1 //===-- ConstraintElimination.cpp - Eliminate conds using constraints. ----===//
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 // Eliminate conditions based on constraints collected from dominating
10 // conditions.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Transforms/Scalar/ConstraintElimination.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/ScopeExit.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/ConstraintSystem.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
23 #include "llvm/Analysis/ScalarEvolution.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstrTypes.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/Verifier.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/DebugCounter.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Transforms/Utils/Cloning.h"
41 #include "llvm/Transforms/Utils/ValueMapper.h"
42
43 #include <cmath>
44 #include <optional>
45 #include <string>
46
47 using namespace llvm;
48 using namespace PatternMatch;
49
50 #define DEBUG_TYPE "constraint-elimination"
51
52 STATISTIC(NumCondsRemoved, "Number of instructions removed");
53 DEBUG_COUNTER(EliminatedCounter, "conds-eliminated",
54 "Controls which conditions are eliminated");
55
56 static cl::opt<unsigned>
57 MaxRows("constraint-elimination-max-rows", cl::init(500), cl::Hidden,
58 cl::desc("Maximum number of rows to keep in constraint system"));
59
60 static cl::opt<bool> DumpReproducers(
61 "constraint-elimination-dump-reproducers", cl::init(false), cl::Hidden,
62 cl::desc("Dump IR to reproduce successful transformations."));
63
64 static int64_t MaxConstraintValue = std::numeric_limits<int64_t>::max();
65 static int64_t MinSignedConstraintValue = std::numeric_limits<int64_t>::min();
66
67 // A helper to multiply 2 signed integers where overflowing is allowed.
multiplyWithOverflow(int64_t A,int64_t B)68 static int64_t multiplyWithOverflow(int64_t A, int64_t B) {
69 int64_t Result;
70 MulOverflow(A, B, Result);
71 return Result;
72 }
73
74 // A helper to add 2 signed integers where overflowing is allowed.
addWithOverflow(int64_t A,int64_t B)75 static int64_t addWithOverflow(int64_t A, int64_t B) {
76 int64_t Result;
77 AddOverflow(A, B, Result);
78 return Result;
79 }
80
getContextInstForUse(Use & U)81 static Instruction *getContextInstForUse(Use &U) {
82 Instruction *UserI = cast<Instruction>(U.getUser());
83 if (auto *Phi = dyn_cast<PHINode>(UserI))
84 UserI = Phi->getIncomingBlock(U)->getTerminator();
85 return UserI;
86 }
87
88 namespace {
89 /// Struct to express a condition of the form %Op0 Pred %Op1.
90 struct ConditionTy {
91 CmpInst::Predicate Pred;
92 Value *Op0;
93 Value *Op1;
94
ConditionTy__anon050fee910111::ConditionTy95 ConditionTy()
96 : Pred(CmpInst::BAD_ICMP_PREDICATE), Op0(nullptr), Op1(nullptr) {}
ConditionTy__anon050fee910111::ConditionTy97 ConditionTy(CmpInst::Predicate Pred, Value *Op0, Value *Op1)
98 : Pred(Pred), Op0(Op0), Op1(Op1) {}
99 };
100
101 /// Represents either
102 /// * a condition that holds on entry to a block (=condition fact)
103 /// * an assume (=assume fact)
104 /// * a use of a compare instruction to simplify.
105 /// It also tracks the Dominator DFS in and out numbers for each entry.
106 struct FactOrCheck {
107 enum class EntryTy {
108 ConditionFact, /// A condition that holds on entry to a block.
109 InstFact, /// A fact that holds after Inst executed (e.g. an assume or
110 /// min/mix intrinsic.
111 InstCheck, /// An instruction to simplify (e.g. an overflow math
112 /// intrinsics).
113 UseCheck /// An use of a compare instruction to simplify.
114 };
115
116 union {
117 Instruction *Inst;
118 Use *U;
119 ConditionTy Cond;
120 };
121
122 /// A pre-condition that must hold for the current fact to be added to the
123 /// system.
124 ConditionTy DoesHold;
125
126 unsigned NumIn;
127 unsigned NumOut;
128 EntryTy Ty;
129
FactOrCheck__anon050fee910111::FactOrCheck130 FactOrCheck(EntryTy Ty, DomTreeNode *DTN, Instruction *Inst)
131 : Inst(Inst), NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()),
132 Ty(Ty) {}
133
FactOrCheck__anon050fee910111::FactOrCheck134 FactOrCheck(DomTreeNode *DTN, Use *U)
135 : U(U), DoesHold(CmpInst::BAD_ICMP_PREDICATE, nullptr, nullptr),
136 NumIn(DTN->getDFSNumIn()), NumOut(DTN->getDFSNumOut()),
137 Ty(EntryTy::UseCheck) {}
138
FactOrCheck__anon050fee910111::FactOrCheck139 FactOrCheck(DomTreeNode *DTN, CmpInst::Predicate Pred, Value *Op0, Value *Op1,
140 ConditionTy Precond = ConditionTy())
141 : Cond(Pred, Op0, Op1), DoesHold(Precond), NumIn(DTN->getDFSNumIn()),
142 NumOut(DTN->getDFSNumOut()), Ty(EntryTy::ConditionFact) {}
143
getConditionFact__anon050fee910111::FactOrCheck144 static FactOrCheck getConditionFact(DomTreeNode *DTN, CmpInst::Predicate Pred,
145 Value *Op0, Value *Op1,
146 ConditionTy Precond = ConditionTy()) {
147 return FactOrCheck(DTN, Pred, Op0, Op1, Precond);
148 }
149
getInstFact__anon050fee910111::FactOrCheck150 static FactOrCheck getInstFact(DomTreeNode *DTN, Instruction *Inst) {
151 return FactOrCheck(EntryTy::InstFact, DTN, Inst);
152 }
153
getCheck__anon050fee910111::FactOrCheck154 static FactOrCheck getCheck(DomTreeNode *DTN, Use *U) {
155 return FactOrCheck(DTN, U);
156 }
157
getCheck__anon050fee910111::FactOrCheck158 static FactOrCheck getCheck(DomTreeNode *DTN, CallInst *CI) {
159 return FactOrCheck(EntryTy::InstCheck, DTN, CI);
160 }
161
isCheck__anon050fee910111::FactOrCheck162 bool isCheck() const {
163 return Ty == EntryTy::InstCheck || Ty == EntryTy::UseCheck;
164 }
165
getContextInst__anon050fee910111::FactOrCheck166 Instruction *getContextInst() const {
167 if (Ty == EntryTy::UseCheck)
168 return getContextInstForUse(*U);
169 return Inst;
170 }
171
getInstructionToSimplify__anon050fee910111::FactOrCheck172 Instruction *getInstructionToSimplify() const {
173 assert(isCheck());
174 if (Ty == EntryTy::InstCheck)
175 return Inst;
176 // The use may have been simplified to a constant already.
177 return dyn_cast<Instruction>(*U);
178 }
179
isConditionFact__anon050fee910111::FactOrCheck180 bool isConditionFact() const { return Ty == EntryTy::ConditionFact; }
181 };
182
183 /// Keep state required to build worklist.
184 struct State {
185 DominatorTree &DT;
186 LoopInfo &LI;
187 ScalarEvolution &SE;
188 SmallVector<FactOrCheck, 64> WorkList;
189
State__anon050fee910111::State190 State(DominatorTree &DT, LoopInfo &LI, ScalarEvolution &SE)
191 : DT(DT), LI(LI), SE(SE) {}
192
193 /// Process block \p BB and add known facts to work-list.
194 void addInfoFor(BasicBlock &BB);
195
196 /// Try to add facts for loop inductions (AddRecs) in EQ/NE compares
197 /// controlling the loop header.
198 void addInfoForInductions(BasicBlock &BB);
199
200 /// Returns true if we can add a known condition from BB to its successor
201 /// block Succ.
canAddSuccessor__anon050fee910111::State202 bool canAddSuccessor(BasicBlock &BB, BasicBlock *Succ) const {
203 return DT.dominates(BasicBlockEdge(&BB, Succ), Succ);
204 }
205 };
206
207 class ConstraintInfo;
208
209 struct StackEntry {
210 unsigned NumIn;
211 unsigned NumOut;
212 bool IsSigned = false;
213 /// Variables that can be removed from the system once the stack entry gets
214 /// removed.
215 SmallVector<Value *, 2> ValuesToRelease;
216
StackEntry__anon050fee910111::StackEntry217 StackEntry(unsigned NumIn, unsigned NumOut, bool IsSigned,
218 SmallVector<Value *, 2> ValuesToRelease)
219 : NumIn(NumIn), NumOut(NumOut), IsSigned(IsSigned),
220 ValuesToRelease(ValuesToRelease) {}
221 };
222
223 struct ConstraintTy {
224 SmallVector<int64_t, 8> Coefficients;
225 SmallVector<ConditionTy, 2> Preconditions;
226
227 SmallVector<SmallVector<int64_t, 8>> ExtraInfo;
228
229 bool IsSigned = false;
230
231 ConstraintTy() = default;
232
ConstraintTy__anon050fee910111::ConstraintTy233 ConstraintTy(SmallVector<int64_t, 8> Coefficients, bool IsSigned, bool IsEq,
234 bool IsNe)
235 : Coefficients(std::move(Coefficients)), IsSigned(IsSigned), IsEq(IsEq),
236 IsNe(IsNe) {}
237
size__anon050fee910111::ConstraintTy238 unsigned size() const { return Coefficients.size(); }
239
empty__anon050fee910111::ConstraintTy240 unsigned empty() const { return Coefficients.empty(); }
241
242 /// Returns true if all preconditions for this list of constraints are
243 /// satisfied given \p CS and the corresponding \p Value2Index mapping.
244 bool isValid(const ConstraintInfo &Info) const;
245
isEq__anon050fee910111::ConstraintTy246 bool isEq() const { return IsEq; }
247
isNe__anon050fee910111::ConstraintTy248 bool isNe() const { return IsNe; }
249
250 /// Check if the current constraint is implied by the given ConstraintSystem.
251 ///
252 /// \return true or false if the constraint is proven to be respectively true,
253 /// or false. When the constraint cannot be proven to be either true or false,
254 /// std::nullopt is returned.
255 std::optional<bool> isImpliedBy(const ConstraintSystem &CS) const;
256
257 private:
258 bool IsEq = false;
259 bool IsNe = false;
260 };
261
262 /// Wrapper encapsulating separate constraint systems and corresponding value
263 /// mappings for both unsigned and signed information. Facts are added to and
264 /// conditions are checked against the corresponding system depending on the
265 /// signed-ness of their predicates. While the information is kept separate
266 /// based on signed-ness, certain conditions can be transferred between the two
267 /// systems.
268 class ConstraintInfo {
269
270 ConstraintSystem UnsignedCS;
271 ConstraintSystem SignedCS;
272
273 const DataLayout &DL;
274
275 public:
ConstraintInfo(const DataLayout & DL,ArrayRef<Value * > FunctionArgs)276 ConstraintInfo(const DataLayout &DL, ArrayRef<Value *> FunctionArgs)
277 : UnsignedCS(FunctionArgs), SignedCS(FunctionArgs), DL(DL) {
278 auto &Value2Index = getValue2Index(false);
279 // Add Arg > -1 constraints to unsigned system for all function arguments.
