clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstCombineVectorOps.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/InstCombine -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/include -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/include -D NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-09-04-040900-46481-1 -x c++ /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
1 | |
2 | |
3 | |
4 | |
5 | |
6 | |
7 | |
8 | |
9 | |
10 | |
11 | |
12 | |
13 | |
14 | #include "InstCombineInternal.h" |
15 | #include "llvm/ADT/APInt.h" |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/DenseMap.h" |
18 | #include "llvm/ADT/STLExtras.h" |
19 | #include "llvm/ADT/SmallBitVector.h" |
20 | #include "llvm/ADT/SmallVector.h" |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/Analysis/InstructionSimplify.h" |
23 | #include "llvm/Analysis/VectorUtils.h" |
24 | #include "llvm/IR/BasicBlock.h" |
25 | #include "llvm/IR/Constant.h" |
26 | #include "llvm/IR/Constants.h" |
27 | #include "llvm/IR/DerivedTypes.h" |
28 | #include "llvm/IR/InstrTypes.h" |
29 | #include "llvm/IR/Instruction.h" |
30 | #include "llvm/IR/Instructions.h" |
31 | #include "llvm/IR/Operator.h" |
32 | #include "llvm/IR/PatternMatch.h" |
33 | #include "llvm/IR/Type.h" |
34 | #include "llvm/IR/User.h" |
35 | #include "llvm/IR/Value.h" |
36 | #include "llvm/Support/Casting.h" |
37 | #include "llvm/Support/ErrorHandling.h" |
38 | #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" |
39 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
40 | #include <cassert> |
41 | #include <cstdint> |
42 | #include <iterator> |
43 | #include <utility> |
44 | |
45 | using namespace llvm; |
46 | using namespace PatternMatch; |
47 | |
48 | #define DEBUG_TYPE "instcombine" |
49 | |
50 | STATISTIC(NumAggregateReconstructionsSimplified, |
51 | "Number of aggregate reconstructions turned into reuse of the " |
52 | "original aggregate"); |
53 | |
54 | |
55 | |
56 | |
57 | |
58 | |
59 | static bool cheapToScalarize(Value *V, Value *EI) { |
60 | ConstantInt *CEI = dyn_cast<ConstantInt>(EI); |
61 | |
62 | |
63 | if (auto *C = dyn_cast<Constant>(V)) |
64 | return CEI || C->getSplatValue(); |
65 | |
66 | if (CEI && match(V, m_Intrinsic<Intrinsic::experimental_stepvector>())) { |
67 | ElementCount EC = cast<VectorType>(V->getType())->getElementCount(); |
68 | |
69 | |
70 | return CEI->getValue().ult(EC.getKnownMinValue()); |
71 | } |
72 | |
73 | |
74 | |
75 | |
76 | if (match(V, m_InsertElt(m_Value(), m_Value(), m_ConstantInt()))) |
77 | return CEI; |
78 | |
79 | if (match(V, m_OneUse(m_Load(m_Value())))) |
80 | return true; |
81 | |
82 | if (match(V, m_OneUse(m_UnOp()))) |
83 | return true; |
84 | |
85 | Value *V0, *V1; |
86 | if (match(V, m_OneUse(m_BinOp(m_Value(V0), m_Value(V1))))) |
87 | if (cheapToScalarize(V0, EI) || cheapToScalarize(V1, EI)) |
88 | return true; |
89 | |
90 | CmpInst::Predicate UnusedPred; |
91 | if (match(V, m_OneUse(m_Cmp(UnusedPred, m_Value(V0), m_Value(V1))))) |
92 | if (cheapToScalarize(V0, EI) || cheapToScalarize(V1, EI)) |
93 | return true; |
94 | |
95 | return false; |
96 | } |
97 | |
98 | |
99 | |
100 | |
101 | Instruction *InstCombinerImpl::scalarizePHI(ExtractElementInst &EI, |
102 | PHINode *PN) { |
103 | SmallVector<Instruction *, 2> Extracts; |
104 | |
105 | |
106 | |
107 | |
108 | Instruction *PHIUser = nullptr; |
109 | for (auto U : PN->users()) { |
110 | if (ExtractElementInst *EU = dyn_cast<ExtractElementInst>(U)) { |
111 | if (EI.getIndexOperand() == EU->getIndexOperand()) |
112 | Extracts.push_back(EU); |
113 | else |
114 | return nullptr; |
115 | } else if (!PHIUser) { |
116 | PHIUser = cast<Instruction>(U); |
117 | } else { |
118 | return nullptr; |
119 | } |
120 | } |
121 | |
122 | if (!PHIUser) |
123 | return nullptr; |
124 | |
125 | |
126 | |
127 | |
128 | if (!PHIUser->hasOneUse() || !(PHIUser->user_back() == PN) || |
129 | !(isa<BinaryOperator>(PHIUser)) || |
130 | !cheapToScalarize(PHIUser, EI.getIndexOperand())) |
131 | return nullptr; |
132 | |
133 | |
134 | |
135 | PHINode *scalarPHI = cast<PHINode>(InsertNewInstWith( |
136 | PHINode::Create(EI.getType(), PN->getNumIncomingValues(), ""), *PN)); |
137 | |
138 | for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) { |
139 | Value *PHIInVal = PN->getIncomingValue(i); |
140 | BasicBlock *inBB = PN->getIncomingBlock(i); |
141 | Value *Elt = EI.getIndexOperand(); |
142 | |
143 | if (PHIInVal == PHIUser) { |
144 | |
145 | |
146 | |
147 | BinaryOperator *B0 = cast<BinaryOperator>(PHIUser); |
148 | unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0; |
149 | Value *Op = InsertNewInstWith( |
150 | ExtractElementInst::Create(B0->getOperand(opId), Elt, |
151 | B0->getOperand(opId)->getName() + ".Elt"), |
152 | *B0); |
153 | Value *newPHIUser = InsertNewInstWith( |
154 | BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(), |
155 | scalarPHI, Op, B0), *B0); |
156 | scalarPHI->addIncoming(newPHIUser, inBB); |
157 | } else { |
158 | |
159 | Instruction *newEI = ExtractElementInst::Create(PHIInVal, Elt, ""); |
160 | |
161 | Instruction *pos = dyn_cast<Instruction>(PHIInVal); |
162 | BasicBlock::iterator InsertPos; |
163 | if (pos && !isa<PHINode>(pos)) { |
164 | InsertPos = ++pos->getIterator(); |
165 | } else { |
166 | InsertPos = inBB->getFirstInsertionPt(); |
167 | } |
168 | |
169 | InsertNewInstWith(newEI, *InsertPos); |
170 | |
171 | scalarPHI->addIncoming(newEI, inBB); |
172 | } |
173 | } |
174 | |
175 | for (auto E : Extracts) |
176 | replaceInstUsesWith(*E, scalarPHI); |
177 | |
178 | return &EI; |
179 | } |
180 | |
181 | static Instruction *foldBitcastExtElt(ExtractElementInst &Ext, |
182 | InstCombiner::BuilderTy &Builder, |
183 | bool IsBigEndian) { |
184 | Value *X; |
185 | uint64_t ExtIndexC; |
186 | if (!match(Ext.getVectorOperand(), m_BitCast(m_Value(X))) || |
| |
187 | !X->getType()->isVectorTy() || |
188 | !match(Ext.getIndexOperand(), m_ConstantInt(ExtIndexC))) |
189 | return nullptr; |
190 | |
191 | |
192 | |
193 | |
194 | auto *SrcTy = cast<VectorType>(X->getType()); |
| 10 | | The object is a 'VectorType' | |
|
195 | Type *DestTy = Ext.getType(); |
196 | ElementCount NumSrcElts = SrcTy->getElementCount(); |
| 11 | | Calling 'VectorType::getElementCount' | |
|
| 17 | | Returning from 'VectorType::getElementCount' | |
|
197 | ElementCount NumElts = |
198 | cast<VectorType>(Ext.getVectorOperandType())->getElementCount(); |
| 18 | | The object is a 'VectorType' | |
|
199 | if (NumSrcElts == NumElts) |
| 19 | | Calling 'UnivariateLinearPolyBase::operator==' | |
|
| 22 | | Returning from 'UnivariateLinearPolyBase::operator==' | |
|
| |
200 | if (Value *Elt = findScalarElement(X, ExtIndexC)) |
201 | return new BitCastInst(Elt, DestTy); |
202 | |
203 | assert(NumSrcElts.isScalable() == NumElts.isScalable() && |
204 | "Src and Dst must be the same sort of vector type"); |
205 | |
206 | |
207 | |
208 | if (NumSrcElts.getKnownMinValue() < NumElts.getKnownMinValue()) { |
| 24 | | Assuming the condition is true | |
|
| |
209 | Value *Scalar; |
210 | uint64_t InsIndexC; |
211 | if (!match(X, m_InsertElt(m_Value(), m_Value(Scalar), |
| 26 | | Calling 'match<llvm::Value, llvm::PatternMatch::ThreeOps_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_const_intval_ty, 62>>' | |
|
| 34 | | Returning from 'match<llvm::Value, llvm::PatternMatch::ThreeOps_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_const_intval_ty, 62>>' | |
|
| |
212 | m_ConstantInt(InsIndexC)))) |
213 | return nullptr; |
214 | |
215 | |
216 | |
217 | |
218 | |
219 | unsigned NarrowingRatio = |
220 | NumElts.getKnownMinValue() / NumSrcElts.getKnownMinValue(); |
| 36 | | Calling 'LinearPolySize::getKnownMinValue' | |
|
| 41 | | Returning from 'LinearPolySize::getKnownMinValue' | |
|
| |
221 | if (ExtIndexC / NarrowingRatio != InsIndexC) |
222 | return nullptr; |
223 | |
224 | |
225 | |
226 | |
227 | |
228 | |
229 | |
230 | |
231 | |
232 | |
233 | |
234 | |
235 | unsigned Chunk = ExtIndexC % NarrowingRatio; |
236 | if (IsBigEndian) |
237 | Chunk = NarrowingRatio - 1 - Chunk; |
238 | |
239 | |
240 | |
241 | |
242 | bool NeedSrcBitcast = SrcTy->getScalarType()->isFloatingPointTy(); |
243 | bool NeedDestBitcast = DestTy->isFloatingPointTy(); |
244 | if (NeedSrcBitcast && NeedDestBitcast) |
245 | return nullptr; |
246 | |
247 | unsigned SrcWidth = SrcTy->getScalarSizeInBits(); |
248 | unsigned DestWidth = DestTy->getPrimitiveSizeInBits(); |
249 | unsigned ShAmt = Chunk * DestWidth; |
250 | |
251 | |
252 | |
253 | |
254 | if (!X->hasOneUse() || !Ext.getVectorOperand()->hasOneUse()) |
255 | if (NeedSrcBitcast || NeedDestBitcast) |
256 | return nullptr; |
257 | |
258 | if (NeedSrcBitcast) { |
259 | Type *SrcIntTy = IntegerType::getIntNTy(Scalar->getContext(), SrcWidth); |
260 | Scalar = Builder.CreateBitCast(Scalar, SrcIntTy); |
261 | } |
262 | |
263 | if (ShAmt) { |
264 | |
265 | if (!Ext.getVectorOperand()->hasOneUse()) |
266 | return nullptr; |
267 | Scalar = Builder.CreateLShr(Scalar, ShAmt); |
268 | } |
269 | |
270 | if (NeedDestBitcast) { |
271 | Type *DestIntTy = IntegerType::getIntNTy(Scalar->getContext(), DestWidth); |
272 | return new BitCastInst(Builder.CreateTrunc(Scalar, DestIntTy), DestTy); |
273 | } |
274 | return new TruncInst(Scalar, DestTy); |
275 | } |
276 | |
277 | return nullptr; |
278 | } |
279 | |
280 | |
281 | static APInt findDemandedEltsBySingleUser(Value *V, Instruction *UserInstr) { |
282 | unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements(); |
283 | |
284 | |
285 | APInt UsedElts(APInt::getAllOnesValue(VWidth)); |
286 | |
287 | switch (UserInstr->getOpcode()) { |
288 | case Instruction::ExtractElement: { |
289 | ExtractElementInst *EEI = cast<ExtractElementInst>(UserInstr); |
290 | assert(EEI->getVectorOperand() == V); |
291 | ConstantInt *EEIIndexC = dyn_cast<ConstantInt>(EEI->getIndexOperand()); |
292 | if (EEIIndexC && EEIIndexC->getValue().ult(VWidth)) { |
293 | UsedElts = APInt::getOneBitSet(VWidth, EEIIndexC->getZExtValue()); |
294 | } |
295 | break; |
296 | } |
297 | case Instruction::ShuffleVector: { |
298 | ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(UserInstr); |
299 | unsigned MaskNumElts = |
300 | cast<FixedVectorType>(UserInstr->getType())->getNumElements(); |
301 | |
302 | UsedElts = APInt(VWidth, 0); |
303 | for (unsigned i = 0; i < MaskNumElts; i++) { |
304 | unsigned MaskVal = Shuffle->getMaskValue(i); |
305 | if (MaskVal == -1u || MaskVal >= 2 * VWidth) |
306 | continue; |
307 | if (Shuffle->getOperand(0) == V && (MaskVal < VWidth)) |
308 | UsedElts.setBit(MaskVal); |
309 | if (Shuffle->getOperand(1) == V && |
310 | ((MaskVal >= VWidth) && (MaskVal < 2 * VWidth))) |
311 | UsedElts.setBit(MaskVal - VWidth); |
312 | } |
313 | break; |
314 | } |
315 | default: |
316 | break; |
317 | } |
318 | return UsedElts; |
319 | } |
320 | |
321 | |
322 | |
323 | |
324 | |
325 | static APInt findDemandedEltsByAllUsers(Value *V) { |
326 | unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements(); |
327 | |
328 | APInt UnionUsedElts(VWidth, 0); |
329 | for (const Use &U : V->uses()) { |
330 | if (Instruction *I = dyn_cast<Instruction>(U.getUser())) { |
331 | UnionUsedElts |= findDemandedEltsBySingleUser(V, I); |
332 | } else { |
333 | UnionUsedElts = APInt::getAllOnesValue(VWidth); |
334 | break; |
335 | } |
336 | |
337 | if (UnionUsedElts.isAllOnesValue()) |
338 | break; |
339 | } |
340 | |
341 | return UnionUsedElts; |
342 | } |
343 | |
344 | Instruction *InstCombinerImpl::visitExtractElementInst(ExtractElementInst &EI) { |
345 | Value *SrcVec = EI.getVectorOperand(); |
346 | Value *Index = EI.getIndexOperand(); |
347 | if (Value *V = SimplifyExtractElementInst(SrcVec, Index, |
| |
| |
348 | SQ.getWithInstruction(&EI))) |
349 | return replaceInstUsesWith(EI, V); |
350 | |
351 | |
352 | |
353 | auto *IndexC = dyn_cast<ConstantInt>(Index); |
| 3 | | Assuming 'Index' is a 'ConstantInt' | |
|
354 | if (IndexC) { |
| |
355 | ElementCount EC = EI.getVectorOperandType()->getElementCount(); |
356 | unsigned NumElts = EC.getKnownMinValue(); |
357 | |
358 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(SrcVec)) { |
| 5 | | Assuming 'SrcVec' is not a 'IntrinsicInst' | |
|
359 | Intrinsic::ID IID = II->getIntrinsicID(); |
360 | |
361 | |
362 | if (IID == Intrinsic::experimental_stepvector && |
363 | IndexC->getValue().ult(NumElts)) { |
364 | Type *Ty = EI.getType(); |
365 | unsigned BitWidth = Ty->getIntegerBitWidth(); |
366 | Value *Idx; |
367 | |
368 | |
369 | if (IndexC->getValue().getActiveBits() <= BitWidth) |
370 | Idx = ConstantInt::get(Ty, IndexC->getValue().zextOrTrunc(BitWidth)); |
371 | else |
372 | Idx = UndefValue::get(Ty); |
373 | return replaceInstUsesWith(EI, Idx); |
374 | } |
375 | } |
376 | |
377 | |
378 | |
379 | if (!EC.isScalable() && IndexC->getValue().uge(NumElts)) |
380 | return nullptr; |
381 | |
382 | |
383 | |
384 | |
385 | if (!EC.isScalable() && NumElts != 1) { |
| 6 | | Assuming 'NumElts' is equal to 1 | |
|
| |
386 | |
387 | |
388 | if (SrcVec->hasOneUse()) { |
389 | APInt UndefElts(NumElts, 0); |
390 | APInt DemandedElts(NumElts, 0); |
391 | DemandedElts.setBit(IndexC->getZExtValue()); |
392 | if (Value *V = |
393 | SimplifyDemandedVectorElts(SrcVec, DemandedElts, UndefElts)) |
394 | return replaceOperand(EI, 0, V); |
395 | } else { |
396 | |
397 | |
398 | APInt DemandedElts = findDemandedEltsByAllUsers(SrcVec); |
399 | if (!DemandedElts.isAllOnesValue()) { |
400 | APInt UndefElts(NumElts, 0); |
401 | if (Value *V = SimplifyDemandedVectorElts( |
402 | SrcVec, DemandedElts, UndefElts, 0 , |
403 | true )) { |
404 | if (V != SrcVec) { |
405 | SrcVec->replaceAllUsesWith(V); |
406 | return &EI; |
407 | } |
408 | } |
409 | } |
410 | } |
411 | } |
412 | |
413 | if (Instruction *I = foldBitcastExtElt(EI, Builder, DL.isBigEndian())) |
| 8 | | Calling 'foldBitcastExtElt' | |
|
414 | return I; |
415 | |
416 | |
417 | |
418 | if (auto *Phi = dyn_cast<PHINode>(SrcVec)) |
419 | if (Instruction *ScalarPHI = scalarizePHI(EI, Phi)) |
420 | return ScalarPHI; |
421 | } |
422 | |
423 | |
424 | |
425 | UnaryOperator *UO; |
426 | if (match(SrcVec, m_UnOp(UO)) && cheapToScalarize(SrcVec, Index)) { |
427 | |
428 | Value *X = UO->getOperand(0); |
429 | Value *E = Builder.CreateExtractElement(X, Index); |
430 | return UnaryOperator::CreateWithCopiedFlags(UO->getOpcode(), E, UO); |
431 | } |
432 | |
433 | BinaryOperator *BO; |
434 | if (match(SrcVec, m_BinOp(BO)) && cheapToScalarize(SrcVec, Index)) { |
435 | |
436 | Value *X = BO->getOperand(0), *Y = BO->getOperand(1); |
437 | Value *E0 = Builder.CreateExtractElement(X, Index); |
438 | Value *E1 = Builder.CreateExtractElement(Y, Index); |
439 | return BinaryOperator::CreateWithCopiedFlags(BO->getOpcode(), E0, E1, BO); |
440 | } |
441 | |
442 | Value *X, *Y; |
443 | CmpInst::Predicate Pred; |
444 | if (match(SrcVec, m_Cmp(Pred, m_Value(X), m_Value(Y))) && |
445 | cheapToScalarize(SrcVec, Index)) { |
446 | |
447 | Value *E0 = Builder.CreateExtractElement(X, Index); |
448 | Value *E1 = Builder.CreateExtractElement(Y, Index); |
449 | return CmpInst::Create(cast<CmpInst>(SrcVec)->getOpcode(), Pred, E0, E1); |
450 | } |
451 | |
452 | if (auto *I = dyn_cast<Instruction>(SrcVec)) { |
453 | if (auto *IE = dyn_cast<InsertElementInst>(I)) { |
454 | |
455 | if (IE->getOperand(2) == Index) |
456 | return replaceInstUsesWith(EI, IE->getOperand(1)); |
457 | |
458 | |
459 | if (isa<Constant>(IE->getOperand(2)) && IndexC) |
460 | return replaceOperand(EI, 0, IE->getOperand(0)); |
461 | } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) { |
462 | auto *VecType = cast<VectorType>(GEP->getType()); |
463 | ElementCount EC = VecType->getElementCount(); |
464 | uint64_t IdxVal = IndexC ? IndexC->getZExtValue() : 0; |
465 | if (IndexC && IdxVal < EC.getKnownMinValue() && GEP->hasOneUse()) { |
466 | |
467 | |
468 | |
469 | |
470 | |
471 | |
472 | |
473 | |
474 | unsigned VectorOps = |
475 | llvm::count_if(GEP->operands(), [](const Value *V) { |
476 | return isa<VectorType>(V->getType()); |
477 | }); |
478 | if (VectorOps > 1) |
479 | return nullptr; |
480 | assert(VectorOps == 1 && "Expected exactly one vector GEP operand!"); |
481 | |
482 | Value *NewPtr = GEP->getPointerOperand(); |
483 | if (isa<VectorType>(NewPtr->getType())) |
484 | NewPtr = Builder.CreateExtractElement(NewPtr, IndexC); |
485 | |
486 | SmallVector<Value *> NewOps; |
487 | for (unsigned I = 1; I != GEP->getNumOperands(); ++I) { |
488 | Value *Op = GEP->getOperand(I); |
489 | if (isa<VectorType>(Op->getType())) |
490 | NewOps.push_back(Builder.CreateExtractElement(Op, IndexC)); |
491 | else |
492 | NewOps.push_back(Op); |
493 | } |
494 | |
495 | GetElementPtrInst *NewGEP = GetElementPtrInst::Create( |
496 | cast<PointerType>(NewPtr->getType())->getElementType(), NewPtr, |
497 | NewOps); |
498 | NewGEP->setIsInBounds(GEP->isInBounds()); |
499 | return NewGEP; |
500 | } |
501 | return nullptr; |
502 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) { |
503 | |
504 | |
505 | |
506 | if (isa<FixedVectorType>(SVI->getType()) && isa<ConstantInt>(Index)) { |
507 | int SrcIdx = |
508 | SVI->getMaskValue(cast<ConstantInt>(Index)->getZExtValue()); |
509 | Value *Src; |
510 | unsigned LHSWidth = cast<FixedVectorType>(SVI->getOperand(0)->getType()) |
511 | ->getNumElements(); |
512 | |
513 | if (SrcIdx < 0) |
514 | return replaceInstUsesWith(EI, UndefValue::get(EI.getType())); |
515 | if (SrcIdx < (int)LHSWidth) |
516 | Src = SVI->getOperand(0); |
517 | else { |
518 | SrcIdx -= LHSWidth; |
519 | Src = SVI->getOperand(1); |
520 | } |
521 | Type *Int32Ty = Type::getInt32Ty(EI.getContext()); |
522 | return ExtractElementInst::Create( |
523 | Src, ConstantInt::get(Int32Ty, SrcIdx, false)); |
524 | } |
525 | } else if (auto *CI = dyn_cast<CastInst>(I)) { |
526 | |
527 | |
528 | |
529 | if (CI->hasOneUse() && (CI->getOpcode() != Instruction::BitCast)) { |
530 | Value *EE = Builder.CreateExtractElement(CI->getOperand(0), Index); |
531 | return CastInst::Create(CI->getOpcode(), EE, EI.getType()); |
532 | } |
533 | } |
534 | } |
535 | return nullptr; |
536 | } |
537 | |
538 | |
539 | |
540 | static bool collectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, |
541 | SmallVectorImpl<int> &Mask) { |
542 | assert(LHS->getType() == RHS->getType() && |
543 | "Invalid CollectSingleShuffleElements"); |
544 | unsigned NumElts = cast<FixedVectorType>(V->getType())->getNumElements(); |
545 | |
546 | if (match(V, m_Undef())) { |
547 | Mask.assign(NumElts, -1); |
548 | return true; |
549 | } |
550 | |
551 | if (V == LHS) { |
552 | for (unsigned i = 0; i != NumElts; ++i) |
553 | Mask.push_back(i); |
554 | return true; |
555 | } |
556 | |
557 | if (V == RHS) { |
558 | for (unsigned i = 0; i != NumElts; ++i) |
559 | Mask.push_back(i + NumElts); |
560 | return true; |
561 | } |
562 | |
563 | if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { |
564 | |
565 | Value *VecOp = IEI->getOperand(0); |
566 | Value *ScalarOp = IEI->getOperand(1); |
567 | Value *IdxOp = IEI->getOperand(2); |
568 | |
569 | if (!isa<ConstantInt>(IdxOp)) |
570 | return false; |
571 | unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); |
572 | |
573 | if (isa<UndefValue>(ScalarOp)) { |
574 | |
575 | |
576 | if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { |
577 | |
578 | Mask[InsertedIdx] = -1; |
579 | return true; |
580 | } |
581 | } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){ |
582 | if (isa<ConstantInt>(EI->getOperand(1))) { |
583 | unsigned ExtractedIdx = |
584 | cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); |
585 | unsigned NumLHSElts = |
586 | cast<FixedVectorType>(LHS->getType())->getNumElements(); |
587 | |
588 | |
589 | if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) { |
590 | |
591 | |
592 | if (collectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { |
593 | |
594 | if (EI->getOperand(0) == LHS) { |
595 | Mask[InsertedIdx % NumElts] = ExtractedIdx; |
596 | } else { |
597 | assert(EI->getOperand(0) == RHS); |
598 | Mask[InsertedIdx % NumElts] = ExtractedIdx + NumLHSElts; |
599 | } |
600 | return true; |
601 | } |
602 | } |
603 | } |
604 | } |
605 | } |
606 | |
607 | return false; |
608 | } |
609 | |
610 | |
611 | |
612 | |
613 | static void replaceExtractElements(InsertElementInst *InsElt, |
614 | ExtractElementInst *ExtElt, |
615 | InstCombinerImpl &IC) { |
616 | auto *InsVecType = cast<FixedVectorType>(InsElt->getType()); |
617 | auto *ExtVecType = cast<FixedVectorType>(ExtElt->getVectorOperandType()); |
618 | unsigned NumInsElts = InsVecType->getNumElements(); |
619 | unsigned NumExtElts = ExtVecType->getNumElements(); |
620 | |
621 | |
622 | if (InsVecType->getElementType() != ExtVecType->getElementType() || |
623 | NumExtElts >= NumInsElts) |
624 | return; |
625 | |
626 | |
627 | |
628 | |
629 | |
630 | SmallVector<int, 16> ExtendMask; |
631 | for (unsigned i = 0; i < NumExtElts; ++i) |
632 | ExtendMask.push_back(i); |
633 | for (unsigned i = NumExtElts; i < NumInsElts; ++i) |
634 | ExtendMask.push_back(-1); |
635 | |
636 | Value *ExtVecOp = ExtElt->getVectorOperand(); |
637 | auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp); |
638 | BasicBlock *InsertionBlock = (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) |
639 | ? ExtVecOpInst->getParent() |
640 | : ExtElt->getParent(); |
641 | |
642 | |
643 | |
644 | |
645 | |
646 | |
647 | |
648 | |
649 | |
650 | |
651 | if (InsertionBlock != InsElt->getParent()) |
652 | return; |
653 | |
654 | |
655 | |
656 | |
657 | |
658 | |
659 | if (InsElt->hasOneUse() && isa<InsertElementInst>(InsElt->user_back())) |
660 | return; |
661 | |
662 | auto *WideVec = |
663 | new ShuffleVectorInst(ExtVecOp, PoisonValue::get(ExtVecType), ExtendMask); |
664 | |
665 | |
666 | |
667 | |
668 | |
669 | if (ExtVecOpInst && !isa<PHINode>(ExtVecOpInst)) |
670 | WideVec->insertAfter(ExtVecOpInst); |
671 | else |
672 | IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt()); |
673 | |
674 | |
675 | |
676 | for (User *U : ExtVecOp->users()) { |
677 | ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U); |
678 | if (!OldExt || OldExt->getParent() != WideVec->getParent()) |
679 | continue; |
680 | auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1)); |
681 | NewExt->insertAfter(OldExt); |
682 | IC.replaceInstUsesWith(*OldExt, NewExt); |
683 | } |
684 | } |
685 | |
686 | |
687 | |
688 | |
689 | |
690 | |
691 | |
692 | |
693 | |
694 | using ShuffleOps = std::pair<Value *, Value *>; |
695 | |
696 | static ShuffleOps collectShuffleElements(Value *V, SmallVectorImpl<int> &Mask, |
697 | Value *PermittedRHS, |
698 | InstCombinerImpl &IC) { |
699 | assert(V->getType()->isVectorTy() && "Invalid shuffle!"); |
700 | unsigned NumElts = cast<FixedVectorType>(V->getType())->getNumElements(); |
