File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Analysis/ValueTracking.cpp |
Warning: | line 211, column 31 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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/None.h" | |||
19 | #include "llvm/ADT/Optional.h" | |||
20 | #include "llvm/ADT/STLExtras.h" | |||
21 | #include "llvm/ADT/SmallPtrSet.h" | |||
22 | #include "llvm/ADT/SmallSet.h" | |||
23 | #include "llvm/ADT/SmallVector.h" | |||
24 | #include "llvm/ADT/StringRef.h" | |||
25 | #include "llvm/ADT/iterator_range.h" | |||
26 | #include "llvm/Analysis/AliasAnalysis.h" | |||
27 | #include "llvm/Analysis/AssumeBundleQueries.h" | |||
28 | #include "llvm/Analysis/AssumptionCache.h" | |||
29 | #include "llvm/Analysis/EHPersonalities.h" | |||
30 | #include "llvm/Analysis/GuardUtils.h" | |||
31 | #include "llvm/Analysis/InstructionSimplify.h" | |||
32 | #include "llvm/Analysis/Loads.h" | |||
33 | #include "llvm/Analysis/LoopInfo.h" | |||
34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
35 | #include "llvm/Analysis/TargetLibraryInfo.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 <utility> | |||
76 | ||||
77 | using namespace llvm; | |||
78 | using namespace llvm::PatternMatch; | |||
79 | ||||
80 | // Controls the number of uses of the value searched for possible | |||
81 | // dominating comparisons. | |||
82 | static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", | |||
83 | cl::Hidden, cl::init(20)); | |||
84 | ||||
85 | // According to the LangRef, branching on a poison condition is absolutely | |||
86 | // immediate full UB. However, historically we haven't implemented that | |||
87 | // consistently as we have an important transformation (non-trivial unswitch) | |||
88 | // which introduces instances of branch on poison/undef to otherwise well | |||
89 | // defined programs. This flag exists to let us test optimization benefit | |||
90 | // of exploiting the specified behavior (in combination with enabling the | |||
91 | // unswitch fix.) | |||
92 | static cl::opt<bool> BranchOnPoisonAsUB("branch-on-poison-as-ub", | |||
93 | cl::Hidden, cl::init(false)); | |||
94 | ||||
95 | ||||
96 | /// Returns the bitwidth of the given scalar or pointer type. For vector types, | |||
97 | /// returns the element type's bitwidth. | |||
98 | static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { | |||
99 | if (unsigned BitWidth = Ty->getScalarSizeInBits()) | |||
100 | return BitWidth; | |||
101 | ||||
102 | return DL.getPointerTypeSizeInBits(Ty); | |||
103 | } | |||
104 | ||||
105 | namespace { | |||
106 | ||||
107 | // Simplifying using an assume can only be done in a particular control-flow | |||
108 | // context (the context instruction provides that context). If an assume and | |||
109 | // the context instruction are not in the same block then the DT helps in | |||
110 | // figuring out if we can use it. | |||
111 | struct Query { | |||
112 | const DataLayout &DL; | |||
113 | AssumptionCache *AC; | |||
114 | const Instruction *CxtI; | |||
115 | const DominatorTree *DT; | |||
116 | ||||
117 | // Unlike the other analyses, this may be a nullptr because not all clients | |||
118 | // provide it currently. | |||
119 | OptimizationRemarkEmitter *ORE; | |||
120 | ||||
121 | /// If true, it is safe to use metadata during simplification. | |||
122 | InstrInfoQuery IIQ; | |||
123 | ||||
124 | Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, | |||
125 | const DominatorTree *DT, bool UseInstrInfo, | |||
126 | OptimizationRemarkEmitter *ORE = nullptr) | |||
127 | : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {} | |||
128 | }; | |||
129 | ||||
130 | } // end anonymous namespace | |||
131 | ||||
132 | // Given the provided Value and, potentially, a context instruction, return | |||
133 | // the preferred context instruction (if any). | |||
134 | static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { | |||
135 | // If we've been provided with a context instruction, then use that (provided | |||
136 | // it has been inserted). | |||
137 | if (CxtI && CxtI->getParent()) | |||
138 | return CxtI; | |||
139 | ||||
140 | // If the value is really an already-inserted instruction, then use that. | |||
141 | CxtI = dyn_cast<Instruction>(V); | |||
142 | if (CxtI && CxtI->getParent()) | |||
143 | return CxtI; | |||
144 | ||||
145 | return nullptr; | |||
146 | } | |||
147 | ||||
148 | static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) { | |||
149 | // If we've been provided with a context instruction, then use that (provided | |||
150 | // it has been inserted). | |||
151 | if (CxtI && CxtI->getParent()) | |||
152 | return CxtI; | |||
153 | ||||
154 | // If the value is really an already-inserted instruction, then use that. | |||
155 | CxtI = dyn_cast<Instruction>(V1); | |||
156 | if (CxtI && CxtI->getParent()) | |||
157 | return CxtI; | |||
158 | ||||
159 | CxtI = dyn_cast<Instruction>(V2); | |||
160 | if (CxtI && CxtI->getParent()) | |||
161 | return CxtI; | |||
162 | ||||
163 | return nullptr; | |||
164 | } | |||
165 | ||||
166 | static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, | |||
167 | const APInt &DemandedElts, | |||
168 | APInt &DemandedLHS, APInt &DemandedRHS) { | |||
169 | // The length of scalable vectors is unknown at compile time, thus we | |||
170 | // cannot check their values | |||
171 | if (isa<ScalableVectorType>(Shuf->getType())) | |||
172 | return false; | |||
173 | ||||
174 | int NumElts = | |||
175 | cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements(); | |||
176 | int NumMaskElts = cast<FixedVectorType>(Shuf->getType())->getNumElements(); | |||
177 | DemandedLHS = DemandedRHS = APInt::getZero(NumElts); | |||
178 | if (DemandedElts.isZero()) | |||
179 | return true; | |||
180 | // Simple case of a shuffle with zeroinitializer. | |||
181 | if (all_of(Shuf->getShuffleMask(), [](int Elt) { return Elt == 0; })) { | |||
182 | DemandedLHS.setBit(0); | |||
183 | return true; | |||
184 | } | |||
185 | for (int i = 0; i != NumMaskElts; ++i) { | |||
186 | if (!DemandedElts[i]) | |||
187 | continue; | |||
188 | int M = Shuf->getMaskValue(i); | |||
189 | assert(M < (NumElts * 2) && "Invalid shuffle mask constant")(static_cast <bool> (M < (NumElts * 2) && "Invalid shuffle mask constant" ) ? void (0) : __assert_fail ("M < (NumElts * 2) && \"Invalid shuffle mask constant\"" , "llvm/lib/Analysis/ValueTracking.cpp", 189, __extension__ __PRETTY_FUNCTION__ )); | |||
190 | ||||
191 | // For undef elements, we don't know anything about the common state of | |||
192 | // the shuffle result. | |||
193 | if (M == -1) | |||
194 | return false; | |||
195 | if (M < NumElts) | |||
196 | DemandedLHS.setBit(M % NumElts); | |||
197 | else | |||
198 | DemandedRHS.setBit(M % NumElts); | |||
199 | } | |||
200 | ||||
201 | return true; | |||
202 | } | |||
203 | ||||
204 | static void computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
205 | KnownBits &Known, unsigned Depth, const Query &Q); | |||
206 | ||||
207 | static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, | |||
208 | const Query &Q) { | |||
209 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | |||
210 | // vector | |||
211 | if (isa<ScalableVectorType>(V->getType())) { | |||
| ||||
212 | Known.resetAll(); | |||
213 | return; | |||
214 | } | |||
215 | ||||
216 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
217 | APInt DemandedElts = | |||
218 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
219 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
220 | } | |||
221 | ||||
222 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, | |||
223 | const DataLayout &DL, unsigned Depth, | |||
224 | AssumptionCache *AC, const Instruction *CxtI, | |||
225 | const DominatorTree *DT, | |||
226 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
227 | ::computeKnownBits(V, Known, Depth, | |||
228 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
229 | } | |||
230 | ||||
231 | void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
232 | KnownBits &Known, const DataLayout &DL, | |||
233 | unsigned Depth, AssumptionCache *AC, | |||
234 | const Instruction *CxtI, const DominatorTree *DT, | |||
235 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
236 | ::computeKnownBits(V, DemandedElts, Known, Depth, | |||
237 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
238 | } | |||
239 | ||||
240 | static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
241 | unsigned Depth, const Query &Q); | |||
242 | ||||
243 | static KnownBits computeKnownBits(const Value *V, unsigned Depth, | |||
244 | const Query &Q); | |||
245 | ||||
246 | KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, | |||
247 | unsigned Depth, AssumptionCache *AC, | |||
248 | const Instruction *CxtI, | |||
249 | const DominatorTree *DT, | |||
250 | OptimizationRemarkEmitter *ORE, | |||
251 | bool UseInstrInfo) { | |||
252 | return ::computeKnownBits( | |||
253 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
254 | } | |||
255 | ||||
256 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
257 | const DataLayout &DL, unsigned Depth, | |||
258 | AssumptionCache *AC, const Instruction *CxtI, | |||
259 | const DominatorTree *DT, | |||
260 | OptimizationRemarkEmitter *ORE, | |||
261 | bool UseInstrInfo) { | |||
262 | return ::computeKnownBits( | |||
263 | V, DemandedElts, Depth, | |||
264 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
265 | } | |||
266 | ||||
267 | bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS, | |||
268 | const DataLayout &DL, AssumptionCache *AC, | |||
269 | const Instruction *CxtI, const DominatorTree *DT, | |||
270 | bool UseInstrInfo) { | |||
271 | assert(LHS->getType() == RHS->getType() &&(static_cast <bool> (LHS->getType() == RHS->getType () && "LHS and RHS should have the same type") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 272, __extension__ __PRETTY_FUNCTION__ )) | |||
272 | "LHS and RHS should have the same type")(static_cast <bool> (LHS->getType() == RHS->getType () && "LHS and RHS should have the same type") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 272, __extension__ __PRETTY_FUNCTION__ )); | |||
273 | assert(LHS->getType()->isIntOrIntVectorTy() &&(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy () && "LHS and RHS should be integers") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\"" , "llvm/lib/Analysis/ValueTracking.cpp", 274, __extension__ __PRETTY_FUNCTION__ )) | |||
274 | "LHS and RHS should be integers")(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy () && "LHS and RHS should be integers") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\"" , "llvm/lib/Analysis/ValueTracking.cpp", 274, __extension__ __PRETTY_FUNCTION__ )); | |||
275 | // Look for an inverted mask: (X & ~M) op (Y & M). | |||
276 | { | |||
277 | Value *M; | |||
278 | if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | |||
279 | match(RHS, m_c_And(m_Specific(M), m_Value()))) | |||
280 | return true; | |||
281 | if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | |||
282 | match(LHS, m_c_And(m_Specific(M), m_Value()))) | |||
283 | return true; | |||
284 | } | |||
285 | // Look for: (A & B) op ~(A | B) | |||
286 | { | |||
287 | Value *A, *B; | |||
288 | if (match(LHS, m_And(m_Value(A), m_Value(B))) && | |||
289 | match(RHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) | |||
290 | return true; | |||
291 | if (match(RHS, m_And(m_Value(A), m_Value(B))) && | |||
292 | match(LHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) | |||
293 | return true; | |||
294 | } | |||
295 | IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); | |||
296 | KnownBits LHSKnown(IT->getBitWidth()); | |||
297 | KnownBits RHSKnown(IT->getBitWidth()); | |||
298 | computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
299 | computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
300 | return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown); | |||
301 | } | |||
302 | ||||
303 | bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *I) { | |||
304 | return !I->user_empty() && all_of(I->users(), [](const User *U) { | |||
305 | ICmpInst::Predicate P; | |||
306 | return match(U, m_ICmp(P, m_Value(), m_Zero())) && ICmpInst::isEquality(P); | |||
307 | }); | |||
308 | } | |||
309 | ||||
310 | static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | |||
311 | const Query &Q); | |||
312 | ||||
313 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, | |||
314 | bool OrZero, unsigned Depth, | |||
315 | AssumptionCache *AC, const Instruction *CxtI, | |||
316 | const DominatorTree *DT, bool UseInstrInfo) { | |||
317 | return ::isKnownToBeAPowerOfTwo( | |||
318 | V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
319 | } | |||
320 | ||||
321 | static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, | |||
322 | unsigned Depth, const Query &Q); | |||
323 | ||||
324 | static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q); | |||
325 | ||||
326 | bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth, | |||
327 | AssumptionCache *AC, const Instruction *CxtI, | |||
328 | const DominatorTree *DT, bool UseInstrInfo) { | |||
329 | return ::isKnownNonZero(V, Depth, | |||
330 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
331 | } | |||
332 | ||||
333 | bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL, | |||
334 | unsigned Depth, AssumptionCache *AC, | |||
335 | const Instruction *CxtI, const DominatorTree *DT, | |||
336 | bool UseInstrInfo) { | |||
337 | KnownBits Known = | |||
338 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
339 | return Known.isNonNegative(); | |||
340 | } | |||
341 | ||||
342 | bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth, | |||
343 | AssumptionCache *AC, const Instruction *CxtI, | |||
344 | const DominatorTree *DT, bool UseInstrInfo) { | |||
345 | if (auto *CI = dyn_cast<ConstantInt>(V)) | |||
346 | return CI->getValue().isStrictlyPositive(); | |||
347 | ||||
348 | // TODO: We'd doing two recursive queries here. We should factor this such | |||
349 | // that only a single query is needed. | |||
350 | return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) && | |||
351 | isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo); | |||
352 | } | |||
353 | ||||
354 | bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth, | |||
355 | AssumptionCache *AC, const Instruction *CxtI, | |||
356 | const DominatorTree *DT, bool UseInstrInfo) { | |||
357 | KnownBits Known = | |||
358 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
359 | return Known.isNegative(); | |||
360 | } | |||
361 | ||||
362 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | |||
363 | const Query &Q); | |||
364 | ||||
365 | bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, | |||
366 | const DataLayout &DL, AssumptionCache *AC, | |||
367 | const Instruction *CxtI, const DominatorTree *DT, | |||
368 | bool UseInstrInfo) { | |||
369 | return ::isKnownNonEqual(V1, V2, 0, | |||
370 | Query(DL, AC, safeCxtI(V2, V1, CxtI), DT, | |||
371 | UseInstrInfo, /*ORE=*/nullptr)); | |||
372 | } | |||
373 | ||||
374 | static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | |||
375 | const Query &Q); | |||
376 | ||||
377 | bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, | |||
378 | const DataLayout &DL, unsigned Depth, | |||
379 | AssumptionCache *AC, const Instruction *CxtI, | |||
380 | const DominatorTree *DT, bool UseInstrInfo) { | |||
381 | return ::MaskedValueIsZero( | |||
382 | V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
383 | } | |||
384 | ||||
385 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | |||
386 | unsigned Depth, const Query &Q); | |||
387 | ||||
388 | static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, | |||
389 | const Query &Q) { | |||
390 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | |||
391 | // vector | |||
392 | if (isa<ScalableVectorType>(V->getType())) | |||
393 | return 1; | |||
394 | ||||
395 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
396 | APInt DemandedElts = | |||
397 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
398 | return ComputeNumSignBits(V, DemandedElts, Depth, Q); | |||
399 | } | |||
400 | ||||
401 | unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, | |||
402 | unsigned Depth, AssumptionCache *AC, | |||
403 | const Instruction *CxtI, | |||
404 | const DominatorTree *DT, bool UseInstrInfo) { | |||
405 | return ::ComputeNumSignBits( | |||
406 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
407 | } | |||
408 | ||||
409 | unsigned llvm::ComputeMaxSignificantBits(const Value *V, const DataLayout &DL, | |||
410 | unsigned Depth, AssumptionCache *AC, | |||
411 | const Instruction *CxtI, | |||
412 | const DominatorTree *DT) { | |||
413 | unsigned SignBits = ComputeNumSignBits(V, DL, Depth, AC, CxtI, DT); | |||
414 | return V->getType()->getScalarSizeInBits() - SignBits + 1; | |||
415 | } | |||
416 | ||||
417 | static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, | |||
418 | bool NSW, const APInt &DemandedElts, | |||
419 | KnownBits &KnownOut, KnownBits &Known2, | |||
420 | unsigned Depth, const Query &Q) { | |||
421 | computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q); | |||
422 | ||||
423 | // If one operand is unknown and we have no nowrap information, | |||
424 | // the result will be unknown independently of the second operand. | |||
425 | if (KnownOut.isUnknown() && !NSW) | |||
426 | return; | |||
427 | ||||
428 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | |||
429 | KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut); | |||
430 | } | |||
431 | ||||
432 | static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, | |||
433 | const APInt &DemandedElts, KnownBits &Known, | |||
434 | KnownBits &Known2, unsigned Depth, | |||
435 | const Query &Q) { | |||
436 | computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q); | |||
437 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | |||
438 | ||||
439 | bool isKnownNegative = false; | |||
440 | bool isKnownNonNegative = false; | |||
441 | // If the multiplication is known not to overflow, compute the sign bit. | |||
442 | if (NSW) { | |||
443 | if (Op0 == Op1) { | |||
444 | // The product of a number with itself is non-negative. | |||
445 | isKnownNonNegative = true; | |||
446 | } else { | |||
447 | bool isKnownNonNegativeOp1 = Known.isNonNegative(); | |||
448 | bool isKnownNonNegativeOp0 = Known2.isNonNegative(); | |||
449 | bool isKnownNegativeOp1 = Known.isNegative(); | |||
450 | bool isKnownNegativeOp0 = Known2.isNegative(); | |||
451 | // The product of two numbers with the same sign is non-negative. | |||
452 | isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || | |||
453 | (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); | |||
454 | // The product of a negative number and a non-negative number is either | |||
455 | // negative or zero. | |||
456 | if (!isKnownNonNegative) | |||
457 | isKnownNegative = | |||
458 | (isKnownNegativeOp1 && isKnownNonNegativeOp0 && | |||
459 | Known2.isNonZero()) || | |||
460 | (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); | |||
461 | } | |||
462 | } | |||
463 | ||||
464 | bool SelfMultiply = Op0 == Op1; | |||
465 | // TODO: SelfMultiply can be poison, but not undef. | |||
466 | if (SelfMultiply) | |||
467 | SelfMultiply &= | |||
468 | isGuaranteedNotToBeUndefOrPoison(Op0, Q.AC, Q.CxtI, Q.DT, Depth + 1); | |||
469 | Known = KnownBits::mul(Known, Known2, SelfMultiply); | |||
470 | ||||
471 | // Only make use of no-wrap flags if we failed to compute the sign bit | |||
472 | // directly. This matters if the multiplication always overflows, in | |||
473 | // which case we prefer to follow the result of the direct computation, | |||
474 | // though as the program is invoking undefined behaviour we can choose | |||
475 | // whatever we like here. | |||
476 | if (isKnownNonNegative && !Known.isNegative()) | |||
477 | Known.makeNonNegative(); | |||
478 | else if (isKnownNegative && !Known.isNonNegative()) | |||
479 | Known.makeNegative(); | |||
480 | } | |||
481 | ||||
482 | void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, | |||
483 | KnownBits &Known) { | |||
484 | unsigned BitWidth = Known.getBitWidth(); | |||
485 | unsigned NumRanges = Ranges.getNumOperands() / 2; | |||
486 | assert(NumRanges >= 1)(static_cast <bool> (NumRanges >= 1) ? void (0) : __assert_fail ("NumRanges >= 1", "llvm/lib/Analysis/ValueTracking.cpp", 486, __extension__ __PRETTY_FUNCTION__)); | |||
487 | ||||
488 | Known.Zero.setAllBits(); | |||
489 | Known.One.setAllBits(); | |||
490 | ||||
491 | for (unsigned i = 0; i < NumRanges; ++i) { | |||
492 | ConstantInt *Lower = | |||
493 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); | |||
494 | ConstantInt *Upper = | |||
495 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); | |||
496 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | |||
497 | ||||
498 | // The first CommonPrefixBits of all values in Range are equal. | |||
499 | unsigned CommonPrefixBits = | |||
500 | (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros(); | |||
501 | APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); | |||
502 | APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth); | |||
503 | Known.One &= UnsignedMax & Mask; | |||
504 | Known.Zero &= ~UnsignedMax & Mask; | |||
505 | } | |||
506 | } | |||
507 | ||||
508 | static bool isEphemeralValueOf(const Instruction *I, const Value *E) { | |||
509 | SmallVector<const Value *, 16> WorkSet(1, I); | |||
510 | SmallPtrSet<const Value *, 32> Visited; | |||
511 | SmallPtrSet<const Value *, 16> EphValues; | |||
512 | ||||
513 | // The instruction defining an assumption's condition itself is always | |||
514 | // considered ephemeral to that assumption (even if it has other | |||
515 | // non-ephemeral users). See r246696's test case for an example. | |||
516 | if (is_contained(I->operands(), E)) | |||
517 | return true; | |||
518 | ||||
519 | while (!WorkSet.empty()) { | |||
520 | const Value *V = WorkSet.pop_back_val(); | |||
521 | if (!Visited.insert(V).second) | |||
522 | continue; | |||
523 | ||||
524 | // If all uses of this value are ephemeral, then so is this value. | |||
525 | if (llvm::all_of(V->users(), [&](const User *U) { | |||
526 | return EphValues.count(U); | |||
527 | })) { | |||
528 | if (V == E) | |||
529 | return true; | |||
530 | ||||
531 | if (V == I || (isa<Instruction>(V) && | |||
532 | !cast<Instruction>(V)->mayHaveSideEffects() && | |||
533 | !cast<Instruction>(V)->isTerminator())) { | |||
534 | EphValues.insert(V); | |||
535 | if (const User *U = dyn_cast<User>(V)) | |||
536 | append_range(WorkSet, U->operands()); | |||
537 | } | |||
538 | } | |||
539 | } | |||
540 | ||||
541 | return false; | |||
542 | } | |||
543 | ||||
544 | // Is this an intrinsic that cannot be speculated but also cannot trap? | |||
545 | bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { | |||
546 | if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I)) | |||
547 | return CI->isAssumeLikeIntrinsic(); | |||
548 | ||||
549 | return false; | |||
550 | } | |||
551 | ||||
552 | bool llvm::isValidAssumeForContext(const Instruction *Inv, | |||
553 | const Instruction *CxtI, | |||
554 | const DominatorTree *DT) { | |||
555 | // There are two restrictions on the use of an assume: | |||
556 | // 1. The assume must dominate the context (or the control flow must | |||
557 | // reach the assume whenever it reaches the context). | |||
558 | // 2. The context must not be in the assume's set of ephemeral values | |||
559 | // (otherwise we will use the assume to prove that the condition | |||
560 | // feeding the assume is trivially true, thus causing the removal of | |||
561 | // the assume). | |||
562 | ||||
563 | if (Inv->getParent() == CxtI->getParent()) { | |||
564 | // If Inv and CtxI are in the same block, check if the assume (Inv) is first | |||
565 | // in the BB. | |||
566 | if (Inv->comesBefore(CxtI)) | |||
567 | return true; | |||
568 | ||||
569 | // Don't let an assume affect itself - this would cause the problems | |||
570 | // `isEphemeralValueOf` is trying to prevent, and it would also make | |||
571 | // the loop below go out of bounds. | |||
572 | if (Inv == CxtI) | |||
573 | return false; | |||
574 | ||||
575 | // The context comes first, but they're both in the same block. | |||
576 | // Make sure there is nothing in between that might interrupt | |||
577 | // the control flow, not even CxtI itself. | |||
578 | // We limit the scan distance between the assume and its context instruction | |||
579 | // to avoid a compile-time explosion. This limit is chosen arbitrarily, so | |||
580 | // it can be adjusted if needed (could be turned into a cl::opt). | |||
581 | auto Range = make_range(CxtI->getIterator(), Inv->getIterator()); | |||
582 | if (!isGuaranteedToTransferExecutionToSuccessor(Range, 15)) | |||
583 | return false; | |||
584 | ||||
585 | return !isEphemeralValueOf(Inv, CxtI); | |||
586 | } | |||
587 | ||||
588 | // Inv and CxtI are in different blocks. | |||
589 | if (DT) { | |||
590 | if (DT->dominates(Inv, CxtI)) | |||
591 | return true; | |||
592 | } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { | |||
593 | // We don't have a DT, but this trivially dominates. | |||
594 | return true; | |||
595 | } | |||
596 | ||||
597 | return false; | |||
598 | } | |||
599 | ||||
600 | static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { | |||
601 | // v u> y implies v != 0. | |||
602 | if (Pred == ICmpInst::ICMP_UGT) | |||
603 | return true; | |||
604 | ||||
605 | // Special-case v != 0 to also handle v != null. | |||
606 | if (Pred == ICmpInst::ICMP_NE) | |||
607 | return match(RHS, m_Zero()); | |||
608 | ||||
609 | // All other predicates - rely on generic ConstantRange handling. | |||
610 | const APInt *C; | |||
611 | if (!match(RHS, m_APInt(C))) | |||
612 | return false; | |||
613 | ||||
614 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C); | |||
615 | return !TrueValues.contains(APInt::getZero(C->getBitWidth())); | |||
616 | } | |||
617 | ||||
618 | static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) { | |||
619 | // Use of assumptions is context-sensitive. If we don't have a context, we | |||
620 | // cannot use them! | |||
621 | if (!Q.AC || !Q.CxtI) | |||
622 | return false; | |||
623 | ||||
624 | if (Q.CxtI && V->getType()->isPointerTy()) { | |||
625 | SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull}; | |||
626 | if (!NullPointerIsDefined(Q.CxtI->getFunction(), | |||
627 | V->getType()->getPointerAddressSpace())) | |||
628 | AttrKinds.push_back(Attribute::Dereferenceable); | |||
629 | ||||
630 | if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC)) | |||
631 | return true; | |||
632 | } | |||
633 | ||||
634 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | |||
635 | if (!AssumeVH) | |||
636 | continue; | |||
637 | CallInst *I = cast<CallInst>(AssumeVH); | |||
638 | assert(I->getFunction() == Q.CxtI->getFunction() &&(static_cast <bool> (I->getFunction() == Q.CxtI-> getFunction() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 639, __extension__ __PRETTY_FUNCTION__ )) | |||
639 | "Got assumption for the wrong function!")(static_cast <bool> (I->getFunction() == Q.CxtI-> getFunction() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 639, __extension__ __PRETTY_FUNCTION__ )); | |||
640 | ||||
641 | // Warning: This loop can end up being somewhat performance sensitive. | |||
642 | // We're running this loop for once for each value queried resulting in a | |||
643 | // runtime of ~O(#assumes * #values). | |||
644 | ||||
645 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 646, __extension__ __PRETTY_FUNCTION__ )) | |||
646 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 646, __extension__ __PRETTY_FUNCTION__ )); | |||
647 | ||||
648 | Value *RHS; | |||
649 | CmpInst::Predicate Pred; | |||
650 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | |||
651 | if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS)))) | |||
652 | return false; | |||
653 | ||||
654 | if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT)) | |||
655 | return true; | |||
656 | } | |||
657 | ||||
658 | return false; | |||
659 | } | |||
660 | ||||
661 | static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known, | |||
662 | unsigned Depth, const Query &Q) { | |||
663 | // Use of assumptions is context-sensitive. If we don't have a context, we | |||
664 | // cannot use them! | |||
665 | if (!Q.AC || !Q.CxtI) | |||
666 | return; | |||
667 | ||||
668 | unsigned BitWidth = Known.getBitWidth(); | |||
669 | ||||
670 | // Refine Known set if the pointer alignment is set by assume bundles. | |||
671 | if (V->getType()->isPointerTy()) { | |||
672 | if (RetainedKnowledge RK = getKnowledgeValidInContext( | |||
673 | V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) { | |||
674 | if (isPowerOf2_64(RK.ArgValue)) | |||
675 | Known.Zero.setLowBits(Log2_64(RK.ArgValue)); | |||
676 | } | |||
677 | } | |||
678 | ||||
679 | // Note that the patterns below need to be kept in sync with the code | |||
680 | // in AssumptionCache::updateAffectedValues. | |||
681 | ||||
682 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | |||
683 | if (!AssumeVH) | |||
684 | continue; | |||
685 | CallInst *I = cast<CallInst>(AssumeVH); | |||
686 | assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&(static_cast <bool> (I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 687, __extension__ __PRETTY_FUNCTION__ )) | |||
687 | "Got assumption for the wrong function!")(static_cast <bool> (I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 687, __extension__ __PRETTY_FUNCTION__ )); | |||
688 | ||||
689 | // Warning: This loop can end up being somewhat performance sensitive. | |||
690 | // We're running this loop for once for each value queried resulting in a | |||
691 | // runtime of ~O(#assumes * #values). | |||
692 | ||||
693 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 694, __extension__ __PRETTY_FUNCTION__ )) | |||
694 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 694, __extension__ __PRETTY_FUNCTION__ )); | |||
695 | ||||
696 | Value *Arg = I->getArgOperand(0); | |||
697 | ||||
698 | if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
699 | assert(BitWidth == 1 && "assume operand is not i1?")(static_cast <bool> (BitWidth == 1 && "assume operand is not i1?" ) ? void (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 699, __extension__ __PRETTY_FUNCTION__ )); | |||
700 | Known.setAllOnes(); | |||
701 | return; | |||
702 | } | |||
703 | if (match(Arg, m_Not(m_Specific(V))) && | |||
704 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
705 | assert(BitWidth == 1 && "assume operand is not i1?")(static_cast <bool> (BitWidth == 1 && "assume operand is not i1?" ) ? void (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 705, __extension__ __PRETTY_FUNCTION__ )); | |||
706 | Known.setAllZero(); | |||
707 | return; | |||
708 | } | |||
709 | ||||
710 | // The remaining tests are all recursive, so bail out if we hit the limit. | |||
711 | if (Depth == MaxAnalysisRecursionDepth) | |||
712 | continue; | |||
713 | ||||
714 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | |||
715 | if (!Cmp) | |||
716 | continue; | |||
717 | ||||
718 | // We are attempting to compute known bits for the operands of an assume. | |||
719 | // Do not try to use other assumptions for those recursive calls because | |||
720 | // that can lead to mutual recursion and a compile-time explosion. | |||
721 | // An example of the mutual recursion: computeKnownBits can call | |||
722 | // isKnownNonZero which calls computeKnownBitsFromAssume (this function) | |||
723 | // and so on. | |||
724 | Query QueryNoAC = Q; | |||
725 | QueryNoAC.AC = nullptr; | |||
726 | ||||
727 | // Note that ptrtoint may change the bitwidth. | |||
728 | Value *A, *B; | |||
729 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | |||
730 | ||||
731 | CmpInst::Predicate Pred; | |||
732 | uint64_t C; | |||
733 | switch (Cmp->getPredicate()) { | |||
734 | default: | |||
735 | break; | |||
736 | case ICmpInst::ICMP_EQ: | |||
737 | // assume(v = a) | |||
738 | if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) && | |||
739 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
740 | KnownBits RHSKnown = | |||
741 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
742 | Known.Zero |= RHSKnown.Zero; | |||
743 | Known.One |= RHSKnown.One; | |||
744 | // assume(v & b = a) | |||
745 | } else if (match(Cmp, | |||
746 | m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && | |||
747 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
748 | KnownBits RHSKnown = | |||
749 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
750 | KnownBits MaskKnown = | |||
751 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
752 | ||||
753 | // For those bits in the mask that are known to be one, we can propagate | |||
754 | // known bits from the RHS to V. | |||
755 | Known.Zero |= RHSKnown.Zero & MaskKnown.One; | |||
756 | Known.One |= RHSKnown.One & MaskKnown.One; | |||
757 | // assume(~(v & b) = a) | |||
758 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), | |||
759 | m_Value(A))) && | |||
760 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
761 | KnownBits RHSKnown = | |||
762 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
763 | KnownBits MaskKnown = | |||
764 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
765 | ||||
766 | // For those bits in the mask that are known to be one, we can propagate | |||
767 | // inverted known bits from the RHS to V. | |||
768 | Known.Zero |= RHSKnown.One & MaskKnown.One; | |||
769 | Known.One |= RHSKnown.Zero & MaskKnown.One; | |||
770 | // assume(v | b = a) | |||
771 | } else if (match(Cmp, | |||
772 | m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && | |||
773 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
774 | KnownBits RHSKnown = | |||
775 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
776 | KnownBits BKnown = | |||
777 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
778 | ||||
779 | // For those bits in B that are known to be zero, we can propagate known | |||
780 | // bits from the RHS to V. | |||
781 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | |||
782 | Known.One |= RHSKnown.One & BKnown.Zero; | |||
783 | // assume(~(v | b) = a) | |||
784 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), | |||
785 | m_Value(A))) && | |||
786 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
787 | KnownBits RHSKnown = | |||
788 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
789 | KnownBits BKnown = | |||
790 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
791 | ||||
792 | // For those bits in B that are known to be zero, we can propagate | |||
793 | // inverted known bits from the RHS to V. | |||
794 | Known.Zero |= RHSKnown.One & BKnown.Zero; | |||
795 | Known.One |= RHSKnown.Zero & BKnown.Zero; | |||
796 | // assume(v ^ b = a) | |||
797 | } else if (match(Cmp, | |||
798 | m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && | |||
799 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
800 | KnownBits RHSKnown = | |||
801 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
802 | KnownBits BKnown = | |||
803 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
804 | ||||
805 | // For those bits in B that are known to be zero, we can propagate known | |||
806 | // bits from the RHS to V. For those bits in B that are known to be one, | |||
807 | // we can propagate inverted known bits from the RHS to V. | |||
808 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | |||
809 | Known.One |= RHSKnown.One & BKnown.Zero; | |||
810 | Known.Zero |= RHSKnown.One & BKnown.One; | |||
811 | Known.One |= RHSKnown.Zero & BKnown.One; | |||
812 | // assume(~(v ^ b) = a) | |||
813 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), | |||
814 | m_Value(A))) && | |||
815 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
816 | KnownBits RHSKnown = | |||
817 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
818 | KnownBits BKnown = | |||
819 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
820 | ||||
821 | // For those bits in B that are known to be zero, we can propagate | |||
822 | // inverted known bits from the RHS to V. For those bits in B that are | |||
823 | // known to be one, we can propagate known bits from the RHS to V. | |||
824 | Known.Zero |= RHSKnown.One & BKnown.Zero; | |||
825 | Known.One |= RHSKnown.Zero & BKnown.Zero; | |||
826 | Known.Zero |= RHSKnown.Zero & BKnown.One; | |||
827 | Known.One |= RHSKnown.One & BKnown.One; | |||
828 | // assume(v << c = a) | |||
829 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), | |||
830 | m_Value(A))) && | |||
831 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | |||
832 | KnownBits RHSKnown = | |||
833 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
834 | ||||
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 | RHSKnown.Zero.lshrInPlace(C); | |||
838 | Known.Zero |= RHSKnown.Zero; | |||
839 | RHSKnown.One.lshrInPlace(C); | |||
840 | Known.One |= RHSKnown.One; | |||
841 | // assume(~(v << c) = a) | |||
842 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), | |||
843 | m_Value(A))) && | |||
844 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | |||
845 | KnownBits RHSKnown = | |||
846 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
847 | // For those bits in RHS that are known, we can propagate them inverted | |||
848 | // to known bits in V shifted to the right by C. | |||
849 | RHSKnown.One.lshrInPlace(C); | |||
850 | Known.Zero |= RHSKnown.One; | |||
851 | RHSKnown.Zero.lshrInPlace(C); | |||
852 | Known.One |= RHSKnown.Zero; | |||
853 | // assume(v >> c = a) | |||
854 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)), | |||
855 | m_Value(A))) && | |||
856 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | |||
857 | KnownBits RHSKnown = | |||
858 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
859 | // For those bits in RHS that are known, we can propagate them to known | |||
860 | // bits in V shifted to the right by C. | |||
861 | Known.Zero |= RHSKnown.Zero << C; | |||
862 | Known.One |= RHSKnown.One << C; | |||
863 | // assume(~(v >> c) = a) | |||
864 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))), | |||
865 | m_Value(A))) && | |||
866 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | |||
867 | KnownBits RHSKnown = | |||
868 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
869 | // For those bits in RHS that are known, we can propagate them inverted | |||
870 | // to known bits in V shifted to the right by C. | |||
871 | Known.Zero |= RHSKnown.One << C; | |||
872 | Known.One |= RHSKnown.Zero << C; | |||
873 | } | |||
874 | break; | |||
875 | case ICmpInst::ICMP_SGE: | |||
876 | // assume(v >=_s c) where c is non-negative | |||
877 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | |||
878 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
879 | KnownBits RHSKnown = | |||
880 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
881 | ||||
882 | if (RHSKnown.isNonNegative()) { | |||
883 | // We know that the sign bit is zero. | |||
884 | Known.makeNonNegative(); | |||
885 | } | |||
886 | } | |||
887 | break; | |||
888 | case ICmpInst::ICMP_SGT: | |||
889 | // assume(v >_s c) where c is at least -1. | |||
890 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | |||
891 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
892 | KnownBits RHSKnown = | |||
893 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
894 | ||||
895 | if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) { | |||
896 | // We know that the sign bit is zero. | |||
897 | Known.makeNonNegative(); | |||
898 | } | |||
899 | } | |||
900 | break; | |||
901 | case ICmpInst::ICMP_SLE: | |||
902 | // assume(v <=_s c) where c is negative | |||
903 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | |||
904 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
905 | KnownBits RHSKnown = | |||
906 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
907 | ||||
908 | if (RHSKnown.isNegative()) { | |||
909 | // We know that the sign bit is one. | |||
910 | Known.makeNegative(); | |||
911 | } | |||
912 | } | |||
913 | break; | |||
914 | case ICmpInst::ICMP_SLT: | |||
915 | // assume(v <_s c) where c is non-positive | |||
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 (RHSKnown.isZero() || RHSKnown.isNegative()) { | |||
922 | // We know that the sign bit is one. | |||
923 | Known.makeNegative(); | |||
924 | } | |||
925 | } | |||
926 | break; | |||
927 | case ICmpInst::ICMP_ULE: | |||
928 | // assume(v <=_u c) | |||
929 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | |||
930 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
931 | KnownBits RHSKnown = | |||
932 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
933 | ||||
934 | // Whatever high bits in c are zero are known to be zero. | |||
935 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | |||
936 | } | |||
937 | break; | |||
938 | case ICmpInst::ICMP_ULT: | |||
939 | // assume(v <_u c) | |||
940 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | |||
941 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
942 | KnownBits RHSKnown = | |||
943 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
944 | ||||
