1 | //===- ConstantFold.cpp - LLVM constant folder ----------------------------===// |
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 implements folding of constants for LLVM. This implements the |
10 | // (internal) ConstantFold.h interface, which is used by the |
11 | // ConstantExpr::get* methods to automatically fold constants when possible. |
12 | // |
13 | // The current constant folding implementation is implemented in two pieces: the |
14 | // pieces that don't need DataLayout, and the pieces that do. This is to avoid |
15 | // a dependence in IR on Target. |
16 | // |
17 | //===----------------------------------------------------------------------===// |
18 | |
19 | #include "llvm/IR/ConstantFold.h" |
20 | #include "llvm/ADT/APSInt.h" |
21 | #include "llvm/ADT/SmallVector.h" |
22 | #include "llvm/IR/Constants.h" |
23 | #include "llvm/IR/DerivedTypes.h" |
24 | #include "llvm/IR/Function.h" |
25 | #include "llvm/IR/GlobalAlias.h" |
26 | #include "llvm/IR/GlobalVariable.h" |
27 | #include "llvm/IR/Instructions.h" |
28 | #include "llvm/IR/Module.h" |
29 | #include "llvm/IR/Operator.h" |
30 | #include "llvm/IR/PatternMatch.h" |
31 | #include "llvm/Support/ErrorHandling.h" |
32 | using namespace llvm; |
33 | using namespace llvm::PatternMatch; |
34 | |
35 | //===----------------------------------------------------------------------===// |
36 | // ConstantFold*Instruction Implementations |
37 | //===----------------------------------------------------------------------===// |
38 | |
39 | /// This function determines which opcode to use to fold two constant cast |
40 | /// expressions together. It uses CastInst::isEliminableCastPair to determine |
41 | /// the opcode. Consequently its just a wrapper around that function. |
42 | /// Determine if it is valid to fold a cast of a cast |
43 | static unsigned |
44 | foldConstantCastPair( |
45 | unsigned opc, ///< opcode of the second cast constant expression |
46 | ConstantExpr *Op, ///< the first cast constant expression |
47 | Type *DstTy ///< destination type of the first cast |
48 | ) { |
49 | assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!" ); |
50 | assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type" ); |
51 | assert(CastInst::isCast(opc) && "Invalid cast opcode" ); |
52 | |
53 | // The types and opcodes for the two Cast constant expressions |
54 | Type *SrcTy = Op->getOperand(i_nocapture: 0)->getType(); |
55 | Type *MidTy = Op->getType(); |
56 | Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode()); |
57 | Instruction::CastOps secondOp = Instruction::CastOps(opc); |
58 | |
59 | // Assume that pointers are never more than 64 bits wide, and only use this |
60 | // for the middle type. Otherwise we could end up folding away illegal |
61 | // bitcasts between address spaces with different sizes. |
62 | IntegerType *FakeIntPtrTy = Type::getInt64Ty(C&: DstTy->getContext()); |
63 | |
64 | // Let CastInst::isEliminableCastPair do the heavy lifting. |
65 | return CastInst::isEliminableCastPair(firstOpcode: firstOp, secondOpcode: secondOp, SrcTy, MidTy, DstTy, |
66 | SrcIntPtrTy: nullptr, MidIntPtrTy: FakeIntPtrTy, DstIntPtrTy: nullptr); |
67 | } |
68 | |
69 | static Constant *FoldBitCast(Constant *V, Type *DestTy) { |
70 | Type *SrcTy = V->getType(); |
71 | if (SrcTy == DestTy) |
72 | return V; // no-op cast |
73 | |
74 | if (V->isAllOnesValue()) |
75 | return Constant::getAllOnesValue(Ty: DestTy); |
76 | |
77 | // Handle ConstantInt -> ConstantFP |
78 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) { |
79 | // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts |
80 | // This allows for other simplifications (although some of them |
81 | // can only be handled by Analysis/ConstantFolding.cpp). |
82 | if (isa<VectorType>(Val: DestTy) && !isa<VectorType>(Val: SrcTy)) |
83 | return ConstantExpr::getBitCast(C: ConstantVector::get(V), Ty: DestTy); |
84 | |
85 | // Make sure dest type is compatible with the folded fp constant. |
86 | // See note below regarding the PPC_FP128 restriction. |
87 | if (!DestTy->isFPOrFPVectorTy() || DestTy->isPPC_FP128Ty() || |
88 | DestTy->getScalarSizeInBits() != SrcTy->getScalarSizeInBits()) |
89 | return nullptr; |
90 | |
91 | return ConstantFP::get( |
92 | Ty: DestTy, |
93 | V: APFloat(DestTy->getScalarType()->getFltSemantics(), CI->getValue())); |
94 | } |
95 | |
96 | // Handle ConstantFP -> ConstantInt |
97 | if (ConstantFP *FP = dyn_cast<ConstantFP>(Val: V)) { |
98 | // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts |
99 | // This allows for other simplifications (although some of them |
100 | // can only be handled by Analysis/ConstantFolding.cpp). |
101 | if (isa<VectorType>(Val: DestTy) && !isa<VectorType>(Val: SrcTy)) |
102 | return ConstantExpr::getBitCast(C: ConstantVector::get(V), Ty: DestTy); |
103 | |
104 | // PPC_FP128 is really the sum of two consecutive doubles, where the first |
105 | // double is always stored first in memory, regardless of the target |
106 | // endianness. The memory layout of i128, however, depends on the target |
107 | // endianness, and so we can't fold this without target endianness |
108 | // information. This should instead be handled by |
109 | // Analysis/ConstantFolding.cpp |
110 | if (SrcTy->isPPC_FP128Ty()) |
111 | return nullptr; |
112 | |
113 | // Make sure dest type is compatible with the folded integer constant. |
114 | if (!DestTy->isIntOrIntVectorTy() || |
115 | DestTy->getScalarSizeInBits() != SrcTy->getScalarSizeInBits()) |
116 | return nullptr; |
117 | |
118 | return ConstantInt::get(Ty: DestTy, V: FP->getValueAPF().bitcastToAPInt()); |
119 | } |
120 | |
121 | return nullptr; |
122 | } |
123 | |
124 | static Constant *foldMaybeUndesirableCast(unsigned opc, Constant *V, |
125 | Type *DestTy) { |
126 | return ConstantExpr::isDesirableCastOp(Opcode: opc) |
127 | ? ConstantExpr::getCast(ops: opc, C: V, Ty: DestTy) |
128 | : ConstantFoldCastInstruction(opcode: opc, V, DestTy); |
129 | } |
130 | |
131 | Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V, |
132 | Type *DestTy) { |
133 | if (isa<PoisonValue>(Val: V)) |
134 | return PoisonValue::get(T: DestTy); |
135 | |
136 | if (isa<UndefValue>(Val: V)) { |
137 | // zext(undef) = 0, because the top bits will be zero. |
138 | // sext(undef) = 0, because the top bits will all be the same. |
139 | // [us]itofp(undef) = 0, because the result value is bounded. |
140 | if (opc == Instruction::ZExt || opc == Instruction::SExt || |
141 | opc == Instruction::UIToFP || opc == Instruction::SIToFP) |
142 | return Constant::getNullValue(Ty: DestTy); |
143 | return UndefValue::get(T: DestTy); |
144 | } |
145 | |
146 | if (V->isNullValue() && !DestTy->isX86_AMXTy() && |
147 | opc != Instruction::AddrSpaceCast) |
148 | return Constant::getNullValue(Ty: DestTy); |
149 | |
150 | // If the cast operand is a constant expression, there's a few things we can |
151 | // do to try to simplify it. |
152 | if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: V)) { |
153 | if (CE->isCast()) { |
154 | // Try hard to fold cast of cast because they are often eliminable. |
155 | if (unsigned newOpc = foldConstantCastPair(opc, Op: CE, DstTy: DestTy)) |
156 | return foldMaybeUndesirableCast(opc: newOpc, V: CE->getOperand(i_nocapture: 0), DestTy); |