701 | |
702 | if (match(V, m_Undef())) { |
703 | Mask.assign(NumElts, -1); |
704 | return std::make_pair( |
705 | PermittedRHS ? UndefValue::get(PermittedRHS->getType()) : V, nullptr); |
706 | } |
707 | |
708 | if (isa<ConstantAggregateZero>(V)) { |
709 | Mask.assign(NumElts, 0); |
710 | return std::make_pair(V, nullptr); |
711 | } |
712 | |
713 | if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { |
714 | |
715 | Value *VecOp = IEI->getOperand(0); |
716 | Value *ScalarOp = IEI->getOperand(1); |
717 | Value *IdxOp = IEI->getOperand(2); |
718 | |
719 | if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { |
720 | if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp)) { |
721 | unsigned ExtractedIdx = |
722 | cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); |
723 | unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); |
724 | |
725 | |
726 | |
727 | if (EI->getOperand(0) == PermittedRHS || PermittedRHS == nullptr) { |
728 | Value *RHS = EI->getOperand(0); |
729 | ShuffleOps LR = collectShuffleElements(VecOp, Mask, RHS, IC); |
730 | assert(LR.second == nullptr || LR.second == RHS); |
731 | |
732 | if (LR.first->getType() != RHS->getType()) { |
733 | |
734 | |
735 | replaceExtractElements(IEI, EI, IC); |
736 | |
737 | |
738 | |
739 | for (unsigned i = 0; i < NumElts; ++i) |
740 | Mask[i] = i; |
741 | return std::make_pair(V, nullptr); |
742 | } |
743 | |
744 | unsigned NumLHSElts = |
745 | cast<FixedVectorType>(RHS->getType())->getNumElements(); |
746 | Mask[InsertedIdx % NumElts] = NumLHSElts + ExtractedIdx; |
747 | return std::make_pair(LR.first, RHS); |
748 | } |
749 | |
750 | if (VecOp == PermittedRHS) { |
751 | |
752 | |
753 | unsigned NumLHSElts = |
754 | cast<FixedVectorType>(EI->getOperand(0)->getType()) |
755 | ->getNumElements(); |
756 | for (unsigned i = 0; i != NumElts; ++i) |
757 | Mask.push_back(i == InsertedIdx ? ExtractedIdx : NumLHSElts + i); |
758 | return std::make_pair(EI->getOperand(0), PermittedRHS); |
759 | } |
760 | |
761 | |
762 | |
763 | if (EI->getOperand(0)->getType() == PermittedRHS->getType() && |
764 | collectSingleShuffleElements(IEI, EI->getOperand(0), PermittedRHS, |
765 | Mask)) |
766 | return std::make_pair(EI->getOperand(0), PermittedRHS); |
767 | } |
768 | } |
769 | } |
770 | |
771 | |
772 | for (unsigned i = 0; i != NumElts; ++i) |
773 | Mask.push_back(i); |
774 | return std::make_pair(V, nullptr); |
775 | } |
776 | |
777 | |
778 | |
779 | |
780 | |
781 | |
782 | Instruction *InstCombinerImpl::foldAggregateConstructionIntoAggregateReuse( |
783 | InsertValueInst &OrigIVI) { |
784 | Type *AggTy = OrigIVI.getType(); |
785 | unsigned NumAggElts; |
786 | switch (AggTy->getTypeID()) { |
787 | case Type::StructTyID: |
788 | NumAggElts = AggTy->getStructNumElements(); |
789 | break; |
790 | case Type::ArrayTyID: |
791 | NumAggElts = AggTy->getArrayNumElements(); |
792 | break; |
793 | default: |
794 | llvm_unreachable("Unhandled aggregate type?"); |
795 | } |
796 | |
797 | |
798 | |
799 | |
800 | assert(NumAggElts > 0 && "Aggregate should have elements."); |
801 | if (NumAggElts > 2) |
802 | return nullptr; |
803 | |
804 | static constexpr auto NotFound = None; |
805 | static constexpr auto FoundMismatch = nullptr; |
806 | |
807 | |
808 | |
809 | SmallVector<Optional<Instruction *>, 2> AggElts(NumAggElts, NotFound); |
810 | |
811 | |
812 | auto KnowAllElts = [&AggElts]() { |
813 | return all_of(AggElts, |
814 | [](Optional<Instruction *> Elt) { return Elt != NotFound; }); |
815 | }; |
816 | |
817 | int Depth = 0; |
818 | |
819 | |
820 | |
821 | static const int DepthLimit = 2 * NumAggElts; |
822 | |
823 | |
824 | |
825 | for (InsertValueInst *CurrIVI = &OrigIVI; |
826 | Depth < DepthLimit && CurrIVI && !KnowAllElts(); |
827 | CurrIVI = dyn_cast<InsertValueInst>(CurrIVI->getAggregateOperand()), |
828 | ++Depth) { |
829 | auto *InsertedValue = |
830 | dyn_cast<Instruction>(CurrIVI->getInsertedValueOperand()); |
831 | if (!InsertedValue) |
832 | return nullptr; |
833 | |
834 | ArrayRef<unsigned int> Indices = CurrIVI->getIndices(); |
835 | |
836 | |
837 | if (Indices.size() != 1) |
838 | return nullptr; |
839 | |
840 | |
841 | |
842 | |
843 | Optional<Instruction *> &Elt = AggElts[Indices.front()]; |
844 | Elt = Elt.getValueOr(InsertedValue); |
845 | |
846 | |
847 | } |
848 | |
849 | |
850 | if (!KnowAllElts()) |
851 | return nullptr; |
852 | |
853 | |
854 | |
855 | |
856 | |
857 | enum class AggregateDescription { |
858 | |
859 | |
860 | NotFound, |
861 | |
862 | |
863 | |
864 | Found, |
865 | |
866 | |
867 | |
868 | |
869 | |
870 | |
871 | FoundMismatch |
872 | }; |
873 | auto Describe = [](Optional<Value *> SourceAggregate) { |
874 | if (SourceAggregate == NotFound) |
875 | return AggregateDescription::NotFound; |
876 | if (*SourceAggregate == FoundMismatch) |
877 | return AggregateDescription::FoundMismatch; |
878 | return AggregateDescription::Found; |
879 | }; |
880 | |
881 | |
882 | |
883 | |
884 | |
885 | |
886 | auto FindSourceAggregate = |
887 | [&](Instruction *Elt, unsigned EltIdx, Optional<BasicBlock *> UseBB, |
888 | Optional<BasicBlock *> PredBB) -> Optional<Value *> { |
889 | |
890 | if (UseBB && PredBB) |
891 | Elt = dyn_cast<Instruction>(Elt->DoPHITranslation(*UseBB, *PredBB)); |
892 | |
893 | |
894 | |
895 | auto *EVI = dyn_cast_or_null<ExtractValueInst>(Elt); |
896 | if (!EVI) |
897 | return NotFound; |
898 | |
899 | Value *SourceAggregate = EVI->getAggregateOperand(); |
900 | |
901 | |
902 | if (SourceAggregate->getType() != AggTy) |
903 | return FoundMismatch; |
904 | |
905 | if (EVI->getNumIndices() != 1 || EltIdx != EVI->getIndices().front()) |
906 | return FoundMismatch; |
907 | |
908 | return SourceAggregate; |
909 | }; |
910 | |
911 | |
912 | |
913 | |
914 | auto FindCommonSourceAggregate = |
915 | [&](Optional<BasicBlock *> UseBB, |
916 | Optional<BasicBlock *> PredBB) -> Optional<Value *> { |
917 | Optional<Value *> SourceAggregate; |
918 | |
919 | for (auto I : enumerate(AggElts)) { |
920 | assert(Describe(SourceAggregate) != AggregateDescription::FoundMismatch && |
921 | "We don't store nullptr in SourceAggregate!"); |
922 | assert((Describe(SourceAggregate) == AggregateDescription::Found) == |
923 | (I.index() != 0) && |
924 | "SourceAggregate should be valid after the first element,"); |
925 | |
926 | |
927 | |
928 | |
929 | Optional<Value *> SourceAggregateForElement = |
930 | FindSourceAggregate(*I.value(), I.index(), UseBB, PredBB); |
931 | |
932 | |
933 | |
934 | |
935 | |
936 | |
937 | if (Describe(SourceAggregateForElement) != AggregateDescription::Found) |
938 | return SourceAggregateForElement; |
939 | |
940 | |
941 | |
942 | switch (Describe(SourceAggregate)) { |
943 | case AggregateDescription::NotFound: |
944 | |
945 | SourceAggregate = SourceAggregateForElement; |
946 | continue; |
947 | case AggregateDescription::Found: |
948 | |
949 | |
950 | if (*SourceAggregateForElement != *SourceAggregate) |
951 | return FoundMismatch; |
952 | continue; |
953 | case AggregateDescription::FoundMismatch: |
954 | llvm_unreachable("Can't happen. We would have early-exited then."); |
955 | }; |
956 | } |
957 | |
958 | assert(Describe(SourceAggregate) == AggregateDescription::Found && |
959 | "Must be a valid Value"); |
960 | return *SourceAggregate; |
961 | }; |
962 | |
963 | Optional<Value *> SourceAggregate; |
964 | |
965 | |
966 | SourceAggregate = FindCommonSourceAggregate(None, None); |
967 | if (Describe(SourceAggregate) != AggregateDescription::NotFound) { |
968 | if (Describe(SourceAggregate) == AggregateDescription::FoundMismatch) |
969 | return nullptr; |
970 | ++NumAggregateReconstructionsSimplified; |
971 | return replaceInstUsesWith(OrigIVI, *SourceAggregate); |
972 | } |
973 | |
974 | |
975 | |
976 | |
977 | |
978 | |
979 | |
980 | |
981 | BasicBlock *UseBB = nullptr; |
982 | |
983 | for (const Optional<Instruction *> &I : AggElts) { |
984 | BasicBlock *BB = (*I)->getParent(); |
985 | |
986 | if (!UseBB) { |
987 | UseBB = BB; |
988 | continue; |
989 | } |
990 | |
991 | if (UseBB != BB) |
992 | return nullptr; |
993 | } |
994 | |
995 | |
996 | |
997 | |
998 | if (!UseBB) |
999 | return nullptr; |
1000 | |
1001 | |
1002 | |
1003 | if (pred_empty(UseBB)) |
1004 | return nullptr; |
1005 | |
1006 | |
1007 | static const int PredCountLimit = 64; |
1008 | |
1009 | |
1010 | |
1011 | SmallVector<BasicBlock *, 4> Preds; |
1012 | for (BasicBlock *Pred : predecessors(UseBB)) { |
1013 | |
1014 | if (Preds.size() >= PredCountLimit) |
1015 | return nullptr; |
1016 | Preds.emplace_back(Pred); |
1017 | } |
1018 | |
1019 | |
1020 | |
1021 | |
1022 | SmallDenseMap<BasicBlock *, Value *, 4> SourceAggregates; |
1023 | for (BasicBlock *Pred : Preds) { |
1024 | std::pair<decltype(SourceAggregates)::iterator, bool> IV = |
1025 | SourceAggregates.insert({Pred, nullptr}); |
1026 | |
1027 | if (!IV.second) |
1028 | continue; |
1029 | |
1030 | |
1031 | |
1032 | |
1033 | SourceAggregate = FindCommonSourceAggregate(UseBB, Pred); |
1034 | if (Describe(SourceAggregate) != AggregateDescription::Found) |
1035 | return nullptr; |
1036 | IV.first->second = *SourceAggregate; |
1037 | } |
1038 | |
1039 | |
1040 | |
1041 | |
1042 | |
1043 | |
1044 | BuilderTy::InsertPointGuard Guard(Builder); |
1045 | Builder.SetInsertPoint(UseBB->getFirstNonPHI()); |
1046 | auto *PHI = |
1047 | Builder.CreatePHI(AggTy, Preds.size(), OrigIVI.getName() + ".merged"); |
1048 | for (BasicBlock *Pred : Preds) |
1049 | PHI->addIncoming(SourceAggregates[Pred], Pred); |
1050 | |
1051 | ++NumAggregateReconstructionsSimplified; |
1052 | return replaceInstUsesWith(OrigIVI, PHI); |
1053 | } |
1054 | |
1055 | |
1056 | |
1057 | |
1058 | |
1059 | |
1060 | |
1061 | |
1062 | Instruction *InstCombinerImpl::visitInsertValueInst(InsertValueInst &I) { |
1063 | bool IsRedundant = false; |
1064 | ArrayRef<unsigned int> FirstIndices = I.getIndices(); |
1065 | |
1066 | |
1067 | |
1068 | |
1069 | |
1070 | Value *V = &I; |
1071 | unsigned Depth = 0; |
1072 | while (V->hasOneUse() && Depth < 10) { |
1073 | User *U = V->user_back(); |
1074 | auto UserInsInst = dyn_cast<InsertValueInst>(U); |
1075 | if (!UserInsInst || U->getOperand(0) != V) |
1076 | break; |
1077 | if (UserInsInst->getIndices() == FirstIndices) { |
1078 | IsRedundant = true; |
1079 | break; |
1080 | } |
1081 | V = UserInsInst; |
1082 | Depth++; |
1083 | } |
1084 | |
1085 | if (IsRedundant) |
1086 | return replaceInstUsesWith(I, I.getOperand(0)); |
1087 | |
1088 | if (Instruction *NewI = foldAggregateConstructionIntoAggregateReuse(I)) |
1089 | return NewI; |
1090 | |
1091 | return nullptr; |
1092 | } |
1093 | |
1094 | static bool isShuffleEquivalentToSelect(ShuffleVectorInst &Shuf) { |
1095 | |
1096 | |
1097 | if (isa<ScalableVectorType>(Shuf.getOperand(0)->getType())) |
1098 | return false; |
1099 | |
1100 | int MaskSize = Shuf.getShuffleMask().size(); |
1101 | int VecSize = |
1102 | cast<FixedVectorType>(Shuf.getOperand(0)->getType())->getNumElements(); |
1103 | |
1104 | |
1105 | if (MaskSize != VecSize) |
1106 | return false; |
1107 | |
1108 | |
1109 | |
1110 | for (int i = 0; i != MaskSize; ++i) { |
1111 | int Elt = Shuf.getMaskValue(i); |
1112 | if (Elt != -1 && Elt != i && Elt != i + VecSize) |
1113 | return false; |
1114 | } |
1115 | |
1116 | return true; |
1117 | } |
1118 | |
1119 | |
1120 | |
1121 | |
1122 | static Instruction *foldInsSequenceIntoSplat(InsertElementInst &InsElt) { |
1123 | |
1124 | |
1125 | if (InsElt.hasOneUse() && isa<InsertElementInst>(InsElt.user_back())) |
1126 | return nullptr; |
1127 | |
1128 | VectorType *VecTy = InsElt.getType(); |
1129 | |
1130 | |
1131 | if (isa<ScalableVectorType>(VecTy)) |
1132 | return nullptr; |
1133 | unsigned NumElements = cast<FixedVectorType>(VecTy)->getNumElements(); |
1134 | |
1135 | |
1136 | |
1137 | if (NumElements == 1) |
1138 | return nullptr; |
1139 | |
1140 | Value *SplatVal = InsElt.getOperand(1); |
1141 | InsertElementInst *CurrIE = &InsElt; |
1142 | SmallBitVector ElementPresent(NumElements, false); |
1143 | InsertElementInst *FirstIE = nullptr; |
1144 | |
1145 | |
1146 | |
1147 | while (CurrIE) { |
1148 | auto *Idx = dyn_cast<ConstantInt>(CurrIE->getOperand(2)); |
1149 | if (!Idx || CurrIE->getOperand(1) != SplatVal) |
1150 | return nullptr; |
1151 | |
1152 | auto *NextIE = dyn_cast<InsertElementInst>(CurrIE->getOperand(0)); |
1153 | |
1154 | |
1155 | |
1156 | if (CurrIE != &InsElt && |
1157 | (!CurrIE->hasOneUse() && (NextIE != nullptr || !Idx->isZero()))) |
1158 | return nullptr; |
1159 | |
1160 | ElementPresent[Idx->getZExtValue()] = true; |
1161 | FirstIE = CurrIE; |
1162 | CurrIE = NextIE; |
1163 | } |
1164 | |
1165 | |
1166 | if (FirstIE == &InsElt) |
1167 | return nullptr; |
1168 | |
1169 | |
1170 | |
1171 | |
1172 | |
1173 | if (!match(FirstIE->getOperand(0), m_Undef())) |
1174 | if (!ElementPresent.all()) |
1175 | return nullptr; |
1176 | |
1177 | |
1178 | Type *Int32Ty = Type::getInt32Ty(InsElt.getContext()); |
1179 | PoisonValue *PoisonVec = PoisonValue::get(VecTy); |
1180 | Constant *Zero = ConstantInt::get(Int32Ty, 0); |
1181 | if (!cast<ConstantInt>(FirstIE->getOperand(2))->isZero()) |
1182 | FirstIE = InsertElementInst::Create(PoisonVec, SplatVal, Zero, "", &InsElt); |
1183 | |
1184 | |
1185 | SmallVector<int, 16> Mask(NumElements, 0); |
1186 | for (unsigned i = 0; i != NumElements; ++i) |
1187 | if (!ElementPresent[i]) |
1188 | Mask[i] = -1; |
1189 | |
1190 | return new ShuffleVectorInst(FirstIE, PoisonVec, Mask); |
1191 | } |
1192 | |
1193 | |
1194 | |
1195 | static Instruction *foldInsEltIntoSplat(InsertElementInst &InsElt) { |
1196 | |
1197 | auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0)); |
1198 | if (!Shuf || !Shuf->isZeroEltSplat()) |
1199 | return nullptr; |
1200 | |
1201 | |
1202 | |
1203 | if (isa<ScalableVectorType>(Shuf->getType())) |
1204 | return nullptr; |
1205 | |
1206 | |
1207 | uint64_t IdxC; |
1208 | if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC))) |
1209 | return nullptr; |
1210 | |
1211 | |
1212 | Value *X = InsElt.getOperand(1); |
1213 | Value *Op0 = Shuf->getOperand(0); |
1214 | if (!match(Op0, m_InsertElt(m_Undef(), m_Specific(X), m_ZeroInt()))) |
1215 | return nullptr; |
1216 | |
1217 | |
1218 | |
1219 | |
1220 | |
1221 | unsigned NumMaskElts = |
1222 | cast<FixedVectorType>(Shuf->getType())->getNumElements(); |
1223 | SmallVector<int, 16> NewMask(NumMaskElts); |
1224 | for (unsigned i = 0; i != NumMaskElts; ++i) |
1225 | NewMask[i] = i == IdxC ? 0 : Shuf->getMaskValue(i); |
1226 | |
1227 | return new ShuffleVectorInst(Op0, UndefValue::get(Op0->getType()), NewMask); |
1228 | } |
1229 | |
1230 | |
1231 | |
1232 | static Instruction *foldInsEltIntoIdentityShuffle(InsertElementInst &InsElt) { |
1233 | |
1234 | auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0)); |
1235 | if (!Shuf || !match(Shuf->getOperand(1), m_Undef()) || |
1236 | !(Shuf->isIdentityWithExtract() || Shuf->isIdentityWithPadding())) |
1237 | return nullptr; |
1238 | |
1239 | |
1240 | |
1241 | if (isa<ScalableVectorType>(Shuf->getType())) |
1242 | return nullptr; |
1243 | |
1244 | |
1245 | uint64_t IdxC; |
1246 | if (!match(InsElt.getOperand(2), m_ConstantInt(IdxC))) |
1247 | return nullptr; |
1248 | |
1249 | |
1250 | |
1251 | Value *Scalar = InsElt.getOperand(1); |
1252 | Value *X = Shuf->getOperand(0); |
1253 | if (!match(Scalar, m_ExtractElt(m_Specific(X), m_SpecificInt(IdxC)))) |
1254 | return nullptr; |
1255 | |
1256 | |
1257 | |
1258 | |
1259 | |
1260 | unsigned NumMaskElts = |
1261 | cast<FixedVectorType>(Shuf->getType())->getNumElements(); |
1262 | SmallVector<int, 16> NewMask(NumMaskElts); |
1263 | ArrayRef<int> OldMask = Shuf->getShuffleMask(); |
1264 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
1265 | if (i != IdxC) { |
1266 | |
1267 | NewMask[i] = OldMask[i]; |
1268 | } else if (OldMask[i] == (int)IdxC) { |
1269 | |
1270 | |
1271 | return nullptr; |
1272 | } else { |
1273 | assert(OldMask[i] == UndefMaskElem && |
1274 | "Unexpected shuffle mask element for identity shuffle"); |
1275 | NewMask[i] = IdxC; |
1276 | } |
1277 | } |
1278 | |
1279 | return new ShuffleVectorInst(X, Shuf->getOperand(1), NewMask); |
1280 | } |
1281 | |
1282 | |
1283 | |
1284 | |
1285 | |
1286 | |
1287 | |
1288 | |
1289 | |
1290 | |
1291 | static Instruction *hoistInsEltConst(InsertElementInst &InsElt2, |
1292 | InstCombiner::BuilderTy &Builder) { |
1293 | auto *InsElt1 = dyn_cast<InsertElementInst>(InsElt2.getOperand(0)); |
1294 | if (!InsElt1 || !InsElt1->hasOneUse()) |
1295 | return nullptr; |
1296 | |
1297 | Value *X, *Y; |
1298 | Constant *ScalarC; |
1299 | ConstantInt *IdxC1, *IdxC2; |
1300 | if (match(InsElt1->getOperand(0), m_Value(X)) && |
1301 | match(InsElt1->getOperand(1), m_Value(Y)) && !isa<Constant>(Y) && |
1302 | match(InsElt1->getOperand(2), m_ConstantInt(IdxC1)) && |
1303 | match(InsElt2.getOperand(1), m_Constant(ScalarC)) && |
1304 | match(InsElt2.getOperand(2), m_ConstantInt(IdxC2)) && IdxC1 != IdxC2) { |
1305 | Value *NewInsElt1 = Builder.CreateInsertElement(X, ScalarC, IdxC2); |
1306 | return InsertElementInst::Create(NewInsElt1, Y, IdxC1); |
1307 | } |
1308 | |
1309 | return nullptr; |
1310 | } |
1311 | |
1312 | |
1313 | |
1314 | static Instruction *foldConstantInsEltIntoShuffle(InsertElementInst &InsElt) { |
1315 | auto *Inst = dyn_cast<Instruction>(InsElt.getOperand(0)); |
1316 | |
1317 | |
1318 | if (!Inst || !Inst->hasOneUse()) |
1319 | return nullptr; |
1320 | if (auto *Shuf = dyn_cast<ShuffleVectorInst>(InsElt.getOperand(0))) { |
1321 | |
1322 | |
1323 | Constant *ShufConstVec, *InsEltScalar; |
1324 | uint64_t InsEltIndex; |
1325 | if (!match(Shuf->getOperand(1), m_Constant(ShufConstVec)) || |
1326 | !match(InsElt.getOperand(1), m_Constant(InsEltScalar)) || |
1327 | !match(InsElt.getOperand(2), m_ConstantInt(InsEltIndex))) |
1328 | return nullptr; |
1329 | |
1330 | |
1331 | |
1332 | |
1333 | |
1334 | |
1335 | if (!isShuffleEquivalentToSelect(*Shuf)) |
1336 | return nullptr; |
1337 | |
1338 | |
1339 | |
1340 | |
1341 | |
1342 | |
1343 | |
1344 | |
1345 | |
1346 | ArrayRef<int> Mask = Shuf->getShuffleMask(); |
1347 | unsigned NumElts = Mask.size(); |
1348 | SmallVector<Constant *, 16> NewShufElts(NumElts); |
1349 | SmallVector<int, 16> NewMaskElts(NumElts); |
1350 | for (unsigned I = 0; I != NumElts; ++I) { |
1351 | if (I == InsEltIndex) { |
1352 | NewShufElts[I] = InsEltScalar; |
1353 | NewMaskElts[I] = InsEltIndex + NumElts; |
1354 | } else { |
1355 | |
1356 | NewShufElts[I] = ShufConstVec->getAggregateElement(I); |
1357 | NewMaskElts[I] = Mask[I]; |
1358 | } |
1359 | } |
1360 | |
1361 | |
1362 | |
1363 | return new ShuffleVectorInst(Shuf->getOperand(0), |
1364 | ConstantVector::get(NewShufElts), NewMaskElts); |
1365 | } else if (auto *IEI = dyn_cast<InsertElementInst>(Inst)) { |
1366 | |
1367 | |
1368 | |
1369 | |
1370 | if (isa<ScalableVectorType>(InsElt.getType())) |
1371 | return nullptr; |
1372 | unsigned NumElts = |
1373 | cast<FixedVectorType>(InsElt.getType())->getNumElements(); |
1374 | |
1375 | uint64_t InsertIdx[2]; |
1376 | Constant *Val[2]; |
1377 | if (!match(InsElt.getOperand(2), m_ConstantInt(InsertIdx[0])) || |
1378 | !match(InsElt.getOperand(1), m_Constant(Val[0])) || |
1379 | !match(IEI->getOperand(2), m_ConstantInt(InsertIdx[1])) || |
1380 | !match(IEI->getOperand(1), m_Constant(Val[1]))) |
1381 | return nullptr; |
1382 | SmallVector<Constant *, 16> Values(NumElts); |
1383 | SmallVector<int, 16> Mask(NumElts); |
1384 | auto ValI = std::begin(Val); |
1385 | |
1386 | |
1387 | |
1388 | for (uint64_t I : InsertIdx) { |
1389 | if (!Values[I]) { |
1390 | Values[I] = *ValI; |
1391 | Mask[I] = NumElts + I; |
1392 | } |
1393 | ++ValI; |
1394 | } |
1395 | |
1396 | for (unsigned I = 0; I < NumElts; ++I) { |
1397 | if (!Values[I]) { |
1398 | Values[I] = UndefValue::get(InsElt.getType()->getElementType()); |
1399 | Mask[I] = I; |
1400 | } |
1401 | } |
1402 | |
1403 | |
1404 | return new ShuffleVectorInst(IEI->getOperand(0), |
1405 | ConstantVector::get(Values), Mask); |
1406 | } |
1407 | return nullptr; |
1408 | } |
1409 | |
1410 | Instruction *InstCombinerImpl::visitInsertElementInst(InsertElementInst &IE) { |
1411 | Value *VecOp = IE.getOperand(0); |
1412 | Value *ScalarOp = IE.getOperand(1); |
1413 | Value *IdxOp = IE.getOperand(2); |
1414 | |
1415 | if (auto *V = SimplifyInsertElementInst( |
1416 | VecOp, ScalarOp, IdxOp, SQ.getWithInstruction(&IE))) |
1417 | return replaceInstUsesWith(IE, V); |
1418 | |
1419 | |
1420 | |
1421 | |
1422 | Value *ScalarSrc; |
1423 | if (match(VecOp, m_Undef()) && |
1424 | match(ScalarOp, m_OneUse(m_BitCast(m_Value(ScalarSrc)))) && |
1425 | (ScalarSrc->getType()->isIntegerTy() || |
1426 | ScalarSrc->getType()->isFloatingPointTy())) { |
1427 | |
1428 | |
1429 | Type *ScalarTy = ScalarSrc->getType(); |
1430 | Type *VecTy = VectorType::get(ScalarTy, IE.getType()->getElementCount()); |
1431 | UndefValue *NewUndef = UndefValue::get(VecTy); |
1432 | Value *NewInsElt = Builder.CreateInsertElement(NewUndef, ScalarSrc, IdxOp); |
1433 | return new BitCastInst(NewInsElt, IE.getType()); |
1434 | } |
1435 | |
1436 | |
1437 | |
1438 | Value *VecSrc; |
1439 | if (match(VecOp, m_BitCast(m_Value(VecSrc))) && |
1440 | match(ScalarOp, m_BitCast(m_Value(ScalarSrc))) && |
1441 | (VecOp->hasOneUse() || ScalarOp->hasOneUse()) && |
1442 | VecSrc->getType()->isVectorTy() && !ScalarSrc->getType()->isVectorTy() && |
1443 | cast<VectorType>(VecSrc->getType())->getElementType() == |
1444 | ScalarSrc->getType()) { |
1445 | |
1446 | |
1447 | Value *NewInsElt = Builder.CreateInsertElement(VecSrc, ScalarSrc, IdxOp); |
1448 | return new BitCastInst(NewInsElt, IE.getType()); |
1449 | } |
1450 | |
1451 | |
1452 | |
1453 | |
1454 | |
1455 | uint64_t InsertedIdx, ExtractedIdx; |
1456 | Value *ExtVecOp; |
1457 | if (isa<FixedVectorType>(IE.getType()) && |
1458 | match(IdxOp, m_ConstantInt(InsertedIdx)) && |
1459 | match(ScalarOp, |
1460 | m_ExtractElt(m_Value(ExtVecOp), m_ConstantInt(ExtractedIdx))) && |
1461 | isa<FixedVectorType>(ExtVecOp->getType()) && |
1462 | ExtractedIdx < |
1463 | cast<FixedVectorType>(ExtVecOp->getType())->getNumElements()) { |
1464 | |
1465 | |
1466 | |
1467 | |
1468 | |
1469 | |
1470 | |
1471 | |
1472 | |
1473 | |
1474 | |
1475 | |
1476 | |
1477 | |
1478 | auto isShuffleRootCandidate = [](InsertElementInst &Insert) { |
1479 | if (!Insert.hasOneUse()) |
1480 | return true; |
1481 | auto *InsertUser = dyn_cast<InsertElementInst>(Insert.user_back()); |
1482 | if (!InsertUser) |
1483 | return true; |
1484 | return false; |
1485 | }; |
1486 | |
1487 | |
1488 | if (isShuffleRootCandidate(IE)) { |
1489 | SmallVector<int, 16> Mask; |
1490 | ShuffleOps LR = collectShuffleElements(&IE, Mask, nullptr, *this); |
1491 | |
1492 | |
1493 | |
1494 | if (LR.first != &IE && LR.second != &IE) { |