945 | // If the RHS is known zero, then this assumption must be wrong (nothing | |||
946 | // is unsigned less than zero). Signal a conflict and get out of here. | |||
947 | if (RHSKnown.isZero()) { | |||
948 | Known.Zero.setAllBits(); | |||
949 | Known.One.setAllBits(); | |||
950 | break; | |||
951 | } | |||
952 | ||||
953 | // Whatever high bits in c are zero are known to be zero (if c is a power | |||
954 | // of 2, then one more). | |||
955 | if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC)) | |||
956 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1); | |||
957 | else | |||
958 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | |||
959 | } | |||
960 | break; | |||
961 | } | |||
962 | } | |||
963 | ||||
964 | // If assumptions conflict with each other or previous known bits, then we | |||
965 | // have a logical fallacy. It's possible that the assumption is not reachable, | |||
966 | // so this isn't a real bug. On the other hand, the program may have undefined | |||
967 | // behavior, or we might have a bug in the compiler. We can't assert/crash, so | |||
968 | // clear out the known bits, try to warn the user, and hope for the best. | |||
969 | if (Known.Zero.intersects(Known.One)) { | |||
970 | Known.resetAll(); | |||
971 | ||||
972 | if (Q.ORE) | |||
973 | Q.ORE->emit([&]() { | |||
974 | auto *CxtI = const_cast<Instruction *>(Q.CxtI); | |||
975 | return OptimizationRemarkAnalysis("value-tracking", "BadAssumption", | |||
976 | CxtI) | |||
977 | << "Detected conflicting code assumptions. Program may " | |||
978 | "have undefined behavior, or compiler may have " | |||
979 | "internal error."; | |||
980 | }); | |||
981 | } | |||
982 | } | |||
983 | ||||
984 | /// Compute known bits from a shift operator, including those with a | |||
985 | /// non-constant shift amount. Known is the output of this function. Known2 is a | |||
986 | /// pre-allocated temporary with the same bit width as Known and on return | |||
987 | /// contains the known bit of the shift value source. KF is an | |||
988 | /// operator-specific function that, given the known-bits and a shift amount, | |||
989 | /// compute the implied known-bits of the shift operator's result respectively | |||
990 | /// for that shift amount. The results from calling KF are conservatively | |||
991 | /// combined for all permitted shift amounts. | |||
992 | static void computeKnownBitsFromShiftOperator( | |||
993 | const Operator *I, const APInt &DemandedElts, KnownBits &Known, | |||
994 | KnownBits &Known2, unsigned Depth, const Query &Q, | |||
995 | function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) { | |||
996 | unsigned BitWidth = Known.getBitWidth(); | |||
997 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
998 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
999 | ||||
1000 | // Note: We cannot use Known.Zero.getLimitedValue() here, because if | |||
1001 | // BitWidth > 64 and any upper bits are known, we'll end up returning the | |||
1002 | // limit value (which implies all bits are known). | |||
1003 | uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue(); | |||
1004 | uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue(); | |||
1005 | bool ShiftAmtIsConstant = Known.isConstant(); | |||
1006 | bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth); | |||
1007 | ||||
1008 | if (ShiftAmtIsConstant) { | |||
1009 | Known = KF(Known2, Known); | |||
1010 | ||||
1011 | // If the known bits conflict, this must be an overflowing left shift, so | |||
1012 | // the shift result is poison. We can return anything we want. Choose 0 for | |||
1013 | // the best folding opportunity. | |||
1014 | if (Known.hasConflict()) | |||
1015 | Known.setAllZero(); | |||
1016 | ||||
1017 | return; | |||
1018 | } | |||
1019 | ||||
1020 | // If the shift amount could be greater than or equal to the bit-width of the | |||
1021 | // LHS, the value could be poison, but bail out because the check below is | |||
1022 | // expensive. | |||
1023 | // TODO: Should we just carry on? | |||
1024 | if (MaxShiftAmtIsOutOfRange) { | |||
1025 | Known.resetAll(); | |||
1026 | return; | |||
1027 | } | |||
1028 | ||||
1029 | // It would be more-clearly correct to use the two temporaries for this | |||
1030 | // calculation. Reusing the APInts here to prevent unnecessary allocations. | |||
1031 | Known.resetAll(); | |||
1032 | ||||
1033 | // If we know the shifter operand is nonzero, we can sometimes infer more | |||
1034 | // known bits. However this is expensive to compute, so be lazy about it and | |||
1035 | // only compute it when absolutely necessary. | |||
1036 | Optional<bool> ShifterOperandIsNonZero; | |||
1037 | ||||
1038 | // Early exit if we can't constrain any well-defined shift amount. | |||
1039 | if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) && | |||
1040 | !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) { | |||
1041 | ShifterOperandIsNonZero = | |||
1042 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | |||
1043 | if (!*ShifterOperandIsNonZero) | |||
1044 | return; | |||
1045 | } | |||
1046 | ||||
1047 | Known.Zero.setAllBits(); | |||
1048 | Known.One.setAllBits(); | |||
1049 | for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { | |||
1050 | // Combine the shifted known input bits only for those shift amounts | |||
1051 | // compatible with its known constraints. | |||
1052 | if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) | |||
1053 | continue; | |||
1054 | if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) | |||
1055 | continue; | |||
1056 | // If we know the shifter is nonzero, we may be able to infer more known | |||
1057 | // bits. This check is sunk down as far as possible to avoid the expensive | |||
1058 | // call to isKnownNonZero if the cheaper checks above fail. | |||
1059 | if (ShiftAmt == 0) { | |||
1060 | if (!ShifterOperandIsNonZero.hasValue()) | |||
1061 | ShifterOperandIsNonZero = | |||
1062 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | |||
1063 | if (*ShifterOperandIsNonZero) | |||
1064 | continue; | |||
1065 | } | |||
1066 | ||||
1067 | Known = KnownBits::commonBits( | |||
1068 | Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt)))); | |||
1069 | } | |||
1070 | ||||
1071 | // If the known bits conflict, the result is poison. Return a 0 and hope the | |||
1072 | // caller can further optimize that. | |||
1073 | if (Known.hasConflict()) | |||
1074 | Known.setAllZero(); | |||
1075 | } | |||
1076 | ||||
1077 | static void computeKnownBitsFromOperator(const Operator *I, | |||
1078 | const APInt &DemandedElts, | |||
1079 | KnownBits &Known, unsigned Depth, | |||
1080 | const Query &Q) { | |||
1081 | unsigned BitWidth = Known.getBitWidth(); | |||
1082 | ||||
1083 | KnownBits Known2(BitWidth); | |||
1084 | switch (I->getOpcode()) { | |||
| ||||
1085 | default: break; | |||
1086 | case Instruction::Load: | |||
1087 | if (MDNode *MD = | |||
1088 | Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range)) | |||
1089 | computeKnownBitsFromRangeMetadata(*MD, Known); | |||
1090 | break; | |||
1091 | case Instruction::And: { | |||
1092 | // If either the LHS or the RHS are Zero, the result is zero. | |||
1093 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
1094 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
1095 | ||||
1096 | Known &= Known2; | |||
1097 | ||||
1098 | // and(x, add (x, -1)) is a common idiom that always clears the low bit; | |||
1099 | // here we handle the more general case of adding any odd number by | |||
1100 | // matching the form add(x, add(x, y)) where y is odd. | |||
1101 | // TODO: This could be generalized to clearing any bit set in y where the | |||
1102 | // following bit is known to be unset in y. | |||
1103 | Value *X = nullptr, *Y = nullptr; | |||
1104 | if (!Known.Zero[0] && !Known.One[0] && | |||
1105 | match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) { | |||
1106 | Known2.resetAll(); | |||
1107 | computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q); | |||
1108 | if (Known2.countMinTrailingOnes() > 0) | |||
1109 | Known.Zero.setBit(0); | |||
1110 | } | |||
1111 | break; | |||
1112 | } | |||
1113 | case Instruction::Or: | |||
1114 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
1115 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
1116 | ||||
1117 | Known |= Known2; | |||
1118 | break; | |||
1119 | case Instruction::Xor: | |||
1120 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
1121 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
1122 | ||||
1123 | Known ^= Known2; | |||
1124 | break; | |||
1125 | case Instruction::Mul: { | |||
1126 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
1127 | computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts, | |||
1128 | Known, Known2, Depth, Q); | |||
1129 | break; | |||
1130 | } | |||
1131 | case Instruction::UDiv: { | |||
1132 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1133 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1134 | Known = KnownBits::udiv(Known, Known2); | |||
1135 | break; | |||
1136 | } | |||
1137 | case Instruction::Select: { | |||
1138 | const Value *LHS = nullptr, *RHS = nullptr; | |||
1139 | SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; | |||
1140 | if (SelectPatternResult::isMinOrMax(SPF)) { | |||
1141 | computeKnownBits(RHS, Known, Depth + 1, Q); | |||
1142 | computeKnownBits(LHS, Known2, Depth + 1, Q); | |||
1143 | switch (SPF) { | |||
1144 | default: | |||
1145 | llvm_unreachable("Unhandled select pattern flavor!")::llvm::llvm_unreachable_internal("Unhandled select pattern flavor!" , "llvm/lib/Analysis/ValueTracking.cpp", 1145); | |||
1146 | case SPF_SMAX: | |||
1147 | Known = KnownBits::smax(Known, Known2); | |||
1148 | break; | |||
1149 | case SPF_SMIN: | |||
1150 | Known = KnownBits::smin(Known, Known2); | |||
1151 | break; | |||
1152 | case SPF_UMAX: | |||
1153 | Known = KnownBits::umax(Known, Known2); | |||
1154 | break; | |||
1155 | case SPF_UMIN: | |||
1156 | Known = KnownBits::umin(Known, Known2); | |||
1157 | break; | |||
1158 | } | |||
1159 | break; | |||
1160 | } | |||
1161 | ||||
1162 | computeKnownBits(I->getOperand(2), Known, Depth + 1, Q); | |||
1163 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1164 | ||||
1165 | // Only known if known in both the LHS and RHS. | |||
1166 | Known = KnownBits::commonBits(Known, Known2); | |||
1167 | ||||
1168 | if (SPF == SPF_ABS) { | |||
1169 | // RHS from matchSelectPattern returns the negation part of abs pattern. | |||
1170 | // If the negate has an NSW flag we can assume the sign bit of the result | |||
1171 | // will be 0 because that makes abs(INT_MIN) undefined. | |||
1172 | if (match(RHS, m_Neg(m_Specific(LHS))) && | |||
1173 | Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RHS))) | |||
1174 | Known.Zero.setSignBit(); | |||
1175 | } | |||
1176 | ||||
1177 | break; | |||
1178 | } | |||
1179 | case Instruction::FPTrunc: | |||
1180 | case Instruction::FPExt: | |||
1181 | case Instruction::FPToUI: | |||
1182 | case Instruction::FPToSI: | |||
1183 | case Instruction::SIToFP: | |||
1184 | case Instruction::UIToFP: | |||
1185 | break; // Can't work with floating point. | |||
1186 | case Instruction::PtrToInt: | |||
1187 | case Instruction::IntToPtr: | |||
1188 | // Fall through and handle them the same as zext/trunc. | |||
1189 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
1190 | case Instruction::ZExt: | |||
1191 | case Instruction::Trunc: { | |||
1192 | Type *SrcTy = I->getOperand(0)->getType(); | |||
1193 | ||||
1194 | unsigned SrcBitWidth; | |||
1195 | // Note that we handle pointer operands here because of inttoptr/ptrtoint | |||
1196 | // which fall through here. | |||
1197 | Type *ScalarTy = SrcTy->getScalarType(); | |||
1198 | SrcBitWidth = ScalarTy->isPointerTy() ? | |||
1199 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | |||
1200 | Q.DL.getTypeSizeInBits(ScalarTy); | |||
1201 | ||||
1202 | assert(SrcBitWidth && "SrcBitWidth can't be zero")(static_cast <bool> (SrcBitWidth && "SrcBitWidth can't be zero" ) ? void (0) : __assert_fail ("SrcBitWidth && \"SrcBitWidth can't be zero\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1202, __extension__ __PRETTY_FUNCTION__ )); | |||
1203 | Known = Known.anyextOrTrunc(SrcBitWidth); | |||
1204 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1205 | Known = Known.zextOrTrunc(BitWidth); | |||
1206 | break; | |||
1207 | } | |||
1208 | case Instruction::BitCast: { | |||
1209 | Type *SrcTy = I->getOperand(0)->getType(); | |||
1210 | if (SrcTy->isIntOrPtrTy() && | |||
1211 | // TODO: For now, not handling conversions like: | |||
1212 | // (bitcast i64 %x to <2 x i32>) | |||
1213 | !I->getType()->isVectorTy()) { | |||
1214 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1215 | break; | |||
1216 | } | |||
1217 | ||||
1218 | // Handle cast from vector integer type to scalar or vector integer. | |||
1219 | auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy); | |||
1220 | if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || | |||
1221 | !I->getType()->isIntOrIntVectorTy()) | |||
1222 | break; | |||
1223 | ||||
1224 | // Look through a cast from narrow vector elements to wider type. | |||
1225 | // Examples: v4i32 -> v2i64, v3i8 -> v24 | |||
1226 | unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); | |||
1227 | if (BitWidth % SubBitWidth == 0) { | |||
1228 | // Known bits are automatically intersected across demanded elements of a | |||
1229 | // vector. So for example, if a bit is computed as known zero, it must be | |||
1230 | // zero across all demanded elements of the vector. | |||
1231 | // | |||
1232 | // For this bitcast, each demanded element of the output is sub-divided | |||
1233 | // across a set of smaller vector elements in the source vector. To get | |||
1234 | // the known bits for an entire element of the output, compute the known | |||
1235 | // bits for each sub-element sequentially. This is done by shifting the | |||
1236 | // one-set-bit demanded elements parameter across the sub-elements for | |||
1237 | // consecutive calls to computeKnownBits. We are using the demanded | |||
1238 | // elements parameter as a mask operator. | |||
1239 | // | |||
1240 | // The known bits of each sub-element are then inserted into place | |||
1241 | // (dependent on endian) to form the full result of known bits. | |||
1242 | unsigned NumElts = DemandedElts.getBitWidth(); | |||
1243 | unsigned SubScale = BitWidth / SubBitWidth; | |||
1244 | APInt SubDemandedElts = APInt::getZero(NumElts * SubScale); | |||
1245 | for (unsigned i = 0; i != NumElts; ++i) { | |||
1246 | if (DemandedElts[i]) | |||
1247 | SubDemandedElts.setBit(i * SubScale); | |||
1248 | } | |||
1249 | ||||
1250 | KnownBits KnownSrc(SubBitWidth); | |||
1251 | for (unsigned i = 0; i != SubScale; ++i) { | |||
1252 | computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc, | |||
1253 | Depth + 1, Q); | |||
1254 | unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i; | |||
1255 | Known.insertBits(KnownSrc, ShiftElt * SubBitWidth); | |||
1256 | } | |||
1257 | } | |||
1258 | break; | |||
1259 | } | |||
1260 | case Instruction::SExt: { | |||
1261 | // Compute the bits in the result that are not present in the input. | |||
1262 | unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); | |||
1263 | ||||
1264 | Known = Known.trunc(SrcBitWidth); | |||
1265 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1266 | // If the sign bit of the input is known set or clear, then we know the | |||
1267 | // top bits of the result. | |||
1268 | Known = Known.sext(BitWidth); | |||
1269 | break; | |||
1270 | } | |||
1271 | case Instruction::Shl: { | |||
1272 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
1273 | auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
1274 | KnownBits Result = KnownBits::shl(KnownVal, KnownAmt); | |||
1275 | // If this shift has "nsw" keyword, then the result is either a poison | |||
1276 | // value or has the same sign bit as the first operand. | |||
1277 | if (NSW) { | |||
1278 | if (KnownVal.Zero.isSignBitSet()) | |||
1279 | Result.Zero.setSignBit(); | |||
1280 | if (KnownVal.One.isSignBitSet()) | |||
1281 | Result.One.setSignBit(); | |||
1282 | } | |||
1283 | return Result; | |||
1284 | }; | |||
1285 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
1286 | KF); | |||
1287 | // Trailing zeros of a right-shifted constant never decrease. | |||
1288 | const APInt *C; | |||
1289 | if (match(I->getOperand(0), m_APInt(C))) | |||
1290 | Known.Zero.setLowBits(C->countTrailingZeros()); | |||
1291 | break; | |||
1292 | } | |||
1293 | case Instruction::LShr: { | |||
1294 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
1295 | return KnownBits::lshr(KnownVal, KnownAmt); | |||
1296 | }; | |||
1297 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
1298 | KF); | |||
1299 | // Leading zeros of a left-shifted constant never decrease. | |||
1300 | const APInt *C; | |||
1301 | if (match(I->getOperand(0), m_APInt(C))) | |||
1302 | Known.Zero.setHighBits(C->countLeadingZeros()); | |||
1303 | break; | |||
1304 | } | |||
1305 | case Instruction::AShr: { | |||
1306 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
1307 | return KnownBits::ashr(KnownVal, KnownAmt); | |||
1308 | }; | |||
1309 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
1310 | KF); | |||
1311 | break; | |||
1312 | } | |||
1313 | case Instruction::Sub: { | |||
1314 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
1315 | computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, | |||
1316 | DemandedElts, Known, Known2, Depth, Q); | |||
1317 | break; | |||
1318 | } | |||
1319 | case Instruction::Add: { | |||
1320 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
1321 | computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, | |||
1322 | DemandedElts, Known, Known2, Depth, Q); | |||
1323 | break; | |||
1324 | } | |||
1325 | case Instruction::SRem: | |||
1326 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1327 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1328 | Known = KnownBits::srem(Known, Known2); | |||
1329 | break; | |||
1330 | ||||
1331 | case Instruction::URem: | |||
1332 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1333 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1334 | Known = KnownBits::urem(Known, Known2); | |||
1335 | break; | |||
1336 | case Instruction::Alloca: | |||
1337 | Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign())); | |||
1338 | break; | |||
1339 | case Instruction::GetElementPtr: { | |||
1340 | // Analyze all of the subscripts of this getelementptr instruction | |||
1341 | // to determine if we can prove known low zero bits. | |||
1342 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1343 | // Accumulate the constant indices in a separate variable | |||
1344 | // to minimize the number of calls to computeForAddSub. | |||
1345 | APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); | |||
1346 | ||||
1347 | gep_type_iterator GTI = gep_type_begin(I); | |||
1348 | for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { | |||
1349 | // TrailZ can only become smaller, short-circuit if we hit zero. | |||
1350 | if (Known.isUnknown()) | |||
1351 | break; | |||
1352 | ||||
1353 | Value *Index = I->getOperand(i); | |||
1354 | ||||
1355 | // Handle case when index is zero. | |||
1356 | Constant *CIndex = dyn_cast<Constant>(Index); | |||
1357 | if (CIndex && CIndex->isZeroValue()) | |||
1358 | continue; | |||
1359 | ||||
1360 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
1361 | // Handle struct member offset arithmetic. | |||
1362 | ||||
1363 | assert(CIndex &&(static_cast <bool> (CIndex && "Access to structure field must be known at compile time" ) ? void (0) : __assert_fail ("CIndex && \"Access to structure field must be known at compile time\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1364, __extension__ __PRETTY_FUNCTION__ )) | |||
1364 | "Access to structure field must be known at compile time")(static_cast <bool> (CIndex && "Access to structure field must be known at compile time" ) ? void (0) : __assert_fail ("CIndex && \"Access to structure field must be known at compile time\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1364, __extension__ __PRETTY_FUNCTION__ )); | |||
1365 | ||||
1366 | if (CIndex->getType()->isVectorTy()) | |||
1367 | Index = CIndex->getSplatValue(); | |||
1368 | ||||
1369 | unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); | |||
1370 | const StructLayout *SL = Q.DL.getStructLayout(STy); | |||
1371 | uint64_t Offset = SL->getElementOffset(Idx); | |||
1372 | AccConstIndices += Offset; | |||
1373 | continue; | |||
1374 | } | |||
1375 | ||||
1376 | // Handle array index arithmetic. | |||
1377 | Type *IndexedTy = GTI.getIndexedType(); | |||
1378 | if (!IndexedTy->isSized()) { | |||
1379 | Known.resetAll(); | |||
1380 | break; | |||
1381 | } | |||
1382 | ||||
1383 | unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); | |||
1384 | KnownBits IndexBits(IndexBitWidth); | |||
1385 | computeKnownBits(Index, IndexBits, Depth + 1, Q); | |||
1386 | TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy); | |||
1387 | uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinSize(); | |||
1388 | KnownBits ScalingFactor(IndexBitWidth); | |||
1389 | // Multiply by current sizeof type. | |||
1390 | // &A[i] == A + i * sizeof(*A[i]). | |||
1391 | if (IndexTypeSize.isScalable()) { | |||
1392 | // For scalable types the only thing we know about sizeof is | |||
1393 | // that this is a multiple of the minimum size. | |||
1394 | ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes)); | |||
1395 | } else if (IndexBits.isConstant()) { | |||
1396 | APInt IndexConst = IndexBits.getConstant(); | |||
1397 | APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); | |||
1398 | IndexConst *= ScalingFactor; | |||
1399 | AccConstIndices += IndexConst.sextOrTrunc(BitWidth); | |||
1400 | continue; | |||
1401 | } else { | |||
1402 | ScalingFactor = | |||
1403 | KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes)); | |||
1404 | } | |||
1405 | IndexBits = KnownBits::mul(IndexBits, ScalingFactor); | |||
1406 | ||||
1407 | // If the offsets have a different width from the pointer, according | |||
1408 | // to the language reference we need to sign-extend or truncate them | |||
1409 | // to the width of the pointer. | |||
1410 | IndexBits = IndexBits.sextOrTrunc(BitWidth); | |||
1411 | ||||
1412 | // Note that inbounds does *not* guarantee nsw for the addition, as only | |||
1413 | // the offset is signed, while the base address is unsigned. | |||
1414 | Known = KnownBits::computeForAddSub( | |||
1415 | /*Add=*/true, /*NSW=*/false, Known, IndexBits); | |||
1416 | } | |||
1417 | if (!Known.isUnknown() && !AccConstIndices.isZero()) { | |||
1418 | KnownBits Index = KnownBits::makeConstant(AccConstIndices); | |||
1419 | Known = KnownBits::computeForAddSub( | |||
1420 | /*Add=*/true, /*NSW=*/false, Known, Index); | |||
1421 | } | |||
1422 | break; | |||
1423 | } | |||
1424 | case Instruction::PHI: { | |||
1425 | const PHINode *P = cast<PHINode>(I); | |||
1426 | BinaryOperator *BO = nullptr; | |||
1427 | Value *R = nullptr, *L = nullptr; | |||
1428 | if (matchSimpleRecurrence(P, BO, R, L)) { | |||
1429 | // Handle the case of a simple two-predecessor recurrence PHI. | |||
1430 | // There's a lot more that could theoretically be done here, but | |||
1431 | // this is sufficient to catch some interesting cases. | |||
1432 | unsigned Opcode = BO->getOpcode(); | |||
1433 | ||||
1434 | // If this is a shift recurrence, we know the bits being shifted in. | |||
1435 | // We can combine that with information about the start value of the | |||
1436 | // recurrence to conclude facts about the result. | |||
1437 | if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr || | |||
1438 | Opcode == Instruction::Shl) && | |||
1439 | BO->getOperand(0) == I) { | |||
1440 | ||||
1441 | // We have matched a recurrence of the form: | |||
1442 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
1443 | // %iv.next = shift_op %iv, L | |||
1444 | ||||
1445 | // Recurse with the phi context to avoid concern about whether facts | |||
1446 | // inferred hold at original context instruction. TODO: It may be | |||
1447 | // correct to use the original context. IF warranted, explore and | |||
1448 | // add sufficient tests to cover. | |||
1449 | Query RecQ = Q; | |||
1450 | RecQ.CxtI = P; | |||
1451 | computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ); | |||
1452 | switch (Opcode) { | |||
1453 | case Instruction::Shl: | |||
1454 | // A shl recurrence will only increase the tailing zeros | |||
1455 | Known.Zero.setLowBits(Known2.countMinTrailingZeros()); | |||
1456 | break; | |||
1457 | case Instruction::LShr: | |||
1458 | // A lshr recurrence will preserve the leading zeros of the | |||
1459 | // start value | |||
1460 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | |||
1461 | break; | |||
1462 | case Instruction::AShr: | |||
1463 | // An ashr recurrence will extend the initial sign bit | |||
1464 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | |||
1465 | Known.One.setHighBits(Known2.countMinLeadingOnes()); | |||
1466 | break; | |||
1467 | }; | |||
1468 | } | |||
1469 | ||||
1470 | // Check for operations that have the property that if | |||
1471 | // both their operands have low zero bits, the result | |||
1472 | // will have low zero bits. | |||
1473 | if (Opcode == Instruction::Add || | |||
1474 | Opcode == Instruction::Sub || | |||
1475 | Opcode == Instruction::And || | |||
1476 | Opcode == Instruction::Or || | |||
1477 | Opcode == Instruction::Mul) { | |||
1478 | // Change the context instruction to the "edge" that flows into the | |||
1479 | // phi. This is important because that is where the value is actually | |||
1480 | // "evaluated" even though it is used later somewhere else. (see also | |||
1481 | // D69571). | |||
1482 | Query RecQ = Q; | |||
1483 | ||||
1484 | unsigned OpNum = P->getOperand(0) == R ? 0 : 1; | |||
1485 | Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator(); | |||
1486 | Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator(); | |||
1487 | ||||
1488 | // Ok, we have a PHI of the form L op= R. Check for low | |||
1489 | // zero bits. | |||
1490 | RecQ.CxtI = RInst; | |||
1491 | computeKnownBits(R, Known2, Depth + 1, RecQ); | |||
1492 | ||||
1493 | // We need to take the minimum number of known bits | |||
1494 | KnownBits Known3(BitWidth); | |||
1495 | RecQ.CxtI = LInst; | |||
1496 | computeKnownBits(L, Known3, Depth + 1, RecQ); | |||
1497 | ||||
1498 | Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(), | |||
1499 | Known3.countMinTrailingZeros())); | |||
1500 | ||||
1501 | auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO); | |||
1502 | if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) { | |||
1503 | // If initial value of recurrence is nonnegative, and we are adding | |||
1504 | // a nonnegative number with nsw, the result can only be nonnegative | |||
1505 | // or poison value regardless of the number of times we execute the | |||
1506 | // add in phi recurrence. If initial value is negative and we are | |||
1507 | // adding a negative number with nsw, the result can only be | |||
1508 | // negative or poison value. Similar arguments apply to sub and mul. | |||
1509 | // | |||
1510 | // (add non-negative, non-negative) --> non-negative | |||
1511 | // (add negative, negative) --> negative | |||
1512 | if (Opcode == Instruction::Add) { | |||
1513 | if (Known2.isNonNegative() && Known3.isNonNegative()) | |||
1514 | Known.makeNonNegative(); | |||
1515 | else if (Known2.isNegative() && Known3.isNegative()) | |||
1516 | Known.makeNegative(); | |||
1517 | } | |||
1518 | ||||
1519 | // (sub nsw non-negative, negative) --> non-negative | |||
1520 | // (sub nsw negative, non-negative) --> negative | |||
1521 | else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) { | |||
1522 | if (Known2.isNonNegative() && Known3.isNegative()) | |||
1523 | Known.makeNonNegative(); | |||
1524 | else if (Known2.isNegative() && Known3.isNonNegative()) | |||
1525 | Known.makeNegative(); | |||
1526 | } | |||
1527 | ||||
1528 | // (mul nsw non-negative, non-negative) --> non-negative | |||
1529 | else if (Opcode == Instruction::Mul && Known2.isNonNegative() && | |||
1530 | Known3.isNonNegative()) | |||
1531 | Known.makeNonNegative(); | |||
1532 | } | |||
1533 | ||||
1534 | break; | |||
1535 | } | |||
1536 | } | |||
1537 | ||||
1538 | // Unreachable blocks may have zero-operand PHI nodes. | |||
1539 | if (P->getNumIncomingValues() == 0) | |||
1540 | break; | |||
1541 | ||||
1542 | // Otherwise take the unions of the known bit sets of the operands, | |||
1543 | // taking conservative care to avoid excessive recursion. | |||
1544 | if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) { | |||
1545 | // Skip if every incoming value references to ourself. | |||
1546 | if (isa_and_nonnull<UndefValue>(P->hasConstantValue())) | |||
1547 | break; | |||
1548 | ||||
1549 | Known.Zero.setAllBits(); | |||
1550 | Known.One.setAllBits(); | |||
1551 | for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { | |||
1552 | Value *IncValue = P->getIncomingValue(u); | |||
1553 | // Skip direct self references. | |||
1554 | if (IncValue == P) continue; | |||
1555 | ||||
1556 | // Change the context instruction to the "edge" that flows into the | |||
1557 | // phi. This is important because that is where the value is actually | |||
1558 | // "evaluated" even though it is used later somewhere else. (see also | |||
1559 | // D69571). | |||
1560 | Query RecQ = Q; | |||
1561 | RecQ.CxtI = P->getIncomingBlock(u)->getTerminator(); | |||
1562 | ||||
1563 | Known2 = KnownBits(BitWidth); | |||
1564 | // Recurse, but cap the recursion to one level, because we don't | |||
1565 | // want to waste time spinning around in loops. | |||
1566 | computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ); | |||
1567 | Known = KnownBits::commonBits(Known, Known2); | |||
1568 | // If all bits have been ruled out, there's no need to check | |||
1569 | // more operands. | |||
1570 | if (Known.isUnknown()) | |||
1571 | break; | |||
1572 | } | |||
1573 | } | |||
1574 | break; | |||
1575 | } | |||
1576 | case Instruction::Call: | |||
1577 | case Instruction::Invoke: | |||
1578 | // If range metadata is attached to this call, set known bits from that, | |||
1579 | // and then intersect with known bits based on other properties of the | |||
1580 | // function. | |||
1581 | if (MDNode *MD = | |||
1582 | Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range)) | |||
1583 | computeKnownBitsFromRangeMetadata(*MD, Known); | |||
1584 | if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) { | |||
1585 | computeKnownBits(RV, Known2, Depth + 1, Q); | |||
1586 | Known.Zero |= Known2.Zero; | |||
1587 | Known.One |= Known2.One; | |||
1588 | } | |||
1589 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { | |||
1590 | switch (II->getIntrinsicID()) { | |||
1591 | default: break; | |||
1592 | case Intrinsic::abs: { | |||
1593 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
1594 | bool IntMinIsPoison = match(II->getArgOperand(1), m_One()); | |||
1595 | Known = Known2.abs(IntMinIsPoison); | |||
1596 | break; | |||
1597 | } | |||
1598 | case Intrinsic::bitreverse: | |||
1599 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
1600 | Known.Zero |= Known2.Zero.reverseBits(); | |||
1601 | Known.One |= Known2.One.reverseBits(); | |||
1602 | break; | |||
1603 | case Intrinsic::bswap: | |||
1604 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
1605 | Known.Zero |= Known2.Zero.byteSwap(); | |||
1606 | Known.One |= Known2.One.byteSwap(); | |||
1607 | break; | |||
1608 | case Intrinsic::ctlz: { | |||
1609 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
1610 | // If we have a known 1, its position is our upper bound. | |||
1611 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); | |||
1612 | // If this call is poison for 0 input, the result will be less than 2^n. | |||
1613 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | |||
1614 | PossibleLZ = std::min(PossibleLZ, BitWidth - 1); | |||
1615 | unsigned LowBits = Log2_32(PossibleLZ)+1; | |||
1616 | Known.Zero.setBitsFrom(LowBits); | |||
1617 | break; | |||
1618 | } | |||
1619 | case Intrinsic::cttz: { | |||
1620 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
1621 | // If we have a known 1, its position is our upper bound. | |||
1622 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); | |||
1623 | // If this call is poison for 0 input, the result will be less than 2^n. | |||
1624 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | |||
1625 | PossibleTZ = std::min(PossibleTZ, BitWidth - 1); | |||
1626 | unsigned LowBits = Log2_32(PossibleTZ)+1; | |||
1627 | Known.Zero.setBitsFrom(LowBits); | |||
1628 | break; | |||
1629 | } | |||
1630 | case Intrinsic::ctpop: { | |||
1631 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
1632 | // We can bound the space the count needs. Also, bits known to be zero | |||
1633 | // can't contribute to the population. | |||
1634 | unsigned BitsPossiblySet = Known2.countMaxPopulation(); | |||
1635 | unsigned LowBits = Log2_32(BitsPossiblySet)+1; | |||
1636 | Known.Zero.setBitsFrom(LowBits); | |||
1637 | // TODO: we could bound KnownOne using the lower bound on the number | |||
1638 | // of bits which might be set provided by popcnt KnownOne2. | |||
1639 | break; | |||
1640 | } | |||
1641 | case Intrinsic::fshr: | |||
1642 | case Intrinsic::fshl: { | |||
1643 | const APInt *SA; | |||
1644 | if (!match(I->getOperand(2), m_APInt(SA))) | |||
1645 | break; | |||
1646 | ||||
1647 | // Normalize to funnel shift left. | |||
1648 | uint64_t ShiftAmt = SA->urem(BitWidth); | |||
1649 | if (II->getIntrinsicID() == Intrinsic::fshr) | |||
1650 | ShiftAmt = BitWidth - ShiftAmt; | |||
1651 | ||||
1652 | KnownBits Known3(BitWidth); | |||
1653 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
1654 | computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q); | |||
1655 | ||||
1656 | Known.Zero = | |||
1657 | Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt); | |||
1658 | Known.One = | |||
1659 | Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt); | |||
1660 | break; | |||
1661 | } | |||
1662 | case Intrinsic::uadd_sat: | |||
1663 | case Intrinsic::usub_sat: { | |||
1664 | bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat; | |||
1665 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1666 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1667 | ||||
1668 | // Add: Leading ones of either operand are preserved. | |||
1669 | // Sub: Leading zeros of LHS and leading ones of RHS are preserved | |||
1670 | // as leading zeros in the result. | |||
1671 | unsigned LeadingKnown; | |||
1672 | if (IsAdd) | |||
1673 | LeadingKnown = std::max(Known.countMinLeadingOnes(), | |||
1674 | Known2.countMinLeadingOnes()); | |||
1675 | else | |||
1676 | LeadingKnown = std::max(Known.countMinLeadingZeros(), | |||
1677 | Known2.countMinLeadingOnes()); | |||
1678 | ||||
1679 | Known = KnownBits::computeForAddSub( | |||
1680 | IsAdd, /* NSW */ false, Known, Known2); | |||
1681 | ||||
1682 | // We select between the operation result and all-ones/zero | |||
1683 | // respectively, so we can preserve known ones/zeros. | |||
1684 | if (IsAdd) { | |||
1685 | Known.One.setHighBits(LeadingKnown); | |||
1686 | Known.Zero.clearAllBits(); | |||
1687 | } else { | |||
1688 | Known.Zero.setHighBits(LeadingKnown); | |||
1689 | Known.One.clearAllBits(); | |||
1690 | } | |||
1691 | break; | |||
1692 | } | |||
1693 | case Intrinsic::umin: | |||
1694 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1695 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1696 | Known = KnownBits::umin(Known, Known2); | |||
1697 | break; | |||
1698 | case Intrinsic::umax: | |||
1699 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1700 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1701 | Known = KnownBits::umax(Known, Known2); | |||
1702 | break; | |||
1703 | case Intrinsic::smin: | |||
1704 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1705 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1706 | Known = KnownBits::smin(Known, Known2); | |||
1707 | break; | |||
1708 | case Intrinsic::smax: | |||
1709 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1710 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
1711 | Known = KnownBits::smax(Known, Known2); | |||
1712 | break; | |||
1713 | case Intrinsic::x86_sse42_crc32_64_64: | |||
1714 | Known.Zero.setBitsFrom(32); | |||
1715 | break; | |||
1716 | case Intrinsic::riscv_vsetvli: | |||
1717 | case Intrinsic::riscv_vsetvlimax: | |||
1718 | // Assume that VL output is positive and would fit in an int32_t. | |||
1719 | // TODO: VLEN might be capped at 16 bits in a future V spec update. | |||
1720 | if (BitWidth >= 32) | |||
1721 | Known.Zero.setBitsFrom(31); | |||
1722 | break; | |||
1723 | case Intrinsic::vscale: { | |||
1724 | if (!II->getParent() || !II->getFunction() || | |||
1725 | !II->getFunction()->hasFnAttribute(Attribute::VScaleRange)) | |||
1726 | break; | |||
1727 | ||||
1728 | auto Attr = II->getFunction()->getFnAttribute(Attribute::VScaleRange); | |||
1729 | Optional<unsigned> VScaleMax = Attr.getVScaleRangeMax(); | |||
1730 | ||||
1731 | if (!VScaleMax) | |||
1732 | break; | |||
1733 | ||||
1734 | unsigned VScaleMin = Attr.getVScaleRangeMin(); | |||
1735 | ||||
1736 | // If vscale min = max then we know the exact value at compile time | |||
1737 | // and hence we know the exact bits. | |||
1738 | if (VScaleMin == VScaleMax) { | |||
1739 | Known.One = VScaleMin; | |||
1740 | Known.Zero = VScaleMin; | |||
1741 | Known.Zero.flipAllBits(); | |||
1742 | break; | |||
1743 | } | |||
1744 | ||||
1745 | unsigned FirstZeroHighBit = | |||
1746 | 32 - countLeadingZeros(VScaleMax.getValue()); | |||
1747 | if (FirstZeroHighBit < BitWidth) | |||
1748 | Known.Zero.setBitsFrom(FirstZeroHighBit); | |||
1749 | ||||
1750 | break; | |||
1751 | } | |||
1752 | } | |||
1753 | } | |||
1754 | break; | |||
1755 | case Instruction::ShuffleVector: { | |||
1756 | auto *Shuf = dyn_cast<ShuffleVectorInst>(I); | |||
1757 | // FIXME: Do we need to handle ConstantExpr involving shufflevectors? | |||
1758 | if (!Shuf) { | |||
1759 | Known.resetAll(); | |||
1760 | return; | |||
1761 | } | |||
1762 | // For undef elements, we don't know anything about the common state of | |||
1763 | // the shuffle result. | |||
1764 | APInt DemandedLHS, DemandedRHS; | |||
1765 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { | |||
1766 | Known.resetAll(); | |||
1767 | return; | |||
1768 | } | |||
1769 | Known.One.setAllBits(); | |||
1770 | Known.Zero.setAllBits(); | |||
1771 | if (!!DemandedLHS) { | |||
1772 | const Value *LHS = Shuf->getOperand(0); | |||
1773 | computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q); | |||
1774 | // If we don't know any bits, early out. | |||
1775 | if (Known.isUnknown()) | |||
1776 | break; | |||
1777 | } | |||
1778 | if (!!DemandedRHS) { | |||
1779 | const Value *RHS = Shuf->getOperand(1); | |||
1780 | computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q); | |||
1781 | Known = KnownBits::commonBits(Known, Known2); | |||
1782 | } | |||
1783 | break; | |||
1784 | } | |||
1785 | case Instruction::InsertElement: { | |||
1786 | const Value *Vec = I->getOperand(0); | |||
1787 | const Value *Elt = I->getOperand(1); | |||
1788 | auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2)); | |||
1789 | // Early out if the index is non-constant or out-of-range. | |||
1790 | unsigned NumElts = DemandedElts.getBitWidth(); | |||
1791 | if (!CIdx || CIdx->getValue().uge(NumElts)) { | |||
1792 | Known.resetAll(); | |||
1793 | return; | |||
1794 | } | |||
1795 | Known.One.setAllBits(); | |||
1796 | Known.Zero.setAllBits(); | |||
1797 | unsigned EltIdx = CIdx->getZExtValue(); | |||
1798 | // Do we demand the inserted element? | |||
1799 | if (DemandedElts[EltIdx]) { | |||
1800 | computeKnownBits(Elt, Known, Depth + 1, Q); | |||
1801 | // If we don't know any bits, early out. | |||
1802 | if (Known.isUnknown()) | |||
1803 | break; | |||
1804 | } | |||
1805 | // We don't need the base vector element that has been inserted. | |||
1806 | APInt DemandedVecElts = DemandedElts; | |||
1807 | DemandedVecElts.clearBit(EltIdx); | |||
1808 | if (!!DemandedVecElts) { | |||
1809 | computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q); | |||
1810 | Known = KnownBits::commonBits(Known, Known2); | |||
1811 | } | |||
1812 | break; | |||
1813 | } | |||
1814 | case Instruction::ExtractElement: { | |||
1815 | // Look through extract element. If the index is non-constant or | |||
1816 | // out-of-range demand all elements, otherwise just the extracted element. | |||
1817 | const Value *Vec = I->getOperand(0); | |||
1818 | const Value *Idx = I->getOperand(1); | |||
1819 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | |||
1820 | if (isa<ScalableVectorType>(Vec->getType())) { | |||
1821 | // FIXME: there's probably *something* we can do with scalable vectors | |||
1822 | Known.resetAll(); | |||
1823 | break; | |||
1824 | } | |||
1825 | unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
1826 | APInt DemandedVecElts = APInt::getAllOnes(NumElts); | |||
1827 | if (CIdx && CIdx->getValue().ult(NumElts)) | |||
