1 | //===- ValueTracking.cpp - Walk computations to compute properties --------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file contains routines that help analyze properties that chains of |
10 | // computations have. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/Analysis/ValueTracking.h" |
15 | #include "llvm/ADT/APFloat.h" |
16 | #include "llvm/ADT/APInt.h" |
17 | #include "llvm/ADT/ArrayRef.h" |
18 | #include "llvm/ADT/STLExtras.h" |
19 | #include "llvm/ADT/ScopeExit.h" |
20 | #include "llvm/ADT/SmallPtrSet.h" |
21 | #include "llvm/ADT/SmallSet.h" |
22 | #include "llvm/ADT/SmallVector.h" |
23 | #include "llvm/ADT/StringRef.h" |
24 | #include "llvm/ADT/iterator_range.h" |
25 | #include "llvm/Analysis/AliasAnalysis.h" |
26 | #include "llvm/Analysis/AssumeBundleQueries.h" |
27 | #include "llvm/Analysis/AssumptionCache.h" |
28 | #include "llvm/Analysis/ConstantFolding.h" |
29 | #include "llvm/Analysis/DomConditionCache.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/Analysis/VectorUtils.h" |
37 | #include "llvm/Analysis/WithCache.h" |
38 | #include "llvm/IR/Argument.h" |
39 | #include "llvm/IR/Attributes.h" |
40 | #include "llvm/IR/BasicBlock.h" |
41 | #include "llvm/IR/Constant.h" |
42 | #include "llvm/IR/ConstantRange.h" |
43 | #include "llvm/IR/Constants.h" |
44 | #include "llvm/IR/DerivedTypes.h" |
45 | #include "llvm/IR/DiagnosticInfo.h" |
46 | #include "llvm/IR/Dominators.h" |
47 | #include "llvm/IR/EHPersonalities.h" |
48 | #include "llvm/IR/Function.h" |
49 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
50 | #include "llvm/IR/GlobalAlias.h" |
51 | #include "llvm/IR/GlobalValue.h" |
52 | #include "llvm/IR/GlobalVariable.h" |
53 | #include "llvm/IR/InstrTypes.h" |
54 | #include "llvm/IR/Instruction.h" |
55 | #include "llvm/IR/Instructions.h" |
56 | #include "llvm/IR/IntrinsicInst.h" |
57 | #include "llvm/IR/Intrinsics.h" |
58 | #include "llvm/IR/IntrinsicsAArch64.h" |
59 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
60 | #include "llvm/IR/IntrinsicsRISCV.h" |
61 | #include "llvm/IR/IntrinsicsX86.h" |
62 | #include "llvm/IR/LLVMContext.h" |
63 | #include "llvm/IR/Metadata.h" |
64 | #include "llvm/IR/Module.h" |
65 | #include "llvm/IR/Operator.h" |
66 | #include "llvm/IR/PatternMatch.h" |
67 | #include "llvm/IR/Type.h" |
68 | #include "llvm/IR/User.h" |
69 | #include "llvm/IR/Value.h" |
70 | #include "llvm/Support/Casting.h" |
71 | #include "llvm/Support/CommandLine.h" |
72 | #include "llvm/Support/Compiler.h" |
73 | #include "llvm/Support/ErrorHandling.h" |
74 | #include "llvm/Support/KnownBits.h" |
75 | #include "llvm/Support/MathExtras.h" |
76 | #include "llvm/TargetParser/RISCVTargetParser.h" |
77 | #include <algorithm> |
78 | #include <cassert> |
79 | #include <cstdint> |
80 | #include <optional> |
81 | #include <utility> |
82 | |
83 | using namespace llvm; |
84 | using namespace llvm::PatternMatch; |
85 | |
86 | // Controls the number of uses of the value searched for possible |
87 | // dominating comparisons. |
88 | static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses" , |
89 | cl::Hidden, cl::init(Val: 20)); |
90 | |
91 | |
92 | /// Returns the bitwidth of the given scalar or pointer type. For vector types, |
93 | /// returns the element type's bitwidth. |
94 | static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { |
95 | if (unsigned BitWidth = Ty->getScalarSizeInBits()) |
96 | return BitWidth; |
97 | |
98 | return DL.getPointerTypeSizeInBits(Ty); |
99 | } |
100 | |
101 | // Given the provided Value and, potentially, a context instruction, return |
102 | // the preferred context instruction (if any). |
103 | static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { |
104 | // If we've been provided with a context instruction, then use that (provided |
105 | // it has been inserted). |
106 | if (CxtI && CxtI->getParent()) |
107 | return CxtI; |
108 | |
109 | // If the value is really an already-inserted instruction, then use that. |
110 | CxtI = dyn_cast<Instruction>(Val: V); |
111 | if (CxtI && CxtI->getParent()) |
112 | return CxtI; |
113 | |
114 | return nullptr; |
115 | } |
116 | |
117 | static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) { |
118 | // If we've been provided with a context instruction, then use that (provided |
119 | // it has been inserted). |
120 | if (CxtI && CxtI->getParent()) |
121 | return CxtI; |
122 | |
123 | // If the value is really an already-inserted instruction, then use that. |
124 | CxtI = dyn_cast<Instruction>(Val: V1); |
125 | if (CxtI && CxtI->getParent()) |
126 | return CxtI; |
127 | |
128 | CxtI = dyn_cast<Instruction>(Val: V2); |
129 | if (CxtI && CxtI->getParent()) |
130 | return CxtI; |
131 | |
132 | return nullptr; |
133 | } |
134 | |
135 | static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, |
136 | const APInt &DemandedElts, |
137 | APInt &DemandedLHS, APInt &DemandedRHS) { |
138 | if (isa<ScalableVectorType>(Val: Shuf->getType())) { |
139 | assert(DemandedElts == APInt(1,1)); |
140 | DemandedLHS = DemandedRHS = DemandedElts; |
141 | return true; |
142 | } |
143 | |
144 | int NumElts = |
145 | cast<FixedVectorType>(Val: Shuf->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
146 | return llvm::getShuffleDemandedElts(SrcWidth: NumElts, Mask: Shuf->getShuffleMask(), |
147 | DemandedElts, DemandedLHS, DemandedRHS); |
148 | } |
149 | |
150 | static void computeKnownBits(const Value *V, const APInt &DemandedElts, |
151 | KnownBits &Known, unsigned Depth, |
152 | const SimplifyQuery &Q); |
153 | |
154 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, |
155 | const SimplifyQuery &Q) { |
156 | // Since the number of lanes in a scalable vector is unknown at compile time, |
157 | // we track one bit which is implicitly broadcast to all lanes. This means |
158 | // that all lanes in a scalable vector are considered demanded. |
159 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
160 | APInt DemandedElts = |
161 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
162 | ::computeKnownBits(V, DemandedElts, Known, Depth, Q); |
163 | } |
164 | |
165 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, |
166 | const DataLayout &DL, unsigned Depth, |
167 | AssumptionCache *AC, const Instruction *CxtI, |
168 | const DominatorTree *DT, bool UseInstrInfo) { |
169 | computeKnownBits( |
170 | V, Known, Depth, |
171 | Q: SimplifyQuery(DL, DT, AC, safeCxtI(V, CxtI), UseInstrInfo)); |
172 | } |
173 | |
174 | KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, |
175 | unsigned Depth, AssumptionCache *AC, |
176 | const Instruction *CxtI, |
177 | const DominatorTree *DT, bool UseInstrInfo) { |
178 | return computeKnownBits( |
179 | V, Depth, Q: SimplifyQuery(DL, DT, AC, safeCxtI(V, CxtI), UseInstrInfo)); |
180 | } |
181 | |
182 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, |
183 | const DataLayout &DL, unsigned Depth, |
184 | AssumptionCache *AC, const Instruction *CxtI, |
185 | const DominatorTree *DT, bool UseInstrInfo) { |
186 | return computeKnownBits( |
187 | V, DemandedElts, Depth, |
188 | Q: SimplifyQuery(DL, DT, AC, safeCxtI(V, CxtI), UseInstrInfo)); |
189 | } |
190 | |
191 | static bool haveNoCommonBitsSetSpecialCases(const Value *LHS, const Value *RHS, |
192 | const SimplifyQuery &SQ) { |
193 | // Look for an inverted mask: (X & ~M) op (Y & M). |
194 | { |
195 | Value *M; |
196 | if (match(V: LHS, P: m_c_And(L: m_Not(V: m_Value(V&: M)), R: m_Value())) && |
197 | match(V: RHS, P: m_c_And(L: m_Specific(V: M), R: m_Value())) && |
198 | isGuaranteedNotToBeUndef(V: M, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
199 | return true; |
200 | } |
201 | |
202 | // X op (Y & ~X) |
203 | if (match(V: RHS, P: m_c_And(L: m_Not(V: m_Specific(V: LHS)), R: m_Value())) && |
204 | isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
205 | return true; |
206 | |
207 | // X op ((X & Y) ^ Y) -- this is the canonical form of the previous pattern |
208 | // for constant Y. |
209 | Value *Y; |
210 | if (match(V: RHS, |
211 | P: m_c_Xor(L: m_c_And(L: m_Specific(V: LHS), R: m_Value(V&: Y)), R: m_Deferred(V: Y))) && |
212 | isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT) && |
213 | isGuaranteedNotToBeUndef(V: Y, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
214 | return true; |
215 | |
216 | // Peek through extends to find a 'not' of the other side: |
217 | // (ext Y) op ext(~Y) |
218 | if (match(V: LHS, P: m_ZExtOrSExt(Op: m_Value(V&: Y))) && |
219 | match(V: RHS, P: m_ZExtOrSExt(Op: m_Not(V: m_Specific(V: Y)))) && |
220 | isGuaranteedNotToBeUndef(V: Y, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
221 | return true; |
222 | |
223 | // Look for: (A & B) op ~(A | B) |
224 | { |
225 | Value *A, *B; |
226 | if (match(V: LHS, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) && |
227 | match(V: RHS, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B)))) && |
228 | isGuaranteedNotToBeUndef(V: A, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT) && |
229 | isGuaranteedNotToBeUndef(V: B, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
230 | return true; |
231 | } |
232 | |
233 | return false; |
234 | } |
235 | |
236 | bool llvm::haveNoCommonBitsSet(const WithCache<const Value *> &LHSCache, |
237 | const WithCache<const Value *> &RHSCache, |
238 | const SimplifyQuery &SQ) { |
239 | const Value *LHS = LHSCache.getValue(); |
240 | const Value *RHS = RHSCache.getValue(); |
241 | |
242 | assert(LHS->getType() == RHS->getType() && |
243 | "LHS and RHS should have the same type" ); |
244 | assert(LHS->getType()->isIntOrIntVectorTy() && |
245 | "LHS and RHS should be integers" ); |
246 | |
247 | if (haveNoCommonBitsSetSpecialCases(LHS, RHS, SQ) || |
248 | haveNoCommonBitsSetSpecialCases(LHS: RHS, RHS: LHS, SQ)) |
249 | return true; |
250 | |
251 | return KnownBits::haveNoCommonBitsSet(LHS: LHSCache.getKnownBits(Q: SQ), |
252 | RHS: RHSCache.getKnownBits(Q: SQ)); |
253 | } |
254 | |
255 | bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *I) { |
256 | return !I->user_empty() && all_of(Range: I->users(), P: [](const User *U) { |
257 | ICmpInst::Predicate P; |
258 | return match(V: U, P: m_ICmp(Pred&: P, L: m_Value(), R: m_Zero())) && ICmpInst::isEquality(P); |
259 | }); |
260 | } |
261 | |
262 | static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, |
263 | const SimplifyQuery &Q); |
264 | |
265 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, |
266 | bool OrZero, unsigned Depth, |
267 | AssumptionCache *AC, const Instruction *CxtI, |
268 | const DominatorTree *DT, bool UseInstrInfo) { |
269 | return ::isKnownToBeAPowerOfTwo( |
270 | V, OrZero, Depth, |
271 | Q: SimplifyQuery(DL, DT, AC, safeCxtI(V, CxtI), UseInstrInfo)); |
272 | } |
273 | |
274 | static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, |
275 | const SimplifyQuery &Q, unsigned Depth); |
276 | |
277 | bool llvm::isKnownNonNegative(const Value *V, const SimplifyQuery &SQ, |
278 | unsigned Depth) { |
279 | return computeKnownBits(V, Depth, Q: SQ).isNonNegative(); |
280 | } |
281 | |
282 | bool llvm::isKnownPositive(const Value *V, const SimplifyQuery &SQ, |
283 | unsigned Depth) { |
284 | if (auto *CI = dyn_cast<ConstantInt>(Val: V)) |
285 | return CI->getValue().isStrictlyPositive(); |
286 | |
287 | // If `isKnownNonNegative` ever becomes more sophisticated, make sure to keep |
288 | // this updated. |
289 | KnownBits Known = computeKnownBits(V, Depth, Q: SQ); |
290 | return Known.isNonNegative() && |
291 | (Known.isNonZero() || isKnownNonZero(V, Q: SQ, Depth)); |
292 | } |
293 | |
294 | bool llvm::isKnownNegative(const Value *V, const SimplifyQuery &SQ, |
295 | unsigned Depth) { |
296 | return computeKnownBits(V, Depth, Q: SQ).isNegative(); |
297 | } |
298 | |
299 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, |
300 | const SimplifyQuery &Q); |
301 | |
302 | bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, |
303 | const DataLayout &DL, AssumptionCache *AC, |
304 | const Instruction *CxtI, const DominatorTree *DT, |
305 | bool UseInstrInfo) { |
306 | return ::isKnownNonEqual( |
307 | V1, V2, Depth: 0, |
308 | Q: SimplifyQuery(DL, DT, AC, safeCxtI(V1: V2, V2: V1, CxtI), UseInstrInfo)); |
309 | } |
310 | |
311 | bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, |
312 | const SimplifyQuery &SQ, unsigned Depth) { |
313 | KnownBits Known(Mask.getBitWidth()); |
314 | computeKnownBits(V, Known, Depth, Q: SQ); |
315 | return Mask.isSubsetOf(RHS: Known.Zero); |
316 | } |
317 | |
318 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, |
319 | unsigned Depth, const SimplifyQuery &Q); |
320 | |
321 | static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, |
322 | const SimplifyQuery &Q) { |
323 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
324 | APInt DemandedElts = |
325 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
326 | return ComputeNumSignBits(V, DemandedElts, Depth, Q); |
327 | } |
328 | |
329 | unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, |
330 | unsigned Depth, AssumptionCache *AC, |
331 | const Instruction *CxtI, |
332 | const DominatorTree *DT, bool UseInstrInfo) { |
333 | return ::ComputeNumSignBits( |
334 | V, Depth, Q: SimplifyQuery(DL, DT, AC, safeCxtI(V, CxtI), UseInstrInfo)); |
335 | } |
336 | |
337 | unsigned llvm::ComputeMaxSignificantBits(const Value *V, const DataLayout &DL, |
338 | unsigned Depth, AssumptionCache *AC, |
339 | const Instruction *CxtI, |
340 | const DominatorTree *DT) { |
341 | unsigned SignBits = ComputeNumSignBits(V, DL, Depth, AC, CxtI, DT); |
342 | return V->getType()->getScalarSizeInBits() - SignBits + 1; |
343 | } |
344 | |
345 | static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, |
346 | bool NSW, bool NUW, |
347 | const APInt &DemandedElts, |
348 | KnownBits &KnownOut, KnownBits &Known2, |
349 | unsigned Depth, const SimplifyQuery &Q) { |
350 | computeKnownBits(V: Op1, DemandedElts, Known&: KnownOut, Depth: Depth + 1, Q); |
351 | |
352 | // If one operand is unknown and we have no nowrap information, |
353 | // the result will be unknown independently of the second operand. |
354 | if (KnownOut.isUnknown() && !NSW && !NUW) |
355 | return; |
356 | |
357 | computeKnownBits(V: Op0, DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
358 | KnownOut = KnownBits::computeForAddSub(Add, NSW, NUW, LHS: Known2, RHS: KnownOut); |
359 | } |
360 | |
361 | static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, |
362 | const APInt &DemandedElts, KnownBits &Known, |
363 | KnownBits &Known2, unsigned Depth, |
364 | const SimplifyQuery &Q) { |
365 | computeKnownBits(V: Op1, DemandedElts, Known, Depth: Depth + 1, Q); |
366 | computeKnownBits(V: Op0, DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
367 | |
368 | bool isKnownNegative = false; |
369 | bool isKnownNonNegative = false; |
370 | // If the multiplication is known not to overflow, compute the sign bit. |
371 | if (NSW) { |
372 | if (Op0 == Op1) { |
373 | // The product of a number with itself is non-negative. |
374 | isKnownNonNegative = true; |
375 | } else { |
376 | bool isKnownNonNegativeOp1 = Known.isNonNegative(); |
377 | bool isKnownNonNegativeOp0 = Known2.isNonNegative(); |
378 | bool isKnownNegativeOp1 = Known.isNegative(); |
379 | bool isKnownNegativeOp0 = Known2.isNegative(); |
380 | // The product of two numbers with the same sign is non-negative. |
381 | isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || |
382 | (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); |
383 | // The product of a negative number and a non-negative number is either |
384 | // negative or zero. |
385 | if (!isKnownNonNegative) |
386 | isKnownNegative = |
387 | (isKnownNegativeOp1 && isKnownNonNegativeOp0 && |
388 | Known2.isNonZero()) || |
389 | (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); |
390 | } |
391 | } |
392 | |
393 | bool SelfMultiply = Op0 == Op1; |
394 | if (SelfMultiply) |
395 | SelfMultiply &= |
396 | isGuaranteedNotToBeUndef(V: Op0, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT, Depth: Depth + 1); |
397 | Known = KnownBits::mul(LHS: Known, RHS: Known2, NoUndefSelfMultiply: SelfMultiply); |
398 | |
399 | // Only make use of no-wrap flags if we failed to compute the sign bit |
400 | // directly. This matters if the multiplication always overflows, in |
401 | // which case we prefer to follow the result of the direct computation, |
402 | // though as the program is invoking undefined behaviour we can choose |
403 | // whatever we like here. |
404 | if (isKnownNonNegative && !Known.isNegative()) |
405 | Known.makeNonNegative(); |
406 | else if (isKnownNegative && !Known.isNonNegative()) |
407 | Known.makeNegative(); |
408 | } |
409 | |
410 | void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, |
411 | KnownBits &Known) { |
412 | unsigned BitWidth = Known.getBitWidth(); |
413 | unsigned NumRanges = Ranges.getNumOperands() / 2; |
414 | assert(NumRanges >= 1); |
415 | |
416 | Known.Zero.setAllBits(); |
417 | Known.One.setAllBits(); |
418 | |
419 | for (unsigned i = 0; i < NumRanges; ++i) { |
420 | ConstantInt *Lower = |
421 | mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 0)); |
422 | ConstantInt *Upper = |
423 | mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 1)); |
424 | ConstantRange Range(Lower->getValue(), Upper->getValue()); |
425 | |
426 | // The first CommonPrefixBits of all values in Range are equal. |
427 | unsigned CommonPrefixBits = |
428 | (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countl_zero(); |
429 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: CommonPrefixBits); |
430 | APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(width: BitWidth); |
431 | Known.One &= UnsignedMax & Mask; |
432 | Known.Zero &= ~UnsignedMax & Mask; |
433 | } |
434 | } |
435 | |
436 | static bool isEphemeralValueOf(const Instruction *I, const Value *E) { |
437 | SmallVector<const Value *, 16> WorkSet(1, I); |
438 | SmallPtrSet<const Value *, 32> Visited; |
439 | SmallPtrSet<const Value *, 16> EphValues; |
440 | |
441 | // The instruction defining an assumption's condition itself is always |
442 | // considered ephemeral to that assumption (even if it has other |
443 | // non-ephemeral users). See r246696's test case for an example. |
444 | if (is_contained(Range: I->operands(), Element: E)) |
445 | return true; |
446 | |
447 | while (!WorkSet.empty()) { |
448 | const Value *V = WorkSet.pop_back_val(); |
449 | if (!Visited.insert(Ptr: V).second) |
450 | continue; |
451 | |
452 | // If all uses of this value are ephemeral, then so is this value. |
453 | if (llvm::all_of(Range: V->users(), P: [&](const User *U) { |
454 | return EphValues.count(Ptr: U); |
455 | })) { |
456 | if (V == E) |
457 | return true; |
458 | |
459 | if (V == I || (isa<Instruction>(Val: V) && |
460 | !cast<Instruction>(Val: V)->mayHaveSideEffects() && |
461 | !cast<Instruction>(Val: V)->isTerminator())) { |
462 | EphValues.insert(Ptr: V); |
463 | if (const User *U = dyn_cast<User>(Val: V)) |
464 | append_range(C&: WorkSet, R: U->operands()); |
465 | } |
466 | } |
467 | } |
468 | |
469 | return false; |
470 | } |
471 | |
472 | // Is this an intrinsic that cannot be speculated but also cannot trap? |
473 | bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { |
474 | if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Val: I)) |
475 | return CI->isAssumeLikeIntrinsic(); |
476 | |
477 | return false; |
478 | } |
479 | |
480 | bool llvm::isValidAssumeForContext(const Instruction *Inv, |
481 | const Instruction *CxtI, |
482 | const DominatorTree *DT, |
483 | bool AllowEphemerals) { |
484 | // There are two restrictions on the use of an assume: |
485 | // 1. The assume must dominate the context (or the control flow must |
486 | // reach the assume whenever it reaches the context). |
487 | // 2. The context must not be in the assume's set of ephemeral values |
488 | // (otherwise we will use the assume to prove that the condition |
489 | // feeding the assume is trivially true, thus causing the removal of |
490 | // the assume). |
491 | |
492 | if (Inv->getParent() == CxtI->getParent()) { |
493 | // If Inv and CtxI are in the same block, check if the assume (Inv) is first |
494 | // in the BB. |
495 | if (Inv->comesBefore(Other: CxtI)) |
496 | return true; |
497 | |
498 | // Don't let an assume affect itself - this would cause the problems |
499 | // `isEphemeralValueOf` is trying to prevent, and it would also make |
500 | // the loop below go out of bounds. |
501 | if (!AllowEphemerals && Inv == CxtI) |
502 | return false; |
503 | |
504 | // The context comes first, but they're both in the same block. |
505 | // Make sure there is nothing in between that might interrupt |
506 | // the control flow, not even CxtI itself. |
507 | // We limit the scan distance between the assume and its context instruction |
508 | // to avoid a compile-time explosion. This limit is chosen arbitrarily, so |
509 | // it can be adjusted if needed (could be turned into a cl::opt). |
510 | auto Range = make_range(x: CxtI->getIterator(), y: Inv->getIterator()); |
511 | if (!isGuaranteedToTransferExecutionToSuccessor(Range, ScanLimit: 15)) |
512 | return false; |
513 | |
514 | return AllowEphemerals || !isEphemeralValueOf(I: Inv, E: CxtI); |
515 | } |
516 | |
517 | // Inv and CxtI are in different blocks. |
518 | if (DT) { |
519 | if (DT->dominates(Def: Inv, User: CxtI)) |
520 | return true; |
521 | } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { |
522 | // We don't have a DT, but this trivially dominates. |
523 | return true; |
524 | } |
525 | |
526 | return false; |
527 | } |
528 | |
529 | // TODO: cmpExcludesZero misses many cases where `RHS` is non-constant but |
530 | // we still have enough information about `RHS` to conclude non-zero. For |
531 | // example Pred=EQ, RHS=isKnownNonZero. cmpExcludesZero is called in loops |
532 | // so the extra compile time may not be worth it, but possibly a second API |
533 | // should be created for use outside of loops. |
534 | static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { |
535 | // v u> y implies v != 0. |
536 | if (Pred == ICmpInst::ICMP_UGT) |
537 | return true; |
538 | |
539 | // Special-case v != 0 to also handle v != null. |
540 | if (Pred == ICmpInst::ICMP_NE) |
541 | return match(V: RHS, P: m_Zero()); |
542 | |
543 | // All other predicates - rely on generic ConstantRange handling. |
544 | const APInt *C; |
545 | auto Zero = APInt::getZero(numBits: RHS->getType()->getScalarSizeInBits()); |
546 | if (match(V: RHS, P: m_APInt(Res&: C))) { |
547 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, Other: *C); |
548 | return !TrueValues.contains(Val: Zero); |
549 | } |
550 | |
551 | auto *VC = dyn_cast<ConstantDataVector>(Val: RHS); |
552 | if (VC == nullptr) |
553 | return false; |
554 | |
555 | for (unsigned ElemIdx = 0, NElem = VC->getNumElements(); ElemIdx < NElem; |
556 | ++ElemIdx) { |
557 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion( |
558 | Pred, Other: VC->getElementAsAPInt(i: ElemIdx)); |
559 | if (TrueValues.contains(Val: Zero)) |
560 | return false; |
561 | } |
562 | return true; |
563 | } |
564 | |
565 | static bool isKnownNonZeroFromAssume(const Value *V, const SimplifyQuery &Q) { |
566 | // Use of assumptions is context-sensitive. If we don't have a context, we |
567 | // cannot use them! |
568 | if (!Q.AC || !Q.CxtI) |
569 | return false; |
570 | |
571 | for (AssumptionCache::ResultElem &Elem : Q.AC->assumptionsFor(V)) { |
572 | if (!Elem.Assume) |
573 | continue; |
574 | |
575 | AssumeInst *I = cast<AssumeInst>(Val&: Elem.Assume); |
576 | assert(I->getFunction() == Q.CxtI->getFunction() && |
577 | "Got assumption for the wrong function!" ); |
578 | |
579 | if (Elem.Index != AssumptionCache::ExprResultIdx) { |
580 | if (!V->getType()->isPointerTy()) |
581 | continue; |
582 | if (RetainedKnowledge RK = getKnowledgeFromBundle( |
583 | Assume&: *I, BOI: I->bundle_op_info_begin()[Elem.Index])) { |
584 | if (RK.WasOn == V && |
585 | (RK.AttrKind == Attribute::NonNull || |
586 | (RK.AttrKind == Attribute::Dereferenceable && |
587 | !NullPointerIsDefined(F: Q.CxtI->getFunction(), |
588 | AS: V->getType()->getPointerAddressSpace()))) && |
589 | isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) |
590 | return true; |
591 | } |
592 | continue; |
593 | } |
594 | |
595 | // Warning: This loop can end up being somewhat performance sensitive. |
596 | // We're running this loop for once for each value queried resulting in a |
597 | // runtime of ~O(#assumes * #values). |
598 | |
599 | Value *RHS; |
600 | CmpInst::Predicate Pred; |
601 | auto m_V = m_CombineOr(L: m_Specific(V), R: m_PtrToInt(Op: m_Specific(V))); |
602 | if (!match(V: I->getArgOperand(i: 0), P: m_c_ICmp(Pred, L: m_V, R: m_Value(V&: RHS)))) |
603 | return false; |
604 | |
605 | if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) |
606 | return true; |
607 | } |
608 | |
609 | return false; |
610 | } |
611 | |
612 | static void computeKnownBitsFromCmp(const Value *V, CmpInst::Predicate Pred, |
613 | Value *LHS, Value *RHS, KnownBits &Known, |
614 | const SimplifyQuery &Q) { |
615 | if (RHS->getType()->isPointerTy()) { |
616 | // Handle comparison of pointer to null explicitly, as it will not be |
617 | // covered by the m_APInt() logic below. |
618 | if (LHS == V && match(V: RHS, P: m_Zero())) { |
619 | switch (Pred) { |
620 | case ICmpInst::ICMP_EQ: |
621 | Known.setAllZero(); |
622 | break; |
623 | case ICmpInst::ICMP_SGE: |
624 | case ICmpInst::ICMP_SGT: |
625 | Known.makeNonNegative(); |
626 | break; |
627 | case ICmpInst::ICMP_SLT: |
628 | Known.makeNegative(); |
629 | break; |
630 | default: |
631 | break; |
632 | } |
633 | } |
634 | return; |
635 | } |
636 | |
637 | unsigned BitWidth = Known.getBitWidth(); |
638 | auto m_V = |
639 | m_CombineOr(L: m_Specific(V), R: m_PtrToIntSameSize(DL: Q.DL, Op: m_Specific(V))); |
640 | |
641 | Value *Y; |
642 | const APInt *Mask, *C; |
643 | uint64_t ShAmt; |
644 | switch (Pred) { |
645 | case ICmpInst::ICMP_EQ: |
646 | // assume(V = C) |
647 | if (match(V: LHS, P: m_V) && match(V: RHS, P: m_APInt(Res&: C))) { |
648 | Known = Known.unionWith(RHS: KnownBits::makeConstant(C: *C)); |
649 | // assume(V & Mask = C) |
650 | } else if (match(V: LHS, P: m_c_And(L: m_V, R: m_Value(V&: Y))) && |
651 | match(V: RHS, P: m_APInt(Res&: C))) { |
652 | // For one bits in Mask, we can propagate bits from C to V. |
653 | Known.One |= *C; |
654 | if (match(V: Y, P: m_APInt(Res&: Mask))) |
655 | Known.Zero |= ~*C & *Mask; |
656 | // assume(V | Mask = C) |
657 | } else if (match(V: LHS, P: m_c_Or(L: m_V, R: m_Value(V&: Y))) && match(V: RHS, P: m_APInt(Res&: C))) { |
658 | // For zero bits in Mask, we can propagate bits from C to V. |
659 | Known.Zero |= ~*C; |
660 | if (match(V: Y, P: m_APInt(Res&: Mask))) |
661 | Known.One |= *C & ~*Mask; |
662 | // assume(V ^ Mask = C) |
663 | } else if (match(V: LHS, P: m_Xor(L: m_V, R: m_APInt(Res&: Mask))) && |
664 | match(V: RHS, P: m_APInt(Res&: C))) { |
665 | // Equivalent to assume(V == Mask ^ C) |
666 | Known = Known.unionWith(RHS: KnownBits::makeConstant(C: *C ^ *Mask)); |
667 | // assume(V << ShAmt = C) |
668 | } else if (match(V: LHS, P: m_Shl(L: m_V, R: m_ConstantInt(V&: ShAmt))) && |
669 | match(V: RHS, P: m_APInt(Res&: C)) && ShAmt < BitWidth) { |
670 | // For those bits in C that are known, we can propagate them to known |
671 | // bits in V shifted to the right by ShAmt. |
672 | KnownBits RHSKnown = KnownBits::makeConstant(C: *C); |
673 | RHSKnown.Zero.lshrInPlace(ShiftAmt: ShAmt); |
674 | RHSKnown.One.lshrInPlace(ShiftAmt: ShAmt); |
675 | Known = Known.unionWith(RHS: RHSKnown); |
676 | // assume(V >> ShAmt = C) |
677 | } else if (match(V: LHS, P: m_Shr(L: m_V, R: m_ConstantInt(V&: ShAmt))) && |
678 | match(V: RHS, P: m_APInt(Res&: C)) && ShAmt < BitWidth) { |
679 | KnownBits RHSKnown = KnownBits::makeConstant(C: *C); |
680 | // For those bits in RHS that are known, we can propagate them to known |
681 | // bits in V shifted to the right by C. |
682 | Known.Zero |= RHSKnown.Zero << ShAmt; |
683 | Known.One |= RHSKnown.One << ShAmt; |
684 | } |
685 | break; |
686 | case ICmpInst::ICMP_NE: { |
687 | // assume (V & B != 0) where B is a power of 2 |
688 | const APInt *BPow2; |
689 | if (match(V: LHS, P: m_And(L: m_V, R: m_Power2(V&: BPow2))) && match(V: RHS, P: m_Zero())) |
690 | Known.One |= *BPow2; |
691 | break; |
692 | } |
693 | default: |
694 | if (match(V: RHS, P: m_APInt(Res&: C))) { |
695 | const APInt *Offset = nullptr; |
696 | if (match(V: LHS, P: m_CombineOr(L: m_V, R: m_AddLike(L: m_V, R: m_APInt(Res&: Offset))))) { |
697 | ConstantRange LHSRange = ConstantRange::makeAllowedICmpRegion(Pred, Other: *C); |
698 | if (Offset) |
699 | LHSRange = LHSRange.sub(Other: *Offset); |
700 | Known = Known.unionWith(RHS: LHSRange.toKnownBits()); |
701 | } |
702 | // X & Y u> C -> X u> C && Y u> C |
703 | if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) && |
704 | match(V: LHS, P: m_c_And(L: m_V, R: m_Value()))) { |
705 | Known.One.setHighBits( |
706 | (*C + (Pred == ICmpInst::ICMP_UGT)).countLeadingOnes()); |
707 | } |
708 | // X | Y u< C -> X u< C && Y u< C |
709 | if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) && |
710 | match(V: LHS, P: m_c_Or(L: m_V, R: m_Value()))) { |
711 | Known.Zero.setHighBits( |
712 | (*C - (Pred == ICmpInst::ICMP_ULT)).countLeadingZeros()); |
713 | } |
714 | } |
715 | break; |
716 | } |
717 | } |
718 | |
719 | static void computeKnownBitsFromICmpCond(const Value *V, ICmpInst *Cmp, |
720 | KnownBits &Known, |
721 | const SimplifyQuery &SQ, bool Invert) { |
722 | ICmpInst::Predicate Pred = |
723 | Invert ? Cmp->getInversePredicate() : Cmp->getPredicate(); |
724 | Value *LHS = Cmp->getOperand(i_nocapture: 0); |
725 | Value *RHS = Cmp->getOperand(i_nocapture: 1); |
726 | |
727 | // Handle icmp pred (trunc V), C |
728 | if (match(V: LHS, P: m_Trunc(Op: m_Specific(V)))) { |
729 | KnownBits DstKnown(LHS->getType()->getScalarSizeInBits()); |
730 | computeKnownBitsFromCmp(V: LHS, Pred, LHS, RHS, Known&: DstKnown, Q: SQ); |
731 | Known = Known.unionWith(RHS: DstKnown.anyext(BitWidth: Known.getBitWidth())); |
732 | return; |
733 | } |
734 | |
735 | computeKnownBitsFromCmp(V, Pred, LHS, RHS, Known, Q: SQ); |
736 | } |
737 | |
738 | static void computeKnownBitsFromCond(const Value *V, Value *Cond, |
739 | KnownBits &Known, unsigned Depth, |
740 | const SimplifyQuery &SQ, bool Invert) { |
741 | Value *A, *B; |
742 | if (Depth < MaxAnalysisRecursionDepth && |
743 | match(V: Cond, P: m_LogicalOp(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
744 | KnownBits Known2(Known.getBitWidth()); |
745 | KnownBits Known3(Known.getBitWidth()); |
746 | computeKnownBitsFromCond(V, Cond: A, Known&: Known2, Depth: Depth + 1, SQ, Invert); |
747 | computeKnownBitsFromCond(V, Cond: B, Known&: Known3, Depth: Depth + 1, SQ, Invert); |
748 | if (Invert ? match(V: Cond, P: m_LogicalOr(L: m_Value(), R: m_Value())) |
749 | : match(V: Cond, P: m_LogicalAnd(L: m_Value(), R: m_Value()))) |
750 | Known2 = Known2.unionWith(RHS: Known3); |
751 | else |
752 | Known2 = Known2.intersectWith(RHS: Known3); |
753 | Known = Known.unionWith(RHS: Known2); |
754 | } |
755 | |
756 | if (auto *Cmp = dyn_cast<ICmpInst>(Val: Cond)) |
757 | computeKnownBitsFromICmpCond(V, Cmp, Known, SQ, Invert); |
758 | } |
759 | |
760 | void llvm::computeKnownBitsFromContext(const Value *V, KnownBits &Known, |
761 | unsigned Depth, const SimplifyQuery &Q) { |
762 | if (!Q.CxtI) |
763 | return; |
764 | |
765 | if (Q.DC && Q.DT) { |
766 | // Handle dominating conditions. |
767 | for (BranchInst *BI : Q.DC->conditionsFor(V)) { |
768 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
769 | if (Q.DT->dominates(BBE: Edge0, BB: Q.CxtI->getParent())) |
770 | computeKnownBitsFromCond(V, Cond: BI->getCondition(), Known, Depth, SQ: Q, |
771 | /*Invert*/ false); |
772 | |
773 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
774 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CxtI->getParent())) |
775 | computeKnownBitsFromCond(V, Cond: BI->getCondition(), Known, Depth, SQ: Q, |
776 | /*Invert*/ true); |
777 | } |
778 | |
779 | if (Known.hasConflict()) |
780 | Known.resetAll(); |
781 | } |
782 | |
783 | if (!Q.AC) |
784 | return; |
785 | |
786 | unsigned BitWidth = Known.getBitWidth(); |
787 | |
788 | // Note that the patterns below need to be kept in sync with the code |
789 | // in AssumptionCache::updateAffectedValues. |
790 | |
791 | for (AssumptionCache::ResultElem &Elem : Q.AC->assumptionsFor(V)) { |
792 | if (!Elem.Assume) |
793 | continue; |
794 | |
795 | AssumeInst *I = cast<AssumeInst>(Val&: Elem.Assume); |
796 | assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && |
797 | "Got assumption for the wrong function!" ); |
798 | |
799 | if (Elem.Index != AssumptionCache::ExprResultIdx) { |
800 | if (!V->getType()->isPointerTy()) |
801 | continue; |
802 | if (RetainedKnowledge RK = getKnowledgeFromBundle( |
803 | Assume&: *I, BOI: I->bundle_op_info_begin()[Elem.Index])) { |
804 | if (RK.WasOn == V && RK.AttrKind == Attribute::Alignment && |
805 | isPowerOf2_64(Value: RK.ArgValue) && |
806 | isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) |
807 | Known.Zero.setLowBits(Log2_64(Value: RK.ArgValue)); |
808 | } |
809 | continue; |
810 | } |
811 | |
812 | // Warning: This loop can end up being somewhat performance sensitive. |
813 | // We're running this loop for once for each value queried resulting in a |
814 | // runtime of ~O(#assumes * #values). |
815 | |
816 | Value *Arg = I->getArgOperand(i: 0); |
817 | |
818 | if (Arg == V && isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) { |
819 | assert(BitWidth == 1 && "assume operand is not i1?" ); |
820 | (void)BitWidth; |
821 | Known.setAllOnes(); |
822 | return; |
823 | } |
824 | if (match(V: Arg, P: m_Not(V: m_Specific(V))) && |
825 | isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) { |
826 | assert(BitWidth == 1 && "assume operand is not i1?" ); |
827 | (void)BitWidth; |
828 | Known.setAllZero(); |
829 | return; |
830 | } |
831 | |
832 | // The remaining tests are all recursive, so bail out if we hit the limit. |
833 | if (Depth == MaxAnalysisRecursionDepth) |
834 | continue; |
835 | |
836 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Arg); |
837 | if (!Cmp) |
838 | continue; |
839 | |
840 | if (!isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) |
841 | continue; |
842 | |
843 | computeKnownBitsFromICmpCond(V, Cmp, Known, SQ: Q, /*Invert=*/false); |
844 | } |
845 | |
846 | // Conflicting assumption: Undefined behavior will occur on this execution |
847 | // path. |
848 | if (Known.hasConflict()) |
849 | Known.resetAll(); |
850 | } |
851 | |
852 | /// Compute known bits from a shift operator, including those with a |
853 | /// non-constant shift amount. Known is the output of this function. Known2 is a |
854 | /// pre-allocated temporary with the same bit width as Known and on return |
855 | /// contains the known bit of the shift value source. KF is an |
856 | /// operator-specific function that, given the known-bits and a shift amount, |
857 | /// compute the implied known-bits of the shift operator's result respectively |
858 | /// for that shift amount. The results from calling KF are conservatively |
859 | /// combined for all permitted shift amounts. |
860 | static void computeKnownBitsFromShiftOperator( |
861 | const Operator *I, const APInt &DemandedElts, KnownBits &Known, |
862 | KnownBits &Known2, unsigned Depth, const SimplifyQuery &Q, |
863 | function_ref<KnownBits(const KnownBits &, const KnownBits &, bool)> KF) { |
864 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
865 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Depth: Depth + 1, Q); |
866 | // To limit compile-time impact, only query isKnownNonZero() if we know at |
867 | // least something about the shift amount. |
868 | bool ShAmtNonZero = |
869 | Known.isNonZero() || |
870 | (Known.getMaxValue().ult(RHS: Known.getBitWidth()) && |
871 | isKnownNonZero(V: I->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1)); |
872 | Known = KF(Known2, Known, ShAmtNonZero); |
873 | } |
874 | |
875 | static KnownBits |
876 | getKnownBitsFromAndXorOr(const Operator *I, const APInt &DemandedElts, |
877 | const KnownBits &KnownLHS, const KnownBits &KnownRHS, |
878 | unsigned Depth, const SimplifyQuery &Q) { |
879 | unsigned BitWidth = KnownLHS.getBitWidth(); |
880 | KnownBits KnownOut(BitWidth); |
881 | bool IsAnd = false; |
882 | bool HasKnownOne = !KnownLHS.One.isZero() || !KnownRHS.One.isZero(); |
883 | Value *X = nullptr, *Y = nullptr; |
884 | |
885 | switch (I->getOpcode()) { |
886 | case Instruction::And: |
887 | KnownOut = KnownLHS & KnownRHS; |
888 | IsAnd = true; |
889 | // and(x, -x) is common idioms that will clear all but lowest set |
890 | // bit. If we have a single known bit in x, we can clear all bits |
891 | // above it. |
892 | // TODO: instcombine often reassociates independent `and` which can hide |
893 | // this pattern. Try to match and(x, and(-x, y)) / and(and(x, y), -x). |
894 | if (HasKnownOne && match(V: I, P: m_c_And(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))))) { |
895 | // -(-x) == x so using whichever (LHS/RHS) gets us a better result. |
896 | if (KnownLHS.countMaxTrailingZeros() <= KnownRHS.countMaxTrailingZeros()) |
897 | KnownOut = KnownLHS.blsi(); |
898 | else |
899 | KnownOut = KnownRHS.blsi(); |
900 | } |
901 | break; |
902 | case Instruction::Or: |
903 | KnownOut = KnownLHS | KnownRHS; |
904 | break; |
905 | case Instruction::Xor: |
906 | KnownOut = KnownLHS ^ KnownRHS; |
907 | // xor(x, x-1) is common idioms that will clear all but lowest set |
908 | // bit. If we have a single known bit in x, we can clear all bits |
909 | // above it. |
910 | // TODO: xor(x, x-1) is often rewritting as xor(x, x-C) where C != |
911 | // -1 but for the purpose of demanded bits (xor(x, x-C) & |
912 | // Demanded) == (xor(x, x-1) & Demanded). Extend the xor pattern |
913 | // to use arbitrary C if xor(x, x-C) as the same as xor(x, x-1). |
914 | if (HasKnownOne && |
915 | match(V: I, P: m_c_Xor(L: m_Value(V&: X), R: m_c_Add(L: m_Deferred(V: X), R: m_AllOnes())))) { |
916 | const KnownBits &XBits = I->getOperand(i: 0) == X ? KnownLHS : KnownRHS; |
917 | KnownOut = XBits.blsmsk(); |
918 | } |
919 | break; |
920 | default: |
921 | llvm_unreachable("Invalid Op used in 'analyzeKnownBitsFromAndXorOr'" ); |
922 | } |
923 | |
924 | // and(x, add (x, -1)) is a common idiom that always clears the low bit; |
925 | // xor/or(x, add (x, -1)) is an idiom that will always set the low bit. |
926 | // here we handle the more general case of adding any odd number by |
927 | // matching the form and/xor/or(x, add(x, y)) where y is odd. |
928 | // TODO: This could be generalized to clearing any bit set in y where the |
929 | // following bit is known to be unset in y. |
930 | if (!KnownOut.Zero[0] && !KnownOut.One[0] && |
931 | (match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_c_Add(L: m_Deferred(V: X), R: m_Value(V&: Y)))) || |
932 | match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_Sub(L: m_Deferred(V: X), R: m_Value(V&: Y)))) || |
933 | match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_Sub(L: m_Value(V&: Y), R: m_Deferred(V: X)))))) { |
934 | KnownBits KnownY(BitWidth); |
935 | computeKnownBits(V: Y, DemandedElts, Known&: KnownY, Depth: Depth + 1, Q); |
936 | if (KnownY.countMinTrailingOnes() > 0) { |
937 | if (IsAnd) |
938 | KnownOut.Zero.setBit(0); |
939 | else |
940 | KnownOut.One.setBit(0); |
941 | } |
942 | } |
943 | return KnownOut; |
944 | } |
945 | |
946 | // Public so this can be used in `SimplifyDemandedUseBits`. |
947 | KnownBits llvm::analyzeKnownBitsFromAndXorOr(const Operator *I, |
948 | const KnownBits &KnownLHS, |
949 | const KnownBits &KnownRHS, |
950 | unsigned Depth, |
951 | const SimplifyQuery &SQ) { |
952 | auto *FVTy = dyn_cast<FixedVectorType>(Val: I->getType()); |
953 | APInt DemandedElts = |
954 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
955 | |
956 | return getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS, KnownRHS, Depth, |
957 | Q: SQ); |
958 | } |
959 | |
960 | ConstantRange llvm::getVScaleRange(const Function *F, unsigned BitWidth) { |
961 | Attribute Attr = F->getFnAttribute(Attribute::VScaleRange); |
962 | // Without vscale_range, we only know that vscale is non-zero. |
963 | if (!Attr.isValid()) |
964 | return ConstantRange(APInt(BitWidth, 1), APInt::getZero(numBits: BitWidth)); |
965 | |
966 | unsigned AttrMin = Attr.getVScaleRangeMin(); |
967 | // Minimum is larger than vscale width, result is always poison. |
968 | if ((unsigned)llvm::bit_width(Value: AttrMin) > BitWidth) |
969 | return ConstantRange::getEmpty(BitWidth); |
970 | |
971 | APInt Min(BitWidth, AttrMin); |
972 | std::optional<unsigned> AttrMax = Attr.getVScaleRangeMax(); |
973 | if (!AttrMax || (unsigned)llvm::bit_width(Value: *AttrMax) > BitWidth) |
974 | return ConstantRange(Min, APInt::getZero(numBits: BitWidth)); |
975 | |
976 | return ConstantRange(Min, APInt(BitWidth, *AttrMax) + 1); |
977 | } |
978 | |
979 | static void computeKnownBitsFromOperator(const Operator *I, |
980 | const APInt &DemandedElts, |
981 | KnownBits &Known, unsigned Depth, |
982 | const SimplifyQuery &Q) { |
983 | unsigned BitWidth = Known.getBitWidth(); |
984 | |
985 | KnownBits Known2(BitWidth); |
986 | switch (I->getOpcode()) { |
987 | default: break; |
988 | case Instruction::Load: |
989 | if (MDNode *MD = |
990 | Q.IIQ.getMetadata(I: cast<LoadInst>(Val: I), KindID: LLVMContext::MD_range)) |
991 | computeKnownBitsFromRangeMetadata(Ranges: *MD, Known); |
992 | break; |
993 | case Instruction::And: |
994 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Depth: Depth + 1, Q); |
995 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
996 | |
997 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Depth, Q); |
998 | break; |
999 | case Instruction::Or: |
1000 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Depth: Depth + 1, Q); |
1001 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
1002 | |
1003 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Depth, Q); |
1004 | break; |
1005 | case Instruction::Xor: |
1006 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Depth: Depth + 1, Q); |
1007 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
1008 | |
1009 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Depth, Q); |
1010 | break; |
1011 | case Instruction::Mul: { |
1012 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1013 | computeKnownBitsMul(Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, DemandedElts, |
1014 | Known, Known2, Depth, Q); |
1015 | break; |
1016 | } |
1017 | case Instruction::UDiv: { |
1018 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1019 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1020 | Known = |
1021 | KnownBits::udiv(LHS: Known, RHS: Known2, Exact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))); |
1022 | break; |
1023 | } |
1024 | case Instruction::SDiv: { |
1025 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1026 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1027 | Known = |
1028 | KnownBits::sdiv(LHS: Known, RHS: Known2, Exact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))); |
1029 | break; |
1030 | } |
1031 | case Instruction::Select: { |
1032 | auto ComputeForArm = [&](Value *Arm, bool Invert) { |
1033 | KnownBits Res(Known.getBitWidth()); |
1034 | computeKnownBits(V: Arm, Known&: Res, Depth: Depth + 1, Q); |
1035 | // If we have a constant arm, we are done. |
1036 | if (Res.isConstant()) |
1037 | return Res; |
1038 | |
1039 | // See what condition implies about the bits of the two select arms. |
1040 | KnownBits CondRes(Res.getBitWidth()); |
1041 | computeKnownBitsFromCond(V: Arm, Cond: I->getOperand(i: 0), Known&: CondRes, Depth: Depth + 1, SQ: Q, |
1042 | Invert); |
1043 | // If we don't get any information from the condition, no reason to |
1044 | // proceed. |
1045 | if (CondRes.isUnknown()) |
1046 | return Res; |
1047 | |
1048 | // We can have conflict if the condition is dead. I.e if we have |
1049 | // (x | 64) < 32 ? (x | 64) : y |
1050 | // we will have conflict at bit 6 from the condition/the `or`. |
1051 | // In that case just return. Its not particularly important |
1052 | // what we do, as this select is going to be simplified soon. |
1053 | CondRes = CondRes.unionWith(RHS: Res); |
1054 | if (CondRes.hasConflict()) |
1055 | return Res; |
1056 | |
1057 | // Finally make sure the information we found is valid. This is relatively |
1058 | // expensive so it's left for the very end. |
1059 | if (!isGuaranteedNotToBeUndef(V: Arm, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT, Depth: Depth + 1)) |
1060 | return Res; |
1061 | |
1062 | // Finally, we know we get information from the condition and its valid, |
1063 | // so return it. |
1064 | return CondRes; |
1065 | }; |
1066 | // Only known if known in both the LHS and RHS. |
1067 | Known = |
1068 | ComputeForArm(I->getOperand(i: 1), /*Invert=*/false) |
1069 | .intersectWith(RHS: ComputeForArm(I->getOperand(i: 2), /*Invert=*/true)); |
1070 | break; |
1071 | } |
1072 | case Instruction::FPTrunc: |
1073 | case Instruction::FPExt: |
1074 | case Instruction::FPToUI: |
1075 | case Instruction::FPToSI: |
1076 | case Instruction::SIToFP: |
1077 | case Instruction::UIToFP: |
1078 | break; // Can't work with floating point. |
1079 | case Instruction::PtrToInt: |
1080 | case Instruction::IntToPtr: |
1081 | // Fall through and handle them the same as zext/trunc. |
1082 | [[fallthrough]]; |
1083 | case Instruction::ZExt: |
1084 | case Instruction::Trunc: { |
1085 | Type *SrcTy = I->getOperand(i: 0)->getType(); |
1086 | |
1087 | unsigned SrcBitWidth; |
1088 | // Note that we handle pointer operands here because of inttoptr/ptrtoint |
1089 | // which fall through here. |
1090 | Type *ScalarTy = SrcTy->getScalarType(); |
1091 | SrcBitWidth = ScalarTy->isPointerTy() ? |
1092 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : |
1093 | Q.DL.getTypeSizeInBits(Ty: ScalarTy); |
1094 | |
1095 | assert(SrcBitWidth && "SrcBitWidth can't be zero" ); |
1096 | Known = Known.anyextOrTrunc(BitWidth: SrcBitWidth); |
1097 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1098 | if (auto *Inst = dyn_cast<PossiblyNonNegInst>(Val: I); |
1099 | Inst && Inst->hasNonNeg() && !Known.isNegative()) |
1100 | Known.makeNonNegative(); |
1101 | Known = Known.zextOrTrunc(BitWidth); |
1102 | break; |
1103 | } |
1104 | case Instruction::BitCast: { |
1105 | Type *SrcTy = I->getOperand(i: 0)->getType(); |
1106 | if (SrcTy->isIntOrPtrTy() && |
1107 | // TODO: For now, not handling conversions like: |
1108 | // (bitcast i64 %x to <2 x i32>) |
1109 | !I->getType()->isVectorTy()) { |
1110 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1111 | break; |
1112 | } |
1113 | |
1114 | // Handle cast from vector integer type to scalar or vector integer. |
1115 | auto *SrcVecTy = dyn_cast<FixedVectorType>(Val: SrcTy); |
1116 | if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || |
1117 | !I->getType()->isIntOrIntVectorTy() || |
1118 | isa<ScalableVectorType>(Val: I->getType())) |
1119 | break; |
1120 | |
1121 | // Look through a cast from narrow vector elements to wider type. |
1122 | // Examples: v4i32 -> v2i64, v3i8 -> v24 |
1123 | unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); |
1124 | if (BitWidth % SubBitWidth == 0) { |
1125 | // Known bits are automatically intersected across demanded elements of a |
1126 | // vector. So for example, if a bit is computed as known zero, it must be |
1127 | // zero across all demanded elements of the vector. |
1128 | // |
1129 | // For this bitcast, each demanded element of the output is sub-divided |
1130 | // across a set of smaller vector elements in the source vector. To get |
1131 | // the known bits for an entire element of the output, compute the known |
1132 | // bits for each sub-element sequentially. This is done by shifting the |
1133 | // one-set-bit demanded elements parameter across the sub-elements for |
1134 | // consecutive calls to computeKnownBits. We are using the demanded |
1135 | // elements parameter as a mask operator. |
1136 | // |
1137 | // The known bits of each sub-element are then inserted into place |
1138 | // (dependent on endian) to form the full result of known bits. |
1139 | unsigned NumElts = DemandedElts.getBitWidth(); |
1140 | unsigned SubScale = BitWidth / SubBitWidth; |
1141 | APInt SubDemandedElts = APInt::getZero(numBits: NumElts * SubScale); |
1142 | for (unsigned i = 0; i != NumElts; ++i) { |
1143 | if (DemandedElts[i]) |
1144 | SubDemandedElts.setBit(i * SubScale); |
1145 | } |
1146 | |
1147 | KnownBits KnownSrc(SubBitWidth); |
1148 | for (unsigned i = 0; i != SubScale; ++i) { |
1149 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts: SubDemandedElts.shl(shiftAmt: i), Known&: KnownSrc, |
1150 | Depth: Depth + 1, Q); |
1151 | unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i; |
1152 | Known.insertBits(SubBits: KnownSrc, BitPosition: ShiftElt * SubBitWidth); |
1153 | } |
1154 | } |
1155 | break; |
1156 | } |
1157 | case Instruction::SExt: { |
1158 | // Compute the bits in the result that are not present in the input. |
1159 | unsigned SrcBitWidth = I->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
1160 | |
1161 | Known = Known.trunc(BitWidth: SrcBitWidth); |
1162 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1163 | // If the sign bit of the input is known set or clear, then we know the |
1164 | // top bits of the result. |
1165 | Known = Known.sext(BitWidth); |
1166 | break; |
1167 | } |
1168 | case Instruction::Shl: { |
1169 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1170 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1171 | auto KF = [NUW, NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
1172 | bool ShAmtNonZero) { |
1173 | return KnownBits::shl(LHS: KnownVal, RHS: KnownAmt, NUW, NSW, ShAmtNonZero); |
1174 | }; |
1175 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, |
1176 | KF); |
1177 | // Trailing zeros of a right-shifted constant never decrease. |
1178 | const APInt *C; |
1179 | if (match(V: I->getOperand(i: 0), P: m_APInt(Res&: C))) |
1180 | Known.Zero.setLowBits(C->countr_zero()); |
1181 | break; |
1182 | } |
1183 | case Instruction::LShr: { |
1184 | bool Exact = Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)); |
