1 | //===- InstCombineCalls.cpp -----------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the visitCall, visitInvoke, and visitCallBr functions. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "InstCombineInternal.h" |
14 | #include "llvm/ADT/APFloat.h" |
15 | #include "llvm/ADT/APInt.h" |
16 | #include "llvm/ADT/APSInt.h" |
17 | #include "llvm/ADT/ArrayRef.h" |
18 | #include "llvm/ADT/STLFunctionalExtras.h" |
19 | #include "llvm/ADT/SmallBitVector.h" |
20 | #include "llvm/ADT/SmallVector.h" |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/Analysis/AliasAnalysis.h" |
23 | #include "llvm/Analysis/AssumeBundleQueries.h" |
24 | #include "llvm/Analysis/AssumptionCache.h" |
25 | #include "llvm/Analysis/InstructionSimplify.h" |
26 | #include "llvm/Analysis/Loads.h" |
27 | #include "llvm/Analysis/MemoryBuiltins.h" |
28 | #include "llvm/Analysis/ValueTracking.h" |
29 | #include "llvm/Analysis/VectorUtils.h" |
30 | #include "llvm/IR/AttributeMask.h" |
31 | #include "llvm/IR/Attributes.h" |
32 | #include "llvm/IR/BasicBlock.h" |
33 | #include "llvm/IR/Constant.h" |
34 | #include "llvm/IR/Constants.h" |
35 | #include "llvm/IR/DataLayout.h" |
36 | #include "llvm/IR/DebugInfo.h" |
37 | #include "llvm/IR/DerivedTypes.h" |
38 | #include "llvm/IR/Function.h" |
39 | #include "llvm/IR/GlobalVariable.h" |
40 | #include "llvm/IR/InlineAsm.h" |
41 | #include "llvm/IR/InstrTypes.h" |
42 | #include "llvm/IR/Instruction.h" |
43 | #include "llvm/IR/Instructions.h" |
44 | #include "llvm/IR/IntrinsicInst.h" |
45 | #include "llvm/IR/Intrinsics.h" |
46 | #include "llvm/IR/IntrinsicsAArch64.h" |
47 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
48 | #include "llvm/IR/IntrinsicsARM.h" |
49 | #include "llvm/IR/IntrinsicsHexagon.h" |
50 | #include "llvm/IR/LLVMContext.h" |
51 | #include "llvm/IR/Metadata.h" |
52 | #include "llvm/IR/PatternMatch.h" |
53 | #include "llvm/IR/Statepoint.h" |
54 | #include "llvm/IR/Type.h" |
55 | #include "llvm/IR/User.h" |
56 | #include "llvm/IR/Value.h" |
57 | #include "llvm/IR/ValueHandle.h" |
58 | #include "llvm/Support/AtomicOrdering.h" |
59 | #include "llvm/Support/Casting.h" |
60 | #include "llvm/Support/CommandLine.h" |
61 | #include "llvm/Support/Compiler.h" |
62 | #include "llvm/Support/Debug.h" |
63 | #include "llvm/Support/ErrorHandling.h" |
64 | #include "llvm/Support/KnownBits.h" |
65 | #include "llvm/Support/KnownFPClass.h" |
66 | #include "llvm/Support/MathExtras.h" |
67 | #include "llvm/Support/raw_ostream.h" |
68 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
69 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
70 | #include "llvm/Transforms/Utils/Local.h" |
71 | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" |
72 | #include <algorithm> |
73 | #include <cassert> |
74 | #include <cstdint> |
75 | #include <optional> |
76 | #include <utility> |
77 | #include <vector> |
78 | |
79 | #define DEBUG_TYPE "instcombine" |
80 | #include "llvm/Transforms/Utils/InstructionWorklist.h" |
81 | |
82 | using namespace llvm; |
83 | using namespace PatternMatch; |
84 | |
85 | STATISTIC(NumSimplified, "Number of library calls simplified" ); |
86 | |
87 | static cl::opt<unsigned> GuardWideningWindow( |
88 | "instcombine-guard-widening-window" , |
89 | cl::init(Val: 3), |
90 | cl::desc("How wide an instruction window to bypass looking for " |
91 | "another guard" )); |
92 | |
93 | /// Return the specified type promoted as it would be to pass though a va_arg |
94 | /// area. |
95 | static Type *getPromotedType(Type *Ty) { |
96 | if (IntegerType* ITy = dyn_cast<IntegerType>(Val: Ty)) { |
97 | if (ITy->getBitWidth() < 32) |
98 | return Type::getInt32Ty(C&: Ty->getContext()); |
99 | } |
100 | return Ty; |
101 | } |
102 | |
103 | /// Recognize a memcpy/memmove from a trivially otherwise unused alloca. |
104 | /// TODO: This should probably be integrated with visitAllocSites, but that |
105 | /// requires a deeper change to allow either unread or unwritten objects. |
106 | static bool hasUndefSource(AnyMemTransferInst *MI) { |
107 | auto *Src = MI->getRawSource(); |
108 | while (isa<GetElementPtrInst>(Val: Src)) { |
109 | if (!Src->hasOneUse()) |
110 | return false; |
111 | Src = cast<Instruction>(Val: Src)->getOperand(i: 0); |
112 | } |
113 | return isa<AllocaInst>(Val: Src) && Src->hasOneUse(); |
114 | } |
115 | |
116 | Instruction *InstCombinerImpl::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) { |
117 | Align DstAlign = getKnownAlignment(V: MI->getRawDest(), DL, CxtI: MI, AC: &AC, DT: &DT); |
118 | MaybeAlign CopyDstAlign = MI->getDestAlign(); |
119 | if (!CopyDstAlign || *CopyDstAlign < DstAlign) { |
120 | MI->setDestAlignment(DstAlign); |
121 | return MI; |
122 | } |
123 | |
124 | Align SrcAlign = getKnownAlignment(V: MI->getRawSource(), DL, CxtI: MI, AC: &AC, DT: &DT); |
125 | MaybeAlign CopySrcAlign = MI->getSourceAlign(); |
126 | if (!CopySrcAlign || *CopySrcAlign < SrcAlign) { |
127 | MI->setSourceAlignment(SrcAlign); |
128 | return MI; |
129 | } |
130 | |
131 | // If we have a store to a location which is known constant, we can conclude |
132 | // that the store must be storing the constant value (else the memory |
133 | // wouldn't be constant), and this must be a noop. |
134 | if (!isModSet(MRI: AA->getModRefInfoMask(P: MI->getDest()))) { |
135 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
136 | MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType())); |
137 | return MI; |
138 | } |
139 | |
140 | // If the source is provably undef, the memcpy/memmove doesn't do anything |
141 | // (unless the transfer is volatile). |
142 | if (hasUndefSource(MI) && !MI->isVolatile()) { |
143 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
144 | MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType())); |
145 | return MI; |
146 | } |
147 | |
148 | // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with |
149 | // load/store. |
150 | ConstantInt *MemOpLength = dyn_cast<ConstantInt>(Val: MI->getLength()); |
151 | if (!MemOpLength) return nullptr; |
152 | |
153 | // Source and destination pointer types are always "i8*" for intrinsic. See |
154 | // if the size is something we can handle with a single primitive load/store. |
155 | // A single load+store correctly handles overlapping memory in the memmove |
156 | // case. |
157 | uint64_t Size = MemOpLength->getLimitedValue(); |
158 | assert(Size && "0-sized memory transferring should be removed already." ); |
159 | |
160 | if (Size > 8 || (Size&(Size-1))) |
161 | return nullptr; // If not 1/2/4/8 bytes, exit. |
162 | |
163 | // If it is an atomic and alignment is less than the size then we will |
164 | // introduce the unaligned memory access which will be later transformed |
165 | // into libcall in CodeGen. This is not evident performance gain so disable |
166 | // it now. |
167 | if (MI->isAtomic()) |
168 | if (*CopyDstAlign < Size || *CopySrcAlign < Size) |
169 | return nullptr; |
170 | |
171 | // Use an integer load+store unless we can find something better. |
172 | IntegerType* IntType = IntegerType::get(C&: MI->getContext(), NumBits: Size<<3); |
173 | |
174 | // If the memcpy has metadata describing the members, see if we can get the |
175 | // TBAA, scope and noalias tags describing our copy. |
176 | AAMDNodes AACopyMD = MI->getAAMetadata().adjustForAccess(AccessSize: Size); |
177 | |
178 | Value *Src = MI->getArgOperand(i: 1); |
179 | Value *Dest = MI->getArgOperand(i: 0); |
180 | LoadInst *L = Builder.CreateLoad(Ty: IntType, Ptr: Src); |
181 | // Alignment from the mem intrinsic will be better, so use it. |
182 | L->setAlignment(*CopySrcAlign); |
183 | L->setAAMetadata(AACopyMD); |
184 | MDNode *LoopMemParallelMD = |
185 | MI->getMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access); |
186 | if (LoopMemParallelMD) |
187 | L->setMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access, Node: LoopMemParallelMD); |
188 | MDNode *AccessGroupMD = MI->getMetadata(KindID: LLVMContext::MD_access_group); |
189 | if (AccessGroupMD) |
190 | L->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessGroupMD); |
191 | |
192 | StoreInst *S = Builder.CreateStore(Val: L, Ptr: Dest); |
193 | // Alignment from the mem intrinsic will be better, so use it. |
194 | S->setAlignment(*CopyDstAlign); |
195 | S->setAAMetadata(AACopyMD); |
196 | if (LoopMemParallelMD) |
197 | S->setMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access, Node: LoopMemParallelMD); |
198 | if (AccessGroupMD) |
199 | S->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessGroupMD); |
200 | S->copyMetadata(SrcInst: *MI, WL: LLVMContext::MD_DIAssignID); |
201 | |
202 | if (auto *MT = dyn_cast<MemTransferInst>(Val: MI)) { |
203 | // non-atomics can be volatile |
204 | L->setVolatile(MT->isVolatile()); |
205 | S->setVolatile(MT->isVolatile()); |
206 | } |
207 | if (MI->isAtomic()) { |
208 | // atomics have to be unordered |
209 | L->setOrdering(AtomicOrdering::Unordered); |
210 | S->setOrdering(AtomicOrdering::Unordered); |
211 | } |
212 | |
213 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
214 | MI->setLength(Constant::getNullValue(Ty: MemOpLength->getType())); |
215 | return MI; |
216 | } |
217 | |
218 | Instruction *InstCombinerImpl::SimplifyAnyMemSet(AnyMemSetInst *MI) { |
219 | const Align KnownAlignment = |
220 | getKnownAlignment(V: MI->getDest(), DL, CxtI: MI, AC: &AC, DT: &DT); |
221 | MaybeAlign MemSetAlign = MI->getDestAlign(); |
222 | if (!MemSetAlign || *MemSetAlign < KnownAlignment) { |
223 | MI->setDestAlignment(KnownAlignment); |
224 | return MI; |
225 | } |
226 | |
227 | // If we have a store to a location which is known constant, we can conclude |
228 | // that the store must be storing the constant value (else the memory |
229 | // wouldn't be constant), and this must be a noop. |
230 | if (!isModSet(MRI: AA->getModRefInfoMask(P: MI->getDest()))) { |
231 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
232 | MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType())); |
233 | return MI; |
234 | } |
235 | |
236 | // Remove memset with an undef value. |
237 | // FIXME: This is technically incorrect because it might overwrite a poison |
238 | // value. Change to PoisonValue once #52930 is resolved. |
239 | if (isa<UndefValue>(Val: MI->getValue())) { |
240 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
241 | MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType())); |
242 | return MI; |
243 | } |
244 | |
245 | // Extract the length and alignment and fill if they are constant. |
246 | ConstantInt *LenC = dyn_cast<ConstantInt>(Val: MI->getLength()); |
247 | ConstantInt *FillC = dyn_cast<ConstantInt>(Val: MI->getValue()); |
248 | if (!LenC || !FillC || !FillC->getType()->isIntegerTy(Bitwidth: 8)) |
249 | return nullptr; |
250 | const uint64_t Len = LenC->getLimitedValue(); |
251 | assert(Len && "0-sized memory setting should be removed already." ); |
252 | const Align Alignment = MI->getDestAlign().valueOrOne(); |
253 | |
254 | // If it is an atomic and alignment is less than the size then we will |
255 | // introduce the unaligned memory access which will be later transformed |
256 | // into libcall in CodeGen. This is not evident performance gain so disable |
257 | // it now. |
258 | if (MI->isAtomic() && Alignment < Len) |
259 | return nullptr; |
260 | |
261 | // memset(s,c,n) -> store s, c (for n=1,2,4,8) |
262 | if (Len <= 8 && isPowerOf2_32(Value: (uint32_t)Len)) { |
263 | Value *Dest = MI->getDest(); |
264 | |
265 | // Extract the fill value and store. |
266 | Constant *FillVal = ConstantInt::get( |
267 | Context&: MI->getContext(), V: APInt::getSplat(NewLen: Len * 8, V: FillC->getValue())); |
268 | StoreInst *S = Builder.CreateStore(Val: FillVal, Ptr: Dest, isVolatile: MI->isVolatile()); |
269 | S->copyMetadata(SrcInst: *MI, WL: LLVMContext::MD_DIAssignID); |
270 | auto replaceOpForAssignmentMarkers = [FillC, FillVal](auto *DbgAssign) { |
271 | if (llvm::is_contained(DbgAssign->location_ops(), FillC)) |
272 | DbgAssign->replaceVariableLocationOp(FillC, FillVal); |
273 | }; |
274 | for_each(Range: at::getAssignmentMarkers(Inst: S), F: replaceOpForAssignmentMarkers); |
275 | for_each(Range: at::getDVRAssignmentMarkers(Inst: S), F: replaceOpForAssignmentMarkers); |
276 | |
277 | S->setAlignment(Alignment); |
278 | if (MI->isAtomic()) |
279 | S->setOrdering(AtomicOrdering::Unordered); |
280 | |
281 | // Set the size of the copy to 0, it will be deleted on the next iteration. |
282 | MI->setLength(Constant::getNullValue(Ty: LenC->getType())); |
283 | return MI; |
284 | } |
285 | |
286 | return nullptr; |
287 | } |
288 | |
289 | // TODO, Obvious Missing Transforms: |
290 | // * Narrow width by halfs excluding zero/undef lanes |
291 | Value *InstCombinerImpl::simplifyMaskedLoad(IntrinsicInst &II) { |
292 | Value *LoadPtr = II.getArgOperand(i: 0); |
293 | const Align Alignment = |
294 | cast<ConstantInt>(Val: II.getArgOperand(i: 1))->getAlignValue(); |
295 | |
296 | // If the mask is all ones or undefs, this is a plain vector load of the 1st |
297 | // argument. |
298 | if (maskIsAllOneOrUndef(Mask: II.getArgOperand(i: 2))) { |
299 | LoadInst *L = Builder.CreateAlignedLoad(Ty: II.getType(), Ptr: LoadPtr, Align: Alignment, |
300 | Name: "unmaskedload" ); |
301 | L->copyMetadata(SrcInst: II); |
302 | return L; |
303 | } |
304 | |
305 | // If we can unconditionally load from this address, replace with a |
306 | // load/select idiom. TODO: use DT for context sensitive query |
307 | if (isDereferenceablePointer(V: LoadPtr, Ty: II.getType(), |
308 | DL: II.getDataLayout(), CtxI: &II, AC: &AC)) { |
309 | LoadInst *LI = Builder.CreateAlignedLoad(Ty: II.getType(), Ptr: LoadPtr, Align: Alignment, |
310 | Name: "unmaskedload" ); |
311 | LI->copyMetadata(SrcInst: II); |
312 | return Builder.CreateSelect(C: II.getArgOperand(i: 2), True: LI, False: II.getArgOperand(i: 3)); |
313 | } |
314 | |
315 | return nullptr; |
316 | } |
317 | |
318 | // TODO, Obvious Missing Transforms: |
319 | // * Single constant active lane -> store |
320 | // * Narrow width by halfs excluding zero/undef lanes |
321 | Instruction *InstCombinerImpl::simplifyMaskedStore(IntrinsicInst &II) { |
322 | auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 3)); |
323 | if (!ConstMask) |
324 | return nullptr; |
325 | |
326 | // If the mask is all zeros, this instruction does nothing. |
327 | if (ConstMask->isNullValue()) |
328 | return eraseInstFromFunction(I&: II); |
329 | |
330 | // If the mask is all ones, this is a plain vector store of the 1st argument. |
331 | if (ConstMask->isAllOnesValue()) { |
332 | Value *StorePtr = II.getArgOperand(i: 1); |
333 | Align Alignment = cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue(); |
334 | StoreInst *S = |
335 | new StoreInst(II.getArgOperand(i: 0), StorePtr, false, Alignment); |
336 | S->copyMetadata(SrcInst: II); |
337 | return S; |
338 | } |
339 | |
340 | if (isa<ScalableVectorType>(Val: ConstMask->getType())) |
341 | return nullptr; |
342 | |
343 | // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts |
344 | APInt DemandedElts = possiblyDemandedEltsInMask(Mask: ConstMask); |
345 | APInt PoisonElts(DemandedElts.getBitWidth(), 0); |
346 | if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 0), DemandedElts, |
347 | PoisonElts)) |
348 | return replaceOperand(I&: II, OpNum: 0, V); |
349 | |
350 | return nullptr; |
351 | } |
352 | |
353 | // TODO, Obvious Missing Transforms: |
354 | // * Single constant active lane load -> load |
355 | // * Dereferenceable address & few lanes -> scalarize speculative load/selects |
356 | // * Adjacent vector addresses -> masked.load |
357 | // * Narrow width by halfs excluding zero/undef lanes |
358 | // * Vector incrementing address -> vector masked load |
359 | Instruction *InstCombinerImpl::simplifyMaskedGather(IntrinsicInst &II) { |
360 | auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 2)); |
361 | if (!ConstMask) |
362 | return nullptr; |
363 | |
364 | // Vector splat address w/known mask -> scalar load |
365 | // Fold the gather to load the source vector first lane |
366 | // because it is reloading the same value each time |
367 | if (ConstMask->isAllOnesValue()) |
368 | if (auto *SplatPtr = getSplatValue(V: II.getArgOperand(i: 0))) { |
369 | auto *VecTy = cast<VectorType>(Val: II.getType()); |
370 | const Align Alignment = |
371 | cast<ConstantInt>(Val: II.getArgOperand(i: 1))->getAlignValue(); |
372 | LoadInst *L = Builder.CreateAlignedLoad(Ty: VecTy->getElementType(), Ptr: SplatPtr, |
373 | Align: Alignment, Name: "load.scalar" ); |
374 | Value *Shuf = |
375 | Builder.CreateVectorSplat(EC: VecTy->getElementCount(), V: L, Name: "broadcast" ); |
376 | return replaceInstUsesWith(I&: II, V: cast<Instruction>(Val: Shuf)); |
377 | } |
378 | |
379 | return nullptr; |
380 | } |
381 | |
382 | // TODO, Obvious Missing Transforms: |
383 | // * Single constant active lane -> store |
384 | // * Adjacent vector addresses -> masked.store |
385 | // * Narrow store width by halfs excluding zero/undef lanes |
386 | // * Vector incrementing address -> vector masked store |
387 | Instruction *InstCombinerImpl::simplifyMaskedScatter(IntrinsicInst &II) { |
388 | auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 3)); |
389 | if (!ConstMask) |
390 | return nullptr; |
391 | |
392 | // If the mask is all zeros, a scatter does nothing. |
393 | if (ConstMask->isNullValue()) |
394 | return eraseInstFromFunction(I&: II); |
395 | |
396 | // Vector splat address -> scalar store |
397 | if (auto *SplatPtr = getSplatValue(V: II.getArgOperand(i: 1))) { |
398 | // scatter(splat(value), splat(ptr), non-zero-mask) -> store value, ptr |
399 | if (auto *SplatValue = getSplatValue(V: II.getArgOperand(i: 0))) { |
400 | if (maskContainsAllOneOrUndef(Mask: ConstMask)) { |
401 | Align Alignment = |
402 | cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue(); |
403 | StoreInst *S = new StoreInst(SplatValue, SplatPtr, /*IsVolatile=*/false, |
404 | Alignment); |
405 | S->copyMetadata(SrcInst: II); |
406 | return S; |
407 | } |
408 | } |
409 | // scatter(vector, splat(ptr), splat(true)) -> store extract(vector, |
410 | // lastlane), ptr |
411 | if (ConstMask->isAllOnesValue()) { |
412 | Align Alignment = cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue(); |
413 | VectorType *WideLoadTy = cast<VectorType>(Val: II.getArgOperand(i: 1)->getType()); |
414 | ElementCount VF = WideLoadTy->getElementCount(); |
415 | Value *RunTimeVF = Builder.CreateElementCount(Ty: Builder.getInt32Ty(), EC: VF); |
416 | Value *LastLane = Builder.CreateSub(LHS: RunTimeVF, RHS: Builder.getInt32(C: 1)); |
417 | Value * = |
418 | Builder.CreateExtractElement(Vec: II.getArgOperand(i: 0), Idx: LastLane); |
419 | StoreInst *S = |
420 | new StoreInst(Extract, SplatPtr, /*IsVolatile=*/false, Alignment); |
421 | S->copyMetadata(SrcInst: II); |
422 | return S; |
423 | } |
424 | } |
425 | if (isa<ScalableVectorType>(Val: ConstMask->getType())) |
426 | return nullptr; |
427 | |
428 | // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts |
429 | APInt DemandedElts = possiblyDemandedEltsInMask(Mask: ConstMask); |
430 | APInt PoisonElts(DemandedElts.getBitWidth(), 0); |
431 | if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 0), DemandedElts, |
432 | PoisonElts)) |
433 | return replaceOperand(I&: II, OpNum: 0, V); |
434 | if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 1), DemandedElts, |
435 | PoisonElts)) |
436 | return replaceOperand(I&: II, OpNum: 1, V); |
437 | |
438 | return nullptr; |
439 | } |
440 | |
441 | /// This function transforms launder.invariant.group and strip.invariant.group |
442 | /// like: |
443 | /// launder(launder(%x)) -> launder(%x) (the result is not the argument) |
444 | /// launder(strip(%x)) -> launder(%x) |
445 | /// strip(strip(%x)) -> strip(%x) (the result is not the argument) |
446 | /// strip(launder(%x)) -> strip(%x) |
447 | /// This is legal because it preserves the most recent information about |
448 | /// the presence or absence of invariant.group. |
449 | static Instruction *simplifyInvariantGroupIntrinsic(IntrinsicInst &II, |
450 | InstCombinerImpl &IC) { |
451 | auto *Arg = II.getArgOperand(i: 0); |
452 | auto *StrippedArg = Arg->stripPointerCasts(); |
453 | auto *StrippedInvariantGroupsArg = StrippedArg; |
454 | while (auto *Intr = dyn_cast<IntrinsicInst>(Val: StrippedInvariantGroupsArg)) { |
455 | if (Intr->getIntrinsicID() != Intrinsic::launder_invariant_group && |
456 | Intr->getIntrinsicID() != Intrinsic::strip_invariant_group) |
457 | break; |
458 | StrippedInvariantGroupsArg = Intr->getArgOperand(i: 0)->stripPointerCasts(); |
459 | } |
460 | if (StrippedArg == StrippedInvariantGroupsArg) |
461 | return nullptr; // No launders/strips to remove. |
462 | |
463 | Value *Result = nullptr; |
464 | |
465 | if (II.getIntrinsicID() == Intrinsic::launder_invariant_group) |
466 | Result = IC.Builder.CreateLaunderInvariantGroup(Ptr: StrippedInvariantGroupsArg); |
467 | else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group) |
468 | Result = IC.Builder.CreateStripInvariantGroup(Ptr: StrippedInvariantGroupsArg); |
469 | else |
470 | llvm_unreachable( |
471 | "simplifyInvariantGroupIntrinsic only handles launder and strip" ); |
472 | if (Result->getType()->getPointerAddressSpace() != |
473 | II.getType()->getPointerAddressSpace()) |
474 | Result = IC.Builder.CreateAddrSpaceCast(V: Result, DestTy: II.getType()); |
475 | |
476 | return cast<Instruction>(Val: Result); |
477 | } |
478 | |
479 | static Instruction *foldCttzCtlz(IntrinsicInst &II, InstCombinerImpl &IC) { |
480 | assert((II.getIntrinsicID() == Intrinsic::cttz || |
481 | II.getIntrinsicID() == Intrinsic::ctlz) && |
482 | "Expected cttz or ctlz intrinsic" ); |
483 | bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz; |
484 | Value *Op0 = II.getArgOperand(i: 0); |
485 | Value *Op1 = II.getArgOperand(i: 1); |
486 | Value *X; |
487 | // ctlz(bitreverse(x)) -> cttz(x) |
488 | // cttz(bitreverse(x)) -> ctlz(x) |
489 | if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X)))) { |
490 | Intrinsic::ID ID = IsTZ ? Intrinsic::ctlz : Intrinsic::cttz; |
491 | Function *F = |
492 | Intrinsic::getOrInsertDeclaration(M: II.getModule(), id: ID, Tys: II.getType()); |
493 | return CallInst::Create(Func: F, Args: {X, II.getArgOperand(i: 1)}); |
494 | } |
495 | |
496 | if (II.getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
497 | // ctlz/cttz i1 Op0 --> not Op0 |
498 | if (match(V: Op1, P: m_Zero())) |
499 | return BinaryOperator::CreateNot(Op: Op0); |
500 | // If zero is poison, then the input can be assumed to be "true", so the |
501 | // instruction simplifies to "false". |
502 | assert(match(Op1, m_One()) && "Expected ctlz/cttz operand to be 0 or 1" ); |
503 | return IC.replaceInstUsesWith(I&: II, V: ConstantInt::getNullValue(Ty: II.getType())); |
504 | } |
505 | |
506 | // If ctlz/cttz is only used as a shift amount, set is_zero_poison to true. |
507 | if (II.hasOneUse() && match(V: Op1, P: m_Zero()) && |
508 | match(V: II.user_back(), P: m_Shift(L: m_Value(), R: m_Specific(V: &II)))) { |
509 | II.dropUBImplyingAttrsAndMetadata(); |
510 | return IC.replaceOperand(I&: II, OpNum: 1, V: IC.Builder.getTrue()); |
511 | } |
512 | |
513 | Constant *C; |
514 | |
515 | if (IsTZ) { |
516 | // cttz(-x) -> cttz(x) |
517 | if (match(V: Op0, P: m_Neg(V: m_Value(V&: X)))) |
518 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
519 | |
520 | // cttz(-x & x) -> cttz(x) |
521 | if (match(V: Op0, P: m_c_And(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) |
522 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
523 | |
524 | // cttz(sext(x)) -> cttz(zext(x)) |
525 | if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))))) { |
526 | auto *Zext = IC.Builder.CreateZExt(V: X, DestTy: II.getType()); |
527 | auto *CttzZext = |
528 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: Zext, RHS: Op1); |
529 | return IC.replaceInstUsesWith(I&: II, V: CttzZext); |
530 | } |
531 | |
532 | // Zext doesn't change the number of trailing zeros, so narrow: |
533 | // cttz(zext(x)) -> zext(cttz(x)) if the 'ZeroIsPoison' parameter is 'true'. |
534 | if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && match(V: Op1, P: m_One())) { |
535 | auto *Cttz = IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: X, |
536 | RHS: IC.Builder.getTrue()); |
537 | auto *ZextCttz = IC.Builder.CreateZExt(V: Cttz, DestTy: II.getType()); |
538 | return IC.replaceInstUsesWith(I&: II, V: ZextCttz); |
539 | } |
540 | |
541 | // cttz(abs(x)) -> cttz(x) |
542 | // cttz(nabs(x)) -> cttz(x) |
543 | Value *Y; |
544 | SelectPatternFlavor SPF = matchSelectPattern(V: Op0, LHS&: X, RHS&: Y).Flavor; |
545 | if (SPF == SPF_ABS || SPF == SPF_NABS) |
546 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
547 | |
548 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X)))) |
549 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
550 | |
551 | // cttz(shl(%const, %val), 1) --> add(cttz(%const, 1), %val) |
552 | if (match(V: Op0, P: m_Shl(L: m_ImmConstant(C), R: m_Value(V&: X))) && |
553 | match(V: Op1, P: m_One())) { |
554 | Value *ConstCttz = |
555 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: C, RHS: Op1); |
556 | return BinaryOperator::CreateAdd(V1: ConstCttz, V2: X); |
557 | } |
558 | |
559 | // cttz(lshr exact (%const, %val), 1) --> sub(cttz(%const, 1), %val) |
560 | if (match(V: Op0, P: m_Exact(SubPattern: m_LShr(L: m_ImmConstant(C), R: m_Value(V&: X)))) && |
561 | match(V: Op1, P: m_One())) { |
562 | Value *ConstCttz = |
563 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: C, RHS: Op1); |
564 | return BinaryOperator::CreateSub(V1: ConstCttz, V2: X); |
565 | } |
566 | |
567 | // cttz(add(lshr(UINT_MAX, %val), 1)) --> sub(width, %val) |
568 | if (match(V: Op0, P: m_Add(L: m_LShr(L: m_AllOnes(), R: m_Value(V&: X)), R: m_One()))) { |
569 | Value *Width = |
570 | ConstantInt::get(Ty: II.getType(), V: II.getType()->getScalarSizeInBits()); |
571 | return BinaryOperator::CreateSub(V1: Width, V2: X); |
572 | } |
573 | } else { |
574 | // ctlz(lshr(%const, %val), 1) --> add(ctlz(%const, 1), %val) |
575 | if (match(V: Op0, P: m_LShr(L: m_ImmConstant(C), R: m_Value(V&: X))) && |
576 | match(V: Op1, P: m_One())) { |
577 | Value *ConstCtlz = |
578 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::ctlz, LHS: C, RHS: Op1); |
579 | return BinaryOperator::CreateAdd(V1: ConstCtlz, V2: X); |
580 | } |
581 | |
582 | // ctlz(shl nuw (%const, %val), 1) --> sub(ctlz(%const, 1), %val) |
583 | if (match(V: Op0, P: m_NUWShl(L: m_ImmConstant(C), R: m_Value(V&: X))) && |
584 | match(V: Op1, P: m_One())) { |
585 | Value *ConstCtlz = |
586 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::ctlz, LHS: C, RHS: Op1); |
