1//===- SCCPSolver.cpp - SCCP Utility --------------------------- *- C++ -*-===//
2//
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
6//
7//===----------------------------------------------------------------------===//
8//
9// \file
10// This file implements the Sparse Conditional Constant Propagation (SCCP)
11// utility.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Transforms/Utils/SCCPSolver.h"
16#include "llvm/ADT/SetVector.h"
17#include "llvm/Analysis/ConstantFolding.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/ValueLattice.h"
20#include "llvm/Analysis/ValueLatticeUtils.h"
21#include "llvm/Analysis/ValueTracking.h"
22#include "llvm/IR/ConstantRange.h"
23#include "llvm/IR/DerivedTypes.h"
24#include "llvm/IR/IRBuilder.h"
25#include "llvm/IR/InstVisitor.h"
26#include "llvm/IR/Instructions.h"
27#include "llvm/IR/NoFolder.h"
28#include "llvm/IR/PatternMatch.h"
29#include "llvm/Support/Casting.h"
30#include "llvm/Support/Debug.h"
31#include "llvm/Support/ErrorHandling.h"
32#include "llvm/Support/raw_ostream.h"
33#include "llvm/Transforms/Utils/Local.h"
34#include <cassert>
35#include <utility>
36#include <vector>
37
38using namespace llvm;
39using namespace PatternMatch;
40
41#define DEBUG_TYPE "sccp"
42
43// The maximum number of range extensions allowed for operations requiring
44// widening.
45static const unsigned MaxNumRangeExtensions = 10;
46
47/// Returns MergeOptions with MaxWidenSteps set to MaxNumRangeExtensions.
48static ValueLatticeElement::MergeOptions getMaxWidenStepsOpts() {
49 return ValueLatticeElement::MergeOptions().setMaxWidenSteps(
50 MaxNumRangeExtensions);
51}
52
53namespace llvm {
54
55bool SCCPSolver::isConstant(const ValueLatticeElement &LV) {
56 return LV.isConstant() ||
57 (LV.isConstantRange() && LV.getConstantRange().isSingleElement());
58}
59
60bool SCCPSolver::isOverdefined(const ValueLatticeElement &LV) {
61 return !LV.isUnknownOrUndef() && !SCCPSolver::isConstant(LV);
62}
63
64bool SCCPSolver::tryToReplaceWithConstant(Value *V) {
65 Constant *Const = getConstantOrNull(V);
66 if (!Const)
67 return false;
68 // Replacing `musttail` instructions with constant breaks `musttail` invariant
69 // unless the call itself can be removed.
70 // Calls with "clang.arc.attachedcall" implicitly use the return value and
71 // those uses cannot be updated with a constant.
72 CallBase *CB = dyn_cast<CallBase>(Val: V);
73 if (CB && ((CB->isMustTailCall() && !wouldInstructionBeTriviallyDead(I: CB)) ||
74 CB->getOperandBundle(ID: LLVMContext::OB_clang_arc_attachedcall))) {
75 Function *F = CB->getCalledFunction();
76
77 // Don't zap returns of the callee
78 if (F)
79 addToMustPreserveReturnsInFunctions(F);
80
81 LLVM_DEBUG(dbgs() << " Can\'t treat the result of call " << *CB
82 << " as a constant\n");
83 return false;
84 }
85
86 LLVM_DEBUG(dbgs() << " Constant: " << *Const << " = " << *V << '\n');
87
88 // Replaces all of the uses of a variable with uses of the constant.
89 V->replaceAllUsesWith(V: Const);
90 return true;
91}
92
93/// Helper for propagting !implicit.ref metadata from callee to caller before
94/// erasing a call instruction. This ensures that references to global objects
95/// (e.g., copyright strings) are preserved even when calls are optimized away.
96static void propagateImplicitRefFromCall(CallBase *CB) {
97 Function *Callee = CB->getCalledFunction();
98 if (!Callee)
99 return;
100
101 if (!Callee->hasMetadata(KindID: LLVMContext::MD_implicit_ref))
102 return;
103
104 Function *Caller = CB->getParent()->getParent();
105 if (!Caller)
106 return;
107
108 SmallVector<MDNode *> MDs;
109 Callee->getMetadata(KindID: LLVMContext::MD_implicit_ref, MDs);
110 for (MDNode *MD : MDs)
111 Caller->addMetadata(KindID: LLVMContext::MD_implicit_ref, MD&: *MD);
112}
113
114/// Helper for getting ranges from \p Solver. Instructions inserted during
115/// simplification are unavailable in the solver, so we return a full range for
116/// them.
117static ConstantRange getRange(Value *Op, SCCPSolver &Solver,
118 const SmallPtrSetImpl<Value *> &InsertedValues) {
119 if (auto *Const = dyn_cast<Constant>(Val: Op))
120 return Const->toConstantRange();
121 if (InsertedValues.contains(Ptr: Op)) {
122 unsigned Bitwidth = Op->getType()->getScalarSizeInBits();
123 return ConstantRange::getFull(BitWidth: Bitwidth);
124 }
125 return Solver.getLatticeValueFor(V: Op).asConstantRange(Ty: Op->getType(),
126 /*UndefAllowed=*/false);
127}
128
129/// Try to use \p Inst's value range from \p Solver to infer the NUW flag.
130static bool refineInstruction(SCCPSolver &Solver,
131 const SmallPtrSetImpl<Value *> &InsertedValues,
132 Instruction &Inst) {
133 bool Changed = false;
134 auto GetRange = [&Solver, &InsertedValues](Value *Op) {
135 return getRange(Op, Solver, InsertedValues);
136 };
137
138 if (isa<OverflowingBinaryOperator>(Val: Inst)) {
139 if (Inst.hasNoSignedWrap() && Inst.hasNoUnsignedWrap())
140 return false;
141
142 auto RangeA = GetRange(Inst.getOperand(i: 0));
143 auto RangeB = GetRange(Inst.getOperand(i: 1));
144 if (!Inst.hasNoUnsignedWrap()) {
145 auto NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
146 BinOp: Instruction::BinaryOps(Inst.getOpcode()), Other: RangeB,
147 NoWrapKind: OverflowingBinaryOperator::NoUnsignedWrap);
148 if (NUWRange.contains(CR: RangeA)) {
149 Inst.setHasNoUnsignedWrap();
150 Changed = true;
151 }
152 }
153 if (!Inst.hasNoSignedWrap()) {
154 auto NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
155 BinOp: Instruction::BinaryOps(Inst.getOpcode()), Other: RangeB,
156 NoWrapKind: OverflowingBinaryOperator::NoSignedWrap);
157 if (NSWRange.contains(CR: RangeA)) {
158 Inst.setHasNoSignedWrap();
159 Changed = true;
160 }
161 }
162 } else if (isa<PossiblyNonNegInst>(Val: Inst) && !Inst.hasNonNeg()) {
163 auto Range = GetRange(Inst.getOperand(i: 0));
164 if (Range.isAllNonNegative()) {
165 Inst.setNonNeg();
166 Changed = true;
167 }
168 } else if (TruncInst *TI = dyn_cast<TruncInst>(Val: &Inst)) {
169 if (TI->hasNoSignedWrap() && TI->hasNoUnsignedWrap())
170 return false;
171
172 auto Range = GetRange(Inst.getOperand(i: 0));
173 uint64_t DestWidth = TI->getDestTy()->getScalarSizeInBits();
174 if (!TI->hasNoUnsignedWrap()) {
175 if (Range.getActiveBits() <= DestWidth) {
176 TI->setHasNoUnsignedWrap(true);
177 Changed = true;
178 }
179 }
180 if (!TI->hasNoSignedWrap()) {
181 if (Range.getMinSignedBits() <= DestWidth) {
182 TI->setHasNoSignedWrap(true);
183 Changed = true;
184 }
185 }
186 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: &Inst)) {
187 if (GEP->hasNoUnsignedWrap() || !GEP->hasNoUnsignedSignedWrap())
188 return false;
189
190 if (all_of(Range: GEP->indices(),
191 P: [&](Value *V) { return GetRange(V).isAllNonNegative(); })) {
192 GEP->setNoWrapFlags(GEP->getNoWrapFlags() |
193 GEPNoWrapFlags::noUnsignedWrap());
194 Changed = true;
195 }
196 }
197
198 return Changed;
199}
200
201/// Try to replace signed instructions with their unsigned equivalent.
202static bool replaceSignedInst(SCCPSolver &Solver,
203 SmallPtrSetImpl<Value *> &InsertedValues,
204 Instruction &Inst) {
205 // Determine if a signed value is known to be >= 0.
206 auto isNonNegative = [&Solver, &InsertedValues](Value *V) {
207 return getRange(Op: V, Solver, InsertedValues).isAllNonNegative();
208 };
209
210 Instruction *NewInst = nullptr;
211 switch (Inst.getOpcode()) {
212 case Instruction::SIToFP:
213 case Instruction::SExt: {
214 // If the source value is not negative, this is a zext/uitofp.
215 Value *Op0 = Inst.getOperand(i: 0);
216 if (!isNonNegative(Op0))
217 return false;
218 NewInst = CastInst::Create(Inst.getOpcode() == Instruction::SExt
219 ? Instruction::ZExt
220 : Instruction::UIToFP,
221 S: Op0, Ty: Inst.getType(), Name: "", InsertBefore: Inst.getIterator());
222 NewInst->setNonNeg();
223 break;
224 }
225 case Instruction::AShr: {
226 // If the shifted value is not negative, this is a logical shift right.
227 Value *Op0 = Inst.getOperand(i: 0);
228 if (!isNonNegative(Op0))
229 return false;
230 NewInst = BinaryOperator::CreateLShr(V1: Op0, V2: Inst.getOperand(i: 1), Name: "", InsertBefore: Inst.getIterator());
231 NewInst->setIsExact(Inst.isExact());
232 break;
233 }
234 case Instruction::SDiv:
235 case Instruction::SRem: {
236 // If both operands are not negative, this is the same as udiv/urem.
237 Value *Op0 = Inst.getOperand(i: 0), *Op1 = Inst.getOperand(i: 1);
238 if (!isNonNegative(Op0) || !isNonNegative(Op1))
239 return false;
240 auto NewOpcode = Inst.getOpcode() == Instruction::SDiv ? Instruction::UDiv
241 : Instruction::URem;
242 NewInst = BinaryOperator::Create(Op: NewOpcode, S1: Op0, S2: Op1, Name: "", InsertBefore: Inst.getIterator());
243 if (Inst.getOpcode() == Instruction::SDiv)
244 NewInst->setIsExact(Inst.isExact());
245 break;
246 }
247 default:
248 return false;
249 }
250
251 // Wire up the new instruction and update state.
252 assert(NewInst && "Expected replacement instruction");
253 NewInst->takeName(V: &Inst);
254 InsertedValues.insert(Ptr: NewInst);
255 Inst.replaceAllUsesWith(V: NewInst);
256 NewInst->setDebugLoc(Inst.getDebugLoc());
257 Solver.removeLatticeValueFor(V: &Inst);
258 Inst.eraseFromParent();
259 return true;
260}
261
262/// Try to use \p Inst's value range from \p Solver to simplify it.
263static Value *simplifyInstruction(SCCPSolver &Solver,
264 SmallPtrSetImpl<Value *> &InsertedValues,
265 Instruction &Inst) {
266 auto GetRange = [&Solver, &InsertedValues](Value *Op) {
267 return getRange(Op, Solver, InsertedValues);
268 };
269
270 Value *X;
271 const APInt *RHSC;
272 // Remove masking operations.
273 if (match(V: &Inst, P: m_And(L: m_Value(V&: X), R: m_LowBitMask(V&: RHSC)))) {
274 ConstantRange LRange = GetRange(X);
275 if (LRange.getUnsignedMax().ule(RHS: *RHSC))
276 return X;
277 }
278
279 // Check if we can simplify [us]cmp(X, Y) to X - Y.
280 if (auto *Cmp = dyn_cast<CmpIntrinsic>(Val: &Inst)) {
281 Value *LHS = Cmp->getOperand(i_nocapture: 0);
282 Value *RHS = Cmp->getOperand(i_nocapture: 1);
283 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
284 // Bail out on 1-bit comparisons.
285 if (BitWidth == 1)
286 return nullptr;
287 ConstantRange LRange = GetRange(LHS);
288 if (LRange.isSizeLargerThan(MaxSize: 3))
289 return nullptr;
290 ConstantRange RRange = GetRange(RHS);
291 if (RRange.isSizeLargerThan(MaxSize: 3))
292 return nullptr;
293 ConstantRange RHSLower = RRange.sub(Other: APInt(BitWidth, 1));
294 ConstantRange RHSUpper = RRange.add(Other: APInt(BitWidth, 1));
295 ICmpInst::Predicate Pred =
296 Cmp->isSigned() ? CmpInst::ICMP_SLE : CmpInst::ICMP_ULE;
297 if (!RHSLower.icmp(Pred, Other: LRange) || !LRange.icmp(Pred, Other: RHSUpper))
298 return nullptr;
299
300 IRBuilder<NoFolder> Builder(&Inst);
301 Value *Sub = Builder.CreateSub(LHS, RHS, Name: Inst.getName(), /*HasNUW=*/false,
302 /*HasNSW=*/Cmp->isSigned());
303 InsertedValues.insert(Ptr: Sub);
304 if (Sub->getType() != Inst.getType()) {
305 Sub = Builder.CreateSExtOrTrunc(V: Sub, DestTy: Inst.getType());
306 InsertedValues.insert(Ptr: Sub);
307 }
308 return Sub;
309 }
310
311 // Relax range checks.