157 | } |
158 | } |
159 | |
160 | // If the cast operand is a constant vector, perform the cast by |
161 | // operating on each element. In the cast of bitcasts, the element |
162 | // count may be mismatched; don't attempt to handle that here. |
163 | if (DestTy->isVectorTy() && V->getType()->isVectorTy() && |
164 | cast<VectorType>(Val: DestTy)->getElementCount() == |
165 | cast<VectorType>(Val: V->getType())->getElementCount()) { |
166 | VectorType *DestVecTy = cast<VectorType>(Val: DestTy); |
167 | Type *DstEltTy = DestVecTy->getElementType(); |
168 | // Fast path for splatted constants. |
169 | if (Constant *Splat = V->getSplatValue()) { |
170 | Constant *Res = foldMaybeUndesirableCast(opc, V: Splat, DestTy: DstEltTy); |
171 | if (!Res) |
172 | return nullptr; |
173 | return ConstantVector::getSplat( |
174 | EC: cast<VectorType>(Val: DestTy)->getElementCount(), Elt: Res); |
175 | } |
176 | if (isa<ScalableVectorType>(Val: DestTy)) |
177 | return nullptr; |
178 | SmallVector<Constant *, 16> res; |
179 | Type *Ty = IntegerType::get(C&: V->getContext(), NumBits: 32); |
180 | for (unsigned i = 0, |
181 | e = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
182 | i != e; ++i) { |
183 | Constant *C = ConstantExpr::getExtractElement(Vec: V, Idx: ConstantInt::get(Ty, V: i)); |
184 | Constant *Casted = foldMaybeUndesirableCast(opc, V: C, DestTy: DstEltTy); |
185 | if (!Casted) |
186 | return nullptr; |
187 | res.push_back(Elt: Casted); |
188 | } |
189 | return ConstantVector::get(V: res); |
190 | } |
191 | |
192 | // We actually have to do a cast now. Perform the cast according to the |
193 | // opcode specified. |
194 | switch (opc) { |
195 | default: |
196 | llvm_unreachable("Failed to cast constant expression" ); |
197 | case Instruction::FPTrunc: |
198 | case Instruction::FPExt: |
199 | if (ConstantFP *FPC = dyn_cast<ConstantFP>(Val: V)) { |
200 | bool ignored; |
201 | APFloat Val = FPC->getValueAPF(); |
202 | Val.convert(ToSemantics: DestTy->getScalarType()->getFltSemantics(), |
203 | RM: APFloat::rmNearestTiesToEven, losesInfo: &ignored); |
204 | return ConstantFP::get(Ty: DestTy, V: Val); |
205 | } |
206 | return nullptr; // Can't fold. |
207 | case Instruction::FPToUI: |
208 | case Instruction::FPToSI: |
209 | if (ConstantFP *FPC = dyn_cast<ConstantFP>(Val: V)) { |
210 | const APFloat &V = FPC->getValueAPF(); |
211 | bool ignored; |
212 | APSInt IntVal(DestTy->getScalarSizeInBits(), opc == Instruction::FPToUI); |
213 | if (APFloat::opInvalidOp == |
214 | V.convertToInteger(Result&: IntVal, RM: APFloat::rmTowardZero, IsExact: &ignored)) { |
215 | // Undefined behavior invoked - the destination type can't represent |
216 | // the input constant. |
217 | return PoisonValue::get(T: DestTy); |
218 | } |
219 | return ConstantInt::get(Ty: DestTy, V: IntVal); |
220 | } |
221 | return nullptr; // Can't fold. |
222 | case Instruction::UIToFP: |
223 | case Instruction::SIToFP: |
224 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) { |
225 | const APInt &api = CI->getValue(); |
226 | APFloat apf(DestTy->getScalarType()->getFltSemantics(), |
227 | APInt::getZero(numBits: DestTy->getScalarSizeInBits())); |
228 | apf.convertFromAPInt(Input: api, IsSigned: opc==Instruction::SIToFP, |
229 | RM: APFloat::rmNearestTiesToEven); |
230 | return ConstantFP::get(Ty: DestTy, V: apf); |
231 | } |
232 | return nullptr; |
233 | case Instruction::ZExt: |
234 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) { |
235 | uint32_t BitWidth = DestTy->getScalarSizeInBits(); |
236 | return ConstantInt::get(Ty: DestTy, V: CI->getValue().zext(width: BitWidth)); |
237 | } |
238 | return nullptr; |
239 | case Instruction::SExt: |
240 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) { |
241 | uint32_t BitWidth = DestTy->getScalarSizeInBits(); |
242 | return ConstantInt::get(Ty: DestTy, V: CI->getValue().sext(width: BitWidth)); |
243 | } |
244 | return nullptr; |
245 | case Instruction::Trunc: { |
246 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) { |
247 | uint32_t BitWidth = DestTy->getScalarSizeInBits(); |
248 | return ConstantInt::get(Ty: DestTy, V: CI->getValue().trunc(width: BitWidth)); |
249 | } |
250 | |
251 | return nullptr; |
252 | } |
253 | case Instruction::BitCast: |
254 | return FoldBitCast(V, DestTy); |
255 | case Instruction::AddrSpaceCast: |
256 | case Instruction::IntToPtr: |
257 | case Instruction::PtrToInt: |
258 | return nullptr; |
259 | } |
260 | } |
261 | |
262 | Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond, |
263 | Constant *V1, Constant *V2) { |
264 | // Check for i1 and vector true/false conditions. |
265 | if (Cond->isNullValue()) return V2; |
266 | if (Cond->isAllOnesValue()) return V1; |
267 | |
268 | // If the condition is a vector constant, fold the result elementwise. |
269 | if (ConstantVector *CondV = dyn_cast<ConstantVector>(Val: Cond)) { |
270 | auto *V1VTy = CondV->getType(); |
271 | SmallVector<Constant*, 16> Result; |
272 | Type *Ty = IntegerType::get(C&: CondV->getContext(), NumBits: 32); |
273 | for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) { |
274 | Constant *V; |
275 | Constant *V1Element = ConstantExpr::getExtractElement(Vec: V1, |
276 | Idx: ConstantInt::get(Ty, V: i)); |
277 | Constant *V2Element = ConstantExpr::getExtractElement(Vec: V2, |
278 | Idx: ConstantInt::get(Ty, V: i)); |
279 | auto *Cond = cast<Constant>(Val: CondV->getOperand(i_nocapture: i)); |
280 | if (isa<PoisonValue>(Val: Cond)) { |
281 | V = PoisonValue::get(T: V1Element->getType()); |
282 | } else if (V1Element == V2Element) { |
283 | V = V1Element; |
284 | } else if (isa<UndefValue>(Val: Cond)) { |
285 | V = isa<UndefValue>(Val: V1Element) ? V1Element : V2Element; |
286 | } else { |
287 | if (!isa<ConstantInt>(Val: Cond)) break; |
288 | V = Cond->isNullValue() ? V2Element : V1Element; |
289 | } |
290 | Result.push_back(Elt: V); |
291 | } |
292 | |
293 | // If we were able to build the vector, return it. |
294 | if (Result.size() == V1VTy->getNumElements()) |
295 | return ConstantVector::get(V: Result); |
296 | } |
297 | |
298 | if (isa<PoisonValue>(Val: Cond)) |
299 | return PoisonValue::get(T: V1->getType()); |
300 | |
301 | if (isa<UndefValue>(Val: Cond)) { |
302 | if (isa<UndefValue>(Val: V1)) return V1; |
303 | return V2; |
304 | } |
305 | |
306 | if (V1 == V2) return V1; |
307 | |
308 | if (isa<PoisonValue>(Val: V1)) |
309 | return V2; |
310 | if (isa<PoisonValue>(Val: V2)) |
311 | return V1; |
312 | |
313 | // If the true or false value is undef, we can fold to the other value as |
314 | // long as the other value isn't poison. |
315 | auto NotPoison = [](Constant *C) { |
316 | if (isa<PoisonValue>(Val: C)) |
317 | return false; |
318 | |
319 | // TODO: We can analyze ConstExpr by opcode to determine if there is any |
320 | // possibility of poison. |
321 | if (isa<ConstantExpr>(Val: C)) |
322 | return false; |
323 | |
324 | if (isa<ConstantInt>(Val: C) || isa<GlobalVariable>(Val: C) || isa<ConstantFP>(Val: C) || |
325 | isa<ConstantPointerNull>(Val: C) || isa<Function>(Val: C)) |
326 | return true; |
327 | |
328 | if (C->getType()->isVectorTy()) |
329 | return !C->containsPoisonElement() && !C->containsConstantExpression(); |
330 | |
331 | // TODO: Recursively analyze aggregates or other constants. |