587 | return BinaryOperator::CreateSub(V1: ConstCtlz, V2: X); |
588 | } |
589 | } |
590 | |
591 | // cttz(Pow2) -> Log2(Pow2) |
592 | // ctlz(Pow2) -> BitWidth - 1 - Log2(Pow2) |
593 | if (auto *R = IC.tryGetLog2(Op: Op0, AssumeNonZero: match(V: Op1, P: m_One()))) { |
594 | if (IsTZ) |
595 | return IC.replaceInstUsesWith(I&: II, V: R); |
596 | BinaryOperator *BO = BinaryOperator::CreateSub( |
597 | V1: ConstantInt::get(Ty: R->getType(), V: R->getType()->getScalarSizeInBits() - 1), |
598 | V2: R); |
599 | BO->setHasNoSignedWrap(); |
600 | BO->setHasNoUnsignedWrap(); |
601 | return BO; |
602 | } |
603 | |
604 | KnownBits Known = IC.computeKnownBits(V: Op0, CxtI: &II); |
605 | |
606 | // Create a mask for bits above (ctlz) or below (cttz) the first known one. |
607 | unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros() |
608 | : Known.countMaxLeadingZeros(); |
609 | unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros() |
610 | : Known.countMinLeadingZeros(); |
611 | |
612 | // If all bits above (ctlz) or below (cttz) the first known one are known |
613 | // zero, this value is constant. |
614 | // FIXME: This should be in InstSimplify because we're replacing an |
615 | // instruction with a constant. |
616 | if (PossibleZeros == DefiniteZeros) { |
617 | auto *C = ConstantInt::get(Ty: Op0->getType(), V: DefiniteZeros); |
618 | return IC.replaceInstUsesWith(I&: II, V: C); |
619 | } |
620 | |
621 | // If the input to cttz/ctlz is known to be non-zero, |
622 | // then change the 'ZeroIsPoison' parameter to 'true' |
623 | // because we know the zero behavior can't affect the result. |
624 | if (!Known.One.isZero() || |
625 | isKnownNonZero(V: Op0, Q: IC.getSimplifyQuery().getWithInstruction(I: &II))) { |
626 | if (!match(V: II.getArgOperand(i: 1), P: m_One())) |
627 | return IC.replaceOperand(I&: II, OpNum: 1, V: IC.Builder.getTrue()); |
628 | } |
629 | |
630 | // Add range attribute since known bits can't completely reflect what we know. |
631 | unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); |
632 | if (BitWidth != 1 && !II.hasRetAttr(Kind: Attribute::Range) && |
633 | !II.getMetadata(KindID: LLVMContext::MD_range)) { |
634 | ConstantRange Range(APInt(BitWidth, DefiniteZeros), |
635 | APInt(BitWidth, PossibleZeros + 1)); |
636 | II.addRangeRetAttr(CR: Range); |
637 | return &II; |
638 | } |
639 | |
640 | return nullptr; |
641 | } |
642 | |
643 | static Instruction *foldCtpop(IntrinsicInst &II, InstCombinerImpl &IC) { |
644 | assert(II.getIntrinsicID() == Intrinsic::ctpop && |
645 | "Expected ctpop intrinsic" ); |
646 | Type *Ty = II.getType(); |
647 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
648 | Value *Op0 = II.getArgOperand(i: 0); |
649 | Value *X, *Y; |
650 | |
651 | // ctpop(bitreverse(x)) -> ctpop(x) |
652 | // ctpop(bswap(x)) -> ctpop(x) |
653 | if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X))) || match(V: Op0, P: m_BSwap(Op0: m_Value(V&: X)))) |
654 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
655 | |
656 | // ctpop(rot(x)) -> ctpop(x) |
657 | if ((match(V: Op0, P: m_FShl(Op0: m_Value(V&: X), Op1: m_Value(V&: Y), Op2: m_Value())) || |
658 | match(V: Op0, P: m_FShr(Op0: m_Value(V&: X), Op1: m_Value(V&: Y), Op2: m_Value()))) && |
659 | X == Y) |
660 | return IC.replaceOperand(I&: II, OpNum: 0, V: X); |
661 | |
662 | // ctpop(x | -x) -> bitwidth - cttz(x, false) |
663 | if (Op0->hasOneUse() && |
664 | match(V: Op0, P: m_c_Or(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))))) { |
665 | auto *Cttz = IC.Builder.CreateIntrinsic(ID: Intrinsic::cttz, Types: Ty, |
666 | Args: {X, IC.Builder.getFalse()}); |
667 | auto *Bw = ConstantInt::get(Ty, V: APInt(BitWidth, BitWidth)); |
668 | return IC.replaceInstUsesWith(I&: II, V: IC.Builder.CreateSub(LHS: Bw, RHS: Cttz)); |
669 | } |
670 | |
671 | // ctpop(~x & (x - 1)) -> cttz(x, false) |
672 | if (match(V: Op0, |
673 | P: m_c_And(L: m_Not(V: m_Value(V&: X)), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes())))) { |
674 | Function *F = |
675 | Intrinsic::getOrInsertDeclaration(M: II.getModule(), id: Intrinsic::cttz, Tys: Ty); |
676 | return CallInst::Create(Func: F, Args: {X, IC.Builder.getFalse()}); |
677 | } |
678 | |
679 | // Zext doesn't change the number of set bits, so narrow: |
680 | // ctpop (zext X) --> zext (ctpop X) |
681 | if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))) { |
682 | Value *NarrowPop = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V: X); |
683 | return CastInst::Create(Instruction::ZExt, S: NarrowPop, Ty); |
684 | } |
685 | |
686 | KnownBits Known(BitWidth); |
687 | IC.computeKnownBits(V: Op0, Known, CxtI: &II); |
688 | |
689 | // If all bits are zero except for exactly one fixed bit, then the result |
690 | // must be 0 or 1, and we can get that answer by shifting to LSB: |
691 | // ctpop (X & 32) --> (X & 32) >> 5 |
692 | // TODO: Investigate removing this as its likely unnecessary given the below |
693 | // `isKnownToBeAPowerOfTwo` check. |
694 | if ((~Known.Zero).isPowerOf2()) |
695 | return BinaryOperator::CreateLShr( |
696 | V1: Op0, V2: ConstantInt::get(Ty, V: (~Known.Zero).exactLogBase2())); |
697 | |
698 | // More generally we can also handle non-constant power of 2 patterns such as |
699 | // shl/shr(Pow2, X), (X & -X), etc... by transforming: |
700 | // ctpop(Pow2OrZero) --> icmp ne X, 0 |
701 | if (IC.isKnownToBeAPowerOfTwo(V: Op0, /* OrZero */ true)) |
702 | return CastInst::Create(Instruction::ZExt, |
703 | S: IC.Builder.CreateICmp(P: ICmpInst::ICMP_NE, LHS: Op0, |
704 | RHS: Constant::getNullValue(Ty)), |
705 | Ty); |
706 | |
707 | // Add range attribute since known bits can't completely reflect what we know. |
708 | if (BitWidth != 1) { |
709 | ConstantRange OldRange = |
710 | II.getRange().value_or(u: ConstantRange::getFull(BitWidth)); |
711 | |
712 | unsigned Lower = Known.countMinPopulation(); |
713 | unsigned Upper = Known.countMaxPopulation() + 1; |
714 | |
715 | if (Lower == 0 && OldRange.contains(Val: APInt::getZero(numBits: BitWidth)) && |
716 | isKnownNonZero(V: Op0, Q: IC.getSimplifyQuery().getWithInstruction(I: &II))) |
717 | Lower = 1; |
718 | |
719 | ConstantRange Range(APInt(BitWidth, Lower), APInt(BitWidth, Upper)); |
720 | Range = Range.intersectWith(CR: OldRange, Type: ConstantRange::Unsigned); |
721 | |
722 | if (Range != OldRange) { |
723 | II.addRangeRetAttr(CR: Range); |
724 | return &II; |
725 | } |
726 | } |
727 | |
728 | return nullptr; |
729 | } |
730 | |
731 | /// Convert a table lookup to shufflevector if the mask is constant. |
732 | /// This could benefit tbl1 if the mask is { 7,6,5,4,3,2,1,0 }, in |
733 | /// which case we could lower the shufflevector with rev64 instructions |
734 | /// as it's actually a byte reverse. |
735 | static Value *simplifyNeonTbl1(const IntrinsicInst &II, |
736 | InstCombiner::BuilderTy &Builder) { |
737 | // Bail out if the mask is not a constant. |
738 | auto *C = dyn_cast<Constant>(Val: II.getArgOperand(i: 1)); |
739 | if (!C) |
740 | return nullptr; |
741 | |
742 | auto *VecTy = cast<FixedVectorType>(Val: II.getType()); |
743 | unsigned NumElts = VecTy->getNumElements(); |
744 | |
745 | // Only perform this transformation for <8 x i8> vector types. |
746 | if (!VecTy->getElementType()->isIntegerTy(Bitwidth: 8) || NumElts != 8) |
747 | return nullptr; |
748 | |
749 | int Indexes[8]; |
750 | |
751 | for (unsigned I = 0; I < NumElts; ++I) { |
752 | Constant *COp = C->getAggregateElement(Elt: I); |
753 | |
754 | if (!COp || !isa<ConstantInt>(Val: COp)) |
755 | return nullptr; |
756 | |
757 | Indexes[I] = cast<ConstantInt>(Val: COp)->getLimitedValue(); |
758 | |
759 | // Make sure the mask indices are in range. |
760 | if ((unsigned)Indexes[I] >= NumElts) |
761 | return nullptr; |
762 | } |
763 | |
764 | auto *V1 = II.getArgOperand(i: 0); |
765 | auto *V2 = Constant::getNullValue(Ty: V1->getType()); |
766 | return Builder.CreateShuffleVector(V1, V2, Mask: ArrayRef(Indexes)); |
767 | } |
768 | |
769 | // Returns true iff the 2 intrinsics have the same operands, limiting the |
770 | // comparison to the first NumOperands. |
771 | static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E, |
772 | unsigned NumOperands) { |
773 | assert(I.arg_size() >= NumOperands && "Not enough operands" ); |
774 | assert(E.arg_size() >= NumOperands && "Not enough operands" ); |
775 | for (unsigned i = 0; i < NumOperands; i++) |
776 | if (I.getArgOperand(i) != E.getArgOperand(i)) |
777 | return false; |
778 | return true; |
779 | } |
780 | |
781 | // Remove trivially empty start/end intrinsic ranges, i.e. a start |
782 | // immediately followed by an end (ignoring debuginfo or other |
783 | // start/end intrinsics in between). As this handles only the most trivial |
784 | // cases, tracking the nesting level is not needed: |
785 | // |
786 | // call @llvm.foo.start(i1 0) |
787 | // call @llvm.foo.start(i1 0) ; This one won't be skipped: it will be removed |
788 | // call @llvm.foo.end(i1 0) |
789 | // call @llvm.foo.end(i1 0) ; &I |
790 | static bool |
791 | removeTriviallyEmptyRange(IntrinsicInst &EndI, InstCombinerImpl &IC, |
792 | std::function<bool(const IntrinsicInst &)> IsStart) { |
793 | // We start from the end intrinsic and scan backwards, so that InstCombine |
794 | // has already processed (and potentially removed) all the instructions |
795 | // before the end intrinsic. |
796 | BasicBlock::reverse_iterator BI(EndI), BE(EndI.getParent()->rend()); |
797 | for (; BI != BE; ++BI) { |
798 | if (auto *I = dyn_cast<IntrinsicInst>(Val: &*BI)) { |
799 | if (I->isDebugOrPseudoInst() || |
800 | I->getIntrinsicID() == EndI.getIntrinsicID()) |
801 | continue; |
802 | if (IsStart(*I)) { |
803 | if (haveSameOperands(I: EndI, E: *I, NumOperands: EndI.arg_size())) { |
804 | IC.eraseInstFromFunction(I&: *I); |
805 | IC.eraseInstFromFunction(I&: EndI); |
806 | return true; |
807 | } |
808 | // Skip start intrinsics that don't pair with this end intrinsic. |
809 | continue; |
810 | } |
811 | } |
812 | break; |
813 | } |
814 | |
815 | return false; |
816 | } |
817 | |
818 | Instruction *InstCombinerImpl::visitVAEndInst(VAEndInst &I) { |
819 | removeTriviallyEmptyRange(EndI&: I, IC&: *this, IsStart: [&I](const IntrinsicInst &II) { |
820 | // Bail out on the case where the source va_list of a va_copy is destroyed |
821 | // immediately by a follow-up va_end. |
822 | return II.getIntrinsicID() == Intrinsic::vastart || |
823 | (II.getIntrinsicID() == Intrinsic::vacopy && |
824 | I.getArgOperand(i: 0) != II.getArgOperand(i: 1)); |
825 | }); |
826 | return nullptr; |
827 | } |
828 | |
829 | static CallInst *canonicalizeConstantArg0ToArg1(CallInst &Call) { |
830 | assert(Call.arg_size() > 1 && "Need at least 2 args to swap" ); |
831 | Value *Arg0 = Call.getArgOperand(i: 0), *Arg1 = Call.getArgOperand(i: 1); |
832 | if (isa<Constant>(Val: Arg0) && !isa<Constant>(Val: Arg1)) { |
833 | Call.setArgOperand(i: 0, v: Arg1); |
834 | Call.setArgOperand(i: 1, v: Arg0); |
835 | return &Call; |
836 | } |
837 | return nullptr; |
838 | } |
839 | |
840 | /// Creates a result tuple for an overflow intrinsic \p II with a given |
841 | /// \p Result and a constant \p Overflow value. |
842 | static Instruction *createOverflowTuple(IntrinsicInst *II, Value *Result, |
843 | Constant *Overflow) { |
844 | Constant *V[] = {PoisonValue::get(T: Result->getType()), Overflow}; |
845 | StructType *ST = cast<StructType>(Val: II->getType()); |
846 | Constant *Struct = ConstantStruct::get(T: ST, V); |
847 | return InsertValueInst::Create(Agg: Struct, Val: Result, Idxs: 0); |
848 | } |
849 | |
850 | Instruction * |
851 | InstCombinerImpl::foldIntrinsicWithOverflowCommon(IntrinsicInst *II) { |
852 | WithOverflowInst *WO = cast<WithOverflowInst>(Val: II); |
853 | Value *OperationResult = nullptr; |
854 | Constant *OverflowResult = nullptr; |
855 | if (OptimizeOverflowCheck(BinaryOp: WO->getBinaryOp(), IsSigned: WO->isSigned(), LHS: WO->getLHS(), |
856 | RHS: WO->getRHS(), CtxI&: *WO, OperationResult, OverflowResult)) |
857 | return createOverflowTuple(II: WO, Result: OperationResult, Overflow: OverflowResult); |
858 | |
859 | // See whether we can optimize the overflow check with assumption information. |
860 | for (User *U : WO->users()) { |
861 | if (!match(V: U, P: m_ExtractValue<1>(V: m_Value()))) |
862 | continue; |
863 | |
864 | for (auto &AssumeVH : AC.assumptionsFor(V: U)) { |
865 | if (!AssumeVH) |
866 | continue; |
867 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
868 | if (!match(V: I->getArgOperand(i: 0), P: m_Not(V: m_Specific(V: U)))) |
869 | continue; |
870 | if (!isValidAssumeForContext(I, CxtI: II, /*DT=*/nullptr, |
871 | /*AllowEphemerals=*/true)) |
872 | continue; |
873 | Value *Result = |
874 | Builder.CreateBinOp(Opc: WO->getBinaryOp(), LHS: WO->getLHS(), RHS: WO->getRHS()); |
875 | Result->takeName(V: WO); |
876 | if (auto *Inst = dyn_cast<Instruction>(Val: Result)) { |
877 | if (WO->isSigned()) |
878 | Inst->setHasNoSignedWrap(); |
879 | else |
880 | Inst->setHasNoUnsignedWrap(); |
881 | } |
882 | return createOverflowTuple(II: WO, Result, |
883 | Overflow: ConstantInt::getFalse(Ty: U->getType())); |
884 | } |
885 | } |
886 | |
887 | return nullptr; |
888 | } |
889 | |
890 | static bool inputDenormalIsIEEE(const Function &F, const Type *Ty) { |
891 | Ty = Ty->getScalarType(); |
892 | return F.getDenormalMode(FPType: Ty->getFltSemantics()).Input == DenormalMode::IEEE; |
893 | } |
894 | |
895 | static bool inputDenormalIsDAZ(const Function &F, const Type *Ty) { |
896 | Ty = Ty->getScalarType(); |
897 | return F.getDenormalMode(FPType: Ty->getFltSemantics()).inputsAreZero(); |
898 | } |
899 | |
900 | /// \returns the compare predicate type if the test performed by |
901 | /// llvm.is.fpclass(x, \p Mask) is equivalent to fcmp o__ x, 0.0 with the |
902 | /// floating-point environment assumed for \p F for type \p Ty |
903 | static FCmpInst::Predicate fpclassTestIsFCmp0(FPClassTest Mask, |
904 | const Function &F, Type *Ty) { |
905 | switch (static_cast<unsigned>(Mask)) { |
906 | case fcZero: |
907 | if (inputDenormalIsIEEE(F, Ty)) |
908 | return FCmpInst::FCMP_OEQ; |
909 | break; |
910 | case fcZero | fcSubnormal: |
911 | if (inputDenormalIsDAZ(F, Ty)) |
912 | return FCmpInst::FCMP_OEQ; |
913 | break; |
914 | case fcPositive | fcNegZero: |
915 | if (inputDenormalIsIEEE(F, Ty)) |
916 | return FCmpInst::FCMP_OGE; |
917 | break; |
918 | case fcPositive | fcNegZero | fcNegSubnormal: |
919 | if (inputDenormalIsDAZ(F, Ty)) |
920 | return FCmpInst::FCMP_OGE; |
921 | break; |
922 | case fcPosSubnormal | fcPosNormal | fcPosInf: |
923 | if (inputDenormalIsIEEE(F, Ty)) |
924 | return FCmpInst::FCMP_OGT; |
925 | break; |
926 | case fcNegative | fcPosZero: |
927 | if (inputDenormalIsIEEE(F, Ty)) |
928 | return FCmpInst::FCMP_OLE; |
929 | break; |
930 | case fcNegative | fcPosZero | fcPosSubnormal: |
931 | if (inputDenormalIsDAZ(F, Ty)) |
932 | return FCmpInst::FCMP_OLE; |
933 | break; |
934 | case fcNegSubnormal | fcNegNormal | fcNegInf: |
935 | if (inputDenormalIsIEEE(F, Ty)) |
936 | return FCmpInst::FCMP_OLT; |
937 | break; |
938 | case fcPosNormal | fcPosInf: |
939 | if (inputDenormalIsDAZ(F, Ty)) |
940 | return FCmpInst::FCMP_OGT; |
941 | break; |
942 | case fcNegNormal | fcNegInf: |
943 | if (inputDenormalIsDAZ(F, Ty)) |
944 | return FCmpInst::FCMP_OLT; |
945 | break; |
946 | case ~fcZero & ~fcNan: |
947 | if (inputDenormalIsIEEE(F, Ty)) |
948 | return FCmpInst::FCMP_ONE; |
949 | break; |
950 | case ~(fcZero | fcSubnormal) & ~fcNan: |
951 | if (inputDenormalIsDAZ(F, Ty)) |
952 | return FCmpInst::FCMP_ONE; |
953 | break; |
954 | default: |
955 | break; |
956 | } |
957 | |
958 | return FCmpInst::BAD_FCMP_PREDICATE; |
959 | } |
960 | |
961 | Instruction *InstCombinerImpl::foldIntrinsicIsFPClass(IntrinsicInst &II) { |
962 | Value *Src0 = II.getArgOperand(i: 0); |
963 | Value *Src1 = II.getArgOperand(i: 1); |
964 | const ConstantInt *CMask = cast<ConstantInt>(Val: Src1); |
965 | FPClassTest Mask = static_cast<FPClassTest>(CMask->getZExtValue()); |
966 | const bool IsUnordered = (Mask & fcNan) == fcNan; |
967 | const bool IsOrdered = (Mask & fcNan) == fcNone; |
968 | const FPClassTest OrderedMask = Mask & ~fcNan; |
969 | const FPClassTest OrderedInvertedMask = ~OrderedMask & ~fcNan; |
970 | |
971 | const bool IsStrict = |
972 | II.getFunction()->getAttributes().hasFnAttr(Kind: Attribute::StrictFP); |
973 | |
974 | Value *FNegSrc; |
975 | if (match(V: Src0, P: m_FNeg(X: m_Value(V&: FNegSrc)))) { |
976 | // is.fpclass (fneg x), mask -> is.fpclass x, (fneg mask) |
977 | |
978 | II.setArgOperand(i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: fneg(Mask))); |
979 | return replaceOperand(I&: II, OpNum: 0, V: FNegSrc); |
980 | } |
981 | |
982 | Value *FAbsSrc; |
983 | if (match(V: Src0, P: m_FAbs(Op0: m_Value(V&: FAbsSrc)))) { |
984 | II.setArgOperand(i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: inverse_fabs(Mask))); |
985 | return replaceOperand(I&: II, OpNum: 0, V: FAbsSrc); |
986 | } |
987 | |
988 | if ((OrderedMask == fcInf || OrderedInvertedMask == fcInf) && |
989 | (IsOrdered || IsUnordered) && !IsStrict) { |
990 | // is.fpclass(x, fcInf) -> fcmp oeq fabs(x), +inf |
991 | // is.fpclass(x, ~fcInf) -> fcmp one fabs(x), +inf |
992 | // is.fpclass(x, fcInf|fcNan) -> fcmp ueq fabs(x), +inf |
993 | // is.fpclass(x, ~(fcInf|fcNan)) -> fcmp une fabs(x), +inf |
994 | Constant *Inf = ConstantFP::getInfinity(Ty: Src0->getType()); |
995 | FCmpInst::Predicate Pred = |
996 | IsUnordered ? FCmpInst::FCMP_UEQ : FCmpInst::FCMP_OEQ; |
997 | if (OrderedInvertedMask == fcInf) |
998 | Pred = IsUnordered ? FCmpInst::FCMP_UNE : FCmpInst::FCMP_ONE; |
999 | |
1000 | Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Src0); |
1001 | Value *CmpInf = Builder.CreateFCmp(P: Pred, LHS: Fabs, RHS: Inf); |
1002 | CmpInf->takeName(V: &II); |
1003 | return replaceInstUsesWith(I&: II, V: CmpInf); |
1004 | } |
1005 | |
1006 | if ((OrderedMask == fcPosInf || OrderedMask == fcNegInf) && |
1007 | (IsOrdered || IsUnordered) && !IsStrict) { |
1008 | // is.fpclass(x, fcPosInf) -> fcmp oeq x, +inf |
1009 | // is.fpclass(x, fcNegInf) -> fcmp oeq x, -inf |
1010 | // is.fpclass(x, fcPosInf|fcNan) -> fcmp ueq x, +inf |
1011 | // is.fpclass(x, fcNegInf|fcNan) -> fcmp ueq x, -inf |
1012 | Constant *Inf = |
1013 | ConstantFP::getInfinity(Ty: Src0->getType(), Negative: OrderedMask == fcNegInf); |
1014 | Value *EqInf = IsUnordered ? Builder.CreateFCmpUEQ(LHS: Src0, RHS: Inf) |
1015 | : Builder.CreateFCmpOEQ(LHS: Src0, RHS: Inf); |
1016 | |
1017 | EqInf->takeName(V: &II); |
1018 | return replaceInstUsesWith(I&: II, V: EqInf); |
1019 | } |
1020 | |
1021 | if ((OrderedInvertedMask == fcPosInf || OrderedInvertedMask == fcNegInf) && |
1022 | (IsOrdered || IsUnordered) && !IsStrict) { |
1023 | // is.fpclass(x, ~fcPosInf) -> fcmp one x, +inf |
1024 | // is.fpclass(x, ~fcNegInf) -> fcmp one x, -inf |
1025 | // is.fpclass(x, ~fcPosInf|fcNan) -> fcmp une x, +inf |
1026 | // is.fpclass(x, ~fcNegInf|fcNan) -> fcmp une x, -inf |
1027 | Constant *Inf = ConstantFP::getInfinity(Ty: Src0->getType(), |
1028 | Negative: OrderedInvertedMask == fcNegInf); |
1029 | Value *NeInf = IsUnordered ? Builder.CreateFCmpUNE(LHS: Src0, RHS: Inf) |
1030 | : Builder.CreateFCmpONE(LHS: Src0, RHS: Inf); |
1031 | NeInf->takeName(V: &II); |
1032 | return replaceInstUsesWith(I&: II, V: NeInf); |
1033 | } |
1034 | |
1035 | if (Mask == fcNan && !IsStrict) { |
1036 | // Equivalent of isnan. Replace with standard fcmp if we don't care about FP |
1037 | // exceptions. |
1038 | Value *IsNan = |
1039 | Builder.CreateFCmpUNO(LHS: Src0, RHS: ConstantFP::getZero(Ty: Src0->getType())); |
1040 | IsNan->takeName(V: &II); |
1041 | return replaceInstUsesWith(I&: II, V: IsNan); |
1042 | } |
1043 | |
1044 | if (Mask == (~fcNan & fcAllFlags) && !IsStrict) { |
1045 | // Equivalent of !isnan. Replace with standard fcmp. |
1046 | Value *FCmp = |
1047 | Builder.CreateFCmpORD(LHS: Src0, RHS: ConstantFP::getZero(Ty: Src0->getType())); |
1048 | FCmp->takeName(V: &II); |
1049 | return replaceInstUsesWith(I&: II, V: FCmp); |
1050 | } |
1051 | |
1052 | FCmpInst::Predicate PredType = FCmpInst::BAD_FCMP_PREDICATE; |
1053 | |
1054 | // Try to replace with an fcmp with 0 |
1055 | // |
1056 | // is.fpclass(x, fcZero) -> fcmp oeq x, 0.0 |
1057 | // is.fpclass(x, fcZero | fcNan) -> fcmp ueq x, 0.0 |
1058 | // is.fpclass(x, ~fcZero & ~fcNan) -> fcmp one x, 0.0 |
1059 | // is.fpclass(x, ~fcZero) -> fcmp une x, 0.0 |
1060 | // |
1061 | // is.fpclass(x, fcPosSubnormal | fcPosNormal | fcPosInf) -> fcmp ogt x, 0.0 |
1062 | // is.fpclass(x, fcPositive | fcNegZero) -> fcmp oge x, 0.0 |
1063 | // |
1064 | // is.fpclass(x, fcNegSubnormal | fcNegNormal | fcNegInf) -> fcmp olt x, 0.0 |
1065 | // is.fpclass(x, fcNegative | fcPosZero) -> fcmp ole x, 0.0 |
1066 | // |
1067 | if (!IsStrict && (IsOrdered || IsUnordered) && |
1068 | (PredType = fpclassTestIsFCmp0(Mask: OrderedMask, F: *II.getFunction(), |
1069 | Ty: Src0->getType())) != |
1070 | FCmpInst::BAD_FCMP_PREDICATE) { |
1071 | Constant *Zero = ConstantFP::getZero(Ty: Src0->getType()); |
1072 | // Equivalent of == 0. |
1073 | Value *FCmp = Builder.CreateFCmp( |
1074 | P: IsUnordered ? FCmpInst::getUnorderedPredicate(Pred: PredType) : PredType, |
1075 | LHS: Src0, RHS: Zero); |
1076 | |
1077 | FCmp->takeName(V: &II); |
1078 | return replaceInstUsesWith(I&: II, V: FCmp); |
1079 | } |
1080 | |
1081 | KnownFPClass Known = computeKnownFPClass(Val: Src0, Interested: Mask, CtxI: &II); |
1082 | |
1083 | // Clear test bits we know must be false from the source value. |
1084 | // fp_class (nnan x), qnan|snan|other -> fp_class (nnan x), other |
1085 | // fp_class (ninf x), ninf|pinf|other -> fp_class (ninf x), other |
1086 | if ((Mask & Known.KnownFPClasses) != Mask) { |
1087 | II.setArgOperand( |
1088 | i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: Mask & Known.KnownFPClasses)); |
1089 | return &II; |
1090 | } |
1091 | |
1092 | // If none of the tests which can return false are possible, fold to true. |
1093 | // fp_class (nnan x), ~(qnan|snan) -> true |
1094 | // fp_class (ninf x), ~(ninf|pinf) -> true |
1095 | if (Mask == Known.KnownFPClasses) |
1096 | return replaceInstUsesWith(I&: II, V: ConstantInt::get(Ty: II.getType(), V: true)); |
1097 | |
1098 | return nullptr; |
1099 | } |
1100 | |
1101 | static std::optional<bool> getKnownSign(Value *Op, const SimplifyQuery &SQ) { |
1102 | KnownBits Known = computeKnownBits(V: Op, Q: SQ); |
1103 | if (Known.isNonNegative()) |
1104 | return false; |
1105 | if (Known.isNegative()) |
1106 | return true; |
1107 | |
1108 | Value *X, *Y; |
1109 | if (match(V: Op, P: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y)))) |
1110 | return isImpliedByDomCondition(Pred: ICmpInst::ICMP_SLT, LHS: X, RHS: Y, ContextI: SQ.CxtI, DL: SQ.DL); |
1111 | |
1112 | return std::nullopt; |
1113 | } |
1114 | |
1115 | static std::optional<bool> getKnownSignOrZero(Value *Op, |
1116 | const SimplifyQuery &SQ) { |
1117 | if (std::optional<bool> Sign = getKnownSign(Op, SQ)) |
1118 | return Sign; |
1119 | |
1120 | Value *X, *Y; |
1121 | if (match(V: Op, P: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y)))) |
1122 | return isImpliedByDomCondition(Pred: ICmpInst::ICMP_SLE, LHS: X, RHS: Y, ContextI: SQ.CxtI, DL: SQ.DL); |
1123 | |
1124 | return std::nullopt; |
1125 | } |
1126 | |
1127 | /// Return true if two values \p Op0 and \p Op1 are known to have the same sign. |
1128 | static bool signBitMustBeTheSame(Value *Op0, Value *Op1, |
1129 | const SimplifyQuery &SQ) { |
1130 | std::optional<bool> Known1 = getKnownSign(Op: Op1, SQ); |
1131 | if (!Known1) |
1132 | return false; |
1133 | std::optional<bool> Known0 = getKnownSign(Op: Op0, SQ); |
1134 | if (!Known0) |
1135 | return false; |
1136 | return *Known0 == *Known1; |
1137 | } |
1138 | |
1139 | /// Try to canonicalize min/max(X + C0, C1) as min/max(X, C1 - C0) + C0. This |
1140 | /// can trigger other combines. |
1141 | static Instruction *moveAddAfterMinMax(IntrinsicInst *II, |
1142 | InstCombiner::BuilderTy &Builder) { |
1143 | Intrinsic::ID MinMaxID = II->getIntrinsicID(); |
1144 | assert((MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin || |
1145 | MinMaxID == Intrinsic::umax || MinMaxID == Intrinsic::umin) && |
1146 | "Expected a min or max intrinsic" ); |
1147 | |
1148 | // TODO: Match vectors with undef elements, but undef may not propagate. |
1149 | Value *Op0 = II->getArgOperand(i: 0), *Op1 = II->getArgOperand(i: 1); |
1150 | Value *X; |
1151 | const APInt *C0, *C1; |
1152 | if (!match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C0)))) || |
1153 | !match(V: Op1, P: m_APInt(Res&: C1))) |
1154 | return nullptr; |
1155 | |
1156 | // Check for necessary no-wrap and overflow constraints. |
1157 | bool IsSigned = MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin; |
1158 | auto *Add = cast<BinaryOperator>(Val: Op0); |
1159 | if ((IsSigned && !Add->hasNoSignedWrap()) || |
1160 | (!IsSigned && !Add->hasNoUnsignedWrap())) |
1161 | return nullptr; |
1162 | |
1163 | // If the constant difference overflows, then instsimplify should reduce the |
1164 | // min/max to the add or C1. |
1165 | bool Overflow; |
1166 | APInt CDiff = |
1167 | IsSigned ? C1->ssub_ov(RHS: *C0, Overflow) : C1->usub_ov(RHS: *C0, Overflow); |
1168 | assert(!Overflow && "Expected simplify of min/max" ); |
1169 | |
1170 | // min/max (add X, C0), C1 --> add (min/max X, C1 - C0), C0 |
1171 | // Note: the "mismatched" no-overflow setting does not propagate. |
1172 | Constant *NewMinMaxC = ConstantInt::get(Ty: II->getType(), V: CDiff); |
1173 | Value *NewMinMax = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS: NewMinMaxC); |
1174 | return IsSigned ? BinaryOperator::CreateNSWAdd(V1: NewMinMax, V2: Add->getOperand(i_nocapture: 1)) |
1175 | : BinaryOperator::CreateNUWAdd(V1: NewMinMax, V2: Add->getOperand(i_nocapture: 1)); |
1176 | } |
1177 | /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value. |
1178 | Instruction *InstCombinerImpl::matchSAddSubSat(IntrinsicInst &MinMax1) { |
1179 | Type *Ty = MinMax1.getType(); |
1180 | |
1181 | // We are looking for a tree of: |
1182 | // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B)))) |
1183 | // Where the min and max could be reversed |
1184 | Instruction *MinMax2; |
1185 | BinaryOperator *AddSub; |
1186 | const APInt *MinValue, *MaxValue; |
1187 | if (match(V: &MinMax1, P: m_SMin(L: m_Instruction(I&: MinMax2), R: m_APInt(Res&: MaxValue)))) { |
1188 | if (!match(V: MinMax2, P: m_SMax(L: m_BinOp(I&: AddSub), R: m_APInt(Res&: MinValue)))) |
1189 | return nullptr; |
1190 | } else if (match(V: &MinMax1, |
1191 | P: m_SMax(L: m_Instruction(I&: MinMax2), R: m_APInt(Res&: MinValue)))) { |
