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/MathExtras.h"
66#include "llvm/Support/raw_ostream.h"
67#include "llvm/Transforms/InstCombine/InstCombiner.h"
68#include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
69#include "llvm/Transforms/Utils/Local.h"
70#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
71#include <algorithm>
72#include <cassert>
73#include <cstdint>
74#include <optional>
75#include <utility>
76#include <vector>
77
78#define DEBUG_TYPE "instcombine"
79#include "llvm/Transforms/Utils/InstructionWorklist.h"
80
81using namespace llvm;
82using namespace PatternMatch;
83
84STATISTIC(NumSimplified, "Number of library calls simplified");
85
86static cl::opt<unsigned> GuardWideningWindow(
87 "instcombine-guard-widening-window",
88 cl::init(Val: 3),
89 cl::desc("How wide an instruction window to bypass looking for "
90 "another guard"));
91
92/// Return the specified type promoted as it would be to pass though a va_arg
93/// area.
94static Type *getPromotedType(Type *Ty) {
95 if (IntegerType* ITy = dyn_cast<IntegerType>(Val: Ty)) {
96 if (ITy->getBitWidth() < 32)
97 return Type::getInt32Ty(C&: Ty->getContext());
98 }
99 return Ty;
100}
101
102/// Recognize a memcpy/memmove from a trivially otherwise unused alloca.
103/// TODO: This should probably be integrated with visitAllocSites, but that
104/// requires a deeper change to allow either unread or unwritten objects.
105static bool hasUndefSource(AnyMemTransferInst *MI) {
106 auto *Src = MI->getRawSource();
107 while (isa<GetElementPtrInst>(Val: Src) || isa<BitCastInst>(Val: Src)) {
108 if (!Src->hasOneUse())
109 return false;
110 Src = cast<Instruction>(Val: Src)->getOperand(i: 0);
111 }
112 return isa<AllocaInst>(Val: Src) && Src->hasOneUse();
113}
114
115Instruction *InstCombinerImpl::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) {
116 Align DstAlign = getKnownAlignment(V: MI->getRawDest(), DL, CxtI: MI, AC: &AC, DT: &DT);
117 MaybeAlign CopyDstAlign = MI->getDestAlign();
118 if (!CopyDstAlign || *CopyDstAlign < DstAlign) {
119 MI->setDestAlignment(DstAlign);
120 return MI;
121 }
122
123 Align SrcAlign = getKnownAlignment(V: MI->getRawSource(), DL, CxtI: MI, AC: &AC, DT: &DT);
124 MaybeAlign CopySrcAlign = MI->getSourceAlign();
125 if (!CopySrcAlign || *CopySrcAlign < SrcAlign) {
126 MI->setSourceAlignment(SrcAlign);
127 return MI;
128 }
129
130 // If we have a store to a location which is known constant, we can conclude
131 // that the store must be storing the constant value (else the memory
132 // wouldn't be constant), and this must be a noop.
133 if (!isModSet(MRI: AA->getModRefInfoMask(P: MI->getDest()))) {
134 // Set the size of the copy to 0, it will be deleted on the next iteration.
135 MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType()));
136 return MI;
137 }
138
139 // If the source is provably undef, the memcpy/memmove doesn't do anything
140 // (unless the transfer is volatile).
141 if (hasUndefSource(MI) && !MI->isVolatile()) {
142 // Set the size of the copy to 0, it will be deleted on the next iteration.
143 MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType()));
144 return MI;
145 }
146
147 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
148 // load/store.
149 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(Val: MI->getLength());
150 if (!MemOpLength) return nullptr;
151
152 // Source and destination pointer types are always "i8*" for intrinsic. See
153 // if the size is something we can handle with a single primitive load/store.
154 // A single load+store correctly handles overlapping memory in the memmove
155 // case.
156 uint64_t Size = MemOpLength->getLimitedValue();
157 assert(Size && "0-sized memory transferring should be removed already.");
158
159 if (Size > 8 || (Size&(Size-1)))
160 return nullptr; // If not 1/2/4/8 bytes, exit.
161
162 // If it is an atomic and alignment is less than the size then we will
163 // introduce the unaligned memory access which will be later transformed
164 // into libcall in CodeGen. This is not evident performance gain so disable
165 // it now.
166 if (isa<AtomicMemTransferInst>(Val: MI))
167 if (*CopyDstAlign < Size || *CopySrcAlign < Size)
168 return nullptr;
169
170 // Use an integer load+store unless we can find something better.
171 IntegerType* IntType = IntegerType::get(C&: MI->getContext(), NumBits: Size<<3);
172
173 // If the memcpy has metadata describing the members, see if we can get the
174 // TBAA, scope and noalias tags describing our copy.
175 AAMDNodes AACopyMD = MI->getAAMetadata().adjustForAccess(AccessSize: Size);
176
177 Value *Src = MI->getArgOperand(i: 1);
178 Value *Dest = MI->getArgOperand(i: 0);
179 LoadInst *L = Builder.CreateLoad(Ty: IntType, Ptr: Src);
180 // Alignment from the mem intrinsic will be better, so use it.
181 L->setAlignment(*CopySrcAlign);
182 L->setAAMetadata(AACopyMD);
183 MDNode *LoopMemParallelMD =
184 MI->getMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access);
185 if (LoopMemParallelMD)
186 L->setMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access, Node: LoopMemParallelMD);
187 MDNode *AccessGroupMD = MI->getMetadata(KindID: LLVMContext::MD_access_group);
188 if (AccessGroupMD)
189 L->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessGroupMD);
190
191 StoreInst *S = Builder.CreateStore(Val: L, Ptr: Dest);
192 // Alignment from the mem intrinsic will be better, so use it.
193 S->setAlignment(*CopyDstAlign);
194 S->setAAMetadata(AACopyMD);
195 if (LoopMemParallelMD)
196 S->setMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access, Node: LoopMemParallelMD);
197 if (AccessGroupMD)
198 S->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessGroupMD);
199 S->copyMetadata(SrcInst: *MI, WL: LLVMContext::MD_DIAssignID);
200
201 if (auto *MT = dyn_cast<MemTransferInst>(Val: MI)) {
202 // non-atomics can be volatile
203 L->setVolatile(MT->isVolatile());
204 S->setVolatile(MT->isVolatile());
205 }
206 if (isa<AtomicMemTransferInst>(Val: MI)) {
207 // atomics have to be unordered
208 L->setOrdering(AtomicOrdering::Unordered);
209 S->setOrdering(AtomicOrdering::Unordered);
210 }
211
212 // Set the size of the copy to 0, it will be deleted on the next iteration.
213 MI->setLength(Constant::getNullValue(Ty: MemOpLength->getType()));
214 return MI;
215}
216
217Instruction *InstCombinerImpl::SimplifyAnyMemSet(AnyMemSetInst *MI) {
218 const Align KnownAlignment =
219 getKnownAlignment(V: MI->getDest(), DL, CxtI: MI, AC: &AC, DT: &DT);
220 MaybeAlign MemSetAlign = MI->getDestAlign();
221 if (!MemSetAlign || *MemSetAlign < KnownAlignment) {
222 MI->setDestAlignment(KnownAlignment);
223 return MI;
224 }
225
226 // If we have a store to a location which is known constant, we can conclude
227 // that the store must be storing the constant value (else the memory
228 // wouldn't be constant), and this must be a noop.
229 if (!isModSet(MRI: AA->getModRefInfoMask(P: MI->getDest()))) {
230 // Set the size of the copy to 0, it will be deleted on the next iteration.
231 MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType()));
232 return MI;
233 }
234
235 // Remove memset with an undef value.
236 // FIXME: This is technically incorrect because it might overwrite a poison
237 // value. Change to PoisonValue once #52930 is resolved.
238 if (isa<UndefValue>(Val: MI->getValue())) {
239 // Set the size of the copy to 0, it will be deleted on the next iteration.
240 MI->setLength(Constant::getNullValue(Ty: MI->getLength()->getType()));
241 return MI;
242 }
243
244 // Extract the length and alignment and fill if they are constant.
245 ConstantInt *LenC = dyn_cast<ConstantInt>(Val: MI->getLength());
246 ConstantInt *FillC = dyn_cast<ConstantInt>(Val: MI->getValue());
247 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(Bitwidth: 8))
248 return nullptr;
249 const uint64_t Len = LenC->getLimitedValue();
250 assert(Len && "0-sized memory setting should be removed already.");
251 const Align Alignment = MI->getDestAlign().valueOrOne();
252
253 // If it is an atomic and alignment is less than the size then we will
254 // introduce the unaligned memory access which will be later transformed
255 // into libcall in CodeGen. This is not evident performance gain so disable
256 // it now.
257 if (isa<AtomicMemSetInst>(Val: MI))
258 if (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 Type *ITy = IntegerType::get(C&: MI->getContext(), NumBits: Len*8); // n=1 -> i8.
264
265 Value *Dest = MI->getDest();
266
267 // Extract the fill value and store.
268 const uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
269 Constant *FillVal = ConstantInt::get(Ty: ITy, V: Fill);
270 StoreInst *S = Builder.CreateStore(Val: FillVal, Ptr: Dest, isVolatile: MI->isVolatile());
271 S->copyMetadata(SrcInst: *MI, WL: LLVMContext::MD_DIAssignID);
272 auto replaceOpForAssignmentMarkers = [FillC, FillVal](auto *DbgAssign) {
273 if (llvm::is_contained(DbgAssign->location_ops(), FillC))
274 DbgAssign->replaceVariableLocationOp(FillC, FillVal);
275 };
276 for_each(Range: at::getAssignmentMarkers(Inst: S), F: replaceOpForAssignmentMarkers);
277 for_each(Range: at::getDVRAssignmentMarkers(Inst: S), F: replaceOpForAssignmentMarkers);
278
279 S->setAlignment(Alignment);
280 if (isa<AtomicMemSetInst>(Val: MI))
281 S->setOrdering(AtomicOrdering::Unordered);
282
283 // Set the size of the copy to 0, it will be deleted on the next iteration.
284 MI->setLength(Constant::getNullValue(Ty: LenC->getType()));
285 return MI;
286 }
287
288 return nullptr;
289}
290
291// TODO, Obvious Missing Transforms:
292// * Narrow width by halfs excluding zero/undef lanes
293Value *InstCombinerImpl::simplifyMaskedLoad(IntrinsicInst &II) {
294 Value *LoadPtr = II.getArgOperand(i: 0);
295 const Align Alignment =
296 cast<ConstantInt>(Val: II.getArgOperand(i: 1))->getAlignValue();
297
298 // If the mask is all ones or undefs, this is a plain vector load of the 1st
299 // argument.
300 if (maskIsAllOneOrUndef(Mask: II.getArgOperand(i: 2))) {
301 LoadInst *L = Builder.CreateAlignedLoad(Ty: II.getType(), Ptr: LoadPtr, Align: Alignment,
302 Name: "unmaskedload");
303 L->copyMetadata(SrcInst: II);
304 return L;
305 }
306
307 // If we can unconditionally load from this address, replace with a
308 // load/select idiom. TODO: use DT for context sensitive query
309 if (isDereferenceablePointer(V: LoadPtr, Ty: II.getType(),
310 DL: II.getDataLayout(), CtxI: &II, AC: &AC)) {
311 LoadInst *LI = Builder.CreateAlignedLoad(Ty: II.getType(), Ptr: LoadPtr, Align: Alignment,
312 Name: "unmaskedload");
313 LI->copyMetadata(SrcInst: II);
314 return Builder.CreateSelect(C: II.getArgOperand(i: 2), True: LI, False: II.getArgOperand(i: 3));
315 }
316
317 return nullptr;
318}
319
320// TODO, Obvious Missing Transforms:
321// * Single constant active lane -> store
322// * Narrow width by halfs excluding zero/undef lanes
323Instruction *InstCombinerImpl::simplifyMaskedStore(IntrinsicInst &II) {
324 auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 3));
325 if (!ConstMask)
326 return nullptr;
327
328 // If the mask is all zeros, this instruction does nothing.
329 if (ConstMask->isNullValue())
330 return eraseInstFromFunction(I&: II);
331
332 // If the mask is all ones, this is a plain vector store of the 1st argument.
333 if (ConstMask->isAllOnesValue()) {
334 Value *StorePtr = II.getArgOperand(i: 1);
335 Align Alignment = cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue();
336 StoreInst *S =
337 new StoreInst(II.getArgOperand(i: 0), StorePtr, false, Alignment);
338 S->copyMetadata(SrcInst: II);
339 return S;
340 }
341
342 if (isa<ScalableVectorType>(Val: ConstMask->getType()))
343 return nullptr;
344
345 // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
346 APInt DemandedElts = possiblyDemandedEltsInMask(Mask: ConstMask);
347 APInt PoisonElts(DemandedElts.getBitWidth(), 0);
348 if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 0), DemandedElts,
349 PoisonElts))
350 return replaceOperand(I&: II, OpNum: 0, V);
351
352 return nullptr;
353}
354
355// TODO, Obvious Missing Transforms:
356// * Single constant active lane load -> load
357// * Dereferenceable address & few lanes -> scalarize speculative load/selects
358// * Adjacent vector addresses -> masked.load
359// * Narrow width by halfs excluding zero/undef lanes
360// * Vector incrementing address -> vector masked load
361Instruction *InstCombinerImpl::simplifyMaskedGather(IntrinsicInst &II) {
362 auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 2));
363 if (!ConstMask)
364 return nullptr;
365
366 // Vector splat address w/known mask -> scalar load
367 // Fold the gather to load the source vector first lane
368 // because it is reloading the same value each time
369 if (ConstMask->isAllOnesValue())
370 if (auto *SplatPtr = getSplatValue(V: II.getArgOperand(i: 0))) {
371 auto *VecTy = cast<VectorType>(Val: II.getType());
372 const Align Alignment =
373 cast<ConstantInt>(Val: II.getArgOperand(i: 1))->getAlignValue();
374 LoadInst *L = Builder.CreateAlignedLoad(Ty: VecTy->getElementType(), Ptr: SplatPtr,
375 Align: Alignment, Name: "load.scalar");
376 Value *Shuf =
377 Builder.CreateVectorSplat(EC: VecTy->getElementCount(), V: L, Name: "broadcast");
378 return replaceInstUsesWith(I&: II, V: cast<Instruction>(Val: Shuf));
379 }
380
381 return nullptr;
382}
383
384// TODO, Obvious Missing Transforms:
385// * Single constant active lane -> store
386// * Adjacent vector addresses -> masked.store
387// * Narrow store width by halfs excluding zero/undef lanes
388// * Vector incrementing address -> vector masked store
389Instruction *InstCombinerImpl::simplifyMaskedScatter(IntrinsicInst &II) {
390 auto *ConstMask = dyn_cast<Constant>(Val: II.getArgOperand(i: 3));
391 if (!ConstMask)
392 return nullptr;
393
394 // If the mask is all zeros, a scatter does nothing.
395 if (ConstMask->isNullValue())
396 return eraseInstFromFunction(I&: II);
397
398 // Vector splat address -> scalar store
399 if (auto *SplatPtr = getSplatValue(V: II.getArgOperand(i: 1))) {
400 // scatter(splat(value), splat(ptr), non-zero-mask) -> store value, ptr
401 if (auto *SplatValue = getSplatValue(V: II.getArgOperand(i: 0))) {
402 if (maskContainsAllOneOrUndef(Mask: ConstMask)) {
403 Align Alignment =
404 cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue();
405 StoreInst *S = new StoreInst(SplatValue, SplatPtr, /*IsVolatile=*/false,
406 Alignment);
407 S->copyMetadata(SrcInst: II);
408 return S;
409 }
410 }
411 // scatter(vector, splat(ptr), splat(true)) -> store extract(vector,
412 // lastlane), ptr
413 if (ConstMask->isAllOnesValue()) {
414 Align Alignment = cast<ConstantInt>(Val: II.getArgOperand(i: 2))->getAlignValue();
415 VectorType *WideLoadTy = cast<VectorType>(Val: II.getArgOperand(i: 1)->getType());
416 ElementCount VF = WideLoadTy->getElementCount();
417 Value *RunTimeVF = Builder.CreateElementCount(DstType: Builder.getInt32Ty(), EC: VF);
418 Value *LastLane = Builder.CreateSub(LHS: RunTimeVF, RHS: Builder.getInt32(C: 1));
419 Value *Extract =
420 Builder.CreateExtractElement(Vec: II.getArgOperand(i: 0), Idx: LastLane);
421 StoreInst *S =
422 new StoreInst(Extract, SplatPtr, /*IsVolatile=*/false, Alignment);
423 S->copyMetadata(SrcInst: II);
424 return S;
425 }
426 }
427 if (isa<ScalableVectorType>(Val: ConstMask->getType()))
428 return nullptr;
429
430 // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
431 APInt DemandedElts = possiblyDemandedEltsInMask(Mask: ConstMask);
432 APInt PoisonElts(DemandedElts.getBitWidth(), 0);
433 if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 0), DemandedElts,
434 PoisonElts))
435 return replaceOperand(I&: II, OpNum: 0, V);
436 if (Value *V = SimplifyDemandedVectorElts(V: II.getOperand(i_nocapture: 1), DemandedElts,
437 PoisonElts))
438 return replaceOperand(I&: II, OpNum: 1, V);
439
440 return nullptr;
441}
442
443/// This function transforms launder.invariant.group and strip.invariant.group
444/// like:
445/// launder(launder(%x)) -> launder(%x) (the result is not the argument)
446/// launder(strip(%x)) -> launder(%x)
447/// strip(strip(%x)) -> strip(%x) (the result is not the argument)
448/// strip(launder(%x)) -> strip(%x)
449/// This is legal because it preserves the most recent information about
450/// the presence or absence of invariant.group.
451static Instruction *simplifyInvariantGroupIntrinsic(IntrinsicInst &II,
452 InstCombinerImpl &IC) {
453 auto *Arg = II.getArgOperand(i: 0);
454 auto *StrippedArg = Arg->stripPointerCasts();
455 auto *StrippedInvariantGroupsArg = StrippedArg;
456 while (auto *Intr = dyn_cast<IntrinsicInst>(Val: StrippedInvariantGroupsArg)) {
457 if (Intr->getIntrinsicID() != Intrinsic::launder_invariant_group &&
458 Intr->getIntrinsicID() != Intrinsic::strip_invariant_group)
459 break;
460 StrippedInvariantGroupsArg = Intr->getArgOperand(i: 0)->stripPointerCasts();
461 }
462 if (StrippedArg == StrippedInvariantGroupsArg)
463 return nullptr; // No launders/strips to remove.
464
465 Value *Result = nullptr;
466
467 if (II.getIntrinsicID() == Intrinsic::launder_invariant_group)
468 Result = IC.Builder.CreateLaunderInvariantGroup(Ptr: StrippedInvariantGroupsArg);
469 else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group)
470 Result = IC.Builder.CreateStripInvariantGroup(Ptr: StrippedInvariantGroupsArg);
471 else
472 llvm_unreachable(
473 "simplifyInvariantGroupIntrinsic only handles launder and strip");
474 if (Result->getType()->getPointerAddressSpace() !=
475 II.getType()->getPointerAddressSpace())
476 Result = IC.Builder.CreateAddrSpaceCast(V: Result, DestTy: II.getType());
477
478 return cast<Instruction>(Val: Result);
479}
480
481static Instruction *foldCttzCtlz(IntrinsicInst &II, InstCombinerImpl &IC) {
482 assert((II.getIntrinsicID() == Intrinsic::cttz ||
483 II.getIntrinsicID() == Intrinsic::ctlz) &&
484 "Expected cttz or ctlz intrinsic");
485 bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz;
486 Value *Op0 = II.getArgOperand(i: 0);
487 Value *Op1 = II.getArgOperand(i: 1);
488 Value *X;
489 // ctlz(bitreverse(x)) -> cttz(x)
490 // cttz(bitreverse(x)) -> ctlz(x)
491 if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X)))) {
492 Intrinsic::ID ID = IsTZ ? Intrinsic::ctlz : Intrinsic::cttz;
493 Function *F = Intrinsic::getDeclaration(M: II.getModule(), id: ID, Tys: II.getType());
494 return CallInst::Create(Func: F, Args: {X, II.getArgOperand(i: 1)});
495 }
496
497 if (II.getType()->isIntOrIntVectorTy(BitWidth: 1)) {
498 // ctlz/cttz i1 Op0 --> not Op0
499 if (match(V: Op1, P: m_Zero()))
500 return BinaryOperator::CreateNot(Op: Op0);
501 // If zero is poison, then the input can be assumed to be "true", so the
502 // instruction simplifies to "false".
503 assert(match(Op1, m_One()) && "Expected ctlz/cttz operand to be 0 or 1");
504 return IC.replaceInstUsesWith(I&: II, V: ConstantInt::getNullValue(Ty: II.getType()));
505 }
506
507 // If ctlz/cttz is only used as a shift amount, set is_zero_poison to true.
508 if (II.hasOneUse() && match(V: Op1, P: m_Zero()) &&
509 match(V: II.user_back(), P: m_Shift(L: m_Value(), R: m_Specific(V: &II))))
510 return IC.replaceOperand(I&: II, OpNum: 1, V: IC.Builder.getTrue());
511
512 Constant *C;
513
514 if (IsTZ) {
515 // cttz(-x) -> cttz(x)
516 if (match(V: Op0, P: m_Neg(V: m_Value(V&: X))))
517 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
518
519 // cttz(-x & x) -> cttz(x)
520 if (match(V: Op0, P: m_c_And(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X))))
521 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
522
523 // cttz(sext(x)) -> cttz(zext(x))
524 if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))))) {
525 auto *Zext = IC.Builder.CreateZExt(V: X, DestTy: II.getType());
526 auto *CttzZext =
527 IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: Zext, RHS: Op1);
528 return IC.replaceInstUsesWith(I&: II, V: CttzZext);
529 }
530
531 // Zext doesn't change the number of trailing zeros, so narrow:
532 // cttz(zext(x)) -> zext(cttz(x)) if the 'ZeroIsPoison' parameter is 'true'.