1495 | |
1496 | if (LR.second == nullptr) |
1497 | LR.second = UndefValue::get(LR.first->getType()); |
1498 | return new ShuffleVectorInst(LR.first, LR.second, Mask); |
1499 | } |
1500 | } |
1501 | } |
1502 | |
1503 | if (auto VecTy = dyn_cast<FixedVectorType>(VecOp->getType())) { |
1504 | unsigned VWidth = VecTy->getNumElements(); |
1505 | APInt UndefElts(VWidth, 0); |
1506 | APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); |
1507 | if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) { |
1508 | if (V != &IE) |
1509 | return replaceInstUsesWith(IE, V); |
1510 | return &IE; |
1511 | } |
1512 | } |
1513 | |
1514 | if (Instruction *Shuf = foldConstantInsEltIntoShuffle(IE)) |
1515 | return Shuf; |
1516 | |
1517 | if (Instruction *NewInsElt = hoistInsEltConst(IE, Builder)) |
1518 | return NewInsElt; |
1519 | |
1520 | if (Instruction *Broadcast = foldInsSequenceIntoSplat(IE)) |
1521 | return Broadcast; |
1522 | |
1523 | if (Instruction *Splat = foldInsEltIntoSplat(IE)) |
1524 | return Splat; |
1525 | |
1526 | if (Instruction *IdentityShuf = foldInsEltIntoIdentityShuffle(IE)) |
1527 | return IdentityShuf; |
1528 | |
1529 | return nullptr; |
1530 | } |
1531 | |
1532 | |
1533 | |
1534 | static bool canEvaluateShuffled(Value *V, ArrayRef<int> Mask, |
1535 | unsigned Depth = 5) { |
1536 | |
1537 | if (isa<Constant>(V)) |
1538 | return true; |
1539 | |
1540 | |
1541 | Instruction *I = dyn_cast<Instruction>(V); |
1542 | if (!I) return false; |
1543 | |
1544 | |
1545 | if (!I->hasOneUse()) |
1546 | return false; |
1547 | |
1548 | if (Depth == 0) return false; |
1549 | |
1550 | switch (I->getOpcode()) { |
1551 | case Instruction::UDiv: |
1552 | case Instruction::SDiv: |
1553 | case Instruction::URem: |
1554 | case Instruction::SRem: |
1555 | |
1556 | |
1557 | |
1558 | if (llvm::is_contained(Mask, -1)) |
1559 | return false; |
1560 | LLVM_FALLTHROUGH; |
1561 | case Instruction::Add: |
1562 | case Instruction::FAdd: |
1563 | case Instruction::Sub: |
1564 | case Instruction::FSub: |
1565 | case Instruction::Mul: |
1566 | case Instruction::FMul: |
1567 | case Instruction::FDiv: |
1568 | case Instruction::FRem: |
1569 | case Instruction::Shl: |
1570 | case Instruction::LShr: |
1571 | case Instruction::AShr: |
1572 | case Instruction::And: |
1573 | case Instruction::Or: |
1574 | case Instruction::Xor: |
1575 | case Instruction::ICmp: |
1576 | case Instruction::FCmp: |
1577 | case Instruction::Trunc: |
1578 | case Instruction::ZExt: |
1579 | case Instruction::SExt: |
1580 | case Instruction::FPToUI: |
1581 | case Instruction::FPToSI: |
1582 | case Instruction::UIToFP: |
1583 | case Instruction::SIToFP: |
1584 | case Instruction::FPTrunc: |
1585 | case Instruction::FPExt: |
1586 | case Instruction::GetElementPtr: { |
1587 | |
1588 | |
1589 | Type *ITy = I->getType(); |
1590 | if (ITy->isVectorTy() && |
1591 | Mask.size() > cast<FixedVectorType>(ITy)->getNumElements()) |
1592 | return false; |
1593 | for (Value *Operand : I->operands()) { |
1594 | if (!canEvaluateShuffled(Operand, Mask, Depth - 1)) |
1595 | return false; |
1596 | } |
1597 | return true; |
1598 | } |
1599 | case Instruction::InsertElement: { |
1600 | ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(2)); |
1601 | if (!CI) return false; |
1602 | int ElementNumber = CI->getLimitedValue(); |
1603 | |
1604 | |
1605 | |
1606 | bool SeenOnce = false; |
1607 | for (int i = 0, e = Mask.size(); i != e; ++i) { |
1608 | if (Mask[i] == ElementNumber) { |
1609 | if (SeenOnce) |
1610 | return false; |
1611 | SeenOnce = true; |
1612 | } |
1613 | } |
1614 | return canEvaluateShuffled(I->getOperand(0), Mask, Depth - 1); |
1615 | } |
1616 | } |
1617 | return false; |
1618 | } |
1619 | |
1620 | |
1621 | |
1622 | static Value *buildNew(Instruction *I, ArrayRef<Value*> NewOps) { |
1623 | |
1624 | |
1625 | switch (I->getOpcode()) { |
1626 | case Instruction::Add: |
1627 | case Instruction::FAdd: |
1628 | case Instruction::Sub: |
1629 | case Instruction::FSub: |
1630 | case Instruction::Mul: |
1631 | case Instruction::FMul: |
1632 | case Instruction::UDiv: |
1633 | case Instruction::SDiv: |
1634 | case Instruction::FDiv: |
1635 | case Instruction::URem: |
1636 | case Instruction::SRem: |
1637 | case Instruction::FRem: |
1638 | case Instruction::Shl: |
1639 | case Instruction::LShr: |
1640 | case Instruction::AShr: |
1641 | case Instruction::And: |
1642 | case Instruction::Or: |
1643 | case Instruction::Xor: { |
1644 | BinaryOperator *BO = cast<BinaryOperator>(I); |
1645 | assert(NewOps.size() == 2 && "binary operator with #ops != 2"); |
1646 | BinaryOperator *New = |
1647 | BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(), |
1648 | NewOps[0], NewOps[1], "", BO); |
1649 | if (isa<OverflowingBinaryOperator>(BO)) { |
1650 | New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap()); |
1651 | New->setHasNoSignedWrap(BO->hasNoSignedWrap()); |
1652 | } |
1653 | if (isa<PossiblyExactOperator>(BO)) { |
1654 | New->setIsExact(BO->isExact()); |
1655 | } |
1656 | if (isa<FPMathOperator>(BO)) |
1657 | New->copyFastMathFlags(I); |
1658 | return New; |
1659 | } |
1660 | case Instruction::ICmp: |
1661 | assert(NewOps.size() == 2 && "icmp with #ops != 2"); |
1662 | return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(), |
1663 | NewOps[0], NewOps[1]); |
1664 | case Instruction::FCmp: |
1665 | assert(NewOps.size() == 2 && "fcmp with #ops != 2"); |
1666 | return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(), |
1667 | NewOps[0], NewOps[1]); |
1668 | case Instruction::Trunc: |
1669 | case Instruction::ZExt: |
1670 | case Instruction::SExt: |
1671 | case Instruction::FPToUI: |
1672 | case Instruction::FPToSI: |
1673 | case Instruction::UIToFP: |
1674 | case Instruction::SIToFP: |
1675 | case Instruction::FPTrunc: |
1676 | case Instruction::FPExt: { |
1677 | |
1678 | |
1679 | Type *DestTy = VectorType::get( |
1680 | I->getType()->getScalarType(), |
1681 | cast<VectorType>(NewOps[0]->getType())->getElementCount()); |
1682 | assert(NewOps.size() == 1 && "cast with #ops != 1"); |
1683 | return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy, |
1684 | "", I); |
1685 | } |
1686 | case Instruction::GetElementPtr: { |
1687 | Value *Ptr = NewOps[0]; |
1688 | ArrayRef<Value*> Idx = NewOps.slice(1); |
1689 | GetElementPtrInst *GEP = GetElementPtrInst::Create( |
1690 | cast<GetElementPtrInst>(I)->getSourceElementType(), Ptr, Idx, "", I); |
1691 | GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds()); |
1692 | return GEP; |
1693 | } |
1694 | } |
1695 | llvm_unreachable("failed to rebuild vector instructions"); |
1696 | } |
1697 | |
1698 | static Value *evaluateInDifferentElementOrder(Value *V, ArrayRef<int> Mask) { |
1699 | |
1700 | |
1701 | assert(V->getType()->isVectorTy() && "can't reorder non-vector elements"); |
1702 | Type *EltTy = V->getType()->getScalarType(); |
1703 | Type *I32Ty = IntegerType::getInt32Ty(V->getContext()); |
1704 | if (match(V, m_Undef())) |
1705 | return UndefValue::get(FixedVectorType::get(EltTy, Mask.size())); |
1706 | |
1707 | if (isa<ConstantAggregateZero>(V)) |
1708 | return ConstantAggregateZero::get(FixedVectorType::get(EltTy, Mask.size())); |
1709 | |
1710 | if (Constant *C = dyn_cast<Constant>(V)) |
1711 | return ConstantExpr::getShuffleVector(C, PoisonValue::get(C->getType()), |
1712 | Mask); |
1713 | |
1714 | Instruction *I = cast<Instruction>(V); |
1715 | switch (I->getOpcode()) { |
1716 | case Instruction::Add: |
1717 | case Instruction::FAdd: |
1718 | case Instruction::Sub: |
1719 | case Instruction::FSub: |
1720 | case Instruction::Mul: |
1721 | case Instruction::FMul: |
1722 | case Instruction::UDiv: |
1723 | case Instruction::SDiv: |
1724 | case Instruction::FDiv: |
1725 | case Instruction::URem: |
1726 | case Instruction::SRem: |
1727 | case Instruction::FRem: |
1728 | case Instruction::Shl: |
1729 | case Instruction::LShr: |
1730 | case Instruction::AShr: |
1731 | case Instruction::And: |
1732 | case Instruction::Or: |
1733 | case Instruction::Xor: |
1734 | case Instruction::ICmp: |
1735 | case Instruction::FCmp: |
1736 | case Instruction::Trunc: |
1737 | case Instruction::ZExt: |
1738 | case Instruction::SExt: |
1739 | case Instruction::FPToUI: |
1740 | case Instruction::FPToSI: |
1741 | case Instruction::UIToFP: |
1742 | case Instruction::SIToFP: |
1743 | case Instruction::FPTrunc: |
1744 | case Instruction::FPExt: |
1745 | case Instruction::Select: |
1746 | case Instruction::GetElementPtr: { |
1747 | SmallVector<Value*, 8> NewOps; |
1748 | bool NeedsRebuild = |
1749 | (Mask.size() != |
1750 | cast<FixedVectorType>(I->getType())->getNumElements()); |
1751 | for (int i = 0, e = I->getNumOperands(); i != e; ++i) { |
1752 | Value *V; |
1753 | |
1754 | |
1755 | |
1756 | if (I->getOperand(i)->getType()->isVectorTy()) |
1757 | V = evaluateInDifferentElementOrder(I->getOperand(i), Mask); |
1758 | else |
1759 | V = I->getOperand(i); |
1760 | NewOps.push_back(V); |
1761 | NeedsRebuild |= (V != I->getOperand(i)); |
1762 | } |
1763 | if (NeedsRebuild) { |
1764 | return buildNew(I, NewOps); |
1765 | } |
1766 | return I; |
1767 | } |
1768 | case Instruction::InsertElement: { |
1769 | int Element = cast<ConstantInt>(I->getOperand(2))->getLimitedValue(); |
1770 | |
1771 | |
1772 | |
1773 | |
1774 | bool Found = false; |
1775 | int Index = 0; |
1776 | for (int e = Mask.size(); Index != e; ++Index) { |
1777 | if (Mask[Index] == Element) { |
1778 | Found = true; |
1779 | break; |
1780 | } |
1781 | } |
1782 | |
1783 | |
1784 | |
1785 | if (!Found) |
1786 | return evaluateInDifferentElementOrder(I->getOperand(0), Mask); |
1787 | |
1788 | Value *V = evaluateInDifferentElementOrder(I->getOperand(0), Mask); |
1789 | return InsertElementInst::Create(V, I->getOperand(1), |
1790 | ConstantInt::get(I32Ty, Index), "", I); |
1791 | } |
1792 | } |
1793 | llvm_unreachable("failed to reorder elements of vector instruction!"); |
1794 | } |
1795 | |
1796 | |
1797 | |
1798 | |
1799 | |
1800 | |
1801 | |
1802 | static bool isShuffleExtractingFromLHS(ShuffleVectorInst &SVI, |
1803 | ArrayRef<int> Mask) { |
1804 | unsigned LHSElems = |
1805 | cast<FixedVectorType>(SVI.getOperand(0)->getType())->getNumElements(); |
1806 | unsigned MaskElems = Mask.size(); |
1807 | unsigned BegIdx = Mask.front(); |
1808 | unsigned EndIdx = Mask.back(); |
1809 | if (BegIdx > EndIdx || EndIdx >= LHSElems || EndIdx - BegIdx != MaskElems - 1) |
1810 | return false; |
1811 | for (unsigned I = 0; I != MaskElems; ++I) |
1812 | if (static_cast<unsigned>(Mask[I]) != BegIdx + I) |
1813 | return false; |
1814 | return true; |
1815 | } |
1816 | |
1817 | |
1818 | |
1819 | struct BinopElts { |
1820 | BinaryOperator::BinaryOps Opcode; |
1821 | Value *Op0; |
1822 | Value *Op1; |
1823 | BinopElts(BinaryOperator::BinaryOps Opc = (BinaryOperator::BinaryOps)0, |
1824 | Value *V0 = nullptr, Value *V1 = nullptr) : |
1825 | Opcode(Opc), Op0(V0), Op1(V1) {} |
1826 | operator bool() const { return Opcode != 0; } |
1827 | }; |
1828 | |
1829 | |
1830 | |
1831 | |
1832 | |
1833 | static BinopElts getAlternateBinop(BinaryOperator *BO, const DataLayout &DL) { |
1834 | Value *BO0 = BO->getOperand(0), *BO1 = BO->getOperand(1); |
1835 | Type *Ty = BO->getType(); |
1836 | switch (BO->getOpcode()) { |
1837 | case Instruction::Shl: { |
1838 | |
1839 | Constant *C; |
1840 | if (match(BO1, m_Constant(C))) { |
1841 | Constant *ShlOne = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C); |
1842 | return { Instruction::Mul, BO0, ShlOne }; |
1843 | } |
1844 | break; |
1845 | } |
1846 | case Instruction::Or: { |
1847 | |
1848 | const APInt *C; |
1849 | if (match(BO1, m_APInt(C)) && MaskedValueIsZero(BO0, *C, DL)) |
1850 | return { Instruction::Add, BO0, BO1 }; |
1851 | break; |
1852 | } |
1853 | default: |
1854 | break; |
1855 | } |
1856 | return {}; |
1857 | } |
1858 | |
1859 | static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf) { |
1860 | assert(Shuf.isSelect() && "Must have select-equivalent shuffle"); |
1861 | |
1862 | |
1863 | |
1864 | Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); |
1865 | Constant *C; |
1866 | bool Op0IsBinop; |
1867 | if (match(Op0, m_BinOp(m_Specific(Op1), m_Constant(C)))) |
1868 | Op0IsBinop = true; |
1869 | else if (match(Op1, m_BinOp(m_Specific(Op0), m_Constant(C)))) |
1870 | Op0IsBinop = false; |
1871 | else |
1872 | return nullptr; |
1873 | |
1874 | |
1875 | |
1876 | |
1877 | auto *BO = cast<BinaryOperator>(Op0IsBinop ? Op0 : Op1); |
1878 | BinaryOperator::BinaryOps BOpcode = BO->getOpcode(); |
1879 | Constant *IdC = ConstantExpr::getBinOpIdentity(BOpcode, Shuf.getType(), true); |
1880 | if (!IdC) |
1881 | return nullptr; |
1882 | |
1883 | |
1884 | |
1885 | |
1886 | |
1887 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
1888 | Constant *NewC = Op0IsBinop ? ConstantExpr::getShuffleVector(C, IdC, Mask) : |
1889 | ConstantExpr::getShuffleVector(IdC, C, Mask); |
1890 | |
1891 | bool MightCreatePoisonOrUB = |
1892 | is_contained(Mask, UndefMaskElem) && |
1893 | (Instruction::isIntDivRem(BOpcode) || Instruction::isShift(BOpcode)); |
1894 | if (MightCreatePoisonOrUB) |
1895 | NewC = InstCombiner::getSafeVectorConstantForBinop(BOpcode, NewC, true); |
1896 | |
1897 | |
1898 | |
1899 | Value *X = Op0IsBinop ? Op1 : Op0; |
1900 | Instruction *NewBO = BinaryOperator::Create(BOpcode, X, NewC); |
1901 | NewBO->copyIRFlags(BO); |
1902 | |
1903 | |
1904 | |
1905 | |
1906 | if (is_contained(Mask, UndefMaskElem) && !MightCreatePoisonOrUB) |
1907 | NewBO->dropPoisonGeneratingFlags(); |
1908 | return NewBO; |
1909 | } |
1910 | |
1911 | |
1912 | |
1913 | |
1914 | |
1915 | static Instruction *canonicalizeInsertSplat(ShuffleVectorInst &Shuf, |
1916 | InstCombiner::BuilderTy &Builder) { |
1917 | Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); |
1918 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
1919 | Value *X; |
1920 | uint64_t IndexC; |
1921 | |
1922 | |
1923 | if (!match(Op0, m_OneUse(m_InsertElt(m_Undef(), m_Value(X), |
1924 | m_ConstantInt(IndexC)))) || |
1925 | !match(Op1, m_Undef()) || match(Mask, m_ZeroMask()) || IndexC == 0) |
1926 | return nullptr; |
1927 | |
1928 | |
1929 | UndefValue *UndefVec = UndefValue::get(Shuf.getType()); |
1930 | Constant *Zero = Builder.getInt32(0); |
1931 | Value *NewIns = Builder.CreateInsertElement(UndefVec, X, Zero); |
1932 | |
1933 | |
1934 | |
1935 | |
1936 | |
1937 | unsigned NumMaskElts = |
1938 | cast<FixedVectorType>(Shuf.getType())->getNumElements(); |
1939 | SmallVector<int, 16> NewMask(NumMaskElts, 0); |
1940 | for (unsigned i = 0; i != NumMaskElts; ++i) |
1941 | if (Mask[i] == UndefMaskElem) |
1942 | NewMask[i] = Mask[i]; |
1943 | |
1944 | return new ShuffleVectorInst(NewIns, UndefVec, NewMask); |
1945 | } |
1946 | |
1947 | |
1948 | static Instruction *foldSelectShuffle(ShuffleVectorInst &Shuf, |
1949 | InstCombiner::BuilderTy &Builder, |
1950 | const DataLayout &DL) { |
1951 | if (!Shuf.isSelect()) |
1952 | return nullptr; |
1953 | |
1954 | |
1955 | |
1956 | unsigned NumElts = cast<FixedVectorType>(Shuf.getType())->getNumElements(); |
1957 | if (!match(Shuf.getOperand(1), m_Undef()) && |
1958 | Shuf.getMaskValue(0) >= (int)NumElts) { |
1959 | |
1960 | |
1961 | Shuf.commute(); |
1962 | return &Shuf; |
1963 | } |
1964 | |
1965 | if (Instruction *I = foldSelectShuffleWith1Binop(Shuf)) |
1966 | return I; |
1967 | |
1968 | BinaryOperator *B0, *B1; |
1969 | if (!match(Shuf.getOperand(0), m_BinOp(B0)) || |
1970 | !match(Shuf.getOperand(1), m_BinOp(B1))) |
1971 | return nullptr; |
1972 | |
1973 | Value *X, *Y; |
1974 | Constant *C0, *C1; |
1975 | bool ConstantsAreOp1; |
1976 | if (match(B0, m_BinOp(m_Value(X), m_Constant(C0))) && |
1977 | match(B1, m_BinOp(m_Value(Y), m_Constant(C1)))) |
1978 | ConstantsAreOp1 = true; |
1979 | else if (match(B0, m_BinOp(m_Constant(C0), m_Value(X))) && |
1980 | match(B1, m_BinOp(m_Constant(C1), m_Value(Y)))) |
1981 | ConstantsAreOp1 = false; |
1982 | else |
1983 | return nullptr; |
1984 | |
1985 | |
1986 | BinaryOperator::BinaryOps Opc0 = B0->getOpcode(); |
1987 | BinaryOperator::BinaryOps Opc1 = B1->getOpcode(); |
1988 | bool DropNSW = false; |
1989 | if (ConstantsAreOp1 && Opc0 != Opc1) { |
1990 | |
1991 | |
1992 | |
1993 | if (Opc0 == Instruction::Shl || Opc1 == Instruction::Shl) |
1994 | DropNSW = true; |
1995 | if (BinopElts AltB0 = getAlternateBinop(B0, DL)) { |
1996 | assert(isa<Constant>(AltB0.Op1) && "Expecting constant with alt binop"); |
1997 | Opc0 = AltB0.Opcode; |
1998 | C0 = cast<Constant>(AltB0.Op1); |
1999 | } else if (BinopElts AltB1 = getAlternateBinop(B1, DL)) { |
2000 | assert(isa<Constant>(AltB1.Op1) && "Expecting constant with alt binop"); |
2001 | Opc1 = AltB1.Opcode; |
2002 | C1 = cast<Constant>(AltB1.Op1); |
2003 | } |
2004 | } |
2005 | |
2006 | if (Opc0 != Opc1) |
2007 | return nullptr; |
2008 | |
2009 | |
2010 | BinaryOperator::BinaryOps BOpc = Opc0; |
2011 | |
2012 | |
2013 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
2014 | Constant *NewC = ConstantExpr::getShuffleVector(C0, C1, Mask); |
2015 | |
2016 | |
2017 | |
2018 | |
2019 | bool MightCreatePoisonOrUB = |
2020 | is_contained(Mask, UndefMaskElem) && |
2021 | (Instruction::isIntDivRem(BOpc) || Instruction::isShift(BOpc)); |
2022 | if (MightCreatePoisonOrUB) |
2023 | NewC = InstCombiner::getSafeVectorConstantForBinop(BOpc, NewC, |
2024 | ConstantsAreOp1); |
2025 | |
2026 | Value *V; |
2027 | if (X == Y) { |
2028 | |
2029 | |
2030 | |
2031 | V = X; |
2032 | } else { |
2033 | |
2034 | |
2035 | |
2036 | if (!B0->hasOneUse() && !B1->hasOneUse()) |
2037 | return nullptr; |
2038 | |
2039 | |
2040 | |
2041 | |
2042 | |
2043 | |
2044 | |
2045 | if (MightCreatePoisonOrUB && !ConstantsAreOp1) |
2046 | return nullptr; |
2047 | |
2048 | |
2049 | |
2050 | |
2051 | |
2052 | |
2053 | |
2054 | |
2055 | V = Builder.CreateShuffleVector(X, Y, Mask); |
2056 | } |
2057 | |
2058 | Instruction *NewBO = ConstantsAreOp1 ? BinaryOperator::Create(BOpc, V, NewC) : |
2059 | BinaryOperator::Create(BOpc, NewC, V); |
2060 | |
2061 | |
2062 | |
2063 | |
2064 | |
2065 | |
2066 | NewBO->copyIRFlags(B0); |
2067 | NewBO->andIRFlags(B1); |
2068 | if (DropNSW) |
2069 | NewBO->setHasNoSignedWrap(false); |
2070 | if (is_contained(Mask, UndefMaskElem) && !MightCreatePoisonOrUB) |
2071 | NewBO->dropPoisonGeneratingFlags(); |
2072 | return NewBO; |
2073 | } |
2074 | |
2075 | |
2076 | |
2077 | |
2078 | static Instruction *foldTruncShuffle(ShuffleVectorInst &Shuf, |
2079 | bool IsBigEndian) { |
2080 | |
2081 | Type *DestType = Shuf.getType(); |
2082 | Value *X; |
2083 | if (!match(Shuf.getOperand(0), m_BitCast(m_Value(X))) || |
2084 | !match(Shuf.getOperand(1), m_Undef()) || !DestType->isIntOrIntVectorTy()) |
2085 | return nullptr; |
2086 | |
2087 | |
2088 | |
2089 | Type *SrcType = X->getType(); |
2090 | if (!SrcType->isVectorTy() || !SrcType->isIntOrIntVectorTy() || |
2091 | cast<FixedVectorType>(SrcType)->getNumElements() != |
2092 | cast<FixedVectorType>(DestType)->getNumElements() || |
2093 | SrcType->getScalarSizeInBits() % DestType->getScalarSizeInBits() != 0) |
2094 | return nullptr; |
2095 | |
2096 | assert(Shuf.changesLength() && !Shuf.increasesLength() && |
2097 | "Expected a shuffle that decreases length"); |
2098 | |
2099 | |
2100 | |
2101 | uint64_t TruncRatio = |
2102 | SrcType->getScalarSizeInBits() / DestType->getScalarSizeInBits(); |
2103 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
2104 | for (unsigned i = 0, e = Mask.size(); i != e; ++i) { |
2105 | if (Mask[i] == UndefMaskElem) |
2106 | continue; |
2107 | uint64_t LSBIndex = IsBigEndian ? (i + 1) * TruncRatio - 1 : i * TruncRatio; |
2108 | assert(LSBIndex <= INT32_MAX && "Overflowed 32-bits"); |
2109 | if (Mask[i] != (int)LSBIndex) |
2110 | return nullptr; |
2111 | } |
2112 | |
2113 | return new TruncInst(X, DestType); |
2114 | } |
2115 | |
2116 | |
2117 | |
2118 | |
2119 | static Instruction *narrowVectorSelect(ShuffleVectorInst &Shuf, |
2120 | InstCombiner::BuilderTy &Builder) { |
2121 | |
2122 | |
2123 | if (!match(Shuf.getOperand(1), m_Undef()) || !Shuf.isIdentityWithExtract()) |
2124 | return nullptr; |
2125 | |
2126 | |
2127 | |
2128 | Value *Cond, *X, *Y; |
2129 | if (!match(Shuf.getOperand(0), |
2130 | m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) |
2131 | return nullptr; |
2132 | |
2133 | |
2134 | |
2135 | unsigned NarrowNumElts = |
2136 | cast<FixedVectorType>(Shuf.getType())->getNumElements(); |
2137 | Value *NarrowCond; |
2138 | if (!match(Cond, m_OneUse(m_Shuffle(m_Value(NarrowCond), m_Undef()))) || |
2139 | cast<FixedVectorType>(NarrowCond->getType())->getNumElements() != |
2140 | NarrowNumElts || |
2141 | !cast<ShuffleVectorInst>(Cond)->isIdentityWithPadding()) |
2142 | return nullptr; |
2143 | |
2144 | |
2145 | |
2146 | Value *NarrowX = Builder.CreateShuffleVector(X, Shuf.getShuffleMask()); |
2147 | Value *NarrowY = Builder.CreateShuffleVector(Y, Shuf.getShuffleMask()); |
2148 | return SelectInst::Create(NarrowCond, NarrowX, NarrowY); |
2149 | } |
2150 | |
2151 | |
2152 | static Instruction *foldIdentityExtractShuffle(ShuffleVectorInst &Shuf) { |
2153 | Value *Op0 = Shuf.getOperand(0), *Op1 = Shuf.getOperand(1); |
2154 | if (!Shuf.isIdentityWithExtract() || !match(Op1, m_Undef())) |
2155 | return nullptr; |
2156 | |
2157 | |
2158 | |
2159 | Value *X; |
2160 | if (match(Op0, m_BitCast(m_InsertElt(m_Value(), m_Value(X), m_Zero()))) && |
2161 | X->getType()->getPrimitiveSizeInBits() == |
2162 | Shuf.getType()->getPrimitiveSizeInBits()) |
2163 | return new BitCastInst(X, Shuf.getType()); |
2164 | |
2165 | |
2166 | Value *Y; |
2167 | ArrayRef<int> Mask; |
2168 | if (!match(Op0, m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) |
2169 | return nullptr; |
2170 | |
2171 | |
2172 | |
2173 | if (!Op0->hasOneUse()) |
2174 | return nullptr; |
2175 | |
2176 | |
2177 | |
2178 | |
2179 | |
2180 | |
2181 | |
2182 | |
2183 | |
2184 | |
2185 | |
2186 | |
2187 | unsigned NumElts = cast<FixedVectorType>(Shuf.getType())->getNumElements(); |
2188 | SmallVector<int, 16> NewMask(NumElts); |
2189 | assert(NumElts < Mask.size() && |
2190 | "Identity with extract must have less elements than its inputs"); |
2191 | |
2192 | for (unsigned i = 0; i != NumElts; ++i) { |
2193 | int ExtractMaskElt = Shuf.getMaskValue(i); |
2194 | int MaskElt = Mask[i]; |
2195 | NewMask[i] = ExtractMaskElt == UndefMaskElem ? ExtractMaskElt : MaskElt; |
2196 | } |
2197 | return new ShuffleVectorInst(X, Y, NewMask); |
2198 | } |
2199 | |
2200 | |
2201 | |
2202 | static Instruction *foldShuffleWithInsert(ShuffleVectorInst &Shuf, |
2203 | InstCombinerImpl &IC) { |
2204 | Value *V0 = Shuf.getOperand(0), *V1 = Shuf.getOperand(1); |
2205 | SmallVector<int, 16> Mask; |
2206 | Shuf.getShuffleMask(Mask); |
2207 | |
2208 | |
2209 | |
2210 | |
2211 | int NumElts = Mask.size(); |
