1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 // This file contains routines that help analyze properties that chains of
10 // computations have.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Analysis/ValueTracking.h"
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/ADT/iterator_range.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/AssumeBundleQueries.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/EHPersonalities.h"
29 #include "llvm/Analysis/GuardUtils.h"
30 #include "llvm/Analysis/InstructionSimplify.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/VectorUtils.h"
36 #include "llvm/IR/Argument.h"
37 #include "llvm/IR/Attributes.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/GetElementPtrTypeIterator.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/IntrinsicsAArch64.h"
56 #include "llvm/IR/IntrinsicsRISCV.h"
57 #include "llvm/IR/IntrinsicsX86.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Metadata.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/PatternMatch.h"
63 #include "llvm/IR/Type.h"
64 #include "llvm/IR/User.h"
65 #include "llvm/IR/Value.h"
66 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Compiler.h"
69 #include "llvm/Support/ErrorHandling.h"
70 #include "llvm/Support/KnownBits.h"
71 #include "llvm/Support/MathExtras.h"
72 #include <algorithm>
73 #include <cassert>
74 #include <cstdint>
75 #include <optional>
76 #include <utility>
77
78 using namespace llvm;
79 using namespace llvm::PatternMatch;
80
81 // Controls the number of uses of the value searched for possible
82 // dominating comparisons.
83 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
84 cl::Hidden, cl::init(20));
85
86
87 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
88 /// returns the element type's bitwidth.
getBitWidth(Type * Ty,const DataLayout & DL)89 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
90 if (unsigned BitWidth = Ty->getScalarSizeInBits())
91 return BitWidth;
92
93 return DL.getPointerTypeSizeInBits(Ty);
94 }
95
96 namespace {
97
98 // Simplifying using an assume can only be done in a particular control-flow
99 // context (the context instruction provides that context). If an assume and
100 // the context instruction are not in the same block then the DT helps in
101 // figuring out if we can use it.
102 struct Query {
103 const DataLayout &DL;
104 AssumptionCache *AC;
105 const Instruction *CxtI;
106 const DominatorTree *DT;
107
108 // Unlike the other analyses, this may be a nullptr because not all clients
109 // provide it currently.
110 OptimizationRemarkEmitter *ORE;
111
112 /// If true, it is safe to use metadata during simplification.
113 InstrInfoQuery IIQ;
114
Query__anona0331a000111::Query115 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
116 const DominatorTree *DT, bool UseInstrInfo,
117 OptimizationRemarkEmitter *ORE = nullptr)
118 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {}
119 };
120
121 } // end anonymous namespace
122
123 // Given the provided Value and, potentially, a context instruction, return
124 // the preferred context instruction (if any).
safeCxtI(const Value * V,const Instruction * CxtI)125 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
126 // If we've been provided with a context instruction, then use that (provided
127 // it has been inserted).
128 if (CxtI && CxtI->getParent())
129 return CxtI;
130
131 // If the value is really an already-inserted instruction, then use that.
132 CxtI = dyn_cast<Instruction>(V);
133 if (CxtI && CxtI->getParent())
134 return CxtI;
135
136 return nullptr;
137 }
138
safeCxtI(const Value * V1,const Value * V2,const Instruction * CxtI)139 static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) {
140 // If we've been provided with a context instruction, then use that (provided
141 // it has been inserted).
142 if (CxtI && CxtI->getParent())
143 return CxtI;
144
145 // If the value is really an already-inserted instruction, then use that.
146 CxtI = dyn_cast<Instruction>(V1);
147 if (CxtI && CxtI->getParent())
148 return CxtI;
149
150 CxtI = dyn_cast<Instruction>(V2);
151 if (CxtI && CxtI->getParent())
152 return CxtI;
153
154 return nullptr;
155 }
156
getShuffleDemandedElts(const ShuffleVectorInst * Shuf,const APInt & DemandedElts,APInt & DemandedLHS,APInt & DemandedRHS)157 static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf,
158 const APInt &DemandedElts,
159 APInt &DemandedLHS, APInt &DemandedRHS) {
160 if (isa<ScalableVectorType>(Shuf->getType())) {
161 assert(DemandedElts == APInt(1,1));
162 DemandedLHS = DemandedRHS = DemandedElts;
163 return true;
164 }
165
166 int NumElts =
167 cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements();
168 return llvm::getShuffleDemandedElts(NumElts, Shuf->getShuffleMask(),
169 DemandedElts, DemandedLHS, DemandedRHS);
170 }
171
172 static void computeKnownBits(const Value *V, const APInt &DemandedElts,
173 KnownBits &Known, unsigned Depth, const Query &Q);
174
computeKnownBits(const Value * V,KnownBits & Known,unsigned Depth,const Query & Q)175 static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
176 const Query &Q) {
177 // Since the number of lanes in a scalable vector is unknown at compile time,
178 // we track one bit which is implicitly broadcast to all lanes. This means
179 // that all lanes in a scalable vector are considered demanded.
180 auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
181 APInt DemandedElts =
182 FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
183 computeKnownBits(V, DemandedElts, Known, Depth, Q);
184 }
185
computeKnownBits(const Value * V,KnownBits & Known,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE,bool UseInstrInfo)186 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
187 const DataLayout &DL, unsigned Depth,
188 AssumptionCache *AC, const Instruction *CxtI,
189 const DominatorTree *DT,
190 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
191 ::computeKnownBits(V, Known, Depth,
192 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
193 }
194
computeKnownBits(const Value * V,const APInt & DemandedElts,KnownBits & Known,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE,bool UseInstrInfo)195 void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts,
196 KnownBits &Known, const DataLayout &DL,
197 unsigned Depth, AssumptionCache *AC,
198 const Instruction *CxtI, const DominatorTree *DT,
199 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
200 ::computeKnownBits(V, DemandedElts, Known, Depth,
201 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
202 }
203
204 static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts,
205 unsigned Depth, const Query &Q);
206
207 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
208 const Query &Q);
209
computeKnownBits(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE,bool UseInstrInfo)210 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL,
211 unsigned Depth, AssumptionCache *AC,
212 const Instruction *CxtI,
213 const DominatorTree *DT,
214 OptimizationRemarkEmitter *ORE,
215 bool UseInstrInfo) {
216 return ::computeKnownBits(
217 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
218 }
219
computeKnownBits(const Value * V,const APInt & DemandedElts,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE,bool UseInstrInfo)220 KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts,
221 const DataLayout &DL, unsigned Depth,
222 AssumptionCache *AC, const Instruction *CxtI,
223 const DominatorTree *DT,
224 OptimizationRemarkEmitter *ORE,
225 bool UseInstrInfo) {
226 return ::computeKnownBits(
227 V, DemandedElts, Depth,
228 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
229 }
230
haveNoCommonBitsSet(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)231 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
232 const DataLayout &DL, AssumptionCache *AC,
233 const Instruction *CxtI, const DominatorTree *DT,
234 bool UseInstrInfo) {
235 assert(LHS->getType() == RHS->getType() &&
236 "LHS and RHS should have the same type");
237 assert(LHS->getType()->isIntOrIntVectorTy() &&
238 "LHS and RHS should be integers");
239 // Look for an inverted mask: (X & ~M) op (Y & M).
240 {
241 Value *M;
242 if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
243 match(RHS, m_c_And(m_Specific(M), m_Value())))
244 return true;
245 if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
246 match(LHS, m_c_And(m_Specific(M), m_Value())))
247 return true;
248 }
249
250 // X op (Y & ~X)
251 if (match(RHS, m_c_And(m_Not(m_Specific(LHS)), m_Value())) ||
252 match(LHS, m_c_And(m_Not(m_Specific(RHS)), m_Value())))
253 return true;
254
255 // X op ((X & Y) ^ Y) -- this is the canonical form of the previous pattern
256 // for constant Y.
257 Value *Y;
258 if (match(RHS,
259 m_c_Xor(m_c_And(m_Specific(LHS), m_Value(Y)), m_Deferred(Y))) ||
260 match(LHS, m_c_Xor(m_c_And(m_Specific(RHS), m_Value(Y)), m_Deferred(Y))))
261 return true;
262
263 // Peek through extends to find a 'not' of the other side:
264 // (ext Y) op ext(~Y)
265 // (ext ~Y) op ext(Y)
266 if ((match(LHS, m_ZExtOrSExt(m_Value(Y))) &&
267 match(RHS, m_ZExtOrSExt(m_Not(m_Specific(Y))))) ||
268 (match(RHS, m_ZExtOrSExt(m_Value(Y))) &&
269 match(LHS, m_ZExtOrSExt(m_Not(m_Specific(Y))))))
270 return true;
271
272 // Look for: (A & B) op ~(A | B)
273 {
274 Value *A, *B;
275 if (match(LHS, m_And(m_Value(A), m_Value(B))) &&
276 match(RHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
277 return true;
278 if (match(RHS, m_And(m_Value(A), m_Value(B))) &&
279 match(LHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
280 return true;
281 }
282 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
283 KnownBits LHSKnown(IT->getBitWidth());
284 KnownBits RHSKnown(IT->getBitWidth());
285 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
286 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
287 return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown);
288 }
289
isOnlyUsedInZeroEqualityComparison(const Instruction * I)290 bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *I) {
291 return !I->user_empty() && all_of(I->users(), [](const User *U) {
292 ICmpInst::Predicate P;
293 return match(U, m_ICmp(P, m_Value(), m_Zero())) && ICmpInst::isEquality(P);
294 });
295 }
296
297 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
298 const Query &Q);
299
isKnownToBeAPowerOfTwo(const Value * V,const DataLayout & DL,bool OrZero,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)300 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
301 bool OrZero, unsigned Depth,
302 AssumptionCache *AC, const Instruction *CxtI,
303 const DominatorTree *DT, bool UseInstrInfo) {
304 return ::isKnownToBeAPowerOfTwo(
305 V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
306 }
307
308 static bool isKnownNonZero(const Value *V, const APInt &DemandedElts,
309 unsigned Depth, const Query &Q);
310
311 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
312
isKnownNonZero(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)313 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
314 AssumptionCache *AC, const Instruction *CxtI,
315 const DominatorTree *DT, bool UseInstrInfo) {
316 return ::isKnownNonZero(V, Depth,
317 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
318 }
319
isKnownNonNegative(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)320 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
321 unsigned Depth, AssumptionCache *AC,
322 const Instruction *CxtI, const DominatorTree *DT,
323 bool UseInstrInfo) {
324 KnownBits Known =
325 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
326 return Known.isNonNegative();
327 }
328
isKnownPositive(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)329 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
330 AssumptionCache *AC, const Instruction *CxtI,
331 const DominatorTree *DT, bool UseInstrInfo) {
332 if (auto *CI = dyn_cast<ConstantInt>(V))
333 return CI->getValue().isStrictlyPositive();
334
335 // TODO: We'd doing two recursive queries here. We should factor this such
336 // that only a single query is needed.
337 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) &&
338 isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo);
339 }
340
isKnownNegative(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)341 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
342 AssumptionCache *AC, const Instruction *CxtI,
343 const DominatorTree *DT, bool UseInstrInfo) {
344 KnownBits Known =
345 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
346 return Known.isNegative();
347 }
348
349 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth,
350 const Query &Q);
351
isKnownNonEqual(const Value * V1,const Value * V2,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)352 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
353 const DataLayout &DL, AssumptionCache *AC,
354 const Instruction *CxtI, const DominatorTree *DT,
355 bool UseInstrInfo) {
356 return ::isKnownNonEqual(V1, V2, 0,
357 Query(DL, AC, safeCxtI(V2, V1, CxtI), DT,
358 UseInstrInfo, /*ORE=*/nullptr));
359 }
360
361 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
362 const Query &Q);
363
MaskedValueIsZero(const Value * V,const APInt & Mask,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)364 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
365 const DataLayout &DL, unsigned Depth,
366 AssumptionCache *AC, const Instruction *CxtI,
367 const DominatorTree *DT, bool UseInstrInfo) {
368 return ::MaskedValueIsZero(
369 V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
370 }
371
372 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts,
373 unsigned Depth, const Query &Q);
374
ComputeNumSignBits(const Value * V,unsigned Depth,const Query & Q)375 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
376 const Query &Q) {
377 auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
378 APInt DemandedElts =
379 FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
380 return ComputeNumSignBits(V, DemandedElts, Depth, Q);
381 }
382
ComputeNumSignBits(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)383 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
384 unsigned Depth, AssumptionCache *AC,
385 const Instruction *CxtI,
386 const DominatorTree *DT, bool UseInstrInfo) {
387 return ::ComputeNumSignBits(
388 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
389 }
390
ComputeMaxSignificantBits(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)391 unsigned llvm::ComputeMaxSignificantBits(const Value *V, const DataLayout &DL,
392 unsigned Depth, AssumptionCache *AC,
393 const Instruction *CxtI,
394 const DominatorTree *DT) {
395 unsigned SignBits = ComputeNumSignBits(V, DL, Depth, AC, CxtI, DT);
396 return V->getType()->getScalarSizeInBits() - SignBits + 1;
397 }
398
computeKnownBitsAddSub(bool Add,const Value * Op0,const Value * Op1,bool NSW,const APInt & DemandedElts,KnownBits & KnownOut,KnownBits & Known2,unsigned Depth,const Query & Q)399 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
400 bool NSW, const APInt &DemandedElts,
401 KnownBits &KnownOut, KnownBits &Known2,
402 unsigned Depth, const Query &Q) {
403 computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q);
404
405 // If one operand is unknown and we have no nowrap information,
406 // the result will be unknown independently of the second operand.
407 if (KnownOut.isUnknown() && !NSW)
408 return;
409
410 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q);
411 KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut);
412 }
413
computeKnownBitsMul(const Value * Op0,const Value * Op1,bool NSW,const APInt & DemandedElts,KnownBits & Known,KnownBits & Known2,unsigned Depth,const Query & Q)414 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
415 const APInt &DemandedElts, KnownBits &Known,
416 KnownBits &Known2, unsigned Depth,
417 const Query &Q) {
418 computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q);
419 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q);
420
421 bool isKnownNegative = false;
422 bool isKnownNonNegative = false;
423 // If the multiplication is known not to overflow, compute the sign bit.
424 if (NSW) {
425 if (Op0 == Op1) {
426 // The product of a number with itself is non-negative.
427 isKnownNonNegative = true;
428 } else {
429 bool isKnownNonNegativeOp1 = Known.isNonNegative();
430 bool isKnownNonNegativeOp0 = Known2.isNonNegative();
431 bool isKnownNegativeOp1 = Known.isNegative();
432 bool isKnownNegativeOp0 = Known2.isNegative();
433 // The product of two numbers with the same sign is non-negative.
434 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
435 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
436 // The product of a negative number and a non-negative number is either
437 // negative or zero.
438 if (!isKnownNonNegative)
439 isKnownNegative =
440 (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
441 Known2.isNonZero()) ||
442 (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero());
443 }
444 }
445
446 bool SelfMultiply = Op0 == Op1;
447 // TODO: SelfMultiply can be poison, but not undef.
448 if (SelfMultiply)
449 SelfMultiply &=
450 isGuaranteedNotToBeUndefOrPoison(Op0, Q.AC, Q.CxtI, Q.DT, Depth + 1);
451 Known = KnownBits::mul(Known, Known2, SelfMultiply);
452
453 // Only make use of no-wrap flags if we failed to compute the sign bit
454 // directly. This matters if the multiplication always overflows, in
455 // which case we prefer to follow the result of the direct computation,
456 // though as the program is invoking undefined behaviour we can choose
457 // whatever we like here.
458 if (isKnownNonNegative && !Known.isNegative())
459 Known.makeNonNegative();
460 else if (isKnownNegative && !Known.isNonNegative())
461 Known.makeNegative();
462 }
463
computeKnownBitsFromRangeMetadata(const MDNode & Ranges,KnownBits & Known)464 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
465 KnownBits &Known) {
466 unsigned BitWidth = Known.getBitWidth();
467 unsigned NumRanges = Ranges.getNumOperands() / 2;
468 assert(NumRanges >= 1);
469
470 Known.Zero.setAllBits();
471 Known.One.setAllBits();
472
473 for (unsigned i = 0; i < NumRanges; ++i) {
474 ConstantInt *Lower =
475 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
476 ConstantInt *Upper =
477 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
478 ConstantRange Range(Lower->getValue(), Upper->getValue());
479
480 // The first CommonPrefixBits of all values in Range are equal.
481 unsigned CommonPrefixBits =
482 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
483 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
484 APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth);
485 Known.One &= UnsignedMax & Mask;
486 Known.Zero &= ~UnsignedMax & Mask;
487 }
488 }
489
isEphemeralValueOf(const Instruction * I,const Value * E)490 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
491 SmallVector<const Value *, 16> WorkSet(1, I);
492 SmallPtrSet<const Value *, 32> Visited;
493 SmallPtrSet<const Value *, 16> EphValues;
494
495 // The instruction defining an assumption's condition itself is always
496 // considered ephemeral to that assumption (even if it has other
497 // non-ephemeral users). See r246696's test case for an example.
498 if (is_contained(I->operands(), E))
499 return true;
500
501 while (!WorkSet.empty()) {
502 const Value *V = WorkSet.pop_back_val();
503 if (!Visited.insert(V).second)
504 continue;
505
506 // If all uses of this value are ephemeral, then so is this value.
507 if (llvm::all_of(V->users(), [&](const User *U) {
508 return EphValues.count(U);
509 })) {
510 if (V == E)
511 return true;
512
513 if (V == I || (isa<Instruction>(V) &&
514 !cast<Instruction>(V)->mayHaveSideEffects() &&
515 !cast<Instruction>(V)->isTerminator())) {
516 EphValues.insert(V);
517 if (const User *U = dyn_cast<User>(V))
518 append_range(WorkSet, U->operands());
519 }
520 }
521 }
522
523 return false;
524 }
525
526 // Is this an intrinsic that cannot be speculated but also cannot trap?
isAssumeLikeIntrinsic(const Instruction * I)527 bool llvm::isAssumeLikeIntrinsic(const Instruction *I) {
528 if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I))
529 return CI->isAssumeLikeIntrinsic();
530
531 return false;
532 }
533
isValidAssumeForContext(const Instruction * Inv,const Instruction * CxtI,const DominatorTree * DT)534 bool llvm::isValidAssumeForContext(const Instruction *Inv,
535 const Instruction *CxtI,
536 const DominatorTree *DT) {
537 // There are two restrictions on the use of an assume:
538 // 1. The assume must dominate the context (or the control flow must
539 // reach the assume whenever it reaches the context).
540 // 2. The context must not be in the assume's set of ephemeral values
541 // (otherwise we will use the assume to prove that the condition
542 // feeding the assume is trivially true, thus causing the removal of
543 // the assume).
544
545 if (Inv->getParent() == CxtI->getParent()) {
546 // If Inv and CtxI are in the same block, check if the assume (Inv) is first
547 // in the BB.
548 if (Inv->comesBefore(CxtI))
549 return true;
550
551 // Don't let an assume affect itself - this would cause the problems
552 // `isEphemeralValueOf` is trying to prevent, and it would also make
553 // the loop below go out of bounds.
554 if (Inv == CxtI)
555 return false;
556
557 // The context comes first, but they're both in the same block.
558 // Make sure there is nothing in between that might interrupt
559 // the control flow, not even CxtI itself.
560 // We limit the scan distance between the assume and its context instruction
561 // to avoid a compile-time explosion. This limit is chosen arbitrarily, so
562 // it can be adjusted if needed (could be turned into a cl::opt).
563 auto Range = make_range(CxtI->getIterator(), Inv->getIterator());
564 if (!isGuaranteedToTransferExecutionToSuccessor(Range, 15))
565 return false;
566
567 return !isEphemeralValueOf(Inv, CxtI);
568 }
569
570 // Inv and CxtI are in different blocks.
571 if (DT) {
572 if (DT->dominates(Inv, CxtI))
573 return true;
574 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
575 // We don't have a DT, but this trivially dominates.
576 return true;
577 }
578
579 return false;
580 }
581
cmpExcludesZero(CmpInst::Predicate Pred,const Value * RHS)582 static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) {
583 // v u> y implies v != 0.
584 if (Pred == ICmpInst::ICMP_UGT)
585 return true;
586
587 // Special-case v != 0 to also handle v != null.
588 if (Pred == ICmpInst::ICMP_NE)
589 return match(RHS, m_Zero());
590
591 // All other predicates - rely on generic ConstantRange handling.
592 const APInt *C;
593 if (!match(RHS, m_APInt(C)))
594 return false;
595
596 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C);
597 return !TrueValues.contains(APInt::getZero(C->getBitWidth()));
598 }
599
isKnownNonZeroFromAssume(const Value * V,const Query & Q)600 static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) {
601 // Use of assumptions is context-sensitive. If we don't have a context, we
602 // cannot use them!
603 if (!Q.AC || !Q.CxtI)
604 return false;
605
606 if (Q.CxtI && V->getType()->isPointerTy()) {
607 SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull};
608 if (!NullPointerIsDefined(Q.CxtI->getFunction(),
609 V->getType()->getPointerAddressSpace()))
610 AttrKinds.push_back(Attribute::Dereferenceable);
611
612 if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC))
613 return true;
614 }
615
616 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
617 if (!AssumeVH)
618 continue;
619 CondGuardInst *I = cast<CondGuardInst>(AssumeVH);
620 assert(I->getFunction() == Q.CxtI->getFunction() &&
621 "Got assumption for the wrong function!");
622
623 // Warning: This loop can end up being somewhat performance sensitive.
624 // We're running this loop for once for each value queried resulting in a
625 // runtime of ~O(#assumes * #values).
626
627 Value *RHS;
628 CmpInst::Predicate Pred;
629 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
630 if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS))))
631 return false;
632
633 if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT))
634 return true;
635 }
636
637 return false;
638 }
639
computeKnownBitsFromAssume(const Value * V,KnownBits & Known,unsigned Depth,const Query & Q)640 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
641 unsigned Depth, const Query &Q) {
642 // Use of assumptions is context-sensitive. If we don't have a context, we
643 // cannot use them!
644 if (!Q.AC || !Q.CxtI)
645 return;
646
647 unsigned BitWidth = Known.getBitWidth();
648
649 // Refine Known set if the pointer alignment is set by assume bundles.
650 if (V->getType()->isPointerTy()) {
651 if (RetainedKnowledge RK = getKnowledgeValidInContext(
652 V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) {
653 if (isPowerOf2_64(RK.ArgValue))
654 Known.Zero.setLowBits(Log2_64(RK.ArgValue));
655 }
656 }
657
658 // Note that the patterns below need to be kept in sync with the code
659 // in AssumptionCache::updateAffectedValues.
660
661 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
662 if (!AssumeVH)
663 continue;
664 CondGuardInst *I = cast<CondGuardInst>(AssumeVH);
665 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
666 "Got assumption for the wrong function!");
667
668 // Warning: This loop can end up being somewhat performance sensitive.
669 // We're running this loop for once for each value queried resulting in a
670 // runtime of ~O(#assumes * #values).
671
672 Value *Arg = I->getArgOperand(0);
673
674 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
675 assert(BitWidth == 1 && "assume operand is not i1?");
676 Known.setAllOnes();
677 return;
678 }
679 if (match(Arg, m_Not(m_Specific(V))) &&
680 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
681 assert(BitWidth == 1 && "assume operand is not i1?");
682 Known.setAllZero();
683 return;
684 }
685
686 // The remaining tests are all recursive, so bail out if we hit the limit.
687 if (Depth == MaxAnalysisRecursionDepth)
688 continue;
689
690 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg);
691 if (!Cmp)
692 continue;
693
694 // We are attempting to compute known bits for the operands of an assume.
695 // Do not try to use other assumptions for those recursive calls because
696 // that can lead to mutual recursion and a compile-time explosion.
697 // An example of the mutual recursion: computeKnownBits can call
698 // isKnownNonZero which calls computeKnownBitsFromAssume (this function)
699 // and so on.
700 Query QueryNoAC = Q;
701 QueryNoAC.AC = nullptr;
702
703 // Note that ptrtoint may change the bitwidth.
704 Value *A, *B;
705 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
706
707 CmpInst::Predicate Pred;
708 uint64_t C;
709 switch (Cmp->getPredicate()) {
710 default:
711 break;
712 case ICmpInst::ICMP_EQ:
713 // assume(v = a)
714 if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) &&
715 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
716 KnownBits RHSKnown =
717 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
718 Known.Zero |= RHSKnown.Zero;
719 Known.One |= RHSKnown.One;
720 // assume(v & b = a)
721 } else if (match(Cmp,
722 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
723 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
724 KnownBits RHSKnown =
725 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
726 KnownBits MaskKnown =
727 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
728
729 // For those bits in the mask that are known to be one, we can propagate
730 // known bits from the RHS to V.
731 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
732 Known.One |= RHSKnown.One & MaskKnown.One;
733 // assume(~(v & b) = a)
734 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
735 m_Value(A))) &&
736 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
737 KnownBits RHSKnown =
738 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
739 KnownBits MaskKnown =
740 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
741
742 // For those bits in the mask that are known to be one, we can propagate
743 // inverted known bits from the RHS to V.
744 Known.Zero |= RHSKnown.One & MaskKnown.One;
745 Known.One |= RHSKnown.Zero & MaskKnown.One;
746 // assume(v | b = a)
747 } else if (match(Cmp,
748 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
749 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
750 KnownBits RHSKnown =
751 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
752 KnownBits BKnown =
753 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
754
755 // For those bits in B that are known to be zero, we can propagate known
756 // bits from the RHS to V.
757 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
758 Known.One |= RHSKnown.One & BKnown.Zero;
759 // assume(~(v | b) = a)
760 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
761 m_Value(A))) &&
762 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
763 KnownBits RHSKnown =
764 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
765 KnownBits BKnown =
766 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
767
768 // For those bits in B that are known to be zero, we can propagate
769 // inverted known bits from the RHS to V.
770 Known.Zero |= RHSKnown.One & BKnown.Zero;
771 Known.One |= RHSKnown.Zero & BKnown.Zero;
772 // assume(v ^ b = a)
773 } else if (match(Cmp,
774 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
775 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
776 KnownBits RHSKnown =
777 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
778 KnownBits BKnown =
779 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
780
781 // For those bits in B that are known to be zero, we can propagate known
782 // bits from the RHS to V. For those bits in B that are known to be one,
783 // we can propagate inverted known bits from the RHS to V.
784 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
785 Known.One |= RHSKnown.One & BKnown.Zero;
786 Known.Zero |= RHSKnown.One & BKnown.One;
787 Known.One |= RHSKnown.Zero & BKnown.One;
788 // assume(~(v ^ b) = a)
789 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
790 m_Value(A))) &&
791 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
792 KnownBits RHSKnown =
793 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
794 KnownBits BKnown =
795 computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
796
797 // For those bits in B that are known to be zero, we can propagate
798 // inverted known bits from the RHS to V. For those bits in B that are
799 // known to be one, we can propagate known bits from the RHS to V.
800 Known.Zero |= RHSKnown.One & BKnown.Zero;
801 Known.One |= RHSKnown.Zero & BKnown.Zero;
802 Known.Zero |= RHSKnown.Zero & BKnown.One;
803 Known.One |= RHSKnown.One & BKnown.One;
804 // assume(v << c = a)
805 } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
806 m_Value(A))) &&
807 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
808 KnownBits RHSKnown =
809 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
810
811 // For those bits in RHS that are known, we can propagate them to known
812 // bits in V shifted to the right by C.
813 RHSKnown.Zero.lshrInPlace(C);
814 Known.Zero |= RHSKnown.Zero;
815 RHSKnown.One.lshrInPlace(C);
816 Known.One |= RHSKnown.One;
817 // assume(~(v << c) = a)
818 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
819 m_Value(A))) &&
820 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
821 KnownBits RHSKnown =
822 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
823 // For those bits in RHS that are known, we can propagate them inverted
824 // to known bits in V shifted to the right by C.
825 RHSKnown.One.lshrInPlace(C);
826 Known.Zero |= RHSKnown.One;
827 RHSKnown.Zero.lshrInPlace(C);
828 Known.One |= RHSKnown.Zero;
829 // assume(v >> c = a)
830 } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
831 m_Value(A))) &&
832 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
833 KnownBits RHSKnown =
834 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
835 // For those bits in RHS that are known, we can propagate them to known
836 // bits in V shifted to the right by C.
837 Known.Zero |= RHSKnown.Zero << C;
838 Known.One |= RHSKnown.One << C;
839 // assume(~(v >> c) = a)
840 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
841 m_Value(A))) &&
842 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
843 KnownBits RHSKnown =
844 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
845 // For those bits in RHS that are known, we can propagate them inverted
846 // to known bits in V shifted to the right by C.
847 Known.Zero |= RHSKnown.One << C;
848 Known.One |= RHSKnown.Zero << C;
849 }
850 break;
851 case ICmpInst::ICMP_SGE:
852 // assume(v >=_s c) where c is non-negative
853 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
854 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
855 KnownBits RHSKnown =
856 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
857
858 if (RHSKnown.isNonNegative()) {
859 // We know that the sign bit is zero.
860 Known.makeNonNegative();
861 }
862 }
863 break;
864 case ICmpInst::ICMP_SGT:
865 // assume(v >_s c) where c is at least -1.
866 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
867 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
868 KnownBits RHSKnown =
869 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
870
871 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
872 // We know that the sign bit is zero.
873 Known.makeNonNegative();
874 }
875 }
876 break;
877 case ICmpInst::ICMP_SLE:
878 // assume(v <=_s c) where c is negative
879 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
880 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
881 KnownBits RHSKnown =
882 computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth);
883
884 if (RHSKnown.isNegative()) {
885 // We know that the sign bit is one.
886 Known.makeNegative();
887 }
888 }
889 break;
890 case ICmpInst::ICMP_SLT:
891 // assume(v <_s c) where c is non-positive
892 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
893 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
894 KnownBits RHSKnown =
895 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
896
897 if (RHSKnown.isZero() || RHSKnown.isNegative()) {
898 // We know that the sign bit is one.
899 Known.makeNegative();
900 }
901 }
902 break;
903 case ICmpInst::ICMP_ULE:
904 // assume(v <=_u c)
905 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
906 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
907 KnownBits RHSKnown =
908 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
909
910 // Whatever high bits in c are zero are known to be zero.
911 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
912 }
913 break;
914 case ICmpInst::ICMP_ULT:
915 // assume(v <_u c)
916 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
917 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
918 KnownBits RHSKnown =
919 computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth);
920
921 // If the RHS is known zero, then this assumption must be wrong (nothing
922 // is unsigned less than zero). Signal a conflict and get out of here.
923 if (RHSKnown.isZero()) {
924 Known.Zero.setAllBits();
925 Known.One.setAllBits();
926 break;
927 }
928
929 // Whatever high bits in c are zero are known to be zero (if c is a power
930 // of 2, then one more).
931 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC))
932 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
933 else
934 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
935 }
936 break;
937 case ICmpInst::ICMP_NE: {
938 // assume (v & b != 0) where b is a power of 2
939 const APInt *BPow2;
940 if (match(Cmp, m_ICmp(Pred, m_c_And(m_V, m_Power2(BPow2)), m_Zero())) &&
941 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
942 Known.One |= BPow2->zextOrTrunc(BitWidth);
943 }
944 } break;
945 }
946 }
947
948 // If assumptions conflict with each other or previous known bits, then we
949 // have a logical fallacy. It's possible that the assumption is not reachable,
950 // so this isn't a real bug. On the other hand, the program may have undefined
951 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
952 // clear out the known bits, try to warn the user, and hope for the best.
953 if (Known.Zero.intersects(Known.One)) {
954 Known.resetAll();
955
956 if (Q.ORE)
957 Q.ORE->emit([&]() {
958 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
959 return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
960 CxtI)
961 << "Detected conflicting code assumptions. Program may "
962 "have undefined behavior, or compiler may have "
963 "internal error.";
964 });
965 }
966 }
967
968 /// Compute known bits from a shift operator, including those with a
969 /// non-constant shift amount. Known is the output of this function. Known2 is a
970 /// pre-allocated temporary with the same bit width as Known and on return
971 /// contains the known bit of the shift value source. KF is an
972 /// operator-specific function that, given the known-bits and a shift amount,
973 /// compute the implied known-bits of the shift operator's result respectively
974 /// for that shift amount. The results from calling KF are conservatively
975 /// combined for all permitted shift amounts.
computeKnownBitsFromShiftOperator(const Operator * I,const APInt & DemandedElts,KnownBits & Known,KnownBits & Known2,unsigned Depth,const Query & Q,function_ref<KnownBits (const KnownBits &,const KnownBits &)> KF)976 static void computeKnownBitsFromShiftOperator(
977 const Operator *I, const APInt &DemandedElts, KnownBits &Known,
978 KnownBits &Known2, unsigned Depth, const Query &Q,
979 function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) {
980 unsigned BitWidth = Known.getBitWidth();
981 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
982 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
983
984 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
985 // BitWidth > 64 and any upper bits are known, we'll end up returning the
986 // limit value (which implies all bits are known).
987 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
988 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
989 bool ShiftAmtIsConstant = Known.isConstant();
990 bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth);
991
992 if (ShiftAmtIsConstant) {
993 Known = KF(Known2, Known);
994
995 // If the known bits conflict, this must be an overflowing left shift, so
996 // the shift result is poison. We can return anything we want. Choose 0 for
997 // the best folding opportunity.
998 if (Known.hasConflict())
999 Known.setAllZero();
1000
1001 return;
1002 }
1003
1004 // If the shift amount could be greater than or equal to the bit-width of the
1005 // LHS, the value could be poison, but bail out because the check below is
1006 // expensive.
1007 // TODO: Should we just carry on?
1008 if (MaxShiftAmtIsOutOfRange) {
1009 Known.resetAll();
1010 return;
1011 }
1012
1013 // It would be more-clearly correct to use the two temporaries for this
1014 // calculation. Reusing the APInts here to prevent unnecessary allocations.
1015 Known.resetAll();
1016
1017 // If we know the shifter operand is nonzero, we can sometimes infer more
1018 // known bits. However this is expensive to compute, so be lazy about it and
1019 // only compute it when absolutely necessary.
1020 std::optional<bool> ShifterOperandIsNonZero;
1021
1022 // Early exit if we can't constrain any well-defined shift amount.
1023 if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
1024 !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
1025 ShifterOperandIsNonZero =
1026 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q);
1027 if (!*ShifterOperandIsNonZero)
1028 return;
1029 }
1030
1031 Known.Zero.setAllBits();
1032 Known.One.setAllBits();
1033 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1034 // Combine the shifted known input bits only for those shift amounts
1035 // compatible with its known constraints.