533 if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && match(V: Op1, P: m_One())) {
534 auto *Cttz = IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: X,
535 RHS: IC.Builder.getTrue());
536 auto *ZextCttz = IC.Builder.CreateZExt(V: Cttz, DestTy: II.getType());
537 return IC.replaceInstUsesWith(I&: II, V: ZextCttz);
538 }
539
540 // cttz(abs(x)) -> cttz(x)
541 // cttz(nabs(x)) -> cttz(x)
542 Value *Y;
543 SelectPatternFlavor SPF = matchSelectPattern(V: Op0, LHS&: X, RHS&: Y).Flavor;
544 if (SPF == SPF_ABS || SPF == SPF_NABS)
545 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
546
547 if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X))))
548 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
549
550 // cttz(shl(%const, %val), 1) --> add(cttz(%const, 1), %val)
551 if (match(V: Op0, P: m_Shl(L: m_ImmConstant(C), R: m_Value(V&: X))) &&
552 match(V: Op1, P: m_One())) {
553 Value *ConstCttz =
554 IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: C, RHS: Op1);
555 return BinaryOperator::CreateAdd(V1: ConstCttz, V2: X);
556 }
557
558 // cttz(lshr exact (%const, %val), 1) --> sub(cttz(%const, 1), %val)
559 if (match(V: Op0, P: m_Exact(SubPattern: m_LShr(L: m_ImmConstant(C), R: m_Value(V&: X)))) &&
560 match(V: Op1, P: m_One())) {
561 Value *ConstCttz =
562 IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::cttz, LHS: C, RHS: Op1);
563 return BinaryOperator::CreateSub(V1: ConstCttz, V2: X);
564 }
565
566 // cttz(add(lshr(UINT_MAX, %val), 1)) --> sub(width, %val)
567 if (match(V: Op0, P: m_Add(L: m_LShr(L: m_AllOnes(), R: m_Value(V&: X)), R: m_One()))) {
568 Value *Width =
569 ConstantInt::get(Ty: II.getType(), V: II.getType()->getScalarSizeInBits());
570 return BinaryOperator::CreateSub(V1: Width, V2: X);
571 }
572 } else {
573 // ctlz(lshr(%const, %val), 1) --> add(ctlz(%const, 1), %val)
574 if (match(V: Op0, P: m_LShr(L: m_ImmConstant(C), R: m_Value(V&: X))) &&
575 match(V: Op1, P: m_One())) {
576 Value *ConstCtlz =
577 IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::ctlz, LHS: C, RHS: Op1);
578 return BinaryOperator::CreateAdd(V1: ConstCtlz, V2: X);
579 }
580
581 // ctlz(shl nuw (%const, %val), 1) --> sub(ctlz(%const, 1), %val)
582 if (match(V: Op0, P: m_NUWShl(L: m_ImmConstant(C), R: m_Value(V&: X))) &&
583 match(V: Op1, P: m_One())) {
584 Value *ConstCtlz =
585 IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::ctlz, LHS: C, RHS: Op1);
586 return BinaryOperator::CreateSub(V1: ConstCtlz, V2: X);
587 }
588 }
589
590 KnownBits Known = IC.computeKnownBits(V: Op0, Depth: 0, CxtI: &II);
591
592 // Create a mask for bits above (ctlz) or below (cttz) the first known one.
593 unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros()
594 : Known.countMaxLeadingZeros();
595 unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros()
596 : Known.countMinLeadingZeros();
597
598 // If all bits above (ctlz) or below (cttz) the first known one are known
599 // zero, this value is constant.
600 // FIXME: This should be in InstSimplify because we're replacing an
601 // instruction with a constant.
602 if (PossibleZeros == DefiniteZeros) {
603 auto *C = ConstantInt::get(Ty: Op0->getType(), V: DefiniteZeros);
604 return IC.replaceInstUsesWith(I&: II, V: C);
605 }
606
607 // If the input to cttz/ctlz is known to be non-zero,
608 // then change the 'ZeroIsPoison' parameter to 'true'
609 // because we know the zero behavior can't affect the result.
610 if (!Known.One.isZero() ||
611 isKnownNonZero(V: Op0, Q: IC.getSimplifyQuery().getWithInstruction(I: &II))) {
612 if (!match(V: II.getArgOperand(i: 1), P: m_One()))
613 return IC.replaceOperand(I&: II, OpNum: 1, V: IC.Builder.getTrue());
614 }
615
616 // Add range attribute since known bits can't completely reflect what we know.
617 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
618 if (BitWidth != 1 && !II.hasRetAttr(Kind: Attribute::Range) &&
619 !II.getMetadata(KindID: LLVMContext::MD_range)) {
620 ConstantRange Range(APInt(BitWidth, DefiniteZeros),
621 APInt(BitWidth, PossibleZeros + 1));
622 II.addRangeRetAttr(CR: Range);
623 return &II;
624 }
625
626 return nullptr;
627}
628
629static Instruction *foldCtpop(IntrinsicInst &II, InstCombinerImpl &IC) {
630 assert(II.getIntrinsicID() == Intrinsic::ctpop &&
631 "Expected ctpop intrinsic");
632 Type *Ty = II.getType();
633 unsigned BitWidth = Ty->getScalarSizeInBits();
634 Value *Op0 = II.getArgOperand(i: 0);
635 Value *X, *Y;
636
637 // ctpop(bitreverse(x)) -> ctpop(x)
638 // ctpop(bswap(x)) -> ctpop(x)
639 if (match(V: Op0, P: m_BitReverse(Op0: m_Value(V&: X))) || match(V: Op0, P: m_BSwap(Op0: m_Value(V&: X))))
640 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
641
642 // ctpop(rot(x)) -> ctpop(x)
643 if ((match(V: Op0, P: m_FShl(Op0: m_Value(V&: X), Op1: m_Value(V&: Y), Op2: m_Value())) ||
644 match(V: Op0, P: m_FShr(Op0: m_Value(V&: X), Op1: m_Value(V&: Y), Op2: m_Value()))) &&
645 X == Y)
646 return IC.replaceOperand(I&: II, OpNum: 0, V: X);
647
648 // ctpop(x | -x) -> bitwidth - cttz(x, false)
649 if (Op0->hasOneUse() &&
650 match(V: Op0, P: m_c_Or(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))))) {
651 Function *F =
652 Intrinsic::getDeclaration(M: II.getModule(), id: Intrinsic::cttz, Tys: Ty);
653 auto *Cttz = IC.Builder.CreateCall(Callee: F, Args: {X, IC.Builder.getFalse()});
654 auto *Bw = ConstantInt::get(Ty, V: APInt(BitWidth, BitWidth));
655 return IC.replaceInstUsesWith(I&: II, V: IC.Builder.CreateSub(LHS: Bw, RHS: Cttz));
656 }
657
658 // ctpop(~x & (x - 1)) -> cttz(x, false)
659 if (match(V: Op0,
660 P: m_c_And(L: m_Not(V: m_Value(V&: X)), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes())))) {
661 Function *F =
662 Intrinsic::getDeclaration(M: II.getModule(), id: Intrinsic::cttz, Tys: Ty);
663 return CallInst::Create(Func: F, Args: {X, IC.Builder.getFalse()});
664 }
665
666 // Zext doesn't change the number of set bits, so narrow:
667 // ctpop (zext X) --> zext (ctpop X)
668 if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))) {
669 Value *NarrowPop = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V: X);
670 return CastInst::Create(Instruction::ZExt, S: NarrowPop, Ty);
671 }
672
673 KnownBits Known(BitWidth);
674 IC.computeKnownBits(V: Op0, Known, Depth: 0, CxtI: &II);
675
676 // If all bits are zero except for exactly one fixed bit, then the result
677 // must be 0 or 1, and we can get that answer by shifting to LSB:
678 // ctpop (X & 32) --> (X & 32) >> 5
679 // TODO: Investigate removing this as its likely unnecessary given the below
680 // `isKnownToBeAPowerOfTwo` check.
681 if ((~Known.Zero).isPowerOf2())
682 return BinaryOperator::CreateLShr(
683 V1: Op0, V2: ConstantInt::get(Ty, V: (~Known.Zero).exactLogBase2()));
684
685 // More generally we can also handle non-constant power of 2 patterns such as
686 // shl/shr(Pow2, X), (X & -X), etc... by transforming:
687 // ctpop(Pow2OrZero) --> icmp ne X, 0
688 if (IC.isKnownToBeAPowerOfTwo(V: Op0, /* OrZero */ true))
689 return CastInst::Create(Instruction::ZExt,
690 S: IC.Builder.CreateICmp(P: ICmpInst::ICMP_NE, LHS: Op0,
691 RHS: Constant::getNullValue(Ty)),
692 Ty);
693
694 // Add range attribute since known bits can't completely reflect what we know.
695 if (BitWidth != 1 && !II.hasRetAttr(Kind: Attribute::Range) &&
696 !II.getMetadata(KindID: LLVMContext::MD_range)) {
697 ConstantRange Range(APInt(BitWidth, Known.countMinPopulation()),
698 APInt(BitWidth, Known.countMaxPopulation() + 1));
699 II.addRangeRetAttr(CR: Range);
700 return &II;
701 }
702
703 return nullptr;
704}
705
706/// Convert a table lookup to shufflevector if the mask is constant.
707/// This could benefit tbl1 if the mask is { 7,6,5,4,3,2,1,0 }, in
708/// which case we could lower the shufflevector with rev64 instructions
709/// as it's actually a byte reverse.
710static Value *simplifyNeonTbl1(const IntrinsicInst &II,
711 InstCombiner::BuilderTy &Builder) {
712 // Bail out if the mask is not a constant.
713 auto *C = dyn_cast<Constant>(Val: II.getArgOperand(i: 1));
714 if (!C)
715 return nullptr;
716
717 auto *VecTy = cast<FixedVectorType>(Val: II.getType());
718 unsigned NumElts = VecTy->getNumElements();
719
720 // Only perform this transformation for <8 x i8> vector types.
721 if (!VecTy->getElementType()->isIntegerTy(Bitwidth: 8) || NumElts != 8)
722 return nullptr;
723
724 int Indexes[8];
725
726 for (unsigned I = 0; I < NumElts; ++I) {
727 Constant *COp = C->getAggregateElement(Elt: I);
728
729 if (!COp || !isa<ConstantInt>(Val: COp))
730 return nullptr;
731
732 Indexes[I] = cast<ConstantInt>(Val: COp)->getLimitedValue();
733
734 // Make sure the mask indices are in range.
735 if ((unsigned)Indexes[I] >= NumElts)
736 return nullptr;
737 }
738
739 auto *V1 = II.getArgOperand(i: 0);
740 auto *V2 = Constant::getNullValue(Ty: V1->getType());
741 return Builder.CreateShuffleVector(V1, V2, Mask: ArrayRef(Indexes));
742}
743
744// Returns true iff the 2 intrinsics have the same operands, limiting the
745// comparison to the first NumOperands.
746static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E,
747 unsigned NumOperands) {
748 assert(I.arg_size() >= NumOperands && "Not enough operands");
749 assert(E.arg_size() >= NumOperands && "Not enough operands");
750 for (unsigned i = 0; i < NumOperands; i++)
751 if (I.getArgOperand(i) != E.getArgOperand(i))
752 return false;
753 return true;
754}
755
756// Remove trivially empty start/end intrinsic ranges, i.e. a start
757// immediately followed by an end (ignoring debuginfo or other
758// start/end intrinsics in between). As this handles only the most trivial
759// cases, tracking the nesting level is not needed:
760//
761// call @llvm.foo.start(i1 0)
762// call @llvm.foo.start(i1 0) ; This one won't be skipped: it will be removed
763// call @llvm.foo.end(i1 0)
764// call @llvm.foo.end(i1 0) ; &I
765static bool
766removeTriviallyEmptyRange(IntrinsicInst &EndI, InstCombinerImpl &IC,
767 std::function<bool(const IntrinsicInst &)> IsStart) {
768 // We start from the end intrinsic and scan backwards, so that InstCombine
769 // has already processed (and potentially removed) all the instructions
770 // before the end intrinsic.
771 BasicBlock::reverse_iterator BI(EndI), BE(EndI.getParent()->rend());
772 for (; BI != BE; ++BI) {
773 if (auto *I = dyn_cast<IntrinsicInst>(Val: &*BI)) {
774 if (I->isDebugOrPseudoInst() ||
775 I->getIntrinsicID() == EndI.getIntrinsicID())
776 continue;
777 if (IsStart(*I)) {
778 if (haveSameOperands(I: EndI, E: *I, NumOperands: EndI.arg_size())) {
779 IC.eraseInstFromFunction(I&: *I);
780 IC.eraseInstFromFunction(I&: EndI);
781 return true;
782 }
783 // Skip start intrinsics that don't pair with this end intrinsic.
784 continue;
785 }
786 }
787 break;
788 }
789
790 return false;
791}
792
793Instruction *InstCombinerImpl::visitVAEndInst(VAEndInst &I) {
794 removeTriviallyEmptyRange(EndI&: I, IC&: *this, IsStart: [](const IntrinsicInst &I) {
795 return I.getIntrinsicID() == Intrinsic::vastart ||
796 I.getIntrinsicID() == Intrinsic::vacopy;
797 });
798 return nullptr;
799}
800
801static CallInst *canonicalizeConstantArg0ToArg1(CallInst &Call) {
802 assert(Call.arg_size() > 1 && "Need at least 2 args to swap");
803 Value *Arg0 = Call.getArgOperand(i: 0), *Arg1 = Call.getArgOperand(i: 1);
804 if (isa<Constant>(Val: Arg0) && !isa<Constant>(Val: Arg1)) {
805 Call.setArgOperand(i: 0, v: Arg1);
806 Call.setArgOperand(i: 1, v: Arg0);
807 return &Call;
808 }
809 return nullptr;
810}
811
812/// Creates a result tuple for an overflow intrinsic \p II with a given
813/// \p Result and a constant \p Overflow value.
814static Instruction *createOverflowTuple(IntrinsicInst *II, Value *Result,
815 Constant *Overflow) {
816 Constant *V[] = {PoisonValue::get(T: Result->getType()), Overflow};
817 StructType *ST = cast<StructType>(Val: II->getType());
818 Constant *Struct = ConstantStruct::get(T: ST, V);
819 return InsertValueInst::Create(Agg: Struct, Val: Result, Idxs: 0);
820}
821
822Instruction *
823InstCombinerImpl::foldIntrinsicWithOverflowCommon(IntrinsicInst *II) {
824 WithOverflowInst *WO = cast<WithOverflowInst>(Val: II);
825 Value *OperationResult = nullptr;
826 Constant *OverflowResult = nullptr;
827 if (OptimizeOverflowCheck(BinaryOp: WO->getBinaryOp(), IsSigned: WO->isSigned(), LHS: WO->getLHS(),
828 RHS: WO->getRHS(), CtxI&: *WO, OperationResult, OverflowResult))
829 return createOverflowTuple(II: WO, Result: OperationResult, Overflow: OverflowResult);
830 return nullptr;
831}
832
833static bool inputDenormalIsIEEE(const Function &F, const Type *Ty) {
834 Ty = Ty->getScalarType();
835 return F.getDenormalMode(FPType: Ty->getFltSemantics()).Input == DenormalMode::IEEE;
836}
837
838static bool inputDenormalIsDAZ(const Function &F, const Type *Ty) {
839 Ty = Ty->getScalarType();
840 return F.getDenormalMode(FPType: Ty->getFltSemantics()).inputsAreZero();
841}
842
843/// \returns the compare predicate type if the test performed by
844/// llvm.is.fpclass(x, \p Mask) is equivalent to fcmp o__ x, 0.0 with the
845/// floating-point environment assumed for \p F for type \p Ty
846static FCmpInst::Predicate fpclassTestIsFCmp0(FPClassTest Mask,
847 const Function &F, Type *Ty) {
848 switch (static_cast<unsigned>(Mask)) {
849 case fcZero:
850 if (inputDenormalIsIEEE(F, Ty))
851 return FCmpInst::FCMP_OEQ;
852 break;
853 case fcZero | fcSubnormal:
854 if (inputDenormalIsDAZ(F, Ty))
855 return FCmpInst::FCMP_OEQ;
856 break;
857 case fcPositive | fcNegZero:
858 if (inputDenormalIsIEEE(F, Ty))
859 return FCmpInst::FCMP_OGE;
860 break;
861 case fcPositive | fcNegZero | fcNegSubnormal:
862 if (inputDenormalIsDAZ(F, Ty))
863 return FCmpInst::FCMP_OGE;
864 break;
865 case fcPosSubnormal | fcPosNormal | fcPosInf:
866 if (inputDenormalIsIEEE(F, Ty))
867 return FCmpInst::FCMP_OGT;
868 break;
869 case fcNegative | fcPosZero:
870 if (inputDenormalIsIEEE(F, Ty))
871 return FCmpInst::FCMP_OLE;
872 break;
873 case fcNegative | fcPosZero | fcPosSubnormal:
874 if (inputDenormalIsDAZ(F, Ty))
875 return FCmpInst::FCMP_OLE;
876 break;
877 case fcNegSubnormal | fcNegNormal | fcNegInf:
878 if (inputDenormalIsIEEE(F, Ty))
879 return FCmpInst::FCMP_OLT;
880 break;
881 case fcPosNormal | fcPosInf:
882 if (inputDenormalIsDAZ(F, Ty))
883 return FCmpInst::FCMP_OGT;
884 break;
885 case fcNegNormal | fcNegInf:
886 if (inputDenormalIsDAZ(F, Ty))
887 return FCmpInst::FCMP_OLT;
888 break;
889 case ~fcZero & ~fcNan:
890 if (inputDenormalIsIEEE(F, Ty))
891 return FCmpInst::FCMP_ONE;
892 break;
893 case ~(fcZero | fcSubnormal) & ~fcNan:
894 if (inputDenormalIsDAZ(F, Ty))
895 return FCmpInst::FCMP_ONE;
896 break;
897 default:
898 break;
899 }
900
901 return FCmpInst::BAD_FCMP_PREDICATE;
902}
903
904Instruction *InstCombinerImpl::foldIntrinsicIsFPClass(IntrinsicInst &II) {
905 Value *Src0 = II.getArgOperand(i: 0);
906 Value *Src1 = II.getArgOperand(i: 1);
907 const ConstantInt *CMask = cast<ConstantInt>(Val: Src1);
908 FPClassTest Mask = static_cast<FPClassTest>(CMask->getZExtValue());
909 const bool IsUnordered = (Mask & fcNan) == fcNan;
910 const bool IsOrdered = (Mask & fcNan) == fcNone;
911 const FPClassTest OrderedMask = Mask & ~fcNan;
912 const FPClassTest OrderedInvertedMask = ~OrderedMask & ~fcNan;
913
914 const bool IsStrict =
915 II.getFunction()->getAttributes().hasFnAttr(Kind: Attribute::StrictFP);
916
917 Value *FNegSrc;
918 if (match(V: Src0, P: m_FNeg(X: m_Value(V&: FNegSrc)))) {
919 // is.fpclass (fneg x), mask -> is.fpclass x, (fneg mask)
920
921 II.setArgOperand(i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: fneg(Mask)));
922 return replaceOperand(I&: II, OpNum: 0, V: FNegSrc);
923 }
924
925 Value *FAbsSrc;
926 if (match(V: Src0, P: m_FAbs(Op0: m_Value(V&: FAbsSrc)))) {
927 II.setArgOperand(i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: inverse_fabs(Mask)));
928 return replaceOperand(I&: II, OpNum: 0, V: FAbsSrc);
929 }
930
931 if ((OrderedMask == fcInf || OrderedInvertedMask == fcInf) &&
932 (IsOrdered || IsUnordered) && !IsStrict) {
933 // is.fpclass(x, fcInf) -> fcmp oeq fabs(x), +inf
934 // is.fpclass(x, ~fcInf) -> fcmp one fabs(x), +inf
935 // is.fpclass(x, fcInf|fcNan) -> fcmp ueq fabs(x), +inf
936 // is.fpclass(x, ~(fcInf|fcNan)) -> fcmp une fabs(x), +inf
937 Constant *Inf = ConstantFP::getInfinity(Ty: Src0->getType());
938 FCmpInst::Predicate Pred =
939 IsUnordered ? FCmpInst::FCMP_UEQ : FCmpInst::FCMP_OEQ;
940 if (OrderedInvertedMask == fcInf)
941 Pred = IsUnordered ? FCmpInst::FCMP_UNE : FCmpInst::FCMP_ONE;
942
943 Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Src0);
944 Value *CmpInf = Builder.CreateFCmp(P: Pred, LHS: Fabs, RHS: Inf);
945 CmpInf->takeName(V: &II);
946 return replaceInstUsesWith(I&: II, V: CmpInf);
947 }
948
949 if ((OrderedMask == fcPosInf || OrderedMask == fcNegInf) &&
950 (IsOrdered || IsUnordered) && !IsStrict) {
951 // is.fpclass(x, fcPosInf) -> fcmp oeq x, +inf
952 // is.fpclass(x, fcNegInf) -> fcmp oeq x, -inf
953 // is.fpclass(x, fcPosInf|fcNan) -> fcmp ueq x, +inf
954 // is.fpclass(x, fcNegInf|fcNan) -> fcmp ueq x, -inf
955 Constant *Inf =
956 ConstantFP::getInfinity(Ty: Src0->getType(), Negative: OrderedMask == fcNegInf);
957 Value *EqInf = IsUnordered ? Builder.CreateFCmpUEQ(LHS: Src0, RHS: Inf)
958 : Builder.CreateFCmpOEQ(LHS: Src0, RHS: Inf);
959
960 EqInf->takeName(V: &II);
961 return replaceInstUsesWith(I&: II, V: EqInf);
962 }
963
964 if ((OrderedInvertedMask == fcPosInf || OrderedInvertedMask == fcNegInf) &&
965 (IsOrdered || IsUnordered) && !IsStrict) {
966 // is.fpclass(x, ~fcPosInf) -> fcmp one x, +inf
967 // is.fpclass(x, ~fcNegInf) -> fcmp one x, -inf
968 // is.fpclass(x, ~fcPosInf|fcNan) -> fcmp une x, +inf
969 // is.fpclass(x, ~fcNegInf|fcNan) -> fcmp une x, -inf
970 Constant *Inf = ConstantFP::getInfinity(Ty: Src0->getType(),
971 Negative: OrderedInvertedMask == fcNegInf);
972 Value *NeInf = IsUnordered ? Builder.CreateFCmpUNE(LHS: Src0, RHS: Inf)
973 : Builder.CreateFCmpONE(LHS: Src0, RHS: Inf);
974 NeInf->takeName(V: &II);
975 return replaceInstUsesWith(I&: II, V: NeInf);
976 }
977
978 if (Mask == fcNan && !IsStrict) {
979 // Equivalent of isnan. Replace with standard fcmp if we don't care about FP
980 // exceptions.