1192 | if (!match(V: MinMax2, P: m_SMin(L: m_BinOp(I&: AddSub), R: m_APInt(Res&: MaxValue)))) |
1193 | return nullptr; |
1194 | } else |
1195 | return nullptr; |
1196 | |
1197 | // Check that the constants clamp a saturate, and that the new type would be |
1198 | // sensible to convert to. |
1199 | if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1) |
1200 | return nullptr; |
1201 | // In what bitwidth can this be treated as saturating arithmetics? |
1202 | unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1; |
1203 | // FIXME: This isn't quite right for vectors, but using the scalar type is a |
1204 | // good first approximation for what should be done there. |
1205 | if (!shouldChangeType(FromBitWidth: Ty->getScalarType()->getIntegerBitWidth(), ToBitWidth: NewBitWidth)) |
1206 | return nullptr; |
1207 | |
1208 | // Also make sure that the inner min/max and the add/sub have one use. |
1209 | if (!MinMax2->hasOneUse() || !AddSub->hasOneUse()) |
1210 | return nullptr; |
1211 | |
1212 | // Create the new type (which can be a vector type) |
1213 | Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth); |
1214 | |
1215 | Intrinsic::ID IntrinsicID; |
1216 | if (AddSub->getOpcode() == Instruction::Add) |
1217 | IntrinsicID = Intrinsic::sadd_sat; |
1218 | else if (AddSub->getOpcode() == Instruction::Sub) |
1219 | IntrinsicID = Intrinsic::ssub_sat; |
1220 | else |
1221 | return nullptr; |
1222 | |
1223 | // The two operands of the add/sub must be nsw-truncatable to the NewTy. This |
1224 | // is usually achieved via a sext from a smaller type. |
1225 | if (ComputeMaxSignificantBits(Op: AddSub->getOperand(i_nocapture: 0), CxtI: AddSub) > NewBitWidth || |
1226 | ComputeMaxSignificantBits(Op: AddSub->getOperand(i_nocapture: 1), CxtI: AddSub) > NewBitWidth) |
1227 | return nullptr; |
1228 | |
1229 | // Finally create and return the sat intrinsic, truncated to the new type |
1230 | Value *AT = Builder.CreateTrunc(V: AddSub->getOperand(i_nocapture: 0), DestTy: NewTy); |
1231 | Value *BT = Builder.CreateTrunc(V: AddSub->getOperand(i_nocapture: 1), DestTy: NewTy); |
1232 | Value *Sat = Builder.CreateIntrinsic(ID: IntrinsicID, Types: NewTy, Args: {AT, BT}); |
1233 | return CastInst::Create(Instruction::SExt, S: Sat, Ty); |
1234 | } |
1235 | |
1236 | |
1237 | /// If we have a clamp pattern like max (min X, 42), 41 -- where the output |
1238 | /// can only be one of two possible constant values -- turn that into a select |
1239 | /// of constants. |
1240 | static Instruction *foldClampRangeOfTwo(IntrinsicInst *II, |
1241 | InstCombiner::BuilderTy &Builder) { |
1242 | Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1); |
1243 | Value *X; |
1244 | const APInt *C0, *C1; |
1245 | if (!match(V: I1, P: m_APInt(Res&: C1)) || !I0->hasOneUse()) |
1246 | return nullptr; |
1247 | |
1248 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; |
1249 | switch (II->getIntrinsicID()) { |
1250 | case Intrinsic::smax: |
1251 | if (match(V: I0, P: m_SMin(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C0 == *C1 + 1) |
1252 | Pred = ICmpInst::ICMP_SGT; |
1253 | break; |
1254 | case Intrinsic::smin: |
1255 | if (match(V: I0, P: m_SMax(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C1 == *C0 + 1) |
1256 | Pred = ICmpInst::ICMP_SLT; |
1257 | break; |
1258 | case Intrinsic::umax: |
1259 | if (match(V: I0, P: m_UMin(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C0 == *C1 + 1) |
1260 | Pred = ICmpInst::ICMP_UGT; |
1261 | break; |
1262 | case Intrinsic::umin: |
1263 | if (match(V: I0, P: m_UMax(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C1 == *C0 + 1) |
1264 | Pred = ICmpInst::ICMP_ULT; |
1265 | break; |
1266 | default: |
1267 | llvm_unreachable("Expected min/max intrinsic" ); |
1268 | } |
1269 | if (Pred == CmpInst::BAD_ICMP_PREDICATE) |
1270 | return nullptr; |
1271 | |
1272 | // max (min X, 42), 41 --> X > 41 ? 42 : 41 |
1273 | // min (max X, 42), 43 --> X < 43 ? 42 : 43 |
1274 | Value *Cmp = Builder.CreateICmp(P: Pred, LHS: X, RHS: I1); |
1275 | return SelectInst::Create(C: Cmp, S1: ConstantInt::get(Ty: II->getType(), V: *C0), S2: I1); |
1276 | } |
1277 | |
1278 | /// If this min/max has a constant operand and an operand that is a matching |
1279 | /// min/max with a constant operand, constant-fold the 2 constant operands. |
1280 | static Value *reassociateMinMaxWithConstants(IntrinsicInst *II, |
1281 | IRBuilderBase &Builder, |
1282 | const SimplifyQuery &SQ) { |
1283 | Intrinsic::ID MinMaxID = II->getIntrinsicID(); |
1284 | auto *LHS = dyn_cast<MinMaxIntrinsic>(Val: II->getArgOperand(i: 0)); |
1285 | if (!LHS) |
1286 | return nullptr; |
1287 | |
1288 | Constant *C0, *C1; |
1289 | if (!match(V: LHS->getArgOperand(i: 1), P: m_ImmConstant(C&: C0)) || |
1290 | !match(V: II->getArgOperand(i: 1), P: m_ImmConstant(C&: C1))) |
1291 | return nullptr; |
1292 | |
1293 | // max (max X, C0), C1 --> max X, (max C0, C1) |
1294 | // min (min X, C0), C1 --> min X, (min C0, C1) |
1295 | // umax (smax X, nneg C0), nneg C1 --> smax X, (umax C0, C1) |
1296 | // smin (umin X, nneg C0), nneg C1 --> umin X, (smin C0, C1) |
1297 | Intrinsic::ID InnerMinMaxID = LHS->getIntrinsicID(); |
1298 | if (InnerMinMaxID != MinMaxID && |
1299 | !(((MinMaxID == Intrinsic::umax && InnerMinMaxID == Intrinsic::smax) || |
1300 | (MinMaxID == Intrinsic::smin && InnerMinMaxID == Intrinsic::umin)) && |
1301 | isKnownNonNegative(V: C0, SQ) && isKnownNonNegative(V: C1, SQ))) |
1302 | return nullptr; |
1303 | |
1304 | ICmpInst::Predicate Pred = MinMaxIntrinsic::getPredicate(ID: MinMaxID); |
1305 | Value *CondC = Builder.CreateICmp(P: Pred, LHS: C0, RHS: C1); |
1306 | Value *NewC = Builder.CreateSelect(C: CondC, True: C0, False: C1); |
1307 | return Builder.CreateIntrinsic(ID: InnerMinMaxID, Types: II->getType(), |
1308 | Args: {LHS->getArgOperand(i: 0), NewC}); |
1309 | } |
1310 | |
1311 | /// If this min/max has a matching min/max operand with a constant, try to push |
1312 | /// the constant operand into this instruction. This can enable more folds. |
1313 | static Instruction * |
1314 | reassociateMinMaxWithConstantInOperand(IntrinsicInst *II, |
1315 | InstCombiner::BuilderTy &Builder) { |
1316 | // Match and capture a min/max operand candidate. |
1317 | Value *X, *Y; |
1318 | Constant *C; |
1319 | Instruction *Inner; |
1320 | if (!match(V: II, P: m_c_MaxOrMin(L: m_OneUse(SubPattern: m_CombineAnd( |
1321 | L: m_Instruction(I&: Inner), |
1322 | R: m_MaxOrMin(L: m_Value(V&: X), R: m_ImmConstant(C)))), |
1323 | R: m_Value(V&: Y)))) |
1324 | return nullptr; |
1325 | |
1326 | // The inner op must match. Check for constants to avoid infinite loops. |
1327 | Intrinsic::ID MinMaxID = II->getIntrinsicID(); |
1328 | auto *InnerMM = dyn_cast<IntrinsicInst>(Val: Inner); |
1329 | if (!InnerMM || InnerMM->getIntrinsicID() != MinMaxID || |
1330 | match(V: X, P: m_ImmConstant()) || match(V: Y, P: m_ImmConstant())) |
1331 | return nullptr; |
1332 | |
1333 | // max (max X, C), Y --> max (max X, Y), C |
1334 | Function *MinMax = Intrinsic::getOrInsertDeclaration(M: II->getModule(), |
1335 | id: MinMaxID, Tys: II->getType()); |
1336 | Value *NewInner = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS: Y); |
1337 | NewInner->takeName(V: Inner); |
1338 | return CallInst::Create(Func: MinMax, Args: {NewInner, C}); |
1339 | } |
1340 | |
1341 | /// Reduce a sequence of min/max intrinsics with a common operand. |
1342 | static Instruction *factorizeMinMaxTree(IntrinsicInst *II) { |
1343 | // Match 3 of the same min/max ops. Example: umin(umin(), umin()). |
1344 | auto *LHS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0)); |
1345 | auto *RHS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 1)); |
1346 | Intrinsic::ID MinMaxID = II->getIntrinsicID(); |
1347 | if (!LHS || !RHS || LHS->getIntrinsicID() != MinMaxID || |
1348 | RHS->getIntrinsicID() != MinMaxID || |
1349 | (!LHS->hasOneUse() && !RHS->hasOneUse())) |
1350 | return nullptr; |
1351 | |
1352 | Value *A = LHS->getArgOperand(i: 0); |
1353 | Value *B = LHS->getArgOperand(i: 1); |
1354 | Value *C = RHS->getArgOperand(i: 0); |
1355 | Value *D = RHS->getArgOperand(i: 1); |
1356 | |
1357 | // Look for a common operand. |
1358 | Value *MinMaxOp = nullptr; |
1359 | Value *ThirdOp = nullptr; |
1360 | if (LHS->hasOneUse()) { |
1361 | // If the LHS is only used in this chain and the RHS is used outside of it, |
1362 | // reuse the RHS min/max because that will eliminate the LHS. |
1363 | if (D == A || C == A) { |
1364 | // min(min(a, b), min(c, a)) --> min(min(c, a), b) |
1365 | // min(min(a, b), min(a, d)) --> min(min(a, d), b) |
1366 | MinMaxOp = RHS; |
1367 | ThirdOp = B; |
1368 | } else if (D == B || C == B) { |
1369 | // min(min(a, b), min(c, b)) --> min(min(c, b), a) |
1370 | // min(min(a, b), min(b, d)) --> min(min(b, d), a) |
1371 | MinMaxOp = RHS; |
1372 | ThirdOp = A; |
1373 | } |
1374 | } else { |
1375 | assert(RHS->hasOneUse() && "Expected one-use operand" ); |
1376 | // Reuse the LHS. This will eliminate the RHS. |
1377 | if (D == A || D == B) { |
1378 | // min(min(a, b), min(c, a)) --> min(min(a, b), c) |
1379 | // min(min(a, b), min(c, b)) --> min(min(a, b), c) |
1380 | MinMaxOp = LHS; |
1381 | ThirdOp = C; |
1382 | } else if (C == A || C == B) { |
1383 | // min(min(a, b), min(b, d)) --> min(min(a, b), d) |
1384 | // min(min(a, b), min(c, b)) --> min(min(a, b), d) |
1385 | MinMaxOp = LHS; |
1386 | ThirdOp = D; |
1387 | } |
1388 | } |
1389 | |
1390 | if (!MinMaxOp || !ThirdOp) |
1391 | return nullptr; |
1392 | |
1393 | Module *Mod = II->getModule(); |
1394 | Function *MinMax = |
1395 | Intrinsic::getOrInsertDeclaration(M: Mod, id: MinMaxID, Tys: II->getType()); |
1396 | return CallInst::Create(Func: MinMax, Args: { MinMaxOp, ThirdOp }); |
1397 | } |
1398 | |
1399 | /// If all arguments of the intrinsic are unary shuffles with the same mask, |
1400 | /// try to shuffle after the intrinsic. |
1401 | Instruction * |
1402 | InstCombinerImpl::foldShuffledIntrinsicOperands(IntrinsicInst *II) { |
1403 | if (!isTriviallyVectorizable(ID: II->getIntrinsicID()) || |
1404 | !II->getCalledFunction()->isSpeculatable()) |
1405 | return nullptr; |
1406 | |
1407 | Value *X; |
1408 | Constant *C; |
1409 | ArrayRef<int> Mask; |
1410 | auto *NonConstArg = find_if_not(Range: II->args(), P: [&II](Use &Arg) { |
1411 | return isa<Constant>(Val: Arg.get()) || |
1412 | isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), |
1413 | ScalarOpdIdx: Arg.getOperandNo(), TTI: nullptr); |
1414 | }); |
1415 | if (!NonConstArg || |
1416 | !match(V: NonConstArg, P: m_Shuffle(v1: m_Value(V&: X), v2: m_Poison(), mask: m_Mask(Mask)))) |
1417 | return nullptr; |
1418 | |
1419 | // At least 1 operand must be a shuffle with 1 use because we are creating 2 |
1420 | // instructions. |
1421 | if (none_of(Range: II->args(), P: [](Value *V) { |
1422 | return isa<ShuffleVectorInst>(Val: V) && V->hasOneUse(); |
1423 | })) |
1424 | return nullptr; |
1425 | |
1426 | // See if all arguments are shuffled with the same mask. |
1427 | SmallVector<Value *, 4> NewArgs; |
1428 | Type *SrcTy = X->getType(); |
1429 | for (Use &Arg : II->args()) { |
1430 | if (isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), |
1431 | ScalarOpdIdx: Arg.getOperandNo(), TTI: nullptr)) |
1432 | NewArgs.push_back(Elt: Arg); |
1433 | else if (match(V: &Arg, |
1434 | P: m_Shuffle(v1: m_Value(V&: X), v2: m_Poison(), mask: m_SpecificMask(Mask))) && |
1435 | X->getType() == SrcTy) |
1436 | NewArgs.push_back(Elt: X); |
1437 | else if (match(V: &Arg, P: m_ImmConstant(C))) { |
1438 | // If it's a constant, try find the constant that would be shuffled to C. |
1439 | if (Constant *ShuffledC = |
1440 | unshuffleConstant(ShMask: Mask, C, NewCTy: cast<VectorType>(Val: SrcTy))) |
1441 | NewArgs.push_back(Elt: ShuffledC); |
1442 | else |
1443 | return nullptr; |
1444 | } else |
1445 | return nullptr; |
1446 | } |
1447 | |
1448 | // intrinsic (shuf X, M), (shuf Y, M), ... --> shuf (intrinsic X, Y, ...), M |
1449 | Instruction *FPI = isa<FPMathOperator>(Val: II) ? II : nullptr; |
1450 | // Result type might be a different vector width. |
1451 | // TODO: Check that the result type isn't widened? |
1452 | VectorType *ResTy = |
1453 | VectorType::get(ElementType: II->getType()->getScalarType(), Other: cast<VectorType>(Val: SrcTy)); |
1454 | Value *NewIntrinsic = |
1455 | Builder.CreateIntrinsic(RetTy: ResTy, ID: II->getIntrinsicID(), Args: NewArgs, FMFSource: FPI); |
1456 | return new ShuffleVectorInst(NewIntrinsic, Mask); |
1457 | } |
1458 | |
1459 | /// If all arguments of the intrinsic are reverses, try to pull the reverse |
1460 | /// after the intrinsic. |
1461 | Value *InstCombinerImpl::foldReversedIntrinsicOperands(IntrinsicInst *II) { |
1462 | if (!isTriviallyVectorizable(ID: II->getIntrinsicID())) |
1463 | return nullptr; |
1464 | |
1465 | // At least 1 operand must be a reverse with 1 use because we are creating 2 |
1466 | // instructions. |
1467 | if (none_of(Range: II->args(), P: [](Value *V) { |
1468 | return match(V, P: m_OneUse(SubPattern: m_VecReverse(Op0: m_Value()))); |
1469 | })) |
1470 | return nullptr; |
1471 | |
1472 | Value *X; |
1473 | Constant *C; |
1474 | SmallVector<Value *> NewArgs; |
1475 | for (Use &Arg : II->args()) { |
1476 | if (isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), |
1477 | ScalarOpdIdx: Arg.getOperandNo(), TTI: nullptr)) |
1478 | NewArgs.push_back(Elt: Arg); |
1479 | else if (match(V: &Arg, P: m_VecReverse(Op0: m_Value(V&: X)))) |
1480 | NewArgs.push_back(Elt: X); |
1481 | else if (isSplatValue(V: Arg)) |
1482 | NewArgs.push_back(Elt: Arg); |
1483 | else if (match(V: &Arg, P: m_ImmConstant(C))) |
1484 | NewArgs.push_back(Elt: Builder.CreateVectorReverse(V: C)); |
1485 | else |
1486 | return nullptr; |
1487 | } |
1488 | |
1489 | // intrinsic (reverse X), (reverse Y), ... --> reverse (intrinsic X, Y, ...) |
1490 | Instruction *FPI = isa<FPMathOperator>(Val: II) ? II : nullptr; |
1491 | Instruction *NewIntrinsic = Builder.CreateIntrinsic( |
1492 | RetTy: II->getType(), ID: II->getIntrinsicID(), Args: NewArgs, FMFSource: FPI); |
1493 | return Builder.CreateVectorReverse(V: NewIntrinsic); |
1494 | } |
1495 | |
1496 | /// Fold the following cases and accepts bswap and bitreverse intrinsics: |
1497 | /// bswap(logic_op(bswap(x), y)) --> logic_op(x, bswap(y)) |
1498 | /// bswap(logic_op(bswap(x), bswap(y))) --> logic_op(x, y) (ignores multiuse) |
1499 | template <Intrinsic::ID IntrID> |
1500 | static Instruction *foldBitOrderCrossLogicOp(Value *V, |
1501 | InstCombiner::BuilderTy &Builder) { |
1502 | static_assert(IntrID == Intrinsic::bswap || IntrID == Intrinsic::bitreverse, |
1503 | "This helper only supports BSWAP and BITREVERSE intrinsics" ); |
1504 | |
1505 | Value *X, *Y; |
1506 | // Find bitwise logic op. Check that it is a BinaryOperator explicitly so we |
1507 | // don't match ConstantExpr that aren't meaningful for this transform. |
1508 | if (match(V, P: m_OneUse(SubPattern: m_BitwiseLogic(L: m_Value(V&: X), R: m_Value(V&: Y)))) && |
1509 | isa<BinaryOperator>(Val: V)) { |
1510 | Value *OldReorderX, *OldReorderY; |
1511 | BinaryOperator::BinaryOps Op = cast<BinaryOperator>(Val: V)->getOpcode(); |
1512 | |
1513 | // If both X and Y are bswap/bitreverse, the transform reduces the number |
1514 | // of instructions even if there's multiuse. |
1515 | // If only one operand is bswap/bitreverse, we need to ensure the operand |
1516 | // have only one use. |
1517 | if (match(X, m_Intrinsic<IntrID>(m_Value(V&: OldReorderX))) && |
1518 | match(Y, m_Intrinsic<IntrID>(m_Value(V&: OldReorderY)))) { |
1519 | return BinaryOperator::Create(Op, S1: OldReorderX, S2: OldReorderY); |
1520 | } |
1521 | |
1522 | if (match(X, m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: OldReorderX))))) { |
1523 | Value *NewReorder = Builder.CreateUnaryIntrinsic(ID: IntrID, V: Y); |
1524 | return BinaryOperator::Create(Op, S1: OldReorderX, S2: NewReorder); |
1525 | } |
1526 | |
1527 | if (match(Y, m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: OldReorderY))))) { |
1528 | Value *NewReorder = Builder.CreateUnaryIntrinsic(ID: IntrID, V: X); |
1529 | return BinaryOperator::Create(Op, S1: NewReorder, S2: OldReorderY); |
1530 | } |
1531 | } |
1532 | return nullptr; |
1533 | } |
1534 | |
1535 | static Value *simplifyReductionOperand(Value *Arg, bool CanReorderLanes) { |
1536 | if (!CanReorderLanes) |
1537 | return nullptr; |
1538 | |
1539 | Value *V; |
1540 | if (match(V: Arg, P: m_VecReverse(Op0: m_Value(V)))) |
1541 | return V; |
1542 | |
1543 | ArrayRef<int> Mask; |
1544 | if (!isa<FixedVectorType>(Val: Arg->getType()) || |
1545 | !match(V: Arg, P: m_Shuffle(v1: m_Value(V), v2: m_Undef(), mask: m_Mask(Mask))) || |
1546 | !cast<ShuffleVectorInst>(Val: Arg)->isSingleSource()) |
1547 | return nullptr; |
1548 | |
1549 | int Sz = Mask.size(); |
1550 | SmallBitVector UsedIndices(Sz); |
1551 | for (int Idx : Mask) { |
1552 | if (Idx == PoisonMaskElem || UsedIndices.test(Idx)) |
1553 | return nullptr; |
1554 | UsedIndices.set(Idx); |
1555 | } |
1556 | |
1557 | // Can remove shuffle iff just shuffled elements, no repeats, undefs, or |
1558 | // other changes. |
1559 | return UsedIndices.all() ? V : nullptr; |
1560 | } |
1561 | |
1562 | /// Fold an unsigned minimum of trailing or leading zero bits counts: |
1563 | /// umin(cttz(CtOp, ZeroUndef), ConstOp) --> cttz(CtOp | (1 << ConstOp)) |
1564 | /// umin(ctlz(CtOp, ZeroUndef), ConstOp) --> ctlz(CtOp | (SignedMin |
1565 | /// >> ConstOp)) |
1566 | template <Intrinsic::ID IntrID> |
1567 | static Value * |
1568 | foldMinimumOverTrailingOrLeadingZeroCount(Value *I0, Value *I1, |
1569 | const DataLayout &DL, |
1570 | InstCombiner::BuilderTy &Builder) { |
1571 | static_assert(IntrID == Intrinsic::cttz || IntrID == Intrinsic::ctlz, |
1572 | "This helper only supports cttz and ctlz intrinsics" ); |
1573 | |
1574 | Value *CtOp; |
1575 | Value *ZeroUndef; |
1576 | if (!match(I0, |
1577 | m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: CtOp), m_Value(V&: ZeroUndef))))) |
1578 | return nullptr; |
1579 | |
1580 | unsigned BitWidth = I1->getType()->getScalarSizeInBits(); |
1581 | auto LessBitWidth = [BitWidth](auto &C) { return C.ult(BitWidth); }; |
1582 | if (!match(I1, m_CheckedInt(LessBitWidth))) |
1583 | // We have a constant >= BitWidth (which can be handled by CVP) |
1584 | // or a non-splat vector with elements < and >= BitWidth |
1585 | return nullptr; |
1586 | |
1587 | Type *Ty = I1->getType(); |
1588 | Constant *NewConst = ConstantFoldBinaryOpOperands( |
1589 | Opcode: IntrID == Intrinsic::cttz ? Instruction::Shl : Instruction::LShr, |
1590 | LHS: IntrID == Intrinsic::cttz |
1591 | ? ConstantInt::get(Ty, V: 1) |
1592 | : ConstantInt::get(Ty, V: APInt::getSignedMinValue(numBits: BitWidth)), |
1593 | RHS: cast<Constant>(Val: I1), DL); |
1594 | return Builder.CreateBinaryIntrinsic( |
1595 | ID: IntrID, LHS: Builder.CreateOr(LHS: CtOp, RHS: NewConst), |
1596 | RHS: ConstantInt::getTrue(Ty: ZeroUndef->getType())); |
1597 | } |
1598 | |
1599 | /// Return whether "X LOp (Y ROp Z)" is always equal to |
1600 | /// "(X LOp Y) ROp (X LOp Z)". |
1601 | static bool leftDistributesOverRight(Instruction::BinaryOps LOp, bool HasNUW, |
1602 | bool HasNSW, Intrinsic::ID ROp) { |
1603 | switch (ROp) { |
1604 | case Intrinsic::umax: |
1605 | case Intrinsic::umin: |
1606 | if (HasNUW && LOp == Instruction::Add) |
1607 | return true; |
1608 | if (HasNUW && LOp == Instruction::Shl) |
1609 | return true; |
1610 | return false; |
1611 | case Intrinsic::smax: |
1612 | case Intrinsic::smin: |
1613 | return HasNSW && LOp == Instruction::Add; |
1614 | default: |
1615 | return false; |
1616 | } |
1617 | } |
1618 | |
1619 | // Attempts to factorise a common term |
1620 | // in an instruction that has the form "(A op' B) op (C op' D) |
1621 | // where op is an intrinsic and op' is a binop |
1622 | static Value * |
1623 | foldIntrinsicUsingDistributiveLaws(IntrinsicInst *II, |
1624 | InstCombiner::BuilderTy &Builder) { |
1625 | Value *LHS = II->getOperand(i_nocapture: 0), *RHS = II->getOperand(i_nocapture: 1); |
1626 | Intrinsic::ID TopLevelOpcode = II->getIntrinsicID(); |
1627 | |
1628 | OverflowingBinaryOperator *Op0 = dyn_cast<OverflowingBinaryOperator>(Val: LHS); |
1629 | OverflowingBinaryOperator *Op1 = dyn_cast<OverflowingBinaryOperator>(Val: RHS); |
1630 | |
1631 | if (!Op0 || !Op1) |
1632 | return nullptr; |
1633 | |
1634 | if (Op0->getOpcode() != Op1->getOpcode()) |
1635 | return nullptr; |
1636 | |
1637 | if (!Op0->hasOneUse() || !Op1->hasOneUse()) |
1638 | return nullptr; |
1639 | |
1640 | Instruction::BinaryOps InnerOpcode = |
1641 | static_cast<Instruction::BinaryOps>(Op0->getOpcode()); |
1642 | bool HasNUW = Op0->hasNoUnsignedWrap() && Op1->hasNoUnsignedWrap(); |
1643 | bool HasNSW = Op0->hasNoSignedWrap() && Op1->hasNoSignedWrap(); |
1644 | |
1645 | if (!leftDistributesOverRight(LOp: InnerOpcode, HasNUW, HasNSW, ROp: TopLevelOpcode)) |
1646 | return nullptr; |
1647 | |
1648 | Value *A = Op0->getOperand(i_nocapture: 0); |
1649 | Value *B = Op0->getOperand(i_nocapture: 1); |
1650 | Value *C = Op1->getOperand(i_nocapture: 0); |
1651 | Value *D = Op1->getOperand(i_nocapture: 1); |
1652 | |
1653 | // Attempts to swap variables such that A equals C or B equals D, |
1654 | // if the inner operation is commutative. |
1655 | if (Op0->isCommutative() && A != C && B != D) { |
1656 | if (A == D || B == C) |
1657 | std::swap(a&: C, b&: D); |
1658 | else |
1659 | return nullptr; |
1660 | } |
1661 | |
1662 | BinaryOperator *NewBinop; |
1663 | if (A == C) { |
1664 | Value *NewIntrinsic = Builder.CreateBinaryIntrinsic(ID: TopLevelOpcode, LHS: B, RHS: D); |
1665 | NewBinop = |
1666 | cast<BinaryOperator>(Val: Builder.CreateBinOp(Opc: InnerOpcode, LHS: A, RHS: NewIntrinsic)); |
1667 | } else if (B == D) { |
1668 | Value *NewIntrinsic = Builder.CreateBinaryIntrinsic(ID: TopLevelOpcode, LHS: A, RHS: C); |
1669 | NewBinop = |
1670 | cast<BinaryOperator>(Val: Builder.CreateBinOp(Opc: InnerOpcode, LHS: NewIntrinsic, RHS: B)); |
1671 | } else { |
1672 | return nullptr; |
1673 | } |
1674 | |
1675 | NewBinop->setHasNoUnsignedWrap(HasNUW); |
1676 | NewBinop->setHasNoSignedWrap(HasNSW); |
1677 | |
1678 | return NewBinop; |
1679 | } |
1680 | |
1681 | /// CallInst simplification. This mostly only handles folding of intrinsic |
1682 | /// instructions. For normal calls, it allows visitCallBase to do the heavy |
1683 | /// lifting. |
1684 | Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) { |
1685 | // Don't try to simplify calls without uses. It will not do anything useful, |
1686 | // but will result in the following folds being skipped. |
1687 | if (!CI.use_empty()) { |
1688 | SmallVector<Value *, 8> Args(CI.args()); |
1689 | if (Value *V = simplifyCall(Call: &CI, Callee: CI.getCalledOperand(), Args, |
1690 | Q: SQ.getWithInstruction(I: &CI))) |
1691 | return replaceInstUsesWith(I&: CI, V); |
1692 | } |
1693 | |
1694 | if (Value *FreedOp = getFreedOperand(CB: &CI, TLI: &TLI)) |
1695 | return visitFree(FI&: CI, FreedOp); |
1696 | |
1697 | // If the caller function (i.e. us, the function that contains this CallInst) |
1698 | // is nounwind, mark the call as nounwind, even if the callee isn't. |
1699 | if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) { |
1700 | CI.setDoesNotThrow(); |
1701 | return &CI; |
1702 | } |
1703 | |
1704 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &CI); |
1705 | if (!II) |
1706 | return visitCallBase(Call&: CI); |
1707 | |
1708 | // Intrinsics cannot occur in an invoke or a callbr, so handle them here |
1709 | // instead of in visitCallBase. |
1710 | if (auto *MI = dyn_cast<AnyMemIntrinsic>(Val: II)) { |
1711 | if (ConstantInt *NumBytes = dyn_cast<ConstantInt>(Val: MI->getLength())) { |
1712 | // memmove/cpy/set of zero bytes is a noop. |
1713 | if (NumBytes->isNullValue()) |
1714 | return eraseInstFromFunction(I&: CI); |
1715 | |
1716 | // For atomic unordered mem intrinsics if len is not a positive or |
1717 | // not a multiple of element size then behavior is undefined. |
1718 | if (MI->isAtomic() && |
1719 | (NumBytes->isNegative() || |
1720 | (NumBytes->getZExtValue() % MI->getElementSizeInBytes() != 0))) { |
1721 | CreateNonTerminatorUnreachable(InsertAt: MI); |
1722 | assert(MI->getType()->isVoidTy() && |
1723 | "non void atomic unordered mem intrinsic" ); |
1724 | return eraseInstFromFunction(I&: *MI); |
1725 | } |
1726 | } |
1727 | |
1728 | // No other transformations apply to volatile transfers. |
1729 | if (MI->isVolatile()) |
1730 | return nullptr; |
1731 | |
1732 | if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(Val: MI)) { |
1733 | // memmove(x,x,size) -> noop. |
1734 | if (MTI->getSource() == MTI->getDest()) |
1735 | return eraseInstFromFunction(I&: CI); |
1736 | } |
1737 | |
1738 | auto IsPointerUndefined = [MI](Value *Ptr) { |
1739 | return isa<ConstantPointerNull>(Val: Ptr) && |
1740 | !NullPointerIsDefined( |
1741 | F: MI->getFunction(), |
1742 | AS: cast<PointerType>(Val: Ptr->getType())->getAddressSpace()); |
1743 | }; |
1744 | bool SrcIsUndefined = false; |
1745 | // If we can determine a pointer alignment that is bigger than currently |
1746 | // set, update the alignment. |
1747 | if (auto *MTI = dyn_cast<AnyMemTransferInst>(Val: MI)) { |
1748 | if (Instruction *I = SimplifyAnyMemTransfer(MI: MTI)) |
1749 | return I; |
1750 | SrcIsUndefined = IsPointerUndefined(MTI->getRawSource()); |
1751 | } else if (auto *MSI = dyn_cast<AnyMemSetInst>(Val: MI)) { |
1752 | if (Instruction *I = SimplifyAnyMemSet(MI: MSI)) |
1753 | return I; |
1754 | } |
1755 | |
1756 | // If src/dest is null, this memory intrinsic must be a noop. |
1757 | if (SrcIsUndefined || IsPointerUndefined(MI->getRawDest())) { |
1758 | Builder.CreateAssumption(Cond: Builder.CreateIsNull(Arg: MI->getLength())); |
1759 | return eraseInstFromFunction(I&: CI); |
1760 | } |
1761 | |
1762 | // If we have a memmove and the source operation is a constant global, |
1763 | // then the source and dest pointers can't alias, so we can change this |
1764 | // into a call to memcpy. |
1765 | if (auto *MMI = dyn_cast<AnyMemMoveInst>(Val: MI)) { |
1766 | if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(Val: MMI->getSource())) |
1767 | if (GVSrc->isConstant()) { |
1768 | Module *M = CI.getModule(); |
1769 | Intrinsic::ID MemCpyID = |
1770 | MMI->isAtomic() |
1771 | ? Intrinsic::memcpy_element_unordered_atomic |
1772 | : Intrinsic::memcpy; |
1773 | Type *Tys[3] = { CI.getArgOperand(i: 0)->getType(), |
1774 | CI.getArgOperand(i: 1)->getType(), |
1775 | CI.getArgOperand(i: 2)->getType() }; |
1776 | CI.setCalledFunction( |
1777 | Intrinsic::getOrInsertDeclaration(M, id: MemCpyID, Tys)); |
1778 | return II; |
1779 | } |
1780 | } |
1781 | } |
1782 | |