1185 | auto KF = [Exact](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
1186 | bool ShAmtNonZero) { |
1187 | return KnownBits::lshr(LHS: KnownVal, RHS: KnownAmt, ShAmtNonZero, Exact); |
1188 | }; |
1189 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, |
1190 | KF); |
1191 | // Leading zeros of a left-shifted constant never decrease. |
1192 | const APInt *C; |
1193 | if (match(V: I->getOperand(i: 0), P: m_APInt(Res&: C))) |
1194 | Known.Zero.setHighBits(C->countl_zero()); |
1195 | break; |
1196 | } |
1197 | case Instruction::AShr: { |
1198 | bool Exact = Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)); |
1199 | auto KF = [Exact](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
1200 | bool ShAmtNonZero) { |
1201 | return KnownBits::ashr(LHS: KnownVal, RHS: KnownAmt, ShAmtNonZero, Exact); |
1202 | }; |
1203 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, |
1204 | KF); |
1205 | break; |
1206 | } |
1207 | case Instruction::Sub: { |
1208 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1209 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1210 | computeKnownBitsAddSub(Add: false, Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, NUW, |
1211 | DemandedElts, KnownOut&: Known, Known2, Depth, Q); |
1212 | break; |
1213 | } |
1214 | case Instruction::Add: { |
1215 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1216 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
1217 | computeKnownBitsAddSub(Add: true, Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, NUW, |
1218 | DemandedElts, KnownOut&: Known, Known2, Depth, Q); |
1219 | break; |
1220 | } |
1221 | case Instruction::SRem: |
1222 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1223 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1224 | Known = KnownBits::srem(LHS: Known, RHS: Known2); |
1225 | break; |
1226 | |
1227 | case Instruction::URem: |
1228 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1229 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1230 | Known = KnownBits::urem(LHS: Known, RHS: Known2); |
1231 | break; |
1232 | case Instruction::Alloca: |
1233 | Known.Zero.setLowBits(Log2(A: cast<AllocaInst>(Val: I)->getAlign())); |
1234 | break; |
1235 | case Instruction::GetElementPtr: { |
1236 | // Analyze all of the subscripts of this getelementptr instruction |
1237 | // to determine if we can prove known low zero bits. |
1238 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1239 | // Accumulate the constant indices in a separate variable |
1240 | // to minimize the number of calls to computeForAddSub. |
1241 | APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); |
1242 | |
1243 | gep_type_iterator GTI = gep_type_begin(GEP: I); |
1244 | for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { |
1245 | // TrailZ can only become smaller, short-circuit if we hit zero. |
1246 | if (Known.isUnknown()) |
1247 | break; |
1248 | |
1249 | Value *Index = I->getOperand(i); |
1250 | |
1251 | // Handle case when index is zero. |
1252 | Constant *CIndex = dyn_cast<Constant>(Val: Index); |
1253 | if (CIndex && CIndex->isZeroValue()) |
1254 | continue; |
1255 | |
1256 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
1257 | // Handle struct member offset arithmetic. |
1258 | |
1259 | assert(CIndex && |
1260 | "Access to structure field must be known at compile time" ); |
1261 | |
1262 | if (CIndex->getType()->isVectorTy()) |
1263 | Index = CIndex->getSplatValue(); |
1264 | |
1265 | unsigned Idx = cast<ConstantInt>(Val: Index)->getZExtValue(); |
1266 | const StructLayout *SL = Q.DL.getStructLayout(Ty: STy); |
1267 | uint64_t Offset = SL->getElementOffset(Idx); |
1268 | AccConstIndices += Offset; |
1269 | continue; |
1270 | } |
1271 | |
1272 | // Handle array index arithmetic. |
1273 | Type *IndexedTy = GTI.getIndexedType(); |
1274 | if (!IndexedTy->isSized()) { |
1275 | Known.resetAll(); |
1276 | break; |
1277 | } |
1278 | |
1279 | unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); |
1280 | KnownBits IndexBits(IndexBitWidth); |
1281 | computeKnownBits(V: Index, Known&: IndexBits, Depth: Depth + 1, Q); |
1282 | TypeSize IndexTypeSize = GTI.getSequentialElementStride(DL: Q.DL); |
1283 | uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinValue(); |
1284 | KnownBits ScalingFactor(IndexBitWidth); |
1285 | // Multiply by current sizeof type. |
1286 | // &A[i] == A + i * sizeof(*A[i]). |
1287 | if (IndexTypeSize.isScalable()) { |
1288 | // For scalable types the only thing we know about sizeof is |
1289 | // that this is a multiple of the minimum size. |
1290 | ScalingFactor.Zero.setLowBits(llvm::countr_zero(Val: TypeSizeInBytes)); |
1291 | } else if (IndexBits.isConstant()) { |
1292 | APInt IndexConst = IndexBits.getConstant(); |
1293 | APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); |
1294 | IndexConst *= ScalingFactor; |
1295 | AccConstIndices += IndexConst.sextOrTrunc(width: BitWidth); |
1296 | continue; |
1297 | } else { |
1298 | ScalingFactor = |
1299 | KnownBits::makeConstant(C: APInt(IndexBitWidth, TypeSizeInBytes)); |
1300 | } |
1301 | IndexBits = KnownBits::mul(LHS: IndexBits, RHS: ScalingFactor); |
1302 | |
1303 | // If the offsets have a different width from the pointer, according |
1304 | // to the language reference we need to sign-extend or truncate them |
1305 | // to the width of the pointer. |
1306 | IndexBits = IndexBits.sextOrTrunc(BitWidth); |
1307 | |
1308 | // Note that inbounds does *not* guarantee nsw for the addition, as only |
1309 | // the offset is signed, while the base address is unsigned. |
1310 | Known = KnownBits::computeForAddSub( |
1311 | /*Add=*/true, /*NSW=*/false, /* NUW=*/false, LHS: Known, RHS: IndexBits); |
1312 | } |
1313 | if (!Known.isUnknown() && !AccConstIndices.isZero()) { |
1314 | KnownBits Index = KnownBits::makeConstant(C: AccConstIndices); |
1315 | Known = KnownBits::computeForAddSub( |
1316 | /*Add=*/true, /*NSW=*/false, /* NUW=*/false, LHS: Known, RHS: Index); |
1317 | } |
1318 | break; |
1319 | } |
1320 | case Instruction::PHI: { |
1321 | const PHINode *P = cast<PHINode>(Val: I); |
1322 | BinaryOperator *BO = nullptr; |
1323 | Value *R = nullptr, *L = nullptr; |
1324 | if (matchSimpleRecurrence(P, BO, Start&: R, Step&: L)) { |
1325 | // Handle the case of a simple two-predecessor recurrence PHI. |
1326 | // There's a lot more that could theoretically be done here, but |
1327 | // this is sufficient to catch some interesting cases. |
1328 | unsigned Opcode = BO->getOpcode(); |
1329 | |
1330 | // If this is a shift recurrence, we know the bits being shifted in. |
1331 | // We can combine that with information about the start value of the |
1332 | // recurrence to conclude facts about the result. |
1333 | if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr || |
1334 | Opcode == Instruction::Shl) && |
1335 | BO->getOperand(i_nocapture: 0) == I) { |
1336 | |
1337 | // We have matched a recurrence of the form: |
1338 | // %iv = [R, %entry], [%iv.next, %backedge] |
1339 | // %iv.next = shift_op %iv, L |
1340 | |
1341 | // Recurse with the phi context to avoid concern about whether facts |
1342 | // inferred hold at original context instruction. TODO: It may be |
1343 | // correct to use the original context. IF warranted, explore and |
1344 | // add sufficient tests to cover. |
1345 | SimplifyQuery RecQ = Q; |
1346 | RecQ.CxtI = P; |
1347 | computeKnownBits(V: R, DemandedElts, Known&: Known2, Depth: Depth + 1, Q: RecQ); |
1348 | switch (Opcode) { |
1349 | case Instruction::Shl: |
1350 | // A shl recurrence will only increase the tailing zeros |
1351 | Known.Zero.setLowBits(Known2.countMinTrailingZeros()); |
1352 | break; |
1353 | case Instruction::LShr: |
1354 | // A lshr recurrence will preserve the leading zeros of the |
1355 | // start value |
1356 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); |
1357 | break; |
1358 | case Instruction::AShr: |
1359 | // An ashr recurrence will extend the initial sign bit |
1360 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); |
1361 | Known.One.setHighBits(Known2.countMinLeadingOnes()); |
1362 | break; |
1363 | }; |
1364 | } |
1365 | |
1366 | // Check for operations that have the property that if |
1367 | // both their operands have low zero bits, the result |
1368 | // will have low zero bits. |
1369 | if (Opcode == Instruction::Add || |
1370 | Opcode == Instruction::Sub || |
1371 | Opcode == Instruction::And || |
1372 | Opcode == Instruction::Or || |
1373 | Opcode == Instruction::Mul) { |
1374 | // Change the context instruction to the "edge" that flows into the |
1375 | // phi. This is important because that is where the value is actually |
1376 | // "evaluated" even though it is used later somewhere else. (see also |
1377 | // D69571). |
1378 | SimplifyQuery RecQ = Q; |
1379 | |
1380 | unsigned OpNum = P->getOperand(i_nocapture: 0) == R ? 0 : 1; |
1381 | Instruction *RInst = P->getIncomingBlock(i: OpNum)->getTerminator(); |
1382 | Instruction *LInst = P->getIncomingBlock(i: 1-OpNum)->getTerminator(); |
1383 | |
1384 | // Ok, we have a PHI of the form L op= R. Check for low |
1385 | // zero bits. |
1386 | RecQ.CxtI = RInst; |
1387 | computeKnownBits(V: R, Known&: Known2, Depth: Depth + 1, Q: RecQ); |
1388 | |
1389 | // We need to take the minimum number of known bits |
1390 | KnownBits Known3(BitWidth); |
1391 | RecQ.CxtI = LInst; |
1392 | computeKnownBits(V: L, Known&: Known3, Depth: Depth + 1, Q: RecQ); |
1393 | |
1394 | Known.Zero.setLowBits(std::min(a: Known2.countMinTrailingZeros(), |
1395 | b: Known3.countMinTrailingZeros())); |
1396 | |
1397 | auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(Val: BO); |
1398 | if (OverflowOp && Q.IIQ.hasNoSignedWrap(Op: OverflowOp)) { |
1399 | // If initial value of recurrence is nonnegative, and we are adding |
1400 | // a nonnegative number with nsw, the result can only be nonnegative |
1401 | // or poison value regardless of the number of times we execute the |
1402 | // add in phi recurrence. If initial value is negative and we are |
1403 | // adding a negative number with nsw, the result can only be |
1404 | // negative or poison value. Similar arguments apply to sub and mul. |
1405 | // |
1406 | // (add non-negative, non-negative) --> non-negative |
1407 | // (add negative, negative) --> negative |
1408 | if (Opcode == Instruction::Add) { |
1409 | if (Known2.isNonNegative() && Known3.isNonNegative()) |
1410 | Known.makeNonNegative(); |
1411 | else if (Known2.isNegative() && Known3.isNegative()) |
1412 | Known.makeNegative(); |
1413 | } |
1414 | |
1415 | // (sub nsw non-negative, negative) --> non-negative |
1416 | // (sub nsw negative, non-negative) --> negative |
1417 | else if (Opcode == Instruction::Sub && BO->getOperand(i_nocapture: 0) == I) { |
1418 | if (Known2.isNonNegative() && Known3.isNegative()) |
1419 | Known.makeNonNegative(); |
1420 | else if (Known2.isNegative() && Known3.isNonNegative()) |
1421 | Known.makeNegative(); |
1422 | } |
1423 | |
1424 | // (mul nsw non-negative, non-negative) --> non-negative |
1425 | else if (Opcode == Instruction::Mul && Known2.isNonNegative() && |
1426 | Known3.isNonNegative()) |
1427 | Known.makeNonNegative(); |
1428 | } |
1429 | |
1430 | break; |
1431 | } |
1432 | } |
1433 | |
1434 | // Unreachable blocks may have zero-operand PHI nodes. |
1435 | if (P->getNumIncomingValues() == 0) |
1436 | break; |
1437 | |
1438 | // Otherwise take the unions of the known bit sets of the operands, |
1439 | // taking conservative care to avoid excessive recursion. |
1440 | if (Depth < MaxAnalysisRecursionDepth - 1 && Known.isUnknown()) { |
1441 | // Skip if every incoming value references to ourself. |
1442 | if (isa_and_nonnull<UndefValue>(Val: P->hasConstantValue())) |
1443 | break; |
1444 | |
1445 | Known.Zero.setAllBits(); |
1446 | Known.One.setAllBits(); |
1447 | for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { |
1448 | Value *IncValue = P->getIncomingValue(i: u); |
1449 | // Skip direct self references. |
1450 | if (IncValue == P) continue; |
1451 | |
1452 | // Change the context instruction to the "edge" that flows into the |
1453 | // phi. This is important because that is where the value is actually |
1454 | // "evaluated" even though it is used later somewhere else. (see also |
1455 | // D69571). |
1456 | SimplifyQuery RecQ = Q; |
1457 | RecQ.CxtI = P->getIncomingBlock(i: u)->getTerminator(); |
1458 | |
1459 | Known2 = KnownBits(BitWidth); |
1460 | |
1461 | // Recurse, but cap the recursion to one level, because we don't |
1462 | // want to waste time spinning around in loops. |
1463 | // TODO: See if we can base recursion limiter on number of incoming phi |
1464 | // edges so we don't overly clamp analysis. |
1465 | computeKnownBits(V: IncValue, Known&: Known2, Depth: MaxAnalysisRecursionDepth - 1, Q: RecQ); |
1466 | |
1467 | // See if we can further use a conditional branch into the phi |
1468 | // to help us determine the range of the value. |
1469 | if (!Known2.isConstant()) { |
1470 | ICmpInst::Predicate Pred; |
1471 | const APInt *RHSC; |
1472 | BasicBlock *TrueSucc, *FalseSucc; |
1473 | // TODO: Use RHS Value and compute range from its known bits. |
1474 | if (match(V: RecQ.CxtI, |
1475 | P: m_Br(C: m_c_ICmp(Pred, L: m_Specific(V: IncValue), R: m_APInt(Res&: RHSC)), |
1476 | T: m_BasicBlock(V&: TrueSucc), F: m_BasicBlock(V&: FalseSucc)))) { |
1477 | // Check for cases of duplicate successors. |
1478 | if ((TrueSucc == P->getParent()) != (FalseSucc == P->getParent())) { |
1479 | // If we're using the false successor, invert the predicate. |
1480 | if (FalseSucc == P->getParent()) |
1481 | Pred = CmpInst::getInversePredicate(pred: Pred); |
1482 | // Get the knownbits implied by the incoming phi condition. |
1483 | auto CR = ConstantRange::makeExactICmpRegion(Pred, Other: *RHSC); |
1484 | KnownBits KnownUnion = Known2.unionWith(RHS: CR.toKnownBits()); |
1485 | // We can have conflicts here if we are analyzing deadcode (its |
1486 | // impossible for us reach this BB based the icmp). |
1487 | if (KnownUnion.hasConflict()) { |
1488 | // No reason to continue analyzing in a known dead region, so |
1489 | // just resetAll and break. This will cause us to also exit the |
1490 | // outer loop. |
1491 | Known.resetAll(); |
1492 | break; |
1493 | } |
1494 | Known2 = KnownUnion; |
1495 | } |
1496 | } |
1497 | } |
1498 | |
1499 | Known = Known.intersectWith(RHS: Known2); |
1500 | // If all bits have been ruled out, there's no need to check |
1501 | // more operands. |
1502 | if (Known.isUnknown()) |
1503 | break; |
1504 | } |
1505 | } |
1506 | break; |
1507 | } |
1508 | case Instruction::Call: |
1509 | case Instruction::Invoke: { |
1510 | // If range metadata is attached to this call, set known bits from that, |
1511 | // and then intersect with known bits based on other properties of the |
1512 | // function. |
1513 | if (MDNode *MD = |
1514 | Q.IIQ.getMetadata(I: cast<Instruction>(Val: I), KindID: LLVMContext::MD_range)) |
1515 | computeKnownBitsFromRangeMetadata(Ranges: *MD, Known); |
1516 | |
1517 | const auto *CB = cast<CallBase>(Val: I); |
1518 | |
1519 | if (std::optional<ConstantRange> Range = CB->getRange()) |
1520 | Known = Known.unionWith(RHS: Range->toKnownBits()); |
1521 | |
1522 | if (const Value *RV = CB->getReturnedArgOperand()) { |
1523 | if (RV->getType() == I->getType()) { |
1524 | computeKnownBits(V: RV, Known&: Known2, Depth: Depth + 1, Q); |
1525 | Known = Known.unionWith(RHS: Known2); |
1526 | // If the function doesn't return properly for all input values |
1527 | // (e.g. unreachable exits) then there might be conflicts between the |
1528 | // argument value and the range metadata. Simply discard the known bits |
1529 | // in case of conflicts. |
1530 | if (Known.hasConflict()) |
1531 | Known.resetAll(); |
1532 | } |
1533 | } |
1534 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) { |
1535 | switch (II->getIntrinsicID()) { |
1536 | default: break; |
1537 | case Intrinsic::abs: { |
1538 | computeKnownBits(V: I->getOperand(i: 0), Known&: Known2, Depth: Depth + 1, Q); |
1539 | bool IntMinIsPoison = match(V: II->getArgOperand(i: 1), P: m_One()); |
1540 | Known = Known2.abs(IntMinIsPoison); |
1541 | break; |
1542 | } |
1543 | case Intrinsic::bitreverse: |
1544 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
1545 | Known.Zero |= Known2.Zero.reverseBits(); |
1546 | Known.One |= Known2.One.reverseBits(); |
1547 | break; |
1548 | case Intrinsic::bswap: |
1549 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Depth: Depth + 1, Q); |
1550 | Known.Zero |= Known2.Zero.byteSwap(); |
1551 | Known.One |= Known2.One.byteSwap(); |
1552 | break; |
1553 | case Intrinsic::ctlz: { |
1554 | computeKnownBits(V: I->getOperand(i: 0), Known&: Known2, Depth: Depth + 1, Q); |
1555 | // If we have a known 1, its position is our upper bound. |
1556 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); |
1557 | // If this call is poison for 0 input, the result will be less than 2^n. |
1558 | if (II->getArgOperand(i: 1) == ConstantInt::getTrue(Context&: II->getContext())) |
1559 | PossibleLZ = std::min(a: PossibleLZ, b: BitWidth - 1); |
1560 | unsigned LowBits = llvm::bit_width(Value: PossibleLZ); |
1561 | Known.Zero.setBitsFrom(LowBits); |
1562 | break; |
1563 | } |
1564 | case Intrinsic::cttz: { |
1565 | computeKnownBits(V: I->getOperand(i: 0), Known&: Known2, Depth: Depth + 1, Q); |
1566 | // If we have a known 1, its position is our upper bound. |
1567 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); |
1568 | // If this call is poison for 0 input, the result will be less than 2^n. |
1569 | if (II->getArgOperand(i: 1) == ConstantInt::getTrue(Context&: II->getContext())) |
1570 | PossibleTZ = std::min(a: PossibleTZ, b: BitWidth - 1); |
1571 | unsigned LowBits = llvm::bit_width(Value: PossibleTZ); |
1572 | Known.Zero.setBitsFrom(LowBits); |
1573 | break; |
1574 | } |
1575 | case Intrinsic::ctpop: { |
1576 | computeKnownBits(V: I->getOperand(i: 0), Known&: Known2, Depth: Depth + 1, Q); |
1577 | // We can bound the space the count needs. Also, bits known to be zero |
1578 | // can't contribute to the population. |
1579 | unsigned BitsPossiblySet = Known2.countMaxPopulation(); |
1580 | unsigned LowBits = llvm::bit_width(Value: BitsPossiblySet); |
1581 | Known.Zero.setBitsFrom(LowBits); |
1582 | // TODO: we could bound KnownOne using the lower bound on the number |
1583 | // of bits which might be set provided by popcnt KnownOne2. |
1584 | break; |
1585 | } |
1586 | case Intrinsic::fshr: |
1587 | case Intrinsic::fshl: { |
1588 | const APInt *SA; |
1589 | if (!match(V: I->getOperand(i: 2), P: m_APInt(Res&: SA))) |
1590 | break; |
1591 | |
1592 | // Normalize to funnel shift left. |
1593 | uint64_t ShiftAmt = SA->urem(RHS: BitWidth); |
1594 | if (II->getIntrinsicID() == Intrinsic::fshr) |
1595 | ShiftAmt = BitWidth - ShiftAmt; |
1596 | |
1597 | KnownBits Known3(BitWidth); |
1598 | computeKnownBits(V: I->getOperand(i: 0), Known&: Known2, Depth: Depth + 1, Q); |
1599 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known3, Depth: Depth + 1, Q); |
1600 | |
1601 | Known.Zero = |
1602 | Known2.Zero.shl(shiftAmt: ShiftAmt) | Known3.Zero.lshr(shiftAmt: BitWidth - ShiftAmt); |
1603 | Known.One = |
1604 | Known2.One.shl(shiftAmt: ShiftAmt) | Known3.One.lshr(shiftAmt: BitWidth - ShiftAmt); |
1605 | break; |
1606 | } |
1607 | case Intrinsic::uadd_sat: |
1608 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1609 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1610 | Known = KnownBits::uadd_sat(LHS: Known, RHS: Known2); |
1611 | break; |
1612 | case Intrinsic::usub_sat: |
1613 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1614 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1615 | Known = KnownBits::usub_sat(LHS: Known, RHS: Known2); |
1616 | break; |
1617 | case Intrinsic::sadd_sat: |
1618 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1619 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1620 | Known = KnownBits::sadd_sat(LHS: Known, RHS: Known2); |
1621 | break; |
1622 | case Intrinsic::ssub_sat: |
1623 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1624 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1625 | Known = KnownBits::ssub_sat(LHS: Known, RHS: Known2); |
1626 | break; |
1627 | // for min/max/and/or reduce, any bit common to each element in the |
1628 | // input vec is set in the output. |
1629 | case Intrinsic::vector_reduce_and: |
1630 | case Intrinsic::vector_reduce_or: |
1631 | case Intrinsic::vector_reduce_umax: |
1632 | case Intrinsic::vector_reduce_umin: |
1633 | case Intrinsic::vector_reduce_smax: |
1634 | case Intrinsic::vector_reduce_smin: |
1635 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1636 | break; |
1637 | case Intrinsic::vector_reduce_xor: { |
1638 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1639 | // The zeros common to all vecs are zero in the output. |
1640 | // If the number of elements is odd, then the common ones remain. If the |
1641 | // number of elements is even, then the common ones becomes zeros. |
1642 | auto *VecTy = cast<VectorType>(Val: I->getOperand(i: 0)->getType()); |
1643 | // Even, so the ones become zeros. |
1644 | bool EvenCnt = VecTy->getElementCount().isKnownEven(); |
1645 | if (EvenCnt) |
1646 | Known.Zero |= Known.One; |
1647 | // Maybe even element count so need to clear ones. |
1648 | if (VecTy->isScalableTy() || EvenCnt) |
1649 | Known.One.clearAllBits(); |
1650 | break; |
1651 | } |
1652 | case Intrinsic::umin: |
1653 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1654 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1655 | Known = KnownBits::umin(LHS: Known, RHS: Known2); |
1656 | break; |
1657 | case Intrinsic::umax: |
1658 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1659 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1660 | Known = KnownBits::umax(LHS: Known, RHS: Known2); |
1661 | break; |
1662 | case Intrinsic::smin: |
1663 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1664 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1665 | Known = KnownBits::smin(LHS: Known, RHS: Known2); |
1666 | break; |
1667 | case Intrinsic::smax: |
1668 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1669 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Depth: Depth + 1, Q); |
1670 | Known = KnownBits::smax(LHS: Known, RHS: Known2); |
1671 | break; |
1672 | case Intrinsic::ptrmask: { |
1673 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1674 | |
1675 | const Value *Mask = I->getOperand(i: 1); |
1676 | Known2 = KnownBits(Mask->getType()->getScalarSizeInBits()); |
1677 | computeKnownBits(V: Mask, Known&: Known2, Depth: Depth + 1, Q); |
1678 | // TODO: 1-extend would be more precise. |
1679 | Known &= Known2.anyextOrTrunc(BitWidth); |
1680 | break; |
1681 | } |
1682 | case Intrinsic::x86_sse42_crc32_64_64: |
1683 | Known.Zero.setBitsFrom(32); |
1684 | break; |
1685 | case Intrinsic::riscv_vsetvli: |
1686 | case Intrinsic::riscv_vsetvlimax: { |
1687 | bool HasAVL = II->getIntrinsicID() == Intrinsic::riscv_vsetvli; |
1688 | const ConstantRange Range = getVScaleRange(F: II->getFunction(), BitWidth); |
1689 | uint64_t SEW = RISCVVType::decodeVSEW( |
1690 | VSEW: cast<ConstantInt>(Val: II->getArgOperand(i: HasAVL))->getZExtValue()); |
1691 | RISCVII::VLMUL VLMUL = static_cast<RISCVII::VLMUL>( |
1692 | cast<ConstantInt>(Val: II->getArgOperand(i: 1 + HasAVL))->getZExtValue()); |
1693 | // The Range is [Lower, Upper), so we need to subtract 1 here to get the |
1694 | // real upper value. |
1695 | uint64_t MaxVLEN = |
1696 | (Range.getUpper().getZExtValue() - 1) * RISCV::RVVBitsPerBlock; |
1697 | uint64_t MaxVL = MaxVLEN / RISCVVType::getSEWLMULRatio(SEW, VLMul: VLMUL); |
1698 | |
1699 | // Result of vsetvli must be not larger than AVL. |
1700 | if (HasAVL) |
1701 | if (auto *CI = dyn_cast<ConstantInt>(Val: II->getArgOperand(i: 0))) |
1702 | MaxVL = std::min(a: MaxVL, b: CI->getZExtValue()); |
1703 | |
1704 | unsigned KnownZeroFirstBit = Log2_32(Value: MaxVL) + 1; |
1705 | if (BitWidth > KnownZeroFirstBit) |
1706 | Known.Zero.setBitsFrom(KnownZeroFirstBit); |
1707 | break; |
1708 | } |
1709 | case Intrinsic::vscale: { |
1710 | if (!II->getParent() || !II->getFunction()) |
1711 | break; |
1712 | |
1713 | Known = getVScaleRange(F: II->getFunction(), BitWidth).toKnownBits(); |
1714 | break; |
1715 | } |
1716 | } |
1717 | } |
1718 | break; |
1719 | } |
1720 | case Instruction::ShuffleVector: { |
1721 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: I); |
1722 | // FIXME: Do we need to handle ConstantExpr involving shufflevectors? |
1723 | if (!Shuf) { |
1724 | Known.resetAll(); |
1725 | return; |
1726 | } |
1727 | // For undef elements, we don't know anything about the common state of |
1728 | // the shuffle result. |
1729 | APInt DemandedLHS, DemandedRHS; |
1730 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { |
1731 | Known.resetAll(); |
1732 | return; |
1733 | } |
1734 | Known.One.setAllBits(); |
1735 | Known.Zero.setAllBits(); |
1736 | if (!!DemandedLHS) { |
1737 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
1738 | computeKnownBits(V: LHS, DemandedElts: DemandedLHS, Known, Depth: Depth + 1, Q); |
1739 | // If we don't know any bits, early out. |
1740 | if (Known.isUnknown()) |
1741 | break; |
1742 | } |
1743 | if (!!DemandedRHS) { |
1744 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
1745 | computeKnownBits(V: RHS, DemandedElts: DemandedRHS, Known&: Known2, Depth: Depth + 1, Q); |
1746 | Known = Known.intersectWith(RHS: Known2); |
1747 | } |
1748 | break; |
1749 | } |
1750 | case Instruction::InsertElement: { |
1751 | if (isa<ScalableVectorType>(Val: I->getType())) { |
1752 | Known.resetAll(); |
1753 | return; |
1754 | } |
1755 | const Value *Vec = I->getOperand(i: 0); |
1756 | const Value *Elt = I->getOperand(i: 1); |
1757 | auto *CIdx = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2)); |
1758 | unsigned NumElts = DemandedElts.getBitWidth(); |
1759 | APInt DemandedVecElts = DemandedElts; |
1760 | bool NeedsElt = true; |
1761 | // If we know the index we are inserting too, clear it from Vec check. |
1762 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
1763 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
1764 | NeedsElt = DemandedElts[CIdx->getZExtValue()]; |
1765 | } |
1766 | |
1767 | Known.One.setAllBits(); |
1768 | Known.Zero.setAllBits(); |
1769 | if (NeedsElt) { |
1770 | computeKnownBits(V: Elt, Known, Depth: Depth + 1, Q); |
1771 | // If we don't know any bits, early out. |
1772 | if (Known.isUnknown()) |
1773 | break; |
1774 | } |
1775 | |
1776 | if (!DemandedVecElts.isZero()) { |
1777 | computeKnownBits(V: Vec, DemandedElts: DemandedVecElts, Known&: Known2, Depth: Depth + 1, Q); |
1778 | Known = Known.intersectWith(RHS: Known2); |
1779 | } |
1780 | break; |
1781 | } |
1782 | case Instruction::ExtractElement: { |
1783 | // Look through extract element. If the index is non-constant or |
1784 | // out-of-range demand all elements, otherwise just the extracted element. |
1785 | const Value *Vec = I->getOperand(i: 0); |
1786 | const Value *Idx = I->getOperand(i: 1); |
1787 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
1788 | if (isa<ScalableVectorType>(Val: Vec->getType())) { |
1789 | // FIXME: there's probably *something* we can do with scalable vectors |
1790 | Known.resetAll(); |
1791 | break; |
1792 | } |
1793 | unsigned NumElts = cast<FixedVectorType>(Val: Vec->getType())->getNumElements(); |
1794 | APInt DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
1795 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
1796 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
1797 | computeKnownBits(V: Vec, DemandedElts: DemandedVecElts, Known, Depth: Depth + 1, Q); |
1798 | break; |
1799 | } |
1800 | case Instruction::ExtractValue: |
1801 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I->getOperand(i: 0))) { |
1802 | const ExtractValueInst *EVI = cast<ExtractValueInst>(Val: I); |
1803 | if (EVI->getNumIndices() != 1) break; |
1804 | if (EVI->getIndices()[0] == 0) { |
1805 | switch (II->getIntrinsicID()) { |
1806 | default: break; |
1807 | case Intrinsic::uadd_with_overflow: |
1808 | case Intrinsic::sadd_with_overflow: |
1809 | computeKnownBitsAddSub( |
1810 | Add: true, Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), /*NSW=*/false, |
1811 | /* NUW=*/false, DemandedElts, KnownOut&: Known, Known2, Depth, Q); |
1812 | break; |
1813 | case Intrinsic::usub_with_overflow: |
1814 | case Intrinsic::ssub_with_overflow: |
1815 | computeKnownBitsAddSub( |
1816 | Add: false, Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), /*NSW=*/false, |
1817 | /* NUW=*/false, DemandedElts, KnownOut&: Known, Known2, Depth, Q); |
1818 | break; |
1819 | case Intrinsic::umul_with_overflow: |
1820 | case Intrinsic::smul_with_overflow: |
1821 | computeKnownBitsMul(Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), NSW: false, |
1822 | DemandedElts, Known, Known2, Depth, Q); |
1823 | break; |
1824 | } |
1825 | } |
1826 | } |
1827 | break; |
1828 | case Instruction::Freeze: |
1829 | if (isGuaranteedNotToBePoison(V: I->getOperand(i: 0), AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT, |
1830 | Depth: Depth + 1)) |
1831 | computeKnownBits(V: I->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
1832 | break; |
1833 | } |
1834 | } |
1835 | |
1836 | /// Determine which bits of V are known to be either zero or one and return |
1837 | /// them. |
1838 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, |
1839 | unsigned Depth, const SimplifyQuery &Q) { |
1840 | KnownBits Known(getBitWidth(Ty: V->getType(), DL: Q.DL)); |
1841 | ::computeKnownBits(V, DemandedElts, Known, Depth, Q); |
1842 | return Known; |
1843 | } |
1844 | |
1845 | /// Determine which bits of V are known to be either zero or one and return |
1846 | /// them. |
1847 | KnownBits llvm::computeKnownBits(const Value *V, unsigned Depth, |
1848 | const SimplifyQuery &Q) { |
1849 | KnownBits Known(getBitWidth(Ty: V->getType(), DL: Q.DL)); |
1850 | computeKnownBits(V, Known, Depth, Q); |
1851 | return Known; |
1852 | } |
1853 | |
1854 | /// Determine which bits of V are known to be either zero or one and return |
1855 | /// them in the Known bit set. |
1856 | /// |
1857 | /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that |
1858 | /// we cannot optimize based on the assumption that it is zero without changing |
1859 | /// it to be an explicit zero. If we don't change it to zero, other code could |
1860 | /// optimized based on the contradictory assumption that it is non-zero. |
1861 | /// Because instcombine aggressively folds operations with undef args anyway, |
1862 | /// this won't lose us code quality. |
1863 | /// |
1864 | /// This function is defined on values with integer type, values with pointer |
1865 | /// type, and vectors of integers. In the case |
1866 | /// where V is a vector, known zero, and known one values are the |
1867 | /// same width as the vector element, and the bit is set only if it is true |
1868 | /// for all of the demanded elements in the vector specified by DemandedElts. |
1869 | void computeKnownBits(const Value *V, const APInt &DemandedElts, |
1870 | KnownBits &Known, unsigned Depth, |
1871 | const SimplifyQuery &Q) { |
1872 | if (!DemandedElts) { |
1873 | // No demanded elts, better to assume we don't know anything. |
1874 | Known.resetAll(); |
1875 | return; |
1876 | } |
1877 | |
1878 | assert(V && "No Value?" ); |
1879 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
1880 | |
1881 | #ifndef NDEBUG |
1882 | Type *Ty = V->getType(); |
1883 | unsigned BitWidth = Known.getBitWidth(); |
1884 | |
1885 | assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && |
1886 | "Not integer or pointer type!" ); |
1887 | |
1888 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
1889 | assert( |
1890 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
1891 | "DemandedElt width should equal the fixed vector number of elements" ); |
1892 | } else { |
1893 | assert(DemandedElts == APInt(1, 1) && |
1894 | "DemandedElt width should be 1 for scalars or scalable vectors" ); |
1895 | } |
1896 | |
1897 | Type *ScalarTy = Ty->getScalarType(); |
1898 | if (ScalarTy->isPointerTy()) { |
1899 | assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && |
1900 | "V and Known should have same BitWidth" ); |
1901 | } else { |
1902 | assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && |
1903 | "V and Known should have same BitWidth" ); |
1904 | } |
1905 | #endif |
1906 | |
1907 | const APInt *C; |
1908 | if (match(V, P: m_APInt(Res&: C))) { |
1909 | // We know all of the bits for a scalar constant or a splat vector constant! |
1910 | Known = KnownBits::makeConstant(C: *C); |
1911 | return; |
1912 | } |
1913 | // Null and aggregate-zero are all-zeros. |
1914 | if (isa<ConstantPointerNull>(Val: V) || isa<ConstantAggregateZero>(Val: V)) { |
1915 | Known.setAllZero(); |
1916 | return; |
1917 | } |
1918 | // Handle a constant vector by taking the intersection of the known bits of |
1919 | // each element. |
1920 | if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(Val: V)) { |
1921 | assert(!isa<ScalableVectorType>(V->getType())); |
1922 | // We know that CDV must be a vector of integers. Take the intersection of |
1923 | // each element. |
1924 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
1925 | for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { |
1926 | if (!DemandedElts[i]) |
1927 | continue; |
1928 | APInt Elt = CDV->getElementAsAPInt(i); |
1929 | Known.Zero &= ~Elt; |
1930 | Known.One &= Elt; |
1931 | } |
1932 | if (Known.hasConflict()) |
1933 | Known.resetAll(); |
1934 | return; |
1935 | } |
1936 | |
1937 | if (const auto *CV = dyn_cast<ConstantVector>(Val: V)) { |
1938 | assert(!isa<ScalableVectorType>(V->getType())); |
1939 | // We know that CV must be a vector of integers. Take the intersection of |
1940 | // each element. |
1941 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
1942 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { |
1943 | if (!DemandedElts[i]) |
1944 | continue; |
1945 | Constant *Element = CV->getAggregateElement(Elt: i); |
1946 | if (isa<PoisonValue>(Val: Element)) |
1947 | continue; |
1948 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element); |
1949 | if (!ElementCI) { |
1950 | Known.resetAll(); |
1951 | return; |
1952 | } |
1953 | const APInt &Elt = ElementCI->getValue(); |
1954 | Known.Zero &= ~Elt; |
1955 | Known.One &= Elt; |
1956 | } |
1957 | if (Known.hasConflict()) |
1958 | Known.resetAll(); |
1959 | return; |
1960 | } |
1961 | |
1962 | // Start out not knowing anything. |
1963 | Known.resetAll(); |
1964 | |
1965 | // We can't imply anything about undefs. |
1966 | if (isa<UndefValue>(Val: V)) |
1967 | return; |
1968 | |
1969 | // There's no point in looking through other users of ConstantData for |
1970 | // assumptions. Confirm that we've handled them all. |
1971 | assert(!isa<ConstantData>(V) && "Unhandled constant data!" ); |
1972 | |
1973 | if (const auto *A = dyn_cast<Argument>(Val: V)) |
1974 | if (std::optional<ConstantRange> Range = A->getRange()) |
1975 | Known = Range->toKnownBits(); |
1976 | |
1977 | // All recursive calls that increase depth must come after this. |
1978 | if (Depth == MaxAnalysisRecursionDepth) |
1979 | return; |
1980 | |
1981 | // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has |
1982 | // the bits of its aliasee. |
1983 | if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(Val: V)) { |
1984 | if (!GA->isInterposable()) |
1985 | computeKnownBits(V: GA->getAliasee(), Known, Depth: Depth + 1, Q); |
1986 | return; |
1987 | } |
1988 | |
1989 | if (const Operator *I = dyn_cast<Operator>(Val: V)) |
1990 | computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); |
1991 | else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
1992 | if (std::optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) |
1993 | Known = CR->toKnownBits(); |
1994 | } |
1995 | |
1996 | // Aligned pointers have trailing zeros - refine Known.Zero set |
1997 | if (isa<PointerType>(Val: V->getType())) { |
1998 | Align Alignment = V->getPointerAlignment(DL: Q.DL); |
1999 | Known.Zero.setLowBits(Log2(A: Alignment)); |
2000 | } |
2001 | |
2002 | // computeKnownBitsFromContext strictly refines Known. |
2003 | // Therefore, we run them after computeKnownBitsFromOperator. |
2004 | |
2005 | // Check whether we can determine known bits from context such as assumes. |
2006 | computeKnownBitsFromContext(V, Known, Depth, Q); |
2007 | |
2008 | assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?" ); |
2009 | } |
2010 | |
2011 | /// Try to detect a recurrence that the value of the induction variable is |
2012 | /// always a power of two (or zero). |
2013 | static bool isPowerOfTwoRecurrence(const PHINode *PN, bool OrZero, |
2014 | unsigned Depth, SimplifyQuery &Q) { |
2015 | BinaryOperator *BO = nullptr; |
2016 | Value *Start = nullptr, *Step = nullptr; |
2017 | if (!matchSimpleRecurrence(P: PN, BO, Start, Step)) |
2018 | return false; |
2019 | |
2020 | // Initial value must be a power of two. |
2021 | for (const Use &U : PN->operands()) { |
2022 | if (U.get() == Start) { |
2023 | // Initial value comes from a different BB, need to adjust context |
2024 | // instruction for analysis. |
2025 | Q.CxtI = PN->getIncomingBlock(U)->getTerminator(); |
2026 | if (!isKnownToBeAPowerOfTwo(V: Start, OrZero, Depth, Q)) |
2027 | return false; |
2028 | } |
2029 | } |
2030 | |
2031 | // Except for Mul, the induction variable must be on the left side of the |
2032 | // increment expression, otherwise its value can be arbitrary. |
2033 | if (BO->getOpcode() != Instruction::Mul && BO->getOperand(i_nocapture: 1) != Step) |
2034 | return false; |
2035 | |
2036 | Q.CxtI = BO->getParent()->getTerminator(); |
2037 | switch (BO->getOpcode()) { |
2038 | case Instruction::Mul: |
2039 | // Power of two is closed under multiplication. |
2040 | return (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: BO) || |
2041 | Q.IIQ.hasNoSignedWrap(Op: BO)) && |
2042 | isKnownToBeAPowerOfTwo(V: Step, OrZero, Depth, Q); |
2043 | case Instruction::SDiv: |
2044 | // Start value must not be signmask for signed division, so simply being a |
2045 | // power of two is not sufficient, and it has to be a constant. |
2046 | if (!match(V: Start, P: m_Power2()) || match(V: Start, P: m_SignMask())) |
2047 | return false; |
2048 | [[fallthrough]]; |
2049 | case Instruction::UDiv: |
2050 | // Divisor must be a power of two. |
2051 | // If OrZero is false, cannot guarantee induction variable is non-zero after |
2052 | // division, same for Shr, unless it is exact division. |
2053 | return (OrZero || Q.IIQ.isExact(Op: BO)) && |
2054 | isKnownToBeAPowerOfTwo(V: Step, OrZero: false, Depth, Q); |
2055 | case Instruction::Shl: |
2056 | return OrZero || Q.IIQ.hasNoUnsignedWrap(Op: BO) || Q.IIQ.hasNoSignedWrap(Op: BO); |
2057 | case Instruction::AShr: |
2058 | if (!match(V: Start, P: m_Power2()) || match(V: Start, P: m_SignMask())) |
2059 | return false; |
2060 | [[fallthrough]]; |
2061 | case Instruction::LShr: |
2062 | return OrZero || Q.IIQ.isExact(Op: BO); |
2063 | default: |
2064 | return false; |
2065 | } |
2066 | } |
2067 | |
2068 | /// Return true if the given value is known to have exactly one |
2069 | /// bit set when defined. For vectors return true if every element is known to |
2070 | /// be a power of two when defined. Supports values with integer or pointer |
2071 | /// types and vectors of integers. |
2072 | bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, |
2073 | const SimplifyQuery &Q) { |
2074 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
2075 | |
2076 | if (isa<Constant>(Val: V)) |
2077 | return OrZero ? match(V, P: m_Power2OrZero()) : match(V, P: m_Power2()); |
2078 | |
2079 | // i1 is by definition a power of 2 or zero. |
2080 | if (OrZero && V->getType()->getScalarSizeInBits() == 1) |
2081 | return true; |
2082 | |
2083 | auto *I = dyn_cast<Instruction>(Val: V); |
2084 | if (!I) |
2085 | return false; |
2086 | |
2087 | if (Q.CxtI && match(V, P: m_VScale())) { |
2088 | const Function *F = Q.CxtI->getFunction(); |
2089 | // The vscale_range indicates vscale is a power-of-two. |
2090 | return F->hasFnAttribute(Attribute::VScaleRange); |
2091 | } |
2092 | |
2093 | // 1 << X is clearly a power of two if the one is not shifted off the end. If |
2094 | // it is shifted off the end then the result is undefined. |
2095 | if (match(V: I, P: m_Shl(L: m_One(), R: m_Value()))) |
2096 | return true; |
2097 | |
2098 | // (signmask) >>l X is clearly a power of two if the one is not shifted off |
2099 | // the bottom. If it is shifted off the bottom then the result is undefined. |
2100 | if (match(V: I, P: m_LShr(L: m_SignMask(), R: m_Value()))) |
2101 | return true; |
2102 | |
2103 | // The remaining tests are all recursive, so bail out if we hit the limit. |
2104 | if (Depth++ == MaxAnalysisRecursionDepth) |
2105 | return false; |
2106 | |
2107 | switch (I->getOpcode()) { |
2108 | case Instruction::ZExt: |
2109 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q); |
2110 | case Instruction::Trunc: |
2111 | return OrZero && isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q); |
2112 | case Instruction::Shl: |
2113 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: I) || Q.IIQ.hasNoSignedWrap(Op: I)) |
2114 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q); |
2115 | return false; |
2116 | case Instruction::LShr: |
2117 | if (OrZero || Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))) |
2118 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q); |
2119 | return false; |
2120 | case Instruction::UDiv: |
2121 | if (Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))) |
2122 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q); |
2123 | return false; |
2124 | case Instruction::Mul: |
2125 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Depth, Q) && |
2126 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q) && |
2127 | (OrZero || isKnownNonZero(V: I, Q, Depth)); |
2128 | case Instruction::And: |
2129 | // A power of two and'd with anything is a power of two or zero. |
2130 | if (OrZero && |
2131 | (isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), /*OrZero*/ true, Depth, Q) || |
2132 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), /*OrZero*/ true, Depth, Q))) |
2133 | return true; |
2134 | // X & (-X) is always a power of two or zero. |
2135 | if (match(V: I->getOperand(i: 0), P: m_Neg(V: m_Specific(V: I->getOperand(i: 1)))) || |
2136 | match(V: I->getOperand(i: 1), P: m_Neg(V: m_Specific(V: I->getOperand(i: 0))))) |
2137 | return OrZero || isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2138 | return false; |
2139 | case Instruction::Add: { |
2140 | // Adding a power-of-two or zero to the same power-of-two or zero yields |
2141 | // either the original power-of-two, a larger power-of-two or zero. |
2142 | const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(Val: V); |
2143 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: VOBO) || |
2144 | Q.IIQ.hasNoSignedWrap(Op: VOBO)) { |
2145 | if (match(V: I->getOperand(i: 0), |
2146 | P: m_c_And(L: m_Specific(V: I->getOperand(i: 1)), R: m_Value())) && |
2147 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Depth, Q)) |
2148 | return true; |
2149 | if (match(V: I->getOperand(i: 1), |
2150 | P: m_c_And(L: m_Specific(V: I->getOperand(i: 0)), R: m_Value())) && |
2151 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Depth, Q)) |
2152 | return true; |
2153 | |
2154 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); |
2155 | KnownBits LHSBits(BitWidth); |
2156 | computeKnownBits(V: I->getOperand(i: 0), Known&: LHSBits, Depth, Q); |
2157 | |
2158 | KnownBits RHSBits(BitWidth); |
2159 | computeKnownBits(V: I->getOperand(i: 1), Known&: RHSBits, Depth, Q); |
2160 | // If i8 V is a power of two or zero: |
2161 | // ZeroBits: 1 1 1 0 1 1 1 1 |
2162 | // ~ZeroBits: 0 0 0 1 0 0 0 0 |
2163 | if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) |
2164 | // If OrZero isn't set, we cannot give back a zero result. |
2165 | // Make sure either the LHS or RHS has a bit set. |
2166 | if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) |
2167 | return true; |
2168 | } |
2169 | return false; |
2170 | } |
2171 | case Instruction::Select: |
2172 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Depth, Q) && |
2173 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 2), OrZero, Depth, Q); |
2174 | case Instruction::PHI: { |
2175 | // A PHI node is power of two if all incoming values are power of two, or if |
2176 | // it is an induction variable where in each step its value is a power of |
2177 | // two. |
2178 | auto *PN = cast<PHINode>(Val: I); |
2179 | SimplifyQuery RecQ = Q; |
2180 | |
2181 | // Check if it is an induction variable and always power of two. |
2182 | if (isPowerOfTwoRecurrence(PN, OrZero, Depth, Q&: RecQ)) |
2183 | return true; |
2184 | |
2185 | // Recursively check all incoming values. Limit recursion to 2 levels, so |
2186 | // that search complexity is limited to number of operands^2. |
2187 | unsigned NewDepth = std::max(a: Depth, b: MaxAnalysisRecursionDepth - 1); |
2188 | return llvm::all_of(Range: PN->operands(), P: [&](const Use &U) { |
2189 | // Value is power of 2 if it is coming from PHI node itself by induction. |
2190 | if (U.get() == PN) |
2191 | return true; |
2192 | |
2193 | // Change the context instruction to the incoming block where it is |
2194 | // evaluated. |
2195 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); |
2196 | return isKnownToBeAPowerOfTwo(V: U.get(), OrZero, Depth: NewDepth, Q: RecQ); |
2197 | }); |
2198 | } |
2199 | case Instruction::Invoke: |
2200 | case Instruction::Call: { |
2201 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
2202 | switch (II->getIntrinsicID()) { |
2203 | case Intrinsic::umax: |
2204 | case Intrinsic::smax: |
2205 | case Intrinsic::umin: |
2206 | case Intrinsic::smin: |
2207 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 1), OrZero, Depth, Q) && |
2208 | isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Depth, Q); |
2209 | // bswap/bitreverse just move around bits, but don't change any 1s/0s |
2210 | // thus dont change pow2/non-pow2 status. |
2211 | case Intrinsic::bitreverse: |
2212 | case Intrinsic::bswap: |
2213 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Depth, Q); |
2214 | case Intrinsic::fshr: |
2215 | case Intrinsic::fshl: |
2216 | // If Op0 == Op1, this is a rotate. is_pow2(rotate(x, y)) == is_pow2(x) |
2217 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) |
2218 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Depth, Q); |
2219 | break; |
2220 | default: |
2221 | break; |
2222 | } |
2223 | } |
2224 | return false; |
2225 | } |
2226 | default: |
2227 | return false; |
2228 | } |
2229 | } |
2230 | |
2231 | /// Test whether a GEP's result is known to be non-null. |
2232 | /// |
2233 | /// Uses properties inherent in a GEP to try to determine whether it is known |
2234 | /// to be non-null. |
2235 | /// |
2236 | /// Currently this routine does not support vector GEPs. |
2237 | static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, |
2238 | const SimplifyQuery &Q) { |
2239 | const Function *F = nullptr; |
2240 | if (const Instruction *I = dyn_cast<Instruction>(Val: GEP)) |
2241 | F = I->getFunction(); |
2242 | |
2243 | if (!GEP->isInBounds() || |
2244 | NullPointerIsDefined(F, AS: GEP->getPointerAddressSpace())) |
2245 | return false; |
2246 | |
2247 | // FIXME: Support vector-GEPs. |
2248 | assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP" ); |
2249 | |
2250 | // If the base pointer is non-null, we cannot walk to a null address with an |
2251 | // inbounds GEP in address space zero. |
2252 | if (isKnownNonZero(V: GEP->getPointerOperand(), Q, Depth)) |
2253 | return true; |
2254 | |
2255 | // Walk the GEP operands and see if any operand introduces a non-zero offset. |
2256 | // If so, then the GEP cannot produce a null pointer, as doing so would |
2257 | // inherently violate the inbounds contract within address space zero. |
2258 | for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); |
2259 | GTI != GTE; ++GTI) { |
2260 | // Struct types are easy -- they must always be indexed by a constant. |
2261 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
2262 | ConstantInt *OpC = cast<ConstantInt>(Val: GTI.getOperand()); |
2263 | unsigned ElementIdx = OpC->getZExtValue(); |
2264 | const StructLayout *SL = Q.DL.getStructLayout(Ty: STy); |
2265 | uint64_t ElementOffset = SL->getElementOffset(Idx: ElementIdx); |
2266 | if (ElementOffset > 0) |
2267 | return true; |
2268 | continue; |
2269 | } |
2270 | |
2271 | // If we have a zero-sized type, the index doesn't matter. Keep looping. |
2272 | if (GTI.getSequentialElementStride(DL: Q.DL).isZero()) |
2273 | continue; |
2274 | |
2275 | // Fast path the constant operand case both for efficiency and so we don't |
2276 | // increment Depth when just zipping down an all-constant GEP. |
2277 | if (ConstantInt *OpC = dyn_cast<ConstantInt>(Val: GTI.getOperand())) { |
2278 | if (!OpC->isZero()) |
2279 | return true; |
2280 | continue; |
2281 | } |
2282 | |
2283 | // We post-increment Depth here because while isKnownNonZero increments it |
2284 | // as well, when we pop back up that increment won't persist. We don't want |
2285 | // to recurse 10k times just because we have 10k GEP operands. We don't |
2286 | // bail completely out because we want to handle constant GEPs regardless |
2287 | // of depth. |
2288 | if (Depth++ >= MaxAnalysisRecursionDepth) |
2289 | continue; |
2290 | |
2291 | if (isKnownNonZero(V: GTI.getOperand(), Q, Depth)) |
2292 | return true; |
2293 | } |
2294 | |
2295 | return false; |
2296 | } |
2297 | |
2298 | static bool isKnownNonNullFromDominatingCondition(const Value *V, |
2299 | const Instruction *CtxI, |
2300 | const DominatorTree *DT) { |
2301 | assert(!isa<Constant>(V) && "Called for constant?" ); |
2302 | |
2303 | if (!CtxI || !DT) |
2304 | return false; |
2305 | |
2306 | unsigned NumUsesExplored = 0; |
2307 | for (const auto *U : V->users()) { |
2308 | // Avoid massive lists |
2309 | if (NumUsesExplored >= DomConditionsMaxUses) |
2310 | break; |
2311 | NumUsesExplored++; |
2312 | |
2313 | // If the value is used as an argument to a call or invoke, then argument |
2314 | // attributes may provide an answer about null-ness. |