981 Value *IsNan =
982 Builder.CreateFCmpUNO(LHS: Src0, RHS: ConstantFP::getZero(Ty: Src0->getType()));
983 IsNan->takeName(V: &II);
984 return replaceInstUsesWith(I&: II, V: IsNan);
985 }
986
987 if (Mask == (~fcNan & fcAllFlags) && !IsStrict) {
988 // Equivalent of !isnan. Replace with standard fcmp.
989 Value *FCmp =
990 Builder.CreateFCmpORD(LHS: Src0, RHS: ConstantFP::getZero(Ty: Src0->getType()));
991 FCmp->takeName(V: &II);
992 return replaceInstUsesWith(I&: II, V: FCmp);
993 }
994
995 FCmpInst::Predicate PredType = FCmpInst::BAD_FCMP_PREDICATE;
996
997 // Try to replace with an fcmp with 0
998 //
999 // is.fpclass(x, fcZero) -> fcmp oeq x, 0.0
1000 // is.fpclass(x, fcZero | fcNan) -> fcmp ueq x, 0.0
1001 // is.fpclass(x, ~fcZero & ~fcNan) -> fcmp one x, 0.0
1002 // is.fpclass(x, ~fcZero) -> fcmp une x, 0.0
1003 //
1004 // is.fpclass(x, fcPosSubnormal | fcPosNormal | fcPosInf) -> fcmp ogt x, 0.0
1005 // is.fpclass(x, fcPositive | fcNegZero) -> fcmp oge x, 0.0
1006 //
1007 // is.fpclass(x, fcNegSubnormal | fcNegNormal | fcNegInf) -> fcmp olt x, 0.0
1008 // is.fpclass(x, fcNegative | fcPosZero) -> fcmp ole x, 0.0
1009 //
1010 if (!IsStrict && (IsOrdered || IsUnordered) &&
1011 (PredType = fpclassTestIsFCmp0(Mask: OrderedMask, F: *II.getFunction(),
1012 Ty: Src0->getType())) !=
1013 FCmpInst::BAD_FCMP_PREDICATE) {
1014 Constant *Zero = ConstantFP::getZero(Ty: Src0->getType());
1015 // Equivalent of == 0.
1016 Value *FCmp = Builder.CreateFCmp(
1017 P: IsUnordered ? FCmpInst::getUnorderedPredicate(Pred: PredType) : PredType,
1018 LHS: Src0, RHS: Zero);
1019
1020 FCmp->takeName(V: &II);
1021 return replaceInstUsesWith(I&: II, V: FCmp);
1022 }
1023
1024 KnownFPClass Known = computeKnownFPClass(Val: Src0, Interested: Mask, CtxI: &II);
1025
1026 // Clear test bits we know must be false from the source value.
1027 // fp_class (nnan x), qnan|snan|other -> fp_class (nnan x), other
1028 // fp_class (ninf x), ninf|pinf|other -> fp_class (ninf x), other
1029 if ((Mask & Known.KnownFPClasses) != Mask) {
1030 II.setArgOperand(
1031 i: 1, v: ConstantInt::get(Ty: Src1->getType(), V: Mask & Known.KnownFPClasses));
1032 return &II;
1033 }
1034
1035 // If none of the tests which can return false are possible, fold to true.
1036 // fp_class (nnan x), ~(qnan|snan) -> true
1037 // fp_class (ninf x), ~(ninf|pinf) -> true
1038 if (Mask == Known.KnownFPClasses)
1039 return replaceInstUsesWith(I&: II, V: ConstantInt::get(Ty: II.getType(), V: true));
1040
1041 return nullptr;
1042}
1043
1044static std::optional<bool> getKnownSign(Value *Op, const SimplifyQuery &SQ) {
1045 KnownBits Known = computeKnownBits(V: Op, /*Depth=*/0, Q: SQ);
1046 if (Known.isNonNegative())
1047 return false;
1048 if (Known.isNegative())
1049 return true;
1050
1051 Value *X, *Y;
1052 if (match(V: Op, P: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y))))
1053 return isImpliedByDomCondition(Pred: ICmpInst::ICMP_SLT, LHS: X, RHS: Y, ContextI: SQ.CxtI, DL: SQ.DL);
1054
1055 return std::nullopt;
1056}
1057
1058static std::optional<bool> getKnownSignOrZero(Value *Op,
1059 const SimplifyQuery &SQ) {
1060 if (std::optional<bool> Sign = getKnownSign(Op, SQ))
1061 return Sign;
1062
1063 Value *X, *Y;
1064 if (match(V: Op, P: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y))))
1065 return isImpliedByDomCondition(Pred: ICmpInst::ICMP_SLE, LHS: X, RHS: Y, ContextI: SQ.CxtI, DL: SQ.DL);
1066
1067 return std::nullopt;
1068}
1069
1070/// Return true if two values \p Op0 and \p Op1 are known to have the same sign.
1071static bool signBitMustBeTheSame(Value *Op0, Value *Op1,
1072 const SimplifyQuery &SQ) {
1073 std::optional<bool> Known1 = getKnownSign(Op: Op1, SQ);
1074 if (!Known1)
1075 return false;
1076 std::optional<bool> Known0 = getKnownSign(Op: Op0, SQ);
1077 if (!Known0)
1078 return false;
1079 return *Known0 == *Known1;
1080}
1081
1082/// Try to canonicalize min/max(X + C0, C1) as min/max(X, C1 - C0) + C0. This
1083/// can trigger other combines.
1084static Instruction *moveAddAfterMinMax(IntrinsicInst *II,
1085 InstCombiner::BuilderTy &Builder) {
1086 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1087 assert((MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin ||
1088 MinMaxID == Intrinsic::umax || MinMaxID == Intrinsic::umin) &&
1089 "Expected a min or max intrinsic");
1090
1091 // TODO: Match vectors with undef elements, but undef may not propagate.
1092 Value *Op0 = II->getArgOperand(i: 0), *Op1 = II->getArgOperand(i: 1);
1093 Value *X;
1094 const APInt *C0, *C1;
1095 if (!match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C0)))) ||
1096 !match(V: Op1, P: m_APInt(Res&: C1)))
1097 return nullptr;
1098
1099 // Check for necessary no-wrap and overflow constraints.
1100 bool IsSigned = MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin;
1101 auto *Add = cast<BinaryOperator>(Val: Op0);
1102 if ((IsSigned && !Add->hasNoSignedWrap()) ||
1103 (!IsSigned && !Add->hasNoUnsignedWrap()))
1104 return nullptr;
1105
1106 // If the constant difference overflows, then instsimplify should reduce the
1107 // min/max to the add or C1.
1108 bool Overflow;
1109 APInt CDiff =
1110 IsSigned ? C1->ssub_ov(RHS: *C0, Overflow) : C1->usub_ov(RHS: *C0, Overflow);
1111 assert(!Overflow && "Expected simplify of min/max");
1112
1113 // min/max (add X, C0), C1 --> add (min/max X, C1 - C0), C0
1114 // Note: the "mismatched" no-overflow setting does not propagate.
1115 Constant *NewMinMaxC = ConstantInt::get(Ty: II->getType(), V: CDiff);
1116 Value *NewMinMax = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS: NewMinMaxC);
1117 return IsSigned ? BinaryOperator::CreateNSWAdd(V1: NewMinMax, V2: Add->getOperand(i_nocapture: 1))
1118 : BinaryOperator::CreateNUWAdd(V1: NewMinMax, V2: Add->getOperand(i_nocapture: 1));
1119}
1120/// Match a sadd_sat or ssub_sat which is using min/max to clamp the value.
1121Instruction *InstCombinerImpl::matchSAddSubSat(IntrinsicInst &MinMax1) {
1122 Type *Ty = MinMax1.getType();
1123
1124 // We are looking for a tree of:
1125 // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B))))
1126 // Where the min and max could be reversed
1127 Instruction *MinMax2;
1128 BinaryOperator *AddSub;
1129 const APInt *MinValue, *MaxValue;
1130 if (match(V: &MinMax1, P: m_SMin(L: m_Instruction(I&: MinMax2), R: m_APInt(Res&: MaxValue)))) {
1131 if (!match(V: MinMax2, P: m_SMax(L: m_BinOp(I&: AddSub), R: m_APInt(Res&: MinValue))))
1132 return nullptr;
1133 } else if (match(V: &MinMax1,
1134 P: m_SMax(L: m_Instruction(I&: MinMax2), R: m_APInt(Res&: MinValue)))) {
1135 if (!match(V: MinMax2, P: m_SMin(L: m_BinOp(I&: AddSub), R: m_APInt(Res&: MaxValue))))
1136 return nullptr;
1137 } else
1138 return nullptr;
1139
1140 // Check that the constants clamp a saturate, and that the new type would be
1141 // sensible to convert to.
1142 if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1)
1143 return nullptr;
1144 // In what bitwidth can this be treated as saturating arithmetics?
1145 unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1;
1146 // FIXME: This isn't quite right for vectors, but using the scalar type is a
1147 // good first approximation for what should be done there.
1148 if (!shouldChangeType(FromBitWidth: Ty->getScalarType()->getIntegerBitWidth(), ToBitWidth: NewBitWidth))
1149 return nullptr;
1150
1151 // Also make sure that the inner min/max and the add/sub have one use.
1152 if (!MinMax2->hasOneUse() || !AddSub->hasOneUse())
1153 return nullptr;
1154
1155 // Create the new type (which can be a vector type)
1156 Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth);
1157
1158 Intrinsic::ID IntrinsicID;
1159 if (AddSub->getOpcode() == Instruction::Add)
1160 IntrinsicID = Intrinsic::sadd_sat;
1161 else if (AddSub->getOpcode() == Instruction::Sub)
1162 IntrinsicID = Intrinsic::ssub_sat;
1163 else
1164 return nullptr;
1165
1166 // The two operands of the add/sub must be nsw-truncatable to the NewTy. This
1167 // is usually achieved via a sext from a smaller type.
1168 if (ComputeMaxSignificantBits(Op: AddSub->getOperand(i_nocapture: 0), Depth: 0, CxtI: AddSub) >
1169 NewBitWidth ||
1170 ComputeMaxSignificantBits(Op: AddSub->getOperand(i_nocapture: 1), Depth: 0, CxtI: AddSub) > NewBitWidth)
1171 return nullptr;
1172
1173 // Finally create and return the sat intrinsic, truncated to the new type
1174 Function *F = Intrinsic::getDeclaration(M: MinMax1.getModule(), id: IntrinsicID, Tys: NewTy);
1175 Value *AT = Builder.CreateTrunc(V: AddSub->getOperand(i_nocapture: 0), DestTy: NewTy);
1176 Value *BT = Builder.CreateTrunc(V: AddSub->getOperand(i_nocapture: 1), DestTy: NewTy);
1177 Value *Sat = Builder.CreateCall(Callee: F, Args: {AT, BT});
1178 return CastInst::Create(Instruction::SExt, S: Sat, Ty);
1179}
1180
1181
1182/// If we have a clamp pattern like max (min X, 42), 41 -- where the output
1183/// can only be one of two possible constant values -- turn that into a select
1184/// of constants.
1185static Instruction *foldClampRangeOfTwo(IntrinsicInst *II,
1186 InstCombiner::BuilderTy &Builder) {
1187 Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1);
1188 Value *X;
1189 const APInt *C0, *C1;
1190 if (!match(V: I1, P: m_APInt(Res&: C1)) || !I0->hasOneUse())
1191 return nullptr;
1192
1193 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
1194 switch (II->getIntrinsicID()) {
1195 case Intrinsic::smax:
1196 if (match(V: I0, P: m_SMin(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C0 == *C1 + 1)
1197 Pred = ICmpInst::ICMP_SGT;
1198 break;
1199 case Intrinsic::smin:
1200 if (match(V: I0, P: m_SMax(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C1 == *C0 + 1)
1201 Pred = ICmpInst::ICMP_SLT;
1202 break;
1203 case Intrinsic::umax:
1204 if (match(V: I0, P: m_UMin(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C0 == *C1 + 1)
1205 Pred = ICmpInst::ICMP_UGT;
1206 break;
1207 case Intrinsic::umin:
1208 if (match(V: I0, P: m_UMax(L: m_Value(V&: X), R: m_APInt(Res&: C0))) && *C1 == *C0 + 1)
1209 Pred = ICmpInst::ICMP_ULT;
1210 break;
1211 default:
1212 llvm_unreachable("Expected min/max intrinsic");
1213 }
1214 if (Pred == CmpInst::BAD_ICMP_PREDICATE)
1215 return nullptr;
1216
1217 // max (min X, 42), 41 --> X > 41 ? 42 : 41
1218 // min (max X, 42), 43 --> X < 43 ? 42 : 43
1219 Value *Cmp = Builder.CreateICmp(P: Pred, LHS: X, RHS: I1);
1220 return SelectInst::Create(C: Cmp, S1: ConstantInt::get(Ty: II->getType(), V: *C0), S2: I1);
1221}
1222
1223/// If this min/max has a constant operand and an operand that is a matching
1224/// min/max with a constant operand, constant-fold the 2 constant operands.
1225static Value *reassociateMinMaxWithConstants(IntrinsicInst *II,
1226 IRBuilderBase &Builder,
1227 const SimplifyQuery &SQ) {
1228 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1229 auto *LHS = dyn_cast<MinMaxIntrinsic>(Val: II->getArgOperand(i: 0));
1230 if (!LHS)
1231 return nullptr;
1232
1233 Constant *C0, *C1;
1234 if (!match(V: LHS->getArgOperand(i: 1), P: m_ImmConstant(C&: C0)) ||
1235 !match(V: II->getArgOperand(i: 1), P: m_ImmConstant(C&: C1)))
1236 return nullptr;
1237
1238 // max (max X, C0), C1 --> max X, (max C0, C1)
1239 // min (min X, C0), C1 --> min X, (min C0, C1)
1240 // umax (smax X, nneg C0), nneg C1 --> smax X, (umax C0, C1)
1241 // smin (umin X, nneg C0), nneg C1 --> umin X, (smin C0, C1)
1242 Intrinsic::ID InnerMinMaxID = LHS->getIntrinsicID();
1243 if (InnerMinMaxID != MinMaxID &&
1244 !(((MinMaxID == Intrinsic::umax && InnerMinMaxID == Intrinsic::smax) ||
1245 (MinMaxID == Intrinsic::smin && InnerMinMaxID == Intrinsic::umin)) &&
1246 isKnownNonNegative(V: C0, SQ) && isKnownNonNegative(V: C1, SQ)))
1247 return nullptr;
1248
1249 ICmpInst::Predicate Pred = MinMaxIntrinsic::getPredicate(ID: MinMaxID);
1250 Value *CondC = Builder.CreateICmp(P: Pred, LHS: C0, RHS: C1);
1251 Value *NewC = Builder.CreateSelect(C: CondC, True: C0, False: C1);
1252 return Builder.CreateIntrinsic(ID: InnerMinMaxID, Types: II->getType(),
1253 Args: {LHS->getArgOperand(i: 0), NewC});
1254}
1255
1256/// If this min/max has a matching min/max operand with a constant, try to push
1257/// the constant operand into this instruction. This can enable more folds.
1258static Instruction *
1259reassociateMinMaxWithConstantInOperand(IntrinsicInst *II,
1260 InstCombiner::BuilderTy &Builder) {
1261 // Match and capture a min/max operand candidate.
1262 Value *X, *Y;
1263 Constant *C;
1264 Instruction *Inner;
1265 if (!match(V: II, P: m_c_MaxOrMin(L: m_OneUse(SubPattern: m_CombineAnd(
1266 L: m_Instruction(I&: Inner),
1267 R: m_MaxOrMin(L: m_Value(V&: X), R: m_ImmConstant(C)))),
1268 R: m_Value(V&: Y))))
1269 return nullptr;
1270
1271 // The inner op must match. Check for constants to avoid infinite loops.
1272 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1273 auto *InnerMM = dyn_cast<IntrinsicInst>(Val: Inner);
1274 if (!InnerMM || InnerMM->getIntrinsicID() != MinMaxID ||
1275 match(V: X, P: m_ImmConstant()) || match(V: Y, P: m_ImmConstant()))
1276 return nullptr;
1277
1278 // max (max X, C), Y --> max (max X, Y), C
1279 Function *MinMax =
1280 Intrinsic::getDeclaration(M: II->getModule(), id: MinMaxID, Tys: II->getType());
1281 Value *NewInner = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS: Y);
1282 NewInner->takeName(V: Inner);
1283 return CallInst::Create(Func: MinMax, Args: {NewInner, C});
1284}
1285
1286/// Reduce a sequence of min/max intrinsics with a common operand.
1287static Instruction *factorizeMinMaxTree(IntrinsicInst *II) {
1288 // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
1289 auto *LHS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0));
1290 auto *RHS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 1));
1291 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1292 if (!LHS || !RHS || LHS->getIntrinsicID() != MinMaxID ||
1293 RHS->getIntrinsicID() != MinMaxID ||
1294 (!LHS->hasOneUse() && !RHS->hasOneUse()))
1295 return nullptr;
1296
1297 Value *A = LHS->getArgOperand(i: 0);
1298 Value *B = LHS->getArgOperand(i: 1);
1299 Value *C = RHS->getArgOperand(i: 0);
1300 Value *D = RHS->getArgOperand(i: 1);
1301
1302 // Look for a common operand.
1303 Value *MinMaxOp = nullptr;
1304 Value *ThirdOp = nullptr;
1305 if (LHS->hasOneUse()) {
1306 // If the LHS is only used in this chain and the RHS is used outside of it,
1307 // reuse the RHS min/max because that will eliminate the LHS.
1308 if (D == A || C == A) {
1309 // min(min(a, b), min(c, a)) --> min(min(c, a), b)
1310 // min(min(a, b), min(a, d)) --> min(min(a, d), b)
1311 MinMaxOp = RHS;
1312 ThirdOp = B;
1313 } else if (D == B || C == B) {
1314 // min(min(a, b), min(c, b)) --> min(min(c, b), a)
1315 // min(min(a, b), min(b, d)) --> min(min(b, d), a)
1316 MinMaxOp = RHS;
1317 ThirdOp = A;
1318 }
1319 } else {
1320 assert(RHS->hasOneUse() && "Expected one-use operand");
1321 // Reuse the LHS. This will eliminate the RHS.
1322 if (D == A || D == B) {
1323 // min(min(a, b), min(c, a)) --> min(min(a, b), c)
1324 // min(min(a, b), min(c, b)) --> min(min(a, b), c)
1325 MinMaxOp = LHS;
1326 ThirdOp = C;
1327 } else if (C == A || C == B) {
1328 // min(min(a, b), min(b, d)) --> min(min(a, b), d)
1329 // min(min(a, b), min(c, b)) --> min(min(a, b), d)
1330 MinMaxOp = LHS;
1331 ThirdOp = D;
1332 }
1333 }
1334
1335 if (!MinMaxOp || !ThirdOp)
1336 return nullptr;
1337
1338 Module *Mod = II->getModule();
1339 Function *MinMax = Intrinsic::getDeclaration(M: Mod, id: MinMaxID, Tys: II->getType());
1340 return CallInst::Create(Func: MinMax, Args: { MinMaxOp, ThirdOp });
1341}
1342
1343/// If all arguments of the intrinsic are unary shuffles with the same mask,
1344/// try to shuffle after the intrinsic.
1345static Instruction *
1346foldShuffledIntrinsicOperands(IntrinsicInst *II,
1347 InstCombiner::BuilderTy &Builder) {
1348 // TODO: This should be extended to handle other intrinsics like fshl, ctpop,
1349 // etc. Use llvm::isTriviallyVectorizable() and related to determine
1350 // which intrinsics are safe to shuffle?