312 if (auto *ICmp = dyn_cast<ICmpInst>(Val: &Inst)) {
313 Value *X;
314 auto MatchTwoInstructionExactRangeCheck =
315 [&]() -> std::optional<ConstantRange> {
316 const APInt *RHSC;
317 if (!match(V: ICmp->getOperand(i_nocapture: 1), P: m_APInt(Res&: RHSC)))
318 return std::nullopt;
319
320 Value *LHS = ICmp->getOperand(i_nocapture: 0);
321 ICmpInst::Predicate Pred = ICmp->getPredicate();
322 const APInt *Offset;
323 if (match(V: LHS, P: m_OneUse(SubPattern: m_AddLike(L: m_Value(V&: X), R: m_APInt(Res&: Offset)))))
324 return ConstantRange::makeExactICmpRegion(Pred, Other: *RHSC).sub(Other: *Offset);
325 // Match icmp eq/ne X & NegPow2, C
326 if (ICmp->isEquality()) {
327 const APInt *Mask;
328 if (match(V: LHS, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_NegatedPower2(V&: Mask)))) &&
329 RHSC->countr_zero() >= Mask->countr_zero()) {
330 ConstantRange CR(*RHSC, *RHSC - *Mask);
331 return Pred == ICmpInst::ICMP_EQ ? CR : CR.inverse();
332 }
333 }
334 return std::nullopt;
335 };
336
337 if (auto CR = MatchTwoInstructionExactRangeCheck()) {
338 ConstantRange LRange = GetRange(X);
339 // Early exit if we know nothing about X.
340 if (LRange.isFullSet())
341 return nullptr;
342 auto ConvertCRToICmp =
343 [&](const std::optional<ConstantRange> &NewCR) -> Value * {
344 ICmpInst::Predicate Pred;
345 APInt RHS;
346 // Check if we can represent NewCR as an icmp predicate.
347 if (NewCR && NewCR->getEquivalentICmp(Pred, RHS)) {
348 IRBuilder<NoFolder> Builder(&Inst);
349 Value *NewICmp =
350 Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: RHS));
351 InsertedValues.insert(Ptr: NewICmp);
352 return NewICmp;
353 }
354 return nullptr;
355 };
356 // We are allowed to refine the comparison to either true or false for out
357 // of range inputs.
358 // Here we refine the comparison to false, and check if we can narrow the
359 // range check to a simpler test.
360 if (auto *V = ConvertCRToICmp(CR->exactIntersectWith(CR: LRange)))
361 return V;
362 // Here we refine the comparison to true, i.e. we relax the range check.
363 if (auto *V = ConvertCRToICmp(CR->exactUnionWith(CR: LRange.inverse())))
364 return V;
365 }
366 }
367
368 return nullptr;
369}
370
371bool SCCPSolver::simplifyInstsInBlock(BasicBlock &BB,
372 SmallPtrSetImpl<Value *> &InsertedValues,
373 Statistic &InstRemovedStat,
374 Statistic &InstReplacedStat) {
375 bool MadeChanges = false;
376 for (Instruction &Inst : make_early_inc_range(Range&: BB)) {
377 if (Inst.getType()->isVoidTy())
378 continue;
379 if (tryToReplaceWithConstant(V: &Inst)) {
380 if (wouldInstructionBeTriviallyDead(I: &Inst)) {
381 // Propagate !implicit.ref before erasing the call.
382 if (auto *CB = dyn_cast<CallBase>(Val: &Inst))
383 propagateImplicitRefFromCall(CB);
384
385 Inst.eraseFromParent();
386 }
387
388 MadeChanges = true;
389 ++InstRemovedStat;
390 } else if (replaceSignedInst(Solver&: *this, InsertedValues, Inst)) {
391 MadeChanges = true;
392 ++InstReplacedStat;
393 } else if (refineInstruction(Solver&: *this, InsertedValues, Inst)) {
394 MadeChanges = true;
395 } else if (auto *V = simplifyInstruction(Solver&: *this, InsertedValues, Inst)) {
396 Inst.replaceAllUsesWith(V);
397 Inst.eraseFromParent();
398 ++InstRemovedStat;
399 MadeChanges = true;
400 }
401 }
402 return MadeChanges;
403}
404
405bool SCCPSolver::removeNonFeasibleEdges(BasicBlock *BB, DomTreeUpdater &DTU,
406 BasicBlock *&NewUnreachableBB) const {
407 SmallPtrSet<BasicBlock *, 8> FeasibleSuccessors;
408 bool HasNonFeasibleEdges = false;
409 for (BasicBlock *Succ : successors(BB)) {
410 if (isEdgeFeasible(From: BB, To: Succ))
411 FeasibleSuccessors.insert(Ptr: Succ);
412 else
413 HasNonFeasibleEdges = true;
414 }
415
416 // All edges feasible, nothing to do.
417 if (!HasNonFeasibleEdges)
418 return false;
419
420 // SCCP can only determine non-feasible edges for br, switch and indirectbr.
421 Instruction *TI = BB->getTerminator();
422 assert((isa<UncondBrInst, CondBrInst, SwitchInst, IndirectBrInst>(TI)) &&
423 "Terminator must be a br, switch or indirectbr");
424
425 if (FeasibleSuccessors.size() == 0) {
426 // Branch on undef/poison, replace with unreachable.
427 SmallPtrSet<BasicBlock *, 8> SeenSuccs;
428 SmallVector<DominatorTree::UpdateType, 8> Updates;
429 for (BasicBlock *Succ : successors(BB)) {
430 Succ->removePredecessor(Pred: BB);
431 if (SeenSuccs.insert(Ptr: Succ).second)
432 Updates.push_back(Elt: {DominatorTree::Delete, BB, Succ});
433 }
434 TI->eraseFromParent();
435 new UnreachableInst(BB->getContext(), BB);
436 DTU.applyUpdatesPermissive(Updates);
437 } else if (FeasibleSuccessors.size() == 1) {
438 // Replace with an unconditional branch to the only feasible successor.
439 BasicBlock *OnlyFeasibleSuccessor = *FeasibleSuccessors.begin();
440 SmallVector<DominatorTree::UpdateType, 8> Updates;
441 bool HaveSeenOnlyFeasibleSuccessor = false;
442 for (BasicBlock *Succ : successors(BB)) {
443 if (Succ == OnlyFeasibleSuccessor && !HaveSeenOnlyFeasibleSuccessor) {
444 // Don't remove the edge to the only feasible successor the first time
445 // we see it. We still do need to remove any multi-edges to it though.
446 HaveSeenOnlyFeasibleSuccessor = true;
447 continue;
448 }
449
450 Succ->removePredecessor(Pred: BB);
451 Updates.push_back(Elt: {DominatorTree::Delete, BB, Succ});
452 }
453
454 Instruction *BI = UncondBrInst::Create(Target: OnlyFeasibleSuccessor, InsertBefore: BB);
455 BI->setDebugLoc(TI->getDebugLoc());
456 TI->eraseFromParent();
457 DTU.applyUpdatesPermissive(Updates);
458 } else if (FeasibleSuccessors.size() > 1) {
459 SwitchInstProfUpdateWrapper SI(*cast<SwitchInst>(Val: TI));
460 SmallVector<DominatorTree::UpdateType, 8> Updates;
461
462 // If the default destination is unfeasible it will never be taken. Replace
463 // it with a new block with a single Unreachable instruction.
464 BasicBlock *DefaultDest = SI->getDefaultDest();
465 if (!FeasibleSuccessors.contains(Ptr: DefaultDest)) {
466 if (!NewUnreachableBB) {
467 NewUnreachableBB =
468 BasicBlock::Create(Context&: DefaultDest->getContext(), Name: "default.unreachable",
469 Parent: DefaultDest->getParent(), InsertBefore: DefaultDest);
470 auto *UI =
471 new UnreachableInst(DefaultDest->getContext(), NewUnreachableBB);
472 UI->setDebugLoc(DebugLoc::getTemporary());
473 }
474
475 DefaultDest->removePredecessor(Pred: BB);
476 SI->setDefaultDest(NewUnreachableBB);
477 Updates.push_back(Elt: {DominatorTree::Delete, BB, DefaultDest});
478 Updates.push_back(Elt: {DominatorTree::Insert, BB, NewUnreachableBB});
479 }
480
481 for (auto CI = SI->case_begin(); CI != SI->case_end();) {
482 if (FeasibleSuccessors.contains(Ptr: CI->getCaseSuccessor())) {
483 ++CI;
484 continue;
485 }
486
487 BasicBlock *Succ = CI->getCaseSuccessor();
488 Succ->removePredecessor(Pred: BB);
489 Updates.push_back(Elt: {DominatorTree::Delete, BB, Succ});
490 SI.removeCase(I: CI);
491 // Don't increment CI, as we removed a case.
492 }
493
494 DTU.applyUpdatesPermissive(Updates);
495 } else {
496 llvm_unreachable("Must have at least one feasible successor");
497 }
498 return true;
499}
500
501static void inferAttribute(Function *F, unsigned AttrIndex,
502 const ValueLatticeElement &Val) {
503 // If there is a known constant range for the value, add range attribute.
504 if (Val.isConstantRange() && !Val.getConstantRange().isSingleElement()) {
505 // Do not add range attribute if the value may include undef.
506 if (Val.isConstantRangeIncludingUndef())
507 return;
508
509 // Take the intersection of the existing attribute and the inferred range.
510 Attribute OldAttr = F->getAttributeAtIndex(i: AttrIndex, Kind: Attribute::Range);
511 ConstantRange CR = Val.getConstantRange();
512 if (OldAttr.isValid())
513 CR = CR.intersectWith(CR: OldAttr.getRange());
514 F->addAttributeAtIndex(
515 i: AttrIndex, Attr: Attribute::get(Context&: F->getContext(), Kind: Attribute::Range, CR));
516 return;
517 }
518 // Infer nonnull attribute.
519 if (Val.isNotConstant() && Val.getNotConstant()->getType()->isPointerTy() &&
520 Val.getNotConstant()->isNullValue() &&
521 !F->hasAttributeAtIndex(Idx: AttrIndex, Kind: Attribute::NonNull)) {
522 F->addAttributeAtIndex(i: AttrIndex,
523 Attr: Attribute::get(Context&: F->getContext(), Kind: Attribute::NonNull));
524 }
525}
526
527void SCCPSolver::inferReturnAttributes() const {
528 for (const auto &[F, ReturnValue] : getTrackedRetVals())
529 inferAttribute(F, AttrIndex: AttributeList::ReturnIndex, Val: ReturnValue);
530}
531
532void SCCPSolver::inferArgAttributes() const {
533 for (Function *F : getArgumentTrackedFunctions()) {
534 if (!isBlockExecutable(BB: &F->front()))
535 continue;
536 for (Argument &A : F->args())
537 if (!A.getType()->isStructTy())
538 inferAttribute(F, AttrIndex: AttributeList::FirstArgIndex + A.getArgNo(),
539 Val: getLatticeValueFor(V: &A));
540 }
541}
542
543/// Helper class for SCCPSolver. This implements the instruction visitor and
544/// holds all the state.
545class SCCPInstVisitor : public InstVisitor<SCCPInstVisitor> {
546 const DataLayout &DL;
547 std::function<const TargetLibraryInfo &(Function &)> GetTLI;
548 /// Basic blocks that are executable (but may not have been visited yet).
549 SmallPtrSet<BasicBlock *, 8> BBExecutable;
550 /// Basic blocks that are executable and have been visited at least once.
551 SmallPtrSet<BasicBlock *, 8> BBVisited;
552 DenseMap<Value *, ValueLatticeElement>
553 ValueState; // The state each value is in.
554
555 /// StructValueState - This maintains ValueState for values that have
556 /// StructType, for example for formal arguments, calls, insertelement, etc.
557 DenseMap<std::pair<Value *, unsigned>, ValueLatticeElement> StructValueState;
558
559 /// GlobalValue - If we are tracking any values for the contents of a global
560 /// variable, we keep a mapping from the constant accessor to the element of
561 /// the global, to the currently known value. If the value becomes
562 /// overdefined, it's entry is simply removed from this map.
563 DenseMap<GlobalVariable *, ValueLatticeElement> TrackedGlobals;
564
565 /// TrackedRetVals - If we are tracking arguments into and the return
566 /// value out of a function, it will have an entry in this map, indicating
567 /// what the known return value for the function is.
568 MapVector<Function *, ValueLatticeElement> TrackedRetVals;
569
570 /// TrackedMultipleRetVals - Same as TrackedRetVals, but used for functions
571 /// that return multiple values.
572 MapVector<std::pair<Function *, unsigned>, ValueLatticeElement>
573 TrackedMultipleRetVals;
574
575 /// The set of values whose lattice has been invalidated.
576 /// Populated by resetLatticeValueFor(), cleared after resolving undefs.
577 DenseSet<Value *> Invalidated;
578
579 /// MRVFunctionsTracked - Each function in TrackedMultipleRetVals is
580 /// represented here for efficient lookup.
581 SmallPtrSet<Function *, 16> MRVFunctionsTracked;
582
583 /// A list of functions whose return cannot be modified.
584 SmallPtrSet<Function *, 16> MustPreserveReturnsInFunctions;
585
586 /// TrackingIncomingArguments - This is the set of functions for whose
587 /// arguments we make optimistic assumptions about and try to prove as
588 /// constants.
589 SmallPtrSet<Function *, 16> TrackingIncomingArguments;
590
591 /// Worklist of instructions to re-visit. This only includes instructions
592 /// in blocks that have already been visited at least once.
593 SmallSetVector<Instruction *, 16> InstWorkList;
594
595 /// Current instruction while visiting a block for the first time, used to
596 /// avoid unnecessary instruction worklist insertions. Null if an instruction
597 /// is visited outside a whole-block visitation.
598 Instruction *CurI = nullptr;
599
600 // The BasicBlock work list
601 SmallVector<BasicBlock *, 64> BBWorkList;
602
603 /// KnownFeasibleEdges - Entries in this set are edges which have already had
604 /// PHI nodes retriggered.