1828 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | |||
1829 | computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q); | |||
1830 | break; | |||
1831 | } | |||
1832 | case Instruction::ExtractValue: | |||
1833 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { | |||
1834 | const ExtractValueInst *EVI = cast<ExtractValueInst>(I); | |||
1835 | if (EVI->getNumIndices() != 1) break; | |||
1836 | if (EVI->getIndices()[0] == 0) { | |||
1837 | switch (II->getIntrinsicID()) { | |||
1838 | default: break; | |||
1839 | case Intrinsic::uadd_with_overflow: | |||
1840 | case Intrinsic::sadd_with_overflow: | |||
1841 | computeKnownBitsAddSub(true, II->getArgOperand(0), | |||
1842 | II->getArgOperand(1), false, DemandedElts, | |||
1843 | Known, Known2, Depth, Q); | |||
1844 | break; | |||
1845 | case Intrinsic::usub_with_overflow: | |||
1846 | case Intrinsic::ssub_with_overflow: | |||
1847 | computeKnownBitsAddSub(false, II->getArgOperand(0), | |||
1848 | II->getArgOperand(1), false, DemandedElts, | |||
1849 | Known, Known2, Depth, Q); | |||
1850 | break; | |||
1851 | case Intrinsic::umul_with_overflow: | |||
1852 | case Intrinsic::smul_with_overflow: | |||
1853 | computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, | |||
1854 | DemandedElts, Known, Known2, Depth, Q); | |||
1855 | break; | |||
1856 | } | |||
1857 | } | |||
1858 | } | |||
1859 | break; | |||
1860 | case Instruction::Freeze: | |||
1861 | if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, | |||
1862 | Depth + 1)) | |||
1863 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
1864 | break; | |||
1865 | } | |||
1866 | } | |||
1867 | ||||
1868 | /// Determine which bits of V are known to be either zero or one and return | |||
1869 | /// them. | |||
1870 | KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
1871 | unsigned Depth, const Query &Q) { | |||
1872 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | |||
1873 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
1874 | return Known; | |||
1875 | } | |||
1876 | ||||
1877 | /// Determine which bits of V are known to be either zero or one and return | |||
1878 | /// them. | |||
1879 | KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) { | |||
1880 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | |||
1881 | computeKnownBits(V, Known, Depth, Q); | |||
1882 | return Known; | |||
1883 | } | |||
1884 | ||||
1885 | /// Determine which bits of V are known to be either zero or one and return | |||
1886 | /// them in the Known bit set. | |||
1887 | /// | |||
1888 | /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that | |||
1889 | /// we cannot optimize based on the assumption that it is zero without changing | |||
1890 | /// it to be an explicit zero. If we don't change it to zero, other code could | |||
1891 | /// optimized based on the contradictory assumption that it is non-zero. | |||
1892 | /// Because instcombine aggressively folds operations with undef args anyway, | |||
1893 | /// this won't lose us code quality. | |||
1894 | /// | |||
1895 | /// This function is defined on values with integer type, values with pointer | |||
1896 | /// type, and vectors of integers. In the case | |||
1897 | /// where V is a vector, known zero, and known one values are the | |||
1898 | /// same width as the vector element, and the bit is set only if it is true | |||
1899 | /// for all of the demanded elements in the vector specified by DemandedElts. | |||
1900 | void computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
1901 | KnownBits &Known, unsigned Depth, const Query &Q) { | |||
1902 | if (!DemandedElts || isa<ScalableVectorType>(V->getType())) { | |||
1903 | // No demanded elts or V is a scalable vector, better to assume we don't | |||
1904 | // know anything. | |||
1905 | Known.resetAll(); | |||
1906 | return; | |||
1907 | } | |||
1908 | ||||
1909 | assert(V && "No Value?")(static_cast <bool> (V && "No Value?") ? void ( 0) : __assert_fail ("V && \"No Value?\"", "llvm/lib/Analysis/ValueTracking.cpp" , 1909, __extension__ __PRETTY_FUNCTION__)); | |||
1910 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1910, __extension__ __PRETTY_FUNCTION__ )); | |||
1911 | ||||
1912 | #ifndef NDEBUG | |||
1913 | Type *Ty = V->getType(); | |||
1914 | unsigned BitWidth = Known.getBitWidth(); | |||
1915 | ||||
1916 | assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) &&(static_cast <bool> ((Ty->isIntOrIntVectorTy(BitWidth ) || Ty->isPtrOrPtrVectorTy()) && "Not integer or pointer type!" ) ? void (0) : __assert_fail ("(Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1917, __extension__ __PRETTY_FUNCTION__ )) | |||
1917 | "Not integer or pointer type!")(static_cast <bool> ((Ty->isIntOrIntVectorTy(BitWidth ) || Ty->isPtrOrPtrVectorTy()) && "Not integer or pointer type!" ) ? void (0) : __assert_fail ("(Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1917, __extension__ __PRETTY_FUNCTION__ )); | |||
1918 | ||||
1919 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | |||
1920 | assert((static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1922, __extension__ __PRETTY_FUNCTION__ )) | |||
1921 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1922, __extension__ __PRETTY_FUNCTION__ )) | |||
1922 | "DemandedElt width should equal the fixed vector number of elements")(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1922, __extension__ __PRETTY_FUNCTION__ )); | |||
1923 | } else { | |||
1924 | assert(DemandedElts == APInt(1, 1) &&(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1925, __extension__ __PRETTY_FUNCTION__ )) | |||
1925 | "DemandedElt width should be 1 for scalars")(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1925, __extension__ __PRETTY_FUNCTION__ )); | |||
1926 | } | |||
1927 | ||||
1928 | Type *ScalarTy = Ty->getScalarType(); | |||
1929 | if (ScalarTy->isPointerTy()) { | |||
1930 | assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) &&(static_cast <bool> (BitWidth == Q.DL.getPointerTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1931, __extension__ __PRETTY_FUNCTION__ )) | |||
1931 | "V and Known should have same BitWidth")(static_cast <bool> (BitWidth == Q.DL.getPointerTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1931, __extension__ __PRETTY_FUNCTION__ )); | |||
1932 | } else { | |||
1933 | assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) &&(static_cast <bool> (BitWidth == Q.DL.getTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1934, __extension__ __PRETTY_FUNCTION__ )) | |||
1934 | "V and Known should have same BitWidth")(static_cast <bool> (BitWidth == Q.DL.getTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1934, __extension__ __PRETTY_FUNCTION__ )); | |||
1935 | } | |||
1936 | #endif | |||
1937 | ||||
1938 | const APInt *C; | |||
1939 | if (match(V, m_APInt(C))) { | |||
1940 | // We know all of the bits for a scalar constant or a splat vector constant! | |||
1941 | Known = KnownBits::makeConstant(*C); | |||
1942 | return; | |||
1943 | } | |||
1944 | // Null and aggregate-zero are all-zeros. | |||
1945 | if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { | |||
1946 | Known.setAllZero(); | |||
1947 | return; | |||
1948 | } | |||
1949 | // Handle a constant vector by taking the intersection of the known bits of | |||
1950 | // each element. | |||
1951 | if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) { | |||
1952 | // We know that CDV must be a vector of integers. Take the intersection of | |||
1953 | // each element. | |||
1954 | Known.Zero.setAllBits(); Known.One.setAllBits(); | |||
1955 | for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { | |||
1956 | if (!DemandedElts[i]) | |||
1957 | continue; | |||
1958 | APInt Elt = CDV->getElementAsAPInt(i); | |||
1959 | Known.Zero &= ~Elt; | |||
1960 | Known.One &= Elt; | |||
1961 | } | |||
1962 | return; | |||
1963 | } | |||
1964 | ||||
1965 | if (const auto *CV = dyn_cast<ConstantVector>(V)) { | |||
1966 | // We know that CV must be a vector of integers. Take the intersection of | |||
1967 | // each element. | |||
1968 | Known.Zero.setAllBits(); Known.One.setAllBits(); | |||
1969 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { | |||
1970 | if (!DemandedElts[i]) | |||
1971 | continue; | |||
1972 | Constant *Element = CV->getAggregateElement(i); | |||
1973 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); | |||
1974 | if (!ElementCI) { | |||
1975 | Known.resetAll(); | |||
1976 | return; | |||
1977 | } | |||
1978 | const APInt &Elt = ElementCI->getValue(); | |||
1979 | Known.Zero &= ~Elt; | |||
1980 | Known.One &= Elt; | |||
1981 | } | |||
1982 | return; | |||
1983 | } | |||
1984 | ||||
1985 | // Start out not knowing anything. | |||
1986 | Known.resetAll(); | |||
1987 | ||||
1988 | // We can't imply anything about undefs. | |||
1989 | if (isa<UndefValue>(V)) | |||
1990 | return; | |||
1991 | ||||
1992 | // There's no point in looking through other users of ConstantData for | |||
1993 | // assumptions. Confirm that we've handled them all. | |||
1994 | assert(!isa<ConstantData>(V) && "Unhandled constant data!")(static_cast <bool> (!isa<ConstantData>(V) && "Unhandled constant data!") ? void (0) : __assert_fail ("!isa<ConstantData>(V) && \"Unhandled constant data!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1994, __extension__ __PRETTY_FUNCTION__ )); | |||
1995 | ||||
1996 | // All recursive calls that increase depth must come after this. | |||
1997 | if (Depth == MaxAnalysisRecursionDepth) | |||
1998 | return; | |||
1999 | ||||
2000 | // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has | |||
2001 | // the bits of its aliasee. | |||
2002 | if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { | |||
2003 | if (!GA->isInterposable()) | |||
2004 | computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q); | |||
2005 | return; | |||
2006 | } | |||
2007 | ||||
2008 | if (const Operator *I = dyn_cast<Operator>(V)) | |||
2009 | computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); | |||
2010 | ||||
2011 | // Aligned pointers have trailing zeros - refine Known.Zero set | |||
2012 | if (isa<PointerType>(V->getType())) { | |||
2013 | Align Alignment = V->getPointerAlignment(Q.DL); | |||
2014 | Known.Zero.setLowBits(Log2(Alignment)); | |||
2015 | } | |||
2016 | ||||
2017 | // computeKnownBitsFromAssume strictly refines Known. | |||
2018 | // Therefore, we run them after computeKnownBitsFromOperator. | |||
2019 | ||||
2020 | // Check whether a nearby assume intrinsic can determine some known bits. | |||
2021 | computeKnownBitsFromAssume(V, Known, Depth, Q); | |||
2022 | ||||
2023 | assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?")(static_cast <bool> ((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?") ? void (0) : __assert_fail ("(Known.Zero & Known.One) == 0 && \"Bits known to be one AND zero?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2023, __extension__ __PRETTY_FUNCTION__ )); | |||
2024 | } | |||
2025 | ||||
2026 | /// Return true if the given value is known to have exactly one | |||
2027 | /// bit set when defined. For vectors return true if every element is known to | |||
2028 | /// be a power of two when defined. Supports values with integer or pointer | |||
2029 | /// types and vectors of integers. | |||
2030 | bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | |||
2031 | const Query &Q) { | |||
2032 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2032, __extension__ __PRETTY_FUNCTION__ )); | |||
2033 | ||||
2034 | // Attempt to match against constants. | |||
2035 | if (OrZero && match(V, m_Power2OrZero())) | |||
2036 | return true; | |||
2037 | if (match(V, m_Power2())) | |||
2038 | return true; | |||
2039 | ||||
2040 | // 1 << X is clearly a power of two if the one is not shifted off the end. If | |||
2041 | // it is shifted off the end then the result is undefined. | |||
2042 | if (match(V, m_Shl(m_One(), m_Value()))) | |||
2043 | return true; | |||
2044 | ||||
2045 | // (signmask) >>l X is clearly a power of two if the one is not shifted off | |||
2046 | // the bottom. If it is shifted off the bottom then the result is undefined. | |||
2047 | if (match(V, m_LShr(m_SignMask(), m_Value()))) | |||
2048 | return true; | |||
2049 | ||||
2050 | // The remaining tests are all recursive, so bail out if we hit the limit. | |||
2051 | if (Depth++ == MaxAnalysisRecursionDepth) | |||
2052 | return false; | |||
2053 | ||||
2054 | Value *X = nullptr, *Y = nullptr; | |||
2055 | // A shift left or a logical shift right of a power of two is a power of two | |||
2056 | // or zero. | |||
2057 | if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || | |||
2058 | match(V, m_LShr(m_Value(X), m_Value())))) | |||
2059 | return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); | |||
2060 | ||||
2061 | if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V)) | |||
2062 | return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); | |||
2063 | ||||
2064 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) | |||
2065 | return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && | |||
2066 | isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); | |||
2067 | ||||
2068 | // Peek through min/max. | |||
2069 | if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) { | |||
2070 | return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) && | |||
2071 | isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q); | |||
2072 | } | |||
2073 | ||||
2074 | if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { | |||
2075 | // A power of two and'd with anything is a power of two or zero. | |||
2076 | if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || | |||
2077 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) | |||
2078 | return true; | |||
2079 | // X & (-X) is always a power of two or zero. | |||
2080 | if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) | |||
2081 | return true; | |||
2082 | return false; | |||
2083 | } | |||
2084 | ||||
2085 | // Adding a power-of-two or zero to the same power-of-two or zero yields | |||
2086 | // either the original power-of-two, a larger power-of-two or zero. | |||
2087 | if (match(V, m_Add(m_Value(X), m_Value(Y)))) { | |||
2088 | const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); | |||
2089 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) || | |||
2090 | Q.IIQ.hasNoSignedWrap(VOBO)) { | |||
2091 | if (match(X, m_And(m_Specific(Y), m_Value())) || | |||
2092 | match(X, m_And(m_Value(), m_Specific(Y)))) | |||
2093 | if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) | |||
2094 | return true; | |||
2095 | if (match(Y, m_And(m_Specific(X), m_Value())) || | |||
2096 | match(Y, m_And(m_Value(), m_Specific(X)))) | |||
2097 | if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) | |||
2098 | return true; | |||
2099 | ||||
2100 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | |||
2101 | KnownBits LHSBits(BitWidth); | |||
2102 | computeKnownBits(X, LHSBits, Depth, Q); | |||
2103 | ||||
2104 | KnownBits RHSBits(BitWidth); | |||
2105 | computeKnownBits(Y, RHSBits, Depth, Q); | |||
2106 | // If i8 V is a power of two or zero: | |||
2107 | // ZeroBits: 1 1 1 0 1 1 1 1 | |||
2108 | // ~ZeroBits: 0 0 0 1 0 0 0 0 | |||
2109 | if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) | |||
2110 | // If OrZero isn't set, we cannot give back a zero result. | |||
2111 | // Make sure either the LHS or RHS has a bit set. | |||
2112 | if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) | |||
2113 | return true; | |||
2114 | } | |||
2115 | } | |||
2116 | ||||
2117 | // An exact divide or right shift can only shift off zero bits, so the result | |||
2118 | // is a power of two only if the first operand is a power of two and not | |||
2119 | // copying a sign bit (sdiv int_min, 2). | |||
2120 | if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || | |||
2121 | match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { | |||
2122 | return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, | |||
2123 | Depth, Q); | |||
2124 | } | |||
2125 | ||||
2126 | return false; | |||
2127 | } | |||
2128 | ||||
2129 | /// Test whether a GEP's result is known to be non-null. | |||
2130 | /// | |||
2131 | /// Uses properties inherent in a GEP to try to determine whether it is known | |||
2132 | /// to be non-null. | |||
2133 | /// | |||
2134 | /// Currently this routine does not support vector GEPs. | |||
2135 | static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, | |||
2136 | const Query &Q) { | |||
2137 | const Function *F = nullptr; | |||
2138 | if (const Instruction *I = dyn_cast<Instruction>(GEP)) | |||
2139 | F = I->getFunction(); | |||
2140 | ||||
2141 | if (!GEP->isInBounds() || | |||
2142 | NullPointerIsDefined(F, GEP->getPointerAddressSpace())) | |||
2143 | return false; | |||
2144 | ||||
2145 | // FIXME: Support vector-GEPs. | |||
2146 | assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP")(static_cast <bool> (GEP->getType()->isPointerTy( ) && "We only support plain pointer GEP") ? void (0) : __assert_fail ("GEP->getType()->isPointerTy() && \"We only support plain pointer GEP\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2146, __extension__ __PRETTY_FUNCTION__ )); | |||
2147 | ||||
2148 | // If the base pointer is non-null, we cannot walk to a null address with an | |||
2149 | // inbounds GEP in address space zero. | |||
2150 | if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) | |||
2151 | return true; | |||
2152 | ||||
2153 | // Walk the GEP operands and see if any operand introduces a non-zero offset. | |||
2154 | // If so, then the GEP cannot produce a null pointer, as doing so would | |||
2155 | // inherently violate the inbounds contract within address space zero. | |||
2156 | for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); | |||
2157 | GTI != GTE; ++GTI) { | |||
2158 | // Struct types are easy -- they must always be indexed by a constant. | |||
2159 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
2160 | ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); | |||
2161 | unsigned ElementIdx = OpC->getZExtValue(); | |||
2162 | const StructLayout *SL = Q.DL.getStructLayout(STy); | |||
2163 | uint64_t ElementOffset = SL->getElementOffset(ElementIdx); | |||
2164 | if (ElementOffset > 0) | |||
2165 | return true; | |||
2166 | continue; | |||
2167 | } | |||
2168 | ||||
2169 | // If we have a zero-sized type, the index doesn't matter. Keep looping. | |||
2170 | if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).getKnownMinSize() == 0) | |||
2171 | continue; | |||
2172 | ||||
2173 | // Fast path the constant operand case both for efficiency and so we don't | |||
2174 | // increment Depth when just zipping down an all-constant GEP. | |||
2175 | if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { | |||
2176 | if (!OpC->isZero()) | |||
2177 | return true; | |||
2178 | continue; | |||
2179 | } | |||
2180 | ||||
2181 | // We post-increment Depth here because while isKnownNonZero increments it | |||
2182 | // as well, when we pop back up that increment won't persist. We don't want | |||
2183 | // to recurse 10k times just because we have 10k GEP operands. We don't | |||
2184 | // bail completely out because we want to handle constant GEPs regardless | |||
2185 | // of depth. | |||
2186 | if (Depth++ >= MaxAnalysisRecursionDepth) | |||
2187 | continue; | |||
2188 | ||||
2189 | if (isKnownNonZero(GTI.getOperand(), Depth, Q)) | |||
2190 | return true; | |||
2191 | } | |||
2192 | ||||
2193 | return false; | |||
2194 | } | |||
2195 | ||||
2196 | static bool isKnownNonNullFromDominatingCondition(const Value *V, | |||
2197 | const Instruction *CtxI, | |||
2198 | const DominatorTree *DT) { | |||
2199 | if (isa<Constant>(V)) | |||
2200 | return false; | |||
2201 | ||||
2202 | if (!CtxI || !DT) | |||
2203 | return false; | |||
2204 | ||||
2205 | unsigned NumUsesExplored = 0; | |||
2206 | for (auto *U : V->users()) { | |||
2207 | // Avoid massive lists | |||
2208 | if (NumUsesExplored >= DomConditionsMaxUses) | |||
2209 | break; | |||
2210 | NumUsesExplored++; | |||
2211 | ||||
2212 | // If the value is used as an argument to a call or invoke, then argument | |||
2213 | // attributes may provide an answer about null-ness. | |||
2214 | if (const auto *CB = dyn_cast<CallBase>(U)) | |||
2215 | if (auto *CalledFunc = CB->getCalledFunction()) | |||
2216 | for (const Argument &Arg : CalledFunc->args()) | |||
2217 | if (CB->getArgOperand(Arg.getArgNo()) == V && | |||
2218 | Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && | |||
2219 | DT->dominates(CB, CtxI)) | |||
2220 | return true; | |||
2221 | ||||
2222 | // If the value is used as a load/store, then the pointer must be non null. | |||
2223 | if (V == getLoadStorePointerOperand(U)) { | |||
2224 | const Instruction *I = cast<Instruction>(U); | |||
2225 | if (!NullPointerIsDefined(I->getFunction(), | |||
2226 | V->getType()->getPointerAddressSpace()) && | |||
2227 | DT->dominates(I, CtxI)) | |||
2228 | return true; | |||
2229 | } | |||
2230 | ||||
2231 | // Consider only compare instructions uniquely controlling a branch | |||
2232 | Value *RHS; | |||
2233 | CmpInst::Predicate Pred; | |||
2234 | if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS)))) | |||
2235 | continue; | |||
2236 | ||||
2237 | bool NonNullIfTrue; | |||
2238 | if (cmpExcludesZero(Pred, RHS)) | |||
2239 | NonNullIfTrue = true; | |||
2240 | else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS)) | |||
2241 | NonNullIfTrue = false; | |||
2242 | else | |||
2243 | continue; | |||
2244 | ||||
2245 | SmallVector<const User *, 4> WorkList; | |||
2246 | SmallPtrSet<const User *, 4> Visited; | |||
2247 | for (auto *CmpU : U->users()) { | |||
2248 | assert(WorkList.empty() && "Should be!")(static_cast <bool> (WorkList.empty() && "Should be!" ) ? void (0) : __assert_fail ("WorkList.empty() && \"Should be!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2248, __extension__ __PRETTY_FUNCTION__ )); | |||
2249 | if (Visited.insert(CmpU).second) | |||
2250 | WorkList.push_back(CmpU); | |||
2251 | ||||
2252 | while (!WorkList.empty()) { | |||
2253 | auto *Curr = WorkList.pop_back_val(); | |||
2254 | ||||
2255 | // If a user is an AND, add all its users to the work list. We only | |||
2256 | // propagate "pred != null" condition through AND because it is only | |||
2257 | // correct to assume that all conditions of AND are met in true branch. | |||
2258 | // TODO: Support similar logic of OR and EQ predicate? | |||
2259 | if (NonNullIfTrue) | |||
2260 | if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) { | |||
2261 | for (auto *CurrU : Curr->users()) | |||
2262 | if (Visited.insert(CurrU).second) | |||
2263 | WorkList.push_back(CurrU); | |||
2264 | continue; | |||
2265 | } | |||
2266 | ||||
2267 | if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) { | |||
2268 | assert(BI->isConditional() && "uses a comparison!")(static_cast <bool> (BI->isConditional() && "uses a comparison!" ) ? void (0) : __assert_fail ("BI->isConditional() && \"uses a comparison!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2268, __extension__ __PRETTY_FUNCTION__ )); | |||
2269 | ||||
2270 | BasicBlock *NonNullSuccessor = | |||
2271 | BI->getSuccessor(NonNullIfTrue ? 0 : 1); | |||
2272 | BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); | |||
2273 | if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) | |||
2274 | return true; | |||
2275 | } else if (NonNullIfTrue && isGuard(Curr) && | |||
2276 | DT->dominates(cast<Instruction>(Curr), CtxI)) { | |||
2277 | return true; | |||
2278 | } | |||
2279 | } | |||
2280 | } | |||
2281 | } | |||
2282 | ||||
2283 | return false; | |||
2284 | } | |||
2285 | ||||
2286 | /// Does the 'Range' metadata (which must be a valid MD_range operand list) | |||
2287 | /// ensure that the value it's attached to is never Value? 'RangeType' is | |||
2288 | /// is the type of the value described by the range. | |||
2289 | static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { | |||
2290 | const unsigned NumRanges = Ranges->getNumOperands() / 2; | |||
2291 | assert(NumRanges >= 1)(static_cast <bool> (NumRanges >= 1) ? void (0) : __assert_fail ("NumRanges >= 1", "llvm/lib/Analysis/ValueTracking.cpp", 2291, __extension__ __PRETTY_FUNCTION__)); | |||
2292 | for (unsigned i = 0; i < NumRanges; ++i) { | |||
2293 | ConstantInt *Lower = | |||
2294 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); | |||
2295 | ConstantInt *Upper = | |||
2296 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); | |||
2297 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | |||
2298 | if (Range.contains(Value)) | |||
2299 | return false; | |||
2300 | } | |||
2301 | return true; | |||
2302 | } | |||
2303 | ||||
2304 | /// Try to detect a recurrence that monotonically increases/decreases from a | |||
2305 | /// non-zero starting value. These are common as induction variables. | |||
2306 | static bool isNonZeroRecurrence(const PHINode *PN) { | |||
2307 | BinaryOperator *BO = nullptr; | |||
2308 | Value *Start = nullptr, *Step = nullptr; | |||
2309 | const APInt *StartC, *StepC; | |||
2310 | if (!matchSimpleRecurrence(PN, BO, Start, Step) || | |||
2311 | !match(Start, m_APInt(StartC)) || StartC->isZero()) | |||
2312 | return false; | |||
2313 | ||||
2314 | switch (BO->getOpcode()) { | |||
2315 | case Instruction::Add: | |||
2316 | // Starting from non-zero and stepping away from zero can never wrap back | |||
2317 | // to zero. | |||
2318 | return BO->hasNoUnsignedWrap() || | |||
2319 | (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) && | |||
2320 | StartC->isNegative() == StepC->isNegative()); | |||
2321 | case Instruction::Mul: | |||
2322 | return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && | |||
2323 | match(Step, m_APInt(StepC)) && !StepC->isZero(); | |||
2324 | case Instruction::Shl: | |||
2325 | return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); | |||
2326 | case Instruction::AShr: | |||
2327 | case Instruction::LShr: | |||
2328 | return BO->isExact(); | |||
2329 | default: | |||
2330 | return false; | |||
2331 | } | |||
2332 | } | |||
2333 | ||||
2334 | /// Return true if the given value is known to be non-zero when defined. For | |||
2335 | /// vectors, return true if every demanded element is known to be non-zero when | |||
2336 | /// defined. For pointers, if the context instruction and dominator tree are | |||
2337 | /// specified, perform context-sensitive analysis and return true if the | |||
2338 | /// pointer couldn't possibly be null at the specified instruction. | |||
2339 | /// Supports values with integer or pointer type and vectors of integers. | |||
2340 | bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth, | |||
2341 | const Query &Q) { | |||
2342 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | |||
2343 | // vector | |||
2344 | if (isa<ScalableVectorType>(V->getType())) | |||
2345 | return false; | |||
2346 | ||||
2347 | if (auto *C = dyn_cast<Constant>(V)) { | |||
2348 | if (C->isNullValue()) | |||
2349 | return false; | |||
2350 | if (isa<ConstantInt>(C)) | |||
2351 | // Must be non-zero due to null test above. | |||
2352 | return true; | |||
2353 | ||||
2354 | if (auto *CE = dyn_cast<ConstantExpr>(C)) { | |||
2355 | // See the comment for IntToPtr/PtrToInt instructions below. | |||
2356 | if (CE->getOpcode() == Instruction::IntToPtr || | |||
2357 | CE->getOpcode() == Instruction::PtrToInt) | |||
2358 | if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType()) | |||
2359 | .getFixedSize() <= | |||
2360 | Q.DL.getTypeSizeInBits(CE->getType()).getFixedSize()) | |||
2361 | return isKnownNonZero(CE->getOperand(0), Depth, Q); | |||
2362 | } | |||
2363 | ||||
2364 | // For constant vectors, check that all elements are undefined or known | |||
2365 | // non-zero to determine that the whole vector is known non-zero. | |||
2366 | if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) { | |||
2367 | for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { | |||
2368 | if (!DemandedElts[i]) | |||
2369 | continue; | |||
2370 | Constant *Elt = C->getAggregateElement(i); | |||
2371 | if (!Elt || Elt->isNullValue()) | |||
2372 | return false; | |||
2373 | if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) | |||
2374 | return false; | |||
2375 | } | |||
2376 | return true; | |||
2377 | } | |||
2378 | ||||
2379 | // A global variable in address space 0 is non null unless extern weak | |||
2380 | // or an absolute symbol reference. Other address spaces may have null as a | |||
2381 | // valid address for a global, so we can't assume anything. | |||
2382 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { | |||
2383 | if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && | |||
2384 | GV->getType()->getAddressSpace() == 0) | |||
2385 | return true; | |||
2386 | } else | |||
2387 | return false; | |||
2388 | } | |||
2389 | ||||
2390 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
2391 | if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) { | |||
2392 | // If the possible ranges don't contain zero, then the value is | |||
2393 | // definitely non-zero. | |||
2394 | if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { | |||
2395 | const APInt ZeroValue(Ty->getBitWidth(), 0); | |||
2396 | if (rangeMetadataExcludesValue(Ranges, ZeroValue)) | |||
2397 | return true; | |||
2398 | } | |||
2399 | } | |||
2400 | } | |||
2401 | ||||
2402 | if (isKnownNonZeroFromAssume(V, Q)) | |||
2403 | return true; | |||
2404 | ||||
2405 | // Some of the tests below are recursive, so bail out if we hit the limit. | |||
2406 | if (Depth++ >= MaxAnalysisRecursionDepth) | |||
2407 | return false; | |||
2408 | ||||
2409 | // Check for pointer simplifications. | |||
2410 | ||||
2411 | if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) { | |||
2412 | // Alloca never returns null, malloc might. | |||
2413 | if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0) | |||
2414 | return true; | |||
2415 | ||||
2416 | // A byval, inalloca may not be null in a non-default addres space. A | |||
2417 | // nonnull argument is assumed never 0. | |||
2418 | if (const Argument *A = dyn_cast<Argument>(V)) { | |||
2419 | if (((A->hasPassPointeeByValueCopyAttr() && | |||
2420 | !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) || | |||
2421 | A->hasNonNullAttr())) | |||
2422 | return true; | |||
2423 | } | |||
2424 | ||||
2425 | // A Load tagged with nonnull metadata is never null. | |||
2426 | if (const LoadInst *LI = dyn_cast<LoadInst>(V)) | |||
2427 | if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull)) | |||
2428 | return true; | |||
2429 | ||||
2430 | if (const auto *Call = dyn_cast<CallBase>(V)) { | |||
2431 | if (Call->isReturnNonNull()) | |||
2432 | return true; | |||
2433 | if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true)) | |||
2434 | return isKnownNonZero(RP, Depth, Q); | |||
2435 | } | |||
2436 | } | |||
2437 | ||||
2438 | if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT)) | |||
2439 | return true; | |||
2440 | ||||
2441 | // Check for recursive pointer simplifications. | |||
2442 | if (V->getType()->isPointerTy()) { | |||
2443 | // Look through bitcast operations, GEPs, and int2ptr instructions as they | |||
2444 | // do not alter the value, or at least not the nullness property of the | |||
2445 | // value, e.g., int2ptr is allowed to zero/sign extend the value. | |||
2446 | // | |||
2447 | // Note that we have to take special care to avoid looking through | |||
2448 | // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well | |||
2449 | // as casts that can alter the value, e.g., AddrSpaceCasts. | |||
2450 | if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) | |||
2451 | return isGEPKnownNonNull(GEP, Depth, Q); | |||
2452 | ||||
2453 | if (auto *BCO = dyn_cast<BitCastOperator>(V)) | |||
2454 | return isKnownNonZero(BCO->getOperand(0), Depth, Q); | |||
2455 | ||||
2456 | if (auto *I2P = dyn_cast<IntToPtrInst>(V)) | |||
2457 | if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()).getFixedSize() <= | |||
2458 | Q.DL.getTypeSizeInBits(I2P->getDestTy()).getFixedSize()) | |||
2459 | return isKnownNonZero(I2P->getOperand(0), Depth, Q); | |||
2460 | } | |||
2461 | ||||
2462 | // Similar to int2ptr above, we can look through ptr2int here if the cast | |||
2463 | // is a no-op or an extend and not a truncate. | |||
2464 | if (auto *P2I = dyn_cast<PtrToIntInst>(V)) | |||
2465 | if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()).getFixedSize() <= | |||
2466 | Q.DL.getTypeSizeInBits(P2I->getDestTy()).getFixedSize()) | |||
2467 | return isKnownNonZero(P2I->getOperand(0), Depth, Q); | |||
2468 | ||||
2469 | unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); | |||
2470 | ||||
2471 | // X | Y != 0 if X != 0 or Y != 0. | |||
2472 | Value *X = nullptr, *Y = nullptr; | |||
2473 | if (match(V, m_Or(m_Value(X), m_Value(Y)))) | |||
2474 | return isKnownNonZero(X, DemandedElts, Depth, Q) || | |||
2475 | isKnownNonZero(Y, DemandedElts, Depth, Q); | |||
2476 | ||||
2477 | // ext X != 0 if X != 0. | |||
2478 | if (isa<SExtInst>(V) || isa<ZExtInst>(V)) | |||
2479 | return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q); | |||
2480 | ||||
2481 | // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined | |||
2482 | // if the lowest bit is shifted off the end. | |||
2483 | if (match(V, m_Shl(m_Value(X), m_Value(Y)))) { | |||
2484 | // shl nuw can't remove any non-zero bits. | |||
2485 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | |||
2486 | if (Q.IIQ.hasNoUnsignedWrap(BO)) | |||
2487 | return isKnownNonZero(X, Depth, Q); | |||
2488 | ||||
2489 | KnownBits Known(BitWidth); | |||
2490 | computeKnownBits(X, DemandedElts, Known, Depth, Q); | |||
2491 | if (Known.One[0]) | |||
2492 | return true; | |||
2493 | } | |||
2494 | // shr X, Y != 0 if X is negative. Note that the value of the shift is not | |||
2495 | // defined if the sign bit is shifted off the end. | |||
2496 | else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { | |||
2497 | // shr exact can only shift out zero bits. | |||
2498 | const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); | |||
2499 | if (BO->isExact()) | |||
2500 | return isKnownNonZero(X, Depth, Q); | |||
2501 | ||||
2502 | KnownBits Known = computeKnownBits(X, DemandedElts, Depth, Q); | |||
2503 | if (Known.isNegative()) | |||
2504 | return true; | |||
2505 | ||||
2506 | // If the shifter operand is a constant, and all of the bits shifted | |||
2507 | // out are known to be zero, and X is known non-zero then at least one | |||
2508 | // non-zero bit must remain. | |||
2509 | if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) { | |||
2510 | auto ShiftVal = Shift->getLimitedValue(BitWidth - 1); | |||
2511 | // Is there a known one in the portion not shifted out? | |||
2512 | if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal) | |||
2513 | return true; | |||
2514 | // Are all the bits to be shifted out known zero? | |||
2515 | if (Known.countMinTrailingZeros() >= ShiftVal) | |||
2516 | return isKnownNonZero(X, DemandedElts, Depth, Q); | |||
2517 | } | |||
2518 | } | |||
2519 | // div exact can only produce a zero if the dividend is zero. | |||
2520 | else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { | |||
2521 | return isKnownNonZero(X, DemandedElts, Depth, Q); | |||
2522 | } | |||
2523 | // X + Y. | |||
2524 | else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { | |||
2525 | KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); | |||
2526 | KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); | |||
2527 | ||||
2528 | // If X and Y are both non-negative (as signed values) then their sum is not | |||
2529 | // zero unless both X and Y are zero. | |||
2530 | if (XKnown.isNonNegative() && YKnown.isNonNegative()) | |||
2531 | if (isKnownNonZero(X, DemandedElts, Depth, Q) || | |||
2532 | isKnownNonZero(Y, DemandedElts, Depth, Q)) | |||
2533 | return true; | |||
2534 | ||||
2535 | // If X and Y are both negative (as signed values) then their sum is not | |||
2536 | // zero unless both X and Y equal INT_MIN. | |||
2537 | if (XKnown.isNegative() && YKnown.isNegative()) { | |||
2538 | APInt Mask = APInt::getSignedMaxValue(BitWidth); | |||
2539 | // The sign bit of X is set. If some other bit is set then X is not equal | |||
2540 | // to INT_MIN. | |||
2541 | if (XKnown.One.intersects(Mask)) | |||
2542 | return true; | |||
2543 | // The sign bit of Y is set. If some other bit is set then Y is not equal | |||
2544 | // to INT_MIN. | |||
2545 | if (YKnown.One.intersects(Mask)) | |||
2546 | return true; | |||
2547 | } | |||
2548 | ||||
2549 | // The sum of a non-negative number and a power of two is not zero. | |||
2550 | if (XKnown.isNonNegative() && | |||
2551 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) | |||
2552 | return true; | |||
2553 | if (YKnown.isNonNegative() && | |||
2554 | isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) | |||
2555 | return true; | |||
2556 | } | |||
2557 | // X * Y. | |||
2558 | else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { | |||
2559 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | |||
2560 | // If X and Y are non-zero then so is X * Y as long as the multiplication | |||
2561 | // does not overflow. | |||
2562 | if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) && | |||
2563 | isKnownNonZero(X, DemandedElts, Depth, Q) && | |||
2564 | isKnownNonZero(Y, DemandedElts, Depth, Q)) | |||
2565 | return true; | |||
2566 | } | |||
2567 | // (C ? X : Y) != 0 if X != 0 and Y != 0. | |||
2568 | else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { | |||
2569 | if (isKnownNonZero(SI->getTrueValue(), DemandedElts, Depth, Q) && | |||
2570 | isKnownNonZero(SI->getFalseValue(), DemandedElts, Depth, Q)) | |||
2571 | return true; | |||
2572 | } | |||
2573 | // PHI | |||
2574 | else if (const PHINode *PN = dyn_cast<PHINode>(V)) { | |||
2575 | if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) | |||
2576 | return true; | |||
2577 | ||||
2578 | // Check if all incoming values are non-zero using recursion. | |||
2579 | Query RecQ = Q; | |||
2580 | unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); | |||
2581 | return llvm::all_of(PN->operands(), [&](const Use &U) { | |||
2582 | if (U.get() == PN) | |||
2583 | return true; | |||
2584 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); | |||
2585 | return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ); | |||
2586 | }); | |||
2587 | } | |||
2588 | // ExtractElement | |||
2589 | else if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) { | |||
2590 | const Value *Vec = EEI->getVectorOperand(); | |||
2591 | const Value *Idx = EEI->getIndexOperand(); | |||
2592 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | |||
2593 | if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { | |||
2594 | unsigned NumElts = VecTy->getNumElements(); | |||
2595 | APInt DemandedVecElts = APInt::getAllOnes(NumElts); | |||
2596 | if (CIdx && CIdx->getValue().ult(NumElts)) | |||
2597 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | |||
2598 | return isKnownNonZero(Vec, DemandedVecElts, Depth, Q); | |||
2599 | } | |||
2600 | } | |||
2601 | // Freeze | |||
2602 | else if (const FreezeInst *FI = dyn_cast<FreezeInst>(V)) { | |||
2603 | auto *Op = FI->getOperand(0); | |||
2604 | if (isKnownNonZero(Op, Depth, Q) && | |||
2605 | isGuaranteedNotToBePoison(Op, Q.AC, Q.CxtI, Q.DT, Depth)) | |||
2606 | return true; | |||
2607 | } | |||
2608 | ||||
2609 | KnownBits Known(BitWidth); | |||
2610 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
2611 | return Known.One != 0; | |||
2612 | } | |||
2613 | ||||
2614 | bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) { | |||
2615 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | |||
2616 | // vector | |||
2617 | if (isa<ScalableVectorType>(V->getType())) | |||
2618 | return false; | |||
2619 | ||||
2620 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
2621 | APInt DemandedElts = | |||
2622 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
2623 | return isKnownNonZero(V, DemandedElts, Depth, Q); | |||
2624 | } | |||
2625 | ||||
2626 | /// If the pair of operators are the same invertible function, return the | |||
2627 | /// the operands of the function corresponding to each input. Otherwise, | |||
2628 | /// return None. An invertible function is one that is 1-to-1 and maps | |||
2629 | /// every input value to exactly one output value. This is equivalent to | |||
2630 | /// saying that Op1 and Op2 are equal exactly when the specified pair of | |||
2631 | /// operands are equal, (except that Op1 and Op2 may be poison more often.) | |||
2632 | static Optional<std::pair<Value*, Value*>> | |||
2633 | getInvertibleOperands(const Operator *Op1, | |||
2634 | const Operator *Op2) { | |||
2635 | if (Op1->getOpcode() != Op2->getOpcode()) | |||
2636 | return None; | |||
2637 | ||||
2638 | auto getOperands = [&](unsigned OpNum) -> auto { | |||
2639 | return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum)); | |||
2640 | }; | |||
2641 | ||||
2642 | switch (Op1->getOpcode()) { | |||
2643 | default: | |||
2644 | break; | |||
2645 | case Instruction::Add: | |||
2646 | case Instruction::Sub: | |||
2647 | if (Op1->getOperand(0) == Op2->getOperand(0)) | |||
2648 | return getOperands(1); | |||
2649 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