1351 switch (II->getIntrinsicID()) {
1352 case Intrinsic::smax:
1353 case Intrinsic::smin:
1354 case Intrinsic::umax:
1355 case Intrinsic::umin:
1356 case Intrinsic::fma:
1357 case Intrinsic::fshl:
1358 case Intrinsic::fshr:
1359 break;
1360 default:
1361 return nullptr;
1362 }
1363
1364 Value *X;
1365 ArrayRef<int> Mask;
1366 if (!match(V: II->getArgOperand(i: 0),
1367 P: m_Shuffle(v1: m_Value(V&: X), v2: m_Undef(), mask: m_Mask(Mask))))
1368 return nullptr;
1369
1370 // At least 1 operand must have 1 use because we are creating 2 instructions.
1371 if (none_of(Range: II->args(), P: [](Value *V) { return V->hasOneUse(); }))
1372 return nullptr;
1373
1374 // See if all arguments are shuffled with the same mask.
1375 SmallVector<Value *, 4> NewArgs(II->arg_size());
1376 NewArgs[0] = X;
1377 Type *SrcTy = X->getType();
1378 for (unsigned i = 1, e = II->arg_size(); i != e; ++i) {
1379 if (!match(V: II->getArgOperand(i),
1380 P: m_Shuffle(v1: m_Value(V&: X), v2: m_Undef(), mask: m_SpecificMask(Mask))) ||
1381 X->getType() != SrcTy)
1382 return nullptr;
1383 NewArgs[i] = X;
1384 }
1385
1386 // intrinsic (shuf X, M), (shuf Y, M), ... --> shuf (intrinsic X, Y, ...), M
1387 Instruction *FPI = isa<FPMathOperator>(Val: II) ? II : nullptr;
1388 Value *NewIntrinsic =
1389 Builder.CreateIntrinsic(ID: II->getIntrinsicID(), Types: SrcTy, Args: NewArgs, FMFSource: FPI);
1390 return new ShuffleVectorInst(NewIntrinsic, Mask);
1391}
1392
1393/// Fold the following cases and accepts bswap and bitreverse intrinsics:
1394/// bswap(logic_op(bswap(x), y)) --> logic_op(x, bswap(y))
1395/// bswap(logic_op(bswap(x), bswap(y))) --> logic_op(x, y) (ignores multiuse)
1396template <Intrinsic::ID IntrID>
1397static Instruction *foldBitOrderCrossLogicOp(Value *V,
1398 InstCombiner::BuilderTy &Builder) {
1399 static_assert(IntrID == Intrinsic::bswap || IntrID == Intrinsic::bitreverse,
1400 "This helper only supports BSWAP and BITREVERSE intrinsics");
1401
1402 Value *X, *Y;
1403 // Find bitwise logic op. Check that it is a BinaryOperator explicitly so we
1404 // don't match ConstantExpr that aren't meaningful for this transform.
1405 if (match(V, P: m_OneUse(SubPattern: m_BitwiseLogic(L: m_Value(V&: X), R: m_Value(V&: Y)))) &&
1406 isa<BinaryOperator>(Val: V)) {
1407 Value *OldReorderX, *OldReorderY;
1408 BinaryOperator::BinaryOps Op = cast<BinaryOperator>(Val: V)->getOpcode();
1409
1410 // If both X and Y are bswap/bitreverse, the transform reduces the number
1411 // of instructions even if there's multiuse.
1412 // If only one operand is bswap/bitreverse, we need to ensure the operand
1413 // have only one use.
1414 if (match(X, m_Intrinsic<IntrID>(m_Value(V&: OldReorderX))) &&
1415 match(Y, m_Intrinsic<IntrID>(m_Value(V&: OldReorderY)))) {
1416 return BinaryOperator::Create(Op, S1: OldReorderX, S2: OldReorderY);
1417 }
1418
1419 if (match(X, m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: OldReorderX))))) {
1420 Value *NewReorder = Builder.CreateUnaryIntrinsic(ID: IntrID, V: Y);
1421 return BinaryOperator::Create(Op, S1: OldReorderX, S2: NewReorder);
1422 }
1423
1424 if (match(Y, m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: OldReorderY))))) {
1425 Value *NewReorder = Builder.CreateUnaryIntrinsic(ID: IntrID, V: X);
1426 return BinaryOperator::Create(Op, S1: NewReorder, S2: OldReorderY);
1427 }
1428 }
1429 return nullptr;
1430}
1431
1432static Value *simplifyReductionOperand(Value *Arg, bool CanReorderLanes) {
1433 if (!CanReorderLanes)
1434 return nullptr;
1435
1436 Value *V;
1437 if (match(V: Arg, P: m_VecReverse(Op0: m_Value(V))))
1438 return V;
1439
1440 ArrayRef<int> Mask;
1441 if (!isa<FixedVectorType>(Val: Arg->getType()) ||
1442 !match(V: Arg, P: m_Shuffle(v1: m_Value(V), v2: m_Undef(), mask: m_Mask(Mask))) ||
1443 !cast<ShuffleVectorInst>(Val: Arg)->isSingleSource())
1444 return nullptr;
1445
1446 int Sz = Mask.size();
1447 SmallBitVector UsedIndices(Sz);
1448 for (int Idx : Mask) {
1449 if (Idx == PoisonMaskElem || UsedIndices.test(Idx))
1450 return nullptr;
1451 UsedIndices.set(Idx);
1452 }
1453
1454 // Can remove shuffle iff just shuffled elements, no repeats, undefs, or
1455 // other changes.
1456 return UsedIndices.all() ? V : nullptr;
1457}
1458
1459/// Fold an unsigned minimum of trailing or leading zero bits counts:
1460/// umin(cttz(CtOp, ZeroUndef), ConstOp) --> cttz(CtOp | (1 << ConstOp))
1461/// umin(ctlz(CtOp, ZeroUndef), ConstOp) --> ctlz(CtOp | (SignedMin
1462/// >> ConstOp))
1463template <Intrinsic::ID IntrID>
1464static Value *
1465foldMinimumOverTrailingOrLeadingZeroCount(Value *I0, Value *I1,
1466 const DataLayout &DL,
1467 InstCombiner::BuilderTy &Builder) {
1468 static_assert(IntrID == Intrinsic::cttz || IntrID == Intrinsic::ctlz,
1469 "This helper only supports cttz and ctlz intrinsics");
1470
1471 Value *CtOp;
1472 Value *ZeroUndef;
1473 if (!match(I0,
1474 m_OneUse(m_Intrinsic<IntrID>(m_Value(V&: CtOp), m_Value(V&: ZeroUndef)))))
1475 return nullptr;
1476
1477 unsigned BitWidth = I1->getType()->getScalarSizeInBits();
1478 auto LessBitWidth = [BitWidth](auto &C) { return C.ult(BitWidth); };
1479 if (!match(I1, m_CheckedInt(LessBitWidth)))
1480 // We have a constant >= BitWidth (which can be handled by CVP)
1481 // or a non-splat vector with elements < and >= BitWidth
1482 return nullptr;
1483
1484 Type *Ty = I1->getType();
1485 Constant *NewConst = ConstantFoldBinaryOpOperands(
1486 Opcode: IntrID == Intrinsic::cttz ? Instruction::Shl : Instruction::LShr,
1487 LHS: IntrID == Intrinsic::cttz
1488 ? ConstantInt::get(Ty, V: 1)
1489 : ConstantInt::get(Ty, V: APInt::getSignedMinValue(numBits: BitWidth)),
1490 RHS: cast<Constant>(Val: I1), DL);
1491 return Builder.CreateBinaryIntrinsic(
1492 ID: IntrID, LHS: Builder.CreateOr(LHS: CtOp, RHS: NewConst),
1493 RHS: ConstantInt::getTrue(Ty: ZeroUndef->getType()));
1494}
1495
1496/// CallInst simplification. This mostly only handles folding of intrinsic
1497/// instructions. For normal calls, it allows visitCallBase to do the heavy
1498/// lifting.
1499Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
1500 // Don't try to simplify calls without uses. It will not do anything useful,
1501 // but will result in the following folds being skipped.
1502 if (!CI.use_empty()) {
1503 SmallVector<Value *, 4> Args;
1504 Args.reserve(N: CI.arg_size());
1505 for (Value *Op : CI.args())
1506 Args.push_back(Elt: Op);
1507 if (Value *V = simplifyCall(Call: &CI, Callee: CI.getCalledOperand(), Args,
1508 Q: SQ.getWithInstruction(I: &CI)))
1509 return replaceInstUsesWith(I&: CI, V);
1510 }
1511
1512 if (Value *FreedOp = getFreedOperand(CB: &CI, TLI: &TLI))
1513 return visitFree(FI&: CI, FreedOp);
1514
1515 // If the caller function (i.e. us, the function that contains this CallInst)
1516 // is nounwind, mark the call as nounwind, even if the callee isn't.
1517 if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) {
1518 CI.setDoesNotThrow();
1519 return &CI;
1520 }
1521
1522 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &CI);
1523 if (!II) return visitCallBase(Call&: CI);
1524
1525 // For atomic unordered mem intrinsics if len is not a positive or
1526 // not a multiple of element size then behavior is undefined.
1527 if (auto *AMI = dyn_cast<AtomicMemIntrinsic>(Val: II))
1528 if (ConstantInt *NumBytes = dyn_cast<ConstantInt>(Val: AMI->getLength()))
1529 if (NumBytes->isNegative() ||
1530 (NumBytes->getZExtValue() % AMI->getElementSizeInBytes() != 0)) {
1531 CreateNonTerminatorUnreachable(InsertAt: AMI);
1532 assert(AMI->getType()->isVoidTy() &&
1533 "non void atomic unordered mem intrinsic");
1534 return eraseInstFromFunction(I&: *AMI);
1535 }
1536
1537 // Intrinsics cannot occur in an invoke or a callbr, so handle them here
1538 // instead of in visitCallBase.
1539 if (auto *MI = dyn_cast<AnyMemIntrinsic>(Val: II)) {
1540 bool Changed = false;
1541
1542 // memmove/cpy/set of zero bytes is a noop.
1543 if (Constant *NumBytes = dyn_cast<Constant>(Val: MI->getLength())) {
1544 if (NumBytes->isNullValue())
1545 return eraseInstFromFunction(I&: CI);
1546 }
1547
1548 // No other transformations apply to volatile transfers.
1549 if (auto *M = dyn_cast<MemIntrinsic>(Val: MI))
1550 if (M->isVolatile())
1551 return nullptr;
1552
1553 // If we have a memmove and the source operation is a constant global,
1554 // then the source and dest pointers can't alias, so we can change this
1555 // into a call to memcpy.
1556 if (auto *MMI = dyn_cast<AnyMemMoveInst>(Val: MI)) {
1557 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(Val: MMI->getSource()))
1558 if (GVSrc->isConstant()) {
1559 Module *M = CI.getModule();
1560 Intrinsic::ID MemCpyID =
1561 isa<AtomicMemMoveInst>(Val: MMI)
1562 ? Intrinsic::memcpy_element_unordered_atomic
1563 : Intrinsic::memcpy;
1564 Type *Tys[3] = { CI.getArgOperand(i: 0)->getType(),
1565 CI.getArgOperand(i: 1)->getType(),
1566 CI.getArgOperand(i: 2)->getType() };
1567 CI.setCalledFunction(Intrinsic::getDeclaration(M, id: MemCpyID, Tys));
1568 Changed = true;
1569 }
1570 }
1571
1572 if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(Val: MI)) {
1573 // memmove(x,x,size) -> noop.
1574 if (MTI->getSource() == MTI->getDest())
1575 return eraseInstFromFunction(I&: CI);
1576 }
1577
1578 // If we can determine a pointer alignment that is bigger than currently
1579 // set, update the alignment.
1580 if (auto *MTI = dyn_cast<AnyMemTransferInst>(Val: MI)) {
1581 if (Instruction *I = SimplifyAnyMemTransfer(MI: MTI))
1582 return I;
1583 } else if (auto *MSI = dyn_cast<AnyMemSetInst>(Val: MI)) {
1584 if (Instruction *I = SimplifyAnyMemSet(MI: MSI))
1585 return I;
1586 }
1587
1588 if (Changed) return II;
1589 }
1590
1591 // For fixed width vector result intrinsics, use the generic demanded vector
1592 // support.
1593 if (auto *IIFVTy = dyn_cast<FixedVectorType>(Val: II->getType())) {
1594 auto VWidth = IIFVTy->getNumElements();
1595 APInt PoisonElts(VWidth, 0);
1596 APInt AllOnesEltMask(APInt::getAllOnes(numBits: VWidth));
1597 if (Value *V = SimplifyDemandedVectorElts(V: II, DemandedElts: AllOnesEltMask, PoisonElts)) {
1598 if (V != II)
1599 return replaceInstUsesWith(I&: *II, V);
1600 return II;
1601 }
1602 }
1603
1604 if (II->isCommutative()) {
1605 if (auto Pair = matchSymmetricPair(LHS: II->getOperand(i_nocapture: 0), RHS: II->getOperand(i_nocapture: 1))) {
1606 replaceOperand(I&: *II, OpNum: 0, V: Pair->first);
1607 replaceOperand(I&: *II, OpNum: 1, V: Pair->second);
1608 return II;
1609 }
1610
1611 if (CallInst *NewCall = canonicalizeConstantArg0ToArg1(Call&: CI))
1612 return NewCall;
1613 }
1614
1615 // Unused constrained FP intrinsic calls may have declared side effect, which
1616 // prevents it from being removed. In some cases however the side effect is
1617 // actually absent. To detect this case, call SimplifyConstrainedFPCall. If it
1618 // returns a replacement, the call may be removed.
1619 if (CI.use_empty() && isa<ConstrainedFPIntrinsic>(Val: CI)) {
1620 if (simplifyConstrainedFPCall(Call: &CI, Q: SQ.getWithInstruction(I: &CI)))
1621 return eraseInstFromFunction(I&: CI);
1622 }
1623
1624 Intrinsic::ID IID = II->getIntrinsicID();
1625 switch (IID) {
1626 case Intrinsic::objectsize: {
1627 SmallVector<Instruction *> InsertedInstructions;
1628 if (Value *V = lowerObjectSizeCall(ObjectSize: II, DL, TLI: &TLI, AA, /*MustSucceed=*/false,
1629 InsertedInstructions: &InsertedInstructions)) {
1630 for (Instruction *Inserted : InsertedInstructions)
1631 Worklist.add(I: Inserted);
1632 return replaceInstUsesWith(I&: CI, V);
1633 }
1634 return nullptr;
1635 }
1636 case Intrinsic::abs: {
1637 Value *IIOperand = II->getArgOperand(i: 0);
1638 bool IntMinIsPoison = cast<Constant>(Val: II->getArgOperand(i: 1))->isOneValue();
1639
1640 // abs(-x) -> abs(x)
1641 // TODO: Copy nsw if it was present on the neg?
1642 Value *X;
1643 if (match(V: IIOperand, P: m_Neg(V: m_Value(V&: X))))
1644 return replaceOperand(I&: *II, OpNum: 0, V: X);
1645 if (match(V: IIOperand, P: m_Select(C: m_Value(), L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X)))))
1646 return replaceOperand(I&: *II, OpNum: 0, V: X);
1647 if (match(V: IIOperand, P: m_Select(C: m_Value(), L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X))))
1648 return replaceOperand(I&: *II, OpNum: 0, V: X);
1649
1650 Value *Y;
1651 // abs(a * abs(b)) -> abs(a * b)
1652 if (match(V: IIOperand,
1653 P: m_OneUse(SubPattern: m_c_Mul(L: m_Value(V&: X),
1654 R: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: Y)))))) {
1655 bool NSW =
1656 cast<Instruction>(Val: IIOperand)->hasNoSignedWrap() && IntMinIsPoison;
1657 auto *XY = NSW ? Builder.CreateNSWMul(LHS: X, RHS: Y) : Builder.CreateMul(LHS: X, RHS: Y);
1658 return replaceOperand(I&: *II, OpNum: 0, V: XY);
1659 }
1660
1661 if (std::optional<bool> Known =
1662 getKnownSignOrZero(Op: IIOperand, SQ: SQ.getWithInstruction(I: II))) {
1663 // abs(x) -> x if x >= 0 (include abs(x-y) --> x - y where x >= y)
1664 // abs(x) -> x if x > 0 (include abs(x-y) --> x - y where x > y)
1665 if (!*Known)
1666 return replaceInstUsesWith(I&: *II, V: IIOperand);
1667
1668 // abs(x) -> -x if x < 0
1669 // abs(x) -> -x if x < = 0 (include abs(x-y) --> y - x where x <= y)
1670 if (IntMinIsPoison)
1671 return BinaryOperator::CreateNSWNeg(Op: IIOperand);
1672 return BinaryOperator::CreateNeg(Op: IIOperand);
1673 }
1674
1675 // abs (sext X) --> zext (abs X*)
1676 // Clear the IsIntMin (nsw) bit on the abs to allow narrowing.
1677 if (match(V: IIOperand, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))))) {
1678 Value *NarrowAbs =
1679 Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: X, RHS: Builder.getFalse());
1680 return CastInst::Create(Instruction::ZExt, S: NarrowAbs, Ty: II->getType());
1681 }
1682
1683 // Match a complicated way to check if a number is odd/even:
1684 // abs (srem X, 2) --> and X, 1
1685 const APInt *C;
1686 if (match(V: IIOperand, P: m_SRem(L: m_Value(V&: X), R: m_APInt(Res&: C))) && *C == 2)
1687 return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty: II->getType(), V: 1));
1688
1689 break;
1690 }
1691 case Intrinsic::umin: {
1692 Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1);
1693 // umin(x, 1) == zext(x != 0)
1694 if (match(V: I1, P: m_One())) {
1695 assert(II->getType()->getScalarSizeInBits() != 1 &&
1696 "Expected simplify of umin with max constant");
1697 Value *Zero = Constant::getNullValue(Ty: I0->getType());
1698 Value *Cmp = Builder.CreateICmpNE(LHS: I0, RHS: Zero);
1699 return CastInst::Create(Instruction::ZExt, S: Cmp, Ty: II->getType());
1700 }
1701 // umin(cttz(x), const) --> cttz(x | (1 << const))
1702 if (Value *FoldedCttz =
1703 foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::cttz>(
1704 I0, I1, DL, Builder))
1705 return replaceInstUsesWith(I&: *II, V: FoldedCttz);
1706 // umin(ctlz(x), const) --> ctlz(x | (SignedMin >> const))
1707 if (Value *FoldedCtlz =
1708 foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::ctlz>(
1709 I0, I1, DL, Builder))
1710 return replaceInstUsesWith(I&: *II, V: FoldedCtlz);
1711 [[fallthrough]];
1712 }
1713 case Intrinsic::umax: {
1714 Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1);
1715 Value *X, *Y;
1716 if (match(V: I0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: I1, P: m_ZExt(Op: m_Value(V&: Y))) &&
1717 (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) {
1718 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y);
1719 return CastInst::Create(Instruction::ZExt, S: NarrowMaxMin, Ty: II->getType());
1720 }
1721 Constant *C;
1722 if (match(V: I0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: I1, P: m_Constant(C)) &&
1723 I0->hasOneUse()) {
1724 if (Constant *NarrowC = getLosslessUnsignedTrunc(C, TruncTy: X->getType())) {
1725 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: NarrowC);
1726 return CastInst::Create(Instruction::ZExt, S: NarrowMaxMin, Ty: II->getType());
1727 }
1728 }
1729 // If both operands of unsigned min/max are sign-extended, it is still ok
1730 // to narrow the operation.
1731 [[fallthrough]];
1732 }
1733 case Intrinsic::smax:
1734 case Intrinsic::smin: {
1735 Value *I0 = II->getArgOperand(i: 0), *I1 = II->getArgOperand(i: 1);
1736 Value *X, *Y;
1737 if (match(V: I0, P: m_SExt(Op: m_Value(V&: X))) && match(V: I1, P: m_SExt(Op: m_Value(V&: Y))) &&
1738 (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) {
1739 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y);
1740 return CastInst::Create(Instruction::SExt, S: NarrowMaxMin, Ty: II->getType());
1741 }
1742
1743 Constant *C;
1744 if (match(V: I0, P: m_SExt(Op: m_Value(V&: X))) && match(V: I1, P: m_Constant(C)) &&
1745 I0->hasOneUse()) {
1746 if (Constant *NarrowC = getLosslessSignedTrunc(C, TruncTy: X->getType())) {
1747 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: NarrowC);
1748 return CastInst::Create(Instruction::SExt, S: NarrowMaxMin, Ty: II->getType());
1749 }
1750 }
1751
1752 // umin(i1 X, i1 Y) -> and i1 X, Y
1753 // smax(i1 X, i1 Y) -> and i1 X, Y
1754 if ((IID == Intrinsic::umin || IID == Intrinsic::smax) &&
1755 II->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
1756 return BinaryOperator::CreateAnd(V1: I0, V2: I1);
1757 }
1758
1759 // umax(i1 X, i1 Y) -> or i1 X, Y
1760 // smin(i1 X, i1 Y) -> or i1 X, Y
1761 if ((IID == Intrinsic::umax || IID == Intrinsic::smin) &&
1762 II->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
1763 return BinaryOperator::CreateOr(V1: I0, V2: I1);
1764 }
1765
1766 if (IID == Intrinsic::smax || IID == Intrinsic::smin) {
1767 // smax (neg nsw X), (neg nsw Y) --> neg nsw (smin X, Y)
1768 // smin (neg nsw X), (neg nsw Y) --> neg nsw (smax X, Y)
1769 // TODO: Canonicalize neg after min/max if I1 is constant.