605 using Edge = std::pair<BasicBlock *, BasicBlock *>;
606 DenseSet<Edge> KnownFeasibleEdges;
607
608 DenseMap<Function *, std::unique_ptr<PredicateInfo>> FnPredicateInfo;
609
610 DenseMap<Value *, SmallSetVector<User *, 2>> AdditionalUsers;
611
612 LLVMContext &Ctx;
613
614 BumpPtrAllocator PredicateInfoAllocator;
615
616private:
617 ConstantInt *getConstantInt(const ValueLatticeElement &IV, Type *Ty) const {
618 return dyn_cast_or_null<ConstantInt>(Val: getConstant(LV: IV, Ty));
619 }
620
621 /// Push instruction \p I to the worklist.
622 void pushToWorkList(Instruction *I);
623
624 /// Push users of value \p V to the worklist.
625 void pushUsersToWorkList(Value *V);
626
627 /// Like pushUsersToWorkList(), but also prints a debug message with the
628 /// updated value.
629 void pushUsersToWorkListMsg(ValueLatticeElement &IV, Value *V);
630
631 // markConstant - Make a value be marked as "constant". If the value
632 // is not already a constant, add it to the instruction work list so that
633 // the users of the instruction are updated later.
634 bool markConstant(ValueLatticeElement &IV, Value *V, Constant *C,
635 bool MayIncludeUndef = false);
636
637 bool markConstant(Value *V, Constant *C) {
638 assert(!V->getType()->isStructTy() && "structs should use mergeInValue");
639 return markConstant(IV&: ValueState[V], V, C);
640 }
641
642 bool markNotConstant(ValueLatticeElement &IV, Value *V, Constant *C);
643
644 bool markNotNull(ValueLatticeElement &IV, Value *V) {
645 return markNotConstant(IV, V, C: Constant::getNullValue(Ty: V->getType()));
646 }
647
648 /// markConstantRange - Mark the object as constant range with \p CR. If the
649 /// object is not a constant range with the range \p CR, add it to the
650 /// instruction work list so that the users of the instruction are updated
651 /// later.
652 bool markConstantRange(ValueLatticeElement &IV, Value *V,
653 const ConstantRange &CR);
654
655 // markOverdefined - Make a value be marked as "overdefined". If the
656 // value is not already overdefined, add it to the overdefined instruction
657 // work list so that the users of the instruction are updated later.
658 bool markOverdefined(ValueLatticeElement &IV, Value *V);
659
660 /// Merge \p MergeWithV into \p IV and push \p V to the worklist, if \p IV
661 /// changes.
662 bool mergeInValue(ValueLatticeElement &IV, Value *V,
663 const ValueLatticeElement &MergeWithV,
664 ValueLatticeElement::MergeOptions Opts = {
665 /*MayIncludeUndef=*/false, /*CheckWiden=*/false});
666
667 /// getValueState - Return the ValueLatticeElement object that corresponds to
668 /// the value. This function handles the case when the value hasn't been seen
669 /// yet by properly seeding constants etc.
670 ValueLatticeElement &getValueState(Value *V) {
671 assert(!V->getType()->isStructTy() && "Should use getStructValueState");
672
673 auto I = ValueState.try_emplace(Key: V);
674 ValueLatticeElement &LV = I.first->second;
675
676 if (!I.second)
677 return LV; // Common case, already in the map.
678
679 if (auto *C = dyn_cast<Constant>(Val: V))
680 LV.markConstant(V: C); // Constants are constant
681
682 // All others are unknown by default.
683 return LV;
684 }
685
686 /// getStructValueState - Return the ValueLatticeElement object that
687 /// corresponds to the value/field pair. This function handles the case when
688 /// the value hasn't been seen yet by properly seeding constants etc.
689 ValueLatticeElement &getStructValueState(Value *V, unsigned i) {
690 assert(V->getType()->isStructTy() && "Should use getValueState");
691 assert(i < cast<StructType>(V->getType())->getNumElements() &&
692 "Invalid element #");
693
694 auto I = StructValueState.insert(
695 KV: std::make_pair(x: std::make_pair(x&: V, y&: i), y: ValueLatticeElement()));
696 ValueLatticeElement &LV = I.first->second;
697
698 if (!I.second)
699 return LV; // Common case, already in the map.
700
701 if (auto *C = dyn_cast<Constant>(Val: V)) {
702 Constant *Elt = C->getAggregateElement(Elt: i);
703
704 if (!Elt)
705 LV.markOverdefined(); // Unknown sort of constant.
706 else
707 LV.markConstant(V: Elt); // Constants are constant.
708 }
709
710 // All others are underdefined by default.
711 return LV;
712 }
713
714 /// Traverse the use-def chain of \p Call, marking itself and its users as
715 /// "unknown" on the way.
716 void invalidate(CallBase *Call) {
717 SmallVector<Instruction *, 64> ToInvalidate;
718 ToInvalidate.push_back(Elt: Call);
719
720 while (!ToInvalidate.empty()) {
721 Instruction *Inst = ToInvalidate.pop_back_val();
722
723 if (!Invalidated.insert(V: Inst).second)
724 continue;
725
726 if (!BBExecutable.count(Ptr: Inst->getParent()))
727 continue;
728
729 Value *V = nullptr;
730 // For return instructions we need to invalidate the tracked returns map.
731 // Anything else has its lattice in the value map.
732 if (auto *RetInst = dyn_cast<ReturnInst>(Val: Inst)) {
733 Function *F = RetInst->getParent()->getParent();
734 if (auto It = TrackedRetVals.find(Key: F); It != TrackedRetVals.end()) {
735 It->second = ValueLatticeElement();
736 V = F;
737 } else if (MRVFunctionsTracked.count(Ptr: F)) {
738 auto *STy = cast<StructType>(Val: F->getReturnType());
739 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I)
740 TrackedMultipleRetVals[{F, I}] = ValueLatticeElement();
741 V = F;
742 }
743 } else if (auto *STy = dyn_cast<StructType>(Val: Inst->getType())) {
744 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) {
745 if (auto It = StructValueState.find(Val: {Inst, I});
746 It != StructValueState.end()) {
747 It->second = ValueLatticeElement();
748 V = Inst;
749 }
750 }
751 } else if (auto It = ValueState.find(Val: Inst); It != ValueState.end()) {
752 It->second = ValueLatticeElement();
753 V = Inst;
754 }
755
756 if (V) {
757 LLVM_DEBUG(dbgs() << "Invalidated lattice for " << *V << "\n");
758
759 for (User *U : V->users())
760 if (auto *UI = dyn_cast<Instruction>(Val: U))
761 ToInvalidate.push_back(Elt: UI);
762
763 auto It = AdditionalUsers.find(Val: V);
764 if (It != AdditionalUsers.end())
765 for (User *U : It->second)
766 if (auto *UI = dyn_cast<Instruction>(Val: U))
767 ToInvalidate.push_back(Elt: UI);
768 }
769 }
770 }
771
772 /// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
773 /// work list if it is not already executable.
774 bool markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest);
775
776 // getFeasibleSuccessors - Return a vector of booleans to indicate which
777 // successors are reachable from a given terminator instruction.
778 void getFeasibleSuccessors(Instruction &TI, SmallVectorImpl<bool> &Succs);
779
780 // Add U as additional user of V.
781 void addAdditionalUser(Value *V, User *U) { AdditionalUsers[V].insert(X: U); }
782
783 void handlePredicate(Instruction *I, Value *CopyOf, const PredicateBase *PI);
784 void handleCallOverdefined(CallBase &CB);
785 void handleCallResult(CallBase &CB);
786 void handleCallArguments(CallBase &CB);
787 void handleExtractOfWithOverflow(ExtractValueInst &EVI,
788 const WithOverflowInst *WO, unsigned Idx);
789 bool isInstFullyOverDefined(Instruction &Inst);
790
791private:
792 friend class InstVisitor<SCCPInstVisitor>;
793
794 // visit implementations - Something changed in this instruction. Either an
795 // operand made a transition, or the instruction is newly executable. Change
796 // the value type of I to reflect these changes if appropriate.
797 void visitPHINode(PHINode &I);
798
799 // Terminators
800
801 void visitReturnInst(ReturnInst &I);
802 void visitTerminator(Instruction &TI);
803
804 void visitCastInst(CastInst &I);
805 void visitSelectInst(SelectInst &I);
806 void visitUnaryOperator(Instruction &I);
807 void visitFreezeInst(FreezeInst &I);
808 void visitBinaryOperator(Instruction &I);
809 void visitCmpInst(CmpInst &I);
810 void visitExtractValueInst(ExtractValueInst &EVI);
811 void visitInsertValueInst(InsertValueInst &IVI);
812
813 void visitCatchSwitchInst(CatchSwitchInst &CPI) {
814 markOverdefined(V: &CPI);
815 visitTerminator(TI&: CPI);
816 }
817
818 // Instructions that cannot be folded away.
819
820 void visitStoreInst(StoreInst &I);
821 void visitLoadInst(LoadInst &I);
822 void visitGetElementPtrInst(GetElementPtrInst &I);
823 void visitAllocaInst(AllocaInst &AI);
824
825 void visitInvokeInst(InvokeInst &II) {
826 visitCallBase(CB&: II);
827 visitTerminator(TI&: II);
828 }
829
830 void visitCallBrInst(CallBrInst &CBI) {
831 visitCallBase(CB&: CBI);
832 visitTerminator(TI&: CBI);
833 }
834
835 void visitCallBase(CallBase &CB);
836 void visitResumeInst(ResumeInst &I) { /*returns void*/
837 }
838 void visitUnreachableInst(UnreachableInst &I) { /*returns void*/
839 }
840 void visitFenceInst(FenceInst &I) { /*returns void*/
841 }
842
843 void visitInstruction(Instruction &I);
844
845public:
846 void addPredicateInfo(Function &F, DominatorTree &DT, AssumptionCache &AC) {
847 FnPredicateInfo.insert(KV: {&F, std::make_unique<PredicateInfo>(
848 args&: F, args&: DT, args&: AC, args&: PredicateInfoAllocator)});
849 }
850
851 void removeSSACopies(Function &F) {
852 auto It = FnPredicateInfo.find(Val: &F);
853 if (It == FnPredicateInfo.end())
854 return;
855
856 for (BasicBlock &BB : F) {
857 for (Instruction &Inst : llvm::make_early_inc_range(Range&: BB)) {
858 if (auto *BC = dyn_cast<BitCastInst>(Val: &Inst)) {
859 if (BC->getType() == BC->getOperand(i_nocapture: 0)->getType()) {
860 if (It->second->getPredicateInfoFor(V: &Inst)) {
861 Value *Op = BC->getOperand(i_nocapture: 0);
862 Inst.replaceAllUsesWith(V: Op);
863 Inst.eraseFromParent();
864 }
865 }
866 }
867 }
868 }
869 }
870
871 void visitCallInst(CallInst &I) { visitCallBase(CB&: I); }
872
873 bool markBlockExecutable(BasicBlock *BB);
874
875 const PredicateBase *getPredicateInfoFor(Instruction *I) {
876 auto It = FnPredicateInfo.find(Val: I->getParent()->getParent());
877 if (It == FnPredicateInfo.end())
878 return nullptr;
879 return It->second->getPredicateInfoFor(V: I);
880 }
881
882 SCCPInstVisitor(const DataLayout &DL,
883 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
884 LLVMContext &Ctx)
885 : DL(DL), GetTLI(GetTLI), Ctx(Ctx) {}
886
887 void trackValueOfGlobalVariable(GlobalVariable *GV) {
888 // We only track the contents of scalar globals.
889 if (GV->getValueType()->isSingleValueType()) {
890 ValueLatticeElement &IV = TrackedGlobals[GV];
891 IV.markConstant(V: GV->getInitializer());
892 }
893 }
894
895 void addTrackedFunction(Function *F) {
896 // Add an entry, F -> undef.
897 if (auto *STy = dyn_cast<StructType>(Val: F->getReturnType())) {
898 MRVFunctionsTracked.insert(Ptr: F);
899 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
900 TrackedMultipleRetVals.try_emplace(Key: std::make_pair(x&: F, y&: i));
901 } else if (!F->getReturnType()->isVoidTy())
902 TrackedRetVals.try_emplace(Key: F);
903 }
904
905 void addToMustPreserveReturnsInFunctions(Function *F) {
906 MustPreserveReturnsInFunctions.insert(Ptr: F);
907 }
908
909 bool mustPreserveReturn(Function *F) {
910 return MustPreserveReturnsInFunctions.count(Ptr: F);
911 }
912
913 void addArgumentTrackedFunction(Function *F) {
914 TrackingIncomingArguments.insert(Ptr: F);
915 }
916
917 bool isArgumentTrackedFunction(Function *F) {
918 return TrackingIncomingArguments.count(Ptr: F);
919 }
920
921 const SmallPtrSetImpl<Function *> &getArgumentTrackedFunctions() const {
922 return TrackingIncomingArguments;
923 }
924
925 void solve();
926
927 bool resolvedUndef(Instruction &I);
928
929 bool resolvedUndefsIn(Function &F);
930
931 bool isBlockExecutable(BasicBlock *BB) const {
932 return BBExecutable.count(Ptr: BB);
933 }
934
935 bool isEdgeFeasible(BasicBlock *From, BasicBlock *To) const;
936
937 std::vector<ValueLatticeElement> getStructLatticeValueFor(Value *V) const {
938 std::vector<ValueLatticeElement> StructValues;
939 auto *STy = dyn_cast<StructType>(Val: V->getType());
940 assert(STy && "getStructLatticeValueFor() can be called only on structs");
941 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
942 auto I = StructValueState.find(Val: std::make_pair(x&: V, y&: i));
943 assert(I != StructValueState.end() && "Value not in valuemap!");
944 StructValues.push_back(x: I->second);
945 }
946 return StructValues;
947 }
948
949 void removeLatticeValueFor(Value *V) { ValueState.erase(Val: V); }
950
951 /// Invalidate the Lattice Value of \p Call and its users after specializing
952 /// the call. Then recompute it.
953 void resetLatticeValueFor(CallBase *Call) {
954 // Calls to void returning functions do not need invalidation.