1783 | // For fixed width vector result intrinsics, use the generic demanded vector |
1784 | // support. |
1785 | if (auto *IIFVTy = dyn_cast<FixedVectorType>(Val: II->getType())) { |
1786 | auto VWidth = IIFVTy->getNumElements(); |
1787 | APInt PoisonElts(VWidth, 0); |
1788 | APInt AllOnesEltMask(APInt::getAllOnes(numBits: VWidth)); |
1789 | if (Value *V = SimplifyDemandedVectorElts(V: II, DemandedElts: AllOnesEltMask, PoisonElts)) { |
1790 | if (V != II) |
1791 | return replaceInstUsesWith(I&: *II, V); |
1792 | return II; |
1793 | } |
1794 | } |
1795 | |
1796 | if (II->isCommutative()) { |
1797 | if (auto Pair = matchSymmetricPair(LHS: II->getOperand(i_nocapture: 0), RHS: II->getOperand(i_nocapture: 1))) { |
1798 | replaceOperand(I&: *II, OpNum: 0, V: Pair->first); |
1799 | replaceOperand(I&: *II, OpNum: 1, V: Pair->second); |
1800 | return II; |
1801 | } |
1802 | |
1803 | if (CallInst *NewCall = canonicalizeConstantArg0ToArg1(Call&: CI)) |
1804 | return NewCall; |
1805 | } |
1806 | |
1807 | // Unused constrained FP intrinsic calls may have declared side effect, which |
1808 | // prevents it from being removed. In some cases however the side effect is |
1809 | // actually absent. To detect this case, call SimplifyConstrainedFPCall. If it |
1810 | // returns a replacement, the call may be removed. |
1811 | if (CI.use_empty() && isa<ConstrainedFPIntrinsic>(Val: CI)) { |
1812 | if (simplifyConstrainedFPCall(Call: &CI, Q: SQ.getWithInstruction(I: &CI))) |
1813 | return eraseInstFromFunction(I&: CI); |
1814 | } |
1815 | |
1816 | Intrinsic::ID IID = II->getIntrinsicID(); |
1817 | switch (IID) { |
1818 | case Intrinsic::objectsize: { |
1819 | SmallVector<Instruction *> InsertedInstructions; |
1820 | if (Value *V = lowerObjectSizeCall(ObjectSize: II, DL, TLI: &TLI, AA, /*MustSucceed=*/false, |
1821 | InsertedInstructions: &InsertedInstructions)) { |
1822 | for (Instruction *Inserted : InsertedInstructions) |
1823 | Worklist.add(I: Inserted); |
1824 | return replaceInstUsesWith(I&: CI, V); |
1825 | } |
1826 | return nullptr; |
1827 | } |
1828 | case Intrinsic::abs: { |
1829 | Value *IIOperand = II->getArgOperand(i: 0); |
1830 | bool IntMinIsPoison = cast<Constant>(Val: II->getArgOperand(i: 1))->isOneValue(); |
1831 | |
1832 | // abs(-x) -> abs(x) |
1833 | // TODO: Copy nsw if it was present on the neg? |
1834 | Value *X; |
1835 | if (match(V: IIOperand, P: m_Neg(V: m_Value(V&: X)))) |
1836 | return replaceOperand(I&: *II, OpNum: 0, V: X); |
1837 | if (match(V: IIOperand, P: m_c_Select(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) |
1838 | return replaceOperand(I&: *II, OpNum: 0, V: X); |
1839 | |
1840 | Value *Y; |
1841 | // abs(a * abs(b)) -> abs(a * b) |
1842 | if (match(V: IIOperand, |
1843 | P: m_OneUse(SubPattern: m_c_Mul(L: m_Value(V&: X), |
1844 | R: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: Y)))))) { |
1845 | bool NSW = |
1846 | cast<Instruction>(Val: IIOperand)->hasNoSignedWrap() && IntMinIsPoison; |
1847 | auto *XY = NSW ? Builder.CreateNSWMul(LHS: X, RHS: Y) : Builder.CreateMul(LHS: X, RHS: Y); |
1848 | return replaceOperand(I&: *II, OpNum: 0, V: XY); |
1849 | } |
1850 | |
1851 | if (std::optional<bool> Known = |
1852 | getKnownSignOrZero(Op: IIOperand, SQ: SQ.getWithInstruction(I: II))) { |
1853 | // abs(x) -> x if x >= 0 (include abs(x-y) --> x - y where x >= y) |
1854 | // abs(x) -> x if x > 0 (include abs(x-y) --> x - y where x > y) |
1855 | if (!*Known) |
1856 | return replaceInstUsesWith(I&: *II, V: IIOperand); |
1857 | |
1858 | // abs(x) -> -x if x < 0 |
1859 | // abs(x) -> -x if x < = 0 (include abs(x-y) --> y - x where x <= y) |
1860 | if (IntMinIsPoison) |
1861 | return BinaryOperator::CreateNSWNeg(Op: IIOperand); |
1862 | return BinaryOperator::CreateNeg(Op: IIOperand); |
1863 | } |
1864 | |
1865 | // abs (sext X) --> zext (abs X*) |
1866 | // Clear the IsIntMin (nsw) bit on the abs to allow narrowing. |
1867 | if (match(V: IIOperand, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))))) { |
1868 | Value *NarrowAbs = |
1869 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: X, RHS: Builder.getFalse()); |
1870 | return CastInst::Create(Instruction::ZExt, S: NarrowAbs, Ty: II->getType()); |
1871 | } |
1872 | |
1873 | // Match a complicated way to check if a number is odd/even: |
1874 | // abs (srem X, 2) --> and X, 1 |
1875 | const APInt *C; |
1876 | if (match(V: IIOperand, P: m_SRem(L: m_Value(V&: X), R: m_APInt(Res&: C))) && *C == 2) |
1877 | return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty: II->getType(), V: 1)); |
1878 | |
1879 | break; |
1880 | } |
1881 | case Intrinsic::umin: { |
1882 | Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1); |
1883 | // umin(x, 1) == zext(x != 0) |
1884 | if (match(V: I1, P: m_One())) { |
1885 | assert(II->getType()->getScalarSizeInBits() != 1 && |
1886 | "Expected simplify of umin with max constant" ); |
1887 | Value *Zero = Constant::getNullValue(Ty: I0->getType()); |
1888 | Value *Cmp = Builder.CreateICmpNE(LHS: I0, RHS: Zero); |
1889 | return CastInst::Create(Instruction::ZExt, S: Cmp, Ty: II->getType()); |
1890 | } |
1891 | // umin(cttz(x), const) --> cttz(x | (1 << const)) |
1892 | if (Value *FoldedCttz = |
1893 | foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::cttz>( |
1894 | I0, I1, DL, Builder)) |
1895 | return replaceInstUsesWith(I&: *II, V: FoldedCttz); |
1896 | // umin(ctlz(x), const) --> ctlz(x | (SignedMin >> const)) |
1897 | if (Value *FoldedCtlz = |
1898 | foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::ctlz>( |
1899 | I0, I1, DL, Builder)) |
1900 | return replaceInstUsesWith(I&: *II, V: FoldedCtlz); |
1901 | [[fallthrough]]; |
1902 | } |
1903 | case Intrinsic::umax: { |
1904 | Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1); |
1905 | Value *X, *Y; |
1906 | if (match(V: I0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: I1, P: m_ZExt(Op: m_Value(V&: Y))) && |
1907 | (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) { |
1908 | Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y); |
1909 | return CastInst::Create(Instruction::ZExt, S: NarrowMaxMin, Ty: II->getType()); |
1910 | } |
1911 | Constant *C; |
1912 | if (match(V: I0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: I1, P: m_Constant(C)) && |
1913 | I0->hasOneUse()) { |
1914 | if (Constant *NarrowC = getLosslessUnsignedTrunc(C, TruncTy: X->getType())) { |
1915 | Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: NarrowC); |
1916 | return CastInst::Create(Instruction::ZExt, S: NarrowMaxMin, Ty: II->getType()); |
1917 | } |
1918 | } |
1919 | // If C is not 0: |
1920 | // umax(nuw_shl(x, C), x + 1) -> x == 0 ? 1 : nuw_shl(x, C) |
1921 | // If C is not 0 or 1: |
1922 | // umax(nuw_mul(x, C), x + 1) -> x == 0 ? 1 : nuw_mul(x, C) |
1923 | auto foldMaxMulShift = [&](Value *A, Value *B) -> Instruction * { |
1924 | const APInt *C; |
1925 | Value *X; |
1926 | if (!match(V: A, P: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: C))) && |
1927 | !(match(V: A, P: m_NUWMul(L: m_Value(V&: X), R: m_APInt(Res&: C))) && !C->isOne())) |
1928 | return nullptr; |
1929 | if (C->isZero()) |
1930 | return nullptr; |
1931 | if (!match(V: B, P: m_OneUse(SubPattern: m_Add(L: m_Specific(V: X), R: m_One())))) |
1932 | return nullptr; |
1933 | |
1934 | Value *Cmp = Builder.CreateICmpEQ(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: 0)); |
1935 | Value *NewSelect = |
1936 | Builder.CreateSelect(C: Cmp, True: ConstantInt::get(Ty: X->getType(), V: 1), False: A); |
1937 | return replaceInstUsesWith(I&: *II, V: NewSelect); |
1938 | }; |
1939 | |
1940 | if (IID == Intrinsic::umax) { |
1941 | if (Instruction *I = foldMaxMulShift(I0, I1)) |
1942 | return I; |
1943 | if (Instruction *I = foldMaxMulShift(I1, I0)) |
1944 | return I; |
1945 | } |
1946 | |
1947 | // If both operands of unsigned min/max are sign-extended, it is still ok |
1948 | // to narrow the operation. |
1949 | [[fallthrough]]; |
1950 | } |
1951 | case Intrinsic::smax: |
1952 | case Intrinsic::smin: { |
1953 | Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1); |
1954 | Value *X, *Y; |
1955 | if (match(V: I0, P: m_SExt(Op: m_Value(V&: X))) && match(V: I1, P: m_SExt(Op: m_Value(V&: Y))) && |
1956 | (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) { |
1957 | Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y); |
1958 | return CastInst::Create(Instruction::SExt, S: NarrowMaxMin, Ty: II->getType()); |
1959 | } |
1960 | |
1961 | Constant *C; |
1962 | if (match(V: I0, P: m_SExt(Op: m_Value(V&: X))) && match(V: I1, P: m_Constant(C)) && |
1963 | I0->hasOneUse()) { |
1964 | if (Constant *NarrowC = getLosslessSignedTrunc(C, TruncTy: X->getType())) { |
1965 | Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: NarrowC); |
1966 | return CastInst::Create(Instruction::SExt, S: NarrowMaxMin, Ty: II->getType()); |
1967 | } |
1968 | } |
1969 | |
1970 | // smax(smin(X, MinC), MaxC) -> smin(smax(X, MaxC), MinC) if MinC s>= MaxC |
1971 | // umax(umin(X, MinC), MaxC) -> umin(umax(X, MaxC), MinC) if MinC u>= MaxC |
1972 | const APInt *MinC, *MaxC; |
1973 | auto CreateCanonicalClampForm = [&](bool IsSigned) { |
1974 | auto MaxIID = IsSigned ? Intrinsic::smax : Intrinsic::umax; |
1975 | auto MinIID = IsSigned ? Intrinsic::smin : Intrinsic::umin; |
1976 | Value *NewMax = Builder.CreateBinaryIntrinsic( |
1977 | ID: MaxIID, LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: *MaxC)); |
1978 | return replaceInstUsesWith( |
1979 | I&: *II, V: Builder.CreateBinaryIntrinsic( |
1980 | ID: MinIID, LHS: NewMax, RHS: ConstantInt::get(Ty: X->getType(), V: *MinC))); |
1981 | }; |
1982 | if (IID == Intrinsic::smax && |
1983 | match(V: I0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::smin>(Op0: m_Value(V&: X), |
1984 | Op1: m_APInt(Res&: MinC)))) && |
1985 | match(V: I1, P: m_APInt(Res&: MaxC)) && MinC->sgt(RHS: *MaxC)) |
1986 | return CreateCanonicalClampForm(true); |
1987 | if (IID == Intrinsic::umax && |
1988 | match(V: I0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::umin>(Op0: m_Value(V&: X), |
1989 | Op1: m_APInt(Res&: MinC)))) && |
1990 | match(V: I1, P: m_APInt(Res&: MaxC)) && MinC->ugt(RHS: *MaxC)) |
1991 | return CreateCanonicalClampForm(false); |
1992 | |
1993 | // umin(i1 X, i1 Y) -> and i1 X, Y |
1994 | // smax(i1 X, i1 Y) -> and i1 X, Y |
1995 | if ((IID == Intrinsic::umin || IID == Intrinsic::smax) && |
1996 | II->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
1997 | return BinaryOperator::CreateAnd(V1: I0, V2: I1); |
1998 | } |
1999 | |
2000 | // umax(i1 X, i1 Y) -> or i1 X, Y |
2001 | // smin(i1 X, i1 Y) -> or i1 X, Y |
2002 | if ((IID == Intrinsic::umax || IID == Intrinsic::smin) && |
2003 | II->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
2004 | return BinaryOperator::CreateOr(V1: I0, V2: I1); |
2005 | } |
2006 | |
2007 | // smin(smax(X, -1), 1) -> scmp(X, 0) |
2008 | // smax(smin(X, 1), -1) -> scmp(X, 0) |
2009 | // At this point, smax(smin(X, 1), -1) is changed to smin(smax(X, -1) |
2010 | // And i1's have been changed to and/ors |
2011 | // So we only need to check for smin |
2012 | if (IID == Intrinsic::smin) { |
2013 | if (match(V: I0, P: m_OneUse(SubPattern: m_SMax(L: m_Value(V&: X), R: m_AllOnes()))) && |
2014 | match(V: I1, P: m_One())) { |
2015 | Value *Zero = ConstantInt::get(Ty: X->getType(), V: 0); |
2016 | return replaceInstUsesWith( |
2017 | I&: CI, |
2018 | V: Builder.CreateIntrinsic(RetTy: II->getType(), ID: Intrinsic::scmp, Args: {X, Zero})); |
2019 | } |
2020 | } |
2021 | |
2022 | if (IID == Intrinsic::smax || IID == Intrinsic::smin) { |
2023 | // smax (neg nsw X), (neg nsw Y) --> neg nsw (smin X, Y) |
2024 | // smin (neg nsw X), (neg nsw Y) --> neg nsw (smax X, Y) |
2025 | // TODO: Canonicalize neg after min/max if I1 is constant. |
2026 | if (match(V: I0, P: m_NSWNeg(V: m_Value(V&: X))) && match(V: I1, P: m_NSWNeg(V: m_Value(V&: Y))) && |
2027 | (I0->hasOneUse() || I1->hasOneUse())) { |
2028 | Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: IID); |
2029 | Value *InvMaxMin = Builder.CreateBinaryIntrinsic(ID: InvID, LHS: X, RHS: Y); |
2030 | return BinaryOperator::CreateNSWNeg(Op: InvMaxMin); |
2031 | } |
2032 | } |
2033 | |
2034 | // (umax X, (xor X, Pow2)) |
2035 | // -> (or X, Pow2) |
2036 | // (umin X, (xor X, Pow2)) |
2037 | // -> (and X, ~Pow2) |
2038 | // (smax X, (xor X, Pos_Pow2)) |
2039 | // -> (or X, Pos_Pow2) |
2040 | // (smin X, (xor X, Pos_Pow2)) |
2041 | // -> (and X, ~Pos_Pow2) |
2042 | // (smax X, (xor X, Neg_Pow2)) |
2043 | // -> (and X, ~Neg_Pow2) |
2044 | // (smin X, (xor X, Neg_Pow2)) |
2045 | // -> (or X, Neg_Pow2) |
2046 | if ((match(V: I0, P: m_c_Xor(L: m_Specific(V: I1), R: m_Value(V&: X))) || |
2047 | match(V: I1, P: m_c_Xor(L: m_Specific(V: I0), R: m_Value(V&: X)))) && |
2048 | isKnownToBeAPowerOfTwo(V: X, /* OrZero */ true)) { |
2049 | bool UseOr = IID == Intrinsic::smax || IID == Intrinsic::umax; |
2050 | bool UseAndN = IID == Intrinsic::smin || IID == Intrinsic::umin; |
2051 | |
2052 | if (IID == Intrinsic::smax || IID == Intrinsic::smin) { |
2053 | auto KnownSign = getKnownSign(Op: X, SQ: SQ.getWithInstruction(I: II)); |
2054 | if (KnownSign == std::nullopt) { |
2055 | UseOr = false; |
2056 | UseAndN = false; |
2057 | } else if (*KnownSign /* true is Signed. */) { |
2058 | UseOr ^= true; |
2059 | UseAndN ^= true; |
2060 | Type *Ty = I0->getType(); |
2061 | // Negative power of 2 must be IntMin. It's possible to be able to |
2062 | // prove negative / power of 2 without actually having known bits, so |
2063 | // just get the value by hand. |
2064 | X = Constant::getIntegerValue( |
2065 | Ty, V: APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits())); |
2066 | } |
2067 | } |
2068 | if (UseOr) |
2069 | return BinaryOperator::CreateOr(V1: I0, V2: X); |
2070 | else if (UseAndN) |
2071 | return BinaryOperator::CreateAnd(V1: I0, V2: Builder.CreateNot(V: X)); |
2072 | } |
2073 | |
2074 | // If we can eliminate ~A and Y is free to invert: |
2075 | // max ~A, Y --> ~(min A, ~Y) |
2076 | // |
2077 | // Examples: |
2078 | // max ~A, ~Y --> ~(min A, Y) |
2079 | // max ~A, C --> ~(min A, ~C) |
2080 | // max ~A, (max ~Y, ~Z) --> ~min( A, (min Y, Z)) |
2081 | auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * { |
2082 | Value *A; |
2083 | if (match(V: X, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: A)))) && |
2084 | !isFreeToInvert(V: A, WillInvertAllUses: A->hasOneUse())) { |
2085 | if (Value *NotY = getFreelyInverted(V: Y, WillInvertAllUses: Y->hasOneUse(), Builder: &Builder)) { |
2086 | Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: IID); |
2087 | Value *InvMaxMin = Builder.CreateBinaryIntrinsic(ID: InvID, LHS: A, RHS: NotY); |
2088 | return BinaryOperator::CreateNot(Op: InvMaxMin); |
2089 | } |
2090 | } |
2091 | return nullptr; |
2092 | }; |
2093 | |
2094 | if (Instruction *I = moveNotAfterMinMax(I0, I1)) |
2095 | return I; |
2096 | if (Instruction *I = moveNotAfterMinMax(I1, I0)) |
2097 | return I; |
2098 | |
2099 | if (Instruction *I = moveAddAfterMinMax(II, Builder)) |
2100 | return I; |
2101 | |
2102 | // minmax (X & NegPow2C, Y & NegPow2C) --> minmax(X, Y) & NegPow2C |
2103 | const APInt *RHSC; |
2104 | if (match(V: I0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_NegatedPower2(V&: RHSC)))) && |
2105 | match(V: I1, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: Y), R: m_SpecificInt(V: *RHSC))))) |
2106 | return BinaryOperator::CreateAnd(V1: Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y), |
2107 | V2: ConstantInt::get(Ty: II->getType(), V: *RHSC)); |
2108 | |
2109 | // smax(X, -X) --> abs(X) |
2110 | // smin(X, -X) --> -abs(X) |
2111 | // umax(X, -X) --> -abs(X) |
2112 | // umin(X, -X) --> abs(X) |
2113 | if (isKnownNegation(X: I0, Y: I1)) { |
2114 | // We can choose either operand as the input to abs(), but if we can |
2115 | // eliminate the only use of a value, that's better for subsequent |
2116 | // transforms/analysis. |
2117 | if (I0->hasOneUse() && !I1->hasOneUse()) |
2118 | std::swap(a&: I0, b&: I1); |
2119 | |
2120 | // This is some variant of abs(). See if we can propagate 'nsw' to the abs |
2121 | // operation and potentially its negation. |
2122 | bool IntMinIsPoison = isKnownNegation(X: I0, Y: I1, /* NeedNSW */ true); |
2123 | Value *Abs = Builder.CreateBinaryIntrinsic( |
2124 | ID: Intrinsic::abs, LHS: I0, |
2125 | RHS: ConstantInt::getBool(Context&: II->getContext(), V: IntMinIsPoison)); |
2126 | |
2127 | // We don't have a "nabs" intrinsic, so negate if needed based on the |
2128 | // max/min operation. |
2129 | if (IID == Intrinsic::smin || IID == Intrinsic::umax) |
2130 | Abs = Builder.CreateNeg(V: Abs, Name: "nabs" , HasNSW: IntMinIsPoison); |
2131 | return replaceInstUsesWith(I&: CI, V: Abs); |
2132 | } |
2133 | |
2134 | if (Instruction *Sel = foldClampRangeOfTwo(II, Builder)) |
2135 | return Sel; |
2136 | |
2137 | if (Instruction *SAdd = matchSAddSubSat(MinMax1&: *II)) |
2138 | return SAdd; |
2139 | |
2140 | if (Value *NewMinMax = reassociateMinMaxWithConstants(II, Builder, SQ)) |
2141 | return replaceInstUsesWith(I&: *II, V: NewMinMax); |
2142 | |
2143 | if (Instruction *R = reassociateMinMaxWithConstantInOperand(II, Builder)) |
2144 | return R; |
2145 | |
2146 | if (Instruction *NewMinMax = factorizeMinMaxTree(II)) |
2147 | return NewMinMax; |
2148 | |
2149 | // Try to fold minmax with constant RHS based on range information |
2150 | if (match(V: I1, P: m_APIntAllowPoison(Res&: RHSC))) { |
2151 | ICmpInst::Predicate Pred = |
2152 | ICmpInst::getNonStrictPredicate(pred: MinMaxIntrinsic::getPredicate(ID: IID)); |
2153 | bool IsSigned = MinMaxIntrinsic::isSigned(ID: IID); |
2154 | ConstantRange LHS_CR = computeConstantRangeIncludingKnownBits( |
2155 | V: I0, ForSigned: IsSigned, SQ: SQ.getWithInstruction(I: II)); |
2156 | if (!LHS_CR.isFullSet()) { |
2157 | if (LHS_CR.icmp(Pred, Other: *RHSC)) |
2158 | return replaceInstUsesWith(I&: *II, V: I0); |
2159 | if (LHS_CR.icmp(Pred: ICmpInst::getSwappedPredicate(pred: Pred), Other: *RHSC)) |
2160 | return replaceInstUsesWith(I&: *II, |
2161 | V: ConstantInt::get(Ty: II->getType(), V: *RHSC)); |
2162 | } |
2163 | } |
2164 | |
2165 | if (Value *V = foldIntrinsicUsingDistributiveLaws(II, Builder)) |
2166 | return replaceInstUsesWith(I&: *II, V); |
2167 | |
2168 | break; |
2169 | } |
2170 | case Intrinsic::scmp: { |
2171 | Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1); |
2172 | Value *LHS, *RHS; |
2173 | if (match(V: I0, P: m_NSWSub(L: m_Value(V&: LHS), R: m_Value(V&: RHS))) && match(V: I1, P: m_Zero())) |
2174 | return replaceInstUsesWith( |
2175 | I&: CI, |
2176 | V: Builder.CreateIntrinsic(RetTy: II->getType(), ID: Intrinsic::scmp, Args: {LHS, RHS})); |
2177 | break; |
2178 | } |
2179 | case Intrinsic::bitreverse: { |
2180 | Value *IIOperand = II->getArgOperand(i: 0); |
2181 | // bitrev (zext i1 X to ?) --> X ? SignBitC : 0 |
2182 | Value *X; |
2183 | if (match(V: IIOperand, P: m_ZExt(Op: m_Value(V&: X))) && |
2184 | X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
2185 | Type *Ty = II->getType(); |
2186 | APInt SignBit = APInt::getSignMask(BitWidth: Ty->getScalarSizeInBits()); |
2187 | return SelectInst::Create(C: X, S1: ConstantInt::get(Ty, V: SignBit), |
2188 | S2: ConstantInt::getNullValue(Ty)); |
2189 | } |
2190 | |
2191 | if (Instruction *crossLogicOpFold = |
2192 | foldBitOrderCrossLogicOp<Intrinsic::bitreverse>(V: IIOperand, Builder)) |
2193 | return crossLogicOpFold; |
2194 | |
2195 | break; |
2196 | } |
2197 | case Intrinsic::bswap: { |
2198 | Value *IIOperand = II->getArgOperand(i: 0); |
2199 | |
2200 | // Try to canonicalize bswap-of-logical-shift-by-8-bit-multiple as |
2201 | // inverse-shift-of-bswap: |
2202 | // bswap (shl X, Y) --> lshr (bswap X), Y |
2203 | // bswap (lshr X, Y) --> shl (bswap X), Y |
2204 | Value *X, *Y; |
2205 | if (match(V: IIOperand, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: X), R: m_Value(V&: Y))))) { |
2206 | unsigned BitWidth = IIOperand->getType()->getScalarSizeInBits(); |
2207 | if (MaskedValueIsZero(V: Y, Mask: APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: 3))) { |
2208 | Value *NewSwap = Builder.CreateUnaryIntrinsic(ID: Intrinsic::bswap, V: X); |
2209 | BinaryOperator::BinaryOps InverseShift = |
2210 | cast<BinaryOperator>(Val: IIOperand)->getOpcode() == Instruction::Shl |
2211 | ? Instruction::LShr |
2212 | : Instruction::Shl; |
2213 | return BinaryOperator::Create(Op: InverseShift, S1: NewSwap, S2: Y); |
2214 | } |
2215 | } |
2216 | |
2217 | KnownBits Known = computeKnownBits(V: IIOperand, CxtI: II); |
2218 | uint64_t LZ = alignDown(Value: Known.countMinLeadingZeros(), Align: 8); |
2219 | uint64_t TZ = alignDown(Value: Known.countMinTrailingZeros(), Align: 8); |
2220 | unsigned BW = Known.getBitWidth(); |
2221 | |
2222 | // bswap(x) -> shift(x) if x has exactly one "active byte" |
2223 | if (BW - LZ - TZ == 8) { |
2224 | assert(LZ != TZ && "active byte cannot be in the middle" ); |
2225 | if (LZ > TZ) // -> shl(x) if the "active byte" is in the low part of x |
2226 | return BinaryOperator::CreateNUWShl( |
2227 | V1: IIOperand, V2: ConstantInt::get(Ty: IIOperand->getType(), V: LZ - TZ)); |
2228 | // -> lshr(x) if the "active byte" is in the high part of x |
2229 | return BinaryOperator::CreateExactLShr( |
2230 | V1: IIOperand, V2: ConstantInt::get(Ty: IIOperand->getType(), V: TZ - LZ)); |
2231 | } |
2232 | |
2233 | // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) |
2234 | if (match(V: IIOperand, P: m_Trunc(Op: m_BSwap(Op0: m_Value(V&: X))))) { |
2235 | unsigned C = X->getType()->getScalarSizeInBits() - BW; |
2236 | Value *CV = ConstantInt::get(Ty: X->getType(), V: C); |
2237 | Value *V = Builder.CreateLShr(LHS: X, RHS: CV); |
2238 | return new TruncInst(V, IIOperand->getType()); |
2239 | } |
2240 | |
2241 | if (Instruction *crossLogicOpFold = |
2242 | foldBitOrderCrossLogicOp<Intrinsic::bswap>(V: IIOperand, Builder)) { |
2243 | return crossLogicOpFold; |
2244 | } |
2245 | |
2246 | // Try to fold into bitreverse if bswap is the root of the expression tree. |
2247 | if (Instruction *BitOp = matchBSwapOrBitReverse(I&: *II, /*MatchBSwaps*/ false, |
2248 | /*MatchBitReversals*/ true)) |
2249 | return BitOp; |
2250 | break; |
2251 | } |
2252 | case Intrinsic::masked_load: |
2253 | if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(II&: *II)) |
2254 | return replaceInstUsesWith(I&: CI, V: SimplifiedMaskedOp); |
2255 | break; |
2256 | case Intrinsic::masked_store: |
2257 | return simplifyMaskedStore(II&: *II); |
2258 | case Intrinsic::masked_gather: |
2259 | return simplifyMaskedGather(II&: *II); |
2260 | case Intrinsic::masked_scatter: |
2261 | return simplifyMaskedScatter(II&: *II); |
2262 | case Intrinsic::launder_invariant_group: |
2263 | case Intrinsic::strip_invariant_group: |
2264 | if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(II&: *II, IC&: *this)) |
2265 | return replaceInstUsesWith(I&: *II, V: SkippedBarrier); |
2266 | break; |
2267 | case Intrinsic::powi: |
2268 | if (ConstantInt *Power = dyn_cast<ConstantInt>(Val: II->getArgOperand(i: 1))) { |
2269 | // 0 and 1 are handled in instsimplify |
2270 | // powi(x, -1) -> 1/x |
2271 | if (Power->isMinusOne()) |
2272 | return BinaryOperator::CreateFDivFMF(V1: ConstantFP::get(Ty: CI.getType(), V: 1.0), |
2273 | V2: II->getArgOperand(i: 0), FMFSource: II); |
2274 | // powi(x, 2) -> x*x |
2275 | if (Power->equalsInt(V: 2)) |
2276 | return BinaryOperator::CreateFMulFMF(V1: II->getArgOperand(i: 0), |
2277 | V2: II->getArgOperand(i: 0), FMFSource: II); |
2278 | |
2279 | if (!Power->getValue()[0]) { |
2280 | Value *X; |
2281 | // If power is even: |
2282 | // powi(-x, p) -> powi(x, p) |
2283 | // powi(fabs(x), p) -> powi(x, p) |
2284 | // powi(copysign(x, y), p) -> powi(x, p) |
2285 | if (match(V: II->getArgOperand(i: 0), P: m_FNeg(X: m_Value(V&: X))) || |
2286 | match(V: II->getArgOperand(i: 0), P: m_FAbs(Op0: m_Value(V&: X))) || |
2287 | match(V: II->getArgOperand(i: 0), |
2288 | P: m_Intrinsic<Intrinsic::copysign>(Op0: m_Value(V&: X), Op1: m_Value()))) |
2289 | return replaceOperand(I&: *II, OpNum: 0, V: X); |
2290 | } |
2291 | } |
2292 | break; |
2293 | |
2294 | case Intrinsic::cttz: |
2295 | case Intrinsic::ctlz: |
2296 | if (auto *I = foldCttzCtlz(II&: *II, IC&: *this)) |
2297 | return I; |
2298 | break; |
2299 | |
2300 | case Intrinsic::ctpop: |
2301 | if (auto *I = foldCtpop(II&: *II, IC&: *this)) |
2302 | return I; |
2303 | break; |
2304 | |
2305 | case Intrinsic::fshl: |
2306 | case Intrinsic::fshr: { |
2307 | Value *Op0 = II->getArgOperand(i: 0), *Op1 = II->getArgOperand(i: 1); |
2308 | Type *Ty = II->getType(); |
2309 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
2310 | Constant *ShAmtC; |
2311 | if (match(V: II->getArgOperand(i: 2), P: m_ImmConstant(C&: ShAmtC))) { |
2312 | // Canonicalize a shift amount constant operand to modulo the bit-width. |
2313 | Constant *WidthC = ConstantInt::get(Ty, V: BitWidth); |
2314 | Constant *ModuloC = |
2315 | ConstantFoldBinaryOpOperands(Opcode: Instruction::URem, LHS: ShAmtC, RHS: WidthC, DL); |
2316 | if (!ModuloC) |
2317 | return nullptr; |
2318 | if (ModuloC != ShAmtC) |
2319 | return replaceOperand(I&: *II, OpNum: 2, V: ModuloC); |
2320 | |
2321 | assert(match(ConstantFoldCompareInstOperands(ICmpInst::ICMP_UGT, WidthC, |
2322 | ShAmtC, DL), |
2323 | m_One()) && |
2324 | "Shift amount expected to be modulo bitwidth" ); |
2325 | |
2326 | // Canonicalize funnel shift right by constant to funnel shift left. This |
2327 | // is not entirely arbitrary. For historical reasons, the backend may |
2328 | // recognize rotate left patterns but miss rotate right patterns. |
2329 | if (IID == Intrinsic::fshr) { |
2330 | // fshr X, Y, C --> fshl X, Y, (BitWidth - C) if C is not zero. |
2331 | if (!isKnownNonZero(V: ShAmtC, Q: SQ.getWithInstruction(I: II))) |
2332 | return nullptr; |
2333 | |
2334 | Constant *LeftShiftC = ConstantExpr::getSub(C1: WidthC, C2: ShAmtC); |
2335 | Module *Mod = II->getModule(); |
2336 | Function *Fshl = |
2337 | Intrinsic::getOrInsertDeclaration(M: Mod, id: Intrinsic::fshl, Tys: Ty); |
2338 | return CallInst::Create(Func: Fshl, Args: { Op0, Op1, LeftShiftC }); |
2339 | } |
2340 | assert(IID == Intrinsic::fshl && |
2341 | "All funnel shifts by simple constants should go left" ); |
2342 | |
2343 | // fshl(X, 0, C) --> shl X, C |
2344 | // fshl(X, undef, C) --> shl X, C |
2345 | if (match(V: Op1, P: m_ZeroInt()) || match(V: Op1, P: m_Undef())) |
2346 | return BinaryOperator::CreateShl(V1: Op0, V2: ShAmtC); |
2347 | |
2348 | // fshl(0, X, C) --> lshr X, (BW-C) |
2349 | // fshl(undef, X, C) --> lshr X, (BW-C) |
2350 | if (match(V: Op0, P: m_ZeroInt()) || match(V: Op0, P: m_Undef())) |
2351 | return BinaryOperator::CreateLShr(V1: Op1, |
2352 | V2: ConstantExpr::getSub(C1: WidthC, C2: ShAmtC)); |