2315 | if (const auto *CB = dyn_cast<CallBase>(Val: U)) |
2316 | if (auto *CalledFunc = CB->getCalledFunction()) |
2317 | for (const Argument &Arg : CalledFunc->args()) |
2318 | if (CB->getArgOperand(i: Arg.getArgNo()) == V && |
2319 | Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && |
2320 | DT->dominates(Def: CB, User: CtxI)) |
2321 | return true; |
2322 | |
2323 | // If the value is used as a load/store, then the pointer must be non null. |
2324 | if (V == getLoadStorePointerOperand(V: U)) { |
2325 | const Instruction *I = cast<Instruction>(Val: U); |
2326 | if (!NullPointerIsDefined(F: I->getFunction(), |
2327 | AS: V->getType()->getPointerAddressSpace()) && |
2328 | DT->dominates(Def: I, User: CtxI)) |
2329 | return true; |
2330 | } |
2331 | |
2332 | if ((match(V: U, P: m_IDiv(L: m_Value(), R: m_Specific(V))) || |
2333 | match(V: U, P: m_IRem(L: m_Value(), R: m_Specific(V)))) && |
2334 | isValidAssumeForContext(Inv: cast<Instruction>(Val: U), CxtI: CtxI, DT)) |
2335 | return true; |
2336 | |
2337 | // Consider only compare instructions uniquely controlling a branch |
2338 | Value *RHS; |
2339 | CmpInst::Predicate Pred; |
2340 | if (!match(V: U, P: m_c_ICmp(Pred, L: m_Specific(V), R: m_Value(V&: RHS)))) |
2341 | continue; |
2342 | |
2343 | bool NonNullIfTrue; |
2344 | if (cmpExcludesZero(Pred, RHS)) |
2345 | NonNullIfTrue = true; |
2346 | else if (cmpExcludesZero(Pred: CmpInst::getInversePredicate(pred: Pred), RHS)) |
2347 | NonNullIfTrue = false; |
2348 | else |
2349 | continue; |
2350 | |
2351 | SmallVector<const User *, 4> WorkList; |
2352 | SmallPtrSet<const User *, 4> Visited; |
2353 | for (const auto *CmpU : U->users()) { |
2354 | assert(WorkList.empty() && "Should be!" ); |
2355 | if (Visited.insert(Ptr: CmpU).second) |
2356 | WorkList.push_back(Elt: CmpU); |
2357 | |
2358 | while (!WorkList.empty()) { |
2359 | auto *Curr = WorkList.pop_back_val(); |
2360 | |
2361 | // If a user is an AND, add all its users to the work list. We only |
2362 | // propagate "pred != null" condition through AND because it is only |
2363 | // correct to assume that all conditions of AND are met in true branch. |
2364 | // TODO: Support similar logic of OR and EQ predicate? |
2365 | if (NonNullIfTrue) |
2366 | if (match(V: Curr, P: m_LogicalAnd(L: m_Value(), R: m_Value()))) { |
2367 | for (const auto *CurrU : Curr->users()) |
2368 | if (Visited.insert(Ptr: CurrU).second) |
2369 | WorkList.push_back(Elt: CurrU); |
2370 | continue; |
2371 | } |
2372 | |
2373 | if (const BranchInst *BI = dyn_cast<BranchInst>(Val: Curr)) { |
2374 | assert(BI->isConditional() && "uses a comparison!" ); |
2375 | |
2376 | BasicBlock *NonNullSuccessor = |
2377 | BI->getSuccessor(i: NonNullIfTrue ? 0 : 1); |
2378 | BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); |
2379 | if (Edge.isSingleEdge() && DT->dominates(BBE: Edge, BB: CtxI->getParent())) |
2380 | return true; |
2381 | } else if (NonNullIfTrue && isGuard(U: Curr) && |
2382 | DT->dominates(Def: cast<Instruction>(Val: Curr), User: CtxI)) { |
2383 | return true; |
2384 | } |
2385 | } |
2386 | } |
2387 | } |
2388 | |
2389 | return false; |
2390 | } |
2391 | |
2392 | /// Does the 'Range' metadata (which must be a valid MD_range operand list) |
2393 | /// ensure that the value it's attached to is never Value? 'RangeType' is |
2394 | /// is the type of the value described by the range. |
2395 | static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { |
2396 | const unsigned NumRanges = Ranges->getNumOperands() / 2; |
2397 | assert(NumRanges >= 1); |
2398 | for (unsigned i = 0; i < NumRanges; ++i) { |
2399 | ConstantInt *Lower = |
2400 | mdconst::extract<ConstantInt>(MD: Ranges->getOperand(I: 2 * i + 0)); |
2401 | ConstantInt *Upper = |
2402 | mdconst::extract<ConstantInt>(MD: Ranges->getOperand(I: 2 * i + 1)); |
2403 | ConstantRange Range(Lower->getValue(), Upper->getValue()); |
2404 | if (Range.contains(Val: Value)) |
2405 | return false; |
2406 | } |
2407 | return true; |
2408 | } |
2409 | |
2410 | /// Try to detect a recurrence that monotonically increases/decreases from a |
2411 | /// non-zero starting value. These are common as induction variables. |
2412 | static bool isNonZeroRecurrence(const PHINode *PN) { |
2413 | BinaryOperator *BO = nullptr; |
2414 | Value *Start = nullptr, *Step = nullptr; |
2415 | const APInt *StartC, *StepC; |
2416 | if (!matchSimpleRecurrence(P: PN, BO, Start, Step) || |
2417 | !match(V: Start, P: m_APInt(Res&: StartC)) || StartC->isZero()) |
2418 | return false; |
2419 | |
2420 | switch (BO->getOpcode()) { |
2421 | case Instruction::Add: |
2422 | // Starting from non-zero and stepping away from zero can never wrap back |
2423 | // to zero. |
2424 | return BO->hasNoUnsignedWrap() || |
2425 | (BO->hasNoSignedWrap() && match(V: Step, P: m_APInt(Res&: StepC)) && |
2426 | StartC->isNegative() == StepC->isNegative()); |
2427 | case Instruction::Mul: |
2428 | return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && |
2429 | match(V: Step, P: m_APInt(Res&: StepC)) && !StepC->isZero(); |
2430 | case Instruction::Shl: |
2431 | return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); |
2432 | case Instruction::AShr: |
2433 | case Instruction::LShr: |
2434 | return BO->isExact(); |
2435 | default: |
2436 | return false; |
2437 | } |
2438 | } |
2439 | |
2440 | static bool isNonZeroAdd(const APInt &DemandedElts, unsigned Depth, |
2441 | const SimplifyQuery &Q, unsigned BitWidth, Value *X, |
2442 | Value *Y, bool NSW, bool NUW) { |
2443 | if (NUW) |
2444 | return isKnownNonZero(V: Y, DemandedElts, Q, Depth) || |
2445 | isKnownNonZero(V: X, DemandedElts, Q, Depth); |
2446 | |
2447 | KnownBits XKnown = computeKnownBits(V: X, DemandedElts, Depth, Q); |
2448 | KnownBits YKnown = computeKnownBits(V: Y, DemandedElts, Depth, Q); |
2449 | |
2450 | // If X and Y are both non-negative (as signed values) then their sum is not |
2451 | // zero unless both X and Y are zero. |
2452 | if (XKnown.isNonNegative() && YKnown.isNonNegative()) |
2453 | if (isKnownNonZero(V: Y, DemandedElts, Q, Depth) || |
2454 | isKnownNonZero(V: X, DemandedElts, Q, Depth)) |
2455 | return true; |
2456 | |
2457 | // If X and Y are both negative (as signed values) then their sum is not |
2458 | // zero unless both X and Y equal INT_MIN. |
2459 | if (XKnown.isNegative() && YKnown.isNegative()) { |
2460 | APInt Mask = APInt::getSignedMaxValue(numBits: BitWidth); |
2461 | // The sign bit of X is set. If some other bit is set then X is not equal |
2462 | // to INT_MIN. |
2463 | if (XKnown.One.intersects(RHS: Mask)) |
2464 | return true; |
2465 | // The sign bit of Y is set. If some other bit is set then Y is not equal |
2466 | // to INT_MIN. |
2467 | if (YKnown.One.intersects(RHS: Mask)) |
2468 | return true; |
2469 | } |
2470 | |
2471 | // The sum of a non-negative number and a power of two is not zero. |
2472 | if (XKnown.isNonNegative() && |
2473 | isKnownToBeAPowerOfTwo(V: Y, /*OrZero*/ false, Depth, Q)) |
2474 | return true; |
2475 | if (YKnown.isNonNegative() && |
2476 | isKnownToBeAPowerOfTwo(V: X, /*OrZero*/ false, Depth, Q)) |
2477 | return true; |
2478 | |
2479 | return KnownBits::computeForAddSub(/*Add=*/true, NSW, NUW, LHS: XKnown, RHS: YKnown) |
2480 | .isNonZero(); |
2481 | } |
2482 | |
2483 | static bool isNonZeroSub(const APInt &DemandedElts, unsigned Depth, |
2484 | const SimplifyQuery &Q, unsigned BitWidth, Value *X, |
2485 | Value *Y) { |
2486 | // TODO: Move this case into isKnownNonEqual(). |
2487 | if (auto *C = dyn_cast<Constant>(Val: X)) |
2488 | if (C->isNullValue() && isKnownNonZero(V: Y, DemandedElts, Q, Depth)) |
2489 | return true; |
2490 | |
2491 | return ::isKnownNonEqual(V1: X, V2: Y, Depth, Q); |
2492 | } |
2493 | |
2494 | static bool isNonZeroMul(const APInt &DemandedElts, unsigned Depth, |
2495 | const SimplifyQuery &Q, unsigned BitWidth, Value *X, |
2496 | Value *Y, bool NSW, bool NUW) { |
2497 | // If X and Y are non-zero then so is X * Y as long as the multiplication |
2498 | // does not overflow. |
2499 | if (NSW || NUW) |
2500 | return isKnownNonZero(V: X, DemandedElts, Q, Depth) && |
2501 | isKnownNonZero(V: Y, DemandedElts, Q, Depth); |
2502 | |
2503 | // If either X or Y is odd, then if the other is non-zero the result can't |
2504 | // be zero. |
2505 | KnownBits XKnown = computeKnownBits(V: X, DemandedElts, Depth, Q); |
2506 | if (XKnown.One[0]) |
2507 | return isKnownNonZero(V: Y, DemandedElts, Q, Depth); |
2508 | |
2509 | KnownBits YKnown = computeKnownBits(V: Y, DemandedElts, Depth, Q); |
2510 | if (YKnown.One[0]) |
2511 | return XKnown.isNonZero() || isKnownNonZero(V: X, DemandedElts, Q, Depth); |
2512 | |
2513 | // If there exists any subset of X (sX) and subset of Y (sY) s.t sX * sY is |
2514 | // non-zero, then X * Y is non-zero. We can find sX and sY by just taking |
2515 | // the lowest known One of X and Y. If they are non-zero, the result |
2516 | // must be non-zero. We can check if LSB(X) * LSB(Y) != 0 by doing |
2517 | // X.CountLeadingZeros + Y.CountLeadingZeros < BitWidth. |
2518 | return (XKnown.countMaxTrailingZeros() + YKnown.countMaxTrailingZeros()) < |
2519 | BitWidth; |
2520 | } |
2521 | |
2522 | static bool isNonZeroShift(const Operator *I, const APInt &DemandedElts, |
2523 | unsigned Depth, const SimplifyQuery &Q, |
2524 | const KnownBits &KnownVal) { |
2525 | auto ShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { |
2526 | switch (I->getOpcode()) { |
2527 | case Instruction::Shl: |
2528 | return Lhs.shl(ShiftAmt: Rhs); |
2529 | case Instruction::LShr: |
2530 | return Lhs.lshr(ShiftAmt: Rhs); |
2531 | case Instruction::AShr: |
2532 | return Lhs.ashr(ShiftAmt: Rhs); |
2533 | default: |
2534 | llvm_unreachable("Unknown Shift Opcode" ); |
2535 | } |
2536 | }; |
2537 | |
2538 | auto InvShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { |
2539 | switch (I->getOpcode()) { |
2540 | case Instruction::Shl: |
2541 | return Lhs.lshr(ShiftAmt: Rhs); |
2542 | case Instruction::LShr: |
2543 | case Instruction::AShr: |
2544 | return Lhs.shl(ShiftAmt: Rhs); |
2545 | default: |
2546 | llvm_unreachable("Unknown Shift Opcode" ); |
2547 | } |
2548 | }; |
2549 | |
2550 | if (KnownVal.isUnknown()) |
2551 | return false; |
2552 | |
2553 | KnownBits KnownCnt = |
2554 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Depth, Q); |
2555 | APInt MaxShift = KnownCnt.getMaxValue(); |
2556 | unsigned NumBits = KnownVal.getBitWidth(); |
2557 | if (MaxShift.uge(RHS: NumBits)) |
2558 | return false; |
2559 | |
2560 | if (!ShiftOp(KnownVal.One, MaxShift).isZero()) |
2561 | return true; |
2562 | |
2563 | // If all of the bits shifted out are known to be zero, and Val is known |
2564 | // non-zero then at least one non-zero bit must remain. |
2565 | if (InvShiftOp(KnownVal.Zero, NumBits - MaxShift) |
2566 | .eq(RHS: InvShiftOp(APInt::getAllOnes(numBits: NumBits), NumBits - MaxShift)) && |
2567 | isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth)) |
2568 | return true; |
2569 | |
2570 | return false; |
2571 | } |
2572 | |
2573 | static bool isKnownNonZeroFromOperator(const Operator *I, |
2574 | const APInt &DemandedElts, |
2575 | unsigned Depth, const SimplifyQuery &Q) { |
2576 | unsigned BitWidth = getBitWidth(Ty: I->getType()->getScalarType(), DL: Q.DL); |
2577 | switch (I->getOpcode()) { |
2578 | case Instruction::Alloca: |
2579 | // Alloca never returns null, malloc might. |
2580 | return I->getType()->getPointerAddressSpace() == 0; |
2581 | case Instruction::GetElementPtr: |
2582 | if (I->getType()->isPointerTy()) |
2583 | return isGEPKnownNonNull(GEP: cast<GEPOperator>(Val: I), Depth, Q); |
2584 | break; |
2585 | case Instruction::BitCast: { |
2586 | // We need to be a bit careful here. We can only peek through the bitcast |
2587 | // if the scalar size of elements in the operand are smaller than and a |
2588 | // multiple of the size they are casting too. Take three cases: |
2589 | // |
2590 | // 1) Unsafe: |
2591 | // bitcast <2 x i16> %NonZero to <4 x i8> |
2592 | // |
2593 | // %NonZero can have 2 non-zero i16 elements, but isKnownNonZero on a |
2594 | // <4 x i8> requires that all 4 i8 elements be non-zero which isn't |
2595 | // guranteed (imagine just sign bit set in the 2 i16 elements). |
2596 | // |
2597 | // 2) Unsafe: |
2598 | // bitcast <4 x i3> %NonZero to <3 x i4> |
2599 | // |
2600 | // Even though the scalar size of the src (`i3`) is smaller than the |
2601 | // scalar size of the dst `i4`, because `i3` is not a multiple of `i4` |
2602 | // its possible for the `3 x i4` elements to be zero because there are |
2603 | // some elements in the destination that don't contain any full src |
2604 | // element. |
2605 | // |
2606 | // 3) Safe: |
2607 | // bitcast <4 x i8> %NonZero to <2 x i16> |
2608 | // |
2609 | // This is always safe as non-zero in the 4 i8 elements implies |
2610 | // non-zero in the combination of any two adjacent ones. Since i8 is a |
2611 | // multiple of i16, each i16 is guranteed to have 2 full i8 elements. |
2612 | // This all implies the 2 i16 elements are non-zero. |
2613 | Type *FromTy = I->getOperand(i: 0)->getType(); |
2614 | if ((FromTy->isIntOrIntVectorTy() || FromTy->isPtrOrPtrVectorTy()) && |
2615 | (BitWidth % getBitWidth(Ty: FromTy->getScalarType(), DL: Q.DL)) == 0) |
2616 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2617 | } break; |
2618 | case Instruction::IntToPtr: |
2619 | // Note that we have to take special care to avoid looking through |
2620 | // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well |
2621 | // as casts that can alter the value, e.g., AddrSpaceCasts. |
2622 | if (!isa<ScalableVectorType>(Val: I->getType()) && |
2623 | Q.DL.getTypeSizeInBits(Ty: I->getOperand(i: 0)->getType()).getFixedValue() <= |
2624 | Q.DL.getTypeSizeInBits(Ty: I->getType()).getFixedValue()) |
2625 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2626 | break; |
2627 | case Instruction::PtrToInt: |
2628 | // Similar to int2ptr above, we can look through ptr2int here if the cast |
2629 | // is a no-op or an extend and not a truncate. |
2630 | if (!isa<ScalableVectorType>(Val: I->getType()) && |
2631 | Q.DL.getTypeSizeInBits(Ty: I->getOperand(i: 0)->getType()).getFixedValue() <= |
2632 | Q.DL.getTypeSizeInBits(Ty: I->getType()).getFixedValue()) |
2633 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2634 | break; |
2635 | case Instruction::Trunc: |
2636 | // nuw/nsw trunc preserves zero/non-zero status of input. |
2637 | if (auto *TI = dyn_cast<TruncInst>(Val: I)) |
2638 | if (TI->hasNoSignedWrap() || TI->hasNoUnsignedWrap()) |
2639 | return isKnownNonZero(V: TI->getOperand(i_nocapture: 0), Q, Depth); |
2640 | break; |
2641 | |
2642 | case Instruction::Sub: |
2643 | return isNonZeroSub(DemandedElts, Depth, Q, BitWidth, X: I->getOperand(i: 0), |
2644 | Y: I->getOperand(i: 1)); |
2645 | case Instruction::Or: |
2646 | // X | Y != 0 if X != 0 or Y != 0. |
2647 | return isKnownNonZero(V: I->getOperand(i: 1), DemandedElts, Q, Depth) || |
2648 | isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
2649 | case Instruction::SExt: |
2650 | case Instruction::ZExt: |
2651 | // ext X != 0 if X != 0. |
2652 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2653 | |
2654 | case Instruction::Shl: { |
2655 | // shl nsw/nuw can't remove any non-zero bits. |
2656 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(Val: I); |
2657 | if (Q.IIQ.hasNoUnsignedWrap(Op: BO) || Q.IIQ.hasNoSignedWrap(Op: BO)) |
2658 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2659 | |
2660 | // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined |
2661 | // if the lowest bit is shifted off the end. |
2662 | KnownBits Known(BitWidth); |
2663 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Depth, Q); |
2664 | if (Known.One[0]) |
2665 | return true; |
2666 | |
2667 | return isNonZeroShift(I, DemandedElts, Depth, Q, KnownVal: Known); |
2668 | } |
2669 | case Instruction::LShr: |
2670 | case Instruction::AShr: { |
2671 | // shr exact can only shift out zero bits. |
2672 | const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(Val: I); |
2673 | if (BO->isExact()) |
2674 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2675 | |
2676 | // shr X, Y != 0 if X is negative. Note that the value of the shift is not |
2677 | // defined if the sign bit is shifted off the end. |
2678 | KnownBits Known = |
2679 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Depth, Q); |
2680 | if (Known.isNegative()) |
2681 | return true; |
2682 | |
2683 | return isNonZeroShift(I, DemandedElts, Depth, Q, KnownVal: Known); |
2684 | } |
2685 | case Instruction::UDiv: |
2686 | case Instruction::SDiv: { |
2687 | // X / Y |
2688 | // div exact can only produce a zero if the dividend is zero. |
2689 | if (cast<PossiblyExactOperator>(Val: I)->isExact()) |
2690 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
2691 | |
2692 | KnownBits XKnown = |
2693 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Depth, Q); |
2694 | // If X is fully unknown we won't be able to figure anything out so don't |
2695 | // both computing knownbits for Y. |
2696 | if (XKnown.isUnknown()) |
2697 | return false; |
2698 | |
2699 | KnownBits YKnown = |
2700 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Depth, Q); |
2701 | if (I->getOpcode() == Instruction::SDiv) { |
2702 | // For signed division need to compare abs value of the operands. |
2703 | XKnown = XKnown.abs(/*IntMinIsPoison*/ false); |
2704 | YKnown = YKnown.abs(/*IntMinIsPoison*/ false); |
2705 | } |
2706 | // If X u>= Y then div is non zero (0/0 is UB). |
2707 | std::optional<bool> XUgeY = KnownBits::uge(LHS: XKnown, RHS: YKnown); |
2708 | // If X is total unknown or X u< Y we won't be able to prove non-zero |
2709 | // with compute known bits so just return early. |
2710 | return XUgeY && *XUgeY; |
2711 | } |
2712 | case Instruction::Add: { |
2713 | // X + Y. |
2714 | |
2715 | // If Add has nuw wrap flag, then if either X or Y is non-zero the result is |
2716 | // non-zero. |
2717 | auto *BO = cast<OverflowingBinaryOperator>(Val: I); |
2718 | return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth, X: I->getOperand(i: 0), |
2719 | Y: I->getOperand(i: 1), NSW: Q.IIQ.hasNoSignedWrap(Op: BO), |
2720 | NUW: Q.IIQ.hasNoUnsignedWrap(Op: BO)); |
2721 | } |
2722 | case Instruction::Mul: { |
2723 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(Val: I); |
2724 | return isNonZeroMul(DemandedElts, Depth, Q, BitWidth, X: I->getOperand(i: 0), |
2725 | Y: I->getOperand(i: 1), NSW: Q.IIQ.hasNoSignedWrap(Op: BO), |
2726 | NUW: Q.IIQ.hasNoUnsignedWrap(Op: BO)); |
2727 | } |
2728 | case Instruction::Select: { |
2729 | // (C ? X : Y) != 0 if X != 0 and Y != 0. |
2730 | |
2731 | // First check if the arm is non-zero using `isKnownNonZero`. If that fails, |
2732 | // then see if the select condition implies the arm is non-zero. For example |
2733 | // (X != 0 ? X : Y), we know the true arm is non-zero as the `X` "return" is |
2734 | // dominated by `X != 0`. |
2735 | auto SelectArmIsNonZero = [&](bool IsTrueArm) { |
2736 | Value *Op; |
2737 | Op = IsTrueArm ? I->getOperand(i: 1) : I->getOperand(i: 2); |
2738 | // Op is trivially non-zero. |
2739 | if (isKnownNonZero(V: Op, DemandedElts, Q, Depth)) |
2740 | return true; |
2741 | |
2742 | // The condition of the select dominates the true/false arm. Check if the |
2743 | // condition implies that a given arm is non-zero. |
2744 | Value *X; |
2745 | CmpInst::Predicate Pred; |
2746 | if (!match(V: I->getOperand(i: 0), P: m_c_ICmp(Pred, L: m_Specific(V: Op), R: m_Value(V&: X)))) |
2747 | return false; |
2748 | |
2749 | if (!IsTrueArm) |
2750 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
2751 | |
2752 | return cmpExcludesZero(Pred, RHS: X); |
2753 | }; |
2754 | |
2755 | if (SelectArmIsNonZero(/* IsTrueArm */ true) && |
2756 | SelectArmIsNonZero(/* IsTrueArm */ false)) |
2757 | return true; |
2758 | break; |
2759 | } |
2760 | case Instruction::PHI: { |
2761 | auto *PN = cast<PHINode>(Val: I); |
2762 | if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) |
2763 | return true; |
2764 | |
2765 | // Check if all incoming values are non-zero using recursion. |
2766 | SimplifyQuery RecQ = Q; |
2767 | unsigned NewDepth = std::max(a: Depth, b: MaxAnalysisRecursionDepth - 1); |
2768 | return llvm::all_of(Range: PN->operands(), P: [&](const Use &U) { |
2769 | if (U.get() == PN) |
2770 | return true; |
2771 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); |
2772 | // Check if the branch on the phi excludes zero. |
2773 | ICmpInst::Predicate Pred; |
2774 | Value *X; |
2775 | BasicBlock *TrueSucc, *FalseSucc; |
2776 | if (match(V: RecQ.CxtI, |
2777 | P: m_Br(C: m_c_ICmp(Pred, L: m_Specific(V: U.get()), R: m_Value(V&: X)), |
2778 | T: m_BasicBlock(V&: TrueSucc), F: m_BasicBlock(V&: FalseSucc)))) { |
2779 | // Check for cases of duplicate successors. |
2780 | if ((TrueSucc == PN->getParent()) != (FalseSucc == PN->getParent())) { |
2781 | // If we're using the false successor, invert the predicate. |
2782 | if (FalseSucc == PN->getParent()) |
2783 | Pred = CmpInst::getInversePredicate(pred: Pred); |
2784 | if (cmpExcludesZero(Pred, RHS: X)) |
2785 | return true; |
2786 | } |
2787 | } |
2788 | // Finally recurse on the edge and check it directly. |
2789 | return isKnownNonZero(V: U.get(), DemandedElts, Q: RecQ, Depth: NewDepth); |
2790 | }); |
2791 | } |
2792 | case Instruction::InsertElement: { |
2793 | if (isa<ScalableVectorType>(Val: I->getType())) |
2794 | break; |
2795 | |
2796 | const Value *Vec = I->getOperand(i: 0); |
2797 | const Value *Elt = I->getOperand(i: 1); |
2798 | auto *CIdx = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2)); |
2799 | |
2800 | unsigned NumElts = DemandedElts.getBitWidth(); |
2801 | APInt DemandedVecElts = DemandedElts; |
2802 | bool SkipElt = false; |
2803 | // If we know the index we are inserting too, clear it from Vec check. |
2804 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
2805 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
2806 | SkipElt = !DemandedElts[CIdx->getZExtValue()]; |
2807 | } |
2808 | |
2809 | // Result is zero if Elt is non-zero and rest of the demanded elts in Vec |
2810 | // are non-zero. |
2811 | return (SkipElt || isKnownNonZero(V: Elt, Q, Depth)) && |
2812 | (DemandedVecElts.isZero() || |
2813 | isKnownNonZero(V: Vec, DemandedElts: DemandedVecElts, Q, Depth)); |
2814 | } |
2815 | case Instruction::ExtractElement: |
2816 | if (const auto *EEI = dyn_cast<ExtractElementInst>(Val: I)) { |
2817 | const Value *Vec = EEI->getVectorOperand(); |
2818 | const Value *Idx = EEI->getIndexOperand(); |
2819 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
2820 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType())) { |
2821 | unsigned NumElts = VecTy->getNumElements(); |
2822 | APInt DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
2823 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
2824 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
2825 | return isKnownNonZero(V: Vec, DemandedElts: DemandedVecElts, Q, Depth); |
2826 | } |
2827 | } |
2828 | break; |
2829 | case Instruction::ShuffleVector: { |
2830 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: I); |
2831 | if (!Shuf) |
2832 | break; |
2833 | APInt DemandedLHS, DemandedRHS; |
2834 | // For undef elements, we don't know anything about the common state of |
2835 | // the shuffle result. |
2836 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
2837 | break; |
2838 | // If demanded elements for both vecs are non-zero, the shuffle is non-zero. |
2839 | return (DemandedRHS.isZero() || |
2840 | isKnownNonZero(V: Shuf->getOperand(i_nocapture: 1), DemandedElts: DemandedRHS, Q, Depth)) && |
2841 | (DemandedLHS.isZero() || |
2842 | isKnownNonZero(V: Shuf->getOperand(i_nocapture: 0), DemandedElts: DemandedLHS, Q, Depth)); |
2843 | } |
2844 | case Instruction::Freeze: |
2845 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth) && |
2846 | isGuaranteedNotToBePoison(V: I->getOperand(i: 0), AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT, |
2847 | Depth); |
2848 | case Instruction::Load: { |
2849 | auto *LI = cast<LoadInst>(Val: I); |
2850 | // A Load tagged with nonnull or dereferenceable with null pointer undefined |
2851 | // is never null. |
2852 | if (auto *PtrT = dyn_cast<PointerType>(Val: I->getType())) { |
2853 | if (Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_nonnull) || |
2854 | (Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_dereferenceable) && |
2855 | !NullPointerIsDefined(F: LI->getFunction(), AS: PtrT->getAddressSpace()))) |
2856 | return true; |
2857 | } else if (MDNode *Ranges = Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_range)) { |
2858 | return rangeMetadataExcludesValue(Ranges, Value: APInt::getZero(numBits: BitWidth)); |
2859 | } |
2860 | |
2861 | // No need to fall through to computeKnownBits as range metadata is already |
2862 | // handled in isKnownNonZero. |
2863 | return false; |
2864 | } |
2865 | case Instruction::ExtractValue: { |
2866 | const WithOverflowInst *WO; |
2867 | if (match(V: I, P: m_ExtractValue<0>(V: m_WithOverflowInst(I&: WO)))) { |
2868 | switch (WO->getBinaryOp()) { |
2869 | default: |
2870 | break; |
2871 | case Instruction::Add: |
2872 | return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth, |
2873 | X: WO->getArgOperand(i: 0), Y: WO->getArgOperand(i: 1), |
2874 | /*NSW=*/false, |
2875 | /*NUW=*/false); |
2876 | case Instruction::Sub: |
2877 | return isNonZeroSub(DemandedElts, Depth, Q, BitWidth, |
2878 | X: WO->getArgOperand(i: 0), Y: WO->getArgOperand(i: 1)); |
2879 | case Instruction::Mul: |
2880 | return isNonZeroMul(DemandedElts, Depth, Q, BitWidth, |
2881 | X: WO->getArgOperand(i: 0), Y: WO->getArgOperand(i: 1), |
2882 | /*NSW=*/false, /*NUW=*/false); |
2883 | break; |
2884 | } |
2885 | } |
2886 | break; |
2887 | } |
2888 | case Instruction::Call: |
2889 | case Instruction::Invoke: { |
2890 | const auto *Call = cast<CallBase>(Val: I); |
2891 | if (I->getType()->isPointerTy()) { |
2892 | if (Call->isReturnNonNull()) |
2893 | return true; |
2894 | if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, MustPreserveNullness: true)) |
2895 | return isKnownNonZero(V: RP, Q, Depth); |
2896 | } else { |
2897 | if (MDNode *Ranges = Q.IIQ.getMetadata(I: Call, KindID: LLVMContext::MD_range)) |
2898 | return rangeMetadataExcludesValue(Ranges, Value: APInt::getZero(numBits: BitWidth)); |
2899 | if (std::optional<ConstantRange> Range = Call->getRange()) { |
2900 | const APInt ZeroValue(Range->getBitWidth(), 0); |
2901 | if (!Range->contains(Val: ZeroValue)) |
2902 | return true; |
2903 | } |
2904 | if (const Value *RV = Call->getReturnedArgOperand()) |
2905 | if (RV->getType() == I->getType() && isKnownNonZero(V: RV, Q, Depth)) |
2906 | return true; |
2907 | } |
2908 | |
2909 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
2910 | switch (II->getIntrinsicID()) { |
2911 | case Intrinsic::sshl_sat: |
2912 | case Intrinsic::ushl_sat: |
2913 | case Intrinsic::abs: |
2914 | case Intrinsic::bitreverse: |
2915 | case Intrinsic::bswap: |
2916 | case Intrinsic::ctpop: |
2917 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
2918 | // NB: We don't do usub_sat here as in any case we can prove its |
2919 | // non-zero, we will fold it to `sub nuw` in InstCombine. |
2920 | case Intrinsic::ssub_sat: |
2921 | return isNonZeroSub(DemandedElts, Depth, Q, BitWidth, |
2922 | X: II->getArgOperand(i: 0), Y: II->getArgOperand(i: 1)); |
2923 | case Intrinsic::sadd_sat: |
2924 | return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth, |
2925 | X: II->getArgOperand(i: 0), Y: II->getArgOperand(i: 1), |
2926 | /*NSW=*/true, /* NUW=*/false); |
2927 | // umin/smin/smax/smin/or of all non-zero elements is always non-zero. |
2928 | case Intrinsic::vector_reduce_or: |
2929 | case Intrinsic::vector_reduce_umax: |
2930 | case Intrinsic::vector_reduce_umin: |
2931 | case Intrinsic::vector_reduce_smax: |
2932 | case Intrinsic::vector_reduce_smin: |
2933 | return isKnownNonZero(V: II->getArgOperand(i: 0), Q, Depth); |
2934 | case Intrinsic::umax: |
2935 | case Intrinsic::uadd_sat: |
2936 | return isKnownNonZero(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth) || |
2937 | isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
2938 | case Intrinsic::smax: { |
2939 | // If either arg is strictly positive the result is non-zero. Otherwise |
2940 | // the result is non-zero if both ops are non-zero. |
2941 | auto IsNonZero = [&](Value *Op, std::optional<bool> &OpNonZero, |
2942 | const KnownBits &OpKnown) { |
2943 | if (!OpNonZero.has_value()) |
2944 | OpNonZero = OpKnown.isNonZero() || |
2945 | isKnownNonZero(V: Op, DemandedElts, Q, Depth); |
2946 | return *OpNonZero; |
2947 | }; |
2948 | // Avoid re-computing isKnownNonZero. |
2949 | std::optional<bool> Op0NonZero, Op1NonZero; |
2950 | KnownBits Op1Known = |
2951 | computeKnownBits(V: II->getArgOperand(i: 1), DemandedElts, Depth, Q); |
2952 | if (Op1Known.isNonNegative() && |
2953 | IsNonZero(II->getArgOperand(i: 1), Op1NonZero, Op1Known)) |
2954 | return true; |
2955 | KnownBits Op0Known = |
2956 | computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Depth, Q); |
2957 | if (Op0Known.isNonNegative() && |
2958 | IsNonZero(II->getArgOperand(i: 0), Op0NonZero, Op0Known)) |
2959 | return true; |
2960 | return IsNonZero(II->getArgOperand(i: 1), Op1NonZero, Op1Known) && |
2961 | IsNonZero(II->getArgOperand(i: 0), Op0NonZero, Op0Known); |
2962 | } |
2963 | case Intrinsic::smin: { |
2964 | // If either arg is negative the result is non-zero. Otherwise |
2965 | // the result is non-zero if both ops are non-zero. |
2966 | KnownBits Op1Known = |
2967 | computeKnownBits(V: II->getArgOperand(i: 1), DemandedElts, Depth, Q); |
2968 | if (Op1Known.isNegative()) |
2969 | return true; |
2970 | KnownBits Op0Known = |
2971 | computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Depth, Q); |
2972 | if (Op0Known.isNegative()) |
2973 | return true; |
2974 | |
2975 | if (Op1Known.isNonZero() && Op0Known.isNonZero()) |
2976 | return true; |
2977 | } |
2978 | [[fallthrough]]; |
2979 | case Intrinsic::umin: |
2980 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth) && |
2981 | isKnownNonZero(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth); |
2982 | case Intrinsic::cttz: |
2983 | return computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Depth, Q) |
2984 | .Zero[0]; |
2985 | case Intrinsic::ctlz: |
2986 | return computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Depth, Q) |
2987 | .isNonNegative(); |
2988 | case Intrinsic::fshr: |
2989 | case Intrinsic::fshl: |
2990 | // If Op0 == Op1, this is a rotate. rotate(x, y) != 0 iff x != 0. |
2991 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) |
2992 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
2993 | break; |
2994 | case Intrinsic::vscale: |
2995 | return true; |
2996 | case Intrinsic::experimental_get_vector_length: |
2997 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
2998 | default: |
2999 | break; |
3000 | } |
3001 | break; |
3002 | } |
3003 | |
3004 | return false; |
3005 | } |
3006 | } |
3007 | |
3008 | KnownBits Known(BitWidth); |
3009 | computeKnownBits(V: I, DemandedElts, Known, Depth, Q); |
3010 | return Known.One != 0; |
3011 | } |
3012 | |
3013 | /// Return true if the given value is known to be non-zero when defined. For |
3014 | /// vectors, return true if every demanded element is known to be non-zero when |
3015 | /// defined. For pointers, if the context instruction and dominator tree are |
3016 | /// specified, perform context-sensitive analysis and return true if the |
3017 | /// pointer couldn't possibly be null at the specified instruction. |
3018 | /// Supports values with integer or pointer type and vectors of integers. |
3019 | bool isKnownNonZero(const Value *V, const APInt &DemandedElts, |
3020 | const SimplifyQuery &Q, unsigned Depth) { |
3021 | Type *Ty = V->getType(); |
3022 | |
3023 | #ifndef NDEBUG |
3024 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
3025 | |
3026 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
3027 | assert( |
3028 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
3029 | "DemandedElt width should equal the fixed vector number of elements" ); |
3030 | } else { |
3031 | assert(DemandedElts == APInt(1, 1) && |
3032 | "DemandedElt width should be 1 for scalars" ); |
3033 | } |
3034 | #endif |
3035 | |
3036 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
3037 | if (C->isNullValue()) |
3038 | return false; |
3039 | if (isa<ConstantInt>(Val: C)) |
3040 | // Must be non-zero due to null test above. |
3041 | return true; |
3042 | |
3043 | // For constant vectors, check that all elements are poison or known |
3044 | // non-zero to determine that the whole vector is known non-zero. |
3045 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
3046 | for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { |
3047 | if (!DemandedElts[i]) |
3048 | continue; |
3049 | Constant *Elt = C->getAggregateElement(Elt: i); |
3050 | if (!Elt || Elt->isNullValue()) |
3051 | return false; |
3052 | if (!isa<PoisonValue>(Val: Elt) && !isa<ConstantInt>(Val: Elt)) |
3053 | return false; |
3054 | } |
3055 | return true; |
3056 | } |
3057 | |
3058 | // A global variable in address space 0 is non null unless extern weak |
3059 | // or an absolute symbol reference. Other address spaces may have null as a |
3060 | // valid address for a global, so we can't assume anything. |
3061 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
3062 | if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && |
3063 | GV->getType()->getAddressSpace() == 0) |
3064 | return true; |
3065 | } |
3066 | |
3067 | // For constant expressions, fall through to the Operator code below. |
3068 | if (!isa<ConstantExpr>(Val: V)) |
3069 | return false; |
3070 | } |
3071 | |
3072 | if (const auto *A = dyn_cast<Argument>(Val: V)) |
3073 | if (std::optional<ConstantRange> Range = A->getRange()) { |
3074 | const APInt ZeroValue(Range->getBitWidth(), 0); |
3075 | if (!Range->contains(Val: ZeroValue)) |
3076 | return true; |
3077 | } |
3078 | |
3079 | if (!isa<Constant>(Val: V) && isKnownNonZeroFromAssume(V, Q)) |
3080 | return true; |
3081 | |
3082 | // Some of the tests below are recursive, so bail out if we hit the limit. |
3083 | if (Depth++ >= MaxAnalysisRecursionDepth) |
3084 | return false; |
3085 | |
3086 | // Check for pointer simplifications. |
3087 | |
3088 | if (PointerType *PtrTy = dyn_cast<PointerType>(Val: Ty)) { |
3089 | // A byval, inalloca may not be null in a non-default addres space. A |
3090 | // nonnull argument is assumed never 0. |
3091 | if (const Argument *A = dyn_cast<Argument>(Val: V)) { |
3092 | if (((A->hasPassPointeeByValueCopyAttr() && |
3093 | !NullPointerIsDefined(F: A->getParent(), AS: PtrTy->getAddressSpace())) || |
3094 | A->hasNonNullAttr())) |
3095 | return true; |
3096 | } |
3097 | } |
3098 | |
3099 | if (const auto *I = dyn_cast<Operator>(Val: V)) |
3100 | if (isKnownNonZeroFromOperator(I, DemandedElts, Depth, Q)) |
3101 | return true; |
3102 | |
3103 | if (!isa<Constant>(Val: V) && |
3104 | isKnownNonNullFromDominatingCondition(V, CtxI: Q.CxtI, DT: Q.DT)) |
3105 | return true; |
3106 | |
3107 | return false; |
3108 | } |
3109 | |
3110 | bool llvm::isKnownNonZero(const Value *V, const SimplifyQuery &Q, |
3111 | unsigned Depth) { |
3112 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
3113 | APInt DemandedElts = |
3114 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
3115 | return ::isKnownNonZero(V, DemandedElts, Q, Depth); |
3116 | } |
3117 | |
3118 | /// If the pair of operators are the same invertible function, return the |
3119 | /// the operands of the function corresponding to each input. Otherwise, |
3120 | /// return std::nullopt. An invertible function is one that is 1-to-1 and maps |
3121 | /// every input value to exactly one output value. This is equivalent to |
3122 | /// saying that Op1 and Op2 are equal exactly when the specified pair of |
3123 | /// operands are equal, (except that Op1 and Op2 may be poison more often.) |
3124 | static std::optional<std::pair<Value*, Value*>> |
3125 | getInvertibleOperands(const Operator *Op1, |
3126 | const Operator *Op2) { |
3127 | if (Op1->getOpcode() != Op2->getOpcode()) |
3128 | return std::nullopt; |
3129 | |
3130 | auto getOperands = [&](unsigned OpNum) -> auto { |
3131 | return std::make_pair(x: Op1->getOperand(i: OpNum), y: Op2->getOperand(i: OpNum)); |
3132 | }; |
3133 | |
3134 | switch (Op1->getOpcode()) { |
3135 | default: |
3136 | break; |
3137 | case Instruction::Or: |
3138 | if (!cast<PossiblyDisjointInst>(Val: Op1)->isDisjoint() || |
3139 | !cast<PossiblyDisjointInst>(Val: Op2)->isDisjoint()) |
3140 | break; |
3141 | [[fallthrough]]; |
3142 | case Instruction::Xor: |
3143 | case Instruction::Add: { |
3144 | Value *Other; |
3145 | if (match(V: Op2, P: m_c_BinOp(L: m_Specific(V: Op1->getOperand(i: 0)), R: m_Value(V&: Other)))) |
3146 | return std::make_pair(x: Op1->getOperand(i: 1), y&: Other); |
3147 | if (match(V: Op2, P: m_c_BinOp(L: m_Specific(V: Op1->getOperand(i: 1)), R: m_Value(V&: Other)))) |
3148 | return std::make_pair(x: Op1->getOperand(i: 0), y&: Other); |
3149 | break; |
3150 | } |
3151 | case Instruction::Sub: |
3152 | if (Op1->getOperand(i: 0) == Op2->getOperand(i: 0)) |
3153 | return getOperands(1); |
3154 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
3155 | return getOperands(0); |
3156 | break; |
3157 | case Instruction::Mul: { |
3158 | // invertible if A * B == (A * B) mod 2^N where A, and B are integers |
3159 | // and N is the bitwdith. The nsw case is non-obvious, but proven by |
3160 | // alive2: https://alive2.llvm.org/ce/z/Z6D5qK |
3161 | auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
3162 | auto *OBO2 = cast<OverflowingBinaryOperator>(Val: Op2); |
3163 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && |
3164 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) |
3165 | break; |
3166 | |
3167 | // Assume operand order has been canonicalized |
3168 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1) && |
3169 | isa<ConstantInt>(Val: Op1->getOperand(i: 1)) && |
3170 | !cast<ConstantInt>(Val: Op1->getOperand(i: 1))->isZero()) |
3171 | return getOperands(0); |
3172 | break; |
3173 | } |
3174 | case Instruction::Shl: { |
3175 | // Same as multiplies, with the difference that we don't need to check |
3176 | // for a non-zero multiply. Shifts always multiply by non-zero. |
3177 | auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
3178 | auto *OBO2 = cast<OverflowingBinaryOperator>(Val: Op2); |
3179 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && |
3180 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) |
3181 | break; |
3182 | |
3183 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
3184 | return getOperands(0); |
3185 | break; |
3186 | } |
3187 | case Instruction::AShr: |
3188 | case Instruction::LShr: { |
3189 | auto *PEO1 = cast<PossiblyExactOperator>(Val: Op1); |
3190 | auto *PEO2 = cast<PossiblyExactOperator>(Val: Op2); |
3191 | if (!PEO1->isExact() || !PEO2->isExact()) |
3192 | break; |
3193 | |
3194 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
3195 | return getOperands(0); |
3196 | break; |
3197 | } |
3198 | case Instruction::SExt: |
3199 | case Instruction::ZExt: |
3200 | if (Op1->getOperand(i: 0)->getType() == Op2->getOperand(i: 0)->getType()) |
3201 | return getOperands(0); |
3202 | break; |
3203 | case Instruction::PHI: { |
3204 | const PHINode *PN1 = cast<PHINode>(Val: Op1); |
3205 | const PHINode *PN2 = cast<PHINode>(Val: Op2); |
3206 | |
3207 | // If PN1 and PN2 are both recurrences, can we prove the entire recurrences |
3208 | // are a single invertible function of the start values? Note that repeated |
3209 | // application of an invertible function is also invertible |
3210 | BinaryOperator *BO1 = nullptr; |
3211 | Value *Start1 = nullptr, *Step1 = nullptr; |
3212 | BinaryOperator *BO2 = nullptr; |
3213 | Value *Start2 = nullptr, *Step2 = nullptr; |
3214 | if (PN1->getParent() != PN2->getParent() || |
3215 | !matchSimpleRecurrence(P: PN1, BO&: BO1, Start&: Start1, Step&: Step1) || |
3216 | !matchSimpleRecurrence(P: PN2, BO&: BO2, Start&: Start2, Step&: Step2)) |
3217 | break; |
3218 | |
3219 | auto Values = getInvertibleOperands(Op1: cast<Operator>(Val: BO1), |
3220 | Op2: cast<Operator>(Val: BO2)); |
3221 | if (!Values) |
3222 | break; |
3223 | |
3224 | // We have to be careful of mutually defined recurrences here. Ex: |
3225 | // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V |
3226 | // * X_i = Y_i = X_(i-1) OP Y_(i-1) |
3227 | // The invertibility of these is complicated, and not worth reasoning |
3228 | // about (yet?). |
3229 | if (Values->first != PN1 || Values->second != PN2) |
3230 | break; |
3231 | |
3232 | return std::make_pair(x&: Start1, y&: Start2); |
3233 | } |
3234 | } |
3235 | return std::nullopt; |
3236 | } |
3237 | |
3238 | /// Return true if V1 == (binop V2, X), where X is known non-zero. |
3239 | /// Only handle a small subset of binops where (binop V2, X) with non-zero X |
3240 | /// implies V2 != V1. |
3241 | static bool isModifyingBinopOfNonZero(const Value *V1, const Value *V2, |
3242 | unsigned Depth, const SimplifyQuery &Q) { |
3243 | const BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: V1); |
3244 | if (!BO) |
3245 | return false; |
3246 | switch (BO->getOpcode()) { |
3247 | default: |
3248 | break; |
3249 | case Instruction::Or: |
3250 | if (!cast<PossiblyDisjointInst>(Val: V1)->isDisjoint()) |
3251 | break; |
3252 | [[fallthrough]]; |
3253 | case Instruction::Xor: |
3254 | case Instruction::Add: |
3255 | Value *Op = nullptr; |
3256 | if (V2 == BO->getOperand(i_nocapture: 0)) |
3257 | Op = BO->getOperand(i_nocapture: 1); |
3258 | else if (V2 == BO->getOperand(i_nocapture: 1)) |
3259 | Op = BO->getOperand(i_nocapture: 0); |
3260 | else |
3261 | return false; |
3262 | return isKnownNonZero(V: Op, Q, Depth: Depth + 1); |
3263 | } |
3264 | return false; |
3265 | } |
3266 | |
3267 | /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and |
3268 | /// the multiplication is nuw or nsw. |
3269 | static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth, |
3270 | const SimplifyQuery &Q) { |
3271 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: V2)) { |
3272 | const APInt *C; |
3273 | return match(V: OBO, P: m_Mul(L: m_Specific(V: V1), R: m_APInt(Res&: C))) && |
3274 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && |
3275 | !C->isZero() && !C->isOne() && isKnownNonZero(V: V1, Q, Depth: Depth + 1); |
3276 | } |
3277 | return false; |
3278 | } |
3279 | |
3280 | /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and |
3281 | /// the shift is nuw or nsw. |
3282 | static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth, |
3283 | const SimplifyQuery &Q) { |
3284 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: V2)) { |
3285 | const APInt *C; |
3286 | return match(V: OBO, P: m_Shl(L: m_Specific(V: V1), R: m_APInt(Res&: C))) && |
3287 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && |
3288 | !C->isZero() && isKnownNonZero(V: V1, Q, Depth: Depth + 1); |
3289 | } |
3290 | return false; |
3291 | } |
3292 | |
3293 | static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, |
3294 | unsigned Depth, const SimplifyQuery &Q) { |
3295 | // Check two PHIs are in same block. |
3296 | if (PN1->getParent() != PN2->getParent()) |
3297 | return false; |
3298 | |
3299 | SmallPtrSet<const BasicBlock *, 8> VisitedBBs; |
3300 | bool UsedFullRecursion = false; |
3301 | for (const BasicBlock *IncomBB : PN1->blocks()) { |
3302 | if (!VisitedBBs.insert(Ptr: IncomBB).second) |
3303 | continue; // Don't reprocess blocks that we have dealt with already. |
3304 | const Value *IV1 = PN1->getIncomingValueForBlock(BB: IncomBB); |
3305 | const Value *IV2 = PN2->getIncomingValueForBlock(BB: IncomBB); |
3306 | const APInt *C1, *C2; |
3307 | if (match(V: IV1, P: m_APInt(Res&: C1)) && match(V: IV2, P: m_APInt(Res&: C2)) && *C1 != *C2) |
3308 | continue; |
3309 | |
3310 | // Only one pair of phi operands is allowed for full recursion. |
3311 | if (UsedFullRecursion) |
3312 | return false; |
3313 | |
3314 | SimplifyQuery RecQ = Q; |
3315 | RecQ.CxtI = IncomBB->getTerminator(); |
3316 | if (!isKnownNonEqual(V1: IV1, V2: IV2, Depth: Depth + 1, Q: RecQ)) |
3317 | return false; |
3318 | UsedFullRecursion = true; |
3319 | } |
3320 | return true; |
3321 | } |
3322 | |
3323 | static bool isNonEqualSelect(const Value *V1, const Value *V2, unsigned Depth, |
3324 | const SimplifyQuery &Q) { |
3325 | const SelectInst *SI1 = dyn_cast<SelectInst>(Val: V1); |
3326 | if (!SI1) |
3327 | return false; |
3328 | |
3329 | if (const SelectInst *SI2 = dyn_cast<SelectInst>(Val: V2)) { |
3330 | const Value *Cond1 = SI1->getCondition(); |
3331 | const Value *Cond2 = SI2->getCondition(); |
3332 | if (Cond1 == Cond2) |
3333 | return isKnownNonEqual(V1: SI1->getTrueValue(), V2: SI2->getTrueValue(), |
3334 | Depth: Depth + 1, Q) && |
3335 | isKnownNonEqual(V1: SI1->getFalseValue(), V2: SI2->getFalseValue(), |
3336 | Depth: Depth + 1, Q); |
3337 | } |
3338 | return isKnownNonEqual(V1: SI1->getTrueValue(), V2, Depth: Depth + 1, Q) && |
3339 | isKnownNonEqual(V1: SI1->getFalseValue(), V2, Depth: Depth + 1, Q); |
3340 | } |
3341 | |
3342 | // Check to see if A is both a GEP and is the incoming value for a PHI in the |
3343 | // loop, and B is either a ptr or another GEP. If the PHI has 2 incoming values, |
3344 | // one of them being the recursive GEP A and the other a ptr at same base and at |
3345 | // the same/higher offset than B we are only incrementing the pointer further in |
3346 | // loop if offset of recursive GEP is greater than 0. |
3347 | static bool isNonEqualPointersWithRecursiveGEP(const Value *A, const Value *B, |
3348 | const SimplifyQuery &Q) { |
3349 | if (!A->getType()->isPointerTy() || !B->getType()->isPointerTy()) |
3350 | return false; |
3351 | |
3352 | auto *GEPA = dyn_cast<GEPOperator>(Val: A); |
3353 | if (!GEPA || GEPA->getNumIndices() != 1 || !isa<Constant>(Val: GEPA->idx_begin())) |
3354 | return false; |
3355 | |
3356 | // Handle 2 incoming PHI values with one being a recursive GEP. |
3357 | auto *PN = dyn_cast<PHINode>(Val: GEPA->getPointerOperand()); |
3358 | if (!PN || PN->getNumIncomingValues() != 2) |
3359 | return false; |
3360 | |
3361 | // Search for the recursive GEP as an incoming operand, and record that as |
3362 | // Step. |
3363 | Value *Start = nullptr; |
3364 | Value *Step = const_cast<Value *>(A); |
3365 | if (PN->getIncomingValue(i: 0) == Step) |
3366 | Start = PN->getIncomingValue(i: 1); |
3367 | else if (PN->getIncomingValue(i: 1) == Step) |
3368 | Start = PN->getIncomingValue(i: 0); |
3369 | else |
3370 | return false; |
3371 | |
3372 | // Other incoming node base should match the B base. |
3373 | // StartOffset >= OffsetB && StepOffset > 0? |
3374 | // StartOffset <= OffsetB && StepOffset < 0? |
3375 | // Is non-equal if above are true. |
3376 | // We use stripAndAccumulateInBoundsConstantOffsets to restrict the |
3377 | // optimisation to inbounds GEPs only. |
3378 | unsigned IndexWidth = Q.DL.getIndexTypeSizeInBits(Ty: Start->getType()); |
3379 | APInt StartOffset(IndexWidth, 0); |
3380 | Start = Start->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: StartOffset); |
3381 | APInt StepOffset(IndexWidth, 0); |
3382 | Step = Step->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: StepOffset); |
3383 | |
3384 | // Check if Base Pointer of Step matches the PHI. |
3385 | if (Step != PN) |
3386 | return false; |
3387 | APInt OffsetB(IndexWidth, 0); |
3388 | B = B->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: OffsetB); |
3389 | return Start == B && |
3390 | ((StartOffset.sge(RHS: OffsetB) && StepOffset.isStrictlyPositive()) || |
3391 | (StartOffset.sle(RHS: OffsetB) && StepOffset.isNegative())); |
3392 | } |
3393 | |
3394 | /// Return true if it is known that V1 != V2. |
3395 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, |
3396 | const SimplifyQuery &Q) { |
3397 | if (V1 == V2) |
3398 | return false; |
3399 | if (V1->getType() != V2->getType()) |
3400 | // We can't look through casts yet. |
3401 | return false; |
3402 | |
3403 | if (Depth >= MaxAnalysisRecursionDepth) |
3404 | return false; |
3405 | |
3406 | // See if we can recurse through (exactly one of) our operands. This |
3407 | // requires our operation be 1-to-1 and map every input value to exactly |
3408 | // one output value. Such an operation is invertible. |
3409 | auto *O1 = dyn_cast<Operator>(Val: V1); |
3410 | auto *O2 = dyn_cast<Operator>(Val: V2); |
3411 | if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { |
3412 | if (auto Values = getInvertibleOperands(Op1: O1, Op2: O2)) |
3413 | return isKnownNonEqual(V1: Values->first, V2: Values->second, Depth: Depth + 1, Q); |
3414 | |
3415 | if (const PHINode *PN1 = dyn_cast<PHINode>(Val: V1)) { |
3416 | const PHINode *PN2 = cast<PHINode>(Val: V2); |
3417 | // FIXME: This is missing a generalization to handle the case where one is |
3418 | // a PHI and another one isn't. |
3419 | if (isNonEqualPHIs(PN1, PN2, Depth, Q)) |
3420 | return true; |
3421 | }; |
3422 | } |
3423 | |
3424 | if (isModifyingBinopOfNonZero(V1, V2, Depth, Q) || |
3425 | isModifyingBinopOfNonZero(V1: V2, V2: V1, Depth, Q)) |
3426 | return true; |
3427 | |
3428 | if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V1: V2, V2: V1, Depth, Q)) |
3429 | return true; |
3430 | |
3431 | if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V1: V2, V2: V1, Depth, Q)) |
3432 | return true; |
3433 | |
3434 | if (V1->getType()->isIntOrIntVectorTy()) { |
3435 | // Are any known bits in V1 contradictory to known bits in V2? If V1 |
3436 | // has a known zero where V2 has a known one, they must not be equal. |
3437 | KnownBits Known1 = computeKnownBits(V: V1, Depth, Q); |
3438 | if (!Known1.isUnknown()) { |
3439 | KnownBits Known2 = computeKnownBits(V: V2, Depth, Q); |
3440 | if (Known1.Zero.intersects(RHS: Known2.One) || |
3441 | Known2.Zero.intersects(RHS: Known1.One)) |
3442 | return true; |
3443 | } |
3444 | } |
3445 | |
3446 | if (isNonEqualSelect(V1, V2, Depth, Q) || isNonEqualSelect(V1: V2, V2: V1, Depth, Q)) |
3447 | return true; |
3448 | |
3449 | if (isNonEqualPointersWithRecursiveGEP(A: V1, B: V2, Q) || |
3450 | isNonEqualPointersWithRecursiveGEP(A: V2, B: V1, Q)) |
3451 | return true; |
3452 | |
3453 | Value *A, *B; |
3454 | // PtrToInts are NonEqual if their Ptrs are NonEqual. |
