 bool NewNSW = NSWRange.contains(LazyRRange());
460 AddOp->setHasNoSignedWrap(NewNSW);
468 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
469 if (Constant *C = LVI->getConstant(V, At->getParent(), At))
472 // TODO: The following really should be sunk inside LVI's core algorithm, or
473 // at least the outer shims around such.
474 auto *C = dyn_cast<CmpInst>(V);
475 if (!C) return nullptr;
477 Value *Op0 = C->getOperand(0);
478 Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
479 if (!Op1) return nullptr;
481 LazyValueInfo::Tristate Result =
482 LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
483 if (Result == LazyValueInfo::Unknown)
486 return (Result == LazyValueInfo::True) ?
487 ConstantInt::getTrue(C->getContext()) :
488 ConstantInt::getFalse(C->getContext());
491 static bool runImpl(Function &F, LazyValueInfo *LVI) {
492 bool FnChanged = false;
494 // Visiting in a pre-order depth-first traversal causes us to simplify early
495 // blocks before querying later blocks (which require us to analyze early
496 // blocks). Eagerly simplifying shallow blocks means there is strictly less
497 // work to do for deep blocks. This also means we don't visit unreachable
499 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
500 bool BBChanged = false;
501 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
502 Instruction *II = &*BI++;
503 switch (II->getOpcode()) {
504 case Instruction::Select:
505 BBChanged |= processSelect(cast<SelectInst>(II), LVI);
507 case Instruction::PHI:
508 BBChanged |= processPHI(cast<PHINode>(II), LVI);
510 case Instruction::ICmp:
511 case Instruction::FCmp:
512 BBChanged |= processCmp(cast<CmpInst>(II), LVI);
514 case Instruction::Load:
515 case Instruction::Store:
516 BBChanged |= processMemAccess(II, LVI);
518 case Instruction::Call:
519 case Instruction::Invoke:
520 BBChanged |= processCallSite(CallSite(II), LVI);
522 case Instruction::SRem:
523 BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
525 case Instruction::SDiv:
526 BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
528 case Instruction::AShr:
529 BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
531 case Instruction::Add:
532 BBChanged |= processAdd(cast<BinaryOperator>(II), LVI);
537 Instruction *Term = BB->getTerminator();
538 switch (Term->getOpcode()) {
539 case Instruction::Switch:
540 BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI);
542 case Instruction::Ret: {
543 auto *RI = cast<ReturnInst>(Term);
544 // Try to determine the return value if we can. This is mainly here to
545 // simplify the writing of unit tests, but also helps to enable IPO by
546 // constant folding the return values of callees.
547 auto *RetVal = RI->getReturnValue();
548 if (!RetVal) break; // handle "ret void"
549 if (isa<Constant>(RetVal)) break; // nothing to do
550 if (auto *C = getConstantAt(RetVal, RI, LVI)) {
552 RI->replaceUsesOfWith(RetVal, C);
558 FnChanged |= BBChanged;
564 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
568 LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
569 return runImpl(F, LVI);
573 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
575 LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
576 bool Changed = runImpl(F, LVI);
579 return PreservedAnalyses::all();
580 PreservedAnalyses PA;
581 PA.preserve<GlobalsAA>();