2353 | |
2354 | // fshl i16 X, X, 8 --> bswap i16 X (reduce to more-specific form) |
2355 | if (Op0 == Op1 && BitWidth == 16 && match(V: ShAmtC, P: m_SpecificInt(V: 8))) { |
2356 | Module *Mod = II->getModule(); |
2357 | Function *Bswap = |
2358 | Intrinsic::getOrInsertDeclaration(M: Mod, id: Intrinsic::bswap, Tys: Ty); |
2359 | return CallInst::Create(Func: Bswap, Args: { Op0 }); |
2360 | } |
2361 | if (Instruction *BitOp = |
2362 | matchBSwapOrBitReverse(I&: *II, /*MatchBSwaps*/ true, |
2363 | /*MatchBitReversals*/ true)) |
2364 | return BitOp; |
2365 | } |
2366 | |
2367 | // fshl(X, X, Neg(Y)) --> fshr(X, X, Y) |
2368 | // fshr(X, X, Neg(Y)) --> fshl(X, X, Y) |
2369 | // if BitWidth is a power-of-2 |
2370 | Value *Y; |
2371 | if (Op0 == Op1 && isPowerOf2_32(Value: BitWidth) && |
2372 | match(V: II->getArgOperand(i: 2), P: m_Neg(V: m_Value(V&: Y)))) { |
2373 | Module *Mod = II->getModule(); |
2374 | Function *OppositeShift = Intrinsic::getOrInsertDeclaration( |
2375 | M: Mod, id: IID == Intrinsic::fshl ? Intrinsic::fshr : Intrinsic::fshl, Tys: Ty); |
2376 | return CallInst::Create(Func: OppositeShift, Args: {Op0, Op1, Y}); |
2377 | } |
2378 | |
2379 | // fshl(X, 0, Y) --> shl(X, and(Y, BitWidth - 1)) if bitwidth is a |
2380 | // power-of-2 |
2381 | if (IID == Intrinsic::fshl && isPowerOf2_32(Value: BitWidth) && |
2382 | match(V: Op1, P: m_ZeroInt())) { |
2383 | Value *Op2 = II->getArgOperand(i: 2); |
2384 | Value *And = Builder.CreateAnd(LHS: Op2, RHS: ConstantInt::get(Ty, V: BitWidth - 1)); |
2385 | return BinaryOperator::CreateShl(V1: Op0, V2: And); |
2386 | } |
2387 | |
2388 | // Left or right might be masked. |
2389 | if (SimplifyDemandedInstructionBits(Inst&: *II)) |
2390 | return &CI; |
2391 | |
2392 | // The shift amount (operand 2) of a funnel shift is modulo the bitwidth, |
2393 | // so only the low bits of the shift amount are demanded if the bitwidth is |
2394 | // a power-of-2. |
2395 | if (!isPowerOf2_32(Value: BitWidth)) |
2396 | break; |
2397 | APInt Op2Demanded = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: Log2_32_Ceil(Value: BitWidth)); |
2398 | KnownBits Op2Known(BitWidth); |
2399 | if (SimplifyDemandedBits(I: II, OpNo: 2, DemandedMask: Op2Demanded, Known&: Op2Known)) |
2400 | return &CI; |
2401 | break; |
2402 | } |
2403 | case Intrinsic::ptrmask: { |
2404 | unsigned BitWidth = DL.getPointerTypeSizeInBits(II->getType()); |
2405 | KnownBits Known(BitWidth); |
2406 | if (SimplifyDemandedInstructionBits(Inst&: *II, Known)) |
2407 | return II; |
2408 | |
2409 | Value *InnerPtr, *InnerMask; |
2410 | bool Changed = false; |
2411 | // Combine: |
2412 | // (ptrmask (ptrmask p, A), B) |
2413 | // -> (ptrmask p, (and A, B)) |
2414 | if (match(V: II->getArgOperand(i: 0), |
2415 | P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ptrmask>(Op0: m_Value(V&: InnerPtr), |
2416 | Op1: m_Value(V&: InnerMask))))) { |
2417 | assert(II->getArgOperand(1)->getType() == InnerMask->getType() && |
2418 | "Mask types must match" ); |
2419 | // TODO: If InnerMask == Op1, we could copy attributes from inner |
2420 | // callsite -> outer callsite. |
2421 | Value *NewMask = Builder.CreateAnd(LHS: II->getArgOperand(i: 1), RHS: InnerMask); |
2422 | replaceOperand(I&: CI, OpNum: 0, V: InnerPtr); |
2423 | replaceOperand(I&: CI, OpNum: 1, V: NewMask); |
2424 | Changed = true; |
2425 | } |
2426 | |
2427 | // See if we can deduce non-null. |
2428 | if (!CI.hasRetAttr(Kind: Attribute::NonNull) && |
2429 | (Known.isNonZero() || |
2430 | isKnownNonZero(V: II, Q: getSimplifyQuery().getWithInstruction(I: II)))) { |
2431 | CI.addRetAttr(Kind: Attribute::NonNull); |
2432 | Changed = true; |
2433 | } |
2434 | |
2435 | unsigned NewAlignmentLog = |
2436 | std::min(a: Value::MaxAlignmentExponent, |
2437 | b: std::min(a: BitWidth - 1, b: Known.countMinTrailingZeros())); |
2438 | // Known bits will capture if we had alignment information associated with |
2439 | // the pointer argument. |
2440 | if (NewAlignmentLog > Log2(A: CI.getRetAlign().valueOrOne())) { |
2441 | CI.addRetAttr(Attr: Attribute::getWithAlignment( |
2442 | Context&: CI.getContext(), Alignment: Align(uint64_t(1) << NewAlignmentLog))); |
2443 | Changed = true; |
2444 | } |
2445 | if (Changed) |
2446 | return &CI; |
2447 | break; |
2448 | } |
2449 | case Intrinsic::uadd_with_overflow: |
2450 | case Intrinsic::sadd_with_overflow: { |
2451 | if (Instruction *I = foldIntrinsicWithOverflowCommon(II)) |
2452 | return I; |
2453 | |
2454 | // Given 2 constant operands whose sum does not overflow: |
2455 | // uaddo (X +nuw C0), C1 -> uaddo X, C0 + C1 |
2456 | // saddo (X +nsw C0), C1 -> saddo X, C0 + C1 |
2457 | Value *X; |
2458 | const APInt *C0, *C1; |
2459 | Value *Arg0 = II->getArgOperand(i: 0); |
2460 | Value *Arg1 = II->getArgOperand(i: 1); |
2461 | bool IsSigned = IID == Intrinsic::sadd_with_overflow; |
2462 | bool HasNWAdd = IsSigned |
2463 | ? match(V: Arg0, P: m_NSWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C0))) |
2464 | : match(V: Arg0, P: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C0))); |
2465 | if (HasNWAdd && match(V: Arg1, P: m_APInt(Res&: C1))) { |
2466 | bool Overflow; |
2467 | APInt NewC = |
2468 | IsSigned ? C1->sadd_ov(RHS: *C0, Overflow) : C1->uadd_ov(RHS: *C0, Overflow); |
2469 | if (!Overflow) |
2470 | return replaceInstUsesWith( |
2471 | I&: *II, V: Builder.CreateBinaryIntrinsic( |
2472 | ID: IID, LHS: X, RHS: ConstantInt::get(Ty: Arg1->getType(), V: NewC))); |
2473 | } |
2474 | break; |
2475 | } |
2476 | |
2477 | case Intrinsic::umul_with_overflow: |
2478 | case Intrinsic::smul_with_overflow: |
2479 | case Intrinsic::usub_with_overflow: |
2480 | if (Instruction *I = foldIntrinsicWithOverflowCommon(II)) |
2481 | return I; |
2482 | break; |
2483 | |
2484 | case Intrinsic::ssub_with_overflow: { |
2485 | if (Instruction *I = foldIntrinsicWithOverflowCommon(II)) |
2486 | return I; |
2487 | |
2488 | Constant *C; |
2489 | Value *Arg0 = II->getArgOperand(i: 0); |
2490 | Value *Arg1 = II->getArgOperand(i: 1); |
2491 | // Given a constant C that is not the minimum signed value |
2492 | // for an integer of a given bit width: |
2493 | // |
2494 | // ssubo X, C -> saddo X, -C |
2495 | if (match(V: Arg1, P: m_Constant(C)) && C->isNotMinSignedValue()) { |
2496 | Value *NegVal = ConstantExpr::getNeg(C); |
2497 | // Build a saddo call that is equivalent to the discovered |
2498 | // ssubo call. |
2499 | return replaceInstUsesWith( |
2500 | I&: *II, V: Builder.CreateBinaryIntrinsic(ID: Intrinsic::sadd_with_overflow, |
2501 | LHS: Arg0, RHS: NegVal)); |
2502 | } |
2503 | |
2504 | break; |
2505 | } |
2506 | |
2507 | case Intrinsic::uadd_sat: |
2508 | case Intrinsic::sadd_sat: |
2509 | case Intrinsic::usub_sat: |
2510 | case Intrinsic::ssub_sat: { |
2511 | SaturatingInst *SI = cast<SaturatingInst>(Val: II); |
2512 | Type *Ty = SI->getType(); |
2513 | Value *Arg0 = SI->getLHS(); |
2514 | Value *Arg1 = SI->getRHS(); |
2515 | |
2516 | // Make use of known overflow information. |
2517 | OverflowResult OR = computeOverflow(BinaryOp: SI->getBinaryOp(), IsSigned: SI->isSigned(), |
2518 | LHS: Arg0, RHS: Arg1, CxtI: SI); |
2519 | switch (OR) { |
2520 | case OverflowResult::MayOverflow: |
2521 | break; |
2522 | case OverflowResult::NeverOverflows: |
2523 | if (SI->isSigned()) |
2524 | return BinaryOperator::CreateNSW(Opc: SI->getBinaryOp(), V1: Arg0, V2: Arg1); |
2525 | else |
2526 | return BinaryOperator::CreateNUW(Opc: SI->getBinaryOp(), V1: Arg0, V2: Arg1); |
2527 | case OverflowResult::AlwaysOverflowsLow: { |
2528 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
2529 | APInt Min = APSInt::getMinValue(numBits: BitWidth, Unsigned: !SI->isSigned()); |
2530 | return replaceInstUsesWith(I&: *SI, V: ConstantInt::get(Ty, V: Min)); |
2531 | } |
2532 | case OverflowResult::AlwaysOverflowsHigh: { |
2533 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
2534 | APInt Max = APSInt::getMaxValue(numBits: BitWidth, Unsigned: !SI->isSigned()); |
2535 | return replaceInstUsesWith(I&: *SI, V: ConstantInt::get(Ty, V: Max)); |
2536 | } |
2537 | } |
2538 | |
2539 | // usub_sat((sub nuw C, A), C1) -> usub_sat(usub_sat(C, C1), A) |
2540 | // which after that: |
2541 | // usub_sat((sub nuw C, A), C1) -> usub_sat(C - C1, A) if C1 u< C |
2542 | // usub_sat((sub nuw C, A), C1) -> 0 otherwise |
2543 | Constant *C, *C1; |
2544 | Value *A; |
2545 | if (IID == Intrinsic::usub_sat && |
2546 | match(V: Arg0, P: m_NUWSub(L: m_ImmConstant(C), R: m_Value(V&: A))) && |
2547 | match(V: Arg1, P: m_ImmConstant(C&: C1))) { |
2548 | auto *NewC = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: C, RHS: C1); |
2549 | auto *NewSub = |
2550 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: NewC, RHS: A); |
2551 | return replaceInstUsesWith(I&: *SI, V: NewSub); |
2552 | } |
2553 | |
2554 | // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN |
2555 | if (IID == Intrinsic::ssub_sat && match(V: Arg1, P: m_Constant(C)) && |
2556 | C->isNotMinSignedValue()) { |
2557 | Value *NegVal = ConstantExpr::getNeg(C); |
2558 | return replaceInstUsesWith( |
2559 | I&: *II, V: Builder.CreateBinaryIntrinsic( |
2560 | ID: Intrinsic::sadd_sat, LHS: Arg0, RHS: NegVal)); |
2561 | } |
2562 | |
2563 | // sat(sat(X + Val2) + Val) -> sat(X + (Val+Val2)) |
2564 | // sat(sat(X - Val2) - Val) -> sat(X - (Val+Val2)) |
2565 | // if Val and Val2 have the same sign |
2566 | if (auto *Other = dyn_cast<IntrinsicInst>(Val: Arg0)) { |
2567 | Value *X; |
2568 | const APInt *Val, *Val2; |
2569 | APInt NewVal; |
2570 | bool IsUnsigned = |
2571 | IID == Intrinsic::uadd_sat || IID == Intrinsic::usub_sat; |
2572 | if (Other->getIntrinsicID() == IID && |
2573 | match(V: Arg1, P: m_APInt(Res&: Val)) && |
2574 | match(V: Other->getArgOperand(i: 0), P: m_Value(V&: X)) && |
2575 | match(V: Other->getArgOperand(i: 1), P: m_APInt(Res&: Val2))) { |
2576 | if (IsUnsigned) |
2577 | NewVal = Val->uadd_sat(RHS: *Val2); |
2578 | else if (Val->isNonNegative() == Val2->isNonNegative()) { |
2579 | bool Overflow; |
2580 | NewVal = Val->sadd_ov(RHS: *Val2, Overflow); |
2581 | if (Overflow) { |
2582 | // Both adds together may add more than SignedMaxValue |
2583 | // without saturating the final result. |
2584 | break; |
2585 | } |
2586 | } else { |
2587 | // Cannot fold saturated addition with different signs. |
2588 | break; |
2589 | } |
2590 | |
2591 | return replaceInstUsesWith( |
2592 | I&: *II, V: Builder.CreateBinaryIntrinsic( |
2593 | ID: IID, LHS: X, RHS: ConstantInt::get(Ty: II->getType(), V: NewVal))); |
2594 | } |
2595 | } |
2596 | break; |
2597 | } |
2598 | |
2599 | case Intrinsic::minnum: |
2600 | case Intrinsic::maxnum: |
2601 | case Intrinsic::minimum: |
2602 | case Intrinsic::maximum: { |
2603 | Value *Arg0 = II->getArgOperand(i: 0); |
2604 | Value *Arg1 = II->getArgOperand(i: 1); |
2605 | Value *X, *Y; |
2606 | if (match(V: Arg0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Arg1, P: m_FNeg(X: m_Value(V&: Y))) && |
2607 | (Arg0->hasOneUse() || Arg1->hasOneUse())) { |
2608 | // If both operands are negated, invert the call and negate the result: |
2609 | // min(-X, -Y) --> -(max(X, Y)) |
2610 | // max(-X, -Y) --> -(min(X, Y)) |
2611 | Intrinsic::ID NewIID; |
2612 | switch (IID) { |
2613 | case Intrinsic::maxnum: |
2614 | NewIID = Intrinsic::minnum; |
2615 | break; |
2616 | case Intrinsic::minnum: |
2617 | NewIID = Intrinsic::maxnum; |
2618 | break; |
2619 | case Intrinsic::maximum: |
2620 | NewIID = Intrinsic::minimum; |
2621 | break; |
2622 | case Intrinsic::minimum: |
2623 | NewIID = Intrinsic::maximum; |
2624 | break; |
2625 | default: |
2626 | llvm_unreachable("unexpected intrinsic ID" ); |
2627 | } |
2628 | Value *NewCall = Builder.CreateBinaryIntrinsic(ID: NewIID, LHS: X, RHS: Y, FMFSource: II); |
2629 | Instruction *FNeg = UnaryOperator::CreateFNeg(V: NewCall); |
2630 | FNeg->copyIRFlags(V: II); |
2631 | return FNeg; |
2632 | } |
2633 | |
2634 | // m(m(X, C2), C1) -> m(X, C) |
2635 | const APFloat *C1, *C2; |
2636 | if (auto *M = dyn_cast<IntrinsicInst>(Val: Arg0)) { |
2637 | if (M->getIntrinsicID() == IID && match(V: Arg1, P: m_APFloat(Res&: C1)) && |
2638 | ((match(V: M->getArgOperand(i: 0), P: m_Value(V&: X)) && |
2639 | match(V: M->getArgOperand(i: 1), P: m_APFloat(Res&: C2))) || |
2640 | (match(V: M->getArgOperand(i: 1), P: m_Value(V&: X)) && |
2641 | match(V: M->getArgOperand(i: 0), P: m_APFloat(Res&: C2))))) { |
2642 | APFloat Res(0.0); |
2643 | switch (IID) { |
2644 | case Intrinsic::maxnum: |
2645 | Res = maxnum(A: *C1, B: *C2); |
2646 | break; |
2647 | case Intrinsic::minnum: |
2648 | Res = minnum(A: *C1, B: *C2); |
2649 | break; |
2650 | case Intrinsic::maximum: |
2651 | Res = maximum(A: *C1, B: *C2); |
2652 | break; |
2653 | case Intrinsic::minimum: |
2654 | Res = minimum(A: *C1, B: *C2); |
2655 | break; |
2656 | default: |
2657 | llvm_unreachable("unexpected intrinsic ID" ); |
2658 | } |
2659 | // TODO: Conservatively intersecting FMF. If Res == C2, the transform |
2660 | // was a simplification (so Arg0 and its original flags could |
2661 | // propagate?) |
2662 | Value *V = Builder.CreateBinaryIntrinsic( |
2663 | ID: IID, LHS: X, RHS: ConstantFP::get(Ty: Arg0->getType(), V: Res), |
2664 | FMFSource: FMFSource::intersect(A: II, B: M)); |
2665 | return replaceInstUsesWith(I&: *II, V); |
2666 | } |
2667 | } |
2668 | |
2669 | // m((fpext X), (fpext Y)) -> fpext (m(X, Y)) |
2670 | if (match(V: Arg0, P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: X)))) && |
2671 | match(V: Arg1, P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: Y)))) && |
2672 | X->getType() == Y->getType()) { |
2673 | Value *NewCall = |
2674 | Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y, FMFSource: II, Name: II->getName()); |
2675 | return new FPExtInst(NewCall, II->getType()); |
2676 | } |
2677 | |
2678 | // max X, -X --> fabs X |
2679 | // min X, -X --> -(fabs X) |
2680 | // TODO: Remove one-use limitation? That is obviously better for max, |
2681 | // hence why we don't check for one-use for that. However, |
2682 | // it would be an extra instruction for min (fnabs), but |
2683 | // that is still likely better for analysis and codegen. |
2684 | auto IsMinMaxOrXNegX = [IID, &X](Value *Op0, Value *Op1) { |
2685 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_Specific(V: X))) |
2686 | return Op0->hasOneUse() || |
2687 | (IID != Intrinsic::minimum && IID != Intrinsic::minnum); |
2688 | return false; |
2689 | }; |
2690 | |
2691 | if (IsMinMaxOrXNegX(Arg0, Arg1) || IsMinMaxOrXNegX(Arg1, Arg0)) { |
2692 | Value *R = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: II); |
2693 | if (IID == Intrinsic::minimum || IID == Intrinsic::minnum) |
2694 | R = Builder.CreateFNegFMF(V: R, FMFSource: II); |
2695 | return replaceInstUsesWith(I&: *II, V: R); |
2696 | } |
2697 | |
2698 | break; |
2699 | } |
2700 | case Intrinsic::matrix_multiply: { |
2701 | // Optimize negation in matrix multiplication. |
2702 | |
2703 | // -A * -B -> A * B |
2704 | Value *A, *B; |
2705 | if (match(V: II->getArgOperand(i: 0), P: m_FNeg(X: m_Value(V&: A))) && |
2706 | match(V: II->getArgOperand(i: 1), P: m_FNeg(X: m_Value(V&: B)))) { |
2707 | replaceOperand(I&: *II, OpNum: 0, V: A); |
2708 | replaceOperand(I&: *II, OpNum: 1, V: B); |
2709 | return II; |
2710 | } |
2711 | |
2712 | Value *Op0 = II->getOperand(i_nocapture: 0); |
2713 | Value *Op1 = II->getOperand(i_nocapture: 1); |
2714 | Value *OpNotNeg, *NegatedOp; |
2715 | unsigned NegatedOpArg, OtherOpArg; |
2716 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: OpNotNeg)))) { |
2717 | NegatedOp = Op0; |
2718 | NegatedOpArg = 0; |
2719 | OtherOpArg = 1; |
2720 | } else if (match(V: Op1, P: m_FNeg(X: m_Value(V&: OpNotNeg)))) { |
2721 | NegatedOp = Op1; |
2722 | NegatedOpArg = 1; |
2723 | OtherOpArg = 0; |
2724 | } else |
2725 | // Multiplication doesn't have a negated operand. |
2726 | break; |
2727 | |
2728 | // Only optimize if the negated operand has only one use. |
2729 | if (!NegatedOp->hasOneUse()) |
2730 | break; |
2731 | |
2732 | Value *OtherOp = II->getOperand(i_nocapture: OtherOpArg); |
2733 | VectorType *RetTy = cast<VectorType>(Val: II->getType()); |
2734 | VectorType *NegatedOpTy = cast<VectorType>(Val: NegatedOp->getType()); |
2735 | VectorType *OtherOpTy = cast<VectorType>(Val: OtherOp->getType()); |
2736 | ElementCount NegatedCount = NegatedOpTy->getElementCount(); |
2737 | ElementCount OtherCount = OtherOpTy->getElementCount(); |
2738 | ElementCount RetCount = RetTy->getElementCount(); |
2739 | // (-A) * B -> A * (-B), if it is cheaper to negate B and vice versa. |
2740 | if (ElementCount::isKnownGT(LHS: NegatedCount, RHS: OtherCount) && |
2741 | ElementCount::isKnownLT(LHS: OtherCount, RHS: RetCount)) { |
2742 | Value *InverseOtherOp = Builder.CreateFNeg(V: OtherOp); |
2743 | replaceOperand(I&: *II, OpNum: NegatedOpArg, V: OpNotNeg); |
2744 | replaceOperand(I&: *II, OpNum: OtherOpArg, V: InverseOtherOp); |
2745 | return II; |
2746 | } |
2747 | // (-A) * B -> -(A * B), if it is cheaper to negate the result |
2748 | if (ElementCount::isKnownGT(LHS: NegatedCount, RHS: RetCount)) { |
2749 | SmallVector<Value *, 5> NewArgs(II->args()); |
2750 | NewArgs[NegatedOpArg] = OpNotNeg; |
2751 | Instruction *NewMul = |
2752 | Builder.CreateIntrinsic(RetTy: II->getType(), ID: IID, Args: NewArgs, FMFSource: II); |
2753 | return replaceInstUsesWith(I&: *II, V: Builder.CreateFNegFMF(V: NewMul, FMFSource: II)); |
2754 | } |
2755 | break; |
2756 | } |
2757 | case Intrinsic::fmuladd: { |
2758 | // Try to simplify the underlying FMul. |
2759 | if (Value *V = |
2760 | simplifyFMulInst(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1), |
2761 | FMF: II->getFastMathFlags(), Q: SQ.getWithInstruction(I: II))) |
2762 | return BinaryOperator::CreateFAddFMF(V1: V, V2: II->getArgOperand(i: 2), |
2763 | FMF: II->getFastMathFlags()); |
2764 | |
2765 | [[fallthrough]]; |
2766 | } |
2767 | case Intrinsic::fma: { |
2768 | // fma fneg(x), fneg(y), z -> fma x, y, z |
2769 | Value *Src0 = II->getArgOperand(i: 0); |
2770 | Value *Src1 = II->getArgOperand(i: 1); |
2771 | Value *Src2 = II->getArgOperand(i: 2); |
2772 | Value *X, *Y; |
2773 | if (match(V: Src0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Src1, P: m_FNeg(X: m_Value(V&: Y)))) { |
2774 | replaceOperand(I&: *II, OpNum: 0, V: X); |
2775 | replaceOperand(I&: *II, OpNum: 1, V: Y); |
2776 | return II; |
2777 | } |
2778 | |
2779 | // fma fabs(x), fabs(x), z -> fma x, x, z |
2780 | if (match(V: Src0, P: m_FAbs(Op0: m_Value(V&: X))) && |
2781 | match(V: Src1, P: m_FAbs(Op0: m_Specific(V: X)))) { |
2782 | replaceOperand(I&: *II, OpNum: 0, V: X); |
2783 | replaceOperand(I&: *II, OpNum: 1, V: X); |
2784 | return II; |
2785 | } |
2786 | |
2787 | // Try to simplify the underlying FMul. We can only apply simplifications |
2788 | // that do not require rounding. |
2789 | if (Value *V = simplifyFMAFMul(LHS: Src0, RHS: Src1, FMF: II->getFastMathFlags(), |
2790 | Q: SQ.getWithInstruction(I: II))) |
2791 | return BinaryOperator::CreateFAddFMF(V1: V, V2: Src2, FMF: II->getFastMathFlags()); |
2792 | |
2793 | // fma x, y, 0 -> fmul x, y |
2794 | // This is always valid for -0.0, but requires nsz for +0.0 as |
2795 | // -0.0 + 0.0 = 0.0, which would not be the same as the fmul on its own. |
2796 | if (match(V: Src2, P: m_NegZeroFP()) || |
2797 | (match(V: Src2, P: m_PosZeroFP()) && II->getFastMathFlags().noSignedZeros())) |
2798 | return BinaryOperator::CreateFMulFMF(V1: Src0, V2: Src1, FMFSource: II); |
2799 | |
2800 | // fma x, -1.0, y -> fsub y, x |
2801 | if (match(V: Src1, P: m_SpecificFP(V: -1.0))) |
2802 | return BinaryOperator::CreateFSubFMF(V1: Src2, V2: Src0, FMFSource: II); |
2803 | |
2804 | break; |
2805 | } |
2806 | case Intrinsic::copysign: { |
2807 | Value *Mag = II->getArgOperand(i: 0), *Sign = II->getArgOperand(i: 1); |
2808 | if (std::optional<bool> KnownSignBit = computeKnownFPSignBit( |
2809 | V: Sign, SQ: getSimplifyQuery().getWithInstruction(I: II))) { |
2810 | if (*KnownSignBit) { |
2811 | // If we know that the sign argument is negative, reduce to FNABS: |
2812 | // copysign Mag, -Sign --> fneg (fabs Mag) |
2813 | Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Mag, FMFSource: II); |
2814 | return replaceInstUsesWith(I&: *II, V: Builder.CreateFNegFMF(V: Fabs, FMFSource: II)); |
2815 | } |
2816 | |
2817 | // If we know that the sign argument is positive, reduce to FABS: |
2818 | // copysign Mag, +Sign --> fabs Mag |
2819 | Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Mag, FMFSource: II); |
2820 | return replaceInstUsesWith(I&: *II, V: Fabs); |
2821 | } |
2822 | |
2823 | // Propagate sign argument through nested calls: |
2824 | // copysign Mag, (copysign ?, X) --> copysign Mag, X |
2825 | Value *X; |
2826 | if (match(V: Sign, P: m_Intrinsic<Intrinsic::copysign>(Op0: m_Value(), Op1: m_Value(V&: X)))) { |
2827 | Value *CopySign = |
2828 | Builder.CreateCopySign(LHS: Mag, RHS: X, FMFSource: FMFSource::intersect(A: II, B: Sign)); |
2829 | return replaceInstUsesWith(I&: *II, V: CopySign); |
2830 | } |
2831 | |
2832 | // Clear sign-bit of constant magnitude: |
2833 | // copysign -MagC, X --> copysign MagC, X |
2834 | // TODO: Support constant folding for fabs |
2835 | const APFloat *MagC; |
2836 | if (match(V: Mag, P: m_APFloat(Res&: MagC)) && MagC->isNegative()) { |
2837 | APFloat PosMagC = *MagC; |
2838 | PosMagC.clearSign(); |
2839 | return replaceOperand(I&: *II, OpNum: 0, V: ConstantFP::get(Ty: Mag->getType(), V: PosMagC)); |
2840 | } |
2841 | |
2842 | // Peek through changes of magnitude's sign-bit. This call rewrites those: |
2843 | // copysign (fabs X), Sign --> copysign X, Sign |
2844 | // copysign (fneg X), Sign --> copysign X, Sign |
2845 | if (match(V: Mag, P: m_FAbs(Op0: m_Value(V&: X))) || match(V: Mag, P: m_FNeg(X: m_Value(V&: X)))) |
2846 | return replaceOperand(I&: *II, OpNum: 0, V: X); |
2847 | |
2848 | break; |
2849 | } |
2850 | case Intrinsic::fabs: { |
2851 | Value *Cond, *TVal, *FVal; |
2852 | Value *Arg = II->getArgOperand(i: 0); |
2853 | Value *X; |
2854 | // fabs (-X) --> fabs (X) |
2855 | if (match(V: Arg, P: m_FNeg(X: m_Value(V&: X)))) { |
2856 | CallInst *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: II); |
2857 | return replaceInstUsesWith(I&: CI, V: Fabs); |
2858 | } |
2859 | |
2860 | if (match(V: Arg, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))) { |
2861 | // fabs (select Cond, TrueC, FalseC) --> select Cond, AbsT, AbsF |
2862 | if (Arg->hasOneUse() ? (isa<Constant>(Val: TVal) || isa<Constant>(Val: FVal)) |
2863 | : (isa<Constant>(Val: TVal) && isa<Constant>(Val: FVal))) { |
2864 | CallInst *AbsT = Builder.CreateCall(Callee: II->getCalledFunction(), Args: {TVal}); |
2865 | CallInst *AbsF = Builder.CreateCall(Callee: II->getCalledFunction(), Args: {FVal}); |
2866 | SelectInst *SI = SelectInst::Create(C: Cond, S1: AbsT, S2: AbsF); |
2867 | FastMathFlags FMF1 = II->getFastMathFlags(); |
2868 | FastMathFlags FMF2 = cast<SelectInst>(Val: Arg)->getFastMathFlags(); |
2869 | FMF2.setNoSignedZeros(false); |
2870 | SI->setFastMathFlags(FMF1 | FMF2); |
2871 | return SI; |
2872 | } |
2873 | // fabs (select Cond, -FVal, FVal) --> fabs FVal |
2874 | if (match(V: TVal, P: m_FNeg(X: m_Specific(V: FVal)))) |
2875 | return replaceOperand(I&: *II, OpNum: 0, V: FVal); |
2876 | // fabs (select Cond, TVal, -TVal) --> fabs TVal |
2877 | if (match(V: FVal, P: m_FNeg(X: m_Specific(V: TVal)))) |
2878 | return replaceOperand(I&: *II, OpNum: 0, V: TVal); |
2879 | } |
2880 | |
2881 | Value *Magnitude, *Sign; |
2882 | if (match(V: II->getArgOperand(i: 0), |
2883 | P: m_CopySign(Op0: m_Value(V&: Magnitude), Op1: m_Value(V&: Sign)))) { |
2884 | // fabs (copysign x, y) -> (fabs x) |
2885 | CallInst *AbsSign = |
2886 | Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Magnitude, FMFSource: II); |
2887 | return replaceInstUsesWith(I&: *II, V: AbsSign); |
2888 | } |
2889 | |
2890 | [[fallthrough]]; |
2891 | } |
2892 | case Intrinsic::ceil: |
2893 | case Intrinsic::floor: |
2894 | case Intrinsic::round: |
2895 | case Intrinsic::roundeven: |
2896 | case Intrinsic::nearbyint: |
2897 | case Intrinsic::rint: |
2898 | case Intrinsic::trunc: { |
2899 | Value *ExtSrc; |
2900 | if (match(V: II->getArgOperand(i: 0), P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: ExtSrc))))) { |
2901 | // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x) |
2902 | Value *NarrowII = Builder.CreateUnaryIntrinsic(ID: IID, V: ExtSrc, FMFSource: II); |
2903 | return new FPExtInst(NarrowII, II->getType()); |
2904 | } |
2905 | break; |
2906 | } |
2907 | case Intrinsic::cos: |
2908 | case Intrinsic::amdgcn_cos: { |
2909 | Value *X, *Sign; |
2910 | Value *Src = II->getArgOperand(i: 0); |
2911 | if (match(V: Src, P: m_FNeg(X: m_Value(V&: X))) || match(V: Src, P: m_FAbs(Op0: m_Value(V&: X))) || |
2912 | match(V: Src, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value(V&: Sign)))) { |
2913 | // cos(-x) --> cos(x) |
2914 | // cos(fabs(x)) --> cos(x) |
2915 | // cos(copysign(x, y)) --> cos(x) |
2916 | return replaceOperand(I&: *II, OpNum: 0, V: X); |
2917 | } |
2918 | break; |
2919 | } |
2920 | case Intrinsic::sin: |
2921 | case Intrinsic::amdgcn_sin: { |
2922 | Value *X; |
2923 | if (match(V: II->getArgOperand(i: 0), P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) { |
2924 | // sin(-x) --> -sin(x) |
2925 | Value *NewSin = Builder.CreateUnaryIntrinsic(ID: IID, V: X, FMFSource: II); |
2926 | return UnaryOperator::CreateFNegFMF(Op: NewSin, FMFSource: II); |
2927 | } |
2928 | break; |
2929 | } |
2930 | case Intrinsic::ldexp: { |
2931 | // ldexp(ldexp(x, a), b) -> ldexp(x, a + b) |
2932 | // |
2933 | // The danger is if the first ldexp would overflow to infinity or underflow |
2934 | // to zero, but the combined exponent avoids it. We ignore this with |
2935 | // reassoc. |
2936 | // |
2937 | // It's also safe to fold if we know both exponents are >= 0 or <= 0 since |
2938 | // it would just double down on the overflow/underflow which would occur |
2939 | // anyway. |
2940 | // |
2941 | // TODO: Could do better if we had range tracking for the input value |
2942 | // exponent. Also could broaden sign check to cover == 0 case. |
2943 | Value *Src = II->getArgOperand(i: 0); |
2944 | Value *Exp = II->getArgOperand(i: 1); |
2945 | Value *InnerSrc; |
2946 | Value *InnerExp; |
2947 | if (match(V: Src, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ldexp>( |
2948 | Op0: m_Value(V&: InnerSrc), Op1: m_Value(V&: InnerExp)))) && |
2949 | Exp->getType() == InnerExp->getType()) { |
2950 | FastMathFlags FMF = II->getFastMathFlags(); |
2951 | FastMathFlags InnerFlags = cast<FPMathOperator>(Val: Src)->getFastMathFlags(); |
2952 | |
2953 | if ((FMF.allowReassoc() && InnerFlags.allowReassoc()) || |
2954 | signBitMustBeTheSame(Op0: Exp, Op1: InnerExp, SQ: SQ.getWithInstruction(I: II))) { |