3455 | // Check PtrToInt type matches the pointer size. |
3456 | if (match(V: V1, P: m_PtrToIntSameSize(DL: Q.DL, Op: m_Value(V&: A))) && |
3457 | match(V: V2, P: m_PtrToIntSameSize(DL: Q.DL, Op: m_Value(V&: B)))) |
3458 | return isKnownNonEqual(V1: A, V2: B, Depth: Depth + 1, Q); |
3459 | |
3460 | return false; |
3461 | } |
3462 | |
3463 | // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). |
3464 | // Returns the input and lower/upper bounds. |
3465 | static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, |
3466 | const APInt *&CLow, const APInt *&CHigh) { |
3467 | assert(isa<Operator>(Select) && |
3468 | cast<Operator>(Select)->getOpcode() == Instruction::Select && |
3469 | "Input should be a Select!" ); |
3470 | |
3471 | const Value *LHS = nullptr, *RHS = nullptr; |
3472 | SelectPatternFlavor SPF = matchSelectPattern(V: Select, LHS, RHS).Flavor; |
3473 | if (SPF != SPF_SMAX && SPF != SPF_SMIN) |
3474 | return false; |
3475 | |
3476 | if (!match(V: RHS, P: m_APInt(Res&: CLow))) |
3477 | return false; |
3478 | |
3479 | const Value *LHS2 = nullptr, *RHS2 = nullptr; |
3480 | SelectPatternFlavor SPF2 = matchSelectPattern(V: LHS, LHS&: LHS2, RHS&: RHS2).Flavor; |
3481 | if (getInverseMinMaxFlavor(SPF) != SPF2) |
3482 | return false; |
3483 | |
3484 | if (!match(V: RHS2, P: m_APInt(Res&: CHigh))) |
3485 | return false; |
3486 | |
3487 | if (SPF == SPF_SMIN) |
3488 | std::swap(a&: CLow, b&: CHigh); |
3489 | |
3490 | In = LHS2; |
3491 | return CLow->sle(RHS: *CHigh); |
3492 | } |
3493 | |
3494 | static bool isSignedMinMaxIntrinsicClamp(const IntrinsicInst *II, |
3495 | const APInt *&CLow, |
3496 | const APInt *&CHigh) { |
3497 | assert((II->getIntrinsicID() == Intrinsic::smin || |
3498 | II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax" ); |
3499 | |
3500 | Intrinsic::ID InverseID = getInverseMinMaxIntrinsic(MinMaxID: II->getIntrinsicID()); |
3501 | auto *InnerII = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0)); |
3502 | if (!InnerII || InnerII->getIntrinsicID() != InverseID || |
3503 | !match(V: II->getArgOperand(i: 1), P: m_APInt(Res&: CLow)) || |
3504 | !match(V: InnerII->getArgOperand(i: 1), P: m_APInt(Res&: CHigh))) |
3505 | return false; |
3506 | |
3507 | if (II->getIntrinsicID() == Intrinsic::smin) |
3508 | std::swap(a&: CLow, b&: CHigh); |
3509 | return CLow->sle(RHS: *CHigh); |
3510 | } |
3511 | |
3512 | /// For vector constants, loop over the elements and find the constant with the |
3513 | /// minimum number of sign bits. Return 0 if the value is not a vector constant |
3514 | /// or if any element was not analyzed; otherwise, return the count for the |
3515 | /// element with the minimum number of sign bits. |
3516 | static unsigned computeNumSignBitsVectorConstant(const Value *V, |
3517 | const APInt &DemandedElts, |
3518 | unsigned TyBits) { |
3519 | const auto *CV = dyn_cast<Constant>(Val: V); |
3520 | if (!CV || !isa<FixedVectorType>(Val: CV->getType())) |
3521 | return 0; |
3522 | |
3523 | unsigned MinSignBits = TyBits; |
3524 | unsigned NumElts = cast<FixedVectorType>(Val: CV->getType())->getNumElements(); |
3525 | for (unsigned i = 0; i != NumElts; ++i) { |
3526 | if (!DemandedElts[i]) |
3527 | continue; |
3528 | // If we find a non-ConstantInt, bail out. |
3529 | auto *Elt = dyn_cast_or_null<ConstantInt>(Val: CV->getAggregateElement(Elt: i)); |
3530 | if (!Elt) |
3531 | return 0; |
3532 | |
3533 | MinSignBits = std::min(a: MinSignBits, b: Elt->getValue().getNumSignBits()); |
3534 | } |
3535 | |
3536 | return MinSignBits; |
3537 | } |
3538 | |
3539 | static unsigned ComputeNumSignBitsImpl(const Value *V, |
3540 | const APInt &DemandedElts, |
3541 | unsigned Depth, const SimplifyQuery &Q); |
3542 | |
3543 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, |
3544 | unsigned Depth, const SimplifyQuery &Q) { |
3545 | unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); |
3546 | assert(Result > 0 && "At least one sign bit needs to be present!" ); |
3547 | return Result; |
3548 | } |
3549 | |
3550 | /// Return the number of times the sign bit of the register is replicated into |
3551 | /// the other bits. We know that at least 1 bit is always equal to the sign bit |
3552 | /// (itself), but other cases can give us information. For example, immediately |
3553 | /// after an "ashr X, 2", we know that the top 3 bits are all equal to each |
3554 | /// other, so we return 3. For vectors, return the number of sign bits for the |
3555 | /// vector element with the minimum number of known sign bits of the demanded |
3556 | /// elements in the vector specified by DemandedElts. |
3557 | static unsigned ComputeNumSignBitsImpl(const Value *V, |
3558 | const APInt &DemandedElts, |
3559 | unsigned Depth, const SimplifyQuery &Q) { |
3560 | Type *Ty = V->getType(); |
3561 | #ifndef NDEBUG |
3562 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
3563 | |
3564 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
3565 | assert( |
3566 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
3567 | "DemandedElt width should equal the fixed vector number of elements" ); |
3568 | } else { |
3569 | assert(DemandedElts == APInt(1, 1) && |
3570 | "DemandedElt width should be 1 for scalars" ); |
3571 | } |
3572 | #endif |
3573 | |
3574 | // We return the minimum number of sign bits that are guaranteed to be present |
3575 | // in V, so for undef we have to conservatively return 1. We don't have the |
3576 | // same behavior for poison though -- that's a FIXME today. |
3577 | |
3578 | Type *ScalarTy = Ty->getScalarType(); |
3579 | unsigned TyBits = ScalarTy->isPointerTy() ? |
3580 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : |
3581 | Q.DL.getTypeSizeInBits(Ty: ScalarTy); |
3582 | |
3583 | unsigned Tmp, Tmp2; |
3584 | unsigned FirstAnswer = 1; |
3585 | |
3586 | // Note that ConstantInt is handled by the general computeKnownBits case |
3587 | // below. |
3588 | |
3589 | if (Depth == MaxAnalysisRecursionDepth) |
3590 | return 1; |
3591 | |
3592 | if (auto *U = dyn_cast<Operator>(Val: V)) { |
3593 | switch (Operator::getOpcode(V)) { |
3594 | default: break; |
3595 | case Instruction::SExt: |
3596 | Tmp = TyBits - U->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
3597 | return ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q) + Tmp; |
3598 | |
3599 | case Instruction::SDiv: { |
3600 | const APInt *Denominator; |
3601 | // sdiv X, C -> adds log(C) sign bits. |
3602 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: Denominator))) { |
3603 | |
3604 | // Ignore non-positive denominator. |
3605 | if (!Denominator->isStrictlyPositive()) |
3606 | break; |
3607 | |
3608 | // Calculate the incoming numerator bits. |
3609 | unsigned NumBits = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3610 | |
3611 | // Add floor(log(C)) bits to the numerator bits. |
3612 | return std::min(a: TyBits, b: NumBits + Denominator->logBase2()); |
3613 | } |
3614 | break; |
3615 | } |
3616 | |
3617 | case Instruction::SRem: { |
3618 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3619 | |
3620 | const APInt *Denominator; |
3621 | // srem X, C -> we know that the result is within [-C+1,C) when C is a |
3622 | // positive constant. This let us put a lower bound on the number of sign |
3623 | // bits. |
3624 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: Denominator))) { |
3625 | |
3626 | // Ignore non-positive denominator. |
3627 | if (Denominator->isStrictlyPositive()) { |
3628 | // Calculate the leading sign bit constraints by examining the |
3629 | // denominator. Given that the denominator is positive, there are two |
3630 | // cases: |
3631 | // |
3632 | // 1. The numerator is positive. The result range is [0,C) and |
3633 | // [0,C) u< (1 << ceilLogBase2(C)). |
3634 | // |
3635 | // 2. The numerator is negative. Then the result range is (-C,0] and |
3636 | // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). |
3637 | // |
3638 | // Thus a lower bound on the number of sign bits is `TyBits - |
3639 | // ceilLogBase2(C)`. |
3640 | |
3641 | unsigned ResBits = TyBits - Denominator->ceilLogBase2(); |
3642 | Tmp = std::max(a: Tmp, b: ResBits); |
3643 | } |
3644 | } |
3645 | return Tmp; |
3646 | } |
3647 | |
3648 | case Instruction::AShr: { |
3649 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3650 | // ashr X, C -> adds C sign bits. Vectors too. |
3651 | const APInt *ShAmt; |
3652 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: ShAmt))) { |
3653 | if (ShAmt->uge(RHS: TyBits)) |
3654 | break; // Bad shift. |
3655 | unsigned ShAmtLimited = ShAmt->getZExtValue(); |
3656 | Tmp += ShAmtLimited; |
3657 | if (Tmp > TyBits) Tmp = TyBits; |
3658 | } |
3659 | return Tmp; |
3660 | } |
3661 | case Instruction::Shl: { |
3662 | const APInt *ShAmt; |
3663 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: ShAmt))) { |
3664 | // shl destroys sign bits. |
3665 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3666 | if (ShAmt->uge(RHS: TyBits) || // Bad shift. |
3667 | ShAmt->uge(RHS: Tmp)) break; // Shifted all sign bits out. |
3668 | Tmp2 = ShAmt->getZExtValue(); |
3669 | return Tmp - Tmp2; |
3670 | } |
3671 | break; |
3672 | } |
3673 | case Instruction::And: |
3674 | case Instruction::Or: |
3675 | case Instruction::Xor: // NOT is handled here. |
3676 | // Logical binary ops preserve the number of sign bits at the worst. |
3677 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3678 | if (Tmp != 1) { |
3679 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), Depth: Depth + 1, Q); |
3680 | FirstAnswer = std::min(a: Tmp, b: Tmp2); |
3681 | // We computed what we know about the sign bits as our first |
3682 | // answer. Now proceed to the generic code that uses |
3683 | // computeKnownBits, and pick whichever answer is better. |
3684 | } |
3685 | break; |
3686 | |
3687 | case Instruction::Select: { |
3688 | // If we have a clamp pattern, we know that the number of sign bits will |
3689 | // be the minimum of the clamp min/max range. |
3690 | const Value *X; |
3691 | const APInt *CLow, *CHigh; |
3692 | if (isSignedMinMaxClamp(Select: U, In&: X, CLow, CHigh)) |
3693 | return std::min(a: CLow->getNumSignBits(), b: CHigh->getNumSignBits()); |
3694 | |
3695 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 1), Depth: Depth + 1, Q); |
3696 | if (Tmp == 1) break; |
3697 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 2), Depth: Depth + 1, Q); |
3698 | return std::min(a: Tmp, b: Tmp2); |
3699 | } |
3700 | |
3701 | case Instruction::Add: |
3702 | // Add can have at most one carry bit. Thus we know that the output |
3703 | // is, at worst, one more bit than the inputs. |
3704 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3705 | if (Tmp == 1) break; |
3706 | |
3707 | // Special case decrementing a value (ADD X, -1): |
3708 | if (const auto *CRHS = dyn_cast<Constant>(Val: U->getOperand(i: 1))) |
3709 | if (CRHS->isAllOnesValue()) { |
3710 | KnownBits Known(TyBits); |
3711 | computeKnownBits(V: U->getOperand(i: 0), Known, Depth: Depth + 1, Q); |
3712 | |
3713 | // If the input is known to be 0 or 1, the output is 0/-1, which is |
3714 | // all sign bits set. |
3715 | if ((Known.Zero | 1).isAllOnes()) |
3716 | return TyBits; |
3717 | |
3718 | // If we are subtracting one from a positive number, there is no carry |
3719 | // out of the result. |
3720 | if (Known.isNonNegative()) |
3721 | return Tmp; |
3722 | } |
3723 | |
3724 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), Depth: Depth + 1, Q); |
3725 | if (Tmp2 == 1) break; |
3726 | return std::min(a: Tmp, b: Tmp2) - 1; |
3727 | |
3728 | case Instruction::Sub: |
3729 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), Depth: Depth + 1, Q); |
3730 | if (Tmp2 == 1) break; |
3731 | |
3732 | // Handle NEG. |
3733 | if (const auto *CLHS = dyn_cast<Constant>(Val: U->getOperand(i: 0))) |
3734 | if (CLHS->isNullValue()) { |
3735 | KnownBits Known(TyBits); |
3736 | computeKnownBits(V: U->getOperand(i: 1), Known, Depth: Depth + 1, Q); |
3737 | // If the input is known to be 0 or 1, the output is 0/-1, which is |
3738 | // all sign bits set. |
3739 | if ((Known.Zero | 1).isAllOnes()) |
3740 | return TyBits; |
3741 | |
3742 | // If the input is known to be positive (the sign bit is known clear), |
3743 | // the output of the NEG has the same number of sign bits as the |
3744 | // input. |
3745 | if (Known.isNonNegative()) |
3746 | return Tmp2; |
3747 | |
3748 | // Otherwise, we treat this like a SUB. |
3749 | } |
3750 | |
3751 | // Sub can have at most one carry bit. Thus we know that the output |
3752 | // is, at worst, one more bit than the inputs. |
3753 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3754 | if (Tmp == 1) break; |
3755 | return std::min(a: Tmp, b: Tmp2) - 1; |
3756 | |
3757 | case Instruction::Mul: { |
3758 | // The output of the Mul can be at most twice the valid bits in the |
3759 | // inputs. |
3760 | unsigned SignBitsOp0 = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3761 | if (SignBitsOp0 == 1) break; |
3762 | unsigned SignBitsOp1 = ComputeNumSignBits(V: U->getOperand(i: 1), Depth: Depth + 1, Q); |
3763 | if (SignBitsOp1 == 1) break; |
3764 | unsigned OutValidBits = |
3765 | (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); |
3766 | return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; |
3767 | } |
3768 | |
3769 | case Instruction::PHI: { |
3770 | const PHINode *PN = cast<PHINode>(Val: U); |
3771 | unsigned NumIncomingValues = PN->getNumIncomingValues(); |
3772 | // Don't analyze large in-degree PHIs. |
3773 | if (NumIncomingValues > 4) break; |
3774 | // Unreachable blocks may have zero-operand PHI nodes. |
3775 | if (NumIncomingValues == 0) break; |
3776 | |
3777 | // Take the minimum of all incoming values. This can't infinitely loop |
3778 | // because of our depth threshold. |
3779 | SimplifyQuery RecQ = Q; |
3780 | Tmp = TyBits; |
3781 | for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { |
3782 | if (Tmp == 1) return Tmp; |
3783 | RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); |
3784 | Tmp = std::min( |
3785 | a: Tmp, b: ComputeNumSignBits(V: PN->getIncomingValue(i), Depth: Depth + 1, Q: RecQ)); |
3786 | } |
3787 | return Tmp; |
3788 | } |
3789 | |
3790 | case Instruction::Trunc: { |
3791 | // If the input contained enough sign bits that some remain after the |
3792 | // truncation, then we can make use of that. Otherwise we don't know |
3793 | // anything. |
3794 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3795 | unsigned OperandTyBits = U->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
3796 | if (Tmp > (OperandTyBits - TyBits)) |
3797 | return Tmp - (OperandTyBits - TyBits); |
3798 | |
3799 | return 1; |
3800 | } |
3801 | |
3802 | case Instruction::ExtractElement: |
3803 | // Look through extract element. At the moment we keep this simple and |
3804 | // skip tracking the specific element. But at least we might find |
3805 | // information valid for all elements of the vector (for example if vector |
3806 | // is sign extended, shifted, etc). |
3807 | return ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3808 | |
3809 | case Instruction::ShuffleVector: { |
3810 | // Collect the minimum number of sign bits that are shared by every vector |
3811 | // element referenced by the shuffle. |
3812 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: U); |
3813 | if (!Shuf) { |
3814 | // FIXME: Add support for shufflevector constant expressions. |
3815 | return 1; |
3816 | } |
3817 | APInt DemandedLHS, DemandedRHS; |
3818 | // For undef elements, we don't know anything about the common state of |
3819 | // the shuffle result. |
3820 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
3821 | return 1; |
3822 | Tmp = std::numeric_limits<unsigned>::max(); |
3823 | if (!!DemandedLHS) { |
3824 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
3825 | Tmp = ComputeNumSignBits(V: LHS, DemandedElts: DemandedLHS, Depth: Depth + 1, Q); |
3826 | } |
3827 | // If we don't know anything, early out and try computeKnownBits |
3828 | // fall-back. |
3829 | if (Tmp == 1) |
3830 | break; |
3831 | if (!!DemandedRHS) { |
3832 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
3833 | Tmp2 = ComputeNumSignBits(V: RHS, DemandedElts: DemandedRHS, Depth: Depth + 1, Q); |
3834 | Tmp = std::min(a: Tmp, b: Tmp2); |
3835 | } |
3836 | // If we don't know anything, early out and try computeKnownBits |
3837 | // fall-back. |
3838 | if (Tmp == 1) |
3839 | break; |
3840 | assert(Tmp <= TyBits && "Failed to determine minimum sign bits" ); |
3841 | return Tmp; |
3842 | } |
3843 | case Instruction::Call: { |
3844 | if (const auto *II = dyn_cast<IntrinsicInst>(Val: U)) { |
3845 | switch (II->getIntrinsicID()) { |
3846 | default: break; |
3847 | case Intrinsic::abs: |
3848 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Depth: Depth + 1, Q); |
3849 | if (Tmp == 1) break; |
3850 | |
3851 | // Absolute value reduces number of sign bits by at most 1. |
3852 | return Tmp - 1; |
3853 | case Intrinsic::smin: |
3854 | case Intrinsic::smax: { |
3855 | const APInt *CLow, *CHigh; |
3856 | if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh)) |
3857 | return std::min(a: CLow->getNumSignBits(), b: CHigh->getNumSignBits()); |
3858 | } |
3859 | } |
3860 | } |
3861 | } |
3862 | } |
3863 | } |
3864 | |
3865 | // Finally, if we can prove that the top bits of the result are 0's or 1's, |
3866 | // use this information. |
3867 | |
3868 | // If we can examine all elements of a vector constant successfully, we're |
3869 | // done (we can't do any better than that). If not, keep trying. |
3870 | if (unsigned VecSignBits = |
3871 | computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) |
3872 | return VecSignBits; |
3873 | |
3874 | KnownBits Known(TyBits); |
3875 | computeKnownBits(V, DemandedElts, Known, Depth, Q); |
3876 | |
3877 | // If we know that the sign bit is either zero or one, determine the number of |
3878 | // identical bits in the top of the input value. |
3879 | return std::max(a: FirstAnswer, b: Known.countMinSignBits()); |
3880 | } |
3881 | |
3882 | Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, |
3883 | const TargetLibraryInfo *TLI) { |
3884 | const Function *F = CB.getCalledFunction(); |
3885 | if (!F) |
3886 | return Intrinsic::not_intrinsic; |
3887 | |
3888 | if (F->isIntrinsic()) |
3889 | return F->getIntrinsicID(); |
3890 | |
3891 | // We are going to infer semantics of a library function based on mapping it |
3892 | // to an LLVM intrinsic. Check that the library function is available from |
3893 | // this callbase and in this environment. |
3894 | LibFunc Func; |
3895 | if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, F&: Func) || |
3896 | !CB.onlyReadsMemory()) |
3897 | return Intrinsic::not_intrinsic; |
3898 | |
3899 | switch (Func) { |
3900 | default: |
3901 | break; |
3902 | case LibFunc_sin: |
3903 | case LibFunc_sinf: |
3904 | case LibFunc_sinl: |
3905 | return Intrinsic::sin; |
3906 | case LibFunc_cos: |
3907 | case LibFunc_cosf: |
3908 | case LibFunc_cosl: |
3909 | return Intrinsic::cos; |
3910 | case LibFunc_exp: |
3911 | case LibFunc_expf: |
3912 | case LibFunc_expl: |
3913 | return Intrinsic::exp; |
3914 | case LibFunc_exp2: |
3915 | case LibFunc_exp2f: |
3916 | case LibFunc_exp2l: |
3917 | return Intrinsic::exp2; |
3918 | case LibFunc_log: |
3919 | case LibFunc_logf: |
3920 | case LibFunc_logl: |
3921 | return Intrinsic::log; |
3922 | case LibFunc_log10: |
3923 | case LibFunc_log10f: |
3924 | case LibFunc_log10l: |
3925 | return Intrinsic::log10; |
3926 | case LibFunc_log2: |
3927 | case LibFunc_log2f: |
3928 | case LibFunc_log2l: |
3929 | return Intrinsic::log2; |
3930 | case LibFunc_fabs: |
3931 | case LibFunc_fabsf: |
3932 | case LibFunc_fabsl: |
3933 | return Intrinsic::fabs; |
3934 | case LibFunc_fmin: |
3935 | case LibFunc_fminf: |
3936 | case LibFunc_fminl: |
3937 | return Intrinsic::minnum; |
3938 | case LibFunc_fmax: |
3939 | case LibFunc_fmaxf: |
3940 | case LibFunc_fmaxl: |
3941 | return Intrinsic::maxnum; |
3942 | case LibFunc_copysign: |
3943 | case LibFunc_copysignf: |
3944 | case LibFunc_copysignl: |
3945 | return Intrinsic::copysign; |
3946 | case LibFunc_floor: |
3947 | case LibFunc_floorf: |
3948 | case LibFunc_floorl: |
3949 | return Intrinsic::floor; |
3950 | case LibFunc_ceil: |
3951 | case LibFunc_ceilf: |
3952 | case LibFunc_ceill: |
3953 | return Intrinsic::ceil; |
3954 | case LibFunc_trunc: |
3955 | case LibFunc_truncf: |
3956 | case LibFunc_truncl: |
3957 | return Intrinsic::trunc; |
3958 | case LibFunc_rint: |
3959 | case LibFunc_rintf: |
3960 | case LibFunc_rintl: |
3961 | return Intrinsic::rint; |
3962 | case LibFunc_nearbyint: |
3963 | case LibFunc_nearbyintf: |
3964 | case LibFunc_nearbyintl: |
3965 | return Intrinsic::nearbyint; |
3966 | case LibFunc_round: |
3967 | case LibFunc_roundf: |
3968 | case LibFunc_roundl: |
3969 | return Intrinsic::round; |
3970 | case LibFunc_roundeven: |
3971 | case LibFunc_roundevenf: |
3972 | case LibFunc_roundevenl: |
3973 | return Intrinsic::roundeven; |
3974 | case LibFunc_pow: |
3975 | case LibFunc_powf: |
3976 | case LibFunc_powl: |
3977 | return Intrinsic::pow; |
3978 | case LibFunc_sqrt: |
3979 | case LibFunc_sqrtf: |
3980 | case LibFunc_sqrtl: |
3981 | return Intrinsic::sqrt; |
3982 | } |
3983 | |
3984 | return Intrinsic::not_intrinsic; |
3985 | } |
3986 | |
3987 | /// Return true if it's possible to assume IEEE treatment of input denormals in |
3988 | /// \p F for \p Val. |
3989 | static bool inputDenormalIsIEEE(const Function &F, const Type *Ty) { |
3990 | Ty = Ty->getScalarType(); |
3991 | return F.getDenormalMode(FPType: Ty->getFltSemantics()).Input == DenormalMode::IEEE; |
3992 | } |
3993 | |
3994 | static bool inputDenormalIsIEEEOrPosZero(const Function &F, const Type *Ty) { |
3995 | Ty = Ty->getScalarType(); |
3996 | DenormalMode Mode = F.getDenormalMode(FPType: Ty->getFltSemantics()); |
3997 | return Mode.Input == DenormalMode::IEEE || |
3998 | Mode.Input == DenormalMode::PositiveZero; |
3999 | } |
4000 | |
4001 | static bool outputDenormalIsIEEEOrPosZero(const Function &F, const Type *Ty) { |
4002 | Ty = Ty->getScalarType(); |
4003 | DenormalMode Mode = F.getDenormalMode(FPType: Ty->getFltSemantics()); |
4004 | return Mode.Output == DenormalMode::IEEE || |
4005 | Mode.Output == DenormalMode::PositiveZero; |
4006 | } |
4007 | |
4008 | bool KnownFPClass::isKnownNeverLogicalZero(const Function &F, Type *Ty) const { |
4009 | return isKnownNeverZero() && |
4010 | (isKnownNeverSubnormal() || inputDenormalIsIEEE(F, Ty)); |
4011 | } |
4012 | |
4013 | bool KnownFPClass::isKnownNeverLogicalNegZero(const Function &F, |
4014 | Type *Ty) const { |
4015 | return isKnownNeverNegZero() && |
4016 | (isKnownNeverNegSubnormal() || inputDenormalIsIEEEOrPosZero(F, Ty)); |
4017 | } |
4018 | |
4019 | bool KnownFPClass::isKnownNeverLogicalPosZero(const Function &F, |
4020 | Type *Ty) const { |
4021 | if (!isKnownNeverPosZero()) |
4022 | return false; |
4023 | |
4024 | // If we know there are no denormals, nothing can be flushed to zero. |
4025 | if (isKnownNeverSubnormal()) |
4026 | return true; |
4027 | |
4028 | DenormalMode Mode = F.getDenormalMode(FPType: Ty->getScalarType()->getFltSemantics()); |
4029 | switch (Mode.Input) { |
4030 | case DenormalMode::IEEE: |
4031 | return true; |
4032 | case DenormalMode::PreserveSign: |
4033 | // Negative subnormal won't flush to +0 |
4034 | return isKnownNeverPosSubnormal(); |
4035 | case DenormalMode::PositiveZero: |
4036 | default: |
4037 | // Both positive and negative subnormal could flush to +0 |
4038 | return false; |
4039 | } |
4040 | |
4041 | llvm_unreachable("covered switch over denormal mode" ); |
4042 | } |
4043 | |
4044 | void KnownFPClass::propagateDenormal(const KnownFPClass &Src, const Function &F, |
4045 | Type *Ty) { |
4046 | KnownFPClasses = Src.KnownFPClasses; |
4047 | // If we aren't assuming the source can't be a zero, we don't have to check if |
4048 | // a denormal input could be flushed. |
4049 | if (!Src.isKnownNeverPosZero() && !Src.isKnownNeverNegZero()) |
4050 | return; |
4051 | |
4052 | // If we know the input can't be a denormal, it can't be flushed to 0. |
4053 | if (Src.isKnownNeverSubnormal()) |
4054 | return; |
4055 | |
4056 | DenormalMode Mode = F.getDenormalMode(FPType: Ty->getScalarType()->getFltSemantics()); |
4057 | |
4058 | if (!Src.isKnownNeverPosSubnormal() && Mode != DenormalMode::getIEEE()) |
4059 | KnownFPClasses |= fcPosZero; |
4060 | |
4061 | if (!Src.isKnownNeverNegSubnormal() && Mode != DenormalMode::getIEEE()) { |
4062 | if (Mode != DenormalMode::getPositiveZero()) |
4063 | KnownFPClasses |= fcNegZero; |
4064 | |
4065 | if (Mode.Input == DenormalMode::PositiveZero || |
4066 | Mode.Output == DenormalMode::PositiveZero || |
4067 | Mode.Input == DenormalMode::Dynamic || |
4068 | Mode.Output == DenormalMode::Dynamic) |
4069 | KnownFPClasses |= fcPosZero; |
4070 | } |
4071 | } |
4072 | |
4073 | void KnownFPClass::propagateCanonicalizingSrc(const KnownFPClass &Src, |
4074 | const Function &F, Type *Ty) { |
4075 | propagateDenormal(Src, F, Ty); |
4076 | propagateNaN(Src, /*PreserveSign=*/true); |
4077 | } |
4078 | |
4079 | /// Given an exploded icmp instruction, return true if the comparison only |
4080 | /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if |
4081 | /// the result of the comparison is true when the input value is signed. |
4082 | bool llvm::isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS, |
4083 | bool &TrueIfSigned) { |
4084 | switch (Pred) { |
4085 | case ICmpInst::ICMP_SLT: // True if LHS s< 0 |
4086 | TrueIfSigned = true; |
4087 | return RHS.isZero(); |
4088 | case ICmpInst::ICMP_SLE: // True if LHS s<= -1 |
4089 | TrueIfSigned = true; |
4090 | return RHS.isAllOnes(); |
4091 | case ICmpInst::ICMP_SGT: // True if LHS s> -1 |
4092 | TrueIfSigned = false; |
4093 | return RHS.isAllOnes(); |
4094 | case ICmpInst::ICMP_SGE: // True if LHS s>= 0 |
4095 | TrueIfSigned = false; |
4096 | return RHS.isZero(); |
4097 | case ICmpInst::ICMP_UGT: |
4098 | // True if LHS u> RHS and RHS == sign-bit-mask - 1 |
4099 | TrueIfSigned = true; |
4100 | return RHS.isMaxSignedValue(); |
4101 | case ICmpInst::ICMP_UGE: |
4102 | // True if LHS u>= RHS and RHS == sign-bit-mask (2^7, 2^15, 2^31, etc) |
4103 | TrueIfSigned = true; |
4104 | return RHS.isMinSignedValue(); |
4105 | case ICmpInst::ICMP_ULT: |
4106 | // True if LHS u< RHS and RHS == sign-bit-mask (2^7, 2^15, 2^31, etc) |
4107 | TrueIfSigned = false; |
4108 | return RHS.isMinSignedValue(); |
4109 | case ICmpInst::ICMP_ULE: |
4110 | // True if LHS u<= RHS and RHS == sign-bit-mask - 1 |
4111 | TrueIfSigned = false; |
4112 | return RHS.isMaxSignedValue(); |
4113 | default: |
4114 | return false; |
4115 | } |
4116 | } |
4117 | |
4118 | /// Returns a pair of values, which if passed to llvm.is.fpclass, returns the |
4119 | /// same result as an fcmp with the given operands. |
4120 | std::pair<Value *, FPClassTest> llvm::fcmpToClassTest(FCmpInst::Predicate Pred, |
4121 | const Function &F, |
4122 | Value *LHS, Value *RHS, |
4123 | bool LookThroughSrc) { |
4124 | const APFloat *ConstRHS; |
4125 | if (!match(V: RHS, P: m_APFloatAllowPoison(Res&: ConstRHS))) |
4126 | return {nullptr, fcAllFlags}; |
4127 | |
4128 | return fcmpToClassTest(Pred, F, LHS, ConstRHS, LookThroughSrc); |
4129 | } |
4130 | |
4131 | std::pair<Value *, FPClassTest> |
4132 | llvm::fcmpToClassTest(FCmpInst::Predicate Pred, const Function &F, Value *LHS, |
4133 | const APFloat *ConstRHS, bool LookThroughSrc) { |
4134 | |
4135 | auto [Src, ClassIfTrue, ClassIfFalse] = |
4136 | fcmpImpliesClass(Pred, F, LHS, RHS: *ConstRHS, LookThroughSrc); |
4137 | if (Src && ClassIfTrue == ~ClassIfFalse) |
4138 | return {Src, ClassIfTrue}; |
4139 | return {nullptr, fcAllFlags}; |
4140 | } |
4141 | |
4142 | /// Return the return value for fcmpImpliesClass for a compare that produces an |
4143 | /// exact class test. |
4144 | static std::tuple<Value *, FPClassTest, FPClassTest> exactClass(Value *V, |
4145 | FPClassTest M) { |
4146 | return {V, M, ~M}; |
4147 | } |
4148 | |
4149 | std::tuple<Value *, FPClassTest, FPClassTest> |
4150 | llvm::fcmpImpliesClass(CmpInst::Predicate Pred, const Function &F, Value *LHS, |
4151 | FPClassTest RHSClass, bool LookThroughSrc) { |
4152 | assert(RHSClass != fcNone); |
4153 | Value *Src = LHS; |
4154 | |
4155 | if (Pred == FCmpInst::FCMP_TRUE) |
4156 | return exactClass(V: Src, M: fcAllFlags); |
4157 | |
4158 | if (Pred == FCmpInst::FCMP_FALSE) |
4159 | return exactClass(V: Src, M: fcNone); |
4160 | |
4161 | const FPClassTest OrigClass = RHSClass; |
4162 | |
4163 | const bool IsNegativeRHS = (RHSClass & fcNegative) == RHSClass; |
4164 | const bool IsPositiveRHS = (RHSClass & fcPositive) == RHSClass; |
4165 | const bool IsNaN = (RHSClass & ~fcNan) == fcNone; |
4166 | |
4167 | if (IsNaN) { |
4168 | // fcmp o__ x, nan -> false |
4169 | // fcmp u__ x, nan -> true |
4170 | return exactClass(V: Src, M: CmpInst::isOrdered(predicate: Pred) ? fcNone : fcAllFlags); |
4171 | } |
4172 | |
4173 | // fcmp ord x, zero|normal|subnormal|inf -> ~fcNan |
4174 | if (Pred == FCmpInst::FCMP_ORD) |
4175 | return exactClass(V: Src, M: ~fcNan); |
4176 | |
4177 | // fcmp uno x, zero|normal|subnormal|inf -> fcNan |
4178 | if (Pred == FCmpInst::FCMP_UNO) |
4179 | return exactClass(V: Src, M: fcNan); |
4180 | |
4181 | const bool IsFabs = LookThroughSrc && match(V: LHS, P: m_FAbs(Op0: m_Value(V&: Src))); |
4182 | if (IsFabs) |
4183 | RHSClass = llvm::inverse_fabs(Mask: RHSClass); |
4184 | |
4185 | const bool IsZero = (OrigClass & fcZero) == OrigClass; |
4186 | if (IsZero) { |
4187 | assert(Pred != FCmpInst::FCMP_ORD && Pred != FCmpInst::FCMP_UNO); |
4188 | // Compares with fcNone are only exactly equal to fcZero if input denormals |
4189 | // are not flushed. |
4190 | // TODO: Handle DAZ by expanding masks to cover subnormal cases. |
4191 | if (!inputDenormalIsIEEE(F, Ty: LHS->getType())) |
4192 | return {nullptr, fcAllFlags, fcAllFlags}; |
4193 | |
4194 | switch (Pred) { |
4195 | case FCmpInst::FCMP_OEQ: // Match x == 0.0 |
4196 | return exactClass(V: Src, M: fcZero); |
4197 | case FCmpInst::FCMP_UEQ: // Match isnan(x) || (x == 0.0) |
4198 | return exactClass(V: Src, M: fcZero | fcNan); |
4199 | case FCmpInst::FCMP_UNE: // Match (x != 0.0) |
4200 | return exactClass(V: Src, M: ~fcZero); |
4201 | case FCmpInst::FCMP_ONE: // Match !isnan(x) && x != 0.0 |
4202 | return exactClass(V: Src, M: ~fcNan & ~fcZero); |
4203 | case FCmpInst::FCMP_ORD: |
4204 | // Canonical form of ord/uno is with a zero. We could also handle |
4205 | // non-canonical other non-NaN constants or LHS == RHS. |
4206 | return exactClass(V: Src, M: ~fcNan); |
4207 | case FCmpInst::FCMP_UNO: |
4208 | return exactClass(V: Src, M: fcNan); |
4209 | case FCmpInst::FCMP_OGT: // x > 0 |
4210 | return exactClass(V: Src, M: fcPosSubnormal | fcPosNormal | fcPosInf); |
4211 | case FCmpInst::FCMP_UGT: // isnan(x) || x > 0 |
4212 | return exactClass(V: Src, M: fcPosSubnormal | fcPosNormal | fcPosInf | fcNan); |
4213 | case FCmpInst::FCMP_OGE: // x >= 0 |
4214 | return exactClass(V: Src, M: fcPositive | fcNegZero); |
4215 | case FCmpInst::FCMP_UGE: // isnan(x) || x >= 0 |
4216 | return exactClass(V: Src, M: fcPositive | fcNegZero | fcNan); |
4217 | case FCmpInst::FCMP_OLT: // x < 0 |
4218 | return exactClass(V: Src, M: fcNegSubnormal | fcNegNormal | fcNegInf); |
4219 | case FCmpInst::FCMP_ULT: // isnan(x) || x < 0 |
4220 | return exactClass(V: Src, M: fcNegSubnormal | fcNegNormal | fcNegInf | fcNan); |
4221 | case FCmpInst::FCMP_OLE: // x <= 0 |
4222 | return exactClass(V: Src, M: fcNegative | fcPosZero); |
4223 | case FCmpInst::FCMP_ULE: // isnan(x) || x <= 0 |
4224 | return exactClass(V: Src, M: fcNegative | fcPosZero | fcNan); |
4225 | default: |
4226 | llvm_unreachable("all compare types are handled" ); |
4227 | } |
4228 | |
4229 | return {nullptr, fcAllFlags, fcAllFlags}; |
4230 | } |
4231 | |
4232 | const bool IsDenormalRHS = (OrigClass & fcSubnormal) == OrigClass; |
4233 | |
4234 | const bool IsInf = (OrigClass & fcInf) == OrigClass; |
4235 | if (IsInf) { |
4236 | FPClassTest Mask = fcAllFlags; |
4237 | |
4238 | switch (Pred) { |
4239 | case FCmpInst::FCMP_OEQ: |
4240 | case FCmpInst::FCMP_UNE: { |
4241 | // Match __builtin_isinf patterns |
4242 | // |
4243 | // fcmp oeq x, +inf -> is_fpclass x, fcPosInf |
4244 | // fcmp oeq fabs(x), +inf -> is_fpclass x, fcInf |
4245 | // fcmp oeq x, -inf -> is_fpclass x, fcNegInf |
4246 | // fcmp oeq fabs(x), -inf -> is_fpclass x, 0 -> false |
4247 | // |
4248 | // fcmp une x, +inf -> is_fpclass x, ~fcPosInf |
4249 | // fcmp une fabs(x), +inf -> is_fpclass x, ~fcInf |
4250 | // fcmp une x, -inf -> is_fpclass x, ~fcNegInf |
4251 | // fcmp une fabs(x), -inf -> is_fpclass x, fcAllFlags -> true |
4252 | if (IsNegativeRHS) { |
4253 | Mask = fcNegInf; |
4254 | if (IsFabs) |
4255 | Mask = fcNone; |
4256 | } else { |
4257 | Mask = fcPosInf; |
4258 | if (IsFabs) |
4259 | Mask |= fcNegInf; |
4260 | } |
4261 | break; |
4262 | } |
4263 | case FCmpInst::FCMP_ONE: |
4264 | case FCmpInst::FCMP_UEQ: { |
4265 | // Match __builtin_isinf patterns |
4266 | // fcmp one x, -inf -> is_fpclass x, fcNegInf |
4267 | // fcmp one fabs(x), -inf -> is_fpclass x, ~fcNegInf & ~fcNan |
4268 | // fcmp one x, +inf -> is_fpclass x, ~fcNegInf & ~fcNan |
4269 | // fcmp one fabs(x), +inf -> is_fpclass x, ~fcInf & fcNan |
4270 | // |
4271 | // fcmp ueq x, +inf -> is_fpclass x, fcPosInf|fcNan |
4272 | // fcmp ueq (fabs x), +inf -> is_fpclass x, fcInf|fcNan |
4273 | // fcmp ueq x, -inf -> is_fpclass x, fcNegInf|fcNan |
4274 | // fcmp ueq fabs(x), -inf -> is_fpclass x, fcNan |
4275 | if (IsNegativeRHS) { |
4276 | Mask = ~fcNegInf & ~fcNan; |
4277 | if (IsFabs) |
4278 | Mask = ~fcNan; |
4279 | } else { |
4280 | Mask = ~fcPosInf & ~fcNan; |
4281 | if (IsFabs) |
4282 | Mask &= ~fcNegInf; |
4283 | } |
4284 | |
4285 | break; |
4286 | } |
4287 | case FCmpInst::FCMP_OLT: |
4288 | case FCmpInst::FCMP_UGE: { |
4289 | if (IsNegativeRHS) { |
4290 | // No value is ordered and less than negative infinity. |
4291 | // All values are unordered with or at least negative infinity. |
4292 | // fcmp olt x, -inf -> false |
4293 | // fcmp uge x, -inf -> true |
4294 | Mask = fcNone; |
4295 | break; |
4296 | } |
4297 | |
4298 | // fcmp olt fabs(x), +inf -> fcFinite |
4299 | // fcmp uge fabs(x), +inf -> ~fcFinite |
4300 | // fcmp olt x, +inf -> fcFinite|fcNegInf |
4301 | // fcmp uge x, +inf -> ~(fcFinite|fcNegInf) |
4302 | Mask = fcFinite; |
4303 | if (!IsFabs) |
4304 | Mask |= fcNegInf; |
4305 | break; |
4306 | } |
4307 | case FCmpInst::FCMP_OGE: |
4308 | case FCmpInst::FCMP_ULT: { |
4309 | if (IsNegativeRHS) { |
4310 | // fcmp oge x, -inf -> ~fcNan |
4311 | // fcmp oge fabs(x), -inf -> ~fcNan |
4312 | // fcmp ult x, -inf -> fcNan |
4313 | // fcmp ult fabs(x), -inf -> fcNan |
4314 | Mask = ~fcNan; |
4315 | break; |
4316 | } |
4317 | |
4318 | // fcmp oge fabs(x), +inf -> fcInf |
4319 | // fcmp oge x, +inf -> fcPosInf |
4320 | // fcmp ult fabs(x), +inf -> ~fcInf |
4321 | // fcmp ult x, +inf -> ~fcPosInf |
4322 | Mask = fcPosInf; |
4323 | if (IsFabs) |
4324 | Mask |= fcNegInf; |
4325 | break; |
4326 | } |
4327 | case FCmpInst::FCMP_OGT: |
4328 | case FCmpInst::FCMP_ULE: { |
4329 | if (IsNegativeRHS) { |
4330 | // fcmp ogt x, -inf -> fcmp one x, -inf |
4331 | // fcmp ogt fabs(x), -inf -> fcmp ord x, x |
4332 | // fcmp ule x, -inf -> fcmp ueq x, -inf |
4333 | // fcmp ule fabs(x), -inf -> fcmp uno x, x |
4334 | Mask = IsFabs ? ~fcNan : ~(fcNegInf | fcNan); |
4335 | break; |
4336 | } |
4337 | |
4338 | // No value is ordered and greater than infinity. |
4339 | Mask = fcNone; |
4340 | break; |
4341 | } |
4342 | case FCmpInst::FCMP_OLE: |
4343 | case FCmpInst::FCMP_UGT: { |
4344 | if (IsNegativeRHS) { |
4345 | Mask = IsFabs ? fcNone : fcNegInf; |
4346 | break; |
4347 | } |
4348 | |
4349 | // fcmp ole x, +inf -> fcmp ord x, x |
4350 | // fcmp ole fabs(x), +inf -> fcmp ord x, x |
4351 | // fcmp ole x, -inf -> fcmp oeq x, -inf |
4352 | // fcmp ole fabs(x), -inf -> false |
4353 | Mask = ~fcNan; |
4354 | break; |
4355 | } |
4356 | default: |
4357 | llvm_unreachable("all compare types are handled" ); |
4358 | } |
4359 | |
4360 | // Invert the comparison for the unordered cases. |
4361 | if (FCmpInst::isUnordered(predicate: Pred)) |
4362 | Mask = ~Mask; |
4363 | |
4364 | return exactClass(V: Src, M: Mask); |
4365 | } |
4366 | |
4367 | if (Pred == FCmpInst::FCMP_OEQ) |
4368 | return {Src, RHSClass, fcAllFlags}; |
4369 | |
4370 | if (Pred == FCmpInst::FCMP_UEQ) { |
4371 | FPClassTest Class = RHSClass | fcNan; |
4372 | return {Src, Class, ~fcNan}; |
4373 | } |
4374 | |
4375 | if (Pred == FCmpInst::FCMP_ONE) |
4376 | return {Src, ~fcNan, RHSClass | fcNan}; |
4377 | |
4378 | if (Pred == FCmpInst::FCMP_UNE) |
4379 | return {Src, fcAllFlags, RHSClass}; |
4380 | |
4381 | assert((RHSClass == fcNone || RHSClass == fcPosNormal || |
4382 | RHSClass == fcNegNormal || RHSClass == fcNormal || |
4383 | RHSClass == fcPosSubnormal || RHSClass == fcNegSubnormal || |
4384 | RHSClass == fcSubnormal) && |
4385 | "should have been recognized as an exact class test" ); |
4386 | |
4387 | if (IsNegativeRHS) { |
4388 | // TODO: Handle fneg(fabs) |
4389 | if (IsFabs) { |
4390 | // fabs(x) o> -k -> fcmp ord x, x |
4391 | // fabs(x) u> -k -> true |
4392 | // fabs(x) o< -k -> false |
4393 | // fabs(x) u< -k -> fcmp uno x, x |
4394 | switch (Pred) { |
4395 | case FCmpInst::FCMP_OGT: |
4396 | case FCmpInst::FCMP_OGE: |
4397 | return {Src, ~fcNan, fcNan}; |
4398 | case FCmpInst::FCMP_UGT: |
4399 | case FCmpInst::FCMP_UGE: |
4400 | return {Src, fcAllFlags, fcNone}; |
4401 | case FCmpInst::FCMP_OLT: |
4402 | case FCmpInst::FCMP_OLE: |
4403 | return {Src, fcNone, fcAllFlags}; |
4404 | case FCmpInst::FCMP_ULT: |
4405 | case FCmpInst::FCMP_ULE: |
4406 | return {Src, fcNan, ~fcNan}; |
4407 | default: |
4408 | break; |
4409 | } |
4410 | |
4411 | return {nullptr, fcAllFlags, fcAllFlags}; |
4412 | } |
4413 | |
4414 | FPClassTest ClassesLE = fcNegInf | fcNegNormal; |
4415 | FPClassTest ClassesGE = fcPositive | fcNegZero | fcNegSubnormal; |
4416 | |
4417 | if (IsDenormalRHS) |
4418 | ClassesLE |= fcNegSubnormal; |
4419 | else |
4420 | ClassesGE |= fcNegNormal; |
4421 | |
4422 | switch (Pred) { |
4423 | case FCmpInst::FCMP_OGT: |
4424 | case FCmpInst::FCMP_OGE: |
4425 | return {Src, ClassesGE, ~ClassesGE | RHSClass}; |
4426 | case FCmpInst::FCMP_UGT: |
4427 | case FCmpInst::FCMP_UGE: |
4428 | return {Src, ClassesGE | fcNan, ~(ClassesGE | fcNan) | RHSClass}; |
4429 | case FCmpInst::FCMP_OLT: |
4430 | case FCmpInst::FCMP_OLE: |
4431 | return {Src, ClassesLE, ~ClassesLE | RHSClass}; |
4432 | case FCmpInst::FCMP_ULT: |
4433 | case FCmpInst::FCMP_ULE: |
4434 | return {Src, ClassesLE | fcNan, ~(ClassesLE | fcNan) | RHSClass}; |
4435 | default: |
4436 | break; |
4437 | } |
4438 | } else if (IsPositiveRHS) { |
4439 | FPClassTest ClassesGE = fcPosNormal | fcPosInf; |
4440 | FPClassTest ClassesLE = fcNegative | fcPosZero | fcPosSubnormal; |
4441 | if (IsDenormalRHS) |
4442 | ClassesGE |= fcPosSubnormal; |
4443 | else |
4444 | ClassesLE |= fcPosNormal; |
4445 | |
4446 | if (IsFabs) { |
4447 | ClassesGE = llvm::inverse_fabs(Mask: ClassesGE); |
4448 | ClassesLE = llvm::inverse_fabs(Mask: ClassesLE); |
4449 | } |
4450 | |
4451 | switch (Pred) { |
4452 | case FCmpInst::FCMP_OGT: |
4453 | case FCmpInst::FCMP_OGE: |
4454 | return {Src, ClassesGE, ~ClassesGE | RHSClass}; |
4455 | case FCmpInst::FCMP_UGT: |
4456 | case FCmpInst::FCMP_UGE: |
4457 | return {Src, ClassesGE | fcNan, ~(ClassesGE | fcNan) | RHSClass}; |
4458 | case FCmpInst::FCMP_OLT: |
4459 | case FCmpInst::FCMP_OLE: |
4460 | return {Src, ClassesLE, ~ClassesLE | RHSClass}; |
4461 | case FCmpInst::FCMP_ULT: |
4462 | case FCmpInst::FCMP_ULE: |
4463 | return {Src, ClassesLE | fcNan, ~(ClassesLE | fcNan) | RHSClass}; |
4464 | default: |
4465 | break; |
4466 | } |
4467 | } |
4468 | |
4469 | return {nullptr, fcAllFlags, fcAllFlags}; |
4470 | } |
4471 | |
4472 | std::tuple<Value *, FPClassTest, FPClassTest> |
4473 | llvm::fcmpImpliesClass(CmpInst::Predicate Pred, const Function &F, Value *LHS, |
4474 | const APFloat &ConstRHS, bool LookThroughSrc) { |
4475 | // We can refine checks against smallest normal / largest denormal to an |
4476 | // exact class test. |
4477 | if (!ConstRHS.isNegative() && ConstRHS.isSmallestNormalized()) { |
4478 | Value *Src = LHS; |
4479 | const bool IsFabs = LookThroughSrc && match(V: LHS, P: m_FAbs(Op0: m_Value(V&: Src))); |
4480 | |
4481 | FPClassTest Mask; |
4482 | // Match pattern that's used in __builtin_isnormal. |
4483 | switch (Pred) { |
4484 | case FCmpInst::FCMP_OLT: |
4485 | case FCmpInst::FCMP_UGE: { |
4486 | // fcmp olt x, smallest_normal -> fcNegInf|fcNegNormal|fcSubnormal|fcZero |
4487 | // fcmp olt fabs(x), smallest_normal -> fcSubnormal|fcZero |
4488 | // fcmp uge x, smallest_normal -> fcNan|fcPosNormal|fcPosInf |
4489 | // fcmp uge fabs(x), smallest_normal -> ~(fcSubnormal|fcZero) |
4490 | Mask = fcZero | fcSubnormal; |
4491 | if (!IsFabs) |
4492 | Mask |= fcNegNormal | fcNegInf; |
4493 | |
4494 | break; |
4495 | } |
4496 | case FCmpInst::FCMP_OGE: |
4497 | case FCmpInst::FCMP_ULT: { |
4498 | // fcmp oge x, smallest_normal -> fcPosNormal | fcPosInf |
4499 | // fcmp oge fabs(x), smallest_normal -> fcInf | fcNormal |
4500 | // fcmp ult x, smallest_normal -> ~(fcPosNormal | fcPosInf) |
4501 | // fcmp ult fabs(x), smallest_normal -> ~(fcInf | fcNormal) |
4502 | Mask = fcPosInf | fcPosNormal; |
4503 | if (IsFabs) |
4504 | Mask |= fcNegInf | fcNegNormal; |
4505 | break; |
4506 | } |
4507 | default: |
4508 | return fcmpImpliesClass(Pred, F, LHS, RHSClass: ConstRHS.classify(), |
4509 | LookThroughSrc); |
4510 | } |
4511 | |
4512 | // Invert the comparison for the unordered cases. |
4513 | if (FCmpInst::isUnordered(predicate: Pred)) |
4514 | Mask = ~Mask; |
4515 | |
4516 | return exactClass(V: Src, M: Mask); |
4517 | } |
4518 | |
4519 | return fcmpImpliesClass(Pred, F, LHS, RHSClass: ConstRHS.classify(), LookThroughSrc); |
4520 | } |
4521 | |
4522 | std::tuple<Value *, FPClassTest, FPClassTest> |
4523 | llvm::fcmpImpliesClass(CmpInst::Predicate Pred, const Function &F, Value *LHS, |
4524 | Value *RHS, bool LookThroughSrc) { |
4525 | const APFloat *ConstRHS; |
4526 | if (!match(V: RHS, P: m_APFloatAllowPoison(Res&: ConstRHS))) |
4527 | return {nullptr, fcAllFlags, fcAllFlags}; |
4528 | |
4529 | // TODO: Just call computeKnownFPClass for RHS to handle non-constants. |
4530 | return fcmpImpliesClass(Pred, F, LHS, ConstRHS: *ConstRHS, LookThroughSrc); |
4531 | } |
4532 | |
4533 | static void computeKnownFPClassFromCond(const Value *V, Value *Cond, |
4534 | bool CondIsTrue, |
4535 | const Instruction *CxtI, |
4536 | KnownFPClass &KnownFromContext) { |
4537 | CmpInst::Predicate Pred; |
4538 | Value *LHS; |
4539 | uint64_t ClassVal = 0; |
4540 | const APFloat *CRHS; |
4541 | const APInt *RHS; |
4542 | if (match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: LHS), R: m_APFloat(Res&: CRHS)))) { |
4543 | auto [CmpVal, MaskIfTrue, MaskIfFalse] = fcmpImpliesClass( |
4544 | Pred, F: *CxtI->getParent()->getParent(), LHS, ConstRHS: *CRHS, LookThroughSrc: LHS != V); |
4545 | if (CmpVal == V) |
4546 | KnownFromContext.knownNot(RuleOut: ~(CondIsTrue ? MaskIfTrue : MaskIfFalse)); |
4547 | } else if (match(Cond, m_Intrinsic<Intrinsic::is_fpclass>( |
4548 | m_Value(LHS), m_ConstantInt(ClassVal)))) { |
4549 | FPClassTest Mask = static_cast<FPClassTest>(ClassVal); |
4550 | KnownFromContext.knownNot(RuleOut: CondIsTrue ? ~Mask : Mask); |
4551 | } else if (match(V: Cond, P: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Value(V&: LHS)), |
4552 | R: m_APInt(Res&: RHS)))) { |
4553 | bool TrueIfSigned; |
4554 | if (!isSignBitCheck(Pred, RHS: *RHS, TrueIfSigned)) |
4555 | return; |
4556 | if (TrueIfSigned == CondIsTrue) |
4557 | KnownFromContext.signBitMustBeOne(); |
4558 | else |
4559 | KnownFromContext.signBitMustBeZero(); |
4560 | } |
4561 | } |
4562 | |
4563 | static KnownFPClass computeKnownFPClassFromContext(const Value *V, |
4564 | const SimplifyQuery &Q) { |
4565 | KnownFPClass KnownFromContext; |
4566 | |
4567 | if (!Q.CxtI) |
4568 | return KnownFromContext; |
4569 | |
4570 | if (Q.DC && Q.DT) { |
4571 | // Handle dominating conditions. |
4572 | for (BranchInst *BI : Q.DC->conditionsFor(V)) { |
4573 | Value *Cond = BI->getCondition(); |
4574 | |
4575 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
4576 | if (Q.DT->dominates(BBE: Edge0, BB: Q.CxtI->getParent())) |
4577 | computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/true, CxtI: Q.CxtI, |
4578 | KnownFromContext); |
4579 | |
4580 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
4581 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CxtI->getParent())) |
4582 | computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/false, CxtI: Q.CxtI, |
4583 | KnownFromContext); |
4584 | } |
4585 | } |
4586 | |
4587 | if (!Q.AC) |
4588 | return KnownFromContext; |
4589 | |
4590 | // Try to restrict the floating-point classes based on information from |
4591 | // assumptions. |
4592 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { |
4593 | if (!AssumeVH) |
4594 | continue; |
4595 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
4596 | |
4597 | assert(I->getFunction() == Q.CxtI->getParent()->getParent() && |
4598 | "Got assumption for the wrong function!" ); |
4599 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && |
4600 | "must be an assume intrinsic" ); |
4601 | |
4602 | if (!isValidAssumeForContext(Inv: I, CxtI: Q.CxtI, DT: Q.DT)) |
4603 | continue; |
4604 | |
4605 | computeKnownFPClassFromCond(V, Cond: I->getArgOperand(i: 0), /*CondIsTrue=*/true, |
4606 | CxtI: Q.CxtI, KnownFromContext); |
4607 | } |
4608 | |
4609 | return KnownFromContext; |
4610 | } |
4611 | |
4612 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, |
4613 | FPClassTest InterestedClasses, KnownFPClass &Known, |
4614 | unsigned Depth, const SimplifyQuery &Q); |
4615 | |
4616 | static void computeKnownFPClass(const Value *V, KnownFPClass &Known, |
4617 | FPClassTest InterestedClasses, unsigned Depth, |
4618 | const SimplifyQuery &Q) { |
4619 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
4620 | APInt DemandedElts = |
4621 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
4622 | computeKnownFPClass(V, DemandedElts, InterestedClasses, Known, Depth, Q); |
4623 | } |
4624 | |
4625 | static void computeKnownFPClassForFPTrunc(const Operator *Op, |
4626 | const APInt &DemandedElts, |
4627 | FPClassTest InterestedClasses, |
4628 | KnownFPClass &Known, unsigned Depth, |
4629 | const SimplifyQuery &Q) { |
4630 | if ((InterestedClasses & |
4631 | (KnownFPClass::OrderedLessThanZeroMask | fcNan)) == fcNone) |
4632 | return; |
4633 | |
4634 | KnownFPClass KnownSrc; |
4635 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
4636 | Known&: KnownSrc, Depth: Depth + 1, Q); |
4637 | |
4638 | // Sign should be preserved |
4639 | // TODO: Handle cannot be ordered greater than zero |
4640 | if (KnownSrc.cannotBeOrderedLessThanZero()) |
4641 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
4642 | |
4643 | Known.propagateNaN(Src: KnownSrc, PreserveSign: true); |
4644 | |
4645 | // Infinity needs a range check. |
4646 | } |
4647 | |
4648 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, |
4649 | FPClassTest InterestedClasses, KnownFPClass &Known, |
4650 | unsigned Depth, const SimplifyQuery &Q) { |
4651 | assert(Known.isUnknown() && "should not be called with known information" ); |
4652 | |
4653 | if (!DemandedElts) { |
4654 | // No demanded elts, better to assume we don't know anything. |
4655 | Known.resetAll(); |
4656 | return; |
4657 | } |
4658 | |
4659 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
4660 | |
4661 | if (auto *CFP = dyn_cast<ConstantFP>(Val: V)) { |
4662 | Known.KnownFPClasses = CFP->getValueAPF().classify(); |
4663 | Known.SignBit = CFP->isNegative(); |
4664 | return; |
4665 | } |
4666 | |
4667 | if (isa<ConstantAggregateZero>(Val: V)) { |
4668 | Known.KnownFPClasses = fcPosZero; |
4669 | Known.SignBit = false; |
4670 | return; |
4671 | } |
4672 | |
4673 | if (isa<PoisonValue>(Val: V)) { |
4674 | Known.KnownFPClasses = fcNone; |
4675 | Known.SignBit = false; |
4676 | return; |
4677 | } |
4678 | |
4679 | // Try to handle fixed width vector constants |
4680 | auto *VFVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
4681 | const Constant *CV = dyn_cast<Constant>(Val: V); |
4682 | if (VFVTy && CV) { |
4683 | Known.KnownFPClasses = fcNone; |
4684 | bool SignBitAllZero = true; |
4685 | bool SignBitAllOne = true; |
4686 | |
4687 | // For vectors, verify that each element is not NaN. |
4688 | unsigned NumElts = VFVTy->getNumElements(); |
4689 | for (unsigned i = 0; i != NumElts; ++i) { |
4690 | if (!DemandedElts[i]) |
4691 | continue; |
4692 | |
4693 | Constant *Elt = CV->getAggregateElement(Elt: i); |
4694 | if (!Elt) { |
4695 | Known = KnownFPClass(); |
4696 | return; |
4697 | } |
4698 | if (isa<UndefValue>(Val: Elt)) |
4699 | continue; |
4700 | auto *CElt = dyn_cast<ConstantFP>(Val: Elt); |
4701 | if (!CElt) { |
4702 | Known = KnownFPClass(); |
4703 | return; |
4704 | } |
4705 | |
4706 | const APFloat &C = CElt->getValueAPF(); |
4707 | Known.KnownFPClasses |= C.classify(); |
4708 | if (C.isNegative()) |
4709 | SignBitAllZero = false; |
4710 | else |
4711 | SignBitAllOne = false; |
4712 | } |
4713 | if (SignBitAllOne != SignBitAllZero) |
4714 | Known.SignBit = SignBitAllOne; |
4715 | return; |
4716 | } |
4717 | |
4718 | FPClassTest KnownNotFromFlags = fcNone; |
4719 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) |
4720 | KnownNotFromFlags |= CB->getRetNoFPClass(); |
4721 | else if (const auto *Arg = dyn_cast<Argument>(Val: V)) |
4722 | KnownNotFromFlags |= Arg->getNoFPClass(); |
4723 | |
4724 | const Operator *Op = dyn_cast<Operator>(Val: V); |
4725 | if (const FPMathOperator *FPOp = dyn_cast_or_null<FPMathOperator>(Val: Op)) { |
4726 | if (FPOp->hasNoNaNs()) |
4727 | KnownNotFromFlags |= fcNan; |
4728 | if (FPOp->hasNoInfs()) |
4729 | KnownNotFromFlags |= fcInf; |
4730 | } |
4731 | |
4732 | KnownFPClass AssumedClasses = computeKnownFPClassFromContext(V, Q); |
4733 | KnownNotFromFlags |= ~AssumedClasses.KnownFPClasses; |
4734 | |
4735 | // We no longer need to find out about these bits from inputs if we can |
4736 | // assume this from flags/attributes. |
4737 | InterestedClasses &= ~KnownNotFromFlags; |