1036 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1037 continue;
1038 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1039 continue;
1040 // If we know the shifter is nonzero, we may be able to infer more known
1041 // bits. This check is sunk down as far as possible to avoid the expensive
1042 // call to isKnownNonZero if the cheaper checks above fail.
1043 if (ShiftAmt == 0) {
1044 if (!ShifterOperandIsNonZero)
1045 ShifterOperandIsNonZero =
1046 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q);
1047 if (*ShifterOperandIsNonZero)
1048 continue;
1049 }
1050
1051 Known = KnownBits::commonBits(
1052 Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt))));
1053 }
1054
1055 // If the known bits conflict, the result is poison. Return a 0 and hope the
1056 // caller can further optimize that.
1057 if (Known.hasConflict())
1058 Known.setAllZero();
1059 }
1060
computeKnownBitsFromOperator(const Operator * I,const APInt & DemandedElts,KnownBits & Known,unsigned Depth,const Query & Q)1061 static void computeKnownBitsFromOperator(const Operator *I,
1062 const APInt &DemandedElts,
1063 KnownBits &Known, unsigned Depth,
1064 const Query &Q) {
1065 unsigned BitWidth = Known.getBitWidth();
1066
1067 KnownBits Known2(BitWidth);
1068 switch (I->getOpcode()) {
1069 default: break;
1070 case Instruction::Load:
1071 if (MDNode *MD =
1072 Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range))
1073 computeKnownBitsFromRangeMetadata(*MD, Known);
1074 break;
1075 case Instruction::And: {
1076 // If either the LHS or the RHS are Zero, the result is zero.
1077 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1078 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1079
1080 Known &= Known2;
1081
1082 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
1083 // here we handle the more general case of adding any odd number by
1084 // matching the form add(x, add(x, y)) where y is odd.
1085 // TODO: This could be generalized to clearing any bit set in y where the
1086 // following bit is known to be unset in y.
1087 Value *X = nullptr, *Y = nullptr;
1088 if (!Known.Zero[0] && !Known.One[0] &&
1089 match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
1090 Known2.resetAll();
1091 computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q);
1092 if (Known2.countMinTrailingOnes() > 0)
1093 Known.Zero.setBit(0);
1094 }
1095 break;
1096 }
1097 case Instruction::Or:
1098 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1099 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1100
1101 Known |= Known2;
1102 break;
1103 case Instruction::Xor:
1104 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q);
1105 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1106
1107 Known ^= Known2;
1108 break;
1109 case Instruction::Mul: {
1110 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1111 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts,
1112 Known, Known2, Depth, Q);
1113 break;
1114 }
1115 case Instruction::UDiv: {
1116 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1117 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1118 Known = KnownBits::udiv(Known, Known2);
1119 break;
1120 }
1121 case Instruction::Select: {
1122 const Value *LHS = nullptr, *RHS = nullptr;
1123 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1124 if (SelectPatternResult::isMinOrMax(SPF)) {
1125 computeKnownBits(RHS, Known, Depth + 1, Q);
1126 computeKnownBits(LHS, Known2, Depth + 1, Q);
1127 switch (SPF) {
1128 default:
1129 llvm_unreachable("Unhandled select pattern flavor!");
1130 case SPF_SMAX:
1131 Known = KnownBits::smax(Known, Known2);
1132 break;
1133 case SPF_SMIN:
1134 Known = KnownBits::smin(Known, Known2);
1135 break;
1136 case SPF_UMAX:
1137 Known = KnownBits::umax(Known, Known2);
1138 break;
1139 case SPF_UMIN:
1140 Known = KnownBits::umin(Known, Known2);
1141 break;
1142 }
1143 break;
1144 }
1145
1146 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1147 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1148
1149 // Only known if known in both the LHS and RHS.
1150 Known = KnownBits::commonBits(Known, Known2);
1151
1152 if (SPF == SPF_ABS) {
1153 // RHS from matchSelectPattern returns the negation part of abs pattern.
1154 // If the negate has an NSW flag we can assume the sign bit of the result
1155 // will be 0 because that makes abs(INT_MIN) undefined.
1156 if (match(RHS, m_Neg(m_Specific(LHS))) &&
1157 Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RHS)))
1158 Known.Zero.setSignBit();
1159 }
1160
1161 break;
1162 }
1163 case Instruction::FPTrunc:
1164 case Instruction::FPExt:
1165 case Instruction::FPToUI:
1166 case Instruction::FPToSI:
1167 case Instruction::SIToFP:
1168 case Instruction::UIToFP:
1169 break; // Can't work with floating point.
1170 case Instruction::PtrToInt:
1171 case Instruction::IntToPtr:
1172 // Fall through and handle them the same as zext/trunc.
1173 [[fallthrough]];
1174 case Instruction::ZExt:
1175 case Instruction::Trunc: {
1176 Type *SrcTy = I->getOperand(0)->getType();
1177
1178 unsigned SrcBitWidth;
1179 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1180 // which fall through here.
1181 Type *ScalarTy = SrcTy->getScalarType();
1182 SrcBitWidth = ScalarTy->isPointerTy() ?
1183 Q.DL.getPointerTypeSizeInBits(ScalarTy) :
1184 Q.DL.getTypeSizeInBits(ScalarTy);
1185
1186 assert(SrcBitWidth && "SrcBitWidth can't be zero");
1187 Known = Known.anyextOrTrunc(SrcBitWidth);
1188 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1189 Known = Known.zextOrTrunc(BitWidth);
1190 break;
1191 }
1192 case Instruction::BitCast: {
1193 Type *SrcTy = I->getOperand(0)->getType();
1194 if (SrcTy->isIntOrPtrTy() &&
1195 // TODO: For now, not handling conversions like:
1196 // (bitcast i64 %x to <2 x i32>)
1197 !I->getType()->isVectorTy()) {
1198 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1199 break;
1200 }
1201
1202 // Handle cast from vector integer type to scalar or vector integer.
1203 auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy);
1204 if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() ||
1205 !I->getType()->isIntOrIntVectorTy() ||
1206 isa<ScalableVectorType>(I->getType()))
1207 break;
1208
1209 // Look through a cast from narrow vector elements to wider type.
1210 // Examples: v4i32 -> v2i64, v3i8 -> v24
1211 unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits();
1212 if (BitWidth % SubBitWidth == 0) {
1213 // Known bits are automatically intersected across demanded elements of a
1214 // vector. So for example, if a bit is computed as known zero, it must be
1215 // zero across all demanded elements of the vector.
1216 //
1217 // For this bitcast, each demanded element of the output is sub-divided
1218 // across a set of smaller vector elements in the source vector. To get
1219 // the known bits for an entire element of the output, compute the known
1220 // bits for each sub-element sequentially. This is done by shifting the
1221 // one-set-bit demanded elements parameter across the sub-elements for
1222 // consecutive calls to computeKnownBits. We are using the demanded
1223 // elements parameter as a mask operator.
1224 //
1225 // The known bits of each sub-element are then inserted into place
1226 // (dependent on endian) to form the full result of known bits.
1227 unsigned NumElts = DemandedElts.getBitWidth();
1228 unsigned SubScale = BitWidth / SubBitWidth;
1229 APInt SubDemandedElts = APInt::getZero(NumElts * SubScale);
1230 for (unsigned i = 0; i != NumElts; ++i) {
1231 if (DemandedElts[i])
1232 SubDemandedElts.setBit(i * SubScale);
1233 }
1234
1235 KnownBits KnownSrc(SubBitWidth);
1236 for (unsigned i = 0; i != SubScale; ++i) {
1237 computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc,
1238 Depth + 1, Q);
1239 unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i;
1240 Known.insertBits(KnownSrc, ShiftElt * SubBitWidth);
1241 }
1242 }
1243 break;
1244 }
1245 case Instruction::SExt: {
1246 // Compute the bits in the result that are not present in the input.
1247 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1248
1249 Known = Known.trunc(SrcBitWidth);
1250 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1251 // If the sign bit of the input is known set or clear, then we know the
1252 // top bits of the result.
1253 Known = Known.sext(BitWidth);
1254 break;
1255 }
1256 case Instruction::Shl: {
1257 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1258 auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1259 KnownBits Result = KnownBits::shl(KnownVal, KnownAmt);
1260 // If this shift has "nsw" keyword, then the result is either a poison
1261 // value or has the same sign bit as the first operand.
1262 if (NSW) {
1263 if (KnownVal.Zero.isSignBitSet())
1264 Result.Zero.setSignBit();
1265 if (KnownVal.One.isSignBitSet())
1266 Result.One.setSignBit();
1267 }
1268 return Result;
1269 };
1270 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1271 KF);
1272 // Trailing zeros of a right-shifted constant never decrease.
1273 const APInt *C;
1274 if (match(I->getOperand(0), m_APInt(C)))
1275 Known.Zero.setLowBits(C->countTrailingZeros());
1276 break;
1277 }
1278 case Instruction::LShr: {
1279 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1280 return KnownBits::lshr(KnownVal, KnownAmt);
1281 };
1282 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1283 KF);
1284 // Leading zeros of a left-shifted constant never decrease.
1285 const APInt *C;
1286 if (match(I->getOperand(0), m_APInt(C)))
1287 Known.Zero.setHighBits(C->countLeadingZeros());
1288 break;
1289 }
1290 case Instruction::AShr: {
1291 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) {
1292 return KnownBits::ashr(KnownVal, KnownAmt);
1293 };
1294 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q,
1295 KF);
1296 break;
1297 }
1298 case Instruction::Sub: {
1299 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1300 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1301 DemandedElts, Known, Known2, Depth, Q);
1302 break;
1303 }
1304 case Instruction::Add: {
1305 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1306 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1307 DemandedElts, Known, Known2, Depth, Q);
1308 break;
1309 }
1310 case Instruction::SRem:
1311 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1312 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1313 Known = KnownBits::srem(Known, Known2);
1314 break;
1315
1316 case Instruction::URem:
1317 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1318 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1319 Known = KnownBits::urem(Known, Known2);
1320 break;
1321 case Instruction::Alloca:
1322 Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign()));
1323 break;
1324 case Instruction::GetElementPtr: {
1325 // Analyze all of the subscripts of this getelementptr instruction
1326 // to determine if we can prove known low zero bits.
1327 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1328 // Accumulate the constant indices in a separate variable
1329 // to minimize the number of calls to computeForAddSub.
1330 APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true);
1331
1332 gep_type_iterator GTI = gep_type_begin(I);
1333 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1334 // TrailZ can only become smaller, short-circuit if we hit zero.
1335 if (Known.isUnknown())
1336 break;
1337
1338 Value *Index = I->getOperand(i);
1339
1340 // Handle case when index is zero.
1341 Constant *CIndex = dyn_cast<Constant>(Index);
1342 if (CIndex && CIndex->isZeroValue())
1343 continue;
1344
1345 if (StructType *STy = GTI.getStructTypeOrNull()) {
1346 // Handle struct member offset arithmetic.
1347
1348 assert(CIndex &&
1349 "Access to structure field must be known at compile time");
1350
1351 if (CIndex->getType()->isVectorTy())
1352 Index = CIndex->getSplatValue();
1353
1354 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1355 const StructLayout *SL = Q.DL.getStructLayout(STy);
1356 uint64_t Offset = SL->getElementOffset(Idx);
1357 AccConstIndices += Offset;
1358 continue;
1359 }
1360
1361 // Handle array index arithmetic.
1362 Type *IndexedTy = GTI.getIndexedType();
1363 if (!IndexedTy->isSized()) {
1364 Known.resetAll();
1365 break;
1366 }
1367
1368 unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits();
1369 KnownBits IndexBits(IndexBitWidth);
1370 computeKnownBits(Index, IndexBits, Depth + 1, Q);
1371 TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1372 uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinValue();
1373 KnownBits ScalingFactor(IndexBitWidth);
1374 // Multiply by current sizeof type.
1375 // &A[i] == A + i * sizeof(*A[i]).
1376 if (IndexTypeSize.isScalable()) {
1377 // For scalable types the only thing we know about sizeof is
1378 // that this is a multiple of the minimum size.
1379 ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes));
1380 } else if (IndexBits.isConstant()) {
1381 APInt IndexConst = IndexBits.getConstant();
1382 APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes);
1383 IndexConst *= ScalingFactor;
1384 AccConstIndices += IndexConst.sextOrTrunc(BitWidth);
1385 continue;
1386 } else {
1387 ScalingFactor =
1388 KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes));
1389 }
1390 IndexBits = KnownBits::mul(IndexBits, ScalingFactor);
1391
1392 // If the offsets have a different width from the pointer, according
1393 // to the language reference we need to sign-extend or truncate them
1394 // to the width of the pointer.
1395 IndexBits = IndexBits.sextOrTrunc(BitWidth);
1396
1397 // Note that inbounds does *not* guarantee nsw for the addition, as only
1398 // the offset is signed, while the base address is unsigned.
1399 Known = KnownBits::computeForAddSub(
1400 /*Add=*/true, /*NSW=*/false, Known, IndexBits);
1401 }
1402 if (!Known.isUnknown() && !AccConstIndices.isZero()) {
1403 KnownBits Index = KnownBits::makeConstant(AccConstIndices);
1404 Known = KnownBits::computeForAddSub(
1405 /*Add=*/true, /*NSW=*/false, Known, Index);
1406 }
1407 break;
1408 }
1409 case Instruction::PHI: {
1410 const PHINode *P = cast<PHINode>(I);
1411 BinaryOperator *BO = nullptr;
1412 Value *R = nullptr, *L = nullptr;
1413 if (matchSimpleRecurrence(P, BO, R, L)) {
1414 // Handle the case of a simple two-predecessor recurrence PHI.
1415 // There's a lot more that could theoretically be done here, but
1416 // this is sufficient to catch some interesting cases.
1417 unsigned Opcode = BO->getOpcode();
1418
1419 // If this is a shift recurrence, we know the bits being shifted in.
1420 // We can combine that with information about the start value of the
1421 // recurrence to conclude facts about the result.
1422 if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr ||
1423 Opcode == Instruction::Shl) &&
1424 BO->getOperand(0) == I) {
1425
1426 // We have matched a recurrence of the form:
1427 // %iv = [R, %entry], [%iv.next, %backedge]
1428 // %iv.next = shift_op %iv, L
1429
1430 // Recurse with the phi context to avoid concern about whether facts
1431 // inferred hold at original context instruction. TODO: It may be
1432 // correct to use the original context. IF warranted, explore and
1433 // add sufficient tests to cover.
1434 Query RecQ = Q;
1435 RecQ.CxtI = P;
1436 computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ);
1437 switch (Opcode) {
1438 case Instruction::Shl:
1439 // A shl recurrence will only increase the tailing zeros
1440 Known.Zero.setLowBits(Known2.countMinTrailingZeros());
1441 break;
1442 case Instruction::LShr:
1443 // A lshr recurrence will preserve the leading zeros of the
1444 // start value
1445 Known.Zero.setHighBits(Known2.countMinLeadingZeros());
1446 break;
1447 case Instruction::AShr:
1448 // An ashr recurrence will extend the initial sign bit
1449 Known.Zero.setHighBits(Known2.countMinLeadingZeros());
1450 Known.One.setHighBits(Known2.countMinLeadingOnes());
1451 break;
1452 };
1453 }
1454
1455 // Check for operations that have the property that if
1456 // both their operands have low zero bits, the result
1457 // will have low zero bits.
1458 if (Opcode == Instruction::Add ||
1459 Opcode == Instruction::Sub ||
1460 Opcode == Instruction::And ||
1461 Opcode == Instruction::Or ||
1462 Opcode == Instruction::Mul) {
1463 // Change the context instruction to the "edge" that flows into the
1464 // phi. This is important because that is where the value is actually
1465 // "evaluated" even though it is used later somewhere else. (see also
1466 // D69571).
1467 Query RecQ = Q;
1468
1469 unsigned OpNum = P->getOperand(0) == R ? 0 : 1;
1470 Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator();
1471 Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator();
1472
1473 // Ok, we have a PHI of the form L op= R. Check for low
1474 // zero bits.
1475 RecQ.CxtI = RInst;
1476 computeKnownBits(R, Known2, Depth + 1, RecQ);
1477
1478 // We need to take the minimum number of known bits
1479 KnownBits Known3(BitWidth);
1480 RecQ.CxtI = LInst;
1481 computeKnownBits(L, Known3, Depth + 1, RecQ);
1482
1483 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1484 Known3.countMinTrailingZeros()));
1485
1486 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO);
1487 if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) {
1488 // If initial value of recurrence is nonnegative, and we are adding
1489 // a nonnegative number with nsw, the result can only be nonnegative
1490 // or poison value regardless of the number of times we execute the
1491 // add in phi recurrence. If initial value is negative and we are
1492 // adding a negative number with nsw, the result can only be
1493 // negative or poison value. Similar arguments apply to sub and mul.
1494 //
1495 // (add non-negative, non-negative) --> non-negative
1496 // (add negative, negative) --> negative
1497 if (Opcode == Instruction::Add) {
1498 if (Known2.isNonNegative() && Known3.isNonNegative())
1499 Known.makeNonNegative();
1500 else if (Known2.isNegative() && Known3.isNegative())
1501 Known.makeNegative();
1502 }
1503
1504 // (sub nsw non-negative, negative) --> non-negative
1505 // (sub nsw negative, non-negative) --> negative
1506 else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) {
1507 if (Known2.isNonNegative() && Known3.isNegative())
1508 Known.makeNonNegative();
1509 else if (Known2.isNegative() && Known3.isNonNegative())
1510 Known.makeNegative();
1511 }
1512
1513 // (mul nsw non-negative, non-negative) --> non-negative
1514 else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1515 Known3.isNonNegative())
1516 Known.makeNonNegative();
1517 }
1518
1519 break;
1520 }
1521 }
1522
1523 // Unreachable blocks may have zero-operand PHI nodes.
1524 if (P->getNumIncomingValues() == 0)
1525 break;
1526
1527 // Otherwise take the unions of the known bit sets of the operands,
1528 // taking conservative care to avoid excessive recursion.
1529 if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) {
1530 // Skip if every incoming value references to ourself.
1531 if (isa_and_nonnull<UndefValue>(P->hasConstantValue()))
1532 break;
1533
1534 Known.Zero.setAllBits();
1535 Known.One.setAllBits();
1536 for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) {
1537 Value *IncValue = P->getIncomingValue(u);
1538 // Skip direct self references.
1539 if (IncValue == P) continue;
1540
1541 // Change the context instruction to the "edge" that flows into the
1542 // phi. This is important because that is where the value is actually
1543 // "evaluated" even though it is used later somewhere else. (see also
1544 // D69571).
1545 Query RecQ = Q;
1546 RecQ.CxtI = P->getIncomingBlock(u)->getTerminator();
1547
1548 Known2 = KnownBits(BitWidth);
1549
1550 // Recurse, but cap the recursion to one level, because we don't
1551 // want to waste time spinning around in loops.
1552 computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ);
1553
1554 // If this failed, see if we can use a conditional branch into the phi
1555 // to help us determine the range of the value.
1556 if (Known2.isUnknown()) {
1557 ICmpInst::Predicate Pred;
1558 const APInt *RHSC;
1559 BasicBlock *TrueSucc, *FalseSucc;
1560 // TODO: Use RHS Value and compute range from its known bits.
1561 if (match(RecQ.CxtI,
1562 m_Br(m_c_ICmp(Pred, m_Specific(IncValue), m_APInt(RHSC)),
1563 m_BasicBlock(TrueSucc), m_BasicBlock(FalseSucc)))) {
1564 // Check for cases of duplicate successors.
1565 if ((TrueSucc == P->getParent()) != (FalseSucc == P->getParent())) {
1566 // If we're using the false successor, invert the predicate.
1567 if (FalseSucc == P->getParent())
1568 Pred = CmpInst::getInversePredicate(Pred);
1569
1570 switch (Pred) {
1571 case CmpInst::Predicate::ICMP_EQ:
1572 Known2 = KnownBits::makeConstant(*RHSC);
1573 break;
1574 case CmpInst::Predicate::ICMP_ULE:
1575 Known2.Zero.setHighBits(RHSC->countLeadingZeros());
1576 break;
1577 case CmpInst::Predicate::ICMP_ULT:
1578 Known2.Zero.setHighBits((*RHSC - 1).countLeadingZeros());
1579 break;
1580 default:
1581 // TODO - add additional integer predicate handling.
1582 break;
1583 }
1584 }
1585 }
1586 }
1587
1588 Known = KnownBits::commonBits(Known, Known2);
1589 // If all bits have been ruled out, there's no need to check
1590 // more operands.
1591 if (Known.isUnknown())
1592 break;
1593 }
1594 }
1595 break;
1596 }
1597 case Instruction::Call:
1598 case Instruction::Invoke:
1599 // If range metadata is attached to this call, set known bits from that,
1600 // and then intersect with known bits based on other properties of the
1601 // function.
1602 if (MDNode *MD =
1603 Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range))
1604 computeKnownBitsFromRangeMetadata(*MD, Known);
1605 if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) {
1606 computeKnownBits(RV, Known2, Depth + 1, Q);
1607 Known.Zero |= Known2.Zero;
1608 Known.One |= Known2.One;
1609 }
1610 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1611 switch (II->getIntrinsicID()) {
1612 default: break;
1613 case Intrinsic::abs: {
1614 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1615 bool IntMinIsPoison = match(II->getArgOperand(1), m_One());
1616 Known = Known2.abs(IntMinIsPoison);
1617 break;
1618 }
1619 case Intrinsic::bitreverse:
1620 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1621 Known.Zero |= Known2.Zero.reverseBits();
1622 Known.One |= Known2.One.reverseBits();
1623 break;
1624 case Intrinsic::bswap:
1625 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q);
1626 Known.Zero |= Known2.Zero.byteSwap();
1627 Known.One |= Known2.One.byteSwap();
1628 break;
1629 case Intrinsic::ctlz: {
1630 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1631 // If we have a known 1, its position is our upper bound.
1632 unsigned PossibleLZ = Known2.countMaxLeadingZeros();
1633 // If this call is poison for 0 input, the result will be less than 2^n.
1634 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1635 PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1636 unsigned LowBits = llvm::bit_width(PossibleLZ);
1637 Known.Zero.setBitsFrom(LowBits);
1638 break;
1639 }
1640 case Intrinsic::cttz: {
1641 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1642 // If we have a known 1, its position is our upper bound.
1643 unsigned PossibleTZ = Known2.countMaxTrailingZeros();
1644 // If this call is poison for 0 input, the result will be less than 2^n.
1645 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1646 PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1647 unsigned LowBits = llvm::bit_width(PossibleTZ);
1648 Known.Zero.setBitsFrom(LowBits);
1649 break;
1650 }
1651 case Intrinsic::ctpop: {
1652 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1653 // We can bound the space the count needs. Also, bits known to be zero
1654 // can't contribute to the population.
1655 unsigned BitsPossiblySet = Known2.countMaxPopulation();
1656 unsigned LowBits = llvm::bit_width(BitsPossiblySet);
1657 Known.Zero.setBitsFrom(LowBits);
1658 // TODO: we could bound KnownOne using the lower bound on the number
1659 // of bits which might be set provided by popcnt KnownOne2.
1660 break;
1661 }
1662 case Intrinsic::fshr:
1663 case Intrinsic::fshl: {
1664 const APInt *SA;
1665 if (!match(I->getOperand(2), m_APInt(SA)))
1666 break;
1667
1668 // Normalize to funnel shift left.
1669 uint64_t ShiftAmt = SA->urem(BitWidth);
1670 if (II->getIntrinsicID() == Intrinsic::fshr)
1671 ShiftAmt = BitWidth - ShiftAmt;
1672
1673 KnownBits Known3(BitWidth);
1674 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1675 computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q);
1676
1677 Known.Zero =
1678 Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt);
1679 Known.One =
1680 Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt);
1681 break;
1682 }
1683 case Intrinsic::uadd_sat:
1684 case Intrinsic::usub_sat: {
1685 bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat;
1686 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1687 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1688
1689 // Add: Leading ones of either operand are preserved.
1690 // Sub: Leading zeros of LHS and leading ones of RHS are preserved
1691 // as leading zeros in the result.
1692 unsigned LeadingKnown;
1693 if (IsAdd)
1694 LeadingKnown = std::max(Known.countMinLeadingOnes(),
1695 Known2.countMinLeadingOnes());
1696 else
1697 LeadingKnown = std::max(Known.countMinLeadingZeros(),
1698 Known2.countMinLeadingOnes());
1699
1700 Known = KnownBits::computeForAddSub(
1701 IsAdd, /* NSW */ false, Known, Known2);
1702
1703 // We select between the operation result and all-ones/zero
1704 // respectively, so we can preserve known ones/zeros.
1705 if (IsAdd) {
1706 Known.One.setHighBits(LeadingKnown);
1707 Known.Zero.clearAllBits();
1708 } else {
1709 Known.Zero.setHighBits(LeadingKnown);
1710 Known.One.clearAllBits();
1711 }
1712 break;
1713 }
1714 case Intrinsic::umin:
1715 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1716 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1717 Known = KnownBits::umin(Known, Known2);
1718 break;
1719 case Intrinsic::umax:
1720 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1721 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1722 Known = KnownBits::umax(Known, Known2);
1723 break;
1724 case Intrinsic::smin:
1725 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1726 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1727 Known = KnownBits::smin(Known, Known2);
1728 break;
1729 case Intrinsic::smax:
1730 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1731 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1732 Known = KnownBits::smax(Known, Known2);
1733 break;
1734 case Intrinsic::x86_sse42_crc32_64_64:
1735 Known.Zero.setBitsFrom(32);
1736 break;
1737 case Intrinsic::riscv_vsetvli:
1738 case Intrinsic::riscv_vsetvlimax:
1739 // Assume that VL output is positive and would fit in an int32_t.
1740 // TODO: VLEN might be capped at 16 bits in a future V spec update.
1741 if (BitWidth >= 32)
1742 Known.Zero.setBitsFrom(31);
1743 break;
1744 case Intrinsic::vscale: {
1745 if (!II->getParent() || !II->getFunction() ||
1746 !II->getFunction()->hasFnAttribute(Attribute::VScaleRange))
1747 break;
1748
1749 auto Attr = II->getFunction()->getFnAttribute(Attribute::VScaleRange);
1750 std::optional<unsigned> VScaleMax = Attr.getVScaleRangeMax();
1751
1752 if (!VScaleMax)
1753 break;
1754
1755 unsigned VScaleMin = Attr.getVScaleRangeMin();
1756
1757 // If vscale min = max then we know the exact value at compile time
1758 // and hence we know the exact bits.
1759 if (VScaleMin == VScaleMax) {
1760 Known.One = VScaleMin;
1761 Known.Zero = VScaleMin;
1762 Known.Zero.flipAllBits();
1763 break;
1764 }
1765
1766 unsigned FirstZeroHighBit = llvm::bit_width(*VScaleMax);
1767 if (FirstZeroHighBit < BitWidth)
1768 Known.Zero.setBitsFrom(FirstZeroHighBit);
1769
1770 break;
1771 }
1772 }
1773 }
1774 break;
1775 case Instruction::ShuffleVector: {
1776 auto *Shuf = dyn_cast<ShuffleVectorInst>(I);
1777 // FIXME: Do we need to handle ConstantExpr involving shufflevectors?
1778 if (!Shuf) {
1779 Known.resetAll();
1780 return;
1781 }
1782 // For undef elements, we don't know anything about the common state of
1783 // the shuffle result.
1784 APInt DemandedLHS, DemandedRHS;
1785 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) {
1786 Known.resetAll();
1787 return;
1788 }
1789 Known.One.setAllBits();
1790 Known.Zero.setAllBits();
1791 if (!!DemandedLHS) {
1792 const Value *LHS = Shuf->getOperand(0);
1793 computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q);
1794 // If we don't know any bits, early out.
1795 if (Known.isUnknown())
1796 break;
1797 }
1798 if (!!DemandedRHS) {
1799 const Value *RHS = Shuf->getOperand(1);
1800 computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q);
1801 Known = KnownBits::commonBits(Known, Known2);
1802 }
1803 break;
1804 }
1805 case Instruction::InsertElement: {
1806 if (isa<ScalableVectorType>(I->getType())) {
1807 Known.resetAll();
1808 return;
1809 }
1810 const Value *Vec = I->getOperand(0);
1811 const Value *Elt = I->getOperand(1);
1812 auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2));
1813 // Early out if the index is non-constant or out-of-range.
1814 unsigned NumElts = DemandedElts.getBitWidth();
1815 if (!CIdx || CIdx->getValue().uge(NumElts)) {
1816 Known.resetAll();
1817 return;
1818 }
1819 Known.One.setAllBits();
1820 Known.Zero.setAllBits();
1821 unsigned EltIdx = CIdx->getZExtValue();
1822 // Do we demand the inserted element?
1823 if (DemandedElts[EltIdx]) {
1824 computeKnownBits(Elt, Known, Depth + 1, Q);
1825 // If we don't know any bits, early out.
1826 if (Known.isUnknown())
1827 break;
1828 }
1829 // We don't need the base vector element that has been inserted.
1830 APInt DemandedVecElts = DemandedElts;
1831 DemandedVecElts.clearBit(EltIdx);
1832 if (!!DemandedVecElts) {
1833 computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q);
1834 Known = KnownBits::commonBits(Known, Known2);
1835 }
1836 break;
1837 }
1838 case Instruction::ExtractElement: {
1839 // Look through extract element. If the index is non-constant or
1840 // out-of-range demand all elements, otherwise just the extracted element.
1841 const Value *Vec = I->getOperand(0);
1842 const Value *Idx = I->getOperand(1);
1843 auto *CIdx = dyn_cast<ConstantInt>(Idx);
1844 if (isa<ScalableVectorType>(Vec->getType())) {
1845 // FIXME: there's probably *something* we can do with scalable vectors
1846 Known.resetAll();
1847 break;
1848 }
1849 unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements();
1850 APInt DemandedVecElts = APInt::getAllOnes(NumElts);
1851 if (CIdx && CIdx->getValue().ult(NumElts))
1852 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue());
1853 computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q);
1854 break;
1855 }
1856 case Instruction::ExtractValue:
1857 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1858 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1859 if (EVI->getNumIndices() != 1) break;
1860 if (EVI->getIndices()[0] == 0) {
1861 switch (II->getIntrinsicID()) {
1862 default: break;
1863 case Intrinsic::uadd_with_overflow:
1864 case Intrinsic::sadd_with_overflow:
1865 computeKnownBitsAddSub(true, II->getArgOperand(0),
1866 II->getArgOperand(1), false, DemandedElts,
1867 Known, Known2, Depth, Q);
1868 break;
1869 case Intrinsic::usub_with_overflow:
1870 case Intrinsic::ssub_with_overflow:
1871 computeKnownBitsAddSub(false, II->getArgOperand(0),
1872 II->getArgOperand(1), false, DemandedElts,
1873 Known, Known2, Depth, Q);
1874 break;
1875 case Intrinsic::umul_with_overflow:
1876 case Intrinsic::smul_with_overflow:
1877 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1878 DemandedElts, Known, Known2, Depth, Q);
1879 break;
1880 }
1881 }
1882 }
1883 break;
1884 case Instruction::Freeze:
1885 if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT,
1886 Depth + 1))
1887 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1888 break;
1889 }
1890 }
1891
1892 /// Determine which bits of V are known to be either zero or one and return
1893 /// them.
computeKnownBits(const Value * V,const APInt & DemandedElts,unsigned Depth,const Query & Q)1894 KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts,
1895 unsigned Depth, const Query &Q) {
1896 KnownBits Known(getBitWidth(V->getType(), Q.DL));
1897 computeKnownBits(V, DemandedElts, Known, Depth, Q);
1898 return Known;
1899 }
1900
1901 /// Determine which bits of V are known to be either zero or one and return
1902 /// them.
computeKnownBits(const Value * V,unsigned Depth,const Query & Q)1903 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1904 KnownBits Known(getBitWidth(V->getType(), Q.DL));
1905 computeKnownBits(V, Known, Depth, Q);
1906 return Known;
1907 }
1908
1909 /// Determine which bits of V are known to be either zero or one and return
1910 /// them in the Known bit set.
1911 ///
1912 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1913 /// we cannot optimize based on the assumption that it is zero without changing
1914 /// it to be an explicit zero. If we don't change it to zero, other code could
1915 /// optimized based on the contradictory assumption that it is non-zero.
1916 /// Because instcombine aggressively folds operations with undef args anyway,
1917 /// this won't lose us code quality.
1918 ///
1919 /// This function is defined on values with integer type, values with pointer
1920 /// type, and vectors of integers. In the case
1921 /// where V is a vector, known zero, and known one values are the
1922 /// same width as the vector element, and the bit is set only if it is true
1923 /// for all of the demanded elements in the vector specified by DemandedElts.
computeKnownBits(const Value * V,const APInt & DemandedElts,KnownBits & Known,unsigned Depth,const Query & Q)1924 void computeKnownBits(const Value *V, const APInt &DemandedElts,
1925 KnownBits &Known, unsigned Depth, const Query &Q) {
1926 if (!DemandedElts) {
1927 // No demanded elts, better to assume we don't know anything.
1928 Known.resetAll();
1929 return;
1930 }
1931
1932 assert(V && "No Value?");
1933 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
1934
1935 #ifndef NDEBUG
1936 Type *Ty = V->getType();
1937 unsigned BitWidth = Known.getBitWidth();
1938
1939 assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) &&
1940 "Not integer or pointer type!");
1941
1942 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {
1943 assert(
1944 FVTy->getNumElements() == DemandedElts.getBitWidth() &&
1945 "DemandedElt width should equal the fixed vector number of elements");
1946 } else {
1947 assert(DemandedElts == APInt(1, 1) &&
1948 "DemandedElt width should be 1 for scalars or scalable vectors");
1949 }
1950
1951 Type *ScalarTy = Ty->getScalarType();
1952 if (ScalarTy->isPointerTy()) {
1953 assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) &&
1954 "V and Known should have same BitWidth");
1955 } else {
1956 assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) &&
1957 "V and Known should have same BitWidth");
1958 }
1959 #endif
1960
1961 const APInt *C;
1962 if (match(V, m_APInt(C))) {
1963 // We know all of the bits for a scalar constant or a splat vector constant!
1964 Known = KnownBits::makeConstant(*C);
1965 return;
1966 }
1967 // Null and aggregate-zero are all-zeros.
1968 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1969 Known.setAllZero();
1970 return;
1971 }
1972 // Handle a constant vector by taking the intersection of the known bits of
1973 // each element.