955 Function *F = Call->getCalledFunction();
956 (void)F;
957 assert(!F->getReturnType()->isVoidTy() &&
958 (TrackedRetVals.count(F) || MRVFunctionsTracked.count(F)) &&
959 "All non void specializations should be tracked");
960 invalidate(Call);
961 handleCallResult(CB&: *Call);
962 }
963
964 const ValueLatticeElement &getLatticeValueFor(Value *V) const {
965 assert(!V->getType()->isStructTy() &&
966 "Should use getStructLatticeValueFor");
967 auto I = ValueState.find(Val: V);
968 assert(I != ValueState.end() &&
969 "V not found in ValueState nor Paramstate map!");
970 return I->second;
971 }
972
973 const MapVector<Function *, ValueLatticeElement> &getTrackedRetVals() const {
974 return TrackedRetVals;
975 }
976
977 const DenseMap<GlobalVariable *, ValueLatticeElement> &
978 getTrackedGlobals() const {
979 return TrackedGlobals;
980 }
981
982 const SmallPtrSet<Function *, 16> &getMRVFunctionsTracked() const {
983 return MRVFunctionsTracked;
984 }
985
986 void markOverdefined(Value *V) {
987 if (auto *STy = dyn_cast<StructType>(Val: V->getType()))
988 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
989 markOverdefined(IV&: getStructValueState(V, i), V);
990 else
991 markOverdefined(IV&: ValueState[V], V);
992 }
993
994 ValueLatticeElement getArgAttributeVL(Argument *A) {
995 if (A->getType()->isIntOrIntVectorTy()) {
996 if (std::optional<ConstantRange> Range = A->getRange())
997 return ValueLatticeElement::getRange(CR: *Range);
998 }
999 if (A->hasNonNullAttr())
1000 return ValueLatticeElement::getNot(C: Constant::getNullValue(Ty: A->getType()));
1001 // Assume nothing about the incoming arguments without attributes.
1002 return ValueLatticeElement::getOverdefined();
1003 }
1004
1005 void trackValueOfArgument(Argument *A) {
1006 if (A->getType()->isStructTy())
1007 return (void)markOverdefined(V: A);
1008 mergeInValue(IV&: ValueState[A], V: A, MergeWithV: getArgAttributeVL(A));
1009 }
1010
1011 bool isStructLatticeConstant(Function *F, StructType *STy);
1012
1013 Constant *getConstant(const ValueLatticeElement &LV, Type *Ty) const;
1014
1015 Constant *getConstantOrNull(Value *V) const;
1016
1017 void setLatticeValueForSpecializationArguments(Function *F,
1018 const SmallVectorImpl<ArgInfo> &Args);
1019
1020 void markFunctionUnreachable(Function *F) {
1021 for (auto &BB : *F)
1022 BBExecutable.erase(Ptr: &BB);
1023 }
1024
1025 void solveWhileResolvedUndefsIn(Module &M) {
1026 bool ResolvedUndefs = true;
1027 while (ResolvedUndefs) {
1028 solve();
1029 ResolvedUndefs = false;
1030 for (Function &F : M)
1031 ResolvedUndefs |= resolvedUndefsIn(F);
1032 }
1033 }
1034
1035 void solveWhileResolvedUndefsIn(SmallVectorImpl<Function *> &WorkList) {
1036 bool ResolvedUndefs = true;
1037 while (ResolvedUndefs) {
1038 solve();
1039 ResolvedUndefs = false;
1040 for (Function *F : WorkList)
1041 ResolvedUndefs |= resolvedUndefsIn(F&: *F);
1042 }
1043 }
1044
1045 void solveWhileResolvedUndefs() {
1046 bool ResolvedUndefs = true;
1047 while (ResolvedUndefs) {
1048 solve();
1049 ResolvedUndefs = false;
1050 for (Value *V : Invalidated)
1051 if (auto *I = dyn_cast<Instruction>(Val: V))
1052 ResolvedUndefs |= resolvedUndef(I&: *I);
1053 }
1054 Invalidated.clear();
1055 }
1056};
1057
1058} // namespace llvm
1059
1060bool SCCPInstVisitor::markBlockExecutable(BasicBlock *BB) {
1061 if (!BBExecutable.insert(Ptr: BB).second)
1062 return false;
1063 LLVM_DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << '\n');
1064 BBWorkList.push_back(Elt: BB); // Add the block to the work list!
1065 return true;
1066}
1067
1068void SCCPInstVisitor::pushToWorkList(Instruction *I) {
1069 // If we're currently visiting a block, do not push any instructions in the
1070 // same blocks that are after the current one, as they will be visited
1071 // anyway. We do have to push updates to earlier instructions (e.g. phi
1072 // nodes or loads of tracked globals).
1073 if (CurI && I->getParent() == CurI->getParent() && !I->comesBefore(Other: CurI))
1074 return;
1075 // Only push instructions in already visited blocks. Otherwise we'll handle
1076 // it when we visit the block for the first time.
1077 if (BBVisited.contains(Ptr: I->getParent()))
1078 InstWorkList.insert(X: I);
1079}
1080
1081void SCCPInstVisitor::pushUsersToWorkList(Value *V) {
1082 for (User *U : V->users())
1083 if (auto *UI = dyn_cast<Instruction>(Val: U))
1084 pushToWorkList(I: UI);
1085
1086 auto Iter = AdditionalUsers.find(Val: V);
1087 if (Iter != AdditionalUsers.end()) {
1088 // Copy additional users before notifying them of changes, because new
1089 // users may be added, potentially invalidating the iterator.
1090 SmallVector<Instruction *, 2> ToNotify;
1091 for (User *U : Iter->second)
1092 if (auto *UI = dyn_cast<Instruction>(Val: U))
1093 ToNotify.push_back(Elt: UI);
1094 for (Instruction *UI : ToNotify)
1095 pushToWorkList(I: UI);
1096 }
1097}
1098
1099void SCCPInstVisitor::pushUsersToWorkListMsg(ValueLatticeElement &IV,
1100 Value *V) {
1101 LLVM_DEBUG(dbgs() << "updated " << IV << ": " << *V << '\n');
1102 pushUsersToWorkList(V);
1103}
1104
1105bool SCCPInstVisitor::markConstant(ValueLatticeElement &IV, Value *V,
1106 Constant *C, bool MayIncludeUndef) {
1107 if (!IV.markConstant(V: C, MayIncludeUndef))
1108 return false;
1109 LLVM_DEBUG(dbgs() << "markConstant: " << *C << ": " << *V << '\n');
1110 pushUsersToWorkList(V);
1111 return true;
1112}
1113
1114bool SCCPInstVisitor::markNotConstant(ValueLatticeElement &IV, Value *V,
1115 Constant *C) {
1116 if (!IV.markNotConstant(V: C))
1117 return false;
1118 LLVM_DEBUG(dbgs() << "markNotConstant: " << *C << ": " << *V << '\n');
1119 pushUsersToWorkList(V);
1120 return true;
1121}
1122
1123bool SCCPInstVisitor::markConstantRange(ValueLatticeElement &IV, Value *V,
1124 const ConstantRange &CR) {
1125 if (!IV.markConstantRange(NewR: CR))
1126 return false;
1127 LLVM_DEBUG(dbgs() << "markConstantRange: " << CR << ": " << *V << '\n');
1128 pushUsersToWorkList(V);
1129 return true;
1130}
1131
1132bool SCCPInstVisitor::markOverdefined(ValueLatticeElement &IV, Value *V) {
1133 if (!IV.markOverdefined())
1134 return false;
1135
1136 LLVM_DEBUG(dbgs() << "markOverdefined: ";
1137 if (auto *F = dyn_cast<Function>(V)) dbgs()
1138 << "Function '" << F->getName() << "'\n";
1139 else dbgs() << *V << '\n');
1140 // Only instructions go on the work list
1141 pushUsersToWorkList(V);
1142 return true;
1143}
1144
1145bool SCCPInstVisitor::isStructLatticeConstant(Function *F, StructType *STy) {
1146 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1147 const auto &It = TrackedMultipleRetVals.find(Key: std::make_pair(x&: F, y&: i));
1148 assert(It != TrackedMultipleRetVals.end());
1149 if (!SCCPSolver::isConstant(LV: It->second))
1150 return false;
1151 }
1152 return true;
1153}
1154
1155Constant *SCCPInstVisitor::getConstant(const ValueLatticeElement &LV,
1156 Type *Ty) const {
1157 if (LV.isConstant()) {
1158 Constant *C = LV.getConstant();
1159 assert(C->getType() == Ty && "Type mismatch");
1160 return C;
1161 }
1162
1163 if (LV.isConstantRange()) {
1164 const auto &CR = LV.getConstantRange();
1165 if (CR.getSingleElement())
1166 return ConstantInt::get(Ty, V: *CR.getSingleElement());
1167 }
1168 return nullptr;
1169}
1170
1171Constant *SCCPInstVisitor::getConstantOrNull(Value *V) const {
1172 Constant *Const = nullptr;
1173 if (V->getType()->isStructTy()) {
1174 std::vector<ValueLatticeElement> LVs = getStructLatticeValueFor(V);
1175 if (any_of(Range&: LVs, P: SCCPSolver::isOverdefined))
1176 return nullptr;
1177 std::vector<Constant *> ConstVals;
1178 auto *ST = cast<StructType>(Val: V->getType());
1179 for (unsigned I = 0, E = ST->getNumElements(); I != E; ++I) {
1180 const ValueLatticeElement &LV = LVs[I];
1181 ConstVals.push_back(x: SCCPSolver::isConstant(LV)
1182 ? getConstant(LV, Ty: ST->getElementType(N: I))
1183 : UndefValue::get(T: ST->getElementType(N: I)));
1184 }
1185 Const = ConstantStruct::get(T: ST, V: ConstVals);
1186 } else {
1187 const ValueLatticeElement &LV = getLatticeValueFor(V);
1188 if (SCCPSolver::isOverdefined(LV))
1189 return nullptr;
1190 Const = SCCPSolver::isConstant(LV) ? getConstant(LV, Ty: V->getType())
1191 : UndefValue::get(T: V->getType());
1192 }
1193 assert(Const && "Constant is nullptr here!");
1194 return Const;
1195}
1196
1197void SCCPInstVisitor::setLatticeValueForSpecializationArguments(Function *F,
1198 const SmallVectorImpl<ArgInfo> &Args) {
1199 assert(!Args.empty() && "Specialization without arguments");
1200 assert(F->arg_size() == Args[0].Formal->getParent()->arg_size() &&
1201 "Functions should have the same number of arguments");
1202
1203 auto Iter = Args.begin();
1204 Function::arg_iterator NewArg = F->arg_begin();
1205 Function::arg_iterator OldArg = Args[0].Formal->getParent()->arg_begin();
1206 for (auto End = F->arg_end(); NewArg != End; ++NewArg, ++OldArg) {
1207
1208 LLVM_DEBUG(dbgs() << "SCCP: Marking argument "
1209 << NewArg->getNameOrAsOperand() << "\n");
1210
1211 // Mark the argument constants in the new function
1212 // or copy the lattice state over from the old function.
1213 if (Iter != Args.end() && Iter->Formal == &*OldArg) {
1214 if (auto *STy = dyn_cast<StructType>(Val: NewArg->getType())) {
1215 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) {
1216 ValueLatticeElement &NewValue = StructValueState[{&*NewArg, I}];
1217 NewValue.markConstant(V: Iter->Actual->getAggregateElement(Elt: I));
1218 }
1219 } else {
1220 ValueState[&*NewArg].markConstant(V: Iter->Actual);
1221 }
1222 ++Iter;
1223 } else {
1224 if (auto *STy = dyn_cast<StructType>(Val: NewArg->getType())) {
1225 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) {
1226 ValueLatticeElement &NewValue = StructValueState[{&*NewArg, I}];
1227 NewValue = StructValueState[{&*OldArg, I}];
1228 }
1229 } else {
1230 ValueLatticeElement &NewValue = ValueState[&*NewArg];
1231 NewValue = ValueState[&*OldArg];
1232 }
1233 }
1234 }
1235}
1236
1237void SCCPInstVisitor::visitInstruction(Instruction &I) {
1238 // All the instructions we don't do any special handling for just
1239 // go to overdefined.
1240 LLVM_DEBUG(dbgs() << "SCCP: Don't know how to handle: " << I << '\n');
1241 markOverdefined(V: &I);
1242}
1243
1244bool SCCPInstVisitor::mergeInValue(ValueLatticeElement &IV, Value *V,
1245 const ValueLatticeElement &MergeWithV,
1246 ValueLatticeElement::MergeOptions Opts) {
1247 if (IV.mergeIn(RHS: MergeWithV, Opts)) {
1248 pushUsersToWorkList(V);
1249 LLVM_DEBUG(dbgs() << "Merged " << MergeWithV << " into " << *V << " : "
1250 << IV << "\n");
1251 return true;
1252 }
1253 return false;
1254}
1255
1256bool SCCPInstVisitor::markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
1257 if (!KnownFeasibleEdges.insert(V: Edge(Source, Dest)).second)
1258 return false; // This edge is already known to be executable!
1259
1260 if (!markBlockExecutable(BB: Dest)) {
1261 // If the destination is already executable, we just made an *edge*
1262 // feasible that wasn't before. Revisit the PHI nodes in the block
1263 // because they have potentially new operands.
1264 LLVM_DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
1265 << " -> " << Dest->getName() << '\n');
1266
1267 for (PHINode &PN : Dest->phis())
1268 pushToWorkList(I: &PN);
1269 }
1270 return true;
1271}
1272
1273// getFeasibleSuccessors - Return a vector of booleans to indicate which
1274// successors are reachable from a given terminator instruction.
1275void SCCPInstVisitor::getFeasibleSuccessors(Instruction &TI,
1276 SmallVectorImpl<bool> &Succs) {
1277 Succs.resize(N: TI.getNumSuccessors());
1278 if (isa<UncondBrInst>(Val: TI)) {
1279 Succs[0] = true;
1280 return;
1281 }
1282
1283 if (auto *BI = dyn_cast<CondBrInst>(Val: &TI)) {
1284 const ValueLatticeElement &BCValue = getValueState(V: BI->getCondition());
1285 ConstantInt *CI = getConstantInt(IV: BCValue, Ty: BI->getCondition()->getType());
1286 if (!CI) {
1287 // Overdefined condition variables, and branches on unfoldable constant
1288 // conditions, mean the branch could go either way.