2650 | return getOperands(0); | |||
2651 | break; | |||
2652 | case Instruction::Mul: { | |||
2653 | // invertible if A * B == (A * B) mod 2^N where A, and B are integers | |||
2654 | // and N is the bitwdith. The nsw case is non-obvious, but proven by | |||
2655 | // alive2: https://alive2.llvm.org/ce/z/Z6D5qK | |||
2656 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | |||
2657 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | |||
2658 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | |||
2659 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | |||
2660 | break; | |||
2661 | ||||
2662 | // Assume operand order has been canonicalized | |||
2663 | if (Op1->getOperand(1) == Op2->getOperand(1) && | |||
2664 | isa<ConstantInt>(Op1->getOperand(1)) && | |||
2665 | !cast<ConstantInt>(Op1->getOperand(1))->isZero()) | |||
2666 | return getOperands(0); | |||
2667 | break; | |||
2668 | } | |||
2669 | case Instruction::Shl: { | |||
2670 | // Same as multiplies, with the difference that we don't need to check | |||
2671 | // for a non-zero multiply. Shifts always multiply by non-zero. | |||
2672 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | |||
2673 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | |||
2674 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | |||
2675 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | |||
2676 | break; | |||
2677 | ||||
2678 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
2679 | return getOperands(0); | |||
2680 | break; | |||
2681 | } | |||
2682 | case Instruction::AShr: | |||
2683 | case Instruction::LShr: { | |||
2684 | auto *PEO1 = cast<PossiblyExactOperator>(Op1); | |||
2685 | auto *PEO2 = cast<PossiblyExactOperator>(Op2); | |||
2686 | if (!PEO1->isExact() || !PEO2->isExact()) | |||
2687 | break; | |||
2688 | ||||
2689 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
2690 | return getOperands(0); | |||
2691 | break; | |||
2692 | } | |||
2693 | case Instruction::SExt: | |||
2694 | case Instruction::ZExt: | |||
2695 | if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType()) | |||
2696 | return getOperands(0); | |||
2697 | break; | |||
2698 | case Instruction::PHI: { | |||
2699 | const PHINode *PN1 = cast<PHINode>(Op1); | |||
2700 | const PHINode *PN2 = cast<PHINode>(Op2); | |||
2701 | ||||
2702 | // If PN1 and PN2 are both recurrences, can we prove the entire recurrences | |||
2703 | // are a single invertible function of the start values? Note that repeated | |||
2704 | // application of an invertible function is also invertible | |||
2705 | BinaryOperator *BO1 = nullptr; | |||
2706 | Value *Start1 = nullptr, *Step1 = nullptr; | |||
2707 | BinaryOperator *BO2 = nullptr; | |||
2708 | Value *Start2 = nullptr, *Step2 = nullptr; | |||
2709 | if (PN1->getParent() != PN2->getParent() || | |||
2710 | !matchSimpleRecurrence(PN1, BO1, Start1, Step1) || | |||
2711 | !matchSimpleRecurrence(PN2, BO2, Start2, Step2)) | |||
2712 | break; | |||
2713 | ||||
2714 | auto Values = getInvertibleOperands(cast<Operator>(BO1), | |||
2715 | cast<Operator>(BO2)); | |||
2716 | if (!Values) | |||
2717 | break; | |||
2718 | ||||
2719 | // We have to be careful of mutually defined recurrences here. Ex: | |||
2720 | // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V | |||
2721 | // * X_i = Y_i = X_(i-1) OP Y_(i-1) | |||
2722 | // The invertibility of these is complicated, and not worth reasoning | |||
2723 | // about (yet?). | |||
2724 | if (Values->first != PN1 || Values->second != PN2) | |||
2725 | break; | |||
2726 | ||||
2727 | return std::make_pair(Start1, Start2); | |||
2728 | } | |||
2729 | } | |||
2730 | return None; | |||
2731 | } | |||
2732 | ||||
2733 | /// Return true if V2 == V1 + X, where X is known non-zero. | |||
2734 | static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth, | |||
2735 | const Query &Q) { | |||
2736 | const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); | |||
2737 | if (!BO || BO->getOpcode() != Instruction::Add) | |||
2738 | return false; | |||
2739 | Value *Op = nullptr; | |||
2740 | if (V2 == BO->getOperand(0)) | |||
2741 | Op = BO->getOperand(1); | |||
2742 | else if (V2 == BO->getOperand(1)) | |||
2743 | Op = BO->getOperand(0); | |||
2744 | else | |||
2745 | return false; | |||
2746 | return isKnownNonZero(Op, Depth + 1, Q); | |||
2747 | } | |||
2748 | ||||
2749 | /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and | |||
2750 | /// the multiplication is nuw or nsw. | |||
2751 | static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth, | |||
2752 | const Query &Q) { | |||
2753 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | |||
2754 | const APInt *C; | |||
2755 | return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) && | |||
2756 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | |||
2757 | !C->isZero() && !C->isOne() && isKnownNonZero(V1, Depth + 1, Q); | |||
2758 | } | |||
2759 | return false; | |||
2760 | } | |||
2761 | ||||
2762 | /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and | |||
2763 | /// the shift is nuw or nsw. | |||
2764 | static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth, | |||
2765 | const Query &Q) { | |||
2766 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | |||
2767 | const APInt *C; | |||
2768 | return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) && | |||
2769 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | |||
2770 | !C->isZero() && isKnownNonZero(V1, Depth + 1, Q); | |||
2771 | } | |||
2772 | return false; | |||
2773 | } | |||
2774 | ||||
2775 | static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, | |||
2776 | unsigned Depth, const Query &Q) { | |||
2777 | // Check two PHIs are in same block. | |||
2778 | if (PN1->getParent() != PN2->getParent()) | |||
2779 | return false; | |||
2780 | ||||
2781 | SmallPtrSet<const BasicBlock *, 8> VisitedBBs; | |||
2782 | bool UsedFullRecursion = false; | |||
2783 | for (const BasicBlock *IncomBB : PN1->blocks()) { | |||
2784 | if (!VisitedBBs.insert(IncomBB).second) | |||
2785 | continue; // Don't reprocess blocks that we have dealt with already. | |||
2786 | const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB); | |||
2787 | const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB); | |||
2788 | const APInt *C1, *C2; | |||
2789 | if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2) | |||
2790 | continue; | |||
2791 | ||||
2792 | // Only one pair of phi operands is allowed for full recursion. | |||
2793 | if (UsedFullRecursion) | |||
2794 | return false; | |||
2795 | ||||
2796 | Query RecQ = Q; | |||
2797 | RecQ.CxtI = IncomBB->getTerminator(); | |||
2798 | if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ)) | |||
2799 | return false; | |||
2800 | UsedFullRecursion = true; | |||
2801 | } | |||
2802 | return true; | |||
2803 | } | |||
2804 | ||||
2805 | /// Return true if it is known that V1 != V2. | |||
2806 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | |||
2807 | const Query &Q) { | |||
2808 | if (V1 == V2) | |||
2809 | return false; | |||
2810 | if (V1->getType() != V2->getType()) | |||
2811 | // We can't look through casts yet. | |||
2812 | return false; | |||
2813 | ||||
2814 | if (Depth >= MaxAnalysisRecursionDepth) | |||
2815 | return false; | |||
2816 | ||||
2817 | // See if we can recurse through (exactly one of) our operands. This | |||
2818 | // requires our operation be 1-to-1 and map every input value to exactly | |||
2819 | // one output value. Such an operation is invertible. | |||
2820 | auto *O1 = dyn_cast<Operator>(V1); | |||
2821 | auto *O2 = dyn_cast<Operator>(V2); | |||
2822 | if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { | |||
2823 | if (auto Values = getInvertibleOperands(O1, O2)) | |||
2824 | return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q); | |||
2825 | ||||
2826 | if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) { | |||
2827 | const PHINode *PN2 = cast<PHINode>(V2); | |||
2828 | // FIXME: This is missing a generalization to handle the case where one is | |||
2829 | // a PHI and another one isn't. | |||
2830 | if (isNonEqualPHIs(PN1, PN2, Depth, Q)) | |||
2831 | return true; | |||
2832 | }; | |||
2833 | } | |||
2834 | ||||
2835 | if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q)) | |||
2836 | return true; | |||
2837 | ||||
2838 | if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q)) | |||
2839 | return true; | |||
2840 | ||||
2841 | if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q)) | |||
2842 | return true; | |||
2843 | ||||
2844 | if (V1->getType()->isIntOrIntVectorTy()) { | |||
2845 | // Are any known bits in V1 contradictory to known bits in V2? If V1 | |||
2846 | // has a known zero where V2 has a known one, they must not be equal. | |||
2847 | KnownBits Known1 = computeKnownBits(V1, Depth, Q); | |||
2848 | KnownBits Known2 = computeKnownBits(V2, Depth, Q); | |||
2849 | ||||
2850 | if (Known1.Zero.intersects(Known2.One) || | |||
2851 | Known2.Zero.intersects(Known1.One)) | |||
2852 | return true; | |||
2853 | } | |||
2854 | return false; | |||
2855 | } | |||
2856 | ||||
2857 | /// Return true if 'V & Mask' is known to be zero. We use this predicate to | |||
2858 | /// simplify operations downstream. Mask is known to be zero for bits that V | |||
2859 | /// cannot have. | |||
2860 | /// | |||
2861 | /// This function is defined on values with integer type, values with pointer | |||
2862 | /// type, and vectors of integers. In the case | |||
2863 | /// where V is a vector, the mask, known zero, and known one values are the | |||
2864 | /// same width as the vector element, and the bit is set only if it is true | |||
2865 | /// for all of the elements in the vector. | |||
2866 | bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | |||
2867 | const Query &Q) { | |||
2868 | KnownBits Known(Mask.getBitWidth()); | |||
2869 | computeKnownBits(V, Known, Depth, Q); | |||
2870 | return Mask.isSubsetOf(Known.Zero); | |||
2871 | } | |||
2872 | ||||
2873 | // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). | |||
2874 | // Returns the input and lower/upper bounds. | |||
2875 | static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, | |||
2876 | const APInt *&CLow, const APInt *&CHigh) { | |||
2877 | assert(isa<Operator>(Select) &&(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2879, __extension__ __PRETTY_FUNCTION__ )) | |||
2878 | cast<Operator>(Select)->getOpcode() == Instruction::Select &&(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2879, __extension__ __PRETTY_FUNCTION__ )) | |||
2879 | "Input should be a Select!")(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2879, __extension__ __PRETTY_FUNCTION__ )); | |||
2880 | ||||
2881 | const Value *LHS = nullptr, *RHS = nullptr; | |||
2882 | SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor; | |||
2883 | if (SPF != SPF_SMAX && SPF != SPF_SMIN) | |||
2884 | return false; | |||
2885 | ||||
2886 | if (!match(RHS, m_APInt(CLow))) | |||
2887 | return false; | |||
2888 | ||||
2889 | const Value *LHS2 = nullptr, *RHS2 = nullptr; | |||
2890 | SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor; | |||
2891 | if (getInverseMinMaxFlavor(SPF) != SPF2) | |||
2892 | return false; | |||
2893 | ||||
2894 | if (!match(RHS2, m_APInt(CHigh))) | |||
2895 | return false; | |||
2896 | ||||
2897 | if (SPF == SPF_SMIN) | |||
2898 | std::swap(CLow, CHigh); | |||
2899 | ||||
2900 | In = LHS2; | |||
2901 | return CLow->sle(*CHigh); | |||
2902 | } | |||
2903 | ||||
2904 | static bool isSignedMinMaxIntrinsicClamp(const IntrinsicInst *II, | |||
2905 | const APInt *&CLow, | |||
2906 | const APInt *&CHigh) { | |||
2907 | assert((II->getIntrinsicID() == Intrinsic::smin ||(static_cast <bool> ((II->getIntrinsicID() == Intrinsic ::smin || II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax") ? void (0) : __assert_fail ("(II->getIntrinsicID() == Intrinsic::smin || II->getIntrinsicID() == Intrinsic::smax) && \"Must be smin/smax\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2908, __extension__ __PRETTY_FUNCTION__ )) | |||
2908 | II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax")(static_cast <bool> ((II->getIntrinsicID() == Intrinsic ::smin || II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax") ? void (0) : __assert_fail ("(II->getIntrinsicID() == Intrinsic::smin || II->getIntrinsicID() == Intrinsic::smax) && \"Must be smin/smax\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2908, __extension__ __PRETTY_FUNCTION__ )); | |||
2909 | ||||
2910 | Intrinsic::ID InverseID = getInverseMinMaxIntrinsic(II->getIntrinsicID()); | |||
2911 | auto *InnerII = dyn_cast<IntrinsicInst>(II->getArgOperand(0)); | |||
2912 | if (!InnerII || InnerII->getIntrinsicID() != InverseID || | |||
2913 | !match(II->getArgOperand(1), m_APInt(CLow)) || | |||
2914 | !match(InnerII->getArgOperand(1), m_APInt(CHigh))) | |||
2915 | return false; | |||
2916 | ||||
2917 | if (II->getIntrinsicID() == Intrinsic::smin) | |||
2918 | std::swap(CLow, CHigh); | |||
2919 | return CLow->sle(*CHigh); | |||
2920 | } | |||
2921 | ||||
2922 | /// For vector constants, loop over the elements and find the constant with the | |||
2923 | /// minimum number of sign bits. Return 0 if the value is not a vector constant | |||
2924 | /// or if any element was not analyzed; otherwise, return the count for the | |||
2925 | /// element with the minimum number of sign bits. | |||
2926 | static unsigned computeNumSignBitsVectorConstant(const Value *V, | |||
2927 | const APInt &DemandedElts, | |||
2928 | unsigned TyBits) { | |||
2929 | const auto *CV = dyn_cast<Constant>(V); | |||
2930 | if (!CV || !isa<FixedVectorType>(CV->getType())) | |||
2931 | return 0; | |||
2932 | ||||
2933 | unsigned MinSignBits = TyBits; | |||
2934 | unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements(); | |||
2935 | for (unsigned i = 0; i != NumElts; ++i) { | |||
2936 | if (!DemandedElts[i]) | |||
2937 | continue; | |||
2938 | // If we find a non-ConstantInt, bail out. | |||
2939 | auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); | |||
2940 | if (!Elt) | |||
2941 | return 0; | |||
2942 | ||||
2943 | MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits()); | |||
2944 | } | |||
2945 | ||||
2946 | return MinSignBits; | |||
2947 | } | |||
2948 | ||||
2949 | static unsigned ComputeNumSignBitsImpl(const Value *V, | |||
2950 | const APInt &DemandedElts, | |||
2951 | unsigned Depth, const Query &Q); | |||
2952 | ||||
2953 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | |||
2954 | unsigned Depth, const Query &Q) { | |||
2955 | unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); | |||
2956 | assert(Result > 0 && "At least one sign bit needs to be present!")(static_cast <bool> (Result > 0 && "At least one sign bit needs to be present!" ) ? void (0) : __assert_fail ("Result > 0 && \"At least one sign bit needs to be present!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2956, __extension__ __PRETTY_FUNCTION__ )); | |||
2957 | return Result; | |||
2958 | } | |||
2959 | ||||
2960 | /// Return the number of times the sign bit of the register is replicated into | |||
2961 | /// the other bits. We know that at least 1 bit is always equal to the sign bit | |||
2962 | /// (itself), but other cases can give us information. For example, immediately | |||
2963 | /// after an "ashr X, 2", we know that the top 3 bits are all equal to each | |||
2964 | /// other, so we return 3. For vectors, return the number of sign bits for the | |||
2965 | /// vector element with the minimum number of known sign bits of the demanded | |||
2966 | /// elements in the vector specified by DemandedElts. | |||
2967 | static unsigned ComputeNumSignBitsImpl(const Value *V, | |||
2968 | const APInt &DemandedElts, | |||
2969 | unsigned Depth, const Query &Q) { | |||
2970 | Type *Ty = V->getType(); | |||
2971 | ||||
2972 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | |||
2973 | // vector | |||
2974 | if (isa<ScalableVectorType>(Ty)) | |||
2975 | return 1; | |||
2976 | ||||
2977 | #ifndef NDEBUG | |||
2978 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2978, __extension__ __PRETTY_FUNCTION__ )); | |||
2979 | ||||
2980 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | |||
2981 | assert((static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2983, __extension__ __PRETTY_FUNCTION__ )) | |||
2982 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2983, __extension__ __PRETTY_FUNCTION__ )) | |||
2983 | "DemandedElt width should equal the fixed vector number of elements")(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2983, __extension__ __PRETTY_FUNCTION__ )); | |||
2984 | } else { | |||
2985 | assert(DemandedElts == APInt(1, 1) &&(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2986, __extension__ __PRETTY_FUNCTION__ )) | |||
2986 | "DemandedElt width should be 1 for scalars")(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2986, __extension__ __PRETTY_FUNCTION__ )); | |||
2987 | } | |||
2988 | #endif | |||
2989 | ||||
2990 | // We return the minimum number of sign bits that are guaranteed to be present | |||
2991 | // in V, so for undef we have to conservatively return 1. We don't have the | |||
2992 | // same behavior for poison though -- that's a FIXME today. | |||
2993 | ||||
2994 | Type *ScalarTy = Ty->getScalarType(); | |||
2995 | unsigned TyBits = ScalarTy->isPointerTy() ? | |||
2996 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | |||
2997 | Q.DL.getTypeSizeInBits(ScalarTy); | |||
2998 | ||||
2999 | unsigned Tmp, Tmp2; | |||
3000 | unsigned FirstAnswer = 1; | |||
3001 | ||||
3002 | // Note that ConstantInt is handled by the general computeKnownBits case | |||
3003 | // below. | |||
3004 | ||||
3005 | if (Depth == MaxAnalysisRecursionDepth) | |||
3006 | return 1; | |||
3007 | ||||
3008 | if (auto *U = dyn_cast<Operator>(V)) { | |||
3009 | switch (Operator::getOpcode(V)) { | |||
3010 | default: break; | |||
3011 | case Instruction::SExt: | |||
3012 | Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); | |||
3013 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; | |||
3014 | ||||
3015 | case Instruction::SDiv: { | |||
3016 | const APInt *Denominator; | |||
3017 | // sdiv X, C -> adds log(C) sign bits. | |||
3018 | if (match(U->getOperand(1), m_APInt(Denominator))) { | |||
3019 | ||||
3020 | // Ignore non-positive denominator. | |||
3021 | if (!Denominator->isStrictlyPositive()) | |||
3022 | break; | |||
3023 | ||||
3024 | // Calculate the incoming numerator bits. | |||
3025 | unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3026 | ||||
3027 | // Add floor(log(C)) bits to the numerator bits. | |||
3028 | return std::min(TyBits, NumBits + Denominator->logBase2()); | |||
3029 | } | |||
3030 | break; | |||
3031 | } | |||
3032 | ||||
3033 | case Instruction::SRem: { | |||
3034 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3035 | ||||
3036 | const APInt *Denominator; | |||
3037 | // srem X, C -> we know that the result is within [-C+1,C) when C is a | |||
3038 | // positive constant. This let us put a lower bound on the number of sign | |||
3039 | // bits. | |||
3040 | if (match(U->getOperand(1), m_APInt(Denominator))) { | |||
3041 | ||||
3042 | // Ignore non-positive denominator. | |||
3043 | if (Denominator->isStrictlyPositive()) { | |||
3044 | // Calculate the leading sign bit constraints by examining the | |||
3045 | // denominator. Given that the denominator is positive, there are two | |||
3046 | // cases: | |||
3047 | // | |||
3048 | // 1. The numerator is positive. The result range is [0,C) and | |||
3049 | // [0,C) u< (1 << ceilLogBase2(C)). | |||
3050 | // | |||
3051 | // 2. The numerator is negative. Then the result range is (-C,0] and | |||
3052 | // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). | |||
3053 | // | |||
3054 | // Thus a lower bound on the number of sign bits is `TyBits - | |||
3055 | // ceilLogBase2(C)`. | |||
3056 | ||||
3057 | unsigned ResBits = TyBits - Denominator->ceilLogBase2(); | |||
3058 | Tmp = std::max(Tmp, ResBits); | |||
3059 | } | |||
3060 | } | |||
3061 | return Tmp; | |||
3062 | } | |||
3063 | ||||
3064 | case Instruction::AShr: { | |||
3065 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3066 | // ashr X, C -> adds C sign bits. Vectors too. | |||
3067 | const APInt *ShAmt; | |||
3068 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | |||
3069 | if (ShAmt->uge(TyBits)) | |||
3070 | break; // Bad shift. | |||
3071 | unsigned ShAmtLimited = ShAmt->getZExtValue(); | |||
3072 | Tmp += ShAmtLimited; | |||
3073 | if (Tmp > TyBits) Tmp = TyBits; | |||
3074 | } | |||
3075 | return Tmp; | |||
3076 | } | |||
3077 | case Instruction::Shl: { | |||
3078 | const APInt *ShAmt; | |||
3079 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | |||
3080 | // shl destroys sign bits. | |||
3081 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3082 | if (ShAmt->uge(TyBits) || // Bad shift. | |||
3083 | ShAmt->uge(Tmp)) break; // Shifted all sign bits out. | |||
3084 | Tmp2 = ShAmt->getZExtValue(); | |||
3085 | return Tmp - Tmp2; | |||
3086 | } | |||
3087 | break; | |||
3088 | } | |||
3089 | case Instruction::And: | |||
3090 | case Instruction::Or: | |||
3091 | case Instruction::Xor: // NOT is handled here. | |||
3092 | // Logical binary ops preserve the number of sign bits at the worst. | |||
3093 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3094 | if (Tmp != 1) { | |||
3095 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
3096 | FirstAnswer = std::min(Tmp, Tmp2); | |||
3097 | // We computed what we know about the sign bits as our first | |||
3098 | // answer. Now proceed to the generic code that uses | |||
3099 | // computeKnownBits, and pick whichever answer is better. | |||
3100 | } | |||
3101 | break; | |||
3102 | ||||
3103 | case Instruction::Select: { | |||
3104 | // If we have a clamp pattern, we know that the number of sign bits will | |||
3105 | // be the minimum of the clamp min/max range. | |||
3106 | const Value *X; | |||
3107 | const APInt *CLow, *CHigh; | |||
3108 | if (isSignedMinMaxClamp(U, X, CLow, CHigh)) | |||
3109 | return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); | |||
3110 | ||||
3111 | Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
3112 | if (Tmp == 1) break; | |||
3113 | Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); | |||
3114 | return std::min(Tmp, Tmp2); | |||
3115 | } | |||
3116 | ||||
3117 | case Instruction::Add: | |||
3118 | // Add can have at most one carry bit. Thus we know that the output | |||
3119 | // is, at worst, one more bit than the inputs. | |||
3120 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3121 | if (Tmp == 1) break; | |||
3122 | ||||
3123 | // Special case decrementing a value (ADD X, -1): | |||
3124 | if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) | |||
3125 | if (CRHS->isAllOnesValue()) { | |||
3126 | KnownBits Known(TyBits); | |||
3127 | computeKnownBits(U->getOperand(0), Known, Depth + 1, Q); | |||
3128 | ||||
3129 | // If the input is known to be 0 or 1, the output is 0/-1, which is | |||
3130 | // all sign bits set. | |||
3131 | if ((Known.Zero | 1).isAllOnes()) | |||
3132 | return TyBits; | |||
3133 | ||||
3134 | // If we are subtracting one from a positive number, there is no carry | |||
3135 | // out of the result. | |||
3136 | if (Known.isNonNegative()) | |||
3137 | return Tmp; | |||
3138 | } | |||
3139 | ||||
3140 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
3141 | if (Tmp2 == 1) break; | |||
3142 | return std::min(Tmp, Tmp2) - 1; | |||
3143 | ||||
3144 | case Instruction::Sub: | |||
3145 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
3146 | if (Tmp2 == 1) break; | |||
3147 | ||||
3148 | // Handle NEG. | |||
3149 | if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) | |||
3150 | if (CLHS->isNullValue()) { | |||
3151 | KnownBits Known(TyBits); | |||
3152 | computeKnownBits(U->getOperand(1), Known, Depth + 1, Q); | |||
3153 | // If the input is known to be 0 or 1, the output is 0/-1, which is | |||
3154 | // all sign bits set. | |||
3155 | if ((Known.Zero | 1).isAllOnes()) | |||
3156 | return TyBits; | |||
3157 | ||||
3158 | // If the input is known to be positive (the sign bit is known clear), | |||
3159 | // the output of the NEG has the same number of sign bits as the | |||
3160 | // input. | |||
3161 | if (Known.isNonNegative()) | |||
3162 | return Tmp2; | |||
3163 | ||||
3164 | // Otherwise, we treat this like a SUB. | |||
3165 | } | |||
3166 | ||||
3167 | // Sub can have at most one carry bit. Thus we know that the output | |||
3168 | // is, at worst, one more bit than the inputs. | |||
3169 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3170 | if (Tmp == 1) break; | |||
3171 | return std::min(Tmp, Tmp2) - 1; | |||
3172 | ||||
3173 | case Instruction::Mul: { | |||
3174 | // The output of the Mul can be at most twice the valid bits in the | |||
3175 | // inputs. | |||
3176 | unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3177 | if (SignBitsOp0 == 1) break; | |||
3178 | unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
3179 | if (SignBitsOp1 == 1) break; | |||
3180 | unsigned OutValidBits = | |||
3181 | (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); | |||
3182 | return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; | |||
3183 | } | |||
3184 | ||||
3185 | case Instruction::PHI: { | |||
3186 | const PHINode *PN = cast<PHINode>(U); | |||
3187 | unsigned NumIncomingValues = PN->getNumIncomingValues(); | |||
3188 | // Don't analyze large in-degree PHIs. | |||
3189 | if (NumIncomingValues > 4) break; | |||
3190 | // Unreachable blocks may have zero-operand PHI nodes. | |||
3191 | if (NumIncomingValues == 0) break; | |||
3192 | ||||
3193 | // Take the minimum of all incoming values. This can't infinitely loop | |||
3194 | // because of our depth threshold. | |||
3195 | Query RecQ = Q; | |||
3196 | Tmp = TyBits; | |||
3197 | for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { | |||
3198 | if (Tmp == 1) return Tmp; | |||
3199 | RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); | |||
3200 | Tmp = std::min( | |||
3201 | Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ)); | |||
3202 | } | |||
3203 | return Tmp; | |||
3204 | } | |||
3205 | ||||
3206 | case Instruction::Trunc: | |||
3207 | // FIXME: it's tricky to do anything useful for this, but it is an | |||
3208 | // important case for targets like X86. | |||
3209 | break; | |||
3210 | ||||
3211 | case Instruction::ExtractElement: | |||
3212 | // Look through extract element. At the moment we keep this simple and | |||
3213 | // skip tracking the specific element. But at least we might find | |||
3214 | // information valid for all elements of the vector (for example if vector | |||
3215 | // is sign extended, shifted, etc). | |||
3216 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3217 | ||||
3218 | case Instruction::ShuffleVector: { | |||
3219 | // Collect the minimum number of sign bits that are shared by every vector | |||
3220 | // element referenced by the shuffle. | |||
3221 | auto *Shuf = dyn_cast<ShuffleVectorInst>(U); | |||
3222 | if (!Shuf) { | |||
3223 | // FIXME: Add support for shufflevector constant expressions. | |||
3224 | return 1; | |||
3225 | } | |||
3226 | APInt DemandedLHS, DemandedRHS; | |||
3227 | // For undef elements, we don't know anything about the common state of | |||
3228 | // the shuffle result. | |||
3229 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) | |||
3230 | return 1; | |||
3231 | Tmp = std::numeric_limits<unsigned>::max(); | |||
3232 | if (!!DemandedLHS) { | |||
3233 | const Value *LHS = Shuf->getOperand(0); | |||
3234 | Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q); | |||
3235 | } | |||
3236 | // If we don't know anything, early out and try computeKnownBits | |||
3237 | // fall-back. | |||
3238 | if (Tmp == 1) | |||
3239 | break; | |||
3240 | if (!!DemandedRHS) { | |||
3241 | const Value *RHS = Shuf->getOperand(1); | |||
3242 | Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q); | |||
3243 | Tmp = std::min(Tmp, Tmp2); | |||
3244 | } | |||
3245 | // If we don't know anything, early out and try computeKnownBits | |||
3246 | // fall-back. | |||
3247 | if (Tmp == 1) | |||
3248 | break; | |||
3249 | assert(Tmp <= TyBits && "Failed to determine minimum sign bits")(static_cast <bool> (Tmp <= TyBits && "Failed to determine minimum sign bits" ) ? void (0) : __assert_fail ("Tmp <= TyBits && \"Failed to determine minimum sign bits\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3249, __extension__ __PRETTY_FUNCTION__ )); | |||
3250 | return Tmp; | |||
3251 | } | |||
3252 | case Instruction::Call: { | |||
3253 | if (const auto *II = dyn_cast<IntrinsicInst>(U)) { | |||
3254 | switch (II->getIntrinsicID()) { | |||
3255 | default: break; | |||
3256 | case Intrinsic::abs: | |||
3257 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
3258 | if (Tmp == 1) break; | |||
3259 | ||||
3260 | // Absolute value reduces number of sign bits by at most 1. | |||
3261 | return Tmp - 1; | |||
3262 | case Intrinsic::smin: | |||
3263 | case Intrinsic::smax: { | |||
3264 | const APInt *CLow, *CHigh; | |||
3265 | if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh)) | |||
3266 | return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); | |||
3267 | } | |||
3268 | } | |||
3269 | } | |||
3270 | } | |||
3271 | } | |||
3272 | } | |||
3273 | ||||
3274 | // Finally, if we can prove that the top bits of the result are 0's or 1's, | |||
3275 | // use this information. | |||
3276 | ||||
3277 | // If we can examine all elements of a vector constant successfully, we're | |||
3278 | // done (we can't do any better than that). If not, keep trying. | |||
3279 | if (unsigned VecSignBits = | |||
3280 | computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) | |||
3281 | return VecSignBits; | |||
3282 | ||||
3283 | KnownBits Known(TyBits); | |||
3284 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
3285 | ||||
3286 | // If we know that the sign bit is either zero or one, determine the number of | |||
3287 | // identical bits in the top of the input value. | |||
3288 | return std::max(FirstAnswer, Known.countMinSignBits()); | |||
3289 | } | |||
3290 | ||||
3291 | Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, | |||
3292 | const TargetLibraryInfo *TLI) { | |||
3293 | const Function *F = CB.getCalledFunction(); | |||
3294 | if (!F) | |||
3295 | return Intrinsic::not_intrinsic; | |||
3296 | ||||
3297 | if (F->isIntrinsic()) | |||
3298 | return F->getIntrinsicID(); | |||
3299 | ||||
3300 | // We are going to infer semantics of a library function based on mapping it | |||
3301 | // to an LLVM intrinsic. Check that the library function is available from | |||
3302 | // this callbase and in this environment. | |||
3303 | LibFunc Func; | |||
3304 | if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) || | |||
3305 | !CB.onlyReadsMemory()) | |||
3306 | return Intrinsic::not_intrinsic; | |||
3307 | ||||
3308 | switch (Func) { | |||
3309 | default: | |||
3310 | break; | |||
3311 | case LibFunc_sin: | |||
3312 | case LibFunc_sinf: | |||
3313 | case LibFunc_sinl: | |||
3314 | return Intrinsic::sin; | |||
3315 | case LibFunc_cos: | |||
3316 | case LibFunc_cosf: | |||
3317 | case LibFunc_cosl: | |||
3318 | return Intrinsic::cos; | |||
3319 | case LibFunc_exp: | |||
3320 | case LibFunc_expf: | |||
3321 | case LibFunc_expl: | |||
3322 | return Intrinsic::exp; | |||
3323 | case LibFunc_exp2: | |||
3324 | case LibFunc_exp2f: | |||
3325 | case LibFunc_exp2l: | |||
3326 | return Intrinsic::exp2; | |||
3327 | case LibFunc_log: | |||
3328 | case LibFunc_logf: | |||
3329 | case LibFunc_logl: | |||
3330 | return Intrinsic::log; | |||
3331 | case LibFunc_log10: | |||
3332 | case LibFunc_log10f: | |||
3333 | case LibFunc_log10l: | |||
3334 | return Intrinsic::log10; | |||
3335 | case LibFunc_log2: | |||
3336 | case LibFunc_log2f: | |||
3337 | case LibFunc_log2l: | |||
3338 | return Intrinsic::log2; | |||
3339 | case LibFunc_fabs: | |||
3340 | case LibFunc_fabsf: | |||
3341 | case LibFunc_fabsl: | |||
3342 | return Intrinsic::fabs; | |||
3343 | case LibFunc_fmin: | |||
3344 | case LibFunc_fminf: | |||
3345 | case LibFunc_fminl: | |||
3346 | return Intrinsic::minnum; | |||
3347 | case LibFunc_fmax: | |||
3348 | case LibFunc_fmaxf: | |||
3349 | case LibFunc_fmaxl: | |||
3350 | return Intrinsic::maxnum; | |||
3351 | case LibFunc_copysign: | |||
3352 | case LibFunc_copysignf: | |||
3353 | case LibFunc_copysignl: | |||
3354 | return Intrinsic::copysign; | |||
3355 | case LibFunc_floor: | |||
3356 | case LibFunc_floorf: | |||
3357 | case LibFunc_floorl: | |||
3358 | return Intrinsic::floor; | |||
3359 | case LibFunc_ceil: | |||
3360 | case LibFunc_ceilf: | |||
3361 | case LibFunc_ceill: | |||
3362 | return Intrinsic::ceil; | |||
3363 | case LibFunc_trunc: | |||
3364 | case LibFunc_truncf: | |||
3365 | case LibFunc_truncl: | |||
3366 | return Intrinsic::trunc; | |||
3367 | case LibFunc_rint: | |||
3368 | case LibFunc_rintf: | |||
3369 | case LibFunc_rintl: | |||
3370 | return Intrinsic::rint; | |||
3371 | case LibFunc_nearbyint: | |||
3372 | case LibFunc_nearbyintf: | |||
3373 | case LibFunc_nearbyintl: | |||
3374 | return Intrinsic::nearbyint; | |||
3375 | case LibFunc_round: | |||
3376 | case LibFunc_roundf: | |||
3377 | case LibFunc_roundl: | |||
3378 | return Intrinsic::round; | |||
3379 | case LibFunc_roundeven: | |||
3380 | case LibFunc_roundevenf: | |||
3381 | case LibFunc_roundevenl: | |||
3382 | return Intrinsic::roundeven; | |||
3383 | case LibFunc_pow: | |||
3384 | case LibFunc_powf: | |||
3385 | case LibFunc_powl: | |||
3386 | return Intrinsic::pow; | |||
3387 | case LibFunc_sqrt: | |||
3388 | case LibFunc_sqrtf: | |||
3389 | case LibFunc_sqrtl: | |||
3390 | return Intrinsic::sqrt; | |||
3391 | } | |||
3392 | ||||
3393 | return Intrinsic::not_intrinsic; | |||
3394 | } | |||
3395 | ||||
3396 | /// Return true if we can prove that the specified FP value is never equal to | |||
3397 | /// -0.0. | |||
3398 | /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee | |||
3399 | /// that a value is not -0.0. It only guarantees that -0.0 may be treated | |||
3400 | /// the same as +0.0 in floating-point ops. | |||
3401 | bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, | |||
3402 | unsigned Depth) { | |||
3403 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
3404 | return !CFP->getValueAPF().isNegZero(); | |||
3405 | ||||
3406 | if (Depth == MaxAnalysisRecursionDepth) | |||
3407 | return false; | |||
3408 | ||||
3409 | auto *Op = dyn_cast<Operator>(V); | |||
3410 | if (!Op) | |||
3411 | return false; | |||
3412 | ||||
3413 | // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. | |||
3414 | if (match(Op, m_FAdd(m_Value(), m_PosZeroFP()))) | |||
3415 | return true; | |||
3416 | ||||
3417 | // sitofp and uitofp turn into +0.0 for zero. | |||
3418 | if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) | |||
3419 | return true; | |||
3420 | ||||
3421 | if (auto *Call = dyn_cast<CallInst>(Op)) { | |||
3422 | Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI); | |||
3423 | switch (IID) { | |||
3424 | default: | |||
3425 | break; | |||
3426 | // sqrt(-0.0) = -0.0, no other negative results are possible. | |||
3427 | case Intrinsic::sqrt: | |||
3428 | case Intrinsic::canonicalize: | |||
3429 | return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); | |||
3430 | case Intrinsic::experimental_constrained_sqrt: { | |||
3431 | // NOTE: This rounding mode restriction may be too strict. | |||
3432 | const auto *CI = cast<ConstrainedFPIntrinsic>(Call); | |||
3433 | if (CI->getRoundingMode() == RoundingMode::NearestTiesToEven) | |||
3434 | return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); | |||
3435 | else | |||
3436 | return false; | |||
3437 | } | |||
3438 | // fabs(x) != -0.0 | |||
3439 | case Intrinsic::fabs: | |||
3440 | return true; | |||
3441 | // sitofp and uitofp turn into +0.0 for zero. | |||
3442 | case Intrinsic::experimental_constrained_sitofp: | |||
3443 | case Intrinsic::experimental_constrained_uitofp: | |||
3444 | return true; | |||
3445 | } | |||
3446 | } | |||
3447 | ||||
3448 | return false; | |||
3449 | } | |||
3450 | ||||
3451 | /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a | |||
3452 | /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign | |||
3453 | /// bit despite comparing equal. | |||
3454 | static bool cannotBeOrderedLessThanZeroImpl(const Value *V, | |||
3455 | const TargetLibraryInfo *TLI, | |||
3456 | bool SignBitOnly, | |||
3457 | unsigned Depth) { | |||
3458 | // TODO: This function does not do the right thing when SignBitOnly is true | |||
3459 | // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform | |||
3460 | // which flips the sign bits of NaNs. See | |||
3461 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | |||
3462 | ||||
3463 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { | |||
3464 | return !CFP->getValueAPF().isNegative() || | |||
3465 | (!SignBitOnly && CFP->getValueAPF().isZero()); | |||
3466 | } | |||
3467 | ||||
3468 | // Handle vector of constants. | |||
3469 | if (auto *CV = dyn_cast<Constant>(V)) { | |||
3470 | if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) { | |||
3471 | unsigned NumElts = CVFVTy->getNumElements(); | |||
3472 | for (unsigned i = 0; i != NumElts; ++i) { | |||
3473 | auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i)); | |||
3474 | if (!CFP) | |||
3475 | return false; | |||
3476 | if (CFP->getValueAPF().isNegative() && | |||
3477 | (SignBitOnly || !CFP->getValueAPF().isZero())) | |||
3478 | return false; | |||
3479 | } | |||
3480 | ||||
3481 | // All non-negative ConstantFPs. | |||
3482 | return true; | |||
3483 | } | |||
3484 | } | |||
3485 | ||||
3486 | if (Depth == MaxAnalysisRecursionDepth) | |||
3487 | return false; | |||
3488 | ||||
3489 | const Operator *I = dyn_cast<Operator>(V); | |||
3490 | if (!I) | |||
3491 | return false; | |||
3492 | ||||
3493 | switch (I->getOpcode()) { | |||
3494 | default: | |||
3495 | break; | |||
3496 | // Unsigned integers are always nonnegative. | |||
3497 | case Instruction::UIToFP: | |||
3498 | return true; | |||
3499 | case Instruction::FMul: | |||
3500 | case Instruction::FDiv: | |||
3501 | // X * X is always non-negative or a NaN. | |||
3502 | // X / X is always exactly 1.0 or a NaN. | |||
3503 | if (I->getOperand(0) == I->getOperand(1) && | |||
3504 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) | |||
3505 | return true; | |||
3506 | ||||
3507 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
3508 | case Instruction::FAdd: | |||
3509 | case Instruction::FRem: | |||
3510 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3511 | Depth + 1) && | |||
3512 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
3513 | Depth + 1); | |||
3514 | case Instruction::Select: | |||
3515 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
3516 | Depth + 1) && | |||
3517 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | |||
3518 | Depth + 1); | |||
3519 | case Instruction::FPExt: | |||
3520 | case Instruction::FPTrunc: | |||
3521 | // Widening/narrowing never change sign. | |||
3522 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3523 | Depth + 1); | |||
3524 | case Instruction::ExtractElement: | |||
3525 | // Look through extract element. At the moment we keep this simple and skip | |||
3526 | // tracking the specific element. But at least we might find information | |||
3527 | // valid for all elements of the vector. | |||
3528 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3529 | Depth + 1); | |||
3530 | case Instruction::Call: | |||
3531 | const auto *CI = cast<CallInst>(I); | |||
3532 | Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI); | |||
3533 | switch (IID) { | |||
3534 | default: | |||
3535 | break; | |||
3536 | case Intrinsic::maxnum: { | |||
3537 | Value *V0 = I->getOperand(0), *V1 = I->getOperand(1); | |||
3538 | auto isPositiveNum = [&](Value *V) { | |||
3539 | if (SignBitOnly) { | |||
3540 | // With SignBitOnly, this is tricky because the result of | |||
3541 | // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is | |||
3542 | // a constant strictly greater than 0.0. | |||
3543 | const APFloat *C; | |||
3544 | return match(V, m_APFloat(C)) && | |||
3545 | *C > APFloat::getZero(C->getSemantics()); | |||
3546 | } | |||
3547 | ||||
3548 | // -0.0 compares equal to 0.0, so if this operand is at least -0.0, | |||
3549 | // maxnum can't be ordered-less-than-zero. | |||
3550 | return isKnownNeverNaN(V, TLI) && | |||
3551 | cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1); | |||
3552 | }; | |||
3553 | ||||
3554 | // TODO: This could be improved. We could also check that neither operand | |||
3555 | // has its sign bit set (and at least 1 is not-NAN?). | |||
3556 | return isPositiveNum(V0) || isPositiveNum(V1); | |||
3557 | } | |||
3558 | ||||
3559 | case Intrinsic::maximum: | |||
3560 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3561 | Depth + 1) || | |||
3562 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
3563 | Depth + 1); | |||
3564 | case Intrinsic::minnum: | |||
3565 | case Intrinsic::minimum: | |||
3566 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3567 | Depth + 1) && | |||
3568 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
3569 | Depth + 1); | |||
3570 | case Intrinsic::exp: | |||