4738 | |
4739 | auto ClearClassesFromFlags = make_scope_exit(F: [=, &Known] { |
4740 | Known.knownNot(RuleOut: KnownNotFromFlags); |
4741 | if (!Known.SignBit && AssumedClasses.SignBit) { |
4742 | if (*AssumedClasses.SignBit) |
4743 | Known.signBitMustBeOne(); |
4744 | else |
4745 | Known.signBitMustBeZero(); |
4746 | } |
4747 | }); |
4748 | |
4749 | if (!Op) |
4750 | return; |
4751 | |
4752 | // All recursive calls that increase depth must come after this. |
4753 | if (Depth == MaxAnalysisRecursionDepth) |
4754 | return; |
4755 | |
4756 | const unsigned Opc = Op->getOpcode(); |
4757 | switch (Opc) { |
4758 | case Instruction::FNeg: { |
4759 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
4760 | Known, Depth: Depth + 1, Q); |
4761 | Known.fneg(); |
4762 | break; |
4763 | } |
4764 | case Instruction::Select: { |
4765 | Value *Cond = Op->getOperand(i: 0); |
4766 | Value *LHS = Op->getOperand(i: 1); |
4767 | Value *RHS = Op->getOperand(i: 2); |
4768 | |
4769 | FPClassTest FilterLHS = fcAllFlags; |
4770 | FPClassTest FilterRHS = fcAllFlags; |
4771 | |
4772 | Value *TestedValue = nullptr; |
4773 | FPClassTest MaskIfTrue = fcAllFlags; |
4774 | FPClassTest MaskIfFalse = fcAllFlags; |
4775 | uint64_t ClassVal = 0; |
4776 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
4777 | CmpInst::Predicate Pred; |
4778 | Value *CmpLHS, *CmpRHS; |
4779 | if (F && match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS)))) { |
4780 | // If the select filters out a value based on the class, it no longer |
4781 | // participates in the class of the result |
4782 | |
4783 | // TODO: In some degenerate cases we can infer something if we try again |
4784 | // without looking through sign operations. |
4785 | bool LookThroughFAbsFNeg = CmpLHS != LHS && CmpLHS != RHS; |
4786 | std::tie(args&: TestedValue, args&: MaskIfTrue, args&: MaskIfFalse) = |
4787 | fcmpImpliesClass(Pred, F: *F, LHS: CmpLHS, RHS: CmpRHS, LookThroughSrc: LookThroughFAbsFNeg); |
4788 | } else if (match(Cond, |
4789 | m_Intrinsic<Intrinsic::is_fpclass>( |
4790 | m_Value(TestedValue), m_ConstantInt(ClassVal)))) { |
4791 | FPClassTest TestedMask = static_cast<FPClassTest>(ClassVal); |
4792 | MaskIfTrue = TestedMask; |
4793 | MaskIfFalse = ~TestedMask; |
4794 | } |
4795 | |
4796 | if (TestedValue == LHS) { |
4797 | // match !isnan(x) ? x : y |
4798 | FilterLHS = MaskIfTrue; |
4799 | } else if (TestedValue == RHS) { // && IsExactClass |
4800 | // match !isnan(x) ? y : x |
4801 | FilterRHS = MaskIfFalse; |
4802 | } |
4803 | |
4804 | KnownFPClass Known2; |
4805 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: InterestedClasses & FilterLHS, Known, |
4806 | Depth: Depth + 1, Q); |
4807 | Known.KnownFPClasses &= FilterLHS; |
4808 | |
4809 | computeKnownFPClass(V: RHS, DemandedElts, InterestedClasses: InterestedClasses & FilterRHS, |
4810 | Known&: Known2, Depth: Depth + 1, Q); |
4811 | Known2.KnownFPClasses &= FilterRHS; |
4812 | |
4813 | Known |= Known2; |
4814 | break; |
4815 | } |
4816 | case Instruction::Call: { |
4817 | const CallInst *II = cast<CallInst>(Val: Op); |
4818 | const Intrinsic::ID IID = II->getIntrinsicID(); |
4819 | switch (IID) { |
4820 | case Intrinsic::fabs: { |
4821 | if ((InterestedClasses & (fcNan | fcPositive)) != fcNone) { |
4822 | // If we only care about the sign bit we don't need to inspect the |
4823 | // operand. |
4824 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
4825 | InterestedClasses, Known, Depth: Depth + 1, Q); |
4826 | } |
4827 | |
4828 | Known.fabs(); |
4829 | break; |
4830 | } |
4831 | case Intrinsic::copysign: { |
4832 | KnownFPClass KnownSign; |
4833 | |
4834 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
4835 | Known, Depth: Depth + 1, Q); |
4836 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses, |
4837 | Known&: KnownSign, Depth: Depth + 1, Q); |
4838 | Known.copysign(Sign: KnownSign); |
4839 | break; |
4840 | } |
4841 | case Intrinsic::fma: |
4842 | case Intrinsic::fmuladd: { |
4843 | if ((InterestedClasses & fcNegative) == fcNone) |
4844 | break; |
4845 | |
4846 | if (II->getArgOperand(i: 0) != II->getArgOperand(i: 1)) |
4847 | break; |
4848 | |
4849 | // The multiply cannot be -0 and therefore the add can't be -0 |
4850 | Known.knownNot(RuleOut: fcNegZero); |
4851 | |
4852 | // x * x + y is non-negative if y is non-negative. |
4853 | KnownFPClass KnownAddend; |
4854 | computeKnownFPClass(V: II->getArgOperand(i: 2), DemandedElts, InterestedClasses, |
4855 | Known&: KnownAddend, Depth: Depth + 1, Q); |
4856 | |
4857 | if (KnownAddend.cannotBeOrderedLessThanZero()) |
4858 | Known.knownNot(RuleOut: fcNegative); |
4859 | break; |
4860 | } |
4861 | case Intrinsic::sqrt: |
4862 | case Intrinsic::experimental_constrained_sqrt: { |
4863 | KnownFPClass KnownSrc; |
4864 | FPClassTest InterestedSrcs = InterestedClasses; |
4865 | if (InterestedClasses & fcNan) |
4866 | InterestedSrcs |= KnownFPClass::OrderedLessThanZeroMask; |
4867 | |
4868 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
4869 | Known&: KnownSrc, Depth: Depth + 1, Q); |
4870 | |
4871 | if (KnownSrc.isKnownNeverPosInfinity()) |
4872 | Known.knownNot(RuleOut: fcPosInf); |
4873 | if (KnownSrc.isKnownNever(Mask: fcSNan)) |
4874 | Known.knownNot(RuleOut: fcSNan); |
4875 | |
4876 | // Any negative value besides -0 returns a nan. |
4877 | if (KnownSrc.isKnownNeverNaN() && KnownSrc.cannotBeOrderedLessThanZero()) |
4878 | Known.knownNot(RuleOut: fcNan); |
4879 | |
4880 | // The only negative value that can be returned is -0 for -0 inputs. |
4881 | Known.knownNot(RuleOut: fcNegInf | fcNegSubnormal | fcNegNormal); |
4882 | |
4883 | // If the input denormal mode could be PreserveSign, a negative |
4884 | // subnormal input could produce a negative zero output. |
4885 | const Function *F = II->getFunction(); |
4886 | if (Q.IIQ.hasNoSignedZeros(Op: II) || |
4887 | (F && KnownSrc.isKnownNeverLogicalNegZero(F: *F, Ty: II->getType()))) { |
4888 | Known.knownNot(RuleOut: fcNegZero); |
4889 | if (KnownSrc.isKnownNeverNaN()) |
4890 | Known.signBitMustBeZero(); |
4891 | } |
4892 | |
4893 | break; |
4894 | } |
4895 | case Intrinsic::sin: |
4896 | case Intrinsic::cos: { |
4897 | // Return NaN on infinite inputs. |
4898 | KnownFPClass KnownSrc; |
4899 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
4900 | Known&: KnownSrc, Depth: Depth + 1, Q); |
4901 | Known.knownNot(RuleOut: fcInf); |
4902 | if (KnownSrc.isKnownNeverNaN() && KnownSrc.isKnownNeverInfinity()) |
4903 | Known.knownNot(RuleOut: fcNan); |
4904 | break; |
4905 | } |
4906 | case Intrinsic::maxnum: |
4907 | case Intrinsic::minnum: |
4908 | case Intrinsic::minimum: |
4909 | case Intrinsic::maximum: { |
4910 | KnownFPClass KnownLHS, KnownRHS; |
4911 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
4912 | Known&: KnownLHS, Depth: Depth + 1, Q); |
4913 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses, |
4914 | Known&: KnownRHS, Depth: Depth + 1, Q); |
4915 | |
4916 | bool NeverNaN = KnownLHS.isKnownNeverNaN() || KnownRHS.isKnownNeverNaN(); |
4917 | Known = KnownLHS | KnownRHS; |
4918 | |
4919 | // If either operand is not NaN, the result is not NaN. |
4920 | if (NeverNaN && (IID == Intrinsic::minnum || IID == Intrinsic::maxnum)) |
4921 | Known.knownNot(RuleOut: fcNan); |
4922 | |
4923 | if (IID == Intrinsic::maxnum) { |
4924 | // If at least one operand is known to be positive, the result must be |
4925 | // positive. |
4926 | if ((KnownLHS.cannotBeOrderedLessThanZero() && |
4927 | KnownLHS.isKnownNeverNaN()) || |
4928 | (KnownRHS.cannotBeOrderedLessThanZero() && |
4929 | KnownRHS.isKnownNeverNaN())) |
4930 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
4931 | } else if (IID == Intrinsic::maximum) { |
4932 | // If at least one operand is known to be positive, the result must be |
4933 | // positive. |
4934 | if (KnownLHS.cannotBeOrderedLessThanZero() || |
4935 | KnownRHS.cannotBeOrderedLessThanZero()) |
4936 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
4937 | } else if (IID == Intrinsic::minnum) { |
4938 | // If at least one operand is known to be negative, the result must be |
4939 | // negative. |
4940 | if ((KnownLHS.cannotBeOrderedGreaterThanZero() && |
4941 | KnownLHS.isKnownNeverNaN()) || |
4942 | (KnownRHS.cannotBeOrderedGreaterThanZero() && |
4943 | KnownRHS.isKnownNeverNaN())) |
4944 | Known.knownNot(RuleOut: KnownFPClass::OrderedGreaterThanZeroMask); |
4945 | } else { |
4946 | // If at least one operand is known to be negative, the result must be |
4947 | // negative. |
4948 | if (KnownLHS.cannotBeOrderedGreaterThanZero() || |
4949 | KnownRHS.cannotBeOrderedGreaterThanZero()) |
4950 | Known.knownNot(RuleOut: KnownFPClass::OrderedGreaterThanZeroMask); |
4951 | } |
4952 | |
4953 | // Fixup zero handling if denormals could be returned as a zero. |
4954 | // |
4955 | // As there's no spec for denormal flushing, be conservative with the |
4956 | // treatment of denormals that could be flushed to zero. For older |
4957 | // subtargets on AMDGPU the min/max instructions would not flush the |
4958 | // output and return the original value. |
4959 | // |
4960 | if ((Known.KnownFPClasses & fcZero) != fcNone && |
4961 | !Known.isKnownNeverSubnormal()) { |
4962 | const Function *Parent = II->getFunction(); |
4963 | if (!Parent) |
4964 | break; |
4965 | |
4966 | DenormalMode Mode = Parent->getDenormalMode( |
4967 | FPType: II->getType()->getScalarType()->getFltSemantics()); |
4968 | if (Mode != DenormalMode::getIEEE()) |
4969 | Known.KnownFPClasses |= fcZero; |
4970 | } |
4971 | |
4972 | if (Known.isKnownNeverNaN()) { |
4973 | if (KnownLHS.SignBit && KnownRHS.SignBit && |
4974 | *KnownLHS.SignBit == *KnownRHS.SignBit) { |
4975 | if (*KnownLHS.SignBit) |
4976 | Known.signBitMustBeOne(); |
4977 | else |
4978 | Known.signBitMustBeZero(); |
4979 | } else if ((IID == Intrinsic::maximum || IID == Intrinsic::minimum) || |
4980 | ((KnownLHS.isKnownNeverNegZero() || |
4981 | KnownRHS.isKnownNeverPosZero()) && |
4982 | (KnownLHS.isKnownNeverPosZero() || |
4983 | KnownRHS.isKnownNeverNegZero()))) { |
4984 | if ((IID == Intrinsic::maximum || IID == Intrinsic::maxnum) && |
4985 | (KnownLHS.SignBit == false || KnownRHS.SignBit == false)) |
4986 | Known.signBitMustBeZero(); |
4987 | else if ((IID == Intrinsic::minimum || IID == Intrinsic::minnum) && |
4988 | (KnownLHS.SignBit == true || KnownRHS.SignBit == true)) |
4989 | Known.signBitMustBeOne(); |
4990 | } |
4991 | } |
4992 | break; |
4993 | } |
4994 | case Intrinsic::canonicalize: { |
4995 | KnownFPClass KnownSrc; |
4996 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
4997 | Known&: KnownSrc, Depth: Depth + 1, Q); |
4998 | |
4999 | // This is essentially a stronger form of |
5000 | // propagateCanonicalizingSrc. Other "canonicalizing" operations don't |
5001 | // actually have an IR canonicalization guarantee. |
5002 | |
5003 | // Canonicalize may flush denormals to zero, so we have to consider the |
5004 | // denormal mode to preserve known-not-0 knowledge. |
5005 | Known.KnownFPClasses = KnownSrc.KnownFPClasses | fcZero | fcQNan; |
5006 | |
5007 | // Stronger version of propagateNaN |
5008 | // Canonicalize is guaranteed to quiet signaling nans. |
5009 | if (KnownSrc.isKnownNeverNaN()) |
5010 | Known.knownNot(RuleOut: fcNan); |
5011 | else |
5012 | Known.knownNot(RuleOut: fcSNan); |
5013 | |
5014 | const Function *F = II->getFunction(); |
5015 | if (!F) |
5016 | break; |
5017 | |
5018 | // If the parent function flushes denormals, the canonical output cannot |
5019 | // be a denormal. |
5020 | const fltSemantics &FPType = |
5021 | II->getType()->getScalarType()->getFltSemantics(); |
5022 | DenormalMode DenormMode = F->getDenormalMode(FPType); |
5023 | if (DenormMode == DenormalMode::getIEEE()) { |
5024 | if (KnownSrc.isKnownNever(Mask: fcPosZero)) |
5025 | Known.knownNot(RuleOut: fcPosZero); |
5026 | if (KnownSrc.isKnownNever(Mask: fcNegZero)) |
5027 | Known.knownNot(RuleOut: fcNegZero); |
5028 | break; |
5029 | } |
5030 | |
5031 | if (DenormMode.inputsAreZero() || DenormMode.outputsAreZero()) |
5032 | Known.knownNot(RuleOut: fcSubnormal); |
5033 | |
5034 | if (DenormMode.Input == DenormalMode::PositiveZero || |
5035 | (DenormMode.Output == DenormalMode::PositiveZero && |
5036 | DenormMode.Input == DenormalMode::IEEE)) |
5037 | Known.knownNot(RuleOut: fcNegZero); |
5038 | |
5039 | break; |
5040 | } |
5041 | case Intrinsic::vector_reduce_fmax: |
5042 | case Intrinsic::vector_reduce_fmin: |
5043 | case Intrinsic::vector_reduce_fmaximum: |
5044 | case Intrinsic::vector_reduce_fminimum: { |
5045 | // reduce min/max will choose an element from one of the vector elements, |
5046 | // so we can infer and class information that is common to all elements. |
5047 | Known = computeKnownFPClass(V: II->getArgOperand(i: 0), FMF: II->getFastMathFlags(), |
5048 | InterestedClasses, Depth: Depth + 1, SQ: Q); |
5049 | // Can only propagate sign if output is never NaN. |
5050 | if (!Known.isKnownNeverNaN()) |
5051 | Known.SignBit.reset(); |
5052 | break; |
5053 | } |
5054 | case Intrinsic::trunc: |
5055 | case Intrinsic::floor: |
5056 | case Intrinsic::ceil: |
5057 | case Intrinsic::rint: |
5058 | case Intrinsic::nearbyint: |
5059 | case Intrinsic::round: |
5060 | case Intrinsic::roundeven: { |
5061 | KnownFPClass KnownSrc; |
5062 | FPClassTest InterestedSrcs = InterestedClasses; |
5063 | if (InterestedSrcs & fcPosFinite) |
5064 | InterestedSrcs |= fcPosFinite; |
5065 | if (InterestedSrcs & fcNegFinite) |
5066 | InterestedSrcs |= fcNegFinite; |
5067 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
5068 | Known&: KnownSrc, Depth: Depth + 1, Q); |
5069 | |
5070 | // Integer results cannot be subnormal. |
5071 | Known.knownNot(RuleOut: fcSubnormal); |
5072 | |
5073 | Known.propagateNaN(Src: KnownSrc, PreserveSign: true); |
5074 | |
5075 | // Pass through infinities, except PPC_FP128 is a special case for |
5076 | // intrinsics other than trunc. |
5077 | if (IID == Intrinsic::trunc || !V->getType()->isMultiUnitFPType()) { |
5078 | if (KnownSrc.isKnownNeverPosInfinity()) |
5079 | Known.knownNot(RuleOut: fcPosInf); |
5080 | if (KnownSrc.isKnownNeverNegInfinity()) |
5081 | Known.knownNot(RuleOut: fcNegInf); |
5082 | } |
5083 | |
5084 | // Negative round ups to 0 produce -0 |
5085 | if (KnownSrc.isKnownNever(Mask: fcPosFinite)) |
5086 | Known.knownNot(RuleOut: fcPosFinite); |
5087 | if (KnownSrc.isKnownNever(Mask: fcNegFinite)) |
5088 | Known.knownNot(RuleOut: fcNegFinite); |
5089 | |
5090 | break; |
5091 | } |
5092 | case Intrinsic::exp: |
5093 | case Intrinsic::exp2: |
5094 | case Intrinsic::exp10: { |
5095 | Known.knownNot(RuleOut: fcNegative); |
5096 | if ((InterestedClasses & fcNan) == fcNone) |
5097 | break; |
5098 | |
5099 | KnownFPClass KnownSrc; |
5100 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
5101 | Known&: KnownSrc, Depth: Depth + 1, Q); |
5102 | if (KnownSrc.isKnownNeverNaN()) { |
5103 | Known.knownNot(RuleOut: fcNan); |
5104 | Known.signBitMustBeZero(); |
5105 | } |
5106 | |
5107 | break; |
5108 | } |
5109 | case Intrinsic::fptrunc_round: { |
5110 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, Known, |
5111 | Depth, Q); |
5112 | break; |
5113 | } |
5114 | case Intrinsic::log: |
5115 | case Intrinsic::log10: |
5116 | case Intrinsic::log2: |
5117 | case Intrinsic::experimental_constrained_log: |
5118 | case Intrinsic::experimental_constrained_log10: |
5119 | case Intrinsic::experimental_constrained_log2: { |
5120 | // log(+inf) -> +inf |
5121 | // log([+-]0.0) -> -inf |
5122 | // log(-inf) -> nan |
5123 | // log(-x) -> nan |
5124 | if ((InterestedClasses & (fcNan | fcInf)) == fcNone) |
5125 | break; |
5126 | |
5127 | FPClassTest InterestedSrcs = InterestedClasses; |
5128 | if ((InterestedClasses & fcNegInf) != fcNone) |
5129 | InterestedSrcs |= fcZero | fcSubnormal; |
5130 | if ((InterestedClasses & fcNan) != fcNone) |
5131 | InterestedSrcs |= fcNan | (fcNegative & ~fcNan); |
5132 | |
5133 | KnownFPClass KnownSrc; |
5134 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
5135 | Known&: KnownSrc, Depth: Depth + 1, Q); |
5136 | |
5137 | if (KnownSrc.isKnownNeverPosInfinity()) |
5138 | Known.knownNot(RuleOut: fcPosInf); |
5139 | |
5140 | if (KnownSrc.isKnownNeverNaN() && KnownSrc.cannotBeOrderedLessThanZero()) |
5141 | Known.knownNot(RuleOut: fcNan); |
5142 | |
5143 | const Function *F = II->getFunction(); |
5144 | if (F && KnownSrc.isKnownNeverLogicalZero(F: *F, Ty: II->getType())) |
5145 | Known.knownNot(RuleOut: fcNegInf); |
5146 | |
5147 | break; |
5148 | } |
5149 | case Intrinsic::powi: { |
5150 | if ((InterestedClasses & fcNegative) == fcNone) |
5151 | break; |
5152 | |
5153 | const Value *Exp = II->getArgOperand(i: 1); |
5154 | Type *ExpTy = Exp->getType(); |
5155 | unsigned BitWidth = ExpTy->getScalarType()->getIntegerBitWidth(); |
5156 | KnownBits ExponentKnownBits(BitWidth); |
5157 | computeKnownBits(V: Exp, DemandedElts: isa<VectorType>(Val: ExpTy) ? DemandedElts : APInt(1, 1), |
5158 | Known&: ExponentKnownBits, Depth: Depth + 1, Q); |
5159 | |
5160 | if (ExponentKnownBits.Zero[0]) { // Is even |
5161 | Known.knownNot(RuleOut: fcNegative); |
5162 | break; |
5163 | } |
5164 | |
5165 | // Given that exp is an integer, here are the |
5166 | // ways that pow can return a negative value: |
5167 | // |
5168 | // pow(-x, exp) --> negative if exp is odd and x is negative. |
5169 | // pow(-0, exp) --> -inf if exp is negative odd. |
5170 | // pow(-0, exp) --> -0 if exp is positive odd. |
5171 | // pow(-inf, exp) --> -0 if exp is negative odd. |
5172 | // pow(-inf, exp) --> -inf if exp is positive odd. |
5173 | KnownFPClass KnownSrc; |
5174 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: fcNegative, |
5175 | Known&: KnownSrc, Depth: Depth + 1, Q); |
5176 | if (KnownSrc.isKnownNever(Mask: fcNegative)) |
5177 | Known.knownNot(RuleOut: fcNegative); |
5178 | break; |
5179 | } |
5180 | case Intrinsic::ldexp: { |
5181 | KnownFPClass KnownSrc; |
5182 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
5183 | Known&: KnownSrc, Depth: Depth + 1, Q); |
5184 | Known.propagateNaN(Src: KnownSrc, /*PropagateSign=*/PreserveSign: true); |
5185 | |
5186 | // Sign is preserved, but underflows may produce zeroes. |
5187 | if (KnownSrc.isKnownNever(Mask: fcNegative)) |
5188 | Known.knownNot(RuleOut: fcNegative); |
5189 | else if (KnownSrc.cannotBeOrderedLessThanZero()) |
5190 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
5191 | |
5192 | if (KnownSrc.isKnownNever(Mask: fcPositive)) |
5193 | Known.knownNot(RuleOut: fcPositive); |
5194 | else if (KnownSrc.cannotBeOrderedGreaterThanZero()) |
5195 | Known.knownNot(RuleOut: KnownFPClass::OrderedGreaterThanZeroMask); |
5196 | |
5197 | // Can refine inf/zero handling based on the exponent operand. |
5198 | const FPClassTest ExpInfoMask = fcZero | fcSubnormal | fcInf; |
5199 | if ((InterestedClasses & ExpInfoMask) == fcNone) |
5200 | break; |
5201 | if ((KnownSrc.KnownFPClasses & ExpInfoMask) == fcNone) |
5202 | break; |
5203 | |
5204 | const fltSemantics &Flt = |
5205 | II->getType()->getScalarType()->getFltSemantics(); |
5206 | unsigned Precision = APFloat::semanticsPrecision(Flt); |
5207 | const Value *ExpArg = II->getArgOperand(i: 1); |
5208 | ConstantRange ExpRange = computeConstantRange( |
5209 | V: ExpArg, ForSigned: true, UseInstrInfo: Q.IIQ.UseInstrInfo, AC: Q.AC, CtxI: Q.CxtI, DT: Q.DT, Depth: Depth + 1); |
5210 | |
5211 | const int MantissaBits = Precision - 1; |
5212 | if (ExpRange.getSignedMin().sge(RHS: static_cast<int64_t>(MantissaBits))) |
5213 | Known.knownNot(RuleOut: fcSubnormal); |
5214 | |
5215 | const Function *F = II->getFunction(); |
5216 | const APInt *ConstVal = ExpRange.getSingleElement(); |
5217 | if (ConstVal && ConstVal->isZero()) { |
5218 | // ldexp(x, 0) -> x, so propagate everything. |
5219 | Known.propagateCanonicalizingSrc(Src: KnownSrc, F: *F, Ty: II->getType()); |
5220 | } else if (ExpRange.isAllNegative()) { |
5221 | // If we know the power is <= 0, can't introduce inf |
5222 | if (KnownSrc.isKnownNeverPosInfinity()) |
5223 | Known.knownNot(RuleOut: fcPosInf); |
5224 | if (KnownSrc.isKnownNeverNegInfinity()) |
5225 | Known.knownNot(RuleOut: fcNegInf); |
5226 | } else if (ExpRange.isAllNonNegative()) { |
5227 | // If we know the power is >= 0, can't introduce subnormal or zero |
5228 | if (KnownSrc.isKnownNeverPosSubnormal()) |
5229 | Known.knownNot(RuleOut: fcPosSubnormal); |
5230 | if (KnownSrc.isKnownNeverNegSubnormal()) |
5231 | Known.knownNot(RuleOut: fcNegSubnormal); |
5232 | if (F && KnownSrc.isKnownNeverLogicalPosZero(F: *F, Ty: II->getType())) |
5233 | Known.knownNot(RuleOut: fcPosZero); |
5234 | if (F && KnownSrc.isKnownNeverLogicalNegZero(F: *F, Ty: II->getType())) |
5235 | Known.knownNot(RuleOut: fcNegZero); |
5236 | } |
5237 | |
5238 | break; |
5239 | } |
5240 | case Intrinsic::arithmetic_fence: { |
5241 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
5242 | Known, Depth: Depth + 1, Q); |
5243 | break; |
5244 | } |
5245 | case Intrinsic::experimental_constrained_sitofp: |
5246 | case Intrinsic::experimental_constrained_uitofp: |
5247 | // Cannot produce nan |
5248 | Known.knownNot(RuleOut: fcNan); |
5249 | |
5250 | // sitofp and uitofp turn into +0.0 for zero. |
5251 | Known.knownNot(RuleOut: fcNegZero); |
5252 | |
5253 | // Integers cannot be subnormal |
5254 | Known.knownNot(RuleOut: fcSubnormal); |
5255 | |
5256 | if (IID == Intrinsic::experimental_constrained_uitofp) |
5257 | Known.signBitMustBeZero(); |
5258 | |
5259 | // TODO: Copy inf handling from instructions |
5260 | break; |
5261 | default: |
5262 | break; |
5263 | } |
5264 | |
5265 | break; |
5266 | } |
5267 | case Instruction::FAdd: |
5268 | case Instruction::FSub: { |
5269 | KnownFPClass KnownLHS, KnownRHS; |
5270 | bool WantNegative = |
5271 | Op->getOpcode() == Instruction::FAdd && |
5272 | (InterestedClasses & KnownFPClass::OrderedLessThanZeroMask) != fcNone; |
5273 | bool WantNaN = (InterestedClasses & fcNan) != fcNone; |
5274 | bool WantNegZero = (InterestedClasses & fcNegZero) != fcNone; |
5275 | |
5276 | if (!WantNaN && !WantNegative && !WantNegZero) |
5277 | break; |
5278 | |
5279 | FPClassTest InterestedSrcs = InterestedClasses; |
5280 | if (WantNegative) |
5281 | InterestedSrcs |= KnownFPClass::OrderedLessThanZeroMask; |
5282 | if (InterestedClasses & fcNan) |
5283 | InterestedSrcs |= fcInf; |
5284 | computeKnownFPClass(V: Op->getOperand(i: 1), DemandedElts, InterestedClasses: InterestedSrcs, |
5285 | Known&: KnownRHS, Depth: Depth + 1, Q); |
5286 | |
5287 | if ((WantNaN && KnownRHS.isKnownNeverNaN()) || |
5288 | (WantNegative && KnownRHS.cannotBeOrderedLessThanZero()) || |
5289 | WantNegZero || Opc == Instruction::FSub) { |
5290 | |
5291 | // RHS is canonically cheaper to compute. Skip inspecting the LHS if |
5292 | // there's no point. |
5293 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
5294 | Known&: KnownLHS, Depth: Depth + 1, Q); |
5295 | // Adding positive and negative infinity produces NaN. |
5296 | // TODO: Check sign of infinities. |
5297 | if (KnownLHS.isKnownNeverNaN() && KnownRHS.isKnownNeverNaN() && |
5298 | (KnownLHS.isKnownNeverInfinity() || KnownRHS.isKnownNeverInfinity())) |
5299 | Known.knownNot(RuleOut: fcNan); |
5300 | |
5301 | // FIXME: Context function should always be passed in separately |
5302 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
5303 | |
5304 | if (Op->getOpcode() == Instruction::FAdd) { |
5305 | if (KnownLHS.cannotBeOrderedLessThanZero() && |
5306 | KnownRHS.cannotBeOrderedLessThanZero()) |
5307 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
5308 | if (!F) |
5309 | break; |
5310 | |
5311 | // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. |
5312 | if ((KnownLHS.isKnownNeverLogicalNegZero(F: *F, Ty: Op->getType()) || |
5313 | KnownRHS.isKnownNeverLogicalNegZero(F: *F, Ty: Op->getType())) && |
5314 | // Make sure output negative denormal can't flush to -0 |
5315 | outputDenormalIsIEEEOrPosZero(F: *F, Ty: Op->getType())) |
5316 | Known.knownNot(RuleOut: fcNegZero); |
5317 | } else { |
5318 | if (!F) |
5319 | break; |
5320 | |
5321 | // Only fsub -0, +0 can return -0 |
5322 | if ((KnownLHS.isKnownNeverLogicalNegZero(F: *F, Ty: Op->getType()) || |
5323 | KnownRHS.isKnownNeverLogicalPosZero(F: *F, Ty: Op->getType())) && |
5324 | // Make sure output negative denormal can't flush to -0 |
5325 | outputDenormalIsIEEEOrPosZero(F: *F, Ty: Op->getType())) |
5326 | Known.knownNot(RuleOut: fcNegZero); |
5327 | } |
5328 | } |
5329 | |
5330 | break; |
5331 | } |
5332 | case Instruction::FMul: { |
5333 | // X * X is always non-negative or a NaN. |
5334 | if (Op->getOperand(i: 0) == Op->getOperand(i: 1)) |
5335 | Known.knownNot(RuleOut: fcNegative); |
5336 | |
5337 | if ((InterestedClasses & fcNan) != fcNan) |
5338 | break; |
5339 | |
5340 | // fcSubnormal is only needed in case of DAZ. |
5341 | const FPClassTest NeedForNan = fcNan | fcInf | fcZero | fcSubnormal; |
5342 | |
5343 | KnownFPClass KnownLHS, KnownRHS; |
5344 | computeKnownFPClass(V: Op->getOperand(i: 1), DemandedElts, InterestedClasses: NeedForNan, Known&: KnownRHS, |
5345 | Depth: Depth + 1, Q); |
5346 | if (!KnownRHS.isKnownNeverNaN()) |
5347 | break; |
5348 | |
5349 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses: NeedForNan, Known&: KnownLHS, |
5350 | Depth: Depth + 1, Q); |
5351 | if (!KnownLHS.isKnownNeverNaN()) |
5352 | break; |
5353 | |
5354 | if (KnownLHS.SignBit && KnownRHS.SignBit) { |
5355 | if (*KnownLHS.SignBit == *KnownRHS.SignBit) |
5356 | Known.signBitMustBeZero(); |
5357 | else |
5358 | Known.signBitMustBeOne(); |
5359 | } |
5360 | |
5361 | // If 0 * +/-inf produces NaN. |
5362 | if (KnownLHS.isKnownNeverInfinity() && KnownRHS.isKnownNeverInfinity()) { |
5363 | Known.knownNot(RuleOut: fcNan); |
5364 | break; |
5365 | } |
5366 | |
5367 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
5368 | if (!F) |
5369 | break; |
5370 | |
5371 | if ((KnownRHS.isKnownNeverInfinity() || |
5372 | KnownLHS.isKnownNeverLogicalZero(F: *F, Ty: Op->getType())) && |
5373 | (KnownLHS.isKnownNeverInfinity() || |
5374 | KnownRHS.isKnownNeverLogicalZero(F: *F, Ty: Op->getType()))) |
5375 | Known.knownNot(RuleOut: fcNan); |
5376 | |
5377 | break; |
5378 | } |
5379 | case Instruction::FDiv: |
5380 | case Instruction::FRem: { |
5381 | if (Op->getOperand(i: 0) == Op->getOperand(i: 1)) { |
5382 | // TODO: Could filter out snan if we inspect the operand |
5383 | if (Op->getOpcode() == Instruction::FDiv) { |
5384 | // X / X is always exactly 1.0 or a NaN. |
5385 | Known.KnownFPClasses = fcNan | fcPosNormal; |
5386 | } else { |
5387 | // X % X is always exactly [+-]0.0 or a NaN. |
5388 | Known.KnownFPClasses = fcNan | fcZero; |
5389 | } |
5390 | |
5391 | break; |
5392 | } |
5393 | |
5394 | const bool WantNan = (InterestedClasses & fcNan) != fcNone; |
5395 | const bool WantNegative = (InterestedClasses & fcNegative) != fcNone; |
5396 | const bool WantPositive = |
5397 | Opc == Instruction::FRem && (InterestedClasses & fcPositive) != fcNone; |
5398 | if (!WantNan && !WantNegative && !WantPositive) |
5399 | break; |
5400 | |
5401 | KnownFPClass KnownLHS, KnownRHS; |
5402 | |
5403 | computeKnownFPClass(V: Op->getOperand(i: 1), DemandedElts, |
5404 | InterestedClasses: fcNan | fcInf | fcZero | fcNegative, Known&: KnownRHS, |
5405 | Depth: Depth + 1, Q); |
5406 | |
5407 | bool KnowSomethingUseful = |
5408 | KnownRHS.isKnownNeverNaN() || KnownRHS.isKnownNever(Mask: fcNegative); |
5409 | |
5410 | if (KnowSomethingUseful || WantPositive) { |
5411 | const FPClassTest InterestedLHS = |
5412 | WantPositive ? fcAllFlags |
5413 | : fcNan | fcInf | fcZero | fcSubnormal | fcNegative; |
5414 | |
5415 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, |
5416 | InterestedClasses: InterestedClasses & InterestedLHS, Known&: KnownLHS, |
5417 | Depth: Depth + 1, Q); |
5418 | } |
5419 | |
5420 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
5421 | |
5422 | if (Op->getOpcode() == Instruction::FDiv) { |
5423 | // Only 0/0, Inf/Inf produce NaN. |
5424 | if (KnownLHS.isKnownNeverNaN() && KnownRHS.isKnownNeverNaN() && |
5425 | (KnownLHS.isKnownNeverInfinity() || |
5426 | KnownRHS.isKnownNeverInfinity()) && |
5427 | ((F && KnownLHS.isKnownNeverLogicalZero(F: *F, Ty: Op->getType())) || |
5428 | (F && KnownRHS.isKnownNeverLogicalZero(F: *F, Ty: Op->getType())))) { |
5429 | Known.knownNot(RuleOut: fcNan); |
5430 | } |
5431 | |
5432 | // X / -0.0 is -Inf (or NaN). |
5433 | // +X / +X is +X |
5434 | if (KnownLHS.isKnownNever(Mask: fcNegative) && KnownRHS.isKnownNever(Mask: fcNegative)) |
5435 | Known.knownNot(RuleOut: fcNegative); |
5436 | } else { |
5437 | // Inf REM x and x REM 0 produce NaN. |
5438 | if (KnownLHS.isKnownNeverNaN() && KnownRHS.isKnownNeverNaN() && |
5439 | KnownLHS.isKnownNeverInfinity() && F && |
5440 | KnownRHS.isKnownNeverLogicalZero(F: *F, Ty: Op->getType())) { |
5441 | Known.knownNot(RuleOut: fcNan); |
5442 | } |
5443 | |
5444 | // The sign for frem is the same as the first operand. |
5445 | if (KnownLHS.cannotBeOrderedLessThanZero()) |
5446 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
5447 | if (KnownLHS.cannotBeOrderedGreaterThanZero()) |
5448 | Known.knownNot(RuleOut: KnownFPClass::OrderedGreaterThanZeroMask); |
5449 | |
5450 | // See if we can be more aggressive about the sign of 0. |
5451 | if (KnownLHS.isKnownNever(Mask: fcNegative)) |
5452 | Known.knownNot(RuleOut: fcNegative); |
5453 | if (KnownLHS.isKnownNever(Mask: fcPositive)) |
5454 | Known.knownNot(RuleOut: fcPositive); |
5455 | } |
5456 | |
5457 | break; |
5458 | } |
5459 | case Instruction::FPExt: { |
5460 | // Infinity, nan and zero propagate from source. |
5461 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
5462 | Known, Depth: Depth + 1, Q); |
5463 | |
5464 | const fltSemantics &DstTy = |
5465 | Op->getType()->getScalarType()->getFltSemantics(); |
5466 | const fltSemantics &SrcTy = |
5467 | Op->getOperand(i: 0)->getType()->getScalarType()->getFltSemantics(); |
5468 | |
5469 | // All subnormal inputs should be in the normal range in the result type. |
5470 | if (APFloat::isRepresentableAsNormalIn(Src: SrcTy, Dst: DstTy)) { |
5471 | if (Known.KnownFPClasses & fcPosSubnormal) |
5472 | Known.KnownFPClasses |= fcPosNormal; |
5473 | if (Known.KnownFPClasses & fcNegSubnormal) |
5474 | Known.KnownFPClasses |= fcNegNormal; |
5475 | Known.knownNot(RuleOut: fcSubnormal); |
5476 | } |
5477 | |
5478 | // Sign bit of a nan isn't guaranteed. |
5479 | if (!Known.isKnownNeverNaN()) |
5480 | Known.SignBit = std::nullopt; |
5481 | break; |
5482 | } |
5483 | case Instruction::FPTrunc: { |
5484 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, Known, |
5485 | Depth, Q); |
5486 | break; |
5487 | } |
5488 | case Instruction::SIToFP: |
5489 | case Instruction::UIToFP: { |
5490 | // Cannot produce nan |
5491 | Known.knownNot(RuleOut: fcNan); |
5492 | |
5493 | // Integers cannot be subnormal |
5494 | Known.knownNot(RuleOut: fcSubnormal); |
5495 | |
5496 | // sitofp and uitofp turn into +0.0 for zero. |
5497 | Known.knownNot(RuleOut: fcNegZero); |
5498 | if (Op->getOpcode() == Instruction::UIToFP) |
5499 | Known.signBitMustBeZero(); |
5500 | |
5501 | if (InterestedClasses & fcInf) { |
5502 | // Get width of largest magnitude integer (remove a bit if signed). |
5503 | // This still works for a signed minimum value because the largest FP |
5504 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). |
5505 | int IntSize = Op->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
5506 | if (Op->getOpcode() == Instruction::SIToFP) |
5507 | --IntSize; |
5508 | |
5509 | // If the exponent of the largest finite FP value can hold the largest |
5510 | // integer, the result of the cast must be finite. |
5511 | Type *FPTy = Op->getType()->getScalarType(); |
5512 | if (ilogb(Arg: APFloat::getLargest(Sem: FPTy->getFltSemantics())) >= IntSize) |
5513 | Known.knownNot(RuleOut: fcInf); |
5514 | } |
5515 | |
5516 | break; |
5517 | } |
5518 | case Instruction::ExtractElement: { |
5519 | // Look through extract element. If the index is non-constant or |
5520 | // out-of-range demand all elements, otherwise just the extracted element. |
5521 | const Value *Vec = Op->getOperand(i: 0); |
5522 | const Value *Idx = Op->getOperand(i: 1); |
5523 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
5524 | |
5525 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType())) { |
5526 | unsigned NumElts = VecTy->getNumElements(); |
5527 | APInt DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
5528 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
5529 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
5530 | return computeKnownFPClass(V: Vec, DemandedElts: DemandedVecElts, InterestedClasses, Known, |
5531 | Depth: Depth + 1, Q); |
5532 | } |
5533 | |
5534 | break; |
5535 | } |
5536 | case Instruction::InsertElement: { |
5537 | if (isa<ScalableVectorType>(Val: Op->getType())) |
5538 | return; |
5539 | |
5540 | const Value *Vec = Op->getOperand(i: 0); |
5541 | const Value *Elt = Op->getOperand(i: 1); |
5542 | auto *CIdx = dyn_cast<ConstantInt>(Val: Op->getOperand(i: 2)); |
5543 | unsigned NumElts = DemandedElts.getBitWidth(); |
5544 | APInt DemandedVecElts = DemandedElts; |
5545 | bool NeedsElt = true; |
5546 | // If we know the index we are inserting to, clear it from Vec check. |
5547 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
5548 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
5549 | NeedsElt = DemandedElts[CIdx->getZExtValue()]; |
5550 | } |
5551 | |
5552 | // Do we demand the inserted element? |
5553 | if (NeedsElt) { |
5554 | computeKnownFPClass(V: Elt, Known, InterestedClasses, Depth: Depth + 1, Q); |
5555 | // If we don't know any bits, early out. |
5556 | if (Known.isUnknown()) |
5557 | break; |
5558 | } else { |
5559 | Known.KnownFPClasses = fcNone; |
5560 | } |
5561 | |
5562 | // Do we need anymore elements from Vec? |
5563 | if (!DemandedVecElts.isZero()) { |
5564 | KnownFPClass Known2; |
5565 | computeKnownFPClass(V: Vec, DemandedElts: DemandedVecElts, InterestedClasses, Known&: Known2, |
5566 | Depth: Depth + 1, Q); |
5567 | Known |= Known2; |
5568 | } |
5569 | |
5570 | break; |
5571 | } |
5572 | case Instruction::ShuffleVector: { |
5573 | // For undef elements, we don't know anything about the common state of |
5574 | // the shuffle result. |
5575 | APInt DemandedLHS, DemandedRHS; |
5576 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: Op); |
5577 | if (!Shuf || !getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
5578 | return; |
5579 | |
5580 | if (!!DemandedLHS) { |
5581 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
5582 | computeKnownFPClass(V: LHS, DemandedElts: DemandedLHS, InterestedClasses, Known, |
5583 | Depth: Depth + 1, Q); |
5584 | |
5585 | // If we don't know any bits, early out. |
5586 | if (Known.isUnknown()) |
5587 | break; |
5588 | } else { |
5589 | Known.KnownFPClasses = fcNone; |
5590 | } |
5591 | |
5592 | if (!!DemandedRHS) { |
5593 | KnownFPClass Known2; |
5594 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
5595 | computeKnownFPClass(V: RHS, DemandedElts: DemandedRHS, InterestedClasses, Known&: Known2, |
5596 | Depth: Depth + 1, Q); |
5597 | Known |= Known2; |
5598 | } |
5599 | |
5600 | break; |
5601 | } |
5602 | case Instruction::ExtractValue: { |
5603 | const ExtractValueInst * = cast<ExtractValueInst>(Val: Op); |
5604 | ArrayRef<unsigned> Indices = Extract->getIndices(); |
5605 | const Value *Src = Extract->getAggregateOperand(); |
5606 | if (isa<StructType>(Val: Src->getType()) && Indices.size() == 1 && |
5607 | Indices[0] == 0) { |
5608 | if (const auto *II = dyn_cast<IntrinsicInst>(Val: Src)) { |
5609 | switch (II->getIntrinsicID()) { |
5610 | case Intrinsic::frexp: { |
5611 | Known.knownNot(RuleOut: fcSubnormal); |
5612 | |
5613 | KnownFPClass KnownSrc; |
5614 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
5615 | InterestedClasses, Known&: KnownSrc, Depth: Depth + 1, Q); |
5616 | |
5617 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
5618 | |
5619 | if (KnownSrc.isKnownNever(Mask: fcNegative)) |
5620 | Known.knownNot(RuleOut: fcNegative); |
5621 | else { |
5622 | if (F && KnownSrc.isKnownNeverLogicalNegZero(F: *F, Ty: Op->getType())) |
5623 | Known.knownNot(RuleOut: fcNegZero); |
5624 | if (KnownSrc.isKnownNever(Mask: fcNegInf)) |
5625 | Known.knownNot(RuleOut: fcNegInf); |
5626 | } |
5627 | |
5628 | if (KnownSrc.isKnownNever(Mask: fcPositive)) |
5629 | Known.knownNot(RuleOut: fcPositive); |
5630 | else { |
5631 | if (F && KnownSrc.isKnownNeverLogicalPosZero(F: *F, Ty: Op->getType())) |
5632 | Known.knownNot(RuleOut: fcPosZero); |
5633 | if (KnownSrc.isKnownNever(Mask: fcPosInf)) |
5634 | Known.knownNot(RuleOut: fcPosInf); |
5635 | } |
5636 | |
5637 | Known.propagateNaN(Src: KnownSrc); |
5638 | return; |
5639 | } |
5640 | default: |
5641 | break; |
5642 | } |
5643 | } |
5644 | } |
5645 | |
5646 | computeKnownFPClass(V: Src, DemandedElts, InterestedClasses, Known, Depth: Depth + 1, |
5647 | Q); |
5648 | break; |
5649 | } |
5650 | case Instruction::PHI: { |
5651 | const PHINode *P = cast<PHINode>(Val: Op); |
5652 | // Unreachable blocks may have zero-operand PHI nodes. |
5653 | if (P->getNumIncomingValues() == 0) |
5654 | break; |
5655 | |
5656 | // Otherwise take the unions of the known bit sets of the operands, |
5657 | // taking conservative care to avoid excessive recursion. |
5658 | const unsigned PhiRecursionLimit = MaxAnalysisRecursionDepth - 2; |
5659 | |
5660 | if (Depth < PhiRecursionLimit) { |
5661 | // Skip if every incoming value references to ourself. |
5662 | if (isa_and_nonnull<UndefValue>(Val: P->hasConstantValue())) |
5663 | break; |
5664 | |
5665 | bool First = true; |
5666 | |
5667 | for (const Use &U : P->operands()) { |
5668 | Value *IncValue = U.get(); |
5669 | // Skip direct self references. |
5670 | if (IncValue == P) |
5671 | continue; |
5672 | |
5673 | KnownFPClass KnownSrc; |
5674 | // Recurse, but cap the recursion to two levels, because we don't want |
5675 | // to waste time spinning around in loops. We need at least depth 2 to |
5676 | // detect known sign bits. |
5677 | computeKnownFPClass( |
5678 | V: IncValue, DemandedElts, InterestedClasses, Known&: KnownSrc, |
5679 | Depth: PhiRecursionLimit, |
5680 | Q: Q.getWithInstruction(I: P->getIncomingBlock(U)->getTerminator())); |
5681 | |
5682 | if (First) { |
5683 | Known = KnownSrc; |
5684 | First = false; |
5685 | } else { |
5686 | Known |= KnownSrc; |
5687 | } |
5688 | |
5689 | if (Known.KnownFPClasses == fcAllFlags) |
5690 | break; |
5691 | } |
5692 | } |
5693 | |
5694 | break; |
5695 | } |
5696 | default: |
5697 | break; |
5698 | } |
5699 | } |
5700 | |
5701 | KnownFPClass llvm::computeKnownFPClass(const Value *V, |
5702 | const APInt &DemandedElts, |
5703 | FPClassTest InterestedClasses, |
5704 | unsigned Depth, |
5705 | const SimplifyQuery &SQ) { |
5706 | KnownFPClass KnownClasses; |
5707 | ::computeKnownFPClass(V, DemandedElts, InterestedClasses, Known&: KnownClasses, Depth, |
5708 | Q: SQ); |
5709 | return KnownClasses; |
5710 | } |
5711 | |
5712 | KnownFPClass llvm::computeKnownFPClass(const Value *V, |
5713 | FPClassTest InterestedClasses, |
5714 | unsigned Depth, |
5715 | const SimplifyQuery &SQ) { |
5716 | KnownFPClass Known; |
5717 | ::computeKnownFPClass(V, Known, InterestedClasses, Depth, Q: SQ); |
5718 | return Known; |
5719 | } |
5720 | |
5721 | Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { |
5722 | |
5723 | // All byte-wide stores are splatable, even of arbitrary variables. |
5724 | if (V->getType()->isIntegerTy(Bitwidth: 8)) |
5725 | return V; |
5726 | |
5727 | LLVMContext &Ctx = V->getContext(); |
5728 | |
5729 | // Undef don't care. |
5730 | auto *UndefInt8 = UndefValue::get(T: Type::getInt8Ty(C&: Ctx)); |
5731 | if (isa<UndefValue>(Val: V)) |
5732 | return UndefInt8; |
5733 | |
5734 | // Return Undef for zero-sized type. |
5735 | if (DL.getTypeStoreSize(Ty: V->getType()).isZero()) |
5736 | return UndefInt8; |
5737 | |
5738 | Constant *C = dyn_cast<Constant>(Val: V); |
5739 | if (!C) { |
5740 | // Conceptually, we could handle things like: |
5741 | // %a = zext i8 %X to i16 |
5742 | // %b = shl i16 %a, 8 |
5743 | // %c = or i16 %a, %b |
5744 | // but until there is an example that actually needs this, it doesn't seem |
5745 | // worth worrying about. |
5746 | return nullptr; |
5747 | } |
5748 | |
5749 | // Handle 'null' ConstantArrayZero etc. |
5750 | if (C->isNullValue()) |
5751 | return Constant::getNullValue(Ty: Type::getInt8Ty(C&: Ctx)); |
5752 | |
5753 | // Constant floating-point values can be handled as integer values if the |
5754 | // corresponding integer value is "byteable". An important case is 0.0. |
5755 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) { |
5756 | Type *Ty = nullptr; |
5757 | if (CFP->getType()->isHalfTy()) |
5758 | Ty = Type::getInt16Ty(C&: Ctx); |
5759 | else if (CFP->getType()->isFloatTy()) |
5760 | Ty = Type::getInt32Ty(C&: Ctx); |
5761 | else if (CFP->getType()->isDoubleTy()) |
5762 | Ty = Type::getInt64Ty(C&: Ctx); |
5763 | // Don't handle long double formats, which have strange constraints. |
5764 | return Ty ? isBytewiseValue(V: ConstantExpr::getBitCast(C: CFP, Ty), DL) |
5765 | : nullptr; |
5766 | } |
5767 | |
5768 | // We can handle constant integers that are multiple of 8 bits. |
5769 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: C)) { |
5770 | if (CI->getBitWidth() % 8 == 0) { |
5771 | assert(CI->getBitWidth() > 8 && "8 bits should be handled above!" ); |
5772 | if (!CI->getValue().isSplat(SplatSizeInBits: 8)) |
5773 | return nullptr; |
5774 | return ConstantInt::get(Context&: Ctx, V: CI->getValue().trunc(width: 8)); |
5775 | } |
5776 | } |
5777 | |
5778 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
5779 | if (CE->getOpcode() == Instruction::IntToPtr) { |
5780 | if (auto *PtrTy = dyn_cast<PointerType>(Val: CE->getType())) { |
5781 | unsigned BitWidth = DL.getPointerSizeInBits(AS: PtrTy->getAddressSpace()); |
5782 | if (Constant *Op = ConstantFoldIntegerCast( |
5783 | C: CE->getOperand(i_nocapture: 0), DestTy: Type::getIntNTy(C&: Ctx, N: BitWidth), IsSigned: false, DL)) |
5784 | return isBytewiseValue(V: Op, DL); |
5785 | } |
5786 | } |
5787 | } |
5788 | |
5789 | auto Merge = [&](Value *LHS, Value *RHS) -> Value * { |
5790 | if (LHS == RHS) |
5791 | return LHS; |
5792 | if (!LHS || !RHS) |
5793 | return nullptr; |
5794 | if (LHS == UndefInt8) |
5795 | return RHS; |
5796 | if (RHS == UndefInt8) |
5797 | return LHS; |
5798 | return nullptr; |
5799 | }; |
5800 | |
5801 | if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(Val: C)) { |
5802 | Value *Val = UndefInt8; |
5803 | for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) |
5804 | if (!(Val = Merge(Val, isBytewiseValue(V: CA->getElementAsConstant(i: I), DL)))) |
5805 | return nullptr; |
5806 | return Val; |
5807 | } |
5808 | |
5809 | if (isa<ConstantAggregate>(Val: C)) { |
5810 | Value *Val = UndefInt8; |
5811 | for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) |
5812 | if (!(Val = Merge(Val, isBytewiseValue(V: C->getOperand(i: I), DL)))) |
5813 | return nullptr; |
5814 | return Val; |
5815 | } |
5816 | |
5817 | // Don't try to handle the handful of other constants. |
5818 | return nullptr; |
5819 | } |
5820 | |
5821 | // This is the recursive version of BuildSubAggregate. It takes a few different |
5822 | // arguments. Idxs is the index within the nested struct From that we are |
5823 | // looking at now (which is of type IndexedType). IdxSkip is the number of |
5824 | // indices from Idxs that should be left out when inserting into the resulting |
5825 | // struct. To is the result struct built so far, new insertvalue instructions |
5826 | // build on that. |
5827 | static Value *BuildSubAggregate(Value *From, Value *To, Type *IndexedType, |
5828 | SmallVectorImpl<unsigned> &Idxs, |
5829 | unsigned IdxSkip, |
5830 | BasicBlock::iterator InsertBefore) { |
5831 | StructType *STy = dyn_cast<StructType>(Val: IndexedType); |
5832 | if (STy) { |
5833 | // Save the original To argument so we can modify it |
5834 | Value *OrigTo = To; |
5835 | // General case, the type indexed by Idxs is a struct |
5836 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
5837 | // Process each struct element recursively |
5838 | Idxs.push_back(Elt: i); |
5839 | Value *PrevTo = To; |
5840 | To = BuildSubAggregate(From, To, IndexedType: STy->getElementType(N: i), Idxs, IdxSkip, |
5841 | InsertBefore); |
5842 | Idxs.pop_back(); |
5843 | if (!To) { |
5844 | // Couldn't find any inserted value for this index? Cleanup |
5845 | while (PrevTo != OrigTo) { |
5846 | InsertValueInst* Del = cast<InsertValueInst>(Val: PrevTo); |
5847 | PrevTo = Del->getAggregateOperand(); |
5848 | Del->eraseFromParent(); |
5849 | } |
5850 | // Stop processing elements |
5851 | break; |
5852 | } |
5853 | } |
5854 | // If we successfully found a value for each of our subaggregates |
5855 | if (To) |
5856 | return To; |
5857 | } |
5858 | // Base case, the type indexed by SourceIdxs is not a struct, or not all of |
5859 | // the struct's elements had a value that was inserted directly. In the latter |
5860 | // case, perhaps we can't determine each of the subelements individually, but |
5861 | // we might be able to find the complete struct somewhere. |
5862 | |
5863 | // Find the value that is at that particular spot |
5864 | Value *V = FindInsertedValue(V: From, idx_range: Idxs); |
5865 | |
5866 | if (!V) |
5867 | return nullptr; |
5868 | |
5869 | // Insert the value in the new (sub) aggregate |
5870 | return InsertValueInst::Create(Agg: To, Val: V, Idxs: ArrayRef(Idxs).slice(N: IdxSkip), NameStr: "tmp" , |
5871 | InsertBefore); |
5872 | } |
5873 | |
5874 | // This helper takes a nested struct and extracts a part of it (which is again a |
5875 | // struct) into a new value. For example, given the struct: |
5876 | // { a, { b, { c, d }, e } } |
5877 | // and the indices "1, 1" this returns |
5878 | // { c, d }. |
5879 | // |
5880 | // It does this by inserting an insertvalue for each element in the resulting |
5881 | // struct, as opposed to just inserting a single struct. This will only work if |
5882 | // each of the elements of the substruct are known (ie, inserted into From by an |
5883 | // insertvalue instruction somewhere). |
5884 | // |
5885 | // All inserted insertvalue instructions are inserted before InsertBefore |
5886 | static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, |
5887 | BasicBlock::iterator InsertBefore) { |
5888 | Type *IndexedType = ExtractValueInst::getIndexedType(Agg: From->getType(), |
5889 | Idxs: idx_range); |
5890 | Value *To = PoisonValue::get(T: IndexedType); |
5891 | SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); |
5892 | unsigned IdxSkip = Idxs.size(); |
5893 | |
5894 | return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); |
5895 | } |
5896 | |
5897 | /// Given an aggregate and a sequence of indices, see if the scalar value |
5898 | /// indexed is already around as a register, for example if it was inserted |
5899 | /// directly into the aggregate. |
5900 | /// |
5901 | /// If InsertBefore is not null, this function will duplicate (modified) |
5902 | /// insertvalues when a part of a nested struct is extracted. |
5903 | Value * |
5904 | llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, |
5905 | std::optional<BasicBlock::iterator> InsertBefore) { |
5906 | // Nothing to index? Just return V then (this is useful at the end of our |
5907 | // recursion). |
5908 | if (idx_range.empty()) |
5909 | return V; |
5910 | // We have indices, so V should have an indexable type. |
5911 | assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && |
5912 | "Not looking at a struct or array?" ); |
5913 | assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && |
5914 | "Invalid indices for type?" ); |
5915 | |
5916 | if (Constant *C = dyn_cast<Constant>(Val: V)) { |
5917 | C = C->getAggregateElement(Elt: idx_range[0]); |
5918 | if (!C) return nullptr; |
5919 | return FindInsertedValue(V: C, idx_range: idx_range.slice(N: 1), InsertBefore); |
5920 | } |
5921 | |
5922 | if (InsertValueInst *I = dyn_cast<InsertValueInst>(Val: V)) { |
5923 | // Loop the indices for the insertvalue instruction in parallel with the |
5924 | // requested indices |
5925 | const unsigned *req_idx = idx_range.begin(); |
5926 | for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); |
5927 | i != e; ++i, ++req_idx) { |
5928 | if (req_idx == idx_range.end()) { |
5929 | // We can't handle this without inserting insertvalues |
5930 | if (!InsertBefore) |
5931 | return nullptr; |
5932 | |
5933 | // The requested index identifies a part of a nested aggregate. Handle |
5934 | // this specially. For example, |
5935 | // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 |
5936 | // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 |
5937 | // %C = extractvalue {i32, { i32, i32 } } %B, 1 |
5938 | // This can be changed into |
5939 | // %A = insertvalue {i32, i32 } undef, i32 10, 0 |
5940 | // %C = insertvalue {i32, i32 } %A, i32 11, 1 |
5941 | // which allows the unused 0,0 element from the nested struct to be |
5942 | // removed. |
5943 | return BuildSubAggregate(From: V, idx_range: ArrayRef(idx_range.begin(), req_idx), |
5944 | InsertBefore: *InsertBefore); |
5945 | } |
5946 | |
5947 | // This insert value inserts something else than what we are looking for. |
5948 | // See if the (aggregate) value inserted into has the value we are |
5949 | // looking for, then. |
5950 | if (*req_idx != *i) |
5951 | return FindInsertedValue(V: I->getAggregateOperand(), idx_range, |
5952 | InsertBefore); |
5953 | } |
5954 | // If we end up here, the indices of the insertvalue match with those |
5955 | // requested (though possibly only partially). Now we recursively look at |
5956 | // the inserted value, passing any remaining indices. |
5957 | return FindInsertedValue(V: I->getInsertedValueOperand(), |
5958 | idx_range: ArrayRef(req_idx, idx_range.end()), InsertBefore); |
5959 | } |
5960 | |
5961 | if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(Val: V)) { |
5962 | // If we're extracting a value from an aggregate that was extracted from |
5963 | // something else, we can extract from that something else directly instead. |
5964 | // However, we will need to chain I's indices with the requested indices. |
5965 | |
5966 | // Calculate the number of indices required |
5967 | unsigned size = I->getNumIndices() + idx_range.size(); |
5968 | // Allocate some space to put the new indices in |