2212 | if (NumElts != (int)(cast<FixedVectorType>(V0->getType())->getNumElements())) |
2213 | return nullptr; |
2214 | |
2215 | |
2216 | |
2217 | |
2218 | |
2219 | |
2220 | Value *X; |
2221 | uint64_t IdxC; |
2222 | if (match(V0, m_InsertElt(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) { |
2223 | |
2224 | if (!is_contained(Mask, (int)IdxC)) |
2225 | return IC.replaceOperand(Shuf, 0, X); |
2226 | } |
2227 | if (match(V1, m_InsertElt(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) { |
2228 | |
2229 | |
2230 | IdxC += NumElts; |
2231 | |
2232 | if (!is_contained(Mask, (int)IdxC)) |
2233 | return IC.replaceOperand(Shuf, 1, X); |
2234 | } |
2235 | |
2236 | |
2237 | auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) { |
2238 | |
2239 | if (!match(V0, m_InsertElt(m_Value(), m_Value(Scalar), |
2240 | m_ConstantInt(IndexC)))) |
2241 | return false; |
2242 | |
2243 | |
2244 | |
2245 | int NewInsIndex = -1; |
2246 | for (int i = 0; i != NumElts; ++i) { |
2247 | |
2248 | if (Mask[i] == -1) |
2249 | continue; |
2250 | |
2251 | |
2252 | if (Mask[i] == NumElts + i) |
2253 | continue; |
2254 | |
2255 | |
2256 | if (NewInsIndex != -1 || Mask[i] != IndexC->getSExtValue()) |
2257 | return false; |
2258 | |
2259 | |
2260 | NewInsIndex = i; |
2261 | } |
2262 | |
2263 | assert(NewInsIndex != -1 && "Did not fold shuffle with unused operand?"); |
2264 | |
2265 | |
2266 | IndexC = ConstantInt::get(IndexC->getType(), NewInsIndex); |
2267 | return true; |
2268 | }; |
2269 | |
2270 | |
2271 | |
2272 | |
2273 | Value *Scalar; |
2274 | ConstantInt *IndexC; |
2275 | if (isShufflingScalarIntoOp1(Scalar, IndexC)) |
2276 | return InsertElementInst::Create(V1, Scalar, IndexC); |
2277 | |
2278 | |
2279 | |
2280 | |
2281 | std::swap(V0, V1); |
2282 | ShuffleVectorInst::commuteShuffleMask(Mask, NumElts); |
2283 | if (isShufflingScalarIntoOp1(Scalar, IndexC)) |
2284 | return InsertElementInst::Create(V1, Scalar, IndexC); |
2285 | |
2286 | return nullptr; |
2287 | } |
2288 | |
2289 | static Instruction *foldIdentityPaddedShuffles(ShuffleVectorInst &Shuf) { |
2290 | |
2291 | |
2292 | |
2293 | auto *Shuffle0 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(0)); |
2294 | auto *Shuffle1 = dyn_cast<ShuffleVectorInst>(Shuf.getOperand(1)); |
2295 | if (!Shuffle0 || !Shuffle0->isIdentityWithPadding() || |
2296 | !Shuffle1 || !Shuffle1->isIdentityWithPadding()) |
2297 | return nullptr; |
2298 | |
2299 | |
2300 | |
2301 | |
2302 | |
2303 | |
2304 | Value *X = Shuffle0->getOperand(0); |
2305 | Value *Y = Shuffle1->getOperand(0); |
2306 | if (X->getType() != Y->getType() || |
2307 | !isPowerOf2_32(cast<FixedVectorType>(Shuf.getType())->getNumElements()) || |
2308 | !isPowerOf2_32( |
2309 | cast<FixedVectorType>(Shuffle0->getType())->getNumElements()) || |
2310 | !isPowerOf2_32(cast<FixedVectorType>(X->getType())->getNumElements()) || |
2311 | match(X, m_Undef()) || match(Y, m_Undef())) |
2312 | return nullptr; |
2313 | assert(match(Shuffle0->getOperand(1), m_Undef()) && |
2314 | match(Shuffle1->getOperand(1), m_Undef()) && |
2315 | "Unexpected operand for identity shuffle"); |
2316 | |
2317 | |
2318 | |
2319 | |
2320 | |
2321 | int NarrowElts = cast<FixedVectorType>(X->getType())->getNumElements(); |
2322 | int WideElts = cast<FixedVectorType>(Shuffle0->getType())->getNumElements(); |
2323 | assert(WideElts > NarrowElts && "Unexpected types for identity with padding"); |
2324 | |
2325 | ArrayRef<int> Mask = Shuf.getShuffleMask(); |
2326 | SmallVector<int, 16> NewMask(Mask.size(), -1); |
2327 | for (int i = 0, e = Mask.size(); i != e; ++i) { |
2328 | if (Mask[i] == -1) |
2329 | continue; |
2330 | |
2331 | |
2332 | |
2333 | if (Mask[i] < WideElts) { |
2334 | if (Shuffle0->getMaskValue(Mask[i]) == -1) |
2335 | continue; |
2336 | } else { |
2337 | if (Shuffle1->getMaskValue(Mask[i] - WideElts) == -1) |
2338 | continue; |
2339 | } |
2340 | |
2341 | |
2342 | |
2343 | |
2344 | if (Mask[i] < WideElts) { |
2345 | assert(Mask[i] < NarrowElts && "Unexpected shuffle mask"); |
2346 | NewMask[i] = Mask[i]; |
2347 | } else { |
2348 | assert(Mask[i] < (WideElts + NarrowElts) && "Unexpected shuffle mask"); |
2349 | NewMask[i] = Mask[i] - (WideElts - NarrowElts); |
2350 | } |
2351 | } |
2352 | return new ShuffleVectorInst(X, Y, NewMask); |
2353 | } |
2354 | |
2355 | Instruction *InstCombinerImpl::visitShuffleVectorInst(ShuffleVectorInst &SVI) { |
2356 | Value *LHS = SVI.getOperand(0); |
2357 | Value *RHS = SVI.getOperand(1); |
2358 | SimplifyQuery ShufQuery = SQ.getWithInstruction(&SVI); |
2359 | if (auto *V = SimplifyShuffleVectorInst(LHS, RHS, SVI.getShuffleMask(), |
2360 | SVI.getType(), ShufQuery)) |
2361 | return replaceInstUsesWith(SVI, V); |
2362 | |
2363 | |
2364 | if (isa<ScalableVectorType>(LHS->getType())) |
2365 | return nullptr; |
2366 | |
2367 | unsigned VWidth = cast<FixedVectorType>(SVI.getType())->getNumElements(); |
2368 | unsigned LHSWidth = cast<FixedVectorType>(LHS->getType())->getNumElements(); |
2369 | |
2370 | |
2371 | |
2372 | |
2373 | |
2374 | |
2375 | |
2376 | |
2377 | Value *X, *Y; |
2378 | if (match(LHS, m_BitCast(m_Value(X))) && match(RHS, m_BitCast(m_Value(Y))) && |
2379 | X->getType()->isVectorTy() && X->getType() == Y->getType() && |
2380 | X->getType()->getScalarSizeInBits() == |
2381 | SVI.getType()->getScalarSizeInBits() && |
2382 | (LHS->hasOneUse() || RHS->hasOneUse())) { |
2383 | Value *V = Builder.CreateShuffleVector(X, Y, SVI.getShuffleMask(), |
2384 | SVI.getName() + ".uncasted"); |
2385 | return new BitCastInst(V, SVI.getType()); |
2386 | } |
2387 | |
2388 | ArrayRef<int> Mask = SVI.getShuffleMask(); |
2389 | Type *Int32Ty = Type::getInt32Ty(SVI.getContext()); |
2390 | |
2391 | |
2392 | |
2393 | |
2394 | |
2395 | |
2396 | |
2397 | if (match(LHS, m_BitCast(m_Value(X))) && match(RHS, m_Undef()) && |
2398 | X->getType()->isVectorTy() && VWidth == LHSWidth) { |
2399 | |
2400 | auto *XType = cast<FixedVectorType>(X->getType()); |
2401 | unsigned XNumElts = XType->getNumElements(); |
2402 | SmallVector<int, 16> ScaledMask; |
2403 | if (XNumElts >= VWidth) { |
2404 | assert(XNumElts % VWidth == 0 && "Unexpected vector bitcast"); |
2405 | narrowShuffleMaskElts(XNumElts / VWidth, Mask, ScaledMask); |
2406 | } else { |
2407 | assert(VWidth % XNumElts == 0 && "Unexpected vector bitcast"); |
2408 | if (!widenShuffleMaskElts(VWidth / XNumElts, Mask, ScaledMask)) |
2409 | ScaledMask.clear(); |
2410 | } |
2411 | if (!ScaledMask.empty()) { |
2412 | |
2413 | |
2414 | if (auto *V = SimplifyShuffleVectorInst(X, UndefValue::get(XType), |
2415 | ScaledMask, XType, ShufQuery)) |
2416 | return BitCastInst::Create(Instruction::BitCast, V, SVI.getType()); |
2417 | } |
2418 | } |
2419 | |
2420 | |
2421 | if (LHS == RHS) { |
2422 | assert(!match(RHS, m_Undef()) && |
2423 | "Shuffle with 2 undef ops not simplified?"); |
2424 | |
2425 | SmallVector<int, 16> Elts; |
2426 | for (unsigned i = 0; i != VWidth; ++i) { |
2427 | |
2428 | if (Mask[i] < 0) |
2429 | Elts.push_back(UndefMaskElem); |
2430 | else |
2431 | Elts.push_back(Mask[i] % LHSWidth); |
2432 | } |
2433 | return new ShuffleVectorInst(LHS, UndefValue::get(RHS->getType()), Elts); |
2434 | } |
2435 | |
2436 | |
2437 | if (match(LHS, m_Undef())) { |
2438 | SVI.commute(); |
2439 | return &SVI; |
2440 | } |
2441 | |
2442 | if (Instruction *I = canonicalizeInsertSplat(SVI, Builder)) |
2443 | return I; |
2444 | |
2445 | if (Instruction *I = foldSelectShuffle(SVI, Builder, DL)) |
2446 | return I; |
2447 | |
2448 | if (Instruction *I = foldTruncShuffle(SVI, DL.isBigEndian())) |
2449 | return I; |
2450 | |
2451 | if (Instruction *I = narrowVectorSelect(SVI, Builder)) |
2452 | return I; |
2453 | |
2454 | APInt UndefElts(VWidth, 0); |
2455 | APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); |
2456 | if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) { |
2457 | if (V != &SVI) |
2458 | return replaceInstUsesWith(SVI, V); |
2459 | return &SVI; |
2460 | } |
2461 | |
2462 | if (Instruction *I = foldIdentityExtractShuffle(SVI)) |
2463 | return I; |
2464 | |
2465 | |
2466 | |
2467 | if (Instruction *I = foldShuffleWithInsert(SVI, *this)) |
2468 | return I; |
2469 | if (Instruction *I = foldIdentityPaddedShuffles(SVI)) |
2470 | return I; |
2471 | |
2472 | if (match(RHS, m_Undef()) && canEvaluateShuffled(LHS, Mask)) { |
2473 | Value *V = evaluateInDifferentElementOrder(LHS, Mask); |
2474 | return replaceInstUsesWith(SVI, V); |
2475 | } |
2476 | |
2477 | |
2478 | |
2479 | |
2480 | |
2481 | |
2482 | |
2483 | |
2484 | |
2485 | |
2486 | |
2487 | |
2488 | |
2489 | |
2490 | |
2491 | |
2492 | |
2493 | |
2494 | |
2495 | |
2496 | |
2497 | |
2498 | |
2499 | |
2500 | |
2501 | |
2502 | |
2503 | |
2504 | |
2505 | |
2506 | bool MadeChange = false; |
2507 | if (isShuffleExtractingFromLHS(SVI, Mask)) { |
2508 | Value *V = LHS; |
2509 | unsigned MaskElems = Mask.size(); |
2510 | auto *SrcTy = cast<FixedVectorType>(V->getType()); |
2511 | unsigned VecBitWidth = SrcTy->getPrimitiveSizeInBits().getFixedSize(); |
2512 | unsigned SrcElemBitWidth = DL.getTypeSizeInBits(SrcTy->getElementType()); |
2513 | assert(SrcElemBitWidth && "vector elements must have a bitwidth"); |
2514 | unsigned SrcNumElems = SrcTy->getNumElements(); |
2515 | SmallVector<BitCastInst *, 8> BCs; |
2516 | DenseMap<Type *, Value *> NewBCs; |
2517 | for (User *U : SVI.users()) |
2518 | if (BitCastInst *BC = dyn_cast<BitCastInst>(U)) |
2519 | if (!BC->use_empty()) |
2520 | |
2521 | BCs.push_back(BC); |
2522 | for (BitCastInst *BC : BCs) { |
2523 | unsigned BegIdx = Mask.front(); |
2524 | Type *TgtTy = BC->getDestTy(); |
2525 | unsigned TgtElemBitWidth = DL.getTypeSizeInBits(TgtTy); |
2526 | if (!TgtElemBitWidth) |
2527 | continue; |
2528 | unsigned TgtNumElems = VecBitWidth / TgtElemBitWidth; |
2529 | bool VecBitWidthsEqual = VecBitWidth == TgtNumElems * TgtElemBitWidth; |
2530 | bool BegIsAligned = 0 == ((SrcElemBitWidth * BegIdx) % TgtElemBitWidth); |
2531 | if (!VecBitWidthsEqual) |
2532 | continue; |
2533 | if (!VectorType::isValidElementType(TgtTy)) |
2534 | continue; |
2535 | auto *CastSrcTy = FixedVectorType::get(TgtTy, TgtNumElems); |
2536 | if (!BegIsAligned) { |
2537 | |
2538 | |
2539 | SmallVector<int, 16> ShuffleMask(SrcNumElems, -1); |
2540 | for (unsigned I = 0, E = MaskElems, Idx = BegIdx; I != E; ++Idx, ++I) |
2541 | ShuffleMask[I] = Idx; |
2542 | V = Builder.CreateShuffleVector(V, ShuffleMask, |
2543 | SVI.getName() + ".extract"); |
2544 | BegIdx = 0; |
2545 | } |
2546 | unsigned SrcElemsPerTgtElem = TgtElemBitWidth / SrcElemBitWidth; |
2547 | assert(SrcElemsPerTgtElem); |
2548 | BegIdx /= SrcElemsPerTgtElem; |
2549 | bool BCAlreadyExists = NewBCs.find(CastSrcTy) != NewBCs.end(); |
2550 | auto *NewBC = |
2551 | BCAlreadyExists |
2552 | ? NewBCs[CastSrcTy] |
2553 | : Builder.CreateBitCast(V, CastSrcTy, SVI.getName() + ".bc"); |
2554 | if (!BCAlreadyExists) |
2555 | NewBCs[CastSrcTy] = NewBC; |
2556 | auto *Ext = Builder.CreateExtractElement( |
2557 | NewBC, ConstantInt::get(Int32Ty, BegIdx), SVI.getName() + ".extract"); |
2558 | |
2559 | |
2560 | replaceInstUsesWith(*BC, Ext); |
2561 | MadeChange = true; |
2562 | } |
2563 | } |
2564 | |
2565 | |
2566 | |
2567 | |
2568 | |
2569 | |
2570 | |
2571 | |
2572 | |
2573 | |
2574 | |
2575 | |
2576 | |
2577 | |
2578 | |
2579 | |
2580 | |
2581 | |
2582 | |
2583 | |
2584 | |
2585 | |
2586 | |
2587 | |
2588 | |
2589 | |
2590 | |
2591 | |
2592 | |
2593 | |
2594 | |
2595 | |
2596 | |
2597 | |
2598 | |
2599 | |
2600 | |
2601 | |
2602 | |
2603 | |
2604 | |
2605 | |
2606 | |
2607 | |
2608 | ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS); |
2609 | ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS); |
2610 | if (LHSShuffle) |
2611 | if (!match(LHSShuffle->getOperand(1), m_Undef()) && !match(RHS, m_Undef())) |
2612 | LHSShuffle = nullptr; |
2613 | if (RHSShuffle) |
2614 | if (!match(RHSShuffle->getOperand(1), m_Undef())) |
2615 | RHSShuffle = nullptr; |
2616 | if (!LHSShuffle && !RHSShuffle) |
2617 | return MadeChange ? &SVI : nullptr; |
2618 | |
2619 | Value* LHSOp0 = nullptr; |
2620 | Value* LHSOp1 = nullptr; |
2621 | Value* RHSOp0 = nullptr; |
2622 | unsigned LHSOp0Width = 0; |
2623 | unsigned RHSOp0Width = 0; |
2624 | if (LHSShuffle) { |
2625 | LHSOp0 = LHSShuffle->getOperand(0); |
2626 | LHSOp1 = LHSShuffle->getOperand(1); |
2627 | LHSOp0Width = cast<FixedVectorType>(LHSOp0->getType())->getNumElements(); |
2628 | } |
2629 | if (RHSShuffle) { |
2630 | RHSOp0 = RHSShuffle->getOperand(0); |
2631 | RHSOp0Width = cast<FixedVectorType>(RHSOp0->getType())->getNumElements(); |
2632 | } |
2633 | Value* newLHS = LHS; |
2634 | Value* newRHS = RHS; |
2635 | if (LHSShuffle) { |
2636 | |
2637 | if (match(RHS, m_Undef())) { |
2638 | newLHS = LHSOp0; |
2639 | newRHS = LHSOp1; |
2640 | } |
2641 | |
2642 | else if (LHSOp0Width == LHSWidth) { |
2643 | newLHS = LHSOp0; |
2644 | } |
2645 | } |
2646 | |
2647 | if (RHSShuffle && RHSOp0Width == LHSWidth) { |
2648 | newRHS = RHSOp0; |
2649 | } |
2650 | |
2651 | if (LHSOp0 == RHSOp0) { |
2652 | newLHS = LHSOp0; |
2653 | newRHS = nullptr; |
2654 | } |
2655 | |
2656 | if (newLHS == LHS && newRHS == RHS) |
2657 | return MadeChange ? &SVI : nullptr; |
2658 | |
2659 | ArrayRef<int> LHSMask; |
2660 | ArrayRef<int> RHSMask; |
2661 | if (newLHS != LHS) |
2662 | LHSMask = LHSShuffle->getShuffleMask(); |
2663 | if (RHSShuffle && newRHS != RHS) |
2664 | RHSMask = RHSShuffle->getShuffleMask(); |
2665 | |
2666 | unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; |
2667 | SmallVector<int, 16> newMask; |
2668 | bool isSplat = true; |
2669 | int SplatElt = -1; |
2670 | |
2671 | |
2672 | for (unsigned i = 0; i < VWidth; ++i) { |
2673 | int eltMask; |
2674 | if (Mask[i] < 0) { |
2675 | |
2676 | eltMask = -1; |
2677 | } else if (Mask[i] < (int)LHSWidth) { |
2678 | |
2679 | |
2680 | |
2681 | |
2682 | if (newLHS != LHS) { |
2683 | eltMask = LHSMask[Mask[i]]; |
2684 | |
2685 | |
2686 | if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1)) |
2687 | eltMask = -1; |
2688 | } else |
2689 | eltMask = Mask[i]; |
2690 | } else { |
2691 | |
2692 | |
2693 | |
2694 | |
2695 | if (match(RHS, m_Undef())) |
2696 | eltMask = -1; |
2697 | |
2698 | |
2699 | else if (newRHS != RHS) { |
2700 | eltMask = RHSMask[Mask[i]-LHSWidth]; |
2701 | |
2702 | |
2703 | if (eltMask >= (int)RHSOp0Width) { |
2704 | assert(match(RHSShuffle->getOperand(1), m_Undef()) && |
2705 | "should have been check above"); |
2706 | eltMask = -1; |
2707 | } |
2708 | } else |
2709 | eltMask = Mask[i]-LHSWidth; |
2710 | |
2711 | |
2712 | |
2713 | |
2714 | |
2715 | |
2716 | |
2717 | if (eltMask >= 0 && newRHS != nullptr && newLHS != newRHS) |
2718 | eltMask += newLHSWidth; |
2719 | } |
2720 | |
2721 | |
2722 | if (eltMask >= 0) { |
2723 | if (SplatElt >= 0 && SplatElt != eltMask) |
2724 | isSplat = false; |
2725 | SplatElt = eltMask; |
2726 | } |
2727 | |
2728 | newMask.push_back(eltMask); |
2729 | } |
2730 | |
2731 | |
2732 | |
2733 | if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { |
2734 | if (!newRHS) |
2735 | newRHS = UndefValue::get(newLHS->getType()); |
2736 | return new ShuffleVectorInst(newLHS, newRHS, newMask); |
2737 | } |
2738 | |
2739 | return MadeChange ? &SVI : nullptr; |
2740 | } |
1 | |
2 | |
3 | |
4 | |
5 | |
6 | |
7 | |
8 | |
9 | |
10 | |
11 | |
12 | |
13 | |
14 | |
15 | |
16 | |
17 | #ifndef LLVM_IR_DERIVEDTYPES_H |
18 | #define LLVM_IR_DERIVEDTYPES_H |
19 | |
20 | #include "llvm/ADT/ArrayRef.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/StringRef.h" |
23 | #include "llvm/IR/Type.h" |
24 | #include "llvm/Support/Casting.h" |
25 | #include "llvm/Support/Compiler.h" |
26 | #include "llvm/Support/TypeSize.h" |
27 | #include <cassert> |
28 | #include <cstdint> |
29 | |
30 | namespace llvm { |
31 | |
32 | class Value; |
33 | class APInt; |
34 | class LLVMContext; |
35 | |
36 | |
37 | |
38 | |
39 | |
40 | class IntegerType : public Type { |
41 | friend class LLVMContextImpl; |
42 | |
43 | protected: |
44 | explicit IntegerType(LLVMContext &C, unsigned NumBits) : Type(C, IntegerTyID){ |
45 | setSubclassData(NumBits); |
46 | } |
47 | |
48 | public: |
49 | |
50 | enum { |
51 | MIN_INT_BITS = 1, |
52 | MAX_INT_BITS = (1<<24)-1 |
53 | |
54 | |
55 | |
56 | }; |
57 | |
58 | |
59 | |
60 | |
61 | |
62 | |
63 | static IntegerType *get(LLVMContext &C, unsigned NumBits); |
64 | |
65 | |
66 | IntegerType *getExtendedType() const { |
67 | return Type::getIntNTy(getContext(), 2 * getScalarSizeInBits()); |
68 | } |
69 | |
70 | |
71 | unsigned getBitWidth() const { return getSubclassData(); } |
72 | |
73 | |
74 | |
75 | uint64_t getBitMask() const { |
76 | return ~uint64_t(0UL) >> (64-getBitWidth()); |
77 | } |
78 | |
79 | |
80 | |
81 | uint64_t getSignBit() const { |
82 | return 1ULL << (getBitWidth()-1); |
83 | } |
84 | |
85 | |
86 | |
87 | |
88 | APInt getMask() const; |
89 | |
90 | |
91 | static bool classof(const Type *T) { |
92 | return T->getTypeID() == IntegerTyID; |
93 | } |
94 | }; |
95 | |
96 | unsigned Type::getIntegerBitWidth() const { |
97 | return cast<IntegerType>(this)->getBitWidth(); |
98 | } |
99 | |
100 | |
101 | |
102 | class FunctionType : public Type { |
103 | FunctionType(Type *Result, ArrayRef<Type*> Params, bool IsVarArgs); |
104 | |
105 | public: |
106 | FunctionType(const FunctionType &) = delete; |
107 | FunctionType &operator=(const FunctionType &) = delete; |
108 | |
109 | |
110 | static FunctionType *get(Type *Result, |
111 | ArrayRef<Type*> Params, bool isVarArg); |
112 | |
113 | |
114 | static FunctionType *get(Type *Result, bool isVarArg); |
115 | |
116 | |
117 | static bool isValidReturnType(Type *RetTy); |
118 | |
119 | |
120 | static bool isValidArgumentType(Type *ArgTy); |
121 | |
122 | bool isVarArg() const { return getSubclassData()!=0; } |
123 | Type *getReturnType() const { return ContainedTys[0]; } |
124 | |
125 | using param_iterator = Type::subtype_iterator; |
126 | |
127 | param_iterator param_begin() const { return ContainedTys + 1; } |
128 | param_iterator param_end() const { return &ContainedTys[NumContainedTys]; } |
129 | ArrayRef<Type *> params() const { |
130 | return makeArrayRef(param_begin(), param_end()); |
131 | } |
132 | |
133 | |
134 | Type *getParamType(unsigned i) const { return ContainedTys[i+1]; } |
135 | |
136 | |
137 | |
138 | unsigned getNumParams() const { return NumContainedTys - 1; } |
139 | |
140 | |
141 | static bool classof(const Type *T) { |
142 | return T->getTypeID() == FunctionTyID; |
143 | } |
144 | }; |
145 | static_assert(alignof(FunctionType) >= alignof(Type *), |
146 | "Alignment sufficient for objects appended to FunctionType"); |
147 | |
148 | bool Type::isFunctionVarArg() const { |
149 | return cast<FunctionType>(this)->isVarArg(); |
150 | } |
151 | |
152 | Type *Type::getFunctionParamType(unsigned i) const { |
153 | return cast<FunctionType>(this)->getParamType(i); |
154 | } |
155 | |
156 | unsigned Type::getFunctionNumParams() const { |
157 | return cast<FunctionType>(this)->getNumParams(); |
158 | } |
159 | |
160 | |
161 | |
162 | |
163 | |
164 | class FunctionCallee { |
165 | public: |
166 | |
167 | |
168 | template <typename T, typename U = decltype(&T::getFunctionType)> |
169 | FunctionCallee(T *Fn) |
170 | : FnTy(Fn ? Fn->getFunctionType() : nullptr), Callee(Fn) {} |
171 | |
172 | FunctionCallee(FunctionType *FnTy, Value *Callee) |
173 | : FnTy(FnTy), Callee(Callee) { |
174 | assert((FnTy == nullptr) == (Callee == nullptr)); |
175 | } |
176 | |
177 | FunctionCallee(std::nullptr_t) {} |
178 | |
179 | FunctionCallee() = default; |
180 | |
181 | FunctionType *getFunctionType() { return FnTy; } |
182 | |
183 | Value *getCallee() { return Callee; } |
184 | |
185 | explicit operator bool() { return Callee; } |
186 | |
187 | private: |
188 | FunctionType *FnTy = nullptr; |
189 | Value *Callee = nullptr; |
190 | }; |
191 | |
192 | |
193 | |
194 | |
195 | |
196 | |
197 | |
198 | |
199 | |
200 | |
201 | |
202 | |
203 | |
204 | |
205 | |
206 | |
207 | |
208 | |
209 | |
210 | |
211 | |
212 | class StructType : public Type { |
213 | StructType(LLVMContext &C) : Type(C, StructTyID) {} |
214 | |
215 | enum { |
216 | |
217 | SCDB_HasBody = 1, |
218 | SCDB_Packed = 2, |
219 | SCDB_IsLiteral = 4, |
220 | SCDB_IsSized = 8 |
221 | }; |
222 | |
223 | |
224 | |
225 | |
226 | |
227 | void *SymbolTableEntry = nullptr; |
228 | |
229 | public: |
230 | StructType(const StructType &) = delete; |
231 | StructType &operator=(const StructType &) = delete; |
232 | |
233 | |
234 | static StructType *create(LLVMContext &Context, StringRef Name); |
235 | static StructType *create(LLVMContext &Context); |
236 | |
237 | static StructType *create(ArrayRef<Type *> Elements, StringRef Name, |
238 | bool isPacked = false); |
239 | static StructType *create(ArrayRef<Type *> Elements); |
240 | static StructType *create(LLVMContext &Context, ArrayRef<Type *> Elements, |
241 | StringRef Name, bool isPacked = false); |
242 | static StructType *create(LLVMContext &Context, ArrayRef<Type *> Elements); |
243 | template <class... Tys> |
244 | static std::enable_if_t<are_base_of<Type, Tys...>::value, StructType *> |
245 | create(StringRef Name, Type *elt1, Tys *... elts) { |
246 | assert(elt1 && "Cannot create a struct type with no elements with this"); |
247 | return create(ArrayRef<Type *>({elt1, elts...}), Name); |
248 | } |
249 | |
250 | |
251 | static StructType *get(LLVMContext &Context, ArrayRef<Type*> Elements, |
252 | bool isPacked = false); |
253 | |
254 | |
255 | static StructType *get(LLVMContext &Context, bool isPacked = false); |
256 | |
257 | |
258 | |
259 | |
260 | template <class... Tys> |
261 | static std::enable_if_t<are_base_of<Type, Tys...>::value, StructType *> |
262 | get(Type *elt1, Tys *... elts) { |
263 | assert(elt1 && "Cannot create a struct type with no elements with this"); |
264 | LLVMContext &Ctx = elt1->getContext(); |
265 | return StructType::get(Ctx, ArrayRef<Type *>({elt1, elts...})); |
266 | } |
267 | |
268 | |
269 | |
270 | static StructType *getTypeByName(LLVMContext &C, StringRef Name); |
271 | |
272 | bool isPacked() const { return (getSubclassData() & SCDB_Packed) != 0; } |
273 | |
274 | |
275 | |
276 | bool isLiteral() const { return (getSubclassData() & SCDB_IsLiteral) != 0; } |
277 | |
278 | |
279 | |
280 | bool isOpaque() const { return (getSubclassData() & SCDB_HasBody) == 0; } |
281 | |
282 | |
283 | bool isSized(SmallPtrSetImpl<Type *> *Visited = nullptr) const; |
284 | |
285 | |
286 | bool containsScalableVectorType() const; |
287 | |
288 | |
289 | bool hasName() const { return SymbolTableEntry != nullptr; } |
290 | |
291 | |
292 | |
293 | |
294 | StringRef getName() const; |
295 | |
296 | |
297 | |
298 | void setName(StringRef Name); |
299 | |
300 | |
301 | void setBody(ArrayRef<Type*> Elements, bool isPacked = false); |