332 | return false; |
333 | }; |
334 | if (isa<UndefValue>(Val: V1) && NotPoison(V2)) return V2; |
335 | if (isa<UndefValue>(Val: V2) && NotPoison(V1)) return V1; |
336 | |
337 | return nullptr; |
338 | } |
339 | |
340 | Constant *llvm::(Constant *Val, |
341 | Constant *Idx) { |
342 | auto *ValVTy = cast<VectorType>(Val: Val->getType()); |
343 | |
344 | // extractelt poison, C -> poison |
345 | // extractelt C, undef -> poison |
346 | if (isa<PoisonValue>(Val) || isa<UndefValue>(Val: Idx)) |
347 | return PoisonValue::get(T: ValVTy->getElementType()); |
348 | |
349 | // extractelt undef, C -> undef |
350 | if (isa<UndefValue>(Val)) |
351 | return UndefValue::get(T: ValVTy->getElementType()); |
352 | |
353 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
354 | if (!CIdx) |
355 | return nullptr; |
356 | |
357 | if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val: Val->getType())) { |
358 | // ee({w,x,y,z}, wrong_value) -> poison |
359 | if (CIdx->uge(Num: ValFVTy->getNumElements())) |
360 | return PoisonValue::get(T: ValFVTy->getElementType()); |
361 | } |
362 | |
363 | // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...) |
364 | if (auto *CE = dyn_cast<ConstantExpr>(Val)) { |
365 | if (auto *GEP = dyn_cast<GEPOperator>(Val: CE)) { |
366 | SmallVector<Constant *, 8> Ops; |
367 | Ops.reserve(N: CE->getNumOperands()); |
368 | for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) { |
369 | Constant *Op = CE->getOperand(i_nocapture: i); |
370 | if (Op->getType()->isVectorTy()) { |
371 | Constant *ScalarOp = ConstantExpr::getExtractElement(Vec: Op, Idx); |
372 | if (!ScalarOp) |
373 | return nullptr; |
374 | Ops.push_back(Elt: ScalarOp); |
375 | } else |
376 | Ops.push_back(Elt: Op); |
377 | } |
378 | return CE->getWithOperands(Ops, Ty: ValVTy->getElementType(), OnlyIfReduced: false, |
379 | SrcTy: GEP->getSourceElementType()); |
380 | } else if (CE->getOpcode() == Instruction::InsertElement) { |
381 | if (const auto *IEIdx = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 2))) { |
382 | if (APSInt::isSameValue(I1: APSInt(IEIdx->getValue()), |
383 | I2: APSInt(CIdx->getValue()))) { |
384 | return CE->getOperand(i_nocapture: 1); |
385 | } else { |
386 | return ConstantExpr::getExtractElement(Vec: CE->getOperand(i_nocapture: 0), Idx: CIdx); |
387 | } |
388 | } |
389 | } |
390 | } |
391 | |
392 | if (Constant *C = Val->getAggregateElement(Elt: CIdx)) |
393 | return C; |
394 | |
395 | // Lane < Splat minimum vector width => extractelt Splat(x), Lane -> x |
396 | if (CIdx->getValue().ult(RHS: ValVTy->getElementCount().getKnownMinValue())) { |
397 | if (Constant *SplatVal = Val->getSplatValue()) |
398 | return SplatVal; |
399 | } |
400 | |
401 | return nullptr; |
402 | } |
403 | |
404 | Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val, |
405 | Constant *Elt, |
406 | Constant *Idx) { |
407 | if (isa<UndefValue>(Val: Idx)) |
408 | return PoisonValue::get(T: Val->getType()); |
409 | |
410 | // Inserting null into all zeros is still all zeros. |
411 | // TODO: This is true for undef and poison splats too. |
412 | if (isa<ConstantAggregateZero>(Val) && Elt->isNullValue()) |
413 | return Val; |
414 | |
415 | ConstantInt *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
416 | if (!CIdx) return nullptr; |
417 | |
418 | // Do not iterate on scalable vector. The num of elements is unknown at |
419 | // compile-time. |
420 | if (isa<ScalableVectorType>(Val: Val->getType())) |
421 | return nullptr; |
422 | |
423 | auto *ValTy = cast<FixedVectorType>(Val: Val->getType()); |
424 | |
425 | unsigned NumElts = ValTy->getNumElements(); |
426 | if (CIdx->uge(Num: NumElts)) |
427 | return PoisonValue::get(T: Val->getType()); |
428 | |
429 | SmallVector<Constant*, 16> Result; |
430 | Result.reserve(N: NumElts); |
431 | auto *Ty = Type::getInt32Ty(C&: Val->getContext()); |
432 | uint64_t IdxVal = CIdx->getZExtValue(); |
433 | for (unsigned i = 0; i != NumElts; ++i) { |
434 | if (i == IdxVal) { |
435 | Result.push_back(Elt); |
436 | continue; |
437 | } |
438 | |
439 | Constant *C = ConstantExpr::getExtractElement(Vec: Val, Idx: ConstantInt::get(Ty, V: i)); |
440 | Result.push_back(Elt: C); |
441 | } |
442 | |
443 | return ConstantVector::get(V: Result); |
444 | } |
445 | |
446 | Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2, |
447 | ArrayRef<int> Mask) { |
448 | auto *V1VTy = cast<VectorType>(Val: V1->getType()); |
449 | unsigned MaskNumElts = Mask.size(); |
450 | auto MaskEltCount = |
451 | ElementCount::get(MinVal: MaskNumElts, Scalable: isa<ScalableVectorType>(Val: V1VTy)); |
452 | Type *EltTy = V1VTy->getElementType(); |
453 | |
454 | // Poison shuffle mask -> poison value. |
455 | if (all_of(Range&: Mask, P: [](int Elt) { return Elt == PoisonMaskElem; })) { |
456 | return PoisonValue::get(T: VectorType::get(ElementType: EltTy, EC: MaskEltCount)); |
457 | } |
458 | |
459 | // If the mask is all zeros this is a splat, no need to go through all |
460 | // elements. |
461 | if (all_of(Range&: Mask, P: [](int Elt) { return Elt == 0; })) { |
462 | Type *Ty = IntegerType::get(C&: V1->getContext(), NumBits: 32); |
463 | Constant *Elt = |
464 | ConstantExpr::getExtractElement(Vec: V1, Idx: ConstantInt::get(Ty, V: 0)); |
465 | |
466 | // For scalable vectors, make sure this doesn't fold back into a |
467 | // shufflevector. |
468 | if (!MaskEltCount.isScalable() || Elt->isNullValue() || isa<UndefValue>(Val: Elt)) |
469 | return ConstantVector::getSplat(EC: MaskEltCount, Elt); |
470 | } |
471 | |
472 | // Do not iterate on scalable vector. The num of elements is unknown at |
473 | // compile-time. |
474 | if (isa<ScalableVectorType>(Val: V1VTy)) |
475 | return nullptr; |
476 | |
477 | unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue(); |
478 | |
479 | // Loop over the shuffle mask, evaluating each element. |
480 | SmallVector<Constant*, 32> Result; |
481 | for (unsigned i = 0; i != MaskNumElts; ++i) { |
482 | int Elt = Mask[i]; |
483 | if (Elt == -1) { |
484 | Result.push_back(Elt: UndefValue::get(T: EltTy)); |
485 | continue; |
486 | } |
487 | Constant *InElt; |
488 | if (unsigned(Elt) >= SrcNumElts*2) |
489 | InElt = UndefValue::get(T: EltTy); |
490 | else if (unsigned(Elt) >= SrcNumElts) { |
491 | Type *Ty = IntegerType::get(C&: V2->getContext(), NumBits: 32); |
492 | InElt = |
493 | ConstantExpr::getExtractElement(Vec: V2, |
494 | Idx: ConstantInt::get(Ty, V: Elt - SrcNumElts)); |
495 | } else { |
496 | Type *Ty = IntegerType::get(C&: V1->getContext(), NumBits: 32); |
497 | InElt = ConstantExpr::getExtractElement(Vec: V1, Idx: ConstantInt::get(Ty, V: Elt)); |
498 | } |
499 | Result.push_back(Elt: InElt); |
500 | } |
501 | |
502 | return ConstantVector::get(V: Result); |
503 | } |
504 | |
505 | Constant *llvm::(Constant *Agg, |
506 | ArrayRef<unsigned> Idxs) { |
507 | // Base case: no indices, so return the entire value. |
508 | if (Idxs.empty()) |
509 | return Agg; |
510 | |
511 | if (Constant *C = Agg->getAggregateElement(Elt: Idxs[0])) |
512 | return ConstantFoldExtractValueInstruction(Agg: C, Idxs: Idxs.slice(N: 1)); |
513 | |
514 | return nullptr; |
515 | } |
516 | |
517 | Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg, |
518 | Constant *Val, |
519 | ArrayRef<unsigned> Idxs) { |