5969 | SmallVector<unsigned, 5> Idxs; |
5970 | Idxs.reserve(N: size); |
5971 | // Add indices from the extract value instruction |
5972 | Idxs.append(in_start: I->idx_begin(), in_end: I->idx_end()); |
5973 | |
5974 | // Add requested indices |
5975 | Idxs.append(in_start: idx_range.begin(), in_end: idx_range.end()); |
5976 | |
5977 | assert(Idxs.size() == size |
5978 | && "Number of indices added not correct?" ); |
5979 | |
5980 | return FindInsertedValue(V: I->getAggregateOperand(), idx_range: Idxs, InsertBefore); |
5981 | } |
5982 | // Otherwise, we don't know (such as, extracting from a function return value |
5983 | // or load instruction) |
5984 | return nullptr; |
5985 | } |
5986 | |
5987 | bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, |
5988 | unsigned CharSize) { |
5989 | // Make sure the GEP has exactly three arguments. |
5990 | if (GEP->getNumOperands() != 3) |
5991 | return false; |
5992 | |
5993 | // Make sure the index-ee is a pointer to array of \p CharSize integers. |
5994 | // CharSize. |
5995 | ArrayType *AT = dyn_cast<ArrayType>(Val: GEP->getSourceElementType()); |
5996 | if (!AT || !AT->getElementType()->isIntegerTy(Bitwidth: CharSize)) |
5997 | return false; |
5998 | |
5999 | // Check to make sure that the first operand of the GEP is an integer and |
6000 | // has value 0 so that we are sure we're indexing into the initializer. |
6001 | const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1)); |
6002 | if (!FirstIdx || !FirstIdx->isZero()) |
6003 | return false; |
6004 | |
6005 | return true; |
6006 | } |
6007 | |
6008 | // If V refers to an initialized global constant, set Slice either to |
6009 | // its initializer if the size of its elements equals ElementSize, or, |
6010 | // for ElementSize == 8, to its representation as an array of unsiged |
6011 | // char. Return true on success. |
6012 | // Offset is in the unit "nr of ElementSize sized elements". |
6013 | bool llvm::getConstantDataArrayInfo(const Value *V, |
6014 | ConstantDataArraySlice &Slice, |
6015 | unsigned ElementSize, uint64_t Offset) { |
6016 | assert(V && "V should not be null." ); |
6017 | assert((ElementSize % 8) == 0 && |
6018 | "ElementSize expected to be a multiple of the size of a byte." ); |
6019 | unsigned ElementSizeInBytes = ElementSize / 8; |
6020 | |
6021 | // Drill down into the pointer expression V, ignoring any intervening |
6022 | // casts, and determine the identity of the object it references along |
6023 | // with the cumulative byte offset into it. |
6024 | const GlobalVariable *GV = |
6025 | dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V)); |
6026 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) |
6027 | // Fail if V is not based on constant global object. |
6028 | return false; |
6029 | |
6030 | const DataLayout &DL = GV->getParent()->getDataLayout(); |
6031 | APInt Off(DL.getIndexTypeSizeInBits(Ty: V->getType()), 0); |
6032 | |
6033 | if (GV != V->stripAndAccumulateConstantOffsets(DL, Offset&: Off, |
6034 | /*AllowNonInbounds*/ true)) |
6035 | // Fail if a constant offset could not be determined. |
6036 | return false; |
6037 | |
6038 | uint64_t StartIdx = Off.getLimitedValue(); |
6039 | if (StartIdx == UINT64_MAX) |
6040 | // Fail if the constant offset is excessive. |
6041 | return false; |
6042 | |
6043 | // Off/StartIdx is in the unit of bytes. So we need to convert to number of |
6044 | // elements. Simply bail out if that isn't possible. |
6045 | if ((StartIdx % ElementSizeInBytes) != 0) |
6046 | return false; |
6047 | |
6048 | Offset += StartIdx / ElementSizeInBytes; |
6049 | ConstantDataArray *Array = nullptr; |
6050 | ArrayType *ArrayTy = nullptr; |
6051 | |
6052 | if (GV->getInitializer()->isNullValue()) { |
6053 | Type *GVTy = GV->getValueType(); |
6054 | uint64_t SizeInBytes = DL.getTypeStoreSize(Ty: GVTy).getFixedValue(); |
6055 | uint64_t Length = SizeInBytes / ElementSizeInBytes; |
6056 | |
6057 | Slice.Array = nullptr; |
6058 | Slice.Offset = 0; |
6059 | // Return an empty Slice for undersized constants to let callers |
6060 | // transform even undefined library calls into simpler, well-defined |
6061 | // expressions. This is preferable to making the calls although it |
6062 | // prevents sanitizers from detecting such calls. |
6063 | Slice.Length = Length < Offset ? 0 : Length - Offset; |
6064 | return true; |
6065 | } |
6066 | |
6067 | auto *Init = const_cast<Constant *>(GV->getInitializer()); |
6068 | if (auto *ArrayInit = dyn_cast<ConstantDataArray>(Val: Init)) { |
6069 | Type *InitElTy = ArrayInit->getElementType(); |
6070 | if (InitElTy->isIntegerTy(Bitwidth: ElementSize)) { |
6071 | // If Init is an initializer for an array of the expected type |
6072 | // and size, use it as is. |
6073 | Array = ArrayInit; |
6074 | ArrayTy = ArrayInit->getType(); |
6075 | } |
6076 | } |
6077 | |
6078 | if (!Array) { |
6079 | if (ElementSize != 8) |
6080 | // TODO: Handle conversions to larger integral types. |
6081 | return false; |
6082 | |
6083 | // Otherwise extract the portion of the initializer starting |
6084 | // at Offset as an array of bytes, and reset Offset. |
6085 | Init = ReadByteArrayFromGlobal(GV, Offset); |
6086 | if (!Init) |
6087 | return false; |
6088 | |
6089 | Offset = 0; |
6090 | Array = dyn_cast<ConstantDataArray>(Val: Init); |
6091 | ArrayTy = dyn_cast<ArrayType>(Val: Init->getType()); |
6092 | } |
6093 | |
6094 | uint64_t NumElts = ArrayTy->getArrayNumElements(); |
6095 | if (Offset > NumElts) |
6096 | return false; |
6097 | |
6098 | Slice.Array = Array; |
6099 | Slice.Offset = Offset; |
6100 | Slice.Length = NumElts - Offset; |
6101 | return true; |
6102 | } |
6103 | |
6104 | /// Extract bytes from the initializer of the constant array V, which need |
6105 | /// not be a nul-terminated string. On success, store the bytes in Str and |
6106 | /// return true. When TrimAtNul is set, Str will contain only the bytes up |
6107 | /// to but not including the first nul. Return false on failure. |
6108 | bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, |
6109 | bool TrimAtNul) { |
6110 | ConstantDataArraySlice Slice; |
6111 | if (!getConstantDataArrayInfo(V, Slice, ElementSize: 8)) |
6112 | return false; |
6113 | |
6114 | if (Slice.Array == nullptr) { |
6115 | if (TrimAtNul) { |
6116 | // Return a nul-terminated string even for an empty Slice. This is |
6117 | // safe because all existing SimplifyLibcalls callers require string |
6118 | // arguments and the behavior of the functions they fold is undefined |
6119 | // otherwise. Folding the calls this way is preferable to making |
6120 | // the undefined library calls, even though it prevents sanitizers |
6121 | // from reporting such calls. |
6122 | Str = StringRef(); |
6123 | return true; |
6124 | } |
6125 | if (Slice.Length == 1) { |
6126 | Str = StringRef("" , 1); |
6127 | return true; |
6128 | } |
6129 | // We cannot instantiate a StringRef as we do not have an appropriate string |
6130 | // of 0s at hand. |
6131 | return false; |
6132 | } |
6133 | |
6134 | // Start out with the entire array in the StringRef. |
6135 | Str = Slice.Array->getAsString(); |
6136 | // Skip over 'offset' bytes. |
6137 | Str = Str.substr(Start: Slice.Offset); |
6138 | |
6139 | if (TrimAtNul) { |
6140 | // Trim off the \0 and anything after it. If the array is not nul |
6141 | // terminated, we just return the whole end of string. The client may know |
6142 | // some other way that the string is length-bound. |
6143 | Str = Str.substr(Start: 0, N: Str.find(C: '\0')); |
6144 | } |
6145 | return true; |
6146 | } |
6147 | |
6148 | // These next two are very similar to the above, but also look through PHI |
6149 | // nodes. |
6150 | // TODO: See if we can integrate these two together. |
6151 | |
6152 | /// If we can compute the length of the string pointed to by |
6153 | /// the specified pointer, return 'len+1'. If we can't, return 0. |
6154 | static uint64_t GetStringLengthH(const Value *V, |
6155 | SmallPtrSetImpl<const PHINode*> &PHIs, |
6156 | unsigned CharSize) { |
6157 | // Look through noop bitcast instructions. |
6158 | V = V->stripPointerCasts(); |
6159 | |
6160 | // If this is a PHI node, there are two cases: either we have already seen it |
6161 | // or we haven't. |
6162 | if (const PHINode *PN = dyn_cast<PHINode>(Val: V)) { |
6163 | if (!PHIs.insert(Ptr: PN).second) |
6164 | return ~0ULL; // already in the set. |
6165 | |
6166 | // If it was new, see if all the input strings are the same length. |
6167 | uint64_t LenSoFar = ~0ULL; |
6168 | for (Value *IncValue : PN->incoming_values()) { |
6169 | uint64_t Len = GetStringLengthH(V: IncValue, PHIs, CharSize); |
6170 | if (Len == 0) return 0; // Unknown length -> unknown. |
6171 | |
6172 | if (Len == ~0ULL) continue; |
6173 | |
6174 | if (Len != LenSoFar && LenSoFar != ~0ULL) |
6175 | return 0; // Disagree -> unknown. |
6176 | LenSoFar = Len; |
6177 | } |
6178 | |
6179 | // Success, all agree. |
6180 | return LenSoFar; |
6181 | } |
6182 | |
6183 | // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) |
6184 | if (const SelectInst *SI = dyn_cast<SelectInst>(Val: V)) { |
6185 | uint64_t Len1 = GetStringLengthH(V: SI->getTrueValue(), PHIs, CharSize); |
6186 | if (Len1 == 0) return 0; |
6187 | uint64_t Len2 = GetStringLengthH(V: SI->getFalseValue(), PHIs, CharSize); |
6188 | if (Len2 == 0) return 0; |
6189 | if (Len1 == ~0ULL) return Len2; |
6190 | if (Len2 == ~0ULL) return Len1; |
6191 | if (Len1 != Len2) return 0; |
6192 | return Len1; |
6193 | } |
6194 | |
6195 | // Otherwise, see if we can read the string. |
6196 | ConstantDataArraySlice Slice; |
6197 | if (!getConstantDataArrayInfo(V, Slice, ElementSize: CharSize)) |
6198 | return 0; |
6199 | |
6200 | if (Slice.Array == nullptr) |
6201 | // Zeroinitializer (including an empty one). |
6202 | return 1; |
6203 | |
6204 | // Search for the first nul character. Return a conservative result even |
6205 | // when there is no nul. This is safe since otherwise the string function |
6206 | // being folded such as strlen is undefined, and can be preferable to |
6207 | // making the undefined library call. |
6208 | unsigned NullIndex = 0; |
6209 | for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { |
6210 | if (Slice.Array->getElementAsInteger(i: Slice.Offset + NullIndex) == 0) |
6211 | break; |
6212 | } |
6213 | |
6214 | return NullIndex + 1; |
6215 | } |
6216 | |
6217 | /// If we can compute the length of the string pointed to by |
6218 | /// the specified pointer, return 'len+1'. If we can't, return 0. |
6219 | uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { |
6220 | if (!V->getType()->isPointerTy()) |
6221 | return 0; |
6222 | |
6223 | SmallPtrSet<const PHINode*, 32> PHIs; |
6224 | uint64_t Len = GetStringLengthH(V, PHIs, CharSize); |
6225 | // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return |
6226 | // an empty string as a length. |
6227 | return Len == ~0ULL ? 1 : Len; |
6228 | } |
6229 | |
6230 | const Value * |
6231 | llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, |
6232 | bool MustPreserveNullness) { |
6233 | assert(Call && |
6234 | "getArgumentAliasingToReturnedPointer only works on nonnull calls" ); |
6235 | if (const Value *RV = Call->getReturnedArgOperand()) |
6236 | return RV; |
6237 | // This can be used only as a aliasing property. |
6238 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
6239 | Call, MustPreserveNullness)) |
6240 | return Call->getArgOperand(i: 0); |
6241 | return nullptr; |
6242 | } |
6243 | |
6244 | bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
6245 | const CallBase *Call, bool MustPreserveNullness) { |
6246 | switch (Call->getIntrinsicID()) { |
6247 | case Intrinsic::launder_invariant_group: |
6248 | case Intrinsic::strip_invariant_group: |
6249 | case Intrinsic::aarch64_irg: |
6250 | case Intrinsic::aarch64_tagp: |
6251 | // The amdgcn_make_buffer_rsrc function does not alter the address of the |
6252 | // input pointer (and thus preserve null-ness for the purposes of escape |
6253 | // analysis, which is where the MustPreserveNullness flag comes in to play). |
6254 | // However, it will not necessarily map ptr addrspace(N) null to ptr |
6255 | // addrspace(8) null, aka the "null descriptor", which has "all loads return |
6256 | // 0, all stores are dropped" semantics. Given the context of this intrinsic |
6257 | // list, no one should be relying on such a strict interpretation of |
6258 | // MustPreserveNullness (and, at time of writing, they are not), but we |
6259 | // document this fact out of an abundance of caution. |
6260 | case Intrinsic::amdgcn_make_buffer_rsrc: |
6261 | return true; |
6262 | case Intrinsic::ptrmask: |
6263 | return !MustPreserveNullness; |
6264 | case Intrinsic::threadlocal_address: |
6265 | // The underlying variable changes with thread ID. The Thread ID may change |
6266 | // at coroutine suspend points. |
6267 | return !Call->getParent()->getParent()->isPresplitCoroutine(); |
6268 | default: |
6269 | return false; |
6270 | } |
6271 | } |
6272 | |
6273 | /// \p PN defines a loop-variant pointer to an object. Check if the |
6274 | /// previous iteration of the loop was referring to the same object as \p PN. |
6275 | static bool isSameUnderlyingObjectInLoop(const PHINode *PN, |
6276 | const LoopInfo *LI) { |
6277 | // Find the loop-defined value. |
6278 | Loop *L = LI->getLoopFor(BB: PN->getParent()); |
6279 | if (PN->getNumIncomingValues() != 2) |
6280 | return true; |
6281 | |
6282 | // Find the value from previous iteration. |
6283 | auto *PrevValue = dyn_cast<Instruction>(Val: PN->getIncomingValue(i: 0)); |
6284 | if (!PrevValue || LI->getLoopFor(BB: PrevValue->getParent()) != L) |
6285 | PrevValue = dyn_cast<Instruction>(Val: PN->getIncomingValue(i: 1)); |
6286 | if (!PrevValue || LI->getLoopFor(BB: PrevValue->getParent()) != L) |
6287 | return true; |
6288 | |
6289 | // If a new pointer is loaded in the loop, the pointer references a different |
6290 | // object in every iteration. E.g.: |
6291 | // for (i) |
6292 | // int *p = a[i]; |
6293 | // ... |
6294 | if (auto *Load = dyn_cast<LoadInst>(Val: PrevValue)) |
6295 | if (!L->isLoopInvariant(V: Load->getPointerOperand())) |
6296 | return false; |
6297 | return true; |
6298 | } |
6299 | |
6300 | const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) { |
6301 | if (!V->getType()->isPointerTy()) |
6302 | return V; |
6303 | for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { |
6304 | if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) { |
6305 | V = GEP->getPointerOperand(); |
6306 | } else if (Operator::getOpcode(V) == Instruction::BitCast || |
6307 | Operator::getOpcode(V) == Instruction::AddrSpaceCast) { |
6308 | V = cast<Operator>(Val: V)->getOperand(i: 0); |
6309 | if (!V->getType()->isPointerTy()) |
6310 | return V; |
6311 | } else if (auto *GA = dyn_cast<GlobalAlias>(Val: V)) { |
6312 | if (GA->isInterposable()) |
6313 | return V; |
6314 | V = GA->getAliasee(); |
6315 | } else { |
6316 | if (auto *PHI = dyn_cast<PHINode>(Val: V)) { |
6317 | // Look through single-arg phi nodes created by LCSSA. |
6318 | if (PHI->getNumIncomingValues() == 1) { |
6319 | V = PHI->getIncomingValue(i: 0); |
6320 | continue; |
6321 | } |
6322 | } else if (auto *Call = dyn_cast<CallBase>(Val: V)) { |
6323 | // CaptureTracking can know about special capturing properties of some |
6324 | // intrinsics like launder.invariant.group, that can't be expressed with |
6325 | // the attributes, but have properties like returning aliasing pointer. |
6326 | // Because some analysis may assume that nocaptured pointer is not |
6327 | // returned from some special intrinsic (because function would have to |
6328 | // be marked with returns attribute), it is crucial to use this function |
6329 | // because it should be in sync with CaptureTracking. Not using it may |
6330 | // cause weird miscompilations where 2 aliasing pointers are assumed to |
6331 | // noalias. |
6332 | if (auto *RP = getArgumentAliasingToReturnedPointer(Call, MustPreserveNullness: false)) { |
6333 | V = RP; |
6334 | continue; |
6335 | } |
6336 | } |
6337 | |
6338 | return V; |
6339 | } |
6340 | assert(V->getType()->isPointerTy() && "Unexpected operand type!" ); |
6341 | } |
6342 | return V; |
6343 | } |
6344 | |
6345 | void llvm::getUnderlyingObjects(const Value *V, |
6346 | SmallVectorImpl<const Value *> &Objects, |
6347 | LoopInfo *LI, unsigned MaxLookup) { |
6348 | SmallPtrSet<const Value *, 4> Visited; |
6349 | SmallVector<const Value *, 4> Worklist; |
6350 | Worklist.push_back(Elt: V); |
6351 | do { |
6352 | const Value *P = Worklist.pop_back_val(); |
6353 | P = getUnderlyingObject(V: P, MaxLookup); |
6354 | |
6355 | if (!Visited.insert(Ptr: P).second) |
6356 | continue; |
6357 | |
6358 | if (auto *SI = dyn_cast<SelectInst>(Val: P)) { |
6359 | Worklist.push_back(Elt: SI->getTrueValue()); |
6360 | Worklist.push_back(Elt: SI->getFalseValue()); |
6361 | continue; |
6362 | } |
6363 | |
6364 | if (auto *PN = dyn_cast<PHINode>(Val: P)) { |
6365 | // If this PHI changes the underlying object in every iteration of the |
6366 | // loop, don't look through it. Consider: |
6367 | // int **A; |
6368 | // for (i) { |
6369 | // Prev = Curr; // Prev = PHI (Prev_0, Curr) |
6370 | // Curr = A[i]; |
6371 | // *Prev, *Curr; |
6372 | // |
6373 | // Prev is tracking Curr one iteration behind so they refer to different |
6374 | // underlying objects. |
6375 | if (!LI || !LI->isLoopHeader(BB: PN->getParent()) || |
6376 | isSameUnderlyingObjectInLoop(PN, LI)) |
6377 | append_range(C&: Worklist, R: PN->incoming_values()); |
6378 | else |
6379 | Objects.push_back(Elt: P); |
6380 | continue; |
6381 | } |
6382 | |
6383 | Objects.push_back(Elt: P); |
6384 | } while (!Worklist.empty()); |
6385 | } |
6386 | |
6387 | /// This is the function that does the work of looking through basic |
6388 | /// ptrtoint+arithmetic+inttoptr sequences. |
6389 | static const Value *getUnderlyingObjectFromInt(const Value *V) { |
6390 | do { |
6391 | if (const Operator *U = dyn_cast<Operator>(Val: V)) { |
6392 | // If we find a ptrtoint, we can transfer control back to the |
6393 | // regular getUnderlyingObjectFromInt. |
6394 | if (U->getOpcode() == Instruction::PtrToInt) |
6395 | return U->getOperand(i: 0); |
6396 | // If we find an add of a constant, a multiplied value, or a phi, it's |
6397 | // likely that the other operand will lead us to the base |
6398 | // object. We don't have to worry about the case where the |
6399 | // object address is somehow being computed by the multiply, |
6400 | // because our callers only care when the result is an |
6401 | // identifiable object. |
6402 | if (U->getOpcode() != Instruction::Add || |
6403 | (!isa<ConstantInt>(Val: U->getOperand(i: 1)) && |
6404 | Operator::getOpcode(V: U->getOperand(i: 1)) != Instruction::Mul && |
6405 | !isa<PHINode>(Val: U->getOperand(i: 1)))) |
6406 | return V; |
6407 | V = U->getOperand(i: 0); |
6408 | } else { |
6409 | return V; |
6410 | } |
6411 | assert(V->getType()->isIntegerTy() && "Unexpected operand type!" ); |
6412 | } while (true); |
6413 | } |
6414 | |
6415 | /// This is a wrapper around getUnderlyingObjects and adds support for basic |
6416 | /// ptrtoint+arithmetic+inttoptr sequences. |
6417 | /// It returns false if unidentified object is found in getUnderlyingObjects. |
6418 | bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, |
6419 | SmallVectorImpl<Value *> &Objects) { |
6420 | SmallPtrSet<const Value *, 16> Visited; |
6421 | SmallVector<const Value *, 4> Working(1, V); |
6422 | do { |
6423 | V = Working.pop_back_val(); |
6424 | |
6425 | SmallVector<const Value *, 4> Objs; |
6426 | getUnderlyingObjects(V, Objects&: Objs); |
6427 | |
6428 | for (const Value *V : Objs) { |
6429 | if (!Visited.insert(Ptr: V).second) |
6430 | continue; |
6431 | if (Operator::getOpcode(V) == Instruction::IntToPtr) { |
6432 | const Value *O = |
6433 | getUnderlyingObjectFromInt(V: cast<User>(Val: V)->getOperand(i: 0)); |
6434 | if (O->getType()->isPointerTy()) { |
6435 | Working.push_back(Elt: O); |
6436 | continue; |
6437 | } |
6438 | } |
6439 | // If getUnderlyingObjects fails to find an identifiable object, |
6440 | // getUnderlyingObjectsForCodeGen also fails for safety. |
6441 | if (!isIdentifiedObject(V)) { |
6442 | Objects.clear(); |
6443 | return false; |
6444 | } |
6445 | Objects.push_back(Elt: const_cast<Value *>(V)); |
6446 | } |
6447 | } while (!Working.empty()); |
6448 | return true; |
6449 | } |
6450 | |
6451 | AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { |
6452 | AllocaInst *Result = nullptr; |
6453 | SmallPtrSet<Value *, 4> Visited; |
6454 | SmallVector<Value *, 4> Worklist; |
6455 | |
6456 | auto AddWork = [&](Value *V) { |
6457 | if (Visited.insert(Ptr: V).second) |
6458 | Worklist.push_back(Elt: V); |
6459 | }; |
6460 | |
6461 | AddWork(V); |
6462 | do { |
6463 | V = Worklist.pop_back_val(); |
6464 | assert(Visited.count(V)); |
6465 | |
6466 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: V)) { |
6467 | if (Result && Result != AI) |
6468 | return nullptr; |
6469 | Result = AI; |
6470 | } else if (CastInst *CI = dyn_cast<CastInst>(Val: V)) { |
6471 | AddWork(CI->getOperand(i_nocapture: 0)); |
6472 | } else if (PHINode *PN = dyn_cast<PHINode>(Val: V)) { |
6473 | for (Value *IncValue : PN->incoming_values()) |
6474 | AddWork(IncValue); |
6475 | } else if (auto *SI = dyn_cast<SelectInst>(Val: V)) { |
6476 | AddWork(SI->getTrueValue()); |
6477 | AddWork(SI->getFalseValue()); |
6478 | } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: V)) { |
6479 | if (OffsetZero && !GEP->hasAllZeroIndices()) |
6480 | return nullptr; |
6481 | AddWork(GEP->getPointerOperand()); |
6482 | } else if (CallBase *CB = dyn_cast<CallBase>(Val: V)) { |
6483 | Value *Returned = CB->getReturnedArgOperand(); |
6484 | if (Returned) |
6485 | AddWork(Returned); |
6486 | else |
6487 | return nullptr; |
6488 | } else { |
6489 | return nullptr; |
6490 | } |
6491 | } while (!Worklist.empty()); |
6492 | |
6493 | return Result; |
6494 | } |
6495 | |
6496 | static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( |
6497 | const Value *V, bool AllowLifetime, bool AllowDroppable) { |
6498 | for (const User *U : V->users()) { |
6499 | const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U); |
6500 | if (!II) |
6501 | return false; |
6502 | |
6503 | if (AllowLifetime && II->isLifetimeStartOrEnd()) |
6504 | continue; |
6505 | |
6506 | if (AllowDroppable && II->isDroppable()) |
6507 | continue; |
6508 | |
6509 | return false; |
6510 | } |
6511 | return true; |
6512 | } |
6513 | |
6514 | bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { |
6515 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( |
6516 | V, /* AllowLifetime */ true, /* AllowDroppable */ false); |
6517 | } |
6518 | bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { |
6519 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( |
6520 | V, /* AllowLifetime */ true, /* AllowDroppable */ true); |
6521 | } |
6522 | |
6523 | bool llvm::mustSuppressSpeculation(const LoadInst &LI) { |
6524 | if (!LI.isUnordered()) |
6525 | return true; |
6526 | const Function &F = *LI.getFunction(); |
6527 | // Speculative load may create a race that did not exist in the source. |
6528 | return F.hasFnAttribute(Attribute::SanitizeThread) || |
6529 | // Speculative load may load data from dirty regions. |
6530 | F.hasFnAttribute(Attribute::SanitizeAddress) || |
6531 | F.hasFnAttribute(Attribute::SanitizeHWAddress); |
6532 | } |
6533 | |
6534 | bool llvm::isSafeToSpeculativelyExecute(const Instruction *Inst, |
6535 | const Instruction *CtxI, |
6536 | AssumptionCache *AC, |
6537 | const DominatorTree *DT, |
6538 | const TargetLibraryInfo *TLI) { |
6539 | return isSafeToSpeculativelyExecuteWithOpcode(Opcode: Inst->getOpcode(), Inst, CtxI, |
6540 | AC, DT, TLI); |
6541 | } |
6542 | |
6543 | bool llvm::isSafeToSpeculativelyExecuteWithOpcode( |
6544 | unsigned Opcode, const Instruction *Inst, const Instruction *CtxI, |
6545 | AssumptionCache *AC, const DominatorTree *DT, |
6546 | const TargetLibraryInfo *TLI) { |
6547 | #ifndef NDEBUG |
6548 | if (Inst->getOpcode() != Opcode) { |
6549 | // Check that the operands are actually compatible with the Opcode override. |
6550 | auto hasEqualReturnAndLeadingOperandTypes = |
6551 | [](const Instruction *Inst, unsigned NumLeadingOperands) { |
6552 | if (Inst->getNumOperands() < NumLeadingOperands) |
6553 | return false; |
6554 | const Type *ExpectedType = Inst->getType(); |
6555 | for (unsigned ItOp = 0; ItOp < NumLeadingOperands; ++ItOp) |
6556 | if (Inst->getOperand(i: ItOp)->getType() != ExpectedType) |
6557 | return false; |
6558 | return true; |
6559 | }; |
6560 | assert(!Instruction::isBinaryOp(Opcode) || |
6561 | hasEqualReturnAndLeadingOperandTypes(Inst, 2)); |
6562 | assert(!Instruction::isUnaryOp(Opcode) || |
6563 | hasEqualReturnAndLeadingOperandTypes(Inst, 1)); |
6564 | } |
6565 | #endif |
6566 | |
6567 | switch (Opcode) { |
6568 | default: |
6569 | return true; |
6570 | case Instruction::UDiv: |
6571 | case Instruction::URem: { |
6572 | // x / y is undefined if y == 0. |
6573 | const APInt *V; |
6574 | if (match(V: Inst->getOperand(i: 1), P: m_APInt(Res&: V))) |
6575 | return *V != 0; |
6576 | return false; |
6577 | } |
6578 | case Instruction::SDiv: |
6579 | case Instruction::SRem: { |
6580 | // x / y is undefined if y == 0 or x == INT_MIN and y == -1 |
6581 | const APInt *Numerator, *Denominator; |
6582 | if (!match(V: Inst->getOperand(i: 1), P: m_APInt(Res&: Denominator))) |
6583 | return false; |
6584 | // We cannot hoist this division if the denominator is 0. |
6585 | if (*Denominator == 0) |
6586 | return false; |
6587 | // It's safe to hoist if the denominator is not 0 or -1. |
6588 | if (!Denominator->isAllOnes()) |
6589 | return true; |
6590 | // At this point we know that the denominator is -1. It is safe to hoist as |
6591 | // long we know that the numerator is not INT_MIN. |
6592 | if (match(V: Inst->getOperand(i: 0), P: m_APInt(Res&: Numerator))) |
6593 | return !Numerator->isMinSignedValue(); |
6594 | // The numerator *might* be MinSignedValue. |
6595 | return false; |
6596 | } |
6597 | case Instruction::Load: { |
6598 | const LoadInst *LI = dyn_cast<LoadInst>(Val: Inst); |
6599 | if (!LI) |
6600 | return false; |
6601 | if (mustSuppressSpeculation(LI: *LI)) |
6602 | return false; |
6603 | const DataLayout &DL = LI->getModule()->getDataLayout(); |
6604 | return isDereferenceableAndAlignedPointer(V: LI->getPointerOperand(), |
6605 | Ty: LI->getType(), Alignment: LI->getAlign(), DL, |
6606 | CtxI, AC, DT, TLI); |
6607 | } |
6608 | case Instruction::Call: { |
6609 | auto *CI = dyn_cast<const CallInst>(Val: Inst); |
6610 | if (!CI) |
6611 | return false; |
6612 | const Function *Callee = CI->getCalledFunction(); |
6613 | |
6614 | // The called function could have undefined behavior or side-effects, even |
6615 | // if marked readnone nounwind. |
6616 | return Callee && Callee->isSpeculatable(); |
6617 | } |
6618 | case Instruction::VAArg: |
6619 | case Instruction::Alloca: |
6620 | case Instruction::Invoke: |
6621 | case Instruction::CallBr: |
6622 | case Instruction::PHI: |
6623 | case Instruction::Store: |
6624 | case Instruction::Ret: |
6625 | case Instruction::Br: |
6626 | case Instruction::IndirectBr: |
6627 | case Instruction::Switch: |
6628 | case Instruction::Unreachable: |
6629 | case Instruction::Fence: |
6630 | case Instruction::AtomicRMW: |
6631 | case Instruction::AtomicCmpXchg: |
6632 | case Instruction::LandingPad: |
6633 | case Instruction::Resume: |
6634 | case Instruction::CatchSwitch: |
6635 | case Instruction::CatchPad: |
6636 | case Instruction::CatchRet: |
6637 | case Instruction::CleanupPad: |
6638 | case Instruction::CleanupRet: |
6639 | return false; // Misc instructions which have effects |
6640 | } |
6641 | } |
6642 | |
6643 | bool llvm::mayHaveNonDefUseDependency(const Instruction &I) { |
6644 | if (I.mayReadOrWriteMemory()) |
6645 | // Memory dependency possible |
6646 | return true; |
6647 | if (!isSafeToSpeculativelyExecute(Inst: &I)) |
6648 | // Can't move above a maythrow call or infinite loop. Or if an |
6649 | // inalloca alloca, above a stacksave call. |
6650 | return true; |
6651 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
6652 | // 1) Can't reorder two inf-loop calls, even if readonly |
6653 | // 2) Also can't reorder an inf-loop call below a instruction which isn't |
6654 | // safe to speculative execute. (Inverse of above) |
6655 | return true; |
6656 | return false; |
6657 | } |
6658 | |
6659 | /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. |
6660 | static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { |
6661 | switch (OR) { |
6662 | case ConstantRange::OverflowResult::MayOverflow: |
6663 | return OverflowResult::MayOverflow; |
6664 | case ConstantRange::OverflowResult::AlwaysOverflowsLow: |
6665 | return OverflowResult::AlwaysOverflowsLow; |
6666 | case ConstantRange::OverflowResult::AlwaysOverflowsHigh: |
6667 | return OverflowResult::AlwaysOverflowsHigh; |
6668 | case ConstantRange::OverflowResult::NeverOverflows: |
6669 | return OverflowResult::NeverOverflows; |
6670 | } |
6671 | llvm_unreachable("Unknown OverflowResult" ); |
6672 | } |
6673 | |
6674 | /// Combine constant ranges from computeConstantRange() and computeKnownBits(). |
6675 | ConstantRange |
6676 | llvm::computeConstantRangeIncludingKnownBits(const WithCache<const Value *> &V, |
6677 | bool ForSigned, |
6678 | const SimplifyQuery &SQ) { |
6679 | ConstantRange CR1 = |
6680 | ConstantRange::fromKnownBits(Known: V.getKnownBits(Q: SQ), IsSigned: ForSigned); |
6681 | ConstantRange CR2 = computeConstantRange(V, ForSigned, UseInstrInfo: SQ.IIQ.UseInstrInfo); |
6682 | ConstantRange::PreferredRangeType RangeType = |
6683 | ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; |
6684 | return CR1.intersectWith(CR: CR2, Type: RangeType); |
6685 | } |
6686 | |
6687 | OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS, |
6688 | const Value *RHS, |
6689 | const SimplifyQuery &SQ) { |
6690 | KnownBits LHSKnown = computeKnownBits(V: LHS, /*Depth=*/0, Q: SQ); |
6691 | KnownBits RHSKnown = computeKnownBits(V: RHS, /*Depth=*/0, Q: SQ); |
6692 | ConstantRange LHSRange = ConstantRange::fromKnownBits(Known: LHSKnown, IsSigned: false); |
6693 | ConstantRange RHSRange = ConstantRange::fromKnownBits(Known: RHSKnown, IsSigned: false); |
6694 | return mapOverflowResult(OR: LHSRange.unsignedMulMayOverflow(Other: RHSRange)); |
6695 | } |
6696 | |
6697 | OverflowResult llvm::computeOverflowForSignedMul(const Value *LHS, |
6698 | const Value *RHS, |
6699 | const SimplifyQuery &SQ) { |
6700 | // Multiplying n * m significant bits yields a result of n + m significant |
6701 | // bits. If the total number of significant bits does not exceed the |
6702 | // result bit width (minus 1), there is no overflow. |
6703 | // This means if we have enough leading sign bits in the operands |
6704 | // we can guarantee that the result does not overflow. |
6705 | // Ref: "Hacker's Delight" by Henry Warren |
6706 | unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); |
6707 | |
6708 | // Note that underestimating the number of sign bits gives a more |
6709 | // conservative answer. |
6710 | unsigned SignBits = |
6711 | ::ComputeNumSignBits(V: LHS, Depth: 0, Q: SQ) + ::ComputeNumSignBits(V: RHS, Depth: 0, Q: SQ); |
6712 | |
6713 | // First handle the easy case: if we have enough sign bits there's |
6714 | // definitely no overflow. |
6715 | if (SignBits > BitWidth + 1) |
6716 | return OverflowResult::NeverOverflows; |
6717 | |
6718 | // There are two ambiguous cases where there can be no overflow: |
6719 | // SignBits == BitWidth + 1 and |
6720 | // SignBits == BitWidth |
6721 | // The second case is difficult to check, therefore we only handle the |
6722 | // first case. |
6723 | if (SignBits == BitWidth + 1) { |
6724 | // It overflows only when both arguments are negative and the true |
6725 | // product is exactly the minimum negative number. |
6726 | // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 |
6727 | // For simplicity we just check if at least one side is not negative. |
6728 | KnownBits LHSKnown = computeKnownBits(V: LHS, /*Depth=*/0, Q: SQ); |
6729 | KnownBits RHSKnown = computeKnownBits(V: RHS, /*Depth=*/0, Q: SQ); |
6730 | if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) |
6731 | return OverflowResult::NeverOverflows; |
6732 | } |
6733 | return OverflowResult::MayOverflow; |
6734 | } |
6735 | |
6736 | OverflowResult |
6737 | llvm::computeOverflowForUnsignedAdd(const WithCache<const Value *> &LHS, |
6738 | const WithCache<const Value *> &RHS, |
6739 | const SimplifyQuery &SQ) { |
6740 | ConstantRange LHSRange = |
6741 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/false, SQ); |
6742 | ConstantRange RHSRange = |
6743 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/false, SQ); |
6744 | return mapOverflowResult(OR: LHSRange.unsignedAddMayOverflow(Other: RHSRange)); |
6745 | } |
6746 | |
6747 | static OverflowResult |
6748 | computeOverflowForSignedAdd(const WithCache<const Value *> &LHS, |
6749 | const WithCache<const Value *> &RHS, |
6750 | const AddOperator *Add, const SimplifyQuery &SQ) { |
6751 | if (Add && Add->hasNoSignedWrap()) { |
6752 | return OverflowResult::NeverOverflows; |
6753 | } |
6754 | |
6755 | // If LHS and RHS each have at least two sign bits, the addition will look |
6756 | // like |
6757 | // |
6758 | // XX..... + |
6759 | // YY..... |
6760 | // |
6761 | // If the carry into the most significant position is 0, X and Y can't both |
6762 | // be 1 and therefore the carry out of the addition is also 0. |
6763 | // |
6764 | // If the carry into the most significant position is 1, X and Y can't both |
6765 | // be 0 and therefore the carry out of the addition is also 1. |
6766 | // |
6767 | // Since the carry into the most significant position is always equal to |
6768 | // the carry out of the addition, there is no signed overflow. |
6769 | if (::ComputeNumSignBits(V: LHS, Depth: 0, Q: SQ) > 1 && |
6770 | ::ComputeNumSignBits(V: RHS, Depth: 0, Q: SQ) > 1) |
6771 | return OverflowResult::NeverOverflows; |
6772 | |
6773 | ConstantRange LHSRange = |
6774 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/true, SQ); |
6775 | ConstantRange RHSRange = |
6776 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/true, SQ); |
6777 | OverflowResult OR = |
6778 | mapOverflowResult(OR: LHSRange.signedAddMayOverflow(Other: RHSRange)); |
6779 | if (OR != OverflowResult::MayOverflow) |
6780 | return OR; |
6781 | |
6782 | // The remaining code needs Add to be available. Early returns if not so. |
6783 | if (!Add) |
6784 | return OverflowResult::MayOverflow; |
6785 | |
6786 | // If the sign of Add is the same as at least one of the operands, this add |
6787 | // CANNOT overflow. If this can be determined from the known bits of the |
6788 | // operands the above signedAddMayOverflow() check will have already done so. |
6789 | // The only other way to improve on the known bits is from an assumption, so |
6790 | // call computeKnownBitsFromContext() directly. |
6791 | bool LHSOrRHSKnownNonNegative = |
6792 | (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); |
6793 | bool LHSOrRHSKnownNegative = |
6794 | (LHSRange.isAllNegative() || RHSRange.isAllNegative()); |
6795 | if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { |
6796 | KnownBits AddKnown(LHSRange.getBitWidth()); |
6797 | computeKnownBitsFromContext(V: Add, Known&: AddKnown, /*Depth=*/0, Q: SQ); |
6798 | if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || |
6799 | (AddKnown.isNegative() && LHSOrRHSKnownNegative)) |
6800 | return OverflowResult::NeverOverflows; |
6801 | } |
6802 | |
6803 | return OverflowResult::MayOverflow; |
6804 | } |
6805 | |
6806 | OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, |
6807 | const Value *RHS, |
6808 | const SimplifyQuery &SQ) { |
6809 | // X - (X % ?) |
6810 | // The remainder of a value can't have greater magnitude than itself, |
6811 | // so the subtraction can't overflow. |
6812 | |
6813 | // X - (X -nuw ?) |
6814 | // In the minimal case, this would simplify to "?", so there's no subtract |
6815 | // at all. But if this analysis is used to peek through casts, for example, |
6816 | // then determining no-overflow may allow other transforms. |
6817 | |
6818 | // TODO: There are other patterns like this. |
6819 | // See simplifyICmpWithBinOpOnLHS() for candidates. |
6820 | if (match(V: RHS, P: m_URem(L: m_Specific(V: LHS), R: m_Value())) || |
6821 | match(V: RHS, P: m_NUWSub(L: m_Specific(V: LHS), R: m_Value()))) |
6822 | if (isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
6823 | return OverflowResult::NeverOverflows; |
6824 | |
6825 | // Checking for conditions implied by dominating conditions may be expensive. |
6826 | // Limit it to usub_with_overflow calls for now. |
6827 | if (match(SQ.CxtI, |
6828 | m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) |
6829 | if (auto C = isImpliedByDomCondition(Pred: CmpInst::ICMP_UGE, LHS, RHS, ContextI: SQ.CxtI, |
6830 | DL: SQ.DL)) { |
6831 | if (*C) |
6832 | return OverflowResult::NeverOverflows; |
6833 | return OverflowResult::AlwaysOverflowsLow; |
6834 | } |
6835 | ConstantRange LHSRange = |
6836 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/false, SQ); |
6837 | ConstantRange RHSRange = |
6838 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/false, SQ); |
6839 | return mapOverflowResult(OR: LHSRange.unsignedSubMayOverflow(Other: RHSRange)); |
6840 | } |
6841 | |
6842 | OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, |
6843 | const Value *RHS, |
6844 | const SimplifyQuery &SQ) { |
6845 | // X - (X % ?) |
6846 | // The remainder of a value can't have greater magnitude than itself, |
6847 | // so the subtraction can't overflow. |
6848 | |
6849 | // X - (X -nsw ?) |
6850 | // In the minimal case, this would simplify to "?", so there's no subtract |
6851 | // at all. But if this analysis is used to peek through casts, for example, |
6852 | // then determining no-overflow may allow other transforms. |
6853 | if (match(V: RHS, P: m_SRem(L: m_Specific(V: LHS), R: m_Value())) || |
6854 | match(V: RHS, P: m_NSWSub(L: m_Specific(V: LHS), R: m_Value()))) |
6855 | if (isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) |
6856 | return OverflowResult::NeverOverflows; |
6857 | |
6858 | // If LHS and RHS each have at least two sign bits, the subtraction |
6859 | // cannot overflow. |
6860 | if (::ComputeNumSignBits(V: LHS, Depth: 0, Q: SQ) > 1 && |
6861 | ::ComputeNumSignBits(V: RHS, Depth: 0, Q: SQ) > 1) |
6862 | return OverflowResult::NeverOverflows; |
6863 | |
6864 | ConstantRange LHSRange = |
6865 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/true, SQ); |
6866 | ConstantRange RHSRange = |
6867 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/true, SQ); |
6868 | return mapOverflowResult(OR: LHSRange.signedSubMayOverflow(Other: RHSRange)); |
6869 | } |
6870 | |
6871 | bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, |
6872 | const DominatorTree &DT) { |
6873 | SmallVector<const BranchInst *, 2> GuardingBranches; |
6874 | SmallVector<const ExtractValueInst *, 2> Results; |
6875 | |
6876 | for (const User *U : WO->users()) { |
6877 | if (const auto *EVI = dyn_cast<ExtractValueInst>(Val: U)) { |
6878 | assert(EVI->getNumIndices() == 1 && "Obvious from CI's type" ); |
6879 | |
6880 | if (EVI->getIndices()[0] == 0) |
6881 | Results.push_back(Elt: EVI); |
6882 | else { |
6883 | assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type" ); |
6884 | |
6885 | for (const auto *U : EVI->users()) |
6886 | if (const auto *B = dyn_cast<BranchInst>(Val: U)) { |
6887 | assert(B->isConditional() && "How else is it using an i1?" ); |
6888 | GuardingBranches.push_back(Elt: B); |
6889 | } |
6890 | } |
6891 | } else { |
6892 | // We are using the aggregate directly in a way we don't want to analyze |
6893 | // here (storing it to a global, say). |
6894 | return false; |
6895 | } |
6896 | } |
6897 | |
6898 | auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { |
6899 | BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(i: 1)); |
6900 | if (!NoWrapEdge.isSingleEdge()) |
6901 | return false; |
6902 | |
6903 | // Check if all users of the add are provably no-wrap. |
6904 | for (const auto *Result : Results) { |
6905 | // If the extractvalue itself is not executed on overflow, the we don't |
6906 | // need to check each use separately, since domination is transitive. |
6907 | if (DT.dominates(BBE: NoWrapEdge, BB: Result->getParent())) |
6908 | continue; |
6909 | |
6910 | for (const auto &RU : Result->uses()) |
6911 | if (!DT.dominates(BBE: NoWrapEdge, U: RU)) |
6912 | return false; |
6913 | } |
6914 | |
6915 | return true; |
6916 | }; |
6917 | |
6918 | return llvm::any_of(Range&: GuardingBranches, P: AllUsesGuardedByBranch); |
6919 | } |
6920 | |
6921 | /// Shifts return poison if shiftwidth is larger than the bitwidth. |
6922 | static bool shiftAmountKnownInRange(const Value *ShiftAmount) { |
6923 | auto *C = dyn_cast<Constant>(Val: ShiftAmount); |
6924 | if (!C) |
6925 | return false; |
6926 | |
6927 | // Shifts return poison if shiftwidth is larger than the bitwidth. |
6928 | SmallVector<const Constant *, 4> ShiftAmounts; |
6929 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: C->getType())) { |
6930 | unsigned NumElts = FVTy->getNumElements(); |
6931 | for (unsigned i = 0; i < NumElts; ++i) |
6932 | ShiftAmounts.push_back(Elt: C->getAggregateElement(Elt: i)); |
6933 | } else if (isa<ScalableVectorType>(Val: C->getType())) |
6934 | return false; // Can't tell, just return false to be safe |
6935 | else |
6936 | ShiftAmounts.push_back(Elt: C); |
6937 | |
6938 | bool Safe = llvm::all_of(Range&: ShiftAmounts, P: [](const Constant *C) { |
6939 | auto *CI = dyn_cast_or_null<ConstantInt>(Val: C); |
6940 | return CI && CI->getValue().ult(RHS: C->getType()->getIntegerBitWidth()); |
6941 | }); |
6942 | |
6943 | return Safe; |
6944 | } |
6945 | |
6946 | enum class UndefPoisonKind { |
6947 | PoisonOnly = (1 << 0), |
6948 | UndefOnly = (1 << 1), |
6949 | UndefOrPoison = PoisonOnly | UndefOnly, |
6950 | }; |
6951 | |
6952 | static bool includesPoison(UndefPoisonKind Kind) { |
6953 | return (unsigned(Kind) & unsigned(UndefPoisonKind::PoisonOnly)) != 0; |
6954 | } |
6955 | |
6956 | static bool includesUndef(UndefPoisonKind Kind) { |
6957 | return (unsigned(Kind) & unsigned(UndefPoisonKind::UndefOnly)) != 0; |
6958 | } |
6959 | |
6960 | static bool canCreateUndefOrPoison(const Operator *Op, UndefPoisonKind Kind, |
6961 | bool ConsiderFlagsAndMetadata) { |
6962 | |
6963 | if (ConsiderFlagsAndMetadata && includesPoison(Kind) && |
6964 | Op->hasPoisonGeneratingAnnotations()) |
6965 | return true; |
6966 | |
6967 | unsigned Opcode = Op->getOpcode(); |
6968 | |
6969 | // Check whether opcode is a poison/undef-generating operation |
6970 | switch (Opcode) { |
6971 | case Instruction::Shl: |
6972 | case Instruction::AShr: |
6973 | case Instruction::LShr: |
6974 | return includesPoison(Kind) && !shiftAmountKnownInRange(ShiftAmount: Op->getOperand(i: 1)); |
6975 | case Instruction::FPToSI: |
6976 | case Instruction::FPToUI: |
6977 | // fptosi/ui yields poison if the resulting value does not fit in the |
6978 | // destination type. |
6979 | return true; |
6980 | case Instruction::Call: |
6981 | if (auto *II = dyn_cast<IntrinsicInst>(Val: Op)) { |
6982 | switch (II->getIntrinsicID()) { |
6983 | // TODO: Add more intrinsics. |
6984 | case Intrinsic::ctlz: |
6985 | case Intrinsic::cttz: |
6986 | case Intrinsic::abs: |
6987 | if (cast<ConstantInt>(Val: II->getArgOperand(i: 1))->isNullValue()) |
6988 | return false; |
6989 | break; |
6990 | case Intrinsic::ctpop: |
6991 | case Intrinsic::bswap: |
6992 | case Intrinsic::bitreverse: |
6993 | case Intrinsic::fshl: |
6994 | case Intrinsic::fshr: |
6995 | case Intrinsic::smax: |
6996 | case Intrinsic::smin: |
6997 | case Intrinsic::umax: |
6998 | case Intrinsic::umin: |
6999 | case Intrinsic::ptrmask: |
7000 | case Intrinsic::fptoui_sat: |
7001 | case Intrinsic::fptosi_sat: |
7002 | case Intrinsic::sadd_with_overflow: |
7003 | case Intrinsic::ssub_with_overflow: |
7004 | case Intrinsic::smul_with_overflow: |
7005 | case Intrinsic::uadd_with_overflow: |
7006 | case Intrinsic::usub_with_overflow: |
7007 | case Intrinsic::umul_with_overflow: |
7008 | case Intrinsic::sadd_sat: |
7009 | case Intrinsic::uadd_sat: |
7010 | case Intrinsic::ssub_sat: |
7011 | case Intrinsic::usub_sat: |
7012 | return false; |
7013 | case Intrinsic::sshl_sat: |
7014 | case Intrinsic::ushl_sat: |
7015 | return includesPoison(Kind) && |
7016 | !shiftAmountKnownInRange(ShiftAmount: II->getArgOperand(i: 1)); |
7017 | case Intrinsic::fma: |
7018 | case Intrinsic::fmuladd: |
7019 | case Intrinsic::sqrt: |
7020 | case Intrinsic::powi: |
7021 | case Intrinsic::sin: |
7022 | case Intrinsic::cos: |
7023 | case Intrinsic::pow: |
7024 | case Intrinsic::log: |
7025 | case Intrinsic::log10: |
7026 | case Intrinsic::log2: |
7027 | case Intrinsic::exp: |
7028 | case Intrinsic::exp2: |
7029 | case Intrinsic::exp10: |
7030 | case Intrinsic::fabs: |
7031 | case Intrinsic::copysign: |
7032 | case Intrinsic::floor: |
7033 | case Intrinsic::ceil: |
7034 | case Intrinsic::trunc: |
7035 | case Intrinsic::rint: |
7036 | case Intrinsic::nearbyint: |
7037 | case Intrinsic::round: |
7038 | case Intrinsic::roundeven: |
7039 | case Intrinsic::fptrunc_round: |
7040 | case Intrinsic::canonicalize: |
7041 | case Intrinsic::arithmetic_fence: |
7042 | case Intrinsic::minnum: |
7043 | case Intrinsic::maxnum: |
7044 | case Intrinsic::minimum: |
7045 | case Intrinsic::maximum: |
7046 | case Intrinsic::is_fpclass: |
7047 | case Intrinsic::ldexp: |
7048 | case Intrinsic::frexp: |
7049 | return false; |
7050 | case Intrinsic::lround: |
7051 | case Intrinsic::llround: |
7052 | case Intrinsic::lrint: |
7053 | case Intrinsic::llrint: |
7054 | // If the value doesn't fit an unspecified value is returned (but this |
7055 | // is not poison). |
7056 | return false; |
7057 | } |
7058 | } |
7059 | [[fallthrough]]; |
7060 | case Instruction::CallBr: |
7061 | case Instruction::Invoke: { |
7062 | const auto *CB = cast<CallBase>(Val: Op); |
7063 | return !CB->hasRetAttr(Attribute::NoUndef); |
7064 | } |
7065 | case Instruction::InsertElement: |
7066 | case Instruction::ExtractElement: { |
7067 | // If index exceeds the length of the vector, it returns poison |
7068 | auto *VTy = cast<VectorType>(Val: Op->getOperand(i: 0)->getType()); |
7069 | unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; |
7070 | auto *Idx = dyn_cast<ConstantInt>(Val: Op->getOperand(i: IdxOp)); |
7071 | if (includesPoison(Kind)) |
7072 | return !Idx || |
7073 | Idx->getValue().uge(RHS: VTy->getElementCount().getKnownMinValue()); |
7074 | return false; |
7075 | } |
7076 | case Instruction::ShuffleVector: { |
7077 | ArrayRef<int> Mask = isa<ConstantExpr>(Val: Op) |
7078 | ? cast<ConstantExpr>(Val: Op)->getShuffleMask() |
7079 | : cast<ShuffleVectorInst>(Val: Op)->getShuffleMask(); |
7080 | return includesPoison(Kind) && is_contained(Range&: Mask, Element: PoisonMaskElem); |
7081 | } |
7082 | case Instruction::FNeg: |
7083 | case Instruction::PHI: |
7084 | case Instruction::Select: |
7085 | case Instruction::URem: |
7086 | case Instruction::SRem: |
7087 | case Instruction::ExtractValue: |
7088 | case Instruction::InsertValue: |
7089 | case Instruction::Freeze: |
7090 | case Instruction::ICmp: |
7091 | case Instruction::FCmp: |
7092 | case Instruction::FAdd: |
7093 | case Instruction::FSub: |
7094 | case Instruction::FMul: |
7095 | case Instruction::FDiv: |
7096 | case Instruction::FRem: |
7097 | return false; |
7098 | case Instruction::GetElementPtr: |
7099 | // inbounds is handled above |
7100 | // TODO: what about inrange on constexpr? |
7101 | return false; |
7102 | default: { |
7103 | const auto *CE = dyn_cast<ConstantExpr>(Val: Op); |
7104 | if (isa<CastInst>(Val: Op) || (CE && CE->isCast())) |
7105 | return false; |
7106 | else if (Instruction::isBinaryOp(Opcode)) |
7107 | return false; |
7108 | // Be conservative and return true. |
7109 | return true; |
7110 | } |
7111 | } |
7112 | } |
7113 | |
7114 | bool llvm::canCreateUndefOrPoison(const Operator *Op, |
7115 | bool ConsiderFlagsAndMetadata) { |
7116 | return ::canCreateUndefOrPoison(Op, Kind: UndefPoisonKind::UndefOrPoison, |
7117 | ConsiderFlagsAndMetadata); |
7118 | } |
7119 | |
7120 | bool llvm::canCreatePoison(const Operator *Op, bool ConsiderFlagsAndMetadata) { |
7121 | return ::canCreateUndefOrPoison(Op, Kind: UndefPoisonKind::PoisonOnly, |
7122 | ConsiderFlagsAndMetadata); |
7123 | } |
7124 | |
7125 | static bool directlyImpliesPoison(const Value *ValAssumedPoison, const Value *V, |
7126 | unsigned Depth) { |
7127 | if (ValAssumedPoison == V) |
7128 | return true; |
7129 | |
7130 | const unsigned MaxDepth = 2; |
7131 | if (Depth >= MaxDepth) |
7132 | return false; |
7133 | |
7134 | if (const auto *I = dyn_cast<Instruction>(Val: V)) { |
7135 | if (any_of(Range: I->operands(), P: [=](const Use &Op) { |
7136 | return propagatesPoison(PoisonOp: Op) && |
7137 | directlyImpliesPoison(ValAssumedPoison, V: Op, Depth: Depth + 1); |
7138 | })) |
7139 | return true; |
7140 | |
7141 | // V = extractvalue V0, idx |
7142 | // V2 = extractvalue V0, idx2 |
7143 | // V0's elements are all poison or not. (e.g., add_with_overflow) |
7144 | const WithOverflowInst *II; |
7145 | if (match(V: I, P: m_ExtractValue(V: m_WithOverflowInst(I&: II))) && |
7146 | (match(V: ValAssumedPoison, P: m_ExtractValue(V: m_Specific(V: II))) || |
7147 | llvm::is_contained(Range: II->args(), Element: ValAssumedPoison))) |
7148 | return true; |
7149 | } |
7150 | return false; |
7151 | } |
7152 | |
7153 | static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, |
7154 | unsigned Depth) { |
7155 | if (isGuaranteedNotToBePoison(V: ValAssumedPoison)) |
7156 | return true; |
7157 | |
7158 | if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) |
7159 | return true; |
7160 | |
7161 | const unsigned MaxDepth = 2; |
7162 | if (Depth >= MaxDepth) |
7163 | return false; |
7164 | |
7165 | const auto *I = dyn_cast<Instruction>(Val: ValAssumedPoison); |
7166 | if (I && !canCreatePoison(Op: cast<Operator>(Val: I))) { |
7167 | return all_of(Range: I->operands(), P: [=](const Value *Op) { |
7168 | return impliesPoison(ValAssumedPoison: Op, V, Depth: Depth + 1); |
7169 | }); |
7170 | } |