302 | |
303 | template <typename... Tys> |
304 | std::enable_if_t<are_base_of<Type, Tys...>::value, void> |
305 | setBody(Type *elt1, Tys *... elts) { |
306 | assert(elt1 && "Cannot create a struct type with no elements with this"); |
307 | setBody(ArrayRef<Type *>({elt1, elts...})); |
308 | } |
309 | |
310 | |
311 | static bool isValidElementType(Type *ElemTy); |
312 | |
313 | |
314 | using element_iterator = Type::subtype_iterator; |
315 | |
316 | element_iterator element_begin() const { return ContainedTys; } |
317 | element_iterator element_end() const { return &ContainedTys[NumContainedTys];} |
318 | ArrayRef<Type *> elements() const { |
319 | return makeArrayRef(element_begin(), element_end()); |
320 | } |
321 | |
322 | |
323 | bool isLayoutIdentical(StructType *Other) const; |
324 | |
325 | |
326 | unsigned getNumElements() const { return NumContainedTys; } |
327 | Type *getElementType(unsigned N) const { |
328 | assert(N < NumContainedTys && "Element number out of range!"); |
329 | return ContainedTys[N]; |
330 | } |
331 | |
332 | Type *getTypeAtIndex(const Value *V) const; |
333 | Type *getTypeAtIndex(unsigned N) const { return getElementType(N); } |
334 | bool indexValid(const Value *V) const; |
335 | bool indexValid(unsigned Idx) const { return Idx < getNumElements(); } |
336 | |
337 | |
338 | static bool classof(const Type *T) { |
339 | return T->getTypeID() == StructTyID; |
340 | } |
341 | }; |
342 | |
343 | StringRef Type::getStructName() const { |
344 | return cast<StructType>(this)->getName(); |
345 | } |
346 | |
347 | unsigned Type::getStructNumElements() const { |
348 | return cast<StructType>(this)->getNumElements(); |
349 | } |
350 | |
351 | Type *Type::getStructElementType(unsigned N) const { |
352 | return cast<StructType>(this)->getElementType(N); |
353 | } |
354 | |
355 | |
356 | class ArrayType : public Type { |
357 | |
358 | Type *ContainedType; |
359 | |
360 | uint64_t NumElements; |
361 | |
362 | ArrayType(Type *ElType, uint64_t NumEl); |
363 | |
364 | public: |
365 | ArrayType(const ArrayType &) = delete; |
366 | ArrayType &operator=(const ArrayType &) = delete; |
367 | |
368 | uint64_t getNumElements() const { return NumElements; } |
369 | Type *getElementType() const { return ContainedType; } |
370 | |
371 | |
372 | static ArrayType *get(Type *ElementType, uint64_t NumElements); |
373 | |
374 | |
375 | static bool isValidElementType(Type *ElemTy); |
376 | |
377 | |
378 | static bool classof(const Type *T) { |
379 | return T->getTypeID() == ArrayTyID; |
380 | } |
381 | }; |
382 | |
383 | uint64_t Type::getArrayNumElements() const { |
384 | return cast<ArrayType>(this)->getNumElements(); |
385 | } |
386 | |
387 | |
388 | class VectorType : public Type { |
389 | |
390 | |
391 | |
392 | |
393 | |
394 | |
395 | |
396 | |
397 | |
398 | |
399 | |
400 | |
401 | |
402 | |
403 | Type *ContainedType; |
404 | |
405 | protected: |
406 | |
407 | |
408 | |
409 | |
410 | |
411 | |
412 | |
413 | const unsigned ElementQuantity; |
414 | |
415 | VectorType(Type *ElType, unsigned EQ, Type::TypeID TID); |
416 | |
417 | public: |
418 | VectorType(const VectorType &) = delete; |
419 | VectorType &operator=(const VectorType &) = delete; |
420 | |
421 | Type *getElementType() const { return ContainedType; } |
422 | |
423 | |
424 | static VectorType *get(Type *ElementType, ElementCount EC); |
425 | |
426 | static VectorType *get(Type *ElementType, unsigned NumElements, |
427 | bool Scalable) { |
428 | return VectorType::get(ElementType, |
429 | ElementCount::get(NumElements, Scalable)); |
430 | } |
431 | |
432 | static VectorType *get(Type *ElementType, const VectorType *Other) { |
433 | return VectorType::get(ElementType, Other->getElementCount()); |
434 | } |
435 | |
436 | |
437 | |
438 | |
439 | static VectorType *getInteger(VectorType *VTy) { |
440 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
441 | assert(EltBits && "Element size must be of a non-zero size"); |
442 | Type *EltTy = IntegerType::get(VTy->getContext(), EltBits); |
443 | return VectorType::get(EltTy, VTy->getElementCount()); |
444 | } |
445 | |
446 | |
447 | |
448 | static VectorType *getExtendedElementVectorType(VectorType *VTy) { |
449 | assert(VTy->isIntOrIntVectorTy() && "VTy expected to be a vector of ints."); |
450 | auto *EltTy = cast<IntegerType>(VTy->getElementType()); |
451 | return VectorType::get(EltTy->getExtendedType(), VTy->getElementCount()); |
452 | } |
453 | |
454 | |
455 | |
456 | |
457 | static VectorType *getTruncatedElementVectorType(VectorType *VTy) { |
458 | Type *EltTy; |
459 | if (VTy->getElementType()->isFloatingPointTy()) { |
460 | switch(VTy->getElementType()->getTypeID()) { |
461 | case DoubleTyID: |
462 | EltTy = Type::getFloatTy(VTy->getContext()); |
463 | break; |
464 | case FloatTyID: |
465 | EltTy = Type::getHalfTy(VTy->getContext()); |
466 | break; |
467 | default: |
468 | llvm_unreachable("Cannot create narrower fp vector element type"); |
469 | } |
470 | } else { |
471 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
472 | assert((EltBits & 1) == 0 && |
473 | "Cannot truncate vector element with odd bit-width"); |
474 | EltTy = IntegerType::get(VTy->getContext(), EltBits / 2); |
475 | } |
476 | return VectorType::get(EltTy, VTy->getElementCount()); |
477 | } |
478 | |
479 | |
480 | |
481 | |
482 | static VectorType *getSubdividedVectorType(VectorType *VTy, int NumSubdivs) { |
483 | for (int i = 0; i < NumSubdivs; ++i) { |
484 | VTy = VectorType::getDoubleElementsVectorType(VTy); |
485 | VTy = VectorType::getTruncatedElementVectorType(VTy); |
486 | } |
487 | return VTy; |
488 | } |
489 | |
490 | |
491 | |
492 | static VectorType *getHalfElementsVectorType(VectorType *VTy) { |
493 | auto EltCnt = VTy->getElementCount(); |
494 | assert(EltCnt.isKnownEven() && |
495 | "Cannot halve vector with odd number of elements."); |
496 | return VectorType::get(VTy->getElementType(), |
497 | EltCnt.divideCoefficientBy(2)); |
498 | } |
499 | |
500 | |
501 | |
502 | static VectorType *getDoubleElementsVectorType(VectorType *VTy) { |
503 | auto EltCnt = VTy->getElementCount(); |
504 | assert((EltCnt.getKnownMinValue() * 2ull) <= UINT_MAX && |
505 | "Too many elements in vector"); |
506 | return VectorType::get(VTy->getElementType(), EltCnt * 2); |
507 | } |
508 | |
509 | |
510 | static bool isValidElementType(Type *ElemTy); |
511 | |
512 | |
513 | |
514 | inline ElementCount getElementCount() const; |
515 | |
516 | |
517 | static bool classof(const Type *T) { |
518 | return T->getTypeID() == FixedVectorTyID || |
519 | T->getTypeID() == ScalableVectorTyID; |
520 | } |
521 | }; |
522 | |
523 | |
524 | class FixedVectorType : public VectorType { |
525 | protected: |
526 | FixedVectorType(Type *ElTy, unsigned NumElts) |
527 | : VectorType(ElTy, NumElts, FixedVectorTyID) {} |
528 | |
529 | public: |
530 | static FixedVectorType *get(Type *ElementType, unsigned NumElts); |
531 | |
532 | static FixedVectorType *get(Type *ElementType, const FixedVectorType *FVTy) { |
533 | return get(ElementType, FVTy->getNumElements()); |
534 | } |
535 | |
536 | static FixedVectorType *getInteger(FixedVectorType *VTy) { |
537 | return cast<FixedVectorType>(VectorType::getInteger(VTy)); |
538 | } |
539 | |
540 | static FixedVectorType *getExtendedElementVectorType(FixedVectorType *VTy) { |
541 | return cast<FixedVectorType>(VectorType::getExtendedElementVectorType(VTy)); |
542 | } |
543 | |
544 | static FixedVectorType *getTruncatedElementVectorType(FixedVectorType *VTy) { |
545 | return cast<FixedVectorType>( |
546 | VectorType::getTruncatedElementVectorType(VTy)); |
547 | } |
548 | |
549 | static FixedVectorType *getSubdividedVectorType(FixedVectorType *VTy, |
550 | int NumSubdivs) { |
551 | return cast<FixedVectorType>( |
552 | VectorType::getSubdividedVectorType(VTy, NumSubdivs)); |
553 | } |
554 | |
555 | static FixedVectorType *getHalfElementsVectorType(FixedVectorType *VTy) { |
556 | return cast<FixedVectorType>(VectorType::getHalfElementsVectorType(VTy)); |
557 | } |
558 | |
559 | static FixedVectorType *getDoubleElementsVectorType(FixedVectorType *VTy) { |
560 | return cast<FixedVectorType>(VectorType::getDoubleElementsVectorType(VTy)); |
561 | } |
562 | |
563 | static bool classof(const Type *T) { |
564 | return T->getTypeID() == FixedVectorTyID; |
565 | } |
566 | |
567 | unsigned getNumElements() const { return ElementQuantity; } |
568 | }; |
569 | |
570 | |
571 | class ScalableVectorType : public VectorType { |
572 | protected: |
573 | ScalableVectorType(Type *ElTy, unsigned MinNumElts) |
574 | : VectorType(ElTy, MinNumElts, ScalableVectorTyID) {} |
575 | |
576 | public: |
577 | static ScalableVectorType *get(Type *ElementType, unsigned MinNumElts); |
578 | |
579 | static ScalableVectorType *get(Type *ElementType, |
580 | const ScalableVectorType *SVTy) { |
581 | return get(ElementType, SVTy->getMinNumElements()); |
582 | } |
583 | |
584 | static ScalableVectorType *getInteger(ScalableVectorType *VTy) { |
585 | return cast<ScalableVectorType>(VectorType::getInteger(VTy)); |
586 | } |
587 | |
588 | static ScalableVectorType * |
589 | getExtendedElementVectorType(ScalableVectorType *VTy) { |
590 | return cast<ScalableVectorType>( |
591 | VectorType::getExtendedElementVectorType(VTy)); |
592 | } |
593 | |
594 | static ScalableVectorType * |
595 | getTruncatedElementVectorType(ScalableVectorType *VTy) { |
596 | return cast<ScalableVectorType>( |
597 | VectorType::getTruncatedElementVectorType(VTy)); |
598 | } |
599 | |
600 | static ScalableVectorType *getSubdividedVectorType(ScalableVectorType *VTy, |
601 | int NumSubdivs) { |
602 | return cast<ScalableVectorType>( |
603 | VectorType::getSubdividedVectorType(VTy, NumSubdivs)); |
604 | } |
605 | |
606 | static ScalableVectorType * |
607 | getHalfElementsVectorType(ScalableVectorType *VTy) { |
608 | return cast<ScalableVectorType>(VectorType::getHalfElementsVectorType(VTy)); |
609 | } |
610 | |
611 | static ScalableVectorType * |
612 | getDoubleElementsVectorType(ScalableVectorType *VTy) { |
613 | return cast<ScalableVectorType>( |
614 | VectorType::getDoubleElementsVectorType(VTy)); |
615 | } |
616 | |
617 | |
618 | |
619 | uint64_t getMinNumElements() const { return ElementQuantity; } |
620 | |
621 | static bool classof(const Type *T) { |
622 | return T->getTypeID() == ScalableVectorTyID; |
623 | } |
624 | }; |
625 | |
626 | inline ElementCount VectorType::getElementCount() const { |
627 | return ElementCount::get(ElementQuantity, isa<ScalableVectorType>(this)); |
| 12 | | Assuming the object is not a 'ScalableVectorType' | |
|
| 13 | | Calling 'LinearPolySize::get' | |
|
| 16 | | Returning from 'LinearPolySize::get' | |
|
628 | } |
629 | |
630 | |
631 | class PointerType : public Type { |
632 | explicit PointerType(Type *ElType, unsigned AddrSpace); |
633 | explicit PointerType(LLVMContext &C, unsigned AddrSpace); |
634 | |
635 | Type *PointeeTy; |
636 | |
637 | public: |
638 | PointerType(const PointerType &) = delete; |
639 | PointerType &operator=(const PointerType &) = delete; |
640 | |
641 | |
642 | |
643 | static PointerType *get(Type *ElementType, unsigned AddressSpace); |
644 | |
645 | |
646 | static PointerType *get(LLVMContext &C, unsigned AddressSpace); |
647 | |
648 | |
649 | |
650 | static PointerType *getUnqual(Type *ElementType) { |
651 | return PointerType::get(ElementType, 0); |
652 | } |
653 | |
654 | |
655 | |
656 | static PointerType *getUnqual(LLVMContext &C) { |
657 | return PointerType::get(C, 0); |
658 | } |
659 | |
660 | |
661 | |
662 | |
663 | |
664 | |
665 | static PointerType *getWithSamePointeeType(PointerType *PT, |
666 | unsigned AddressSpace) { |
667 | if (PT->isOpaque()) |
668 | return get(PT->getContext(), AddressSpace); |
669 | return get(PT->getElementType(), AddressSpace); |
670 | } |
671 | |
672 | Type *getElementType() const { |
673 | assert(!isOpaque() && "Attempting to get element type of opaque pointer"); |
674 | return PointeeTy; |
675 | } |
676 | |
677 | bool isOpaque() const { return !PointeeTy; } |
678 | |
679 | |
680 | static bool isValidElementType(Type *ElemTy); |
681 | |
682 | |
683 | static bool isLoadableOrStorableType(Type *ElemTy); |
684 | |
685 | |
686 | inline unsigned getAddressSpace() const { return getSubclassData(); } |
687 | |
688 | |
689 | |
690 | |
691 | |
692 | bool isOpaqueOrPointeeTypeMatches(Type *Ty) { |
693 | return isOpaque() || PointeeTy == Ty; |
694 | } |
695 | |
696 | |
697 | |
698 | |
699 | |
700 | bool hasSameElementTypeAs(PointerType *Other) { |
701 | return PointeeTy == Other->PointeeTy; |
702 | } |
703 | |
704 | |
705 | static bool classof(const Type *T) { |
706 | return T->getTypeID() == PointerTyID; |
707 | } |
708 | }; |
709 | |
710 | Type *Type::getExtendedType() const { |
711 | assert( |
712 | isIntOrIntVectorTy() && |
713 | "Original type expected to be a vector of integers or a scalar integer."); |
714 | if (auto *VTy = dyn_cast<VectorType>(this)) |
715 | return VectorType::getExtendedElementVectorType( |
716 | const_cast<VectorType *>(VTy)); |
717 | return cast<IntegerType>(this)->getExtendedType(); |
718 | } |
719 | |
720 | Type *Type::getWithNewType(Type *EltTy) const { |
721 | if (auto *VTy = dyn_cast<VectorType>(this)) |
722 | return VectorType::get(EltTy, VTy->getElementCount()); |
723 | return EltTy; |
724 | } |
725 | |
726 | Type *Type::getWithNewBitWidth(unsigned NewBitWidth) const { |
727 | assert( |
728 | isIntOrIntVectorTy() && |
729 | "Original type expected to be a vector of integers or a scalar integer."); |
730 | return getWithNewType(getIntNTy(getContext(), NewBitWidth)); |
731 | } |
732 | |
733 | unsigned Type::getPointerAddressSpace() const { |
734 | return cast<PointerType>(getScalarType())->getAddressSpace(); |
735 | } |
736 | |
737 | } |
738 | |
739 | #endif // LLVM_IR_DERIVEDTYPES_H |
1 | |
2 | |
3 | |
4 | |
5 | |
6 | |
7 | |
8 | |
9 | |
10 | |
11 | |
12 | |
13 | |
14 | |
15 | |
16 | |
17 | |
18 | |
19 | |
20 | |
21 | |
22 | |
23 | |
24 | |
25 | |
26 | |
27 | |
28 | #ifndef LLVM_IR_PATTERNMATCH_H |
29 | #define LLVM_IR_PATTERNMATCH_H |
30 | |
31 | #include "llvm/ADT/APFloat.h" |
32 | #include "llvm/ADT/APInt.h" |
33 | #include "llvm/IR/Constant.h" |
34 | #include "llvm/IR/Constants.h" |
35 | #include "llvm/IR/DataLayout.h" |
36 | #include "llvm/IR/InstrTypes.h" |
37 | #include "llvm/IR/Instruction.h" |
38 | #include "llvm/IR/Instructions.h" |
39 | #include "llvm/IR/IntrinsicInst.h" |
40 | #include "llvm/IR/Intrinsics.h" |
41 | #include "llvm/IR/Operator.h" |
42 | #include "llvm/IR/Value.h" |
43 | #include "llvm/Support/Casting.h" |
44 | #include <cstdint> |
45 | |
46 | namespace llvm { |
47 | namespace PatternMatch { |
48 | |
49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { |
50 | return const_cast<Pattern &>(P).match(V); |
| 27 | | Calling 'ThreeOps_match::match' | |
|
| 32 | | Returning from 'ThreeOps_match::match' | |
|
| 33 | | Returning the value 1, which participates in a condition later | |
|
51 | } |
52 | |
53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { |
54 | return const_cast<Pattern &>(P).match(Mask); |
55 | } |
56 | |
57 | template <typename SubPattern_t> struct OneUse_match { |
58 | SubPattern_t SubPattern; |
59 | |
60 | OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {} |
61 | |
62 | template <typename OpTy> bool match(OpTy *V) { |
63 | return V->hasOneUse() && SubPattern.match(V); |
64 | } |
65 | }; |
66 | |
67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { |
68 | return SubPattern; |
69 | } |
70 | |
71 | template <typename Class> struct class_match { |
72 | template <typename ITy> bool match(ITy *V) { return isa<Class>(V); } |
73 | }; |
74 | |
75 | |
76 | inline class_match<Value> m_Value() { return class_match<Value>(); } |
77 | |
78 | |
79 | inline class_match<UnaryOperator> m_UnOp() { |
80 | return class_match<UnaryOperator>(); |
81 | } |
82 | |
83 | |
84 | inline class_match<BinaryOperator> m_BinOp() { |
85 | return class_match<BinaryOperator>(); |
86 | } |
87 | |
88 | |
89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } |
90 | |
91 | struct undef_match { |
92 | static bool check(const Value *V) { |
93 | if (isa<UndefValue>(V)) |
94 | return true; |
95 | |
96 | const auto *CA = dyn_cast<ConstantAggregate>(V); |
97 | if (!CA) |
98 | return false; |
99 | |
100 | SmallPtrSet<const ConstantAggregate *, 8> Seen; |
101 | SmallVector<const ConstantAggregate *, 8> Worklist; |
102 | |
103 | |
104 | |
105 | |
106 | auto CheckValue = [&](const ConstantAggregate *CA) { |
107 | for (const Value *Op : CA->operand_values()) { |
108 | if (isa<UndefValue>(Op)) |
109 | continue; |
110 | |
111 | const auto *CA = dyn_cast<ConstantAggregate>(Op); |
112 | if (!CA) |
113 | return false; |
114 | if (Seen.insert(CA).second) |
115 | Worklist.emplace_back(CA); |
116 | } |
117 | |
118 | return true; |
119 | }; |
120 | |
121 | if (!CheckValue(CA)) |
122 | return false; |
123 | |
124 | while (!Worklist.empty()) { |
125 | if (!CheckValue(Worklist.pop_back_val())) |
126 | return false; |
127 | } |
128 | return true; |
129 | } |
130 | template <typename ITy> bool match(ITy *V) { return check(V); } |
131 | }; |
132 | |
133 | |
134 | |
135 | |
136 | inline auto m_Undef() { return undef_match(); } |
137 | |
138 | |
139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } |
140 | |
141 | |
142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } |
143 | |
144 | |
145 | inline class_match<ConstantInt> m_ConstantInt() { |
146 | return class_match<ConstantInt>(); |
147 | } |
148 | |
149 | |
150 | inline class_match<ConstantFP> m_ConstantFP() { |
151 | return class_match<ConstantFP>(); |
152 | } |
153 | |
154 | |
155 | inline class_match<ConstantExpr> m_ConstantExpr() { |
156 | return class_match<ConstantExpr>(); |
157 | } |
158 | |
159 | |
160 | inline class_match<BasicBlock> m_BasicBlock() { |
161 | return class_match<BasicBlock>(); |
162 | } |
163 | |
164 | |
165 | template <typename Ty> struct match_unless { |
166 | Ty M; |
167 | |
168 | match_unless(const Ty &Matcher) : M(Matcher) {} |
169 | |
170 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } |
171 | }; |
172 | |
173 | |
174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { |
175 | return match_unless<Ty>(M); |
176 | } |
177 | |
178 | |
179 | template <typename LTy, typename RTy> struct match_combine_or { |
180 | LTy L; |
181 | RTy R; |
182 | |
183 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
184 | |
185 | template <typename ITy> bool match(ITy *V) { |
186 | if (L.match(V)) |
187 | return true; |
188 | if (R.match(V)) |
189 | return true; |
190 | return false; |
191 | } |
192 | }; |
193 | |
194 | template <typename LTy, typename RTy> struct match_combine_and { |
195 | LTy L; |
196 | RTy R; |
197 | |
198 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} |
199 | |
200 | template <typename ITy> bool match(ITy *V) { |
201 | if (L.match(V)) |
202 | if (R.match(V)) |
203 | return true; |
204 | return false; |
205 | } |
206 | }; |
207 | |
208 | |
209 | template <typename LTy, typename RTy> |
210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { |
211 | return match_combine_or<LTy, RTy>(L, R); |
212 | } |
213 | |
214 | |
215 | template <typename LTy, typename RTy> |
216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { |
217 | return match_combine_and<LTy, RTy>(L, R); |
218 | } |
219 | |
220 | struct apint_match { |
221 | const APInt *&Res; |
222 | bool AllowUndef; |
223 | |
224 | apint_match(const APInt *&Res, bool AllowUndef) |
225 | : Res(Res), AllowUndef(AllowUndef) {} |
226 | |
227 | template <typename ITy> bool match(ITy *V) { |
228 | if (auto *CI = dyn_cast<ConstantInt>(V)) { |
229 | Res = &CI->getValue(); |
230 | return true; |
231 | } |
232 | if (V->getType()->isVectorTy()) |
233 | if (const auto *C = dyn_cast<Constant>(V)) |
234 | if (auto *CI = dyn_cast_or_null<ConstantInt>( |
235 | C->getSplatValue(AllowUndef))) { |
236 | Res = &CI->getValue(); |
237 | return true; |
238 | } |
239 | return false; |
240 | } |
241 | }; |
242 | |
243 | |
244 | |
245 | struct apfloat_match { |
246 | const APFloat *&Res; |
247 | bool AllowUndef; |
248 | |
249 | apfloat_match(const APFloat *&Res, bool AllowUndef) |
250 | : Res(Res), AllowUndef(AllowUndef) {} |
251 | |
252 | template <typename ITy> bool match(ITy *V) { |
253 | if (auto *CI = dyn_cast<ConstantFP>(V)) { |
254 | Res = &CI->getValueAPF(); |
255 | return true; |
256 | } |
257 | if (V->getType()->isVectorTy()) |
258 | if (const auto *C = dyn_cast<Constant>(V)) |
259 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
260 | C->getSplatValue(AllowUndef))) { |
261 | Res = &CI->getValueAPF(); |
262 | return true; |
263 | } |
264 | return false; |
265 | } |
266 | }; |
267 | |
268 | |
269 | |
270 | inline apint_match m_APInt(const APInt *&Res) { |
271 | |
272 | return apint_match(Res, false); |
273 | } |
274 | |
275 | |
276 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { |
277 | return apint_match(Res, true); |
278 | } |
279 | |
280 | |
281 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { |
282 | return apint_match(Res, false); |
283 | } |
284 | |
285 | |
286 | |
287 | inline apfloat_match m_APFloat(const APFloat *&Res) { |
288 | |
289 | return apfloat_match(Res, false); |
290 | } |
291 | |
292 | |
293 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { |
294 | return apfloat_match(Res, true); |
295 | } |
296 | |
297 | |
298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { |
299 | return apfloat_match(Res, false); |
300 | } |
301 | |
302 | template <int64_t Val> struct constantint_match { |
303 | template <typename ITy> bool match(ITy *V) { |
304 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { |
305 | const APInt &CIV = CI->getValue(); |
306 | if (Val >= 0) |
307 | return CIV == static_cast<uint64_t>(Val); |
308 | |
309 | |
310 | |
311 | return -CIV == -Val; |
312 | } |
313 | return false; |
314 | } |
315 | }; |
316 | |
317 | |
318 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { |
319 | return constantint_match<Val>(); |
320 | } |
321 | |
322 | |
323 | |
324 | |
325 | template <typename Predicate, typename ConstantVal> |
326 | struct cstval_pred_ty : public Predicate { |
327 | template <typename ITy> bool match(ITy *V) { |
328 | if (const auto *CV = dyn_cast<ConstantVal>(V)) |
329 | return this->isValue(CV->getValue()); |
330 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { |
331 | if (const auto *C = dyn_cast<Constant>(V)) { |
332 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) |
333 | return this->isValue(CV->getValue()); |
334 | |
335 | |
336 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); |
337 | if (!FVTy) |
338 | return false; |
339 | |
340 | |
341 | unsigned NumElts = FVTy->getNumElements(); |
342 | assert(NumElts != 0 && "Constant vector with no elements?"); |