1289 if (!BCValue.isUnknownOrUndef())
1290 Succs[0] = Succs[1] = true;
1291 return;
1292 }
1293
1294 // Constant condition variables mean the branch can only go a single way.
1295 Succs[CI->isZero()] = true;
1296 return;
1297 }
1298
1299 // We cannot analyze special terminators, so consider all successors
1300 // executable.
1301 if (TI.isSpecialTerminator()) {
1302 Succs.assign(NumElts: TI.getNumSuccessors(), Elt: true);
1303 return;
1304 }
1305
1306 if (auto *SI = dyn_cast<SwitchInst>(Val: &TI)) {
1307 if (!SI->getNumCases()) {
1308 Succs[0] = true;
1309 return;
1310 }
1311 const ValueLatticeElement &SCValue = getValueState(V: SI->getCondition());
1312 if (ConstantInt *CI =
1313 getConstantInt(IV: SCValue, Ty: SI->getCondition()->getType())) {
1314 Succs[SI->findCaseValue(C: CI)->getSuccessorIndex()] = true;
1315 return;
1316 }
1317
1318 // TODO: Switch on undef is UB. Stop passing false once the rest of LLVM
1319 // is ready.
1320 if (SCValue.isConstantRange(/*UndefAllowed=*/false)) {
1321 const ConstantRange &Range = SCValue.getConstantRange();
1322 unsigned ReachableCaseCount = 0;
1323 for (const auto &Case : SI->cases()) {
1324 const APInt &CaseValue = Case.getCaseValue()->getValue();
1325 if (Range.contains(Val: CaseValue)) {
1326 Succs[Case.getSuccessorIndex()] = true;
1327 ++ReachableCaseCount;
1328 }
1329 }
1330
1331 Succs[SI->case_default()->getSuccessorIndex()] =
1332 Range.isSizeLargerThan(MaxSize: ReachableCaseCount);
1333 return;
1334 }
1335
1336 // Overdefined or unknown condition? All destinations are executable!
1337 if (!SCValue.isUnknownOrUndef())
1338 Succs.assign(NumElts: TI.getNumSuccessors(), Elt: true);
1339 return;
1340 }
1341
1342 // In case of indirect branch and its address is a blockaddress, we mark
1343 // the target as executable.
1344 if (auto *IBR = dyn_cast<IndirectBrInst>(Val: &TI)) {
1345 // Casts are folded by visitCastInst.
1346 const ValueLatticeElement &IBRValue = getValueState(V: IBR->getAddress());
1347 BlockAddress *Addr = dyn_cast_or_null<BlockAddress>(
1348 Val: getConstant(LV: IBRValue, Ty: IBR->getAddress()->getType()));
1349 if (!Addr) { // Overdefined or unknown condition?
1350 // All destinations are executable!
1351 if (!IBRValue.isUnknownOrUndef())
1352 Succs.assign(NumElts: TI.getNumSuccessors(), Elt: true);
1353 return;
1354 }
1355
1356 BasicBlock *T = Addr->getBasicBlock();
1357 assert(Addr->getFunction() == T->getParent() &&
1358 "Block address of a different function ?");
1359 for (unsigned i = 0; i < IBR->getNumSuccessors(); ++i) {
1360 // This is the target.
1361 if (IBR->getDestination(i) == T) {
1362 Succs[i] = true;
1363 return;
1364 }
1365 }
1366
1367 // If we didn't find our destination in the IBR successor list, then we
1368 // have undefined behavior. Its ok to assume no successor is executable.
1369 return;
1370 }
1371
1372 LLVM_DEBUG(dbgs() << "Unknown terminator instruction: " << TI << '\n');
1373 llvm_unreachable("SCCP: Don't know how to handle this terminator!");
1374}
1375
1376// isEdgeFeasible - Return true if the control flow edge from the 'From' basic
1377// block to the 'To' basic block is currently feasible.
1378bool SCCPInstVisitor::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const {
1379 // Check if we've called markEdgeExecutable on the edge yet. (We could
1380 // be more aggressive and try to consider edges which haven't been marked
1381 // yet, but there isn't any need.)
1382 return KnownFeasibleEdges.count(V: Edge(From, To));
1383}
1384
1385// visit Implementations - Something changed in this instruction, either an
1386// operand made a transition, or the instruction is newly executable. Change
1387// the value type of I to reflect these changes if appropriate. This method
1388// makes sure to do the following actions:
1389//
1390// 1. If a phi node merges two constants in, and has conflicting value coming
1391// from different branches, or if the PHI node merges in an overdefined
1392// value, then the PHI node becomes overdefined.
1393// 2. If a phi node merges only constants in, and they all agree on value, the
1394// PHI node becomes a constant value equal to that.
1395// 3. If V <- x (op) y && isConstant(x) && isConstant(y) V = Constant
1396// 4. If V <- x (op) y && (isOverdefined(x) || isOverdefined(y)) V = Overdefined
1397// 5. If V <- MEM or V <- CALL or V <- (unknown) then V = Overdefined
1398// 6. If a conditional branch has a value that is constant, make the selected
1399// destination executable
1400// 7. If a conditional branch has a value that is overdefined, make all
1401// successors executable.
1402void SCCPInstVisitor::visitPHINode(PHINode &PN) {
1403 // Super-extra-high-degree PHI nodes are unlikely to ever be marked constant,
1404 // and slow us down a lot. Just mark them overdefined.
1405 if (PN.getNumIncomingValues() > 64)
1406 return (void)markOverdefined(V: &PN);
1407
1408 if (isInstFullyOverDefined(Inst&: PN))
1409 return;
1410 SmallVector<unsigned> FeasibleIncomingIndices;
1411 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
1412 if (!isEdgeFeasible(From: PN.getIncomingBlock(i), To: PN.getParent()))
1413 continue;
1414 FeasibleIncomingIndices.push_back(Elt: i);
1415 }
1416
1417 // Look at all of the executable operands of the PHI node. If any of them
1418 // are overdefined, the PHI becomes overdefined as well. If they are all
1419 // constant, and they agree with each other, the PHI becomes the identical
1420 // constant. If they are constant and don't agree, the PHI is a constant
1421 // range. If there are no executable operands, the PHI remains unknown.
1422 if (StructType *STy = dyn_cast<StructType>(Val: PN.getType())) {
1423 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1424 ValueLatticeElement PhiState = getStructValueState(V: &PN, i);
1425 if (PhiState.isOverdefined())
1426 continue;
1427 for (unsigned j : FeasibleIncomingIndices) {
1428 const ValueLatticeElement &IV =
1429 getStructValueState(V: PN.getIncomingValue(i: j), i);
1430 PhiState.mergeIn(RHS: IV);
1431 if (PhiState.isOverdefined())
1432 break;
1433 }
1434 ValueLatticeElement &PhiStateRef = getStructValueState(V: &PN, i);
1435 mergeInValue(IV&: PhiStateRef, V: &PN, MergeWithV: PhiState,
1436 Opts: ValueLatticeElement::MergeOptions().setMaxWidenSteps(
1437 FeasibleIncomingIndices.size() + 1));
1438 PhiStateRef.setNumRangeExtensions(
1439 std::max(a: (unsigned)FeasibleIncomingIndices.size(),
1440 b: PhiStateRef.getNumRangeExtensions()));
1441 }
1442 } else {
1443 ValueLatticeElement PhiState = getValueState(V: &PN);
1444 for (unsigned i : FeasibleIncomingIndices) {
1445 const ValueLatticeElement &IV = getValueState(V: PN.getIncomingValue(i));
1446 PhiState.mergeIn(RHS: IV);
1447 if (PhiState.isOverdefined())
1448 break;
1449 }
1450 // We allow up to 1 range extension per active incoming value and one
1451 // additional extension. Note that we manually adjust the number of range
1452 // extensions to match the number of active incoming values. This helps to
1453 // limit multiple extensions caused by the same incoming value, if other
1454 // incoming values are equal.
1455 ValueLatticeElement &PhiStateRef = ValueState[&PN];
1456 mergeInValue(IV&: PhiStateRef, V: &PN, MergeWithV: PhiState,
1457 Opts: ValueLatticeElement::MergeOptions().setMaxWidenSteps(
1458 FeasibleIncomingIndices.size() + 1));
1459 PhiStateRef.setNumRangeExtensions(
1460 std::max(a: (unsigned)FeasibleIncomingIndices.size(),
1461 b: PhiStateRef.getNumRangeExtensions()));
1462 }
1463}
1464
1465void SCCPInstVisitor::visitReturnInst(ReturnInst &I) {
1466 if (I.getNumOperands() == 0)
1467 return; // ret void
1468
1469 Function *F = I.getParent()->getParent();
1470 Value *ResultOp = I.getOperand(i_nocapture: 0);
1471
1472 // If we are tracking the return value of this function, merge it in.
1473 if (!TrackedRetVals.empty() && !ResultOp->getType()->isStructTy()) {
1474 auto TFRVI = TrackedRetVals.find(Key: F);
1475 if (TFRVI != TrackedRetVals.end()) {
1476 mergeInValue(IV&: TFRVI->second, V: F, MergeWithV: getValueState(V: ResultOp));
1477 return;
1478 }
1479 }
1480
1481 // Handle functions that return multiple values.
1482 if (!TrackedMultipleRetVals.empty()) {
1483 if (auto *STy = dyn_cast<StructType>(Val: ResultOp->getType()))
1484 if (MRVFunctionsTracked.count(Ptr: F))
1485 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1486 mergeInValue(IV&: TrackedMultipleRetVals[std::make_pair(x&: F, y&: i)], V: F,
1487 MergeWithV: getStructValueState(V: ResultOp, i));
1488 }
1489}
1490
1491void SCCPInstVisitor::visitTerminator(Instruction &TI) {
1492 SmallVector<bool, 16> SuccFeasible;
1493 getFeasibleSuccessors(TI, Succs&: SuccFeasible);
1494
1495 BasicBlock *BB = TI.getParent();
1496
1497 // Mark all feasible successors executable.
1498 for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
1499 if (SuccFeasible[i])
1500 markEdgeExecutable(Source: BB, Dest: TI.getSuccessor(Idx: i));
1501}
1502
1503void SCCPInstVisitor::visitCastInst(CastInst &I) {
1504 // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1505 // discover a concrete value later.
1506 if (ValueState[&I].isOverdefined())
1507 return;
1508
1509 if (auto *BC = dyn_cast<BitCastInst>(Val: &I)) {
1510 if (BC->getType() == BC->getOperand(i_nocapture: 0)->getType()) {
1511 if (const PredicateBase *PI = getPredicateInfoFor(I: &I)) {
1512 handlePredicate(I: &I, CopyOf: I.getOperand(i_nocapture: 0), PI);
1513 return;
1514 }
1515 }
1516 }
1517
1518 const ValueLatticeElement &OpSt = getValueState(V: I.getOperand(i_nocapture: 0));
1519 if (OpSt.isUnknownOrUndef())
1520 return;
1521
1522 if (Constant *OpC = getConstant(LV: OpSt, Ty: I.getOperand(i_nocapture: 0)->getType())) {
1523 // Fold the constant as we build.
1524 if (Constant *C =
1525 ConstantFoldCastOperand(Opcode: I.getOpcode(), C: OpC, DestTy: I.getType(), DL)) {
1526 auto &LV = ValueState[&I];
1527 mergeInValue(IV&: LV, V: &I, MergeWithV: ValueLatticeElement::get(C));
1528 return;
1529 }
1530 }
1531
1532 // Ignore bitcasts, as they may change the number of vector elements.
1533 if (I.getDestTy()->isIntOrIntVectorTy() &&
1534 I.getSrcTy()->isIntOrIntVectorTy() &&
1535 I.getOpcode() != Instruction::BitCast) {
1536 ConstantRange OpRange =
1537 OpSt.asConstantRange(Ty: I.getSrcTy(), /*UndefAllowed=*/false);
1538 auto &LV = getValueState(V: &I);
1539
1540 Type *DestTy = I.getDestTy();
1541 ConstantRange Res = ConstantRange::getEmpty(BitWidth: DestTy->getScalarSizeInBits());
1542 if (auto *Trunc = dyn_cast<TruncInst>(Val: &I))
1543 Res = OpRange.truncate(BitWidth: DestTy->getScalarSizeInBits(),
1544 NoWrapKind: Trunc->getNoWrapKind());
1545 else
1546 Res = OpRange.castOp(CastOp: I.getOpcode(), BitWidth: DestTy->getScalarSizeInBits());
1547 mergeInValue(IV&: LV, V: &I, MergeWithV: ValueLatticeElement::getRange(CR: Res));
1548 } else
1549 markOverdefined(V: &I);
1550}
1551
1552void SCCPInstVisitor::handleExtractOfWithOverflow(ExtractValueInst &EVI,
1553 const WithOverflowInst *WO,
1554 unsigned Idx) {
1555 Value *LHS = WO->getLHS(), *RHS = WO->getRHS();
1556 Type *Ty = LHS->getType();
1557
1558 addAdditionalUser(V: LHS, U: &EVI);
1559 addAdditionalUser(V: RHS, U: &EVI);
1560
1561 const ValueLatticeElement &L = getValueState(V: LHS);
1562 if (L.isUnknownOrUndef())
1563 return; // Wait to resolve.
1564 ConstantRange LR = L.asConstantRange(Ty, /*UndefAllowed=*/false);
1565
1566 const ValueLatticeElement &R = getValueState(V: RHS);
1567 if (R.isUnknownOrUndef())
1568 return; // Wait to resolve.