1974 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) {
1975 assert(!isa<ScalableVectorType>(V->getType()));
1976 // We know that CDV must be a vector of integers. Take the intersection of
1977 // each element.
1978 Known.Zero.setAllBits(); Known.One.setAllBits();
1979 for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) {
1980 if (!DemandedElts[i])
1981 continue;
1982 APInt Elt = CDV->getElementAsAPInt(i);
1983 Known.Zero &= ~Elt;
1984 Known.One &= Elt;
1985 }
1986 return;
1987 }
1988
1989 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1990 assert(!isa<ScalableVectorType>(V->getType()));
1991 // We know that CV must be a vector of integers. Take the intersection of
1992 // each element.
1993 Known.Zero.setAllBits(); Known.One.setAllBits();
1994 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1995 if (!DemandedElts[i])
1996 continue;
1997 Constant *Element = CV->getAggregateElement(i);
1998 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1999 if (!ElementCI) {
2000 Known.resetAll();
2001 return;
2002 }
2003 const APInt &Elt = ElementCI->getValue();
2004 Known.Zero &= ~Elt;
2005 Known.One &= Elt;
2006 }
2007 return;
2008 }
2009
2010 // Start out not knowing anything.
2011 Known.resetAll();
2012
2013 // We can't imply anything about undefs.
2014 if (isa<UndefValue>(V))
2015 return;
2016
2017 // There's no point in looking through other users of ConstantData for
2018 // assumptions. Confirm that we've handled them all.
2019 assert(!isa<ConstantData>(V) && "Unhandled constant data!");
2020
2021 // All recursive calls that increase depth must come after this.
2022 if (Depth == MaxAnalysisRecursionDepth)
2023 return;
2024
2025 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
2026 // the bits of its aliasee.
2027 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
2028 if (!GA->isInterposable())
2029 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
2030 return;
2031 }
2032
2033 if (const Operator *I = dyn_cast<Operator>(V))
2034 computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q);
2035
2036 // Aligned pointers have trailing zeros - refine Known.Zero set
2037 if (isa<PointerType>(V->getType())) {
2038 Align Alignment = V->getPointerAlignment(Q.DL);
2039 Known.Zero.setLowBits(Log2(Alignment));
2040 }
2041
2042 // computeKnownBitsFromAssume strictly refines Known.
2043 // Therefore, we run them after computeKnownBitsFromOperator.
2044
2045 // Check whether a nearby assume intrinsic can determine some known bits.
2046 computeKnownBitsFromAssume(V, Known, Depth, Q);
2047
2048 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
2049 }
2050
2051 /// Try to detect a recurrence that the value of the induction variable is
2052 /// always a power of two (or zero).
isPowerOfTwoRecurrence(const PHINode * PN,bool OrZero,unsigned Depth,Query & Q)2053 static bool isPowerOfTwoRecurrence(const PHINode *PN, bool OrZero,
2054 unsigned Depth, Query &Q) {
2055 BinaryOperator *BO = nullptr;
2056 Value *Start = nullptr, *Step = nullptr;
2057 if (!matchSimpleRecurrence(PN, BO, Start, Step))
2058 return false;
2059
2060 // Initial value must be a power of two.
2061 for (const Use &U : PN->operands()) {
2062 if (U.get() == Start) {
2063 // Initial value comes from a different BB, need to adjust context
2064 // instruction for analysis.
2065 Q.CxtI = PN->getIncomingBlock(U)->getTerminator();
2066 if (!isKnownToBeAPowerOfTwo(Start, OrZero, Depth, Q))
2067 return false;
2068 }
2069 }
2070
2071 // Except for Mul, the induction variable must be on the left side of the
2072 // increment expression, otherwise its value can be arbitrary.
2073 if (BO->getOpcode() != Instruction::Mul && BO->getOperand(1) != Step)
2074 return false;
2075
2076 Q.CxtI = BO->getParent()->getTerminator();
2077 switch (BO->getOpcode()) {
2078 case Instruction::Mul:
2079 // Power of two is closed under multiplication.
2080 return (OrZero || Q.IIQ.hasNoUnsignedWrap(BO) ||
2081 Q.IIQ.hasNoSignedWrap(BO)) &&
2082 isKnownToBeAPowerOfTwo(Step, OrZero, Depth, Q);
2083 case Instruction::SDiv:
2084 // Start value must not be signmask for signed division, so simply being a
2085 // power of two is not sufficient, and it has to be a constant.
2086 if (!match(Start, m_Power2()) || match(Start, m_SignMask()))
2087 return false;
2088 [[fallthrough]];
2089 case Instruction::UDiv:
2090 // Divisor must be a power of two.
2091 // If OrZero is false, cannot guarantee induction variable is non-zero after
2092 // division, same for Shr, unless it is exact division.
2093 return (OrZero || Q.IIQ.isExact(BO)) &&
2094 isKnownToBeAPowerOfTwo(Step, false, Depth, Q);
2095 case Instruction::Shl:
2096 return OrZero || Q.IIQ.hasNoUnsignedWrap(BO) || Q.IIQ.hasNoSignedWrap(BO);
2097 case Instruction::AShr:
2098 if (!match(Start, m_Power2()) || match(Start, m_SignMask()))
2099 return false;
2100 [[fallthrough]];
2101 case Instruction::LShr:
2102 return OrZero || Q.IIQ.isExact(BO);
2103 default:
2104 return false;
2105 }
2106 }
2107
2108 /// Return true if the given value is known to have exactly one
2109 /// bit set when defined. For vectors return true if every element is known to
2110 /// be a power of two when defined. Supports values with integer or pointer
2111 /// types and vectors of integers.
isKnownToBeAPowerOfTwo(const Value * V,bool OrZero,unsigned Depth,const Query & Q)2112 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
2113 const Query &Q) {
2114 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
2115
2116 // Attempt to match against constants.
2117 if (OrZero && match(V, m_Power2OrZero()))
2118 return true;
2119 if (match(V, m_Power2()))
2120 return true;
2121
2122 // 1 << X is clearly a power of two if the one is not shifted off the end. If
2123 // it is shifted off the end then the result is undefined.
2124 if (match(V, m_Shl(m_One(), m_Value())))
2125 return true;
2126
2127 // (signmask) >>l X is clearly a power of two if the one is not shifted off
2128 // the bottom. If it is shifted off the bottom then the result is undefined.
2129 if (match(V, m_LShr(m_SignMask(), m_Value())))
2130 return true;
2131
2132 // The remaining tests are all recursive, so bail out if we hit the limit.
2133 if (Depth++ == MaxAnalysisRecursionDepth)
2134 return false;
2135
2136 Value *X = nullptr, *Y = nullptr;
2137 // A shift left or a logical shift right of a power of two is a power of two
2138 // or zero.
2139 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
2140 match(V, m_LShr(m_Value(X), m_Value()))))
2141 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
2142
2143 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
2144 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
2145
2146 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
2147 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
2148 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
2149
2150 // Peek through min/max.
2151 if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) {
2152 return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) &&
2153 isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q);
2154 }
2155
2156 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
2157 // A power of two and'd with anything is a power of two or zero.
2158 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
2159 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
2160 return true;
2161 // X & (-X) is always a power of two or zero.
2162 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
2163 return true;
2164 return false;
2165 }
2166
2167 // Adding a power-of-two or zero to the same power-of-two or zero yields
2168 // either the original power-of-two, a larger power-of-two or zero.
2169 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2170 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
2171 if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
2172 Q.IIQ.hasNoSignedWrap(VOBO)) {
2173 if (match(X, m_And(m_Specific(Y), m_Value())) ||
2174 match(X, m_And(m_Value(), m_Specific(Y))))
2175 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
2176 return true;
2177 if (match(Y, m_And(m_Specific(X), m_Value())) ||
2178 match(Y, m_And(m_Value(), m_Specific(X))))
2179 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
2180 return true;
2181
2182 unsigned BitWidth = V->getType()->getScalarSizeInBits();
2183 KnownBits LHSBits(BitWidth);
2184 computeKnownBits(X, LHSBits, Depth, Q);
2185
2186 KnownBits RHSBits(BitWidth);
2187 computeKnownBits(Y, RHSBits, Depth, Q);
2188 // If i8 V is a power of two or zero:
2189 // ZeroBits: 1 1 1 0 1 1 1 1
2190 // ~ZeroBits: 0 0 0 1 0 0 0 0
2191 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
2192 // If OrZero isn't set, we cannot give back a zero result.
2193 // Make sure either the LHS or RHS has a bit set.
2194 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
2195 return true;
2196 }
2197 }
2198
2199 // A PHI node is power of two if all incoming values are power of two, or if
2200 // it is an induction variable where in each step its value is a power of two.
2201 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2202 Query RecQ = Q;
2203
2204 // Check if it is an induction variable and always power of two.
2205 if (isPowerOfTwoRecurrence(PN, OrZero, Depth, RecQ))
2206 return true;
2207
2208 // Recursively check all incoming values. Limit recursion to 2 levels, so
2209 // that search complexity is limited to number of operands^2.
2210 unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1);
2211 return llvm::all_of(PN->operands(), [&](const Use &U) {
2212 // Value is power of 2 if it is coming from PHI node itself by induction.
2213 if (U.get() == PN)
2214 return true;
2215
2216 // Change the context instruction to the incoming block where it is
2217 // evaluated.
2218 RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator();
2219 return isKnownToBeAPowerOfTwo(U.get(), OrZero, NewDepth, RecQ);
2220 });
2221 }
2222
2223 // An exact divide or right shift can only shift off zero bits, so the result
2224 // is a power of two only if the first operand is a power of two and not
2225 // copying a sign bit (sdiv int_min, 2).
2226 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
2227 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
2228 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
2229 Depth, Q);
2230 }
2231
2232 return false;
2233 }
2234
2235 /// Test whether a GEP's result is known to be non-null.
2236 ///
2237 /// Uses properties inherent in a GEP to try to determine whether it is known
2238 /// to be non-null.
2239 ///
2240 /// Currently this routine does not support vector GEPs.
isGEPKnownNonNull(const GEPOperator * GEP,unsigned Depth,const Query & Q)2241 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
2242 const Query &Q) {
2243 const Function *F = nullptr;
2244 if (const Instruction *I = dyn_cast<Instruction>(GEP))
2245 F = I->getFunction();
2246
2247 if (!GEP->isInBounds() ||
2248 NullPointerIsDefined(F, GEP->getPointerAddressSpace()))
2249 return false;
2250
2251 // FIXME: Support vector-GEPs.
2252 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
2253
2254 // If the base pointer is non-null, we cannot walk to a null address with an
2255 // inbounds GEP in address space zero.
2256 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
2257 return true;
2258
2259 // Walk the GEP operands and see if any operand introduces a non-zero offset.
2260 // If so, then the GEP cannot produce a null pointer, as doing so would
2261 // inherently violate the inbounds contract within address space zero.
2262 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
2263 GTI != GTE; ++GTI) {
2264 // Struct types are easy -- they must always be indexed by a constant.
2265 if (StructType *STy = GTI.getStructTypeOrNull()) {
2266 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
2267 unsigned ElementIdx = OpC->getZExtValue();
2268 const StructLayout *SL = Q.DL.getStructLayout(STy);
2269 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
2270 if (ElementOffset > 0)
2271 return true;
2272 continue;
2273 }
2274
2275 // If we have a zero-sized type, the index doesn't matter. Keep looping.
2276 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).isZero())
2277 continue;
2278
2279 // Fast path the constant operand case both for efficiency and so we don't
2280 // increment Depth when just zipping down an all-constant GEP.
2281 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
2282 if (!OpC->isZero())
2283 return true;
2284 continue;
2285 }
2286
2287 // We post-increment Depth here because while isKnownNonZero increments it
2288 // as well, when we pop back up that increment won't persist. We don't want
2289 // to recurse 10k times just because we have 10k GEP operands. We don't
2290 // bail completely out because we want to handle constant GEPs regardless
2291 // of depth.
2292 if (Depth++ >= MaxAnalysisRecursionDepth)
2293 continue;
2294
2295 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
2296 return true;
2297 }
2298
2299 return false;
2300 }
2301
isKnownNonNullFromDominatingCondition(const Value * V,const Instruction * CtxI,const DominatorTree * DT)2302 static bool isKnownNonNullFromDominatingCondition(const Value *V,
2303 const Instruction *CtxI,
2304 const DominatorTree *DT) {
2305 if (isa<Constant>(V))
2306 return false;
2307
2308 if (!CtxI || !DT)
2309 return false;
2310
2311 unsigned NumUsesExplored = 0;
2312 for (const auto *U : V->users()) {
2313 // Avoid massive lists
2314 if (NumUsesExplored >= DomConditionsMaxUses)
2315 break;
2316 NumUsesExplored++;
2317
2318 // If the value is used as an argument to a call or invoke, then argument
2319 // attributes may provide an answer about null-ness.
2320 if (const auto *CB = dyn_cast<CallBase>(U))
2321 if (auto *CalledFunc = CB->getCalledFunction())
2322 for (const Argument &Arg : CalledFunc->args())
2323 if (CB->getArgOperand(Arg.getArgNo()) == V &&
2324 Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) &&
2325 DT->dominates(CB, CtxI))
2326 return true;
2327
2328 // If the value is used as a load/store, then the pointer must be non null.
2329 if (V == getLoadStorePointerOperand(U)) {
2330 const Instruction *I = cast<Instruction>(U);
2331 if (!NullPointerIsDefined(I->getFunction(),
2332 V->getType()->getPointerAddressSpace()) &&
2333 DT->dominates(I, CtxI))
2334 return true;
2335 }
2336
2337 // Consider only compare instructions uniquely controlling a branch
2338 Value *RHS;
2339 CmpInst::Predicate Pred;
2340 if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS))))
2341 continue;
2342
2343 bool NonNullIfTrue;
2344 if (cmpExcludesZero(Pred, RHS))
2345 NonNullIfTrue = true;
2346 else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS))
2347 NonNullIfTrue = false;
2348 else
2349 continue;
2350
2351 SmallVector<const User *, 4> WorkList;
2352 SmallPtrSet<const User *, 4> Visited;
2353 for (const auto *CmpU : U->users()) {
2354 assert(WorkList.empty() && "Should be!");
2355 if (Visited.insert(CmpU).second)
2356 WorkList.push_back(CmpU);
2357
2358 while (!WorkList.empty()) {
2359 auto *Curr = WorkList.pop_back_val();
2360
2361 // If a user is an AND, add all its users to the work list. We only
2362 // propagate "pred != null" condition through AND because it is only
2363 // correct to assume that all conditions of AND are met in true branch.
2364 // TODO: Support similar logic of OR and EQ predicate?
2365 if (NonNullIfTrue)
2366 if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) {
2367 for (const auto *CurrU : Curr->users())
2368 if (Visited.insert(CurrU).second)
2369 WorkList.push_back(CurrU);
2370 continue;
2371 }
2372
2373 if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
2374 assert(BI->isConditional() && "uses a comparison!");
2375
2376 BasicBlock *NonNullSuccessor =
2377 BI->getSuccessor(NonNullIfTrue ? 0 : 1);
2378 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
2379 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
2380 return true;
2381 } else if (NonNullIfTrue && isGuard(Curr) &&
2382 DT->dominates(cast<Instruction>(Curr), CtxI)) {
2383 return true;
2384 }
2385 }
2386 }
2387 }
2388
2389 return false;
2390 }
2391
2392 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
2393 /// ensure that the value it's attached to is never Value? 'RangeType' is
2394 /// is the type of the value described by the range.
rangeMetadataExcludesValue(const MDNode * Ranges,const APInt & Value)2395 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
2396 const unsigned NumRanges = Ranges->getNumOperands() / 2;
2397 assert(NumRanges >= 1);
2398 for (unsigned i = 0; i < NumRanges; ++i) {
2399 ConstantInt *Lower =
2400 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
2401 ConstantInt *Upper =
2402 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
2403 ConstantRange Range(Lower->getValue(), Upper->getValue());
2404 if (Range.contains(Value))
2405 return false;
2406 }
2407 return true;
2408 }
2409
2410 /// Try to detect a recurrence that monotonically increases/decreases from a
2411 /// non-zero starting value. These are common as induction variables.
isNonZeroRecurrence(const PHINode * PN)2412 static bool isNonZeroRecurrence(const PHINode *PN) {
2413 BinaryOperator *BO = nullptr;
2414 Value *Start = nullptr, *Step = nullptr;
2415 const APInt *StartC, *StepC;
2416 if (!matchSimpleRecurrence(PN, BO, Start, Step) ||
2417 !match(Start, m_APInt(StartC)) || StartC->isZero())
2418 return false;
2419
2420 switch (BO->getOpcode()) {
2421 case Instruction::Add:
2422 // Starting from non-zero and stepping away from zero can never wrap back
2423 // to zero.
2424 return BO->hasNoUnsignedWrap() ||
2425 (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) &&
2426 StartC->isNegative() == StepC->isNegative());
2427 case Instruction::Mul:
2428 return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) &&
2429 match(Step, m_APInt(StepC)) && !StepC->isZero();
2430 case Instruction::Shl:
2431 return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap();
2432 case Instruction::AShr:
2433 case Instruction::LShr:
2434 return BO->isExact();
2435 default:
2436 return false;
2437 }
2438 }
2439
2440 /// Return true if the given value is known to be non-zero when defined. For
2441 /// vectors, return true if every demanded element is known to be non-zero when
2442 /// defined. For pointers, if the context instruction and dominator tree are
2443 /// specified, perform context-sensitive analysis and return true if the
2444 /// pointer couldn't possibly be null at the specified instruction.
2445 /// Supports values with integer or pointer type and vectors of integers.
isKnownNonZero(const Value * V,const APInt & DemandedElts,unsigned Depth,const Query & Q)2446 bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth,
2447 const Query &Q) {
2448
2449 #ifndef NDEBUG
2450 Type *Ty = V->getType();
2451 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
2452
2453 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {
2454 assert(
2455 FVTy->getNumElements() == DemandedElts.getBitWidth() &&
2456 "DemandedElt width should equal the fixed vector number of elements");
2457 } else {
2458 assert(DemandedElts == APInt(1, 1) &&
2459 "DemandedElt width should be 1 for scalars");
2460 }
2461 #endif
2462
2463 if (auto *C = dyn_cast<Constant>(V)) {
2464 if (C->isNullValue())
2465 return false;
2466 if (isa<ConstantInt>(C))
2467 // Must be non-zero due to null test above.
2468 return true;
2469
2470 // For constant vectors, check that all elements are undefined or known
2471 // non-zero to determine that the whole vector is known non-zero.
2472 if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) {
2473 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
2474 if (!DemandedElts[i])
2475 continue;
2476 Constant *Elt = C->getAggregateElement(i);
2477 if (!Elt || Elt->isNullValue())
2478 return false;
2479 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
2480 return false;
2481 }
2482 return true;
2483 }
2484
2485 // A global variable in address space 0 is non null unless extern weak
2486 // or an absolute symbol reference. Other address spaces may have null as a
2487 // valid address for a global, so we can't assume anything.
2488 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
2489 if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
2490 GV->getType()->getAddressSpace() == 0)
2491 return true;
2492 }
2493
2494 // For constant expressions, fall through to the Operator code below.
2495 if (!isa<ConstantExpr>(V))
2496 return false;
2497 }
2498
2499 if (auto *I = dyn_cast<Instruction>(V)) {
2500 if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
2501 // If the possible ranges don't contain zero, then the value is
2502 // definitely non-zero.
2503 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
2504 const APInt ZeroValue(Ty->getBitWidth(), 0);
2505 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
2506 return true;
2507 }
2508 }
2509 }
2510
2511 if (!isa<Constant>(V) && isKnownNonZeroFromAssume(V, Q))
2512 return true;
2513
2514 // Some of the tests below are recursive, so bail out if we hit the limit.
2515 if (Depth++ >= MaxAnalysisRecursionDepth)
2516 return false;
2517
2518 // Check for pointer simplifications.
2519
2520 if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) {
2521 // Alloca never returns null, malloc might.
2522 if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
2523 return true;
2524
2525 // A byval, inalloca may not be null in a non-default addres space. A
2526 // nonnull argument is assumed never 0.
2527 if (const Argument *A = dyn_cast<Argument>(V)) {
2528 if (((A->hasPassPointeeByValueCopyAttr() &&
2529 !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) ||
2530 A->hasNonNullAttr()))
2531 return true;
2532 }
2533
2534 // A Load tagged with nonnull metadata is never null.
2535 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2536 if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2537 return true;
2538
2539 if (const auto *Call = dyn_cast<CallBase>(V)) {
2540 if (Call->isReturnNonNull())
2541 return true;
2542 if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true))
2543 return isKnownNonZero(RP, Depth, Q);
2544 }
2545 }
2546
2547 if (!isa<Constant>(V) &&
2548 isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2549 return true;
2550
2551 const Operator *I = dyn_cast<Operator>(V);
2552 if (!I)
2553 return false;
2554
2555 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2556 switch (I->getOpcode()) {
2557 case Instruction::GetElementPtr:
2558 if (I->getType()->isPointerTy())
2559 return isGEPKnownNonNull(cast<GEPOperator>(I), Depth, Q);
2560 break;
2561 case Instruction::BitCast:
2562 if (I->getType()->isPointerTy())
2563 return isKnownNonZero(I->getOperand(0), Depth, Q);
2564 break;
2565 case Instruction::IntToPtr:
2566 // Note that we have to take special care to avoid looking through
2567 // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well
2568 // as casts that can alter the value, e.g., AddrSpaceCasts.
2569 if (!isa<ScalableVectorType>(I->getType()) &&
2570 Q.DL.getTypeSizeInBits(I->getOperand(0)->getType()).getFixedValue() <=
2571 Q.DL.getTypeSizeInBits(I->getType()).getFixedValue())
2572 return isKnownNonZero(I->getOperand(0), Depth, Q);
2573 break;
2574 case Instruction::PtrToInt:
2575 // Similar to int2ptr above, we can look through ptr2int here if the cast
2576 // is a no-op or an extend and not a truncate.
2577 if (!isa<ScalableVectorType>(I->getType()) &&
2578 Q.DL.getTypeSizeInBits(I->getOperand(0)->getType()).getFixedValue() <=
2579 Q.DL.getTypeSizeInBits(I->getType()).getFixedValue())
2580 return isKnownNonZero(I->getOperand(0), Depth, Q);
2581 break;
2582 case Instruction::Or:
2583 // X | Y != 0 if X != 0 or Y != 0.
2584 return isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q) ||
2585 isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q);
2586 case Instruction::SExt:
2587 case Instruction::ZExt:
2588 // ext X != 0 if X != 0.
2589 return isKnownNonZero(I->getOperand(0), Depth, Q);
2590
2591 case Instruction::Shl: {
2592 // shl nuw can't remove any non-zero bits.
2593 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2594 if (Q.IIQ.hasNoUnsignedWrap(BO))
2595 return isKnownNonZero(I->getOperand(0), Depth, Q);
2596
2597 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2598 // if the lowest bit is shifted off the end.
2599 KnownBits Known(BitWidth);
2600 computeKnownBits(I->getOperand(0), DemandedElts, Known, Depth, Q);
2601 if (Known.One[0])
2602 return true;
2603 break;
2604 }
2605 case Instruction::LShr:
2606 case Instruction::AShr: {
2607 // shr exact can only shift out zero bits.
2608 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2609 if (BO->isExact())
2610 return isKnownNonZero(I->getOperand(0), Depth, Q);
2611
2612 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2613 // defined if the sign bit is shifted off the end.
2614 KnownBits Known =
2615 computeKnownBits(I->getOperand(0), DemandedElts, Depth, Q);
2616 if (Known.isNegative())
2617 return true;
2618
2619 // If the shifter operand is a constant, and all of the bits shifted
2620 // out are known to be zero, and X is known non-zero then at least one
2621 // non-zero bit must remain.
2622 if (ConstantInt *Shift = dyn_cast<ConstantInt>(I->getOperand(1))) {
2623 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2624 // Is there a known one in the portion not shifted out?
2625 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2626 return true;
2627 // Are all the bits to be shifted out known zero?
2628 if (Known.countMinTrailingZeros() >= ShiftVal)
2629 return isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q);
2630 }
2631 break;
2632 }
2633 case Instruction::UDiv:
2634 case Instruction::SDiv:
2635 // div exact can only produce a zero if the dividend is zero.
2636 if (cast<PossiblyExactOperator>(I)->isExact())
2637 return isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q);
2638 break;
2639 case Instruction::Add: {
2640 // X + Y.
2641 KnownBits XKnown =
2642 computeKnownBits(I->getOperand(0), DemandedElts, Depth, Q);
2643 KnownBits YKnown =
2644 computeKnownBits(I->getOperand(1), DemandedElts, Depth, Q);
2645
2646 // If X and Y are both non-negative (as signed values) then their sum is not
2647 // zero unless both X and Y are zero.
2648 if (XKnown.isNonNegative() && YKnown.isNonNegative())
2649 if (isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q) ||
2650 isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q))
2651 return true;
2652
2653 // If X and Y are both negative (as signed values) then their sum is not
2654 // zero unless both X and Y equal INT_MIN.
2655 if (XKnown.isNegative() && YKnown.isNegative()) {
2656 APInt Mask = APInt::getSignedMaxValue(BitWidth);
2657 // The sign bit of X is set. If some other bit is set then X is not equal
2658 // to INT_MIN.
2659 if (XKnown.One.intersects(Mask))
2660 return true;
2661 // The sign bit of Y is set. If some other bit is set then Y is not equal
2662 // to INT_MIN.
2663 if (YKnown.One.intersects(Mask))
2664 return true;
2665 }
2666
2667 // The sum of a non-negative number and a power of two is not zero.
2668 if (XKnown.isNonNegative() &&
2669 isKnownToBeAPowerOfTwo(I->getOperand(1), /*OrZero*/ false, Depth, Q))
2670 return true;
2671 if (YKnown.isNonNegative() &&
2672 isKnownToBeAPowerOfTwo(I->getOperand(0), /*OrZero*/ false, Depth, Q))
2673 return true;
2674 break;
2675 }
2676 case Instruction::Mul: {
2677 // If X and Y are non-zero then so is X * Y as long as the multiplication
2678 // does not overflow.
2679 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2680 if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2681 isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q) &&
2682 isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q))
2683 return true;
2684 break;
2685 }
2686 case Instruction::Select:
2687 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2688 if (isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q) &&
2689 isKnownNonZero(I->getOperand(2), DemandedElts, Depth, Q))
2690 return true;
2691 break;
2692 case Instruction::PHI: {
2693 auto *PN = cast<PHINode>(I);
2694 if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN))
2695 return true;
2696
2697 // Check if all incoming values are non-zero using recursion.
2698 Query RecQ = Q;
2699 unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1);
2700 return llvm::all_of(PN->operands(), [&](const Use &U) {
2701 if (U.get() == PN)
2702 return true;
2703 RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator();
2704 return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ);
2705 });
2706 }
2707 case Instruction::ExtractElement:
2708 if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) {
2709 const Value *Vec = EEI->getVectorOperand();
2710 const Value *Idx = EEI->getIndexOperand();
2711 auto *CIdx = dyn_cast<ConstantInt>(Idx);
2712 if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) {
2713 unsigned NumElts = VecTy->getNumElements();
2714 APInt DemandedVecElts = APInt::getAllOnes(NumElts);
2715 if (CIdx && CIdx->getValue().ult(NumElts))
2716 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue());
2717 return isKnownNonZero(Vec, DemandedVecElts, Depth, Q);
2718 }
2719 }
2720 break;
2721 case Instruction::Freeze:
2722 return isKnownNonZero(I->getOperand(0), Depth, Q) &&
2723 isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT,
2724 Depth);
2725 case Instruction::Call:
2726 if (cast<CallInst>(I)->getIntrinsicID() == Intrinsic::vscale)
2727 return true;
2728 break;
2729 }
2730
2731 KnownBits Known(BitWidth);
2732 computeKnownBits(V, DemandedElts, Known, Depth, Q);
2733 return Known.One != 0;
2734 }
2735
isKnownNonZero(const Value * V,unsigned Depth,const Query & Q)2736 bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) {
2737 auto *FVTy = dyn_cast<FixedVectorType>(V->getType());
2738 APInt DemandedElts =
2739 FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1);
2740 return isKnownNonZero(V, DemandedElts, Depth, Q);
2741 }
2742
2743 /// If the pair of operators are the same invertible function, return the
2744 /// the operands of the function corresponding to each input. Otherwise,
2745 /// return std::nullopt. An invertible function is one that is 1-to-1 and maps
2746 /// every input value to exactly one output value. This is equivalent to
2747 /// saying that Op1 and Op2 are equal exactly when the specified pair of
2748 /// operands are equal, (except that Op1 and Op2 may be poison more often.)
2749 static std::optional<std::pair<Value*, Value*>>
getInvertibleOperands(const Operator * Op1,const Operator * Op2)2750 getInvertibleOperands(const Operator *Op1,
2751 const Operator *Op2) {
2752 if (Op1->getOpcode() != Op2->getOpcode())
2753 return std::nullopt;
2754
2755 auto getOperands = [&](unsigned OpNum) -> auto {
2756 return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum));
2757 };
2758
2759 switch (Op1->getOpcode()) {
2760 default:
2761 break;
2762 case Instruction::Add:
2763 case Instruction::Sub:
2764 if (Op1->getOperand(0) == Op2->getOperand(0))
2765 return getOperands(1);
2766 if (Op1->getOperand(1) == Op2->getOperand(1))
2767 return getOperands(0);
2768 break;
2769 case Instruction::Mul: {
2770 // invertible if A * B == (A * B) mod 2^N where A, and B are integers
2771 // and N is the bitwdith. The nsw case is non-obvious, but proven by
2772 // alive2: https://alive2.llvm.org/ce/z/Z6D5qK
2773 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2774 auto *OBO2 = cast<OverflowingBinaryOperator>(Op2);
2775 if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) &&
2776 (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap()))
2777 break;
2778
2779 // Assume operand order has been canonicalized
2780 if (Op1->getOperand(1) == Op2->getOperand(1) &&
2781 isa<ConstantInt>(Op1->getOperand(1)) &&
2782 !cast<ConstantInt>(Op1->getOperand(1))->isZero())
2783 return getOperands(0);
2784 break;
2785 }
2786 case Instruction::Shl: {
2787 // Same as multiplies, with the difference that we don't need to check
2788 // for a non-zero multiply. Shifts always multiply by non-zero.
2789 auto *OBO1 = cast<OverflowingBinaryOperator>(Op1);
2790 auto *OBO2 = cast<OverflowingBinaryOperator>(Op2);
2791 if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) &&
2792 (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap()))
2793 break;
2794
2795 if (Op1->getOperand(1) == Op2->getOperand(1))
2796 return getOperands(0);
2797 break;
2798 }
2799 case Instruction::AShr:
2800 case Instruction::LShr: {
2801 auto *PEO1 = cast<PossiblyExactOperator>(Op1);
2802 auto *PEO2 = cast<PossiblyExactOperator>(Op2);
2803 if (!PEO1->isExact() || !PEO2->isExact())
2804 break;
2805
2806 if (Op1->getOperand(1) == Op2->getOperand(1))
2807 return getOperands(0);
2808 break;
2809 }
2810 case Instruction::SExt:
2811 case Instruction::ZExt:
2812 if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType())
2813 return getOperands(0);
2814 break;
2815 case Instruction::PHI: {
2816 const PHINode *PN1 = cast<PHINode>(Op1);
2817 const PHINode *PN2 = cast<PHINode>(Op2);
2818
2819 // If PN1 and PN2 are both recurrences, can we prove the entire recurrences
2820 // are a single invertible function of the start values? Note that repeated
2821 // application of an invertible function is also invertible
2822 BinaryOperator *BO1 = nullptr;
2823 Value *Start1 = nullptr, *Step1 = nullptr;
2824 BinaryOperator *BO2 = nullptr;
2825 Value *Start2 = nullptr, *Step2 = nullptr;
2826 if (PN1->getParent() != PN2->getParent() ||
2827 !matchSimpleRecurrence(PN1, BO1, Start1, Step1) ||
2828 !matchSimpleRecurrence(PN2, BO2, Start2, Step2))
2829 break;
2830
2831 auto Values = getInvertibleOperands(cast<Operator>(BO1),
2832 cast<Operator>(BO2));
2833 if (!Values)
2834 break;
2835
2836 // We have to be careful of mutually defined recurrences here. Ex:
2837 // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V
2838 // * X_i = Y_i = X_(i-1) OP Y_(i-1)
2839 // The invertibility of these is complicated, and not worth reasoning
2840 // about (yet?).
2841 if (Values->first != PN1 || Values->second != PN2)
2842 break;
2843
2844 return std::make_pair(Start1, Start2);
2845 }
2846 }
2847 return std::nullopt;
2848 }
2849
2850 /// Return true if V2 == V1 + X, where X is known non-zero.
isAddOfNonZero(const Value * V1,const Value * V2,unsigned Depth,const Query & Q)2851 static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth,
2852 const Query &Q) {
2853 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2854 if (!BO || BO->getOpcode() != Instruction::Add)
2855 return false;
2856 Value *Op = nullptr;
2857 if (V2 == BO->getOperand(0))
2858 Op = BO->getOperand(1);
2859 else if (V2 == BO->getOperand(1))
2860 Op = BO->getOperand(0);
2861 else
2862 return false;
2863 return isKnownNonZero(Op, Depth + 1, Q);
2864 }
2865
2866 /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and
2867 /// the multiplication is nuw or nsw.
isNonEqualMul(const Value * V1,const Value * V2,unsigned Depth,const Query & Q)2868 static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth,
2869 const Query &Q) {
2870 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) {
2871 const APInt *C;
2872 return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) &&
2873 (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) &&
2874 !C->isZero() && !C->isOne() && isKnownNonZero(V1, Depth + 1, Q);
2875 }
2876 return false;
2877 }
2878
2879 /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and
2880 /// the shift is nuw or nsw.
isNonEqualShl(const Value * V1,const Value * V2,unsigned Depth,const Query & Q)2881 static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth,
2882 const Query &Q) {
2883 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) {
2884 const APInt *C;
2885 return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) &&
2886 (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) &&
2887 !C->isZero() && isKnownNonZero(V1, Depth + 1, Q);
2888 }
2889 return false;
2890 }
2891
isNonEqualPHIs(const PHINode * PN1,const PHINode * PN2,unsigned Depth,const Query & Q)2892 static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2,
2893 unsigned Depth, const Query &Q) {
2894 // Check two PHIs are in same block.