520 | // Base case: no indices, so replace the entire value. |
521 | if (Idxs.empty()) |
522 | return Val; |
523 | |
524 | unsigned NumElts; |
525 | if (StructType *ST = dyn_cast<StructType>(Val: Agg->getType())) |
526 | NumElts = ST->getNumElements(); |
527 | else |
528 | NumElts = cast<ArrayType>(Val: Agg->getType())->getNumElements(); |
529 | |
530 | SmallVector<Constant*, 32> Result; |
531 | for (unsigned i = 0; i != NumElts; ++i) { |
532 | Constant *C = Agg->getAggregateElement(Elt: i); |
533 | if (!C) return nullptr; |
534 | |
535 | if (Idxs[0] == i) |
536 | C = ConstantFoldInsertValueInstruction(Agg: C, Val, Idxs: Idxs.slice(N: 1)); |
537 | |
538 | Result.push_back(Elt: C); |
539 | } |
540 | |
541 | if (StructType *ST = dyn_cast<StructType>(Val: Agg->getType())) |
542 | return ConstantStruct::get(T: ST, V: Result); |
543 | return ConstantArray::get(T: cast<ArrayType>(Val: Agg->getType()), V: Result); |
544 | } |
545 | |
546 | Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) { |
547 | assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected" ); |
548 | |
549 | // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length |
550 | // vectors are always evaluated per element. |
551 | bool IsScalableVector = isa<ScalableVectorType>(Val: C->getType()); |
552 | bool HasScalarUndefOrScalableVectorUndef = |
553 | (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(Val: C); |
554 | |
555 | if (HasScalarUndefOrScalableVectorUndef) { |
556 | switch (static_cast<Instruction::UnaryOps>(Opcode)) { |
557 | case Instruction::FNeg: |
558 | return C; // -undef -> undef |
559 | case Instruction::UnaryOpsEnd: |
560 | llvm_unreachable("Invalid UnaryOp" ); |
561 | } |
562 | } |
563 | |
564 | // Constant should not be UndefValue, unless these are vector constants. |
565 | assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue" ); |
566 | // We only have FP UnaryOps right now. |
567 | assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp" ); |
568 | |
569 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) { |
570 | const APFloat &CV = CFP->getValueAPF(); |
571 | switch (Opcode) { |
572 | default: |
573 | break; |
574 | case Instruction::FNeg: |
575 | return ConstantFP::get(Ty: C->getType(), V: neg(X: CV)); |
576 | } |
577 | } else if (auto *VTy = dyn_cast<VectorType>(Val: C->getType())) { |
578 | // Fast path for splatted constants. |
579 | if (Constant *Splat = C->getSplatValue()) |
580 | if (Constant *Elt = ConstantFoldUnaryInstruction(Opcode, C: Splat)) |
581 | return ConstantVector::getSplat(EC: VTy->getElementCount(), Elt); |
582 | |
583 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: VTy)) { |
584 | // Fold each element and create a vector constant from those constants. |
585 | Type *Ty = IntegerType::get(C&: FVTy->getContext(), NumBits: 32); |
586 | SmallVector<Constant *, 16> Result; |
587 | for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) { |
588 | Constant * = ConstantInt::get(Ty, V: i); |
589 | Constant *Elt = ConstantExpr::getExtractElement(Vec: C, Idx: ExtractIdx); |
590 | Constant *Res = ConstantFoldUnaryInstruction(Opcode, C: Elt); |
591 | if (!Res) |
592 | return nullptr; |
593 | Result.push_back(Elt: Res); |
594 | } |
595 | |
596 | return ConstantVector::get(V: Result); |
597 | } |
598 | } |
599 | |
600 | // We don't know how to fold this. |
601 | return nullptr; |
602 | } |
603 | |
604 | Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1, |
605 | Constant *C2) { |
606 | assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected" ); |
607 | |
608 | // Simplify BinOps with their identity values first. They are no-ops and we |
609 | // can always return the other value, including undef or poison values. |
610 | if (Constant *Identity = ConstantExpr::getBinOpIdentity( |
611 | Opcode, Ty: C1->getType(), /*AllowRHSIdentity*/ AllowRHSConstant: false)) { |
612 | if (C1 == Identity) |
613 | return C2; |
614 | if (C2 == Identity) |
615 | return C1; |
616 | } else if (Constant *Identity = ConstantExpr::getBinOpIdentity( |
617 | Opcode, Ty: C1->getType(), /*AllowRHSIdentity*/ AllowRHSConstant: true)) { |
618 | if (C2 == Identity) |
619 | return C1; |
620 | } |
621 | |
622 | // Binary operations propagate poison. |
623 | if (isa<PoisonValue>(Val: C1) || isa<PoisonValue>(Val: C2)) |
624 | return PoisonValue::get(T: C1->getType()); |
625 | |
626 | // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length |
627 | // vectors are always evaluated per element. |
628 | bool IsScalableVector = isa<ScalableVectorType>(Val: C1->getType()); |
629 | bool HasScalarUndefOrScalableVectorUndef = |
630 | (!C1->getType()->isVectorTy() || IsScalableVector) && |
631 | (isa<UndefValue>(Val: C1) || isa<UndefValue>(Val: C2)); |
632 | if (HasScalarUndefOrScalableVectorUndef) { |
633 | switch (static_cast<Instruction::BinaryOps>(Opcode)) { |
634 | case Instruction::Xor: |
635 | if (isa<UndefValue>(Val: C1) && isa<UndefValue>(Val: C2)) |
636 | // Handle undef ^ undef -> 0 special case. This is a common |
637 | // idiom (misuse). |
638 | return Constant::getNullValue(Ty: C1->getType()); |
639 | [[fallthrough]]; |
640 | case Instruction::Add: |
641 | case Instruction::Sub: |
642 | return UndefValue::get(T: C1->getType()); |
643 | case Instruction::And: |
644 | if (isa<UndefValue>(Val: C1) && isa<UndefValue>(Val: C2)) // undef & undef -> undef |
645 | return C1; |
646 | return Constant::getNullValue(Ty: C1->getType()); // undef & X -> 0 |
647 | case Instruction::Mul: { |
648 | // undef * undef -> undef |
649 | if (isa<UndefValue>(Val: C1) && isa<UndefValue>(Val: C2)) |
650 | return C1; |
651 | const APInt *CV; |
652 | // X * undef -> undef if X is odd |
653 | if (match(V: C1, P: m_APInt(Res&: CV)) || match(V: C2, P: m_APInt(Res&: CV))) |
654 | if ((*CV)[0]) |
655 | return UndefValue::get(T: C1->getType()); |
656 | |
657 | // X * undef -> 0 otherwise |
658 | return Constant::getNullValue(Ty: C1->getType()); |
659 | } |
660 | case Instruction::SDiv: |
661 | case Instruction::UDiv: |
662 | // X / undef -> poison |
663 | // X / 0 -> poison |
664 | if (match(V: C2, P: m_CombineOr(L: m_Undef(), R: m_Zero()))) |
665 | return PoisonValue::get(T: C2->getType()); |
666 | // undef / X -> 0 otherwise |
667 | return Constant::getNullValue(Ty: C1->getType()); |
668 | case Instruction::URem: |
669 | case Instruction::SRem: |
670 | // X % undef -> poison |
671 | // X % 0 -> poison |
672 | if (match(V: C2, P: m_CombineOr(L: m_Undef(), R: m_Zero()))) |
673 | return PoisonValue::get(T: C2->getType()); |
674 | // undef % X -> 0 otherwise |
675 | return Constant::getNullValue(Ty: C1->getType()); |
676 | case Instruction::Or: // X | undef -> -1 |
677 | if (isa<UndefValue>(Val: C1) && isa<UndefValue>(Val: C2)) // undef | undef -> undef |
678 | return C1; |
679 | return Constant::getAllOnesValue(Ty: C1->getType()); // undef | X -> ~0 |
680 | case Instruction::LShr: |
681 | // X >>l undef -> poison |
682 | if (isa<UndefValue>(Val: C2)) |
683 | return PoisonValue::get(T: C2->getType()); |
684 | // undef >>l X -> 0 |
685 | return Constant::getNullValue(Ty: C1->getType()); |
686 | case Instruction::AShr: |
687 | // X >>a undef -> poison |
688 | if (isa<UndefValue>(Val: C2)) |
689 | return PoisonValue::get(T: C2->getType()); |
690 | // TODO: undef >>a X -> poison if the shift is exact |