343 | bool HasNonUndefElements = false; |
344 | for (unsigned i = 0; i != NumElts; ++i) { |
345 | Constant *Elt = C->getAggregateElement(i); |
346 | if (!Elt) |
347 | return false; |
348 | if (isa<UndefValue>(Elt)) |
349 | continue; |
350 | auto *CV = dyn_cast<ConstantVal>(Elt); |
351 | if (!CV || !this->isValue(CV->getValue())) |
352 | return false; |
353 | HasNonUndefElements = true; |
354 | } |
355 | return HasNonUndefElements; |
356 | } |
357 | } |
358 | return false; |
359 | } |
360 | }; |
361 | |
362 | |
363 | template <typename Predicate> |
364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; |
365 | |
366 | |
367 | template <typename Predicate> |
368 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; |
369 | |
370 | |
371 | |
372 | template <typename Predicate> struct api_pred_ty : public Predicate { |
373 | const APInt *&Res; |
374 | |
375 | api_pred_ty(const APInt *&R) : Res(R) {} |
376 | |
377 | template <typename ITy> bool match(ITy *V) { |
378 | if (const auto *CI = dyn_cast<ConstantInt>(V)) |
379 | if (this->isValue(CI->getValue())) { |
380 | Res = &CI->getValue(); |
381 | return true; |
382 | } |
383 | if (V->getType()->isVectorTy()) |
384 | if (const auto *C = dyn_cast<Constant>(V)) |
385 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) |
386 | if (this->isValue(CI->getValue())) { |
387 | Res = &CI->getValue(); |
388 | return true; |
389 | } |
390 | |
391 | return false; |
392 | } |
393 | }; |
394 | |
395 | |
396 | |
397 | |
398 | template <typename Predicate> struct apf_pred_ty : public Predicate { |
399 | const APFloat *&Res; |
400 | |
401 | apf_pred_ty(const APFloat *&R) : Res(R) {} |
402 | |
403 | template <typename ITy> bool match(ITy *V) { |
404 | if (const auto *CI = dyn_cast<ConstantFP>(V)) |
405 | if (this->isValue(CI->getValue())) { |
406 | Res = &CI->getValue(); |
407 | return true; |
408 | } |
409 | if (V->getType()->isVectorTy()) |
410 | if (const auto *C = dyn_cast<Constant>(V)) |
411 | if (auto *CI = dyn_cast_or_null<ConstantFP>( |
412 | C->getSplatValue( true))) |
413 | if (this->isValue(CI->getValue())) { |
414 | Res = &CI->getValue(); |
415 | return true; |
416 | } |
417 | |
418 | return false; |
419 | } |
420 | }; |
421 | |
422 | |
423 | |
424 | |
425 | |
426 | |
427 | |
428 | |
429 | |
430 | |
431 | struct is_any_apint { |
432 | bool isValue(const APInt &C) { return true; } |
433 | }; |
434 | |
435 | |
436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { |
437 | return cst_pred_ty<is_any_apint>(); |
438 | } |
439 | |
440 | struct is_all_ones { |
441 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } |
442 | }; |
443 | |
444 | |
445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { |
446 | return cst_pred_ty<is_all_ones>(); |
447 | } |
448 | |
449 | struct is_maxsignedvalue { |
450 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } |
451 | }; |
452 | |
453 | |
454 | |
455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { |
456 | return cst_pred_ty<is_maxsignedvalue>(); |
457 | } |
458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { |
459 | return V; |
460 | } |
461 | |
462 | struct is_negative { |
463 | bool isValue(const APInt &C) { return C.isNegative(); } |
464 | }; |
465 | |
466 | |
467 | inline cst_pred_ty<is_negative> m_Negative() { |
468 | return cst_pred_ty<is_negative>(); |
469 | } |
470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { |
471 | return V; |
472 | } |
473 | |
474 | struct is_nonnegative { |
475 | bool isValue(const APInt &C) { return C.isNonNegative(); } |
476 | }; |
477 | |
478 | |
479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { |
480 | return cst_pred_ty<is_nonnegative>(); |
481 | } |
482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { |
483 | return V; |
484 | } |
485 | |
486 | struct is_strictlypositive { |
487 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } |
488 | }; |
489 | |
490 | |
491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { |
492 | return cst_pred_ty<is_strictlypositive>(); |
493 | } |
494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { |
495 | return V; |
496 | } |
497 | |
498 | struct is_nonpositive { |
499 | bool isValue(const APInt &C) { return C.isNonPositive(); } |
500 | }; |
501 | |
502 | |
503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { |
504 | return cst_pred_ty<is_nonpositive>(); |
505 | } |
506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } |
507 | |
508 | struct is_one { |
509 | bool isValue(const APInt &C) { return C.isOneValue(); } |
510 | }; |
511 | |
512 | |
513 | inline cst_pred_ty<is_one> m_One() { |
514 | return cst_pred_ty<is_one>(); |
515 | } |
516 | |
517 | struct is_zero_int { |
518 | bool isValue(const APInt &C) { return C.isNullValue(); } |
519 | }; |
520 | |
521 | |
522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { |
523 | return cst_pred_ty<is_zero_int>(); |
524 | } |
525 | |
526 | struct is_zero { |
527 | template <typename ITy> bool match(ITy *V) { |
528 | auto *C = dyn_cast<Constant>(V); |
529 | |
530 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); |
531 | } |
532 | }; |
533 | |
534 | |
535 | inline is_zero m_Zero() { |
536 | return is_zero(); |
537 | } |
538 | |
539 | struct is_power2 { |
540 | bool isValue(const APInt &C) { return C.isPowerOf2(); } |
541 | }; |
542 | |
543 | |
544 | inline cst_pred_ty<is_power2> m_Power2() { |
545 | return cst_pred_ty<is_power2>(); |
546 | } |
547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { |
548 | return V; |
549 | } |
550 | |
551 | struct is_negated_power2 { |
552 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } |
553 | }; |
554 | |
555 | |
556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { |
557 | return cst_pred_ty<is_negated_power2>(); |
558 | } |
559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { |
560 | return V; |
561 | } |
562 | |
563 | struct is_power2_or_zero { |
564 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } |
565 | }; |
566 | |
567 | |
568 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { |
569 | return cst_pred_ty<is_power2_or_zero>(); |
570 | } |
571 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { |
572 | return V; |
573 | } |
574 | |
575 | struct is_sign_mask { |
576 | bool isValue(const APInt &C) { return C.isSignMask(); } |
577 | }; |
578 | |
579 | |
580 | inline cst_pred_ty<is_sign_mask> m_SignMask() { |
581 | return cst_pred_ty<is_sign_mask>(); |
582 | } |
583 | |
584 | struct is_lowbit_mask { |
585 | bool isValue(const APInt &C) { return C.isMask(); } |
586 | }; |
587 | |
588 | |
589 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { |
590 | return cst_pred_ty<is_lowbit_mask>(); |
591 | } |
592 | |
593 | struct icmp_pred_with_threshold { |
594 | ICmpInst::Predicate Pred; |
595 | const APInt *Thr; |
596 | bool isValue(const APInt &C) { |
597 | switch (Pred) { |
598 | case ICmpInst::Predicate::ICMP_EQ: |
599 | return C.eq(*Thr); |
600 | case ICmpInst::Predicate::ICMP_NE: |
601 | return C.ne(*Thr); |
602 | case ICmpInst::Predicate::ICMP_UGT: |
603 | return C.ugt(*Thr); |
604 | case ICmpInst::Predicate::ICMP_UGE: |
605 | return C.uge(*Thr); |
606 | case ICmpInst::Predicate::ICMP_ULT: |
607 | return C.ult(*Thr); |
608 | case ICmpInst::Predicate::ICMP_ULE: |
609 | return C.ule(*Thr); |
610 | case ICmpInst::Predicate::ICMP_SGT: |
611 | return C.sgt(*Thr); |
612 | case ICmpInst::Predicate::ICMP_SGE: |
613 | return C.sge(*Thr); |
614 | case ICmpInst::Predicate::ICMP_SLT: |
615 | return C.slt(*Thr); |
616 | case ICmpInst::Predicate::ICMP_SLE: |
617 | return C.sle(*Thr); |
618 | default: |
619 | llvm_unreachable("Unhandled ICmp predicate"); |
620 | } |
621 | } |
622 | }; |
623 | |
624 | |
625 | inline cst_pred_ty<icmp_pred_with_threshold> |
626 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { |
627 | cst_pred_ty<icmp_pred_with_threshold> P; |
628 | P.Pred = Predicate; |
629 | P.Thr = &Threshold; |
630 | return P; |
631 | } |
632 | |
633 | struct is_nan { |
634 | bool isValue(const APFloat &C) { return C.isNaN(); } |
635 | }; |
636 | |
637 | |
638 | inline cstfp_pred_ty<is_nan> m_NaN() { |
639 | return cstfp_pred_ty<is_nan>(); |
640 | } |
641 | |
642 | struct is_nonnan { |
643 | bool isValue(const APFloat &C) { return !C.isNaN(); } |
644 | }; |
645 | |
646 | |
647 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { |
648 | return cstfp_pred_ty<is_nonnan>(); |
649 | } |
650 | |
651 | struct is_inf { |
652 | bool isValue(const APFloat &C) { return C.isInfinity(); } |
653 | }; |
654 | |
655 | |
656 | inline cstfp_pred_ty<is_inf> m_Inf() { |
657 | return cstfp_pred_ty<is_inf>(); |
658 | } |
659 | |
660 | struct is_noninf { |
661 | bool isValue(const APFloat &C) { return !C.isInfinity(); } |
662 | }; |
663 | |
664 | |
665 | inline cstfp_pred_ty<is_noninf> m_NonInf() { |
666 | return cstfp_pred_ty<is_noninf>(); |
667 | } |
668 | |
669 | struct is_finite { |
670 | bool isValue(const APFloat &C) { return C.isFinite(); } |
671 | }; |
672 | |
673 | |
674 | inline cstfp_pred_ty<is_finite> m_Finite() { |
675 | return cstfp_pred_ty<is_finite>(); |
676 | } |
677 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } |
678 | |
679 | struct is_finitenonzero { |
680 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } |
681 | }; |
682 | |
683 | |
684 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { |
685 | return cstfp_pred_ty<is_finitenonzero>(); |
686 | } |
687 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { |
688 | return V; |
689 | } |
690 | |
691 | struct is_any_zero_fp { |
692 | bool isValue(const APFloat &C) { return C.isZero(); } |
693 | }; |
694 | |
695 | |
696 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { |
697 | return cstfp_pred_ty<is_any_zero_fp>(); |
698 | } |
699 | |
700 | struct is_pos_zero_fp { |
701 | bool isValue(const APFloat &C) { return C.isPosZero(); } |
702 | }; |
703 | |
704 | |
705 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { |
706 | return cstfp_pred_ty<is_pos_zero_fp>(); |
707 | } |
708 | |
709 | struct is_neg_zero_fp { |
710 | bool isValue(const APFloat &C) { return C.isNegZero(); } |
711 | }; |
712 | |
713 | |
714 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { |
715 | return cstfp_pred_ty<is_neg_zero_fp>(); |
716 | } |
717 | |
718 | struct is_non_zero_fp { |
719 | bool isValue(const APFloat &C) { return C.isNonZero(); } |
720 | }; |
721 | |
722 | |
723 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { |
724 | return cstfp_pred_ty<is_non_zero_fp>(); |
725 | } |
726 | |
727 | |
728 | |
729 | template <typename Class> struct bind_ty { |
730 | Class *&VR; |
731 | |
732 | bind_ty(Class *&V) : VR(V) {} |
733 | |
734 | template <typename ITy> bool match(ITy *V) { |
735 | if (auto *CV = dyn_cast<Class>(V)) { |
736 | VR = CV; |
737 | return true; |
738 | } |
739 | return false; |
740 | } |
741 | }; |
742 | |
743 | |
744 | inline bind_ty<Value> m_Value(Value *&V) { return V; } |
745 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } |
746 | |
747 | |
748 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } |
749 | |
750 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } |
751 | |
752 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } |
753 | |
754 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } |
755 | inline bind_ty<const WithOverflowInst> |
756 | m_WithOverflowInst(const WithOverflowInst *&I) { |
757 | return I; |
758 | } |
759 | |
760 | |
761 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } |
762 | |
763 | |
764 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } |
765 | |
766 | |
767 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } |
768 | |
769 | |
770 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } |
771 | |
772 | |
773 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } |
774 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { |
775 | return V; |
776 | } |
777 | |
778 | |
779 | inline match_combine_and<class_match<Constant>, |
780 | match_unless<class_match<ConstantExpr>>> |
781 | m_ImmConstant() { |
782 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); |
783 | } |
784 | |
785 | |
786 | inline match_combine_and<bind_ty<Constant>, |
787 | match_unless<class_match<ConstantExpr>>> |
788 | m_ImmConstant(Constant *&C) { |
789 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); |
790 | } |
791 | |
792 | |
793 | struct specificval_ty { |
794 | const Value *Val; |
795 | |
796 | specificval_ty(const Value *V) : Val(V) {} |
797 | |
798 | template <typename ITy> bool match(ITy *V) { return V == Val; } |
799 | }; |
800 | |
801 | |
802 | inline specificval_ty m_Specific(const Value *V) { return V; } |
803 | |
804 | |
805 | |
806 | template <typename Class> struct deferredval_ty { |
807 | Class *const &Val; |
808 | |
809 | deferredval_ty(Class *const &V) : Val(V) {} |
810 | |
811 | template <typename ITy> bool match(ITy *const V) { return V == Val; } |
812 | }; |
813 | |
814 | |
815 | |
816 | |
817 | |
818 | |
819 | |
820 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } |
821 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { |
822 | return V; |
823 | } |
824 | |
825 | |
826 | |
827 | struct specific_fpval { |
828 | double Val; |
829 | |
830 | specific_fpval(double V) : Val(V) {} |
831 | |
832 | template <typename ITy> bool match(ITy *V) { |
833 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) |
834 | return CFP->isExactlyValue(Val); |
835 | if (V->getType()->isVectorTy()) |
836 | if (const auto *C = dyn_cast<Constant>(V)) |
837 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) |
838 | return CFP->isExactlyValue(Val); |
839 | return false; |
840 | } |
841 | }; |
842 | |
843 | |
844 | |
845 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } |
846 | |
847 | |
848 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } |
849 | |
850 | struct bind_const_intval_ty { |
851 | uint64_t &VR; |
852 | |
853 | bind_const_intval_ty(uint64_t &V) : VR(V) {} |
854 | |
855 | template <typename ITy> bool match(ITy *V) { |
856 | if (const auto *CV = dyn_cast<ConstantInt>(V)) |
857 | if (CV->getValue().ule(UINT64_MAX)) { |
858 | VR = CV->getZExtValue(); |
859 | return true; |
860 | } |
861 | return false; |
862 | } |
863 | }; |
864 | |
865 | |
866 | |
867 | template <bool AllowUndefs> |
868 | struct specific_intval { |
869 | APInt Val; |
870 | |
871 | specific_intval(APInt V) : Val(std::move(V)) {} |
872 | |
873 | template <typename ITy> bool match(ITy *V) { |
874 | const auto *CI = dyn_cast<ConstantInt>(V); |
875 | if (!CI && V->getType()->isVectorTy()) |
876 | if (const auto *C = dyn_cast<Constant>(V)) |
877 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); |
878 | |
879 | return CI && APInt::isSameValue(CI->getValue(), Val); |
880 | } |
881 | }; |
882 | |
883 | |
884 | |
885 | inline specific_intval<false> m_SpecificInt(APInt V) { |
886 | return specific_intval<false>(std::move(V)); |
887 | } |
888 | |
889 | inline specific_intval<false> m_SpecificInt(uint64_t V) { |
890 | return m_SpecificInt(APInt(64, V)); |
891 | } |
892 | |
893 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { |
894 | return specific_intval<true>(std::move(V)); |
895 | } |
896 | |
897 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { |
898 | return m_SpecificIntAllowUndef(APInt(64, V)); |
899 | } |
900 | |
901 | |
902 | |
903 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } |
904 | |
905 | |
906 | struct specific_bbval { |
907 | BasicBlock *Val; |
908 | |
909 | specific_bbval(BasicBlock *Val) : Val(Val) {} |
910 | |
911 | template <typename ITy> bool match(ITy *V) { |
912 | const auto *BB = dyn_cast<BasicBlock>(V); |
913 | return BB && BB == Val; |
914 | } |
915 | }; |
916 | |
917 | |
918 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { |
919 | return specific_bbval(BB); |
920 | } |
921 | |
922 | |
923 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { |
924 | return BB; |
925 | } |
926 | inline deferredval_ty<const BasicBlock> |
927 | m_Deferred(const BasicBlock *const &BB) { |
928 | return BB; |
929 | } |
930 | |
931 | |
932 | |
933 | |
934 | template <typename LHS_t, typename RHS_t, bool Commutable = false> |
935 | struct AnyBinaryOp_match { |
936 | LHS_t L; |
937 | RHS_t R; |
938 | |
939 | |
940 | |
941 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
942 | |
943 | template <typename OpTy> bool match(OpTy *V) { |
944 | if (auto *I = dyn_cast<BinaryOperator>(V)) |
945 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || |
946 | (Commutable && L.match(I->getOperand(1)) && |
947 | R.match(I->getOperand(0))); |
948 | return false; |
949 | } |
950 | }; |
951 | |
952 | template <typename LHS, typename RHS> |
953 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { |
954 | return AnyBinaryOp_match<LHS, RHS>(L, R); |
955 | } |
956 | |
957 | |
958 | |
959 | |
960 | |
961 | template <typename OP_t> struct AnyUnaryOp_match { |
962 | OP_t X; |
963 | |
964 | AnyUnaryOp_match(const OP_t &X) : X(X) {} |
965 | |
966 | template <typename OpTy> bool match(OpTy *V) { |
967 | if (auto *I = dyn_cast<UnaryOperator>(V)) |
968 | return X.match(I->getOperand(0)); |
969 | return false; |
970 | } |
971 | }; |
972 | |
973 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { |
974 | return AnyUnaryOp_match<OP_t>(X); |
975 | } |
976 | |
977 | |
978 | |
979 | |
980 | |
981 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
982 | bool Commutable = false> |
983 | struct BinaryOp_match { |
984 | LHS_t L; |
985 | RHS_t R; |
986 | |
987 | |
988 | |
989 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
990 | |
991 | template <typename OpTy> bool match(OpTy *V) { |
992 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
993 | auto *I = cast<BinaryOperator>(V); |
994 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || |
995 | (Commutable && L.match(I->getOperand(1)) && |
996 | R.match(I->getOperand(0))); |
997 | } |
998 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
999 | return CE->getOpcode() == Opcode && |
1000 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || |
1001 | (Commutable && L.match(CE->getOperand(1)) && |
1002 | R.match(CE->getOperand(0)))); |
1003 | return false; |
1004 | } |
1005 | }; |
1006 | |
1007 | template <typename LHS, typename RHS> |
1008 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, |
1009 | const RHS &R) { |
1010 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); |
1011 | } |
1012 | |
1013 | template <typename LHS, typename RHS> |
1014 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, |
1015 | const RHS &R) { |
1016 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); |
1017 | } |
1018 | |
1019 | template <typename LHS, typename RHS> |
1020 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, |
1021 | const RHS &R) { |
1022 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); |
1023 | } |
1024 | |
1025 | template <typename LHS, typename RHS> |
1026 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, |
1027 | const RHS &R) { |
1028 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); |
1029 | } |
1030 | |
1031 | template <typename Op_t> struct FNeg_match { |
1032 | Op_t X; |
1033 | |
1034 | FNeg_match(const Op_t &Op) : X(Op) {} |
1035 | template <typename OpTy> bool match(OpTy *V) { |
1036 | auto *FPMO = dyn_cast<FPMathOperator>(V); |
1037 | if (!FPMO) return false; |
1038 | |
1039 | if (FPMO->getOpcode() == Instruction::FNeg) |
1040 | return X.match(FPMO->getOperand(0)); |
1041 | |
1042 | if (FPMO->getOpcode() == Instruction::FSub) { |
1043 | if (FPMO->hasNoSignedZeros()) { |
1044 | |
1045 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) |
1046 | return false; |
1047 | } else { |
1048 | |
1049 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) |
1050 | return false; |
1051 | } |
1052 | |
1053 | return X.match(FPMO->getOperand(1)); |
1054 | } |
1055 | |
1056 | return false; |
1057 | } |
1058 | }; |
1059 | |
1060 | |
1061 | template <typename OpTy> |
1062 | inline FNeg_match<OpTy> |
1063 | m_FNeg(const OpTy &X) { |
1064 | return FNeg_match<OpTy>(X); |
1065 | } |
1066 | |
1067 | |
1068 | template <typename RHS> |
1069 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> |
1070 | m_FNegNSZ(const RHS &X) { |
1071 | return m_FSub(m_AnyZeroFP(), X); |
1072 | } |
1073 | |
1074 | template <typename LHS, typename RHS> |
1075 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, |
1076 | const RHS &R) { |
1077 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); |
1078 | } |
1079 | |
1080 | template <typename LHS, typename RHS> |
1081 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, |
1082 | const RHS &R) { |
1083 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); |
1084 | } |
1085 | |
1086 | template <typename LHS, typename RHS> |
1087 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, |
1088 | const RHS &R) { |
1089 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); |
1090 | } |
1091 | |
1092 | template <typename LHS, typename RHS> |
1093 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, |
1094 | const RHS &R) { |
1095 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); |
1096 | } |
1097 | |
1098 | template <typename LHS, typename RHS> |
1099 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, |
1100 | const RHS &R) { |
1101 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); |
1102 | } |
1103 | |
1104 | template <typename LHS, typename RHS> |
1105 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, |
1106 | const RHS &R) { |
1107 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); |
1108 | } |
1109 | |
1110 | template <typename LHS, typename RHS> |
1111 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, |
1112 | const RHS &R) { |
1113 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); |
1114 | } |
1115 | |
1116 | template <typename LHS, typename RHS> |
1117 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, |
1118 | const RHS &R) { |
1119 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); |
1120 | } |
1121 | |
1122 | template <typename LHS, typename RHS> |
1123 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, |
1124 | const RHS &R) { |
1125 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); |
1126 | } |
1127 | |
1128 | template <typename LHS, typename RHS> |
1129 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, |
1130 | const RHS &R) { |
1131 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); |
1132 | } |
1133 | |
1134 | template <typename LHS, typename RHS> |
1135 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, |
1136 | const RHS &R) { |
1137 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); |
1138 | } |
1139 | |
1140 | template <typename LHS, typename RHS> |
1141 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, |
1142 | const RHS &R) { |
1143 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); |