3571 | case Intrinsic::exp2: | |||
3572 | case Intrinsic::fabs: | |||
3573 | return true; | |||
3574 | ||||
3575 | case Intrinsic::sqrt: | |||
3576 | // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0. | |||
3577 | if (!SignBitOnly) | |||
3578 | return true; | |||
3579 | return CI->hasNoNaNs() && (CI->hasNoSignedZeros() || | |||
3580 | CannotBeNegativeZero(CI->getOperand(0), TLI)); | |||
3581 | ||||
3582 | case Intrinsic::powi: | |||
3583 | if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) { | |||
3584 | // powi(x,n) is non-negative if n is even. | |||
3585 | if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0) | |||
3586 | return true; | |||
3587 | } | |||
3588 | // TODO: This is not correct. Given that exp is an integer, here are the | |||
3589 | // ways that pow can return a negative value: | |||
3590 | // | |||
3591 | // pow(x, exp) --> negative if exp is odd and x is negative. | |||
3592 | // pow(-0, exp) --> -inf if exp is negative odd. | |||
3593 | // pow(-0, exp) --> -0 if exp is positive odd. | |||
3594 | // pow(-inf, exp) --> -0 if exp is negative odd. | |||
3595 | // pow(-inf, exp) --> -inf if exp is positive odd. | |||
3596 | // | |||
3597 | // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN, | |||
3598 | // but we must return false if x == -0. Unfortunately we do not currently | |||
3599 | // have a way of expressing this constraint. See details in | |||
3600 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | |||
3601 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
3602 | Depth + 1); | |||
3603 | ||||
3604 | case Intrinsic::fma: | |||
3605 | case Intrinsic::fmuladd: | |||
3606 | // x*x+y is non-negative if y is non-negative. | |||
3607 | return I->getOperand(0) == I->getOperand(1) && | |||
3608 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) && | |||
3609 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | |||
3610 | Depth + 1); | |||
3611 | } | |||
3612 | break; | |||
3613 | } | |||
3614 | return false; | |||
3615 | } | |||
3616 | ||||
3617 | bool llvm::CannotBeOrderedLessThanZero(const Value *V, | |||
3618 | const TargetLibraryInfo *TLI) { | |||
3619 | return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0); | |||
3620 | } | |||
3621 | ||||
3622 | bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) { | |||
3623 | return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0); | |||
3624 | } | |||
3625 | ||||
3626 | bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI, | |||
3627 | unsigned Depth) { | |||
3628 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type")(static_cast <bool> (V->getType()->isFPOrFPVectorTy () && "Querying for Inf on non-FP type") ? void (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for Inf on non-FP type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3628, __extension__ __PRETTY_FUNCTION__ )); | |||
3629 | ||||
3630 | // If we're told that infinities won't happen, assume they won't. | |||
3631 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | |||
3632 | if (FPMathOp->hasNoInfs()) | |||
3633 | return true; | |||
3634 | ||||
3635 | // Handle scalar constants. | |||
3636 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
3637 | return !CFP->isInfinity(); | |||
3638 | ||||
3639 | if (Depth == MaxAnalysisRecursionDepth) | |||
3640 | return false; | |||
3641 | ||||
3642 | if (auto *Inst = dyn_cast<Instruction>(V)) { | |||
3643 | switch (Inst->getOpcode()) { | |||
3644 | case Instruction::Select: { | |||
3645 | return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) && | |||
3646 | isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1); | |||
3647 | } | |||
3648 | case Instruction::SIToFP: | |||
3649 | case Instruction::UIToFP: { | |||
3650 | // Get width of largest magnitude integer (remove a bit if signed). | |||
3651 | // This still works for a signed minimum value because the largest FP | |||
3652 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). | |||
3653 | int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits(); | |||
3654 | if (Inst->getOpcode() == Instruction::SIToFP) | |||
3655 | --IntSize; | |||
3656 | ||||
3657 | // If the exponent of the largest finite FP value can hold the largest | |||
3658 | // integer, the result of the cast must be finite. | |||
3659 | Type *FPTy = Inst->getType()->getScalarType(); | |||
3660 | return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize; | |||
3661 | } | |||
3662 | default: | |||
3663 | break; | |||
3664 | } | |||
3665 | } | |||
3666 | ||||
3667 | // try to handle fixed width vector constants | |||
3668 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
3669 | if (VFVTy && isa<Constant>(V)) { | |||
3670 | // For vectors, verify that each element is not infinity. | |||
3671 | unsigned NumElts = VFVTy->getNumElements(); | |||
3672 | for (unsigned i = 0; i != NumElts; ++i) { | |||
3673 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | |||
3674 | if (!Elt) | |||
3675 | return false; | |||
3676 | if (isa<UndefValue>(Elt)) | |||
3677 | continue; | |||
3678 | auto *CElt = dyn_cast<ConstantFP>(Elt); | |||
3679 | if (!CElt || CElt->isInfinity()) | |||
3680 | return false; | |||
3681 | } | |||
3682 | // All elements were confirmed non-infinity or undefined. | |||
3683 | return true; | |||
3684 | } | |||
3685 | ||||
3686 | // was not able to prove that V never contains infinity | |||
3687 | return false; | |||
3688 | } | |||
3689 | ||||
3690 | bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI, | |||
3691 | unsigned Depth) { | |||
3692 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type")(static_cast <bool> (V->getType()->isFPOrFPVectorTy () && "Querying for NaN on non-FP type") ? void (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for NaN on non-FP type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3692, __extension__ __PRETTY_FUNCTION__ )); | |||
3693 | ||||
3694 | // If we're told that NaNs won't happen, assume they won't. | |||
3695 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | |||
3696 | if (FPMathOp->hasNoNaNs()) | |||
3697 | return true; | |||
3698 | ||||
3699 | // Handle scalar constants. | |||
3700 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
3701 | return !CFP->isNaN(); | |||
3702 | ||||
3703 | if (Depth == MaxAnalysisRecursionDepth) | |||
3704 | return false; | |||
3705 | ||||
3706 | if (auto *Inst = dyn_cast<Instruction>(V)) { | |||
3707 | switch (Inst->getOpcode()) { | |||
3708 | case Instruction::FAdd: | |||
3709 | case Instruction::FSub: | |||
3710 | // Adding positive and negative infinity produces NaN. | |||
3711 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | |||
3712 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
3713 | (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) || | |||
3714 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1)); | |||
3715 | ||||
3716 | case Instruction::FMul: | |||
3717 | // Zero multiplied with infinity produces NaN. | |||
3718 | // FIXME: If neither side can be zero fmul never produces NaN. | |||
3719 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | |||
3720 | isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && | |||
3721 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
3722 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); | |||
3723 | ||||
3724 | case Instruction::FDiv: | |||
3725 | case Instruction::FRem: | |||
3726 | // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN. | |||
3727 | return false; | |||
3728 | ||||
3729 | case Instruction::Select: { | |||
3730 | return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
3731 | isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1); | |||
3732 | } | |||
3733 | case Instruction::SIToFP: | |||
3734 | case Instruction::UIToFP: | |||
3735 | return true; | |||
3736 | case Instruction::FPTrunc: | |||
3737 | case Instruction::FPExt: | |||
3738 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1); | |||
3739 | default: | |||
3740 | break; | |||
3741 | } | |||
3742 | } | |||
3743 | ||||
3744 | if (const auto *II = dyn_cast<IntrinsicInst>(V)) { | |||
3745 | switch (II->getIntrinsicID()) { | |||
3746 | case Intrinsic::canonicalize: | |||
3747 | case Intrinsic::fabs: | |||
3748 | case Intrinsic::copysign: | |||
3749 | case Intrinsic::exp: | |||
3750 | case Intrinsic::exp2: | |||
3751 | case Intrinsic::floor: | |||
3752 | case Intrinsic::ceil: | |||
3753 | case Intrinsic::trunc: | |||
3754 | case Intrinsic::rint: | |||
3755 | case Intrinsic::nearbyint: | |||
3756 | case Intrinsic::round: | |||
3757 | case Intrinsic::roundeven: | |||
3758 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1); | |||
3759 | case Intrinsic::sqrt: | |||
3760 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) && | |||
3761 | CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI); | |||
3762 | case Intrinsic::minnum: | |||
3763 | case Intrinsic::maxnum: | |||
3764 | // If either operand is not NaN, the result is not NaN. | |||
3765 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) || | |||
3766 | isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1); | |||
3767 | default: | |||
3768 | return false; | |||
3769 | } | |||
3770 | } | |||
3771 | ||||
3772 | // Try to handle fixed width vector constants | |||
3773 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
3774 | if (VFVTy && isa<Constant>(V)) { | |||
3775 | // For vectors, verify that each element is not NaN. | |||
3776 | unsigned NumElts = VFVTy->getNumElements(); | |||
3777 | for (unsigned i = 0; i != NumElts; ++i) { | |||
3778 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | |||
3779 | if (!Elt) | |||
3780 | return false; | |||
3781 | if (isa<UndefValue>(Elt)) | |||
3782 | continue; | |||
3783 | auto *CElt = dyn_cast<ConstantFP>(Elt); | |||
3784 | if (!CElt || CElt->isNaN()) | |||
3785 | return false; | |||
3786 | } | |||
3787 | // All elements were confirmed not-NaN or undefined. | |||
3788 | return true; | |||
3789 | } | |||
3790 | ||||
3791 | // Was not able to prove that V never contains NaN | |||
3792 | return false; | |||
3793 | } | |||
3794 | ||||
3795 | Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { | |||
3796 | ||||
3797 | // All byte-wide stores are splatable, even of arbitrary variables. | |||
3798 | if (V->getType()->isIntegerTy(8)) | |||
3799 | return V; | |||
3800 | ||||
3801 | LLVMContext &Ctx = V->getContext(); | |||
3802 | ||||
3803 | // Undef don't care. | |||
3804 | auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx)); | |||
3805 | if (isa<UndefValue>(V)) | |||
3806 | return UndefInt8; | |||
3807 | ||||
3808 | // Return Undef for zero-sized type. | |||
3809 | if (!DL.getTypeStoreSize(V->getType()).isNonZero()) | |||
3810 | return UndefInt8; | |||
3811 | ||||
3812 | Constant *C = dyn_cast<Constant>(V); | |||
3813 | if (!C) { | |||
3814 | // Conceptually, we could handle things like: | |||
3815 | // %a = zext i8 %X to i16 | |||
3816 | // %b = shl i16 %a, 8 | |||
3817 | // %c = or i16 %a, %b | |||
3818 | // but until there is an example that actually needs this, it doesn't seem | |||
3819 | // worth worrying about. | |||
3820 | return nullptr; | |||
3821 | } | |||
3822 | ||||
3823 | // Handle 'null' ConstantArrayZero etc. | |||
3824 | if (C->isNullValue()) | |||
3825 | return Constant::getNullValue(Type::getInt8Ty(Ctx)); | |||
3826 | ||||
3827 | // Constant floating-point values can be handled as integer values if the | |||
3828 | // corresponding integer value is "byteable". An important case is 0.0. | |||
3829 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { | |||
3830 | Type *Ty = nullptr; | |||
3831 | if (CFP->getType()->isHalfTy()) | |||
3832 | Ty = Type::getInt16Ty(Ctx); | |||
3833 | else if (CFP->getType()->isFloatTy()) | |||
3834 | Ty = Type::getInt32Ty(Ctx); | |||
3835 | else if (CFP->getType()->isDoubleTy()) | |||
3836 | Ty = Type::getInt64Ty(Ctx); | |||
3837 | // Don't handle long double formats, which have strange constraints. | |||
3838 | return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL) | |||
3839 | : nullptr; | |||
3840 | } | |||
3841 | ||||
3842 | // We can handle constant integers that are multiple of 8 bits. | |||
3843 | if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { | |||
3844 | if (CI->getBitWidth() % 8 == 0) { | |||
3845 | assert(CI->getBitWidth() > 8 && "8 bits should be handled above!")(static_cast <bool> (CI->getBitWidth() > 8 && "8 bits should be handled above!") ? void (0) : __assert_fail ("CI->getBitWidth() > 8 && \"8 bits should be handled above!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3845, __extension__ __PRETTY_FUNCTION__ )); | |||
3846 | if (!CI->getValue().isSplat(8)) | |||
3847 | return nullptr; | |||
3848 | return ConstantInt::get(Ctx, CI->getValue().trunc(8)); | |||
3849 | } | |||
3850 | } | |||
3851 | ||||
3852 | if (auto *CE = dyn_cast<ConstantExpr>(C)) { | |||
3853 | if (CE->getOpcode() == Instruction::IntToPtr) { | |||
3854 | if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) { | |||
3855 | unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace()); | |||
3856 | return isBytewiseValue( | |||
3857 | ConstantExpr::getIntegerCast(CE->getOperand(0), | |||
3858 | Type::getIntNTy(Ctx, BitWidth), false), | |||
3859 | DL); | |||
3860 | } | |||
3861 | } | |||
3862 | } | |||
3863 | ||||
3864 | auto Merge = [&](Value *LHS, Value *RHS) -> Value * { | |||
3865 | if (LHS == RHS) | |||
3866 | return LHS; | |||
3867 | if (!LHS || !RHS) | |||
3868 | return nullptr; | |||
3869 | if (LHS == UndefInt8) | |||
3870 | return RHS; | |||
3871 | if (RHS == UndefInt8) | |||
3872 | return LHS; | |||
3873 | return nullptr; | |||
3874 | }; | |||
3875 | ||||
3876 | if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) { | |||
3877 | Value *Val = UndefInt8; | |||
3878 | for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) | |||
3879 | if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL)))) | |||
3880 | return nullptr; | |||
3881 | return Val; | |||
3882 | } | |||
3883 | ||||
3884 | if (isa<ConstantAggregate>(C)) { | |||
3885 | Value *Val = UndefInt8; | |||
3886 | for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) | |||
3887 | if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL)))) | |||
3888 | return nullptr; | |||
3889 | return Val; | |||
3890 | } | |||
3891 | ||||
3892 | // Don't try to handle the handful of other constants. | |||
3893 | return nullptr; | |||
3894 | } | |||
3895 | ||||
3896 | // This is the recursive version of BuildSubAggregate. It takes a few different | |||
3897 | // arguments. Idxs is the index within the nested struct From that we are | |||
3898 | // looking at now (which is of type IndexedType). IdxSkip is the number of | |||
3899 | // indices from Idxs that should be left out when inserting into the resulting | |||
3900 | // struct. To is the result struct built so far, new insertvalue instructions | |||
3901 | // build on that. | |||
3902 | static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, | |||
3903 | SmallVectorImpl<unsigned> &Idxs, | |||
3904 | unsigned IdxSkip, | |||
3905 | Instruction *InsertBefore) { | |||
3906 | StructType *STy = dyn_cast<StructType>(IndexedType); | |||
3907 | if (STy) { | |||
3908 | // Save the original To argument so we can modify it | |||
3909 | Value *OrigTo = To; | |||
3910 | // General case, the type indexed by Idxs is a struct | |||
3911 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { | |||
3912 | // Process each struct element recursively | |||
3913 | Idxs.push_back(i); | |||
3914 | Value *PrevTo = To; | |||
3915 | To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, | |||
3916 | InsertBefore); | |||
3917 | Idxs.pop_back(); | |||
3918 | if (!To) { | |||
3919 | // Couldn't find any inserted value for this index? Cleanup | |||
3920 | while (PrevTo != OrigTo) { | |||
3921 | InsertValueInst* Del = cast<InsertValueInst>(PrevTo); | |||
3922 | PrevTo = Del->getAggregateOperand(); | |||
3923 | Del->eraseFromParent(); | |||
3924 | } | |||
3925 | // Stop processing elements | |||
3926 | break; | |||
3927 | } | |||
3928 | } | |||
3929 | // If we successfully found a value for each of our subaggregates | |||
3930 | if (To) | |||
3931 | return To; | |||
3932 | } | |||
3933 | // Base case, the type indexed by SourceIdxs is not a struct, or not all of | |||
3934 | // the struct's elements had a value that was inserted directly. In the latter | |||
3935 | // case, perhaps we can't determine each of the subelements individually, but | |||
3936 | // we might be able to find the complete struct somewhere. | |||
3937 | ||||
3938 | // Find the value that is at that particular spot | |||
3939 | Value *V = FindInsertedValue(From, Idxs); | |||
3940 | ||||
3941 | if (!V) | |||
3942 | return nullptr; | |||
3943 | ||||
3944 | // Insert the value in the new (sub) aggregate | |||
3945 | return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), | |||
3946 | "tmp", InsertBefore); | |||
3947 | } | |||
3948 | ||||
3949 | // This helper takes a nested struct and extracts a part of it (which is again a | |||
3950 | // struct) into a new value. For example, given the struct: | |||
3951 | // { a, { b, { c, d }, e } } | |||
3952 | // and the indices "1, 1" this returns | |||
3953 | // { c, d }. | |||
3954 | // | |||
3955 | // It does this by inserting an insertvalue for each element in the resulting | |||
3956 | // struct, as opposed to just inserting a single struct. This will only work if | |||
3957 | // each of the elements of the substruct are known (ie, inserted into From by an | |||
3958 | // insertvalue instruction somewhere). | |||
3959 | // | |||
3960 | // All inserted insertvalue instructions are inserted before InsertBefore | |||
3961 | static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, | |||
3962 | Instruction *InsertBefore) { | |||
3963 | assert(InsertBefore && "Must have someplace to insert!")(static_cast <bool> (InsertBefore && "Must have someplace to insert!" ) ? void (0) : __assert_fail ("InsertBefore && \"Must have someplace to insert!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3963, __extension__ __PRETTY_FUNCTION__ )); | |||
3964 | Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), | |||
3965 | idx_range); | |||
3966 | Value *To = UndefValue::get(IndexedType); | |||
3967 | SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); | |||
3968 | unsigned IdxSkip = Idxs.size(); | |||
3969 | ||||
3970 | return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); | |||
3971 | } | |||
3972 | ||||
3973 | /// Given an aggregate and a sequence of indices, see if the scalar value | |||
3974 | /// indexed is already around as a register, for example if it was inserted | |||
3975 | /// directly into the aggregate. | |||
3976 | /// | |||
3977 | /// If InsertBefore is not null, this function will duplicate (modified) | |||
3978 | /// insertvalues when a part of a nested struct is extracted. | |||
3979 | Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, | |||
3980 | Instruction *InsertBefore) { | |||
3981 | // Nothing to index? Just return V then (this is useful at the end of our | |||
3982 | // recursion). | |||
3983 | if (idx_range.empty()) | |||
3984 | return V; | |||
3985 | // We have indices, so V should have an indexable type. | |||
3986 | assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&(static_cast <bool> ((V->getType()->isStructTy() || V->getType()->isArrayTy()) && "Not looking at a struct or array?" ) ? void (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3987, __extension__ __PRETTY_FUNCTION__ )) | |||
3987 | "Not looking at a struct or array?")(static_cast <bool> ((V->getType()->isStructTy() || V->getType()->isArrayTy()) && "Not looking at a struct or array?" ) ? void (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3987, __extension__ __PRETTY_FUNCTION__ )); | |||
3988 | assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&(static_cast <bool> (ExtractValueInst::getIndexedType(V ->getType(), idx_range) && "Invalid indices for type?" ) ? void (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3989, __extension__ __PRETTY_FUNCTION__ )) | |||
3989 | "Invalid indices for type?")(static_cast <bool> (ExtractValueInst::getIndexedType(V ->getType(), idx_range) && "Invalid indices for type?" ) ? void (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3989, __extension__ __PRETTY_FUNCTION__ )); | |||
3990 | ||||
3991 | if (Constant *C = dyn_cast<Constant>(V)) { | |||
3992 | C = C->getAggregateElement(idx_range[0]); | |||
3993 | if (!C) return nullptr; | |||
3994 | return FindInsertedValue(C, idx_range.slice(1), InsertBefore); | |||
3995 | } | |||
3996 | ||||
3997 | if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { | |||
3998 | // Loop the indices for the insertvalue instruction in parallel with the | |||
3999 | // requested indices | |||
4000 | const unsigned *req_idx = idx_range.begin(); | |||
4001 | for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); | |||
4002 | i != e; ++i, ++req_idx) { | |||
4003 | if (req_idx == idx_range.end()) { | |||
4004 | // We can't handle this without inserting insertvalues | |||
4005 | if (!InsertBefore) | |||
4006 | return nullptr; | |||
4007 | ||||
4008 | // The requested index identifies a part of a nested aggregate. Handle | |||
4009 | // this specially. For example, | |||
4010 | // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 | |||
4011 | // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 | |||
4012 | // %C = extractvalue {i32, { i32, i32 } } %B, 1 | |||
4013 | // This can be changed into | |||
4014 | // %A = insertvalue {i32, i32 } undef, i32 10, 0 | |||
4015 | // %C = insertvalue {i32, i32 } %A, i32 11, 1 | |||
4016 | // which allows the unused 0,0 element from the nested struct to be | |||
4017 | // removed. | |||
4018 | return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), | |||
4019 | InsertBefore); | |||
4020 | } | |||
4021 | ||||
4022 | // This insert value inserts something else than what we are looking for. | |||
4023 | // See if the (aggregate) value inserted into has the value we are | |||
4024 | // looking for, then. | |||
4025 | if (*req_idx != *i) | |||
4026 | return FindInsertedValue(I->getAggregateOperand(), idx_range, | |||
4027 | InsertBefore); | |||
4028 | } | |||
4029 | // If we end up here, the indices of the insertvalue match with those | |||
4030 | // requested (though possibly only partially). Now we recursively look at | |||
4031 | // the inserted value, passing any remaining indices. | |||
4032 | return FindInsertedValue(I->getInsertedValueOperand(), | |||
4033 | makeArrayRef(req_idx, idx_range.end()), | |||
4034 | InsertBefore); | |||
4035 | } | |||
4036 | ||||
4037 | if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { | |||
4038 | // If we're extracting a value from an aggregate that was extracted from | |||
4039 | // something else, we can extract from that something else directly instead. | |||
4040 | // However, we will need to chain I's indices with the requested indices. | |||
4041 | ||||
4042 | // Calculate the number of indices required | |||
4043 | unsigned size = I->getNumIndices() + idx_range.size(); | |||
4044 | // Allocate some space to put the new indices in | |||
4045 | SmallVector<unsigned, 5> Idxs; | |||
4046 | Idxs.reserve(size); | |||
4047 | // Add indices from the extract value instruction | |||
4048 | Idxs.append(I->idx_begin(), I->idx_end()); | |||
4049 | ||||
4050 | // Add requested indices | |||
4051 | Idxs.append(idx_range.begin(), idx_range.end()); | |||
4052 | ||||
4053 | assert(Idxs.size() == size(static_cast <bool> (Idxs.size() == size && "Number of indices added not correct?" ) ? void (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4054, __extension__ __PRETTY_FUNCTION__ )) | |||
4054 | && "Number of indices added not correct?")(static_cast <bool> (Idxs.size() == size && "Number of indices added not correct?" ) ? void (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4054, __extension__ __PRETTY_FUNCTION__ )); | |||
4055 | ||||
4056 | return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); | |||
4057 | } | |||
4058 | // Otherwise, we don't know (such as, extracting from a function return value | |||
4059 | // or load instruction) | |||
4060 | return nullptr; | |||
4061 | } | |||
4062 | ||||
4063 | bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, | |||
4064 | unsigned CharSize) { | |||
4065 | // Make sure the GEP has exactly three arguments. | |||
4066 | if (GEP->getNumOperands() != 3) | |||
4067 | return false; | |||
4068 | ||||
4069 | // Make sure the index-ee is a pointer to array of \p CharSize integers. | |||
4070 | // CharSize. | |||
4071 | ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); | |||
4072 | if (!AT || !AT->getElementType()->isIntegerTy(CharSize)) | |||
4073 | return false; | |||
4074 | ||||
4075 | // Check to make sure that the first operand of the GEP is an integer and | |||
4076 | // has value 0 so that we are sure we're indexing into the initializer. | |||
4077 | const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); | |||
4078 | if (!FirstIdx || !FirstIdx->isZero()) | |||
4079 | return false; | |||
4080 | ||||
4081 | return true; | |||
4082 | } | |||
4083 | ||||
4084 | bool llvm::getConstantDataArrayInfo(const Value *V, | |||
4085 | ConstantDataArraySlice &Slice, | |||
4086 | unsigned ElementSize, uint64_t Offset) { | |||
4087 | assert(V)(static_cast <bool> (V) ? void (0) : __assert_fail ("V" , "llvm/lib/Analysis/ValueTracking.cpp", 4087, __extension__ __PRETTY_FUNCTION__ )); | |||
4088 | ||||
4089 | // Look through bitcast instructions and geps. | |||
4090 | V = V->stripPointerCasts(); | |||
4091 | ||||
4092 | // If the value is a GEP instruction or constant expression, treat it as an | |||
4093 | // offset. | |||
4094 | if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { | |||
4095 | // The GEP operator should be based on a pointer to string constant, and is | |||
4096 | // indexing into the string constant. | |||
4097 | if (!isGEPBasedOnPointerToString(GEP, ElementSize)) | |||
4098 | return false; | |||
4099 | ||||
4100 | // If the second index isn't a ConstantInt, then this is a variable index | |||
4101 | // into the array. If this occurs, we can't say anything meaningful about | |||
4102 | // the string. | |||
4103 | uint64_t StartIdx = 0; | |||
4104 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) | |||
4105 | StartIdx = CI->getZExtValue(); | |||
4106 | else | |||
4107 | return false; | |||
4108 | return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize, | |||
4109 | StartIdx + Offset); | |||
4110 | } | |||
4111 | ||||
4112 | // The GEP instruction, constant or instruction, must reference a global | |||
4113 | // variable that is a constant and is initialized. The referenced constant | |||
4114 | // initializer is the array that we'll use for optimization. | |||
4115 | const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); | |||
4116 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) | |||
4117 | return false; | |||
4118 | ||||
4119 | const ConstantDataArray *Array; | |||
4120 | ArrayType *ArrayTy; | |||
4121 | if (GV->getInitializer()->isNullValue()) { | |||
4122 | Type *GVTy = GV->getValueType(); | |||
4123 | if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) { | |||
4124 | // A zeroinitializer for the array; there is no ConstantDataArray. | |||
4125 | Array = nullptr; | |||
4126 | } else { | |||
4127 | const DataLayout &DL = GV->getParent()->getDataLayout(); | |||
4128 | uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize(); | |||
4129 | uint64_t Length = SizeInBytes / (ElementSize / 8); | |||
4130 | if (Length <= Offset) | |||
4131 | return false; | |||
4132 | ||||
4133 | Slice.Array = nullptr; | |||
4134 | Slice.Offset = 0; | |||
4135 | Slice.Length = Length - Offset; | |||
4136 | return true; | |||
4137 | } | |||
4138 | } else { | |||
4139 | // This must be a ConstantDataArray. | |||
4140 | Array = dyn_cast<ConstantDataArray>(GV->getInitializer()); | |||
4141 | if (!Array) | |||
4142 | return false; | |||
4143 | ArrayTy = Array->getType(); | |||
4144 | } | |||
4145 | if (!ArrayTy->getElementType()->isIntegerTy(ElementSize)) | |||
4146 | return false; | |||
4147 | ||||
4148 | uint64_t NumElts = ArrayTy->getArrayNumElements(); | |||
4149 | if (Offset > NumElts) | |||
4150 | return false; | |||
4151 | ||||
4152 | Slice.Array = Array; | |||
4153 | Slice.Offset = Offset; | |||
4154 | Slice.Length = NumElts - Offset; | |||
4155 | return true; | |||
4156 | } | |||
4157 | ||||
4158 | /// This function computes the length of a null-terminated C string pointed to | |||
4159 | /// by V. If successful, it returns true and returns the string in Str. | |||
4160 | /// If unsuccessful, it returns false. | |||
4161 | bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, | |||
4162 | uint64_t Offset, bool TrimAtNul) { | |||
4163 | ConstantDataArraySlice Slice; | |||
4164 | if (!getConstantDataArrayInfo(V, Slice, 8, Offset)) | |||
4165 | return false; | |||
4166 | ||||
4167 | if (Slice.Array == nullptr) { | |||
4168 | if (TrimAtNul) { | |||
4169 | Str = StringRef(); | |||
4170 | return true; | |||
4171 | } | |||
4172 | if (Slice.Length == 1) { | |||
4173 | Str = StringRef("", 1); | |||
4174 | return true; | |||
4175 | } | |||
4176 | // We cannot instantiate a StringRef as we do not have an appropriate string | |||
4177 | // of 0s at hand. | |||
4178 | return false; | |||
4179 | } | |||
4180 | ||||
4181 | // Start out with the entire array in the StringRef. | |||
4182 | Str = Slice.Array->getAsString(); | |||
4183 | // Skip over 'offset' bytes. | |||
4184 | Str = Str.substr(Slice.Offset); | |||
4185 | ||||
4186 | if (TrimAtNul) { | |||
4187 | // Trim off the \0 and anything after it. If the array is not nul | |||
4188 | // terminated, we just return the whole end of string. The client may know | |||
4189 | // some other way that the string is length-bound. | |||
4190 | Str = Str.substr(0, Str.find('\0')); | |||
4191 | } | |||
4192 | return true; | |||
4193 | } | |||
4194 | ||||
4195 | // These next two are very similar to the above, but also look through PHI | |||
4196 | // nodes. | |||
4197 | // TODO: See if we can integrate these two together. | |||
4198 | ||||
4199 | /// If we can compute the length of the string pointed to by | |||
4200 | /// the specified pointer, return 'len+1'. If we can't, return 0. | |||
4201 | static uint64_t GetStringLengthH(const Value *V, | |||
4202 | SmallPtrSetImpl<const PHINode*> &PHIs, | |||
4203 | unsigned CharSize) { | |||
4204 | // Look through noop bitcast instructions. | |||
4205 | V = V->stripPointerCasts(); | |||
4206 | ||||
4207 | // If this is a PHI node, there are two cases: either we have already seen it | |||
4208 | // or we haven't. | |||
4209 | if (const PHINode *PN = dyn_cast<PHINode>(V)) { | |||
4210 | if (!PHIs.insert(PN).second) | |||
4211 | return ~0ULL; // already in the set. | |||
4212 | ||||
4213 | // If it was new, see if all the input strings are the same length. | |||
4214 | uint64_t LenSoFar = ~0ULL; | |||
4215 | for (Value *IncValue : PN->incoming_values()) { | |||
4216 | uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize); | |||
4217 | if (Len == 0) return 0; // Unknown length -> unknown. | |||
4218 | ||||
4219 | if (Len == ~0ULL) continue; | |||
4220 | ||||
4221 | if (Len != LenSoFar && LenSoFar != ~0ULL) | |||
4222 | return 0; // Disagree -> unknown. | |||
4223 | LenSoFar = Len; | |||
4224 | } | |||
4225 | ||||
4226 | // Success, all agree. | |||
4227 | return LenSoFar; | |||
4228 | } | |||
4229 | ||||
4230 | // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) | |||
4231 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { | |||
4232 | uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize); | |||
4233 | if (Len1 == 0) return 0; | |||
4234 | uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize); | |||
4235 | if (Len2 == 0) return 0; | |||
4236 | if (Len1 == ~0ULL) return Len2; | |||
4237 | if (Len2 == ~0ULL) return Len1; | |||
4238 | if (Len1 != Len2) return 0; | |||
4239 | return Len1; | |||
4240 | } | |||
4241 | ||||
4242 | // Otherwise, see if we can read the string. | |||
4243 | ConstantDataArraySlice Slice; | |||
4244 | if (!getConstantDataArrayInfo(V, Slice, CharSize)) | |||
4245 | return 0; | |||
4246 | ||||
4247 | if (Slice.Array == nullptr) | |||
4248 | return 1; | |||
4249 | ||||
4250 | // Search for nul characters | |||
4251 | unsigned NullIndex = 0; | |||
4252 | for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { | |||
4253 | if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0) | |||
4254 | break; | |||
4255 | } | |||
4256 | ||||
4257 | return NullIndex + 1; | |||
4258 | } | |||
4259 | ||||
4260 | /// If we can compute the length of the string pointed to by | |||
4261 | /// the specified pointer, return 'len+1'. If we can't, return 0. | |||
4262 | uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { | |||
4263 | if (!V->getType()->isPointerTy()) | |||
4264 | return 0; | |||
4265 | ||||
4266 | SmallPtrSet<const PHINode*, 32> PHIs; | |||
4267 | uint64_t Len = GetStringLengthH(V, PHIs, CharSize); | |||
4268 | // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return | |||
4269 | // an empty string as a length. | |||
4270 | return Len == ~0ULL ? 1 : Len; | |||
4271 | } | |||
4272 | ||||
4273 | const Value * | |||
4274 | llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, | |||
4275 | bool MustPreserveNullness) { | |||
4276 | assert(Call &&(static_cast <bool> (Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls" ) ? void (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4277, __extension__ __PRETTY_FUNCTION__ )) | |||
4277 | "getArgumentAliasingToReturnedPointer only works on nonnull calls")(static_cast <bool> (Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls" ) ? void (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4277, __extension__ __PRETTY_FUNCTION__ )); | |||
4278 | if (const Value *RV = Call->getReturnedArgOperand()) | |||
4279 | return RV; | |||
4280 | // This can be used only as a aliasing property. | |||
4281 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | |||
4282 | Call, MustPreserveNullness)) | |||
4283 | return Call->getArgOperand(0); | |||
4284 | return nullptr; | |||
4285 | } | |||
4286 | ||||
4287 | bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | |||
4288 | const CallBase *Call, bool MustPreserveNullness) { | |||
4289 | switch (Call->getIntrinsicID()) { | |||
4290 | case Intrinsic::launder_invariant_group: | |||
4291 | case Intrinsic::strip_invariant_group: | |||
4292 | case Intrinsic::aarch64_irg: | |||
4293 | case Intrinsic::aarch64_tagp: | |||
4294 | return true; | |||
4295 | case Intrinsic::ptrmask: | |||
4296 | return !MustPreserveNullness; | |||
4297 | default: | |||
4298 | return false; | |||
4299 | } | |||
4300 | } | |||
4301 | ||||
4302 | /// \p PN defines a loop-variant pointer to an object. Check if the | |||
4303 | /// previous iteration of the loop was referring to the same object as \p PN. | |||
4304 | static bool isSameUnderlyingObjectInLoop(const PHINode *PN, | |||
4305 | const LoopInfo *LI) { | |||
4306 | // Find the loop-defined value. | |||
4307 | Loop *L = LI->getLoopFor(PN->getParent()); | |||
4308 | if (PN->getNumIncomingValues() != 2) | |||
4309 | return true; | |||
4310 | ||||
4311 | // Find the value from previous iteration. | |||
4312 | auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); | |||
4313 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | |||
4314 | PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); | |||
4315 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | |||
4316 | return true; | |||
4317 | ||||
4318 | // If a new pointer is loaded in the loop, the pointer references a different | |||
4319 | // object in every iteration. E.g.: | |||
4320 | // for (i) | |||
4321 | // int *p = a[i]; | |||
4322 | // ... | |||
4323 | if (auto *Load = dyn_cast<LoadInst>(PrevValue)) | |||
4324 | if (!L->isLoopInvariant(Load->getPointerOperand())) | |||
4325 | return false; | |||
4326 | return true; | |||
4327 | } | |||
4328 | ||||
4329 | const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) { | |||
4330 | if (!V->getType()->isPointerTy()) | |||
4331 | return V; | |||
4332 | for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { | |||
4333 | if (auto *GEP = dyn_cast<GEPOperator>(V)) { | |||
4334 | V = GEP->getPointerOperand(); | |||
4335 | } else if (Operator::getOpcode(V) == Instruction::BitCast || | |||
4336 | Operator::getOpcode(V) == Instruction::AddrSpaceCast) { | |||
4337 | V = cast<Operator>(V)->getOperand(0); | |||
4338 | if (!V->getType()->isPointerTy()) | |||
4339 | return V; | |||
4340 | } else if (auto *GA = dyn_cast<GlobalAlias>(V)) { | |||
4341 | if (GA->isInterposable()) | |||
4342 | return V; | |||
4343 | V = GA->getAliasee(); | |||
4344 | } else { | |||
4345 | if (auto *PHI = dyn_cast<PHINode>(V)) { | |||
4346 | // Look through single-arg phi nodes created by LCSSA. | |||
4347 | if (PHI->getNumIncomingValues() == 1) { | |||
4348 | V = PHI->getIncomingValue(0); | |||
4349 | continue; | |||
4350 | } | |||
4351 | } else if (auto *Call = dyn_cast<CallBase>(V)) { | |||
4352 | // CaptureTracking can know about special capturing properties of some | |||
4353 | // intrinsics like launder.invariant.group, that can't be expressed with | |||
4354 | // the attributes, but have properties like returning aliasing pointer. | |||
4355 | // Because some analysis may assume that nocaptured pointer is not | |||
4356 | // returned from some special intrinsic (because function would have to | |||
4357 | // be marked with returns attribute), it is crucial to use this function | |||
4358 | // because it should be in sync with CaptureTracking. Not using it may | |||
4359 | // cause weird miscompilations where 2 aliasing pointers are assumed to | |||
4360 | // noalias. | |||
4361 | if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) { | |||
4362 | V = RP; | |||
4363 | continue; | |||
4364 | } | |||
4365 | } | |||
4366 | ||||
4367 | return V; | |||
4368 | } | |||
4369 | assert(V->getType()->isPointerTy() && "Unexpected operand type!")(static_cast <bool> (V->getType()->isPointerTy() && "Unexpected operand type!") ? void (0) : __assert_fail ("V->getType()->isPointerTy() && \"Unexpected operand type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4369, __extension__ __PRETTY_FUNCTION__ )); | |||
4370 | } | |||
4371 | return V; | |||
4372 | } | |||
4373 | ||||
4374 | void llvm::getUnderlyingObjects(const Value *V, | |||
4375 | SmallVectorImpl<const Value *> &Objects, | |||
4376 | LoopInfo *LI, unsigned MaxLookup) { | |||
4377 | SmallPtrSet<const Value *, 4> Visited; | |||
4378 | SmallVector<const Value *, 4> Worklist; | |||
4379 | Worklist.push_back(V); | |||
4380 | do { | |||
4381 | const Value *P = Worklist.pop_back_val(); | |||
4382 | P = getUnderlyingObject(P, MaxLookup); | |||
4383 | ||||
4384 | if (!Visited.insert(P).second) | |||
4385 | continue; | |||
4386 | ||||
4387 | if (auto *SI = dyn_cast<SelectInst>(P)) { | |||
4388 | Worklist.push_back(SI->getTrueValue()); | |||
4389 | Worklist.push_back(SI->getFalseValue()); | |||
4390 | continue; | |||
4391 | } | |||
4392 | ||||
4393 | if (auto *PN = dyn_cast<PHINode>(P)) { | |||
4394 | // If this PHI changes the underlying object in every iteration of the | |||
4395 | // loop, don't look through it. Consider: | |||
4396 | // int **A; | |||
4397 | // for (i) { | |||
4398 | // Prev = Curr; // Prev = PHI (Prev_0, Curr) | |||
4399 | // Curr = A[i]; | |||
4400 | // *Prev, *Curr; | |||
4401 | // | |||
4402 | // Prev is tracking Curr one iteration behind so they refer to different | |||
4403 | // underlying objects. | |||
4404 | if (!LI || !LI->isLoopHeader(PN->getParent()) || | |||
4405 | isSameUnderlyingObjectInLoop(PN, LI)) | |||
4406 | append_range(Worklist, PN->incoming_values()); | |||
4407 | continue; | |||
4408 | } | |||
4409 | ||||
4410 | Objects.push_back(P); | |||
4411 | } while (!Worklist.empty()); | |||
4412 | } | |||
4413 | ||||
4414 | /// This is the function that does the work of looking through basic | |||
4415 | /// ptrtoint+arithmetic+inttoptr sequences. | |||
4416 | static const Value *getUnderlyingObjectFromInt(const Value *V) { | |||
4417 | do { | |||
4418 | if (const Operator *U = dyn_cast<Operator>(V)) { | |||
4419 | // If we find a ptrtoint, we can transfer control back to the | |||
4420 | // regular getUnderlyingObjectFromInt. | |||
4421 | if (U->getOpcode() == Instruction::PtrToInt) | |||
4422 | return U->getOperand(0); | |||
4423 | // If we find an add of a constant, a multiplied value, or a phi, it's | |||
4424 | // likely that the other operand will lead us to the base | |||
4425 | // object. We don't have to worry about the case where the | |||
4426 | // object address is somehow being computed by the multiply, | |||
4427 | // because our callers only care when the result is an | |||
4428 | // identifiable object. | |||
4429 | if (U->getOpcode() != Instruction::Add || | |||
4430 | (!isa<ConstantInt>(U->getOperand(1)) && | |||
4431 | Operator::getOpcode(U->getOperand(1)) != Instruction::Mul && | |||
4432 | !isa<PHINode>(U->getOperand(1)))) | |||
4433 | return V; | |||
4434 | V = U->getOperand(0); | |||
4435 | } else { | |||
4436 | return V; | |||
4437 | } | |||