1569
1570 ConstantRange RR = R.asConstantRange(Ty, /*UndefAllowed=*/false);
1571 if (Idx == 0) {
1572 ConstantRange Res = LR.binaryOp(BinOp: WO->getBinaryOp(), Other: RR);
1573 mergeInValue(IV&: ValueState[&EVI], V: &EVI, MergeWithV: ValueLatticeElement::getRange(CR: Res));
1574 } else {
1575 assert(Idx == 1 && "Index can only be 0 or 1");
1576 ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
1577 BinOp: WO->getBinaryOp(), Other: RR, NoWrapKind: WO->getNoWrapKind());
1578 if (NWRegion.contains(CR: LR))
1579 return (void)markConstant(V: &EVI, C: ConstantInt::getFalse(Ty: EVI.getType()));
1580 markOverdefined(V: &EVI);
1581 }
1582}
1583
1584void SCCPInstVisitor::visitExtractValueInst(ExtractValueInst &EVI) {
1585 // If this returns a struct, mark all elements over defined, we don't track
1586 // structs in structs.
1587 if (EVI.getType()->isStructTy())
1588 return (void)markOverdefined(V: &EVI);
1589
1590 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1591 // discover a concrete value later.
1592 if (ValueState[&EVI].isOverdefined())
1593 return (void)markOverdefined(V: &EVI);
1594
1595 // If this is extracting from more than one level of struct, we don't know.
1596 if (EVI.getNumIndices() != 1)
1597 return (void)markOverdefined(V: &EVI);
1598
1599 Value *AggVal = EVI.getAggregateOperand();
1600 if (AggVal->getType()->isStructTy()) {
1601 unsigned i = *EVI.idx_begin();
1602 if (auto *WO = dyn_cast<WithOverflowInst>(Val: AggVal))
1603 return handleExtractOfWithOverflow(EVI, WO, Idx: i);
1604 ValueLatticeElement EltVal = getStructValueState(V: AggVal, i);
1605 mergeInValue(IV&: ValueState[&EVI], V: &EVI, MergeWithV: EltVal);
1606 } else {
1607 // Otherwise, must be extracting from an array.
1608 return (void)markOverdefined(V: &EVI);
1609 }
1610}
1611
1612void SCCPInstVisitor::visitInsertValueInst(InsertValueInst &IVI) {
1613 auto *STy = dyn_cast<StructType>(Val: IVI.getType());
1614 if (!STy)
1615 return (void)markOverdefined(V: &IVI);
1616
1617 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1618 // discover a concrete value later.
1619 if (ValueState[&IVI].isOverdefined())
1620 return (void)markOverdefined(V: &IVI);
1621
1622 // If this has more than one index, we can't handle it, drive all results to
1623 // undef.
1624 if (IVI.getNumIndices() != 1)
1625 return (void)markOverdefined(V: &IVI);
1626
1627 Value *Aggr = IVI.getAggregateOperand();
1628 unsigned Idx = *IVI.idx_begin();
1629
1630 // Compute the result based on what we're inserting.
1631 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1632 // This passes through all values that aren't the inserted element.
1633 if (i != Idx) {
1634 ValueLatticeElement EltVal = getStructValueState(V: Aggr, i);
1635 mergeInValue(IV&: getStructValueState(V: &IVI, i), V: &IVI, MergeWithV: EltVal);
1636 continue;
1637 }
1638
1639 Value *Val = IVI.getInsertedValueOperand();
1640 if (Val->getType()->isStructTy())
1641 // We don't track structs in structs.
1642 markOverdefined(IV&: getStructValueState(V: &IVI, i), V: &IVI);
1643 else {
1644 ValueLatticeElement InVal = getValueState(V: Val);
1645 mergeInValue(IV&: getStructValueState(V: &IVI, i), V: &IVI, MergeWithV: InVal);
1646 }
1647 }
1648}
1649
1650void SCCPInstVisitor::visitSelectInst(SelectInst &I) {
1651 // If this select returns a struct, just mark the result overdefined.
1652 // TODO: We could do a lot better than this if code actually uses this.
1653 if (I.getType()->isStructTy())
1654 return (void)markOverdefined(V: &I);
1655
1656 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1657 // discover a concrete value later.
1658 if (ValueState[&I].isOverdefined())
1659 return (void)markOverdefined(V: &I);
1660
1661 const ValueLatticeElement &CondValue = getValueState(V: I.getCondition());
1662 if (CondValue.isUnknownOrUndef())
1663 return;
1664
1665 if (ConstantInt *CondCB =
1666 getConstantInt(IV: CondValue, Ty: I.getCondition()->getType())) {
1667 Value *OpVal = CondCB->isZero() ? I.getFalseValue() : I.getTrueValue();
1668 const ValueLatticeElement &OpValState = getValueState(V: OpVal);
1669 // Safety: ValueState[&I] doesn't invalidate OpValState since it is already
1670 // in the map.
1671 assert(ValueState.contains(&I) && "&I is not in ValueState map.");
1672 mergeInValue(IV&: ValueState[&I], V: &I, MergeWithV: OpValState);
1673 return;
1674 }
1675
1676 // Otherwise, the condition is overdefined or a constant we can't evaluate.
1677 // See if we can produce something better than overdefined based on the T/F
1678 // value.
1679 ValueLatticeElement TVal = getValueState(V: I.getTrueValue());
1680 ValueLatticeElement FVal = getValueState(V: I.getFalseValue());
1681
1682 ValueLatticeElement &State = ValueState[&I];
1683 bool Changed = State.mergeIn(RHS: TVal);
1684 Changed |= State.mergeIn(RHS: FVal);
1685 if (Changed)
1686 pushUsersToWorkListMsg(IV&: State, V: &I);
1687}
1688
1689// Handle Unary Operators.
1690void SCCPInstVisitor::visitUnaryOperator(Instruction &I) {
1691 ValueLatticeElement V0State = getValueState(V: I.getOperand(i: 0));
1692
1693 ValueLatticeElement &IV = ValueState[&I];
1694 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1695 // discover a concrete value later.
1696 if (IV.isOverdefined())
1697 return (void)markOverdefined(V: &I);
1698
1699 // If something is unknown/undef, wait for it to resolve.
1700 if (V0State.isUnknownOrUndef())
1701 return;
1702
1703 if (SCCPSolver::isConstant(LV: V0State))
1704 if (Constant *C = ConstantFoldUnaryOpOperand(
1705 Opcode: I.getOpcode(), Op: getConstant(LV: V0State, Ty: I.getType()), DL))
1706 return (void)markConstant(IV, V: &I, C);
1707
1708 markOverdefined(V: &I);
1709}
1710
1711void SCCPInstVisitor::visitFreezeInst(FreezeInst &I) {
1712 // If this freeze returns a struct, just mark the result overdefined.
1713 // TODO: We could do a lot better than this.
1714 if (I.getType()->isStructTy())
1715 return (void)markOverdefined(V: &I);
1716
1717 ValueLatticeElement V0State = getValueState(V: I.getOperand(i_nocapture: 0));
1718 ValueLatticeElement &IV = ValueState[&I];
1719 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1720 // discover a concrete value later.
1721 if (IV.isOverdefined())
1722 return (void)markOverdefined(V: &I);
1723
1724 // If something is unknown/undef, wait for it to resolve.
1725 if (V0State.isUnknownOrUndef())
1726 return;
1727
1728 if (SCCPSolver::isConstant(LV: V0State) &&
1729 isGuaranteedNotToBeUndefOrPoison(V: getConstant(LV: V0State, Ty: I.getType())))
1730 return (void)markConstant(IV, V: &I, C: getConstant(LV: V0State, Ty: I.getType()));
1731
1732 markOverdefined(V: &I);
1733}
1734
1735// Handle Binary Operators.
1736void SCCPInstVisitor::visitBinaryOperator(Instruction &I) {
1737 ValueLatticeElement V1State = getValueState(V: I.getOperand(i: 0));
1738 ValueLatticeElement V2State = getValueState(V: I.getOperand(i: 1));
1739
1740 ValueLatticeElement &IV = ValueState[&I];
1741 if (IV.isOverdefined())
1742 return;
1743
1744 // If something is undef, wait for it to resolve.
1745 if (V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef())
1746 return;
1747
1748 if (V1State.isOverdefined() && V2State.isOverdefined())
1749 return (void)markOverdefined(V: &I);
1750
1751 // If either of the operands is a constant, try to fold it to a constant.
1752 // TODO: Use information from notconstant better.
1753 if ((V1State.isConstant() || V2State.isConstant())) {
1754 Value *V1 = SCCPSolver::isConstant(LV: V1State)
1755 ? getConstant(LV: V1State, Ty: I.getOperand(i: 0)->getType())
1756 : I.getOperand(i: 0);
1757 Value *V2 = SCCPSolver::isConstant(LV: V2State)
1758 ? getConstant(LV: V2State, Ty: I.getOperand(i: 1)->getType())
1759 : I.getOperand(i: 1);
1760 Value *R = simplifyBinOp(Opcode: I.getOpcode(), LHS: V1, RHS: V2, Q: SimplifyQuery(DL, &I));
1761 auto *C = dyn_cast_or_null<Constant>(Val: R);
1762 if (C) {
1763 // Conservatively assume that the result may be based on operands that may
1764 // be undef. Note that we use mergeInValue to combine the constant with
1765 // the existing lattice value for I, as different constants might be found
1766 // after one of the operands go to overdefined, e.g. due to one operand
1767 // being a special floating value.
1768 ValueLatticeElement NewV;
1769 NewV.markConstant(V: C, /*MayIncludeUndef=*/true);
1770 return (void)mergeInValue(IV&: ValueState[&I], V: &I, MergeWithV: NewV);
1771 }
1772 }
1773
1774 // Only use ranges for binary operators on integers.
1775 if (!I.getType()->isIntOrIntVectorTy())
1776 return markOverdefined(V: &I);
1777
1778 // Try to simplify to a constant range.
1779 ConstantRange A =
1780 V1State.asConstantRange(Ty: I.getType(), /*UndefAllowed=*/false);
1781 ConstantRange B =
1782 V2State.asConstantRange(Ty: I.getType(), /*UndefAllowed=*/false);
1783
1784 auto *BO = cast<BinaryOperator>(Val: &I);
1785 ConstantRange R = ConstantRange::getEmpty(BitWidth: I.getType()->getScalarSizeInBits());
1786 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: BO))
1787 R = A.overflowingBinaryOp(BinOp: BO->getOpcode(), Other: B, NoWrapKind: OBO->getNoWrapKind());
1788 else
1789 R = A.binaryOp(BinOp: BO->getOpcode(), Other: B);
1790 mergeInValue(IV&: ValueState[&I], V: &I, MergeWithV: ValueLatticeElement::getRange(CR: R));
1791
1792 // TODO: Currently we do not exploit special values that produce something
1793 // better than overdefined with an overdefined operand for vector or floating
1794 // point types, like and <4 x i32> overdefined, zeroinitializer.
1795}
1796
1797// Handle ICmpInst instruction.
1798void SCCPInstVisitor::visitCmpInst(CmpInst &I) {
1799 // Do not cache this lookup, getValueState calls later in the function might
1800 // invalidate the reference.
1801 if (ValueState[&I].isOverdefined())
1802 return (void)markOverdefined(V: &I);
1803
1804 Value *Op1 = I.getOperand(i_nocapture: 0);
1805 Value *Op2 = I.getOperand(i_nocapture: 1);
1806
1807 // For parameters, use ParamState which includes constant range info if
1808 // available.
1809 auto V1State = getValueState(V: Op1);
1810 auto V2State = getValueState(V: Op2);
1811
1812 Constant *C = V1State.getCompare(Pred: I.getPredicate(), Ty: I.getType(), Other: V2State, DL);
1813 if (C) {
1814 ValueLatticeElement CV;
1815 CV.markConstant(V: C);
1816 mergeInValue(IV&: ValueState[&I], V: &I, MergeWithV: CV);
1817 return;
1818 }
1819
1820 // If operands are still unknown, wait for it to resolve.
1821 if ((V1State.isUnknownOrUndef() || V2State.isUnknownOrUndef()) &&
1822 !SCCPSolver::isConstant(LV: ValueState[&I]))
1823 return;
1824
1825 markOverdefined(V: &I);
1826}
1827
1828// Handle getelementptr instructions. If all operands are constants then we
1829// can turn this into a getelementptr ConstantExpr.
1830void SCCPInstVisitor::visitGetElementPtrInst(GetElementPtrInst &I) {
1831 if (ValueState[&I].isOverdefined())
1832 return (void)markOverdefined(V: &I);
1833
1834 const ValueLatticeElement &PtrState = getValueState(V: I.getPointerOperand());
1835 if (PtrState.isUnknownOrUndef())
1836 return;
1837
1838 // gep inbounds/nuw of non-null is non-null.
1839 if (PtrState.isNotConstant() && PtrState.getNotConstant()->isNullValue()) {
1840 if (I.hasNoUnsignedWrap() ||
1841 (I.isInBounds() &&
1842 !NullPointerIsDefined(F: I.getFunction(), AS: I.getAddressSpace())))
1843 return (void)markNotNull(IV&: ValueState[&I], V: &I);
1844 return (void)markOverdefined(V: &I);
1845 }
1846
1847 SmallVector<Constant *, 8> Operands;
1848 Operands.reserve(N: I.getNumOperands());
1849
1850 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1851 const ValueLatticeElement &State = getValueState(V: I.getOperand(i_nocapture: i));
1852 if (State.isUnknownOrUndef())
1853 return; // Operands are not resolved yet.
1854
1855 if (Constant *C = getConstant(LV: State, Ty: I.getOperand(i_nocapture: i)->getType())) {
1856 Operands.push_back(Elt: C);
1857 continue;
1858 }
1859
1860 return (void)markOverdefined(V: &I);
1861 }
1862
1863 if (Constant *C = ConstantFoldInstOperands(I: &I, Ops: Operands, DL))
1864 markConstant(V: &I, C);
1865 else
1866 markOverdefined(V: &I);
1867}
1868
1869void SCCPInstVisitor::visitAllocaInst(AllocaInst &I) {
1870 if (!NullPointerIsDefined(F: I.getFunction(), AS: I.getAddressSpace()))
1871 return (void)markNotNull(IV&: ValueState[&I], V: &I);
1872
1873 markOverdefined(V: &I);
1874}
1875
1876void SCCPInstVisitor::visitStoreInst(StoreInst &SI) {
1877 // If this store is of a struct, ignore it.