1770 if (match(V: I0, P: m_NSWNeg(V: m_Value(V&: X))) && match(V: I1, P: m_NSWNeg(V: m_Value(V&: Y))) &&
1771 (I0->hasOneUse() || I1->hasOneUse())) {
1772 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: IID);
1773 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(ID: InvID, LHS: X, RHS: Y);
1774 return BinaryOperator::CreateNSWNeg(Op: InvMaxMin);
1775 }
1776 }
1777
1778 // (umax X, (xor X, Pow2))
1779 // -> (or X, Pow2)
1780 // (umin X, (xor X, Pow2))
1781 // -> (and X, ~Pow2)
1782 // (smax X, (xor X, Pos_Pow2))
1783 // -> (or X, Pos_Pow2)
1784 // (smin X, (xor X, Pos_Pow2))
1785 // -> (and X, ~Pos_Pow2)
1786 // (smax X, (xor X, Neg_Pow2))
1787 // -> (and X, ~Neg_Pow2)
1788 // (smin X, (xor X, Neg_Pow2))
1789 // -> (or X, Neg_Pow2)
1790 if ((match(V: I0, P: m_c_Xor(L: m_Specific(V: I1), R: m_Value(V&: X))) ||
1791 match(V: I1, P: m_c_Xor(L: m_Specific(V: I0), R: m_Value(V&: X)))) &&
1792 isKnownToBeAPowerOfTwo(V: X, /* OrZero */ true)) {
1793 bool UseOr = IID == Intrinsic::smax || IID == Intrinsic::umax;
1794 bool UseAndN = IID == Intrinsic::smin || IID == Intrinsic::umin;
1795
1796 if (IID == Intrinsic::smax || IID == Intrinsic::smin) {
1797 auto KnownSign = getKnownSign(Op: X, SQ: SQ.getWithInstruction(I: II));
1798 if (KnownSign == std::nullopt) {
1799 UseOr = false;
1800 UseAndN = false;
1801 } else if (*KnownSign /* true is Signed. */) {
1802 UseOr ^= true;
1803 UseAndN ^= true;
1804 Type *Ty = I0->getType();
1805 // Negative power of 2 must be IntMin. It's possible to be able to
1806 // prove negative / power of 2 without actually having known bits, so
1807 // just get the value by hand.
1808 X = Constant::getIntegerValue(
1809 Ty, V: APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()));
1810 }
1811 }
1812 if (UseOr)
1813 return BinaryOperator::CreateOr(V1: I0, V2: X);
1814 else if (UseAndN)
1815 return BinaryOperator::CreateAnd(V1: I0, V2: Builder.CreateNot(V: X));
1816 }
1817
1818 // If we can eliminate ~A and Y is free to invert:
1819 // max ~A, Y --> ~(min A, ~Y)
1820 //
1821 // Examples:
1822 // max ~A, ~Y --> ~(min A, Y)
1823 // max ~A, C --> ~(min A, ~C)
1824 // max ~A, (max ~Y, ~Z) --> ~min( A, (min Y, Z))
1825 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * {
1826 Value *A;
1827 if (match(V: X, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: A)))) &&
1828 !isFreeToInvert(V: A, WillInvertAllUses: A->hasOneUse())) {
1829 if (Value *NotY = getFreelyInverted(V: Y, WillInvertAllUses: Y->hasOneUse(), Builder: &Builder)) {
1830 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: IID);
1831 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(ID: InvID, LHS: A, RHS: NotY);
1832 return BinaryOperator::CreateNot(Op: InvMaxMin);
1833 }
1834 }
1835 return nullptr;
1836 };
1837
1838 if (Instruction *I = moveNotAfterMinMax(I0, I1))
1839 return I;
1840 if (Instruction *I = moveNotAfterMinMax(I1, I0))
1841 return I;
1842
1843 if (Instruction *I = moveAddAfterMinMax(II, Builder))
1844 return I;
1845
1846 // minmax (X & NegPow2C, Y & NegPow2C) --> minmax(X, Y) & NegPow2C
1847 const APInt *RHSC;
1848 if (match(V: I0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_NegatedPower2(V&: RHSC)))) &&
1849 match(V: I1, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: Y), R: m_SpecificInt(V: *RHSC)))))
1850 return BinaryOperator::CreateAnd(V1: Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y),
1851 V2: ConstantInt::get(Ty: II->getType(), V: *RHSC));
1852
1853 // smax(X, -X) --> abs(X)
1854 // smin(X, -X) --> -abs(X)
1855 // umax(X, -X) --> -abs(X)
1856 // umin(X, -X) --> abs(X)
1857 if (isKnownNegation(X: I0, Y: I1)) {
1858 // We can choose either operand as the input to abs(), but if we can
1859 // eliminate the only use of a value, that's better for subsequent
1860 // transforms/analysis.
1861 if (I0->hasOneUse() && !I1->hasOneUse())
1862 std::swap(a&: I0, b&: I1);
1863
1864 // This is some variant of abs(). See if we can propagate 'nsw' to the abs
1865 // operation and potentially its negation.
1866 bool IntMinIsPoison = isKnownNegation(X: I0, Y: I1, /* NeedNSW */ true);
1867 Value *Abs = Builder.CreateBinaryIntrinsic(
1868 ID: Intrinsic::abs, LHS: I0,
1869 RHS: ConstantInt::getBool(Context&: II->getContext(), V: IntMinIsPoison));
1870
1871 // We don't have a "nabs" intrinsic, so negate if needed based on the
1872 // max/min operation.
1873 if (IID == Intrinsic::smin || IID == Intrinsic::umax)
1874 Abs = Builder.CreateNeg(V: Abs, Name: "nabs", HasNSW: IntMinIsPoison);
1875 return replaceInstUsesWith(I&: CI, V: Abs);
1876 }
1877
1878 if (Instruction *Sel = foldClampRangeOfTwo(II, Builder))
1879 return Sel;
1880
1881 if (Instruction *SAdd = matchSAddSubSat(MinMax1&: *II))
1882 return SAdd;
1883
1884 if (Value *NewMinMax = reassociateMinMaxWithConstants(II, Builder, SQ))
1885 return replaceInstUsesWith(I&: *II, V: NewMinMax);
1886
1887 if (Instruction *R = reassociateMinMaxWithConstantInOperand(II, Builder))
1888 return R;
1889
1890 if (Instruction *NewMinMax = factorizeMinMaxTree(II))
1891 return NewMinMax;
1892
1893 // Try to fold minmax with constant RHS based on range information
1894 if (match(V: I1, P: m_APIntAllowPoison(Res&: RHSC))) {
1895 ICmpInst::Predicate Pred =
1896 ICmpInst::getNonStrictPredicate(pred: MinMaxIntrinsic::getPredicate(ID: IID));
1897 bool IsSigned = MinMaxIntrinsic::isSigned(ID: IID);
1898 ConstantRange LHS_CR = computeConstantRangeIncludingKnownBits(
1899 V: I0, ForSigned: IsSigned, SQ: SQ.getWithInstruction(I: II));
1900 if (!LHS_CR.isFullSet()) {
1901 if (LHS_CR.icmp(Pred, Other: *RHSC))
1902 return replaceInstUsesWith(I&: *II, V: I0);
1903 if (LHS_CR.icmp(Pred: ICmpInst::getSwappedPredicate(pred: Pred), Other: *RHSC))
1904 return replaceInstUsesWith(I&: *II,
1905 V: ConstantInt::get(Ty: II->getType(), V: *RHSC));
1906 }
1907 }
1908
1909 break;
1910 }
1911 case Intrinsic::bitreverse: {
1912 Value *IIOperand = II->getArgOperand(i: 0);
1913 // bitrev (zext i1 X to ?) --> X ? SignBitC : 0
1914 Value *X;
1915 if (match(V: IIOperand, P: m_ZExt(Op: m_Value(V&: X))) &&
1916 X->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
1917 Type *Ty = II->getType();
1918 APInt SignBit = APInt::getSignMask(BitWidth: Ty->getScalarSizeInBits());
1919 return SelectInst::Create(C: X, S1: ConstantInt::get(Ty, V: SignBit),
1920 S2: ConstantInt::getNullValue(Ty));
1921 }
1922
1923 if (Instruction *crossLogicOpFold =
1924 foldBitOrderCrossLogicOp<Intrinsic::bitreverse>(V: IIOperand, Builder))
1925 return crossLogicOpFold;
1926
1927 break;
1928 }
1929 case Intrinsic::bswap: {
1930 Value *IIOperand = II->getArgOperand(i: 0);
1931
1932 // Try to canonicalize bswap-of-logical-shift-by-8-bit-multiple as
1933 // inverse-shift-of-bswap:
1934 // bswap (shl X, Y) --> lshr (bswap X), Y
1935 // bswap (lshr X, Y) --> shl (bswap X), Y
1936 Value *X, *Y;
1937 if (match(V: IIOperand, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
1938 unsigned BitWidth = IIOperand->getType()->getScalarSizeInBits();
1939 if (MaskedValueIsZero(V: Y, Mask: APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: 3))) {
1940 Value *NewSwap = Builder.CreateUnaryIntrinsic(ID: Intrinsic::bswap, V: X);
1941 BinaryOperator::BinaryOps InverseShift =
1942 cast<BinaryOperator>(Val: IIOperand)->getOpcode() == Instruction::Shl
1943 ? Instruction::LShr
1944 : Instruction::Shl;
1945 return BinaryOperator::Create(Op: InverseShift, S1: NewSwap, S2: Y);
1946 }
1947 }
1948
1949 KnownBits Known = computeKnownBits(V: IIOperand, Depth: 0, CxtI: II);
1950 uint64_t LZ = alignDown(Value: Known.countMinLeadingZeros(), Align: 8);
1951 uint64_t TZ = alignDown(Value: Known.countMinTrailingZeros(), Align: 8);
1952 unsigned BW = Known.getBitWidth();
1953
1954 // bswap(x) -> shift(x) if x has exactly one "active byte"
1955 if (BW - LZ - TZ == 8) {
1956 assert(LZ != TZ && "active byte cannot be in the middle");
1957 if (LZ > TZ) // -> shl(x) if the "active byte" is in the low part of x
1958 return BinaryOperator::CreateNUWShl(
1959 V1: IIOperand, V2: ConstantInt::get(Ty: IIOperand->getType(), V: LZ - TZ));
1960 // -> lshr(x) if the "active byte" is in the high part of x
1961 return BinaryOperator::CreateExactLShr(
1962 V1: IIOperand, V2: ConstantInt::get(Ty: IIOperand->getType(), V: TZ - LZ));
1963 }
1964
1965 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
1966 if (match(V: IIOperand, P: m_Trunc(Op: m_BSwap(Op0: m_Value(V&: X))))) {
1967 unsigned C = X->getType()->getScalarSizeInBits() - BW;
1968 Value *CV = ConstantInt::get(Ty: X->getType(), V: C);
1969 Value *V = Builder.CreateLShr(LHS: X, RHS: CV);
1970 return new TruncInst(V, IIOperand->getType());
1971 }
1972
1973 if (Instruction *crossLogicOpFold =
1974 foldBitOrderCrossLogicOp<Intrinsic::bswap>(V: IIOperand, Builder)) {
1975 return crossLogicOpFold;
1976 }
1977
1978 // Try to fold into bitreverse if bswap is the root of the expression tree.
1979 if (Instruction *BitOp = matchBSwapOrBitReverse(I&: *II, /*MatchBSwaps*/ false,
1980 /*MatchBitReversals*/ true))
1981 return BitOp;
1982 break;
1983 }
1984 case Intrinsic::masked_load:
1985 if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(II&: *II))
1986 return replaceInstUsesWith(I&: CI, V: SimplifiedMaskedOp);
1987 break;
1988 case Intrinsic::masked_store:
1989 return simplifyMaskedStore(II&: *II);
1990 case Intrinsic::masked_gather:
1991 return simplifyMaskedGather(II&: *II);
1992 case Intrinsic::masked_scatter:
1993 return simplifyMaskedScatter(II&: *II);
1994 case Intrinsic::launder_invariant_group:
1995 case Intrinsic::strip_invariant_group:
1996 if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(II&: *II, IC&: *this))
1997 return replaceInstUsesWith(I&: *II, V: SkippedBarrier);
1998 break;
1999 case Intrinsic::powi:
2000 if (ConstantInt *Power = dyn_cast<ConstantInt>(Val: II->getArgOperand(i: 1))) {
2001 // 0 and 1 are handled in instsimplify
2002 // powi(x, -1) -> 1/x
2003 if (Power->isMinusOne())
2004 return BinaryOperator::CreateFDivFMF(V1: ConstantFP::get(Ty: CI.getType(), V: 1.0),
2005 V2: II->getArgOperand(i: 0), FMFSource: II);
2006 // powi(x, 2) -> x*x
2007 if (Power->equalsInt(V: 2))
2008 return BinaryOperator::CreateFMulFMF(V1: II->getArgOperand(i: 0),
2009 V2: II->getArgOperand(i: 0), FMFSource: II);
2010
2011 if (!Power->getValue()[0]) {
2012 Value *X;
2013 // If power is even:
2014 // powi(-x, p) -> powi(x, p)
2015 // powi(fabs(x), p) -> powi(x, p)
2016 // powi(copysign(x, y), p) -> powi(x, p)
2017 if (match(V: II->getArgOperand(i: 0), P: m_FNeg(X: m_Value(V&: X))) ||
2018 match(V: II->getArgOperand(i: 0), P: m_FAbs(Op0: m_Value(V&: X))) ||
2019 match(V: II->getArgOperand(i: 0),
2020 P: m_Intrinsic<Intrinsic::copysign>(Op0: m_Value(V&: X), Op1: m_Value())))
2021 return replaceOperand(I&: *II, OpNum: 0, V: X);
2022 }
2023 }
2024 break;
2025
2026 case Intrinsic::cttz:
2027 case Intrinsic::ctlz:
2028 if (auto *I = foldCttzCtlz(II&: *II, IC&: *this))
2029 return I;
2030 break;
2031
2032 case Intrinsic::ctpop:
2033 if (auto *I = foldCtpop(II&: *II, IC&: *this))
2034 return I;
2035 break;
2036
2037 case Intrinsic::fshl:
2038 case Intrinsic::fshr: {
2039 Value *Op0 = II->getArgOperand(i: 0), *Op1 = II->getArgOperand(i: 1);
2040 Type *Ty = II->getType();
2041 unsigned BitWidth = Ty->getScalarSizeInBits();
2042 Constant *ShAmtC;
2043 if (match(V: II->getArgOperand(i: 2), P: m_ImmConstant(C&: ShAmtC))) {
2044 // Canonicalize a shift amount constant operand to modulo the bit-width.
2045 Constant *WidthC = ConstantInt::get(Ty, V: BitWidth);
2046 Constant *ModuloC =
2047 ConstantFoldBinaryOpOperands(Opcode: Instruction::URem, LHS: ShAmtC, RHS: WidthC, DL);
2048 if (!ModuloC)
2049 return nullptr;
2050 if (ModuloC != ShAmtC)
2051 return replaceOperand(I&: *II, OpNum: 2, V: ModuloC);
2052
2053 assert(match(ConstantFoldCompareInstOperands(ICmpInst::ICMP_UGT, WidthC,
2054 ShAmtC, DL),
2055 m_One()) &&
2056 "Shift amount expected to be modulo bitwidth");
2057
2058 // Canonicalize funnel shift right by constant to funnel shift left. This
2059 // is not entirely arbitrary. For historical reasons, the backend may
2060 // recognize rotate left patterns but miss rotate right patterns.
2061 if (IID == Intrinsic::fshr) {
2062 // fshr X, Y, C --> fshl X, Y, (BitWidth - C) if C is not zero.
2063 if (!isKnownNonZero(V: ShAmtC, Q: SQ.getWithInstruction(I: II)))
2064 return nullptr;
2065
2066 Constant *LeftShiftC = ConstantExpr::getSub(C1: WidthC, C2: ShAmtC);
2067 Module *Mod = II->getModule();
2068 Function *Fshl = Intrinsic::getDeclaration(M: Mod, id: Intrinsic::fshl, Tys: Ty);
2069 return CallInst::Create(Func: Fshl, Args: { Op0, Op1, LeftShiftC });
2070 }
2071 assert(IID == Intrinsic::fshl &&
2072 "All funnel shifts by simple constants should go left");
2073
2074 // fshl(X, 0, C) --> shl X, C
2075 // fshl(X, undef, C) --> shl X, C
2076 if (match(V: Op1, P: m_ZeroInt()) || match(V: Op1, P: m_Undef()))
2077 return BinaryOperator::CreateShl(V1: Op0, V2: ShAmtC);
2078
2079 // fshl(0, X, C) --> lshr X, (BW-C)
2080 // fshl(undef, X, C) --> lshr X, (BW-C)
2081 if (match(V: Op0, P: m_ZeroInt()) || match(V: Op0, P: m_Undef()))
2082 return BinaryOperator::CreateLShr(V1: Op1,
2083 V2: ConstantExpr::getSub(C1: WidthC, C2: ShAmtC));
2084
2085 // fshl i16 X, X, 8 --> bswap i16 X (reduce to more-specific form)
2086 if (Op0 == Op1 && BitWidth == 16 && match(V: ShAmtC, P: m_SpecificInt(V: 8))) {
2087 Module *Mod = II->getModule();
2088 Function *Bswap = Intrinsic::getDeclaration(M: Mod, id: Intrinsic::bswap, Tys: Ty);
2089 return CallInst::Create(Func: Bswap, Args: { Op0 });
2090 }
2091 if (Instruction *BitOp =
2092 matchBSwapOrBitReverse(I&: *II, /*MatchBSwaps*/ true,
2093 /*MatchBitReversals*/ true))
2094 return BitOp;
2095 }
2096
2097 // Left or right might be masked.
2098 if (SimplifyDemandedInstructionBits(Inst&: *II))
2099 return &CI;
2100
2101 // The shift amount (operand 2) of a funnel shift is modulo the bitwidth,
2102 // so only the low bits of the shift amount are demanded if the bitwidth is
2103 // a power-of-2.
2104 if (!isPowerOf2_32(Value: BitWidth))
2105 break;
2106 APInt Op2Demanded = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: Log2_32_Ceil(Value: BitWidth));
2107 KnownBits Op2Known(BitWidth);
2108 if (SimplifyDemandedBits(I: II, OpNo: 2, DemandedMask: Op2Demanded, Known&: Op2Known))
2109 return &CI;
2110 break;
2111 }
2112 case Intrinsic::ptrmask: {
2113 unsigned BitWidth = DL.getPointerTypeSizeInBits(II->getType());
2114 KnownBits Known(BitWidth);
2115 if (SimplifyDemandedInstructionBits(Inst&: *II, Known))
2116 return II;
2117
2118 Value *InnerPtr, *InnerMask;
2119 bool Changed = false;
2120 // Combine:
2121 // (ptrmask (ptrmask p, A), B)
2122 // -> (ptrmask p, (and A, B))
2123 if (match(V: II->getArgOperand(i: 0),
2124 P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ptrmask>(Op0: m_Value(V&: InnerPtr),
2125 Op1: m_Value(V&: InnerMask))))) {
2126 assert(II->getArgOperand(1)->getType() == InnerMask->getType() &&
2127 "Mask types must match");
2128 // TODO: If InnerMask == Op1, we could copy attributes from inner
2129 // callsite -> outer callsite.
2130 Value *NewMask = Builder.CreateAnd(LHS: II->getArgOperand(i: 1), RHS: InnerMask);
2131 replaceOperand(I&: CI, OpNum: 0, V: InnerPtr);
2132 replaceOperand(I&: CI, OpNum: 1, V: NewMask);
2133 Changed = true;
2134 }
2135
2136 // See if we can deduce non-null.
2137 if (!CI.hasRetAttr(Kind: Attribute::NonNull) &&
2138 (Known.isNonZero() ||
2139 isKnownNonZero(V: II, Q: getSimplifyQuery().getWithInstruction(I: II)))) {
2140 CI.addRetAttr(Kind: Attribute::NonNull);
2141 Changed = true;
2142 }
2143
2144 unsigned NewAlignmentLog =
2145 std::min(a: Value::MaxAlignmentExponent,
2146 b: std::min(a: BitWidth - 1, b: Known.countMinTrailingZeros()));
2147 // Known bits will capture if we had alignment information associated with
2148 // the pointer argument.
2149 if (NewAlignmentLog > Log2(A: CI.getRetAlign().valueOrOne())) {
2150 CI.addRetAttr(Attr: Attribute::getWithAlignment(
2151 Context&: CI.getContext(), Alignment: Align(uint64_t(1) << NewAlignmentLog)));
2152 Changed = true;
2153 }
2154 if (Changed)
2155 return &CI;
2156 break;
2157 }
2158 case Intrinsic::uadd_with_overflow:
2159 case Intrinsic::sadd_with_overflow: {
2160 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2161 return I;
2162
2163 // Given 2 constant operands whose sum does not overflow:
2164 // uaddo (X +nuw C0), C1 -> uaddo X, C0 + C1
2165 // saddo (X +nsw C0), C1 -> saddo X, C0 + C1
2166 Value *X;
2167 const APInt *C0, *C1;
2168 Value *Arg0 = II->getArgOperand(i: 0);
2169 Value *Arg1 = II->getArgOperand(i: 1);
2170 bool IsSigned = IID == Intrinsic::sadd_with_overflow;
2171 bool HasNWAdd = IsSigned
2172 ? match(V: Arg0, P: m_NSWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C0)))
2173 : match(V: Arg0, P: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: C0)));
2174 if (HasNWAdd && match(V: Arg1, P: m_APInt(Res&: C1))) {
2175 bool Overflow;
2176 APInt NewC =
2177 IsSigned ? C1->sadd_ov(RHS: *C0, Overflow) : C1->uadd_ov(RHS: *C0, Overflow);
2178 if (!Overflow)
2179 return replaceInstUsesWith(
2180 I&: *II, V: Builder.CreateBinaryIntrinsic(
2181 ID: IID, LHS: X, RHS: ConstantInt::get(Ty: Arg1->getType(), V: NewC)));
2182 }
2183 break;
2184 }
2185
2186 case Intrinsic::umul_with_overflow:
2187 case Intrinsic::smul_with_overflow:
2188 case Intrinsic::usub_with_overflow:
2189 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2190 return I;
2191 break;
2192
2193 case Intrinsic::ssub_with_overflow: {
2194 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2195 return I;
2196
2197 Constant *C;
2198 Value *Arg0 = II->getArgOperand(i: 0);
2199 Value *Arg1 = II->getArgOperand(i: 1);
2200 // Given a constant C that is not the minimum signed value
2201 // for an integer of a given bit width:
2202 //
2203 // ssubo X, C -> saddo X, -C
2204 if (match(V: Arg1, P: m_Constant(C)) && C->isNotMinSignedValue()) {
2205 Value *NegVal = ConstantExpr::getNeg(C);
2206 // Build a saddo call that is equivalent to the discovered
2207 // ssubo call.