2895 if (PN1->getParent() != PN2->getParent())
2896 return false;
2897
2898 SmallPtrSet<const BasicBlock *, 8> VisitedBBs;
2899 bool UsedFullRecursion = false;
2900 for (const BasicBlock *IncomBB : PN1->blocks()) {
2901 if (!VisitedBBs.insert(IncomBB).second)
2902 continue; // Don't reprocess blocks that we have dealt with already.
2903 const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB);
2904 const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB);
2905 const APInt *C1, *C2;
2906 if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2)
2907 continue;
2908
2909 // Only one pair of phi operands is allowed for full recursion.
2910 if (UsedFullRecursion)
2911 return false;
2912
2913 Query RecQ = Q;
2914 RecQ.CxtI = IncomBB->getTerminator();
2915 if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ))
2916 return false;
2917 UsedFullRecursion = true;
2918 }
2919 return true;
2920 }
2921
2922 /// Return true if it is known that V1 != V2.
isKnownNonEqual(const Value * V1,const Value * V2,unsigned Depth,const Query & Q)2923 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth,
2924 const Query &Q) {
2925 if (V1 == V2)
2926 return false;
2927 if (V1->getType() != V2->getType())
2928 // We can't look through casts yet.
2929 return false;
2930
2931 if (Depth >= MaxAnalysisRecursionDepth)
2932 return false;
2933
2934 // See if we can recurse through (exactly one of) our operands. This
2935 // requires our operation be 1-to-1 and map every input value to exactly
2936 // one output value. Such an operation is invertible.
2937 auto *O1 = dyn_cast<Operator>(V1);
2938 auto *O2 = dyn_cast<Operator>(V2);
2939 if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) {
2940 if (auto Values = getInvertibleOperands(O1, O2))
2941 return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q);
2942
2943 if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) {
2944 const PHINode *PN2 = cast<PHINode>(V2);
2945 // FIXME: This is missing a generalization to handle the case where one is
2946 // a PHI and another one isn't.
2947 if (isNonEqualPHIs(PN1, PN2, Depth, Q))
2948 return true;
2949 };
2950 }
2951
2952 if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q))
2953 return true;
2954
2955 if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q))
2956 return true;
2957
2958 if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q))
2959 return true;
2960
2961 if (V1->getType()->isIntOrIntVectorTy()) {
2962 // Are any known bits in V1 contradictory to known bits in V2? If V1
2963 // has a known zero where V2 has a known one, they must not be equal.
2964 KnownBits Known1 = computeKnownBits(V1, Depth, Q);
2965 KnownBits Known2 = computeKnownBits(V2, Depth, Q);
2966
2967 if (Known1.Zero.intersects(Known2.One) ||
2968 Known2.Zero.intersects(Known1.One))
2969 return true;
2970 }
2971 return false;
2972 }
2973
2974 /// Return true if 'V & Mask' is known to be zero. We use this predicate to
2975 /// simplify operations downstream. Mask is known to be zero for bits that V
2976 /// cannot have.
2977 ///
2978 /// This function is defined on values with integer type, values with pointer
2979 /// type, and vectors of integers. In the case
2980 /// where V is a vector, the mask, known zero, and known one values are the
2981 /// same width as the vector element, and the bit is set only if it is true
2982 /// for all of the elements in the vector.
MaskedValueIsZero(const Value * V,const APInt & Mask,unsigned Depth,const Query & Q)2983 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2984 const Query &Q) {
2985 KnownBits Known(Mask.getBitWidth());
2986 computeKnownBits(V, Known, Depth, Q);
2987 return Mask.isSubsetOf(Known.Zero);
2988 }
2989
2990 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2991 // Returns the input and lower/upper bounds.
isSignedMinMaxClamp(const Value * Select,const Value * & In,const APInt * & CLow,const APInt * & CHigh)2992 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2993 const APInt *&CLow, const APInt *&CHigh) {
2994 assert(isa<Operator>(Select) &&
2995 cast<Operator>(Select)->getOpcode() == Instruction::Select &&
2996 "Input should be a Select!");
2997
2998 const Value *LHS = nullptr, *RHS = nullptr;
2999 SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor;
3000 if (SPF != SPF_SMAX && SPF != SPF_SMIN)
3001 return false;
3002
3003 if (!match(RHS, m_APInt(CLow)))
3004 return false;
3005
3006 const Value *LHS2 = nullptr, *RHS2 = nullptr;
3007 SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
3008 if (getInverseMinMaxFlavor(SPF) != SPF2)
3009 return false;
3010
3011 if (!match(RHS2, m_APInt(CHigh)))
3012 return false;
3013
3014 if (SPF == SPF_SMIN)
3015 std::swap(CLow, CHigh);
3016
3017 In = LHS2;
3018 return CLow->sle(*CHigh);
3019 }
3020
isSignedMinMaxIntrinsicClamp(const IntrinsicInst * II,const APInt * & CLow,const APInt * & CHigh)3021 static bool isSignedMinMaxIntrinsicClamp(const IntrinsicInst *II,
3022 const APInt *&CLow,
3023 const APInt *&CHigh) {
3024 assert((II->getIntrinsicID() == Intrinsic::smin ||
3025 II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax");
3026
3027 Intrinsic::ID InverseID = getInverseMinMaxIntrinsic(II->getIntrinsicID());
3028 auto *InnerII = dyn_cast<IntrinsicInst>(II->getArgOperand(0));
3029 if (!InnerII || InnerII->getIntrinsicID() != InverseID ||
3030 !match(II->getArgOperand(1), m_APInt(CLow)) ||
3031 !match(InnerII->getArgOperand(1), m_APInt(CHigh)))
3032 return false;
3033
3034 if (II->getIntrinsicID() == Intrinsic::smin)
3035 std::swap(CLow, CHigh);
3036 return CLow->sle(*CHigh);
3037 }
3038
3039 /// For vector constants, loop over the elements and find the constant with the
3040 /// minimum number of sign bits. Return 0 if the value is not a vector constant
3041 /// or if any element was not analyzed; otherwise, return the count for the
3042 /// element with the minimum number of sign bits.
computeNumSignBitsVectorConstant(const Value * V,const APInt & DemandedElts,unsigned TyBits)3043 static unsigned computeNumSignBitsVectorConstant(const Value *V,
3044 const APInt &DemandedElts,
3045 unsigned TyBits) {
3046 const auto *CV = dyn_cast<Constant>(V);
3047 if (!CV || !isa<FixedVectorType>(CV->getType()))
3048 return 0;
3049
3050 unsigned MinSignBits = TyBits;
3051 unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements();
3052 for (unsigned i = 0; i != NumElts; ++i) {
3053 if (!DemandedElts[i])
3054 continue;
3055 // If we find a non-ConstantInt, bail out.
3056 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
3057 if (!Elt)
3058 return 0;
3059
3060 MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
3061 }
3062
3063 return MinSignBits;
3064 }
3065
3066 static unsigned ComputeNumSignBitsImpl(const Value *V,
3067 const APInt &DemandedElts,
3068 unsigned Depth, const Query &Q);
3069
ComputeNumSignBits(const Value * V,const APInt & DemandedElts,unsigned Depth,const Query & Q)3070 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts,
3071 unsigned Depth, const Query &Q) {
3072 unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q);
3073 assert(Result > 0 && "At least one sign bit needs to be present!");
3074 return Result;
3075 }
3076
3077 /// Return the number of times the sign bit of the register is replicated into
3078 /// the other bits. We know that at least 1 bit is always equal to the sign bit
3079 /// (itself), but other cases can give us information. For example, immediately
3080 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
3081 /// other, so we return 3. For vectors, return the number of sign bits for the
3082 /// vector element with the minimum number of known sign bits of the demanded
3083 /// elements in the vector specified by DemandedElts.
ComputeNumSignBitsImpl(const Value * V,const APInt & DemandedElts,unsigned Depth,const Query & Q)3084 static unsigned ComputeNumSignBitsImpl(const Value *V,
3085 const APInt &DemandedElts,
3086 unsigned Depth, const Query &Q) {
3087 Type *Ty = V->getType();
3088 #ifndef NDEBUG
3089 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
3090
3091 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) {
3092 assert(
3093 FVTy->getNumElements() == DemandedElts.getBitWidth() &&
3094 "DemandedElt width should equal the fixed vector number of elements");
3095 } else {
3096 assert(DemandedElts == APInt(1, 1) &&
3097 "DemandedElt width should be 1 for scalars");
3098 }
3099 #endif
3100
3101 // We return the minimum number of sign bits that are guaranteed to be present
3102 // in V, so for undef we have to conservatively return 1. We don't have the
3103 // same behavior for poison though -- that's a FIXME today.
3104
3105 Type *ScalarTy = Ty->getScalarType();
3106 unsigned TyBits = ScalarTy->isPointerTy() ?
3107 Q.DL.getPointerTypeSizeInBits(ScalarTy) :
3108 Q.DL.getTypeSizeInBits(ScalarTy);
3109
3110 unsigned Tmp, Tmp2;
3111 unsigned FirstAnswer = 1;
3112
3113 // Note that ConstantInt is handled by the general computeKnownBits case
3114 // below.
3115
3116 if (Depth == MaxAnalysisRecursionDepth)
3117 return 1;
3118
3119 if (auto *U = dyn_cast<Operator>(V)) {
3120 switch (Operator::getOpcode(V)) {
3121 default: break;
3122 case Instruction::SExt:
3123 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
3124 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
3125
3126 case Instruction::SDiv: {
3127 const APInt *Denominator;
3128 // sdiv X, C -> adds log(C) sign bits.
3129 if (match(U->getOperand(1), m_APInt(Denominator))) {
3130
3131 // Ignore non-positive denominator.
3132 if (!Denominator->isStrictlyPositive())
3133 break;
3134
3135 // Calculate the incoming numerator bits.
3136 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3137
3138 // Add floor(log(C)) bits to the numerator bits.
3139 return std::min(TyBits, NumBits + Denominator->logBase2());
3140 }
3141 break;
3142 }
3143
3144 case Instruction::SRem: {
3145 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3146
3147 const APInt *Denominator;
3148 // srem X, C -> we know that the result is within [-C+1,C) when C is a
3149 // positive constant. This let us put a lower bound on the number of sign
3150 // bits.
3151 if (match(U->getOperand(1), m_APInt(Denominator))) {
3152
3153 // Ignore non-positive denominator.
3154 if (Denominator->isStrictlyPositive()) {
3155 // Calculate the leading sign bit constraints by examining the
3156 // denominator. Given that the denominator is positive, there are two
3157 // cases:
3158 //
3159 // 1. The numerator is positive. The result range is [0,C) and
3160 // [0,C) u< (1 << ceilLogBase2(C)).
3161 //
3162 // 2. The numerator is negative. Then the result range is (-C,0] and
3163 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
3164 //
3165 // Thus a lower bound on the number of sign bits is `TyBits -
3166 // ceilLogBase2(C)`.
3167
3168 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
3169 Tmp = std::max(Tmp, ResBits);
3170 }
3171 }
3172 return Tmp;
3173 }
3174
3175 case Instruction::AShr: {
3176 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3177 // ashr X, C -> adds C sign bits. Vectors too.
3178 const APInt *ShAmt;
3179 if (match(U->getOperand(1), m_APInt(ShAmt))) {
3180 if (ShAmt->uge(TyBits))
3181 break; // Bad shift.
3182 unsigned ShAmtLimited = ShAmt->getZExtValue();
3183 Tmp += ShAmtLimited;
3184 if (Tmp > TyBits) Tmp = TyBits;
3185 }
3186 return Tmp;
3187 }
3188 case Instruction::Shl: {
3189 const APInt *ShAmt;
3190 if (match(U->getOperand(1), m_APInt(ShAmt))) {
3191 // shl destroys sign bits.
3192 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3193 if (ShAmt->uge(TyBits) || // Bad shift.
3194 ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
3195 Tmp2 = ShAmt->getZExtValue();
3196 return Tmp - Tmp2;
3197 }
3198 break;
3199 }
3200 case Instruction::And:
3201 case Instruction::Or:
3202 case Instruction::Xor: // NOT is handled here.
3203 // Logical binary ops preserve the number of sign bits at the worst.
3204 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3205 if (Tmp != 1) {
3206 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3207 FirstAnswer = std::min(Tmp, Tmp2);
3208 // We computed what we know about the sign bits as our first
3209 // answer. Now proceed to the generic code that uses
3210 // computeKnownBits, and pick whichever answer is better.
3211 }
3212 break;
3213
3214 case Instruction::Select: {
3215 // If we have a clamp pattern, we know that the number of sign bits will
3216 // be the minimum of the clamp min/max range.
3217 const Value *X;
3218 const APInt *CLow, *CHigh;
3219 if (isSignedMinMaxClamp(U, X, CLow, CHigh))
3220 return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
3221
3222 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3223 if (Tmp == 1) break;
3224 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
3225 return std::min(Tmp, Tmp2);
3226 }
3227
3228 case Instruction::Add:
3229 // Add can have at most one carry bit. Thus we know that the output
3230 // is, at worst, one more bit than the inputs.
3231 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3232 if (Tmp == 1) break;
3233
3234 // Special case decrementing a value (ADD X, -1):
3235 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
3236 if (CRHS->isAllOnesValue()) {
3237 KnownBits Known(TyBits);
3238 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
3239
3240 // If the input is known to be 0 or 1, the output is 0/-1, which is
3241 // all sign bits set.
3242 if ((Known.Zero | 1).isAllOnes())
3243 return TyBits;
3244
3245 // If we are subtracting one from a positive number, there is no carry
3246 // out of the result.
3247 if (Known.isNonNegative())
3248 return Tmp;
3249 }
3250
3251 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3252 if (Tmp2 == 1) break;
3253 return std::min(Tmp, Tmp2) - 1;
3254
3255 case Instruction::Sub:
3256 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3257 if (Tmp2 == 1) break;
3258
3259 // Handle NEG.
3260 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
3261 if (CLHS->isNullValue()) {
3262 KnownBits Known(TyBits);
3263 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
3264 // If the input is known to be 0 or 1, the output is 0/-1, which is
3265 // all sign bits set.
3266 if ((Known.Zero | 1).isAllOnes())
3267 return TyBits;
3268
3269 // If the input is known to be positive (the sign bit is known clear),
3270 // the output of the NEG has the same number of sign bits as the
3271 // input.
3272 if (Known.isNonNegative())
3273 return Tmp2;
3274
3275 // Otherwise, we treat this like a SUB.
3276 }
3277
3278 // Sub can have at most one carry bit. Thus we know that the output
3279 // is, at worst, one more bit than the inputs.
3280 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3281 if (Tmp == 1) break;
3282 return std::min(Tmp, Tmp2) - 1;
3283
3284 case Instruction::Mul: {
3285 // The output of the Mul can be at most twice the valid bits in the
3286 // inputs.
3287 unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3288 if (SignBitsOp0 == 1) break;
3289 unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
3290 if (SignBitsOp1 == 1) break;
3291 unsigned OutValidBits =
3292 (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
3293 return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
3294 }
3295
3296 case Instruction::PHI: {
3297 const PHINode *PN = cast<PHINode>(U);
3298 unsigned NumIncomingValues = PN->getNumIncomingValues();
3299 // Don't analyze large in-degree PHIs.
3300 if (NumIncomingValues > 4) break;
3301 // Unreachable blocks may have zero-operand PHI nodes.
3302 if (NumIncomingValues == 0) break;
3303
3304 // Take the minimum of all incoming values. This can't infinitely loop
3305 // because of our depth threshold.
3306 Query RecQ = Q;
3307 Tmp = TyBits;
3308 for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) {
3309 if (Tmp == 1) return Tmp;
3310 RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator();
3311 Tmp = std::min(
3312 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ));
3313 }
3314 return Tmp;
3315 }
3316
3317 case Instruction::Trunc: {
3318 // If the input contained enough sign bits that some remain after the
3319 // truncation, then we can make use of that. Otherwise we don't know
3320 // anything.
3321 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3322 unsigned OperandTyBits = U->getOperand(0)->getType()->getScalarSizeInBits();
3323 if (Tmp > (OperandTyBits - TyBits))
3324 return Tmp - (OperandTyBits - TyBits);
3325
3326 return 1;
3327 }
3328
3329 case Instruction::ExtractElement:
3330 // Look through extract element. At the moment we keep this simple and
3331 // skip tracking the specific element. But at least we might find
3332 // information valid for all elements of the vector (for example if vector
3333 // is sign extended, shifted, etc).
3334 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3335
3336 case Instruction::ShuffleVector: {
3337 // Collect the minimum number of sign bits that are shared by every vector
3338 // element referenced by the shuffle.
3339 auto *Shuf = dyn_cast<ShuffleVectorInst>(U);
3340 if (!Shuf) {
3341 // FIXME: Add support for shufflevector constant expressions.
3342 return 1;
3343 }
3344 APInt DemandedLHS, DemandedRHS;
3345 // For undef elements, we don't know anything about the common state of
3346 // the shuffle result.
3347 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS))
3348 return 1;
3349 Tmp = std::numeric_limits<unsigned>::max();
3350 if (!!DemandedLHS) {
3351 const Value *LHS = Shuf->getOperand(0);
3352 Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q);
3353 }
3354 // If we don't know anything, early out and try computeKnownBits
3355 // fall-back.
3356 if (Tmp == 1)
3357 break;
3358 if (!!DemandedRHS) {
3359 const Value *RHS = Shuf->getOperand(1);
3360 Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q);
3361 Tmp = std::min(Tmp, Tmp2);
3362 }
3363 // If we don't know anything, early out and try computeKnownBits
3364 // fall-back.
3365 if (Tmp == 1)
3366 break;
3367 assert(Tmp <= TyBits && "Failed to determine minimum sign bits");
3368 return Tmp;
3369 }
3370 case Instruction::Call: {
3371 if (const auto *II = dyn_cast<IntrinsicInst>(U)) {
3372 switch (II->getIntrinsicID()) {
3373 default: break;
3374 case Intrinsic::abs:
3375 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
3376 if (Tmp == 1) break;
3377
3378 // Absolute value reduces number of sign bits by at most 1.
3379 return Tmp - 1;
3380 case Intrinsic::smin:
3381 case Intrinsic::smax: {
3382 const APInt *CLow, *CHigh;
3383 if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh))
3384 return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
3385 }
3386 }
3387 }
3388 }
3389 }
3390 }
3391
3392 // Finally, if we can prove that the top bits of the result are 0's or 1's,
3393 // use this information.
3394
3395 // If we can examine all elements of a vector constant successfully, we're
3396 // done (we can't do any better than that). If not, keep trying.
3397 if (unsigned VecSignBits =
3398 computeNumSignBitsVectorConstant(V, DemandedElts, TyBits))
3399 return VecSignBits;
3400
3401 KnownBits Known(TyBits);
3402 computeKnownBits(V, DemandedElts, Known, Depth, Q);
3403
3404 // If we know that the sign bit is either zero or one, determine the number of
3405 // identical bits in the top of the input value.
3406 return std::max(FirstAnswer, Known.countMinSignBits());
3407 }
3408
getIntrinsicForCallSite(const CallBase & CB,const TargetLibraryInfo * TLI)3409 Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB,
3410 const TargetLibraryInfo *TLI) {
3411 const Function *F = CB.getCalledFunction();
3412 if (!F)
3413 return Intrinsic::not_intrinsic;
3414
3415 if (F->isIntrinsic())
3416 return F->getIntrinsicID();
3417
3418 // We are going to infer semantics of a library function based on mapping it
3419 // to an LLVM intrinsic. Check that the library function is available from
3420 // this callbase and in this environment.
3421 LibFunc Func;
3422 if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) ||
3423 !CB.onlyReadsMemory())
3424 return Intrinsic::not_intrinsic;
3425
3426 switch (Func) {
3427 default:
3428 break;
3429 case LibFunc_sin:
3430 case LibFunc_sinf:
3431 case LibFunc_sinl:
3432 return Intrinsic::sin;
3433 case LibFunc_cos:
3434 case LibFunc_cosf:
3435 case LibFunc_cosl:
3436 return Intrinsic::cos;
3437 case LibFunc_exp:
3438 case LibFunc_expf:
3439 case LibFunc_expl:
3440 return Intrinsic::exp;
3441 case LibFunc_exp2:
3442 case LibFunc_exp2f:
3443 case LibFunc_exp2l:
3444 return Intrinsic::exp2;
3445 case LibFunc_log:
3446 case LibFunc_logf:
3447 case LibFunc_logl:
3448 return Intrinsic::log;
3449 case LibFunc_log10:
3450 case LibFunc_log10f:
3451 case LibFunc_log10l:
3452 return Intrinsic::log10;
3453 case LibFunc_log2:
3454 case LibFunc_log2f:
3455 case LibFunc_log2l:
3456 return Intrinsic::log2;
3457 case LibFunc_fabs:
3458 case LibFunc_fabsf:
3459 case LibFunc_fabsl:
3460 return Intrinsic::fabs;
3461 case LibFunc_fmin:
3462 case LibFunc_fminf:
3463 case LibFunc_fminl:
3464 return Intrinsic::minnum;
3465 case LibFunc_fmax:
3466 case LibFunc_fmaxf:
3467 case LibFunc_fmaxl:
3468 return Intrinsic::maxnum;
3469 case LibFunc_copysign:
3470 case LibFunc_copysignf:
3471 case LibFunc_copysignl:
3472 return Intrinsic::copysign;
3473 case LibFunc_floor:
3474 case LibFunc_floorf:
3475 case LibFunc_floorl:
3476 return Intrinsic::floor;
3477 case LibFunc_ceil:
3478 case LibFunc_ceilf:
3479 case LibFunc_ceill:
3480 return Intrinsic::ceil;
3481 case LibFunc_trunc:
3482 case LibFunc_truncf:
3483 case LibFunc_truncl:
3484 return Intrinsic::trunc;
3485 case LibFunc_rint:
3486 case LibFunc_rintf:
3487 case LibFunc_rintl:
3488 return Intrinsic::rint;
3489 case LibFunc_nearbyint:
3490 case LibFunc_nearbyintf:
3491 case LibFunc_nearbyintl:
3492 return Intrinsic::nearbyint;
3493 case LibFunc_round:
3494 case LibFunc_roundf:
3495 case LibFunc_roundl:
3496 return Intrinsic::round;
3497 case LibFunc_roundeven:
3498 case LibFunc_roundevenf:
3499 case LibFunc_roundevenl:
3500 return Intrinsic::roundeven;
3501 case LibFunc_pow:
3502 case LibFunc_powf:
3503 case LibFunc_powl:
3504 return Intrinsic::pow;
3505 case LibFunc_sqrt:
3506 case LibFunc_sqrtf:
3507 case LibFunc_sqrtl:
3508 return Intrinsic::sqrt;
3509 }
3510
3511 return Intrinsic::not_intrinsic;
3512 }
3513
3514 /// Return true if we can prove that the specified FP value is never equal to
3515 /// -0.0.
3516 /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee
3517 /// that a value is not -0.0. It only guarantees that -0.0 may be treated
3518 /// the same as +0.0 in floating-point ops.
CannotBeNegativeZero(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)3519 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
3520 unsigned Depth) {
3521 if (auto *CFP = dyn_cast<ConstantFP>(V))
3522 return !CFP->getValueAPF().isNegZero();
3523
3524 if (Depth == MaxAnalysisRecursionDepth)
3525 return false;
3526
3527 auto *Op = dyn_cast<Operator>(V);
3528 if (!Op)
3529 return false;
3530
3531 // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
3532 if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
3533 return true;
3534
3535 // sitofp and uitofp turn into +0.0 for zero.
3536 if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
3537 return true;
3538
3539 if (auto *Call = dyn_cast<CallInst>(Op)) {
3540 Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI);
3541 switch (IID) {
3542 default:
3543 break;
3544 // sqrt(-0.0) = -0.0, no other negative results are possible.
3545 case Intrinsic::sqrt:
3546 case Intrinsic::canonicalize:
3547 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
3548 case Intrinsic::experimental_constrained_sqrt: {
3549 // NOTE: This rounding mode restriction may be too strict.
3550 const auto *CI = cast<ConstrainedFPIntrinsic>(Call);
3551 if (CI->getRoundingMode() == RoundingMode::NearestTiesToEven)
3552 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
3553 else
3554 return false;
3555 }
3556 // fabs(x) != -0.0
3557 case Intrinsic::fabs:
3558 return true;
3559 // sitofp and uitofp turn into +0.0 for zero.
3560 case Intrinsic::experimental_constrained_sitofp:
3561 case Intrinsic::experimental_constrained_uitofp:
3562 return true;
3563 }
3564 }
3565
3566 return false;
3567 }
3568
3569 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
3570 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
3571 /// bit despite comparing equal.
cannotBeOrderedLessThanZeroImpl(const Value * V,const TargetLibraryInfo * TLI,bool SignBitOnly,unsigned Depth)3572 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
3573 const TargetLibraryInfo *TLI,
3574 bool SignBitOnly,
3575 unsigned Depth) {
3576 // TODO: This function does not do the right thing when SignBitOnly is true
3577 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
3578 // which flips the sign bits of NaNs. See
3579 // https://llvm.org/bugs/show_bug.cgi?id=31702.
3580
3581 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
3582 return !CFP->getValueAPF().isNegative() ||
3583 (!SignBitOnly && CFP->getValueAPF().isZero());
3584 }
3585
3586 // Handle vector of constants.
3587 if (auto *CV = dyn_cast<Constant>(V)) {
3588 if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) {
3589 unsigned NumElts = CVFVTy->getNumElements();
3590 for (unsigned i = 0; i != NumElts; ++i) {
3591 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
3592 if (!CFP)
3593 return false;
3594 if (CFP->getValueAPF().isNegative() &&
3595 (SignBitOnly || !CFP->getValueAPF().isZero()))
3596 return false;
3597 }
3598
3599 // All non-negative ConstantFPs.
3600 return true;
3601 }
3602 }
3603
3604 if (Depth == MaxAnalysisRecursionDepth)
3605 return false;
3606
3607 const Operator *I = dyn_cast<Operator>(V);
3608 if (!I)
3609 return false;
3610
3611 switch (I->getOpcode()) {
3612 default:
3613 break;
3614 // Unsigned integers are always nonnegative.
3615 case Instruction::UIToFP:
3616 return true;
3617 case Instruction::FDiv:
3618 // X / X is always exactly 1.0 or a NaN.
3619 if (I->getOperand(0) == I->getOperand(1) &&
3620 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
3621 return true;
3622
3623 // Set SignBitOnly for RHS, because X / -0.0 is -Inf (or NaN).
3624 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3625 Depth + 1) &&
3626 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI,
3627 /*SignBitOnly*/ true, Depth + 1);
3628 case Instruction::FMul:
3629 // X * X is always non-negative or a NaN.
3630 if (I->getOperand(0) == I->getOperand(1) &&
3631 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
3632 return true;
3633
3634 [[fallthrough]];
3635 case Instruction::FAdd:
3636 case Instruction::FRem:
3637 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3638 Depth + 1) &&
3639 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3640 Depth + 1);
3641 case Instruction::Select:
3642 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3643 Depth + 1) &&
3644 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3645 Depth + 1);
3646 case Instruction::FPExt:
3647 case Instruction::FPTrunc:
3648 // Widening/narrowing never change sign.
3649 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3650 Depth + 1);
3651 case Instruction::ExtractElement:
3652 // Look through extract element. At the moment we keep this simple and skip
3653 // tracking the specific element. But at least we might find information
3654 // valid for all elements of the vector.
3655 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3656 Depth + 1);
3657 case Instruction::Call:
3658 const auto *CI = cast<CallInst>(I);
3659 Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI);
3660 switch (IID) {
3661 default:
3662 break;
3663 case Intrinsic::canonicalize:
3664 case Intrinsic::arithmetic_fence:
3665 case Intrinsic::floor:
3666 case Intrinsic::ceil:
3667 case Intrinsic::trunc:
3668 case Intrinsic::rint:
3669 case Intrinsic::nearbyint:
3670 case Intrinsic::round:
3671 case Intrinsic::roundeven:
3672 case Intrinsic::fptrunc_round:
3673 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, Depth + 1);
3674 case Intrinsic::maxnum: {
3675 Value *V0 = I->getOperand(0), *V1 = I->getOperand(1);
3676 auto isPositiveNum = [&](Value *V) {
3677 if (SignBitOnly) {
3678 // With SignBitOnly, this is tricky because the result of
3679 // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is
3680 // a constant strictly greater than 0.0.
3681 const APFloat *C;
3682 return match(V, m_APFloat(C)) &&
3683 *C > APFloat::getZero(C->getSemantics());
3684 }
3685
3686 // -0.0 compares equal to 0.0, so if this operand is at least -0.0,
3687 // maxnum can't be ordered-less-than-zero.
3688 return isKnownNeverNaN(V, TLI) &&
3689 cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1);
3690 };
3691
3692 // TODO: This could be improved. We could also check that neither operand
3693 // has its sign bit set (and at least 1 is not-NAN?).
3694 return isPositiveNum(V0) || isPositiveNum(V1);
3695 }
3696
3697 case Intrinsic::maximum:
3698 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3699 Depth + 1) ||
3700 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3701 Depth + 1);
3702 case Intrinsic::minnum:
3703 case Intrinsic::minimum:
3704 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3705 Depth + 1) &&
3706 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3707 Depth + 1);
3708 case Intrinsic::exp:
3709 case Intrinsic::exp2:
3710 case Intrinsic::fabs:
3711 return true;
3712 case Intrinsic::copysign:
3713 // Only the sign operand matters.
3714 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, true,
3715 Depth + 1);
3716 case Intrinsic::sqrt:
3717 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
3718 if (!SignBitOnly)
3719 return true;
3720 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
3721 CannotBeNegativeZero(CI->getOperand(0), TLI));
3722
3723 case Intrinsic::powi:
3724 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
3725 // powi(x,n) is non-negative if n is even.
3726 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
3727 return true;
3728 }
3729 // TODO: This is not correct. Given that exp is an integer, here are the
3730 // ways that pow can return a negative value:
3731 //
3732 // pow(x, exp) --> negative if exp is odd and x is negative.
3733 // pow(-0, exp) --> -inf if exp is negative odd.
3734 // pow(-0, exp) --> -0 if exp is positive odd.
3735 // pow(-inf, exp) --> -0 if exp is negative odd.
3736 // pow(-inf, exp) --> -inf if exp is positive odd.
3737 //
3738 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
3739 // but we must return false if x == -0. Unfortunately we do not currently
3740 // have a way of expressing this constraint. See details in
3741 // https://llvm.org/bugs/show_bug.cgi?id=31702.
3742 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3743 Depth + 1);
3744
3745 case Intrinsic::fma:
3746 case Intrinsic::fmuladd:
3747 // x*x+y is non-negative if y is non-negative.
3748 return I->getOperand(0) == I->getOperand(1) &&
3749 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3750 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3751 Depth + 1);
3752 }
3753 break;
3754 }
3755 return false;
3756 }
3757
CannotBeOrderedLessThanZero(const Value * V,const TargetLibraryInfo * TLI)3758 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
3759 const TargetLibraryInfo *TLI) {
3760 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3761 }
3762
SignBitMustBeZero(const Value * V,const TargetLibraryInfo * TLI)3763 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
3764 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3765 }
3766
isKnownNeverInfinity(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)3767 bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI,
3768 unsigned Depth) {
3769 assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type");
3770
3771 // If we're told that infinities won't happen, assume they won't.
3772 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3773 if (FPMathOp->hasNoInfs())
3774 return true;
3775
3776 // Handle scalar constants.
3777 if (auto *CFP = dyn_cast<ConstantFP>(V))
3778 return !CFP->isInfinity();
3779
3780 if (Depth == MaxAnalysisRecursionDepth)
3781 return false;
3782
3783 if (auto *Inst = dyn_cast<Instruction>(V)) {
3784 switch (Inst->getOpcode()) {
3785 case Instruction::Select: {
3786 return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) &&
3787 isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1);
3788 }
3789 case Instruction::SIToFP:
3790 case Instruction::UIToFP: {
3791 // Get width of largest magnitude integer (remove a bit if signed).
3792 // This still works for a signed minimum value because the largest FP
3793 // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx).
3794 int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits();
3795 if (Inst->getOpcode() == Instruction::SIToFP)
3796 --IntSize;
3797
3798 // If the exponent of the largest finite FP value can hold the largest
3799 // integer, the result of the cast must be finite.
3800 Type *FPTy = Inst->getType()->getScalarType();
3801 return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize;
3802 }
3803 case Instruction::FNeg:
3804 case Instruction::FPExt: {
3805 // Peek through to source op. If it is not infinity, this is not infinity.
3806 return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1);
3807 }
3808 case Instruction::FPTrunc: {
3809 // Need a range check.
3810 return false;
3811 }
3812 default:
3813 break;
3814 }
3815
3816 if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3817 switch (II->getIntrinsicID()) {
3818 case Intrinsic::sin:
3819 case Intrinsic::cos:
3820 // Return NaN on infinite inputs.
3821 return true;
3822 case Intrinsic::fabs:
3823 case Intrinsic::sqrt:
3824 case Intrinsic::canonicalize:
3825 case Intrinsic::copysign:
3826 case Intrinsic::arithmetic_fence:
3827 case Intrinsic::trunc:
3828 return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1);
3829 case Intrinsic::floor:
3830 case Intrinsic::ceil:
3831 case Intrinsic::rint:
3832 case Intrinsic::nearbyint:
3833 case Intrinsic::round:
3834 case Intrinsic::roundeven:
3835 // PPC_FP128 is a special case.
3836 if (V->getType()->isMultiUnitFPType())
3837 return false;
3838 return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1);
3839 case Intrinsic::fptrunc_round:
3840 // Requires knowing the value range.
3841 return false;
3842 case Intrinsic::minnum:
3843 case Intrinsic::maxnum:
3844 case Intrinsic::minimum:
3845 case Intrinsic::maximum:
3846 return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) &&
3847 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1);
3848 case Intrinsic::log:
3849 case Intrinsic::log10:
3850 case Intrinsic::log2:
3851 // log(+inf) -> +inf
3852 // log([+-]0.0) -> -inf
3853 // log(-inf) -> nan
3854 // log(-x) -> nan
3855 // TODO: We lack API to check the == 0 case.
3856 return false;
3857 case Intrinsic::exp:
3858 case Intrinsic::exp2:
3859 case Intrinsic::pow:
3860 case Intrinsic::powi:
3861 case Intrinsic::fma:
3862 case Intrinsic::fmuladd:
3863 // These can return infinities on overflow cases, so it's hard to prove
3864 // anything about it.
3865 return false;
3866 default:
3867 break;
3868 }
3869 }
3870 }
3871
3872 // try to handle fixed width vector constants
3873 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3874 if (VFVTy && isa<Constant>(V)) {
3875 // For vectors, verify that each element is not infinity.