2955 | // TODO: Add nsw/nuw probably safe if integer type exceeds exponent |
2956 | // width. |
2957 | Value *NewExp = Builder.CreateAdd(LHS: InnerExp, RHS: Exp); |
2958 | II->setArgOperand(i: 1, v: NewExp); |
2959 | II->setFastMathFlags(InnerFlags); // Or the inner flags. |
2960 | return replaceOperand(I&: *II, OpNum: 0, V: InnerSrc); |
2961 | } |
2962 | } |
2963 | |
2964 | // ldexp(x, zext(i1 y)) -> fmul x, (select y, 2.0, 1.0) |
2965 | // ldexp(x, sext(i1 y)) -> fmul x, (select y, 0.5, 1.0) |
2966 | Value *ExtSrc; |
2967 | if (match(V: Exp, P: m_ZExt(Op: m_Value(V&: ExtSrc))) && |
2968 | ExtSrc->getType()->getScalarSizeInBits() == 1) { |
2969 | Value *Select = |
2970 | Builder.CreateSelect(C: ExtSrc, True: ConstantFP::get(Ty: II->getType(), V: 2.0), |
2971 | False: ConstantFP::get(Ty: II->getType(), V: 1.0)); |
2972 | return BinaryOperator::CreateFMulFMF(V1: Src, V2: Select, FMFSource: II); |
2973 | } |
2974 | if (match(V: Exp, P: m_SExt(Op: m_Value(V&: ExtSrc))) && |
2975 | ExtSrc->getType()->getScalarSizeInBits() == 1) { |
2976 | Value *Select = |
2977 | Builder.CreateSelect(C: ExtSrc, True: ConstantFP::get(Ty: II->getType(), V: 0.5), |
2978 | False: ConstantFP::get(Ty: II->getType(), V: 1.0)); |
2979 | return BinaryOperator::CreateFMulFMF(V1: Src, V2: Select, FMFSource: II); |
2980 | } |
2981 | |
2982 | // ldexp(x, c ? exp : 0) -> c ? ldexp(x, exp) : x |
2983 | // ldexp(x, c ? 0 : exp) -> c ? x : ldexp(x, exp) |
2984 | /// |
2985 | // TODO: If we cared, should insert a canonicalize for x |
2986 | Value *SelectCond, *SelectLHS, *SelectRHS; |
2987 | if (match(V: II->getArgOperand(i: 1), |
2988 | P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: SelectCond), L: m_Value(V&: SelectLHS), |
2989 | R: m_Value(V&: SelectRHS))))) { |
2990 | Value *NewLdexp = nullptr; |
2991 | Value *Select = nullptr; |
2992 | if (match(V: SelectRHS, P: m_ZeroInt())) { |
2993 | NewLdexp = Builder.CreateLdexp(Src, Exp: SelectLHS, FMFSource: II); |
2994 | Select = Builder.CreateSelect(C: SelectCond, True: NewLdexp, False: Src); |
2995 | } else if (match(V: SelectLHS, P: m_ZeroInt())) { |
2996 | NewLdexp = Builder.CreateLdexp(Src, Exp: SelectRHS, FMFSource: II); |
2997 | Select = Builder.CreateSelect(C: SelectCond, True: Src, False: NewLdexp); |
2998 | } |
2999 | |
3000 | if (NewLdexp) { |
3001 | Select->takeName(V: II); |
3002 | return replaceInstUsesWith(I&: *II, V: Select); |
3003 | } |
3004 | } |
3005 | |
3006 | break; |
3007 | } |
3008 | case Intrinsic::ptrauth_auth: |
3009 | case Intrinsic::ptrauth_resign: { |
3010 | // (sign|resign) + (auth|resign) can be folded by omitting the middle |
3011 | // sign+auth component if the key and discriminator match. |
3012 | bool NeedSign = II->getIntrinsicID() == Intrinsic::ptrauth_resign; |
3013 | Value *Ptr = II->getArgOperand(i: 0); |
3014 | Value *Key = II->getArgOperand(i: 1); |
3015 | Value *Disc = II->getArgOperand(i: 2); |
3016 | |
3017 | // AuthKey will be the key we need to end up authenticating against in |
3018 | // whatever we replace this sequence with. |
3019 | Value *AuthKey = nullptr, *AuthDisc = nullptr, *BasePtr; |
3020 | if (const auto *CI = dyn_cast<CallBase>(Val: Ptr)) { |
3021 | BasePtr = CI->getArgOperand(i: 0); |
3022 | if (CI->getIntrinsicID() == Intrinsic::ptrauth_sign) { |
3023 | if (CI->getArgOperand(i: 1) != Key || CI->getArgOperand(i: 2) != Disc) |
3024 | break; |
3025 | } else if (CI->getIntrinsicID() == Intrinsic::ptrauth_resign) { |
3026 | if (CI->getArgOperand(i: 3) != Key || CI->getArgOperand(i: 4) != Disc) |
3027 | break; |
3028 | AuthKey = CI->getArgOperand(i: 1); |
3029 | AuthDisc = CI->getArgOperand(i: 2); |
3030 | } else |
3031 | break; |
3032 | } else if (const auto *PtrToInt = dyn_cast<PtrToIntOperator>(Val: Ptr)) { |
3033 | // ptrauth constants are equivalent to a call to @llvm.ptrauth.sign for |
3034 | // our purposes, so check for that too. |
3035 | const auto *CPA = dyn_cast<ConstantPtrAuth>(Val: PtrToInt->getOperand(i_nocapture: 0)); |
3036 | if (!CPA || !CPA->isKnownCompatibleWith(Key, Discriminator: Disc, DL)) |
3037 | break; |
3038 | |
3039 | // resign(ptrauth(p,ks,ds),ks,ds,kr,dr) -> ptrauth(p,kr,dr) |
3040 | if (NeedSign && isa<ConstantInt>(Val: II->getArgOperand(i: 4))) { |
3041 | auto *SignKey = cast<ConstantInt>(Val: II->getArgOperand(i: 3)); |
3042 | auto *SignDisc = cast<ConstantInt>(Val: II->getArgOperand(i: 4)); |
3043 | auto *SignAddrDisc = ConstantPointerNull::get(T: Builder.getPtrTy()); |
3044 | auto *NewCPA = ConstantPtrAuth::get(Ptr: CPA->getPointer(), Key: SignKey, |
3045 | Disc: SignDisc, AddrDisc: SignAddrDisc); |
3046 | replaceInstUsesWith( |
3047 | I&: *II, V: ConstantExpr::getPointerCast(C: NewCPA, Ty: II->getType())); |
3048 | return eraseInstFromFunction(I&: *II); |
3049 | } |
3050 | |
3051 | // auth(ptrauth(p,k,d),k,d) -> p |
3052 | BasePtr = Builder.CreatePtrToInt(V: CPA->getPointer(), DestTy: II->getType()); |
3053 | } else |
3054 | break; |
3055 | |
3056 | unsigned NewIntrin; |
3057 | if (AuthKey && NeedSign) { |
3058 | // resign(0,1) + resign(1,2) = resign(0, 2) |
3059 | NewIntrin = Intrinsic::ptrauth_resign; |
3060 | } else if (AuthKey) { |
3061 | // resign(0,1) + auth(1) = auth(0) |
3062 | NewIntrin = Intrinsic::ptrauth_auth; |
3063 | } else if (NeedSign) { |
3064 | // sign(0) + resign(0, 1) = sign(1) |
3065 | NewIntrin = Intrinsic::ptrauth_sign; |
3066 | } else { |
3067 | // sign(0) + auth(0) = nop |
3068 | replaceInstUsesWith(I&: *II, V: BasePtr); |
3069 | return eraseInstFromFunction(I&: *II); |
3070 | } |
3071 | |
3072 | SmallVector<Value *, 4> CallArgs; |
3073 | CallArgs.push_back(Elt: BasePtr); |
3074 | if (AuthKey) { |
3075 | CallArgs.push_back(Elt: AuthKey); |
3076 | CallArgs.push_back(Elt: AuthDisc); |
3077 | } |
3078 | |
3079 | if (NeedSign) { |
3080 | CallArgs.push_back(Elt: II->getArgOperand(i: 3)); |
3081 | CallArgs.push_back(Elt: II->getArgOperand(i: 4)); |
3082 | } |
3083 | |
3084 | Function *NewFn = |
3085 | Intrinsic::getOrInsertDeclaration(M: II->getModule(), id: NewIntrin); |
3086 | return CallInst::Create(Func: NewFn, Args: CallArgs); |
3087 | } |
3088 | case Intrinsic::arm_neon_vtbl1: |
3089 | case Intrinsic::aarch64_neon_tbl1: |
3090 | if (Value *V = simplifyNeonTbl1(II: *II, Builder)) |
3091 | return replaceInstUsesWith(I&: *II, V); |
3092 | break; |
3093 | |
3094 | case Intrinsic::arm_neon_vmulls: |
3095 | case Intrinsic::arm_neon_vmullu: |
3096 | case Intrinsic::aarch64_neon_smull: |
3097 | case Intrinsic::aarch64_neon_umull: { |
3098 | Value *Arg0 = II->getArgOperand(i: 0); |
3099 | Value *Arg1 = II->getArgOperand(i: 1); |
3100 | |
3101 | // Handle mul by zero first: |
3102 | if (isa<ConstantAggregateZero>(Val: Arg0) || isa<ConstantAggregateZero>(Val: Arg1)) { |
3103 | return replaceInstUsesWith(I&: CI, V: ConstantAggregateZero::get(Ty: II->getType())); |
3104 | } |
3105 | |
3106 | // Check for constant LHS & RHS - in this case we just simplify. |
3107 | bool Zext = (IID == Intrinsic::arm_neon_vmullu || |
3108 | IID == Intrinsic::aarch64_neon_umull); |
3109 | VectorType *NewVT = cast<VectorType>(Val: II->getType()); |
3110 | if (Constant *CV0 = dyn_cast<Constant>(Val: Arg0)) { |
3111 | if (Constant *CV1 = dyn_cast<Constant>(Val: Arg1)) { |
3112 | Value *V0 = Builder.CreateIntCast(V: CV0, DestTy: NewVT, /*isSigned=*/!Zext); |
3113 | Value *V1 = Builder.CreateIntCast(V: CV1, DestTy: NewVT, /*isSigned=*/!Zext); |
3114 | return replaceInstUsesWith(I&: CI, V: Builder.CreateMul(LHS: V0, RHS: V1)); |
3115 | } |
3116 | |
3117 | // Couldn't simplify - canonicalize constant to the RHS. |
3118 | std::swap(a&: Arg0, b&: Arg1); |
3119 | } |
3120 | |
3121 | // Handle mul by one: |
3122 | if (Constant *CV1 = dyn_cast<Constant>(Val: Arg1)) |
3123 | if (ConstantInt *Splat = |
3124 | dyn_cast_or_null<ConstantInt>(Val: CV1->getSplatValue())) |
3125 | if (Splat->isOne()) |
3126 | return CastInst::CreateIntegerCast(S: Arg0, Ty: II->getType(), |
3127 | /*isSigned=*/!Zext); |
3128 | |
3129 | break; |
3130 | } |
3131 | case Intrinsic::arm_neon_aesd: |
3132 | case Intrinsic::arm_neon_aese: |
3133 | case Intrinsic::aarch64_crypto_aesd: |
3134 | case Intrinsic::aarch64_crypto_aese: |
3135 | case Intrinsic::aarch64_sve_aesd: |
3136 | case Intrinsic::aarch64_sve_aese: { |
3137 | Value *DataArg = II->getArgOperand(i: 0); |
3138 | Value *KeyArg = II->getArgOperand(i: 1); |
3139 | |
3140 | // Accept zero on either operand. |
3141 | if (!match(V: KeyArg, P: m_ZeroInt())) |
3142 | std::swap(a&: KeyArg, b&: DataArg); |
3143 | |
3144 | // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR |
3145 | Value *Data, *Key; |
3146 | if (match(V: KeyArg, P: m_ZeroInt()) && |
3147 | match(V: DataArg, P: m_Xor(L: m_Value(V&: Data), R: m_Value(V&: Key)))) { |
3148 | replaceOperand(I&: *II, OpNum: 0, V: Data); |
3149 | replaceOperand(I&: *II, OpNum: 1, V: Key); |
3150 | return II; |
3151 | } |
3152 | break; |
3153 | } |
3154 | case Intrinsic::hexagon_V6_vandvrt: |
3155 | case Intrinsic::hexagon_V6_vandvrt_128B: { |
3156 | // Simplify Q -> V -> Q conversion. |
3157 | if (auto Op0 = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0))) { |
3158 | Intrinsic::ID ID0 = Op0->getIntrinsicID(); |
3159 | if (ID0 != Intrinsic::hexagon_V6_vandqrt && |
3160 | ID0 != Intrinsic::hexagon_V6_vandqrt_128B) |
3161 | break; |
3162 | Value *Bytes = Op0->getArgOperand(i: 1), *Mask = II->getArgOperand(i: 1); |
3163 | uint64_t Bytes1 = computeKnownBits(V: Bytes, CxtI: Op0).One.getZExtValue(); |
3164 | uint64_t Mask1 = computeKnownBits(V: Mask, CxtI: II).One.getZExtValue(); |
3165 | // Check if every byte has common bits in Bytes and Mask. |
3166 | uint64_t C = Bytes1 & Mask1; |
3167 | if ((C & 0xFF) && (C & 0xFF00) && (C & 0xFF0000) && (C & 0xFF000000)) |
3168 | return replaceInstUsesWith(I&: *II, V: Op0->getArgOperand(i: 0)); |
3169 | } |
3170 | break; |
3171 | } |
3172 | case Intrinsic::stackrestore: { |
3173 | enum class ClassifyResult { |
3174 | None, |
3175 | Alloca, |
3176 | StackRestore, |
3177 | CallWithSideEffects, |
3178 | }; |
3179 | auto Classify = [](const Instruction *I) { |
3180 | if (isa<AllocaInst>(Val: I)) |
3181 | return ClassifyResult::Alloca; |
3182 | |
3183 | if (auto *CI = dyn_cast<CallInst>(Val: I)) { |
3184 | if (auto *II = dyn_cast<IntrinsicInst>(Val: CI)) { |
3185 | if (II->getIntrinsicID() == Intrinsic::stackrestore) |
3186 | return ClassifyResult::StackRestore; |
3187 | |
3188 | if (II->mayHaveSideEffects()) |
3189 | return ClassifyResult::CallWithSideEffects; |
3190 | } else { |
3191 | // Consider all non-intrinsic calls to be side effects |
3192 | return ClassifyResult::CallWithSideEffects; |
3193 | } |
3194 | } |
3195 | |
3196 | return ClassifyResult::None; |
3197 | }; |
3198 | |
3199 | // If the stacksave and the stackrestore are in the same BB, and there is |
3200 | // no intervening call, alloca, or stackrestore of a different stacksave, |
3201 | // remove the restore. This can happen when variable allocas are DCE'd. |
3202 | if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0))) { |
3203 | if (SS->getIntrinsicID() == Intrinsic::stacksave && |
3204 | SS->getParent() == II->getParent()) { |
3205 | BasicBlock::iterator BI(SS); |
3206 | bool CannotRemove = false; |
3207 | for (++BI; &*BI != II; ++BI) { |
3208 | switch (Classify(&*BI)) { |
3209 | case ClassifyResult::None: |
3210 | // So far so good, look at next instructions. |
3211 | break; |
3212 | |
3213 | case ClassifyResult::StackRestore: |
3214 | // If we found an intervening stackrestore for a different |
3215 | // stacksave, we can't remove the stackrestore. Otherwise, continue. |
3216 | if (cast<IntrinsicInst>(Val&: *BI).getArgOperand(i: 0) != SS) |
3217 | CannotRemove = true; |
3218 | break; |
3219 | |
3220 | case ClassifyResult::Alloca: |
3221 | case ClassifyResult::CallWithSideEffects: |
3222 | // If we found an alloca, a non-intrinsic call, or an intrinsic |
3223 | // call with side effects, we can't remove the stackrestore. |
3224 | CannotRemove = true; |
3225 | break; |
3226 | } |
3227 | if (CannotRemove) |
3228 | break; |
3229 | } |
3230 | |
3231 | if (!CannotRemove) |
3232 | return eraseInstFromFunction(I&: CI); |
3233 | } |
3234 | } |
3235 | |
3236 | // Scan down this block to see if there is another stack restore in the |
3237 | // same block without an intervening call/alloca. |
3238 | BasicBlock::iterator BI(II); |
3239 | Instruction *TI = II->getParent()->getTerminator(); |
3240 | bool CannotRemove = false; |
3241 | for (++BI; &*BI != TI; ++BI) { |
3242 | switch (Classify(&*BI)) { |
3243 | case ClassifyResult::None: |
3244 | // So far so good, look at next instructions. |
3245 | break; |
3246 | |
3247 | case ClassifyResult::StackRestore: |
3248 | // If there is a stackrestore below this one, remove this one. |
3249 | return eraseInstFromFunction(I&: CI); |
3250 | |
3251 | case ClassifyResult::Alloca: |
3252 | case ClassifyResult::CallWithSideEffects: |
3253 | // If we found an alloca, a non-intrinsic call, or an intrinsic call |
3254 | // with side effects (such as llvm.stacksave and llvm.read_register), |
3255 | // we can't remove the stack restore. |
3256 | CannotRemove = true; |
3257 | break; |
3258 | } |
3259 | if (CannotRemove) |
3260 | break; |
3261 | } |
3262 | |
3263 | // If the stack restore is in a return, resume, or unwind block and if there |
3264 | // are no allocas or calls between the restore and the return, nuke the |
3265 | // restore. |
3266 | if (!CannotRemove && (isa<ReturnInst>(Val: TI) || isa<ResumeInst>(Val: TI))) |
3267 | return eraseInstFromFunction(I&: CI); |
3268 | break; |
3269 | } |
3270 | case Intrinsic::lifetime_end: |
3271 | // Asan needs to poison memory to detect invalid access which is possible |
3272 | // even for empty lifetime range. |
3273 | if (II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
3274 | II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeMemory) || |
3275 | II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeHWAddress)) |
3276 | break; |
3277 | |
3278 | if (removeTriviallyEmptyRange(EndI&: *II, IC&: *this, IsStart: [](const IntrinsicInst &I) { |
3279 | return I.getIntrinsicID() == Intrinsic::lifetime_start; |
3280 | })) |
3281 | return nullptr; |
3282 | break; |
3283 | case Intrinsic::assume: { |
3284 | Value *IIOperand = II->getArgOperand(i: 0); |
3285 | SmallVector<OperandBundleDef, 4> OpBundles; |
3286 | II->getOperandBundlesAsDefs(Defs&: OpBundles); |
3287 | |
3288 | /// This will remove the boolean Condition from the assume given as |
3289 | /// argument and remove the assume if it becomes useless. |
3290 | /// always returns nullptr for use as a return values. |
3291 | auto RemoveConditionFromAssume = [&](Instruction *Assume) -> Instruction * { |
3292 | assert(isa<AssumeInst>(Assume)); |
3293 | if (isAssumeWithEmptyBundle(Assume: *cast<AssumeInst>(Val: II))) |
3294 | return eraseInstFromFunction(I&: CI); |
3295 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: ConstantInt::getTrue(Context&: II->getContext())); |
3296 | return nullptr; |
3297 | }; |
3298 | // Remove an assume if it is followed by an identical assume. |
3299 | // TODO: Do we need this? Unless there are conflicting assumptions, the |
3300 | // computeKnownBits(IIOperand) below here eliminates redundant assumes. |
3301 | Instruction *Next = II->getNextNonDebugInstruction(); |
3302 | if (match(V: Next, P: m_Intrinsic<Intrinsic::assume>(Op0: m_Specific(V: IIOperand)))) |
3303 | return RemoveConditionFromAssume(Next); |
3304 | |
3305 | // Canonicalize assume(a && b) -> assume(a); assume(b); |
3306 | // Note: New assumption intrinsics created here are registered by |
3307 | // the InstCombineIRInserter object. |
3308 | FunctionType *AssumeIntrinsicTy = II->getFunctionType(); |
3309 | Value *AssumeIntrinsic = II->getCalledOperand(); |
3310 | Value *A, *B; |
3311 | if (match(V: IIOperand, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
3312 | Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, Args: A, OpBundles, |
3313 | Name: II->getName()); |
3314 | Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, Args: B, Name: II->getName()); |
3315 | return eraseInstFromFunction(I&: *II); |
3316 | } |
3317 | // assume(!(a || b)) -> assume(!a); assume(!b); |
3318 | if (match(V: IIOperand, P: m_Not(V: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))))) { |
3319 | Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, |
3320 | Args: Builder.CreateNot(V: A), OpBundles, Name: II->getName()); |
3321 | Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, |
3322 | Args: Builder.CreateNot(V: B), Name: II->getName()); |
3323 | return eraseInstFromFunction(I&: *II); |
3324 | } |
3325 | |
3326 | // assume( (load addr) != null ) -> add 'nonnull' metadata to load |
3327 | // (if assume is valid at the load) |
3328 | Instruction *LHS; |
3329 | if (match(V: IIOperand, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: m_Instruction(I&: LHS), |
3330 | R: m_Zero())) && |
3331 | LHS->getOpcode() == Instruction::Load && |
3332 | LHS->getType()->isPointerTy() && |
3333 | isValidAssumeForContext(I: II, CxtI: LHS, DT: &DT)) { |
3334 | MDNode *MD = MDNode::get(Context&: II->getContext(), MDs: {}); |
3335 | LHS->setMetadata(KindID: LLVMContext::MD_nonnull, Node: MD); |
3336 | LHS->setMetadata(KindID: LLVMContext::MD_noundef, Node: MD); |
3337 | return RemoveConditionFromAssume(II); |
3338 | |
3339 | // TODO: apply nonnull return attributes to calls and invokes |
3340 | // TODO: apply range metadata for range check patterns? |
3341 | } |
3342 | |
3343 | // Separate storage assumptions apply to the underlying allocations, not any |
3344 | // particular pointer within them. When evaluating the hints for AA purposes |
3345 | // we getUnderlyingObject them; by precomputing the answers here we can |
3346 | // avoid having to do so repeatedly there. |
3347 | for (unsigned Idx = 0; Idx < II->getNumOperandBundles(); Idx++) { |
3348 | OperandBundleUse OBU = II->getOperandBundleAt(Index: Idx); |
3349 | if (OBU.getTagName() == "separate_storage" ) { |
3350 | assert(OBU.Inputs.size() == 2); |
3351 | auto MaybeSimplifyHint = [&](const Use &U) { |
3352 | Value *Hint = U.get(); |
3353 | // Not having a limit is safe because InstCombine removes unreachable |
3354 | // code. |
3355 | Value *UnderlyingObject = getUnderlyingObject(V: Hint, /*MaxLookup*/ 0); |
3356 | if (Hint != UnderlyingObject) |
3357 | replaceUse(U&: const_cast<Use &>(U), NewValue: UnderlyingObject); |
3358 | }; |
3359 | MaybeSimplifyHint(OBU.Inputs[0]); |
3360 | MaybeSimplifyHint(OBU.Inputs[1]); |
3361 | } |
3362 | } |
3363 | |
3364 | // Convert nonnull assume like: |
3365 | // %A = icmp ne i32* %PTR, null |
3366 | // call void @llvm.assume(i1 %A) |
3367 | // into |
3368 | // call void @llvm.assume(i1 true) [ "nonnull"(i32* %PTR) ] |
3369 | if (EnableKnowledgeRetention && |
3370 | match(V: IIOperand, |
3371 | P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: m_Value(V&: A), R: m_Zero())) && |
3372 | A->getType()->isPointerTy()) { |
3373 | if (auto *Replacement = buildAssumeFromKnowledge( |
3374 | Knowledge: {RetainedKnowledge{.AttrKind: Attribute::NonNull, .ArgValue: 0, .WasOn: A}}, CtxI: Next, AC: &AC, DT: &DT)) { |
3375 | |
3376 | Replacement->insertBefore(InsertPos: Next->getIterator()); |
3377 | AC.registerAssumption(CI: Replacement); |
3378 | return RemoveConditionFromAssume(II); |
3379 | } |
3380 | } |
3381 | |
3382 | // Convert alignment assume like: |
3383 | // %B = ptrtoint i32* %A to i64 |
3384 | // %C = and i64 %B, Constant |
3385 | // %D = icmp eq i64 %C, 0 |
3386 | // call void @llvm.assume(i1 %D) |
3387 | // into |
3388 | // call void @llvm.assume(i1 true) [ "align"(i32* [[A]], i64 Constant + 1)] |
3389 | uint64_t AlignMask = 1; |
3390 | if (EnableKnowledgeRetention && |
3391 | (match(V: IIOperand, P: m_Not(V: m_Trunc(Op: m_Value(V&: A)))) || |
3392 | match(V: IIOperand, |
3393 | P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, |
3394 | L: m_And(L: m_Value(V&: A), R: m_ConstantInt(V&: AlignMask)), |
3395 | R: m_Zero())))) { |
3396 | if (isPowerOf2_64(Value: AlignMask + 1)) { |
3397 | uint64_t Offset = 0; |
3398 | match(V: A, P: m_Add(L: m_Value(V&: A), R: m_ConstantInt(V&: Offset))); |
3399 | if (match(V: A, P: m_PtrToInt(Op: m_Value(V&: A)))) { |
3400 | /// Note: this doesn't preserve the offset information but merges |
3401 | /// offset and alignment. |
3402 | /// TODO: we can generate a GEP instead of merging the alignment with |
3403 | /// the offset. |
3404 | RetainedKnowledge RK{.AttrKind: Attribute::Alignment, |
3405 | .ArgValue: (unsigned)MinAlign(A: Offset, B: AlignMask + 1), .WasOn: A}; |
3406 | if (auto *Replacement = |
3407 | buildAssumeFromKnowledge(Knowledge: RK, CtxI: Next, AC: &AC, DT: &DT)) { |
3408 | |
3409 | Replacement->insertAfter(InsertPos: II->getIterator()); |
3410 | AC.registerAssumption(CI: Replacement); |
3411 | } |
3412 | return RemoveConditionFromAssume(II); |
3413 | } |
3414 | } |
3415 | } |
3416 | |
3417 | /// Canonicalize Knowledge in operand bundles. |
3418 | if (EnableKnowledgeRetention && II->hasOperandBundles()) { |
3419 | for (unsigned Idx = 0; Idx < II->getNumOperandBundles(); Idx++) { |
3420 | auto &BOI = II->bundle_op_info_begin()[Idx]; |
3421 | RetainedKnowledge RK = |
3422 | llvm::getKnowledgeFromBundle(Assume&: cast<AssumeInst>(Val&: *II), BOI); |
3423 | if (BOI.End - BOI.Begin > 2) |
3424 | continue; // Prevent reducing knowledge in an align with offset since |
3425 | // extracting a RetainedKnowledge from them looses offset |
3426 | // information |
3427 | RetainedKnowledge CanonRK = |
3428 | llvm::simplifyRetainedKnowledge(Assume: cast<AssumeInst>(Val: II), RK, |
3429 | AC: &getAssumptionCache(), |
3430 | DT: &getDominatorTree()); |
3431 | if (CanonRK == RK) |
3432 | continue; |
3433 | if (!CanonRK) { |
3434 | if (BOI.End - BOI.Begin > 0) { |
3435 | Worklist.pushValue(V: II->op_begin()[BOI.Begin]); |
3436 | Value::dropDroppableUse(U&: II->op_begin()[BOI.Begin]); |
3437 | } |
3438 | continue; |
3439 | } |
3440 | assert(RK.AttrKind == CanonRK.AttrKind); |
3441 | if (BOI.End - BOI.Begin > 0) |
3442 | II->op_begin()[BOI.Begin].set(CanonRK.WasOn); |
3443 | if (BOI.End - BOI.Begin > 1) |
3444 | II->op_begin()[BOI.Begin + 1].set(ConstantInt::get( |
3445 | Ty: Type::getInt64Ty(C&: II->getContext()), V: CanonRK.ArgValue)); |
3446 | if (RK.WasOn) |
3447 | Worklist.pushValue(V: RK.WasOn); |
3448 | return II; |
3449 | } |
3450 | } |
3451 | |
3452 | // If there is a dominating assume with the same condition as this one, |
3453 | // then this one is redundant, and should be removed. |
3454 | KnownBits Known(1); |
3455 | computeKnownBits(V: IIOperand, Known, CxtI: II); |
3456 | if (Known.isAllOnes() && isAssumeWithEmptyBundle(Assume: cast<AssumeInst>(Val&: *II))) |
3457 | return eraseInstFromFunction(I&: *II); |
3458 | |
3459 | // assume(false) is unreachable. |
3460 | if (match(V: IIOperand, P: m_CombineOr(L: m_Zero(), R: m_Undef()))) { |
3461 | CreateNonTerminatorUnreachable(InsertAt: II); |
3462 | return eraseInstFromFunction(I&: *II); |
3463 | } |
3464 | |
3465 | // Update the cache of affected values for this assumption (we might be |
3466 | // here because we just simplified the condition). |
3467 | AC.updateAffectedValues(CI: cast<AssumeInst>(Val: II)); |
3468 | break; |
3469 | } |
3470 | case Intrinsic::experimental_guard: { |
3471 | // Is this guard followed by another guard? We scan forward over a small |
3472 | // fixed window of instructions to handle common cases with conditions |
3473 | // computed between guards. |
3474 | Instruction *NextInst = II->getNextNonDebugInstruction(); |
3475 | for (unsigned i = 0; i < GuardWideningWindow; i++) { |
3476 | // Note: Using context-free form to avoid compile time blow up |
3477 | if (!isSafeToSpeculativelyExecute(I: NextInst)) |
3478 | break; |
3479 | NextInst = NextInst->getNextNonDebugInstruction(); |
3480 | } |
3481 | Value *NextCond = nullptr; |
3482 | if (match(V: NextInst, |
3483 | P: m_Intrinsic<Intrinsic::experimental_guard>(Op0: m_Value(V&: NextCond)))) { |
3484 | Value *CurrCond = II->getArgOperand(i: 0); |
3485 | |
3486 | // Remove a guard that it is immediately preceded by an identical guard. |
3487 | // Otherwise canonicalize guard(a); guard(b) -> guard(a & b). |
3488 | if (CurrCond != NextCond) { |
3489 | Instruction *MoveI = II->getNextNonDebugInstruction(); |
3490 | while (MoveI != NextInst) { |
3491 | auto *Temp = MoveI; |
3492 | MoveI = MoveI->getNextNonDebugInstruction(); |
3493 | Temp->moveBefore(InsertPos: II->getIterator()); |
3494 | } |
3495 | replaceOperand(I&: *II, OpNum: 0, V: Builder.CreateAnd(LHS: CurrCond, RHS: NextCond)); |
3496 | } |
3497 | eraseInstFromFunction(I&: *NextInst); |
3498 | return II; |
3499 | } |
3500 | break; |
3501 | } |
3502 | case Intrinsic::vector_insert: { |
3503 | Value *Vec = II->getArgOperand(i: 0); |
3504 | Value *SubVec = II->getArgOperand(i: 1); |
3505 | Value *Idx = II->getArgOperand(i: 2); |
3506 | auto *DstTy = dyn_cast<FixedVectorType>(Val: II->getType()); |
3507 | auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType()); |
3508 | auto *SubVecTy = dyn_cast<FixedVectorType>(Val: SubVec->getType()); |
3509 | |
3510 | // Only canonicalize if the destination vector, Vec, and SubVec are all |
3511 | // fixed vectors. |
3512 | if (DstTy && VecTy && SubVecTy) { |
3513 | unsigned DstNumElts = DstTy->getNumElements(); |
3514 | unsigned VecNumElts = VecTy->getNumElements(); |
3515 | unsigned SubVecNumElts = SubVecTy->getNumElements(); |
3516 | unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue(); |
3517 | |
3518 | // An insert that entirely overwrites Vec with SubVec is a nop. |
3519 | if (VecNumElts == SubVecNumElts) |
3520 | return replaceInstUsesWith(I&: CI, V: SubVec); |
3521 | |
3522 | // Widen SubVec into a vector of the same width as Vec, since |
3523 | // shufflevector requires the two input vectors to be the same width. |
3524 | // Elements beyond the bounds of SubVec within the widened vector are |
3525 | // undefined. |
3526 | SmallVector<int, 8> WidenMask; |
3527 | unsigned i; |
3528 | for (i = 0; i != SubVecNumElts; ++i) |
3529 | WidenMask.push_back(Elt: i); |
3530 | for (; i != VecNumElts; ++i) |
3531 | WidenMask.push_back(Elt: PoisonMaskElem); |
3532 | |
3533 | Value *WidenShuffle = Builder.CreateShuffleVector(V: SubVec, Mask: WidenMask); |
3534 | |
3535 | SmallVector<int, 8> Mask; |
3536 | for (unsigned i = 0; i != IdxN; ++i) |
3537 | Mask.push_back(Elt: i); |
3538 | for (unsigned i = DstNumElts; i != DstNumElts + SubVecNumElts; ++i) |
3539 | Mask.push_back(Elt: i); |
3540 | for (unsigned i = IdxN + SubVecNumElts; i != DstNumElts; ++i) |
3541 | Mask.push_back(Elt: i); |
3542 | |
3543 | Value *Shuffle = Builder.CreateShuffleVector(V1: Vec, V2: WidenShuffle, Mask); |