4438 | assert(V->getType()->isIntegerTy() && "Unexpected operand type!")(static_cast <bool> (V->getType()->isIntegerTy() && "Unexpected operand type!") ? void (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Unexpected operand type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4438, __extension__ __PRETTY_FUNCTION__ )); | |||
4439 | } while (true); | |||
4440 | } | |||
4441 | ||||
4442 | /// This is a wrapper around getUnderlyingObjects and adds support for basic | |||
4443 | /// ptrtoint+arithmetic+inttoptr sequences. | |||
4444 | /// It returns false if unidentified object is found in getUnderlyingObjects. | |||
4445 | bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, | |||
4446 | SmallVectorImpl<Value *> &Objects) { | |||
4447 | SmallPtrSet<const Value *, 16> Visited; | |||
4448 | SmallVector<const Value *, 4> Working(1, V); | |||
4449 | do { | |||
4450 | V = Working.pop_back_val(); | |||
4451 | ||||
4452 | SmallVector<const Value *, 4> Objs; | |||
4453 | getUnderlyingObjects(V, Objs); | |||
4454 | ||||
4455 | for (const Value *V : Objs) { | |||
4456 | if (!Visited.insert(V).second) | |||
4457 | continue; | |||
4458 | if (Operator::getOpcode(V) == Instruction::IntToPtr) { | |||
4459 | const Value *O = | |||
4460 | getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); | |||
4461 | if (O->getType()->isPointerTy()) { | |||
4462 | Working.push_back(O); | |||
4463 | continue; | |||
4464 | } | |||
4465 | } | |||
4466 | // If getUnderlyingObjects fails to find an identifiable object, | |||
4467 | // getUnderlyingObjectsForCodeGen also fails for safety. | |||
4468 | if (!isIdentifiedObject(V)) { | |||
4469 | Objects.clear(); | |||
4470 | return false; | |||
4471 | } | |||
4472 | Objects.push_back(const_cast<Value *>(V)); | |||
4473 | } | |||
4474 | } while (!Working.empty()); | |||
4475 | return true; | |||
4476 | } | |||
4477 | ||||
4478 | AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { | |||
4479 | AllocaInst *Result = nullptr; | |||
4480 | SmallPtrSet<Value *, 4> Visited; | |||
4481 | SmallVector<Value *, 4> Worklist; | |||
4482 | ||||
4483 | auto AddWork = [&](Value *V) { | |||
4484 | if (Visited.insert(V).second) | |||
4485 | Worklist.push_back(V); | |||
4486 | }; | |||
4487 | ||||
4488 | AddWork(V); | |||
4489 | do { | |||
4490 | V = Worklist.pop_back_val(); | |||
4491 | assert(Visited.count(V))(static_cast <bool> (Visited.count(V)) ? void (0) : __assert_fail ("Visited.count(V)", "llvm/lib/Analysis/ValueTracking.cpp", 4491 , __extension__ __PRETTY_FUNCTION__)); | |||
4492 | ||||
4493 | if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { | |||
4494 | if (Result && Result != AI) | |||
4495 | return nullptr; | |||
4496 | Result = AI; | |||
4497 | } else if (CastInst *CI = dyn_cast<CastInst>(V)) { | |||
4498 | AddWork(CI->getOperand(0)); | |||
4499 | } else if (PHINode *PN = dyn_cast<PHINode>(V)) { | |||
4500 | for (Value *IncValue : PN->incoming_values()) | |||
4501 | AddWork(IncValue); | |||
4502 | } else if (auto *SI = dyn_cast<SelectInst>(V)) { | |||
4503 | AddWork(SI->getTrueValue()); | |||
4504 | AddWork(SI->getFalseValue()); | |||
4505 | } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { | |||
4506 | if (OffsetZero && !GEP->hasAllZeroIndices()) | |||
4507 | return nullptr; | |||
4508 | AddWork(GEP->getPointerOperand()); | |||
4509 | } else if (CallBase *CB = dyn_cast<CallBase>(V)) { | |||
4510 | Value *Returned = CB->getReturnedArgOperand(); | |||
4511 | if (Returned) | |||
4512 | AddWork(Returned); | |||
4513 | else | |||
4514 | return nullptr; | |||
4515 | } else { | |||
4516 | return nullptr; | |||
4517 | } | |||
4518 | } while (!Worklist.empty()); | |||
4519 | ||||
4520 | return Result; | |||
4521 | } | |||
4522 | ||||
4523 | static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
4524 | const Value *V, bool AllowLifetime, bool AllowDroppable) { | |||
4525 | for (const User *U : V->users()) { | |||
4526 | const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | |||
4527 | if (!II) | |||
4528 | return false; | |||
4529 | ||||
4530 | if (AllowLifetime && II->isLifetimeStartOrEnd()) | |||
4531 | continue; | |||
4532 | ||||
4533 | if (AllowDroppable && II->isDroppable()) | |||
4534 | continue; | |||
4535 | ||||
4536 | return false; | |||
4537 | } | |||
4538 | return true; | |||
4539 | } | |||
4540 | ||||
4541 | bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { | |||
4542 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
4543 | V, /* AllowLifetime */ true, /* AllowDroppable */ false); | |||
4544 | } | |||
4545 | bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { | |||
4546 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
4547 | V, /* AllowLifetime */ true, /* AllowDroppable */ true); | |||
4548 | } | |||
4549 | ||||
4550 | bool llvm::mustSuppressSpeculation(const LoadInst &LI) { | |||
4551 | if (!LI.isUnordered()) | |||
4552 | return true; | |||
4553 | const Function &F = *LI.getFunction(); | |||
4554 | // Speculative load may create a race that did not exist in the source. | |||
4555 | return F.hasFnAttribute(Attribute::SanitizeThread) || | |||
4556 | // Speculative load may load data from dirty regions. | |||
4557 | F.hasFnAttribute(Attribute::SanitizeAddress) || | |||
4558 | F.hasFnAttribute(Attribute::SanitizeHWAddress); | |||
4559 | } | |||
4560 | ||||
4561 | ||||
4562 | bool llvm::isSafeToSpeculativelyExecute(const Value *V, | |||
4563 | const Instruction *CtxI, | |||
4564 | const DominatorTree *DT, | |||
4565 | const TargetLibraryInfo *TLI) { | |||
4566 | const Operator *Inst = dyn_cast<Operator>(V); | |||
4567 | if (!Inst) | |||
4568 | return false; | |||
4569 | ||||
4570 | for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) | |||
4571 | if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) | |||
4572 | if (C->canTrap()) | |||
4573 | return false; | |||
4574 | ||||
4575 | switch (Inst->getOpcode()) { | |||
4576 | default: | |||
4577 | return true; | |||
4578 | case Instruction::UDiv: | |||
4579 | case Instruction::URem: { | |||
4580 | // x / y is undefined if y == 0. | |||
4581 | const APInt *V; | |||
4582 | if (match(Inst->getOperand(1), m_APInt(V))) | |||
4583 | return *V != 0; | |||
4584 | return false; | |||
4585 | } | |||
4586 | case Instruction::SDiv: | |||
4587 | case Instruction::SRem: { | |||
4588 | // x / y is undefined if y == 0 or x == INT_MIN and y == -1 | |||
4589 | const APInt *Numerator, *Denominator; | |||
4590 | if (!match(Inst->getOperand(1), m_APInt(Denominator))) | |||
4591 | return false; | |||
4592 | // We cannot hoist this division if the denominator is 0. | |||
4593 | if (*Denominator == 0) | |||
4594 | return false; | |||
4595 | // It's safe to hoist if the denominator is not 0 or -1. | |||
4596 | if (!Denominator->isAllOnes()) | |||
4597 | return true; | |||
4598 | // At this point we know that the denominator is -1. It is safe to hoist as | |||
4599 | // long we know that the numerator is not INT_MIN. | |||
4600 | if (match(Inst->getOperand(0), m_APInt(Numerator))) | |||
4601 | return !Numerator->isMinSignedValue(); | |||
4602 | // The numerator *might* be MinSignedValue. | |||
4603 | return false; | |||
4604 | } | |||
4605 | case Instruction::Load: { | |||
4606 | const LoadInst *LI = cast<LoadInst>(Inst); | |||
4607 | if (mustSuppressSpeculation(*LI)) | |||
4608 | return false; | |||
4609 | const DataLayout &DL = LI->getModule()->getDataLayout(); | |||
4610 | return isDereferenceableAndAlignedPointer( | |||
4611 | LI->getPointerOperand(), LI->getType(), LI->getAlign(), DL, CtxI, DT, | |||
4612 | TLI); | |||
4613 | } | |||
4614 | case Instruction::Call: { | |||
4615 | auto *CI = cast<const CallInst>(Inst); | |||
4616 | const Function *Callee = CI->getCalledFunction(); | |||
4617 | ||||
4618 | // The called function could have undefined behavior or side-effects, even | |||
4619 | // if marked readnone nounwind. | |||
4620 | return Callee && Callee->isSpeculatable(); | |||
4621 | } | |||
4622 | case Instruction::VAArg: | |||
4623 | case Instruction::Alloca: | |||
4624 | case Instruction::Invoke: | |||
4625 | case Instruction::CallBr: | |||
4626 | case Instruction::PHI: | |||
4627 | case Instruction::Store: | |||
4628 | case Instruction::Ret: | |||
4629 | case Instruction::Br: | |||
4630 | case Instruction::IndirectBr: | |||
4631 | case Instruction::Switch: | |||
4632 | case Instruction::Unreachable: | |||
4633 | case Instruction::Fence: | |||
4634 | case Instruction::AtomicRMW: | |||
4635 | case Instruction::AtomicCmpXchg: | |||
4636 | case Instruction::LandingPad: | |||
4637 | case Instruction::Resume: | |||
4638 | case Instruction::CatchSwitch: | |||
4639 | case Instruction::CatchPad: | |||
4640 | case Instruction::CatchRet: | |||
4641 | case Instruction::CleanupPad: | |||
4642 | case Instruction::CleanupRet: | |||
4643 | return false; // Misc instructions which have effects | |||
4644 | } | |||
4645 | } | |||
4646 | ||||
4647 | bool llvm::mayHaveNonDefUseDependency(const Instruction &I) { | |||
4648 | if (I.mayReadOrWriteMemory()) | |||
4649 | // Memory dependency possible | |||
4650 | return true; | |||
4651 | if (!isSafeToSpeculativelyExecute(&I)) | |||
4652 | // Can't move above a maythrow call or infinite loop. Or if an | |||
4653 | // inalloca alloca, above a stacksave call. | |||
4654 | return true; | |||
4655 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
4656 | // 1) Can't reorder two inf-loop calls, even if readonly | |||
4657 | // 2) Also can't reorder an inf-loop call below a instruction which isn't | |||
4658 | // safe to speculative execute. (Inverse of above) | |||
4659 | return true; | |||
4660 | return false; | |||
4661 | } | |||
4662 | ||||
4663 | /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. | |||
4664 | static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { | |||
4665 | switch (OR) { | |||
4666 | case ConstantRange::OverflowResult::MayOverflow: | |||
4667 | return OverflowResult::MayOverflow; | |||
4668 | case ConstantRange::OverflowResult::AlwaysOverflowsLow: | |||
4669 | return OverflowResult::AlwaysOverflowsLow; | |||
4670 | case ConstantRange::OverflowResult::AlwaysOverflowsHigh: | |||
4671 | return OverflowResult::AlwaysOverflowsHigh; | |||
4672 | case ConstantRange::OverflowResult::NeverOverflows: | |||
4673 | return OverflowResult::NeverOverflows; | |||
4674 | } | |||
4675 | llvm_unreachable("Unknown OverflowResult")::llvm::llvm_unreachable_internal("Unknown OverflowResult", "llvm/lib/Analysis/ValueTracking.cpp" , 4675); | |||
4676 | } | |||
4677 | ||||
4678 | /// Combine constant ranges from computeConstantRange() and computeKnownBits(). | |||
4679 | static ConstantRange computeConstantRangeIncludingKnownBits( | |||
4680 | const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth, | |||
4681 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
4682 | OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) { | |||
4683 | KnownBits Known = computeKnownBits( | |||
4684 | V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo); | |||
4685 | ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned); | |||
4686 | ConstantRange CR2 = computeConstantRange(V, UseInstrInfo); | |||
4687 | ConstantRange::PreferredRangeType RangeType = | |||
4688 | ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; | |||
4689 | return CR1.intersectWith(CR2, RangeType); | |||
4690 | } | |||
4691 | ||||
4692 | OverflowResult llvm::computeOverflowForUnsignedMul( | |||
4693 | const Value *LHS, const Value *RHS, const DataLayout &DL, | |||
4694 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
4695 | bool UseInstrInfo) { | |||
4696 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4697 | nullptr, UseInstrInfo); | |||
4698 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4699 | nullptr, UseInstrInfo); | |||
4700 | ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false); | |||
4701 | ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false); | |||
4702 | return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange)); | |||
4703 | } | |||
4704 | ||||
4705 | OverflowResult | |||
4706 | llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS, | |||
4707 | const DataLayout &DL, AssumptionCache *AC, | |||
4708 | const Instruction *CxtI, | |||
4709 | const DominatorTree *DT, bool UseInstrInfo) { | |||
4710 | // Multiplying n * m significant bits yields a result of n + m significant | |||
4711 | // bits. If the total number of significant bits does not exceed the | |||
4712 | // result bit width (minus 1), there is no overflow. | |||
4713 | // This means if we have enough leading sign bits in the operands | |||
4714 | // we can guarantee that the result does not overflow. | |||
4715 | // Ref: "Hacker's Delight" by Henry Warren | |||
4716 | unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); | |||
4717 | ||||
4718 | // Note that underestimating the number of sign bits gives a more | |||
4719 | // conservative answer. | |||
4720 | unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) + | |||
4721 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT); | |||
4722 | ||||
4723 | // First handle the easy case: if we have enough sign bits there's | |||
4724 | // definitely no overflow. | |||
4725 | if (SignBits > BitWidth + 1) | |||
4726 | return OverflowResult::NeverOverflows; | |||
4727 | ||||
4728 | // There are two ambiguous cases where there can be no overflow: | |||
4729 | // SignBits == BitWidth + 1 and | |||
4730 | // SignBits == BitWidth | |||
4731 | // The second case is difficult to check, therefore we only handle the | |||
4732 | // first case. | |||
4733 | if (SignBits == BitWidth + 1) { | |||
4734 | // It overflows only when both arguments are negative and the true | |||
4735 | // product is exactly the minimum negative number. | |||
4736 | // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 | |||
4737 | // For simplicity we just check if at least one side is not negative. | |||
4738 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4739 | nullptr, UseInstrInfo); | |||
4740 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4741 | nullptr, UseInstrInfo); | |||
4742 | if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) | |||
4743 | return OverflowResult::NeverOverflows; | |||
4744 | } | |||
4745 | return OverflowResult::MayOverflow; | |||
4746 | } | |||
4747 | ||||
4748 | OverflowResult llvm::computeOverflowForUnsignedAdd( | |||
4749 | const Value *LHS, const Value *RHS, const DataLayout &DL, | |||
4750 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
4751 | bool UseInstrInfo) { | |||
4752 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
4753 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4754 | nullptr, UseInstrInfo); | |||
4755 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
4756 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | |||
4757 | nullptr, UseInstrInfo); | |||
4758 | return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange)); | |||
4759 | } | |||
4760 | ||||
4761 | static OverflowResult computeOverflowForSignedAdd(const Value *LHS, | |||
4762 | const Value *RHS, | |||
4763 | const AddOperator *Add, | |||
4764 | const DataLayout &DL, | |||
4765 | AssumptionCache *AC, | |||
4766 | const Instruction *CxtI, | |||
4767 | const DominatorTree *DT) { | |||
4768 | if (Add && Add->hasNoSignedWrap()) { | |||
4769 | return OverflowResult::NeverOverflows; | |||
4770 | } | |||
4771 | ||||
4772 | // If LHS and RHS each have at least two sign bits, the addition will look | |||
4773 | // like | |||
4774 | // | |||
4775 | // XX..... + | |||
4776 | // YY..... | |||
4777 | // | |||
4778 | // If the carry into the most significant position is 0, X and Y can't both | |||
4779 | // be 1 and therefore the carry out of the addition is also 0. | |||
4780 | // | |||
4781 | // If the carry into the most significant position is 1, X and Y can't both | |||
4782 | // be 0 and therefore the carry out of the addition is also 1. | |||
4783 | // | |||
4784 | // Since the carry into the most significant position is always equal to | |||
4785 | // the carry out of the addition, there is no signed overflow. | |||
4786 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | |||
4787 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | |||
4788 | return OverflowResult::NeverOverflows; | |||
4789 | ||||
4790 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
4791 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4792 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
4793 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4794 | OverflowResult OR = | |||
4795 | mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange)); | |||
4796 | if (OR != OverflowResult::MayOverflow) | |||
4797 | return OR; | |||
4798 | ||||
4799 | // The remaining code needs Add to be available. Early returns if not so. | |||
4800 | if (!Add) | |||
4801 | return OverflowResult::MayOverflow; | |||
4802 | ||||
4803 | // If the sign of Add is the same as at least one of the operands, this add | |||
4804 | // CANNOT overflow. If this can be determined from the known bits of the | |||
4805 | // operands the above signedAddMayOverflow() check will have already done so. | |||
4806 | // The only other way to improve on the known bits is from an assumption, so | |||
4807 | // call computeKnownBitsFromAssume() directly. | |||
4808 | bool LHSOrRHSKnownNonNegative = | |||
4809 | (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); | |||
4810 | bool LHSOrRHSKnownNegative = | |||
4811 | (LHSRange.isAllNegative() || RHSRange.isAllNegative()); | |||
4812 | if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { | |||
4813 | KnownBits AddKnown(LHSRange.getBitWidth()); | |||
4814 | computeKnownBitsFromAssume( | |||
4815 | Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true)); | |||
4816 | if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || | |||
4817 | (AddKnown.isNegative() && LHSOrRHSKnownNegative)) | |||
4818 | return OverflowResult::NeverOverflows; | |||
4819 | } | |||
4820 | ||||
4821 | return OverflowResult::MayOverflow; | |||
4822 | } | |||
4823 | ||||
4824 | OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, | |||
4825 | const Value *RHS, | |||
4826 | const DataLayout &DL, | |||
4827 | AssumptionCache *AC, | |||
4828 | const Instruction *CxtI, | |||
4829 | const DominatorTree *DT) { | |||
4830 | // Checking for conditions implied by dominating conditions may be expensive. | |||
4831 | // Limit it to usub_with_overflow calls for now. | |||
4832 | if (match(CxtI, | |||
4833 | m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) | |||
4834 | if (auto C = | |||
4835 | isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) { | |||
4836 | if (*C) | |||
4837 | return OverflowResult::NeverOverflows; | |||
4838 | return OverflowResult::AlwaysOverflowsLow; | |||
4839 | } | |||
4840 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
4841 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4842 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
4843 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4844 | return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange)); | |||
4845 | } | |||
4846 | ||||
4847 | OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, | |||
4848 | const Value *RHS, | |||
4849 | const DataLayout &DL, | |||
4850 | AssumptionCache *AC, | |||
4851 | const Instruction *CxtI, | |||
4852 | const DominatorTree *DT) { | |||
4853 | // If LHS and RHS each have at least two sign bits, the subtraction | |||
4854 | // cannot overflow. | |||
4855 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | |||
4856 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | |||
4857 | return OverflowResult::NeverOverflows; | |||
4858 | ||||
4859 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
4860 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4861 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
4862 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
4863 | return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange)); | |||
4864 | } | |||
4865 | ||||
4866 | bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, | |||
4867 | const DominatorTree &DT) { | |||
4868 | SmallVector<const BranchInst *, 2> GuardingBranches; | |||
4869 | SmallVector<const ExtractValueInst *, 2> Results; | |||
4870 | ||||
4871 | for (const User *U : WO->users()) { | |||
4872 | if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) { | |||
4873 | assert(EVI->getNumIndices() == 1 && "Obvious from CI's type")(static_cast <bool> (EVI->getNumIndices() == 1 && "Obvious from CI's type") ? void (0) : __assert_fail ("EVI->getNumIndices() == 1 && \"Obvious from CI's type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4873, __extension__ __PRETTY_FUNCTION__ )); | |||
4874 | ||||
4875 | if (EVI->getIndices()[0] == 0) | |||
4876 | Results.push_back(EVI); | |||
4877 | else { | |||
4878 | assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type")(static_cast <bool> (EVI->getIndices()[0] == 1 && "Obvious from CI's type") ? void (0) : __assert_fail ("EVI->getIndices()[0] == 1 && \"Obvious from CI's type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4878, __extension__ __PRETTY_FUNCTION__ )); | |||
4879 | ||||
4880 | for (const auto *U : EVI->users()) | |||
4881 | if (const auto *B = dyn_cast<BranchInst>(U)) { | |||
4882 | assert(B->isConditional() && "How else is it using an i1?")(static_cast <bool> (B->isConditional() && "How else is it using an i1?" ) ? void (0) : __assert_fail ("B->isConditional() && \"How else is it using an i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4882, __extension__ __PRETTY_FUNCTION__ )); | |||
4883 | GuardingBranches.push_back(B); | |||
4884 | } | |||
4885 | } | |||
4886 | } else { | |||
4887 | // We are using the aggregate directly in a way we don't want to analyze | |||
4888 | // here (storing it to a global, say). | |||
4889 | return false; | |||
4890 | } | |||
4891 | } | |||
4892 | ||||
4893 | auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { | |||
4894 | BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); | |||
4895 | if (!NoWrapEdge.isSingleEdge()) | |||
4896 | return false; | |||
4897 | ||||
4898 | // Check if all users of the add are provably no-wrap. | |||
4899 | for (const auto *Result : Results) { | |||
4900 | // If the extractvalue itself is not executed on overflow, the we don't | |||
4901 | // need to check each use separately, since domination is transitive. | |||
4902 | if (DT.dominates(NoWrapEdge, Result->getParent())) | |||
4903 | continue; | |||
4904 | ||||
4905 | for (auto &RU : Result->uses()) | |||
4906 | if (!DT.dominates(NoWrapEdge, RU)) | |||
4907 | return false; | |||
4908 | } | |||
4909 | ||||
4910 | return true; | |||
4911 | }; | |||
4912 | ||||
4913 | return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch); | |||
4914 | } | |||
4915 | ||||
4916 | static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly, | |||
4917 | bool ConsiderFlags) { | |||
4918 | ||||
4919 | if (ConsiderFlags && Op->hasPoisonGeneratingFlags()) | |||
4920 | return true; | |||
4921 | ||||
4922 | unsigned Opcode = Op->getOpcode(); | |||
4923 | ||||
4924 | // Check whether opcode is a poison/undef-generating operation | |||
4925 | switch (Opcode) { | |||
4926 | case Instruction::Shl: | |||
4927 | case Instruction::AShr: | |||
4928 | case Instruction::LShr: { | |||
4929 | // Shifts return poison if shiftwidth is larger than the bitwidth. | |||
4930 | if (auto *C = dyn_cast<Constant>(Op->getOperand(1))) { | |||
4931 | SmallVector<Constant *, 4> ShiftAmounts; | |||
4932 | if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) { | |||
4933 | unsigned NumElts = FVTy->getNumElements(); | |||
4934 | for (unsigned i = 0; i < NumElts; ++i) | |||
4935 | ShiftAmounts.push_back(C->getAggregateElement(i)); | |||
4936 | } else if (isa<ScalableVectorType>(C->getType())) | |||
4937 | return true; // Can't tell, just return true to be safe | |||
4938 | else | |||
4939 | ShiftAmounts.push_back(C); | |||
4940 | ||||
4941 | bool Safe = llvm::all_of(ShiftAmounts, [](Constant *C) { | |||
4942 | auto *CI = dyn_cast_or_null<ConstantInt>(C); | |||
4943 | return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth()); | |||
4944 | }); | |||
4945 | return !Safe; | |||
4946 | } | |||
4947 | return true; | |||
4948 | } | |||
4949 | case Instruction::FPToSI: | |||
4950 | case Instruction::FPToUI: | |||
4951 | // fptosi/ui yields poison if the resulting value does not fit in the | |||
4952 | // destination type. | |||
4953 | return true; | |||
4954 | case Instruction::Call: | |||
4955 | if (auto *II = dyn_cast<IntrinsicInst>(Op)) { | |||
4956 | switch (II->getIntrinsicID()) { | |||
4957 | // TODO: Add more intrinsics. | |||
4958 | case Intrinsic::ctpop: | |||
4959 | case Intrinsic::sadd_with_overflow: | |||
4960 | case Intrinsic::ssub_with_overflow: | |||
4961 | case Intrinsic::smul_with_overflow: | |||
4962 | case Intrinsic::uadd_with_overflow: | |||
4963 | case Intrinsic::usub_with_overflow: | |||
4964 | case Intrinsic::umul_with_overflow: | |||
4965 | return false; | |||
4966 | } | |||
4967 | } | |||
4968 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
4969 | case Instruction::CallBr: | |||
4970 | case Instruction::Invoke: { | |||
4971 | const auto *CB = cast<CallBase>(Op); | |||
4972 | return !CB->hasRetAttr(Attribute::NoUndef); | |||
4973 | } | |||
4974 | case Instruction::InsertElement: | |||
4975 | case Instruction::ExtractElement: { | |||
4976 | // If index exceeds the length of the vector, it returns poison | |||
4977 | auto *VTy = cast<VectorType>(Op->getOperand(0)->getType()); | |||
4978 | unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; | |||
4979 | auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp)); | |||
4980 | if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue())) | |||
4981 | return true; | |||
4982 | return false; | |||
4983 | } | |||
4984 | case Instruction::ShuffleVector: { | |||
4985 | // shufflevector may return undef. | |||
4986 | if (PoisonOnly) | |||
4987 | return false; | |||
4988 | ArrayRef<int> Mask = isa<ConstantExpr>(Op) | |||
4989 | ? cast<ConstantExpr>(Op)->getShuffleMask() | |||
4990 | : cast<ShuffleVectorInst>(Op)->getShuffleMask(); | |||
4991 | return is_contained(Mask, UndefMaskElem); | |||
4992 | } | |||
4993 | case Instruction::FNeg: | |||
4994 | case Instruction::PHI: | |||
4995 | case Instruction::Select: | |||
4996 | case Instruction::URem: | |||
4997 | case Instruction::SRem: | |||
4998 | case Instruction::ExtractValue: | |||
4999 | case Instruction::InsertValue: | |||
5000 | case Instruction::Freeze: | |||
5001 | case Instruction::ICmp: | |||
5002 | case Instruction::FCmp: | |||
5003 | return false; | |||
5004 | case Instruction::GetElementPtr: | |||
5005 | // inbounds is handled above | |||
5006 | // TODO: what about inrange on constexpr? | |||
5007 | return false; | |||
5008 | default: { | |||
5009 | const auto *CE = dyn_cast<ConstantExpr>(Op); | |||
5010 | if (isa<CastInst>(Op) || (CE && CE->isCast())) | |||
5011 | return false; | |||
5012 | else if (Instruction::isBinaryOp(Opcode)) | |||
5013 | return false; | |||
5014 | // Be conservative and return true. | |||
5015 | return true; | |||
5016 | } | |||
5017 | } | |||
5018 | } | |||
5019 | ||||
5020 | bool llvm::canCreateUndefOrPoison(const Operator *Op, bool ConsiderFlags) { | |||
5021 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false, ConsiderFlags); | |||
5022 | } | |||
5023 | ||||
5024 | bool llvm::canCreatePoison(const Operator *Op, bool ConsiderFlags) { | |||
5025 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true, ConsiderFlags); | |||
5026 | } | |||
5027 | ||||
5028 | static bool directlyImpliesPoison(const Value *ValAssumedPoison, | |||
5029 | const Value *V, unsigned Depth) { | |||
5030 | if (ValAssumedPoison == V) | |||
5031 | return true; | |||
5032 | ||||
5033 | const unsigned MaxDepth = 2; | |||
5034 | if (Depth >= MaxDepth) | |||
5035 | return false; | |||
5036 | ||||
5037 | if (const auto *I = dyn_cast<Instruction>(V)) { | |||
5038 | if (propagatesPoison(cast<Operator>(I))) | |||
5039 | return any_of(I->operands(), [=](const Value *Op) { | |||
5040 | return directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1); | |||
5041 | }); | |||
5042 | ||||
5043 | // 'select ValAssumedPoison, _, _' is poison. | |||
5044 | if (const auto *SI = dyn_cast<SelectInst>(I)) | |||
5045 | return directlyImpliesPoison(ValAssumedPoison, SI->getCondition(), | |||
5046 | Depth + 1); | |||
5047 | // V = extractvalue V0, idx | |||
5048 | // V2 = extractvalue V0, idx2 | |||
5049 | // V0's elements are all poison or not. (e.g., add_with_overflow) | |||
5050 | const WithOverflowInst *II; | |||
5051 | if (match(I, m_ExtractValue(m_WithOverflowInst(II))) && | |||
5052 | (match(ValAssumedPoison, m_ExtractValue(m_Specific(II))) || | |||
5053 | llvm::is_contained(II->args(), ValAssumedPoison))) | |||
5054 | return true; | |||
5055 | } | |||
5056 | return false; | |||
5057 | } | |||
5058 | ||||
5059 | static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, | |||
5060 | unsigned Depth) { | |||
5061 | if (isGuaranteedNotToBeUndefOrPoison(ValAssumedPoison)) | |||
5062 | return true; | |||
5063 | ||||
5064 | if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) | |||
5065 | return true; | |||
5066 | ||||
5067 | const unsigned MaxDepth = 2; | |||
5068 | if (Depth >= MaxDepth) | |||
5069 | return false; | |||
5070 | ||||
5071 | const auto *I = dyn_cast<Instruction>(ValAssumedPoison); | |||
5072 | if (I && !canCreatePoison(cast<Operator>(I))) { | |||
5073 | return all_of(I->operands(), [=](const Value *Op) { | |||
5074 | return impliesPoison(Op, V, Depth + 1); | |||
5075 | }); | |||
5076 | } | |||
5077 | return false; | |||
5078 | } | |||
5079 | ||||
5080 | bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { | |||
5081 | return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); | |||
5082 | } | |||
5083 | ||||
5084 | static bool programUndefinedIfUndefOrPoison(const Value *V, | |||
5085 | bool PoisonOnly); | |||
5086 | ||||
5087 | static bool isGuaranteedNotToBeUndefOrPoison(const Value *V, | |||
5088 | AssumptionCache *AC, | |||
5089 | const Instruction *CtxI, | |||
5090 | const DominatorTree *DT, | |||
5091 | unsigned Depth, bool PoisonOnly) { | |||
5092 | if (Depth >= MaxAnalysisRecursionDepth) | |||
5093 | return false; | |||
5094 | ||||
5095 | if (isa<MetadataAsValue>(V)) | |||
5096 | return false; | |||
5097 | ||||
5098 | if (const auto *A = dyn_cast<Argument>(V)) { | |||
5099 | if (A->hasAttribute(Attribute::NoUndef)) | |||
5100 | return true; | |||
5101 | } | |||
5102 | ||||
5103 | if (auto *C = dyn_cast<Constant>(V)) { | |||
5104 | if (isa<UndefValue>(C)) | |||
5105 | return PoisonOnly && !isa<PoisonValue>(C); | |||
5106 | ||||
5107 | if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) || | |||
5108 | isa<ConstantPointerNull>(C) || isa<Function>(C)) | |||
5109 | return true; | |||
5110 | ||||
5111 | if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C)) | |||
5112 | return (PoisonOnly ? !C->containsPoisonElement() | |||
5113 | : !C->containsUndefOrPoisonElement()) && | |||
5114 | !C->containsConstantExpression(); | |||
5115 | } | |||
5116 | ||||
5117 | // Strip cast operations from a pointer value. | |||
5118 | // Note that stripPointerCastsSameRepresentation can strip off getelementptr | |||
5119 | // inbounds with zero offset. To guarantee that the result isn't poison, the | |||
5120 | // stripped pointer is checked as it has to be pointing into an allocated | |||
5121 | // object or be null `null` to ensure `inbounds` getelement pointers with a | |||
5122 | // zero offset could not produce poison. | |||
5123 | // It can strip off addrspacecast that do not change bit representation as | |||
5124 | // well. We believe that such addrspacecast is equivalent to no-op. | |||
5125 | auto *StrippedV = V->stripPointerCastsSameRepresentation(); | |||
5126 | if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) || | |||
5127 | isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV)) | |||
5128 | return true; | |||
5129 | ||||
5130 | auto OpCheck = [&](const Value *V) { | |||
5131 | return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1, | |||
5132 | PoisonOnly); | |||
5133 | }; | |||
5134 | ||||
5135 | if (auto *Opr = dyn_cast<Operator>(V)) { | |||
5136 | // If the value is a freeze instruction, then it can never | |||
5137 | // be undef or poison. | |||
5138 | if (isa<FreezeInst>(V)) | |||
5139 | return true; | |||
5140 | ||||
5141 | if (const auto *CB = dyn_cast<CallBase>(V)) { | |||
5142 | if (CB->hasRetAttr(Attribute::NoUndef)) | |||
5143 | return true; | |||
5144 | } | |||
5145 | ||||
5146 | if (const auto *PN = dyn_cast<PHINode>(V)) { | |||
5147 | unsigned Num = PN->getNumIncomingValues(); | |||
5148 | bool IsWellDefined = true; | |||
5149 | for (unsigned i = 0; i < Num; ++i) { | |||
5150 | auto *TI = PN->getIncomingBlock(i)->getTerminator(); | |||
5151 | if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI, | |||
5152 | DT, Depth + 1, PoisonOnly)) { | |||
5153 | IsWellDefined = false; | |||
5154 | break; | |||
5155 | } | |||
5156 | } | |||
5157 | if (IsWellDefined) | |||
5158 | return true; | |||
5159 | } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck)) | |||
5160 | return true; | |||
5161 | } | |||
5162 | ||||
5163 | if (auto *I = dyn_cast<LoadInst>(V)) | |||
5164 | if (I->getMetadata(LLVMContext::MD_noundef)) | |||
5165 | return true; | |||
5166 | ||||
5167 | if (programUndefinedIfUndefOrPoison(V, PoisonOnly)) | |||
5168 | return true; | |||
5169 | ||||
5170 | // CxtI may be null or a cloned instruction. | |||
5171 | if (!CtxI || !CtxI->getParent() || !DT) | |||
5172 | return false; | |||
5173 | ||||
5174 | auto *DNode = DT->getNode(CtxI->getParent()); | |||
5175 | if (!DNode) | |||
5176 | // Unreachable block | |||
5177 | return false; | |||
5178 | ||||
5179 | // If V is used as a branch condition before reaching CtxI, V cannot be | |||
5180 | // undef or poison. | |||
5181 | // br V, BB1, BB2 | |||
5182 | // BB1: | |||
5183 | // CtxI ; V cannot be undef or poison here | |||
5184 | auto *Dominator = DNode->getIDom(); | |||
5185 | while (Dominator) { | |||
5186 | auto *TI = Dominator->getBlock()->getTerminator(); | |||
5187 | ||||
5188 | Value *Cond = nullptr; | |||
5189 | if (auto BI = dyn_cast_or_null<BranchInst>(TI)) { | |||
5190 | if (BI->isConditional()) | |||
5191 | Cond = BI->getCondition(); | |||
5192 | } else if (auto SI = dyn_cast_or_null<SwitchInst>(TI)) { | |||
5193 | Cond = SI->getCondition(); | |||
5194 | } | |||
5195 | ||||
5196 | if (Cond) { | |||
5197 | if (Cond == V) | |||
5198 | return true; | |||
5199 | else if (PoisonOnly && isa<Operator>(Cond)) { | |||
5200 | // For poison, we can analyze further | |||
5201 | auto *Opr = cast<Operator>(Cond); | |||
5202 | if (propagatesPoison(Opr) && is_contained(Opr->operand_values(), V)) | |||
5203 | return true; | |||
5204 | } | |||
5205 | } | |||
5206 | ||||
5207 | Dominator = Dominator->getIDom(); | |||
5208 | } | |||
5209 | ||||
5210 | if (getKnowledgeValidInContext(V, {Attribute::NoUndef}, CtxI, DT, AC)) | |||
5211 | return true; | |||
5212 | ||||
5213 | return false; | |||
5214 | } | |||
5215 | ||||
5216 | bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, | |||
5217 | const Instruction *CtxI, | |||
5218 | const DominatorTree *DT, | |||
5219 | unsigned Depth) { | |||
5220 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false); | |||
5221 | } | |||
5222 | ||||
5223 | bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, | |||
5224 | const Instruction *CtxI, | |||
5225 | const DominatorTree *DT, unsigned Depth) { | |||
5226 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true); | |||
5227 | } | |||
5228 | ||||
5229 | OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, | |||
5230 | const DataLayout &DL, | |||
5231 | AssumptionCache *AC, | |||
5232 | const Instruction *CxtI, | |||
5233 | const DominatorTree *DT) { | |||
5234 | return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), | |||
5235 | Add, DL, AC, CxtI, DT); | |||
5236 | } | |||
5237 | ||||
5238 | OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS, | |||
5239 | const Value *RHS, | |||
5240 | const DataLayout &DL, | |||
5241 | AssumptionCache *AC, | |||
5242 | const Instruction *CxtI, | |||
5243 | const DominatorTree *DT) { | |||
5244 | return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); | |||
5245 | } | |||
5246 | ||||
5247 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { | |||
5248 | // Note: An atomic operation isn't guaranteed to return in a reasonable amount | |||
5249 | // of time because it's possible for another thread to interfere with it for an | |||
5250 | // arbitrary length of time, but programs aren't allowed to rely on that. | |||
5251 | ||||
5252 | // If there is no successor, then execution can't transfer to it. | |||
5253 | if (isa<ReturnInst>(I)) | |||
5254 | return false; | |||
5255 | if (isa<UnreachableInst>(I)) | |||
5256 | return false; | |||
5257 | ||||
5258 | // Note: Do not add new checks here; instead, change Instruction::mayThrow or | |||
5259 | // Instruction::willReturn. | |||
5260 | // | |||
5261 | // FIXME: Move this check into Instruction::willReturn. | |||
5262 | if (isa<CatchPadInst>(I)) { | |||
5263 | switch (classifyEHPersonality(I->getFunction()->getPersonalityFn())) { | |||
5264 | default: | |||
5265 | // A catchpad may invoke exception object constructors and such, which | |||
5266 | // in some languages can be arbitrary code, so be conservative by default. | |||
5267 | return false; | |||
5268 | case EHPersonality::CoreCLR: | |||
5269 | // For CoreCLR, it just involves a type test. | |||
5270 | return true; | |||
5271 | } | |||
5272 | } | |||
5273 | ||||
5274 | // An instruction that returns without throwing must transfer control flow | |||
5275 | // to a successor. | |||
5276 | return !I->mayThrow() && I->willReturn(); | |||
5277 | } | |||
5278 | ||||
5279 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { | |||
5280 | // TODO: This is slightly conservative for invoke instruction since exiting | |||
5281 | // via an exception *is* normal control for them. | |||
5282 | for (const Instruction &I : *BB) | |||
5283 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
5284 | return false; | |||
5285 | return true; | |||
5286 | } | |||
5287 | ||||
5288 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( | |||
5289 | BasicBlock::const_iterator Begin, BasicBlock::const_iterator End, | |||
5290 | unsigned ScanLimit) { | |||
5291 | return isGuaranteedToTransferExecutionToSuccessor(make_range(Begin, End), | |||
5292 | ScanLimit); | |||
5293 | } | |||
5294 | ||||
5295 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( | |||
5296 | iterator_range<BasicBlock::const_iterator> Range, unsigned ScanLimit) { | |||
5297 | assert(ScanLimit && "scan limit must be non-zero")(static_cast <bool> (ScanLimit && "scan limit must be non-zero" ) ? void (0) : __assert_fail ("ScanLimit && \"scan limit must be non-zero\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5297, __extension__ __PRETTY_FUNCTION__ )); | |||
5298 | for (const Instruction &I : Range) { | |||
5299 | if (isa<DbgInfoIntrinsic>(I)) | |||
5300 | continue; | |||
5301 | if (--ScanLimit == 0) | |||
5302 | return false; | |||
5303 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
5304 | return false; | |||
5305 | } | |||
5306 | return true; | |||
5307 | } | |||
5308 | ||||
5309 | bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, | |||
5310 | const Loop *L) { | |||
5311 | // The loop header is guaranteed to be executed for every iteration. | |||
5312 | // | |||
5313 | // FIXME: Relax this constraint to cover all basic blocks that are | |||
5314 | // guaranteed to be executed at every iteration. | |||
5315 | if (I->getParent() != L->getHeader()) return false; | |||
5316 | ||||
5317 | for (const Instruction &LI : *L->getHeader()) { | |||
5318 | if (&LI == I) return true; | |||
5319 | if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; | |||
5320 | } | |||
5321 | llvm_unreachable("Instruction not contained in its own parent basic block.")::llvm::llvm_unreachable_internal("Instruction not contained in its own parent basic block." , "llvm/lib/Analysis/ValueTracking.cpp", 5321); | |||
5322 | } | |||
5323 | ||||
5324 | bool llvm::propagatesPoison(const Operator *I) { | |||
5325 | switch (I->getOpcode()) { | |||
5326 | case Instruction::Freeze: | |||
5327 | case Instruction::Select: | |||
5328 | case Instruction::PHI: | |||
5329 | case Instruction::Invoke: | |||
5330 | return false; | |||
5331 | case Instruction::Call: | |||
5332 | if (auto *II = dyn_cast<IntrinsicInst>(I)) { | |||
5333 | switch (II->getIntrinsicID()) { | |||
5334 | // TODO: Add more intrinsics. | |||
5335 | case Intrinsic::sadd_with_overflow: | |||
5336 | case Intrinsic::ssub_with_overflow: | |||
5337 | case Intrinsic::smul_with_overflow: | |||
5338 | case Intrinsic::uadd_with_overflow: | |||
5339 | case Intrinsic::usub_with_overflow: | |||
5340 | case Intrinsic::umul_with_overflow: | |||
5341 | // If an input is a vector containing a poison element, the | |||
5342 | // two output vectors (calculated results, overflow bits)' | |||
5343 | // corresponding lanes are poison. | |||
5344 | return true; | |||
5345 | case Intrinsic::ctpop: | |||
5346 | return true; | |||
5347 | } | |||
5348 | } | |||
5349 | return false; | |||
5350 | case Instruction::ICmp: | |||
5351 | case Instruction::FCmp: | |||
5352 | case Instruction::GetElementPtr: | |||