1144 | } |
1145 | |
1146 | template <typename LHS, typename RHS> |
1147 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, |
1148 | const RHS &R) { |
1149 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); |
1150 | } |
1151 | |
1152 | template <typename LHS, typename RHS> |
1153 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, |
1154 | const RHS &R) { |
1155 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); |
1156 | } |
1157 | |
1158 | template <typename LHS_t, typename RHS_t, unsigned Opcode, |
1159 | unsigned WrapFlags = 0> |
1160 | struct OverflowingBinaryOp_match { |
1161 | LHS_t L; |
1162 | RHS_t R; |
1163 | |
1164 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) |
1165 | : L(LHS), R(RHS) {} |
1166 | |
1167 | template <typename OpTy> bool match(OpTy *V) { |
1168 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { |
1169 | if (Op->getOpcode() != Opcode) |
1170 | return false; |
1171 | if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) && |
1172 | !Op->hasNoUnsignedWrap()) |
1173 | return false; |
1174 | if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) && |
1175 | !Op->hasNoSignedWrap()) |
1176 | return false; |
1177 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); |
1178 | } |
1179 | return false; |
1180 | } |
1181 | }; |
1182 | |
1183 | template <typename LHS, typename RHS> |
1184 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1185 | OverflowingBinaryOperator::NoSignedWrap> |
1186 | m_NSWAdd(const LHS &L, const RHS &R) { |
1187 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1188 | OverflowingBinaryOperator::NoSignedWrap>( |
1189 | L, R); |
1190 | } |
1191 | template <typename LHS, typename RHS> |
1192 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1193 | OverflowingBinaryOperator::NoSignedWrap> |
1194 | m_NSWSub(const LHS &L, const RHS &R) { |
1195 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1196 | OverflowingBinaryOperator::NoSignedWrap>( |
1197 | L, R); |
1198 | } |
1199 | template <typename LHS, typename RHS> |
1200 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1201 | OverflowingBinaryOperator::NoSignedWrap> |
1202 | m_NSWMul(const LHS &L, const RHS &R) { |
1203 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1204 | OverflowingBinaryOperator::NoSignedWrap>( |
1205 | L, R); |
1206 | } |
1207 | template <typename LHS, typename RHS> |
1208 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1209 | OverflowingBinaryOperator::NoSignedWrap> |
1210 | m_NSWShl(const LHS &L, const RHS &R) { |
1211 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1212 | OverflowingBinaryOperator::NoSignedWrap>( |
1213 | L, R); |
1214 | } |
1215 | |
1216 | template <typename LHS, typename RHS> |
1217 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1218 | OverflowingBinaryOperator::NoUnsignedWrap> |
1219 | m_NUWAdd(const LHS &L, const RHS &R) { |
1220 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, |
1221 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1222 | L, R); |
1223 | } |
1224 | template <typename LHS, typename RHS> |
1225 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1226 | OverflowingBinaryOperator::NoUnsignedWrap> |
1227 | m_NUWSub(const LHS &L, const RHS &R) { |
1228 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, |
1229 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1230 | L, R); |
1231 | } |
1232 | template <typename LHS, typename RHS> |
1233 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1234 | OverflowingBinaryOperator::NoUnsignedWrap> |
1235 | m_NUWMul(const LHS &L, const RHS &R) { |
1236 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, |
1237 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1238 | L, R); |
1239 | } |
1240 | template <typename LHS, typename RHS> |
1241 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1242 | OverflowingBinaryOperator::NoUnsignedWrap> |
1243 | m_NUWShl(const LHS &L, const RHS &R) { |
1244 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, |
1245 | OverflowingBinaryOperator::NoUnsignedWrap>( |
1246 | L, R); |
1247 | } |
1248 | |
1249 | |
1250 | |
1251 | |
1252 | template <typename LHS_t, typename RHS_t, typename Predicate> |
1253 | struct BinOpPred_match : Predicate { |
1254 | LHS_t L; |
1255 | RHS_t R; |
1256 | |
1257 | BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
1258 | |
1259 | template <typename OpTy> bool match(OpTy *V) { |
1260 | if (auto *I = dyn_cast<Instruction>(V)) |
1261 | return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) && |
1262 | R.match(I->getOperand(1)); |
1263 | if (auto *CE = dyn_cast<ConstantExpr>(V)) |
1264 | return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) && |
1265 | R.match(CE->getOperand(1)); |
1266 | return false; |
1267 | } |
1268 | }; |
1269 | |
1270 | struct is_shift_op { |
1271 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } |
1272 | }; |
1273 | |
1274 | struct is_right_shift_op { |
1275 | bool isOpType(unsigned Opcode) { |
1276 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; |
1277 | } |
1278 | }; |
1279 | |
1280 | struct is_logical_shift_op { |
1281 | bool isOpType(unsigned Opcode) { |
1282 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; |
1283 | } |
1284 | }; |
1285 | |
1286 | struct is_bitwiselogic_op { |
1287 | bool isOpType(unsigned Opcode) { |
1288 | return Instruction::isBitwiseLogicOp(Opcode); |
1289 | } |
1290 | }; |
1291 | |
1292 | struct is_idiv_op { |
1293 | bool isOpType(unsigned Opcode) { |
1294 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; |
1295 | } |
1296 | }; |
1297 | |
1298 | struct is_irem_op { |
1299 | bool isOpType(unsigned Opcode) { |
1300 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; |
1301 | } |
1302 | }; |
1303 | |
1304 | |
1305 | template <typename LHS, typename RHS> |
1306 | inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L, |
1307 | const RHS &R) { |
1308 | return BinOpPred_match<LHS, RHS, is_shift_op>(L, R); |
1309 | } |
1310 | |
1311 | |
1312 | template <typename LHS, typename RHS> |
1313 | inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L, |
1314 | const RHS &R) { |
1315 | return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R); |
1316 | } |
1317 | |
1318 | |
1319 | template <typename LHS, typename RHS> |
1320 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> |
1321 | m_LogicalShift(const LHS &L, const RHS &R) { |
1322 | return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R); |
1323 | } |
1324 | |
1325 | |
1326 | template <typename LHS, typename RHS> |
1327 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> |
1328 | m_BitwiseLogic(const LHS &L, const RHS &R) { |
1329 | return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R); |
1330 | } |
1331 | |
1332 | |
1333 | template <typename LHS, typename RHS> |
1334 | inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L, |
1335 | const RHS &R) { |
1336 | return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R); |
1337 | } |
1338 | |
1339 | |
1340 | template <typename LHS, typename RHS> |
1341 | inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L, |
1342 | const RHS &R) { |
1343 | return BinOpPred_match<LHS, RHS, is_irem_op>(L, R); |
1344 | } |
1345 | |
1346 | |
1347 | |
1348 | |
1349 | template <typename SubPattern_t> struct Exact_match { |
1350 | SubPattern_t SubPattern; |
1351 | |
1352 | Exact_match(const SubPattern_t &SP) : SubPattern(SP) {} |
1353 | |
1354 | template <typename OpTy> bool match(OpTy *V) { |
1355 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(V)) |
1356 | return PEO->isExact() && SubPattern.match(V); |
1357 | return false; |
1358 | } |
1359 | }; |
1360 | |
1361 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { |
1362 | return SubPattern; |
1363 | } |
1364 | |
1365 | |
1366 | |
1367 | |
1368 | |
1369 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, |
1370 | bool Commutable = false> |
1371 | struct CmpClass_match { |
1372 | PredicateTy &Predicate; |
1373 | LHS_t L; |
1374 | RHS_t R; |
1375 | |
1376 | |
1377 | |
1378 | CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS) |
1379 | : Predicate(Pred), L(LHS), R(RHS) {} |
1380 | |
1381 | template <typename OpTy> bool match(OpTy *V) { |
1382 | if (auto *I = dyn_cast<Class>(V)) { |
1383 | if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) { |
1384 | Predicate = I->getPredicate(); |
1385 | return true; |
1386 | } else if (Commutable && L.match(I->getOperand(1)) && |
1387 | R.match(I->getOperand(0))) { |
1388 | Predicate = I->getSwappedPredicate(); |
1389 | return true; |
1390 | } |
1391 | } |
1392 | return false; |
1393 | } |
1394 | }; |
1395 | |
1396 | template <typename LHS, typename RHS> |
1397 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> |
1398 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1399 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); |
1400 | } |
1401 | |
1402 | template <typename LHS, typename RHS> |
1403 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> |
1404 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1405 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); |
1406 | } |
1407 | |
1408 | template <typename LHS, typename RHS> |
1409 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> |
1410 | m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
1411 | return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R); |
1412 | } |
1413 | |
1414 | |
1415 | |
1416 | |
1417 | |
1418 | |
1419 | template <typename T0, unsigned Opcode> struct OneOps_match { |
1420 | T0 Op1; |
1421 | |
1422 | OneOps_match(const T0 &Op1) : Op1(Op1) {} |
1423 | |
1424 | template <typename OpTy> bool match(OpTy *V) { |
1425 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
1426 | auto *I = cast<Instruction>(V); |
1427 | return Op1.match(I->getOperand(0)); |
1428 | } |
1429 | return false; |
1430 | } |
1431 | }; |
1432 | |
1433 | |
1434 | template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match { |
1435 | T0 Op1; |
1436 | T1 Op2; |
1437 | |
1438 | TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {} |
1439 | |
1440 | template <typename OpTy> bool match(OpTy *V) { |
1441 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
1442 | auto *I = cast<Instruction>(V); |
1443 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)); |
1444 | } |
1445 | return false; |
1446 | } |
1447 | }; |
1448 | |
1449 | |
1450 | template <typename T0, typename T1, typename T2, unsigned Opcode> |
1451 | struct ThreeOps_match { |
1452 | T0 Op1; |
1453 | T1 Op2; |
1454 | T2 Op3; |
1455 | |
1456 | ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3) |
1457 | : Op1(Op1), Op2(Op2), Op3(Op3) {} |
1458 | |
1459 | template <typename OpTy> bool match(OpTy *V) { |
1460 | if (V->getValueID() == Value::InstructionVal + Opcode) { |
| 28 | | Assuming the condition is true | |
|
| |
1461 | auto *I = cast<Instruction>(V); |
| |
1462 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
| 31 | | Returning the value 1, which participates in a condition later | |
|
1463 | Op3.match(I->getOperand(2)); |
1464 | } |
1465 | return false; |
1466 | } |
1467 | }; |
1468 | |
1469 | |
1470 | template <typename Cond, typename LHS, typename RHS> |
1471 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> |
1472 | m_Select(const Cond &C, const LHS &L, const RHS &R) { |
1473 | return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R); |
1474 | } |
1475 | |
1476 | |
1477 | |
1478 | template <int64_t L, int64_t R, typename Cond> |
1479 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, |
1480 | Instruction::Select> |
1481 | m_SelectCst(const Cond &C) { |
1482 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); |
1483 | } |
1484 | |
1485 | |
1486 | template <typename OpTy> |
1487 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { |
1488 | return OneOps_match<OpTy, Instruction::Freeze>(Op); |
1489 | } |
1490 | |
1491 | |
1492 | template <typename Val_t, typename Elt_t, typename Idx_t> |
1493 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> |
1494 | m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) { |
1495 | return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>( |
1496 | Val, Elt, Idx); |
1497 | } |
1498 | |
1499 | |
1500 | template <typename Val_t, typename Idx_t> |
1501 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> |
1502 | m_ExtractElt(const Val_t &Val, const Idx_t &Idx) { |
1503 | return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx); |
1504 | } |
1505 | |
1506 | |
1507 | template <typename T0, typename T1, typename T2> struct Shuffle_match { |
1508 | T0 Op1; |
1509 | T1 Op2; |
1510 | T2 Mask; |
1511 | |
1512 | Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask) |
1513 | : Op1(Op1), Op2(Op2), Mask(Mask) {} |
1514 | |
1515 | template <typename OpTy> bool match(OpTy *V) { |
1516 | if (auto *I = dyn_cast<ShuffleVectorInst>(V)) { |
1517 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && |
1518 | Mask.match(I->getShuffleMask()); |
1519 | } |
1520 | return false; |
1521 | } |
1522 | }; |
1523 | |
1524 | struct m_Mask { |
1525 | ArrayRef<int> &MaskRef; |
1526 | m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
1527 | bool match(ArrayRef<int> Mask) { |
1528 | MaskRef = Mask; |
1529 | return true; |
1530 | } |
1531 | }; |
1532 | |
1533 | struct m_ZeroMask { |
1534 | bool match(ArrayRef<int> Mask) { |
1535 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); |
1536 | } |
1537 | }; |
1538 | |
1539 | struct m_SpecificMask { |
1540 | ArrayRef<int> &MaskRef; |
1541 | m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} |
1542 | bool match(ArrayRef<int> Mask) { return MaskRef == Mask; } |
1543 | }; |
1544 | |
1545 | struct m_SplatOrUndefMask { |
1546 | int &SplatIndex; |
1547 | m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {} |
1548 | bool match(ArrayRef<int> Mask) { |
1549 | auto First = find_if(Mask, [](int Elem) { return Elem != -1; }); |
1550 | if (First == Mask.end()) |
1551 | return false; |
1552 | SplatIndex = *First; |
1553 | return all_of(Mask, |
1554 | [First](int Elem) { return Elem == *First || Elem == -1; }); |
1555 | } |
1556 | }; |
1557 | |
1558 | |
1559 | template <typename V1_t, typename V2_t> |
1560 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> |
1561 | m_Shuffle(const V1_t &v1, const V2_t &v2) { |
1562 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); |
1563 | } |
1564 | |
1565 | template <typename V1_t, typename V2_t, typename Mask_t> |
1566 | inline Shuffle_match<V1_t, V2_t, Mask_t> |
1567 | m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) { |
1568 | return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask); |
1569 | } |
1570 | |
1571 | |
1572 | template <typename OpTy> |
1573 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { |
1574 | return OneOps_match<OpTy, Instruction::Load>(Op); |
1575 | } |
1576 | |
1577 | |
1578 | template <typename ValueOpTy, typename PointerOpTy> |
1579 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> |
1580 | m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) { |
1581 | return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp, |
1582 | PointerOp); |
1583 | } |
1584 | |
1585 | |
1586 | |
1587 | |
1588 | |
1589 | template <typename Op_t, unsigned Opcode> struct CastClass_match { |
1590 | Op_t Op; |
1591 | |
1592 | CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {} |
1593 | |
1594 | template <typename OpTy> bool match(OpTy *V) { |
1595 | if (auto *O = dyn_cast<Operator>(V)) |
1596 | return O->getOpcode() == Opcode && Op.match(O->getOperand(0)); |
1597 | return false; |
1598 | } |
1599 | }; |
1600 | |
1601 | |
1602 | template <typename OpTy> |
1603 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { |
1604 | return CastClass_match<OpTy, Instruction::BitCast>(Op); |
1605 | } |
1606 | |
1607 | |
1608 | template <typename OpTy> |
1609 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { |
1610 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); |
1611 | } |
1612 | |
1613 | |
1614 | template <typename OpTy> |
1615 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { |
1616 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); |
1617 | } |
1618 | |
1619 | |
1620 | template <typename OpTy> |
1621 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { |
1622 | return CastClass_match<OpTy, Instruction::Trunc>(Op); |
1623 | } |
1624 | |
1625 | template <typename OpTy> |
1626 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> |
1627 | m_TruncOrSelf(const OpTy &Op) { |
1628 | return m_CombineOr(m_Trunc(Op), Op); |
1629 | } |
1630 | |
1631 | |
1632 | template <typename OpTy> |
1633 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { |
1634 | return CastClass_match<OpTy, Instruction::SExt>(Op); |
1635 | } |
1636 | |
1637 | |
1638 | template <typename OpTy> |
1639 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { |
1640 | return CastClass_match<OpTy, Instruction::ZExt>(Op); |
1641 | } |
1642 | |
1643 | template <typename OpTy> |
1644 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> |
1645 | m_ZExtOrSelf(const OpTy &Op) { |
1646 | return m_CombineOr(m_ZExt(Op), Op); |
1647 | } |
1648 | |
1649 | template <typename OpTy> |
1650 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> |
1651 | m_SExtOrSelf(const OpTy &Op) { |
1652 | return m_CombineOr(m_SExt(Op), Op); |
1653 | } |
1654 | |
1655 | template <typename OpTy> |
1656 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
1657 | CastClass_match<OpTy, Instruction::SExt>> |
1658 | m_ZExtOrSExt(const OpTy &Op) { |
1659 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); |
1660 | } |
1661 | |
1662 | template <typename OpTy> |
1663 | inline match_combine_or< |
1664 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, |
1665 | CastClass_match<OpTy, Instruction::SExt>>, |
1666 | OpTy> |
1667 | m_ZExtOrSExtOrSelf(const OpTy &Op) { |
1668 | return m_CombineOr(m_ZExtOrSExt(Op), Op); |
1669 | } |
1670 | |
1671 | template <typename OpTy> |
1672 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { |
1673 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); |
1674 | } |
1675 | |
1676 | template <typename OpTy> |
1677 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { |
1678 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); |
1679 | } |
1680 | |
1681 | template <typename OpTy> |
1682 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { |
1683 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); |
1684 | } |
1685 | |
1686 | template <typename OpTy> |
1687 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { |
1688 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); |
1689 | } |
1690 | |
1691 | template <typename OpTy> |
1692 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { |
1693 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); |
1694 | } |
1695 | |
1696 | template <typename OpTy> |
1697 | inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) { |
1698 | return CastClass_match<OpTy, Instruction::FPExt>(Op); |
1699 | } |
1700 | |
1701 | |
1702 | |
1703 | |
1704 | |
1705 | struct br_match { |
1706 | BasicBlock *&Succ; |
1707 | |
1708 | br_match(BasicBlock *&Succ) : Succ(Succ) {} |
1709 | |
1710 | template <typename OpTy> bool match(OpTy *V) { |
1711 | if (auto *BI = dyn_cast<BranchInst>(V)) |
1712 | if (BI->isUnconditional()) { |
1713 | Succ = BI->getSuccessor(0); |
1714 | return true; |
1715 | } |
1716 | return false; |
1717 | } |
1718 | }; |
1719 | |
1720 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } |
1721 | |
1722 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
1723 | struct brc_match { |
1724 | Cond_t Cond; |
1725 | TrueBlock_t T; |
1726 | FalseBlock_t F; |
1727 | |
1728 | brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f) |
1729 | : Cond(C), T(t), F(f) {} |
1730 | |
1731 | template <typename OpTy> bool match(OpTy *V) { |
1732 | if (auto *BI = dyn_cast<BranchInst>(V)) |
1733 | if (BI->isConditional() && Cond.match(BI->getCondition())) |
1734 | return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1)); |
1735 | return false; |
1736 | } |
1737 | }; |
1738 | |
1739 | template <typename Cond_t> |
1740 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> |
1741 | m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) { |
1742 | return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>( |
1743 | C, m_BasicBlock(T), m_BasicBlock(F)); |
1744 | } |
1745 | |
1746 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> |
1747 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> |
1748 | m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) { |
1749 | return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F); |
1750 | } |
1751 | |
1752 | |
1753 | |
1754 | |
1755 | |
1756 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, |
1757 | bool Commutable = false> |
1758 | struct MaxMin_match { |
1759 | using PredType = Pred_t; |
1760 | LHS_t L; |
1761 | RHS_t R; |
1762 | |
1763 | |
1764 | |
1765 | MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} |
1766 | |
1767 | template <typename OpTy> bool match(OpTy *V) { |
1768 | if (auto *II = dyn_cast<IntrinsicInst>(V)) { |
1769 | Intrinsic::ID IID = II->getIntrinsicID(); |
1770 | if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) || |
1771 | (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) || |
1772 | (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) || |
1773 | (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) { |
1774 | Value *LHS = II->getOperand(0), *RHS = II->getOperand(1); |
1775 | return (L.match(LHS) && R.match(RHS)) || |
1776 | (Commutable && L.match(RHS) && R.match(LHS)); |
1777 | } |
1778 | } |
1779 | |
1780 | auto *SI = dyn_cast<SelectInst>(V); |
1781 | if (!SI) |
1782 | return false; |
1783 | auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition()); |
1784 | if (!Cmp) |
1785 | return false; |
1786 | |
1787 | |
1788 | auto *TrueVal = SI->getTrueValue(); |
1789 | auto *FalseVal = SI->getFalseValue(); |
1790 | auto *LHS = Cmp->getOperand(0); |
1791 | auto *RHS = Cmp->getOperand(1); |
1792 | if ((TrueVal != LHS || FalseVal != RHS) && |
1793 | (TrueVal != RHS || FalseVal != LHS)) |
1794 | return false; |
1795 | typename CmpInst_t::Predicate Pred = |
1796 | LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate(); |
1797 | |
1798 | if (!Pred_t::match(Pred)) |
1799 | return false; |
1800 | |
1801 | return (L.match(LHS) && R.match(RHS)) || |
1802 | (Commutable && L.match(RHS) && R.match(LHS)); |
1803 | } |
1804 | }; |
1805 | |
1806 | |
1807 | struct smax_pred_ty { |
1808 | static bool match(ICmpInst::Predicate Pred) { |
1809 | return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE; |
1810 | } |
1811 | }; |
1812 | |
1813 | |
1814 | struct smin_pred_ty { |
1815 | static bool match(ICmpInst::Predicate Pred) { |