3544 | return replaceInstUsesWith(I&: CI, V: Shuffle); |
3545 | } |
3546 | break; |
3547 | } |
3548 | case Intrinsic::vector_extract: { |
3549 | Value *Vec = II->getArgOperand(i: 0); |
3550 | Value *Idx = II->getArgOperand(i: 1); |
3551 | |
3552 | Type *ReturnType = II->getType(); |
3553 | // (extract_vector (insert_vector InsertTuple, InsertValue, InsertIdx), |
3554 | // ExtractIdx) |
3555 | unsigned = cast<ConstantInt>(Val: Idx)->getZExtValue(); |
3556 | Value *InsertTuple, *InsertIdx, *InsertValue; |
3557 | if (match(V: Vec, P: m_Intrinsic<Intrinsic::vector_insert>(Op0: m_Value(V&: InsertTuple), |
3558 | Op1: m_Value(V&: InsertValue), |
3559 | Op2: m_Value(V&: InsertIdx))) && |
3560 | InsertValue->getType() == ReturnType) { |
3561 | unsigned Index = cast<ConstantInt>(Val: InsertIdx)->getZExtValue(); |
3562 | // Case where we get the same index right after setting it. |
3563 | // extract.vector(insert.vector(InsertTuple, InsertValue, Idx), Idx) --> |
3564 | // InsertValue |
3565 | if (ExtractIdx == Index) |
3566 | return replaceInstUsesWith(I&: CI, V: InsertValue); |
3567 | // If we are getting a different index than what was set in the |
3568 | // insert.vector intrinsic. We can just set the input tuple to the one up |
3569 | // in the chain. extract.vector(insert.vector(InsertTuple, InsertValue, |
3570 | // InsertIndex), ExtractIndex) |
3571 | // --> extract.vector(InsertTuple, ExtractIndex) |
3572 | else |
3573 | return replaceOperand(I&: CI, OpNum: 0, V: InsertTuple); |
3574 | } |
3575 | |
3576 | auto *DstTy = dyn_cast<VectorType>(Val: ReturnType); |
3577 | auto *VecTy = dyn_cast<VectorType>(Val: Vec->getType()); |
3578 | |
3579 | if (DstTy && VecTy) { |
3580 | auto DstEltCnt = DstTy->getElementCount(); |
3581 | auto VecEltCnt = VecTy->getElementCount(); |
3582 | unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue(); |
3583 | |
3584 | // Extracting the entirety of Vec is a nop. |
3585 | if (DstEltCnt == VecTy->getElementCount()) { |
3586 | replaceInstUsesWith(I&: CI, V: Vec); |
3587 | return eraseInstFromFunction(I&: CI); |
3588 | } |
3589 | |
3590 | // Only canonicalize to shufflevector if the destination vector and |
3591 | // Vec are fixed vectors. |
3592 | if (VecEltCnt.isScalable() || DstEltCnt.isScalable()) |
3593 | break; |
3594 | |
3595 | SmallVector<int, 8> Mask; |
3596 | for (unsigned i = 0; i != DstEltCnt.getKnownMinValue(); ++i) |
3597 | Mask.push_back(Elt: IdxN + i); |
3598 | |
3599 | Value *Shuffle = Builder.CreateShuffleVector(V: Vec, Mask); |
3600 | return replaceInstUsesWith(I&: CI, V: Shuffle); |
3601 | } |
3602 | break; |
3603 | } |
3604 | case Intrinsic::experimental_vp_reverse: { |
3605 | Value *X; |
3606 | Value *Vec = II->getArgOperand(i: 0); |
3607 | Value *Mask = II->getArgOperand(i: 1); |
3608 | if (!match(V: Mask, P: m_AllOnes())) |
3609 | break; |
3610 | Value *EVL = II->getArgOperand(i: 2); |
3611 | // TODO: Canonicalize experimental.vp.reverse after unop/binops? |
3612 | // rev(unop rev(X)) --> unop X |
3613 | if (match(V: Vec, |
3614 | P: m_OneUse(SubPattern: m_UnOp(X: m_Intrinsic<Intrinsic::experimental_vp_reverse>( |
3615 | Op0: m_Value(V&: X), Op1: m_AllOnes(), Op2: m_Specific(V: EVL)))))) { |
3616 | auto *OldUnOp = cast<UnaryOperator>(Val: Vec); |
3617 | auto *NewUnOp = UnaryOperator::CreateWithCopiedFlags( |
3618 | Opc: OldUnOp->getOpcode(), V: X, CopyO: OldUnOp, Name: OldUnOp->getName(), |
3619 | InsertBefore: II->getIterator()); |
3620 | return replaceInstUsesWith(I&: CI, V: NewUnOp); |
3621 | } |
3622 | break; |
3623 | } |
3624 | case Intrinsic::vector_reduce_or: |
3625 | case Intrinsic::vector_reduce_and: { |
3626 | // Canonicalize logical or/and reductions: |
3627 | // Or reduction for i1 is represented as: |
3628 | // %val = bitcast <ReduxWidth x i1> to iReduxWidth |
3629 | // %res = cmp ne iReduxWidth %val, 0 |
3630 | // And reduction for i1 is represented as: |
3631 | // %val = bitcast <ReduxWidth x i1> to iReduxWidth |
3632 | // %res = cmp eq iReduxWidth %val, 11111 |
3633 | Value *Arg = II->getArgOperand(i: 0); |
3634 | Value *Vect; |
3635 | |
3636 | if (Value *NewOp = |
3637 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3638 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3639 | return II; |
3640 | } |
3641 | |
3642 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3643 | if (auto *FTy = dyn_cast<FixedVectorType>(Val: Vect->getType())) |
3644 | if (FTy->getElementType() == Builder.getInt1Ty()) { |
3645 | Value *Res = Builder.CreateBitCast( |
3646 | V: Vect, DestTy: Builder.getIntNTy(N: FTy->getNumElements())); |
3647 | if (IID == Intrinsic::vector_reduce_and) { |
3648 | Res = Builder.CreateICmpEQ( |
3649 | LHS: Res, RHS: ConstantInt::getAllOnesValue(Ty: Res->getType())); |
3650 | } else { |
3651 | assert(IID == Intrinsic::vector_reduce_or && |
3652 | "Expected or reduction." ); |
3653 | Res = Builder.CreateIsNotNull(Arg: Res); |
3654 | } |
3655 | if (Arg != Vect) |
3656 | Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res, |
3657 | DestTy: II->getType()); |
3658 | return replaceInstUsesWith(I&: CI, V: Res); |
3659 | } |
3660 | } |
3661 | [[fallthrough]]; |
3662 | } |
3663 | case Intrinsic::vector_reduce_add: { |
3664 | if (IID == Intrinsic::vector_reduce_add) { |
3665 | // Convert vector_reduce_add(ZExt(<n x i1>)) to |
3666 | // ZExtOrTrunc(ctpop(bitcast <n x i1> to in)). |
3667 | // Convert vector_reduce_add(SExt(<n x i1>)) to |
3668 | // -ZExtOrTrunc(ctpop(bitcast <n x i1> to in)). |
3669 | // Convert vector_reduce_add(<n x i1>) to |
3670 | // Trunc(ctpop(bitcast <n x i1> to in)). |
3671 | Value *Arg = II->getArgOperand(i: 0); |
3672 | Value *Vect; |
3673 | |
3674 | if (Value *NewOp = |
3675 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3676 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3677 | return II; |
3678 | } |
3679 | |
3680 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3681 | if (auto *FTy = dyn_cast<FixedVectorType>(Val: Vect->getType())) |
3682 | if (FTy->getElementType() == Builder.getInt1Ty()) { |
3683 | Value *V = Builder.CreateBitCast( |
3684 | V: Vect, DestTy: Builder.getIntNTy(N: FTy->getNumElements())); |
3685 | Value *Res = Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V); |
3686 | if (Res->getType() != II->getType()) |
3687 | Res = Builder.CreateZExtOrTrunc(V: Res, DestTy: II->getType()); |
3688 | if (Arg != Vect && |
3689 | cast<Instruction>(Val: Arg)->getOpcode() == Instruction::SExt) |
3690 | Res = Builder.CreateNeg(V: Res); |
3691 | return replaceInstUsesWith(I&: CI, V: Res); |
3692 | } |
3693 | } |
3694 | } |
3695 | [[fallthrough]]; |
3696 | } |
3697 | case Intrinsic::vector_reduce_xor: { |
3698 | if (IID == Intrinsic::vector_reduce_xor) { |
3699 | // Exclusive disjunction reduction over the vector with |
3700 | // (potentially-extended) i1 element type is actually a |
3701 | // (potentially-extended) arithmetic `add` reduction over the original |
3702 | // non-extended value: |
3703 | // vector_reduce_xor(?ext(<n x i1>)) |
3704 | // --> |
3705 | // ?ext(vector_reduce_add(<n x i1>)) |
3706 | Value *Arg = II->getArgOperand(i: 0); |
3707 | Value *Vect; |
3708 | |
3709 | if (Value *NewOp = |
3710 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3711 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3712 | return II; |
3713 | } |
3714 | |
3715 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3716 | if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType())) |
3717 | if (VTy->getElementType() == Builder.getInt1Ty()) { |
3718 | Value *Res = Builder.CreateAddReduce(Src: Vect); |
3719 | if (Arg != Vect) |
3720 | Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res, |
3721 | DestTy: II->getType()); |
3722 | return replaceInstUsesWith(I&: CI, V: Res); |
3723 | } |
3724 | } |
3725 | } |
3726 | [[fallthrough]]; |
3727 | } |
3728 | case Intrinsic::vector_reduce_mul: { |
3729 | if (IID == Intrinsic::vector_reduce_mul) { |
3730 | // Multiplicative reduction over the vector with (potentially-extended) |
3731 | // i1 element type is actually a (potentially zero-extended) |
3732 | // logical `and` reduction over the original non-extended value: |
3733 | // vector_reduce_mul(?ext(<n x i1>)) |
3734 | // --> |
3735 | // zext(vector_reduce_and(<n x i1>)) |
3736 | Value *Arg = II->getArgOperand(i: 0); |
3737 | Value *Vect; |
3738 | |
3739 | if (Value *NewOp = |
3740 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3741 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3742 | return II; |
3743 | } |
3744 | |
3745 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3746 | if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType())) |
3747 | if (VTy->getElementType() == Builder.getInt1Ty()) { |
3748 | Value *Res = Builder.CreateAndReduce(Src: Vect); |
3749 | if (Res->getType() != II->getType()) |
3750 | Res = Builder.CreateZExt(V: Res, DestTy: II->getType()); |
3751 | return replaceInstUsesWith(I&: CI, V: Res); |
3752 | } |
3753 | } |
3754 | } |
3755 | [[fallthrough]]; |
3756 | } |
3757 | case Intrinsic::vector_reduce_umin: |
3758 | case Intrinsic::vector_reduce_umax: { |
3759 | if (IID == Intrinsic::vector_reduce_umin || |
3760 | IID == Intrinsic::vector_reduce_umax) { |
3761 | // UMin/UMax reduction over the vector with (potentially-extended) |
3762 | // i1 element type is actually a (potentially-extended) |
3763 | // logical `and`/`or` reduction over the original non-extended value: |
3764 | // vector_reduce_u{min,max}(?ext(<n x i1>)) |
3765 | // --> |
3766 | // ?ext(vector_reduce_{and,or}(<n x i1>)) |
3767 | Value *Arg = II->getArgOperand(i: 0); |
3768 | Value *Vect; |
3769 | |
3770 | if (Value *NewOp = |
3771 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3772 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3773 | return II; |
3774 | } |
3775 | |
3776 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3777 | if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType())) |
3778 | if (VTy->getElementType() == Builder.getInt1Ty()) { |
3779 | Value *Res = IID == Intrinsic::vector_reduce_umin |
3780 | ? Builder.CreateAndReduce(Src: Vect) |
3781 | : Builder.CreateOrReduce(Src: Vect); |
3782 | if (Arg != Vect) |
3783 | Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res, |
3784 | DestTy: II->getType()); |
3785 | return replaceInstUsesWith(I&: CI, V: Res); |
3786 | } |
3787 | } |
3788 | } |
3789 | [[fallthrough]]; |
3790 | } |
3791 | case Intrinsic::vector_reduce_smin: |
3792 | case Intrinsic::vector_reduce_smax: { |
3793 | if (IID == Intrinsic::vector_reduce_smin || |
3794 | IID == Intrinsic::vector_reduce_smax) { |
3795 | // SMin/SMax reduction over the vector with (potentially-extended) |
3796 | // i1 element type is actually a (potentially-extended) |
3797 | // logical `and`/`or` reduction over the original non-extended value: |
3798 | // vector_reduce_s{min,max}(<n x i1>) |
3799 | // --> |
3800 | // vector_reduce_{or,and}(<n x i1>) |
3801 | // and |
3802 | // vector_reduce_s{min,max}(sext(<n x i1>)) |
3803 | // --> |
3804 | // sext(vector_reduce_{or,and}(<n x i1>)) |
3805 | // and |
3806 | // vector_reduce_s{min,max}(zext(<n x i1>)) |
3807 | // --> |
3808 | // zext(vector_reduce_{and,or}(<n x i1>)) |
3809 | Value *Arg = II->getArgOperand(i: 0); |
3810 | Value *Vect; |
3811 | |
3812 | if (Value *NewOp = |
3813 | simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) { |
3814 | replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp); |
3815 | return II; |
3816 | } |
3817 | |
3818 | if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) { |
3819 | if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType())) |
3820 | if (VTy->getElementType() == Builder.getInt1Ty()) { |
3821 | Instruction::CastOps ExtOpc = Instruction::CastOps::CastOpsEnd; |
3822 | if (Arg != Vect) |
3823 | ExtOpc = cast<CastInst>(Val: Arg)->getOpcode(); |
3824 | Value *Res = ((IID == Intrinsic::vector_reduce_smin) == |
3825 | (ExtOpc == Instruction::CastOps::ZExt)) |
3826 | ? Builder.CreateAndReduce(Src: Vect) |
3827 | : Builder.CreateOrReduce(Src: Vect); |
3828 | if (Arg != Vect) |
3829 | Res = Builder.CreateCast(Op: ExtOpc, V: Res, DestTy: II->getType()); |
3830 | return replaceInstUsesWith(I&: CI, V: Res); |
3831 | } |
3832 | } |
3833 | } |
3834 | [[fallthrough]]; |
3835 | } |
3836 | case Intrinsic::vector_reduce_fmax: |
3837 | case Intrinsic::vector_reduce_fmin: |
3838 | case Intrinsic::vector_reduce_fadd: |
3839 | case Intrinsic::vector_reduce_fmul: { |
3840 | bool CanReorderLanes = (IID != Intrinsic::vector_reduce_fadd && |
3841 | IID != Intrinsic::vector_reduce_fmul) || |
3842 | II->hasAllowReassoc(); |
3843 | const unsigned ArgIdx = (IID == Intrinsic::vector_reduce_fadd || |
3844 | IID == Intrinsic::vector_reduce_fmul) |
3845 | ? 1 |
3846 | : 0; |
3847 | Value *Arg = II->getArgOperand(i: ArgIdx); |
3848 | if (Value *NewOp = simplifyReductionOperand(Arg, CanReorderLanes)) { |
3849 | replaceUse(U&: II->getOperandUse(i: ArgIdx), NewValue: NewOp); |
3850 | return nullptr; |
3851 | } |
3852 | break; |
3853 | } |
3854 | case Intrinsic::is_fpclass: { |
3855 | if (Instruction *I = foldIntrinsicIsFPClass(II&: *II)) |
3856 | return I; |
3857 | break; |
3858 | } |
3859 | case Intrinsic::threadlocal_address: { |
3860 | Align MinAlign = getKnownAlignment(V: II->getArgOperand(i: 0), DL, CxtI: II, AC: &AC, DT: &DT); |
3861 | MaybeAlign Align = II->getRetAlign(); |
3862 | if (MinAlign > Align.valueOrOne()) { |
3863 | II->addRetAttr(Attr: Attribute::getWithAlignment(Context&: II->getContext(), Alignment: MinAlign)); |
3864 | return II; |
3865 | } |
3866 | break; |
3867 | } |
3868 | case Intrinsic::frexp: { |
3869 | Value *X; |
3870 | // The first result is idempotent with the added complication of the struct |
3871 | // return, and the second result is zero because the value is already |
3872 | // normalized. |
3873 | if (match(V: II->getArgOperand(i: 0), P: m_ExtractValue<0>(V: m_Value(V&: X)))) { |
3874 | if (match(V: X, P: m_Intrinsic<Intrinsic::frexp>(Op0: m_Value()))) { |
3875 | X = Builder.CreateInsertValue( |
3876 | Agg: X, Val: Constant::getNullValue(Ty: II->getType()->getStructElementType(N: 1)), |
3877 | Idxs: 1); |
3878 | return replaceInstUsesWith(I&: *II, V: X); |
3879 | } |
3880 | } |
3881 | break; |
3882 | } |
3883 | default: { |
3884 | // Handle target specific intrinsics |
3885 | std::optional<Instruction *> V = targetInstCombineIntrinsic(II&: *II); |
3886 | if (V) |
3887 | return *V; |
3888 | break; |
3889 | } |
3890 | } |
3891 | |
3892 | // Try to fold intrinsic into select operands. This is legal if: |
3893 | // * The intrinsic is speculatable. |
3894 | // * The select condition is not a vector, or the intrinsic does not |
3895 | // perform cross-lane operations. |
3896 | if (isSafeToSpeculativelyExecuteWithVariableReplaced(I: &CI) && |
3897 | isNotCrossLaneOperation(I: II)) |
3898 | for (Value *Op : II->args()) |
3899 | if (auto *Sel = dyn_cast<SelectInst>(Val: Op)) |
3900 | if (Instruction *R = FoldOpIntoSelect(Op&: *II, SI: Sel)) |
3901 | return R; |
3902 | |
3903 | if (Instruction *Shuf = foldShuffledIntrinsicOperands(II)) |
3904 | return Shuf; |
3905 | |
3906 | if (Value *Reverse = foldReversedIntrinsicOperands(II)) |
3907 | return replaceInstUsesWith(I&: *II, V: Reverse); |
3908 | |
3909 | // Some intrinsics (like experimental_gc_statepoint) can be used in invoke |
3910 | // context, so it is handled in visitCallBase and we should trigger it. |
3911 | return visitCallBase(Call&: *II); |
3912 | } |
3913 | |
3914 | // Fence instruction simplification |
3915 | Instruction *InstCombinerImpl::visitFenceInst(FenceInst &FI) { |
3916 | auto *NFI = dyn_cast<FenceInst>(Val: FI.getNextNonDebugInstruction()); |
3917 | // This check is solely here to handle arbitrary target-dependent syncscopes. |
3918 | // TODO: Can remove if does not matter in practice. |
3919 | if (NFI && FI.isIdenticalTo(I: NFI)) |
3920 | return eraseInstFromFunction(I&: FI); |
3921 | |
3922 | // Returns true if FI1 is identical or stronger fence than FI2. |
3923 | auto isIdenticalOrStrongerFence = [](FenceInst *FI1, FenceInst *FI2) { |
3924 | auto FI1SyncScope = FI1->getSyncScopeID(); |
3925 | // Consider same scope, where scope is global or single-thread. |
3926 | if (FI1SyncScope != FI2->getSyncScopeID() || |
3927 | (FI1SyncScope != SyncScope::System && |
3928 | FI1SyncScope != SyncScope::SingleThread)) |
3929 | return false; |
3930 | |
3931 | return isAtLeastOrStrongerThan(AO: FI1->getOrdering(), Other: FI2->getOrdering()); |
3932 | }; |
3933 | if (NFI && isIdenticalOrStrongerFence(NFI, &FI)) |
3934 | return eraseInstFromFunction(I&: FI); |
3935 | |
3936 | if (auto *PFI = dyn_cast_or_null<FenceInst>(Val: FI.getPrevNonDebugInstruction())) |
3937 | if (isIdenticalOrStrongerFence(PFI, &FI)) |
3938 | return eraseInstFromFunction(I&: FI); |
3939 | return nullptr; |
3940 | } |
3941 | |
3942 | // InvokeInst simplification |
3943 | Instruction *InstCombinerImpl::visitInvokeInst(InvokeInst &II) { |
3944 | return visitCallBase(Call&: II); |
3945 | } |
3946 | |
3947 | // CallBrInst simplification |
3948 | Instruction *InstCombinerImpl::visitCallBrInst(CallBrInst &CBI) { |
3949 | return visitCallBase(Call&: CBI); |
3950 | } |
3951 | |
3952 | Instruction *InstCombinerImpl::tryOptimizeCall(CallInst *CI) { |
3953 | if (!CI->getCalledFunction()) return nullptr; |
3954 | |
3955 | // Skip optimizing notail and musttail calls so |
3956 | // LibCallSimplifier::optimizeCall doesn't have to preserve those invariants. |
3957 | // LibCallSimplifier::optimizeCall should try to preserve tail calls though. |
3958 | if (CI->isMustTailCall() || CI->isNoTailCall()) |
3959 | return nullptr; |
3960 | |
3961 | auto InstCombineRAUW = [this](Instruction *From, Value *With) { |
3962 | replaceInstUsesWith(I&: *From, V: With); |
3963 | }; |
3964 | auto InstCombineErase = [this](Instruction *I) { |
3965 | eraseInstFromFunction(I&: *I); |
3966 | }; |
3967 | LibCallSimplifier Simplifier(DL, &TLI, &DT, &DC, &AC, ORE, BFI, PSI, |
3968 | InstCombineRAUW, InstCombineErase); |
3969 | if (Value *With = Simplifier.optimizeCall(CI, B&: Builder)) { |
3970 | ++NumSimplified; |
3971 | return CI->use_empty() ? CI : replaceInstUsesWith(I&: *CI, V: With); |
3972 | } |
3973 | |
3974 | return nullptr; |
3975 | } |
3976 | |
3977 | static IntrinsicInst *findInitTrampolineFromAlloca(Value *TrampMem) { |
3978 | // Strip off at most one level of pointer casts, looking for an alloca. This |
3979 | // is good enough in practice and simpler than handling any number of casts. |
3980 | Value *Underlying = TrampMem->stripPointerCasts(); |
3981 | if (Underlying != TrampMem && |
3982 | (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem)) |
3983 | return nullptr; |
3984 | if (!isa<AllocaInst>(Val: Underlying)) |
3985 | return nullptr; |
3986 | |
3987 | IntrinsicInst *InitTrampoline = nullptr; |
3988 | for (User *U : TrampMem->users()) { |
3989 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U); |
3990 | if (!II) |
3991 | return nullptr; |
3992 | if (II->getIntrinsicID() == Intrinsic::init_trampoline) { |
3993 | if (InitTrampoline) |
3994 | // More than one init_trampoline writes to this value. Give up. |
3995 | return nullptr; |
3996 | InitTrampoline = II; |
3997 | continue; |
3998 | } |
3999 | if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) |
4000 | // Allow any number of calls to adjust.trampoline. |
4001 | continue; |
4002 | return nullptr; |
4003 | } |
4004 | |
4005 | // No call to init.trampoline found. |
4006 | if (!InitTrampoline) |
4007 | return nullptr; |
4008 | |
4009 | // Check that the alloca is being used in the expected way. |
4010 | if (InitTrampoline->getOperand(i_nocapture: 0) != TrampMem) |
4011 | return nullptr; |
4012 | |
4013 | return InitTrampoline; |
4014 | } |
4015 | |
4016 | static IntrinsicInst *findInitTrampolineFromBB(IntrinsicInst *AdjustTramp, |
4017 | Value *TrampMem) { |
4018 | // Visit all the previous instructions in the basic block, and try to find a |
4019 | // init.trampoline which has a direct path to the adjust.trampoline. |
4020 | for (BasicBlock::iterator I = AdjustTramp->getIterator(), |
4021 | E = AdjustTramp->getParent()->begin(); |
4022 | I != E;) { |
4023 | Instruction *Inst = &*--I; |
4024 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val&: I)) |
4025 | if (II->getIntrinsicID() == Intrinsic::init_trampoline && |
4026 | II->getOperand(i_nocapture: 0) == TrampMem) |
4027 | return II; |
4028 | if (Inst->mayWriteToMemory()) |
4029 | return nullptr; |
4030 | } |
4031 | return nullptr; |
4032 | } |
4033 | |
4034 | // Given a call to llvm.adjust.trampoline, find and return the corresponding |
4035 | // call to llvm.init.trampoline if the call to the trampoline can be optimized |
4036 | // to a direct call to a function. Otherwise return NULL. |
4037 | static IntrinsicInst *findInitTrampoline(Value *Callee) { |
4038 | Callee = Callee->stripPointerCasts(); |
4039 | IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Val: Callee); |
4040 | if (!AdjustTramp || |
4041 | AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) |
4042 | return nullptr; |
4043 | |
4044 | Value *TrampMem = AdjustTramp->getOperand(i_nocapture: 0); |
4045 | |
4046 | if (IntrinsicInst *IT = findInitTrampolineFromAlloca(TrampMem)) |
4047 | return IT; |
4048 | if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem)) |
4049 | return IT; |
4050 | return nullptr; |
4051 | } |
4052 | |
4053 | bool InstCombinerImpl::annotateAnyAllocSite(CallBase &Call, |
4054 | const TargetLibraryInfo *TLI) { |
4055 | // Note: We only handle cases which can't be driven from generic attributes |
4056 | // here. So, for example, nonnull and noalias (which are common properties |
4057 | // of some allocation functions) are expected to be handled via annotation |
4058 | // of the respective allocator declaration with generic attributes. |
4059 | bool Changed = false; |
4060 | |
4061 | if (!Call.getType()->isPointerTy()) |
4062 | return Changed; |
4063 | |
4064 | std::optional<APInt> Size = getAllocSize(CB: &Call, TLI); |
4065 | if (Size && *Size != 0) { |
4066 | // TODO: We really should just emit deref_or_null here and then |
4067 | // let the generic inference code combine that with nonnull. |
4068 | if (Call.hasRetAttr(Kind: Attribute::NonNull)) { |
4069 | Changed = !Call.hasRetAttr(Kind: Attribute::Dereferenceable); |
4070 | Call.addRetAttr(Attr: Attribute::getWithDereferenceableBytes( |
4071 | Context&: Call.getContext(), Bytes: Size->getLimitedValue())); |
4072 | } else { |
4073 | Changed = !Call.hasRetAttr(Kind: Attribute::DereferenceableOrNull); |
4074 | Call.addRetAttr(Attr: Attribute::getWithDereferenceableOrNullBytes( |
4075 | Context&: Call.getContext(), Bytes: Size->getLimitedValue())); |
4076 | } |
4077 | } |
4078 | |
4079 | // Add alignment attribute if alignment is a power of two constant. |
4080 | Value *Alignment = getAllocAlignment(V: &Call, TLI); |
4081 | if (!Alignment) |
4082 | return Changed; |
4083 | |
4084 | ConstantInt *AlignOpC = dyn_cast<ConstantInt>(Val: Alignment); |
4085 | if (AlignOpC && AlignOpC->getValue().ult(RHS: llvm::Value::MaximumAlignment)) { |
4086 | uint64_t AlignmentVal = AlignOpC->getZExtValue(); |
4087 | if (llvm::isPowerOf2_64(Value: AlignmentVal)) { |
4088 | Align ExistingAlign = Call.getRetAlign().valueOrOne(); |
4089 | Align NewAlign = Align(AlignmentVal); |
4090 | if (NewAlign > ExistingAlign) { |
4091 | Call.addRetAttr( |
4092 | Attr: Attribute::getWithAlignment(Context&: Call.getContext(), Alignment: NewAlign)); |
4093 | Changed = true; |
4094 | } |
4095 | } |
4096 | } |
4097 | return Changed; |
4098 | } |
4099 | |
4100 | /// Improvements for call, callbr and invoke instructions. |
4101 | Instruction *InstCombinerImpl::visitCallBase(CallBase &Call) { |
4102 | bool Changed = annotateAnyAllocSite(Call, TLI: &TLI); |
4103 | |
4104 | // Mark any parameters that are known to be non-null with the nonnull |
4105 | // attribute. This is helpful for inlining calls to functions with null |
4106 | // checks on their arguments. |
4107 | SmallVector<unsigned, 4> ArgNos; |
4108 | unsigned ArgNo = 0; |
4109 | |
4110 | for (Value *V : Call.args()) { |
4111 | if (V->getType()->isPointerTy()) { |
4112 | // Simplify the nonnull operand if the parameter is known to be nonnull. |
4113 | // Otherwise, try to infer nonnull for it. |
4114 | bool HasDereferenceable = Call.getParamDereferenceableBytes(i: ArgNo) > 0; |
4115 | if (Call.paramHasAttr(ArgNo, Kind: Attribute::NonNull) || |
4116 | (HasDereferenceable && |
4117 | !NullPointerIsDefined(F: Call.getFunction(), |
4118 | AS: V->getType()->getPointerAddressSpace()))) { |
4119 | if (Value *Res = simplifyNonNullOperand(V, HasDereferenceable)) { |
4120 | replaceOperand(I&: Call, OpNum: ArgNo, V: Res); |
4121 | Changed = true; |
4122 | } |
4123 | } else if (isKnownNonZero(V, |
4124 | Q: getSimplifyQuery().getWithInstruction(I: &Call))) { |
4125 | ArgNos.push_back(Elt: ArgNo); |
4126 | } |
4127 | } |
4128 | ArgNo++; |
4129 | } |
4130 | |
4131 | assert(ArgNo == Call.arg_size() && "Call arguments not processed correctly." ); |
4132 | |
4133 | if (!ArgNos.empty()) { |
4134 | AttributeList AS = Call.getAttributes(); |
4135 | LLVMContext &Ctx = Call.getContext(); |
4136 | AS = AS.addParamAttribute(C&: Ctx, ArgNos, |
4137 | A: Attribute::get(Context&: Ctx, Kind: Attribute::NonNull)); |
4138 | Call.setAttributes(AS); |
4139 | Changed = true; |
4140 | } |
4141 | |
4142 | // If the callee is a pointer to a function, attempt to move any casts to the |
4143 | // arguments of the call/callbr/invoke. |
4144 | Value *Callee = Call.getCalledOperand(); |
4145 | Function *CalleeF = dyn_cast<Function>(Val: Callee); |
4146 | if ((!CalleeF || CalleeF->getFunctionType() != Call.getFunctionType()) && |
4147 | transformConstExprCastCall(Call)) |
4148 | return nullptr; |
4149 | |
4150 | if (CalleeF) { |
4151 | // Remove the convergent attr on calls when the callee is not convergent. |
4152 | if (Call.isConvergent() && !CalleeF->isConvergent() && |
4153 | !CalleeF->isIntrinsic()) { |
4154 | LLVM_DEBUG(dbgs() << "Removing convergent attr from instr " << Call |
4155 | << "\n" ); |
4156 | Call.setNotConvergent(); |
4157 | return &Call; |
4158 | } |
4159 | |
4160 | // If the call and callee calling conventions don't match, and neither one |
4161 | // of the calling conventions is compatible with C calling convention |
4162 | // this call must be unreachable, as the call is undefined. |
4163 | if ((CalleeF->getCallingConv() != Call.getCallingConv() && |
4164 | !(CalleeF->getCallingConv() == llvm::CallingConv::C && |
4165 | TargetLibraryInfoImpl::isCallingConvCCompatible(CI: &Call)) && |
4166 | !(Call.getCallingConv() == llvm::CallingConv::C && |
4167 | TargetLibraryInfoImpl::isCallingConvCCompatible(Callee: CalleeF))) && |