5353 | return true; | |||
5354 | default: | |||
5355 | if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I)) | |||
5356 | return true; | |||
5357 | ||||
5358 | // Be conservative and return false. | |||
5359 | return false; | |||
5360 | } | |||
5361 | } | |||
5362 | ||||
5363 | void llvm::getGuaranteedWellDefinedOps( | |||
5364 | const Instruction *I, SmallPtrSetImpl<const Value *> &Operands) { | |||
5365 | switch (I->getOpcode()) { | |||
5366 | case Instruction::Store: | |||
5367 | Operands.insert(cast<StoreInst>(I)->getPointerOperand()); | |||
5368 | break; | |||
5369 | ||||
5370 | case Instruction::Load: | |||
5371 | Operands.insert(cast<LoadInst>(I)->getPointerOperand()); | |||
5372 | break; | |||
5373 | ||||
5374 | // Since dereferenceable attribute imply noundef, atomic operations | |||
5375 | // also implicitly have noundef pointers too | |||
5376 | case Instruction::AtomicCmpXchg: | |||
5377 | Operands.insert(cast<AtomicCmpXchgInst>(I)->getPointerOperand()); | |||
5378 | break; | |||
5379 | ||||
5380 | case Instruction::AtomicRMW: | |||
5381 | Operands.insert(cast<AtomicRMWInst>(I)->getPointerOperand()); | |||
5382 | break; | |||
5383 | ||||
5384 | case Instruction::Call: | |||
5385 | case Instruction::Invoke: { | |||
5386 | const CallBase *CB = cast<CallBase>(I); | |||
5387 | if (CB->isIndirectCall()) | |||
5388 | Operands.insert(CB->getCalledOperand()); | |||
5389 | for (unsigned i = 0; i < CB->arg_size(); ++i) { | |||
5390 | if (CB->paramHasAttr(i, Attribute::NoUndef) || | |||
5391 | CB->paramHasAttr(i, Attribute::Dereferenceable)) | |||
5392 | Operands.insert(CB->getArgOperand(i)); | |||
5393 | } | |||
5394 | break; | |||
5395 | } | |||
5396 | case Instruction::Ret: | |||
5397 | if (I->getFunction()->hasRetAttribute(Attribute::NoUndef)) | |||
5398 | Operands.insert(I->getOperand(0)); | |||
5399 | break; | |||
5400 | default: | |||
5401 | break; | |||
5402 | } | |||
5403 | } | |||
5404 | ||||
5405 | void llvm::getGuaranteedNonPoisonOps(const Instruction *I, | |||
5406 | SmallPtrSetImpl<const Value *> &Operands) { | |||
5407 | getGuaranteedWellDefinedOps(I, Operands); | |||
5408 | switch (I->getOpcode()) { | |||
5409 | // Divisors of these operations are allowed to be partially undef. | |||
5410 | case Instruction::UDiv: | |||
5411 | case Instruction::SDiv: | |||
5412 | case Instruction::URem: | |||
5413 | case Instruction::SRem: | |||
5414 | Operands.insert(I->getOperand(1)); | |||
5415 | break; | |||
5416 | case Instruction::Switch: | |||
5417 | if (BranchOnPoisonAsUB) | |||
5418 | Operands.insert(cast<SwitchInst>(I)->getCondition()); | |||
5419 | break; | |||
5420 | case Instruction::Br: { | |||
5421 | auto *BR = cast<BranchInst>(I); | |||
5422 | if (BranchOnPoisonAsUB && BR->isConditional()) | |||
5423 | Operands.insert(BR->getCondition()); | |||
5424 | break; | |||
5425 | } | |||
5426 | default: | |||
5427 | break; | |||
5428 | } | |||
5429 | } | |||
5430 | ||||
5431 | bool llvm::mustTriggerUB(const Instruction *I, | |||
5432 | const SmallSet<const Value *, 16>& KnownPoison) { | |||
5433 | SmallPtrSet<const Value *, 4> NonPoisonOps; | |||
5434 | getGuaranteedNonPoisonOps(I, NonPoisonOps); | |||
5435 | ||||
5436 | for (const auto *V : NonPoisonOps) | |||
5437 | if (KnownPoison.count(V)) | |||
5438 | return true; | |||
5439 | ||||
5440 | return false; | |||
5441 | } | |||
5442 | ||||
5443 | static bool programUndefinedIfUndefOrPoison(const Value *V, | |||
5444 | bool PoisonOnly) { | |||
5445 | // We currently only look for uses of values within the same basic | |||
5446 | // block, as that makes it easier to guarantee that the uses will be | |||
5447 | // executed given that Inst is executed. | |||
5448 | // | |||
5449 | // FIXME: Expand this to consider uses beyond the same basic block. To do | |||
5450 | // this, look out for the distinction between post-dominance and strong | |||
5451 | // post-dominance. | |||
5452 | const BasicBlock *BB = nullptr; | |||
5453 | BasicBlock::const_iterator Begin; | |||
5454 | if (const auto *Inst = dyn_cast<Instruction>(V)) { | |||
5455 | BB = Inst->getParent(); | |||
5456 | Begin = Inst->getIterator(); | |||
5457 | Begin++; | |||
5458 | } else if (const auto *Arg = dyn_cast<Argument>(V)) { | |||
5459 | BB = &Arg->getParent()->getEntryBlock(); | |||
5460 | Begin = BB->begin(); | |||
5461 | } else { | |||
5462 | return false; | |||
5463 | } | |||
5464 | ||||
5465 | // Limit number of instructions we look at, to avoid scanning through large | |||
5466 | // blocks. The current limit is chosen arbitrarily. | |||
5467 | unsigned ScanLimit = 32; | |||
5468 | BasicBlock::const_iterator End = BB->end(); | |||
5469 | ||||
5470 | if (!PoisonOnly) { | |||
5471 | // Since undef does not propagate eagerly, be conservative & just check | |||
5472 | // whether a value is directly passed to an instruction that must take | |||
5473 | // well-defined operands. | |||
5474 | ||||
5475 | for (auto &I : make_range(Begin, End)) { | |||
5476 | if (isa<DbgInfoIntrinsic>(I)) | |||
5477 | continue; | |||
5478 | if (--ScanLimit == 0) | |||
5479 | break; | |||
5480 | ||||
5481 | SmallPtrSet<const Value *, 4> WellDefinedOps; | |||
5482 | getGuaranteedWellDefinedOps(&I, WellDefinedOps); | |||
5483 | if (WellDefinedOps.contains(V)) | |||
5484 | return true; | |||
5485 | ||||
5486 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
5487 | break; | |||
5488 | } | |||
5489 | return false; | |||
5490 | } | |||
5491 | ||||
5492 | // Set of instructions that we have proved will yield poison if Inst | |||
5493 | // does. | |||
5494 | SmallSet<const Value *, 16> YieldsPoison; | |||
5495 | SmallSet<const BasicBlock *, 4> Visited; | |||
5496 | ||||
5497 | YieldsPoison.insert(V); | |||
5498 | auto Propagate = [&](const User *User) { | |||
5499 | if (propagatesPoison(cast<Operator>(User))) | |||
5500 | YieldsPoison.insert(User); | |||
5501 | }; | |||
5502 | for_each(V->users(), Propagate); | |||
5503 | Visited.insert(BB); | |||
5504 | ||||
5505 | while (true) { | |||
5506 | for (auto &I : make_range(Begin, End)) { | |||
5507 | if (isa<DbgInfoIntrinsic>(I)) | |||
5508 | continue; | |||
5509 | if (--ScanLimit == 0) | |||
5510 | return false; | |||
5511 | if (mustTriggerUB(&I, YieldsPoison)) | |||
5512 | return true; | |||
5513 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
5514 | return false; | |||
5515 | ||||
5516 | // Mark poison that propagates from I through uses of I. | |||
5517 | if (YieldsPoison.count(&I)) | |||
5518 | for_each(I.users(), Propagate); | |||
5519 | } | |||
5520 | ||||
5521 | BB = BB->getSingleSuccessor(); | |||
5522 | if (!BB || !Visited.insert(BB).second) | |||
5523 | break; | |||
5524 | ||||
5525 | Begin = BB->getFirstNonPHI()->getIterator(); | |||
5526 | End = BB->end(); | |||
5527 | } | |||
5528 | return false; | |||
5529 | } | |||
5530 | ||||
5531 | bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { | |||
5532 | return ::programUndefinedIfUndefOrPoison(Inst, false); | |||
5533 | } | |||
5534 | ||||
5535 | bool llvm::programUndefinedIfPoison(const Instruction *Inst) { | |||
5536 | return ::programUndefinedIfUndefOrPoison(Inst, true); | |||
5537 | } | |||
5538 | ||||
5539 | static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { | |||
5540 | if (FMF.noNaNs()) | |||
5541 | return true; | |||
5542 | ||||
5543 | if (auto *C = dyn_cast<ConstantFP>(V)) | |||
5544 | return !C->isNaN(); | |||
5545 | ||||
5546 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | |||
5547 | if (!C->getElementType()->isFloatingPointTy()) | |||
5548 | return false; | |||
5549 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | |||
5550 | if (C->getElementAsAPFloat(I).isNaN()) | |||
5551 | return false; | |||
5552 | } | |||
5553 | return true; | |||
5554 | } | |||
5555 | ||||
5556 | if (isa<ConstantAggregateZero>(V)) | |||
5557 | return true; | |||
5558 | ||||
5559 | return false; | |||
5560 | } | |||
5561 | ||||
5562 | static bool isKnownNonZero(const Value *V) { | |||
5563 | if (auto *C = dyn_cast<ConstantFP>(V)) | |||
5564 | return !C->isZero(); | |||
5565 | ||||
5566 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | |||
5567 | if (!C->getElementType()->isFloatingPointTy()) | |||
5568 | return false; | |||
5569 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | |||
5570 | if (C->getElementAsAPFloat(I).isZero()) | |||
5571 | return false; | |||
5572 | } | |||
5573 | return true; | |||
5574 | } | |||
5575 | ||||
5576 | return false; | |||
5577 | } | |||
5578 | ||||
5579 | /// Match clamp pattern for float types without care about NaNs or signed zeros. | |||
5580 | /// Given non-min/max outer cmp/select from the clamp pattern this | |||
5581 | /// function recognizes if it can be substitued by a "canonical" min/max | |||
5582 | /// pattern. | |||
5583 | static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, | |||
5584 | Value *CmpLHS, Value *CmpRHS, | |||
5585 | Value *TrueVal, Value *FalseVal, | |||
5586 | Value *&LHS, Value *&RHS) { | |||
5587 | // Try to match | |||
5588 | // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) | |||
5589 | // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) | |||
5590 | // and return description of the outer Max/Min. | |||
5591 | ||||
5592 | // First, check if select has inverse order: | |||
5593 | if (CmpRHS == FalseVal) { | |||
5594 | std::swap(TrueVal, FalseVal); | |||
5595 | Pred = CmpInst::getInversePredicate(Pred); | |||
5596 | } | |||
5597 | ||||
5598 | // Assume success now. If there's no match, callers should not use these anyway. | |||
5599 | LHS = TrueVal; | |||
5600 | RHS = FalseVal; | |||
5601 | ||||
5602 | const APFloat *FC1; | |||
5603 | if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite()) | |||
5604 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5605 | ||||
5606 | const APFloat *FC2; | |||
5607 | switch (Pred) { | |||
5608 | case CmpInst::FCMP_OLT: | |||
5609 | case CmpInst::FCMP_OLE: | |||
5610 | case CmpInst::FCMP_ULT: | |||
5611 | case CmpInst::FCMP_ULE: | |||
5612 | if (match(FalseVal, | |||
5613 | m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)), | |||
5614 | m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) && | |||
5615 | *FC1 < *FC2) | |||
5616 | return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false}; | |||
5617 | break; | |||
5618 | case CmpInst::FCMP_OGT: | |||
5619 | case CmpInst::FCMP_OGE: | |||
5620 | case CmpInst::FCMP_UGT: | |||
5621 | case CmpInst::FCMP_UGE: | |||
5622 | if (match(FalseVal, | |||
5623 | m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)), | |||
5624 | m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) && | |||
5625 | *FC1 > *FC2) | |||
5626 | return {SPF_FMINNUM, SPNB_RETURNS_ANY, false}; | |||
5627 | break; | |||
5628 | default: | |||
5629 | break; | |||
5630 | } | |||
5631 | ||||
5632 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5633 | } | |||
5634 | ||||
5635 | /// Recognize variations of: | |||
5636 | /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) | |||
5637 | static SelectPatternResult matchClamp(CmpInst::Predicate Pred, | |||
5638 | Value *CmpLHS, Value *CmpRHS, | |||
5639 | Value *TrueVal, Value *FalseVal) { | |||
5640 | // Swap the select operands and predicate to match the patterns below. | |||
5641 | if (CmpRHS != TrueVal) { | |||
5642 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
5643 | std::swap(TrueVal, FalseVal); | |||
5644 | } | |||
5645 | const APInt *C1; | |||
5646 | if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) { | |||
5647 | const APInt *C2; | |||
5648 | // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) | |||
5649 | if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) && | |||
5650 | C1->slt(*C2) && Pred == CmpInst::ICMP_SLT) | |||
5651 | return {SPF_SMAX, SPNB_NA, false}; | |||
5652 | ||||
5653 | // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) | |||
5654 | if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) && | |||
5655 | C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT) | |||
5656 | return {SPF_SMIN, SPNB_NA, false}; | |||
5657 | ||||
5658 | // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) | |||
5659 | if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) && | |||
5660 | C1->ult(*C2) && Pred == CmpInst::ICMP_ULT) | |||
5661 | return {SPF_UMAX, SPNB_NA, false}; | |||
5662 | ||||
5663 | // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) | |||
5664 | if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) && | |||
5665 | C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT) | |||
5666 | return {SPF_UMIN, SPNB_NA, false}; | |||
5667 | } | |||
5668 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5669 | } | |||
5670 | ||||
5671 | /// Recognize variations of: | |||
5672 | /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) | |||
5673 | static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, | |||
5674 | Value *CmpLHS, Value *CmpRHS, | |||
5675 | Value *TVal, Value *FVal, | |||
5676 | unsigned Depth) { | |||
5677 | // TODO: Allow FP min/max with nnan/nsz. | |||
5678 | assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison")(static_cast <bool> (CmpInst::isIntPredicate(Pred) && "Expected integer comparison") ? void (0) : __assert_fail ("CmpInst::isIntPredicate(Pred) && \"Expected integer comparison\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5678, __extension__ __PRETTY_FUNCTION__ )); | |||
5679 | ||||
5680 | Value *A = nullptr, *B = nullptr; | |||
5681 | SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1); | |||
5682 | if (!SelectPatternResult::isMinOrMax(L.Flavor)) | |||
5683 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5684 | ||||
5685 | Value *C = nullptr, *D = nullptr; | |||
5686 | SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1); | |||
5687 | if (L.Flavor != R.Flavor) | |||
5688 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5689 | ||||
5690 | // We have something like: x Pred y ? min(a, b) : min(c, d). | |||
5691 | // Try to match the compare to the min/max operations of the select operands. | |||
5692 | // First, make sure we have the right compare predicate. | |||
5693 | switch (L.Flavor) { | |||
5694 | case SPF_SMIN: | |||
5695 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { | |||
5696 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
5697 | std::swap(CmpLHS, CmpRHS); | |||
5698 | } | |||
5699 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) | |||
5700 | break; | |||
5701 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5702 | case SPF_SMAX: | |||
5703 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { | |||
5704 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
5705 | std::swap(CmpLHS, CmpRHS); | |||
5706 | } | |||
5707 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) | |||
5708 | break; | |||
5709 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5710 | case SPF_UMIN: | |||
5711 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { | |||
5712 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
5713 | std::swap(CmpLHS, CmpRHS); | |||
5714 | } | |||
5715 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) | |||
5716 | break; | |||
5717 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5718 | case SPF_UMAX: | |||
5719 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { | |||
5720 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
5721 | std::swap(CmpLHS, CmpRHS); | |||
5722 | } | |||
5723 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) | |||
5724 | break; | |||
5725 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5726 | default: | |||
5727 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5728 | } | |||
5729 | ||||
5730 | // If there is a common operand in the already matched min/max and the other | |||
5731 | // min/max operands match the compare operands (either directly or inverted), | |||
5732 | // then this is min/max of the same flavor. | |||
5733 | ||||
5734 | // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | |||
5735 | // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | |||
5736 | if (D == B) { | |||
5737 | if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | |||
5738 | match(A, m_Not(m_Specific(CmpRHS))))) | |||
5739 | return {L.Flavor, SPNB_NA, false}; | |||
5740 | } | |||
5741 | // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | |||
5742 | // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | |||
5743 | if (C == B) { | |||
5744 | if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | |||
5745 | match(A, m_Not(m_Specific(CmpRHS))))) | |||
5746 | return {L.Flavor, SPNB_NA, false}; | |||
5747 | } | |||
5748 | // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | |||
5749 | // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | |||
5750 | if (D == A) { | |||
5751 | if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | |||
5752 | match(B, m_Not(m_Specific(CmpRHS))))) | |||
5753 | return {L.Flavor, SPNB_NA, false}; | |||
5754 | } | |||
5755 | // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | |||
5756 | // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | |||
5757 | if (C == A) { | |||
5758 | if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | |||
5759 | match(B, m_Not(m_Specific(CmpRHS))))) | |||
5760 | return {L.Flavor, SPNB_NA, false}; | |||
5761 | } | |||
5762 | ||||
5763 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5764 | } | |||
5765 | ||||
5766 | /// If the input value is the result of a 'not' op, constant integer, or vector | |||
5767 | /// splat of a constant integer, return the bitwise-not source value. | |||
5768 | /// TODO: This could be extended to handle non-splat vector integer constants. | |||
5769 | static Value *getNotValue(Value *V) { | |||
5770 | Value *NotV; | |||
5771 | if (match(V, m_Not(m_Value(NotV)))) | |||
5772 | return NotV; | |||
5773 | ||||
5774 | const APInt *C; | |||
5775 | if (match(V, m_APInt(C))) | |||
5776 | return ConstantInt::get(V->getType(), ~(*C)); | |||
5777 | ||||
5778 | return nullptr; | |||
5779 | } | |||
5780 | ||||
5781 | /// Match non-obvious integer minimum and maximum sequences. | |||
5782 | static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, | |||
5783 | Value *CmpLHS, Value *CmpRHS, | |||
5784 | Value *TrueVal, Value *FalseVal, | |||
5785 | Value *&LHS, Value *&RHS, | |||
5786 | unsigned Depth) { | |||
5787 | // Assume success. If there's no match, callers should not use these anyway. | |||
5788 | LHS = TrueVal; | |||
5789 | RHS = FalseVal; | |||
5790 | ||||
5791 | SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); | |||
5792 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | |||
5793 | return SPR; | |||
5794 | ||||
5795 | SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth); | |||
5796 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | |||
5797 | return SPR; | |||
5798 | ||||
5799 | // Look through 'not' ops to find disguised min/max. | |||
5800 | // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) | |||
5801 | // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) | |||
5802 | if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) { | |||
5803 | switch (Pred) { | |||
5804 | case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false}; | |||
5805 | case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false}; | |||
5806 | case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false}; | |||
5807 | case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false}; | |||
5808 | default: break; | |||
5809 | } | |||
5810 | } | |||
5811 | ||||
5812 | // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) | |||
5813 | // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) | |||
5814 | if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) { | |||
5815 | switch (Pred) { | |||
5816 | case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false}; | |||
5817 | case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false}; | |||
5818 | case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false}; | |||
5819 | case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false}; | |||
5820 | default: break; | |||
5821 | } | |||
5822 | } | |||
5823 | ||||
5824 | if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) | |||
5825 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5826 | ||||
5827 | const APInt *C1; | |||
5828 | if (!match(CmpRHS, m_APInt(C1))) | |||
5829 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5830 | ||||
5831 | // An unsigned min/max can be written with a signed compare. | |||
5832 | const APInt *C2; | |||
5833 | if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) || | |||
5834 | (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) { | |||
5835 | // Is the sign bit set? | |||
5836 | // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX | |||
5837 | // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN | |||
5838 | if (Pred == CmpInst::ICMP_SLT && C1->isZero() && C2->isMaxSignedValue()) | |||
5839 | return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | |||
5840 | ||||
5841 | // Is the sign bit clear? | |||
5842 | // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX | |||
5843 | // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN | |||
5844 | if (Pred == CmpInst::ICMP_SGT && C1->isAllOnes() && C2->isMinSignedValue()) | |||
5845 | return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | |||
5846 | } | |||
5847 | ||||
5848 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5849 | } | |||
5850 | ||||
5851 | bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) { | |||
5852 | assert(X && Y && "Invalid operand")(static_cast <bool> (X && Y && "Invalid operand" ) ? void (0) : __assert_fail ("X && Y && \"Invalid operand\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5852, __extension__ __PRETTY_FUNCTION__ )); | |||
5853 | ||||
5854 | // X = sub (0, Y) || X = sub nsw (0, Y) | |||
5855 | if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) || | |||
5856 | (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y))))) | |||
5857 | return true; | |||
5858 | ||||
5859 | // Y = sub (0, X) || Y = sub nsw (0, X) | |||
5860 | if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) || | |||
5861 | (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X))))) | |||
5862 | return true; | |||
5863 | ||||
5864 | // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) | |||
5865 | Value *A, *B; | |||
5866 | return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) && | |||
5867 | match(Y, m_Sub(m_Specific(B), m_Specific(A))))) || | |||
5868 | (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) && | |||
5869 | match(Y, m_NSWSub(m_Specific(B), m_Specific(A))))); | |||
5870 | } | |||
5871 | ||||
5872 | static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, | |||
5873 | FastMathFlags FMF, | |||
5874 | Value *CmpLHS, Value *CmpRHS, | |||
5875 | Value *TrueVal, Value *FalseVal, | |||
5876 | Value *&LHS, Value *&RHS, | |||
5877 | unsigned Depth) { | |||
5878 | if (CmpInst::isFPPredicate(Pred)) { | |||
5879 | // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one | |||
5880 | // 0.0 operand, set the compare's 0.0 operands to that same value for the | |||
5881 | // purpose of identifying min/max. Disregard vector constants with undefined | |||
5882 | // elements because those can not be back-propagated for analysis. | |||
5883 | Value *OutputZeroVal = nullptr; | |||
5884 | if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) && | |||
5885 | !cast<Constant>(TrueVal)->containsUndefOrPoisonElement()) | |||
5886 | OutputZeroVal = TrueVal; | |||
5887 | else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) && | |||
5888 | !cast<Constant>(FalseVal)->containsUndefOrPoisonElement()) | |||
5889 | OutputZeroVal = FalseVal; | |||
5890 | ||||
5891 | if (OutputZeroVal) { | |||
5892 | if (match(CmpLHS, m_AnyZeroFP())) | |||
5893 | CmpLHS = OutputZeroVal; | |||
5894 | if (match(CmpRHS, m_AnyZeroFP())) | |||
5895 | CmpRHS = OutputZeroVal; | |||
5896 | } | |||
5897 | } | |||
5898 | ||||
5899 | LHS = CmpLHS; | |||
5900 | RHS = CmpRHS; | |||
5901 | ||||
5902 | // Signed zero may return inconsistent results between implementations. | |||
5903 | // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 | |||
5904 | // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) | |||
5905 | // Therefore, we behave conservatively and only proceed if at least one of the | |||
5906 | // operands is known to not be zero or if we don't care about signed zero. | |||
5907 | switch (Pred) { | |||
5908 | default: break; | |||
5909 | // FIXME: Include OGT/OLT/UGT/ULT. | |||
5910 | case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: | |||
5911 | case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: | |||
5912 | if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | |||
5913 | !isKnownNonZero(CmpRHS)) | |||
5914 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5915 | } | |||
5916 | ||||
5917 | SelectPatternNaNBehavior NaNBehavior = SPNB_NA; | |||
5918 | bool Ordered = false; | |||
5919 | ||||
5920 | // When given one NaN and one non-NaN input: | |||
5921 | // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. | |||
5922 | // - A simple C99 (a < b ? a : b) construction will return 'b' (as the | |||
5923 | // ordered comparison fails), which could be NaN or non-NaN. | |||
5924 | // so here we discover exactly what NaN behavior is required/accepted. | |||
5925 | if (CmpInst::isFPPredicate(Pred)) { | |||
5926 | bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); | |||
5927 | bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); | |||
5928 | ||||
5929 | if (LHSSafe && RHSSafe) { | |||
5930 | // Both operands are known non-NaN. | |||
5931 | NaNBehavior = SPNB_RETURNS_ANY; | |||
5932 | } else if (CmpInst::isOrdered(Pred)) { | |||
5933 | // An ordered comparison will return false when given a NaN, so it | |||
5934 | // returns the RHS. | |||
5935 | Ordered = true; | |||
5936 | if (LHSSafe) | |||
5937 | // LHS is non-NaN, so if RHS is NaN then NaN will be returned. | |||
5938 | NaNBehavior = SPNB_RETURNS_NAN; | |||
5939 | else if (RHSSafe) | |||
5940 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
5941 | else | |||
5942 | // Completely unsafe. | |||
5943 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5944 | } else { | |||
5945 | Ordered = false; | |||
5946 | // An unordered comparison will return true when given a NaN, so it | |||
5947 | // returns the LHS. | |||
5948 | if (LHSSafe) | |||
5949 | // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. | |||
5950 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
5951 | else if (RHSSafe) | |||
5952 | NaNBehavior = SPNB_RETURNS_NAN; | |||
5953 | else | |||
5954 | // Completely unsafe. | |||
5955 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
5956 | } | |||
5957 | } | |||
5958 | ||||
5959 | if (TrueVal == CmpRHS && FalseVal == CmpLHS) { | |||
5960 | std::swap(CmpLHS, CmpRHS); | |||
5961 | Pred = CmpInst::getSwappedPredicate(Pred); | |||
5962 | if (NaNBehavior == SPNB_RETURNS_NAN) | |||
5963 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
5964 | else if (NaNBehavior == SPNB_RETURNS_OTHER) | |||
5965 | NaNBehavior = SPNB_RETURNS_NAN; | |||
5966 | Ordered = !Ordered; | |||
5967 | } | |||
5968 | ||||
5969 | // ([if]cmp X, Y) ? X : Y | |||
5970 | if (TrueVal == CmpLHS && FalseVal == CmpRHS) { | |||
5971 | switch (Pred) { | |||
5972 | default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. | |||
5973 | case ICmpInst::ICMP_UGT: | |||
5974 | case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; | |||
5975 | case ICmpInst::ICMP_SGT: | |||
5976 | case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; | |||
5977 | case ICmpInst::ICMP_ULT: | |||
5978 | case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; | |||
5979 | case ICmpInst::ICMP_SLT: | |||
5980 | case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; | |||
5981 | case FCmpInst::FCMP_UGT: | |||
5982 | case FCmpInst::FCMP_UGE: | |||
5983 | case FCmpInst::FCMP_OGT: | |||
5984 | case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; | |||
5985 | case FCmpInst::FCMP_ULT: | |||
5986 | case FCmpInst::FCMP_ULE: | |||
5987 | case FCmpInst::FCMP_OLT: | |||
5988 | case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; | |||
5989 | } | |||
5990 | } | |||
5991 | ||||
5992 | if (isKnownNegation(TrueVal, FalseVal)) { | |||
5993 | // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can | |||
5994 | // match against either LHS or sext(LHS). | |||
5995 | auto MaybeSExtCmpLHS = | |||
5996 | m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS))); | |||
5997 | auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes()); | |||
5998 | auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One()); | |||
5999 | if (match(TrueVal, MaybeSExtCmpLHS)) { | |||
6000 | // Set the return values. If the compare uses the negated value (-X >s 0), | |||
6001 | // swap the return values because the negated value is always 'RHS'. | |||
6002 | LHS = TrueVal; | |||
6003 | RHS = FalseVal; | |||
6004 | if (match(CmpLHS, m_Neg(m_Specific(FalseVal)))) | |||
6005 | std::swap(LHS, RHS); | |||
6006 | ||||
6007 | // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) | |||
6008 | // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) | |||
6009 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | |||
6010 | return {SPF_ABS, SPNB_NA, false}; | |||
6011 | ||||
6012 | // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) | |||
6013 | if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne)) | |||
6014 | return {SPF_ABS, SPNB_NA, false}; | |||
6015 | ||||
6016 | // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) | |||
6017 | // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) | |||
6018 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | |||
6019 | return {SPF_NABS, SPNB_NA, false}; | |||
6020 | } | |||
6021 | else if (match(FalseVal, MaybeSExtCmpLHS)) { | |||
6022 | // Set the return values. If the compare uses the negated value (-X >s 0), | |||
6023 | // swap the return values because the negated value is always 'RHS'. | |||
6024 | LHS = FalseVal; | |||
6025 | RHS = TrueVal; | |||
6026 | if (match(CmpLHS, m_Neg(m_Specific(TrueVal)))) | |||
6027 | std::swap(LHS, RHS); | |||
6028 | ||||
6029 | // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) | |||
6030 | // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) | |||
6031 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | |||
6032 | return {SPF_NABS, SPNB_NA, false}; | |||
6033 | ||||
6034 | // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) | |||
6035 | // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) | |||
6036 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | |||
6037 | return {SPF_ABS, SPNB_NA, false}; | |||
6038 | } | |||
6039 | } | |||
6040 | ||||
6041 | if (CmpInst::isIntPredicate(Pred)) | |||
6042 | return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); | |||
6043 | ||||
6044 | // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar | |||
6045 | // may return either -0.0 or 0.0, so fcmp/select pair has stricter | |||
6046 | // semantics than minNum. Be conservative in such case. | |||
6047 | if (NaNBehavior != SPNB_RETURNS_ANY || | |||
6048 | (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | |||
6049 | !isKnownNonZero(CmpRHS))) | |||
6050 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
6051 | ||||
6052 | return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); | |||
6053 | } | |||
6054 | ||||
6055 | /// Helps to match a select pattern in case of a type mismatch. | |||
6056 | /// | |||
6057 | /// The function processes the case when type of true and false values of a | |||
6058 | /// select instruction differs from type of the cmp instruction operands because | |||
6059 | /// of a cast instruction. The function checks if it is legal to move the cast | |||
6060 | /// operation after "select". If yes, it returns the new second value of | |||
6061 | /// "select" (with the assumption that cast is moved): | |||
6062 | /// 1. As operand of cast instruction when both values of "select" are same cast | |||
6063 | /// instructions. | |||
6064 | /// 2. As restored constant (by applying reverse cast operation) when the first | |||
6065 | /// value of the "select" is a cast operation and the second value is a | |||
6066 | /// constant. | |||
6067 | /// NOTE: We return only the new second value because the first value could be | |||
6068 | /// accessed as operand of cast instruction. | |||
6069 | static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, | |||
6070 | Instruction::CastOps *CastOp) { | |||
6071 | auto *Cast1 = dyn_cast<CastInst>(V1); | |||
6072 | if (!Cast1) | |||
6073 | return nullptr; | |||
6074 | ||||
6075 | *CastOp = Cast1->getOpcode(); | |||
6076 | Type *SrcTy = Cast1->getSrcTy(); | |||
6077 | if (auto *Cast2 = dyn_cast<CastInst>(V2)) { | |||
6078 | // If V1 and V2 are both the same cast from the same type, look through V1. | |||
6079 | if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) | |||
6080 | return Cast2->getOperand(0); | |||
6081 | return nullptr; | |||
6082 | } | |||
6083 | ||||
6084 | auto *C = dyn_cast<Constant>(V2); | |||
6085 | if (!C) | |||
6086 | return nullptr; | |||
6087 | ||||
6088 | Constant *CastedTo = nullptr; | |||
6089 | switch (*CastOp) { | |||
6090 | case Instruction::ZExt: | |||
6091 | if (CmpI->isUnsigned()) | |||
6092 | CastedTo = ConstantExpr::getTrunc(C, SrcTy); | |||
6093 | break; | |||
6094 | case Instruction::SExt: | |||
6095 | if (CmpI->isSigned()) | |||
6096 | CastedTo = ConstantExpr::getTrunc(C, SrcTy, true); | |||
6097 | break; | |||
6098 | case Instruction::Trunc: | |||
6099 | Constant *CmpConst; | |||
6100 | if (match(CmpI->getOperand(1), m_Constant(CmpConst)) && | |||
6101 | CmpConst->getType() == SrcTy) { | |||
6102 | // Here we have the following case: | |||
6103 | // | |||
6104 | // %cond = cmp iN %x, CmpConst | |||
6105 | // %tr = trunc iN %x to iK | |||
6106 | // %narrowsel = select i1 %cond, iK %t, iK C | |||
6107 | // | |||
6108 | // We can always move trunc after select operation: | |||
6109 | // | |||
6110 | // %cond = cmp iN %x, CmpConst | |||
6111 | // %widesel = select i1 %cond, iN %x, iN CmpConst | |||
6112 | // %tr = trunc iN %widesel to iK | |||
6113 | // | |||
6114 | // Note that C could be extended in any way because we don't care about | |||
6115 | // upper bits after truncation. It can't be abs pattern, because it would | |||
6116 | // look like: | |||
6117 | // | |||
6118 | // select i1 %cond, x, -x. | |||
6119 | // | |||
6120 | // So only min/max pattern could be matched. Such match requires widened C | |||
6121 | // == CmpConst. That is why set widened C = CmpConst, condition trunc | |||
6122 | // CmpConst == C is checked below. | |||
6123 | CastedTo = CmpConst; | |||
6124 | } else { | |||
6125 | CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned()); | |||
6126 | } | |||
6127 | break; | |||
6128 | case Instruction::FPTrunc: | |||
6129 | CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true); | |||
6130 | break; | |||
6131 | case Instruction::FPExt: | |||
6132 | CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true); | |||
6133 | break; | |||
6134 | case Instruction::FPToUI: | |||
6135 | CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true); | |||
6136 | break; | |||
6137 | case Instruction::FPToSI: | |||
6138 | CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true); | |||
6139 | break; | |||
6140 | case Instruction::UIToFP: | |||
6141 | CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true); | |||
6142 | break; | |||
6143 | case Instruction::SIToFP: | |||
6144 | CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true); | |||
6145 | break; | |||
6146 | default: | |||
6147 | break; | |||
6148 | } | |||
6149 | ||||
6150 | if (!CastedTo) | |||
6151 | return nullptr; | |||
6152 | ||||
6153 | // Make sure the cast doesn't lose any information. | |||
6154 | Constant *CastedBack = | |||
6155 | ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true); | |||
6156 | if (CastedBack != C) | |||
6157 | return nullptr; | |||
6158 | ||||
6159 | return CastedTo; | |||
6160 | } | |||
6161 | ||||
6162 | SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, | |||
6163 | Instruction::CastOps *CastOp, | |||
6164 | unsigned Depth) { | |||
6165 | if (Depth >= MaxAnalysisRecursionDepth) | |||
6166 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
6167 | ||||
6168 | SelectInst *SI = dyn_cast<SelectInst>(V); | |||
6169 | if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; | |||
6170 | ||||
6171 | CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); | |||
6172 | if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; | |||
6173 | ||||
6174 | Value *TrueVal = SI->getTrueValue(); | |||
6175 | Value *FalseVal = SI->getFalseValue(); | |||
6176 | ||||
6177 | return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, | |||
6178 | CastOp, Depth); | |||
6179 | } | |||
6180 | ||||
6181 | SelectPatternResult llvm::matchDecomposedSelectPattern( | |||
6182 | CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, | |||
6183 | Instruction::CastOps *CastOp, unsigned Depth) { | |||
6184 | CmpInst::Predicate Pred = CmpI->getPredicate(); | |||
6185 | Value *CmpLHS = CmpI->getOperand(0); | |||
6186 | Value *CmpRHS = CmpI->getOperand(1); | |||
6187 | FastMathFlags FMF; | |||
6188 | if (isa<FPMathOperator>(CmpI)) | |||
6189 | FMF = CmpI->getFastMathFlags(); | |||
6190 | ||||
6191 | // Bail out early. | |||
6192 | if (CmpI->isEquality()) | |||
6193 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
6194 | ||||
6195 | // Deal with type mismatches. | |||
6196 | if (CastOp && CmpLHS->getType() != TrueVal->getType()) { | |||
6197 | if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) { | |||
6198 | // If this is a potential fmin/fmax with a cast to integer, then ignore | |||
6199 | // -0.0 because there is no corresponding integer value. | |||
6200 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | |||
6201 | FMF.setNoSignedZeros(); | |||
6202 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | |||
6203 | cast<CastInst>(TrueVal)->getOperand(0), C, | |||
6204 | LHS, RHS, Depth); | |||
6205 | } | |||
6206 | if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) { | |||
6207 | // If this is a potential fmin/fmax with a cast to integer, then ignore | |||
6208 | // -0.0 because there is no corresponding integer value. | |||
6209 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | |||
6210 | FMF.setNoSignedZeros(); | |||
6211 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | |||
6212 | C, cast<CastInst>(FalseVal)->getOperand(0), | |||
6213 | LHS, RHS, Depth); | |||
6214 | } | |||
6215 | } | |||
6216 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, | |||
6217 | LHS, RHS, Depth); | |||
6218 | } | |||
6219 | ||||
6220 | CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { | |||
6221 | if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; | |||
6222 | if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; | |||
6223 | if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; | |||
6224 | if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; | |||
6225 | if (SPF == SPF_FMINNUM) | |||
6226 | return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; | |||
6227 | if (SPF == SPF_FMAXNUM) | |||
6228 | return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; | |||
6229 | llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "llvm/lib/Analysis/ValueTracking.cpp" , 6229); | |||
6230 | } | |||
6231 | ||||
6232 | SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { | |||
6233 | if (SPF == SPF_SMIN) return SPF_SMAX; | |||
6234 | if (SPF == SPF_UMIN) return SPF_UMAX; | |||
6235 | if (SPF == SPF_SMAX) return SPF_SMIN; | |||
6236 | if (SPF == SPF_UMAX) return SPF_UMIN; | |||
6237 | llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "llvm/lib/Analysis/ValueTracking.cpp" , 6237); | |||
6238 | } | |||
6239 | ||||
6240 | Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { | |||
6241 | switch (MinMaxID) { | |||
6242 | case Intrinsic::smax: return Intrinsic::smin; | |||
6243 | case Intrinsic::smin: return Intrinsic::smax; | |||
6244 | case Intrinsic::umax: return Intrinsic::umin; | |||
6245 | case Intrinsic::umin: return Intrinsic::umax; | |||
6246 | default: llvm_unreachable("Unexpected intrinsic")::llvm::llvm_unreachable_internal("Unexpected intrinsic", "llvm/lib/Analysis/ValueTracking.cpp" , 6246); | |||
6247 | } | |||
6248 | } | |||
6249 | ||||
6250 | CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) { | |||
6251 | return getMinMaxPred(getInverseMinMaxFlavor(SPF)); | |||
6252 | } | |||
6253 | ||||
6254 | APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { | |||
6255 | switch (SPF) { | |||
6256 | case SPF_SMAX: return APInt::getSignedMaxValue(BitWidth); | |||
6257 | case SPF_SMIN: return APInt::getSignedMinValue(BitWidth); | |||
6258 | case SPF_UMAX: return APInt::getMaxValue(BitWidth); | |||
6259 | case SPF_UMIN: return APInt::getMinValue(BitWidth); | |||
6260 | default: llvm_unreachable("Unexpected flavor")::llvm::llvm_unreachable_internal("Unexpected flavor", "llvm/lib/Analysis/ValueTracking.cpp" , 6260); | |||
6261 | } | |||
6262 | } | |||
6263 | ||||
6264 | std::pair<Intrinsic::ID, bool> | |||
6265 | llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { | |||
6266 | // Check if VL contains select instructions that can be folded into a min/max | |||