280 for (Value *Arg : FunctionArgs) {
281 ConstraintTy VarPos(SmallVector<int64_t, 8>(Value2Index.size() + 1, 0),
282 false, false, false);
283 VarPos.Coefficients[Value2Index[Arg]] = -1;
284 UnsignedCS.addVariableRow(VarPos.Coefficients);
285 }
286 }
287
getValue2Index(bool Signed)288 DenseMap<Value *, unsigned> &getValue2Index(bool Signed) {
289 return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index();
290 }
getValue2Index(bool Signed) const291 const DenseMap<Value *, unsigned> &getValue2Index(bool Signed) const {
292 return Signed ? SignedCS.getValue2Index() : UnsignedCS.getValue2Index();
293 }
294
getCS(bool Signed)295 ConstraintSystem &getCS(bool Signed) {
296 return Signed ? SignedCS : UnsignedCS;
297 }
getCS(bool Signed) const298 const ConstraintSystem &getCS(bool Signed) const {
299 return Signed ? SignedCS : UnsignedCS;
300 }
301
popLastConstraint(bool Signed)302 void popLastConstraint(bool Signed) { getCS(Signed).popLastConstraint(); }
popLastNVariables(bool Signed,unsigned N)303 void popLastNVariables(bool Signed, unsigned N) {
304 getCS(Signed).popLastNVariables(N);
305 }
306
307 bool doesHold(CmpInst::Predicate Pred, Value *A, Value *B) const;
308
309 void addFact(CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn,
310 unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack);
311
312 /// Turn a comparison of the form \p Op0 \p Pred \p Op1 into a vector of
313 /// constraints, using indices from the corresponding constraint system.
314 /// New variables that need to be added to the system are collected in
315 /// \p NewVariables.
316 ConstraintTy getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1,
317 SmallVectorImpl<Value *> &NewVariables) const;
318
319 /// Turns a comparison of the form \p Op0 \p Pred \p Op1 into a vector of
320 /// constraints using getConstraint. Returns an empty constraint if the result
321 /// cannot be used to query the existing constraint system, e.g. because it
322 /// would require adding new variables. Also tries to convert signed
323 /// predicates to unsigned ones if possible to allow using the unsigned system
324 /// which increases the effectiveness of the signed <-> unsigned transfer
325 /// logic.
326 ConstraintTy getConstraintForSolving(CmpInst::Predicate Pred, Value *Op0,
327 Value *Op1) const;
328
329 /// Try to add information from \p A \p Pred \p B to the unsigned/signed
330 /// system if \p Pred is signed/unsigned.
331 void transferToOtherSystem(CmpInst::Predicate Pred, Value *A, Value *B,
332 unsigned NumIn, unsigned NumOut,
333 SmallVectorImpl<StackEntry> &DFSInStack);
334 };
335
336 /// Represents a (Coefficient * Variable) entry after IR decomposition.
337 struct DecompEntry {
338 int64_t Coefficient;
339 Value *Variable;
340 /// True if the variable is known positive in the current constraint.
341 bool IsKnownNonNegative;
342
DecompEntry__anon050fee910111::DecompEntry343 DecompEntry(int64_t Coefficient, Value *Variable,
344 bool IsKnownNonNegative = false)
345 : Coefficient(Coefficient), Variable(Variable),
346 IsKnownNonNegative(IsKnownNonNegative) {}
347 };
348
349 /// Represents an Offset + Coefficient1 * Variable1 + ... decomposition.
350 struct Decomposition {
351 int64_t Offset = 0;
352 SmallVector<DecompEntry, 3> Vars;
353
Decomposition__anon050fee910111::Decomposition354 Decomposition(int64_t Offset) : Offset(Offset) {}
Decomposition__anon050fee910111::Decomposition355 Decomposition(Value *V, bool IsKnownNonNegative = false) {
356 Vars.emplace_back(1, V, IsKnownNonNegative);
357 }
Decomposition__anon050fee910111::Decomposition358 Decomposition(int64_t Offset, ArrayRef<DecompEntry> Vars)
359 : Offset(Offset), Vars(Vars) {}
360
add__anon050fee910111::Decomposition361 void add(int64_t OtherOffset) {
362 Offset = addWithOverflow(Offset, OtherOffset);
363 }
364
add__anon050fee910111::Decomposition365 void add(const Decomposition &Other) {
366 add(Other.Offset);
367 append_range(Vars, Other.Vars);
368 }
369
sub__anon050fee910111::Decomposition370 void sub(const Decomposition &Other) {
371 Decomposition Tmp = Other;
372 Tmp.mul(-1);
373 add(Tmp.Offset);
374 append_range(Vars, Tmp.Vars);
375 }
376
mul__anon050fee910111::Decomposition377 void mul(int64_t Factor) {
378 Offset = multiplyWithOverflow(Offset, Factor);
379 for (auto &Var : Vars)
380 Var.Coefficient = multiplyWithOverflow(Var.Coefficient, Factor);
381 }
382 };
383
384 // Variable and constant offsets for a chain of GEPs, with base pointer BasePtr.
385 struct OffsetResult {
386 Value *BasePtr;
387 APInt ConstantOffset;
388 MapVector<Value *, APInt> VariableOffsets;
389 bool AllInbounds;
390
OffsetResult__anon050fee910111::OffsetResult391 OffsetResult() : BasePtr(nullptr), ConstantOffset(0, uint64_t(0)) {}
392
OffsetResult__anon050fee910111::OffsetResult393 OffsetResult(GEPOperator &GEP, const DataLayout &DL)
394 : BasePtr(GEP.getPointerOperand()), AllInbounds(GEP.isInBounds()) {
395 ConstantOffset = APInt(DL.getIndexTypeSizeInBits(BasePtr->getType()), 0);
396 }
397 };
398 } // namespace
399
400 // Try to collect variable and constant offsets for \p GEP, partly traversing
401 // nested GEPs. Returns an OffsetResult with nullptr as BasePtr of collecting
402 // the offset fails.
collectOffsets(GEPOperator & GEP,const DataLayout & DL)403 static OffsetResult collectOffsets(GEPOperator &GEP, const DataLayout &DL) {
404 OffsetResult Result(GEP, DL);
405 unsigned BitWidth = Result.ConstantOffset.getBitWidth();
406 if (!GEP.collectOffset(DL, BitWidth, Result.VariableOffsets,
407 Result.ConstantOffset))
408 return {};
409
410 // If we have a nested GEP, check if we can combine the constant offset of the
411 // inner GEP with the outer GEP.
412 if (auto *InnerGEP = dyn_cast<GetElementPtrInst>(Result.BasePtr)) {
413 MapVector<Value *, APInt> VariableOffsets2;
414 APInt ConstantOffset2(BitWidth, 0);
415 bool CanCollectInner = InnerGEP->collectOffset(
416 DL, BitWidth, VariableOffsets2, ConstantOffset2);
417 // TODO: Support cases with more than 1 variable offset.
418 if (!CanCollectInner || Result.VariableOffsets.size() > 1 ||
419 VariableOffsets2.size() > 1 ||
420 (Result.VariableOffsets.size() >= 1 && VariableOffsets2.size() >= 1)) {
421 // More than 1 variable index, use outer result.
422 return Result;
423 }
424 Result.BasePtr = InnerGEP->getPointerOperand();
425 Result.ConstantOffset += ConstantOffset2;
426 if (Result.VariableOffsets.size() == 0 && VariableOffsets2.size() == 1)
427 Result.VariableOffsets = VariableOffsets2;
428 Result.AllInbounds &= InnerGEP->isInBounds();
429 }
430 return Result;
431 }
432
433 static Decomposition decompose(Value *V,
434 SmallVectorImpl<ConditionTy> &Preconditions,
435 bool IsSigned, const DataLayout &DL);
436
canUseSExt(ConstantInt * CI)437 static bool canUseSExt(ConstantInt *CI) {
438 const APInt &Val = CI->getValue();
439 return Val.sgt(MinSignedConstraintValue) && Val.slt(MaxConstraintValue);
440 }
441
decomposeGEP(GEPOperator & GEP,SmallVectorImpl<ConditionTy> & Preconditions,bool IsSigned,const DataLayout & DL)442 static Decomposition decomposeGEP(GEPOperator &GEP,
443 SmallVectorImpl<ConditionTy> &Preconditions,
444 bool IsSigned, const DataLayout &DL) {
445 // Do not reason about pointers where the index size is larger than 64 bits,
446 // as the coefficients used to encode constraints are 64 bit integers.
447 if (DL.getIndexTypeSizeInBits(GEP.getPointerOperand()->getType()) > 64)
448 return &GEP;
449
450 assert(!IsSigned && "The logic below only supports decomposition for "
451 "unsigned predicates at the moment.");
452 const auto &[BasePtr, ConstantOffset, VariableOffsets, AllInbounds] =
453 collectOffsets(GEP, DL);
454 if (!BasePtr || !AllInbounds)
455 return &GEP;
456
457 Decomposition Result(ConstantOffset.getSExtValue(), DecompEntry(1, BasePtr));
458 for (auto [Index, Scale] : VariableOffsets) {
459 auto IdxResult = decompose(Index, Preconditions, IsSigned, DL);
460 IdxResult.mul(Scale.getSExtValue());
461 Result.add(IdxResult);
462
463 // If Op0 is signed non-negative, the GEP is increasing monotonically and
464 // can be de-composed.
465 if (!isKnownNonNegative(Index, DL))
466 Preconditions.emplace_back(CmpInst::ICMP_SGE, Index,
467 ConstantInt::get(Index->getType(), 0));
468 }
469 return Result;
470 }
471
472 // Decomposes \p V into a constant offset + list of pairs { Coefficient,
473 // Variable } where Coefficient * Variable. The sum of the constant offset and
474 // pairs equals \p V.
decompose(Value * V,SmallVectorImpl<ConditionTy> & Preconditions,bool IsSigned,const DataLayout & DL)475 static Decomposition decompose(Value *V,
476 SmallVectorImpl<ConditionTy> &Preconditions,
477 bool IsSigned, const DataLayout &DL) {
478
479 auto MergeResults = [&Preconditions, IsSigned, &DL](Value *A, Value *B,
480 bool IsSignedB) {
481 auto ResA = decompose(A, Preconditions, IsSigned, DL);
482 auto ResB = decompose(B, Preconditions, IsSignedB, DL);
483 ResA.add(ResB);
484 return ResA;
485 };
486
487 Type *Ty = V->getType()->getScalarType();
488 if (Ty->isPointerTy() && !IsSigned) {
489 if (auto *GEP = dyn_cast<GEPOperator>(V))
490 return decomposeGEP(*GEP, Preconditions, IsSigned, DL);
491 if (isa<ConstantPointerNull>(V))
492 return int64_t(0);
493
494 return V;
495 }
496
497 // Don't handle integers > 64 bit. Our coefficients are 64-bit large, so
498 // coefficient add/mul may wrap, while the operation in the full bit width
499 // would not.
500 if (!Ty->isIntegerTy() || Ty->getIntegerBitWidth() > 64)
501 return V;
502
503 bool IsKnownNonNegative = false;
504
505 // Decompose \p V used with a signed predicate.