2208 return replaceInstUsesWith(
2209 I&: *II, V: Builder.CreateBinaryIntrinsic(ID: Intrinsic::sadd_with_overflow,
2210 LHS: Arg0, RHS: NegVal));
2211 }
2212
2213 break;
2214 }
2215
2216 case Intrinsic::uadd_sat:
2217 case Intrinsic::sadd_sat:
2218 case Intrinsic::usub_sat:
2219 case Intrinsic::ssub_sat: {
2220 SaturatingInst *SI = cast<SaturatingInst>(Val: II);
2221 Type *Ty = SI->getType();
2222 Value *Arg0 = SI->getLHS();
2223 Value *Arg1 = SI->getRHS();
2224
2225 // Make use of known overflow information.
2226 OverflowResult OR = computeOverflow(BinaryOp: SI->getBinaryOp(), IsSigned: SI->isSigned(),
2227 LHS: Arg0, RHS: Arg1, CxtI: SI);
2228 switch (OR) {
2229 case OverflowResult::MayOverflow:
2230 break;
2231 case OverflowResult::NeverOverflows:
2232 if (SI->isSigned())
2233 return BinaryOperator::CreateNSW(Opc: SI->getBinaryOp(), V1: Arg0, V2: Arg1);
2234 else
2235 return BinaryOperator::CreateNUW(Opc: SI->getBinaryOp(), V1: Arg0, V2: Arg1);
2236 case OverflowResult::AlwaysOverflowsLow: {
2237 unsigned BitWidth = Ty->getScalarSizeInBits();
2238 APInt Min = APSInt::getMinValue(numBits: BitWidth, Unsigned: !SI->isSigned());
2239 return replaceInstUsesWith(I&: *SI, V: ConstantInt::get(Ty, V: Min));
2240 }
2241 case OverflowResult::AlwaysOverflowsHigh: {
2242 unsigned BitWidth = Ty->getScalarSizeInBits();
2243 APInt Max = APSInt::getMaxValue(numBits: BitWidth, Unsigned: !SI->isSigned());
2244 return replaceInstUsesWith(I&: *SI, V: ConstantInt::get(Ty, V: Max));
2245 }
2246 }
2247
2248 // usub_sat((sub nuw C, A), C1) -> usub_sat(usub_sat(C, C1), A)
2249 // which after that:
2250 // usub_sat((sub nuw C, A), C1) -> usub_sat(C - C1, A) if C1 u< C
2251 // usub_sat((sub nuw C, A), C1) -> 0 otherwise
2252 Constant *C, *C1;
2253 Value *A;
2254 if (IID == Intrinsic::usub_sat &&
2255 match(V: Arg0, P: m_NUWSub(L: m_ImmConstant(C), R: m_Value(V&: A))) &&
2256 match(V: Arg1, P: m_ImmConstant(C&: C1))) {
2257 auto *NewC = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: C, RHS: C1);
2258 auto *NewSub =
2259 Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: NewC, RHS: A);
2260 return replaceInstUsesWith(I&: *SI, V: NewSub);
2261 }
2262
2263 // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN
2264 if (IID == Intrinsic::ssub_sat && match(V: Arg1, P: m_Constant(C)) &&
2265 C->isNotMinSignedValue()) {
2266 Value *NegVal = ConstantExpr::getNeg(C);
2267 return replaceInstUsesWith(
2268 I&: *II, V: Builder.CreateBinaryIntrinsic(
2269 ID: Intrinsic::sadd_sat, LHS: Arg0, RHS: NegVal));
2270 }
2271
2272 // sat(sat(X + Val2) + Val) -> sat(X + (Val+Val2))
2273 // sat(sat(X - Val2) - Val) -> sat(X - (Val+Val2))
2274 // if Val and Val2 have the same sign
2275 if (auto *Other = dyn_cast<IntrinsicInst>(Val: Arg0)) {
2276 Value *X;
2277 const APInt *Val, *Val2;
2278 APInt NewVal;
2279 bool IsUnsigned =
2280 IID == Intrinsic::uadd_sat || IID == Intrinsic::usub_sat;
2281 if (Other->getIntrinsicID() == IID &&
2282 match(V: Arg1, P: m_APInt(Res&: Val)) &&
2283 match(V: Other->getArgOperand(i: 0), P: m_Value(V&: X)) &&
2284 match(V: Other->getArgOperand(i: 1), P: m_APInt(Res&: Val2))) {
2285 if (IsUnsigned)
2286 NewVal = Val->uadd_sat(RHS: *Val2);
2287 else if (Val->isNonNegative() == Val2->isNonNegative()) {
2288 bool Overflow;
2289 NewVal = Val->sadd_ov(RHS: *Val2, Overflow);
2290 if (Overflow) {
2291 // Both adds together may add more than SignedMaxValue
2292 // without saturating the final result.
2293 break;
2294 }
2295 } else {
2296 // Cannot fold saturated addition with different signs.
2297 break;
2298 }
2299
2300 return replaceInstUsesWith(
2301 I&: *II, V: Builder.CreateBinaryIntrinsic(
2302 ID: IID, LHS: X, RHS: ConstantInt::get(Ty: II->getType(), V: NewVal)));
2303 }
2304 }
2305 break;
2306 }
2307
2308 case Intrinsic::minnum:
2309 case Intrinsic::maxnum:
2310 case Intrinsic::minimum:
2311 case Intrinsic::maximum: {
2312 Value *Arg0 = II->getArgOperand(i: 0);
2313 Value *Arg1 = II->getArgOperand(i: 1);
2314 Value *X, *Y;
2315 if (match(V: Arg0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Arg1, P: m_FNeg(X: m_Value(V&: Y))) &&
2316 (Arg0->hasOneUse() || Arg1->hasOneUse())) {
2317 // If both operands are negated, invert the call and negate the result:
2318 // min(-X, -Y) --> -(max(X, Y))
2319 // max(-X, -Y) --> -(min(X, Y))
2320 Intrinsic::ID NewIID;
2321 switch (IID) {
2322 case Intrinsic::maxnum:
2323 NewIID = Intrinsic::minnum;
2324 break;
2325 case Intrinsic::minnum:
2326 NewIID = Intrinsic::maxnum;
2327 break;
2328 case Intrinsic::maximum:
2329 NewIID = Intrinsic::minimum;
2330 break;
2331 case Intrinsic::minimum:
2332 NewIID = Intrinsic::maximum;
2333 break;
2334 default:
2335 llvm_unreachable("unexpected intrinsic ID");
2336 }
2337 Value *NewCall = Builder.CreateBinaryIntrinsic(ID: NewIID, LHS: X, RHS: Y, FMFSource: II);
2338 Instruction *FNeg = UnaryOperator::CreateFNeg(V: NewCall);
2339 FNeg->copyIRFlags(V: II);
2340 return FNeg;
2341 }
2342
2343 // m(m(X, C2), C1) -> m(X, C)
2344 const APFloat *C1, *C2;
2345 if (auto *M = dyn_cast<IntrinsicInst>(Val: Arg0)) {
2346 if (M->getIntrinsicID() == IID && match(V: Arg1, P: m_APFloat(Res&: C1)) &&
2347 ((match(V: M->getArgOperand(i: 0), P: m_Value(V&: X)) &&
2348 match(V: M->getArgOperand(i: 1), P: m_APFloat(Res&: C2))) ||
2349 (match(V: M->getArgOperand(i: 1), P: m_Value(V&: X)) &&
2350 match(V: M->getArgOperand(i: 0), P: m_APFloat(Res&: C2))))) {
2351 APFloat Res(0.0);
2352 switch (IID) {
2353 case Intrinsic::maxnum:
2354 Res = maxnum(A: *C1, B: *C2);
2355 break;
2356 case Intrinsic::minnum:
2357 Res = minnum(A: *C1, B: *C2);
2358 break;
2359 case Intrinsic::maximum:
2360 Res = maximum(A: *C1, B: *C2);
2361 break;
2362 case Intrinsic::minimum:
2363 Res = minimum(A: *C1, B: *C2);
2364 break;
2365 default:
2366 llvm_unreachable("unexpected intrinsic ID");
2367 }
2368 Value *V = Builder.CreateBinaryIntrinsic(
2369 ID: IID, LHS: X, RHS: ConstantFP::get(Ty: Arg0->getType(), V: Res), FMFSource: II);
2370 // TODO: Conservatively intersecting FMF. If Res == C2, the transform
2371 // was a simplification (so Arg0 and its original flags could
2372 // propagate?)
2373 if (auto *CI = dyn_cast<CallInst>(Val: V))
2374 CI->andIRFlags(V: M);
2375 return replaceInstUsesWith(I&: *II, V);
2376 }
2377 }
2378
2379 // m((fpext X), (fpext Y)) -> fpext (m(X, Y))
2380 if (match(V: Arg0, P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: X)))) &&
2381 match(V: Arg1, P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: Y)))) &&
2382 X->getType() == Y->getType()) {
2383 Value *NewCall =
2384 Builder.CreateBinaryIntrinsic(ID: IID, LHS: X, RHS: Y, FMFSource: II, Name: II->getName());
2385 return new FPExtInst(NewCall, II->getType());
2386 }
2387
2388 // max X, -X --> fabs X
2389 // min X, -X --> -(fabs X)
2390 // TODO: Remove one-use limitation? That is obviously better for max,
2391 // hence why we don't check for one-use for that. However,
2392 // it would be an extra instruction for min (fnabs), but
2393 // that is still likely better for analysis and codegen.
2394 auto IsMinMaxOrXNegX = [IID, &X](Value *Op0, Value *Op1) {
2395 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_Specific(V: X)))
2396 return Op0->hasOneUse() ||
2397 (IID != Intrinsic::minimum && IID != Intrinsic::minnum);
2398 return false;
2399 };
2400
2401 if (IsMinMaxOrXNegX(Arg0, Arg1) || IsMinMaxOrXNegX(Arg1, Arg0)) {
2402 Value *R = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: II);
2403 if (IID == Intrinsic::minimum || IID == Intrinsic::minnum)
2404 R = Builder.CreateFNegFMF(V: R, FMFSource: II);
2405 return replaceInstUsesWith(I&: *II, V: R);
2406 }
2407
2408 break;
2409 }
2410 case Intrinsic::matrix_multiply: {
2411 // Optimize negation in matrix multiplication.
2412
2413 // -A * -B -> A * B
2414 Value *A, *B;
2415 if (match(V: II->getArgOperand(i: 0), P: m_FNeg(X: m_Value(V&: A))) &&
2416 match(V: II->getArgOperand(i: 1), P: m_FNeg(X: m_Value(V&: B)))) {
2417 replaceOperand(I&: *II, OpNum: 0, V: A);
2418 replaceOperand(I&: *II, OpNum: 1, V: B);
2419 return II;
2420 }
2421
2422 Value *Op0 = II->getOperand(i_nocapture: 0);
2423 Value *Op1 = II->getOperand(i_nocapture: 1);
2424 Value *OpNotNeg, *NegatedOp;
2425 unsigned NegatedOpArg, OtherOpArg;
2426 if (match(V: Op0, P: m_FNeg(X: m_Value(V&: OpNotNeg)))) {
2427 NegatedOp = Op0;
2428 NegatedOpArg = 0;
2429 OtherOpArg = 1;
2430 } else if (match(V: Op1, P: m_FNeg(X: m_Value(V&: OpNotNeg)))) {
2431 NegatedOp = Op1;
2432 NegatedOpArg = 1;
2433 OtherOpArg = 0;
2434 } else
2435 // Multiplication doesn't have a negated operand.
2436 break;
2437
2438 // Only optimize if the negated operand has only one use.
2439 if (!NegatedOp->hasOneUse())
2440 break;
2441
2442 Value *OtherOp = II->getOperand(i_nocapture: OtherOpArg);
2443 VectorType *RetTy = cast<VectorType>(Val: II->getType());
2444 VectorType *NegatedOpTy = cast<VectorType>(Val: NegatedOp->getType());
2445 VectorType *OtherOpTy = cast<VectorType>(Val: OtherOp->getType());
2446 ElementCount NegatedCount = NegatedOpTy->getElementCount();
2447 ElementCount OtherCount = OtherOpTy->getElementCount();
2448 ElementCount RetCount = RetTy->getElementCount();
2449 // (-A) * B -> A * (-B), if it is cheaper to negate B and vice versa.
2450 if (ElementCount::isKnownGT(LHS: NegatedCount, RHS: OtherCount) &&
2451 ElementCount::isKnownLT(LHS: OtherCount, RHS: RetCount)) {
2452 Value *InverseOtherOp = Builder.CreateFNeg(V: OtherOp);
2453 replaceOperand(I&: *II, OpNum: NegatedOpArg, V: OpNotNeg);
2454 replaceOperand(I&: *II, OpNum: OtherOpArg, V: InverseOtherOp);
2455 return II;
2456 }
2457 // (-A) * B -> -(A * B), if it is cheaper to negate the result
2458 if (ElementCount::isKnownGT(LHS: NegatedCount, RHS: RetCount)) {
2459 SmallVector<Value *, 5> NewArgs(II->args());
2460 NewArgs[NegatedOpArg] = OpNotNeg;
2461 Instruction *NewMul =
2462 Builder.CreateIntrinsic(RetTy: II->getType(), ID: IID, Args: NewArgs, FMFSource: II);
2463 return replaceInstUsesWith(I&: *II, V: Builder.CreateFNegFMF(V: NewMul, FMFSource: II));
2464 }
2465 break;
2466 }
2467 case Intrinsic::fmuladd: {
2468 // Try to simplify the underlying FMul.
2469 if (Value *V = simplifyFMulInst(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1),
2470 FMF: II->getFastMathFlags(),
2471 Q: SQ.getWithInstruction(I: II))) {
2472 auto *FAdd = BinaryOperator::CreateFAdd(V1: V, V2: II->getArgOperand(i: 2));
2473 FAdd->copyFastMathFlags(I: II);
2474 return FAdd;
2475 }
2476
2477 [[fallthrough]];
2478 }
2479 case Intrinsic::fma: {
2480 // fma fneg(x), fneg(y), z -> fma x, y, z
2481 Value *Src0 = II->getArgOperand(i: 0);
2482 Value *Src1 = II->getArgOperand(i: 1);
2483 Value *X, *Y;
2484 if (match(V: Src0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Src1, P: m_FNeg(X: m_Value(V&: Y)))) {
2485 replaceOperand(I&: *II, OpNum: 0, V: X);
2486 replaceOperand(I&: *II, OpNum: 1, V: Y);
2487 return II;
2488 }
2489
2490 // fma fabs(x), fabs(x), z -> fma x, x, z
2491 if (match(V: Src0, P: m_FAbs(Op0: m_Value(V&: X))) &&
2492 match(V: Src1, P: m_FAbs(Op0: m_Specific(V: X)))) {
2493 replaceOperand(I&: *II, OpNum: 0, V: X);
2494 replaceOperand(I&: *II, OpNum: 1, V: X);
2495 return II;
2496 }
2497
2498 // Try to simplify the underlying FMul. We can only apply simplifications
2499 // that do not require rounding.
2500 if (Value *V = simplifyFMAFMul(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1),
2501 FMF: II->getFastMathFlags(),
2502 Q: SQ.getWithInstruction(I: II))) {
2503 auto *FAdd = BinaryOperator::CreateFAdd(V1: V, V2: II->getArgOperand(i: 2));
2504 FAdd->copyFastMathFlags(I: II);
2505 return FAdd;
2506 }
2507
2508 // fma x, y, 0 -> fmul x, y
2509 // This is always valid for -0.0, but requires nsz for +0.0 as
2510 // -0.0 + 0.0 = 0.0, which would not be the same as the fmul on its own.
2511 if (match(V: II->getArgOperand(i: 2), P: m_NegZeroFP()) ||
2512 (match(V: II->getArgOperand(i: 2), P: m_PosZeroFP()) &&
2513 II->getFastMathFlags().noSignedZeros()))
2514 return BinaryOperator::CreateFMulFMF(V1: Src0, V2: Src1, FMFSource: II);
2515
2516 break;
2517 }
2518 case Intrinsic::copysign: {
2519 Value *Mag = II->getArgOperand(i: 0), *Sign = II->getArgOperand(i: 1);
2520 if (std::optional<bool> KnownSignBit = computeKnownFPSignBit(
2521 V: Sign, /*Depth=*/0, SQ: getSimplifyQuery().getWithInstruction(I: II))) {
2522 if (*KnownSignBit) {
2523 // If we know that the sign argument is negative, reduce to FNABS:
2524 // copysign Mag, -Sign --> fneg (fabs Mag)
2525 Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Mag, FMFSource: II);
2526 return replaceInstUsesWith(I&: *II, V: Builder.CreateFNegFMF(V: Fabs, FMFSource: II));
2527 }
2528
2529 // If we know that the sign argument is positive, reduce to FABS:
2530 // copysign Mag, +Sign --> fabs Mag
2531 Value *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: Mag, FMFSource: II);
2532 return replaceInstUsesWith(I&: *II, V: Fabs);
2533 }
2534
2535 // Propagate sign argument through nested calls:
2536 // copysign Mag, (copysign ?, X) --> copysign Mag, X
2537 Value *X;
2538 if (match(V: Sign, P: m_Intrinsic<Intrinsic::copysign>(Op0: m_Value(), Op1: m_Value(V&: X))))
2539 return replaceOperand(I&: *II, OpNum: 1, V: X);
2540
2541 // Clear sign-bit of constant magnitude:
2542 // copysign -MagC, X --> copysign MagC, X
2543 // TODO: Support constant folding for fabs
2544 const APFloat *MagC;
2545 if (match(V: Mag, P: m_APFloat(Res&: MagC)) && MagC->isNegative()) {
2546 APFloat PosMagC = *MagC;
2547 PosMagC.clearSign();
2548 return replaceOperand(I&: *II, OpNum: 0, V: ConstantFP::get(Ty: Mag->getType(), V: PosMagC));
2549 }
2550
2551 // Peek through changes of magnitude's sign-bit. This call rewrites those:
2552 // copysign (fabs X), Sign --> copysign X, Sign
2553 // copysign (fneg X), Sign --> copysign X, Sign
2554 if (match(V: Mag, P: m_FAbs(Op0: m_Value(V&: X))) || match(V: Mag, P: m_FNeg(X: m_Value(V&: X))))
2555 return replaceOperand(I&: *II, OpNum: 0, V: X);
2556
2557 break;
2558 }
2559 case Intrinsic::fabs: {
2560 Value *Cond, *TVal, *FVal;
2561 Value *Arg = II->getArgOperand(i: 0);
2562 Value *X;
2563 // fabs (-X) --> fabs (X)
2564 if (match(V: Arg, P: m_FNeg(X: m_Value(V&: X)))) {
2565 CallInst *Fabs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: II);
2566 return replaceInstUsesWith(I&: CI, V: Fabs);
2567 }
2568
2569 if (match(V: Arg, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TVal), R: m_Value(V&: FVal)))) {
2570 // fabs (select Cond, TrueC, FalseC) --> select Cond, AbsT, AbsF
2571 if (isa<Constant>(Val: TVal) || isa<Constant>(Val: FVal)) {
2572 CallInst *AbsT = Builder.CreateCall(Callee: II->getCalledFunction(), Args: {TVal});
2573 CallInst *AbsF = Builder.CreateCall(Callee: II->getCalledFunction(), Args: {FVal});
2574 SelectInst *SI = SelectInst::Create(C: Cond, S1: AbsT, S2: AbsF);
2575 FastMathFlags FMF1 = II->getFastMathFlags();
2576 FastMathFlags FMF2 = cast<SelectInst>(Val: Arg)->getFastMathFlags();
2577 FMF2.setNoSignedZeros(false);
2578 SI->setFastMathFlags(FMF1 | FMF2);
2579 return SI;
2580 }
2581 // fabs (select Cond, -FVal, FVal) --> fabs FVal
2582 if (match(V: TVal, P: m_FNeg(X: m_Specific(V: FVal))))
2583 return replaceOperand(I&: *II, OpNum: 0, V: FVal);
2584 // fabs (select Cond, TVal, -TVal) --> fabs TVal
2585 if (match(V: FVal, P: m_FNeg(X: m_Specific(V: TVal))))
2586 return replaceOperand(I&: *II, OpNum: 0, V: TVal);
2587 }
2588
2589 Value *Magnitude, *Sign;
2590 if (match(V: II->getArgOperand(i: 0),
2591 P: m_CopySign(Op0: m_Value(V&: Magnitude), Op1: m_Value(V&: Sign)))) {
2592 // fabs (copysign x, y) -> (fabs x)
2593 CallInst *AbsSign =
2594 Builder.CreateCall(Callee: II->getCalledFunction(), Args: {Magnitude});
2595 AbsSign->copyFastMathFlags(I: II);
2596 return replaceInstUsesWith(I&: *II, V: AbsSign);
2597 }
2598
2599 [[fallthrough]];
2600 }
2601 case Intrinsic::ceil:
2602 case Intrinsic::floor:
2603 case Intrinsic::round:
2604 case Intrinsic::roundeven:
2605 case Intrinsic::nearbyint:
2606 case Intrinsic::rint:
2607 case Intrinsic::trunc: {
2608 Value *ExtSrc;
2609 if (match(V: II->getArgOperand(i: 0), P: m_OneUse(SubPattern: m_FPExt(Op: m_Value(V&: ExtSrc))))) {
2610 // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x)
2611 Value *NarrowII = Builder.CreateUnaryIntrinsic(ID: IID, V: ExtSrc, FMFSource: II);
2612 return new FPExtInst(NarrowII, II->getType());
2613 }
2614 break;
2615 }
2616 case Intrinsic::cos:
2617 case Intrinsic::amdgcn_cos: {
2618 Value *X, *Sign;
2619 Value *Src = II->getArgOperand(i: 0);
2620 if (match(V: Src, P: m_FNeg(X: m_Value(V&: X))) || match(V: Src, P: m_FAbs(Op0: m_Value(V&: X))) ||
2621 match(V: Src, P: m_CopySign(Op0: m_Value(V&: X), Op1: m_Value(V&: Sign)))) {
2622 // cos(-x) --> cos(x)
2623 // cos(fabs(x)) --> cos(x)
2624 // cos(copysign(x, y)) --> cos(x)
2625 return replaceOperand(I&: *II, OpNum: 0, V: X);
2626 }
2627 break;
2628 }
2629 case Intrinsic::sin:
2630 case Intrinsic::amdgcn_sin: {
2631 Value *X;
2632 if (match(V: II->getArgOperand(i: 0), P: m_OneUse(SubPattern: m_FNeg(X: m_Value(V&: X))))) {
2633 // sin(-x) --> -sin(x)
2634 Value *NewSin = Builder.CreateUnaryIntrinsic(ID: IID, V: X, FMFSource: II);
2635 return UnaryOperator::CreateFNegFMF(Op: NewSin, FMFSource: II);
2636 }
2637 break;
2638 }
2639 case Intrinsic::ldexp: {
2640 // ldexp(ldexp(x, a), b) -> ldexp(x, a + b)
2641 //
2642 // The danger is if the first ldexp would overflow to infinity or underflow
2643 // to zero, but the combined exponent avoids it. We ignore this with
2644 // reassoc.