3876 unsigned NumElts = VFVTy->getNumElements();
3877 for (unsigned i = 0; i != NumElts; ++i) {
3878 Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3879 if (!Elt)
3880 return false;
3881 if (isa<UndefValue>(Elt))
3882 continue;
3883 auto *CElt = dyn_cast<ConstantFP>(Elt);
3884 if (!CElt || CElt->isInfinity())
3885 return false;
3886 }
3887 // All elements were confirmed non-infinity or undefined.
3888 return true;
3889 }
3890
3891 // was not able to prove that V never contains infinity
3892 return false;
3893 }
3894
isKnownNeverNaN(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)3895 bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI,
3896 unsigned Depth) {
3897 assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
3898
3899 // If we're told that NaNs won't happen, assume they won't.
3900 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3901 if (FPMathOp->hasNoNaNs())
3902 return true;
3903
3904 // Handle scalar constants.
3905 if (auto *CFP = dyn_cast<ConstantFP>(V))
3906 return !CFP->isNaN();
3907
3908 if (Depth == MaxAnalysisRecursionDepth)
3909 return false;
3910
3911 if (auto *Inst = dyn_cast<Instruction>(V)) {
3912 switch (Inst->getOpcode()) {
3913 case Instruction::FAdd:
3914 case Instruction::FSub:
3915 // Adding positive and negative infinity produces NaN.
3916 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3917 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3918 (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) ||
3919 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1));
3920
3921 case Instruction::FMul:
3922 // Zero multiplied with infinity produces NaN.
3923 // FIXME: If neither side can be zero fmul never produces NaN.
3924 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3925 isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) &&
3926 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3927 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1);
3928
3929 case Instruction::FDiv:
3930 case Instruction::FRem:
3931 // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN.
3932 return false;
3933
3934 case Instruction::Select: {
3935 return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3936 isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3937 }
3938 case Instruction::SIToFP:
3939 case Instruction::UIToFP:
3940 return true;
3941 case Instruction::FPTrunc:
3942 case Instruction::FPExt:
3943 case Instruction::FNeg:
3944 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3945 default:
3946 break;
3947 }
3948 }
3949
3950 if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3951 switch (II->getIntrinsicID()) {
3952 case Intrinsic::canonicalize:
3953 case Intrinsic::fabs:
3954 case Intrinsic::copysign:
3955 case Intrinsic::exp:
3956 case Intrinsic::exp2:
3957 case Intrinsic::floor:
3958 case Intrinsic::ceil:
3959 case Intrinsic::trunc:
3960 case Intrinsic::rint:
3961 case Intrinsic::nearbyint:
3962 case Intrinsic::round:
3963 case Intrinsic::roundeven:
3964 case Intrinsic::arithmetic_fence:
3965 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3966 case Intrinsic::sqrt:
3967 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3968 CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3969 case Intrinsic::minnum:
3970 case Intrinsic::maxnum:
3971 // If either operand is not NaN, the result is not NaN.
3972 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) ||
3973 isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1);
3974 default:
3975 return false;
3976 }
3977 }
3978
3979 // Try to handle fixed width vector constants
3980 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
3981 if (VFVTy && isa<Constant>(V)) {
3982 // For vectors, verify that each element is not NaN.
3983 unsigned NumElts = VFVTy->getNumElements();
3984 for (unsigned i = 0; i != NumElts; ++i) {
3985 Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3986 if (!Elt)
3987 return false;
3988 if (isa<UndefValue>(Elt))
3989 continue;
3990 auto *CElt = dyn_cast<ConstantFP>(Elt);
3991 if (!CElt || CElt->isNaN())
3992 return false;
3993 }
3994 // All elements were confirmed not-NaN or undefined.
3995 return true;
3996 }
3997
3998 // Was not able to prove that V never contains NaN
3999 return false;
4000 }
4001
isBytewiseValue(Value * V,const DataLayout & DL)4002 Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) {
4003
4004 // All byte-wide stores are splatable, even of arbitrary variables.
4005 if (V->getType()->isIntegerTy(8))
4006 return V;
4007
4008 LLVMContext &Ctx = V->getContext();
4009
4010 // Undef don't care.
4011 auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
4012 if (isa<UndefValue>(V))
4013 return UndefInt8;
4014
4015 // Return Undef for zero-sized type.
4016 if (!DL.getTypeStoreSize(V->getType()).isNonZero())
4017 return UndefInt8;
4018
4019 Constant *C = dyn_cast<Constant>(V);
4020 if (!C) {
4021 // Conceptually, we could handle things like:
4022 // %a = zext i8 %X to i16
4023 // %b = shl i16 %a, 8
4024 // %c = or i16 %a, %b
4025 // but until there is an example that actually needs this, it doesn't seem
4026 // worth worrying about.
4027 return nullptr;
4028 }
4029
4030 // Handle 'null' ConstantArrayZero etc.
4031 if (C->isNullValue())
4032 return Constant::getNullValue(Type::getInt8Ty(Ctx));
4033
4034 // Constant floating-point values can be handled as integer values if the
4035 // corresponding integer value is "byteable". An important case is 0.0.
4036 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
4037 Type *Ty = nullptr;
4038 if (CFP->getType()->isHalfTy())
4039 Ty = Type::getInt16Ty(Ctx);
4040 else if (CFP->getType()->isFloatTy())
4041 Ty = Type::getInt32Ty(Ctx);
4042 else if (CFP->getType()->isDoubleTy())
4043 Ty = Type::getInt64Ty(Ctx);
4044 // Don't handle long double formats, which have strange constraints.
4045 return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL)
4046 : nullptr;
4047 }
4048
4049 // We can handle constant integers that are multiple of 8 bits.
4050 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
4051 if (CI->getBitWidth() % 8 == 0) {
4052 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
4053 if (!CI->getValue().isSplat(8))
4054 return nullptr;
4055 return ConstantInt::get(Ctx, CI->getValue().trunc(8));
4056 }
4057 }
4058
4059 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
4060 if (CE->getOpcode() == Instruction::IntToPtr) {
4061 if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) {
4062 unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace());
4063 return isBytewiseValue(
4064 ConstantExpr::getIntegerCast(CE->getOperand(0),
4065 Type::getIntNTy(Ctx, BitWidth), false),
4066 DL);
4067 }
4068 }
4069 }
4070
4071 auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
4072 if (LHS == RHS)
4073 return LHS;
4074 if (!LHS || !RHS)
4075 return nullptr;
4076 if (LHS == UndefInt8)
4077 return RHS;
4078 if (RHS == UndefInt8)
4079 return LHS;
4080 return nullptr;
4081 };
4082
4083 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
4084 Value *Val = UndefInt8;
4085 for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
4086 if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL))))
4087 return nullptr;
4088 return Val;
4089 }
4090
4091 if (isa<ConstantAggregate>(C)) {
4092 Value *Val = UndefInt8;
4093 for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
4094 if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL))))
4095 return nullptr;
4096 return Val;
4097 }
4098
4099 // Don't try to handle the handful of other constants.
4100 return nullptr;
4101 }
4102
4103 // This is the recursive version of BuildSubAggregate. It takes a few different
4104 // arguments. Idxs is the index within the nested struct From that we are
4105 // looking at now (which is of type IndexedType). IdxSkip is the number of
4106 // indices from Idxs that should be left out when inserting into the resulting
4107 // struct. To is the result struct built so far, new insertvalue instructions
4108 // build on that.
BuildSubAggregate(Value * From,Value * To,Type * IndexedType,SmallVectorImpl<unsigned> & Idxs,unsigned IdxSkip,Instruction * InsertBefore)4109 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
4110 SmallVectorImpl<unsigned> &Idxs,
4111 unsigned IdxSkip,
4112 Instruction *InsertBefore) {
4113 StructType *STy = dyn_cast<StructType>(IndexedType);
4114 if (STy) {
4115 // Save the original To argument so we can modify it
4116 Value *OrigTo = To;
4117 // General case, the type indexed by Idxs is a struct
4118 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
4119 // Process each struct element recursively
4120 Idxs.push_back(i);
4121 Value *PrevTo = To;
4122 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
4123 InsertBefore);
4124 Idxs.pop_back();
4125 if (!To) {
4126 // Couldn't find any inserted value for this index? Cleanup
4127 while (PrevTo != OrigTo) {
4128 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
4129 PrevTo = Del->getAggregateOperand();
4130 Del->eraseFromParent();
4131 }
4132 // Stop processing elements
4133 break;
4134 }
4135 }
4136 // If we successfully found a value for each of our subaggregates
4137 if (To)
4138 return To;
4139 }
4140 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
4141 // the struct's elements had a value that was inserted directly. In the latter
4142 // case, perhaps we can't determine each of the subelements individually, but
4143 // we might be able to find the complete struct somewhere.
4144
4145 // Find the value that is at that particular spot
4146 Value *V = FindInsertedValue(From, Idxs);
4147
4148 if (!V)
4149 return nullptr;
4150
4151 // Insert the value in the new (sub) aggregate
4152 return InsertValueInst::Create(To, V, ArrayRef(Idxs).slice(IdxSkip), "tmp",
4153 InsertBefore);
4154 }
4155
4156 // This helper takes a nested struct and extracts a part of it (which is again a
4157 // struct) into a new value. For example, given the struct:
4158 // { a, { b, { c, d }, e } }
4159 // and the indices "1, 1" this returns
4160 // { c, d }.
4161 //
4162 // It does this by inserting an insertvalue for each element in the resulting
4163 // struct, as opposed to just inserting a single struct. This will only work if
4164 // each of the elements of the substruct are known (ie, inserted into From by an
4165 // insertvalue instruction somewhere).
4166 //
4167 // All inserted insertvalue instructions are inserted before InsertBefore
BuildSubAggregate(Value * From,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)4168 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
4169 Instruction *InsertBefore) {
4170 assert(InsertBefore && "Must have someplace to insert!");
4171 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
4172 idx_range);
4173 Value *To = PoisonValue::get(IndexedType);
4174 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
4175 unsigned IdxSkip = Idxs.size();
4176
4177 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
4178 }
4179
4180 /// Given an aggregate and a sequence of indices, see if the scalar value
4181 /// indexed is already around as a register, for example if it was inserted
4182 /// directly into the aggregate.
4183 ///
4184 /// If InsertBefore is not null, this function will duplicate (modified)
4185 /// insertvalues when a part of a nested struct is extracted.
FindInsertedValue(Value * V,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)4186 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
4187 Instruction *InsertBefore) {
4188 // Nothing to index? Just return V then (this is useful at the end of our
4189 // recursion).
4190 if (idx_range.empty())
4191 return V;
4192 // We have indices, so V should have an indexable type.
4193 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
4194 "Not looking at a struct or array?");
4195 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
4196 "Invalid indices for type?");
4197
4198 if (Constant *C = dyn_cast<Constant>(V)) {
4199 C = C->getAggregateElement(idx_range[0]);
4200 if (!C) return nullptr;
4201 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
4202 }
4203
4204 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
4205 // Loop the indices for the insertvalue instruction in parallel with the
4206 // requested indices
4207 const unsigned *req_idx = idx_range.begin();
4208 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
4209 i != e; ++i, ++req_idx) {
4210 if (req_idx == idx_range.end()) {
4211 // We can't handle this without inserting insertvalues
4212 if (!InsertBefore)
4213 return nullptr;
4214
4215 // The requested index identifies a part of a nested aggregate. Handle
4216 // this specially. For example,
4217 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
4218 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
4219 // %C = extractvalue {i32, { i32, i32 } } %B, 1
4220 // This can be changed into
4221 // %A = insertvalue {i32, i32 } undef, i32 10, 0
4222 // %C = insertvalue {i32, i32 } %A, i32 11, 1
4223 // which allows the unused 0,0 element from the nested struct to be
4224 // removed.
4225 return BuildSubAggregate(V, ArrayRef(idx_range.begin(), req_idx),
4226 InsertBefore);
4227 }
4228
4229 // This insert value inserts something else than what we are looking for.
4230 // See if the (aggregate) value inserted into has the value we are
4231 // looking for, then.
4232 if (*req_idx != *i)
4233 return FindInsertedValue(I->getAggregateOperand(), idx_range,
4234 InsertBefore);
4235 }
4236 // If we end up here, the indices of the insertvalue match with those
4237 // requested (though possibly only partially). Now we recursively look at
4238 // the inserted value, passing any remaining indices.
4239 return FindInsertedValue(I->getInsertedValueOperand(),
4240 ArrayRef(req_idx, idx_range.end()), InsertBefore);
4241 }
4242
4243 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
4244 // If we're extracting a value from an aggregate that was extracted from
4245 // something else, we can extract from that something else directly instead.
4246 // However, we will need to chain I's indices with the requested indices.
4247
4248 // Calculate the number of indices required
4249 unsigned size = I->getNumIndices() + idx_range.size();
4250 // Allocate some space to put the new indices in
4251 SmallVector<unsigned, 5> Idxs;
4252 Idxs.reserve(size);
4253 // Add indices from the extract value instruction
4254 Idxs.append(I->idx_begin(), I->idx_end());
4255
4256 // Add requested indices
4257 Idxs.append(idx_range.begin(), idx_range.end());
4258
4259 assert(Idxs.size() == size
4260 && "Number of indices added not correct?");
4261
4262 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
4263 }
4264 // Otherwise, we don't know (such as, extracting from a function return value
4265 // or load instruction)
4266 return nullptr;
4267 }
4268
isGEPBasedOnPointerToString(const GEPOperator * GEP,unsigned CharSize)4269 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
4270 unsigned CharSize) {
4271 // Make sure the GEP has exactly three arguments.
4272 if (GEP->getNumOperands() != 3)
4273 return false;
4274
4275 // Make sure the index-ee is a pointer to array of \p CharSize integers.
4276 // CharSize.
4277 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
4278 if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
4279 return false;
4280
4281 // Check to make sure that the first operand of the GEP is an integer and
4282 // has value 0 so that we are sure we're indexing into the initializer.
4283 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
4284 if (!FirstIdx || !FirstIdx->isZero())
4285 return false;
4286
4287 return true;
4288 }
4289
4290 // If V refers to an initialized global constant, set Slice either to
4291 // its initializer if the size of its elements equals ElementSize, or,
4292 // for ElementSize == 8, to its representation as an array of unsiged
4293 // char. Return true on success.
4294 // Offset is in the unit "nr of ElementSize sized elements".
getConstantDataArrayInfo(const Value * V,ConstantDataArraySlice & Slice,unsigned ElementSize,uint64_t Offset)4295 bool llvm::getConstantDataArrayInfo(const Value *V,
4296 ConstantDataArraySlice &Slice,
4297 unsigned ElementSize, uint64_t Offset) {
4298 assert(V && "V should not be null.");
4299 assert((ElementSize % 8) == 0 &&
4300 "ElementSize expected to be a multiple of the size of a byte.");
4301 unsigned ElementSizeInBytes = ElementSize / 8;
4302
4303 // Drill down into the pointer expression V, ignoring any intervening
4304 // casts, and determine the identity of the object it references along
4305 // with the cumulative byte offset into it.
4306 const GlobalVariable *GV =
4307 dyn_cast<GlobalVariable>(getUnderlyingObject(V));
4308 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
4309 // Fail if V is not based on constant global object.
4310 return false;
4311
4312 const DataLayout &DL = GV->getParent()->getDataLayout();
4313 APInt Off(DL.getIndexTypeSizeInBits(V->getType()), 0);
4314
4315 if (GV != V->stripAndAccumulateConstantOffsets(DL, Off,
4316 /*AllowNonInbounds*/ true))
4317 // Fail if a constant offset could not be determined.
4318 return false;
4319
4320 uint64_t StartIdx = Off.getLimitedValue();
4321 if (StartIdx == UINT64_MAX)
4322 // Fail if the constant offset is excessive.
4323 return false;
4324
4325 // Off/StartIdx is in the unit of bytes. So we need to convert to number of
4326 // elements. Simply bail out if that isn't possible.
4327 if ((StartIdx % ElementSizeInBytes) != 0)
4328 return false;
4329
4330 Offset += StartIdx / ElementSizeInBytes;
4331 ConstantDataArray *Array = nullptr;
4332 ArrayType *ArrayTy = nullptr;
4333
4334 if (GV->getInitializer()->isNullValue()) {
4335 Type *GVTy = GV->getValueType();
4336 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedValue();
4337 uint64_t Length = SizeInBytes / ElementSizeInBytes;
4338
4339 Slice.Array = nullptr;
4340 Slice.Offset = 0;
4341 // Return an empty Slice for undersized constants to let callers
4342 // transform even undefined library calls into simpler, well-defined
4343 // expressions. This is preferable to making the calls although it
4344 // prevents sanitizers from detecting such calls.
4345 Slice.Length = Length < Offset ? 0 : Length - Offset;
4346 return true;
4347 }
4348
4349 auto *Init = const_cast<Constant *>(GV->getInitializer());
4350 if (auto *ArrayInit = dyn_cast<ConstantDataArray>(Init)) {
4351 Type *InitElTy = ArrayInit->getElementType();
4352 if (InitElTy->isIntegerTy(ElementSize)) {
4353 // If Init is an initializer for an array of the expected type
4354 // and size, use it as is.
4355 Array = ArrayInit;
4356 ArrayTy = ArrayInit->getType();
4357 }
4358 }
4359
4360 if (!Array) {
4361 if (ElementSize != 8)
4362 // TODO: Handle conversions to larger integral types.
4363 return false;
4364
4365 // Otherwise extract the portion of the initializer starting
4366 // at Offset as an array of bytes, and reset Offset.
4367 Init = ReadByteArrayFromGlobal(GV, Offset);
4368 if (!Init)
4369 return false;
4370
4371 Offset = 0;
4372 Array = dyn_cast<ConstantDataArray>(Init);
4373 ArrayTy = dyn_cast<ArrayType>(Init->getType());
4374 }
4375
4376 uint64_t NumElts = ArrayTy->getArrayNumElements();
4377 if (Offset > NumElts)
4378 return false;
4379
4380 Slice.Array = Array;
4381 Slice.Offset = Offset;
4382 Slice.Length = NumElts - Offset;
4383 return true;
4384 }
4385
4386 /// Extract bytes from the initializer of the constant array V, which need
4387 /// not be a nul-terminated string. On success, store the bytes in Str and
4388 /// return true. When TrimAtNul is set, Str will contain only the bytes up
4389 /// to but not including the first nul. Return false on failure.
getConstantStringInfo(const Value * V,StringRef & Str,bool TrimAtNul)4390 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
4391 bool TrimAtNul) {
4392 ConstantDataArraySlice Slice;
4393 if (!getConstantDataArrayInfo(V, Slice, 8))
4394 return false;
4395
4396 if (Slice.Array == nullptr) {
4397 if (TrimAtNul) {
4398 // Return a nul-terminated string even for an empty Slice. This is
4399 // safe because all existing SimplifyLibcalls callers require string
4400 // arguments and the behavior of the functions they fold is undefined
4401 // otherwise. Folding the calls this way is preferable to making
4402 // the undefined library calls, even though it prevents sanitizers
4403 // from reporting such calls.
4404 Str = StringRef();
4405 return true;
4406 }
4407 if (Slice.Length == 1) {
4408 Str = StringRef("", 1);
4409 return true;
4410 }
4411 // We cannot instantiate a StringRef as we do not have an appropriate string
4412 // of 0s at hand.
4413 return false;
4414 }
4415
4416 // Start out with the entire array in the StringRef.
4417 Str = Slice.Array->getAsString();
4418 // Skip over 'offset' bytes.
4419 Str = Str.substr(Slice.Offset);
4420
4421 if (TrimAtNul) {
4422 // Trim off the \0 and anything after it. If the array is not nul
4423 // terminated, we just return the whole end of string. The client may know
4424 // some other way that the string is length-bound.
4425 Str = Str.substr(0, Str.find('\0'));
4426 }
4427 return true;
4428 }
4429
4430 // These next two are very similar to the above, but also look through PHI
4431 // nodes.
4432 // TODO: See if we can integrate these two together.
4433
4434 /// If we can compute the length of the string pointed to by
4435 /// the specified pointer, return 'len+1'. If we can't, return 0.
GetStringLengthH(const Value * V,SmallPtrSetImpl<const PHINode * > & PHIs,unsigned CharSize)4436 static uint64_t GetStringLengthH(const Value *V,
4437 SmallPtrSetImpl<const PHINode*> &PHIs,
4438 unsigned CharSize) {
4439 // Look through noop bitcast instructions.
4440 V = V->stripPointerCasts();
4441
4442 // If this is a PHI node, there are two cases: either we have already seen it
4443 // or we haven't.
4444 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
4445 if (!PHIs.insert(PN).second)
4446 return ~0ULL; // already in the set.
4447
4448 // If it was new, see if all the input strings are the same length.
4449 uint64_t LenSoFar = ~0ULL;
4450 for (Value *IncValue : PN->incoming_values()) {
4451 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
4452 if (Len == 0) return 0; // Unknown length -> unknown.
4453
4454 if (Len == ~0ULL) continue;
4455
4456 if (Len != LenSoFar && LenSoFar != ~0ULL)
4457 return 0; // Disagree -> unknown.
4458 LenSoFar = Len;
4459 }
4460
4461 // Success, all agree.
4462 return LenSoFar;
4463 }
4464
4465 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
4466 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
4467 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
4468 if (Len1 == 0) return 0;
4469 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
4470 if (Len2 == 0) return 0;
4471 if (Len1 == ~0ULL) return Len2;
4472 if (Len2 == ~0ULL) return Len1;
4473 if (Len1 != Len2) return 0;
4474 return Len1;
4475 }
4476
4477 // Otherwise, see if we can read the string.
4478 ConstantDataArraySlice Slice;
4479 if (!getConstantDataArrayInfo(V, Slice, CharSize))
4480 return 0;
4481
4482 if (Slice.Array == nullptr)
4483 // Zeroinitializer (including an empty one).
4484 return 1;
4485
4486 // Search for the first nul character. Return a conservative result even
4487 // when there is no nul. This is safe since otherwise the string function
4488 // being folded such as strlen is undefined, and can be preferable to
4489 // making the undefined library call.
4490 unsigned NullIndex = 0;
4491 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
4492 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
4493 break;
4494 }
4495
4496 return NullIndex + 1;
4497 }
4498
4499 /// If we can compute the length of the string pointed to by
4500 /// the specified pointer, return 'len+1'. If we can't, return 0.
GetStringLength(const Value * V,unsigned CharSize)4501 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
4502 if (!V->getType()->isPointerTy())
4503 return 0;
4504
4505 SmallPtrSet<const PHINode*, 32> PHIs;
4506 uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
4507 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
4508 // an empty string as a length.
4509 return Len == ~0ULL ? 1 : Len;
4510 }
4511
4512 const Value *
getArgumentAliasingToReturnedPointer(const CallBase * Call,bool MustPreserveNullness)4513 llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call,
4514 bool MustPreserveNullness) {
4515 assert(Call &&
4516 "getArgumentAliasingToReturnedPointer only works on nonnull calls");
4517 if (const Value *RV = Call->getReturnedArgOperand())
4518 return RV;
4519 // This can be used only as a aliasing property.
4520 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
4521 Call, MustPreserveNullness))
4522 return Call->getArgOperand(0);
4523 return nullptr;
4524 }
4525
isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(const CallBase * Call,bool MustPreserveNullness)4526 bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
4527 const CallBase *Call, bool MustPreserveNullness) {
4528 switch (Call->getIntrinsicID()) {
4529 case Intrinsic::launder_invariant_group:
4530 case Intrinsic::strip_invariant_group:
4531 case Intrinsic::aarch64_irg:
4532 case Intrinsic::aarch64_tagp:
4533 return true;
4534 case Intrinsic::ptrmask:
4535 return !MustPreserveNullness;
4536 default:
4537 return false;
4538 }
4539 }
4540
4541 /// \p PN defines a loop-variant pointer to an object. Check if the
4542 /// previous iteration of the loop was referring to the same object as \p PN.
isSameUnderlyingObjectInLoop(const PHINode * PN,const LoopInfo * LI)4543 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
4544 const LoopInfo *LI) {
4545 // Find the loop-defined value.
4546 Loop *L = LI->getLoopFor(PN->getParent());
4547 if (PN->getNumIncomingValues() != 2)
4548 return true;
4549
4550 // Find the value from previous iteration.
4551 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
4552 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4553 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
4554 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
4555 return true;
4556
4557 // If a new pointer is loaded in the loop, the pointer references a different
4558 // object in every iteration. E.g.:
4559 // for (i)
4560 // int *p = a[i];
4561 // ...
4562 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
4563 if (!L->isLoopInvariant(Load->getPointerOperand()))
4564 return false;
4565 return true;
4566 }
4567
getUnderlyingObject(const Value * V,unsigned MaxLookup)4568 const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) {
4569 if (!V->getType()->isPointerTy())
4570 return V;
4571 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
4572 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
4573 V = GEP->getPointerOperand();
4574 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
4575 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
4576 V = cast<Operator>(V)->getOperand(0);
4577 if (!V->getType()->isPointerTy())
4578 return V;
4579 } else if (auto *GA = dyn_cast<GlobalAlias>(V)) {
4580 if (GA->isInterposable())
4581 return V;
4582 V = GA->getAliasee();
4583 } else {
4584 if (auto *PHI = dyn_cast<PHINode>(V)) {
4585 // Look through single-arg phi nodes created by LCSSA.
4586 if (PHI->getNumIncomingValues() == 1) {
4587 V = PHI->getIncomingValue(0);
4588 continue;
4589 }
4590 } else if (auto *Call = dyn_cast<CallBase>(V)) {
4591 // CaptureTracking can know about special capturing properties of some
4592 // intrinsics like launder.invariant.group, that can't be expressed with
4593 // the attributes, but have properties like returning aliasing pointer.
4594 // Because some analysis may assume that nocaptured pointer is not
4595 // returned from some special intrinsic (because function would have to
4596 // be marked with returns attribute), it is crucial to use this function
4597 // because it should be in sync with CaptureTracking. Not using it may
4598 // cause weird miscompilations where 2 aliasing pointers are assumed to
4599 // noalias.
4600 if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
4601 V = RP;
4602 continue;
4603 }
4604 }
4605
4606 return V;
4607 }
4608 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
4609 }
4610 return V;
4611 }
4612
getUnderlyingObjects(const Value * V,SmallVectorImpl<const Value * > & Objects,LoopInfo * LI,unsigned MaxLookup)4613 void llvm::getUnderlyingObjects(const Value *V,
4614 SmallVectorImpl<const Value *> &Objects,
4615 LoopInfo *LI, unsigned MaxLookup) {
4616 SmallPtrSet<const Value *, 4> Visited;
4617 SmallVector<const Value *, 4> Worklist;
4618 Worklist.push_back(V);
4619 do {
4620 const Value *P = Worklist.pop_back_val();
4621 P = getUnderlyingObject(P, MaxLookup);
4622
4623 if (!Visited.insert(P).second)
4624 continue;
4625
4626 if (auto *SI = dyn_cast<SelectInst>(P)) {
4627 Worklist.push_back(SI->getTrueValue());
4628 Worklist.push_back(SI->getFalseValue());
4629 continue;
4630 }
4631
4632 if (auto *PN = dyn_cast<PHINode>(P)) {
4633 // If this PHI changes the underlying object in every iteration of the
4634 // loop, don't look through it. Consider:
4635 // int **A;
4636 // for (i) {
4637 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
4638 // Curr = A[i];
4639 // *Prev, *Curr;
4640 //
4641 // Prev is tracking Curr one iteration behind so they refer to different
4642 // underlying objects.
4643 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
4644 isSameUnderlyingObjectInLoop(PN, LI))
4645 append_range(Worklist, PN->incoming_values());
4646 continue;
4647 }
4648
4649 Objects.push_back(P);
4650 } while (!Worklist.empty());
4651 }
4652
4653 /// This is the function that does the work of looking through basic
4654 /// ptrtoint+arithmetic+inttoptr sequences.
getUnderlyingObjectFromInt(const Value * V)4655 static const Value *getUnderlyingObjectFromInt(const Value *V) {
4656 do {
4657 if (const Operator *U = dyn_cast<Operator>(V)) {
4658 // If we find a ptrtoint, we can transfer control back to the
4659 // regular getUnderlyingObjectFromInt.
4660 if (U->getOpcode() == Instruction::PtrToInt)
4661 return U->getOperand(0);
4662 // If we find an add of a constant, a multiplied value, or a phi, it's
4663 // likely that the other operand will lead us to the base
4664 // object. We don't have to worry about the case where the
4665 // object address is somehow being computed by the multiply,
4666 // because our callers only care when the result is an
4667 // identifiable object.
4668 if (U->getOpcode() != Instruction::Add ||
4669 (!isa<ConstantInt>(U->getOperand(1)) &&
4670 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
4671 !isa<PHINode>(U->getOperand(1))))
4672 return V;
4673 V = U->getOperand(0);
4674 } else {
4675 return V;
4676 }
4677 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
4678 } while (true);
4679 }
4680
4681 /// This is a wrapper around getUnderlyingObjects and adds support for basic
4682 /// ptrtoint+arithmetic+inttoptr sequences.
4683 /// It returns false if unidentified object is found in getUnderlyingObjects.
getUnderlyingObjectsForCodeGen(const Value * V,SmallVectorImpl<Value * > & Objects)4684 bool llvm::getUnderlyingObjectsForCodeGen(const Value *V,
4685 SmallVectorImpl<Value *> &Objects) {
4686 SmallPtrSet<const Value *, 16> Visited;
4687 SmallVector<const Value *, 4> Working(1, V);
4688 do {
4689 V = Working.pop_back_val();
4690
4691 SmallVector<const Value *, 4> Objs;
4692 getUnderlyingObjects(V, Objs);
4693
4694 for (const Value *V : Objs) {
4695 if (!Visited.insert(V).second)
4696 continue;
4697 if (Operator::getOpcode(V) == Instruction::IntToPtr) {
4698 const Value *O =
4699 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
4700 if (O->getType()->isPointerTy()) {
4701 Working.push_back(O);
4702 continue;
4703 }
4704 }
4705 // If getUnderlyingObjects fails to find an identifiable object,
4706 // getUnderlyingObjectsForCodeGen also fails for safety.
4707 if (!isIdentifiedObject(V)) {
4708 Objects.clear();
4709 return false;
4710 }
4711 Objects.push_back(const_cast<Value *>(V));
4712 }
4713 } while (!Working.empty());
4714 return true;
4715 }
4716
findAllocaForValue(Value * V,bool OffsetZero)4717 AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) {
4718 AllocaInst *Result = nullptr;
4719 SmallPtrSet<Value *, 4> Visited;
4720 SmallVector<Value *, 4> Worklist;
4721
4722 auto AddWork = [&](Value *V) {
4723 if (Visited.insert(V).second)
4724 Worklist.push_back(V);
4725 };
4726
4727 AddWork(V);
4728 do {
4729 V = Worklist.pop_back_val();
4730 assert(Visited.count(V));
4731
4732 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
4733 if (Result && Result != AI)
4734 return nullptr;
4735 Result = AI;
4736 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
4737 AddWork(CI->getOperand(0));
4738 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
4739 for (Value *IncValue : PN->incoming_values())
4740 AddWork(IncValue);
4741 } else if (auto *SI = dyn_cast<SelectInst>(V)) {
4742 AddWork(SI->getTrueValue());
4743 AddWork(SI->getFalseValue());
4744 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
4745 if (OffsetZero && !GEP->hasAllZeroIndices())
4746 return nullptr;
4747 AddWork(GEP->getPointerOperand());
4748 } else if (CallBase *CB = dyn_cast<CallBase>(V)) {
4749 Value *Returned = CB->getReturnedArgOperand();
4750 if (Returned)
4751 AddWork(Returned);
4752 else
4753 return nullptr;
4754 } else {
4755 return nullptr;
4756 }
4757 } while (!Worklist.empty());
4758
4759 return Result;
4760 }
4761
onlyUsedByLifetimeMarkersOrDroppableInstsHelper(const Value * V,bool AllowLifetime,bool AllowDroppable)4762 static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper(
4763 const Value *V, bool AllowLifetime, bool AllowDroppable) {
4764 for (const User *U : V->users()) {
4765 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
4766 if (!II)
4767 return false;
4768
4769 if (AllowLifetime && II->isLifetimeStartOrEnd())
4770 continue;
4771
4772 if (AllowDroppable && II->isDroppable())
4773 continue;
4774
4775 return false;
4776 }
4777 return true;
4778 }
4779
onlyUsedByLifetimeMarkers(const Value * V)4780 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
4781 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper(
4782 V, /* AllowLifetime */ true, /* AllowDroppable */ false);
4783 }
onlyUsedByLifetimeMarkersOrDroppableInsts(const Value * V)4784 bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) {
4785 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper(
4786 V, /* AllowLifetime */ true, /* AllowDroppable */ true);
4787 }
4788
mustSuppressSpeculation(const LoadInst & LI)4789 bool llvm::mustSuppressSpeculation(const LoadInst &LI) {
4790 if (!LI.isUnordered())
4791 return true;
4792 const Function &F = *LI.getFunction();
4793 // Speculative load may create a race that did not exist in the source.
4794 return F.hasFnAttribute(Attribute::SanitizeThread) ||
4795 // Speculative load may load data from dirty regions.
4796 F.hasFnAttribute(Attribute::SanitizeAddress) ||
4797 F.hasFnAttribute(Attribute::SanitizeHWAddress);
4798 }
4799
isSafeToSpeculativelyExecute(const Instruction * Inst,const Instruction * CtxI,AssumptionCache * AC,const DominatorTree * DT,const TargetLibraryInfo * TLI)4800 bool llvm::isSafeToSpeculativelyExecute(const Instruction *Inst,
4801 const Instruction *CtxI,
4802 AssumptionCache *AC,
4803 const DominatorTree *DT,
4804 const TargetLibraryInfo *TLI) {
4805 return isSafeToSpeculativelyExecuteWithOpcode(Inst->getOpcode(), Inst, CtxI,
4806 AC, DT, TLI);
4807 }
4808
isSafeToSpeculativelyExecuteWithOpcode(unsigned Opcode,const Instruction * Inst,const Instruction * CtxI,AssumptionCache * AC,const DominatorTree * DT,const TargetLibraryInfo * TLI)4809 bool llvm::isSafeToSpeculativelyExecuteWithOpcode(
4810 unsigned Opcode, const Instruction *Inst, const Instruction *CtxI,
4811 AssumptionCache *AC, const DominatorTree *DT,
4812 const TargetLibraryInfo *TLI) {
4813 #ifndef NDEBUG
4814 if (Inst->getOpcode() != Opcode) {
4815 // Check that the operands are actually compatible with the Opcode override.