506 if (IsSigned) {
507 if (auto *CI = dyn_cast<ConstantInt>(V)) {
508 if (canUseSExt(CI))
509 return CI->getSExtValue();
510 }
511 Value *Op0;
512 Value *Op1;
513
514 if (match(V, m_SExt(m_Value(Op0))))
515 V = Op0;
516 else if (match(V, m_NNegZExt(m_Value(Op0)))) {
517 V = Op0;
518 IsKnownNonNegative = true;
519 }
520
521 if (match(V, m_NSWAdd(m_Value(Op0), m_Value(Op1))))
522 return MergeResults(Op0, Op1, IsSigned);
523
524 ConstantInt *CI;
525 if (match(V, m_NSWMul(m_Value(Op0), m_ConstantInt(CI))) && canUseSExt(CI)) {
526 auto Result = decompose(Op0, Preconditions, IsSigned, DL);
527 Result.mul(CI->getSExtValue());
528 return Result;
529 }
530
531 // (shl nsw x, shift) is (mul nsw x, (1<<shift)), with the exception of
532 // shift == bw-1.
533 if (match(V, m_NSWShl(m_Value(Op0), m_ConstantInt(CI)))) {
534 uint64_t Shift = CI->getValue().getLimitedValue();
535 if (Shift < Ty->getIntegerBitWidth() - 1) {
536 assert(Shift < 64 && "Would overflow");
537 auto Result = decompose(Op0, Preconditions, IsSigned, DL);
538 Result.mul(int64_t(1) << Shift);
539 return Result;
540 }
541 }
542
543 return {V, IsKnownNonNegative};
544 }
545
546 if (auto *CI = dyn_cast<ConstantInt>(V)) {
547 if (CI->uge(MaxConstraintValue))
548 return V;
549 return int64_t(CI->getZExtValue());
550 }
551
552 Value *Op0;
553 if (match(V, m_ZExt(m_Value(Op0)))) {
554 IsKnownNonNegative = true;
555 V = Op0;
556 }
557
558 if (match(V, m_SExt(m_Value(Op0)))) {
559 V = Op0;
560 Preconditions.emplace_back(CmpInst::ICMP_SGE, Op0,
561 ConstantInt::get(Op0->getType(), 0));
562 }
563
564 Value *Op1;
565 ConstantInt *CI;
566 if (match(V, m_NUWAdd(m_Value(Op0), m_Value(Op1)))) {
567 return MergeResults(Op0, Op1, IsSigned);
568 }
569 if (match(V, m_NSWAdd(m_Value(Op0), m_Value(Op1)))) {
570 if (!isKnownNonNegative(Op0, DL))
571 Preconditions.emplace_back(CmpInst::ICMP_SGE, Op0,
572 ConstantInt::get(Op0->getType(), 0));
573 if (!isKnownNonNegative(Op1, DL))
574 Preconditions.emplace_back(CmpInst::ICMP_SGE, Op1,
575 ConstantInt::get(Op1->getType(), 0));
576
577 return MergeResults(Op0, Op1, IsSigned);
578 }
579
580 if (match(V, m_Add(m_Value(Op0), m_ConstantInt(CI))) && CI->isNegative() &&
581 canUseSExt(CI)) {
582 Preconditions.emplace_back(
583 CmpInst::ICMP_UGE, Op0,
584 ConstantInt::get(Op0->getType(), CI->getSExtValue() * -1));
585 return MergeResults(Op0, CI, true);
586 }
587
588 // Decompose or as an add if there are no common bits between the operands.
589 if (match(V, m_DisjointOr(m_Value(Op0), m_ConstantInt(CI))))
590 return MergeResults(Op0, CI, IsSigned);
591
592 if (match(V, m_NUWShl(m_Value(Op1), m_ConstantInt(CI))) && canUseSExt(CI)) {
593 if (CI->getSExtValue() < 0 || CI->getSExtValue() >= 64)
594 return {V, IsKnownNonNegative};
595 auto Result = decompose(Op1, Preconditions, IsSigned, DL);
596 Result.mul(int64_t{1} << CI->getSExtValue());
597 return Result;
598 }
599
600 if (match(V, m_NUWMul(m_Value(Op1), m_ConstantInt(CI))) && canUseSExt(CI) &&
601 (!CI->isNegative())) {
602 auto Result = decompose(Op1, Preconditions, IsSigned, DL);
603 Result.mul(CI->getSExtValue());
604 return Result;
605 }
606
607 if (match(V, m_NUWSub(m_Value(Op0), m_Value(Op1)))) {
608 auto ResA = decompose(Op0, Preconditions, IsSigned, DL);
609 auto ResB = decompose(Op1, Preconditions, IsSigned, DL);
610 ResA.sub(ResB);
611 return ResA;
612 }
613
614 return {V, IsKnownNonNegative};
615 }
616
617 ConstraintTy
getConstraint(CmpInst::Predicate Pred,Value * Op0,Value * Op1,SmallVectorImpl<Value * > & NewVariables) const618 ConstraintInfo::getConstraint(CmpInst::Predicate Pred, Value *Op0, Value *Op1,
619 SmallVectorImpl<Value *> &NewVariables) const {
620 assert(NewVariables.empty() && "NewVariables must be empty when passed in");
621 bool IsEq = false;
622 bool IsNe = false;
623
624 // Try to convert Pred to one of ULE/SLT/SLE/SLT.
625 switch (Pred) {
626 case CmpInst::ICMP_UGT:
627 case CmpInst::ICMP_UGE:
628 case CmpInst::ICMP_SGT:
629 case CmpInst::ICMP_SGE: {
630 Pred = CmpInst::getSwappedPredicate(Pred);
631 std::swap(Op0, Op1);
632 break;
633 }
634 case CmpInst::ICMP_EQ:
635 if (match(Op1, m_Zero())) {
636 Pred = CmpInst::ICMP_ULE;
637 } else {
638 IsEq = true;
639 Pred = CmpInst::ICMP_ULE;
640 }
641 break;
642 case CmpInst::ICMP_NE:
643 if (match(Op1, m_Zero())) {
644 Pred = CmpInst::getSwappedPredicate(CmpInst::ICMP_UGT);
645 std::swap(Op0, Op1);
646 } else {
647 IsNe = true;
648 Pred = CmpInst::ICMP_ULE;
649 }
650 break;
651 default:
652 break;
653 }
654
655 if (Pred != CmpInst::ICMP_ULE && Pred != CmpInst::ICMP_ULT &&
656 Pred != CmpInst::ICMP_SLE && Pred != CmpInst::ICMP_SLT)
657 return {};
658
659 SmallVector<ConditionTy, 4> Preconditions;
660 bool IsSigned = CmpInst::isSigned(Pred);
661 auto &Value2Index = getValue2Index(IsSigned);
662 auto ADec = decompose(Op0->stripPointerCastsSameRepresentation(),
663 Preconditions, IsSigned, DL);
664 auto BDec = decompose(Op1->stripPointerCastsSameRepresentation(),
665 Preconditions, IsSigned, DL);
666 int64_t Offset1 = ADec.Offset;
667 int64_t Offset2 = BDec.Offset;
668 Offset1 *= -1;
669
670 auto &VariablesA = ADec.Vars;
671 auto &VariablesB = BDec.Vars;
672
673 // First try to look up \p V in Value2Index and NewVariables. Otherwise add a
674 // new entry to NewVariables.
675 SmallDenseMap<Value *, unsigned> NewIndexMap;
676 auto GetOrAddIndex = [&Value2Index, &NewVariables,
677 &NewIndexMap](Value *V) -> unsigned {
678 auto V2I = Value2Index.find(V);
679 if (V2I != Value2Index.end())
680 return V2I->second;
681 auto Insert =
682 NewIndexMap.insert({V, Value2Index.size() + NewVariables.size() + 1});
683 if (Insert.second)
684 NewVariables.push_back(V);
685 return Insert.first->second;
686 };
687
688 // Make sure all variables have entries in Value2Index or NewVariables.
689 for (const auto &KV : concat<DecompEntry>(VariablesA, VariablesB))
690 GetOrAddIndex(KV.Variable);
691
692 // Build result constraint, by first adding all coefficients from A and then
693 // subtracting all coefficients from B.
694 ConstraintTy Res(
695 SmallVector<int64_t, 8>(Value2Index.size() + NewVariables.size() + 1, 0),
696 IsSigned, IsEq, IsNe);
697 // Collect variables that are known to be positive in all uses in the
698 // constraint.
699 SmallDenseMap<Value *, bool> KnownNonNegativeVariables;
700 auto &R = Res.Coefficients;
701 for (const auto &KV : VariablesA) {
702 R[GetOrAddIndex(KV.Variable)] += KV.Coefficient;
703 auto I =
704 KnownNonNegativeVariables.insert({KV.Variable, KV.IsKnownNonNegative});
705 I.first->second &= KV.IsKnownNonNegative;
706 }
707
708 for (const auto &KV : VariablesB) {
709 if (SubOverflow(R[GetOrAddIndex(KV.Variable)], KV.Coefficient,
710 R[GetOrAddIndex(KV.Variable)]))
711 return {};
712 auto I =
713 KnownNonNegativeVariables.insert({KV.Variable, KV.IsKnownNonNegative});
714 I.first->second &= KV.IsKnownNonNegative;
715 }
716
717 int64_t OffsetSum;
718 if (AddOverflow(Offset1, Offset2, OffsetSum))
719 return {};
720 if (Pred == (IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT))
721 if (AddOverflow(OffsetSum, int64_t(-1), OffsetSum))
722 return {};
723 R[0] = OffsetSum;
724 Res.Preconditions = std::move(Preconditions);
725
726 // Remove any (Coefficient, Variable) entry where the Coefficient is 0 for new
727 // variables.
728 while (!NewVariables.empty()) {
729 int64_t Last = R.back();
730 if (Last != 0)
731 break;
732 R.pop_back();
733 Value *RemovedV = NewVariables.pop_back_val();
734 NewIndexMap.erase(RemovedV);
735 }
736
737 // Add extra constraints for variables that are known positive.
738 for (auto &KV : KnownNonNegativeVariables) {
739 if (!KV.second ||
740 (!Value2Index.contains(KV.first) && !NewIndexMap.contains(KV.first)))
741 continue;
742 SmallVector<int64_t, 8> C(Value2Index.size() + NewVariables.size() + 1, 0);
743 C[GetOrAddIndex(KV.first)] = -1;
744 Res.ExtraInfo.push_back(C);
745 }
746 return Res;
747 }
748
getConstraintForSolving(CmpInst::Predicate Pred,Value * Op0,Value * Op1) const749 ConstraintTy ConstraintInfo::getConstraintForSolving(CmpInst::Predicate Pred,
750 Value *Op0,
751 Value *Op1) const {
752 Constant *NullC = Constant::getNullValue(Op0->getType());
753 // Handle trivially true compares directly to avoid adding V UGE 0 constraints
754 // for all variables in the unsigned system.
755 if ((Pred == CmpInst::ICMP_ULE && Op0 == NullC) ||
756 (Pred == CmpInst::ICMP_UGE && Op1 == NullC)) {
757 auto &Value2Index = getValue2Index(false);
758 // Return constraint that's trivially true.
759 return ConstraintTy(SmallVector<int64_t, 8>(Value2Index.size(), 0), false,
760 false, false);
761 }
762
763 // If both operands are known to be non-negative, change signed predicates to
764 // unsigned ones. This increases the reasoning effectiveness in combination
765 // with the signed <-> unsigned transfer logic.