1878 if (SI.getOperand(i_nocapture: 0)->getType()->isStructTy())
1879 return;
1880
1881 if (TrackedGlobals.empty() || !isa<GlobalVariable>(Val: SI.getOperand(i_nocapture: 1)))
1882 return;
1883
1884 GlobalVariable *GV = cast<GlobalVariable>(Val: SI.getOperand(i_nocapture: 1));
1885 auto I = TrackedGlobals.find(Val: GV);
1886 if (I == TrackedGlobals.end())
1887 return;
1888
1889 // Get the value we are storing into the global, then merge it.
1890 mergeInValue(IV&: I->second, V: GV, MergeWithV: getValueState(V: SI.getOperand(i_nocapture: 0)),
1891 Opts: ValueLatticeElement::MergeOptions().setCheckWiden(false));
1892 if (I->second.isOverdefined())
1893 TrackedGlobals.erase(I); // No need to keep tracking this!
1894}
1895
1896static ValueLatticeElement getValueFromMetadata(const Instruction *I) {
1897 if (const auto *CB = dyn_cast<CallBase>(Val: I)) {
1898 if (CB->getType()->isIntOrIntVectorTy())
1899 if (std::optional<ConstantRange> Range = CB->getRange())
1900 return ValueLatticeElement::getRange(CR: *Range);
1901 if (CB->getType()->isPointerTy() && CB->isReturnNonNull())
1902 return ValueLatticeElement::getNot(
1903 C: ConstantPointerNull::get(T: cast<PointerType>(Val: I->getType())));
1904 }
1905
1906 if (I->getType()->isIntOrIntVectorTy())
1907 if (MDNode *Ranges = I->getMetadata(KindID: LLVMContext::MD_range))
1908 return ValueLatticeElement::getRange(
1909 CR: getConstantRangeFromMetadata(RangeMD: *Ranges));
1910 if (I->hasMetadata(KindID: LLVMContext::MD_nonnull))
1911 return ValueLatticeElement::getNot(
1912 C: ConstantPointerNull::get(T: cast<PointerType>(Val: I->getType())));
1913
1914 return ValueLatticeElement::getOverdefined();
1915}
1916
1917// Handle load instructions. If the operand is a constant pointer to a constant
1918// global, we can replace the load with the loaded constant value!
1919void SCCPInstVisitor::visitLoadInst(LoadInst &I) {
1920 // If this load is of a struct or the load is volatile, just mark the result
1921 // as overdefined.
1922 if (I.getType()->isStructTy() || I.isVolatile())
1923 return (void)markOverdefined(V: &I);
1924
1925 // resolvedUndefsIn might mark I as overdefined. Bail out, even if we would
1926 // discover a concrete value later.
1927 if (ValueState[&I].isOverdefined())
1928 return (void)markOverdefined(V: &I);
1929
1930 const ValueLatticeElement &PtrVal = getValueState(V: I.getOperand(i_nocapture: 0));
1931 if (PtrVal.isUnknownOrUndef())
1932 return; // The pointer is not resolved yet!
1933
1934 if (SCCPSolver::isConstant(LV: PtrVal)) {
1935 Constant *Ptr = getConstant(LV: PtrVal, Ty: I.getOperand(i_nocapture: 0)->getType());
1936 ValueLatticeElement &IV = ValueState[&I];
1937
1938 // load null is undefined.
1939 if (isa<ConstantPointerNull>(Val: Ptr)) {
1940 if (NullPointerIsDefined(F: I.getFunction(), AS: I.getPointerAddressSpace()))
1941 return (void)markOverdefined(IV, V: &I);
1942 else
1943 return;
1944 }
1945
1946 // Transform load (constant global) into the value loaded.
1947 if (auto *GV = dyn_cast<GlobalVariable>(Val: Ptr)) {
1948 if (!TrackedGlobals.empty()) {
1949 // If we are tracking this global, merge in the known value for it.
1950 auto It = TrackedGlobals.find(Val: GV);
1951 if (It != TrackedGlobals.end()) {
1952 mergeInValue(IV, V: &I, MergeWithV: It->second, Opts: getMaxWidenStepsOpts());
1953 return;
1954 }
1955 }
1956 }
1957
1958 // Transform load from a constant into a constant if possible.
1959 if (Constant *C = ConstantFoldLoadFromConstPtr(C: Ptr, Ty: I.getType(), DL))
1960 return (void)markConstant(IV, V: &I, C);
1961 }
1962
1963 // Fall back to metadata.
1964 mergeInValue(IV&: ValueState[&I], V: &I, MergeWithV: getValueFromMetadata(I: &I));
1965}
1966
1967void SCCPInstVisitor::visitCallBase(CallBase &CB) {
1968 handleCallResult(CB);
1969 handleCallArguments(CB);
1970}
1971
1972void SCCPInstVisitor::handleCallOverdefined(CallBase &CB) {
1973 Function *F = CB.getCalledFunction();
1974
1975 // Void return and not tracking callee, just bail.
1976 if (CB.getType()->isVoidTy())
1977 return;
1978
1979 // Always mark struct return as overdefined.
1980 if (CB.getType()->isStructTy())
1981 return (void)markOverdefined(V: &CB);
1982
1983 // Otherwise, if we have a single return value case, and if the function is
1984 // a declaration, maybe we can constant fold it.
1985 if (F && F->isDeclaration() && canConstantFoldCallTo(Call: &CB, F)) {
1986 SmallVector<Constant *, 8> Operands;
1987 for (const Use &A : CB.args()) {
1988 if (A.get()->getType()->isStructTy())
1989 return markOverdefined(V: &CB); // Can't handle struct args.
1990 if (A.get()->getType()->isMetadataTy())
1991 continue; // Carried in CB, not allowed in Operands.
1992 const ValueLatticeElement &State = getValueState(V: A);
1993
1994 if (State.isUnknownOrUndef())
1995 return; // Operands are not resolved yet.
1996 if (SCCPSolver::isOverdefined(LV: State))
1997 return (void)markOverdefined(V: &CB);
1998 assert(SCCPSolver::isConstant(State) && "Unknown state!");
1999 Operands.push_back(Elt: getConstant(LV: State, Ty: A->getType()));
2000 }
2001
2002 if (SCCPSolver::isOverdefined(LV: getValueState(V: &CB)))
2003 return (void)markOverdefined(V: &CB);
2004
2005 // If we can constant fold this, mark the result of the call as a
2006 // constant.
2007 if (Constant *C = ConstantFoldCall(Call: &CB, F, Operands, TLI: &GetTLI(*F)))
2008 return (void)markConstant(V: &CB, C);
2009 }
2010
2011 // Fall back to metadata.
2012 mergeInValue(IV&: ValueState[&CB], V: &CB, MergeWithV: getValueFromMetadata(I: &CB));
2013}
2014
2015void SCCPInstVisitor::handleCallArguments(CallBase &CB) {
2016 Function *F = CB.getCalledFunction();
2017 // If this is a local function that doesn't have its address taken, mark its
2018 // entry block executable and merge in the actual arguments to the call into
2019 // the formal arguments of the function.
2020 if (TrackingIncomingArguments.count(Ptr: F)) {
2021 markBlockExecutable(BB: &F->front());
2022
2023 // Propagate information from this call site into the callee.
2024 auto CAI = CB.arg_begin();
2025 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2026 ++AI, ++CAI) {
2027 // If this argument is byval, and if the function is not readonly, there
2028 // will be an implicit copy formed of the input aggregate.
2029 if (AI->hasByValAttr() && !F->onlyReadsMemory()) {
2030 markOverdefined(V: &*AI);
2031 continue;
2032 }
2033
2034 if (auto *STy = dyn_cast<StructType>(Val: AI->getType())) {
2035 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2036 ValueLatticeElement CallArg = getStructValueState(V: *CAI, i);
2037 mergeInValue(IV&: getStructValueState(V: &*AI, i), V: &*AI, MergeWithV: CallArg,
2038 Opts: getMaxWidenStepsOpts());
2039 }
2040 } else {
2041 ValueLatticeElement CallArg =
2042 getValueState(V: *CAI).intersect(Other: getArgAttributeVL(A: &*AI));
2043 mergeInValue(IV&: ValueState[&*AI], V: &*AI, MergeWithV: CallArg, Opts: getMaxWidenStepsOpts());
2044 }
2045 }
2046 }
2047}
2048
2049void SCCPInstVisitor::handlePredicate(Instruction *I, Value *CopyOf,
2050 const PredicateBase *PI) {
2051 ValueLatticeElement CopyOfVal = getValueState(V: CopyOf);
2052 const std::optional<PredicateConstraint> &Constraint = PI->getConstraint();
2053 if (!Constraint) {
2054 mergeInValue(IV&: ValueState[I], V: I, MergeWithV: CopyOfVal);
2055 return;
2056 }
2057
2058 CmpInst::Predicate Pred = Constraint->Predicate;
2059 Value *OtherOp = Constraint->OtherOp;
2060
2061 // Wait until OtherOp is resolved.
2062 if (getValueState(V: OtherOp).isUnknown()) {
2063 addAdditionalUser(V: OtherOp, U: I);
2064 return;
2065 }
2066
2067 ValueLatticeElement CondVal = getValueState(V: OtherOp);
2068 ValueLatticeElement &IV = ValueState[I];
2069 if (CondVal.isConstantRange() || CopyOfVal.isConstantRange()) {
2070 auto ImposedCR =
2071 ConstantRange::getFull(BitWidth: DL.getTypeSizeInBits(Ty: CopyOf->getType()));
2072
2073 // Get the range imposed by the condition.
2074 if (CondVal.isConstantRange())
2075 ImposedCR = ConstantRange::makeAllowedICmpRegion(
2076 Pred, Other: CondVal.getConstantRange());
2077
2078 // Combine range info for the original value with the new range from the
2079 // condition.
2080 auto CopyOfCR = CopyOfVal.asConstantRange(Ty: CopyOf->getType(),
2081 /*UndefAllowed=*/true);
2082 // Treat an unresolved input like a full range.
2083 if (CopyOfCR.isEmptySet())
2084 CopyOfCR = ConstantRange::getFull(BitWidth: CopyOfCR.getBitWidth());
2085 auto NewCR = ImposedCR.intersectWith(CR: CopyOfCR);
2086 // If the existing information is != x, do not use the information from
2087 // a chained predicate, as the != x information is more likely to be
2088 // helpful in practice.
2089 if (!CopyOfCR.contains(CR: NewCR) && CopyOfCR.getSingleMissingElement())
2090 NewCR = std::move(CopyOfCR);
2091
2092 // The new range is based on a branch condition. That guarantees that
2093 // neither of the compare operands can be undef in the branch targets,
2094 // unless we have conditions that are always true/false (e.g. icmp ule
2095 // i32, %a, i32_max). For the latter overdefined/empty range will be
2096 // inferred, but the branch will get folded accordingly anyways.
2097 addAdditionalUser(V: OtherOp, U: I);
2098 mergeInValue(
2099 IV, V: I, MergeWithV: ValueLatticeElement::getRange(CR: NewCR, /*MayIncludeUndef*/ false));
2100 return;
2101 } else if (Pred == CmpInst::ICMP_EQ &&
2102 (CondVal.isConstant() || CondVal.isNotConstant())) {
2103 // For non-integer values or integer constant expressions, only
2104 // propagate equal constants or not-constants.
2105 addAdditionalUser(V: OtherOp, U: I);
2106 mergeInValue(IV, V: I, MergeWithV: CondVal);
2107 return;
2108 } else if (Pred == CmpInst::ICMP_NE && CondVal.isConstant()) {
2109 // Propagate inequalities.
2110 addAdditionalUser(V: OtherOp, U: I);
2111 mergeInValue(IV, V: I, MergeWithV: ValueLatticeElement::getNot(C: CondVal.getConstant()));
2112 return;
2113 }
2114
2115 return (void)mergeInValue(IV, V: I, MergeWithV: CopyOfVal);
2116}
2117
2118void SCCPInstVisitor::handleCallResult(CallBase &CB) {
2119 Function *F = CB.getCalledFunction();
2120
2121 if (auto *II = dyn_cast<IntrinsicInst>(Val: &CB)) {
2122 if (II->getIntrinsicID() == Intrinsic::vscale) {
2123 unsigned BitWidth = CB.getType()->getScalarSizeInBits();
2124 const ConstantRange Result = getVScaleRange(F: II->getFunction(), BitWidth);
2125 return (void)mergeInValue(IV&: ValueState[II], V: II,
2126 MergeWithV: ValueLatticeElement::getRange(CR: Result));
2127 }
2128 if (II->getIntrinsicID() == Intrinsic::experimental_get_vector_length) {
2129 Value *CountArg = II->getArgOperand(i: 0);
2130 Value *VF = II->getArgOperand(i: 1);
2131 bool Scalable = cast<ConstantInt>(Val: II->getArgOperand(i: 2))->isOne();
2132
2133 // Computation happens in the larger type.
2134 unsigned BitWidth = std::max(a: CountArg->getType()->getScalarSizeInBits(),
2135 b: VF->getType()->getScalarSizeInBits());
2136
2137 ConstantRange Count = getValueState(V: CountArg)
2138 .asConstantRange(Ty: CountArg->getType(), UndefAllowed: false)
2139 .zeroExtend(BitWidth);
2140 ConstantRange MaxLanes = getValueState(V: VF)
2141 .asConstantRange(Ty: VF->getType(), UndefAllowed: false)
2142 .zeroExtend(BitWidth);
2143 if (Scalable)
2144 MaxLanes =
2145 MaxLanes.multiply(Other: getVScaleRange(F: II->getFunction(), BitWidth));
2146
2147 // The result is always less than both Count and MaxLanes.