4816 auto hasEqualReturnAndLeadingOperandTypes =
4817 [](const Instruction *Inst, unsigned NumLeadingOperands) {
4818 if (Inst->getNumOperands() < NumLeadingOperands)
4819 return false;
4820 const Type *ExpectedType = Inst->getType();
4821 for (unsigned ItOp = 0; ItOp < NumLeadingOperands; ++ItOp)
4822 if (Inst->getOperand(ItOp)->getType() != ExpectedType)
4823 return false;
4824 return true;
4825 };
4826 assert(!Instruction::isBinaryOp(Opcode) ||
4827 hasEqualReturnAndLeadingOperandTypes(Inst, 2));
4828 assert(!Instruction::isUnaryOp(Opcode) ||
4829 hasEqualReturnAndLeadingOperandTypes(Inst, 1));
4830 }
4831 #endif
4832
4833 switch (Opcode) {
4834 default:
4835 return true;
4836 case Instruction::UDiv:
4837 case Instruction::URem: {
4838 // x / y is undefined if y == 0.
4839 const APInt *V;
4840 if (match(Inst->getOperand(1), m_APInt(V)))
4841 return *V != 0;
4842 return false;
4843 }
4844 case Instruction::SDiv:
4845 case Instruction::SRem: {
4846 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
4847 const APInt *Numerator, *Denominator;
4848 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
4849 return false;
4850 // We cannot hoist this division if the denominator is 0.
4851 if (*Denominator == 0)
4852 return false;
4853 // It's safe to hoist if the denominator is not 0 or -1.
4854 if (!Denominator->isAllOnes())
4855 return true;
4856 // At this point we know that the denominator is -1. It is safe to hoist as
4857 // long we know that the numerator is not INT_MIN.
4858 if (match(Inst->getOperand(0), m_APInt(Numerator)))
4859 return !Numerator->isMinSignedValue();
4860 // The numerator *might* be MinSignedValue.
4861 return false;
4862 }
4863 case Instruction::Load: {
4864 const LoadInst *LI = dyn_cast<LoadInst>(Inst);
4865 if (!LI)
4866 return false;
4867 if (mustSuppressSpeculation(*LI))
4868 return false;
4869 const DataLayout &DL = LI->getModule()->getDataLayout();
4870 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
4871 LI->getType(), LI->getAlign(), DL,
4872 CtxI, AC, DT, TLI);
4873 }
4874 case Instruction::Call: {
4875 auto *CI = dyn_cast<const CallInst>(Inst);
4876 if (!CI)
4877 return false;
4878 const Function *Callee = CI->getCalledFunction();
4879
4880 // The called function could have undefined behavior or side-effects, even
4881 // if marked readnone nounwind.
4882 return Callee && Callee->isSpeculatable();
4883 }
4884 case Instruction::VAArg:
4885 case Instruction::Alloca:
4886 case Instruction::Invoke:
4887 case Instruction::CallBr:
4888 case Instruction::PHI:
4889 case Instruction::Store:
4890 case Instruction::Ret:
4891 case Instruction::Br:
4892 case Instruction::IndirectBr:
4893 case Instruction::Switch:
4894 case Instruction::Unreachable:
4895 case Instruction::Fence:
4896 case Instruction::AtomicRMW:
4897 case Instruction::AtomicCmpXchg:
4898 case Instruction::LandingPad:
4899 case Instruction::Resume:
4900 case Instruction::CatchSwitch:
4901 case Instruction::CatchPad:
4902 case Instruction::CatchRet:
4903 case Instruction::CleanupPad:
4904 case Instruction::CleanupRet:
4905 return false; // Misc instructions which have effects
4906 }
4907 }
4908
mayHaveNonDefUseDependency(const Instruction & I)4909 bool llvm::mayHaveNonDefUseDependency(const Instruction &I) {
4910 if (I.mayReadOrWriteMemory())
4911 // Memory dependency possible
4912 return true;
4913 if (!isSafeToSpeculativelyExecute(&I))
4914 // Can't move above a maythrow call or infinite loop. Or if an
4915 // inalloca alloca, above a stacksave call.
4916 return true;
4917 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
4918 // 1) Can't reorder two inf-loop calls, even if readonly
4919 // 2) Also can't reorder an inf-loop call below a instruction which isn't
4920 // safe to speculative execute. (Inverse of above)
4921 return true;
4922 return false;
4923 }
4924
4925 /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult.
mapOverflowResult(ConstantRange::OverflowResult OR)4926 static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) {
4927 switch (OR) {
4928 case ConstantRange::OverflowResult::MayOverflow:
4929 return OverflowResult::MayOverflow;
4930 case ConstantRange::OverflowResult::AlwaysOverflowsLow:
4931 return OverflowResult::AlwaysOverflowsLow;
4932 case ConstantRange::OverflowResult::AlwaysOverflowsHigh:
4933 return OverflowResult::AlwaysOverflowsHigh;
4934 case ConstantRange::OverflowResult::NeverOverflows:
4935 return OverflowResult::NeverOverflows;
4936 }
4937 llvm_unreachable("Unknown OverflowResult");
4938 }
4939
4940 /// Combine constant ranges from computeConstantRange() and computeKnownBits().
computeConstantRangeIncludingKnownBits(const Value * V,bool ForSigned,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE=nullptr,bool UseInstrInfo=true)4941 static ConstantRange computeConstantRangeIncludingKnownBits(
4942 const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth,
4943 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4944 OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) {
4945 KnownBits Known = computeKnownBits(
4946 V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo);
4947 ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned);
4948 ConstantRange CR2 = computeConstantRange(V, UseInstrInfo);
4949 ConstantRange::PreferredRangeType RangeType =
4950 ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned;
4951 return CR1.intersectWith(CR2, RangeType);
4952 }
4953
computeOverflowForUnsignedMul(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)4954 OverflowResult llvm::computeOverflowForUnsignedMul(
4955 const Value *LHS, const Value *RHS, const DataLayout &DL,
4956 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4957 bool UseInstrInfo) {
4958 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4959 nullptr, UseInstrInfo);
4960 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4961 nullptr, UseInstrInfo);
4962 ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false);
4963 ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false);
4964 return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange));
4965 }
4966
4967 OverflowResult
computeOverflowForSignedMul(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)4968 llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS,
4969 const DataLayout &DL, AssumptionCache *AC,
4970 const Instruction *CxtI,
4971 const DominatorTree *DT, bool UseInstrInfo) {
4972 // Multiplying n * m significant bits yields a result of n + m significant
4973 // bits. If the total number of significant bits does not exceed the
4974 // result bit width (minus 1), there is no overflow.
4975 // This means if we have enough leading sign bits in the operands
4976 // we can guarantee that the result does not overflow.
4977 // Ref: "Hacker's Delight" by Henry Warren
4978 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
4979
4980 // Note that underestimating the number of sign bits gives a more
4981 // conservative answer.
4982 unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
4983 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
4984
4985 // First handle the easy case: if we have enough sign bits there's
4986 // definitely no overflow.
4987 if (SignBits > BitWidth + 1)
4988 return OverflowResult::NeverOverflows;
4989
4990 // There are two ambiguous cases where there can be no overflow:
4991 // SignBits == BitWidth + 1 and
4992 // SignBits == BitWidth
4993 // The second case is difficult to check, therefore we only handle the
4994 // first case.
4995 if (SignBits == BitWidth + 1) {
4996 // It overflows only when both arguments are negative and the true
4997 // product is exactly the minimum negative number.
4998 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4999 // For simplicity we just check if at least one side is not negative.
5000 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
5001 nullptr, UseInstrInfo);
5002 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
5003 nullptr, UseInstrInfo);
5004 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
5005 return OverflowResult::NeverOverflows;
5006 }
5007 return OverflowResult::MayOverflow;
5008 }
5009
computeOverflowForUnsignedAdd(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,bool UseInstrInfo)5010 OverflowResult llvm::computeOverflowForUnsignedAdd(
5011 const Value *LHS, const Value *RHS, const DataLayout &DL,
5012 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
5013 bool UseInstrInfo) {
5014 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
5015 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
5016 nullptr, UseInstrInfo);
5017 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
5018 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
5019 nullptr, UseInstrInfo);
5020 return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange));
5021 }
5022
computeOverflowForSignedAdd(const Value * LHS,const Value * RHS,const AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)5023 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
5024 const Value *RHS,
5025 const AddOperator *Add,
5026 const DataLayout &DL,
5027 AssumptionCache *AC,
5028 const Instruction *CxtI,
5029 const DominatorTree *DT) {
5030 if (Add && Add->hasNoSignedWrap()) {
5031 return OverflowResult::NeverOverflows;
5032 }
5033
5034 // If LHS and RHS each have at least two sign bits, the addition will look
5035 // like
5036 //
5037 // XX..... +
5038 // YY.....
5039 //
5040 // If the carry into the most significant position is 0, X and Y can't both
5041 // be 1 and therefore the carry out of the addition is also 0.
5042 //
5043 // If the carry into the most significant position is 1, X and Y can't both
5044 // be 0 and therefore the carry out of the addition is also 1.
5045 //
5046 // Since the carry into the most significant position is always equal to
5047 // the carry out of the addition, there is no signed overflow.
5048 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
5049 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
5050 return OverflowResult::NeverOverflows;
5051
5052 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
5053 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
5054 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
5055 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
5056 OverflowResult OR =
5057 mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange));
5058 if (OR != OverflowResult::MayOverflow)
5059 return OR;
5060
5061 // The remaining code needs Add to be available. Early returns if not so.
5062 if (!Add)
5063 return OverflowResult::MayOverflow;
5064
5065 // If the sign of Add is the same as at least one of the operands, this add
5066 // CANNOT overflow. If this can be determined from the known bits of the
5067 // operands the above signedAddMayOverflow() check will have already done so.
5068 // The only other way to improve on the known bits is from an assumption, so
5069 // call computeKnownBitsFromAssume() directly.
5070 bool LHSOrRHSKnownNonNegative =
5071 (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative());
5072 bool LHSOrRHSKnownNegative =
5073 (LHSRange.isAllNegative() || RHSRange.isAllNegative());
5074 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
5075 KnownBits AddKnown(LHSRange.getBitWidth());
5076 computeKnownBitsFromAssume(
5077 Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true));
5078 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
5079 (AddKnown.isNegative() && LHSOrRHSKnownNegative))
5080 return OverflowResult::NeverOverflows;
5081 }
5082
5083 return OverflowResult::MayOverflow;
5084 }
5085
computeOverflowForUnsignedSub(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)5086 OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS,
5087 const Value *RHS,
5088 const DataLayout &DL,
5089 AssumptionCache *AC,
5090 const Instruction *CxtI,
5091 const DominatorTree *DT) {
5092 // X - (X % ?)
5093 // The remainder of a value can't have greater magnitude than itself,
5094 // so the subtraction can't overflow.
5095
5096 // X - (X -nuw ?)
5097 // In the minimal case, this would simplify to "?", so there's no subtract
5098 // at all. But if this analysis is used to peek through casts, for example,
5099 // then determining no-overflow may allow other transforms.
5100
5101 // TODO: There are other patterns like this.
5102 // See simplifyICmpWithBinOpOnLHS() for candidates.
5103 if (match(RHS, m_URem(m_Specific(LHS), m_Value())) ||
5104 match(RHS, m_NUWSub(m_Specific(LHS), m_Value())))
5105 if (isGuaranteedNotToBeUndefOrPoison(LHS, AC, CxtI, DT))
5106 return OverflowResult::NeverOverflows;
5107
5108 // Checking for conditions implied by dominating conditions may be expensive.
5109 // Limit it to usub_with_overflow calls for now.
5110 if (match(CxtI,
5111 m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value())))
5112 if (auto C =
5113 isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) {
5114 if (*C)
5115 return OverflowResult::NeverOverflows;
5116 return OverflowResult::AlwaysOverflowsLow;
5117 }
5118 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
5119 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
5120 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
5121 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
5122 return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange));
5123 }
5124
computeOverflowForSignedSub(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)5125 OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS,
5126 const Value *RHS,
5127 const DataLayout &DL,
5128 AssumptionCache *AC,
5129 const Instruction *CxtI,
5130 const DominatorTree *DT) {
5131 // X - (X % ?)
5132 // The remainder of a value can't have greater magnitude than itself,
5133 // so the subtraction can't overflow.
5134
5135 // X - (X -nsw ?)
5136 // In the minimal case, this would simplify to "?", so there's no subtract
5137 // at all. But if this analysis is used to peek through casts, for example,
5138 // then determining no-overflow may allow other transforms.
5139 if (match(RHS, m_SRem(m_Specific(LHS), m_Value())) ||
5140 match(RHS, m_NSWSub(m_Specific(LHS), m_Value())))
5141 if (isGuaranteedNotToBeUndefOrPoison(LHS, AC, CxtI, DT))
5142 return OverflowResult::NeverOverflows;
5143
5144 // If LHS and RHS each have at least two sign bits, the subtraction
5145 // cannot overflow.
5146 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
5147 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
5148 return OverflowResult::NeverOverflows;
5149
5150 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
5151 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
5152 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
5153 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
5154 return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange));
5155 }
5156
isOverflowIntrinsicNoWrap(const WithOverflowInst * WO,const DominatorTree & DT)5157 bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO,
5158 const DominatorTree &DT) {
5159 SmallVector<const BranchInst *, 2> GuardingBranches;
5160 SmallVector<const ExtractValueInst *, 2> Results;
5161
5162 for (const User *U : WO->users()) {
5163 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
5164 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
5165
5166 if (EVI->getIndices()[0] == 0)
5167 Results.push_back(EVI);
5168 else {
5169 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
5170
5171 for (const auto *U : EVI->users())
5172 if (const auto *B = dyn_cast<BranchInst>(U)) {
5173 assert(B->isConditional() && "How else is it using an i1?");
5174 GuardingBranches.push_back(B);
5175 }
5176 }
5177 } else {
5178 // We are using the aggregate directly in a way we don't want to analyze
5179 // here (storing it to a global, say).
5180 return false;
5181 }
5182 }
5183
5184 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
5185 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
5186 if (!NoWrapEdge.isSingleEdge())
5187 return false;
5188
5189 // Check if all users of the add are provably no-wrap.
5190 for (const auto *Result : Results) {
5191 // If the extractvalue itself is not executed on overflow, the we don't
5192 // need to check each use separately, since domination is transitive.
5193 if (DT.dominates(NoWrapEdge, Result->getParent()))
5194 continue;
5195
5196 for (const auto &RU : Result->uses())
5197 if (!DT.dominates(NoWrapEdge, RU))
5198 return false;
5199 }
5200
5201 return true;
5202 };
5203
5204 return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
5205 }
5206
5207 /// Shifts return poison if shiftwidth is larger than the bitwidth.
shiftAmountKnownInRange(const Value * ShiftAmount)5208 static bool shiftAmountKnownInRange(const Value *ShiftAmount) {
5209 auto *C = dyn_cast<Constant>(ShiftAmount);
5210 if (!C)
5211 return false;
5212
5213 // Shifts return poison if shiftwidth is larger than the bitwidth.
5214 SmallVector<const Constant *, 4> ShiftAmounts;
5215 if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) {
5216 unsigned NumElts = FVTy->getNumElements();
5217 for (unsigned i = 0; i < NumElts; ++i)
5218 ShiftAmounts.push_back(C->getAggregateElement(i));
5219 } else if (isa<ScalableVectorType>(C->getType()))
5220 return false; // Can't tell, just return false to be safe
5221 else
5222 ShiftAmounts.push_back(C);
5223
5224 bool Safe = llvm::all_of(ShiftAmounts, [](const Constant *C) {
5225 auto *CI = dyn_cast_or_null<ConstantInt>(C);
5226 return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth());
5227 });
5228
5229 return Safe;
5230 }
5231
canCreateUndefOrPoison(const Operator * Op,bool PoisonOnly,bool ConsiderFlagsAndMetadata)5232 static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly,
5233 bool ConsiderFlagsAndMetadata) {
5234
5235 if (ConsiderFlagsAndMetadata && Op->hasPoisonGeneratingFlagsOrMetadata())
5236 return true;
5237
5238 unsigned Opcode = Op->getOpcode();
5239
5240 // Check whether opcode is a poison/undef-generating operation
5241 switch (Opcode) {
5242 case Instruction::Shl:
5243 case Instruction::AShr:
5244 case Instruction::LShr:
5245 return !shiftAmountKnownInRange(Op->getOperand(1));
5246 case Instruction::FPToSI:
5247 case Instruction::FPToUI:
5248 // fptosi/ui yields poison if the resulting value does not fit in the
5249 // destination type.
5250 return true;
5251 case Instruction::Call:
5252 if (auto *II = dyn_cast<IntrinsicInst>(Op)) {
5253 switch (II->getIntrinsicID()) {
5254 // TODO: Add more intrinsics.
5255 case Intrinsic::ctlz:
5256 case Intrinsic::cttz:
5257 case Intrinsic::abs:
5258 if (cast<ConstantInt>(II->getArgOperand(1))->isNullValue())
5259 return false;
5260 break;
5261 case Intrinsic::ctpop:
5262 case Intrinsic::bswap:
5263 case Intrinsic::bitreverse:
5264 case Intrinsic::fshl:
5265 case Intrinsic::fshr:
5266 case Intrinsic::smax:
5267 case Intrinsic::smin:
5268 case Intrinsic::umax:
5269 case Intrinsic::umin:
5270 case Intrinsic::ptrmask:
5271 case Intrinsic::fptoui_sat:
5272 case Intrinsic::fptosi_sat:
5273 case Intrinsic::sadd_with_overflow:
5274 case Intrinsic::ssub_with_overflow:
5275 case Intrinsic::smul_with_overflow:
5276 case Intrinsic::uadd_with_overflow:
5277 case Intrinsic::usub_with_overflow:
5278 case Intrinsic::umul_with_overflow:
5279 case Intrinsic::sadd_sat:
5280 case Intrinsic::uadd_sat:
5281 case Intrinsic::ssub_sat:
5282 case Intrinsic::usub_sat:
5283 return false;
5284 case Intrinsic::sshl_sat:
5285 case Intrinsic::ushl_sat:
5286 return !shiftAmountKnownInRange(II->getArgOperand(1));
5287 case Intrinsic::fma:
5288 case Intrinsic::fmuladd:
5289 case Intrinsic::sqrt:
5290 case Intrinsic::powi:
5291 case Intrinsic::sin:
5292 case Intrinsic::cos:
5293 case Intrinsic::pow:
5294 case Intrinsic::log:
5295 case Intrinsic::log10:
5296 case Intrinsic::log2:
5297 case Intrinsic::exp:
5298 case Intrinsic::exp2:
5299 case Intrinsic::fabs:
5300 case Intrinsic::copysign:
5301 case Intrinsic::floor:
5302 case Intrinsic::ceil:
5303 case Intrinsic::trunc:
5304 case Intrinsic::rint:
5305 case Intrinsic::nearbyint:
5306 case Intrinsic::round:
5307 case Intrinsic::roundeven:
5308 case Intrinsic::fptrunc_round:
5309 case Intrinsic::canonicalize:
5310 case Intrinsic::arithmetic_fence:
5311 case Intrinsic::minnum:
5312 case Intrinsic::maxnum:
5313 case Intrinsic::minimum:
5314 case Intrinsic::maximum:
5315 case Intrinsic::is_fpclass:
5316 return false;
5317 case Intrinsic::lround:
5318 case Intrinsic::llround:
5319 case Intrinsic::lrint:
5320 case Intrinsic::llrint:
5321 // If the value doesn't fit an unspecified value is returned (but this
5322 // is not poison).
5323 return false;
5324 }
5325 }
5326 [[fallthrough]];
5327 case Instruction::CallBr:
5328 case Instruction::Invoke: {
5329 const auto *CB = cast<CallBase>(Op);
5330 return !CB->hasRetAttr(Attribute::NoUndef);
5331 }
5332 case Instruction::InsertElement:
5333 case Instruction::ExtractElement: {
5334 // If index exceeds the length of the vector, it returns poison
5335 auto *VTy = cast<VectorType>(Op->getOperand(0)->getType());
5336 unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1;
5337 auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp));
5338 if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue()))
5339 return true;
5340 return false;
5341 }
5342 case Instruction::ShuffleVector: {
5343 // shufflevector may return undef.
5344 if (PoisonOnly)
5345 return false;
5346 ArrayRef<int> Mask = isa<ConstantExpr>(Op)
5347 ? cast<ConstantExpr>(Op)->getShuffleMask()
5348 : cast<ShuffleVectorInst>(Op)->getShuffleMask();
5349 return is_contained(Mask, UndefMaskElem);
5350 }
5351 case Instruction::FNeg:
5352 case Instruction::PHI:
5353 case Instruction::Select:
5354 case Instruction::URem:
5355 case Instruction::SRem:
5356 case Instruction::ExtractValue:
5357 case Instruction::InsertValue:
5358 case Instruction::Freeze:
5359 case Instruction::ICmp:
5360 case Instruction::FCmp:
5361 case Instruction::FAdd:
5362 case Instruction::FSub:
5363 case Instruction::FMul:
5364 case Instruction::FDiv:
5365 case Instruction::FRem:
5366 return false;
5367 case Instruction::GetElementPtr:
5368 // inbounds is handled above
5369 // TODO: what about inrange on constexpr?
5370 return false;
5371 default: {
5372 const auto *CE = dyn_cast<ConstantExpr>(Op);
5373 if (isa<CastInst>(Op) || (CE && CE->isCast()))
5374 return false;
5375 else if (Instruction::isBinaryOp(Opcode))
5376 return false;
5377 // Be conservative and return true.
5378 return true;
5379 }
5380 }
5381 }
5382
canCreateUndefOrPoison(const Operator * Op,bool ConsiderFlagsAndMetadata)5383 bool llvm::canCreateUndefOrPoison(const Operator *Op,
5384 bool ConsiderFlagsAndMetadata) {
5385 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false,
5386 ConsiderFlagsAndMetadata);
5387 }
5388
canCreatePoison(const Operator * Op,bool ConsiderFlagsAndMetadata)5389 bool llvm::canCreatePoison(const Operator *Op, bool ConsiderFlagsAndMetadata) {
5390 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true,
5391 ConsiderFlagsAndMetadata);
5392 }
5393
directlyImpliesPoison(const Value * ValAssumedPoison,const Value * V,unsigned Depth)5394 static bool directlyImpliesPoison(const Value *ValAssumedPoison,
5395 const Value *V, unsigned Depth) {
5396 if (ValAssumedPoison == V)
5397 return true;
5398
5399 const unsigned MaxDepth = 2;
5400 if (Depth >= MaxDepth)
5401 return false;
5402
5403 if (const auto *I = dyn_cast<Instruction>(V)) {
5404 if (any_of(I->operands(), [=](const Use &Op) {
5405 return propagatesPoison(Op) &&
5406 directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1);
5407 }))
5408 return true;
5409
5410 // V = extractvalue V0, idx
5411 // V2 = extractvalue V0, idx2
5412 // V0's elements are all poison or not. (e.g., add_with_overflow)
5413 const WithOverflowInst *II;
5414 if (match(I, m_ExtractValue(m_WithOverflowInst(II))) &&
5415 (match(ValAssumedPoison, m_ExtractValue(m_Specific(II))) ||
5416 llvm::is_contained(II->args(), ValAssumedPoison)))
5417 return true;
5418 }
5419 return false;
5420 }
5421
impliesPoison(const Value * ValAssumedPoison,const Value * V,unsigned Depth)5422 static bool impliesPoison(const Value *ValAssumedPoison, const Value *V,
5423 unsigned Depth) {
5424 if (isGuaranteedNotToBeUndefOrPoison(ValAssumedPoison))
5425 return true;
5426
5427 if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0))
5428 return true;
5429
5430 const unsigned MaxDepth = 2;
5431 if (Depth >= MaxDepth)
5432 return false;
5433
5434 const auto *I = dyn_cast<Instruction>(ValAssumedPoison);
5435 if (I && !canCreatePoison(cast<Operator>(I))) {
5436 return all_of(I->operands(), [=](const Value *Op) {
5437 return impliesPoison(Op, V, Depth + 1);
5438 });
5439 }
5440 return false;
5441 }
5442
impliesPoison(const Value * ValAssumedPoison,const Value * V)5443 bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) {
5444 return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0);
5445 }
5446
5447 static bool programUndefinedIfUndefOrPoison(const Value *V,
5448 bool PoisonOnly);
5449
isGuaranteedNotToBeUndefOrPoison(const Value * V,AssumptionCache * AC,const Instruction * CtxI,const DominatorTree * DT,unsigned Depth,bool PoisonOnly)5450 static bool isGuaranteedNotToBeUndefOrPoison(const Value *V,
5451 AssumptionCache *AC,
5452 const Instruction *CtxI,
5453 const DominatorTree *DT,
5454 unsigned Depth, bool PoisonOnly) {
5455 if (Depth >= MaxAnalysisRecursionDepth)
5456 return false;
5457
5458 if (isa<MetadataAsValue>(V))
5459 return false;
5460
5461 if (const auto *A = dyn_cast<Argument>(V)) {
5462 if (A->hasAttribute(Attribute::NoUndef))
5463 return true;
5464 }
5465
5466 if (auto *C = dyn_cast<Constant>(V)) {
5467 if (isa<UndefValue>(C))
5468 return PoisonOnly && !isa<PoisonValue>(C);
5469
5470 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) ||
5471 isa<ConstantPointerNull>(C) || isa<Function>(C))
5472 return true;
5473
5474 if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C))
5475 return (PoisonOnly ? !C->containsPoisonElement()
5476 : !C->containsUndefOrPoisonElement()) &&
5477 !C->containsConstantExpression();
5478 }
5479
5480 // Strip cast operations from a pointer value.
5481 // Note that stripPointerCastsSameRepresentation can strip off getelementptr
5482 // inbounds with zero offset. To guarantee that the result isn't poison, the
5483 // stripped pointer is checked as it has to be pointing into an allocated
5484 // object or be null `null` to ensure `inbounds` getelement pointers with a
5485 // zero offset could not produce poison.
5486 // It can strip off addrspacecast that do not change bit representation as
5487 // well. We believe that such addrspacecast is equivalent to no-op.
5488 auto *StrippedV = V->stripPointerCastsSameRepresentation();
5489 if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) ||
5490 isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV))
5491 return true;
5492
5493 auto OpCheck = [&](const Value *V) {
5494 return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1,
5495 PoisonOnly);
5496 };
5497
5498 if (auto *Opr = dyn_cast<Operator>(V)) {
5499 // If the value is a freeze instruction, then it can never
5500 // be undef or poison.
5501 if (isa<FreezeInst>(V))
5502 return true;
5503
5504 if (const auto *CB = dyn_cast<CallBase>(V)) {
5505 if (CB->hasRetAttr(Attribute::NoUndef))
5506 return true;
5507 }
5508
5509 if (const auto *PN = dyn_cast<PHINode>(V)) {
5510 unsigned Num = PN->getNumIncomingValues();
5511 bool IsWellDefined = true;
5512 for (unsigned i = 0; i < Num; ++i) {
5513 auto *TI = PN->getIncomingBlock(i)->getTerminator();
5514 if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI,
5515 DT, Depth + 1, PoisonOnly)) {
5516 IsWellDefined = false;
5517 break;
5518 }
5519 }
5520 if (IsWellDefined)
5521 return true;
5522 } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck))
5523 return true;
5524 }
5525
5526 if (auto *I = dyn_cast<LoadInst>(V))
5527 if (I->hasMetadata(LLVMContext::MD_noundef) ||
5528 I->hasMetadata(LLVMContext::MD_dereferenceable) ||
5529 I->hasMetadata(LLVMContext::MD_dereferenceable_or_null))
5530 return true;
5531
5532 if (programUndefinedIfUndefOrPoison(V, PoisonOnly))
5533 return true;
5534
5535 // CxtI may be null or a cloned instruction.
5536 if (!CtxI || !CtxI->getParent() || !DT)
5537 return false;
5538
5539 auto *DNode = DT->getNode(CtxI->getParent());
5540 if (!DNode)
5541 // Unreachable block
5542 return false;
5543
5544 // If V is used as a branch condition before reaching CtxI, V cannot be
5545 // undef or poison.
5546 // br V, BB1, BB2
5547 // BB1:
5548 // CtxI ; V cannot be undef or poison here
5549 auto *Dominator = DNode->getIDom();
5550 while (Dominator) {
5551 auto *TI = Dominator->getBlock()->getTerminator();
5552
5553 Value *Cond = nullptr;
5554 if (auto BI = dyn_cast_or_null<BranchInst>(TI)) {
5555 if (BI->isConditional())
5556 Cond = BI->getCondition();
5557 } else if (auto SI = dyn_cast_or_null<SwitchInst>(TI)) {
5558 Cond = SI->getCondition();
5559 }
5560
5561 if (Cond) {
5562 if (Cond == V)
5563 return true;
5564 else if (PoisonOnly && isa<Operator>(Cond)) {
5565 // For poison, we can analyze further
5566 auto *Opr = cast<Operator>(Cond);
5567 if (any_of(Opr->operands(),
5568 [V](const Use &U) { return V == U && propagatesPoison(U); }))
5569 return true;
5570 }
5571 }
5572
5573 Dominator = Dominator->getIDom();
5574 }
5575
5576 if (getKnowledgeValidInContext(V, {Attribute::NoUndef}, CtxI, DT, AC))
5577 return true;
5578
5579 return false;
5580 }
5581
isGuaranteedNotToBeUndefOrPoison(const Value * V,AssumptionCache * AC,const Instruction * CtxI,const DominatorTree * DT,unsigned Depth)5582 bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC,
5583 const Instruction *CtxI,
5584 const DominatorTree *DT,
5585 unsigned Depth) {
5586 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false);
5587 }
5588
isGuaranteedNotToBePoison(const Value * V,AssumptionCache * AC,const Instruction * CtxI,const DominatorTree * DT,unsigned Depth)5589 bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC,
5590 const Instruction *CtxI,
5591 const DominatorTree *DT, unsigned Depth) {
5592 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true);
5593 }
5594
computeOverflowForSignedAdd(const AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)5595 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
5596 const DataLayout &DL,
5597 AssumptionCache *AC,
5598 const Instruction *CxtI,
5599 const DominatorTree *DT) {
5600 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
5601 Add, DL, AC, CxtI, DT);
5602 }
5603
computeOverflowForSignedAdd(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)5604 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
5605 const Value *RHS,
5606 const DataLayout &DL,
5607 AssumptionCache *AC,
5608 const Instruction *CxtI,
5609 const DominatorTree *DT) {
5610 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
5611 }
5612
isGuaranteedToTransferExecutionToSuccessor(const Instruction * I)5613 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
5614 // Note: An atomic operation isn't guaranteed to return in a reasonable amount
5615 // of time because it's possible for another thread to interfere with it for an
5616 // arbitrary length of time, but programs aren't allowed to rely on that.
5617
5618 // If there is no successor, then execution can't transfer to it.
5619 if (isa<ReturnInst>(I))
5620 return false;
5621 if (isa<UnreachableInst>(I))
5622 return false;
5623
5624 // Note: Do not add new checks here; instead, change Instruction::mayThrow or
5625 // Instruction::willReturn.
5626 //
5627 // FIXME: Move this check into Instruction::willReturn.
5628 if (isa<CatchPadInst>(I)) {
5629 switch (classifyEHPersonality(I->getFunction()->getPersonalityFn())) {
5630 default:
5631 // A catchpad may invoke exception object constructors and such, which
5632 // in some languages can be arbitrary code, so be conservative by default.
5633 return false;
5634 case EHPersonality::CoreCLR:
5635 // For CoreCLR, it just involves a type test.
5636 return true;
5637 }
5638 }
5639
5640 // An instruction that returns without throwing must transfer control flow
5641 // to a successor.
5642 return !I->mayThrow() && I->willReturn();
5643 }
5644
isGuaranteedToTransferExecutionToSuccessor(const BasicBlock * BB)5645 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) {
5646 // TODO: This is slightly conservative for invoke instruction since exiting
5647 // via an exception *is* normal control for them.
5648 for (const Instruction &I : *BB)
5649 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
5650 return false;
5651 return true;
5652 }
5653
isGuaranteedToTransferExecutionToSuccessor(BasicBlock::const_iterator Begin,BasicBlock::const_iterator End,unsigned ScanLimit)5654 bool llvm::isGuaranteedToTransferExecutionToSuccessor(
5655 BasicBlock::const_iterator Begin, BasicBlock::const_iterator End,
5656 unsigned ScanLimit) {
5657 return isGuaranteedToTransferExecutionToSuccessor(make_range(Begin, End),
5658 ScanLimit);
5659 }
5660
isGuaranteedToTransferExecutionToSuccessor(iterator_range<BasicBlock::const_iterator> Range,unsigned ScanLimit)5661 bool llvm::isGuaranteedToTransferExecutionToSuccessor(
5662 iterator_range<BasicBlock::const_iterator> Range, unsigned ScanLimit) {
5663 assert(ScanLimit && "scan limit must be non-zero");
5664 for (const Instruction &I : Range) {
5665 if (isa<DbgInfoIntrinsic>(I))
5666 continue;
5667 if (--ScanLimit == 0)
5668 return false;
5669 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
5670 return false;
5671 }
5672 return true;
5673 }
5674
isGuaranteedToExecuteForEveryIteration(const Instruction * I,const Loop * L)5675 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
5676 const Loop *L) {
5677 // The loop header is guaranteed to be executed for every iteration.
5678 //
5679 // FIXME: Relax this constraint to cover all basic blocks that are
5680 // guaranteed to be executed at every iteration.
5681 if (I->getParent() != L->getHeader()) return false;
5682
5683 for (const Instruction &LI : *L->getHeader()) {
5684 if (&LI == I) return true;
5685 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
5686 }
5687 llvm_unreachable("Instruction not contained in its own parent basic block.");
5688 }
5689
propagatesPoison(const Use & PoisonOp)5690 bool llvm::propagatesPoison(const Use &PoisonOp) {
5691 const Operator *I = cast<Operator>(PoisonOp.getUser());
5692 switch (I->getOpcode()) {
5693 case Instruction::Freeze:
5694 case Instruction::PHI:
5695 case Instruction::Invoke:
5696 return false;
5697 case Instruction::Select:
5698 return PoisonOp.getOperandNo() == 0;
5699 case Instruction::Call:
5700 if (auto *II = dyn_cast<IntrinsicInst>(I)) {
5701 switch (II->getIntrinsicID()) {
5702 // TODO: Add more intrinsics.