766 if (CmpInst::isSigned(Pred) &&
767 isKnownNonNegative(Op0, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1) &&
768 isKnownNonNegative(Op1, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1))
769 Pred = CmpInst::getUnsignedPredicate(Pred);
770
771 SmallVector<Value *> NewVariables;
772 ConstraintTy R = getConstraint(Pred, Op0, Op1, NewVariables);
773 if (!NewVariables.empty())
774 return {};
775 return R;
776 }
777
isValid(const ConstraintInfo & Info) const778 bool ConstraintTy::isValid(const ConstraintInfo &Info) const {
779 return Coefficients.size() > 0 &&
780 all_of(Preconditions, [&Info](const ConditionTy &C) {
781 return Info.doesHold(C.Pred, C.Op0, C.Op1);
782 });
783 }
784
785 std::optional<bool>
isImpliedBy(const ConstraintSystem & CS) const786 ConstraintTy::isImpliedBy(const ConstraintSystem &CS) const {
787 bool IsConditionImplied = CS.isConditionImplied(Coefficients);
788
789 if (IsEq || IsNe) {
790 auto NegatedOrEqual = ConstraintSystem::negateOrEqual(Coefficients);
791 bool IsNegatedOrEqualImplied =
792 !NegatedOrEqual.empty() && CS.isConditionImplied(NegatedOrEqual);
793
794 // In order to check that `%a == %b` is true (equality), both conditions `%a
795 // >= %b` and `%a <= %b` must hold true. When checking for equality (`IsEq`
796 // is true), we return true if they both hold, false in the other cases.
797 if (IsConditionImplied && IsNegatedOrEqualImplied)
798 return IsEq;
799
800 auto Negated = ConstraintSystem::negate(Coefficients);
801 bool IsNegatedImplied = !Negated.empty() && CS.isConditionImplied(Negated);
802
803 auto StrictLessThan = ConstraintSystem::toStrictLessThan(Coefficients);
804 bool IsStrictLessThanImplied =
805 !StrictLessThan.empty() && CS.isConditionImplied(StrictLessThan);
806
807 // In order to check that `%a != %b` is true (non-equality), either
808 // condition `%a > %b` or `%a < %b` must hold true. When checking for
809 // non-equality (`IsNe` is true), we return true if one of the two holds,
810 // false in the other cases.
811 if (IsNegatedImplied || IsStrictLessThanImplied)
812 return IsNe;
813
814 return std::nullopt;
815 }
816
817 if (IsConditionImplied)
818 return true;
819
820 auto Negated = ConstraintSystem::negate(Coefficients);
821 auto IsNegatedImplied = !Negated.empty() && CS.isConditionImplied(Negated);
822 if (IsNegatedImplied)
823 return false;
824
825 // Neither the condition nor its negated holds, did not prove anything.
826 return std::nullopt;
827 }
828
doesHold(CmpInst::Predicate Pred,Value * A,Value * B) const829 bool ConstraintInfo::doesHold(CmpInst::Predicate Pred, Value *A,
830 Value *B) const {
831 auto R = getConstraintForSolving(Pred, A, B);
832 return R.isValid(*this) &&
833 getCS(R.IsSigned).isConditionImplied(R.Coefficients);
834 }
835
transferToOtherSystem(CmpInst::Predicate Pred,Value * A,Value * B,unsigned NumIn,unsigned NumOut,SmallVectorImpl<StackEntry> & DFSInStack)836 void ConstraintInfo::transferToOtherSystem(
837 CmpInst::Predicate Pred, Value *A, Value *B, unsigned NumIn,
838 unsigned NumOut, SmallVectorImpl<StackEntry> &DFSInStack) {
839 auto IsKnownNonNegative = [this](Value *V) {
840 return doesHold(CmpInst::ICMP_SGE, V, ConstantInt::get(V->getType(), 0)) ||
841 isKnownNonNegative(V, DL, /*Depth=*/MaxAnalysisRecursionDepth - 1);
842 };
843 // Check if we can combine facts from the signed and unsigned systems to
844 // derive additional facts.
845 if (!A->getType()->isIntegerTy())
846 return;
847 // FIXME: This currently depends on the order we add facts. Ideally we
848 // would first add all known facts and only then try to add additional
849 // facts.
850 switch (Pred) {
851 default:
852 break;
853 case CmpInst::ICMP_ULT:
854 case CmpInst::ICMP_ULE:
855 // If B is a signed positive constant, then A >=s 0 and A <s (or <=s) B.
856 if (IsKnownNonNegative(B)) {
857 addFact(CmpInst::ICMP_SGE, A, ConstantInt::get(B->getType(), 0), NumIn,
858 NumOut, DFSInStack);
859 addFact(CmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut,
860 DFSInStack);
861 }
862 break;
863 case CmpInst::ICMP_UGE:
864 case CmpInst::ICMP_UGT:
865 // If A is a signed positive constant, then B >=s 0 and A >s (or >=s) B.
866 if (IsKnownNonNegative(A)) {
867 addFact(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), 0), NumIn,
868 NumOut, DFSInStack);
869 addFact(CmpInst::getSignedPredicate(Pred), A, B, NumIn, NumOut,
870 DFSInStack);
871 }
872 break;
873 case CmpInst::ICMP_SLT:
874 if (IsKnownNonNegative(A))
875 addFact(CmpInst::ICMP_ULT, A, B, NumIn, NumOut, DFSInStack);
876 break;
877 case CmpInst::ICMP_SGT: {
878 if (doesHold(CmpInst::ICMP_SGE, B, ConstantInt::get(B->getType(), -1)))
879 addFact(CmpInst::ICMP_UGE, A, ConstantInt::get(B->getType(), 0), NumIn,
880 NumOut, DFSInStack);
881 if (IsKnownNonNegative(B))
882 addFact(CmpInst::ICMP_UGT, A, B, NumIn, NumOut, DFSInStack);
883
884 break;
885 }
886 case CmpInst::ICMP_SGE:
887 if (IsKnownNonNegative(B))
888 addFact(CmpInst::ICMP_UGE, A, B, NumIn, NumOut, DFSInStack);
889 break;
890 }
891 }
892
893 #ifndef NDEBUG
894
dumpConstraint(ArrayRef<int64_t> C,const DenseMap<Value *,unsigned> & Value2Index)895 static void dumpConstraint(ArrayRef<int64_t> C,
896 const DenseMap<Value *, unsigned> &Value2Index) {
897 ConstraintSystem CS(Value2Index);
898 CS.addVariableRowFill(C);
899 CS.dump();
900 }
901 #endif
902
addInfoForInductions(BasicBlock & BB)903 void State::addInfoForInductions(BasicBlock &BB) {
904 auto *L = LI.getLoopFor(&BB);
905 if (!L || L->getHeader() != &BB)
906 return;
907
908 Value *A;
909 Value *B;
910 CmpInst::Predicate Pred;
911
912 if (!match(BB.getTerminator(),
913 m_Br(m_ICmp(Pred, m_Value(A), m_Value(B)), m_Value(), m_Value())))
914 return;
915 PHINode *PN = dyn_cast<PHINode>(A);
916 if (!PN) {
917 Pred = CmpInst::getSwappedPredicate(Pred);
918 std::swap(A, B);
919 PN = dyn_cast<PHINode>(A);
920 }
921
922 if (!PN || PN->getParent() != &BB || PN->getNumIncomingValues() != 2 ||
923 !SE.isSCEVable(PN->getType()))
924 return;
925
926 BasicBlock *InLoopSucc = nullptr;
927 if (Pred == CmpInst::ICMP_NE)
928 InLoopSucc = cast<BranchInst>(BB.getTerminator())->getSuccessor(0);
929 else if (Pred == CmpInst::ICMP_EQ)
930 InLoopSucc = cast<BranchInst>(BB.getTerminator())->getSuccessor(1);
931 else
932 return;
933
934 if (!L->contains(InLoopSucc) || !L->isLoopExiting(&BB) || InLoopSucc == &BB)
935 return;
936
937 auto *AR = dyn_cast_or_null<SCEVAddRecExpr>(SE.getSCEV(PN));
938 BasicBlock *LoopPred = L->getLoopPredecessor();
939 if (!AR || AR->getLoop() != L || !LoopPred)
940 return;
941
942 const SCEV *StartSCEV = AR->getStart();
943 Value *StartValue = nullptr;
944 if (auto *C = dyn_cast<SCEVConstant>(StartSCEV)) {
945 StartValue = C->getValue();
946 } else {
947 StartValue = PN->getIncomingValueForBlock(LoopPred);
948 assert(SE.getSCEV(StartValue) == StartSCEV && "inconsistent start value");
949 }
950
951 DomTreeNode *DTN = DT.getNode(InLoopSucc);
952 auto IncUnsigned = SE.getMonotonicPredicateType(AR, CmpInst::ICMP_UGT);
953 auto IncSigned = SE.getMonotonicPredicateType(AR, CmpInst::ICMP_SGT);
954 bool MonotonicallyIncreasingUnsigned =
955 IncUnsigned && *IncUnsigned == ScalarEvolution::MonotonicallyIncreasing;
956 bool MonotonicallyIncreasingSigned =
957 IncSigned && *IncSigned == ScalarEvolution::MonotonicallyIncreasing;
958 // If SCEV guarantees that AR does not wrap, PN >= StartValue can be added
959 // unconditionally.
960 if (MonotonicallyIncreasingUnsigned)
961 WorkList.push_back(
962 FactOrCheck::getConditionFact(DTN, CmpInst::ICMP_UGE, PN, StartValue));
963 if (MonotonicallyIncreasingSigned)
964 WorkList.push_back(
965 FactOrCheck::getConditionFact(DTN, CmpInst::ICMP_SGE, PN, StartValue));
966
967 APInt StepOffset;
968 if (auto *C = dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
969 StepOffset = C->getAPInt();
970 else
971 return;
972
973 // Make sure the bound B is loop-invariant.
974 if (!L->isLoopInvariant(B))
975 return;
976
977 // Handle negative steps.
978 if (StepOffset.isNegative()) {
979 // TODO: Extend to allow steps > -1.
980 if (!(-StepOffset).isOne())
981 return;
982
983 // AR may wrap.
984 // Add StartValue >= PN conditional on B <= StartValue which guarantees that
985 // the loop exits before wrapping with a step of -1.
986 WorkList.push_back(FactOrCheck::getConditionFact(
987 DTN, CmpInst::ICMP_UGE, StartValue, PN,
988 ConditionTy(CmpInst::ICMP_ULE, B, StartValue)));
989 WorkList.push_back(FactOrCheck::getConditionFact(
990 DTN, CmpInst::ICMP_SGE, StartValue, PN,
991 ConditionTy(CmpInst::ICMP_SLE, B, StartValue)));
992 // Add PN > B conditional on B <= StartValue which guarantees that the loop
993 // exits when reaching B with a step of -1.
994 WorkList.push_back(FactOrCheck::getConditionFact(
995 DTN, CmpInst::ICMP_UGT, PN, B,
996 ConditionTy(CmpInst::ICMP_ULE, B, StartValue)));
997 WorkList.push_back(FactOrCheck::getConditionFact(
998 DTN, CmpInst::ICMP_SGT, PN, B,
999 ConditionTy(CmpInst::ICMP_SLE, B, StartValue)));
1000 return;
1001 }
1002
1003 // Make sure AR either steps by 1 or that the value we compare against is a
1004 // GEP based on the same start value and all offsets are a multiple of the
1005 // step size, to guarantee that the induction will reach the value.