2645 //
2646 // It's also safe to fold if we know both exponents are >= 0 or <= 0 since
2647 // it would just double down on the overflow/underflow which would occur
2648 // anyway.
2649 //
2650 // TODO: Could do better if we had range tracking for the input value
2651 // exponent. Also could broaden sign check to cover == 0 case.
2652 Value *Src = II->getArgOperand(i: 0);
2653 Value *Exp = II->getArgOperand(i: 1);
2654 Value *InnerSrc;
2655 Value *InnerExp;
2656 if (match(V: Src, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::ldexp>(
2657 Op0: m_Value(V&: InnerSrc), Op1: m_Value(V&: InnerExp)))) &&
2658 Exp->getType() == InnerExp->getType()) {
2659 FastMathFlags FMF = II->getFastMathFlags();
2660 FastMathFlags InnerFlags = cast<FPMathOperator>(Val: Src)->getFastMathFlags();
2661
2662 if ((FMF.allowReassoc() && InnerFlags.allowReassoc()) ||
2663 signBitMustBeTheSame(Op0: Exp, Op1: InnerExp, SQ: SQ.getWithInstruction(I: II))) {
2664 // TODO: Add nsw/nuw probably safe if integer type exceeds exponent
2665 // width.
2666 Value *NewExp = Builder.CreateAdd(LHS: InnerExp, RHS: Exp);
2667 II->setArgOperand(i: 1, v: NewExp);
2668 II->setFastMathFlags(InnerFlags); // Or the inner flags.
2669 return replaceOperand(I&: *II, OpNum: 0, V: InnerSrc);
2670 }
2671 }
2672
2673 // ldexp(x, zext(i1 y)) -> fmul x, (select y, 2.0, 1.0)
2674 // ldexp(x, sext(i1 y)) -> fmul x, (select y, 0.5, 1.0)
2675 Value *ExtSrc;
2676 if (match(V: Exp, P: m_ZExt(Op: m_Value(V&: ExtSrc))) &&
2677 ExtSrc->getType()->getScalarSizeInBits() == 1) {
2678 Value *Select =
2679 Builder.CreateSelect(C: ExtSrc, True: ConstantFP::get(Ty: II->getType(), V: 2.0),
2680 False: ConstantFP::get(Ty: II->getType(), V: 1.0));
2681 return BinaryOperator::CreateFMulFMF(V1: Src, V2: Select, FMFSource: II);
2682 }
2683 if (match(V: Exp, P: m_SExt(Op: m_Value(V&: ExtSrc))) &&
2684 ExtSrc->getType()->getScalarSizeInBits() == 1) {
2685 Value *Select =
2686 Builder.CreateSelect(C: ExtSrc, True: ConstantFP::get(Ty: II->getType(), V: 0.5),
2687 False: ConstantFP::get(Ty: II->getType(), V: 1.0));
2688 return BinaryOperator::CreateFMulFMF(V1: Src, V2: Select, FMFSource: II);
2689 }
2690
2691 // ldexp(x, c ? exp : 0) -> c ? ldexp(x, exp) : x
2692 // ldexp(x, c ? 0 : exp) -> c ? x : ldexp(x, exp)
2693 ///
2694 // TODO: If we cared, should insert a canonicalize for x
2695 Value *SelectCond, *SelectLHS, *SelectRHS;
2696 if (match(V: II->getArgOperand(i: 1),
2697 P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: SelectCond), L: m_Value(V&: SelectLHS),
2698 R: m_Value(V&: SelectRHS))))) {
2699 Value *NewLdexp = nullptr;
2700 Value *Select = nullptr;
2701 if (match(V: SelectRHS, P: m_ZeroInt())) {
2702 NewLdexp = Builder.CreateLdexp(Src, Exp: SelectLHS);
2703 Select = Builder.CreateSelect(C: SelectCond, True: NewLdexp, False: Src);
2704 } else if (match(V: SelectLHS, P: m_ZeroInt())) {
2705 NewLdexp = Builder.CreateLdexp(Src, Exp: SelectRHS);
2706 Select = Builder.CreateSelect(C: SelectCond, True: Src, False: NewLdexp);
2707 }
2708
2709 if (NewLdexp) {
2710 Select->takeName(V: II);
2711 cast<Instruction>(Val: NewLdexp)->copyFastMathFlags(I: II);
2712 return replaceInstUsesWith(I&: *II, V: Select);
2713 }
2714 }
2715
2716 break;
2717 }
2718 case Intrinsic::ptrauth_auth:
2719 case Intrinsic::ptrauth_resign: {
2720 // (sign|resign) + (auth|resign) can be folded by omitting the middle
2721 // sign+auth component if the key and discriminator match.
2722 bool NeedSign = II->getIntrinsicID() == Intrinsic::ptrauth_resign;
2723 Value *Ptr = II->getArgOperand(i: 0);
2724 Value *Key = II->getArgOperand(i: 1);
2725 Value *Disc = II->getArgOperand(i: 2);
2726
2727 // AuthKey will be the key we need to end up authenticating against in
2728 // whatever we replace this sequence with.
2729 Value *AuthKey = nullptr, *AuthDisc = nullptr, *BasePtr;
2730 if (const auto *CI = dyn_cast<CallBase>(Val: Ptr)) {
2731 BasePtr = CI->getArgOperand(i: 0);
2732 if (CI->getIntrinsicID() == Intrinsic::ptrauth_sign) {
2733 if (CI->getArgOperand(i: 1) != Key || CI->getArgOperand(i: 2) != Disc)
2734 break;
2735 } else if (CI->getIntrinsicID() == Intrinsic::ptrauth_resign) {
2736 if (CI->getArgOperand(i: 3) != Key || CI->getArgOperand(i: 4) != Disc)
2737 break;
2738 AuthKey = CI->getArgOperand(i: 1);
2739 AuthDisc = CI->getArgOperand(i: 2);
2740 } else
2741 break;
2742 } else if (const auto *PtrToInt = dyn_cast<PtrToIntOperator>(Val: Ptr)) {
2743 // ptrauth constants are equivalent to a call to @llvm.ptrauth.sign for
2744 // our purposes, so check for that too.
2745 const auto *CPA = dyn_cast<ConstantPtrAuth>(Val: PtrToInt->getOperand(i_nocapture: 0));
2746 if (!CPA || !CPA->isKnownCompatibleWith(Key, Discriminator: Disc, DL))
2747 break;
2748
2749 // resign(ptrauth(p,ks,ds),ks,ds,kr,dr) -> ptrauth(p,kr,dr)
2750 if (NeedSign && isa<ConstantInt>(Val: II->getArgOperand(i: 4))) {
2751 auto *SignKey = cast<ConstantInt>(Val: II->getArgOperand(i: 3));
2752 auto *SignDisc = cast<ConstantInt>(Val: II->getArgOperand(i: 4));
2753 auto *SignAddrDisc = ConstantPointerNull::get(T: Builder.getPtrTy());
2754 auto *NewCPA = ConstantPtrAuth::get(Ptr: CPA->getPointer(), Key: SignKey,
2755 Disc: SignDisc, AddrDisc: SignAddrDisc);
2756 replaceInstUsesWith(
2757 I&: *II, V: ConstantExpr::getPointerCast(C: NewCPA, Ty: II->getType()));
2758 return eraseInstFromFunction(I&: *II);
2759 }
2760
2761 // auth(ptrauth(p,k,d),k,d) -> p
2762 BasePtr = Builder.CreatePtrToInt(V: CPA->getPointer(), DestTy: II->getType());
2763 } else
2764 break;
2765
2766 unsigned NewIntrin;
2767 if (AuthKey && NeedSign) {
2768 // resign(0,1) + resign(1,2) = resign(0, 2)
2769 NewIntrin = Intrinsic::ptrauth_resign;
2770 } else if (AuthKey) {
2771 // resign(0,1) + auth(1) = auth(0)
2772 NewIntrin = Intrinsic::ptrauth_auth;
2773 } else if (NeedSign) {
2774 // sign(0) + resign(0, 1) = sign(1)
2775 NewIntrin = Intrinsic::ptrauth_sign;
2776 } else {
2777 // sign(0) + auth(0) = nop
2778 replaceInstUsesWith(I&: *II, V: BasePtr);
2779 return eraseInstFromFunction(I&: *II);
2780 }
2781
2782 SmallVector<Value *, 4> CallArgs;
2783 CallArgs.push_back(Elt: BasePtr);
2784 if (AuthKey) {
2785 CallArgs.push_back(Elt: AuthKey);
2786 CallArgs.push_back(Elt: AuthDisc);
2787 }
2788
2789 if (NeedSign) {
2790 CallArgs.push_back(Elt: II->getArgOperand(i: 3));
2791 CallArgs.push_back(Elt: II->getArgOperand(i: 4));
2792 }
2793
2794 Function *NewFn = Intrinsic::getDeclaration(M: II->getModule(), id: NewIntrin);
2795 return CallInst::Create(Func: NewFn, Args: CallArgs);
2796 }
2797 case Intrinsic::arm_neon_vtbl1:
2798 case Intrinsic::aarch64_neon_tbl1:
2799 if (Value *V = simplifyNeonTbl1(II: *II, Builder))
2800 return replaceInstUsesWith(I&: *II, V);
2801 break;
2802
2803 case Intrinsic::arm_neon_vmulls:
2804 case Intrinsic::arm_neon_vmullu:
2805 case Intrinsic::aarch64_neon_smull:
2806 case Intrinsic::aarch64_neon_umull: {
2807 Value *Arg0 = II->getArgOperand(i: 0);
2808 Value *Arg1 = II->getArgOperand(i: 1);
2809
2810 // Handle mul by zero first:
2811 if (isa<ConstantAggregateZero>(Val: Arg0) || isa<ConstantAggregateZero>(Val: Arg1)) {
2812 return replaceInstUsesWith(I&: CI, V: ConstantAggregateZero::get(Ty: II->getType()));
2813 }
2814
2815 // Check for constant LHS & RHS - in this case we just simplify.
2816 bool Zext = (IID == Intrinsic::arm_neon_vmullu ||
2817 IID == Intrinsic::aarch64_neon_umull);
2818 VectorType *NewVT = cast<VectorType>(Val: II->getType());
2819 if (Constant *CV0 = dyn_cast<Constant>(Val: Arg0)) {
2820 if (Constant *CV1 = dyn_cast<Constant>(Val: Arg1)) {
2821 Value *V0 = Builder.CreateIntCast(V: CV0, DestTy: NewVT, /*isSigned=*/!Zext);
2822 Value *V1 = Builder.CreateIntCast(V: CV1, DestTy: NewVT, /*isSigned=*/!Zext);
2823 return replaceInstUsesWith(I&: CI, V: Builder.CreateMul(LHS: V0, RHS: V1));
2824 }
2825
2826 // Couldn't simplify - canonicalize constant to the RHS.
2827 std::swap(a&: Arg0, b&: Arg1);
2828 }
2829
2830 // Handle mul by one:
2831 if (Constant *CV1 = dyn_cast<Constant>(Val: Arg1))
2832 if (ConstantInt *Splat =
2833 dyn_cast_or_null<ConstantInt>(Val: CV1->getSplatValue()))
2834 if (Splat->isOne())
2835 return CastInst::CreateIntegerCast(S: Arg0, Ty: II->getType(),
2836 /*isSigned=*/!Zext);
2837
2838 break;
2839 }
2840 case Intrinsic::arm_neon_aesd:
2841 case Intrinsic::arm_neon_aese:
2842 case Intrinsic::aarch64_crypto_aesd:
2843 case Intrinsic::aarch64_crypto_aese: {
2844 Value *DataArg = II->getArgOperand(i: 0);
2845 Value *KeyArg = II->getArgOperand(i: 1);
2846
2847 // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR
2848 Value *Data, *Key;
2849 if (match(V: KeyArg, P: m_ZeroInt()) &&
2850 match(V: DataArg, P: m_Xor(L: m_Value(V&: Data), R: m_Value(V&: Key)))) {
2851 replaceOperand(I&: *II, OpNum: 0, V: Data);
2852 replaceOperand(I&: *II, OpNum: 1, V: Key);
2853 return II;
2854 }
2855 break;
2856 }
2857 case Intrinsic::hexagon_V6_vandvrt:
2858 case Intrinsic::hexagon_V6_vandvrt_128B: {
2859 // Simplify Q -> V -> Q conversion.
2860 if (auto Op0 = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0))) {
2861 Intrinsic::ID ID0 = Op0->getIntrinsicID();
2862 if (ID0 != Intrinsic::hexagon_V6_vandqrt &&
2863 ID0 != Intrinsic::hexagon_V6_vandqrt_128B)
2864 break;
2865 Value *Bytes = Op0->getArgOperand(i: 1), *Mask = II->getArgOperand(i: 1);
2866 uint64_t Bytes1 = computeKnownBits(V: Bytes, Depth: 0, CxtI: Op0).One.getZExtValue();
2867 uint64_t Mask1 = computeKnownBits(V: Mask, Depth: 0, CxtI: II).One.getZExtValue();
2868 // Check if every byte has common bits in Bytes and Mask.
2869 uint64_t C = Bytes1 & Mask1;
2870 if ((C & 0xFF) && (C & 0xFF00) && (C & 0xFF0000) && (C & 0xFF000000))
2871 return replaceInstUsesWith(I&: *II, V: Op0->getArgOperand(i: 0));
2872 }
2873 break;
2874 }
2875 case Intrinsic::stackrestore: {
2876 enum class ClassifyResult {
2877 None,
2878 Alloca,
2879 StackRestore,
2880 CallWithSideEffects,
2881 };
2882 auto Classify = [](const Instruction *I) {
2883 if (isa<AllocaInst>(Val: I))
2884 return ClassifyResult::Alloca;
2885
2886 if (auto *CI = dyn_cast<CallInst>(Val: I)) {
2887 if (auto *II = dyn_cast<IntrinsicInst>(Val: CI)) {
2888 if (II->getIntrinsicID() == Intrinsic::stackrestore)
2889 return ClassifyResult::StackRestore;
2890
2891 if (II->mayHaveSideEffects())
2892 return ClassifyResult::CallWithSideEffects;
2893 } else {
2894 // Consider all non-intrinsic calls to be side effects
2895 return ClassifyResult::CallWithSideEffects;
2896 }
2897 }
2898
2899 return ClassifyResult::None;
2900 };
2901
2902 // If the stacksave and the stackrestore are in the same BB, and there is
2903 // no intervening call, alloca, or stackrestore of a different stacksave,
2904 // remove the restore. This can happen when variable allocas are DCE'd.
2905 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0))) {
2906 if (SS->getIntrinsicID() == Intrinsic::stacksave &&
2907 SS->getParent() == II->getParent()) {
2908 BasicBlock::iterator BI(SS);
2909 bool CannotRemove = false;
2910 for (++BI; &*BI != II; ++BI) {
2911 switch (Classify(&*BI)) {
2912 case ClassifyResult::None:
2913 // So far so good, look at next instructions.
2914 break;
2915
2916 case ClassifyResult::StackRestore:
2917 // If we found an intervening stackrestore for a different
2918 // stacksave, we can't remove the stackrestore. Otherwise, continue.
2919 if (cast<IntrinsicInst>(Val&: *BI).getArgOperand(i: 0) != SS)
2920 CannotRemove = true;
2921 break;
2922
2923 case ClassifyResult::Alloca:
2924 case ClassifyResult::CallWithSideEffects:
2925 // If we found an alloca, a non-intrinsic call, or an intrinsic
2926 // call with side effects, we can't remove the stackrestore.
2927 CannotRemove = true;
2928 break;
2929 }
2930 if (CannotRemove)
2931 break;
2932 }
2933
2934 if (!CannotRemove)
2935 return eraseInstFromFunction(I&: CI);
2936 }
2937 }
2938
2939 // Scan down this block to see if there is another stack restore in the
2940 // same block without an intervening call/alloca.
2941 BasicBlock::iterator BI(II);
2942 Instruction *TI = II->getParent()->getTerminator();
2943 bool CannotRemove = false;
2944 for (++BI; &*BI != TI; ++BI) {
2945 switch (Classify(&*BI)) {
2946 case ClassifyResult::None:
2947 // So far so good, look at next instructions.
2948 break;
2949
2950 case ClassifyResult::StackRestore:
2951 // If there is a stackrestore below this one, remove this one.
2952 return eraseInstFromFunction(I&: CI);
2953
2954 case ClassifyResult::Alloca:
2955 case ClassifyResult::CallWithSideEffects:
2956 // If we found an alloca, a non-intrinsic call, or an intrinsic call
2957 // with side effects (such as llvm.stacksave and llvm.read_register),
2958 // we can't remove the stack restore.
2959 CannotRemove = true;
2960 break;
2961 }
2962 if (CannotRemove)
2963 break;
2964 }
2965
2966 // If the stack restore is in a return, resume, or unwind block and if there
2967 // are no allocas or calls between the restore and the return, nuke the
2968 // restore.
2969 if (!CannotRemove && (isa<ReturnInst>(Val: TI) || isa<ResumeInst>(Val: TI)))
2970 return eraseInstFromFunction(I&: CI);
2971 break;
2972 }
2973 case Intrinsic::lifetime_end:
2974 // Asan needs to poison memory to detect invalid access which is possible
2975 // even for empty lifetime range.
2976 if (II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeAddress) ||
2977 II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeMemory) ||
2978 II->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeHWAddress))
2979 break;
2980
2981 if (removeTriviallyEmptyRange(EndI&: *II, IC&: *this, IsStart: [](const IntrinsicInst &I) {
2982 return I.getIntrinsicID() == Intrinsic::lifetime_start;
2983 }))
2984 return nullptr;
2985 break;
2986 case Intrinsic::assume: {
2987 Value *IIOperand = II->getArgOperand(i: 0);
2988 SmallVector<OperandBundleDef, 4> OpBundles;
2989 II->getOperandBundlesAsDefs(Defs&: OpBundles);
2990
2991 /// This will remove the boolean Condition from the assume given as
2992 /// argument and remove the assume if it becomes useless.
2993 /// always returns nullptr for use as a return values.
2994 auto RemoveConditionFromAssume = [&](Instruction *Assume) -> Instruction * {
2995 assert(isa<AssumeInst>(Assume));
2996 if (isAssumeWithEmptyBundle(Assume: *cast<AssumeInst>(Val: II)))
2997 return eraseInstFromFunction(I&: CI);
2998 replaceUse(U&: II->getOperandUse(i: 0), NewValue: ConstantInt::getTrue(Context&: II->getContext()));
2999 return nullptr;
3000 };
3001 // Remove an assume if it is followed by an identical assume.
3002 // TODO: Do we need this? Unless there are conflicting assumptions, the
3003 // computeKnownBits(IIOperand) below here eliminates redundant assumes.
3004 Instruction *Next = II->getNextNonDebugInstruction();
3005 if (match(V: Next, P: m_Intrinsic<Intrinsic::assume>(Op0: m_Specific(V: IIOperand))))
3006 return RemoveConditionFromAssume(Next);
3007
3008 // Canonicalize assume(a && b) -> assume(a); assume(b);
3009 // Note: New assumption intrinsics created here are registered by
3010 // the InstCombineIRInserter object.