4168 | // Only do this for calls to a function with a body. A prototype may |
4169 | // not actually end up matching the implementation's calling conv for a |
4170 | // variety of reasons (e.g. it may be written in assembly). |
4171 | !CalleeF->isDeclaration()) { |
4172 | Instruction *OldCall = &Call; |
4173 | CreateNonTerminatorUnreachable(InsertAt: OldCall); |
4174 | // If OldCall does not return void then replaceInstUsesWith poison. |
4175 | // This allows ValueHandlers and custom metadata to adjust itself. |
4176 | if (!OldCall->getType()->isVoidTy()) |
4177 | replaceInstUsesWith(I&: *OldCall, V: PoisonValue::get(T: OldCall->getType())); |
4178 | if (isa<CallInst>(Val: OldCall)) |
4179 | return eraseInstFromFunction(I&: *OldCall); |
4180 | |
4181 | // We cannot remove an invoke or a callbr, because it would change thexi |
4182 | // CFG, just change the callee to a null pointer. |
4183 | cast<CallBase>(Val: OldCall)->setCalledFunction( |
4184 | FTy: CalleeF->getFunctionType(), |
4185 | Fn: Constant::getNullValue(Ty: CalleeF->getType())); |
4186 | return nullptr; |
4187 | } |
4188 | } |
4189 | |
4190 | // Calling a null function pointer is undefined if a null address isn't |
4191 | // dereferenceable. |
4192 | if ((isa<ConstantPointerNull>(Val: Callee) && |
4193 | !NullPointerIsDefined(F: Call.getFunction())) || |
4194 | isa<UndefValue>(Val: Callee)) { |
4195 | // If Call does not return void then replaceInstUsesWith poison. |
4196 | // This allows ValueHandlers and custom metadata to adjust itself. |
4197 | if (!Call.getType()->isVoidTy()) |
4198 | replaceInstUsesWith(I&: Call, V: PoisonValue::get(T: Call.getType())); |
4199 | |
4200 | if (Call.isTerminator()) { |
4201 | // Can't remove an invoke or callbr because we cannot change the CFG. |
4202 | return nullptr; |
4203 | } |
4204 | |
4205 | // This instruction is not reachable, just remove it. |
4206 | CreateNonTerminatorUnreachable(InsertAt: &Call); |
4207 | return eraseInstFromFunction(I&: Call); |
4208 | } |
4209 | |
4210 | if (IntrinsicInst *II = findInitTrampoline(Callee)) |
4211 | return transformCallThroughTrampoline(Call, Tramp&: *II); |
4212 | |
4213 | if (isa<InlineAsm>(Val: Callee) && !Call.doesNotThrow()) { |
4214 | InlineAsm *IA = cast<InlineAsm>(Val: Callee); |
4215 | if (!IA->canThrow()) { |
4216 | // Normal inline asm calls cannot throw - mark them |
4217 | // 'nounwind'. |
4218 | Call.setDoesNotThrow(); |
4219 | Changed = true; |
4220 | } |
4221 | } |
4222 | |
4223 | // Try to optimize the call if possible, we require DataLayout for most of |
4224 | // this. None of these calls are seen as possibly dead so go ahead and |
4225 | // delete the instruction now. |
4226 | if (CallInst *CI = dyn_cast<CallInst>(Val: &Call)) { |
4227 | Instruction *I = tryOptimizeCall(CI); |
4228 | // If we changed something return the result, etc. Otherwise let |
4229 | // the fallthrough check. |
4230 | if (I) return eraseInstFromFunction(I&: *I); |
4231 | } |
4232 | |
4233 | if (!Call.use_empty() && !Call.isMustTailCall()) |
4234 | if (Value *ReturnedArg = Call.getReturnedArgOperand()) { |
4235 | Type *CallTy = Call.getType(); |
4236 | Type *RetArgTy = ReturnedArg->getType(); |
4237 | if (RetArgTy->canLosslesslyBitCastTo(Ty: CallTy)) |
4238 | return replaceInstUsesWith( |
4239 | I&: Call, V: Builder.CreateBitOrPointerCast(V: ReturnedArg, DestTy: CallTy)); |
4240 | } |
4241 | |
4242 | // Drop unnecessary kcfi operand bundles from calls that were converted |
4243 | // into direct calls. |
4244 | auto Bundle = Call.getOperandBundle(ID: LLVMContext::OB_kcfi); |
4245 | if (Bundle && !Call.isIndirectCall()) { |
4246 | DEBUG_WITH_TYPE(DEBUG_TYPE "-kcfi" , { |
4247 | if (CalleeF) { |
4248 | ConstantInt *FunctionType = nullptr; |
4249 | ConstantInt *ExpectedType = cast<ConstantInt>(Bundle->Inputs[0]); |
4250 | |
4251 | if (MDNode *MD = CalleeF->getMetadata(LLVMContext::MD_kcfi_type)) |
4252 | FunctionType = mdconst::extract<ConstantInt>(MD->getOperand(0)); |
4253 | |
4254 | if (FunctionType && |
4255 | FunctionType->getZExtValue() != ExpectedType->getZExtValue()) |
4256 | dbgs() << Call.getModule()->getName() |
4257 | << ": warning: kcfi: " << Call.getCaller()->getName() |
4258 | << ": call to " << CalleeF->getName() |
4259 | << " using a mismatching function pointer type\n" ; |
4260 | } |
4261 | }); |
4262 | |
4263 | return CallBase::removeOperandBundle(CB: &Call, ID: LLVMContext::OB_kcfi); |
4264 | } |
4265 | |
4266 | if (isRemovableAlloc(V: &Call, TLI: &TLI)) |
4267 | return visitAllocSite(FI&: Call); |
4268 | |
4269 | // Handle intrinsics which can be used in both call and invoke context. |
4270 | switch (Call.getIntrinsicID()) { |
4271 | case Intrinsic::experimental_gc_statepoint: { |
4272 | GCStatepointInst &GCSP = *cast<GCStatepointInst>(Val: &Call); |
4273 | SmallPtrSet<Value *, 32> LiveGcValues; |
4274 | for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) { |
4275 | GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc); |
4276 | |
4277 | // Remove the relocation if unused. |
4278 | if (GCR.use_empty()) { |
4279 | eraseInstFromFunction(I&: GCR); |
4280 | continue; |
4281 | } |
4282 | |
4283 | Value *DerivedPtr = GCR.getDerivedPtr(); |
4284 | Value *BasePtr = GCR.getBasePtr(); |
4285 | |
4286 | // Undef is undef, even after relocation. |
4287 | if (isa<UndefValue>(Val: DerivedPtr) || isa<UndefValue>(Val: BasePtr)) { |
4288 | replaceInstUsesWith(I&: GCR, V: UndefValue::get(T: GCR.getType())); |
4289 | eraseInstFromFunction(I&: GCR); |
4290 | continue; |
4291 | } |
4292 | |
4293 | if (auto *PT = dyn_cast<PointerType>(Val: GCR.getType())) { |
4294 | // The relocation of null will be null for most any collector. |
4295 | // TODO: provide a hook for this in GCStrategy. There might be some |
4296 | // weird collector this property does not hold for. |
4297 | if (isa<ConstantPointerNull>(Val: DerivedPtr)) { |
4298 | // Use null-pointer of gc_relocate's type to replace it. |
4299 | replaceInstUsesWith(I&: GCR, V: ConstantPointerNull::get(T: PT)); |
4300 | eraseInstFromFunction(I&: GCR); |
4301 | continue; |
4302 | } |
4303 | |
4304 | // isKnownNonNull -> nonnull attribute |
4305 | if (!GCR.hasRetAttr(Kind: Attribute::NonNull) && |
4306 | isKnownNonZero(V: DerivedPtr, |
4307 | Q: getSimplifyQuery().getWithInstruction(I: &Call))) { |
4308 | GCR.addRetAttr(Kind: Attribute::NonNull); |
4309 | // We discovered new fact, re-check users. |
4310 | Worklist.pushUsersToWorkList(I&: GCR); |
4311 | } |
4312 | } |
4313 | |
4314 | // If we have two copies of the same pointer in the statepoint argument |
4315 | // list, canonicalize to one. This may let us common gc.relocates. |
4316 | if (GCR.getBasePtr() == GCR.getDerivedPtr() && |
4317 | GCR.getBasePtrIndex() != GCR.getDerivedPtrIndex()) { |
4318 | auto *OpIntTy = GCR.getOperand(i_nocapture: 2)->getType(); |
4319 | GCR.setOperand(i_nocapture: 2, Val_nocapture: ConstantInt::get(Ty: OpIntTy, V: GCR.getBasePtrIndex())); |
4320 | } |
4321 | |
4322 | // TODO: bitcast(relocate(p)) -> relocate(bitcast(p)) |
4323 | // Canonicalize on the type from the uses to the defs |
4324 | |
4325 | // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...) |
4326 | LiveGcValues.insert(Ptr: BasePtr); |
4327 | LiveGcValues.insert(Ptr: DerivedPtr); |
4328 | } |
4329 | std::optional<OperandBundleUse> Bundle = |
4330 | GCSP.getOperandBundle(ID: LLVMContext::OB_gc_live); |
4331 | unsigned NumOfGCLives = LiveGcValues.size(); |
4332 | if (!Bundle || NumOfGCLives == Bundle->Inputs.size()) |
4333 | break; |
4334 | // We can reduce the size of gc live bundle. |
4335 | DenseMap<Value *, unsigned> Val2Idx; |
4336 | std::vector<Value *> NewLiveGc; |
4337 | for (Value *V : Bundle->Inputs) { |
4338 | auto [It, Inserted] = Val2Idx.try_emplace(Key: V); |
4339 | if (!Inserted) |
4340 | continue; |
4341 | if (LiveGcValues.count(Ptr: V)) { |
4342 | It->second = NewLiveGc.size(); |
4343 | NewLiveGc.push_back(x: V); |
4344 | } else |
4345 | It->second = NumOfGCLives; |
4346 | } |
4347 | // Update all gc.relocates |
4348 | for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) { |
4349 | GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc); |
4350 | Value *BasePtr = GCR.getBasePtr(); |
4351 | assert(Val2Idx.count(BasePtr) && Val2Idx[BasePtr] != NumOfGCLives && |
4352 | "Missed live gc for base pointer" ); |
4353 | auto *OpIntTy1 = GCR.getOperand(i_nocapture: 1)->getType(); |
4354 | GCR.setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: OpIntTy1, V: Val2Idx[BasePtr])); |
4355 | Value *DerivedPtr = GCR.getDerivedPtr(); |
4356 | assert(Val2Idx.count(DerivedPtr) && Val2Idx[DerivedPtr] != NumOfGCLives && |
4357 | "Missed live gc for derived pointer" ); |
4358 | auto *OpIntTy2 = GCR.getOperand(i_nocapture: 2)->getType(); |
4359 | GCR.setOperand(i_nocapture: 2, Val_nocapture: ConstantInt::get(Ty: OpIntTy2, V: Val2Idx[DerivedPtr])); |
4360 | } |
4361 | // Create new statepoint instruction. |
4362 | OperandBundleDef NewBundle("gc-live" , NewLiveGc); |
4363 | return CallBase::Create(CB: &Call, Bundle: NewBundle); |
4364 | } |
4365 | default: { break; } |
4366 | } |
4367 | |
4368 | return Changed ? &Call : nullptr; |
4369 | } |
4370 | |
4371 | /// If the callee is a constexpr cast of a function, attempt to move the cast to |
4372 | /// the arguments of the call/invoke. |
4373 | /// CallBrInst is not supported. |
4374 | bool InstCombinerImpl::transformConstExprCastCall(CallBase &Call) { |
4375 | auto *Callee = |
4376 | dyn_cast<Function>(Val: Call.getCalledOperand()->stripPointerCasts()); |
4377 | if (!Callee) |
4378 | return false; |
4379 | |
4380 | assert(!isa<CallBrInst>(Call) && |
4381 | "CallBr's don't have a single point after a def to insert at" ); |
4382 | |
4383 | // Don't perform the transform for declarations, which may not be fully |
4384 | // accurate. For example, void @foo() is commonly used as a placeholder for |
4385 | // unknown prototypes. |
4386 | if (Callee->isDeclaration()) |
4387 | return false; |
4388 | |
4389 | // If this is a call to a thunk function, don't remove the cast. Thunks are |
4390 | // used to transparently forward all incoming parameters and outgoing return |
4391 | // values, so it's important to leave the cast in place. |
4392 | if (Callee->hasFnAttribute(Kind: "thunk" )) |
4393 | return false; |
4394 | |
4395 | // If this is a call to a naked function, the assembly might be |
4396 | // using an argument, or otherwise rely on the frame layout, |
4397 | // the function prototype will mismatch. |
4398 | if (Callee->hasFnAttribute(Kind: Attribute::Naked)) |
4399 | return false; |
4400 | |
4401 | // If this is a musttail call, the callee's prototype must match the caller's |
4402 | // prototype with the exception of pointee types. The code below doesn't |
4403 | // implement that, so we can't do this transform. |
4404 | // TODO: Do the transform if it only requires adding pointer casts. |
4405 | if (Call.isMustTailCall()) |
4406 | return false; |
4407 | |
4408 | Instruction *Caller = &Call; |
4409 | const AttributeList &CallerPAL = Call.getAttributes(); |
4410 | |
4411 | // Okay, this is a cast from a function to a different type. Unless doing so |
4412 | // would cause a type conversion of one of our arguments, change this call to |
4413 | // be a direct call with arguments casted to the appropriate types. |
4414 | FunctionType *FT = Callee->getFunctionType(); |
4415 | Type *OldRetTy = Caller->getType(); |
4416 | Type *NewRetTy = FT->getReturnType(); |
4417 | |
4418 | // Check to see if we are changing the return type... |
4419 | if (OldRetTy != NewRetTy) { |
4420 | |
4421 | if (NewRetTy->isStructTy()) |
4422 | return false; // TODO: Handle multiple return values. |
4423 | |
4424 | if (!CastInst::isBitOrNoopPointerCastable(SrcTy: NewRetTy, DestTy: OldRetTy, DL)) { |
4425 | if (!Caller->use_empty()) |
4426 | return false; // Cannot transform this return value. |
4427 | } |
4428 | |
4429 | if (!CallerPAL.isEmpty() && !Caller->use_empty()) { |
4430 | AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs()); |
4431 | if (RAttrs.overlaps(AM: AttributeFuncs::typeIncompatible( |
4432 | Ty: NewRetTy, AS: CallerPAL.getRetAttrs()))) |
4433 | return false; // Attribute not compatible with transformed value. |
4434 | } |
4435 | |
4436 | // If the callbase is an invoke instruction, and the return value is |
4437 | // used by a PHI node in a successor, we cannot change the return type of |
4438 | // the call because there is no place to put the cast instruction (without |
4439 | // breaking the critical edge). Bail out in this case. |
4440 | if (!Caller->use_empty()) { |
4441 | BasicBlock *PhisNotSupportedBlock = nullptr; |
4442 | if (auto *II = dyn_cast<InvokeInst>(Val: Caller)) |
4443 | PhisNotSupportedBlock = II->getNormalDest(); |
4444 | if (PhisNotSupportedBlock) |
4445 | for (User *U : Caller->users()) |
4446 | if (PHINode *PN = dyn_cast<PHINode>(Val: U)) |
4447 | if (PN->getParent() == PhisNotSupportedBlock) |
4448 | return false; |
4449 | } |
4450 | } |
4451 | |
4452 | unsigned NumActualArgs = Call.arg_size(); |
4453 | unsigned NumCommonArgs = std::min(a: FT->getNumParams(), b: NumActualArgs); |
4454 | |
4455 | // Prevent us turning: |
4456 | // declare void @takes_i32_inalloca(i32* inalloca) |
4457 | // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0) |
4458 | // |
4459 | // into: |
4460 | // call void @takes_i32_inalloca(i32* null) |
4461 | // |
4462 | // Similarly, avoid folding away bitcasts of byval calls. |
4463 | if (Callee->getAttributes().hasAttrSomewhere(Kind: Attribute::InAlloca) || |
4464 | Callee->getAttributes().hasAttrSomewhere(Kind: Attribute::Preallocated)) |
4465 | return false; |
4466 | |
4467 | auto AI = Call.arg_begin(); |
4468 | for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { |
4469 | Type *ParamTy = FT->getParamType(i); |
4470 | Type *ActTy = (*AI)->getType(); |
4471 | |
4472 | if (!CastInst::isBitOrNoopPointerCastable(SrcTy: ActTy, DestTy: ParamTy, DL)) |
4473 | return false; // Cannot transform this parameter value. |
4474 | |
4475 | // Check if there are any incompatible attributes we cannot drop safely. |
4476 | if (AttrBuilder(FT->getContext(), CallerPAL.getParamAttrs(ArgNo: i)) |
4477 | .overlaps(AM: AttributeFuncs::typeIncompatible( |
4478 | Ty: ParamTy, AS: CallerPAL.getParamAttrs(ArgNo: i), |
4479 | ASK: AttributeFuncs::ASK_UNSAFE_TO_DROP))) |
4480 | return false; // Attribute not compatible with transformed value. |
4481 | |
4482 | if (Call.isInAllocaArgument(ArgNo: i) || |
4483 | CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::Preallocated)) |
4484 | return false; // Cannot transform to and from inalloca/preallocated. |
4485 | |
4486 | if (CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::SwiftError)) |
4487 | return false; |
4488 | |
4489 | if (CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::ByVal) != |
4490 | Callee->getAttributes().hasParamAttr(ArgNo: i, Kind: Attribute::ByVal)) |
4491 | return false; // Cannot transform to or from byval. |
4492 | } |
4493 | |
4494 | if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && |
4495 | !CallerPAL.isEmpty()) { |
4496 | // In this case we have more arguments than the new function type, but we |
4497 | // won't be dropping them. Check that these extra arguments have attributes |
4498 | // that are compatible with being a vararg call argument. |
4499 | unsigned SRetIdx; |
4500 | if (CallerPAL.hasAttrSomewhere(Kind: Attribute::StructRet, Index: &SRetIdx) && |
4501 | SRetIdx - AttributeList::FirstArgIndex >= FT->getNumParams()) |
4502 | return false; |
4503 | } |
4504 | |
4505 | // Okay, we decided that this is a safe thing to do: go ahead and start |
4506 | // inserting cast instructions as necessary. |
4507 | SmallVector<Value *, 8> Args; |
4508 | SmallVector<AttributeSet, 8> ArgAttrs; |
4509 | Args.reserve(N: NumActualArgs); |
4510 | ArgAttrs.reserve(N: NumActualArgs); |
4511 | |
4512 | // Get any return attributes. |
4513 | AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs()); |
4514 | |
4515 | // If the return value is not being used, the type may not be compatible |
4516 | // with the existing attributes. Wipe out any problematic attributes. |
4517 | RAttrs.remove( |
4518 | AM: AttributeFuncs::typeIncompatible(Ty: NewRetTy, AS: CallerPAL.getRetAttrs())); |
4519 | |
4520 | LLVMContext &Ctx = Call.getContext(); |
4521 | AI = Call.arg_begin(); |
4522 | for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { |
4523 | Type *ParamTy = FT->getParamType(i); |
4524 | |
4525 | Value *NewArg = *AI; |
4526 | if ((*AI)->getType() != ParamTy) |
4527 | NewArg = Builder.CreateBitOrPointerCast(V: *AI, DestTy: ParamTy); |
4528 | Args.push_back(Elt: NewArg); |
4529 | |
4530 | // Add any parameter attributes except the ones incompatible with the new |
4531 | // type. Note that we made sure all incompatible ones are safe to drop. |
4532 | AttributeMask IncompatibleAttrs = AttributeFuncs::typeIncompatible( |
4533 | Ty: ParamTy, AS: CallerPAL.getParamAttrs(ArgNo: i), ASK: AttributeFuncs::ASK_SAFE_TO_DROP); |
4534 | ArgAttrs.push_back( |
4535 | Elt: CallerPAL.getParamAttrs(ArgNo: i).removeAttributes(C&: Ctx, AttrsToRemove: IncompatibleAttrs)); |
4536 | } |
4537 | |
4538 | // If the function takes more arguments than the call was taking, add them |
4539 | // now. |
4540 | for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) { |
4541 | Args.push_back(Elt: Constant::getNullValue(Ty: FT->getParamType(i))); |
4542 | ArgAttrs.push_back(Elt: AttributeSet()); |
4543 | } |
4544 | |
4545 | // If we are removing arguments to the function, emit an obnoxious warning. |
4546 | if (FT->getNumParams() < NumActualArgs) { |
4547 | // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722 |
4548 | if (FT->isVarArg()) { |
4549 | // Add all of the arguments in their promoted form to the arg list. |
4550 | for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { |
4551 | Type *PTy = getPromotedType(Ty: (*AI)->getType()); |
4552 | Value *NewArg = *AI; |
4553 | if (PTy != (*AI)->getType()) { |
4554 | // Must promote to pass through va_arg area! |
4555 | Instruction::CastOps opcode = |
4556 | CastInst::getCastOpcode(Val: *AI, SrcIsSigned: false, Ty: PTy, DstIsSigned: false); |
4557 | NewArg = Builder.CreateCast(Op: opcode, V: *AI, DestTy: PTy); |
4558 | } |
4559 | Args.push_back(Elt: NewArg); |
4560 | |
4561 | // Add any parameter attributes. |
4562 | ArgAttrs.push_back(Elt: CallerPAL.getParamAttrs(ArgNo: i)); |
4563 | } |
4564 | } |
4565 | } |
4566 | |
4567 | AttributeSet FnAttrs = CallerPAL.getFnAttrs(); |
4568 | |
4569 | if (NewRetTy->isVoidTy()) |
4570 | Caller->setName("" ); // Void type should not have a name. |
4571 | |
4572 | assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) && |
4573 | "missing argument attributes" ); |
4574 | AttributeList NewCallerPAL = AttributeList::get( |
4575 | C&: Ctx, FnAttrs, RetAttrs: AttributeSet::get(C&: Ctx, B: RAttrs), ArgAttrs); |
4576 | |
4577 | SmallVector<OperandBundleDef, 1> OpBundles; |
4578 | Call.getOperandBundlesAsDefs(Defs&: OpBundles); |
4579 | |
4580 | CallBase *NewCall; |
4581 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: Caller)) { |
4582 | NewCall = Builder.CreateInvoke(Callee, NormalDest: II->getNormalDest(), |
4583 | UnwindDest: II->getUnwindDest(), Args, OpBundles); |
4584 | } else { |
4585 | NewCall = Builder.CreateCall(Callee, Args, OpBundles); |
4586 | cast<CallInst>(Val: NewCall)->setTailCallKind( |
4587 | cast<CallInst>(Val: Caller)->getTailCallKind()); |
4588 | } |
4589 | NewCall->takeName(V: Caller); |
4590 | NewCall->setCallingConv(Call.getCallingConv()); |
4591 | NewCall->setAttributes(NewCallerPAL); |
4592 | |
4593 | // Preserve prof metadata if any. |
4594 | NewCall->copyMetadata(SrcInst: *Caller, WL: {LLVMContext::MD_prof}); |
4595 | |
4596 | // Insert a cast of the return type as necessary. |
4597 | Instruction *NC = NewCall; |
4598 | Value *NV = NC; |
4599 | if (OldRetTy != NV->getType() && !Caller->use_empty()) { |
4600 | assert(!NV->getType()->isVoidTy()); |
4601 | NV = NC = CastInst::CreateBitOrPointerCast(S: NC, Ty: OldRetTy); |
4602 | NC->setDebugLoc(Caller->getDebugLoc()); |
4603 | |
4604 | auto OptInsertPt = NewCall->getInsertionPointAfterDef(); |
4605 | assert(OptInsertPt && "No place to insert cast" ); |
4606 | InsertNewInstBefore(New: NC, Old: *OptInsertPt); |
4607 | Worklist.pushUsersToWorkList(I&: *Caller); |
4608 | } |
4609 | |
4610 | if (!Caller->use_empty()) |
4611 | replaceInstUsesWith(I&: *Caller, V: NV); |
4612 | else if (Caller->hasValueHandle()) { |
4613 | if (OldRetTy == NV->getType()) |
4614 | ValueHandleBase::ValueIsRAUWd(Old: Caller, New: NV); |
4615 | else |
4616 | // We cannot call ValueIsRAUWd with a different type, and the |
4617 | // actual tracked value will disappear. |
4618 | ValueHandleBase::ValueIsDeleted(V: Caller); |
4619 | } |
4620 | |
4621 | eraseInstFromFunction(I&: *Caller); |
4622 | return true; |
4623 | } |
4624 | |
4625 | /// Turn a call to a function created by init_trampoline / adjust_trampoline |
4626 | /// intrinsic pair into a direct call to the underlying function. |
4627 | Instruction * |
4628 | InstCombinerImpl::transformCallThroughTrampoline(CallBase &Call, |
4629 | IntrinsicInst &Tramp) { |
4630 | FunctionType *FTy = Call.getFunctionType(); |
4631 | AttributeList Attrs = Call.getAttributes(); |
4632 | |
4633 | // If the call already has the 'nest' attribute somewhere then give up - |
4634 | // otherwise 'nest' would occur twice after splicing in the chain. |
4635 | if (Attrs.hasAttrSomewhere(Kind: Attribute::Nest)) |
4636 | return nullptr; |
4637 | |
4638 | Function *NestF = cast<Function>(Val: Tramp.getArgOperand(i: 1)->stripPointerCasts()); |
4639 | FunctionType *NestFTy = NestF->getFunctionType(); |
4640 | |
4641 | AttributeList NestAttrs = NestF->getAttributes(); |
4642 | if (!NestAttrs.isEmpty()) { |
4643 | unsigned NestArgNo = 0; |
4644 | Type *NestTy = nullptr; |
4645 | AttributeSet NestAttr; |
4646 | |
4647 | // Look for a parameter marked with the 'nest' attribute. |
4648 | for (FunctionType::param_iterator I = NestFTy->param_begin(), |
4649 | E = NestFTy->param_end(); |
4650 | I != E; ++NestArgNo, ++I) { |
4651 | AttributeSet AS = NestAttrs.getParamAttrs(ArgNo: NestArgNo); |
4652 | if (AS.hasAttribute(Kind: Attribute::Nest)) { |
4653 | // Record the parameter type and any other attributes. |
4654 | NestTy = *I; |
4655 | NestAttr = AS; |
4656 | break; |
4657 | } |
4658 | } |
4659 | |
4660 | if (NestTy) { |
4661 | std::vector<Value*> NewArgs; |
4662 | std::vector<AttributeSet> NewArgAttrs; |
4663 | NewArgs.reserve(n: Call.arg_size() + 1); |
4664 | NewArgAttrs.reserve(n: Call.arg_size()); |
4665 | |
4666 | // Insert the nest argument into the call argument list, which may |
4667 | // mean appending it. Likewise for attributes. |
4668 | |
4669 | { |
4670 | unsigned ArgNo = 0; |
4671 | auto I = Call.arg_begin(), E = Call.arg_end(); |
4672 | do { |
4673 | if (ArgNo == NestArgNo) { |
4674 | // Add the chain argument and attributes. |
4675 | Value *NestVal = Tramp.getArgOperand(i: 2); |
4676 | if (NestVal->getType() != NestTy) |
4677 | NestVal = Builder.CreateBitCast(V: NestVal, DestTy: NestTy, Name: "nest" ); |
4678 | NewArgs.push_back(x: NestVal); |
4679 | NewArgAttrs.push_back(x: NestAttr); |
4680 | } |
4681 | |
4682 | if (I == E) |
4683 | break; |
4684 | |
4685 | // Add the original argument and attributes. |
4686 | NewArgs.push_back(x: *I); |
4687 | NewArgAttrs.push_back(x: Attrs.getParamAttrs(ArgNo)); |
4688 | |
4689 | ++ArgNo; |
4690 | ++I; |
4691 | } while (true); |
4692 | } |
4693 | |
4694 | // The trampoline may have been bitcast to a bogus type (FTy). |
4695 | // Handle this by synthesizing a new function type, equal to FTy |
4696 | // with the chain parameter inserted. |
4697 | |
4698 | std::vector<Type*> NewTypes; |
4699 | NewTypes.reserve(n: FTy->getNumParams()+1); |
4700 | |
4701 | // Insert the chain's type into the list of parameter types, which may |
4702 | // mean appending it. |
4703 | { |
4704 | unsigned ArgNo = 0; |
4705 | FunctionType::param_iterator I = FTy->param_begin(), |
4706 | E = FTy->param_end(); |
4707 | |
4708 | do { |
4709 | if (ArgNo == NestArgNo) |
4710 | // Add the chain's type. |
4711 | NewTypes.push_back(x: NestTy); |
4712 | |
4713 | if (I == E) |
4714 | break; |
4715 | |
4716 | // Add the original type. |
4717 | NewTypes.push_back(x: *I); |
4718 | |
4719 | ++ArgNo; |
4720 | ++I; |
4721 | } while (true); |
4722 | } |
4723 | |
4724 | // Replace the trampoline call with a direct call. Let the generic |
4725 | // code sort out any function type mismatches. |
4726 | FunctionType *NewFTy = |
4727 | FunctionType::get(Result: FTy->getReturnType(), Params: NewTypes, isVarArg: FTy->isVarArg()); |
4728 | AttributeList NewPAL = |
4729 | AttributeList::get(C&: FTy->getContext(), FnAttrs: Attrs.getFnAttrs(), |
4730 | RetAttrs: Attrs.getRetAttrs(), ArgAttrs: NewArgAttrs); |
4731 | |
4732 | SmallVector<OperandBundleDef, 1> OpBundles; |
4733 | Call.getOperandBundlesAsDefs(Defs&: OpBundles); |
4734 | |
4735 | Instruction *NewCaller; |
4736 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &Call)) { |
4737 | NewCaller = InvokeInst::Create(Ty: NewFTy, Func: NestF, IfNormal: II->getNormalDest(), |
4738 | IfException: II->getUnwindDest(), Args: NewArgs, Bundles: OpBundles); |
4739 | cast<InvokeInst>(Val: NewCaller)->setCallingConv(II->getCallingConv()); |
4740 | cast<InvokeInst>(Val: NewCaller)->setAttributes(NewPAL); |
4741 | } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(Val: &Call)) { |
4742 | NewCaller = |
4743 | CallBrInst::Create(Ty: NewFTy, Func: NestF, DefaultDest: CBI->getDefaultDest(), |
4744 | IndirectDests: CBI->getIndirectDests(), Args: NewArgs, Bundles: OpBundles); |
4745 | cast<CallBrInst>(Val: NewCaller)->setCallingConv(CBI->getCallingConv()); |
4746 | cast<CallBrInst>(Val: NewCaller)->setAttributes(NewPAL); |
4747 | } else { |
4748 | NewCaller = CallInst::Create(Ty: NewFTy, Func: NestF, Args: NewArgs, Bundles: OpBundles); |
4749 | cast<CallInst>(Val: NewCaller)->setTailCallKind( |
4750 | cast<CallInst>(Val&: Call).getTailCallKind()); |
4751 | cast<CallInst>(Val: NewCaller)->setCallingConv( |
4752 | cast<CallInst>(Val&: Call).getCallingConv()); |
4753 | cast<CallInst>(Val: NewCaller)->setAttributes(NewPAL); |
4754 | } |
4755 | NewCaller->setDebugLoc(Call.getDebugLoc()); |
4756 | |
4757 | return NewCaller; |
4758 | } |
4759 | } |
4760 | |
4761 | // Replace the trampoline call with a direct call. Since there is no 'nest' |
4762 | // parameter, there is no need to adjust the argument list. Let the generic |
4763 | // code sort out any function type mismatches. |
4764 | Call.setCalledFunction(FTy, Fn: NestF); |
4765 | return &Call; |
4766 | } |
4767 | |