1006 if (StepOffset.isZero() || StepOffset.isNegative())
1007 return;
1008
1009 if (!StepOffset.isOne()) {
1010 // Check whether B-Start is known to be a multiple of StepOffset.
1011 const SCEV *BMinusStart = SE.getMinusSCEV(SE.getSCEV(B), StartSCEV);
1012 if (isa<SCEVCouldNotCompute>(BMinusStart) ||
1013 !SE.getConstantMultiple(BMinusStart).urem(StepOffset).isZero())
1014 return;
1015 }
1016
1017 // AR may wrap. Add PN >= StartValue conditional on StartValue <= B which
1018 // guarantees that the loop exits before wrapping in combination with the
1019 // restrictions on B and the step above.
1020 if (!MonotonicallyIncreasingUnsigned)
1021 WorkList.push_back(FactOrCheck::getConditionFact(
1022 DTN, CmpInst::ICMP_UGE, PN, StartValue,
1023 ConditionTy(CmpInst::ICMP_ULE, StartValue, B)));
1024 if (!MonotonicallyIncreasingSigned)
1025 WorkList.push_back(FactOrCheck::getConditionFact(
1026 DTN, CmpInst::ICMP_SGE, PN, StartValue,
1027 ConditionTy(CmpInst::ICMP_SLE, StartValue, B)));
1028
1029 WorkList.push_back(FactOrCheck::getConditionFact(
1030 DTN, CmpInst::ICMP_ULT, PN, B,
1031 ConditionTy(CmpInst::ICMP_ULE, StartValue, B)));
1032 WorkList.push_back(FactOrCheck::getConditionFact(
1033 DTN, CmpInst::ICMP_SLT, PN, B,
1034 ConditionTy(CmpInst::ICMP_SLE, StartValue, B)));
1035
1036 // Try to add condition from header to the dedicated exit blocks. When exiting
1037 // either with EQ or NE in the header, we know that the induction value must
1038 // be u<= B, as other exits may only exit earlier.
1039 assert(!StepOffset.isNegative() && "induction must be increasing");
1040 assert((Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) &&
1041 "unsupported predicate");
1042 ConditionTy Precond = {CmpInst::ICMP_ULE, StartValue, B};
1043 SmallVector<BasicBlock *> ExitBBs;
1044 L->getExitBlocks(ExitBBs);
1045 for (BasicBlock *EB : ExitBBs) {
1046 // Bail out on non-dedicated exits.
1047 if (DT.dominates(&BB, EB)) {
1048 WorkList.emplace_back(FactOrCheck::getConditionFact(
1049 DT.getNode(EB), CmpInst::ICMP_ULE, A, B, Precond));
1050 }
1051 }
1052 }
1053
addInfoFor(BasicBlock & BB)1054 void State::addInfoFor(BasicBlock &BB) {
1055 addInfoForInductions(BB);
1056
1057 // True as long as long as the current instruction is guaranteed to execute.
1058 bool GuaranteedToExecute = true;
1059 // Queue conditions and assumes.
1060 for (Instruction &I : BB) {
1061 if (auto Cmp = dyn_cast<ICmpInst>(&I)) {
1062 for (Use &U : Cmp->uses()) {
1063 auto *UserI = getContextInstForUse(U);
1064 auto *DTN = DT.getNode(UserI->getParent());
1065 if (!DTN)
1066 continue;
1067 WorkList.push_back(FactOrCheck::getCheck(DTN, &U));
1068 }
1069 continue;
1070 }
1071
1072 auto *II = dyn_cast<IntrinsicInst>(&I);
1073 Intrinsic::ID ID = II ? II->getIntrinsicID() : Intrinsic::not_intrinsic;
1074 switch (ID) {
1075 case Intrinsic::assume: {
1076 Value *A, *B;
1077 CmpInst::Predicate Pred;
1078 if (!match(I.getOperand(0), m_ICmp(Pred, m_Value(A), m_Value(B))))
1079 break;
1080 if (GuaranteedToExecute) {
1081 // The assume is guaranteed to execute when BB is entered, hence Cond
1082 // holds on entry to BB.
1083 WorkList.emplace_back(FactOrCheck::getConditionFact(
1084 DT.getNode(I.getParent()), Pred, A, B));
1085 } else {
1086 WorkList.emplace_back(
1087 FactOrCheck::getInstFact(DT.getNode(I.getParent()), &I));
1088 }
1089 break;
1090 }
1091 // Enqueue ssub_with_overflow for simplification.
1092 case Intrinsic::ssub_with_overflow:
1093 case Intrinsic::ucmp:
1094 case Intrinsic::scmp:
1095 WorkList.push_back(
1096 FactOrCheck::getCheck(DT.getNode(&BB), cast<CallInst>(&I)));
1097 break;
1098 // Enqueue the intrinsics to add extra info.
1099 case Intrinsic::umin:
1100 case Intrinsic::umax:
1101 case Intrinsic::smin:
1102 case Intrinsic::smax:
1103 // TODO: handle llvm.abs as well
1104 WorkList.push_back(
1105 FactOrCheck::getCheck(DT.getNode(&BB), cast<CallInst>(&I)));
1106 // TODO: Check if it is possible to instead only added the min/max facts
1107 // when simplifying uses of the min/max intrinsics.
1108 if (!isGuaranteedNotToBePoison(&I))
1109 break;
1110 [[fallthrough]];
1111 case Intrinsic::abs:
1112 WorkList.push_back(FactOrCheck::getInstFact(DT.getNode(&BB), &I));
1113 break;
1114 }
1115
1116 GuaranteedToExecute &= isGuaranteedToTransferExecutionToSuccessor(&I);
1117 }
1118
1119 if (auto *Switch = dyn_cast<SwitchInst>(BB.getTerminator())) {
1120 for (auto &Case : Switch->cases()) {
1121 BasicBlock *Succ = Case.getCaseSuccessor();
1122 Value *V = Case.getCaseValue();
1123 if (!canAddSuccessor(BB, Succ))
1124 continue;
1125 WorkList.emplace_back(FactOrCheck::getConditionFact(
1126 DT.getNode(Succ), CmpInst::ICMP_EQ, Switch->getCondition(), V));
1127 }
1128 return;
1129 }
1130
1131 auto *Br = dyn_cast<BranchInst>(BB.getTerminator());
1132 if (!Br || !Br->isConditional())
1133 return;
1134
1135 Value *Cond = Br->getCondition();
1136
1137 // If the condition is a chain of ORs/AND and the successor only has the
1138 // current block as predecessor, queue conditions for the successor.
1139 Value *Op0, *Op1;
1140 if (match(Cond, m_LogicalOr(m_Value(Op0), m_Value(Op1))) ||
1141 match(Cond, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
1142 bool IsOr = match(Cond, m_LogicalOr());
1143 bool IsAnd = match(Cond, m_LogicalAnd());
1144 // If there's a select that matches both AND and OR, we need to commit to
1145 // one of the options. Arbitrarily pick OR.
1146 if (IsOr && IsAnd)
1147 IsAnd = false;
1148
1149 BasicBlock *Successor = Br->getSuccessor(IsOr ? 1 : 0);
1150 if (canAddSuccessor(BB, Successor)) {
1151 SmallVector<Value *> CondWorkList;
1152 SmallPtrSet<Value *, 8> SeenCond;
1153 auto QueueValue = [&CondWorkList, &SeenCond](Value *V) {
1154 if (SeenCond.insert(V).second)
1155 CondWorkList.push_back(V);
1156 };
1157 QueueValue(Op1);
1158 QueueValue(Op0);
1159 while (!CondWorkList.empty()) {
1160 Value *Cur = CondWorkList.pop_back_val();
1161 if (auto *Cmp = dyn_cast<ICmpInst>(Cur)) {
1162 WorkList.emplace_back(FactOrCheck::getConditionFact(
1163 DT.getNode(Successor),
1164 IsOr ? CmpInst::getInversePredicate(Cmp->getPredicate())
1165 : Cmp->getPredicate(),
1166 Cmp->getOperand(0), Cmp->getOperand(1)));
1167 continue;
1168 }
1169 if (IsOr && match(Cur, m_LogicalOr(m_Value(Op0), m_Value(Op1)))) {
1170 QueueValue(Op1);
1171 QueueValue(Op0);
1172 continue;
1173 }
1174 if (IsAnd && match(Cur, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
1175 QueueValue(Op1);
1176 QueueValue(Op0);
1177 continue;
1178 }
1179 }
1180 }
1181 return;
1182 }
1183
1184 auto *CmpI = dyn_cast<ICmpInst>(Br->getCondition());
1185 if (!CmpI)
1186 return;
1187 if (canAddSuccessor(BB, Br->getSuccessor(0)))
1188 WorkList.emplace_back(FactOrCheck::getConditionFact(
1189 DT.getNode(Br->getSuccessor(0)), CmpI->getPredicate(),
1190 CmpI->getOperand(0), CmpI->getOperand(1)));
1191 if (canAddSuccessor(BB, Br->getSuccessor(1)))
1192 WorkList.emplace_back(FactOrCheck::getConditionFact(
1193 DT.getNode(Br->getSuccessor(1)),
1194 CmpInst::getInversePredicate(CmpI->getPredicate()), CmpI->getOperand(0),
1195 CmpI->getOperand(1)));
1196 }
1197
1198 #ifndef NDEBUG
dumpUnpackedICmp(raw_ostream & OS,ICmpInst::Predicate Pred,Value * LHS,Value * RHS)1199 static void dumpUnpackedICmp(raw_ostream &OS, ICmpInst::Predicate Pred,
1200 Value *LHS, Value *RHS) {
1201 OS << "icmp " << Pred << ' ';
1202 LHS->printAsOperand(OS, /*PrintType=*/true);
1203 OS << ", ";
1204 RHS->printAsOperand(OS, /*PrintType=*/false);
1205 }
1206 #endif
1207
1208 namespace {
1209 /// Helper to keep track of a condition and if it should be treated as negated
1210 /// for reproducer construction.
1211 /// Pred == Predicate::BAD_ICMP_PREDICATE indicates that this entry is a
1212 /// placeholder to keep the ReproducerCondStack in sync with DFSInStack.
1213 struct ReproducerEntry {
1214 ICmpInst::Predicate Pred;
1215 Value *LHS;
1216 Value *RHS;
1217
ReproducerEntry__anon050fee910811::ReproducerEntry1218 ReproducerEntry(ICmpInst::Predicate Pred, Value *LHS, Value *RHS)
1219 : Pred(Pred), LHS(LHS), RHS(RHS) {}
1220 };
1221 } // namespace
1222
1223 /// Helper function to generate a reproducer function for simplifying \p Cond.