7171 | return false; |
7172 | } |
7173 | |
7174 | bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { |
7175 | return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); |
7176 | } |
7177 | |
7178 | static bool programUndefinedIfUndefOrPoison(const Value *V, bool PoisonOnly); |
7179 | |
7180 | static bool isGuaranteedNotToBeUndefOrPoison( |
7181 | const Value *V, AssumptionCache *AC, const Instruction *CtxI, |
7182 | const DominatorTree *DT, unsigned Depth, UndefPoisonKind Kind) { |
7183 | if (Depth >= MaxAnalysisRecursionDepth) |
7184 | return false; |
7185 | |
7186 | if (isa<MetadataAsValue>(Val: V)) |
7187 | return false; |
7188 | |
7189 | if (const auto *A = dyn_cast<Argument>(Val: V)) { |
7190 | if (A->hasAttribute(Attribute::NoUndef) || |
7191 | A->hasAttribute(Attribute::Dereferenceable) || |
7192 | A->hasAttribute(Attribute::DereferenceableOrNull)) |
7193 | return true; |
7194 | } |
7195 | |
7196 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
7197 | if (isa<PoisonValue>(Val: C)) |
7198 | return !includesPoison(Kind); |
7199 | |
7200 | if (isa<UndefValue>(Val: C)) |
7201 | return !includesUndef(Kind); |
7202 | |
7203 | if (isa<ConstantInt>(Val: C) || isa<GlobalVariable>(Val: C) || isa<ConstantFP>(Val: V) || |
7204 | isa<ConstantPointerNull>(Val: C) || isa<Function>(Val: C)) |
7205 | return true; |
7206 | |
7207 | if (C->getType()->isVectorTy() && !isa<ConstantExpr>(Val: C)) |
7208 | return (!includesUndef(Kind) ? !C->containsPoisonElement() |
7209 | : !C->containsUndefOrPoisonElement()) && |
7210 | !C->containsConstantExpression(); |
7211 | } |
7212 | |
7213 | // Strip cast operations from a pointer value. |
7214 | // Note that stripPointerCastsSameRepresentation can strip off getelementptr |
7215 | // inbounds with zero offset. To guarantee that the result isn't poison, the |
7216 | // stripped pointer is checked as it has to be pointing into an allocated |
7217 | // object or be null `null` to ensure `inbounds` getelement pointers with a |
7218 | // zero offset could not produce poison. |
7219 | // It can strip off addrspacecast that do not change bit representation as |
7220 | // well. We believe that such addrspacecast is equivalent to no-op. |
7221 | auto *StrippedV = V->stripPointerCastsSameRepresentation(); |
7222 | if (isa<AllocaInst>(Val: StrippedV) || isa<GlobalVariable>(Val: StrippedV) || |
7223 | isa<Function>(Val: StrippedV) || isa<ConstantPointerNull>(Val: StrippedV)) |
7224 | return true; |
7225 | |
7226 | auto OpCheck = [&](const Value *V) { |
7227 | return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth: Depth + 1, Kind); |
7228 | }; |
7229 | |
7230 | if (auto *Opr = dyn_cast<Operator>(Val: V)) { |
7231 | // If the value is a freeze instruction, then it can never |
7232 | // be undef or poison. |
7233 | if (isa<FreezeInst>(Val: V)) |
7234 | return true; |
7235 | |
7236 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) { |
7237 | if (CB->hasRetAttr(Attribute::NoUndef) || |
7238 | CB->hasRetAttr(Attribute::Dereferenceable) || |
7239 | CB->hasRetAttr(Attribute::DereferenceableOrNull)) |
7240 | return true; |
7241 | } |
7242 | |
7243 | if (const auto *PN = dyn_cast<PHINode>(Val: V)) { |
7244 | unsigned Num = PN->getNumIncomingValues(); |
7245 | bool IsWellDefined = true; |
7246 | for (unsigned i = 0; i < Num; ++i) { |
7247 | auto *TI = PN->getIncomingBlock(i)->getTerminator(); |
7248 | if (!isGuaranteedNotToBeUndefOrPoison(V: PN->getIncomingValue(i), AC, CtxI: TI, |
7249 | DT, Depth: Depth + 1, Kind)) { |
7250 | IsWellDefined = false; |
7251 | break; |
7252 | } |
7253 | } |
7254 | if (IsWellDefined) |
7255 | return true; |
7256 | } else if (!::canCreateUndefOrPoison(Op: Opr, Kind, |
7257 | /*ConsiderFlagsAndMetadata*/ true) && |
7258 | all_of(Range: Opr->operands(), P: OpCheck)) |
7259 | return true; |
7260 | } |
7261 | |
7262 | if (auto *I = dyn_cast<LoadInst>(Val: V)) |
7263 | if (I->hasMetadata(KindID: LLVMContext::MD_noundef) || |
7264 | I->hasMetadata(KindID: LLVMContext::MD_dereferenceable) || |
7265 | I->hasMetadata(KindID: LLVMContext::MD_dereferenceable_or_null)) |
7266 | return true; |
7267 | |
7268 | if (programUndefinedIfUndefOrPoison(V, PoisonOnly: !includesUndef(Kind))) |
7269 | return true; |
7270 | |
7271 | // CxtI may be null or a cloned instruction. |
7272 | if (!CtxI || !CtxI->getParent() || !DT) |
7273 | return false; |
7274 | |
7275 | auto *DNode = DT->getNode(BB: CtxI->getParent()); |
7276 | if (!DNode) |
7277 | // Unreachable block |
7278 | return false; |
7279 | |
7280 | // If V is used as a branch condition before reaching CtxI, V cannot be |
7281 | // undef or poison. |
7282 | // br V, BB1, BB2 |
7283 | // BB1: |
7284 | // CtxI ; V cannot be undef or poison here |
7285 | auto *Dominator = DNode->getIDom(); |
7286 | while (Dominator) { |
7287 | auto *TI = Dominator->getBlock()->getTerminator(); |
7288 | |
7289 | Value *Cond = nullptr; |
7290 | if (auto BI = dyn_cast_or_null<BranchInst>(Val: TI)) { |
7291 | if (BI->isConditional()) |
7292 | Cond = BI->getCondition(); |
7293 | } else if (auto SI = dyn_cast_or_null<SwitchInst>(Val: TI)) { |
7294 | Cond = SI->getCondition(); |
7295 | } |
7296 | |
7297 | if (Cond) { |
7298 | if (Cond == V) |
7299 | return true; |
7300 | else if (!includesUndef(Kind) && isa<Operator>(Val: Cond)) { |
7301 | // For poison, we can analyze further |
7302 | auto *Opr = cast<Operator>(Val: Cond); |
7303 | if (any_of(Range: Opr->operands(), |
7304 | P: [V](const Use &U) { return V == U && propagatesPoison(PoisonOp: U); })) |
7305 | return true; |
7306 | } |
7307 | } |
7308 | |
7309 | Dominator = Dominator->getIDom(); |
7310 | } |
7311 | |
7312 | if (getKnowledgeValidInContext(V, {Attribute::NoUndef}, CtxI, DT, AC)) |
7313 | return true; |
7314 | |
7315 | return false; |
7316 | } |
7317 | |
7318 | bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, |
7319 | const Instruction *CtxI, |
7320 | const DominatorTree *DT, |
7321 | unsigned Depth) { |
7322 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
7323 | Kind: UndefPoisonKind::UndefOrPoison); |
7324 | } |
7325 | |
7326 | bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, |
7327 | const Instruction *CtxI, |
7328 | const DominatorTree *DT, unsigned Depth) { |
7329 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
7330 | Kind: UndefPoisonKind::PoisonOnly); |
7331 | } |
7332 | |
7333 | bool llvm::isGuaranteedNotToBeUndef(const Value *V, AssumptionCache *AC, |
7334 | const Instruction *CtxI, |
7335 | const DominatorTree *DT, unsigned Depth) { |
7336 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
7337 | Kind: UndefPoisonKind::UndefOnly); |
7338 | } |
7339 | |
7340 | /// Return true if undefined behavior would provably be executed on the path to |
7341 | /// OnPathTo if Root produced a posion result. Note that this doesn't say |
7342 | /// anything about whether OnPathTo is actually executed or whether Root is |
7343 | /// actually poison. This can be used to assess whether a new use of Root can |
7344 | /// be added at a location which is control equivalent with OnPathTo (such as |
7345 | /// immediately before it) without introducing UB which didn't previously |
7346 | /// exist. Note that a false result conveys no information. |
7347 | bool llvm::mustExecuteUBIfPoisonOnPathTo(Instruction *Root, |
7348 | Instruction *OnPathTo, |
7349 | DominatorTree *DT) { |
7350 | // Basic approach is to assume Root is poison, propagate poison forward |
7351 | // through all users we can easily track, and then check whether any of those |
7352 | // users are provable UB and must execute before out exiting block might |
7353 | // exit. |
7354 | |
7355 | // The set of all recursive users we've visited (which are assumed to all be |
7356 | // poison because of said visit) |
7357 | SmallSet<const Value *, 16> KnownPoison; |
7358 | SmallVector<const Instruction*, 16> Worklist; |
7359 | Worklist.push_back(Elt: Root); |
7360 | while (!Worklist.empty()) { |
7361 | const Instruction *I = Worklist.pop_back_val(); |
7362 | |
7363 | // If we know this must trigger UB on a path leading our target. |
7364 | if (mustTriggerUB(I, KnownPoison) && DT->dominates(Def: I, User: OnPathTo)) |
7365 | return true; |
7366 | |
7367 | // If we can't analyze propagation through this instruction, just skip it |
7368 | // and transitive users. Safe as false is a conservative result. |
7369 | if (I != Root && !any_of(Range: I->operands(), P: [&KnownPoison](const Use &U) { |
7370 | return KnownPoison.contains(Ptr: U) && propagatesPoison(PoisonOp: U); |
7371 | })) |
7372 | continue; |
7373 | |
7374 | if (KnownPoison.insert(Ptr: I).second) |
7375 | for (const User *User : I->users()) |
7376 | Worklist.push_back(Elt: cast<Instruction>(Val: User)); |
7377 | } |
7378 | |
7379 | // Might be non-UB, or might have a path we couldn't prove must execute on |
7380 | // way to exiting bb. |
7381 | return false; |
7382 | } |
7383 | |
7384 | OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, |
7385 | const SimplifyQuery &SQ) { |
7386 | return ::computeOverflowForSignedAdd(LHS: Add->getOperand(i_nocapture: 0), RHS: Add->getOperand(i_nocapture: 1), |
7387 | Add, SQ); |
7388 | } |
7389 | |
7390 | OverflowResult |
7391 | llvm::computeOverflowForSignedAdd(const WithCache<const Value *> &LHS, |
7392 | const WithCache<const Value *> &RHS, |
7393 | const SimplifyQuery &SQ) { |
7394 | return ::computeOverflowForSignedAdd(LHS, RHS, Add: nullptr, SQ); |
7395 | } |
7396 | |
7397 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { |
7398 | // Note: An atomic operation isn't guaranteed to return in a reasonable amount |
7399 | // of time because it's possible for another thread to interfere with it for an |
7400 | // arbitrary length of time, but programs aren't allowed to rely on that. |
7401 | |
7402 | // If there is no successor, then execution can't transfer to it. |
7403 | if (isa<ReturnInst>(Val: I)) |
7404 | return false; |
7405 | if (isa<UnreachableInst>(Val: I)) |
7406 | return false; |
7407 | |
7408 | // Note: Do not add new checks here; instead, change Instruction::mayThrow or |
7409 | // Instruction::willReturn. |
7410 | // |
7411 | // FIXME: Move this check into Instruction::willReturn. |
7412 | if (isa<CatchPadInst>(Val: I)) { |
7413 | switch (classifyEHPersonality(Pers: I->getFunction()->getPersonalityFn())) { |
7414 | default: |
7415 | // A catchpad may invoke exception object constructors and such, which |
7416 | // in some languages can be arbitrary code, so be conservative by default. |
7417 | return false; |
7418 | case EHPersonality::CoreCLR: |
7419 | // For CoreCLR, it just involves a type test. |
7420 | return true; |
7421 | } |
7422 | } |
7423 | |
7424 | // An instruction that returns without throwing must transfer control flow |
7425 | // to a successor. |
7426 | return !I->mayThrow() && I->willReturn(); |
7427 | } |
7428 | |
7429 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { |
7430 | // TODO: This is slightly conservative for invoke instruction since exiting |
7431 | // via an exception *is* normal control for them. |
7432 | for (const Instruction &I : *BB) |
7433 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
7434 | return false; |
7435 | return true; |
7436 | } |
7437 | |
7438 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( |
7439 | BasicBlock::const_iterator Begin, BasicBlock::const_iterator End, |
7440 | unsigned ScanLimit) { |
7441 | return isGuaranteedToTransferExecutionToSuccessor(Range: make_range(x: Begin, y: End), |
7442 | ScanLimit); |
7443 | } |
7444 | |
7445 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( |
7446 | iterator_range<BasicBlock::const_iterator> Range, unsigned ScanLimit) { |
7447 | assert(ScanLimit && "scan limit must be non-zero" ); |
7448 | for (const Instruction &I : Range) { |
7449 | if (isa<DbgInfoIntrinsic>(Val: I)) |
7450 | continue; |
7451 | if (--ScanLimit == 0) |
7452 | return false; |
7453 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
7454 | return false; |
7455 | } |
7456 | return true; |
7457 | } |
7458 | |
7459 | bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, |
7460 | const Loop *L) { |
7461 | // The loop header is guaranteed to be executed for every iteration. |
7462 | // |
7463 | // FIXME: Relax this constraint to cover all basic blocks that are |
7464 | // guaranteed to be executed at every iteration. |
7465 | if (I->getParent() != L->getHeader()) return false; |
7466 | |
7467 | for (const Instruction &LI : *L->getHeader()) { |
7468 | if (&LI == I) return true; |
7469 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &LI)) return false; |
7470 | } |
7471 | llvm_unreachable("Instruction not contained in its own parent basic block." ); |
7472 | } |
7473 | |
7474 | bool llvm::propagatesPoison(const Use &PoisonOp) { |
7475 | const Operator *I = cast<Operator>(Val: PoisonOp.getUser()); |
7476 | switch (I->getOpcode()) { |
7477 | case Instruction::Freeze: |
7478 | case Instruction::PHI: |
7479 | case Instruction::Invoke: |
7480 | return false; |
7481 | case Instruction::Select: |
7482 | return PoisonOp.getOperandNo() == 0; |
7483 | case Instruction::Call: |
7484 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
7485 | switch (II->getIntrinsicID()) { |
7486 | // TODO: Add more intrinsics. |
7487 | case Intrinsic::sadd_with_overflow: |
7488 | case Intrinsic::ssub_with_overflow: |
7489 | case Intrinsic::smul_with_overflow: |
7490 | case Intrinsic::uadd_with_overflow: |
7491 | case Intrinsic::usub_with_overflow: |
7492 | case Intrinsic::umul_with_overflow: |
7493 | // If an input is a vector containing a poison element, the |
7494 | // two output vectors (calculated results, overflow bits)' |
7495 | // corresponding lanes are poison. |
7496 | return true; |
7497 | case Intrinsic::ctpop: |
7498 | case Intrinsic::ctlz: |
7499 | case Intrinsic::cttz: |
7500 | case Intrinsic::abs: |
7501 | case Intrinsic::smax: |
7502 | case Intrinsic::smin: |
7503 | case Intrinsic::umax: |
7504 | case Intrinsic::umin: |
7505 | case Intrinsic::bitreverse: |
7506 | case Intrinsic::bswap: |
7507 | case Intrinsic::sadd_sat: |
7508 | case Intrinsic::ssub_sat: |
7509 | case Intrinsic::sshl_sat: |
7510 | case Intrinsic::uadd_sat: |
7511 | case Intrinsic::usub_sat: |
7512 | case Intrinsic::ushl_sat: |
7513 | return true; |
7514 | } |
7515 | } |
7516 | return false; |
7517 | case Instruction::ICmp: |
7518 | case Instruction::FCmp: |
7519 | case Instruction::GetElementPtr: |
7520 | return true; |
7521 | default: |
7522 | if (isa<BinaryOperator>(Val: I) || isa<UnaryOperator>(Val: I) || isa<CastInst>(Val: I)) |
7523 | return true; |
7524 | |
7525 | // Be conservative and return false. |
7526 | return false; |
7527 | } |
7528 | } |
7529 | |
7530 | /// Enumerates all operands of \p I that are guaranteed to not be undef or |
7531 | /// poison. If the callback \p Handle returns true, stop processing and return |
7532 | /// true. Otherwise, return false. |
7533 | template <typename CallableT> |
7534 | static bool handleGuaranteedWellDefinedOps(const Instruction *I, |
7535 | const CallableT &Handle) { |
7536 | switch (I->getOpcode()) { |
7537 | case Instruction::Store: |
7538 | if (Handle(cast<StoreInst>(Val: I)->getPointerOperand())) |
7539 | return true; |
7540 | break; |
7541 | |
7542 | case Instruction::Load: |
7543 | if (Handle(cast<LoadInst>(Val: I)->getPointerOperand())) |
7544 | return true; |
7545 | break; |
7546 | |
7547 | // Since dereferenceable attribute imply noundef, atomic operations |
7548 | // also implicitly have noundef pointers too |
7549 | case Instruction::AtomicCmpXchg: |
7550 | if (Handle(cast<AtomicCmpXchgInst>(Val: I)->getPointerOperand())) |
7551 | return true; |
7552 | break; |
7553 | |
7554 | case Instruction::AtomicRMW: |
7555 | if (Handle(cast<AtomicRMWInst>(Val: I)->getPointerOperand())) |
7556 | return true; |
7557 | break; |
7558 | |
7559 | case Instruction::Call: |
7560 | case Instruction::Invoke: { |
7561 | const CallBase *CB = cast<CallBase>(Val: I); |
7562 | if (CB->isIndirectCall() && Handle(CB->getCalledOperand())) |
7563 | return true; |
7564 | for (unsigned i = 0; i < CB->arg_size(); ++i) |
7565 | if ((CB->paramHasAttr(i, Attribute::NoUndef) || |
7566 | CB->paramHasAttr(i, Attribute::Dereferenceable) || |
7567 | CB->paramHasAttr(i, Attribute::DereferenceableOrNull)) && |
7568 | Handle(CB->getArgOperand(i))) |
7569 | return true; |
7570 | break; |
7571 | } |
7572 | case Instruction::Ret: |
7573 | if (I->getFunction()->hasRetAttribute(Attribute::NoUndef) && |
7574 | Handle(I->getOperand(0))) |
7575 | return true; |
7576 | break; |
7577 | case Instruction::Switch: |
7578 | if (Handle(cast<SwitchInst>(Val: I)->getCondition())) |
7579 | return true; |
7580 | break; |
7581 | case Instruction::Br: { |
7582 | auto *BR = cast<BranchInst>(Val: I); |
7583 | if (BR->isConditional() && Handle(BR->getCondition())) |
7584 | return true; |
7585 | break; |
7586 | } |
7587 | default: |
7588 | break; |
7589 | } |
7590 | |
7591 | return false; |
7592 | } |
7593 | |
7594 | void llvm::getGuaranteedWellDefinedOps( |
7595 | const Instruction *I, SmallVectorImpl<const Value *> &Operands) { |
7596 | handleGuaranteedWellDefinedOps(I, Handle: [&](const Value *V) { |
7597 | Operands.push_back(Elt: V); |
7598 | return false; |
7599 | }); |
7600 | } |
7601 | |
7602 | /// Enumerates all operands of \p I that are guaranteed to not be poison. |
7603 | template <typename CallableT> |
7604 | static bool handleGuaranteedNonPoisonOps(const Instruction *I, |
7605 | const CallableT &Handle) { |
7606 | if (handleGuaranteedWellDefinedOps(I, Handle)) |
7607 | return true; |
7608 | switch (I->getOpcode()) { |
7609 | // Divisors of these operations are allowed to be partially undef. |
7610 | case Instruction::UDiv: |
7611 | case Instruction::SDiv: |
7612 | case Instruction::URem: |
7613 | case Instruction::SRem: |
7614 | return Handle(I->getOperand(i: 1)); |
7615 | default: |
7616 | return false; |
7617 | } |
7618 | } |
7619 | |
7620 | void llvm::getGuaranteedNonPoisonOps(const Instruction *I, |
7621 | SmallVectorImpl<const Value *> &Operands) { |
7622 | handleGuaranteedNonPoisonOps(I, Handle: [&](const Value *V) { |
7623 | Operands.push_back(Elt: V); |
7624 | return false; |
7625 | }); |
7626 | } |
7627 | |
7628 | bool llvm::mustTriggerUB(const Instruction *I, |
7629 | const SmallPtrSetImpl<const Value *> &KnownPoison) { |
7630 | return handleGuaranteedNonPoisonOps( |
7631 | I, Handle: [&](const Value *V) { return KnownPoison.count(Ptr: V); }); |
7632 | } |
7633 | |
7634 | static bool programUndefinedIfUndefOrPoison(const Value *V, |
7635 | bool PoisonOnly) { |
7636 | // We currently only look for uses of values within the same basic |
7637 | // block, as that makes it easier to guarantee that the uses will be |
7638 | // executed given that Inst is executed. |
7639 | // |
7640 | // FIXME: Expand this to consider uses beyond the same basic block. To do |
7641 | // this, look out for the distinction between post-dominance and strong |
7642 | // post-dominance. |
7643 | const BasicBlock *BB = nullptr; |
7644 | BasicBlock::const_iterator Begin; |
7645 | if (const auto *Inst = dyn_cast<Instruction>(Val: V)) { |
7646 | BB = Inst->getParent(); |
7647 | Begin = Inst->getIterator(); |
7648 | Begin++; |
7649 | } else if (const auto *Arg = dyn_cast<Argument>(Val: V)) { |
7650 | if (Arg->getParent()->isDeclaration()) |
7651 | return false; |
7652 | BB = &Arg->getParent()->getEntryBlock(); |
7653 | Begin = BB->begin(); |
7654 | } else { |
7655 | return false; |
7656 | } |
7657 | |
7658 | // Limit number of instructions we look at, to avoid scanning through large |
7659 | // blocks. The current limit is chosen arbitrarily. |
7660 | unsigned ScanLimit = 32; |
7661 | BasicBlock::const_iterator End = BB->end(); |
7662 | |
7663 | if (!PoisonOnly) { |
7664 | // Since undef does not propagate eagerly, be conservative & just check |
7665 | // whether a value is directly passed to an instruction that must take |
7666 | // well-defined operands. |
7667 | |
7668 | for (const auto &I : make_range(x: Begin, y: End)) { |
7669 | if (isa<DbgInfoIntrinsic>(Val: I)) |
7670 | continue; |
7671 | if (--ScanLimit == 0) |
7672 | break; |
7673 | |
7674 | if (handleGuaranteedWellDefinedOps(I: &I, Handle: [V](const Value *WellDefinedOp) { |
7675 | return WellDefinedOp == V; |
7676 | })) |
7677 | return true; |
7678 | |
7679 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
7680 | break; |
7681 | } |
7682 | return false; |
7683 | } |
7684 | |
7685 | // Set of instructions that we have proved will yield poison if Inst |
7686 | // does. |
7687 | SmallSet<const Value *, 16> YieldsPoison; |
7688 | SmallSet<const BasicBlock *, 4> Visited; |
7689 | |
7690 | YieldsPoison.insert(Ptr: V); |
7691 | Visited.insert(Ptr: BB); |
7692 | |
7693 | while (true) { |
7694 | for (const auto &I : make_range(x: Begin, y: End)) { |
7695 | if (isa<DbgInfoIntrinsic>(Val: I)) |
7696 | continue; |
7697 | if (--ScanLimit == 0) |
7698 | return false; |
7699 | if (mustTriggerUB(I: &I, KnownPoison: YieldsPoison)) |
7700 | return true; |
7701 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
7702 | return false; |
7703 | |
7704 | // If an operand is poison and propagates it, mark I as yielding poison. |
7705 | for (const Use &Op : I.operands()) { |
7706 | if (YieldsPoison.count(Ptr: Op) && propagatesPoison(PoisonOp: Op)) { |
7707 | YieldsPoison.insert(Ptr: &I); |
7708 | break; |
7709 | } |
7710 | } |
7711 | |
7712 | // Special handling for select, which returns poison if its operand 0 is |
7713 | // poison (handled in the loop above) *or* if both its true/false operands |
7714 | // are poison (handled here). |
7715 | if (I.getOpcode() == Instruction::Select && |
7716 | YieldsPoison.count(Ptr: I.getOperand(i: 1)) && |
7717 | YieldsPoison.count(Ptr: I.getOperand(i: 2))) { |
7718 | YieldsPoison.insert(Ptr: &I); |
7719 | } |
7720 | } |
7721 | |
7722 | BB = BB->getSingleSuccessor(); |
7723 | if (!BB || !Visited.insert(Ptr: BB).second) |
7724 | break; |
7725 | |
7726 | Begin = BB->getFirstNonPHI()->getIterator(); |
7727 | End = BB->end(); |
7728 | } |
7729 | return false; |
7730 | } |
7731 | |
7732 | bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { |
7733 | return ::programUndefinedIfUndefOrPoison(V: Inst, PoisonOnly: false); |
7734 | } |
7735 | |
7736 | bool llvm::programUndefinedIfPoison(const Instruction *Inst) { |
7737 | return ::programUndefinedIfUndefOrPoison(V: Inst, PoisonOnly: true); |
7738 | } |
7739 | |
7740 | static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { |
7741 | if (FMF.noNaNs()) |
7742 | return true; |
7743 | |
7744 | if (auto *C = dyn_cast<ConstantFP>(Val: V)) |
7745 | return !C->isNaN(); |
7746 | |
7747 | if (auto *C = dyn_cast<ConstantDataVector>(Val: V)) { |
7748 | if (!C->getElementType()->isFloatingPointTy()) |
7749 | return false; |
7750 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { |
7751 | if (C->getElementAsAPFloat(i: I).isNaN()) |
7752 | return false; |
7753 | } |
7754 | return true; |
7755 | } |
7756 | |
7757 | if (isa<ConstantAggregateZero>(Val: V)) |
7758 | return true; |
7759 | |
7760 | return false; |
7761 | } |
7762 | |
7763 | static bool isKnownNonZero(const Value *V) { |
7764 | if (auto *C = dyn_cast<ConstantFP>(Val: V)) |
7765 | return !C->isZero(); |
7766 | |
7767 | if (auto *C = dyn_cast<ConstantDataVector>(Val: V)) { |
7768 | if (!C->getElementType()->isFloatingPointTy()) |
7769 | return false; |
7770 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { |
7771 | if (C->getElementAsAPFloat(i: I).isZero()) |
7772 | return false; |
7773 | } |
7774 | return true; |
7775 | } |
7776 | |
7777 | return false; |
7778 | } |
7779 | |
7780 | /// Match clamp pattern for float types without care about NaNs or signed zeros. |
7781 | /// Given non-min/max outer cmp/select from the clamp pattern this |
7782 | /// function recognizes if it can be substitued by a "canonical" min/max |
7783 | /// pattern. |
7784 | static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, |
7785 | Value *CmpLHS, Value *CmpRHS, |
7786 | Value *TrueVal, Value *FalseVal, |
7787 | Value *&LHS, Value *&RHS) { |
7788 | // Try to match |
7789 | // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) |
7790 | // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) |
7791 | // and return description of the outer Max/Min. |
7792 | |
7793 | // First, check if select has inverse order: |
7794 | if (CmpRHS == FalseVal) { |
7795 | std::swap(a&: TrueVal, b&: FalseVal); |
7796 | Pred = CmpInst::getInversePredicate(pred: Pred); |
7797 | } |
7798 | |
7799 | // Assume success now. If there's no match, callers should not use these anyway. |
7800 | LHS = TrueVal; |
7801 | RHS = FalseVal; |
7802 | |
7803 | const APFloat *FC1; |
7804 | if (CmpRHS != TrueVal || !match(V: CmpRHS, P: m_APFloat(Res&: FC1)) || !FC1->isFinite()) |
7805 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7806 | |
7807 | const APFloat *FC2; |
7808 | switch (Pred) { |
7809 | case CmpInst::FCMP_OLT: |
7810 | case CmpInst::FCMP_OLE: |
7811 | case CmpInst::FCMP_ULT: |
7812 | case CmpInst::FCMP_ULE: |
7813 | if (match(V: FalseVal, |
7814 | P: m_CombineOr(L: m_OrdFMin(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2)), |
7815 | R: m_UnordFMin(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2)))) && |
7816 | *FC1 < *FC2) |
7817 | return {.Flavor: SPF_FMAXNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
7818 | break; |
7819 | case CmpInst::FCMP_OGT: |
7820 | case CmpInst::FCMP_OGE: |
7821 | case CmpInst::FCMP_UGT: |
7822 | case CmpInst::FCMP_UGE: |
7823 | if (match(V: FalseVal, |
7824 | P: m_CombineOr(L: m_OrdFMax(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2)), |
7825 | R: m_UnordFMax(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2)))) && |
7826 | *FC1 > *FC2) |
7827 | return {.Flavor: SPF_FMINNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
7828 | break; |
7829 | default: |
7830 | break; |
7831 | } |
7832 | |
7833 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7834 | } |
7835 | |
7836 | /// Recognize variations of: |
7837 | /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) |
7838 | static SelectPatternResult matchClamp(CmpInst::Predicate Pred, |
7839 | Value *CmpLHS, Value *CmpRHS, |
7840 | Value *TrueVal, Value *FalseVal) { |
7841 | // Swap the select operands and predicate to match the patterns below. |
7842 | if (CmpRHS != TrueVal) { |
7843 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
7844 | std::swap(a&: TrueVal, b&: FalseVal); |
7845 | } |
7846 | const APInt *C1; |
7847 | if (CmpRHS == TrueVal && match(V: CmpRHS, P: m_APInt(Res&: C1))) { |
7848 | const APInt *C2; |
7849 | // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) |
7850 | if (match(V: FalseVal, P: m_SMin(L: m_Specific(V: CmpLHS), R: m_APInt(Res&: C2))) && |
7851 | C1->slt(RHS: *C2) && Pred == CmpInst::ICMP_SLT) |
7852 | return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7853 | |
7854 | // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) |
7855 | if (match(V: FalseVal, P: m_SMax(L: m_Specific(V: CmpLHS), R: m_APInt(Res&: C2))) && |
7856 | C1->sgt(RHS: *C2) && Pred == CmpInst::ICMP_SGT) |
7857 | return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7858 | |
7859 | // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) |
7860 | if (match(V: FalseVal, P: m_UMin(L: m_Specific(V: CmpLHS), R: m_APInt(Res&: C2))) && |
7861 | C1->ult(RHS: *C2) && Pred == CmpInst::ICMP_ULT) |
7862 | return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7863 | |
7864 | // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) |
7865 | if (match(V: FalseVal, P: m_UMax(L: m_Specific(V: CmpLHS), R: m_APInt(Res&: C2))) && |
7866 | C1->ugt(RHS: *C2) && Pred == CmpInst::ICMP_UGT) |
7867 | return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7868 | } |
7869 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7870 | } |
7871 | |
7872 | /// Recognize variations of: |
7873 | /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) |
7874 | static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, |
7875 | Value *CmpLHS, Value *CmpRHS, |
7876 | Value *TVal, Value *FVal, |
7877 | unsigned Depth) { |
7878 | // TODO: Allow FP min/max with nnan/nsz. |
7879 | assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison" ); |
7880 | |
7881 | Value *A = nullptr, *B = nullptr; |
7882 | SelectPatternResult L = matchSelectPattern(V: TVal, LHS&: A, RHS&: B, CastOp: nullptr, Depth: Depth + 1); |
7883 | if (!SelectPatternResult::isMinOrMax(SPF: L.Flavor)) |
7884 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7885 | |
7886 | Value *C = nullptr, *D = nullptr; |
7887 | SelectPatternResult R = matchSelectPattern(V: FVal, LHS&: C, RHS&: D, CastOp: nullptr, Depth: Depth + 1); |
7888 | if (L.Flavor != R.Flavor) |
7889 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7890 | |
7891 | // We have something like: x Pred y ? min(a, b) : min(c, d). |
7892 | // Try to match the compare to the min/max operations of the select operands. |
7893 | // First, make sure we have the right compare predicate. |
7894 | switch (L.Flavor) { |
7895 | case SPF_SMIN: |
7896 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { |
7897 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
7898 | std::swap(a&: CmpLHS, b&: CmpRHS); |
7899 | } |
7900 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) |
7901 | break; |
7902 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7903 | case SPF_SMAX: |
7904 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { |
7905 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
7906 | std::swap(a&: CmpLHS, b&: CmpRHS); |
7907 | } |
7908 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) |
7909 | break; |
7910 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7911 | case SPF_UMIN: |
7912 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { |
7913 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
7914 | std::swap(a&: CmpLHS, b&: CmpRHS); |
7915 | } |
7916 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) |
7917 | break; |
7918 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7919 | case SPF_UMAX: |
7920 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { |
7921 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
7922 | std::swap(a&: CmpLHS, b&: CmpRHS); |
7923 | } |
7924 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) |
7925 | break; |
7926 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7927 | default: |
7928 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7929 | } |
7930 | |
7931 | // If there is a common operand in the already matched min/max and the other |
7932 | // min/max operands match the compare operands (either directly or inverted), |
7933 | // then this is min/max of the same flavor. |
7934 | |
7935 | // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) |
7936 | // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) |
7937 | if (D == B) { |
7938 | if ((CmpLHS == A && CmpRHS == C) || (match(V: C, P: m_Not(V: m_Specific(V: CmpLHS))) && |
7939 | match(V: A, P: m_Not(V: m_Specific(V: CmpRHS))))) |
7940 | return {.Flavor: L.Flavor, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7941 | } |
7942 | // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) |
7943 | // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) |
7944 | if (C == B) { |
7945 | if ((CmpLHS == A && CmpRHS == D) || (match(V: D, P: m_Not(V: m_Specific(V: CmpLHS))) && |
7946 | match(V: A, P: m_Not(V: m_Specific(V: CmpRHS))))) |
7947 | return {.Flavor: L.Flavor, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7948 | } |
7949 | // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) |
7950 | // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) |
7951 | if (D == A) { |
7952 | if ((CmpLHS == B && CmpRHS == C) || (match(V: C, P: m_Not(V: m_Specific(V: CmpLHS))) && |
7953 | match(V: B, P: m_Not(V: m_Specific(V: CmpRHS))))) |
7954 | return {.Flavor: L.Flavor, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7955 | } |
7956 | // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) |
7957 | // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) |
7958 | if (C == A) { |
7959 | if ((CmpLHS == B && CmpRHS == D) || (match(V: D, P: m_Not(V: m_Specific(V: CmpLHS))) && |
7960 | match(V: B, P: m_Not(V: m_Specific(V: CmpRHS))))) |
7961 | return {.Flavor: L.Flavor, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7962 | } |
7963 | |
7964 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
7965 | } |
7966 | |
7967 | /// If the input value is the result of a 'not' op, constant integer, or vector |
7968 | /// splat of a constant integer, return the bitwise-not source value. |
7969 | /// TODO: This could be extended to handle non-splat vector integer constants. |
7970 | static Value *getNotValue(Value *V) { |
7971 | Value *NotV; |
7972 | if (match(V, P: m_Not(V: m_Value(V&: NotV)))) |
7973 | return NotV; |
7974 | |
7975 | const APInt *C; |
7976 | if (match(V, P: m_APInt(Res&: C))) |
7977 | return ConstantInt::get(Ty: V->getType(), V: ~(*C)); |
7978 | |
7979 | return nullptr; |
7980 | } |
7981 | |
7982 | /// Match non-obvious integer minimum and maximum sequences. |
7983 | static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, |
7984 | Value *CmpLHS, Value *CmpRHS, |
7985 | Value *TrueVal, Value *FalseVal, |
7986 | Value *&LHS, Value *&RHS, |
7987 | unsigned Depth) { |
7988 | // Assume success. If there's no match, callers should not use these anyway. |
7989 | LHS = TrueVal; |
7990 | RHS = FalseVal; |
7991 | |
7992 | SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); |
7993 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) |
7994 | return SPR; |
7995 | |
7996 | SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TVal: TrueVal, FVal: FalseVal, Depth); |
7997 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) |
7998 | return SPR; |
7999 | |
8000 | // Look through 'not' ops to find disguised min/max. |
8001 | // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) |
8002 | // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) |
8003 | if (CmpLHS == getNotValue(V: TrueVal) && CmpRHS == getNotValue(V: FalseVal)) { |
8004 | switch (Pred) { |
8005 | case CmpInst::ICMP_SGT: return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8006 | case CmpInst::ICMP_SLT: return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8007 | case CmpInst::ICMP_UGT: return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8008 | case CmpInst::ICMP_ULT: return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8009 | default: break; |
8010 | } |
8011 | } |
8012 | |
8013 | // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) |
8014 | // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) |
8015 | if (CmpLHS == getNotValue(V: FalseVal) && CmpRHS == getNotValue(V: TrueVal)) { |
8016 | switch (Pred) { |
8017 | case CmpInst::ICMP_SGT: return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8018 | case CmpInst::ICMP_SLT: return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8019 | case CmpInst::ICMP_UGT: return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8020 | case CmpInst::ICMP_ULT: return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8021 | default: break; |
8022 | } |
8023 | } |
8024 | |
8025 | if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) |
8026 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8027 | |
8028 | const APInt *C1; |
8029 | if (!match(V: CmpRHS, P: m_APInt(Res&: C1))) |
8030 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8031 | |
8032 | // An unsigned min/max can be written with a signed compare. |
8033 | const APInt *C2; |
8034 | if ((CmpLHS == TrueVal && match(V: FalseVal, P: m_APInt(Res&: C2))) || |
8035 | (CmpLHS == FalseVal && match(V: TrueVal, P: m_APInt(Res&: C2)))) { |
8036 | // Is the sign bit set? |
8037 | // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX |
8038 | // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN |
8039 | if (Pred == CmpInst::ICMP_SLT && C1->isZero() && C2->isMaxSignedValue()) |
8040 | return {.Flavor: CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8041 | |
8042 | // Is the sign bit clear? |
8043 | // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX |
8044 | // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN |
8045 | if (Pred == CmpInst::ICMP_SGT && C1->isAllOnes() && C2->isMinSignedValue()) |
8046 | return {.Flavor: CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8047 | } |
8048 | |
8049 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8050 | } |
8051 | |
8052 | bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW, |
8053 | bool AllowPoison) { |
8054 | assert(X && Y && "Invalid operand" ); |
8055 | |
8056 | auto IsNegationOf = [&](const Value *X, const Value *Y) { |
8057 | if (!match(V: X, P: m_Neg(V: m_Specific(V: Y)))) |
8058 | return false; |
8059 | |
8060 | auto *BO = cast<BinaryOperator>(Val: X); |
8061 | if (NeedNSW && !BO->hasNoSignedWrap()) |
8062 | return false; |
8063 | |
8064 | auto *Zero = cast<Constant>(Val: BO->getOperand(i_nocapture: 0)); |
8065 | if (!AllowPoison && !Zero->isNullValue()) |
8066 | return false; |
8067 | |
8068 | return true; |
8069 | }; |
8070 | |
8071 | // X = -Y or Y = -X |
8072 | if (IsNegationOf(X, Y) || IsNegationOf(Y, X)) |
8073 | return true; |
8074 | |
8075 | // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) |
8076 | Value *A, *B; |
8077 | return (!NeedNSW && (match(V: X, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B))) && |
8078 | match(V: Y, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))))) || |
8079 | (NeedNSW && (match(V: X, P: m_NSWSub(L: m_Value(V&: A), R: m_Value(V&: B))) && |
8080 | match(V: Y, P: m_NSWSub(L: m_Specific(V: B), R: m_Specific(V: A))))); |
8081 | } |
8082 | |
8083 | static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, |
8084 | FastMathFlags FMF, |
8085 | Value *CmpLHS, Value *CmpRHS, |
8086 | Value *TrueVal, Value *FalseVal, |
8087 | Value *&LHS, Value *&RHS, |
8088 | unsigned Depth) { |
8089 | bool HasMismatchedZeros = false; |
8090 | if (CmpInst::isFPPredicate(P: Pred)) { |
8091 | // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one |
8092 | // 0.0 operand, set the compare's 0.0 operands to that same value for the |
8093 | // purpose of identifying min/max. Disregard vector constants with undefined |
8094 | // elements because those can not be back-propagated for analysis. |
8095 | Value *OutputZeroVal = nullptr; |
8096 | if (match(V: TrueVal, P: m_AnyZeroFP()) && !match(V: FalseVal, P: m_AnyZeroFP()) && |
8097 | !cast<Constant>(Val: TrueVal)->containsUndefOrPoisonElement()) |
8098 | OutputZeroVal = TrueVal; |
8099 | else if (match(V: FalseVal, P: m_AnyZeroFP()) && !match(V: TrueVal, P: m_AnyZeroFP()) && |
8100 | !cast<Constant>(Val: FalseVal)->containsUndefOrPoisonElement()) |
8101 | OutputZeroVal = FalseVal; |
8102 | |
8103 | if (OutputZeroVal) { |
8104 | if (match(V: CmpLHS, P: m_AnyZeroFP()) && CmpLHS != OutputZeroVal) { |
8105 | HasMismatchedZeros = true; |
8106 | CmpLHS = OutputZeroVal; |
8107 | } |
8108 | if (match(V: CmpRHS, P: m_AnyZeroFP()) && CmpRHS != OutputZeroVal) { |
8109 | HasMismatchedZeros = true; |
8110 | CmpRHS = OutputZeroVal; |
8111 | } |
8112 | } |
8113 | } |
8114 | |
8115 | LHS = CmpLHS; |
8116 | RHS = CmpRHS; |
8117 | |
8118 | // Signed zero may return inconsistent results between implementations. |
8119 | // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 |
8120 | // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) |
8121 | // Therefore, we behave conservatively and only proceed if at least one of the |
8122 | // operands is known to not be zero or if we don't care about signed zero. |
8123 | switch (Pred) { |
8124 | default: break; |
8125 | case CmpInst::FCMP_OGT: case CmpInst::FCMP_OLT: |
8126 | case CmpInst::FCMP_UGT: case CmpInst::FCMP_ULT: |
8127 | if (!HasMismatchedZeros) |
8128 | break; |
8129 | [[fallthrough]]; |
8130 | case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: |
8131 | case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: |
8132 | if (!FMF.noSignedZeros() && !isKnownNonZero(V: CmpLHS) && |
8133 | !isKnownNonZero(V: CmpRHS)) |
8134 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8135 | } |
8136 | |
8137 | SelectPatternNaNBehavior NaNBehavior = SPNB_NA; |
8138 | bool Ordered = false; |
8139 | |
8140 | // When given one NaN and one non-NaN input: |
8141 | // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. |
8142 | // - A simple C99 (a < b ? a : b) construction will return 'b' (as the |
8143 | // ordered comparison fails), which could be NaN or non-NaN. |
8144 | // so here we discover exactly what NaN behavior is required/accepted. |
8145 | if (CmpInst::isFPPredicate(P: Pred)) { |
8146 | bool LHSSafe = isKnownNonNaN(V: CmpLHS, FMF); |
8147 | bool RHSSafe = isKnownNonNaN(V: CmpRHS, FMF); |
8148 | |
8149 | if (LHSSafe && RHSSafe) { |
8150 | // Both operands are known non-NaN. |
8151 | NaNBehavior = SPNB_RETURNS_ANY; |
8152 | } else if (CmpInst::isOrdered(predicate: Pred)) { |
8153 | // An ordered comparison will return false when given a NaN, so it |
8154 | // returns the RHS. |
8155 | Ordered = true; |
8156 | if (LHSSafe) |
8157 | // LHS is non-NaN, so if RHS is NaN then NaN will be returned. |
8158 | NaNBehavior = SPNB_RETURNS_NAN; |
8159 | else if (RHSSafe) |
8160 | NaNBehavior = SPNB_RETURNS_OTHER; |
8161 | else |
8162 | // Completely unsafe. |
8163 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8164 | } else { |
8165 | Ordered = false; |
8166 | // An unordered comparison will return true when given a NaN, so it |
8167 | // returns the LHS. |
8168 | if (LHSSafe) |
8169 | // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. |
8170 | NaNBehavior = SPNB_RETURNS_OTHER; |
8171 | else if (RHSSafe) |
8172 | NaNBehavior = SPNB_RETURNS_NAN; |
8173 | else |
8174 | // Completely unsafe. |
8175 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8176 | } |
8177 | } |
8178 | |
8179 | if (TrueVal == CmpRHS && FalseVal == CmpLHS) { |
8180 | std::swap(a&: CmpLHS, b&: CmpRHS); |
8181 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
8182 | if (NaNBehavior == SPNB_RETURNS_NAN) |
8183 | NaNBehavior = SPNB_RETURNS_OTHER; |
8184 | else if (NaNBehavior == SPNB_RETURNS_OTHER) |
8185 | NaNBehavior = SPNB_RETURNS_NAN; |
8186 | Ordered = !Ordered; |
8187 | } |
8188 | |
8189 | // ([if]cmp X, Y) ? X : Y |
8190 | if (TrueVal == CmpLHS && FalseVal == CmpRHS) { |
8191 | switch (Pred) { |
8192 | default: return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; // Equality. |
8193 | case ICmpInst::ICMP_UGT: |
8194 | case ICmpInst::ICMP_UGE: return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8195 | case ICmpInst::ICMP_SGT: |
8196 | case ICmpInst::ICMP_SGE: return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8197 | case ICmpInst::ICMP_ULT: |
8198 | case ICmpInst::ICMP_ULE: return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8199 | case ICmpInst::ICMP_SLT: |
8200 | case ICmpInst::ICMP_SLE: return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8201 | case FCmpInst::FCMP_UGT: |
8202 | case FCmpInst::FCMP_UGE: |
8203 | case FCmpInst::FCMP_OGT: |
8204 | case FCmpInst::FCMP_OGE: return {.Flavor: SPF_FMAXNUM, .NaNBehavior: NaNBehavior, .Ordered: Ordered}; |
8205 | case FCmpInst::FCMP_ULT: |
8206 | case FCmpInst::FCMP_ULE: |
8207 | case FCmpInst::FCMP_OLT: |
8208 | case FCmpInst::FCMP_OLE: return {.Flavor: SPF_FMINNUM, .NaNBehavior: NaNBehavior, .Ordered: Ordered}; |
8209 | } |
8210 | } |
8211 | |
8212 | if (isKnownNegation(X: TrueVal, Y: FalseVal)) { |
8213 | // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can |
8214 | // match against either LHS or sext(LHS). |
8215 | auto MaybeSExtCmpLHS = |
8216 | m_CombineOr(L: m_Specific(V: CmpLHS), R: m_SExt(Op: m_Specific(V: CmpLHS))); |
8217 | auto ZeroOrAllOnes = m_CombineOr(L: m_ZeroInt(), R: m_AllOnes()); |
8218 | auto ZeroOrOne = m_CombineOr(L: m_ZeroInt(), R: m_One()); |
8219 | if (match(V: TrueVal, P: MaybeSExtCmpLHS)) { |
8220 | // Set the return values. If the compare uses the negated value (-X >s 0), |
8221 | // swap the return values because the negated value is always 'RHS'. |
8222 | LHS = TrueVal; |
8223 | RHS = FalseVal; |
8224 | if (match(V: CmpLHS, P: m_Neg(V: m_Specific(V: FalseVal)))) |
8225 | std::swap(a&: LHS, b&: RHS); |
8226 | |
8227 | // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) |
8228 | // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) |
8229 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: ZeroOrAllOnes)) |
8230 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8231 | |
8232 | // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) |
8233 | if (Pred == ICmpInst::ICMP_SGE && match(V: CmpRHS, P: ZeroOrOne)) |
8234 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8235 | |
8236 | // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) |
8237 | // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) |
8238 | if (Pred == ICmpInst::ICMP_SLT && match(V: CmpRHS, P: ZeroOrOne)) |
8239 | return {.Flavor: SPF_NABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8240 | } |
8241 | else if (match(V: FalseVal, P: MaybeSExtCmpLHS)) { |
8242 | // Set the return values. If the compare uses the negated value (-X >s 0), |
8243 | // swap the return values because the negated value is always 'RHS'. |
8244 | LHS = FalseVal; |
8245 | RHS = TrueVal; |
8246 | if (match(V: CmpLHS, P: m_Neg(V: m_Specific(V: TrueVal)))) |
8247 | std::swap(a&: LHS, b&: RHS); |
8248 | |
8249 | // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) |
8250 | // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) |
8251 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: ZeroOrAllOnes)) |
8252 | return {.Flavor: SPF_NABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8253 | |
8254 | // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) |
8255 | // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) |
8256 | if (Pred == ICmpInst::ICMP_SLT && match(V: CmpRHS, P: ZeroOrOne)) |
8257 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8258 | } |
8259 | } |
8260 | |
8261 | if (CmpInst::isIntPredicate(P: Pred)) |
8262 | return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); |
8263 | |
8264 | // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar |
8265 | // may return either -0.0 or 0.0, so fcmp/select pair has stricter |
8266 | // semantics than minNum. Be conservative in such case. |
8267 | if (NaNBehavior != SPNB_RETURNS_ANY || |
8268 | (!FMF.noSignedZeros() && !isKnownNonZero(V: CmpLHS) && |
8269 | !isKnownNonZero(V: CmpRHS))) |
8270 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8271 | |
8272 | return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); |
8273 | } |
8274 | |
8275 | /// Helps to match a select pattern in case of a type mismatch. |
8276 | /// |
8277 | /// The function processes the case when type of true and false values of a |
8278 | /// select instruction differs from type of the cmp instruction operands because |
8279 | /// of a cast instruction. The function checks if it is legal to move the cast |
8280 | /// operation after "select". If yes, it returns the new second value of |
8281 | /// "select" (with the assumption that cast is moved): |
8282 | /// 1. As operand of cast instruction when both values of "select" are same cast |
8283 | /// instructions. |
8284 | /// 2. As restored constant (by applying reverse cast operation) when the first |
8285 | /// value of the "select" is a cast operation and the second value is a |
8286 | /// constant. |
8287 | /// NOTE: We return only the new second value because the first value could be |
8288 | /// accessed as operand of cast instruction. |
8289 | static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, |
8290 | Instruction::CastOps *CastOp) { |
8291 | auto *Cast1 = dyn_cast<CastInst>(Val: V1); |
8292 | if (!Cast1) |
8293 | return nullptr; |
8294 | |
8295 | *CastOp = Cast1->getOpcode(); |
8296 | Type *SrcTy = Cast1->getSrcTy(); |
8297 | if (auto *Cast2 = dyn_cast<CastInst>(Val: V2)) { |
8298 | // If V1 and V2 are both the same cast from the same type, look through V1. |
8299 | if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) |
8300 | return Cast2->getOperand(i_nocapture: 0); |
8301 | return nullptr; |
8302 | } |
8303 | |
8304 | auto *C = dyn_cast<Constant>(Val: V2); |
8305 | if (!C) |
8306 | return nullptr; |
8307 | |
8308 | const DataLayout &DL = CmpI->getModule()->getDataLayout(); |
8309 | Constant *CastedTo = nullptr; |
8310 | switch (*CastOp) { |
8311 | case Instruction::ZExt: |
8312 | if (CmpI->isUnsigned()) |
8313 | CastedTo = ConstantExpr::getTrunc(C, Ty: SrcTy); |
8314 | break; |
8315 | case Instruction::SExt: |
8316 | if (CmpI->isSigned()) |
8317 | CastedTo = ConstantExpr::getTrunc(C, Ty: SrcTy, OnlyIfReduced: true); |
8318 | break; |
8319 | case Instruction::Trunc: |
8320 | Constant *CmpConst; |
8321 | if (match(V: CmpI->getOperand(i_nocapture: 1), P: m_Constant(C&: CmpConst)) && |
8322 | CmpConst->getType() == SrcTy) { |
8323 | // Here we have the following case: |
8324 | // |
8325 | // %cond = cmp iN %x, CmpConst |
8326 | // %tr = trunc iN %x to iK |
8327 | // %narrowsel = select i1 %cond, iK %t, iK C |
8328 | // |
8329 | // We can always move trunc after select operation: |
8330 | // |
8331 | // %cond = cmp iN %x, CmpConst |
8332 | // %widesel = select i1 %cond, iN %x, iN CmpConst |
8333 | // %tr = trunc iN %widesel to iK |
8334 | // |
8335 | // Note that C could be extended in any way because we don't care about |
8336 | // upper bits after truncation. It can't be abs pattern, because it would |
8337 | // look like: |
8338 | // |
8339 | // select i1 %cond, x, -x. |
8340 | // |
8341 | // So only min/max pattern could be matched. Such match requires widened C |
8342 | // == CmpConst. That is why set widened C = CmpConst, condition trunc |
8343 | // CmpConst == C is checked below. |
8344 | CastedTo = CmpConst; |
8345 | } else { |
8346 | unsigned ExtOp = CmpI->isSigned() ? Instruction::SExt : Instruction::ZExt; |
8347 | CastedTo = ConstantFoldCastOperand(Opcode: ExtOp, C, DestTy: SrcTy, DL); |
8348 | } |
8349 | break; |
8350 | case Instruction::FPTrunc: |
8351 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPExt, C, DestTy: SrcTy, DL); |
8352 | break; |
8353 | case Instruction::FPExt: |
8354 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPTrunc, C, DestTy: SrcTy, DL); |
8355 | break; |
8356 | case Instruction::FPToUI: |
8357 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::UIToFP, C, DestTy: SrcTy, DL); |