2148 ConstantRange Result = ConstantRange::getNonEmpty(
2149 Lower: APInt::getZero(numBits: BitWidth),
2150 Upper: APIntOps::umin(A: Count.getUnsignedMax(), B: MaxLanes.getUnsignedMax()) +
2151 1);
2152
2153 // If Count <= MaxLanes, getvectorlength(Count, MaxLanes) = Count
2154 if (Count.icmp(Pred: CmpInst::ICMP_ULE, Other: MaxLanes))
2155 Result = std::move(Count);
2156
2157 Result = Result.truncate(BitWidth: II->getType()->getScalarSizeInBits());
2158 return (void)mergeInValue(IV&: ValueState[II], V: II,
2159 MergeWithV: ValueLatticeElement::getRange(CR: Result));
2160 }
2161
2162 if (ConstantRange::isIntrinsicSupported(IntrinsicID: II->getIntrinsicID())) {
2163 // Compute result range for intrinsics supported by ConstantRange.
2164 // Do this even if we don't know a range for all operands, as we may
2165 // still know something about the result range, e.g. of abs(x).
2166 SmallVector<ConstantRange, 2> OpRanges;
2167 for (Value *Op : II->args()) {
2168 const ValueLatticeElement &State = getValueState(V: Op);
2169 if (State.isUnknownOrUndef())
2170 return;
2171 OpRanges.push_back(
2172 Elt: State.asConstantRange(Ty: Op->getType(), /*UndefAllowed=*/false));
2173 }
2174
2175 ConstantRange Result =
2176 ConstantRange::intrinsic(IntrinsicID: II->getIntrinsicID(), Ops: OpRanges);
2177 return (void)mergeInValue(IV&: ValueState[II], V: II,
2178 MergeWithV: ValueLatticeElement::getRange(CR: Result));
2179 }
2180 }
2181
2182 // The common case is that we aren't tracking the callee, either because we
2183 // are not doing interprocedural analysis or the callee is indirect, or is
2184 // external. Handle these cases first.
2185 if (!F || F->isDeclaration())
2186 return handleCallOverdefined(CB);
2187
2188 // If this is a single/zero retval case, see if we're tracking the function.
2189 if (auto *STy = dyn_cast<StructType>(Val: F->getReturnType())) {
2190 if (!MRVFunctionsTracked.count(Ptr: F))
2191 return handleCallOverdefined(CB); // Not tracking this callee.
2192
2193 // If we are tracking this callee, propagate the result of the function
2194 // into this call site.
2195 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2196 mergeInValue(IV&: getStructValueState(V: &CB, i), V: &CB,
2197 MergeWithV: TrackedMultipleRetVals[std::make_pair(x&: F, y&: i)],
2198 Opts: getMaxWidenStepsOpts());
2199 } else {
2200 auto TFRVI = TrackedRetVals.find(Key: F);
2201 if (TFRVI == TrackedRetVals.end())
2202 return handleCallOverdefined(CB); // Not tracking this callee.
2203
2204 // If so, propagate the return value of the callee into this call result.
2205 mergeInValue(IV&: ValueState[&CB], V: &CB, MergeWithV: TFRVI->second, Opts: getMaxWidenStepsOpts());
2206 }
2207}
2208
2209bool SCCPInstVisitor::isInstFullyOverDefined(Instruction &Inst) {
2210 // For structure Type, we handle each member separately.
2211 // A structure object won't be considered as overdefined when
2212 // there is at least one member that is not overdefined.
2213 if (StructType *STy = dyn_cast<StructType>(Val: Inst.getType())) {
2214 for (unsigned i = 0, e = STy->getNumElements(); i < e; ++i) {
2215 if (!getStructValueState(V: &Inst, i).isOverdefined())
2216 return false;
2217 }
2218 return true;
2219 }
2220
2221 return getValueState(V: &Inst).isOverdefined();
2222}
2223
2224void SCCPInstVisitor::solve() {
2225 // Process the work lists until they are empty!
2226 while (!BBWorkList.empty() || !InstWorkList.empty()) {
2227 // Process the instruction work list.
2228 while (!InstWorkList.empty()) {
2229 Instruction *I = InstWorkList.pop_back_val();
2230 Invalidated.erase(V: I);
2231
2232 LLVM_DEBUG(dbgs() << "\nPopped off I-WL: " << *I << '\n');
2233
2234 visit(I);
2235 }
2236
2237 // Process the basic block work list.
2238 while (!BBWorkList.empty()) {
2239 BasicBlock *BB = BBWorkList.pop_back_val();
2240 BBVisited.insert(Ptr: BB);
2241
2242 LLVM_DEBUG(dbgs() << "\nPopped off BBWL: " << *BB << '\n');
2243 for (Instruction &I : *BB) {
2244 CurI = &I;
2245 visit(I);
2246 }
2247 CurI = nullptr;
2248 }
2249 }
2250}
2251
2252bool SCCPInstVisitor::resolvedUndef(Instruction &I) {
2253 // Look for instructions which produce undef values.
2254 if (I.getType()->isVoidTy())
2255 return false;
2256
2257 if (auto *STy = dyn_cast<StructType>(Val: I.getType())) {
2258 // Only a few things that can be structs matter for undef.
2259
2260 // Tracked calls must never be marked overdefined in resolvedUndefsIn.
2261 if (auto *CB = dyn_cast<CallBase>(Val: &I))
2262 if (Function *F = CB->getCalledFunction())
2263 if (MRVFunctionsTracked.count(Ptr: F))
2264 return false;
2265
2266 // extractvalue and insertvalue don't need to be marked; they are
2267 // tracked as precisely as their operands.
2268 if (isa<ExtractValueInst>(Val: I) || isa<InsertValueInst>(Val: I))
2269 return false;
2270 // Send the results of everything else to overdefined. We could be
2271 // more precise than this but it isn't worth bothering.
2272 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2273 ValueLatticeElement &LV = getStructValueState(V: &I, i);
2274 if (LV.isUnknown()) {
2275 markOverdefined(IV&: LV, V: &I);
2276 return true;
2277 }
2278 }
2279 return false;
2280 }
2281
2282 ValueLatticeElement &LV = getValueState(V: &I);
2283 if (!LV.isUnknown())
2284 return false;
2285
2286 // There are two reasons a call can have an undef result
2287 // 1. It could be tracked.
2288 // 2. It could be constant-foldable.
2289 // Because of the way we solve return values, tracked calls must
2290 // never be marked overdefined in resolvedUndefsIn.
2291 if (auto *CB = dyn_cast<CallBase>(Val: &I))
2292 if (Function *F = CB->getCalledFunction())
2293 if (TrackedRetVals.count(Key: F))
2294 return false;
2295
2296 if (isa<LoadInst>(Val: I)) {
2297 // A load here means one of two things: a load of undef from a global,
2298 // a load from an unknown pointer. Either way, having it return undef
2299 // is okay.
2300 return false;
2301 }
2302
2303 markOverdefined(V: &I);
2304 return true;
2305}
2306
2307/// While solving the dataflow for a function, we don't compute a result for
2308/// operations with an undef operand, to allow undef to be lowered to a
2309/// constant later. For example, constant folding of "zext i8 undef to i16"
2310/// would result in "i16 0", and if undef is later lowered to "i8 1", then the
2311/// zext result would become "i16 1" and would result into an overdefined
2312/// lattice value once merged with the previous result. Not computing the
2313/// result of the zext (treating undef the same as unknown) allows us to handle
2314/// a later undef->constant lowering more optimally.
2315///
2316/// However, if the operand remains undef when the solver returns, we do need
2317/// to assign some result to the instruction (otherwise we would treat it as
2318/// unreachable). For simplicity, we mark any instructions that are still
2319/// unknown as overdefined.
2320bool SCCPInstVisitor::resolvedUndefsIn(Function &F) {
2321 bool MadeChange = false;
2322 for (BasicBlock &BB : F) {
2323 if (!BBExecutable.count(Ptr: &BB))
2324 continue;
2325
2326 for (Instruction &I : BB)
2327 MadeChange |= resolvedUndef(I);
2328 }
2329
2330 LLVM_DEBUG(if (MadeChange) dbgs()
2331 << "\nResolved undefs in " << F.getName() << '\n');
2332
2333 return MadeChange;
2334}
2335
2336//===----------------------------------------------------------------------===//
2337//
2338// SCCPSolver implementations
2339//
2340SCCPSolver::SCCPSolver(
2341 const DataLayout &DL,
2342 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
2343 LLVMContext &Ctx)
2344 : Visitor(new SCCPInstVisitor(DL, std::move(GetTLI), Ctx)) {}
2345
2346SCCPSolver::~SCCPSolver() = default;
2347
2348void SCCPSolver::addPredicateInfo(Function &F, DominatorTree &DT,
2349 AssumptionCache &AC) {
2350 Visitor->addPredicateInfo(F, DT, AC);
2351}
2352
2353void SCCPSolver::removeSSACopies(Function &F) {
2354 Visitor->removeSSACopies(F);
2355}
2356
2357bool SCCPSolver::markBlockExecutable(BasicBlock *BB) {
2358 return Visitor->markBlockExecutable(BB);
2359}
2360
2361const PredicateBase *SCCPSolver::getPredicateInfoFor(Instruction *I) {
2362 return Visitor->getPredicateInfoFor(I);
2363}
2364
2365void SCCPSolver::trackValueOfGlobalVariable(GlobalVariable *GV) {
2366 Visitor->trackValueOfGlobalVariable(GV);
2367}
2368
2369void SCCPSolver::addTrackedFunction(Function *F) {
2370 Visitor->addTrackedFunction(F);
2371}
2372
2373void SCCPSolver::addToMustPreserveReturnsInFunctions(Function *F) {
2374 Visitor->addToMustPreserveReturnsInFunctions(F);
2375}
2376
2377bool SCCPSolver::mustPreserveReturn(Function *F) {
2378 return Visitor->mustPreserveReturn(F);
2379}
2380
2381void SCCPSolver::addArgumentTrackedFunction(Function *F) {
2382 Visitor->addArgumentTrackedFunction(F);
2383}
2384
2385bool SCCPSolver::isArgumentTrackedFunction(Function *F) {
2386 return Visitor->isArgumentTrackedFunction(F);
2387}
2388
2389const SmallPtrSetImpl<Function *> &
2390SCCPSolver::getArgumentTrackedFunctions() const {
2391 return Visitor->getArgumentTrackedFunctions();
2392}
2393
2394void SCCPSolver::solve() { Visitor->solve(); }
2395
2396bool SCCPSolver::resolvedUndefsIn(Function &F) {
2397 return Visitor->resolvedUndefsIn(F);
2398}
2399
2400void SCCPSolver::solveWhileResolvedUndefsIn(Module &M) {
2401 Visitor->solveWhileResolvedUndefsIn(M);
2402}
2403
2404void
2405SCCPSolver::solveWhileResolvedUndefsIn(SmallVectorImpl<Function *> &WorkList) {
2406 Visitor->solveWhileResolvedUndefsIn(WorkList);
2407}
2408
2409void SCCPSolver::solveWhileResolvedUndefs() {
2410 Visitor->solveWhileResolvedUndefs();
2411}
2412
2413bool SCCPSolver::isBlockExecutable(BasicBlock *BB) const {
2414 return Visitor->isBlockExecutable(BB);
2415}
2416
2417bool SCCPSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To) const {
2418 return Visitor->isEdgeFeasible(From, To);
2419}
2420
2421std::vector<ValueLatticeElement>
2422SCCPSolver::getStructLatticeValueFor(Value *V) const {
2423 return Visitor->getStructLatticeValueFor(V);
2424}
2425
2426void SCCPSolver::removeLatticeValueFor(Value *V) {
2427 return Visitor->removeLatticeValueFor(V);
2428}
2429
2430void SCCPSolver::resetLatticeValueFor(CallBase *Call) {
2431 Visitor->resetLatticeValueFor(Call);
2432}
2433
2434const ValueLatticeElement &SCCPSolver::getLatticeValueFor(Value *V) const {
2435 return Visitor->getLatticeValueFor(V);
2436}
2437
2438const MapVector<Function *, ValueLatticeElement> &
2439SCCPSolver::getTrackedRetVals() const {
2440 return Visitor->getTrackedRetVals();
2441}
2442
2443const DenseMap<GlobalVariable *, ValueLatticeElement> &
2444SCCPSolver::getTrackedGlobals() const {
2445 return Visitor->getTrackedGlobals();
2446}
2447
2448const SmallPtrSet<Function *, 16> &SCCPSolver::getMRVFunctionsTracked() const {
2449 return Visitor->getMRVFunctionsTracked();
2450}
2451
2452void SCCPSolver::markOverdefined(Value *V) { Visitor->markOverdefined(V); }
2453
2454void SCCPSolver::trackValueOfArgument(Argument *V) {
2455 Visitor->trackValueOfArgument(A: V);
2456}
2457
2458bool SCCPSolver::isStructLatticeConstant(Function *F, StructType *STy) {
2459 return Visitor->isStructLatticeConstant(F, STy);
2460}
2461
2462Constant *SCCPSolver::getConstant(const ValueLatticeElement &LV,
2463 Type *Ty) const {
2464 return Visitor->getConstant(LV, Ty);
2465}
2466
2467Constant *SCCPSolver::getConstantOrNull(Value *V) const {
2468 return Visitor->getConstantOrNull(V);
2469}
2470
2471void SCCPSolver::setLatticeValueForSpecializationArguments(Function *F,
2472 const SmallVectorImpl<ArgInfo> &Args) {
2473 Visitor->setLatticeValueForSpecializationArguments(F, Args);
2474}
2475
2476void SCCPSolver::markFunctionUnreachable(Function *F) {
2477 Visitor->markFunctionUnreachable(F);
2478}
2479
2480void SCCPSolver::visit(Instruction *I) { Visitor->visit(I); }
2481
2482void SCCPSolver::visitCall(CallInst &I) { Visitor->visitCall(I); }
2483