691 | // undef >>a X -> 0 |
692 | return Constant::getNullValue(Ty: C1->getType()); |
693 | case Instruction::Shl: |
694 | // X << undef -> undef |
695 | if (isa<UndefValue>(Val: C2)) |
696 | return PoisonValue::get(T: C2->getType()); |
697 | // undef << X -> 0 |
698 | return Constant::getNullValue(Ty: C1->getType()); |
699 | case Instruction::FSub: |
700 | // -0.0 - undef --> undef (consistent with "fneg undef") |
701 | if (match(V: C1, P: m_NegZeroFP()) && isa<UndefValue>(Val: C2)) |
702 | return C2; |
703 | [[fallthrough]]; |
704 | case Instruction::FAdd: |
705 | case Instruction::FMul: |
706 | case Instruction::FDiv: |
707 | case Instruction::FRem: |
708 | // [any flop] undef, undef -> undef |
709 | if (isa<UndefValue>(Val: C1) && isa<UndefValue>(Val: C2)) |
710 | return C1; |
711 | // [any flop] C, undef -> NaN |
712 | // [any flop] undef, C -> NaN |
713 | // We could potentially specialize NaN/Inf constants vs. 'normal' |
714 | // constants (possibly differently depending on opcode and operand). This |
715 | // would allow returning undef sometimes. But it is always safe to fold to |
716 | // NaN because we can choose the undef operand as NaN, and any FP opcode |
717 | // with a NaN operand will propagate NaN. |
718 | return ConstantFP::getNaN(Ty: C1->getType()); |
719 | case Instruction::BinaryOpsEnd: |
720 | llvm_unreachable("Invalid BinaryOp" ); |
721 | } |
722 | } |
723 | |
724 | // Neither constant should be UndefValue, unless these are vector constants. |
725 | assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue" ); |
726 | |
727 | // Handle simplifications when the RHS is a constant int. |
728 | if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Val: C2)) { |
729 | if (C2 == ConstantExpr::getBinOpAbsorber(Opcode, Ty: C2->getType(), |
730 | /*AllowLHSConstant*/ false)) |
731 | return C2; |
732 | |
733 | switch (Opcode) { |
734 | case Instruction::UDiv: |
735 | case Instruction::SDiv: |
736 | if (CI2->isZero()) |
737 | return PoisonValue::get(T: CI2->getType()); // X / 0 == poison |
738 | break; |
739 | case Instruction::URem: |
740 | case Instruction::SRem: |
741 | if (CI2->isOne()) |
742 | return Constant::getNullValue(Ty: CI2->getType()); // X % 1 == 0 |
743 | if (CI2->isZero()) |
744 | return PoisonValue::get(T: CI2->getType()); // X % 0 == poison |
745 | break; |
746 | case Instruction::And: |
747 | assert(!CI2->isZero() && "And zero handled above" ); |
748 | if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Val: C1)) { |
749 | // If and'ing the address of a global with a constant, fold it. |
750 | if (CE1->getOpcode() == Instruction::PtrToInt && |
751 | isa<GlobalValue>(Val: CE1->getOperand(i_nocapture: 0))) { |
752 | GlobalValue *GV = cast<GlobalValue>(Val: CE1->getOperand(i_nocapture: 0)); |
753 | |
754 | Align GVAlign; // defaults to 1 |
755 | |
756 | if (Module *TheModule = GV->getParent()) { |
757 | const DataLayout &DL = TheModule->getDataLayout(); |
758 | GVAlign = GV->getPointerAlignment(DL); |
759 | |
760 | // If the function alignment is not specified then assume that it |
761 | // is 4. |
762 | // This is dangerous; on x86, the alignment of the pointer |
763 | // corresponds to the alignment of the function, but might be less |
764 | // than 4 if it isn't explicitly specified. |
765 | // However, a fix for this behaviour was reverted because it |
766 | // increased code size (see https://reviews.llvm.org/D55115) |
767 | // FIXME: This code should be deleted once existing targets have |
768 | // appropriate defaults |
769 | if (isa<Function>(Val: GV) && !DL.getFunctionPtrAlign()) |
770 | GVAlign = Align(4); |
771 | } else if (isa<GlobalVariable>(Val: GV)) { |
772 | GVAlign = cast<GlobalVariable>(Val: GV)->getAlign().valueOrOne(); |
773 | } |
774 | |
775 | if (GVAlign > 1) { |
776 | unsigned DstWidth = CI2->getBitWidth(); |
777 | unsigned SrcWidth = std::min(a: DstWidth, b: Log2(A: GVAlign)); |
778 | APInt BitsNotSet(APInt::getLowBitsSet(numBits: DstWidth, loBitsSet: SrcWidth)); |
779 | |
780 | // If checking bits we know are clear, return zero. |
781 | if ((CI2->getValue() & BitsNotSet) == CI2->getValue()) |
782 | return Constant::getNullValue(Ty: CI2->getType()); |
783 | } |
784 | } |
785 | } |
786 | break; |
787 | } |
788 | } else if (isa<ConstantInt>(Val: C1)) { |
789 | // If C1 is a ConstantInt and C2 is not, swap the operands. |
790 | if (Instruction::isCommutative(Opcode)) |
791 | return ConstantExpr::isDesirableBinOp(Opcode) |
792 | ? ConstantExpr::get(Opcode, C1: C2, C2: C1) |
793 | : ConstantFoldBinaryInstruction(Opcode, C1: C2, C2: C1); |
794 | } |
795 | |
796 | if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Val: C1)) { |
797 | if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Val: C2)) { |
798 | const APInt &C1V = CI1->getValue(); |
799 | const APInt &C2V = CI2->getValue(); |
800 | switch (Opcode) { |
801 | default: |
802 | break; |
803 | case Instruction::Add: |
804 | return ConstantInt::get(Ty: C1->getType(), V: C1V + C2V); |
805 | case Instruction::Sub: |
806 | return ConstantInt::get(Ty: C1->getType(), V: C1V - C2V); |
807 | case Instruction::Mul: |
808 | return ConstantInt::get(Ty: C1->getType(), V: C1V * C2V); |
809 | case Instruction::UDiv: |
810 | assert(!CI2->isZero() && "Div by zero handled above" ); |
811 | return ConstantInt::get(Ty: CI1->getType(), V: C1V.udiv(RHS: C2V)); |
812 | case Instruction::SDiv: |
813 | assert(!CI2->isZero() && "Div by zero handled above" ); |
814 | if (C2V.isAllOnes() && C1V.isMinSignedValue()) |
815 | return PoisonValue::get(T: CI1->getType()); // MIN_INT / -1 -> poison |
816 | return ConstantInt::get(Ty: CI1->getType(), V: C1V.sdiv(RHS: C2V)); |
817 | case Instruction::URem: |
818 | assert(!CI2->isZero() && "Div by zero handled above" ); |
819 | return ConstantInt::get(Ty: C1->getType(), V: C1V.urem(RHS: C2V)); |
820 | case Instruction::SRem: |
821 | assert(!CI2->isZero() && "Div by zero handled above" ); |
822 | if (C2V.isAllOnes() && C1V.isMinSignedValue()) |
823 | return PoisonValue::get(T: C1->getType()); // MIN_INT % -1 -> poison |
824 | return ConstantInt::get(Ty: C1->getType(), V: C1V.srem(RHS: C2V)); |
825 | case Instruction::And: |
826 | return ConstantInt::get(Ty: C1->getType(), V: C1V & C2V); |
827 | case Instruction::Or: |
828 | return ConstantInt::get(Ty: C1->getType(), V: C1V | C2V); |
829 | case Instruction::Xor: |
830 | return ConstantInt::get(Ty: C1->getType(), V: C1V ^ C2V); |
831 | case Instruction::Shl: |
832 | if (C2V.ult(RHS: C1V.getBitWidth())) |
833 | return ConstantInt::get(Ty: C1->getType(), V: C1V.shl(ShiftAmt: C2V)); |
834 | return PoisonValue::get(T: C1->getType()); // too big shift is poison |
835 | case Instruction::LShr: |
836 | if (C2V.ult(RHS: C1V.getBitWidth())) |
837 | return ConstantInt::get(Ty: C1->getType(), V: C1V.lshr(ShiftAmt: C2V)); |
838 | return PoisonValue::get(T: C1->getType()); // too big shift is poison |
839 | case Instruction::AShr: |
840 | if (C2V.ult(RHS: C1V.getBitWidth())) |
841 | return ConstantInt::get(Ty: C1->getType(), V: C1V.ashr(ShiftAmt: C2V)); |
842 | return PoisonValue::get(T: C1->getType()); // too big shift is poison |
843 | } |
844 | } |
845 | |
846 | if (C1 == ConstantExpr::getBinOpAbsorber(Opcode, Ty: C1->getType(), |