8358 | break; |
8359 | case Instruction::FPToSI: |
8360 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::SIToFP, C, DestTy: SrcTy, DL); |
8361 | break; |
8362 | case Instruction::UIToFP: |
8363 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPToUI, C, DestTy: SrcTy, DL); |
8364 | break; |
8365 | case Instruction::SIToFP: |
8366 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPToSI, C, DestTy: SrcTy, DL); |
8367 | break; |
8368 | default: |
8369 | break; |
8370 | } |
8371 | |
8372 | if (!CastedTo) |
8373 | return nullptr; |
8374 | |
8375 | // Make sure the cast doesn't lose any information. |
8376 | Constant *CastedBack = |
8377 | ConstantFoldCastOperand(Opcode: *CastOp, C: CastedTo, DestTy: C->getType(), DL); |
8378 | if (CastedBack && CastedBack != C) |
8379 | return nullptr; |
8380 | |
8381 | return CastedTo; |
8382 | } |
8383 | |
8384 | SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, |
8385 | Instruction::CastOps *CastOp, |
8386 | unsigned Depth) { |
8387 | if (Depth >= MaxAnalysisRecursionDepth) |
8388 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8389 | |
8390 | SelectInst *SI = dyn_cast<SelectInst>(Val: V); |
8391 | if (!SI) return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8392 | |
8393 | CmpInst *CmpI = dyn_cast<CmpInst>(Val: SI->getCondition()); |
8394 | if (!CmpI) return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8395 | |
8396 | Value *TrueVal = SI->getTrueValue(); |
8397 | Value *FalseVal = SI->getFalseValue(); |
8398 | |
8399 | return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, |
8400 | CastOp, Depth); |
8401 | } |
8402 | |
8403 | SelectPatternResult llvm::matchDecomposedSelectPattern( |
8404 | CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, |
8405 | Instruction::CastOps *CastOp, unsigned Depth) { |
8406 | CmpInst::Predicate Pred = CmpI->getPredicate(); |
8407 | Value *CmpLHS = CmpI->getOperand(i_nocapture: 0); |
8408 | Value *CmpRHS = CmpI->getOperand(i_nocapture: 1); |
8409 | FastMathFlags FMF; |
8410 | if (isa<FPMathOperator>(Val: CmpI)) |
8411 | FMF = CmpI->getFastMathFlags(); |
8412 | |
8413 | // Bail out early. |
8414 | if (CmpI->isEquality()) |
8415 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
8416 | |
8417 | // Deal with type mismatches. |
8418 | if (CastOp && CmpLHS->getType() != TrueVal->getType()) { |
8419 | if (Value *C = lookThroughCast(CmpI, V1: TrueVal, V2: FalseVal, CastOp)) { |
8420 | // If this is a potential fmin/fmax with a cast to integer, then ignore |
8421 | // -0.0 because there is no corresponding integer value. |
8422 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) |
8423 | FMF.setNoSignedZeros(); |
8424 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, |
8425 | TrueVal: cast<CastInst>(Val: TrueVal)->getOperand(i_nocapture: 0), FalseVal: C, |
8426 | LHS, RHS, Depth); |
8427 | } |
8428 | if (Value *C = lookThroughCast(CmpI, V1: FalseVal, V2: TrueVal, CastOp)) { |
8429 | // If this is a potential fmin/fmax with a cast to integer, then ignore |
8430 | // -0.0 because there is no corresponding integer value. |
8431 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) |
8432 | FMF.setNoSignedZeros(); |
8433 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, |
8434 | TrueVal: C, FalseVal: cast<CastInst>(Val: FalseVal)->getOperand(i_nocapture: 0), |
8435 | LHS, RHS, Depth); |
8436 | } |
8437 | } |
8438 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, |
8439 | LHS, RHS, Depth); |
8440 | } |
8441 | |
8442 | CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { |
8443 | if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; |
8444 | if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; |
8445 | if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; |
8446 | if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; |
8447 | if (SPF == SPF_FMINNUM) |
8448 | return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; |
8449 | if (SPF == SPF_FMAXNUM) |
8450 | return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; |
8451 | llvm_unreachable("unhandled!" ); |
8452 | } |
8453 | |
8454 | SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { |
8455 | if (SPF == SPF_SMIN) return SPF_SMAX; |
8456 | if (SPF == SPF_UMIN) return SPF_UMAX; |
8457 | if (SPF == SPF_SMAX) return SPF_SMIN; |
8458 | if (SPF == SPF_UMAX) return SPF_UMIN; |
8459 | llvm_unreachable("unhandled!" ); |
8460 | } |
8461 | |
8462 | Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { |
8463 | switch (MinMaxID) { |
8464 | case Intrinsic::smax: return Intrinsic::smin; |
8465 | case Intrinsic::smin: return Intrinsic::smax; |
8466 | case Intrinsic::umax: return Intrinsic::umin; |
8467 | case Intrinsic::umin: return Intrinsic::umax; |
8468 | // Please note that next four intrinsics may produce the same result for |
8469 | // original and inverted case even if X != Y due to NaN is handled specially. |
8470 | case Intrinsic::maximum: return Intrinsic::minimum; |
8471 | case Intrinsic::minimum: return Intrinsic::maximum; |
8472 | case Intrinsic::maxnum: return Intrinsic::minnum; |
8473 | case Intrinsic::minnum: return Intrinsic::maxnum; |
8474 | default: llvm_unreachable("Unexpected intrinsic" ); |
8475 | } |
8476 | } |
8477 | |
8478 | APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { |
8479 | switch (SPF) { |
8480 | case SPF_SMAX: return APInt::getSignedMaxValue(numBits: BitWidth); |
8481 | case SPF_SMIN: return APInt::getSignedMinValue(numBits: BitWidth); |
8482 | case SPF_UMAX: return APInt::getMaxValue(numBits: BitWidth); |
8483 | case SPF_UMIN: return APInt::getMinValue(numBits: BitWidth); |
8484 | default: llvm_unreachable("Unexpected flavor" ); |
8485 | } |
8486 | } |
8487 | |
8488 | std::pair<Intrinsic::ID, bool> |
8489 | llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { |
8490 | // Check if VL contains select instructions that can be folded into a min/max |
8491 | // vector intrinsic and return the intrinsic if it is possible. |
8492 | // TODO: Support floating point min/max. |
8493 | bool AllCmpSingleUse = true; |
8494 | SelectPatternResult SelectPattern; |
8495 | SelectPattern.Flavor = SPF_UNKNOWN; |
8496 | if (all_of(Range&: VL, P: [&SelectPattern, &AllCmpSingleUse](Value *I) { |
8497 | Value *LHS, *RHS; |
8498 | auto CurrentPattern = matchSelectPattern(V: I, LHS, RHS); |
8499 | if (!SelectPatternResult::isMinOrMax(SPF: CurrentPattern.Flavor) || |
8500 | CurrentPattern.Flavor == SPF_FMINNUM || |
8501 | CurrentPattern.Flavor == SPF_FMAXNUM || |
8502 | !I->getType()->isIntOrIntVectorTy()) |
8503 | return false; |
8504 | if (SelectPattern.Flavor != SPF_UNKNOWN && |
8505 | SelectPattern.Flavor != CurrentPattern.Flavor) |
8506 | return false; |
8507 | SelectPattern = CurrentPattern; |
8508 | AllCmpSingleUse &= |
8509 | match(V: I, P: m_Select(C: m_OneUse(SubPattern: m_Value()), L: m_Value(), R: m_Value())); |
8510 | return true; |
8511 | })) { |
8512 | switch (SelectPattern.Flavor) { |
8513 | case SPF_SMIN: |
8514 | return {Intrinsic::smin, AllCmpSingleUse}; |
8515 | case SPF_UMIN: |
8516 | return {Intrinsic::umin, AllCmpSingleUse}; |
8517 | case SPF_SMAX: |
8518 | return {Intrinsic::smax, AllCmpSingleUse}; |
8519 | case SPF_UMAX: |
8520 | return {Intrinsic::umax, AllCmpSingleUse}; |
8521 | default: |
8522 | llvm_unreachable("unexpected select pattern flavor" ); |
8523 | } |
8524 | } |
8525 | return {Intrinsic::not_intrinsic, false}; |
8526 | } |
8527 | |
8528 | bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, |
8529 | Value *&Start, Value *&Step) { |
8530 | // Handle the case of a simple two-predecessor recurrence PHI. |
8531 | // There's a lot more that could theoretically be done here, but |
8532 | // this is sufficient to catch some interesting cases. |
8533 | if (P->getNumIncomingValues() != 2) |
8534 | return false; |
8535 | |
8536 | for (unsigned i = 0; i != 2; ++i) { |
8537 | Value *L = P->getIncomingValue(i); |
8538 | Value *R = P->getIncomingValue(i: !i); |
8539 | auto *LU = dyn_cast<BinaryOperator>(Val: L); |
8540 | if (!LU) |
8541 | continue; |
8542 | unsigned Opcode = LU->getOpcode(); |
8543 | |
8544 | switch (Opcode) { |
8545 | default: |
8546 | continue; |
8547 | // TODO: Expand list -- xor, div, gep, uaddo, etc.. |
8548 | case Instruction::LShr: |
8549 | case Instruction::AShr: |
8550 | case Instruction::Shl: |
8551 | case Instruction::Add: |
8552 | case Instruction::Sub: |
8553 | case Instruction::And: |
8554 | case Instruction::Or: |
8555 | case Instruction::Mul: |
8556 | case Instruction::FMul: { |
8557 | Value *LL = LU->getOperand(i_nocapture: 0); |
8558 | Value *LR = LU->getOperand(i_nocapture: 1); |
8559 | // Find a recurrence. |
8560 | if (LL == P) |
8561 | L = LR; |
8562 | else if (LR == P) |
8563 | L = LL; |
8564 | else |
8565 | continue; // Check for recurrence with L and R flipped. |
8566 | |
8567 | break; // Match! |
8568 | } |
8569 | }; |
8570 | |
8571 | // We have matched a recurrence of the form: |
8572 | // %iv = [R, %entry], [%iv.next, %backedge] |
8573 | // %iv.next = binop %iv, L |
8574 | // OR |
8575 | // %iv = [R, %entry], [%iv.next, %backedge] |
8576 | // %iv.next = binop L, %iv |
8577 | BO = LU; |
8578 | Start = R; |
8579 | Step = L; |
8580 | return true; |
8581 | } |
8582 | return false; |
8583 | } |
8584 | |
8585 | bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, |
8586 | Value *&Start, Value *&Step) { |
8587 | BinaryOperator *BO = nullptr; |
8588 | P = dyn_cast<PHINode>(Val: I->getOperand(i_nocapture: 0)); |
8589 | if (!P) |
8590 | P = dyn_cast<PHINode>(Val: I->getOperand(i_nocapture: 1)); |
8591 | return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I; |
8592 | } |
8593 | |
8594 | /// Return true if "icmp Pred LHS RHS" is always true. |
8595 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, |
8596 | const Value *RHS) { |
8597 | if (ICmpInst::isTrueWhenEqual(predicate: Pred) && LHS == RHS) |
8598 | return true; |
8599 | |
8600 | switch (Pred) { |
8601 | default: |
8602 | return false; |
8603 | |
8604 | case CmpInst::ICMP_SLE: { |
8605 | const APInt *C; |
8606 | |
8607 | // LHS s<= LHS +_{nsw} C if C >= 0 |
8608 | // LHS s<= LHS | C if C >= 0 |
8609 | if (match(V: RHS, P: m_NSWAdd(L: m_Specific(V: LHS), R: m_APInt(Res&: C))) || |
8610 | match(V: RHS, P: m_Or(L: m_Specific(V: LHS), R: m_APInt(Res&: C)))) |
8611 | return !C->isNegative(); |
8612 | |
8613 | // LHS s<= smax(LHS, V) for any V |
8614 | if (match(V: RHS, P: m_c_SMax(L: m_Specific(V: LHS), R: m_Value()))) |
8615 | return true; |
8616 | |
8617 | // smin(RHS, V) s<= RHS for any V |
8618 | if (match(V: LHS, P: m_c_SMin(L: m_Specific(V: RHS), R: m_Value()))) |
8619 | return true; |
8620 | |
8621 | // Match A to (X +_{nsw} CA) and B to (X +_{nsw} CB) |
8622 | const Value *X; |
8623 | const APInt *CLHS, *CRHS; |
8624 | if (match(V: LHS, P: m_NSWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: CLHS))) && |
8625 | match(V: RHS, P: m_NSWAddLike(L: m_Specific(V: X), R: m_APInt(Res&: CRHS)))) |
8626 | return CLHS->sle(RHS: *CRHS); |
8627 | |
8628 | return false; |
8629 | } |
8630 | |
8631 | case CmpInst::ICMP_ULE: { |
8632 | // LHS u<= LHS +_{nuw} V for any V |
8633 | if (match(V: RHS, P: m_c_Add(L: m_Specific(V: LHS), R: m_Value())) && |
8634 | cast<OverflowingBinaryOperator>(Val: RHS)->hasNoUnsignedWrap()) |
8635 | return true; |
8636 | |
8637 | // LHS u<= LHS | V for any V |
8638 | if (match(V: RHS, P: m_c_Or(L: m_Specific(V: LHS), R: m_Value()))) |
8639 | return true; |
8640 | |
8641 | // LHS u<= umax(LHS, V) for any V |
8642 | if (match(V: RHS, P: m_c_UMax(L: m_Specific(V: LHS), R: m_Value()))) |
8643 | return true; |
8644 | |
8645 | // RHS >> V u<= RHS for any V |
8646 | if (match(V: LHS, P: m_LShr(L: m_Specific(V: RHS), R: m_Value()))) |
8647 | return true; |
8648 | |
8649 | // RHS u/ C_ugt_1 u<= RHS |
8650 | const APInt *C; |
8651 | if (match(V: LHS, P: m_UDiv(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && C->ugt(RHS: 1)) |
8652 | return true; |
8653 | |
8654 | // RHS & V u<= RHS for any V |
8655 | if (match(V: LHS, P: m_c_And(L: m_Specific(V: RHS), R: m_Value()))) |
8656 | return true; |
8657 | |
8658 | // umin(RHS, V) u<= RHS for any V |
8659 | if (match(V: LHS, P: m_c_UMin(L: m_Specific(V: RHS), R: m_Value()))) |
8660 | return true; |
8661 | |
8662 | // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) |
8663 | const Value *X; |
8664 | const APInt *CLHS, *CRHS; |
8665 | if (match(V: LHS, P: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: CLHS))) && |
8666 | match(V: RHS, P: m_NUWAddLike(L: m_Specific(V: X), R: m_APInt(Res&: CRHS)))) |
8667 | return CLHS->ule(RHS: *CRHS); |
8668 | |
8669 | return false; |
8670 | } |
8671 | } |
8672 | } |
8673 | |
8674 | /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred |
8675 | /// ALHS ARHS" is true. Otherwise, return std::nullopt. |
8676 | static std::optional<bool> |
8677 | isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, |
8678 | const Value *ARHS, const Value *BLHS, const Value *BRHS) { |
8679 | switch (Pred) { |
8680 | default: |
8681 | return std::nullopt; |
8682 | |
8683 | case CmpInst::ICMP_SLT: |
8684 | case CmpInst::ICMP_SLE: |
8685 | if (isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: BLHS, RHS: ALHS) && |
8686 | isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: ARHS, RHS: BRHS)) |
8687 | return true; |
8688 | return std::nullopt; |
8689 | |
8690 | case CmpInst::ICMP_SGT: |
8691 | case CmpInst::ICMP_SGE: |
8692 | if (isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: ALHS, RHS: BLHS) && |
8693 | isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: BRHS, RHS: ARHS)) |
8694 | return true; |
8695 | return std::nullopt; |
8696 | |
8697 | case CmpInst::ICMP_ULT: |
8698 | case CmpInst::ICMP_ULE: |
8699 | if (isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: BLHS, RHS: ALHS) && |
8700 | isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: ARHS, RHS: BRHS)) |
8701 | return true; |
8702 | return std::nullopt; |
8703 | |
8704 | case CmpInst::ICMP_UGT: |
8705 | case CmpInst::ICMP_UGE: |
8706 | if (isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: ALHS, RHS: BLHS) && |
8707 | isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: BRHS, RHS: ARHS)) |
8708 | return true; |
8709 | return std::nullopt; |
8710 | } |
8711 | } |
8712 | |
8713 | /// Return true if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is true. |
8714 | /// Return false if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is false. |
8715 | /// Otherwise, return std::nullopt if we can't infer anything. |
8716 | static std::optional<bool> |
8717 | isImpliedCondMatchingOperands(CmpInst::Predicate LPred, |
8718 | CmpInst::Predicate RPred) { |
8719 | if (CmpInst::isImpliedTrueByMatchingCmp(Pred1: LPred, Pred2: RPred)) |
8720 | return true; |
8721 | if (CmpInst::isImpliedFalseByMatchingCmp(Pred1: LPred, Pred2: RPred)) |
8722 | return false; |
8723 | |
8724 | return std::nullopt; |
8725 | } |
8726 | |
8727 | /// Return true if "icmp LPred X, LC" implies "icmp RPred X, RC" is true. |
8728 | /// Return false if "icmp LPred X, LC" implies "icmp RPred X, RC" is false. |
8729 | /// Otherwise, return std::nullopt if we can't infer anything. |
8730 | static std::optional<bool> isImpliedCondCommonOperandWithConstants( |
8731 | CmpInst::Predicate LPred, const APInt &LC, CmpInst::Predicate RPred, |
8732 | const APInt &RC) { |
8733 | ConstantRange DomCR = ConstantRange::makeExactICmpRegion(Pred: LPred, Other: LC); |
8734 | ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred: RPred, Other: RC); |
8735 | ConstantRange Intersection = DomCR.intersectWith(CR); |
8736 | ConstantRange Difference = DomCR.difference(CR); |
8737 | if (Intersection.isEmptySet()) |
8738 | return false; |
8739 | if (Difference.isEmptySet()) |
8740 | return true; |
8741 | return std::nullopt; |
8742 | } |
8743 | |
8744 | /// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1") |
8745 | /// is true. Return false if LHS implies RHS is false. Otherwise, return |
8746 | /// std::nullopt if we can't infer anything. |
8747 | static std::optional<bool> isImpliedCondICmps(const ICmpInst *LHS, |
8748 | CmpInst::Predicate RPred, |
8749 | const Value *R0, const Value *R1, |
8750 | const DataLayout &DL, |
8751 | bool LHSIsTrue) { |
8752 | Value *L0 = LHS->getOperand(i_nocapture: 0); |
8753 | Value *L1 = LHS->getOperand(i_nocapture: 1); |
8754 | |
8755 | // The rest of the logic assumes the LHS condition is true. If that's not the |
8756 | // case, invert the predicate to make it so. |
8757 | CmpInst::Predicate LPred = |
8758 | LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); |
8759 | |
8760 | // We can have non-canonical operands, so try to normalize any common operand |
8761 | // to L0/R0. |
8762 | if (L0 == R1) { |
8763 | std::swap(a&: R0, b&: R1); |
8764 | RPred = ICmpInst::getSwappedPredicate(pred: RPred); |
8765 | } |
8766 | if (R0 == L1) { |
8767 | std::swap(a&: L0, b&: L1); |
8768 | LPred = ICmpInst::getSwappedPredicate(pred: LPred); |
8769 | } |
8770 | if (L1 == R1) { |
8771 | // If we have L0 == R0 and L1 == R1, then make L1/R1 the constants. |
8772 | if (L0 != R0 || match(V: L0, P: m_ImmConstant())) { |
8773 | std::swap(a&: L0, b&: L1); |
8774 | LPred = ICmpInst::getSwappedPredicate(pred: LPred); |
8775 | std::swap(a&: R0, b&: R1); |
8776 | RPred = ICmpInst::getSwappedPredicate(pred: RPred); |
8777 | } |
8778 | } |
8779 | |
8780 | // Can we infer anything when the 0-operands match and the 1-operands are |
8781 | // constants (not necessarily matching)? |
8782 | const APInt *LC, *RC; |
8783 | if (L0 == R0 && match(V: L1, P: m_APInt(Res&: LC)) && match(V: R1, P: m_APInt(Res&: RC))) |
8784 | return isImpliedCondCommonOperandWithConstants(LPred, LC: *LC, RPred, RC: *RC); |
8785 | |
8786 | // Can we infer anything when the two compares have matching operands? |
8787 | if (L0 == R0 && L1 == R1) |
8788 | return isImpliedCondMatchingOperands(LPred, RPred); |
8789 | |
8790 | // L0 = R0 = L1 + R1, L0 >=u L1 implies R0 >=u R1, L0 <u L1 implies R0 <u R1 |
8791 | if (L0 == R0 && |
8792 | (LPred == ICmpInst::ICMP_ULT || LPred == ICmpInst::ICMP_UGE) && |
8793 | (RPred == ICmpInst::ICMP_ULT || RPred == ICmpInst::ICMP_UGE) && |
8794 | match(V: L0, P: m_c_Add(L: m_Specific(V: L1), R: m_Specific(V: R1)))) |
8795 | return LPred == RPred; |
8796 | |
8797 | if (LPred == RPred) |
8798 | return isImpliedCondOperands(Pred: LPred, ALHS: L0, ARHS: L1, BLHS: R0, BRHS: R1); |
8799 | |
8800 | return std::nullopt; |
8801 | } |
8802 | |
8803 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is |
8804 | /// false. Otherwise, return std::nullopt if we can't infer anything. We |
8805 | /// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' |
8806 | /// instruction. |
8807 | static std::optional<bool> |
8808 | isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, |
8809 | const Value *RHSOp0, const Value *RHSOp1, |
8810 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { |
8811 | // The LHS must be an 'or', 'and', or a 'select' instruction. |
8812 | assert((LHS->getOpcode() == Instruction::And || |
8813 | LHS->getOpcode() == Instruction::Or || |
8814 | LHS->getOpcode() == Instruction::Select) && |
8815 | "Expected LHS to be 'and', 'or', or 'select'." ); |
8816 | |
8817 | assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit" ); |
8818 | |
8819 | // If the result of an 'or' is false, then we know both legs of the 'or' are |
8820 | // false. Similarly, if the result of an 'and' is true, then we know both |
8821 | // legs of the 'and' are true. |
8822 | const Value *ALHS, *ARHS; |
8823 | if ((!LHSIsTrue && match(V: LHS, P: m_LogicalOr(L: m_Value(V&: ALHS), R: m_Value(V&: ARHS)))) || |
8824 | (LHSIsTrue && match(V: LHS, P: m_LogicalAnd(L: m_Value(V&: ALHS), R: m_Value(V&: ARHS))))) { |
8825 | // FIXME: Make this non-recursion. |
8826 | if (std::optional<bool> Implication = isImpliedCondition( |
8827 | LHS: ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth: Depth + 1)) |
8828 | return Implication; |
8829 | if (std::optional<bool> Implication = isImpliedCondition( |
8830 | LHS: ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth: Depth + 1)) |
8831 | return Implication; |
8832 | return std::nullopt; |
8833 | } |
8834 | return std::nullopt; |
8835 | } |
8836 | |
8837 | std::optional<bool> |
8838 | llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, |
8839 | const Value *RHSOp0, const Value *RHSOp1, |
8840 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { |
8841 | // Bail out when we hit the limit. |
8842 | if (Depth == MaxAnalysisRecursionDepth) |
8843 | return std::nullopt; |
8844 | |
8845 | // A mismatch occurs when we compare a scalar cmp to a vector cmp, for |
8846 | // example. |
8847 | if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) |
8848 | return std::nullopt; |
8849 | |
8850 | assert(LHS->getType()->isIntOrIntVectorTy(1) && |
8851 | "Expected integer type only!" ); |
8852 | |
8853 | // Match not |
8854 | if (match(V: LHS, P: m_Not(V: m_Value(V&: LHS)))) |
8855 | LHSIsTrue = !LHSIsTrue; |
8856 | |
8857 | // Both LHS and RHS are icmps. |
8858 | const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(Val: LHS); |
8859 | if (LHSCmp) |
8860 | return isImpliedCondICmps(LHS: LHSCmp, RPred: RHSPred, R0: RHSOp0, R1: RHSOp1, DL, LHSIsTrue); |
8861 | |
8862 | /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect |
8863 | /// the RHS to be an icmp. |
8864 | /// FIXME: Add support for and/or/select on the RHS. |
8865 | if (const Instruction *LHSI = dyn_cast<Instruction>(Val: LHS)) { |
8866 | if ((LHSI->getOpcode() == Instruction::And || |
8867 | LHSI->getOpcode() == Instruction::Or || |
8868 | LHSI->getOpcode() == Instruction::Select)) |
8869 | return isImpliedCondAndOr(LHS: LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, |
8870 | Depth); |
8871 | } |
8872 | return std::nullopt; |
8873 | } |
8874 | |
8875 | std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, |
8876 | const DataLayout &DL, |
8877 | bool LHSIsTrue, unsigned Depth) { |
8878 | // LHS ==> RHS by definition |
8879 | if (LHS == RHS) |
8880 | return LHSIsTrue; |
8881 | |
8882 | // Match not |
8883 | bool InvertRHS = false; |
8884 | if (match(V: RHS, P: m_Not(V: m_Value(V&: RHS)))) { |
8885 | if (LHS == RHS) |
8886 | return !LHSIsTrue; |
8887 | InvertRHS = true; |
8888 | } |
8889 | |
8890 | if (const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(Val: RHS)) { |
8891 | if (auto Implied = isImpliedCondition( |
8892 | LHS, RHSPred: RHSCmp->getPredicate(), RHSOp0: RHSCmp->getOperand(i_nocapture: 0), |
8893 | RHSOp1: RHSCmp->getOperand(i_nocapture: 1), DL, LHSIsTrue, Depth)) |
8894 | return InvertRHS ? !*Implied : *Implied; |
8895 | return std::nullopt; |
8896 | } |
8897 | |
8898 | if (Depth == MaxAnalysisRecursionDepth) |
8899 | return std::nullopt; |
8900 | |
8901 | // LHS ==> (RHS1 || RHS2) if LHS ==> RHS1 or LHS ==> RHS2 |
8902 | // LHS ==> !(RHS1 && RHS2) if LHS ==> !RHS1 or LHS ==> !RHS2 |
8903 | const Value *RHS1, *RHS2; |
8904 | if (match(V: RHS, P: m_LogicalOr(L: m_Value(V&: RHS1), R: m_Value(V&: RHS2)))) { |
8905 | if (std::optional<bool> Imp = |
8906 | isImpliedCondition(LHS, RHS: RHS1, DL, LHSIsTrue, Depth: Depth + 1)) |
8907 | if (*Imp == true) |
8908 | return !InvertRHS; |
8909 | if (std::optional<bool> Imp = |
8910 | isImpliedCondition(LHS, RHS: RHS2, DL, LHSIsTrue, Depth: Depth + 1)) |
8911 | if (*Imp == true) |
8912 | return !InvertRHS; |
8913 | } |
8914 | if (match(V: RHS, P: m_LogicalAnd(L: m_Value(V&: RHS1), R: m_Value(V&: RHS2)))) { |
8915 | if (std::optional<bool> Imp = |
8916 | isImpliedCondition(LHS, RHS: RHS1, DL, LHSIsTrue, Depth: Depth + 1)) |
8917 | if (*Imp == false) |
8918 | return InvertRHS; |
8919 | if (std::optional<bool> Imp = |
8920 | isImpliedCondition(LHS, RHS: RHS2, DL, LHSIsTrue, Depth: Depth + 1)) |
8921 | if (*Imp == false) |
8922 | return InvertRHS; |
8923 | } |
8924 | |
8925 | return std::nullopt; |
8926 | } |
8927 | |
8928 | // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch |
8929 | // condition dominating ContextI or nullptr, if no condition is found. |
8930 | static std::pair<Value *, bool> |
8931 | getDomPredecessorCondition(const Instruction *ContextI) { |
8932 | if (!ContextI || !ContextI->getParent()) |
8933 | return {nullptr, false}; |
8934 | |
8935 | // TODO: This is a poor/cheap way to determine dominance. Should we use a |
8936 | // dominator tree (eg, from a SimplifyQuery) instead? |
8937 | const BasicBlock *ContextBB = ContextI->getParent(); |
8938 | const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); |
8939 | if (!PredBB) |
8940 | return {nullptr, false}; |
8941 | |
8942 | // We need a conditional branch in the predecessor. |
8943 | Value *PredCond; |
8944 | BasicBlock *TrueBB, *FalseBB; |
8945 | if (!match(V: PredBB->getTerminator(), P: m_Br(C: m_Value(V&: PredCond), T&: TrueBB, F&: FalseBB))) |
8946 | return {nullptr, false}; |
8947 | |
8948 | // The branch should get simplified. Don't bother simplifying this condition. |
8949 | if (TrueBB == FalseBB) |
8950 | return {nullptr, false}; |
8951 | |
8952 | assert((TrueBB == ContextBB || FalseBB == ContextBB) && |
8953 | "Predecessor block does not point to successor?" ); |
8954 | |
8955 | // Is this condition implied by the predecessor condition? |
8956 | return {PredCond, TrueBB == ContextBB}; |
8957 | } |
8958 | |
8959 | std::optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, |
8960 | const Instruction *ContextI, |
8961 | const DataLayout &DL) { |
8962 | assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool" ); |
8963 | auto PredCond = getDomPredecessorCondition(ContextI); |
8964 | if (PredCond.first) |
8965 | return isImpliedCondition(LHS: PredCond.first, RHS: Cond, DL, LHSIsTrue: PredCond.second); |
8966 | return std::nullopt; |
8967 | } |
8968 | |
8969 | std::optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, |
8970 | const Value *LHS, |
8971 | const Value *RHS, |
8972 | const Instruction *ContextI, |
8973 | const DataLayout &DL) { |
8974 | auto PredCond = getDomPredecessorCondition(ContextI); |
8975 | if (PredCond.first) |
8976 | return isImpliedCondition(LHS: PredCond.first, RHSPred: Pred, RHSOp0: LHS, RHSOp1: RHS, DL, |
8977 | LHSIsTrue: PredCond.second); |
8978 | return std::nullopt; |
8979 | } |
8980 | |
8981 | static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, |
8982 | APInt &Upper, const InstrInfoQuery &IIQ, |
8983 | bool PreferSignedRange) { |
8984 | unsigned Width = Lower.getBitWidth(); |
8985 | const APInt *C; |
8986 | switch (BO.getOpcode()) { |
8987 | case Instruction::Add: |
8988 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero()) { |
8989 | bool HasNSW = IIQ.hasNoSignedWrap(Op: &BO); |
8990 | bool HasNUW = IIQ.hasNoUnsignedWrap(Op: &BO); |
8991 | |
8992 | // If the caller expects a signed compare, then try to use a signed range. |
8993 | // Otherwise if both no-wraps are set, use the unsigned range because it |
8994 | // is never larger than the signed range. Example: |
8995 | // "add nuw nsw i8 X, -2" is unsigned [254,255] vs. signed [-128, 125]. |
8996 | if (PreferSignedRange && HasNSW && HasNUW) |
8997 | HasNUW = false; |
8998 | |
8999 | if (HasNUW) { |
9000 | // 'add nuw x, C' produces [C, UINT_MAX]. |
9001 | Lower = *C; |
9002 | } else if (HasNSW) { |
9003 | if (C->isNegative()) { |
9004 | // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. |
9005 | Lower = APInt::getSignedMinValue(numBits: Width); |
9006 | Upper = APInt::getSignedMaxValue(numBits: Width) + *C + 1; |
9007 | } else { |
9008 | // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. |
9009 | Lower = APInt::getSignedMinValue(numBits: Width) + *C; |
9010 | Upper = APInt::getSignedMaxValue(numBits: Width) + 1; |
9011 | } |
9012 | } |
9013 | } |
9014 | break; |
9015 | |
9016 | case Instruction::And: |
9017 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9018 | // 'and x, C' produces [0, C]. |
9019 | Upper = *C + 1; |
9020 | // X & -X is a power of two or zero. So we can cap the value at max power of |
9021 | // two. |
9022 | if (match(V: BO.getOperand(i_nocapture: 0), P: m_Neg(V: m_Specific(V: BO.getOperand(i_nocapture: 1)))) || |
9023 | match(V: BO.getOperand(i_nocapture: 1), P: m_Neg(V: m_Specific(V: BO.getOperand(i_nocapture: 0))))) |
9024 | Upper = APInt::getSignedMinValue(numBits: Width) + 1; |
9025 | break; |
9026 | |
9027 | case Instruction::Or: |
9028 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9029 | // 'or x, C' produces [C, UINT_MAX]. |
9030 | Lower = *C; |
9031 | break; |
9032 | |
9033 | case Instruction::AShr: |
9034 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
9035 | // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. |
9036 | Lower = APInt::getSignedMinValue(numBits: Width).ashr(ShiftAmt: *C); |
9037 | Upper = APInt::getSignedMaxValue(numBits: Width).ashr(ShiftAmt: *C) + 1; |
9038 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9039 | unsigned ShiftAmount = Width - 1; |
9040 | if (!C->isZero() && IIQ.isExact(Op: &BO)) |
9041 | ShiftAmount = C->countr_zero(); |
9042 | if (C->isNegative()) { |
9043 | // 'ashr C, x' produces [C, C >> (Width-1)] |
9044 | Lower = *C; |
9045 | Upper = C->ashr(ShiftAmt: ShiftAmount) + 1; |
9046 | } else { |
9047 | // 'ashr C, x' produces [C >> (Width-1), C] |
9048 | Lower = C->ashr(ShiftAmt: ShiftAmount); |
9049 | Upper = *C + 1; |
9050 | } |
9051 | } |
9052 | break; |
9053 | |
9054 | case Instruction::LShr: |
9055 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
9056 | // 'lshr x, C' produces [0, UINT_MAX >> C]. |
9057 | Upper = APInt::getAllOnes(numBits: Width).lshr(ShiftAmt: *C) + 1; |
9058 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9059 | // 'lshr C, x' produces [C >> (Width-1), C]. |
9060 | unsigned ShiftAmount = Width - 1; |
9061 | if (!C->isZero() && IIQ.isExact(Op: &BO)) |
9062 | ShiftAmount = C->countr_zero(); |
9063 | Lower = C->lshr(shiftAmt: ShiftAmount); |
9064 | Upper = *C + 1; |
9065 | } |
9066 | break; |
9067 | |
9068 | case Instruction::Shl: |
9069 | if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9070 | if (IIQ.hasNoUnsignedWrap(Op: &BO)) { |
9071 | // 'shl nuw C, x' produces [C, C << CLZ(C)] |
9072 | Lower = *C; |
9073 | Upper = Lower.shl(shiftAmt: Lower.countl_zero()) + 1; |
9074 | } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? |
9075 | if (C->isNegative()) { |
9076 | // 'shl nsw C, x' produces [C << CLO(C)-1, C] |
9077 | unsigned ShiftAmount = C->countl_one() - 1; |
9078 | Lower = C->shl(shiftAmt: ShiftAmount); |
9079 | Upper = *C + 1; |
9080 | } else { |
9081 | // 'shl nsw C, x' produces [C, C << CLZ(C)-1] |
9082 | unsigned ShiftAmount = C->countl_zero() - 1; |
9083 | Lower = *C; |
9084 | Upper = C->shl(shiftAmt: ShiftAmount) + 1; |
9085 | } |
9086 | } else { |
9087 | // If lowbit is set, value can never be zero. |
9088 | if ((*C)[0]) |
9089 | Lower = APInt::getOneBitSet(numBits: Width, BitNo: 0); |
9090 | // If we are shifting a constant the largest it can be is if the longest |
9091 | // sequence of consecutive ones is shifted to the highbits (breaking |
9092 | // ties for which sequence is higher). At the moment we take a liberal |
9093 | // upper bound on this by just popcounting the constant. |
9094 | // TODO: There may be a bitwise trick for it longest/highest |
9095 | // consecutative sequence of ones (naive method is O(Width) loop). |
9096 | Upper = APInt::getHighBitsSet(numBits: Width, hiBitsSet: C->popcount()) + 1; |
9097 | } |
9098 | } else if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
9099 | Upper = APInt::getBitsSetFrom(numBits: Width, loBit: C->getZExtValue()) + 1; |
9100 | } |
9101 | break; |
9102 | |
9103 | case Instruction::SDiv: |
9104 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
9105 | APInt IntMin = APInt::getSignedMinValue(numBits: Width); |
9106 | APInt IntMax = APInt::getSignedMaxValue(numBits: Width); |
9107 | if (C->isAllOnes()) { |
9108 | // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] |
9109 | // where C != -1 and C != 0 and C != 1 |
9110 | Lower = IntMin + 1; |
9111 | Upper = IntMax + 1; |
9112 | } else if (C->countl_zero() < Width - 1) { |
9113 | // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] |
9114 | // where C != -1 and C != 0 and C != 1 |
9115 | Lower = IntMin.sdiv(RHS: *C); |
9116 | Upper = IntMax.sdiv(RHS: *C); |
9117 | if (Lower.sgt(RHS: Upper)) |
9118 | std::swap(a&: Lower, b&: Upper); |
9119 | Upper = Upper + 1; |
9120 | assert(Upper != Lower && "Upper part of range has wrapped!" ); |
9121 | } |
9122 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9123 | if (C->isMinSignedValue()) { |
9124 | // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. |
9125 | Lower = *C; |
9126 | Upper = Lower.lshr(shiftAmt: 1) + 1; |
9127 | } else { |
9128 | // 'sdiv C, x' produces [-|C|, |C|]. |
9129 | Upper = C->abs() + 1; |
9130 | Lower = (-Upper) + 1; |
9131 | } |
9132 | } |
9133 | break; |
9134 | |
9135 | case Instruction::UDiv: |
9136 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero()) { |
9137 | // 'udiv x, C' produces [0, UINT_MAX / C]. |
9138 | Upper = APInt::getMaxValue(numBits: Width).udiv(RHS: *C) + 1; |
9139 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9140 | // 'udiv C, x' produces [0, C]. |
9141 | Upper = *C + 1; |
9142 | } |
9143 | break; |
9144 | |
9145 | case Instruction::SRem: |
9146 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
9147 | // 'srem x, C' produces (-|C|, |C|). |
9148 | Upper = C->abs(); |
9149 | Lower = (-Upper) + 1; |
9150 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9151 | if (C->isNegative()) { |
9152 | // 'srem -|C|, x' produces [-|C|, 0]. |
9153 | Upper = 1; |
9154 | Lower = *C; |
9155 | } else { |
9156 | // 'srem |C|, x' produces [0, |C|]. |
9157 | Upper = *C + 1; |
9158 | } |
9159 | } |
9160 | break; |
9161 | |
9162 | case Instruction::URem: |
9163 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9164 | // 'urem x, C' produces [0, C). |
9165 | Upper = *C; |
9166 | else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) |
9167 | // 'urem C, x' produces [0, C]. |
9168 | Upper = *C + 1; |
9169 | break; |
9170 | |
9171 | default: |
9172 | break; |
9173 | } |
9174 | } |
9175 | |
9176 | static ConstantRange getRangeForIntrinsic(const IntrinsicInst &II) { |
9177 | unsigned Width = II.getType()->getScalarSizeInBits(); |
9178 | const APInt *C; |
9179 | switch (II.getIntrinsicID()) { |
9180 | case Intrinsic::ctpop: |
9181 | case Intrinsic::ctlz: |
9182 | case Intrinsic::cttz: |
9183 | // Maximum of set/clear bits is the bit width. |
9184 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
9185 | Upper: APInt(Width, Width + 1)); |
9186 | case Intrinsic::uadd_sat: |
9187 | // uadd.sat(x, C) produces [C, UINT_MAX]. |
9188 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) || |
9189 | match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9190 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: Width)); |
9191 | break; |
9192 | case Intrinsic::sadd_sat: |
9193 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) || |
9194 | match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
9195 | if (C->isNegative()) |
9196 | // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. |
9197 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
9198 | Upper: APInt::getSignedMaxValue(numBits: Width) + *C + |
9199 | 1); |
9200 | |
9201 | // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. |
9202 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width) + *C, |
9203 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
9204 | } |
9205 | break; |
9206 | case Intrinsic::usub_sat: |
9207 | // usub.sat(C, x) produces [0, C]. |
9208 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) |
9209 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), Upper: *C + 1); |
9210 | |
9211 | // usub.sat(x, C) produces [0, UINT_MAX - C]. |
9212 | if (match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9213 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
9214 | Upper: APInt::getMaxValue(numBits: Width) - *C + 1); |
9215 | break; |
9216 | case Intrinsic::ssub_sat: |
9217 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
9218 | if (C->isNegative()) |
9219 | // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. |
9220 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
9221 | Upper: *C - APInt::getSignedMinValue(numBits: Width) + |
9222 | 1); |
9223 | |
9224 | // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. |
9225 | return ConstantRange::getNonEmpty(Lower: *C - APInt::getSignedMaxValue(numBits: Width), |
9226 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
9227 | } else if (match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
9228 | if (C->isNegative()) |
9229 | // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: |
9230 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width) - *C, |
9231 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
9232 | |
9233 | // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. |
9234 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
9235 | Upper: APInt::getSignedMaxValue(numBits: Width) - *C + |
9236 | 1); |
9237 | } |
9238 | break; |
9239 | case Intrinsic::umin: |
9240 | case Intrinsic::umax: |
9241 | case Intrinsic::smin: |
9242 | case Intrinsic::smax: |
9243 | if (!match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) && |
9244 | !match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
9245 | break; |
9246 | |
9247 | switch (II.getIntrinsicID()) { |
9248 | case Intrinsic::umin: |
9249 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), Upper: *C + 1); |
9250 | case Intrinsic::umax: |
9251 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: Width)); |
9252 | case Intrinsic::smin: |
9253 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
9254 | Upper: *C + 1); |
9255 | case Intrinsic::smax: |
9256 | return ConstantRange::getNonEmpty(Lower: *C, |
9257 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
9258 | default: |
9259 | llvm_unreachable("Must be min/max intrinsic" ); |
9260 | } |
9261 | break; |
9262 | case Intrinsic::abs: |
9263 | // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], |
9264 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. |
9265 | if (match(V: II.getOperand(i_nocapture: 1), P: m_One())) |
9266 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
9267 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
9268 | |
9269 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
9270 | Upper: APInt::getSignedMinValue(numBits: Width) + 1); |
9271 | case Intrinsic::vscale: |
9272 | if (!II.getParent() || !II.getFunction()) |
9273 | break; |
9274 | return getVScaleRange(F: II.getFunction(), BitWidth: Width); |
9275 | default: |
9276 | break; |
9277 | } |
9278 | |
9279 | return ConstantRange::getFull(BitWidth: Width); |
9280 | } |
9281 | |
9282 | static ConstantRange getRangeForSelectPattern(const SelectInst &SI, |
9283 | const InstrInfoQuery &IIQ) { |
9284 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); |
9285 | const Value *LHS = nullptr, *RHS = nullptr; |
9286 | SelectPatternResult R = matchSelectPattern(V: &SI, LHS, RHS); |
9287 | if (R.Flavor == SPF_UNKNOWN) |
9288 | return ConstantRange::getFull(BitWidth); |
9289 | |
9290 | if (R.Flavor == SelectPatternFlavor::SPF_ABS) { |
9291 | // If the negation part of the abs (in RHS) has the NSW flag, |
9292 | // then the result of abs(X) is [0..SIGNED_MAX], |
9293 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. |
9294 | if (match(V: RHS, P: m_Neg(V: m_Specific(V: LHS))) && |
9295 | IIQ.hasNoSignedWrap(Op: cast<Instruction>(Val: RHS))) |
9296 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), |
9297 | Upper: APInt::getSignedMaxValue(numBits: BitWidth) + 1); |
9298 | |
9299 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), |
9300 | Upper: APInt::getSignedMinValue(numBits: BitWidth) + 1); |
9301 | } |
9302 | |
9303 | if (R.Flavor == SelectPatternFlavor::SPF_NABS) { |
9304 | // The result of -abs(X) is <= 0. |
9305 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: BitWidth), |
9306 | Upper: APInt(BitWidth, 1)); |
9307 | } |
9308 | |
9309 | const APInt *C; |
9310 | if (!match(V: LHS, P: m_APInt(Res&: C)) && !match(V: RHS, P: m_APInt(Res&: C))) |
9311 | return ConstantRange::getFull(BitWidth); |
9312 | |
9313 | switch (R.Flavor) { |
9314 | case SPF_UMIN: |
9315 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), Upper: *C + 1); |
9316 | case SPF_UMAX: |
9317 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: BitWidth)); |
9318 | case SPF_SMIN: |
9319 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: BitWidth), |
9320 | Upper: *C + 1); |
9321 | case SPF_SMAX: |
9322 | return ConstantRange::getNonEmpty(Lower: *C, |
9323 | Upper: APInt::getSignedMaxValue(numBits: BitWidth) + 1); |
9324 | default: |
9325 | return ConstantRange::getFull(BitWidth); |
9326 | } |
9327 | } |
9328 | |
9329 | static void setLimitForFPToI(const Instruction *I, APInt &Lower, APInt &Upper) { |
9330 | // The maximum representable value of a half is 65504. For floats the maximum |
9331 | // value is 3.4e38 which requires roughly 129 bits. |
9332 | unsigned BitWidth = I->getType()->getScalarSizeInBits(); |
9333 | if (!I->getOperand(i: 0)->getType()->getScalarType()->isHalfTy()) |
9334 | return; |
9335 | if (isa<FPToSIInst>(Val: I) && BitWidth >= 17) { |
9336 | Lower = APInt(BitWidth, -65504); |
9337 | Upper = APInt(BitWidth, 65505); |
9338 | } |
9339 | |
9340 | if (isa<FPToUIInst>(Val: I) && BitWidth >= 16) { |
9341 | // For a fptoui the lower limit is left as 0. |
9342 | Upper = APInt(BitWidth, 65505); |
9343 | } |
9344 | } |
9345 | |
9346 | ConstantRange llvm::computeConstantRange(const Value *V, bool ForSigned, |
9347 | bool UseInstrInfo, AssumptionCache *AC, |
9348 | const Instruction *CtxI, |
9349 | const DominatorTree *DT, |
9350 | unsigned Depth) { |
9351 | assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction" ); |
9352 | |
9353 | if (Depth == MaxAnalysisRecursionDepth) |
9354 | return ConstantRange::getFull(BitWidth: V->getType()->getScalarSizeInBits()); |
9355 | |
9356 | const APInt *C; |
9357 | if (match(V, P: m_APInt(Res&: C))) |
9358 | return ConstantRange(*C); |
9359 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); |
9360 | |
9361 | if (auto *VC = dyn_cast<ConstantDataVector>(Val: V)) { |
9362 | ConstantRange CR = ConstantRange::getEmpty(BitWidth); |
9363 | for (unsigned ElemIdx = 0, NElem = VC->getNumElements(); ElemIdx < NElem; |
9364 | ++ElemIdx) |
9365 | CR = CR.unionWith(CR: VC->getElementAsAPInt(i: ElemIdx)); |
9366 | return CR; |
9367 | } |
9368 | |
9369 | InstrInfoQuery IIQ(UseInstrInfo); |
9370 | ConstantRange CR = ConstantRange::getFull(BitWidth); |
9371 | if (auto *BO = dyn_cast<BinaryOperator>(Val: V)) { |
9372 | APInt Lower = APInt(BitWidth, 0); |
9373 | APInt Upper = APInt(BitWidth, 0); |
9374 | // TODO: Return ConstantRange. |
9375 | setLimitsForBinOp(BO: *BO, Lower, Upper, IIQ, PreferSignedRange: ForSigned); |
9376 | CR = ConstantRange::getNonEmpty(Lower, Upper); |
9377 | } else if (auto *II = dyn_cast<IntrinsicInst>(Val: V)) |
9378 | CR = getRangeForIntrinsic(II: *II); |
9379 | else if (auto *SI = dyn_cast<SelectInst>(Val: V)) { |
9380 | ConstantRange CRTrue = computeConstantRange( |
9381 | V: SI->getTrueValue(), ForSigned, UseInstrInfo, AC, CtxI, DT, Depth: Depth + 1); |
9382 | ConstantRange CRFalse = computeConstantRange( |
9383 | V: SI->getFalseValue(), ForSigned, UseInstrInfo, AC, CtxI, DT, Depth: Depth + 1); |
9384 | CR = CRTrue.unionWith(CR: CRFalse); |
9385 | CR = CR.intersectWith(CR: getRangeForSelectPattern(SI: *SI, IIQ)); |
9386 | } else if (isa<FPToUIInst>(Val: V) || isa<FPToSIInst>(Val: V)) { |
9387 | APInt Lower = APInt(BitWidth, 0); |
9388 | APInt Upper = APInt(BitWidth, 0); |
9389 | // TODO: Return ConstantRange. |
9390 | setLimitForFPToI(I: cast<Instruction>(Val: V), Lower, Upper); |
9391 | CR = ConstantRange::getNonEmpty(Lower, Upper); |
9392 | } else if (const auto *A = dyn_cast<Argument>(Val: V)) |
9393 | if (std::optional<ConstantRange> Range = A->getRange()) |
9394 | CR = *Range; |
9395 | |
9396 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
9397 | if (auto *Range = IIQ.getMetadata(I, KindID: LLVMContext::MD_range)) |
9398 | CR = CR.intersectWith(CR: getConstantRangeFromMetadata(RangeMD: *Range)); |
9399 | |
9400 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) |
9401 | if (std::optional<ConstantRange> Range = CB->getRange()) |
9402 | CR = CR.intersectWith(CR: *Range); |
9403 | } |
9404 | |
9405 | if (CtxI && AC) { |
9406 | // Try to restrict the range based on information from assumptions. |
9407 | for (auto &AssumeVH : AC->assumptionsFor(V)) { |
9408 | if (!AssumeVH) |
9409 | continue; |
9410 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
9411 | assert(I->getParent()->getParent() == CtxI->getParent()->getParent() && |
9412 | "Got assumption for the wrong function!" ); |
9413 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && |
9414 | "must be an assume intrinsic" ); |
9415 | |
9416 | if (!isValidAssumeForContext(Inv: I, CxtI: CtxI, DT)) |
9417 | continue; |
9418 | Value *Arg = I->getArgOperand(i: 0); |
9419 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Arg); |
9420 | // Currently we just use information from comparisons. |
9421 | if (!Cmp || Cmp->getOperand(i_nocapture: 0) != V) |
9422 | continue; |
9423 | // TODO: Set "ForSigned" parameter via Cmp->isSigned()? |
9424 | ConstantRange RHS = |
9425 | computeConstantRange(V: Cmp->getOperand(i_nocapture: 1), /* ForSigned */ false, |
9426 | UseInstrInfo, AC, CtxI: I, DT, Depth: Depth + 1); |
9427 | CR = CR.intersectWith( |
9428 | CR: ConstantRange::makeAllowedICmpRegion(Pred: Cmp->getPredicate(), Other: RHS)); |
9429 | } |
9430 | } |
9431 | |
9432 | return CR; |
9433 | } |
9434 | |
9435 | static void |
9436 | addValueAffectedByCondition(Value *V, |
9437 | function_ref<void(Value *)> InsertAffected) { |
9438 | assert(V != nullptr); |
9439 | if (isa<Argument>(Val: V) || isa<GlobalValue>(Val: V)) { |
9440 | InsertAffected(V); |
9441 | } else if (auto *I = dyn_cast<Instruction>(Val: V)) { |
9442 | InsertAffected(V); |
9443 | |
9444 | // Peek through unary operators to find the source of the condition. |
9445 | Value *Op; |
9446 | if (match(V: I, P: m_CombineOr(L: m_PtrToInt(Op: m_Value(V&: Op)), R: m_Trunc(Op: m_Value(V&: Op))))) { |
9447 | if (isa<Instruction>(Val: Op) || isa<Argument>(Val: Op)) |
9448 | InsertAffected(Op); |
9449 | } |
9450 | } |
9451 | } |
9452 | |
9453 | void llvm::findValuesAffectedByCondition( |
9454 | Value *Cond, bool IsAssume, function_ref<void(Value *)> InsertAffected) { |
9455 | auto AddAffected = [&InsertAffected](Value *V) { |
9456 | addValueAffectedByCondition(V, InsertAffected); |
9457 | }; |
9458 | |
9459 | auto AddCmpOperands = [&AddAffected, IsAssume](Value *LHS, Value *RHS) { |
9460 | if (IsAssume) { |
9461 | AddAffected(LHS); |
9462 | AddAffected(RHS); |
9463 | } else if (match(V: RHS, P: m_Constant())) |
9464 | AddAffected(LHS); |
9465 | }; |
9466 | |
9467 | SmallVector<Value *, 8> Worklist; |
9468 | SmallPtrSet<Value *, 8> Visited; |
9469 | Worklist.push_back(Elt: Cond); |
9470 | while (!Worklist.empty()) { |
9471 | Value *V = Worklist.pop_back_val(); |
9472 | if (!Visited.insert(Ptr: V).second) |
9473 | continue; |
9474 | |
9475 | CmpInst::Predicate Pred; |
9476 | Value *A, *B, *X; |
9477 | |
9478 | if (IsAssume) { |
9479 | AddAffected(V); |
9480 | if (match(V, P: m_Not(V: m_Value(V&: X)))) |
9481 | AddAffected(X); |
9482 | } |
9483 | |
9484 | if (match(V, P: m_LogicalOp(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
9485 | // assume(A && B) is split to -> assume(A); assume(B); |
9486 | // assume(!(A || B)) is split to -> assume(!A); assume(!B); |
9487 | // Finally, assume(A || B) / assume(!(A && B)) generally don't provide |
9488 | // enough information to be worth handling (intersection of information as |
9489 | // opposed to union). |
9490 | if (!IsAssume) { |
9491 | Worklist.push_back(Elt: A); |
9492 | Worklist.push_back(Elt: B); |
9493 | } |
9494 | } else if (match(V, P: m_ICmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
9495 | AddCmpOperands(A, B); |
9496 | |
9497 | if (ICmpInst::isEquality(P: Pred)) { |
9498 | if (match(V: B, P: m_ConstantInt())) { |
9499 | Value *Y; |
9500 | // (X & C) or (X | C) or (X ^ C). |
9501 | // (X << C) or (X >>_s C) or (X >>_u C). |
9502 | if (match(V: A, P: m_BitwiseLogic(L: m_Value(V&: X), R: m_ConstantInt())) || |
9503 | match(V: A, P: m_Shift(L: m_Value(V&: X), R: m_ConstantInt()))) |
9504 | AddAffected(X); |
9505 | else if (match(V: A, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) || |
9506 | match(V: A, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
9507 | AddAffected(X); |
9508 | AddAffected(Y); |
9509 | } |
9510 | } |
9511 | } else { |
9512 | if (match(V: B, P: m_ConstantInt())) { |
9513 | // Handle (A + C1) u< C2, which is the canonical form of |
9514 | // A > C3 && A < C4. |
9515 | if (match(V: A, P: m_AddLike(L: m_Value(V&: X), R: m_ConstantInt()))) |
9516 | AddAffected(X); |
9517 | |
9518 | Value *Y; |
9519 | // X & Y u> C -> X >u C && Y >u C |
9520 | // X | Y u< C -> X u< C && Y u< C |
9521 | if (ICmpInst::isUnsigned(predicate: Pred) && |
9522 | (match(V: A, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) || |
9523 | match(V: A, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))))) { |
9524 | AddAffected(X); |
9525 | AddAffected(Y); |
9526 | } |
9527 | } |
9528 | |
9529 | // Handle icmp slt/sgt (bitcast X to int), 0/-1, which is supported |
9530 | // by computeKnownFPClass(). |
9531 | if (match(V: A, P: m_ElementWiseBitCast(Op: m_Value(V&: X)))) { |
9532 | if (Pred == ICmpInst::ICMP_SLT && match(V: B, P: m_Zero())) |
9533 | InsertAffected(X); |
9534 | else if (Pred == ICmpInst::ICMP_SGT && match(V: B, P: m_AllOnes())) |
9535 | InsertAffected(X); |
9536 | } |
9537 | } |
9538 | } else if (match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
9539 | AddCmpOperands(A, B); |
9540 | |
9541 | // fcmp fneg(x), y |
9542 | // fcmp fabs(x), y |
9543 | // fcmp fneg(fabs(x)), y |
9544 | if (match(V: A, P: m_FNeg(X: m_Value(V&: A)))) |
9545 | AddAffected(A); |
9546 | if (match(V: A, P: m_FAbs(Op0: m_Value(V&: A)))) |
9547 | AddAffected(A); |
9548 | |
9549 | } else if (match(V, m_Intrinsic<Intrinsic::is_fpclass>(m_Value(A), |
9550 | m_Value()))) { |
9551 | // Handle patterns that computeKnownFPClass() support. |
9552 | AddAffected(A); |
9553 | } |
9554 | } |
9555 | } |
9556 | |