1816 | return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE; |
1817 | } |
1818 | }; |
1819 | |
1820 | |
1821 | struct umax_pred_ty { |
1822 | static bool match(ICmpInst::Predicate Pred) { |
1823 | return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE; |
1824 | } |
1825 | }; |
1826 | |
1827 | |
1828 | struct umin_pred_ty { |
1829 | static bool match(ICmpInst::Predicate Pred) { |
1830 | return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE; |
1831 | } |
1832 | }; |
1833 | |
1834 | |
1835 | struct ofmax_pred_ty { |
1836 | static bool match(FCmpInst::Predicate Pred) { |
1837 | return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE; |
1838 | } |
1839 | }; |
1840 | |
1841 | |
1842 | struct ofmin_pred_ty { |
1843 | static bool match(FCmpInst::Predicate Pred) { |
1844 | return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE; |
1845 | } |
1846 | }; |
1847 | |
1848 | |
1849 | struct ufmax_pred_ty { |
1850 | static bool match(FCmpInst::Predicate Pred) { |
1851 | return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE; |
1852 | } |
1853 | }; |
1854 | |
1855 | |
1856 | struct ufmin_pred_ty { |
1857 | static bool match(FCmpInst::Predicate Pred) { |
1858 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; |
1859 | } |
1860 | }; |
1861 | |
1862 | template <typename LHS, typename RHS> |
1863 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L, |
1864 | const RHS &R) { |
1865 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R); |
1866 | } |
1867 | |
1868 | template <typename LHS, typename RHS> |
1869 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L, |
1870 | const RHS &R) { |
1871 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R); |
1872 | } |
1873 | |
1874 | template <typename LHS, typename RHS> |
1875 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L, |
1876 | const RHS &R) { |
1877 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R); |
1878 | } |
1879 | |
1880 | template <typename LHS, typename RHS> |
1881 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L, |
1882 | const RHS &R) { |
1883 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R); |
1884 | } |
1885 | |
1886 | template <typename LHS, typename RHS> |
1887 | inline match_combine_or< |
1888 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>, |
1889 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>, |
1890 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>, |
1891 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>> |
1892 | m_MaxOrMin(const LHS &L, const RHS &R) { |
1893 | return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)), |
1894 | m_CombineOr(m_UMax(L, R), m_UMin(L, R))); |
1895 | } |
1896 | |
1897 | |
1898 | |
1899 | |
1900 | |
1901 | |
1902 | |
1903 | |
1904 | |
1905 | |
1906 | template <typename LHS, typename RHS> |
1907 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L, |
1908 | const RHS &R) { |
1909 | return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R); |
1910 | } |
1911 | |
1912 | |
1913 | |
1914 | |
1915 | |
1916 | |
1917 | |
1918 | |
1919 | |
1920 | |
1921 | template <typename LHS, typename RHS> |
1922 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L, |
1923 | const RHS &R) { |
1924 | return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R); |
1925 | } |
1926 | |
1927 | |
1928 | |
1929 | |
1930 | |
1931 | |
1932 | |
1933 | |
1934 | |
1935 | |
1936 | template <typename LHS, typename RHS> |
1937 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> |
1938 | m_UnordFMax(const LHS &L, const RHS &R) { |
1939 | return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R); |
1940 | } |
1941 | |
1942 | |
1943 | |
1944 | |
1945 | |
1946 | |
1947 | |
1948 | |
1949 | |
1950 | |
1951 | template <typename LHS, typename RHS> |
1952 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> |
1953 | m_UnordFMin(const LHS &L, const RHS &R) { |
1954 | return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R); |
1955 | } |
1956 | |
1957 | |
1958 | |
1959 | |
1960 | |
1961 | |
1962 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
1963 | struct UAddWithOverflow_match { |
1964 | LHS_t L; |
1965 | RHS_t R; |
1966 | Sum_t S; |
1967 | |
1968 | UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S) |
1969 | : L(L), R(R), S(S) {} |
1970 | |
1971 | template <typename OpTy> bool match(OpTy *V) { |
1972 | Value *ICmpLHS, *ICmpRHS; |
1973 | ICmpInst::Predicate Pred; |
1974 | if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V)) |
1975 | return false; |
1976 | |
1977 | Value *AddLHS, *AddRHS; |
1978 | auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS)); |
1979 | |
1980 | |
1981 | if (Pred == ICmpInst::ICMP_ULT) |
1982 | if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS)) |
1983 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
1984 | |
1985 | |
1986 | if (Pred == ICmpInst::ICMP_UGT) |
1987 | if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS)) |
1988 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
1989 | |
1990 | Value *Op1; |
1991 | auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes())); |
1992 | |
1993 | if (Pred == ICmpInst::ICMP_ULT) { |
1994 | if (XorExpr.match(ICmpLHS)) |
1995 | return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS); |
1996 | } |
1997 | |
1998 | if (Pred == ICmpInst::ICMP_UGT) { |
1999 | if (XorExpr.match(ICmpRHS)) |
2000 | return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS); |
2001 | } |
2002 | |
2003 | |
2004 | if (Pred == ICmpInst::ICMP_EQ) { |
2005 | |
2006 | |
2007 | if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) && |
2008 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
2009 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); |
2010 | |
2011 | |
2012 | if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) && |
2013 | (m_One().match(AddLHS) || m_One().match(AddRHS))) |
2014 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); |
2015 | } |
2016 | |
2017 | return false; |
2018 | } |
2019 | }; |
2020 | |
2021 | |
2022 | |
2023 | |
2024 | |
2025 | template <typename LHS_t, typename RHS_t, typename Sum_t> |
2026 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> |
2027 | m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) { |
2028 | return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S); |
2029 | } |
2030 | |
2031 | template <typename Opnd_t> struct Argument_match { |
2032 | unsigned OpI; |
2033 | Opnd_t Val; |
2034 | |
2035 | Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {} |
2036 | |
2037 | template <typename OpTy> bool match(OpTy *V) { |
2038 | |
2039 | if (const auto *CI = dyn_cast<CallInst>(V)) |
2040 | return Val.match(CI->getArgOperand(OpI)); |
2041 | return false; |
2042 | } |
2043 | }; |
2044 | |
2045 | |
2046 | template <unsigned OpI, typename Opnd_t> |
2047 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { |
2048 | return Argument_match<Opnd_t>(OpI, Op); |
2049 | } |
2050 | |
2051 | |
2052 | struct IntrinsicID_match { |
2053 | unsigned ID; |
2054 | |
2055 | IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {} |
2056 | |
2057 | template <typename OpTy> bool match(OpTy *V) { |
2058 | if (const auto *CI = dyn_cast<CallInst>(V)) |
2059 | if (const auto *F = CI->getCalledFunction()) |
2060 | return F->getIntrinsicID() == ID; |
2061 | return false; |
2062 | } |
2063 | }; |
2064 | |
2065 | |
2066 | |
2067 | |
2068 | |
2069 | template <typename T0 = void, typename T1 = void, typename T2 = void, |
2070 | typename T3 = void, typename T4 = void, typename T5 = void, |
2071 | typename T6 = void, typename T7 = void, typename T8 = void, |
2072 | typename T9 = void, typename T10 = void> |
2073 | struct m_Intrinsic_Ty; |
2074 | template <typename T0> struct m_Intrinsic_Ty<T0> { |
2075 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; |
2076 | }; |
2077 | template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> { |
2078 | using Ty = |
2079 | match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>; |
2080 | }; |
2081 | template <typename T0, typename T1, typename T2> |
2082 | struct m_Intrinsic_Ty<T0, T1, T2> { |
2083 | using Ty = |
2084 | match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty, |
2085 | Argument_match<T2>>; |
2086 | }; |
2087 | template <typename T0, typename T1, typename T2, typename T3> |
2088 | struct m_Intrinsic_Ty<T0, T1, T2, T3> { |
2089 | using Ty = |
2090 | match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty, |
2091 | Argument_match<T3>>; |
2092 | }; |
2093 | |
2094 | template <typename T0, typename T1, typename T2, typename T3, typename T4> |
2095 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> { |
2096 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty, |
2097 | Argument_match<T4>>; |
2098 | }; |
2099 | |
2100 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> |
2101 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> { |
2102 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty, |
2103 | Argument_match<T5>>; |
2104 | }; |
2105 | |
2106 | |
2107 | |
2108 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { |
2109 | return IntrinsicID_match(IntrID); |
2110 | } |
2111 | |
2112 | |
2113 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> |
2114 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty |
2115 | m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, |
2116 | const Opnd3 &Op3) { |
2117 | return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3); |
2118 | } |
2119 | |
2120 | template <Intrinsic::ID IntrID, typename T0> |
2121 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { |
2122 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); |
2123 | } |
2124 | |
2125 | template <Intrinsic::ID IntrID, typename T0, typename T1> |
2126 | inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0, |
2127 | const T1 &Op1) { |
2128 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1)); |
2129 | } |
2130 | |
2131 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> |
2132 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty |
2133 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) { |
2134 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2)); |
2135 | } |
2136 | |
2137 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2138 | typename T3> |
2139 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty |
2140 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) { |
2141 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3)); |
2142 | } |
2143 | |
2144 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2145 | typename T3, typename T4> |
2146 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty |
2147 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
2148 | const T4 &Op4) { |
2149 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3), |
2150 | m_Argument<4>(Op4)); |
2151 | } |
2152 | |
2153 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, |
2154 | typename T3, typename T4, typename T5> |
2155 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty |
2156 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, |
2157 | const T4 &Op4, const T5 &Op5) { |
2158 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4), |
2159 | m_Argument<5>(Op5)); |
2160 | } |
2161 | |
2162 | |
2163 | template <typename Opnd0> |
2164 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { |
2165 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); |
2166 | } |
2167 | |
2168 | template <typename Opnd0> |
2169 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { |
2170 | return m_Intrinsic<Intrinsic::bswap>(Op0); |
2171 | } |
2172 | |
2173 | template <typename Opnd0> |
2174 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { |
2175 | return m_Intrinsic<Intrinsic::fabs>(Op0); |
2176 | } |
2177 | |
2178 | template <typename Opnd0> |
2179 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { |
2180 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); |
2181 | } |
2182 | |
2183 | template <typename Opnd0, typename Opnd1> |
2184 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0, |
2185 | const Opnd1 &Op1) { |
2186 | return m_Intrinsic<Intrinsic::minnum>(Op0, Op1); |
2187 | } |
2188 | |
2189 | template <typename Opnd0, typename Opnd1> |
2190 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0, |
2191 | const Opnd1 &Op1) { |
2192 | return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1); |
2193 | } |
2194 | |
2195 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
2196 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
2197 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
2198 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); |
2199 | } |
2200 | |
2201 | template <typename Opnd0, typename Opnd1, typename Opnd2> |
2202 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty |
2203 | m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { |
2204 | return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2); |
2205 | } |
2206 | |
2207 | |
2208 | |
2209 | |
2210 | |
2211 | |
2212 | template <typename LHS, typename RHS> |
2213 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { |
2214 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); |
2215 | } |
2216 | |
2217 | |
2218 | |
2219 | template <typename LHS, typename RHS> |
2220 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> |
2221 | m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { |
2222 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L, |
2223 | R); |
2224 | } |
2225 | |
2226 | |
2227 | template <typename LHS, typename RHS> |
2228 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, |
2229 | const RHS &R) { |
2230 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); |
2231 | } |
2232 | |
2233 | |
2234 | template <typename LHS, typename RHS> |
2235 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, |
2236 | const RHS &R) { |
2237 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); |
2238 | } |
2239 | |
2240 | |
2241 | template <typename LHS, typename RHS> |
2242 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, |
2243 | const RHS &R) { |
2244 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); |
2245 | } |
2246 | |
2247 | |
2248 | template <typename LHS, typename RHS> |
2249 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, |
2250 | const RHS &R) { |
2251 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); |
2252 | } |
2253 | |
2254 | |
2255 | template <typename LHS, typename RHS> |
2256 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, |
2257 | const RHS &R) { |
2258 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); |
2259 | } |
2260 | |
2261 | |
2262 | template <typename ValTy> |
2263 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> |
2264 | m_Neg(const ValTy &V) { |
2265 | return m_Sub(m_ZeroInt(), V); |
2266 | } |
2267 | |
2268 | |
2269 | template <typename ValTy> |
2270 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, |
2271 | Instruction::Sub, |
2272 | OverflowingBinaryOperator::NoSignedWrap> |
2273 | m_NSWNeg(const ValTy &V) { |
2274 | return m_NSWSub(m_ZeroInt(), V); |
2275 | } |
2276 | |
2277 | |
2278 | template <typename ValTy> |
2279 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> |
2280 | m_Not(const ValTy &V) { |
2281 | return m_c_Xor(V, m_AllOnes()); |
2282 | } |
2283 | |
2284 | |
2285 | template <typename LHS, typename RHS> |
2286 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> |
2287 | m_c_SMin(const LHS &L, const RHS &R) { |
2288 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); |
2289 | } |
2290 | |
2291 | template <typename LHS, typename RHS> |
2292 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> |
2293 | m_c_SMax(const LHS &L, const RHS &R) { |
2294 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); |
2295 | } |
2296 | |
2297 | template <typename LHS, typename RHS> |
2298 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> |
2299 | m_c_UMin(const LHS &L, const RHS &R) { |
2300 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); |
2301 | } |
2302 | |
2303 | template <typename LHS, typename RHS> |
2304 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> |
2305 | m_c_UMax(const LHS &L, const RHS &R) { |
2306 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); |
2307 | } |
2308 | |
2309 | template <typename LHS, typename RHS> |
2310 | inline match_combine_or< |
2311 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, |
2312 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, |
2313 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, |
2314 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> |
2315 | m_c_MaxOrMin(const LHS &L, const RHS &R) { |
2316 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), |
2317 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); |
2318 | } |
2319 | |
2320 | |
2321 | template <typename LHS, typename RHS> |
2322 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> |
2323 | m_c_FAdd(const LHS &L, const RHS &R) { |
2324 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); |
2325 | } |
2326 | |
2327 | |
2328 | template <typename LHS, typename RHS> |
2329 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> |
2330 | m_c_FMul(const LHS &L, const RHS &R) { |
2331 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); |
2332 | } |
2333 | |
2334 | template <typename Opnd_t> struct Signum_match { |
2335 | Opnd_t Val; |
2336 | Signum_match(const Opnd_t &V) : Val(V) {} |
2337 | |
2338 | template <typename OpTy> bool match(OpTy *V) { |
2339 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); |
2340 | if (TypeSize == 0) |
2341 | return false; |
2342 | |
2343 | unsigned ShiftWidth = TypeSize - 1; |
2344 | Value *OpL = nullptr, *OpR = nullptr; |
2345 | |
2346 | |
2347 | |
2348 | |
2349 | |
2350 | |
2351 | |
2352 | |
2353 | |
2354 | |
2355 | |
2356 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); |
2357 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); |
2358 | auto Signum = m_Or(LHS, RHS); |
2359 | |
2360 | return Signum.match(V) && OpL == OpR && Val.match(OpL); |
2361 | } |
2362 | }; |
2363 | |
2364 | |
2365 | |
2366 | |
2367 | |
2368 | |
2369 | |
2370 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { |
2371 | return Signum_match<Val_t>(V); |
2372 | } |
2373 | |
2374 | template <int Ind, typename Opnd_t> struct ExtractValue_match { |
2375 | Opnd_t Val; |
2376 | ExtractValue_match(const Opnd_t &V) : Val(V) {} |
2377 | |
2378 | template <typename OpTy> bool match(OpTy *V) { |
2379 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { |
2380 | |
2381 | if (Ind != -1 && |
2382 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) |
2383 | return false; |
2384 | return Val.match(I->getAggregateOperand()); |
2385 | } |
2386 | return false; |
2387 | } |
2388 | }; |
2389 | |
2390 | |
2391 | |
2392 | template <int Ind, typename Val_t> |
2393 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { |
2394 | return ExtractValue_match<Ind, Val_t>(V); |
2395 | } |
2396 | |
2397 | |
2398 | |
2399 | template <typename Val_t> |
2400 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { |
2401 | return ExtractValue_match<-1, Val_t>(V); |
2402 | } |
2403 | |
2404 | |
2405 | template <int Ind, typename T0, typename T1> struct InsertValue_match { |
2406 | T0 Op0; |
2407 | T1 Op1; |
2408 | |
2409 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} |
2410 | |
2411 | template <typename OpTy> bool match(OpTy *V) { |
2412 | if (auto *I = dyn_cast<InsertValueInst>(V)) { |
2413 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && |
2414 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; |
2415 | } |
2416 | return false; |
2417 | } |
2418 | }; |
2419 | |
2420 | |
2421 | template <int Ind, typename Val_t, typename Elt_t> |
2422 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, |
2423 | const Elt_t &Elt) { |
2424 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); |
2425 | } |
2426 | |
2427 | |
2428 | |
2429 | |
2430 | |
2431 | struct VScaleVal_match { |
2432 | const DataLayout &DL; |
2433 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} |
2434 | |
2435 | template <typename ITy> bool match(ITy *V) { |
2436 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) |
2437 | return true; |
2438 | |
2439 | Value *Ptr; |
2440 | if (m_PtrToInt(m_Value(Ptr)).match(V)) { |
2441 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { |
2442 | auto *DerefTy = GEP->getSourceElementType(); |
2443 | if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) && |
2444 | m_Zero().match(GEP->getPointerOperand()) && |
2445 | m_SpecificInt(1).match(GEP->idx_begin()->get()) && |
2446 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) |
2447 | return true; |
2448 | } |
2449 | } |
2450 | |
2451 | return false; |
2452 | } |
2453 | }; |
2454 | |
2455 | inline VScaleVal_match m_VScale(const DataLayout &DL) { |
2456 | return VScaleVal_match(DL); |
2457 | } |
2458 | |
2459 | template <typename LHS, typename RHS, unsigned Opcode> |
2460 | struct LogicalOp_match { |
2461 | LHS L; |
2462 | RHS R; |
2463 | |
2464 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} |
2465 | |
2466 | template <typename T> bool match(T *V) { |
2467 | if (auto *I = dyn_cast<Instruction>(V)) { |
2468 | if (!I->getType()->isIntOrIntVectorTy(1)) |
2469 | return false; |
2470 | |
2471 | if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) && |
2472 | R.match(I->getOperand(1))) |
2473 | return true; |
2474 | |
2475 | if (auto *SI = dyn_cast<SelectInst>(I)) { |
2476 | if (Opcode == Instruction::And) { |
2477 | if (const auto *C = dyn_cast<Constant>(SI->getFalseValue())) |
2478 | if (C->isNullValue() && L.match(SI->getCondition()) && |
2479 | R.match(SI->getTrueValue())) |
2480 | return true; |
2481 | } else { |
2482 | assert(Opcode == Instruction::Or); |
2483 | if (const auto *C = dyn_cast<Constant>(SI->getTrueValue())) |
2484 | if (C->isOneValue() && L.match(SI->getCondition()) && |
2485 | R.match(SI->getFalseValue())) |
2486 | return true; |
2487 | } |
2488 | } |
2489 | } |
2490 | |
2491 | return false; |
2492 | } |
2493 | }; |
2494 | |
2495 | |
2496 | |
2497 | template <typename LHS, typename RHS> |
2498 | inline LogicalOp_match<LHS, RHS, Instruction::And> |
2499 | m_LogicalAnd(const LHS &L, const RHS &R) { |
2500 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); |
2501 | } |
2502 | |
2503 | |
2504 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } |
2505 | |
2506 | |
2507 | |
2508 | template <typename LHS, typename RHS> |
2509 | inline LogicalOp_match<LHS, RHS, Instruction::Or> |
2510 | m_LogicalOr(const LHS &L, const RHS &R) { |
2511 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); |
2512 | } |
2513 | |
2514 | |
2515 | inline auto m_LogicalOr() { |
2516 | return m_LogicalOr(m_Value(), m_Value()); |
2517 | } |
2518 | |
2519 | } |
2520 | } |
2521 | |
2522 | #endif // LLVM_IR_PATTERNMATCH_H |