1224 /// The reproducer function contains a series of @llvm.assume calls, one for
1225 /// each condition in \p Stack. For each condition, the operand instruction are
1226 /// cloned until we reach operands that have an entry in \p Value2Index. Those
1227 /// will then be added as function arguments. \p DT is used to order cloned
1228 /// instructions. The reproducer function will get added to \p M, if it is
1229 /// non-null. Otherwise no reproducer function is generated.
generateReproducer(CmpInst * Cond,Module * M,ArrayRef<ReproducerEntry> Stack,ConstraintInfo & Info,DominatorTree & DT)1230 static void generateReproducer(CmpInst *Cond, Module *M,
1231 ArrayRef<ReproducerEntry> Stack,
1232 ConstraintInfo &Info, DominatorTree &DT) {
1233 if (!M)
1234 return;
1235
1236 LLVMContext &Ctx = Cond->getContext();
1237
1238 LLVM_DEBUG(dbgs() << "Creating reproducer for " << *Cond << "\n");
1239
1240 ValueToValueMapTy Old2New;
1241 SmallVector<Value *> Args;
1242 SmallPtrSet<Value *, 8> Seen;
1243 // Traverse Cond and its operands recursively until we reach a value that's in
1244 // Value2Index or not an instruction, or not a operation that
1245 // ConstraintElimination can decompose. Such values will be considered as
1246 // external inputs to the reproducer, they are collected and added as function
1247 // arguments later.
1248 auto CollectArguments = [&](ArrayRef<Value *> Ops, bool IsSigned) {
1249 auto &Value2Index = Info.getValue2Index(IsSigned);
1250 SmallVector<Value *, 4> WorkList(Ops);
1251 while (!WorkList.empty()) {
1252 Value *V = WorkList.pop_back_val();
1253 if (!Seen.insert(V).second)
1254 continue;
1255 if (Old2New.find(V) != Old2New.end())
1256 continue;
1257 if (isa<Constant>(V))
1258 continue;
1259
1260 auto *I = dyn_cast<Instruction>(V);
1261 if (Value2Index.contains(V) || !I ||
1262 !isa<CmpInst, BinaryOperator, GEPOperator, CastInst>(V)) {
1263 Old2New[V] = V;
1264 Args.push_back(V);
1265 LLVM_DEBUG(dbgs() << " found external input " << *V << "\n");
1266 } else {
1267 append_range(WorkList, I->operands());
1268 }
1269 }
1270 };
1271
1272 for (auto &Entry : Stack)
1273 if (Entry.Pred != ICmpInst::BAD_ICMP_PREDICATE)
1274 CollectArguments({Entry.LHS, Entry.RHS}, ICmpInst::isSigned(Entry.Pred));
1275 CollectArguments(Cond, ICmpInst::isSigned(Cond->getPredicate()));
1276
1277 SmallVector<Type *> ParamTys;
1278 for (auto *P : Args)
1279 ParamTys.push_back(P->getType());
1280
1281 FunctionType *FTy = FunctionType::get(Cond->getType(), ParamTys,
1282 /*isVarArg=*/false);
1283 Function *F = Function::Create(FTy, Function::ExternalLinkage,
1284 Cond->getModule()->getName() +
1285 Cond->getFunction()->getName() + "repro",
1286 M);
1287 // Add arguments to the reproducer function for each external value collected.
1288 for (unsigned I = 0; I < Args.size(); ++I) {
1289 F->getArg(I)->setName(Args[I]->getName());
1290 Old2New[Args[I]] = F->getArg(I);
1291 }
1292
1293 BasicBlock *Entry = BasicBlock::Create(Ctx, "entry", F);
1294 IRBuilder<> Builder(Entry);
1295 Builder.CreateRet(Builder.getTrue());
1296 Builder.SetInsertPoint(Entry->getTerminator());
1297
1298 // Clone instructions in \p Ops and their operands recursively until reaching
1299 // an value in Value2Index (external input to the reproducer). Update Old2New
1300 // mapping for the original and cloned instructions. Sort instructions to
1301 // clone by dominance, then insert the cloned instructions in the function.
1302 auto CloneInstructions = [&](ArrayRef<Value *> Ops, bool IsSigned) {
1303 SmallVector<Value *, 4> WorkList(Ops);
1304 SmallVector<Instruction *> ToClone;
1305 auto &Value2Index = Info.getValue2Index(IsSigned);
1306 while (!WorkList.empty()) {
1307 Value *V = WorkList.pop_back_val();
1308 if (Old2New.find(V) != Old2New.end())
1309 continue;
1310
1311 auto *I = dyn_cast<Instruction>(V);
1312 if (!Value2Index.contains(V) && I) {
1313 Old2New[V] = nullptr;
1314 ToClone.push_back(I);
1315 append_range(WorkList, I->operands());
1316 }
1317 }
1318
1319 sort(ToClone,
1320 [&DT](Instruction *A, Instruction *B) { return DT.dominates(A, B); });
1321 for (Instruction *I : ToClone) {
1322 Instruction *Cloned = I->clone();
1323 Old2New[I] = Cloned;
1324 Old2New[I]->setName(I->getName());
1325 Cloned->insertBefore(&*Builder.GetInsertPoint());
1326 Cloned->dropUnknownNonDebugMetadata();
1327 Cloned->setDebugLoc({});
1328 }
1329 };
1330
1331 // Materialize the assumptions for the reproducer using the entries in Stack.
1332 // That is, first clone the operands of the condition recursively until we
1333 // reach an external input to the reproducer and add them to the reproducer
1334 // function. Then add an ICmp for the condition (with the inverse predicate if
1335 // the entry is negated) and an assert using the ICmp.
1336 for (auto &Entry : Stack) {
1337 if (Entry.Pred == ICmpInst::BAD_ICMP_PREDICATE)
1338 continue;
1339
1340 LLVM_DEBUG(dbgs() << " Materializing assumption ";
1341 dumpUnpackedICmp(dbgs(), Entry.Pred, Entry.LHS, Entry.RHS);
1342 dbgs() << "\n");
1343 CloneInstructions({Entry.LHS, Entry.RHS}, CmpInst::isSigned(Entry.Pred));
1344
1345 auto *Cmp = Builder.CreateICmp(Entry.Pred, Entry.LHS, Entry.RHS);
1346 Builder.CreateAssumption(Cmp);
1347 }
1348
1349 // Finally, clone the condition to reproduce and remap instruction operands in
1350 // the reproducer using Old2New.
1351 CloneInstructions(Cond, CmpInst::isSigned(Cond->getPredicate()));
1352 Entry->getTerminator()->setOperand(0, Cond);
1353 remapInstructionsInBlocks({Entry}, Old2New);
1354
1355 assert(!verifyFunction(*F, &dbgs()));
1356 }
1357
checkCondition(CmpInst::Predicate Pred,Value * A,Value * B,Instruction * CheckInst,ConstraintInfo & Info)1358 static std::optional<bool> checkCondition(CmpInst::Predicate Pred, Value *A,
1359 Value *B, Instruction *CheckInst,
1360 ConstraintInfo &Info) {
1361 LLVM_DEBUG(dbgs() << "Checking " << *CheckInst << "\n");
1362
1363 auto R = Info.getConstraintForSolving(Pred, A, B);
1364 if (R.empty() || !R.isValid(Info)){
1365 LLVM_DEBUG(dbgs() << " failed to decompose condition\n");
1366 return std::nullopt;
1367 }
1368
1369 auto &CSToUse = Info.getCS(R.IsSigned);
1370
1371 // If there was extra information collected during decomposition, apply
1372 // it now and remove it immediately once we are done with reasoning
1373 // about the constraint.
1374 for (auto &Row : R.ExtraInfo)
1375 CSToUse.addVariableRow(Row);
1376 auto InfoRestorer = make_scope_exit([&]() {
1377 for (unsigned I = 0; I < R.ExtraInfo.size(); ++I)
1378 CSToUse.popLastConstraint();
1379 });
1380
1381 if (auto ImpliedCondition = R.isImpliedBy(CSToUse)) {
1382 if (!DebugCounter::shouldExecute(EliminatedCounter))
1383 return std::nullopt;
1384
1385 LLVM_DEBUG({
1386 dbgs() << "Condition ";
1387 dumpUnpackedICmp(
1388 dbgs(), *ImpliedCondition ? Pred : CmpInst::getInversePredicate(Pred),
1389 A, B);
1390 dbgs() << " implied by dominating constraints\n";
1391 CSToUse.dump();
1392 });
1393 return ImpliedCondition;
1394 }
1395
1396 return std::nullopt;
1397 }
1398
checkAndReplaceCondition(CmpInst * Cmp,ConstraintInfo & Info,unsigned NumIn,unsigned NumOut,Instruction * ContextInst,Module * ReproducerModule,ArrayRef<ReproducerEntry> ReproducerCondStack,DominatorTree & DT,SmallVectorImpl<Instruction * > & ToRemove)1399 static bool checkAndReplaceCondition(
1400 CmpInst *Cmp, ConstraintInfo &Info, unsigned NumIn, unsigned NumOut,
1401 Instruction *ContextInst, Module *ReproducerModule,
1402 ArrayRef<ReproducerEntry> ReproducerCondStack, DominatorTree &DT,
1403 SmallVectorImpl<Instruction *> &ToRemove) {
1404 auto ReplaceCmpWithConstant = [&](CmpInst *Cmp, bool IsTrue) {
1405 generateReproducer(Cmp, ReproducerModule, ReproducerCondStack, Info, DT);
1406 Constant *ConstantC = ConstantInt::getBool(
1407 CmpInst::makeCmpResultType(Cmp->getType()), IsTrue);
1408 Cmp->replaceUsesWithIf(ConstantC, [&DT, NumIn, NumOut,
1409 ContextInst](Use &U) {
1410 auto *UserI = getContextInstForUse(U);
1411 auto *DTN = DT.getNode(UserI->getParent());
1412 if (!DTN || DTN->getDFSNumIn() < NumIn || DTN->getDFSNumOut() > NumOut)
1413 return false;
1414 if (UserI->getParent() == ContextInst->getParent() &&
1415 UserI->comesBefore(ContextInst))
1416 return false;
1417
1418 // Conditions in an assume trivially simplify to true. Skip uses
1419 // in assume calls to not destroy the available information.
1420 auto *II = dyn_cast<IntrinsicInst>(U.getUser());
1421 return !II || II->getIntrinsicID() != Intrinsic::assume;
1422 });
1423 NumCondsRemoved++;
1424 if (Cmp->use_empty())
1425 ToRemove.push_back(Cmp);
1426 return true;
1427 };
1428
1429 if (auto ImpliedCondition =
1430 checkCondition(Cmp->getPredicate(), Cmp->getOperand(0),
1431 Cmp->getOperand(1), Cmp, Info))
1432 return ReplaceCmpWithConstant(Cmp, *ImpliedCondition);
1433 return false;
1434 }
1435
checkAndReplaceMinMax(MinMaxIntrinsic * MinMax,ConstraintInfo & Info,SmallVectorImpl<Instruction * > & ToRemove)1436 static bool checkAndReplaceMinMax(MinMaxIntrinsic *MinMax, ConstraintInfo &Info,
1437 SmallVectorImpl<Instruction *> &ToRemove) {
1438 auto ReplaceMinMaxWithOperand = [&](MinMaxIntrinsic *MinMax, bool UseLHS) {
1439 // TODO: generate reproducer for min/max.