5703 case Intrinsic::sadd_with_overflow:
5704 case Intrinsic::ssub_with_overflow:
5705 case Intrinsic::smul_with_overflow:
5706 case Intrinsic::uadd_with_overflow:
5707 case Intrinsic::usub_with_overflow:
5708 case Intrinsic::umul_with_overflow:
5709 // If an input is a vector containing a poison element, the
5710 // two output vectors (calculated results, overflow bits)'
5711 // corresponding lanes are poison.
5712 return true;
5713 case Intrinsic::ctpop:
5714 return true;
5715 }
5716 }
5717 return false;
5718 case Instruction::ICmp:
5719 case Instruction::FCmp:
5720 case Instruction::GetElementPtr:
5721 return true;
5722 default:
5723 if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I))
5724 return true;
5725
5726 // Be conservative and return false.
5727 return false;
5728 }
5729 }
5730
getGuaranteedWellDefinedOps(const Instruction * I,SmallVectorImpl<const Value * > & Operands)5731 void llvm::getGuaranteedWellDefinedOps(
5732 const Instruction *I, SmallVectorImpl<const Value *> &Operands) {
5733 switch (I->getOpcode()) {
5734 case Instruction::Store:
5735 Operands.push_back(cast<StoreInst>(I)->getPointerOperand());
5736 break;
5737
5738 case Instruction::Load:
5739 Operands.push_back(cast<LoadInst>(I)->getPointerOperand());
5740 break;
5741
5742 // Since dereferenceable attribute imply noundef, atomic operations
5743 // also implicitly have noundef pointers too
5744 case Instruction::AtomicCmpXchg:
5745 Operands.push_back(cast<AtomicCmpXchgInst>(I)->getPointerOperand());
5746 break;
5747
5748 case Instruction::AtomicRMW:
5749 Operands.push_back(cast<AtomicRMWInst>(I)->getPointerOperand());
5750 break;
5751
5752 case Instruction::Call:
5753 case Instruction::Invoke: {
5754 const CallBase *CB = cast<CallBase>(I);
5755 if (CB->isIndirectCall())
5756 Operands.push_back(CB->getCalledOperand());
5757 for (unsigned i = 0; i < CB->arg_size(); ++i) {
5758 if (CB->paramHasAttr(i, Attribute::NoUndef) ||
5759 CB->paramHasAttr(i, Attribute::Dereferenceable))
5760 Operands.push_back(CB->getArgOperand(i));
5761 }
5762 break;
5763 }
5764 case Instruction::Ret:
5765 if (I->getFunction()->hasRetAttribute(Attribute::NoUndef))
5766 Operands.push_back(I->getOperand(0));
5767 break;
5768 case Instruction::Switch:
5769 Operands.push_back(cast<SwitchInst>(I)->getCondition());
5770 break;
5771 case Instruction::Br: {
5772 auto *BR = cast<BranchInst>(I);
5773 if (BR->isConditional())
5774 Operands.push_back(BR->getCondition());
5775 break;
5776 }
5777 default:
5778 break;
5779 }
5780 }
5781
getGuaranteedNonPoisonOps(const Instruction * I,SmallVectorImpl<const Value * > & Operands)5782 void llvm::getGuaranteedNonPoisonOps(const Instruction *I,
5783 SmallVectorImpl<const Value *> &Operands) {
5784 getGuaranteedWellDefinedOps(I, Operands);
5785 switch (I->getOpcode()) {
5786 // Divisors of these operations are allowed to be partially undef.
5787 case Instruction::UDiv:
5788 case Instruction::SDiv:
5789 case Instruction::URem:
5790 case Instruction::SRem:
5791 Operands.push_back(I->getOperand(1));
5792 break;
5793 default:
5794 break;
5795 }
5796 }
5797
mustTriggerUB(const Instruction * I,const SmallSet<const Value *,16> & KnownPoison)5798 bool llvm::mustTriggerUB(const Instruction *I,
5799 const SmallSet<const Value *, 16>& KnownPoison) {
5800 SmallVector<const Value *, 4> NonPoisonOps;
5801 getGuaranteedNonPoisonOps(I, NonPoisonOps);
5802
5803 for (const auto *V : NonPoisonOps)
5804 if (KnownPoison.count(V))
5805 return true;
5806
5807 return false;
5808 }
5809
programUndefinedIfUndefOrPoison(const Value * V,bool PoisonOnly)5810 static bool programUndefinedIfUndefOrPoison(const Value *V,
5811 bool PoisonOnly) {
5812 // We currently only look for uses of values within the same basic
5813 // block, as that makes it easier to guarantee that the uses will be
5814 // executed given that Inst is executed.
5815 //
5816 // FIXME: Expand this to consider uses beyond the same basic block. To do
5817 // this, look out for the distinction between post-dominance and strong
5818 // post-dominance.
5819 const BasicBlock *BB = nullptr;
5820 BasicBlock::const_iterator Begin;
5821 if (const auto *Inst = dyn_cast<Instruction>(V)) {
5822 BB = Inst->getParent();
5823 Begin = Inst->getIterator();
5824 Begin++;
5825 } else if (const auto *Arg = dyn_cast<Argument>(V)) {
5826 BB = &Arg->getParent()->getEntryBlock();
5827 Begin = BB->begin();
5828 } else {
5829 return false;
5830 }
5831
5832 // Limit number of instructions we look at, to avoid scanning through large
5833 // blocks. The current limit is chosen arbitrarily.
5834 unsigned ScanLimit = 32;
5835 BasicBlock::const_iterator End = BB->end();
5836
5837 if (!PoisonOnly) {
5838 // Since undef does not propagate eagerly, be conservative & just check
5839 // whether a value is directly passed to an instruction that must take
5840 // well-defined operands.
5841
5842 for (const auto &I : make_range(Begin, End)) {
5843 if (isa<DbgInfoIntrinsic>(I))
5844 continue;
5845 if (--ScanLimit == 0)
5846 break;
5847
5848 SmallVector<const Value *, 4> WellDefinedOps;
5849 getGuaranteedWellDefinedOps(&I, WellDefinedOps);
5850 if (is_contained(WellDefinedOps, V))
5851 return true;
5852
5853 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
5854 break;
5855 }
5856 return false;
5857 }
5858
5859 // Set of instructions that we have proved will yield poison if Inst
5860 // does.
5861 SmallSet<const Value *, 16> YieldsPoison;
5862 SmallSet<const BasicBlock *, 4> Visited;
5863
5864 YieldsPoison.insert(V);
5865 Visited.insert(BB);
5866
5867 while (true) {
5868 for (const auto &I : make_range(Begin, End)) {
5869 if (isa<DbgInfoIntrinsic>(I))
5870 continue;
5871 if (--ScanLimit == 0)
5872 return false;
5873 if (mustTriggerUB(&I, YieldsPoison))
5874 return true;
5875 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
5876 return false;
5877
5878 // If an operand is poison and propagates it, mark I as yielding poison.
5879 for (const Use &Op : I.operands()) {
5880 if (YieldsPoison.count(Op) && propagatesPoison(Op)) {
5881 YieldsPoison.insert(&I);
5882 break;
5883 }
5884 }
5885 }
5886
5887 BB = BB->getSingleSuccessor();
5888 if (!BB || !Visited.insert(BB).second)
5889 break;
5890
5891 Begin = BB->getFirstNonPHI()->getIterator();
5892 End = BB->end();
5893 }
5894 return false;
5895 }
5896
programUndefinedIfUndefOrPoison(const Instruction * Inst)5897 bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) {
5898 return ::programUndefinedIfUndefOrPoison(Inst, false);
5899 }
5900
programUndefinedIfPoison(const Instruction * Inst)5901 bool llvm::programUndefinedIfPoison(const Instruction *Inst) {
5902 return ::programUndefinedIfUndefOrPoison(Inst, true);
5903 }
5904
isKnownNonNaN(const Value * V,FastMathFlags FMF)5905 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
5906 if (FMF.noNaNs())
5907 return true;
5908
5909 if (auto *C = dyn_cast<ConstantFP>(V))
5910 return !C->isNaN();
5911
5912 if (auto *C = dyn_cast<ConstantDataVector>(V)) {
5913 if (!C->getElementType()->isFloatingPointTy())
5914 return false;
5915 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
5916 if (C->getElementAsAPFloat(I).isNaN())
5917 return false;
5918 }
5919 return true;
5920 }
5921
5922 if (isa<ConstantAggregateZero>(V))
5923 return true;
5924
5925 return false;
5926 }
5927
isKnownNonZero(const Value * V)5928 static bool isKnownNonZero(const Value *V) {
5929 if (auto *C = dyn_cast<ConstantFP>(V))
5930 return !C->isZero();
5931
5932 if (auto *C = dyn_cast<ConstantDataVector>(V)) {
5933 if (!C->getElementType()->isFloatingPointTy())
5934 return false;
5935 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
5936 if (C->getElementAsAPFloat(I).isZero())
5937 return false;
5938 }
5939 return true;
5940 }
5941
5942 return false;
5943 }
5944
5945 /// Match clamp pattern for float types without care about NaNs or signed zeros.
5946 /// Given non-min/max outer cmp/select from the clamp pattern this
5947 /// function recognizes if it can be substitued by a "canonical" min/max
5948 /// pattern.
matchFastFloatClamp(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS)5949 static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred,
5950 Value *CmpLHS, Value *CmpRHS,
5951 Value *TrueVal, Value *FalseVal,
5952 Value *&LHS, Value *&RHS) {
5953 // Try to match
5954 // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
5955 // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
5956 // and return description of the outer Max/Min.
5957
5958 // First, check if select has inverse order:
5959 if (CmpRHS == FalseVal) {
5960 std::swap(TrueVal, FalseVal);
5961 Pred = CmpInst::getInversePredicate(Pred);
5962 }
5963
5964 // Assume success now. If there's no match, callers should not use these anyway.
5965 LHS = TrueVal;
5966 RHS = FalseVal;
5967
5968 const APFloat *FC1;
5969 if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
5970 return {SPF_UNKNOWN, SPNB_NA, false};
5971
5972 const APFloat *FC2;
5973 switch (Pred) {
5974 case CmpInst::FCMP_OLT:
5975 case CmpInst::FCMP_OLE:
5976 case CmpInst::FCMP_ULT:
5977 case CmpInst::FCMP_ULE:
5978 if (match(FalseVal,
5979 m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
5980 m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
5981 *FC1 < *FC2)
5982 return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
5983 break;
5984 case CmpInst::FCMP_OGT:
5985 case CmpInst::FCMP_OGE:
5986 case CmpInst::FCMP_UGT:
5987 case CmpInst::FCMP_UGE:
5988 if (match(FalseVal,
5989 m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
5990 m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
5991 *FC1 > *FC2)
5992 return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
5993 break;
5994 default:
5995 break;
5996 }
5997
5998 return {SPF_UNKNOWN, SPNB_NA, false};
5999 }
6000
6001 /// Recognize variations of:
6002 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
matchClamp(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal)6003 static SelectPatternResult matchClamp(CmpInst::Predicate Pred,
6004 Value *CmpLHS, Value *CmpRHS,
6005 Value *TrueVal, Value *FalseVal) {
6006 // Swap the select operands and predicate to match the patterns below.
6007 if (CmpRHS != TrueVal) {
6008 Pred = ICmpInst::getSwappedPredicate(Pred);
6009 std::swap(TrueVal, FalseVal);
6010 }
6011 const APInt *C1;
6012 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
6013 const APInt *C2;
6014 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
6015 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
6016 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
6017 return {SPF_SMAX, SPNB_NA, false};
6018
6019 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
6020 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
6021 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
6022 return {SPF_SMIN, SPNB_NA, false};
6023
6024 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
6025 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
6026 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
6027 return {SPF_UMAX, SPNB_NA, false};
6028
6029 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
6030 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
6031 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
6032 return {SPF_UMIN, SPNB_NA, false};
6033 }
6034 return {SPF_UNKNOWN, SPNB_NA, false};
6035 }
6036
6037 /// Recognize variations of:
6038 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
matchMinMaxOfMinMax(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TVal,Value * FVal,unsigned Depth)6039 static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred,
6040 Value *CmpLHS, Value *CmpRHS,
6041 Value *TVal, Value *FVal,
6042 unsigned Depth) {
6043 // TODO: Allow FP min/max with nnan/nsz.
6044 assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
6045
6046 Value *A = nullptr, *B = nullptr;
6047 SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
6048 if (!SelectPatternResult::isMinOrMax(L.Flavor))
6049 return {SPF_UNKNOWN, SPNB_NA, false};
6050
6051 Value *C = nullptr, *D = nullptr;
6052 SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
6053 if (L.Flavor != R.Flavor)
6054 return {SPF_UNKNOWN, SPNB_NA, false};
6055
6056 // We have something like: x Pred y ? min(a, b) : min(c, d).
6057 // Try to match the compare to the min/max operations of the select operands.
6058 // First, make sure we have the right compare predicate.
6059 switch (L.Flavor) {
6060 case SPF_SMIN:
6061 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
6062 Pred = ICmpInst::getSwappedPredicate(Pred);
6063 std::swap(CmpLHS, CmpRHS);
6064 }
6065 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
6066 break;
6067 return {SPF_UNKNOWN, SPNB_NA, false};
6068 case SPF_SMAX:
6069 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
6070 Pred = ICmpInst::getSwappedPredicate(Pred);
6071 std::swap(CmpLHS, CmpRHS);
6072 }
6073 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
6074 break;
6075 return {SPF_UNKNOWN, SPNB_NA, false};
6076 case SPF_UMIN:
6077 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
6078 Pred = ICmpInst::getSwappedPredicate(Pred);
6079 std::swap(CmpLHS, CmpRHS);
6080 }
6081 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
6082 break;
6083 return {SPF_UNKNOWN, SPNB_NA, false};
6084 case SPF_UMAX:
6085 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
6086 Pred = ICmpInst::getSwappedPredicate(Pred);
6087 std::swap(CmpLHS, CmpRHS);
6088 }
6089 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
6090 break;
6091 return {SPF_UNKNOWN, SPNB_NA, false};
6092 default:
6093 return {SPF_UNKNOWN, SPNB_NA, false};
6094 }
6095
6096 // If there is a common operand in the already matched min/max and the other
6097 // min/max operands match the compare operands (either directly or inverted),
6098 // then this is min/max of the same flavor.
6099
6100 // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
6101 // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
6102 if (D == B) {
6103 if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
6104 match(A, m_Not(m_Specific(CmpRHS)))))
6105 return {L.Flavor, SPNB_NA, false};
6106 }
6107 // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
6108 // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
6109 if (C == B) {
6110 if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
6111 match(A, m_Not(m_Specific(CmpRHS)))))
6112 return {L.Flavor, SPNB_NA, false};
6113 }
6114 // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
6115 // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
6116 if (D == A) {
6117 if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
6118 match(B, m_Not(m_Specific(CmpRHS)))))
6119 return {L.Flavor, SPNB_NA, false};
6120 }
6121 // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
6122 // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
6123 if (C == A) {
6124 if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
6125 match(B, m_Not(m_Specific(CmpRHS)))))
6126 return {L.Flavor, SPNB_NA, false};
6127 }
6128
6129 return {SPF_UNKNOWN, SPNB_NA, false};
6130 }
6131
6132 /// If the input value is the result of a 'not' op, constant integer, or vector
6133 /// splat of a constant integer, return the bitwise-not source value.
6134 /// TODO: This could be extended to handle non-splat vector integer constants.
getNotValue(Value * V)6135 static Value *getNotValue(Value *V) {
6136 Value *NotV;
6137 if (match(V, m_Not(m_Value(NotV))))
6138 return NotV;
6139
6140 const APInt *C;
6141 if (match(V, m_APInt(C)))
6142 return ConstantInt::get(V->getType(), ~(*C));
6143
6144 return nullptr;
6145 }
6146
6147 /// Match non-obvious integer minimum and maximum sequences.
matchMinMax(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS,unsigned Depth)6148 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
6149 Value *CmpLHS, Value *CmpRHS,
6150 Value *TrueVal, Value *FalseVal,
6151 Value *&LHS, Value *&RHS,
6152 unsigned Depth) {
6153 // Assume success. If there's no match, callers should not use these anyway.
6154 LHS = TrueVal;
6155 RHS = FalseVal;
6156
6157 SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
6158 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
6159 return SPR;
6160
6161 SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
6162 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
6163 return SPR;
6164
6165 // Look through 'not' ops to find disguised min/max.
6166 // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y)
6167 // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y)
6168 if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) {
6169 switch (Pred) {
6170 case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false};
6171 case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false};
6172 case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false};
6173 case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false};
6174 default: break;
6175 }
6176 }
6177
6178 // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X)
6179 // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X)
6180 if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) {
6181 switch (Pred) {
6182 case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false};
6183 case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false};
6184 case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false};
6185 case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false};
6186 default: break;
6187 }
6188 }
6189
6190 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
6191 return {SPF_UNKNOWN, SPNB_NA, false};
6192
6193 const APInt *C1;
6194 if (!match(CmpRHS, m_APInt(C1)))
6195 return {SPF_UNKNOWN, SPNB_NA, false};
6196
6197 // An unsigned min/max can be written with a signed compare.
6198 const APInt *C2;
6199 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
6200 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
6201 // Is the sign bit set?
6202 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
6203 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
6204 if (Pred == CmpInst::ICMP_SLT && C1->isZero() && C2->isMaxSignedValue())
6205 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
6206
6207 // Is the sign bit clear?
6208 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
6209 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
6210 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnes() && C2->isMinSignedValue())
6211 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
6212 }
6213
6214 return {SPF_UNKNOWN, SPNB_NA, false};
6215 }
6216
isKnownNegation(const Value * X,const Value * Y,bool NeedNSW)6217 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
6218 assert(X && Y && "Invalid operand");
6219
6220 // X = sub (0, Y) || X = sub nsw (0, Y)
6221 if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
6222 (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
6223 return true;
6224
6225 // Y = sub (0, X) || Y = sub nsw (0, X)
6226 if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
6227 (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
6228 return true;
6229
6230 // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
6231 Value *A, *B;
6232 return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
6233 match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
6234 (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
6235 match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
6236 }
6237
matchSelectPattern(CmpInst::Predicate Pred,FastMathFlags FMF,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS,unsigned Depth)6238 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
6239 FastMathFlags FMF,
6240 Value *CmpLHS, Value *CmpRHS,
6241 Value *TrueVal, Value *FalseVal,
6242 Value *&LHS, Value *&RHS,
6243 unsigned Depth) {
6244 bool HasMismatchedZeros = false;
6245 if (CmpInst::isFPPredicate(Pred)) {
6246 // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one
6247 // 0.0 operand, set the compare's 0.0 operands to that same value for the
6248 // purpose of identifying min/max. Disregard vector constants with undefined
6249 // elements because those can not be back-propagated for analysis.
6250 Value *OutputZeroVal = nullptr;
6251 if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) &&
6252 !cast<Constant>(TrueVal)->containsUndefOrPoisonElement())
6253 OutputZeroVal = TrueVal;
6254 else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) &&
6255 !cast<Constant>(FalseVal)->containsUndefOrPoisonElement())
6256 OutputZeroVal = FalseVal;
6257
6258 if (OutputZeroVal) {
6259 if (match(CmpLHS, m_AnyZeroFP()) && CmpLHS != OutputZeroVal) {
6260 HasMismatchedZeros = true;
6261 CmpLHS = OutputZeroVal;
6262 }
6263 if (match(CmpRHS, m_AnyZeroFP()) && CmpRHS != OutputZeroVal) {
6264 HasMismatchedZeros = true;
6265 CmpRHS = OutputZeroVal;
6266 }
6267 }
6268 }
6269
6270 LHS = CmpLHS;
6271 RHS = CmpRHS;
6272
6273 // Signed zero may return inconsistent results between implementations.
6274 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
6275 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
6276 // Therefore, we behave conservatively and only proceed if at least one of the
6277 // operands is known to not be zero or if we don't care about signed zero.
6278 switch (Pred) {
6279 default: break;
6280 case CmpInst::FCMP_OGT: case CmpInst::FCMP_OLT:
6281 case CmpInst::FCMP_UGT: case CmpInst::FCMP_ULT:
6282 if (!HasMismatchedZeros)
6283 break;
6284 [[fallthrough]];
6285 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
6286 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
6287 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
6288 !isKnownNonZero(CmpRHS))
6289 return {SPF_UNKNOWN, SPNB_NA, false};
6290 }
6291
6292 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
6293 bool Ordered = false;
6294
6295 // When given one NaN and one non-NaN input:
6296 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
6297 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
6298 // ordered comparison fails), which could be NaN or non-NaN.
6299 // so here we discover exactly what NaN behavior is required/accepted.
6300 if (CmpInst::isFPPredicate(Pred)) {
6301 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
6302 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
6303
6304 if (LHSSafe && RHSSafe) {
6305 // Both operands are known non-NaN.
6306 NaNBehavior = SPNB_RETURNS_ANY;
6307 } else if (CmpInst::isOrdered(Pred)) {
6308 // An ordered comparison will return false when given a NaN, so it
6309 // returns the RHS.
6310 Ordered = true;
6311 if (LHSSafe)
6312 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
6313 NaNBehavior = SPNB_RETURNS_NAN;
6314 else if (RHSSafe)
6315 NaNBehavior = SPNB_RETURNS_OTHER;
6316 else
6317 // Completely unsafe.
6318 return {SPF_UNKNOWN, SPNB_NA, false};
6319 } else {
6320 Ordered = false;
6321 // An unordered comparison will return true when given a NaN, so it
6322 // returns the LHS.
6323 if (LHSSafe)
6324 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
6325 NaNBehavior = SPNB_RETURNS_OTHER;
6326 else if (RHSSafe)
6327 NaNBehavior = SPNB_RETURNS_NAN;
6328 else
6329 // Completely unsafe.
6330 return {SPF_UNKNOWN, SPNB_NA, false};
6331 }
6332 }
6333
6334 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
6335 std::swap(CmpLHS, CmpRHS);
6336 Pred = CmpInst::getSwappedPredicate(Pred);
6337 if (NaNBehavior == SPNB_RETURNS_NAN)
6338 NaNBehavior = SPNB_RETURNS_OTHER;
6339 else if (NaNBehavior == SPNB_RETURNS_OTHER)
6340 NaNBehavior = SPNB_RETURNS_NAN;
6341 Ordered = !Ordered;
6342 }
6343
6344 // ([if]cmp X, Y) ? X : Y
6345 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
6346 switch (Pred) {
6347 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
6348 case ICmpInst::ICMP_UGT:
6349 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
6350 case ICmpInst::ICMP_SGT:
6351 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
6352 case ICmpInst::ICMP_ULT:
6353 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
6354 case ICmpInst::ICMP_SLT:
6355 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
6356 case FCmpInst::FCMP_UGT:
6357 case FCmpInst::FCMP_UGE:
6358 case FCmpInst::FCMP_OGT:
6359 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
6360 case FCmpInst::FCMP_ULT:
6361 case FCmpInst::FCMP_ULE:
6362 case FCmpInst::FCMP_OLT:
6363 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
6364 }
6365 }
6366
6367 if (isKnownNegation(TrueVal, FalseVal)) {
6368 // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
6369 // match against either LHS or sext(LHS).
6370 auto MaybeSExtCmpLHS =
6371 m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
6372 auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
6373 auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
6374 if (match(TrueVal, MaybeSExtCmpLHS)) {
6375 // Set the return values. If the compare uses the negated value (-X >s 0),
6376 // swap the return values because the negated value is always 'RHS'.
6377 LHS = TrueVal;
6378 RHS = FalseVal;
6379 if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
6380 std::swap(LHS, RHS);
6381
6382 // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
6383 // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
6384 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
6385 return {SPF_ABS, SPNB_NA, false};
6386
6387 // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X)
6388 if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne))
6389 return {SPF_ABS, SPNB_NA, false};
6390
6391 // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
6392 // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
6393 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
6394 return {SPF_NABS, SPNB_NA, false};
6395 }
6396 else if (match(FalseVal, MaybeSExtCmpLHS)) {
6397 // Set the return values. If the compare uses the negated value (-X >s 0),
6398 // swap the return values because the negated value is always 'RHS'.
6399 LHS = FalseVal;
6400 RHS = TrueVal;
6401 if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
6402 std::swap(LHS, RHS);
6403
6404 // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
6405 // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
6406 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
6407 return {SPF_NABS, SPNB_NA, false};
6408
6409 // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
6410 // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
6411 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
6412 return {SPF_ABS, SPNB_NA, false};
6413 }
6414 }
6415
6416 if (CmpInst::isIntPredicate(Pred))
6417 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
6418
6419 // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
6420 // may return either -0.0 or 0.0, so fcmp/select pair has stricter
6421 // semantics than minNum. Be conservative in such case.
6422 if (NaNBehavior != SPNB_RETURNS_ANY ||
6423 (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
6424 !isKnownNonZero(CmpRHS)))
6425 return {SPF_UNKNOWN, SPNB_NA, false};
6426
6427 return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
6428 }
6429
6430 /// Helps to match a select pattern in case of a type mismatch.
6431 ///
6432 /// The function processes the case when type of true and false values of a
6433 /// select instruction differs from type of the cmp instruction operands because
6434 /// of a cast instruction. The function checks if it is legal to move the cast
6435 /// operation after "select". If yes, it returns the new second value of
6436 /// "select" (with the assumption that cast is moved):
6437 /// 1. As operand of cast instruction when both values of "select" are same cast
6438 /// instructions.
6439 /// 2. As restored constant (by applying reverse cast operation) when the first
6440 /// value of the "select" is a cast operation and the second value is a
6441 /// constant.
6442 /// NOTE: We return only the new second value because the first value could be
6443 /// accessed as operand of cast instruction.
lookThroughCast(CmpInst * CmpI,Value * V1,Value * V2,Instruction::CastOps * CastOp)6444 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
6445 Instruction::CastOps *CastOp) {
6446 auto *Cast1 = dyn_cast<CastInst>(V1);
6447 if (!Cast1)
6448 return nullptr;
6449
6450 *CastOp = Cast1->getOpcode();
6451 Type *SrcTy = Cast1->getSrcTy();
6452 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
6453 // If V1 and V2 are both the same cast from the same type, look through V1.
6454 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
6455 return Cast2->getOperand(0);
6456 return nullptr;
6457 }
6458
6459 auto *C = dyn_cast<Constant>(V2);
6460 if (!C)
6461 return nullptr;
6462
6463 Constant *CastedTo = nullptr;
6464 switch (*CastOp) {
6465 case Instruction::ZExt:
6466 if (CmpI->isUnsigned())
6467 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
6468 break;
6469 case Instruction::SExt:
6470 if (CmpI->isSigned())
6471 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
6472 break;
6473 case Instruction::Trunc:
6474 Constant *CmpConst;
6475 if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
6476 CmpConst->getType() == SrcTy) {
6477 // Here we have the following case:
6478 //
6479 // %cond = cmp iN %x, CmpConst
6480 // %tr = trunc iN %x to iK
6481 // %narrowsel = select i1 %cond, iK %t, iK C
6482 //
6483 // We can always move trunc after select operation:
6484 //
6485 // %cond = cmp iN %x, CmpConst
6486 // %widesel = select i1 %cond, iN %x, iN CmpConst
6487 // %tr = trunc iN %widesel to iK
6488 //
6489 // Note that C could be extended in any way because we don't care about
6490 // upper bits after truncation. It can't be abs pattern, because it would
6491 // look like:
6492 //
6493 // select i1 %cond, x, -x.
6494 //
6495 // So only min/max pattern could be matched. Such match requires widened C
6496 // == CmpConst. That is why set widened C = CmpConst, condition trunc
6497 // CmpConst == C is checked below.
6498 CastedTo = CmpConst;
6499 } else {
6500 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
6501 }
6502 break;
6503 case Instruction::FPTrunc:
6504 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
6505 break;
6506 case Instruction::FPExt:
6507 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
6508 break;
6509 case Instruction::FPToUI:
6510 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
6511 break;
6512 case Instruction::FPToSI:
6513 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
6514 break;
6515 case Instruction::UIToFP:
6516 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
6517 break;
6518 case Instruction::SIToFP:
6519 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
6520 break;
6521 default:
6522 break;
6523 }
6524
6525 if (!CastedTo)
6526 return nullptr;
6527
6528 // Make sure the cast doesn't lose any information.
6529 Constant *CastedBack =
6530 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
6531 if (CastedBack != C)
6532 return nullptr;
6533
6534 return CastedTo;
6535 }
6536
matchSelectPattern(Value * V,Value * & LHS,Value * & RHS,Instruction::CastOps * CastOp,unsigned Depth)6537 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
6538 Instruction::CastOps *CastOp,
6539 unsigned Depth) {
6540 if (Depth >= MaxAnalysisRecursionDepth)
6541 return {SPF_UNKNOWN, SPNB_NA, false};
6542
6543 SelectInst *SI = dyn_cast<SelectInst>(V);
6544 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
6545
6546 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
6547 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
6548
6549 Value *TrueVal = SI->getTrueValue();
6550 Value *FalseVal = SI->getFalseValue();
6551
6552 return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS,
6553 CastOp, Depth);
6554 }
6555
matchDecomposedSelectPattern(CmpInst * CmpI,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS,Instruction::CastOps * CastOp,unsigned Depth)6556 SelectPatternResult llvm::matchDecomposedSelectPattern(
6557 CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS,
6558 Instruction::CastOps *CastOp, unsigned Depth) {
6559 CmpInst::Predicate Pred = CmpI->getPredicate();
6560 Value *CmpLHS = CmpI->getOperand(0);
6561 Value *CmpRHS = CmpI->getOperand(1);
6562 FastMathFlags FMF;
6563 if (isa<FPMathOperator>(CmpI))
6564 FMF = CmpI->getFastMathFlags();
6565
6566 // Bail out early.
6567 if (CmpI->isEquality())
6568 return {SPF_UNKNOWN, SPNB_NA, false};
6569
6570 // Deal with type mismatches.
6571 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
6572 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
6573 // If this is a potential fmin/fmax with a cast to integer, then ignore
6574 // -0.0 because there is no corresponding integer value.
6575 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
6576 FMF.setNoSignedZeros();
6577 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
6578 cast<CastInst>(TrueVal)->getOperand(0), C,
6579 LHS, RHS, Depth);
6580 }
6581 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
6582 // If this is a potential fmin/fmax with a cast to integer, then ignore
6583 // -0.0 because there is no corresponding integer value.
6584 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
6585 FMF.setNoSignedZeros();
6586 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
6587 C, cast<CastInst>(FalseVal)->getOperand(0),
6588 LHS, RHS, Depth);
6589 }
6590 }
6591 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
6592 LHS, RHS, Depth);
6593 }
6594
getMinMaxPred(SelectPatternFlavor SPF,bool Ordered)6595 CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) {
6596 if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
6597 if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
6598 if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
6599 if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
6600 if (SPF == SPF_FMINNUM)
6601 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
6602 if (SPF == SPF_FMAXNUM)
6603 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
6604 llvm_unreachable("unhandled!");
6605 }
6606
getInverseMinMaxFlavor(SelectPatternFlavor SPF)6607 SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) {
6608 if (SPF == SPF_SMIN) return SPF_SMAX;
6609 if (SPF == SPF_UMIN) return SPF_UMAX;
6610 if (SPF == SPF_SMAX) return SPF_SMIN;
6611 if (SPF == SPF_UMAX) return SPF_UMIN;
6612 llvm_unreachable("unhandled!");
6613 }
6614
getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID)6615 Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) {
6616 switch (MinMaxID) {
6617 case Intrinsic::smax: return Intrinsic::smin;
6618 case Intrinsic::smin: return Intrinsic::smax;
6619 case Intrinsic::umax: return Intrinsic::umin;
6620 case Intrinsic::umin: return Intrinsic::umax;
6621 default: llvm_unreachable("Unexpected intrinsic");
6622 }
6623 }
6624
getMinMaxLimit(SelectPatternFlavor SPF,unsigned BitWidth)6625 APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) {
6626 switch (SPF) {
6627 case SPF_SMAX: return APInt::getSignedMaxValue(BitWidth);
6628 case SPF_SMIN: return APInt::getSignedMinValue(BitWidth);
6629 case SPF_UMAX: return APInt::getMaxValue(BitWidth);
6630 case SPF_UMIN: return APInt::getMinValue(BitWidth);
6631 default: llvm_unreachable("Unexpected flavor");
6632 }
6633 }
6634
6635 std::pair<Intrinsic::ID, bool>
canConvertToMinOrMaxIntrinsic(ArrayRef<Value * > VL)6636 llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) {
6637 // Check if VL contains select instructions that can be folded into a min/max
6638 // vector intrinsic and return the intrinsic if it is possible.
6639 // TODO: Support floating point min/max.
6640 bool AllCmpSingleUse = true;
6641 SelectPatternResult SelectPattern;
6642 SelectPattern.Flavor = SPF_UNKNOWN;
6643 if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) {
6644 Value *LHS, *RHS;
6645 auto CurrentPattern = matchSelectPattern(I, LHS, RHS);
6646 if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) ||
6647 CurrentPattern.Flavor == SPF_FMINNUM ||
6648 CurrentPattern.Flavor == SPF_FMAXNUM ||
6649 !I->getType()->isIntOrIntVectorTy())
6650 return false;
6651 if (SelectPattern.Flavor != SPF_UNKNOWN &&
6652 SelectPattern.Flavor != CurrentPattern.Flavor)
6653 return false;
6654 SelectPattern = CurrentPattern;
6655 AllCmpSingleUse &=
6656 match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value()));
6657 return true;
6658 })) {
6659 switch (SelectPattern.Flavor) {
6660 case SPF_SMIN:
6661 return {Intrinsic::smin, AllCmpSingleUse};
6662 case SPF_UMIN:
6663 return {Intrinsic::umin, AllCmpSingleUse};
6664 case SPF_SMAX:
6665 return {Intrinsic::smax, AllCmpSingleUse};
6666 case SPF_UMAX:
6667 return {Intrinsic::umax, AllCmpSingleUse};
6668 default:
6669 llvm_unreachable("unexpected select pattern flavor");
6670 }
6671 }
6672 return {Intrinsic::not_intrinsic, false};
6673 }
6674
matchSimpleRecurrence(const PHINode * P,BinaryOperator * & BO,Value * & Start,Value * & Step)6675 bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO,
6676 Value *&Start, Value *&Step) {
6677 // Handle the case of a simple two-predecessor recurrence PHI.
6678 // There's a lot more that could theoretically be done here, but
6679 // this is sufficient to catch some interesting cases.