6267 | // vector intrinsic and return the intrinsic if it is possible. | |||
6268 | // TODO: Support floating point min/max. | |||
6269 | bool AllCmpSingleUse = true; | |||
6270 | SelectPatternResult SelectPattern; | |||
6271 | SelectPattern.Flavor = SPF_UNKNOWN; | |||
6272 | if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) { | |||
6273 | Value *LHS, *RHS; | |||
6274 | auto CurrentPattern = matchSelectPattern(I, LHS, RHS); | |||
6275 | if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) || | |||
6276 | CurrentPattern.Flavor == SPF_FMINNUM || | |||
6277 | CurrentPattern.Flavor == SPF_FMAXNUM || | |||
6278 | !I->getType()->isIntOrIntVectorTy()) | |||
6279 | return false; | |||
6280 | if (SelectPattern.Flavor != SPF_UNKNOWN && | |||
6281 | SelectPattern.Flavor != CurrentPattern.Flavor) | |||
6282 | return false; | |||
6283 | SelectPattern = CurrentPattern; | |||
6284 | AllCmpSingleUse &= | |||
6285 | match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value())); | |||
6286 | return true; | |||
6287 | })) { | |||
6288 | switch (SelectPattern.Flavor) { | |||
6289 | case SPF_SMIN: | |||
6290 | return {Intrinsic::smin, AllCmpSingleUse}; | |||
6291 | case SPF_UMIN: | |||
6292 | return {Intrinsic::umin, AllCmpSingleUse}; | |||
6293 | case SPF_SMAX: | |||
6294 | return {Intrinsic::smax, AllCmpSingleUse}; | |||
6295 | case SPF_UMAX: | |||
6296 | return {Intrinsic::umax, AllCmpSingleUse}; | |||
6297 | default: | |||
6298 | llvm_unreachable("unexpected select pattern flavor")::llvm::llvm_unreachable_internal("unexpected select pattern flavor" , "llvm/lib/Analysis/ValueTracking.cpp", 6298); | |||
6299 | } | |||
6300 | } | |||
6301 | return {Intrinsic::not_intrinsic, false}; | |||
6302 | } | |||
6303 | ||||
6304 | bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, | |||
6305 | Value *&Start, Value *&Step) { | |||
6306 | // Handle the case of a simple two-predecessor recurrence PHI. | |||
6307 | // There's a lot more that could theoretically be done here, but | |||
6308 | // this is sufficient to catch some interesting cases. | |||
6309 | if (P->getNumIncomingValues() != 2) | |||
6310 | return false; | |||
6311 | ||||
6312 | for (unsigned i = 0; i != 2; ++i) { | |||
6313 | Value *L = P->getIncomingValue(i); | |||
6314 | Value *R = P->getIncomingValue(!i); | |||
6315 | Operator *LU = dyn_cast<Operator>(L); | |||
6316 | if (!LU) | |||
6317 | continue; | |||
6318 | unsigned Opcode = LU->getOpcode(); | |||
6319 | ||||
6320 | switch (Opcode) { | |||
6321 | default: | |||
6322 | continue; | |||
6323 | // TODO: Expand list -- xor, div, gep, uaddo, etc.. | |||
6324 | case Instruction::LShr: | |||
6325 | case Instruction::AShr: | |||
6326 | case Instruction::Shl: | |||
6327 | case Instruction::Add: | |||
6328 | case Instruction::Sub: | |||
6329 | case Instruction::And: | |||
6330 | case Instruction::Or: | |||
6331 | case Instruction::Mul: { | |||
6332 | Value *LL = LU->getOperand(0); | |||
6333 | Value *LR = LU->getOperand(1); | |||
6334 | // Find a recurrence. | |||
6335 | if (LL == P) | |||
6336 | L = LR; | |||
6337 | else if (LR == P) | |||
6338 | L = LL; | |||
6339 | else | |||
6340 | continue; // Check for recurrence with L and R flipped. | |||
6341 | ||||
6342 | break; // Match! | |||
6343 | } | |||
6344 | }; | |||
6345 | ||||
6346 | // We have matched a recurrence of the form: | |||
6347 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
6348 | // %iv.next = binop %iv, L | |||
6349 | // OR | |||
6350 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
6351 | // %iv.next = binop L, %iv | |||
6352 | BO = cast<BinaryOperator>(LU); | |||
6353 | Start = R; | |||
6354 | Step = L; | |||
6355 | return true; | |||
6356 | } | |||
6357 | return false; | |||
6358 | } | |||
6359 | ||||
6360 | bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, | |||
6361 | Value *&Start, Value *&Step) { | |||
6362 | BinaryOperator *BO = nullptr; | |||
6363 | P = dyn_cast<PHINode>(I->getOperand(0)); | |||
6364 | if (!P) | |||
6365 | P = dyn_cast<PHINode>(I->getOperand(1)); | |||
6366 | return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I; | |||
6367 | } | |||
6368 | ||||
6369 | /// Return true if "icmp Pred LHS RHS" is always true. | |||
6370 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, | |||
6371 | const Value *RHS, const DataLayout &DL, | |||
6372 | unsigned Depth) { | |||
6373 | assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!")(static_cast <bool> (!LHS->getType()->isVectorTy( ) && "TODO: extend to handle vectors!") ? void (0) : __assert_fail ("!LHS->getType()->isVectorTy() && \"TODO: extend to handle vectors!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6373, __extension__ __PRETTY_FUNCTION__ )); | |||
6374 | if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) | |||
6375 | return true; | |||
6376 | ||||
6377 | switch (Pred) { | |||
6378 | default: | |||
6379 | return false; | |||
6380 | ||||
6381 | case CmpInst::ICMP_SLE: { | |||
6382 | const APInt *C; | |||
6383 | ||||
6384 | // LHS s<= LHS +_{nsw} C if C >= 0 | |||
6385 | if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) | |||
6386 | return !C->isNegative(); | |||
6387 | return false; | |||
6388 | } | |||
6389 | ||||
6390 | case CmpInst::ICMP_ULE: { | |||
6391 | const APInt *C; | |||
6392 | ||||
6393 | // LHS u<= LHS +_{nuw} C for any C | |||
6394 | if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) | |||
6395 | return true; | |||
6396 | ||||
6397 | // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) | |||
6398 | auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B, | |||
6399 | const Value *&X, | |||
6400 | const APInt *&CA, const APInt *&CB) { | |||
6401 | if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && | |||
6402 | match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) | |||
6403 | return true; | |||
6404 | ||||
6405 | // If X & C == 0 then (X | C) == X +_{nuw} C | |||
6406 | if (match(A, m_Or(m_Value(X), m_APInt(CA))) && | |||
6407 | match(B, m_Or(m_Specific(X), m_APInt(CB)))) { | |||
6408 | KnownBits Known(CA->getBitWidth()); | |||
6409 | computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr, | |||
6410 | /*CxtI*/ nullptr, /*DT*/ nullptr); | |||
6411 | if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero)) | |||
6412 | return true; | |||
6413 | } | |||
6414 | ||||
6415 | return false; | |||
6416 | }; | |||
6417 | ||||
6418 | const Value *X; | |||
6419 | const APInt *CLHS, *CRHS; | |||
6420 | if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) | |||
6421 | return CLHS->ule(*CRHS); | |||
6422 | ||||
6423 | return false; | |||
6424 | } | |||
6425 | } | |||
6426 | } | |||
6427 | ||||
6428 | /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred | |||
6429 | /// ALHS ARHS" is true. Otherwise, return None. | |||
6430 | static Optional<bool> | |||
6431 | isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, | |||
6432 | const Value *ARHS, const Value *BLHS, const Value *BRHS, | |||
6433 | const DataLayout &DL, unsigned Depth) { | |||
6434 | switch (Pred) { | |||
6435 | default: | |||
6436 | return None; | |||
6437 | ||||
6438 | case CmpInst::ICMP_SLT: | |||
6439 | case CmpInst::ICMP_SLE: | |||
6440 | if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) && | |||
6441 | isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth)) | |||
6442 | return true; | |||
6443 | return None; | |||
6444 | ||||
6445 | case CmpInst::ICMP_ULT: | |||
6446 | case CmpInst::ICMP_ULE: | |||
6447 | if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) && | |||
6448 | isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth)) | |||
6449 | return true; | |||
6450 | return None; | |||
6451 | } | |||
6452 | } | |||
6453 | ||||
6454 | /// Return true if the operands of the two compares match. IsSwappedOps is true | |||
6455 | /// when the operands match, but are swapped. | |||
6456 | static bool isMatchingOps(const Value *ALHS, const Value *ARHS, | |||
6457 | const Value *BLHS, const Value *BRHS, | |||
6458 | bool &IsSwappedOps) { | |||
6459 | ||||
6460 | bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS); | |||
6461 | IsSwappedOps = (ALHS == BRHS && ARHS == BLHS); | |||
6462 | return IsMatchingOps || IsSwappedOps; | |||
6463 | } | |||
6464 | ||||
6465 | /// Return true if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is true. | |||
6466 | /// Return false if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is false. | |||
6467 | /// Otherwise, return None if we can't infer anything. | |||
6468 | static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred, | |||
6469 | CmpInst::Predicate BPred, | |||
6470 | bool AreSwappedOps) { | |||
6471 | // Canonicalize the predicate as if the operands were not commuted. | |||
6472 | if (AreSwappedOps) | |||
6473 | BPred = ICmpInst::getSwappedPredicate(BPred); | |||
6474 | ||||
6475 | if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred)) | |||
6476 | return true; | |||
6477 | if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred)) | |||
6478 | return false; | |||
6479 | ||||
6480 | return None; | |||
6481 | } | |||
6482 | ||||
6483 | /// Return true if "icmp APred X, C1" implies "icmp BPred X, C2" is true. | |||
6484 | /// Return false if "icmp APred X, C1" implies "icmp BPred X, C2" is false. | |||
6485 | /// Otherwise, return None if we can't infer anything. | |||
6486 | static Optional<bool> | |||
6487 | isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, | |||
6488 | const ConstantInt *C1, | |||
6489 | CmpInst::Predicate BPred, | |||
6490 | const ConstantInt *C2) { | |||
6491 | ConstantRange DomCR = | |||
6492 | ConstantRange::makeExactICmpRegion(APred, C1->getValue()); | |||
6493 | ConstantRange CR = ConstantRange::makeExactICmpRegion(BPred, C2->getValue()); | |||
6494 | ConstantRange Intersection = DomCR.intersectWith(CR); | |||
6495 | ConstantRange Difference = DomCR.difference(CR); | |||
6496 | if (Intersection.isEmptySet()) | |||
6497 | return false; | |||
6498 | if (Difference.isEmptySet()) | |||
6499 | return true; | |||
6500 | return None; | |||
6501 | } | |||
6502 | ||||
6503 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is | |||
6504 | /// false. Otherwise, return None if we can't infer anything. | |||
6505 | static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS, | |||
6506 | CmpInst::Predicate BPred, | |||
6507 | const Value *BLHS, const Value *BRHS, | |||
6508 | const DataLayout &DL, bool LHSIsTrue, | |||
6509 | unsigned Depth) { | |||
6510 | Value *ALHS = LHS->getOperand(0); | |||
6511 | Value *ARHS = LHS->getOperand(1); | |||
6512 | ||||
6513 | // The rest of the logic assumes the LHS condition is true. If that's not the | |||
6514 | // case, invert the predicate to make it so. | |||
6515 | CmpInst::Predicate APred = | |||
6516 | LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); | |||
6517 | ||||
6518 | // Can we infer anything when the two compares have matching operands? | |||
6519 | bool AreSwappedOps; | |||
6520 | if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, AreSwappedOps)) { | |||
6521 | if (Optional<bool> Implication = isImpliedCondMatchingOperands( | |||
6522 | APred, BPred, AreSwappedOps)) | |||
6523 | return Implication; | |||
6524 | // No amount of additional analysis will infer the second condition, so | |||
6525 | // early exit. | |||
6526 | return None; | |||
6527 | } | |||
6528 | ||||
6529 | // Can we infer anything when the LHS operands match and the RHS operands are | |||
6530 | // constants (not necessarily matching)? | |||
6531 | if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) { | |||
6532 | if (Optional<bool> Implication = isImpliedCondMatchingImmOperands( | |||
6533 | APred, cast<ConstantInt>(ARHS), BPred, cast<ConstantInt>(BRHS))) | |||
6534 | return Implication; | |||
6535 | // No amount of additional analysis will infer the second condition, so | |||
6536 | // early exit. | |||
6537 | return None; | |||
6538 | } | |||
6539 | ||||
6540 | if (APred == BPred) | |||
6541 | return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth); | |||
6542 | return None; | |||
6543 | } | |||
6544 | ||||
6545 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is | |||
6546 | /// false. Otherwise, return None if we can't infer anything. We expect the | |||
6547 | /// RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' instruction. | |||
6548 | static Optional<bool> | |||
6549 | isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, | |||
6550 | const Value *RHSOp0, const Value *RHSOp1, | |||
6551 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | |||
6552 | // The LHS must be an 'or', 'and', or a 'select' instruction. | |||
6553 | assert((LHS->getOpcode() == Instruction::And ||(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6556, __extension__ __PRETTY_FUNCTION__ )) | |||
6554 | LHS->getOpcode() == Instruction::Or ||(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6556, __extension__ __PRETTY_FUNCTION__ )) | |||
6555 | LHS->getOpcode() == Instruction::Select) &&(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6556, __extension__ __PRETTY_FUNCTION__ )) | |||
6556 | "Expected LHS to be 'and', 'or', or 'select'.")(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6556, __extension__ __PRETTY_FUNCTION__ )); | |||
6557 | ||||
6558 | assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit") ? void (0) : __assert_fail ("Depth <= MaxAnalysisRecursionDepth && \"Hit recursion limit\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6558, __extension__ __PRETTY_FUNCTION__ )); | |||
6559 | ||||
6560 | // If the result of an 'or' is false, then we know both legs of the 'or' are | |||
6561 | // false. Similarly, if the result of an 'and' is true, then we know both | |||
6562 | // legs of the 'and' are true. | |||
6563 | const Value *ALHS, *ARHS; | |||
6564 | if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) || | |||
6565 | (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) { | |||
6566 | // FIXME: Make this non-recursion. | |||
6567 | if (Optional<bool> Implication = isImpliedCondition( | |||
6568 | ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | |||
6569 | return Implication; | |||
6570 | if (Optional<bool> Implication = isImpliedCondition( | |||
6571 | ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | |||
6572 | return Implication; | |||
6573 | return None; | |||
6574 | } | |||
6575 | return None; | |||
6576 | } | |||
6577 | ||||
6578 | Optional<bool> | |||
6579 | llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, | |||
6580 | const Value *RHSOp0, const Value *RHSOp1, | |||
6581 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | |||
6582 | // Bail out when we hit the limit. | |||
6583 | if (Depth == MaxAnalysisRecursionDepth) | |||
6584 | return None; | |||
6585 | ||||
6586 | // A mismatch occurs when we compare a scalar cmp to a vector cmp, for | |||
6587 | // example. | |||
6588 | if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) | |||
6589 | return None; | |||
6590 | ||||
6591 | Type *OpTy = LHS->getType(); | |||
6592 | assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!")(static_cast <bool> (OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!") ? void (0) : __assert_fail ("OpTy->isIntOrIntVectorTy(1) && \"Expected integer type only!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6592, __extension__ __PRETTY_FUNCTION__ )); | |||
6593 | ||||
6594 | // FIXME: Extending the code below to handle vectors. | |||
6595 | if (OpTy->isVectorTy()) | |||
6596 | return None; | |||
6597 | ||||
6598 | assert(OpTy->isIntegerTy(1) && "implied by above")(static_cast <bool> (OpTy->isIntegerTy(1) && "implied by above") ? void (0) : __assert_fail ("OpTy->isIntegerTy(1) && \"implied by above\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6598, __extension__ __PRETTY_FUNCTION__ )); | |||
6599 | ||||
6600 | // Both LHS and RHS are icmps. | |||
6601 | const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS); | |||
6602 | if (LHSCmp) | |||
6603 | return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | |||
6604 | Depth); | |||
6605 | ||||
6606 | /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect | |||
6607 | /// the RHS to be an icmp. | |||
6608 | /// FIXME: Add support for and/or/select on the RHS. | |||
6609 | if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) { | |||
6610 | if ((LHSI->getOpcode() == Instruction::And || | |||
6611 | LHSI->getOpcode() == Instruction::Or || | |||
6612 | LHSI->getOpcode() == Instruction::Select)) | |||
6613 | return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | |||
6614 | Depth); | |||
6615 | } | |||
6616 | return None; | |||
6617 | } | |||
6618 | ||||
6619 | Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, | |||
6620 | const DataLayout &DL, bool LHSIsTrue, | |||
6621 | unsigned Depth) { | |||
6622 | // LHS ==> RHS by definition | |||
6623 | if (LHS == RHS) | |||
6624 | return LHSIsTrue; | |||
6625 | ||||
6626 | const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS); | |||
6627 | if (RHSCmp) | |||
6628 | return isImpliedCondition(LHS, RHSCmp->getPredicate(), | |||
6629 | RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL, | |||
6630 | LHSIsTrue, Depth); | |||
6631 | return None; | |||
6632 | } | |||
6633 | ||||
6634 | // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch | |||
6635 | // condition dominating ContextI or nullptr, if no condition is found. | |||
6636 | static std::pair<Value *, bool> | |||
6637 | getDomPredecessorCondition(const Instruction *ContextI) { | |||
6638 | if (!ContextI || !ContextI->getParent()) | |||
6639 | return {nullptr, false}; | |||
6640 | ||||
6641 | // TODO: This is a poor/cheap way to determine dominance. Should we use a | |||
6642 | // dominator tree (eg, from a SimplifyQuery) instead? | |||
6643 | const BasicBlock *ContextBB = ContextI->getParent(); | |||
6644 | const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); | |||
6645 | if (!PredBB) | |||
6646 | return {nullptr, false}; | |||
6647 | ||||
6648 | // We need a conditional branch in the predecessor. | |||
6649 | Value *PredCond; | |||
6650 | BasicBlock *TrueBB, *FalseBB; | |||
6651 | if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB))) | |||
6652 | return {nullptr, false}; | |||
6653 | ||||
6654 | // The branch should get simplified. Don't bother simplifying this condition. | |||
6655 | if (TrueBB == FalseBB) | |||
6656 | return {nullptr, false}; | |||
6657 | ||||
6658 | assert((TrueBB == ContextBB || FalseBB == ContextBB) &&(static_cast <bool> ((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?" ) ? void (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6659, __extension__ __PRETTY_FUNCTION__ )) | |||
6659 | "Predecessor block does not point to successor?")(static_cast <bool> ((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?" ) ? void (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6659, __extension__ __PRETTY_FUNCTION__ )); | |||
6660 | ||||
6661 | // Is this condition implied by the predecessor condition? | |||
6662 | return {PredCond, TrueBB == ContextBB}; | |||
6663 | } | |||
6664 | ||||
6665 | Optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, | |||
6666 | const Instruction *ContextI, | |||
6667 | const DataLayout &DL) { | |||
6668 | assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool")(static_cast <bool> (Cond->getType()->isIntOrIntVectorTy (1) && "Condition must be bool") ? void (0) : __assert_fail ("Cond->getType()->isIntOrIntVectorTy(1) && \"Condition must be bool\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6668, __extension__ __PRETTY_FUNCTION__ )); | |||
6669 | auto PredCond = getDomPredecessorCondition(ContextI); | |||
6670 | if (PredCond.first) | |||
6671 | return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second); | |||
6672 | return None; | |||
6673 | } | |||
6674 | ||||
6675 | Optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, | |||
6676 | const Value *LHS, const Value *RHS, | |||
6677 | const Instruction *ContextI, | |||
6678 | const DataLayout &DL) { | |||
6679 | auto PredCond = getDomPredecessorCondition(ContextI); | |||
6680 | if (PredCond.first) | |||
6681 | return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL, | |||
6682 | PredCond.second); | |||
6683 | return None; | |||
6684 | } | |||
6685 | ||||
6686 | static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, | |||
6687 | APInt &Upper, const InstrInfoQuery &IIQ, | |||
6688 | bool PreferSignedRange) { | |||
6689 | unsigned Width = Lower.getBitWidth(); | |||
6690 | const APInt *C; | |||
6691 | switch (BO.getOpcode()) { | |||
6692 | case Instruction::Add: | |||
6693 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) { | |||
6694 | bool HasNSW = IIQ.hasNoSignedWrap(&BO); | |||
6695 | bool HasNUW = IIQ.hasNoUnsignedWrap(&BO); | |||
6696 | ||||
6697 | // If the caller expects a signed compare, then try to use a signed range. | |||
6698 | // Otherwise if both no-wraps are set, use the unsigned range because it | |||
6699 | // is never larger than the signed range. Example: | |||
6700 | // "add nuw nsw i8 X, -2" is unsigned [254,255] vs. signed [-128, 125]. | |||
6701 | if (PreferSignedRange && HasNSW && HasNUW) | |||
6702 | HasNUW = false; | |||
6703 | ||||
6704 | if (HasNUW) { | |||
6705 | // 'add nuw x, C' produces [C, UINT_MAX]. | |||
6706 | Lower = *C; | |||
6707 | } else if (HasNSW) { | |||
6708 | if (C->isNegative()) { | |||
6709 | // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. | |||
6710 | Lower = APInt::getSignedMinValue(Width); | |||
6711 | Upper = APInt::getSignedMaxValue(Width) + *C + 1; | |||
6712 | } else { | |||
6713 | // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. | |||
6714 | Lower = APInt::getSignedMinValue(Width) + *C; | |||
6715 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6716 | } | |||
6717 | } | |||
6718 | } | |||
6719 | break; | |||
6720 | ||||
6721 | case Instruction::And: | |||
6722 | if (match(BO.getOperand(1), m_APInt(C))) | |||
6723 | // 'and x, C' produces [0, C]. | |||
6724 | Upper = *C + 1; | |||
6725 | break; | |||
6726 | ||||
6727 | case Instruction::Or: | |||
6728 | if (match(BO.getOperand(1), m_APInt(C))) | |||
6729 | // 'or x, C' produces [C, UINT_MAX]. | |||
6730 | Lower = *C; | |||
6731 | break; | |||
6732 | ||||
6733 | case Instruction::AShr: | |||
6734 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | |||
6735 | // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. | |||
6736 | Lower = APInt::getSignedMinValue(Width).ashr(*C); | |||
6737 | Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; | |||
6738 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
6739 | unsigned ShiftAmount = Width - 1; | |||
6740 | if (!C->isZero() && IIQ.isExact(&BO)) | |||
6741 | ShiftAmount = C->countTrailingZeros(); | |||
6742 | if (C->isNegative()) { | |||
6743 | // 'ashr C, x' produces [C, C >> (Width-1)] | |||
6744 | Lower = *C; | |||
6745 | Upper = C->ashr(ShiftAmount) + 1; | |||
6746 | } else { | |||
6747 | // 'ashr C, x' produces [C >> (Width-1), C] | |||
6748 | Lower = C->ashr(ShiftAmount); | |||
6749 | Upper = *C + 1; | |||
6750 | } | |||
6751 | } | |||
6752 | break; | |||
6753 | ||||
6754 | case Instruction::LShr: | |||
6755 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | |||
6756 | // 'lshr x, C' produces [0, UINT_MAX >> C]. | |||
6757 | Upper = APInt::getAllOnes(Width).lshr(*C) + 1; | |||
6758 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
6759 | // 'lshr C, x' produces [C >> (Width-1), C]. | |||
6760 | unsigned ShiftAmount = Width - 1; | |||
6761 | if (!C->isZero() && IIQ.isExact(&BO)) | |||
6762 | ShiftAmount = C->countTrailingZeros(); | |||
6763 | Lower = C->lshr(ShiftAmount); | |||
6764 | Upper = *C + 1; | |||
6765 | } | |||
6766 | break; | |||
6767 | ||||
6768 | case Instruction::Shl: | |||
6769 | if (match(BO.getOperand(0), m_APInt(C))) { | |||
6770 | if (IIQ.hasNoUnsignedWrap(&BO)) { | |||
6771 | // 'shl nuw C, x' produces [C, C << CLZ(C)] | |||
6772 | Lower = *C; | |||
6773 | Upper = Lower.shl(Lower.countLeadingZeros()) + 1; | |||
6774 | } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? | |||
6775 | if (C->isNegative()) { | |||
6776 | // 'shl nsw C, x' produces [C << CLO(C)-1, C] | |||
6777 | unsigned ShiftAmount = C->countLeadingOnes() - 1; | |||
6778 | Lower = C->shl(ShiftAmount); | |||
6779 | Upper = *C + 1; | |||
6780 | } else { | |||
6781 | // 'shl nsw C, x' produces [C, C << CLZ(C)-1] | |||
6782 | unsigned ShiftAmount = C->countLeadingZeros() - 1; | |||
6783 | Lower = *C; | |||
6784 | Upper = C->shl(ShiftAmount) + 1; | |||
6785 | } | |||
6786 | } | |||
6787 | } | |||
6788 | break; | |||
6789 | ||||
6790 | case Instruction::SDiv: | |||
6791 | if (match(BO.getOperand(1), m_APInt(C))) { | |||
6792 | APInt IntMin = APInt::getSignedMinValue(Width); | |||
6793 | APInt IntMax = APInt::getSignedMaxValue(Width); | |||
6794 | if (C->isAllOnes()) { | |||
6795 | // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] | |||
6796 | // where C != -1 and C != 0 and C != 1 | |||
6797 | Lower = IntMin + 1; | |||
6798 | Upper = IntMax + 1; | |||
6799 | } else if (C->countLeadingZeros() < Width - 1) { | |||
6800 | // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] | |||
6801 | // where C != -1 and C != 0 and C != 1 | |||
6802 | Lower = IntMin.sdiv(*C); | |||
6803 | Upper = IntMax.sdiv(*C); | |||
6804 | if (Lower.sgt(Upper)) | |||
6805 | std::swap(Lower, Upper); | |||
6806 | Upper = Upper + 1; | |||
6807 | assert(Upper != Lower && "Upper part of range has wrapped!")(static_cast <bool> (Upper != Lower && "Upper part of range has wrapped!" ) ? void (0) : __assert_fail ("Upper != Lower && \"Upper part of range has wrapped!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6807, __extension__ __PRETTY_FUNCTION__ )); | |||
6808 | } | |||
6809 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
6810 | if (C->isMinSignedValue()) { | |||
6811 | // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. | |||
6812 | Lower = *C; | |||
6813 | Upper = Lower.lshr(1) + 1; | |||
6814 | } else { | |||
6815 | // 'sdiv C, x' produces [-|C|, |C|]. | |||
6816 | Upper = C->abs() + 1; | |||
6817 | Lower = (-Upper) + 1; | |||
6818 | } | |||
6819 | } | |||
6820 | break; | |||
6821 | ||||
6822 | case Instruction::UDiv: | |||
6823 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) { | |||
6824 | // 'udiv x, C' produces [0, UINT_MAX / C]. | |||
6825 | Upper = APInt::getMaxValue(Width).udiv(*C) + 1; | |||
6826 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
6827 | // 'udiv C, x' produces [0, C]. | |||
6828 | Upper = *C + 1; | |||
6829 | } | |||
6830 | break; | |||
6831 | ||||
6832 | case Instruction::SRem: | |||
6833 | if (match(BO.getOperand(1), m_APInt(C))) { | |||
6834 | // 'srem x, C' produces (-|C|, |C|). | |||
6835 | Upper = C->abs(); | |||
6836 | Lower = (-Upper) + 1; | |||
6837 | } | |||
6838 | break; | |||
6839 | ||||
6840 | case Instruction::URem: | |||
6841 | if (match(BO.getOperand(1), m_APInt(C))) | |||
6842 | // 'urem x, C' produces [0, C). | |||
6843 | Upper = *C; | |||
6844 | break; | |||
6845 | ||||
6846 | default: | |||
6847 | break; | |||
6848 | } | |||
6849 | } | |||
6850 | ||||
6851 | static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower, | |||
6852 | APInt &Upper) { | |||
6853 | unsigned Width = Lower.getBitWidth(); | |||
6854 | const APInt *C; | |||
6855 | switch (II.getIntrinsicID()) { | |||
6856 | case Intrinsic::ctpop: | |||
6857 | case Intrinsic::ctlz: | |||
6858 | case Intrinsic::cttz: | |||
6859 | // Maximum of set/clear bits is the bit width. | |||
6860 | assert(Lower == 0 && "Expected lower bound to be zero")(static_cast <bool> (Lower == 0 && "Expected lower bound to be zero" ) ? void (0) : __assert_fail ("Lower == 0 && \"Expected lower bound to be zero\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6860, __extension__ __PRETTY_FUNCTION__ )); | |||
6861 | Upper = Width + 1; | |||
6862 | break; | |||
6863 | case Intrinsic::uadd_sat: | |||
6864 | // uadd.sat(x, C) produces [C, UINT_MAX]. | |||
6865 | if (match(II.getOperand(0), m_APInt(C)) || | |||
6866 | match(II.getOperand(1), m_APInt(C))) | |||
6867 | Lower = *C; | |||
6868 | break; | |||
6869 | case Intrinsic::sadd_sat: | |||
6870 | if (match(II.getOperand(0), m_APInt(C)) || | |||
6871 | match(II.getOperand(1), m_APInt(C))) { | |||
6872 | if (C->isNegative()) { | |||
6873 | // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. | |||
6874 | Lower = APInt::getSignedMinValue(Width); | |||
6875 | Upper = APInt::getSignedMaxValue(Width) + *C + 1; | |||
6876 | } else { | |||
6877 | // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. | |||
6878 | Lower = APInt::getSignedMinValue(Width) + *C; | |||
6879 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6880 | } | |||
6881 | } | |||
6882 | break; | |||
6883 | case Intrinsic::usub_sat: | |||
6884 | // usub.sat(C, x) produces [0, C]. | |||
6885 | if (match(II.getOperand(0), m_APInt(C))) | |||
6886 | Upper = *C + 1; | |||
6887 | // usub.sat(x, C) produces [0, UINT_MAX - C]. | |||
6888 | else if (match(II.getOperand(1), m_APInt(C))) | |||
6889 | Upper = APInt::getMaxValue(Width) - *C + 1; | |||
6890 | break; | |||
6891 | case Intrinsic::ssub_sat: | |||
6892 | if (match(II.getOperand(0), m_APInt(C))) { | |||
6893 | if (C->isNegative()) { | |||
6894 | // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. | |||
6895 | Lower = APInt::getSignedMinValue(Width); | |||
6896 | Upper = *C - APInt::getSignedMinValue(Width) + 1; | |||
6897 | } else { | |||
6898 | // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. | |||
6899 | Lower = *C - APInt::getSignedMaxValue(Width); | |||
6900 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6901 | } | |||
6902 | } else if (match(II.getOperand(1), m_APInt(C))) { | |||
6903 | if (C->isNegative()) { | |||
6904 | // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: | |||
6905 | Lower = APInt::getSignedMinValue(Width) - *C; | |||
6906 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6907 | } else { | |||
6908 | // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. | |||
6909 | Lower = APInt::getSignedMinValue(Width); | |||
6910 | Upper = APInt::getSignedMaxValue(Width) - *C + 1; | |||
6911 | } | |||
6912 | } | |||
6913 | break; | |||
6914 | case Intrinsic::umin: | |||
6915 | case Intrinsic::umax: | |||
6916 | case Intrinsic::smin: | |||
6917 | case Intrinsic::smax: | |||
6918 | if (!match(II.getOperand(0), m_APInt(C)) && | |||
6919 | !match(II.getOperand(1), m_APInt(C))) | |||
6920 | break; | |||
6921 | ||||
6922 | switch (II.getIntrinsicID()) { | |||
6923 | case Intrinsic::umin: | |||
6924 | Upper = *C + 1; | |||
6925 | break; | |||
6926 | case Intrinsic::umax: | |||
6927 | Lower = *C; | |||
6928 | break; | |||
6929 | case Intrinsic::smin: | |||
6930 | Lower = APInt::getSignedMinValue(Width); | |||
6931 | Upper = *C + 1; | |||
6932 | break; | |||
6933 | case Intrinsic::smax: | |||
6934 | Lower = *C; | |||
6935 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6936 | break; | |||
6937 | default: | |||
6938 | llvm_unreachable("Must be min/max intrinsic")::llvm::llvm_unreachable_internal("Must be min/max intrinsic" , "llvm/lib/Analysis/ValueTracking.cpp", 6938); | |||
6939 | } | |||
6940 | break; | |||
6941 | case Intrinsic::abs: | |||
6942 | // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], | |||
6943 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | |||
6944 | if (match(II.getOperand(1), m_One())) | |||
6945 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
6946 | else | |||
6947 | Upper = APInt::getSignedMinValue(Width) + 1; | |||
6948 | break; | |||
6949 | default: | |||
6950 | break; | |||
6951 | } | |||
6952 | } | |||
6953 | ||||
6954 | static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower, | |||
6955 | APInt &Upper, const InstrInfoQuery &IIQ) { | |||
6956 | const Value *LHS = nullptr, *RHS = nullptr; | |||
6957 | SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS); | |||
6958 | if (R.Flavor == SPF_UNKNOWN) | |||
6959 | return; | |||
6960 | ||||
6961 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); | |||
6962 | ||||
6963 | if (R.Flavor == SelectPatternFlavor::SPF_ABS) { | |||
6964 | // If the negation part of the abs (in RHS) has the NSW flag, | |||
6965 | // then the result of abs(X) is [0..SIGNED_MAX], | |||
6966 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | |||
6967 | Lower = APInt::getZero(BitWidth); | |||
6968 | if (match(RHS, m_Neg(m_Specific(LHS))) && | |||
6969 | IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) | |||
6970 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | |||
6971 | else | |||
6972 | Upper = APInt::getSignedMinValue(BitWidth) + 1; | |||
6973 | return; | |||
6974 | } | |||
6975 | ||||
6976 | if (R.Flavor == SelectPatternFlavor::SPF_NABS) { | |||
6977 | // The result of -abs(X) is <= 0. | |||
6978 | Lower = APInt::getSignedMinValue(BitWidth); | |||
6979 | Upper = APInt(BitWidth, 1); | |||
6980 | return; | |||
6981 | } | |||
6982 | ||||
6983 | const APInt *C; | |||
6984 | if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C))) | |||
6985 | return; | |||
6986 | ||||
6987 | switch (R.Flavor) { | |||
6988 | case SPF_UMIN: | |||
6989 | Upper = *C + 1; | |||
6990 | break; | |||
6991 | case SPF_UMAX: | |||
6992 | Lower = *C; | |||
6993 | break; | |||
6994 | case SPF_SMIN: | |||
6995 | Lower = APInt::getSignedMinValue(BitWidth); | |||
6996 | Upper = *C + 1; | |||
6997 | break; | |||
6998 | case SPF_SMAX: | |||
6999 | Lower = *C; | |||
7000 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | |||
7001 | break; | |||
7002 | default: | |||
7003 | break; | |||
7004 | } | |||
7005 | } | |||
7006 | ||||
7007 | static void setLimitForFPToI(const Instruction *I, APInt &Lower, APInt &Upper) { | |||
7008 | // The maximum representable value of a half is 65504. For floats the maximum | |||
7009 | // value is 3.4e38 which requires roughly 129 bits. | |||
7010 | unsigned BitWidth = I->getType()->getScalarSizeInBits(); | |||
7011 | if (!I->getOperand(0)->getType()->getScalarType()->isHalfTy()) | |||
7012 | return; | |||
7013 | if (isa<FPToSIInst>(I) && BitWidth >= 17) { | |||
7014 | Lower = APInt(BitWidth, -65504); | |||
7015 | Upper = APInt(BitWidth, 65505); | |||
7016 | } | |||
7017 | ||||
7018 | if (isa<FPToUIInst>(I) && BitWidth >= 16) { | |||
7019 | // For a fptoui the lower limit is left as 0. | |||
7020 | Upper = APInt(BitWidth, 65505); | |||
7021 | } | |||
7022 | } | |||
7023 | ||||
7024 | ConstantRange llvm::computeConstantRange(const Value *V, bool ForSigned, | |||
7025 | bool UseInstrInfo, AssumptionCache *AC, | |||
7026 | const Instruction *CtxI, | |||
7027 | const DominatorTree *DT, | |||
7028 | unsigned Depth) { | |||
7029 | assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction")(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && "Expected integer instruction") ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && \"Expected integer instruction\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7029, __extension__ __PRETTY_FUNCTION__ )); | |||
7030 | ||||
7031 | if (Depth == MaxAnalysisRecursionDepth) | |||
7032 | return ConstantRange::getFull(V->getType()->getScalarSizeInBits()); | |||
7033 | ||||
7034 | const APInt *C; | |||
7035 | if (match(V, m_APInt(C))) | |||
7036 | return ConstantRange(*C); | |||
7037 | ||||
7038 | InstrInfoQuery IIQ(UseInstrInfo); | |||
7039 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | |||
7040 | APInt Lower = APInt(BitWidth, 0); | |||
7041 | APInt Upper = APInt(BitWidth, 0); | |||
7042 | if (auto *BO = dyn_cast<BinaryOperator>(V)) | |||
7043 | setLimitsForBinOp(*BO, Lower, Upper, IIQ, ForSigned); | |||
7044 | else if (auto *II = dyn_cast<IntrinsicInst>(V)) | |||
7045 | setLimitsForIntrinsic(*II, Lower, Upper); | |||
7046 | else if (auto *SI = dyn_cast<SelectInst>(V)) | |||
7047 | setLimitsForSelectPattern(*SI, Lower, Upper, IIQ); | |||
7048 | else if (isa<FPToUIInst>(V) || isa<FPToSIInst>(V)) | |||
7049 | setLimitForFPToI(cast<Instruction>(V), Lower, Upper); | |||
7050 | ||||
7051 | ConstantRange CR = ConstantRange::getNonEmpty(Lower, Upper); | |||
7052 | ||||
7053 | if (auto *I = dyn_cast<Instruction>(V)) | |||
7054 | if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range)) | |||
7055 | CR = CR.intersectWith(getConstantRangeFromMetadata(*Range)); | |||
7056 | ||||
7057 | if (CtxI && AC) { | |||
7058 | // Try to restrict the range based on information from assumptions. | |||
7059 | for (auto &AssumeVH : AC->assumptionsFor(V)) { | |||
7060 | if (!AssumeVH) | |||
7061 | continue; | |||
7062 | CallInst *I = cast<CallInst>(AssumeVH); | |||
7063 | assert(I->getParent()->getParent() == CtxI->getParent()->getParent() &&(static_cast <bool> (I->getParent()->getParent() == CtxI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == CtxI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7064, __extension__ __PRETTY_FUNCTION__ )) | |||
7064 | "Got assumption for the wrong function!")(static_cast <bool> (I->getParent()->getParent() == CtxI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == CtxI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7064, __extension__ __PRETTY_FUNCTION__ )); | |||
7065 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7066, __extension__ __PRETTY_FUNCTION__ )) | |||
7066 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7066, __extension__ __PRETTY_FUNCTION__ )); | |||
7067 | ||||
7068 | if (!isValidAssumeForContext(I, CtxI, DT)) | |||
7069 | continue; | |||
7070 | Value *Arg = I->getArgOperand(0); | |||
7071 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | |||
7072 | // Currently we just use information from comparisons. | |||
7073 | if (!Cmp || Cmp->getOperand(0) != V) | |||
7074 | continue; | |||
7075 | // TODO: Set "ForSigned" parameter via Cmp->isSigned()? | |||
7076 | ConstantRange RHS = | |||
7077 | computeConstantRange(Cmp->getOperand(1), /* ForSigned */ false, | |||
7078 | UseInstrInfo, AC, I, DT, Depth + 1); | |||
7079 | CR = CR.intersectWith( | |||
7080 | ConstantRange::makeAllowedICmpRegion(Cmp->getPredicate(), RHS)); | |||
7081 | } | |||
7082 | } | |||
7083 | ||||
7084 | return CR; | |||
7085 | } | |||
7086 | ||||
7087 | static Optional<int64_t> | |||
7088 | getOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, const DataLayout &DL) { | |||
7089 | // Skip over the first indices. | |||
7090 | gep_type_iterator GTI = gep_type_begin(GEP); | |||
7091 | for (unsigned i = 1; i != Idx; ++i, ++GTI) | |||
7092 | /*skip along*/; | |||
7093 | ||||
7094 | // Compute the offset implied by the rest of the indices. | |||
7095 | int64_t Offset = 0; | |||
7096 | for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { | |||
7097 | ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); | |||
7098 | if (!OpC) | |||
7099 | return None; | |||
7100 | if (OpC->isZero()) | |||
7101 | continue; // No offset. | |||
7102 | ||||
7103 | // Handle struct indices, which add their field offset to the pointer. | |||
7104 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
7105 | Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); | |||
7106 | continue; | |||
7107 | } | |||
7108 | ||||
7109 | // Otherwise, we have a sequential type like an array or fixed-length | |||
7110 | // vector. Multiply the index by the ElementSize. | |||
7111 | TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType()); | |||
7112 | if (Size.isScalable()) | |||
7113 | return None; | |||
7114 | Offset += Size.getFixedSize() * OpC->getSExtValue(); | |||
7115 | } | |||
7116 | ||||
7117 | return Offset; | |||
7118 | } | |||
7119 | ||||
7120 | Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2, | |||
7121 | const DataLayout &DL) { | |||
7122 | APInt Offset1(DL.getIndexTypeSizeInBits(Ptr1->getType()), 0); | |||
7123 | APInt Offset2(DL.getIndexTypeSizeInBits(Ptr2->getType()), 0); | |||
7124 | Ptr1 = Ptr1->stripAndAccumulateConstantOffsets(DL, Offset1, true); | |||
7125 | Ptr2 = Ptr2->stripAndAccumulateConstantOffsets(DL, Offset2, true); | |||
7126 | ||||
7127 | // Handle the trivial case first. | |||
7128 | if (Ptr1 == Ptr2) | |||
7129 | return Offset2.getSExtValue() - Offset1.getSExtValue(); | |||
7130 | ||||
7131 | const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); | |||
7132 | const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); | |||
7133 | ||||
7134 | // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical | |||
7135 | // base. After that base, they may have some number of common (and | |||
7136 | // potentially variable) indices. After that they handle some constant | |||
7137 | // offset, which determines their offset from each other. At this point, we | |||
7138 | // handle no other case. | |||
7139 | if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0) || | |||
7140 | GEP1->getSourceElementType() != GEP2->getSourceElementType()) | |||
7141 | return None; | |||
7142 | ||||
7143 | // Skip any common indices and track the GEP types. | |||
7144 | unsigned Idx = 1; | |||
7145 | for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) | |||
7146 | if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) | |||
7147 | break; | |||
7148 | ||||
7149 | auto IOffset1 = getOffsetFromIndex(GEP1, Idx, DL); | |||
7150 | auto IOffset2 = getOffsetFromIndex(GEP2, Idx, DL); | |||
7151 | if (!IOffset1 || !IOffset2) | |||
7152 | return None; | |||
7153 | return *IOffset2 - *IOffset1 + Offset2.getSExtValue() - | |||
7154 | Offset1.getSExtValue(); | |||
7155 | } |