1440 MinMax->replaceAllUsesWith(MinMax->getOperand(UseLHS ? 0 : 1));
1441 ToRemove.push_back(MinMax);
1442 return true;
1443 };
1444
1445 ICmpInst::Predicate Pred =
1446 ICmpInst::getNonStrictPredicate(MinMax->getPredicate());
1447 if (auto ImpliedCondition = checkCondition(
1448 Pred, MinMax->getOperand(0), MinMax->getOperand(1), MinMax, Info))
1449 return ReplaceMinMaxWithOperand(MinMax, *ImpliedCondition);
1450 if (auto ImpliedCondition = checkCondition(
1451 Pred, MinMax->getOperand(1), MinMax->getOperand(0), MinMax, Info))
1452 return ReplaceMinMaxWithOperand(MinMax, !*ImpliedCondition);
1453 return false;
1454 }
1455
checkAndReplaceCmp(CmpIntrinsic * I,ConstraintInfo & Info,SmallVectorImpl<Instruction * > & ToRemove)1456 static bool checkAndReplaceCmp(CmpIntrinsic *I, ConstraintInfo &Info,
1457 SmallVectorImpl<Instruction *> &ToRemove) {
1458 Value *LHS = I->getOperand(0);
1459 Value *RHS = I->getOperand(1);
1460 if (checkCondition(I->getGTPredicate(), LHS, RHS, I, Info).value_or(false)) {
1461 I->replaceAllUsesWith(ConstantInt::get(I->getType(), 1));
1462 ToRemove.push_back(I);
1463 return true;
1464 }
1465 if (checkCondition(I->getLTPredicate(), LHS, RHS, I, Info).value_or(false)) {
1466 I->replaceAllUsesWith(ConstantInt::getSigned(I->getType(), -1));
1467 ToRemove.push_back(I);
1468 return true;
1469 }
1470 if (checkCondition(ICmpInst::ICMP_EQ, LHS, RHS, I, Info).value_or(false)) {
1471 I->replaceAllUsesWith(ConstantInt::get(I->getType(), 0));
1472 ToRemove.push_back(I);
1473 return true;
1474 }
1475 return false;
1476 }
1477
1478 static void
removeEntryFromStack(const StackEntry & E,ConstraintInfo & Info,Module * ReproducerModule,SmallVectorImpl<ReproducerEntry> & ReproducerCondStack,SmallVectorImpl<StackEntry> & DFSInStack)1479 removeEntryFromStack(const StackEntry &E, ConstraintInfo &Info,
1480 Module *ReproducerModule,
1481 SmallVectorImpl<ReproducerEntry> &ReproducerCondStack,
1482 SmallVectorImpl<StackEntry> &DFSInStack) {
1483 Info.popLastConstraint(E.IsSigned);
1484 // Remove variables in the system that went out of scope.
1485 auto &Mapping = Info.getValue2Index(E.IsSigned);
1486 for (Value *V : E.ValuesToRelease)
1487 Mapping.erase(V);
1488 Info.popLastNVariables(E.IsSigned, E.ValuesToRelease.size());
1489 DFSInStack.pop_back();
1490 if (ReproducerModule)
1491 ReproducerCondStack.pop_back();
1492 }
1493
1494 /// Check if either the first condition of an AND or OR is implied by the
1495 /// (negated in case of OR) second condition or vice versa.
checkOrAndOpImpliedByOther(FactOrCheck & CB,ConstraintInfo & Info,Module * ReproducerModule,SmallVectorImpl<ReproducerEntry> & ReproducerCondStack,SmallVectorImpl<StackEntry> & DFSInStack)1496 static bool checkOrAndOpImpliedByOther(
1497 FactOrCheck &CB, ConstraintInfo &Info, Module *ReproducerModule,
1498 SmallVectorImpl<ReproducerEntry> &ReproducerCondStack,
1499 SmallVectorImpl<StackEntry> &DFSInStack) {
1500
1501 CmpInst::Predicate Pred;
1502 Value *A, *B;
1503 Instruction *JoinOp = CB.getContextInst();
1504 CmpInst *CmpToCheck = cast<CmpInst>(CB.getInstructionToSimplify());
1505 unsigned OtherOpIdx = JoinOp->getOperand(0) == CmpToCheck ? 1 : 0;
1506
1507 // Don't try to simplify the first condition of a select by the second, as
1508 // this may make the select more poisonous than the original one.
1509 // TODO: check if the first operand may be poison.
1510 if (OtherOpIdx != 0 && isa<SelectInst>(JoinOp))
1511 return false;
1512
1513 if (!match(JoinOp->getOperand(OtherOpIdx),
1514 m_ICmp(Pred, m_Value(A), m_Value(B))))
1515 return false;
1516
1517 // For OR, check if the negated condition implies CmpToCheck.
1518 bool IsOr = match(JoinOp, m_LogicalOr());
1519 if (IsOr)
1520 Pred = CmpInst::getInversePredicate(Pred);
1521
1522 // Optimistically add fact from first condition.
1523 unsigned OldSize = DFSInStack.size();
1524 Info.addFact(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack);
1525 if (OldSize == DFSInStack.size())
1526 return false;
1527
1528 bool Changed = false;
1529 // Check if the second condition can be simplified now.
1530 if (auto ImpliedCondition =
1531 checkCondition(CmpToCheck->getPredicate(), CmpToCheck->getOperand(0),
1532 CmpToCheck->getOperand(1), CmpToCheck, Info)) {
1533 if (IsOr && isa<SelectInst>(JoinOp)) {
1534 JoinOp->setOperand(
1535 OtherOpIdx == 0 ? 2 : 0,
1536 ConstantInt::getBool(JoinOp->getType(), *ImpliedCondition));
1537 } else
1538 JoinOp->setOperand(
1539 1 - OtherOpIdx,
1540 ConstantInt::getBool(JoinOp->getType(), *ImpliedCondition));
1541
1542 Changed = true;
1543 }
1544
1545 // Remove entries again.
1546 while (OldSize < DFSInStack.size()) {
1547 StackEntry E = DFSInStack.back();
1548 removeEntryFromStack(E, Info, ReproducerModule, ReproducerCondStack,
1549 DFSInStack);
1550 }
1551 return Changed;
1552 }
1553
addFact(CmpInst::Predicate Pred,Value * A,Value * B,unsigned NumIn,unsigned NumOut,SmallVectorImpl<StackEntry> & DFSInStack)1554 void ConstraintInfo::addFact(CmpInst::Predicate Pred, Value *A, Value *B,
1555 unsigned NumIn, unsigned NumOut,
1556 SmallVectorImpl<StackEntry> &DFSInStack) {
1557 // If the constraint has a pre-condition, skip the constraint if it does not
1558 // hold.
1559 SmallVector<Value *> NewVariables;
1560 auto R = getConstraint(Pred, A, B, NewVariables);
1561
1562 // TODO: Support non-equality for facts as well.
1563 if (!R.isValid(*this) || R.isNe())
1564 return;
1565
1566 LLVM_DEBUG(dbgs() << "Adding '"; dumpUnpackedICmp(dbgs(), Pred, A, B);
1567 dbgs() << "'\n");
1568 bool Added = false;
1569 auto &CSToUse = getCS(R.IsSigned);
1570 if (R.Coefficients.empty())
1571 return;
1572
1573 Added |= CSToUse.addVariableRowFill(R.Coefficients);
1574
1575 // If R has been added to the system, add the new variables and queue it for
1576 // removal once it goes out-of-scope.
1577 if (Added) {
1578 SmallVector<Value *, 2> ValuesToRelease;
1579 auto &Value2Index = getValue2Index(R.IsSigned);
1580 for (Value *V : NewVariables) {
1581 Value2Index.insert({V, Value2Index.size() + 1});
1582 ValuesToRelease.push_back(V);
1583 }
1584
1585 LLVM_DEBUG({
1586 dbgs() << " constraint: ";
1587 dumpConstraint(R.Coefficients, getValue2Index(R.IsSigned));
1588 dbgs() << "\n";
1589 });
1590
1591 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned,
1592 std::move(ValuesToRelease));
1593
1594 if (!R.IsSigned) {
1595 for (Value *V : NewVariables) {
1596 ConstraintTy VarPos(SmallVector<int64_t, 8>(Value2Index.size() + 1, 0),
1597 false, false, false);
1598 VarPos.Coefficients[Value2Index[V]] = -1;
1599 CSToUse.addVariableRow(VarPos.Coefficients);
1600 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned,
1601 SmallVector<Value *, 2>());
1602 }
1603 }
1604
1605 if (R.isEq()) {
1606 // Also add the inverted constraint for equality constraints.
1607 for (auto &Coeff : R.Coefficients)
1608 Coeff *= -1;
1609 CSToUse.addVariableRowFill(R.Coefficients);
1610
1611 DFSInStack.emplace_back(NumIn, NumOut, R.IsSigned,
1612 SmallVector<Value *, 2>());
1613 }
1614 }
1615 }
1616
replaceSubOverflowUses(IntrinsicInst * II,Value * A,Value * B,SmallVectorImpl<Instruction * > & ToRemove)1617 static bool replaceSubOverflowUses(IntrinsicInst *II, Value *A, Value *B,
1618 SmallVectorImpl<Instruction *> &ToRemove) {
1619 bool Changed = false;
1620 IRBuilder<> Builder(II->getParent(), II->getIterator());
1621 Value *Sub = nullptr;
1622 for (User *U : make_early_inc_range(II->users())) {
1623 if (match(U, m_ExtractValue<0>(m_Value()))) {
1624 if (!Sub)
1625 Sub = Builder.CreateSub(A, B);
1626 U->replaceAllUsesWith(Sub);
1627 Changed = true;
1628 } else if (match(U, m_ExtractValue<1>(m_Value()))) {
1629 U->replaceAllUsesWith(Builder.getFalse());
1630 Changed = true;
1631 } else
1632 continue;
1633
1634 if (U->use_empty()) {
1635 auto *I = cast<Instruction>(U);
1636 ToRemove.push_back(I);
1637 I->setOperand(0, PoisonValue::get(II->getType()));
1638 Changed = true;
1639 }
1640 }
1641
1642 if (II->use_empty()) {
1643 II->eraseFromParent();
1644 Changed = true;
1645 }
1646 return Changed;
1647 }
1648
1649 static bool
tryToSimplifyOverflowMath(IntrinsicInst * II,ConstraintInfo & Info,SmallVectorImpl<Instruction * > & ToRemove)1650 tryToSimplifyOverflowMath(IntrinsicInst *II, ConstraintInfo &Info,
1651 SmallVectorImpl<Instruction *> &ToRemove) {
1652 auto DoesConditionHold = [](CmpInst::Predicate Pred, Value *A, Value *B,
1653 ConstraintInfo &Info) {
1654 auto R = Info.getConstraintForSolving(Pred, A, B);
1655 if (R.size() < 2 || !R.isValid(Info))
1656 return false;
1657
1658 auto &CSToUse = Info.getCS(R.IsSigned);
1659 return CSToUse.isConditionImplied(R.Coefficients);
1660 };
1661
1662 bool Changed = false;
1663 if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow) {
1664 // If A s>= B && B s>= 0, ssub.with.overflow(a, b) should not overflow and
1665 // can be simplified to a regular sub.