847 | /*AllowLHSConstant*/ true)) |
848 | return C1; |
849 | } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(Val: C1)) { |
850 | if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(Val: C2)) { |
851 | const APFloat &C1V = CFP1->getValueAPF(); |
852 | const APFloat &C2V = CFP2->getValueAPF(); |
853 | APFloat C3V = C1V; // copy for modification |
854 | switch (Opcode) { |
855 | default: |
856 | break; |
857 | case Instruction::FAdd: |
858 | (void)C3V.add(RHS: C2V, RM: APFloat::rmNearestTiesToEven); |
859 | return ConstantFP::get(Ty: C1->getType(), V: C3V); |
860 | case Instruction::FSub: |
861 | (void)C3V.subtract(RHS: C2V, RM: APFloat::rmNearestTiesToEven); |
862 | return ConstantFP::get(Ty: C1->getType(), V: C3V); |
863 | case Instruction::FMul: |
864 | (void)C3V.multiply(RHS: C2V, RM: APFloat::rmNearestTiesToEven); |
865 | return ConstantFP::get(Ty: C1->getType(), V: C3V); |
866 | case Instruction::FDiv: |
867 | (void)C3V.divide(RHS: C2V, RM: APFloat::rmNearestTiesToEven); |
868 | return ConstantFP::get(Ty: C1->getType(), V: C3V); |
869 | case Instruction::FRem: |
870 | (void)C3V.mod(RHS: C2V); |
871 | return ConstantFP::get(Ty: C1->getType(), V: C3V); |
872 | } |
873 | } |
874 | } |
875 | |
876 | if (auto *VTy = dyn_cast<VectorType>(Val: C1->getType())) { |
877 | // Fast path for splatted constants. |
878 | if (Constant *C2Splat = C2->getSplatValue()) { |
879 | if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue()) |
880 | return PoisonValue::get(T: VTy); |
881 | if (Constant *C1Splat = C1->getSplatValue()) { |
882 | Constant *Res = |
883 | ConstantExpr::isDesirableBinOp(Opcode) |
884 | ? ConstantExpr::get(Opcode, C1: C1Splat, C2: C2Splat) |
885 | : ConstantFoldBinaryInstruction(Opcode, C1: C1Splat, C2: C2Splat); |
886 | if (!Res) |
887 | return nullptr; |
888 | return ConstantVector::getSplat(EC: VTy->getElementCount(), Elt: Res); |
889 | } |
890 | } |
891 | |
892 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: VTy)) { |
893 | // Fold each element and create a vector constant from those constants. |
894 | SmallVector<Constant*, 16> Result; |
895 | Type *Ty = IntegerType::get(C&: FVTy->getContext(), NumBits: 32); |
896 | for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) { |
897 | Constant * = ConstantInt::get(Ty, V: i); |
898 | Constant *LHS = ConstantExpr::getExtractElement(Vec: C1, Idx: ExtractIdx); |
899 | Constant *RHS = ConstantExpr::getExtractElement(Vec: C2, Idx: ExtractIdx); |
900 | Constant *Res = ConstantExpr::isDesirableBinOp(Opcode) |
901 | ? ConstantExpr::get(Opcode, C1: LHS, C2: RHS) |
902 | : ConstantFoldBinaryInstruction(Opcode, C1: LHS, C2: RHS); |
903 | if (!Res) |
904 | return nullptr; |
905 | Result.push_back(Elt: Res); |
906 | } |
907 | |
908 | return ConstantVector::get(V: Result); |
909 | } |
910 | } |
911 | |
912 | if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Val: C1)) { |
913 | // There are many possible foldings we could do here. We should probably |
914 | // at least fold add of a pointer with an integer into the appropriate |
915 | // getelementptr. This will improve alias analysis a bit. |
916 | |
917 | // Given ((a + b) + c), if (b + c) folds to something interesting, return |
918 | // (a + (b + c)). |
919 | if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) { |
920 | Constant *T = ConstantExpr::get(Opcode, C1: CE1->getOperand(i_nocapture: 1), C2); |
921 | if (!isa<ConstantExpr>(Val: T) || cast<ConstantExpr>(Val: T)->getOpcode() != Opcode) |
922 | return ConstantExpr::get(Opcode, C1: CE1->getOperand(i_nocapture: 0), C2: T); |
923 | } |
924 | } else if (isa<ConstantExpr>(Val: C2)) { |
925 | // If C2 is a constant expr and C1 isn't, flop them around and fold the |
926 | // other way if possible. |
927 | if (Instruction::isCommutative(Opcode)) |
928 | return ConstantFoldBinaryInstruction(Opcode, C1: C2, C2: C1); |
929 | } |
930 | |
931 | // i1 can be simplified in many cases. |
932 | if (C1->getType()->isIntegerTy(Bitwidth: 1)) { |
933 | switch (Opcode) { |
934 | case Instruction::Add: |
935 | case Instruction::Sub: |
936 | return ConstantExpr::getXor(C1, C2); |
937 | case Instruction::Shl: |
938 | case Instruction::LShr: |
939 | case Instruction::AShr: |
940 | // We can assume that C2 == 0. If it were one the result would be |
941 | // undefined because the shift value is as large as the bitwidth. |
942 | return C1; |
943 | case Instruction::SDiv: |
944 | case Instruction::UDiv: |
945 | // We can assume that C2 == 1. If it were zero the result would be |
946 | // undefined through division by zero. |
947 | return C1; |
948 | case Instruction::URem: |
949 | case Instruction::SRem: |
950 | // We can assume that C2 == 1. If it were zero the result would be |
951 | // undefined through division by zero. |
952 | return ConstantInt::getFalse(Context&: C1->getContext()); |
953 | default: |
954 | break; |
955 | } |
956 | } |
957 | |
958 | // We don't know how to fold this. |
959 | return nullptr; |
960 | } |
961 | |
962 | static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1, |
963 | const GlobalValue *GV2) { |
964 | auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) { |
965 | if (GV->isInterposable() || GV->hasGlobalUnnamedAddr()) |
966 | return true; |
967 | if (const auto *GVar = dyn_cast<GlobalVariable>(Val: GV)) { |
968 | Type *Ty = GVar->getValueType(); |
969 | // A global with opaque type might end up being zero sized. |
970 | if (!Ty->isSized()) |
971 | return true; |
972 | // A global with an empty type might lie at the address of any other |
973 | // global. |
974 | if (Ty->isEmptyTy()) |
975 | return true; |
976 | } |
977 | return false; |
978 | }; |
979 | // Don't try to decide equality of aliases. |
980 | if (!isa<GlobalAlias>(Val: GV1) && !isa<GlobalAlias>(Val: GV2)) |
981 | if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2)) |
982 | return ICmpInst::ICMP_NE; |
983 | return ICmpInst::BAD_ICMP_PREDICATE; |
984 | } |
985 | |
986 | /// This function determines if there is anything we can decide about the two |
987 | /// constants provided. This doesn't need to handle simple things like integer |
988 | /// comparisons, but should instead handle ConstantExprs and GlobalValues. |
989 | /// If we can determine that the two constants have a particular relation to |
990 | /// each other, we should return the corresponding ICmp predicate, otherwise |
991 | /// return ICmpInst::BAD_ICMP_PREDICATE. |
992 | static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2) { |
993 | assert(V1->getType() == V2->getType() && |
994 | "Cannot compare different types of values!" ); |
995 | if (V1 == V2) return ICmpInst::ICMP_EQ; |
996 | |
997 | // The following folds only apply to pointers. |
998 | if (!V1->getType()->isPointerTy()) |
999 | return ICmpInst::BAD_ICMP_PREDICATE; |
1000 | |
1001 | // To simplify this code we canonicalize the relation so that the first |
1002 | // operand is always the most "complex" of the two. We consider simple |
1003 | // constants (like ConstantPointerNull) to be the simplest, followed by |
1004 | // BlockAddress, GlobalValues, and ConstantExpr's (the most complex). |
1005 | auto GetComplexity = [](Constant *V) { |
1006 | if (isa<ConstantExpr>(Val: V)) |
1007 | return 3; |
1008 | if (isa<GlobalValue>(Val: V)) |
1009 | return 2; |