6680 if (P->getNumIncomingValues() != 2)
6681 return false;
6682
6683 for (unsigned i = 0; i != 2; ++i) {
6684 Value *L = P->getIncomingValue(i);
6685 Value *R = P->getIncomingValue(!i);
6686 Operator *LU = dyn_cast<Operator>(L);
6687 if (!LU)
6688 continue;
6689 unsigned Opcode = LU->getOpcode();
6690
6691 switch (Opcode) {
6692 default:
6693 continue;
6694 // TODO: Expand list -- xor, div, gep, uaddo, etc..
6695 case Instruction::LShr:
6696 case Instruction::AShr:
6697 case Instruction::Shl:
6698 case Instruction::Add:
6699 case Instruction::Sub:
6700 case Instruction::And:
6701 case Instruction::Or:
6702 case Instruction::Mul:
6703 case Instruction::FMul: {
6704 Value *LL = LU->getOperand(0);
6705 Value *LR = LU->getOperand(1);
6706 // Find a recurrence.
6707 if (LL == P)
6708 L = LR;
6709 else if (LR == P)
6710 L = LL;
6711 else
6712 continue; // Check for recurrence with L and R flipped.
6713
6714 break; // Match!
6715 }
6716 };
6717
6718 // We have matched a recurrence of the form:
6719 // %iv = [R, %entry], [%iv.next, %backedge]
6720 // %iv.next = binop %iv, L
6721 // OR
6722 // %iv = [R, %entry], [%iv.next, %backedge]
6723 // %iv.next = binop L, %iv
6724 BO = cast<BinaryOperator>(LU);
6725 Start = R;
6726 Step = L;
6727 return true;
6728 }
6729 return false;
6730 }
6731
matchSimpleRecurrence(const BinaryOperator * I,PHINode * & P,Value * & Start,Value * & Step)6732 bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P,
6733 Value *&Start, Value *&Step) {
6734 BinaryOperator *BO = nullptr;
6735 P = dyn_cast<PHINode>(I->getOperand(0));
6736 if (!P)
6737 P = dyn_cast<PHINode>(I->getOperand(1));
6738 return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I;
6739 }
6740
6741 /// Return true if "icmp Pred LHS RHS" is always true.
isTruePredicate(CmpInst::Predicate Pred,const Value * LHS,const Value * RHS,const DataLayout & DL,unsigned Depth)6742 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
6743 const Value *RHS, const DataLayout &DL,
6744 unsigned Depth) {
6745 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
6746 return true;
6747
6748 switch (Pred) {
6749 default:
6750 return false;
6751
6752 case CmpInst::ICMP_SLE: {
6753 const APInt *C;
6754
6755 // LHS s<= LHS +_{nsw} C if C >= 0
6756 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
6757 return !C->isNegative();
6758 return false;
6759 }
6760
6761 case CmpInst::ICMP_ULE: {
6762 const APInt *C;
6763
6764 // LHS u<= LHS +_{nuw} C for any C
6765 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
6766 return true;
6767
6768 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
6769 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
6770 const Value *&X,
6771 const APInt *&CA, const APInt *&CB) {
6772 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
6773 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
6774 return true;
6775
6776 // If X & C == 0 then (X | C) == X +_{nuw} C
6777 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
6778 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
6779 KnownBits Known(CA->getBitWidth());
6780 computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr,
6781 /*CxtI*/ nullptr, /*DT*/ nullptr);
6782 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
6783 return true;
6784 }
6785
6786 return false;
6787 };
6788
6789 const Value *X;
6790 const APInt *CLHS, *CRHS;
6791 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
6792 return CLHS->ule(*CRHS);
6793
6794 return false;
6795 }
6796 }
6797 }
6798
6799 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
6800 /// ALHS ARHS" is true. Otherwise, return std::nullopt.
6801 static std::optional<bool>
isImpliedCondOperands(CmpInst::Predicate Pred,const Value * ALHS,const Value * ARHS,const Value * BLHS,const Value * BRHS,const DataLayout & DL,unsigned Depth)6802 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
6803 const Value *ARHS, const Value *BLHS, const Value *BRHS,
6804 const DataLayout &DL, unsigned Depth) {
6805 switch (Pred) {
6806 default:
6807 return std::nullopt;
6808
6809 case CmpInst::ICMP_SLT:
6810 case CmpInst::ICMP_SLE:
6811 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) &&
6812 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth))
6813 return true;
6814 return std::nullopt;
6815
6816 case CmpInst::ICMP_ULT:
6817 case CmpInst::ICMP_ULE:
6818 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) &&
6819 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth))
6820 return true;
6821 return std::nullopt;
6822 }
6823 }
6824
6825 /// Return true if the operands of two compares (expanded as "L0 pred L1" and
6826 /// "R0 pred R1") match. IsSwappedOps is true when the operands match, but are
6827 /// swapped.
areMatchingOperands(const Value * L0,const Value * L1,const Value * R0,const Value * R1,bool & AreSwappedOps)6828 static bool areMatchingOperands(const Value *L0, const Value *L1, const Value *R0,
6829 const Value *R1, bool &AreSwappedOps) {
6830 bool AreMatchingOps = (L0 == R0 && L1 == R1);
6831 AreSwappedOps = (L0 == R1 && L1 == R0);
6832 return AreMatchingOps || AreSwappedOps;
6833 }
6834
6835 /// Return true if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is true.
6836 /// Return false if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is false.
6837 /// Otherwise, return std::nullopt if we can't infer anything.
6838 static std::optional<bool>
isImpliedCondMatchingOperands(CmpInst::Predicate LPred,CmpInst::Predicate RPred,bool AreSwappedOps)6839 isImpliedCondMatchingOperands(CmpInst::Predicate LPred,
6840 CmpInst::Predicate RPred, bool AreSwappedOps) {
6841 // Canonicalize the predicate as if the operands were not commuted.
6842 if (AreSwappedOps)
6843 RPred = ICmpInst::getSwappedPredicate(RPred);
6844
6845 if (CmpInst::isImpliedTrueByMatchingCmp(LPred, RPred))
6846 return true;
6847 if (CmpInst::isImpliedFalseByMatchingCmp(LPred, RPred))
6848 return false;
6849
6850 return std::nullopt;
6851 }
6852
6853 /// Return true if "icmp LPred X, LC" implies "icmp RPred X, RC" is true.
6854 /// Return false if "icmp LPred X, LC" implies "icmp RPred X, RC" is false.
6855 /// Otherwise, return std::nullopt if we can't infer anything.
isImpliedCondCommonOperandWithConstants(CmpInst::Predicate LPred,const APInt & LC,CmpInst::Predicate RPred,const APInt & RC)6856 static std::optional<bool> isImpliedCondCommonOperandWithConstants(
6857 CmpInst::Predicate LPred, const APInt &LC, CmpInst::Predicate RPred,
6858 const APInt &RC) {
6859 ConstantRange DomCR = ConstantRange::makeExactICmpRegion(LPred, LC);
6860 ConstantRange CR = ConstantRange::makeExactICmpRegion(RPred, RC);
6861 ConstantRange Intersection = DomCR.intersectWith(CR);
6862 ConstantRange Difference = DomCR.difference(CR);
6863 if (Intersection.isEmptySet())
6864 return false;
6865 if (Difference.isEmptySet())
6866 return true;
6867 return std::nullopt;
6868 }
6869
6870 /// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1")
6871 /// is true. Return false if LHS implies RHS is false. Otherwise, return
6872 /// std::nullopt if we can't infer anything.
isImpliedCondICmps(const ICmpInst * LHS,CmpInst::Predicate RPred,const Value * R0,const Value * R1,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)6873 static std::optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
6874 CmpInst::Predicate RPred,
6875 const Value *R0, const Value *R1,
6876 const DataLayout &DL,
6877 bool LHSIsTrue, unsigned Depth) {
6878 Value *L0 = LHS->getOperand(0);
6879 Value *L1 = LHS->getOperand(1);
6880
6881 // The rest of the logic assumes the LHS condition is true. If that's not the
6882 // case, invert the predicate to make it so.
6883 CmpInst::Predicate LPred =
6884 LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate();
6885
6886 // Can we infer anything when the two compares have matching operands?
6887 bool AreSwappedOps;
6888 if (areMatchingOperands(L0, L1, R0, R1, AreSwappedOps))
6889 return isImpliedCondMatchingOperands(LPred, RPred, AreSwappedOps);
6890
6891 // Can we infer anything when the 0-operands match and the 1-operands are
6892 // constants (not necessarily matching)?
6893 const APInt *LC, *RC;
6894 if (L0 == R0 && match(L1, m_APInt(LC)) && match(R1, m_APInt(RC)))
6895 return isImpliedCondCommonOperandWithConstants(LPred, *LC, RPred, *RC);
6896
6897 if (LPred == RPred)
6898 return isImpliedCondOperands(LPred, L0, L1, R0, R1, DL, Depth);
6899
6900 return std::nullopt;
6901 }
6902
6903 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
6904 /// false. Otherwise, return std::nullopt if we can't infer anything. We
6905 /// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select'
6906 /// instruction.
6907 static std::optional<bool>
isImpliedCondAndOr(const Instruction * LHS,CmpInst::Predicate RHSPred,const Value * RHSOp0,const Value * RHSOp1,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)6908 isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred,
6909 const Value *RHSOp0, const Value *RHSOp1,
6910 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) {
6911 // The LHS must be an 'or', 'and', or a 'select' instruction.
6912 assert((LHS->getOpcode() == Instruction::And ||
6913 LHS->getOpcode() == Instruction::Or ||
6914 LHS->getOpcode() == Instruction::Select) &&
6915 "Expected LHS to be 'and', 'or', or 'select'.");
6916
6917 assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit");
6918
6919 // If the result of an 'or' is false, then we know both legs of the 'or' are
6920 // false. Similarly, if the result of an 'and' is true, then we know both
6921 // legs of the 'and' are true.
6922 const Value *ALHS, *ARHS;
6923 if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) ||
6924 (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) {
6925 // FIXME: Make this non-recursion.
6926 if (std::optional<bool> Implication = isImpliedCondition(
6927 ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1))
6928 return Implication;
6929 if (std::optional<bool> Implication = isImpliedCondition(
6930 ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1))
6931 return Implication;
6932 return std::nullopt;
6933 }
6934 return std::nullopt;
6935 }
6936
6937 std::optional<bool>
isImpliedCondition(const Value * LHS,CmpInst::Predicate RHSPred,const Value * RHSOp0,const Value * RHSOp1,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)6938 llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred,
6939 const Value *RHSOp0, const Value *RHSOp1,
6940 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) {
6941 // Bail out when we hit the limit.
6942 if (Depth == MaxAnalysisRecursionDepth)
6943 return std::nullopt;
6944
6945 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for
6946 // example.
6947 if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy())
6948 return std::nullopt;
6949
6950 assert(LHS->getType()->isIntOrIntVectorTy(1) &&
6951 "Expected integer type only!");
6952
6953 // Both LHS and RHS are icmps.
6954 const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS);
6955 if (LHSCmp)
6956 return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue,
6957 Depth);
6958
6959 /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect
6960 /// the RHS to be an icmp.
6961 /// FIXME: Add support for and/or/select on the RHS.
6962 if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
6963 if ((LHSI->getOpcode() == Instruction::And ||
6964 LHSI->getOpcode() == Instruction::Or ||
6965 LHSI->getOpcode() == Instruction::Select))
6966 return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue,
6967 Depth);
6968 }
6969 return std::nullopt;
6970 }
6971
isImpliedCondition(const Value * LHS,const Value * RHS,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)6972 std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
6973 const DataLayout &DL,
6974 bool LHSIsTrue, unsigned Depth) {
6975 // LHS ==> RHS by definition
6976 if (LHS == RHS)
6977 return LHSIsTrue;
6978
6979 if (const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS))
6980 return isImpliedCondition(LHS, RHSCmp->getPredicate(),
6981 RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL,
6982 LHSIsTrue, Depth);
6983
6984 if (Depth == MaxAnalysisRecursionDepth)
6985 return std::nullopt;
6986
6987 // LHS ==> (RHS1 || RHS2) if LHS ==> RHS1 or LHS ==> RHS2
6988 // LHS ==> !(RHS1 && RHS2) if LHS ==> !RHS1 or LHS ==> !RHS2
6989 const Value *RHS1, *RHS2;
6990 if (match(RHS, m_LogicalOr(m_Value(RHS1), m_Value(RHS2)))) {
6991 if (std::optional<bool> Imp =
6992 isImpliedCondition(LHS, RHS1, DL, LHSIsTrue, Depth + 1))
6993 if (*Imp == true)
6994 return true;
6995 if (std::optional<bool> Imp =
6996 isImpliedCondition(LHS, RHS2, DL, LHSIsTrue, Depth + 1))
6997 if (*Imp == true)
6998 return true;
6999 }
7000 if (match(RHS, m_LogicalAnd(m_Value(RHS1), m_Value(RHS2)))) {
7001 if (std::optional<bool> Imp =
7002 isImpliedCondition(LHS, RHS1, DL, LHSIsTrue, Depth + 1))
7003 if (*Imp == false)
7004 return false;
7005 if (std::optional<bool> Imp =
7006 isImpliedCondition(LHS, RHS2, DL, LHSIsTrue, Depth + 1))
7007 if (*Imp == false)
7008 return false;
7009 }
7010
7011 return std::nullopt;
7012 }
7013
7014 // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch
7015 // condition dominating ContextI or nullptr, if no condition is found.
7016 static std::pair<Value *, bool>
getDomPredecessorCondition(const Instruction * ContextI)7017 getDomPredecessorCondition(const Instruction *ContextI) {
7018 if (!ContextI || !ContextI->getParent())
7019 return {nullptr, false};
7020
7021 // TODO: This is a poor/cheap way to determine dominance. Should we use a
7022 // dominator tree (eg, from a SimplifyQuery) instead?
7023 const BasicBlock *ContextBB = ContextI->getParent();
7024 const BasicBlock *PredBB = ContextBB->getSinglePredecessor();
7025 if (!PredBB)
7026 return {nullptr, false};
7027
7028 // We need a conditional branch in the predecessor.
7029 Value *PredCond;
7030 BasicBlock *TrueBB, *FalseBB;
7031 if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB)))
7032 return {nullptr, false};
7033
7034 // The branch should get simplified. Don't bother simplifying this condition.
7035 if (TrueBB == FalseBB)
7036 return {nullptr, false};
7037
7038 assert((TrueBB == ContextBB || FalseBB == ContextBB) &&
7039 "Predecessor block does not point to successor?");
7040
7041 // Is this condition implied by the predecessor condition?
7042 return {PredCond, TrueBB == ContextBB};
7043 }
7044
isImpliedByDomCondition(const Value * Cond,const Instruction * ContextI,const DataLayout & DL)7045 std::optional<bool> llvm::isImpliedByDomCondition(const Value *Cond,
7046 const Instruction *ContextI,
7047 const DataLayout &DL) {
7048 assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool");
7049 auto PredCond = getDomPredecessorCondition(ContextI);
7050 if (PredCond.first)
7051 return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second);
7052 return std::nullopt;
7053 }
7054
isImpliedByDomCondition(CmpInst::Predicate Pred,const Value * LHS,const Value * RHS,const Instruction * ContextI,const DataLayout & DL)7055 std::optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred,
7056 const Value *LHS,
7057 const Value *RHS,
7058 const Instruction *ContextI,
7059 const DataLayout &DL) {
7060 auto PredCond = getDomPredecessorCondition(ContextI);
7061 if (PredCond.first)
7062 return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL,
7063 PredCond.second);
7064 return std::nullopt;
7065 }
7066
setLimitsForBinOp(const BinaryOperator & BO,APInt & Lower,APInt & Upper,const InstrInfoQuery & IIQ,bool PreferSignedRange)7067 static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower,
7068 APInt &Upper, const InstrInfoQuery &IIQ,
7069 bool PreferSignedRange) {
7070 unsigned Width = Lower.getBitWidth();
7071 const APInt *C;
7072 switch (BO.getOpcode()) {
7073 case Instruction::Add:
7074 if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) {
7075 bool HasNSW = IIQ.hasNoSignedWrap(&BO);
7076 bool HasNUW = IIQ.hasNoUnsignedWrap(&BO);
7077
7078 // If the caller expects a signed compare, then try to use a signed range.
7079 // Otherwise if both no-wraps are set, use the unsigned range because it
7080 // is never larger than the signed range. Example:
7081 // "add nuw nsw i8 X, -2" is unsigned [254,255] vs. signed [-128, 125].
7082 if (PreferSignedRange && HasNSW && HasNUW)
7083 HasNUW = false;
7084
7085 if (HasNUW) {
7086 // 'add nuw x, C' produces [C, UINT_MAX].
7087 Lower = *C;
7088 } else if (HasNSW) {
7089 if (C->isNegative()) {
7090 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
7091 Lower = APInt::getSignedMinValue(Width);
7092 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
7093 } else {
7094 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
7095 Lower = APInt::getSignedMinValue(Width) + *C;
7096 Upper = APInt::getSignedMaxValue(Width) + 1;
7097 }
7098 }
7099 }
7100 break;
7101
7102 case Instruction::And:
7103 if (match(BO.getOperand(1), m_APInt(C)))
7104 // 'and x, C' produces [0, C].
7105 Upper = *C + 1;
7106 break;
7107
7108 case Instruction::Or:
7109 if (match(BO.getOperand(1), m_APInt(C)))
7110 // 'or x, C' produces [C, UINT_MAX].
7111 Lower = *C;
7112 break;
7113
7114 case Instruction::AShr:
7115 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
7116 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
7117 Lower = APInt::getSignedMinValue(Width).ashr(*C);
7118 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
7119 } else if (match(BO.getOperand(0), m_APInt(C))) {
7120 unsigned ShiftAmount = Width - 1;
7121 if (!C->isZero() && IIQ.isExact(&BO))
7122 ShiftAmount = C->countTrailingZeros();
7123 if (C->isNegative()) {
7124 // 'ashr C, x' produces [C, C >> (Width-1)]
7125 Lower = *C;
7126 Upper = C->ashr(ShiftAmount) + 1;
7127 } else {
7128 // 'ashr C, x' produces [C >> (Width-1), C]
7129 Lower = C->ashr(ShiftAmount);
7130 Upper = *C + 1;
7131 }
7132 }
7133 break;
7134
7135 case Instruction::LShr:
7136 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
7137 // 'lshr x, C' produces [0, UINT_MAX >> C].
7138 Upper = APInt::getAllOnes(Width).lshr(*C) + 1;
7139 } else if (match(BO.getOperand(0), m_APInt(C))) {
7140 // 'lshr C, x' produces [C >> (Width-1), C].
7141 unsigned ShiftAmount = Width - 1;
7142 if (!C->isZero() && IIQ.isExact(&BO))
7143 ShiftAmount = C->countTrailingZeros();
7144 Lower = C->lshr(ShiftAmount);
7145 Upper = *C + 1;
7146 }
7147 break;
7148
7149 case Instruction::Shl:
7150 if (match(BO.getOperand(0), m_APInt(C))) {
7151 if (IIQ.hasNoUnsignedWrap(&BO)) {
7152 // 'shl nuw C, x' produces [C, C << CLZ(C)]
7153 Lower = *C;
7154 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
7155 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
7156 if (C->isNegative()) {
7157 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
7158 unsigned ShiftAmount = C->countLeadingOnes() - 1;
7159 Lower = C->shl(ShiftAmount);
7160 Upper = *C + 1;
7161 } else {
7162 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
7163 unsigned ShiftAmount = C->countLeadingZeros() - 1;
7164 Lower = *C;
7165 Upper = C->shl(ShiftAmount) + 1;
7166 }
7167 }
7168 }
7169 break;
7170
7171 case Instruction::SDiv:
7172 if (match(BO.getOperand(1), m_APInt(C))) {
7173 APInt IntMin = APInt::getSignedMinValue(Width);
7174 APInt IntMax = APInt::getSignedMaxValue(Width);
7175 if (C->isAllOnes()) {
7176 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
7177 // where C != -1 and C != 0 and C != 1
7178 Lower = IntMin + 1;
7179 Upper = IntMax + 1;
7180 } else if (C->countLeadingZeros() < Width - 1) {
7181 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
7182 // where C != -1 and C != 0 and C != 1
7183 Lower = IntMin.sdiv(*C);
7184 Upper = IntMax.sdiv(*C);
7185 if (Lower.sgt(Upper))
7186 std::swap(Lower, Upper);
7187 Upper = Upper + 1;
7188 assert(Upper != Lower && "Upper part of range has wrapped!");
7189 }
7190 } else if (match(BO.getOperand(0), m_APInt(C))) {
7191 if (C->isMinSignedValue()) {
7192 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
7193 Lower = *C;
7194 Upper = Lower.lshr(1) + 1;
7195 } else {
7196 // 'sdiv C, x' produces [-|C|, |C|].
7197 Upper = C->abs() + 1;
7198 Lower = (-Upper) + 1;
7199 }
7200 }
7201 break;
7202
7203 case Instruction::UDiv:
7204 if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) {
7205 // 'udiv x, C' produces [0, UINT_MAX / C].
7206 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
7207 } else if (match(BO.getOperand(0), m_APInt(C))) {
7208 // 'udiv C, x' produces [0, C].
7209 Upper = *C + 1;
7210 }
7211 break;
7212
7213 case Instruction::SRem:
7214 if (match(BO.getOperand(1), m_APInt(C))) {
7215 // 'srem x, C' produces (-|C|, |C|).
7216 Upper = C->abs();
7217 Lower = (-Upper) + 1;
7218 }
7219 break;
7220
7221 case Instruction::URem:
7222 if (match(BO.getOperand(1), m_APInt(C)))
7223 // 'urem x, C' produces [0, C).
7224 Upper = *C;
7225 break;
7226
7227 default:
7228 break;
7229 }
7230 }
7231
setLimitsForIntrinsic(const IntrinsicInst & II,APInt & Lower,APInt & Upper)7232 static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower,
7233 APInt &Upper) {
7234 unsigned Width = Lower.getBitWidth();
7235 const APInt *C;
7236 switch (II.getIntrinsicID()) {
7237 case Intrinsic::ctpop:
7238 case Intrinsic::ctlz:
7239 case Intrinsic::cttz:
7240 // Maximum of set/clear bits is the bit width.
7241 assert(Lower == 0 && "Expected lower bound to be zero");
7242 Upper = Width + 1;
7243 break;
7244 case Intrinsic::uadd_sat:
7245 // uadd.sat(x, C) produces [C, UINT_MAX].
7246 if (match(II.getOperand(0), m_APInt(C)) ||
7247 match(II.getOperand(1), m_APInt(C)))
7248 Lower = *C;
7249 break;
7250 case Intrinsic::sadd_sat:
7251 if (match(II.getOperand(0), m_APInt(C)) ||
7252 match(II.getOperand(1), m_APInt(C))) {
7253 if (C->isNegative()) {
7254 // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)].
7255 Lower = APInt::getSignedMinValue(Width);
7256 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
7257 } else {
7258 // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX].
7259 Lower = APInt::getSignedMinValue(Width) + *C;
7260 Upper = APInt::getSignedMaxValue(Width) + 1;
7261 }
7262 }
7263 break;
7264 case Intrinsic::usub_sat:
7265 // usub.sat(C, x) produces [0, C].
7266 if (match(II.getOperand(0), m_APInt(C)))
7267 Upper = *C + 1;
7268 // usub.sat(x, C) produces [0, UINT_MAX - C].
7269 else if (match(II.getOperand(1), m_APInt(C)))
7270 Upper = APInt::getMaxValue(Width) - *C + 1;
7271 break;
7272 case Intrinsic::ssub_sat:
7273 if (match(II.getOperand(0), m_APInt(C))) {
7274 if (C->isNegative()) {
7275 // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)].
7276 Lower = APInt::getSignedMinValue(Width);
7277 Upper = *C - APInt::getSignedMinValue(Width) + 1;
7278 } else {
7279 // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX].
7280 Lower = *C - APInt::getSignedMaxValue(Width);
7281 Upper = APInt::getSignedMaxValue(Width) + 1;
7282 }
7283 } else if (match(II.getOperand(1), m_APInt(C))) {
7284 if (C->isNegative()) {
7285 // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]:
7286 Lower = APInt::getSignedMinValue(Width) - *C;
7287 Upper = APInt::getSignedMaxValue(Width) + 1;
7288 } else {
7289 // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C].
7290 Lower = APInt::getSignedMinValue(Width);
7291 Upper = APInt::getSignedMaxValue(Width) - *C + 1;
7292 }
7293 }
7294 break;
7295 case Intrinsic::umin:
7296 case Intrinsic::umax:
7297 case Intrinsic::smin:
7298 case Intrinsic::smax:
7299 if (!match(II.getOperand(0), m_APInt(C)) &&
7300 !match(II.getOperand(1), m_APInt(C)))
7301 break;
7302
7303 switch (II.getIntrinsicID()) {
7304 case Intrinsic::umin:
7305 Upper = *C + 1;
7306 break;
7307 case Intrinsic::umax:
7308 Lower = *C;
7309 break;
7310 case Intrinsic::smin:
7311 Lower = APInt::getSignedMinValue(Width);
7312 Upper = *C + 1;
7313 break;
7314 case Intrinsic::smax:
7315 Lower = *C;
7316 Upper = APInt::getSignedMaxValue(Width) + 1;
7317 break;
7318 default:
7319 llvm_unreachable("Must be min/max intrinsic");
7320 }
7321 break;
7322 case Intrinsic::abs:
7323 // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX],
7324 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN.
7325 if (match(II.getOperand(1), m_One()))
7326 Upper = APInt::getSignedMaxValue(Width) + 1;
7327 else
7328 Upper = APInt::getSignedMinValue(Width) + 1;
7329 break;
7330 default:
7331 break;
7332 }
7333 }
7334
setLimitsForSelectPattern(const SelectInst & SI,APInt & Lower,APInt & Upper,const InstrInfoQuery & IIQ)7335 static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower,
7336 APInt &Upper, const InstrInfoQuery &IIQ) {
7337 const Value *LHS = nullptr, *RHS = nullptr;
7338 SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS);
7339 if (R.Flavor == SPF_UNKNOWN)
7340 return;
7341
7342 unsigned BitWidth = SI.getType()->getScalarSizeInBits();
7343
7344 if (R.Flavor == SelectPatternFlavor::SPF_ABS) {
7345 // If the negation part of the abs (in RHS) has the NSW flag,
7346 // then the result of abs(X) is [0..SIGNED_MAX],
7347 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN.
7348 Lower = APInt::getZero(BitWidth);
7349 if (match(RHS, m_Neg(m_Specific(LHS))) &&
7350 IIQ.hasNoSignedWrap(cast<Instruction>(RHS)))
7351 Upper = APInt::getSignedMaxValue(BitWidth) + 1;
7352 else
7353 Upper = APInt::getSignedMinValue(BitWidth) + 1;
7354 return;
7355 }
7356
7357 if (R.Flavor == SelectPatternFlavor::SPF_NABS) {
7358 // The result of -abs(X) is <= 0.
7359 Lower = APInt::getSignedMinValue(BitWidth);
7360 Upper = APInt(BitWidth, 1);
7361 return;
7362 }
7363
7364 const APInt *C;
7365 if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C)))
7366 return;
7367
7368 switch (R.Flavor) {
7369 case SPF_UMIN:
7370 Upper = *C + 1;
7371 break;
7372 case SPF_UMAX:
7373 Lower = *C;
7374 break;
7375 case SPF_SMIN:
7376 Lower = APInt::getSignedMinValue(BitWidth);
7377 Upper = *C + 1;
7378 break;
7379 case SPF_SMAX:
7380 Lower = *C;
7381 Upper = APInt::getSignedMaxValue(BitWidth) + 1;
7382 break;
7383 default:
7384 break;
7385 }
7386 }
7387
setLimitForFPToI(const Instruction * I,APInt & Lower,APInt & Upper)7388 static void setLimitForFPToI(const Instruction *I, APInt &Lower, APInt &Upper) {
7389 // The maximum representable value of a half is 65504. For floats the maximum
7390 // value is 3.4e38 which requires roughly 129 bits.
7391 unsigned BitWidth = I->getType()->getScalarSizeInBits();
7392 if (!I->getOperand(0)->getType()->getScalarType()->isHalfTy())
7393 return;
7394 if (isa<FPToSIInst>(I) && BitWidth >= 17) {
7395 Lower = APInt(BitWidth, -65504);
7396 Upper = APInt(BitWidth, 65505);
7397 }
7398
7399 if (isa<FPToUIInst>(I) && BitWidth >= 16) {
7400 // For a fptoui the lower limit is left as 0.
7401 Upper = APInt(BitWidth, 65505);
7402 }
7403 }
7404
computeConstantRange(const Value * V,bool ForSigned,bool UseInstrInfo,AssumptionCache * AC,const Instruction * CtxI,const DominatorTree * DT,unsigned Depth)7405 ConstantRange llvm::computeConstantRange(const Value *V, bool ForSigned,
7406 bool UseInstrInfo, AssumptionCache *AC,
7407 const Instruction *CtxI,
7408 const DominatorTree *DT,
7409 unsigned Depth) {
7410 assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction");
7411
7412 if (Depth == MaxAnalysisRecursionDepth)
7413 return ConstantRange::getFull(V->getType()->getScalarSizeInBits());
7414
7415 const APInt *C;
7416 if (match(V, m_APInt(C)))
7417 return ConstantRange(*C);
7418
7419 InstrInfoQuery IIQ(UseInstrInfo);
7420 unsigned BitWidth = V->getType()->getScalarSizeInBits();
7421 APInt Lower = APInt(BitWidth, 0);
7422 APInt Upper = APInt(BitWidth, 0);
7423 if (auto *BO = dyn_cast<BinaryOperator>(V))
7424 setLimitsForBinOp(*BO, Lower, Upper, IIQ, ForSigned);
7425 else if (auto *II = dyn_cast<IntrinsicInst>(V))
7426 setLimitsForIntrinsic(*II, Lower, Upper);
7427 else if (auto *SI = dyn_cast<SelectInst>(V))
7428 setLimitsForSelectPattern(*SI, Lower, Upper, IIQ);
7429 else if (isa<FPToUIInst>(V) || isa<FPToSIInst>(V))
7430 setLimitForFPToI(cast<Instruction>(V), Lower, Upper);
7431
7432 ConstantRange CR = ConstantRange::getNonEmpty(Lower, Upper);
7433
7434 if (auto *I = dyn_cast<Instruction>(V))
7435 if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range))
7436 CR = CR.intersectWith(getConstantRangeFromMetadata(*Range));
7437
7438 if (CtxI && AC) {
7439 // Try to restrict the range based on information from assumptions.
7440 for (auto &AssumeVH : AC->assumptionsFor(V)) {
7441 if (!AssumeVH)
7442 continue;
7443 IntrinsicInst *I = cast<IntrinsicInst>(AssumeVH);
7444 assert(I->getParent()->getParent() == CtxI->getParent()->getParent() &&
7445 "Got assumption for the wrong function!");
7446
7447 if (!isValidAssumeForContext(I, CtxI, DT))
7448 continue;
7449 Value *Arg = I->getArgOperand(0);
7450 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg);
7451 // Currently we just use information from comparisons.
7452 if (!Cmp || Cmp->getOperand(0) != V)
7453 continue;
7454 // TODO: Set "ForSigned" parameter via Cmp->isSigned()?
7455 ConstantRange RHS =
7456 computeConstantRange(Cmp->getOperand(1), /* ForSigned */ false,
7457 UseInstrInfo, AC, I, DT, Depth + 1);
7458 CR = CR.intersectWith(
7459 ConstantRange::makeAllowedICmpRegion(Cmp->getPredicate(), RHS));
7460 }
7461 }
7462
7463 return CR;
7464 }
7465
7466 static std::optional<int64_t>
getOffsetFromIndex(const GEPOperator * GEP,unsigned Idx,const DataLayout & DL)7467 getOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, const DataLayout &DL) {
7468 // Skip over the first indices.
7469 gep_type_iterator GTI = gep_type_begin(GEP);
7470 for (unsigned i = 1; i != Idx; ++i, ++GTI)
7471 /*skip along*/;
7472
7473 // Compute the offset implied by the rest of the indices.
7474 int64_t Offset = 0;
7475 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
7476 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
7477 if (!OpC)
7478 return std::nullopt;
7479 if (OpC->isZero())
7480 continue; // No offset.
7481
7482 // Handle struct indices, which add their field offset to the pointer.
7483 if (StructType *STy = GTI.getStructTypeOrNull()) {
7484 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
7485 continue;
7486 }
7487
7488 // Otherwise, we have a sequential type like an array or fixed-length
7489 // vector. Multiply the index by the ElementSize.
7490 TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType());
7491 if (Size.isScalable())
7492 return std::nullopt;
7493 Offset += Size.getFixedValue() * OpC->getSExtValue();
7494 }
7495
7496 return Offset;
7497 }
7498
isPointerOffset(const Value * Ptr1,const Value * Ptr2,const DataLayout & DL)7499 std::optional<int64_t> llvm::isPointerOffset(const Value *Ptr1,
7500 const Value *Ptr2,
7501 const DataLayout &DL) {
7502 APInt Offset1(DL.getIndexTypeSizeInBits(Ptr1->getType()), 0);
7503 APInt Offset2(DL.getIndexTypeSizeInBits(Ptr2->getType()), 0);
7504 Ptr1 = Ptr1->stripAndAccumulateConstantOffsets(DL, Offset1, true);
7505 Ptr2 = Ptr2->stripAndAccumulateConstantOffsets(DL, Offset2, true);
7506
7507 // Handle the trivial case first.
7508 if (Ptr1 == Ptr2)
7509 return Offset2.getSExtValue() - Offset1.getSExtValue();
7510
7511 const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
7512 const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
7513
7514 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
7515 // base. After that base, they may have some number of common (and
7516 // potentially variable) indices. After that they handle some constant
7517 // offset, which determines their offset from each other. At this point, we
7518 // handle no other case.
7519 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0) ||
7520 GEP1->getSourceElementType() != GEP2->getSourceElementType())
7521 return std::nullopt;
7522
7523 // Skip any common indices and track the GEP types.
7524 unsigned Idx = 1;
7525 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
7526 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
7527 break;
7528
7529 auto IOffset1 = getOffsetFromIndex(GEP1, Idx, DL);
7530 auto IOffset2 = getOffsetFromIndex(GEP2, Idx, DL);
7531 if (!IOffset1 || !IOffset2)
7532 return std::nullopt;
7533 return *IOffset2 - *IOffset1 + Offset2.getSExtValue() -
7534 Offset1.getSExtValue();
7535 }
7536