1666 Value *A = II->getArgOperand(0);
1667 Value *B = II->getArgOperand(1);
1668 if (!DoesConditionHold(CmpInst::ICMP_SGE, A, B, Info) ||
1669 !DoesConditionHold(CmpInst::ICMP_SGE, B,
1670 ConstantInt::get(A->getType(), 0), Info))
1671 return false;
1672 Changed = replaceSubOverflowUses(II, A, B, ToRemove);
1673 }
1674 return Changed;
1675 }
1676
eliminateConstraints(Function & F,DominatorTree & DT,LoopInfo & LI,ScalarEvolution & SE,OptimizationRemarkEmitter & ORE)1677 static bool eliminateConstraints(Function &F, DominatorTree &DT, LoopInfo &LI,
1678 ScalarEvolution &SE,
1679 OptimizationRemarkEmitter &ORE) {
1680 bool Changed = false;
1681 DT.updateDFSNumbers();
1682 SmallVector<Value *> FunctionArgs;
1683 for (Value &Arg : F.args())
1684 FunctionArgs.push_back(&Arg);
1685 ConstraintInfo Info(F.getDataLayout(), FunctionArgs);
1686 State S(DT, LI, SE);
1687 std::unique_ptr<Module> ReproducerModule(
1688 DumpReproducers ? new Module(F.getName(), F.getContext()) : nullptr);
1689
1690 // First, collect conditions implied by branches and blocks with their
1691 // Dominator DFS in and out numbers.
1692 for (BasicBlock &BB : F) {
1693 if (!DT.getNode(&BB))
1694 continue;
1695 S.addInfoFor(BB);
1696 }
1697
1698 // Next, sort worklist by dominance, so that dominating conditions to check
1699 // and facts come before conditions and facts dominated by them. If a
1700 // condition to check and a fact have the same numbers, conditional facts come
1701 // first. Assume facts and checks are ordered according to their relative
1702 // order in the containing basic block. Also make sure conditions with
1703 // constant operands come before conditions without constant operands. This
1704 // increases the effectiveness of the current signed <-> unsigned fact
1705 // transfer logic.
1706 stable_sort(S.WorkList, [](const FactOrCheck &A, const FactOrCheck &B) {
1707 auto HasNoConstOp = [](const FactOrCheck &B) {
1708 Value *V0 = B.isConditionFact() ? B.Cond.Op0 : B.Inst->getOperand(0);
1709 Value *V1 = B.isConditionFact() ? B.Cond.Op1 : B.Inst->getOperand(1);
1710 return !isa<ConstantInt>(V0) && !isa<ConstantInt>(V1);
1711 };
1712 // If both entries have the same In numbers, conditional facts come first.
1713 // Otherwise use the relative order in the basic block.
1714 if (A.NumIn == B.NumIn) {
1715 if (A.isConditionFact() && B.isConditionFact()) {
1716 bool NoConstOpA = HasNoConstOp(A);
1717 bool NoConstOpB = HasNoConstOp(B);
1718 return NoConstOpA < NoConstOpB;
1719 }
1720 if (A.isConditionFact())
1721 return true;
1722 if (B.isConditionFact())
1723 return false;
1724 auto *InstA = A.getContextInst();
1725 auto *InstB = B.getContextInst();
1726 return InstA->comesBefore(InstB);
1727 }
1728 return A.NumIn < B.NumIn;
1729 });
1730
1731 SmallVector<Instruction *> ToRemove;
1732
1733 // Finally, process ordered worklist and eliminate implied conditions.
1734 SmallVector<StackEntry, 16> DFSInStack;
1735 SmallVector<ReproducerEntry> ReproducerCondStack;
1736 for (FactOrCheck &CB : S.WorkList) {
1737 // First, pop entries from the stack that are out-of-scope for CB. Remove
1738 // the corresponding entry from the constraint system.
1739 while (!DFSInStack.empty()) {
1740 auto &E = DFSInStack.back();
1741 LLVM_DEBUG(dbgs() << "Top of stack : " << E.NumIn << " " << E.NumOut
1742 << "\n");
1743 LLVM_DEBUG(dbgs() << "CB: " << CB.NumIn << " " << CB.NumOut << "\n");
1744 assert(E.NumIn <= CB.NumIn);
1745 if (CB.NumOut <= E.NumOut)
1746 break;
1747 LLVM_DEBUG({
1748 dbgs() << "Removing ";
1749 dumpConstraint(Info.getCS(E.IsSigned).getLastConstraint(),
1750 Info.getValue2Index(E.IsSigned));
1751 dbgs() << "\n";
1752 });
1753 removeEntryFromStack(E, Info, ReproducerModule.get(), ReproducerCondStack,
1754 DFSInStack);
1755 }
1756
1757 // For a block, check if any CmpInsts become known based on the current set
1758 // of constraints.
1759 if (CB.isCheck()) {
1760 Instruction *Inst = CB.getInstructionToSimplify();
1761 if (!Inst)
1762 continue;
1763 LLVM_DEBUG(dbgs() << "Processing condition to simplify: " << *Inst
1764 << "\n");
1765 if (auto *II = dyn_cast<WithOverflowInst>(Inst)) {
1766 Changed |= tryToSimplifyOverflowMath(II, Info, ToRemove);
1767 } else if (auto *Cmp = dyn_cast<ICmpInst>(Inst)) {
1768 bool Simplified = checkAndReplaceCondition(
1769 Cmp, Info, CB.NumIn, CB.NumOut, CB.getContextInst(),
1770 ReproducerModule.get(), ReproducerCondStack, S.DT, ToRemove);
1771 if (!Simplified &&
1772 match(CB.getContextInst(), m_LogicalOp(m_Value(), m_Value()))) {
1773 Simplified =
1774 checkOrAndOpImpliedByOther(CB, Info, ReproducerModule.get(),
1775 ReproducerCondStack, DFSInStack);
1776 }
1777 Changed |= Simplified;
1778 } else if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Inst)) {
1779 Changed |= checkAndReplaceMinMax(MinMax, Info, ToRemove);
1780 } else if (auto *CmpIntr = dyn_cast<CmpIntrinsic>(Inst)) {
1781 Changed |= checkAndReplaceCmp(CmpIntr, Info, ToRemove);
1782 }
1783 continue;
1784 }
1785
1786 auto AddFact = [&](CmpInst::Predicate Pred, Value *A, Value *B) {
1787 LLVM_DEBUG(dbgs() << "Processing fact to add to the system: ";
1788 dumpUnpackedICmp(dbgs(), Pred, A, B); dbgs() << "\n");
1789 if (Info.getCS(CmpInst::isSigned(Pred)).size() > MaxRows) {
1790 LLVM_DEBUG(
1791 dbgs()
1792 << "Skip adding constraint because system has too many rows.\n");
1793 return;
1794 }
1795
1796 Info.addFact(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack);
1797 if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size())
1798 ReproducerCondStack.emplace_back(Pred, A, B);
1799
1800 Info.transferToOtherSystem(Pred, A, B, CB.NumIn, CB.NumOut, DFSInStack);
1801 if (ReproducerModule && DFSInStack.size() > ReproducerCondStack.size()) {
1802 // Add dummy entries to ReproducerCondStack to keep it in sync with
1803 // DFSInStack.
1804 for (unsigned I = 0,
1805 E = (DFSInStack.size() - ReproducerCondStack.size());
1806 I < E; ++I) {
1807 ReproducerCondStack.emplace_back(ICmpInst::BAD_ICMP_PREDICATE,
1808 nullptr, nullptr);
1809 }
1810 }
1811 };
1812
1813 ICmpInst::Predicate Pred;
1814 if (!CB.isConditionFact()) {
1815 Value *X;
1816 if (match(CB.Inst, m_Intrinsic<Intrinsic::abs>(m_Value(X)))) {
1817 // If is_int_min_poison is true then we may assume llvm.abs >= 0.
1818 if (cast<ConstantInt>(CB.Inst->getOperand(1))->isOne())
1819 AddFact(CmpInst::ICMP_SGE, CB.Inst,
1820 ConstantInt::get(CB.Inst->getType(), 0));
1821 AddFact(CmpInst::ICMP_SGE, CB.Inst, X);
1822 continue;
1823 }
1824
1825 if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(CB.Inst)) {
1826 Pred = ICmpInst::getNonStrictPredicate(MinMax->getPredicate());
1827 AddFact(Pred, MinMax, MinMax->getLHS());
1828 AddFact(Pred, MinMax, MinMax->getRHS());
1829 continue;
1830 }
1831 }
1832
1833 Value *A = nullptr, *B = nullptr;
1834 if (CB.isConditionFact()) {
1835 Pred = CB.Cond.Pred;
1836 A = CB.Cond.Op0;
1837 B = CB.Cond.Op1;
1838 if (CB.DoesHold.Pred != CmpInst::BAD_ICMP_PREDICATE &&
1839 !Info.doesHold(CB.DoesHold.Pred, CB.DoesHold.Op0, CB.DoesHold.Op1)) {
1840 LLVM_DEBUG({
1841 dbgs() << "Not adding fact ";
1842 dumpUnpackedICmp(dbgs(), Pred, A, B);
1843 dbgs() << " because precondition ";
1844 dumpUnpackedICmp(dbgs(), CB.DoesHold.Pred, CB.DoesHold.Op0,
1845 CB.DoesHold.Op1);
1846 dbgs() << " does not hold.\n";
1847 });
1848 continue;
1849 }
1850 } else {
1851 bool Matched = match(CB.Inst, m_Intrinsic<Intrinsic::assume>(
1852 m_ICmp(Pred, m_Value(A), m_Value(B))));
1853 (void)Matched;
1854 assert(Matched && "Must have an assume intrinsic with a icmp operand");
1855 }
1856 AddFact(Pred, A, B);
1857 }
1858
1859 if (ReproducerModule && !ReproducerModule->functions().empty()) {
1860 std::string S;
1861 raw_string_ostream StringS(S);
1862 ReproducerModule->print(StringS, nullptr);
1863 StringS.flush();
1864 OptimizationRemark Rem(DEBUG_TYPE, "Reproducer", &F);
1865 Rem << ore::NV("module") << S;
1866 ORE.emit(Rem);
1867 }
1868
1869 #ifndef NDEBUG
1870 unsigned SignedEntries =
1871 count_if(DFSInStack, [](const StackEntry &E) { return E.IsSigned; });
1872 assert(Info.getCS(false).size() - FunctionArgs.size() ==
1873 DFSInStack.size() - SignedEntries &&
1874 "updates to CS and DFSInStack are out of sync");
1875 assert(Info.getCS(true).size() == SignedEntries &&
1876 "updates to CS and DFSInStack are out of sync");
1877 #endif
1878
1879 for (Instruction *I : ToRemove)
1880 I->eraseFromParent();
1881 return Changed;
1882 }
1883
run(Function & F,FunctionAnalysisManager & AM)1884 PreservedAnalyses ConstraintEliminationPass::run(Function &F,
1885 FunctionAnalysisManager &AM) {
1886 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1887 auto &LI = AM.getResult<LoopAnalysis>(F);
1888 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1889 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1890 if (!eliminateConstraints(F, DT, LI, SE, ORE))
1891 return PreservedAnalyses::all();
1892
1893 PreservedAnalyses PA;
1894 PA.preserve<DominatorTreeAnalysis>();
1895 PA.preserve<LoopAnalysis>();
1896 PA.preserve<ScalarEvolutionAnalysis>();
1897 PA.preserveSet<CFGAnalyses>();
1898 return PA;
1899 }
1900