1010 | if (isa<BlockAddress>(Val: V)) |
1011 | return 1; |
1012 | return 0; |
1013 | }; |
1014 | if (GetComplexity(V1) < GetComplexity(V2)) { |
1015 | ICmpInst::Predicate SwappedRelation = evaluateICmpRelation(V1: V2, V2: V1); |
1016 | if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE) |
1017 | return ICmpInst::getSwappedPredicate(pred: SwappedRelation); |
1018 | return ICmpInst::BAD_ICMP_PREDICATE; |
1019 | } |
1020 | |
1021 | if (const BlockAddress *BA = dyn_cast<BlockAddress>(Val: V1)) { |
1022 | // Now we know that the RHS is a BlockAddress or simple constant. |
1023 | if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(Val: V2)) { |
1024 | // Block address in another function can't equal this one, but block |
1025 | // addresses in the current function might be the same if blocks are |
1026 | // empty. |
1027 | if (BA2->getFunction() != BA->getFunction()) |
1028 | return ICmpInst::ICMP_NE; |
1029 | } else if (isa<ConstantPointerNull>(Val: V2)) { |
1030 | return ICmpInst::ICMP_NE; |
1031 | } |
1032 | } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V1)) { |
1033 | // Now we know that the RHS is a GlobalValue, BlockAddress or simple |
1034 | // constant. |
1035 | if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(Val: V2)) { |
1036 | return areGlobalsPotentiallyEqual(GV1: GV, GV2); |
1037 | } else if (isa<BlockAddress>(Val: V2)) { |
1038 | return ICmpInst::ICMP_NE; // Globals never equal labels. |
1039 | } else if (isa<ConstantPointerNull>(Val: V2)) { |
1040 | // GlobalVals can never be null unless they have external weak linkage. |
1041 | // We don't try to evaluate aliases here. |
1042 | // NOTE: We should not be doing this constant folding if null pointer |
1043 | // is considered valid for the function. But currently there is no way to |
1044 | // query it from the Constant type. |
1045 | if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(Val: GV) && |
1046 | !NullPointerIsDefined(F: nullptr /* F */, |
1047 | AS: GV->getType()->getAddressSpace())) |
1048 | return ICmpInst::ICMP_UGT; |
1049 | } |
1050 | } else if (auto *CE1 = dyn_cast<ConstantExpr>(Val: V1)) { |
1051 | // Ok, the LHS is known to be a constantexpr. The RHS can be any of a |
1052 | // constantexpr, a global, block address, or a simple constant. |
1053 | Constant *CE1Op0 = CE1->getOperand(i_nocapture: 0); |
1054 | |
1055 | switch (CE1->getOpcode()) { |
1056 | case Instruction::GetElementPtr: { |
1057 | GEPOperator *CE1GEP = cast<GEPOperator>(Val: CE1); |
1058 | // Ok, since this is a getelementptr, we know that the constant has a |
1059 | // pointer type. Check the various cases. |
1060 | if (isa<ConstantPointerNull>(Val: V2)) { |
1061 | // If we are comparing a GEP to a null pointer, check to see if the base |
1062 | // of the GEP equals the null pointer. |
1063 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: CE1Op0)) { |
1064 | // If its not weak linkage, the GVal must have a non-zero address |
1065 | // so the result is greater-than |
1066 | if (!GV->hasExternalWeakLinkage() && CE1GEP->isInBounds()) |
1067 | return ICmpInst::ICMP_UGT; |
1068 | } |
1069 | } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(Val: V2)) { |
1070 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: CE1Op0)) { |
1071 | if (GV != GV2) { |
1072 | if (CE1GEP->hasAllZeroIndices()) |
1073 | return areGlobalsPotentiallyEqual(GV1: GV, GV2); |
1074 | return ICmpInst::BAD_ICMP_PREDICATE; |
1075 | } |
1076 | } |
1077 | } else if (const auto *CE2GEP = dyn_cast<GEPOperator>(Val: V2)) { |
1078 | // By far the most common case to handle is when the base pointers are |
1079 | // obviously to the same global. |
1080 | const Constant *CE2Op0 = cast<Constant>(Val: CE2GEP->getPointerOperand()); |
1081 | if (isa<GlobalValue>(Val: CE1Op0) && isa<GlobalValue>(Val: CE2Op0)) { |
1082 | // Don't know relative ordering, but check for inequality. |
1083 | if (CE1Op0 != CE2Op0) { |
1084 | if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices()) |
1085 | return areGlobalsPotentiallyEqual(GV1: cast<GlobalValue>(Val: CE1Op0), |
1086 | GV2: cast<GlobalValue>(Val: CE2Op0)); |
1087 | return ICmpInst::BAD_ICMP_PREDICATE; |
1088 | } |
1089 | } |
1090 | } |
1091 | break; |
1092 | } |
1093 | default: |
1094 | break; |
1095 | } |
1096 | } |
1097 | |
1098 | return ICmpInst::BAD_ICMP_PREDICATE; |
1099 | } |
1100 | |
1101 | Constant *llvm::ConstantFoldCompareInstruction(CmpInst::Predicate Predicate, |
1102 | Constant *C1, Constant *C2) { |
1103 | Type *ResultTy; |
1104 | if (VectorType *VT = dyn_cast<VectorType>(Val: C1->getType())) |
1105 | ResultTy = VectorType::get(ElementType: Type::getInt1Ty(C&: C1->getContext()), |
1106 | EC: VT->getElementCount()); |
1107 | else |
1108 | ResultTy = Type::getInt1Ty(C&: C1->getContext()); |
1109 | |
1110 | // Fold FCMP_FALSE/FCMP_TRUE unconditionally. |
1111 | if (Predicate == FCmpInst::FCMP_FALSE) |
1112 | return Constant::getNullValue(Ty: ResultTy); |
1113 | |
1114 | if (Predicate == FCmpInst::FCMP_TRUE) |
1115 | return Constant::getAllOnesValue(Ty: ResultTy); |
1116 | |
1117 | // Handle some degenerate cases first |
1118 | if (isa<PoisonValue>(Val: C1) || isa<PoisonValue>(Val: C2)) |
1119 | return PoisonValue::get(T: ResultTy); |
1120 | |
1121 | if (isa<UndefValue>(Val: C1) || isa<UndefValue>(Val: C2)) { |
1122 | bool isIntegerPredicate = ICmpInst::isIntPredicate(P: Predicate); |
1123 | // For EQ and NE, we can always pick a value for the undef to make the |
1124 | // predicate pass or fail, so we can return undef. |
1125 | // Also, if both operands are undef, we can return undef for int comparison. |
1126 | if (ICmpInst::isEquality(P: Predicate) || (isIntegerPredicate && C1 == C2)) |
1127 | return UndefValue::get(T: ResultTy); |
1128 | |
1129 | // Otherwise, for integer compare, pick the same value as the non-undef |
1130 | // operand, and fold it to true or false. |
1131 | if (isIntegerPredicate) |
1132 | return ConstantInt::get(Ty: ResultTy, V: CmpInst::isTrueWhenEqual(predicate: Predicate)); |
1133 | |
1134 | // Choosing NaN for the undef will always make unordered comparison succeed |
1135 | // and ordered comparison fails. |
1136 | return ConstantInt::get(Ty: ResultTy, V: CmpInst::isUnordered(predicate: Predicate)); |
1137 | } |
1138 | |
1139 | if (C2->isNullValue()) { |
1140 | // The caller is expected to commute the operands if the constant expression |
1141 | // is C2. |
1142 | // C1 >= 0 --> true |
1143 | if (Predicate == ICmpInst::ICMP_UGE) |
1144 | return Constant::getAllOnesValue(Ty: ResultTy); |
1145 | // C1 < 0 --> false |
1146 | if (Predicate == ICmpInst::ICMP_ULT) |
1147 | return Constant::getNullValue(Ty: ResultTy); |
1148 | } |
1149 | |
1150 | // If the comparison is a comparison between two i1's, simplify it. |
1151 | if (C1->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
1152 | switch (Predicate) { |
1153 | case ICmpInst::ICMP_EQ: |
1154 | if (isa<ConstantExpr>(Val: C1)) |
1155 | return ConstantExpr::getXor(C1, C2: ConstantExpr::getNot(C: C2)); |
1156 | return ConstantExpr::getXor(C1: ConstantExpr::getNot(C: C1), C2); |
1157 | case ICmpInst::ICMP_NE: |
1158 | return ConstantExpr::getXor(C1, C2); |
1159 | default: |
1160 | break; |
1161 | } |
1162 | } |
1163 | |
1164 | if (isa<ConstantInt>(Val: C1) && isa<ConstantInt>(Val: C2)) { |
1165 | const APInt &V1 = cast<ConstantInt>(Val: C1)->getValue(); |
1166 | const APInt &V2 = cast<ConstantInt>(Val: C2)->getValue(); |
1167 | return ConstantInt::get(Ty: ResultTy, V: ICmpInst::compare(LHS: V1, RHS: V2, Pred: Predicate)); |
1168 | } else if (isa<ConstantFP>(Val: C1) && isa<ConstantFP>(Val: C2)) { |
1169 | const APFloat &C1V = cast<ConstantFP>(Val: C1)->getValueAPF(); |
1170 | const APFloat &C2V = cast<ConstantFP>(Val: C2)->getValueAPF(); |
1171 | return ConstantInt::get(Ty: ResultTy, V: FCmpInst::compare(LHS: C1V, RHS: C2V, Pred: Predicate)); |
1172 | } else if (auto *C1VTy = dyn_cast<VectorType>(Val: C1->getType())) { |
1173 | |
1174 | // Fast path for splatted constants. |
1175 | if (Constant *C1Splat = C1->getSplatValue()) |
1176 | if (Constant *C2Splat = C2->getSplatValue()) |
1177 | if (Constant *Elt = |
1178 | ConstantFoldCompareInstruction(Predicate, C1: C1Splat, C2: C2Splat)) |
1179 | return ConstantVector::getSplat(EC: C1VTy->getElementCount(), Elt); |
1180 | |
1181 | // Do not iterate on scalable vector. The number of elements is unknown at |
1182 | // compile-time. |
1183 | if (isa<ScalableVectorType>(Val: C1VTy)) |
1184 | return nullptr; |
1185 | |
1186 | // If we can constant fold the comparison of each element, constant fold |
1187 | // the whole vector comparison. |
1188 | SmallVector<Constant*, 4> ResElts; |
1189 | Type *Ty = IntegerType::get(C&: C1->getContext(), NumBits: 32); |
1190 | // Compare the elements, producing an i1 result or constant expr. |
1191 | for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue(); |
1192 | I != E; ++I) { |
1193 | Constant *C1E = |
1194 | ConstantExpr::getExtractElement(Vec: C1, Idx: ConstantInt::get(Ty, V: I)); |
1195 | Constant *C2E = |
1196 | ConstantExpr::getExtractElement(Vec: C2, Idx: ConstantInt::get(Ty, V: I)); |
1197 | Constant *Elt = ConstantFoldCompareInstruction(Predicate, C1: C1E, C2: C2E); |
1198 | if (!Elt) |
1199 | return nullptr; |
1200 | |
1201 | ResElts.push_back(Elt); |
1202 | } |
1203 | |
1204 | return ConstantVector::get(V: ResElts); |
1205 | } |
1206 | |
1207 | if (C1->getType()->isFPOrFPVectorTy()) { |
1208 | if (C1 == C2) { |
1209 | // We know that C1 == C2 || isUnordered(C1, C2). |
1210 | if (Predicate == FCmpInst::FCMP_ONE) |
1211 | return ConstantInt::getFalse(Ty: ResultTy); |
1212 | else if (Predicate == FCmpInst::FCMP_UEQ) |
1213 | return ConstantInt::getTrue(Ty: ResultTy); |
1214 | } |
1215 | } else { |
1216 | // Evaluate the relation between the two constants, per the predicate. |
1217 | int Result = -1; // -1 = unknown, 0 = known false, 1 = known true. |
1218 | switch (evaluateICmpRelation(V1: C1, V2: C2)) { |
1219 | default: llvm_unreachable("Unknown relational!" ); |
1220 | case ICmpInst::BAD_ICMP_PREDICATE: |
1221 | break; // Couldn't determine anything about these constants. |
1222 | case ICmpInst::ICMP_EQ: // We know the constants are equal! |
1223 | // If we know the constants are equal, we can decide the result of this |
1224 | // computation precisely. |
1225 | Result = ICmpInst::isTrueWhenEqual(predicate: Predicate); |
1226 | break; |
1227 | case ICmpInst::ICMP_ULT: |
1228 | switch (Predicate) { |
1229 | case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE: |
1230 | Result = 1; break; |
1231 | case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE: |
1232 | Result = 0; break; |
1233 | default: |
1234 | break; |
1235 | } |
1236 | break; |
1237 | case ICmpInst::ICMP_SLT: |
1238 | switch (Predicate) { |
1239 | case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE: |
1240 | Result = 1; break; |
1241 | case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE: |
1242 | Result = 0; break; |
1243 | default: |
1244 | break; |
1245 | } |
1246 | break; |
1247 | case ICmpInst::ICMP_UGT: |
1248 | switch (Predicate) { |
1249 | case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE: |
1250 | Result = 1; break; |
1251 | case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE: |
1252 | Result = 0; break; |
1253 | default: |
1254 | break; |
1255 | } |
1256 | break; |
1257 | case ICmpInst::ICMP_SGT: |
1258 | switch (Predicate) { |
1259 | case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE: |
1260 | Result = 1; break; |
1261 | case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE: |
1262 | Result = 0; break; |
1263 | default: |
1264 | break; |
1265 | } |
1266 | break; |
1267 | case ICmpInst::ICMP_ULE: |
1268 | if (Predicate == ICmpInst::ICMP_UGT) |
1269 | Result = 0; |
1270 | if (Predicate == ICmpInst::ICMP_ULT || Predicate == ICmpInst::ICMP_ULE) |
1271 | Result = 1; |
1272 | break; |
1273 | case ICmpInst::ICMP_SLE: |
1274 | if (Predicate == ICmpInst::ICMP_SGT) |
1275 | Result = 0; |
1276 | if (Predicate == ICmpInst::ICMP_SLT || Predicate == ICmpInst::ICMP_SLE) |
1277 | Result = 1; |
1278 | break; |
1279 | case ICmpInst::ICMP_UGE: |
1280 | if (Predicate == ICmpInst::ICMP_ULT) |
1281 | Result = 0; |
1282 | if (Predicate == ICmpInst::ICMP_UGT || Predicate == ICmpInst::ICMP_UGE) |
1283 | Result = 1; |
1284 | break; |
1285 | case ICmpInst::ICMP_SGE: |
1286 | if (Predicate == ICmpInst::ICMP_SLT) |
1287 | Result = 0; |
1288 | if (Predicate == ICmpInst::ICMP_SGT || Predicate == ICmpInst::ICMP_SGE) |
1289 | Result = 1; |
1290 | break; |
1291 | case ICmpInst::ICMP_NE: |
1292 | if (Predicate == ICmpInst::ICMP_EQ) |
1293 | Result = 0; |
1294 | if (Predicate == ICmpInst::ICMP_NE) |
1295 | Result = 1; |
1296 | break; |
1297 | } |
1298 | |
1299 | // If we evaluated the result, return it now. |
1300 | if (Result != -1) |
1301 | return ConstantInt::get(Ty: ResultTy, V: Result); |
1302 | |
1303 | if ((!isa<ConstantExpr>(Val: C1) && isa<ConstantExpr>(Val: C2)) || |
1304 | (C1->isNullValue() && !C2->isNullValue())) { |
1305 | // If C2 is a constant expr and C1 isn't, flip them around and fold the |
1306 | // other way if possible. |
1307 | // Also, if C1 is null and C2 isn't, flip them around. |
1308 | Predicate = ICmpInst::getSwappedPredicate(pred: Predicate); |
1309 | return ConstantFoldCompareInstruction(Predicate, C1: C2, C2: C1); |
1310 | } |
1311 | } |
1312 | return nullptr; |
1313 | } |
1314 | |
1315 | Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C, |
1316 | std::optional<ConstantRange> InRange, |
1317 | ArrayRef<Value *> Idxs) { |
1318 | if (Idxs.empty()) return C; |
1319 | |
1320 | Type *GEPTy = GetElementPtrInst::getGEPReturnType( |
1321 | Ptr: C, IdxList: ArrayRef((Value *const *)Idxs.data(), Idxs.size())); |
1322 | |
1323 | if (isa<PoisonValue>(Val: C)) |
1324 | return PoisonValue::get(T: GEPTy); |
1325 | |
1326 | if (isa<UndefValue>(Val: C)) |
1327 | return UndefValue::get(T: GEPTy); |
1328 | |
1329 | auto IsNoOp = [&]() { |
1330 | // Avoid losing inrange information. |
1331 | if (InRange) |
1332 | return false; |
1333 | |
1334 | return all_of(Range&: Idxs, P: [](Value *Idx) { |
1335 | Constant *IdxC = cast<Constant>(Val: Idx); |
1336 | return IdxC->isNullValue() || isa<UndefValue>(Val: IdxC); |
1337 | }); |
1338 | }; |
1339 | if (IsNoOp()) |
1340 | return GEPTy->isVectorTy() && !C->getType()->isVectorTy() |
1341 | ? ConstantVector::getSplat( |
1342 | EC: cast<VectorType>(Val: GEPTy)->getElementCount(), Elt: C) |
1343 | : C; |
1344 | |
1345 | return nullptr; |
1346 | } |
1347 | |