3011 FunctionType *AssumeIntrinsicTy = II->getFunctionType();
3012 Value *AssumeIntrinsic = II->getCalledOperand();
3013 Value *A, *B;
3014 if (match(V: IIOperand, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) {
3015 Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, Args: A, OpBundles,
3016 Name: II->getName());
3017 Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic, Args: B, Name: II->getName());
3018 return eraseInstFromFunction(I&: *II);
3019 }
3020 // assume(!(a || b)) -> assume(!a); assume(!b);
3021 if (match(V: IIOperand, P: m_Not(V: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))))) {
3022 Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic,
3023 Args: Builder.CreateNot(V: A), OpBundles, Name: II->getName());
3024 Builder.CreateCall(FTy: AssumeIntrinsicTy, Callee: AssumeIntrinsic,
3025 Args: Builder.CreateNot(V: B), Name: II->getName());
3026 return eraseInstFromFunction(I&: *II);
3027 }
3028
3029 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
3030 // (if assume is valid at the load)
3031 CmpInst::Predicate Pred;
3032 Instruction *LHS;
3033 if (match(V: IIOperand, P: m_ICmp(Pred, L: m_Instruction(I&: LHS), R: m_Zero())) &&
3034 Pred == ICmpInst::ICMP_NE && LHS->getOpcode() == Instruction::Load &&
3035 LHS->getType()->isPointerTy() &&
3036 isValidAssumeForContext(I: II, CxtI: LHS, DT: &DT)) {
3037 MDNode *MD = MDNode::get(Context&: II->getContext(), MDs: std::nullopt);
3038 LHS->setMetadata(KindID: LLVMContext::MD_nonnull, Node: MD);
3039 LHS->setMetadata(KindID: LLVMContext::MD_noundef, Node: MD);
3040 return RemoveConditionFromAssume(II);
3041
3042 // TODO: apply nonnull return attributes to calls and invokes
3043 // TODO: apply range metadata for range check patterns?
3044 }
3045
3046 // Separate storage assumptions apply to the underlying allocations, not any
3047 // particular pointer within them. When evaluating the hints for AA purposes
3048 // we getUnderlyingObject them; by precomputing the answers here we can
3049 // avoid having to do so repeatedly there.
3050 for (unsigned Idx = 0; Idx < II->getNumOperandBundles(); Idx++) {
3051 OperandBundleUse OBU = II->getOperandBundleAt(Index: Idx);
3052 if (OBU.getTagName() == "separate_storage") {
3053 assert(OBU.Inputs.size() == 2);
3054 auto MaybeSimplifyHint = [&](const Use &U) {
3055 Value *Hint = U.get();
3056 // Not having a limit is safe because InstCombine removes unreachable
3057 // code.
3058 Value *UnderlyingObject = getUnderlyingObject(V: Hint, /*MaxLookup*/ 0);
3059 if (Hint != UnderlyingObject)
3060 replaceUse(U&: const_cast<Use &>(U), NewValue: UnderlyingObject);
3061 };
3062 MaybeSimplifyHint(OBU.Inputs[0]);
3063 MaybeSimplifyHint(OBU.Inputs[1]);
3064 }
3065 }
3066
3067 // Convert nonnull assume like:
3068 // %A = icmp ne i32* %PTR, null
3069 // call void @llvm.assume(i1 %A)
3070 // into
3071 // call void @llvm.assume(i1 true) [ "nonnull"(i32* %PTR) ]
3072 if (EnableKnowledgeRetention &&
3073 match(V: IIOperand, P: m_Cmp(Pred, L: m_Value(V&: A), R: m_Zero())) &&
3074 Pred == CmpInst::ICMP_NE && A->getType()->isPointerTy()) {
3075 if (auto *Replacement = buildAssumeFromKnowledge(
3076 Knowledge: {RetainedKnowledge{.AttrKind: Attribute::NonNull, .ArgValue: 0, .WasOn: A}}, CtxI: Next, AC: &AC, DT: &DT)) {
3077
3078 Replacement->insertBefore(InsertPos: Next);
3079 AC.registerAssumption(CI: Replacement);
3080 return RemoveConditionFromAssume(II);
3081 }
3082 }
3083
3084 // Convert alignment assume like:
3085 // %B = ptrtoint i32* %A to i64
3086 // %C = and i64 %B, Constant
3087 // %D = icmp eq i64 %C, 0
3088 // call void @llvm.assume(i1 %D)
3089 // into
3090 // call void @llvm.assume(i1 true) [ "align"(i32* [[A]], i64 Constant + 1)]
3091 uint64_t AlignMask;
3092 if (EnableKnowledgeRetention &&
3093 match(V: IIOperand,
3094 P: m_Cmp(Pred, L: m_And(L: m_Value(V&: A), R: m_ConstantInt(V&: AlignMask)),
3095 R: m_Zero())) &&
3096 Pred == CmpInst::ICMP_EQ) {
3097 if (isPowerOf2_64(Value: AlignMask + 1)) {
3098 uint64_t Offset = 0;
3099 match(V: A, P: m_Add(L: m_Value(V&: A), R: m_ConstantInt(V&: Offset)));
3100 if (match(V: A, P: m_PtrToInt(Op: m_Value(V&: A)))) {
3101 /// Note: this doesn't preserve the offset information but merges
3102 /// offset and alignment.
3103 /// TODO: we can generate a GEP instead of merging the alignment with
3104 /// the offset.
3105 RetainedKnowledge RK{.AttrKind: Attribute::Alignment,
3106 .ArgValue: (unsigned)MinAlign(A: Offset, B: AlignMask + 1), .WasOn: A};
3107 if (auto *Replacement =
3108 buildAssumeFromKnowledge(Knowledge: RK, CtxI: Next, AC: &AC, DT: &DT)) {
3109
3110 Replacement->insertAfter(InsertPos: II);
3111 AC.registerAssumption(CI: Replacement);
3112 }
3113 return RemoveConditionFromAssume(II);
3114 }
3115 }
3116 }
3117
3118 /// Canonicalize Knowledge in operand bundles.
3119 if (EnableKnowledgeRetention && II->hasOperandBundles()) {
3120 for (unsigned Idx = 0; Idx < II->getNumOperandBundles(); Idx++) {
3121 auto &BOI = II->bundle_op_info_begin()[Idx];
3122 RetainedKnowledge RK =
3123 llvm::getKnowledgeFromBundle(Assume&: cast<AssumeInst>(Val&: *II), BOI);
3124 if (BOI.End - BOI.Begin > 2)
3125 continue; // Prevent reducing knowledge in an align with offset since
3126 // extracting a RetainedKnowledge from them looses offset
3127 // information
3128 RetainedKnowledge CanonRK =
3129 llvm::simplifyRetainedKnowledge(Assume: cast<AssumeInst>(Val: II), RK,
3130 AC: &getAssumptionCache(),
3131 DT: &getDominatorTree());
3132 if (CanonRK == RK)
3133 continue;
3134 if (!CanonRK) {
3135 if (BOI.End - BOI.Begin > 0) {
3136 Worklist.pushValue(V: II->op_begin()[BOI.Begin]);
3137 Value::dropDroppableUse(U&: II->op_begin()[BOI.Begin]);
3138 }
3139 continue;
3140 }
3141 assert(RK.AttrKind == CanonRK.AttrKind);
3142 if (BOI.End - BOI.Begin > 0)
3143 II->op_begin()[BOI.Begin].set(CanonRK.WasOn);
3144 if (BOI.End - BOI.Begin > 1)
3145 II->op_begin()[BOI.Begin + 1].set(ConstantInt::get(
3146 Ty: Type::getInt64Ty(C&: II->getContext()), V: CanonRK.ArgValue));
3147 if (RK.WasOn)
3148 Worklist.pushValue(V: RK.WasOn);
3149 return II;
3150 }
3151 }
3152
3153 // If there is a dominating assume with the same condition as this one,
3154 // then this one is redundant, and should be removed.
3155 KnownBits Known(1);
3156 computeKnownBits(V: IIOperand, Known, Depth: 0, CxtI: II);
3157 if (Known.isAllOnes() && isAssumeWithEmptyBundle(Assume: cast<AssumeInst>(Val&: *II)))
3158 return eraseInstFromFunction(I&: *II);
3159
3160 // assume(false) is unreachable.
3161 if (match(V: IIOperand, P: m_CombineOr(L: m_Zero(), R: m_Undef()))) {
3162 CreateNonTerminatorUnreachable(InsertAt: II);
3163 return eraseInstFromFunction(I&: *II);
3164 }
3165
3166 // Update the cache of affected values for this assumption (we might be
3167 // here because we just simplified the condition).
3168 AC.updateAffectedValues(CI: cast<AssumeInst>(Val: II));
3169 break;
3170 }
3171 case Intrinsic::experimental_guard: {
3172 // Is this guard followed by another guard? We scan forward over a small
3173 // fixed window of instructions to handle common cases with conditions
3174 // computed between guards.
3175 Instruction *NextInst = II->getNextNonDebugInstruction();
3176 for (unsigned i = 0; i < GuardWideningWindow; i++) {
3177 // Note: Using context-free form to avoid compile time blow up
3178 if (!isSafeToSpeculativelyExecute(I: NextInst))
3179 break;
3180 NextInst = NextInst->getNextNonDebugInstruction();
3181 }
3182 Value *NextCond = nullptr;
3183 if (match(V: NextInst,
3184 P: m_Intrinsic<Intrinsic::experimental_guard>(Op0: m_Value(V&: NextCond)))) {
3185 Value *CurrCond = II->getArgOperand(i: 0);
3186
3187 // Remove a guard that it is immediately preceded by an identical guard.
3188 // Otherwise canonicalize guard(a); guard(b) -> guard(a & b).
3189 if (CurrCond != NextCond) {
3190 Instruction *MoveI = II->getNextNonDebugInstruction();
3191 while (MoveI != NextInst) {
3192 auto *Temp = MoveI;
3193 MoveI = MoveI->getNextNonDebugInstruction();
3194 Temp->moveBefore(MovePos: II);
3195 }
3196 replaceOperand(I&: *II, OpNum: 0, V: Builder.CreateAnd(LHS: CurrCond, RHS: NextCond));
3197 }
3198 eraseInstFromFunction(I&: *NextInst);
3199 return II;
3200 }
3201 break;
3202 }
3203 case Intrinsic::vector_insert: {
3204 Value *Vec = II->getArgOperand(i: 0);
3205 Value *SubVec = II->getArgOperand(i: 1);
3206 Value *Idx = II->getArgOperand(i: 2);
3207 auto *DstTy = dyn_cast<FixedVectorType>(Val: II->getType());
3208 auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType());
3209 auto *SubVecTy = dyn_cast<FixedVectorType>(Val: SubVec->getType());
3210
3211 // Only canonicalize if the destination vector, Vec, and SubVec are all
3212 // fixed vectors.
3213 if (DstTy && VecTy && SubVecTy) {
3214 unsigned DstNumElts = DstTy->getNumElements();
3215 unsigned VecNumElts = VecTy->getNumElements();
3216 unsigned SubVecNumElts = SubVecTy->getNumElements();
3217 unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue();
3218
3219 // An insert that entirely overwrites Vec with SubVec is a nop.
3220 if (VecNumElts == SubVecNumElts)
3221 return replaceInstUsesWith(I&: CI, V: SubVec);
3222
3223 // Widen SubVec into a vector of the same width as Vec, since
3224 // shufflevector requires the two input vectors to be the same width.
3225 // Elements beyond the bounds of SubVec within the widened vector are
3226 // undefined.
3227 SmallVector<int, 8> WidenMask;
3228 unsigned i;
3229 for (i = 0; i != SubVecNumElts; ++i)
3230 WidenMask.push_back(Elt: i);
3231 for (; i != VecNumElts; ++i)
3232 WidenMask.push_back(Elt: PoisonMaskElem);
3233
3234 Value *WidenShuffle = Builder.CreateShuffleVector(V: SubVec, Mask: WidenMask);
3235
3236 SmallVector<int, 8> Mask;
3237 for (unsigned i = 0; i != IdxN; ++i)
3238 Mask.push_back(Elt: i);
3239 for (unsigned i = DstNumElts; i != DstNumElts + SubVecNumElts; ++i)
3240 Mask.push_back(Elt: i);
3241 for (unsigned i = IdxN + SubVecNumElts; i != DstNumElts; ++i)
3242 Mask.push_back(Elt: i);
3243
3244 Value *Shuffle = Builder.CreateShuffleVector(V1: Vec, V2: WidenShuffle, Mask);
3245 return replaceInstUsesWith(I&: CI, V: Shuffle);
3246 }
3247 break;
3248 }
3249 case Intrinsic::vector_extract: {
3250 Value *Vec = II->getArgOperand(i: 0);
3251 Value *Idx = II->getArgOperand(i: 1);
3252
3253 Type *ReturnType = II->getType();
3254 // (extract_vector (insert_vector InsertTuple, InsertValue, InsertIdx),
3255 // ExtractIdx)
3256 unsigned ExtractIdx = cast<ConstantInt>(Val: Idx)->getZExtValue();
3257 Value *InsertTuple, *InsertIdx, *InsertValue;
3258 if (match(V: Vec, P: m_Intrinsic<Intrinsic::vector_insert>(Op0: m_Value(V&: InsertTuple),
3259 Op1: m_Value(V&: InsertValue),
3260 Op2: m_Value(V&: InsertIdx))) &&
3261 InsertValue->getType() == ReturnType) {
3262 unsigned Index = cast<ConstantInt>(Val: InsertIdx)->getZExtValue();
3263 // Case where we get the same index right after setting it.
3264 // extract.vector(insert.vector(InsertTuple, InsertValue, Idx), Idx) -->
3265 // InsertValue
3266 if (ExtractIdx == Index)
3267 return replaceInstUsesWith(I&: CI, V: InsertValue);
3268 // If we are getting a different index than what was set in the
3269 // insert.vector intrinsic. We can just set the input tuple to the one up
3270 // in the chain. extract.vector(insert.vector(InsertTuple, InsertValue,
3271 // InsertIndex), ExtractIndex)
3272 // --> extract.vector(InsertTuple, ExtractIndex)
3273 else
3274 return replaceOperand(I&: CI, OpNum: 0, V: InsertTuple);
3275 }
3276
3277 auto *DstTy = dyn_cast<VectorType>(Val: ReturnType);
3278 auto *VecTy = dyn_cast<VectorType>(Val: Vec->getType());
3279
3280 if (DstTy && VecTy) {
3281 auto DstEltCnt = DstTy->getElementCount();
3282 auto VecEltCnt = VecTy->getElementCount();
3283 unsigned IdxN = cast<ConstantInt>(Val: Idx)->getZExtValue();
3284
3285 // Extracting the entirety of Vec is a nop.
3286 if (DstEltCnt == VecTy->getElementCount()) {
3287 replaceInstUsesWith(I&: CI, V: Vec);
3288 return eraseInstFromFunction(I&: CI);
3289 }
3290
3291 // Only canonicalize to shufflevector if the destination vector and
3292 // Vec are fixed vectors.
3293 if (VecEltCnt.isScalable() || DstEltCnt.isScalable())
3294 break;
3295
3296 SmallVector<int, 8> Mask;
3297 for (unsigned i = 0; i != DstEltCnt.getKnownMinValue(); ++i)
3298 Mask.push_back(Elt: IdxN + i);
3299
3300 Value *Shuffle = Builder.CreateShuffleVector(V: Vec, Mask);
3301 return replaceInstUsesWith(I&: CI, V: Shuffle);
3302 }
3303 break;
3304 }
3305 case Intrinsic::vector_reverse: {
3306 Value *BO0, *BO1, *X, *Y;
3307 Value *Vec = II->getArgOperand(i: 0);
3308 if (match(V: Vec, P: m_OneUse(SubPattern: m_BinOp(L: m_Value(V&: BO0), R: m_Value(V&: BO1))))) {
3309 auto *OldBinOp = cast<BinaryOperator>(Val: Vec);
3310 if (match(V: BO0, P: m_VecReverse(Op0: m_Value(V&: X)))) {
3311 // rev(binop rev(X), rev(Y)) --> binop X, Y
3312 if (match(V: BO1, P: m_VecReverse(Op0: m_Value(V&: Y))))
3313 return replaceInstUsesWith(I&: CI, V: BinaryOperator::CreateWithCopiedFlags(
3314 Opc: OldBinOp->getOpcode(), V1: X, V2: Y,
3315 CopyO: OldBinOp, Name: OldBinOp->getName(),
3316 InsertBefore: II->getIterator()));
3317 // rev(binop rev(X), BO1Splat) --> binop X, BO1Splat
3318 if (isSplatValue(V: BO1))
3319 return replaceInstUsesWith(I&: CI, V: BinaryOperator::CreateWithCopiedFlags(
3320 Opc: OldBinOp->getOpcode(), V1: X, V2: BO1,
3321 CopyO: OldBinOp, Name: OldBinOp->getName(),
3322 InsertBefore: II->getIterator()));
3323 }
3324 // rev(binop BO0Splat, rev(Y)) --> binop BO0Splat, Y
3325 if (match(V: BO1, P: m_VecReverse(Op0: m_Value(V&: Y))) && isSplatValue(V: BO0))
3326 return replaceInstUsesWith(I&: CI,
3327 V: BinaryOperator::CreateWithCopiedFlags(
3328 Opc: OldBinOp->getOpcode(), V1: BO0, V2: Y, CopyO: OldBinOp,
3329 Name: OldBinOp->getName(), InsertBefore: II->getIterator()));
3330 }
3331 // rev(unop rev(X)) --> unop X
3332 if (match(V: Vec, P: m_OneUse(SubPattern: m_UnOp(X: m_VecReverse(Op0: m_Value(V&: X)))))) {
3333 auto *OldUnOp = cast<UnaryOperator>(Val: Vec);
3334 auto *NewUnOp = UnaryOperator::CreateWithCopiedFlags(
3335 Opc: OldUnOp->getOpcode(), V: X, CopyO: OldUnOp, Name: OldUnOp->getName(),
3336 InsertBefore: II->getIterator());
3337 return replaceInstUsesWith(I&: CI, V: NewUnOp);
3338 }
3339 break;
3340 }
3341 case Intrinsic::vector_reduce_or:
3342 case Intrinsic::vector_reduce_and: {
3343 // Canonicalize logical or/and reductions:
3344 // Or reduction for i1 is represented as:
3345 // %val = bitcast <ReduxWidth x i1> to iReduxWidth
3346 // %res = cmp ne iReduxWidth %val, 0
3347 // And reduction for i1 is represented as:
3348 // %val = bitcast <ReduxWidth x i1> to iReduxWidth
3349 // %res = cmp eq iReduxWidth %val, 11111
3350 Value *Arg = II->getArgOperand(i: 0);
3351 Value *Vect;
3352
3353 if (Value *NewOp =
3354 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3355 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3356 return II;
3357 }
3358
3359 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3360 if (auto *FTy = dyn_cast<FixedVectorType>(Val: Vect->getType()))
3361 if (FTy->getElementType() == Builder.getInt1Ty()) {
3362 Value *Res = Builder.CreateBitCast(
3363 V: Vect, DestTy: Builder.getIntNTy(N: FTy->getNumElements()));
3364 if (IID == Intrinsic::vector_reduce_and) {
3365 Res = Builder.CreateICmpEQ(
3366 LHS: Res, RHS: ConstantInt::getAllOnesValue(Ty: Res->getType()));
3367 } else {
3368 assert(IID == Intrinsic::vector_reduce_or &&
3369 "Expected or reduction.");
3370 Res = Builder.CreateIsNotNull(Arg: Res);
3371 }
3372 if (Arg != Vect)
3373 Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res,
3374 DestTy: II->getType());
3375 return replaceInstUsesWith(I&: CI, V: Res);
3376 }
3377 }
3378 [[fallthrough]];
3379 }
3380 case Intrinsic::vector_reduce_add: {
3381 if (IID == Intrinsic::vector_reduce_add) {
3382 // Convert vector_reduce_add(ZExt(<n x i1>)) to
3383 // ZExtOrTrunc(ctpop(bitcast <n x i1> to in)).
3384 // Convert vector_reduce_add(SExt(<n x i1>)) to
3385 // -ZExtOrTrunc(ctpop(bitcast <n x i1> to in)).
3386 // Convert vector_reduce_add(<n x i1>) to
3387 // Trunc(ctpop(bitcast <n x i1> to in)).
3388 Value *Arg = II->getArgOperand(i: 0);
3389 Value *Vect;
3390
3391 if (Value *NewOp =
3392 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3393 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3394 return II;
3395 }
3396
3397 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3398 if (auto *FTy = dyn_cast<FixedVectorType>(Val: Vect->getType()))
3399 if (FTy->getElementType() == Builder.getInt1Ty()) {
3400 Value *V = Builder.CreateBitCast(
3401 V: Vect, DestTy: Builder.getIntNTy(N: FTy->getNumElements()));
3402 Value *Res = Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V);
3403 if (Res->getType() != II->getType())
3404 Res = Builder.CreateZExtOrTrunc(V: Res, DestTy: II->getType());
3405 if (Arg != Vect &&
3406 cast<Instruction>(Val: Arg)->getOpcode() == Instruction::SExt)
3407 Res = Builder.CreateNeg(V: Res);
3408 return replaceInstUsesWith(I&: CI, V: Res);
3409 }
3410 }
3411 }
3412 [[fallthrough]];
3413 }
3414 case Intrinsic::vector_reduce_xor: {
3415 if (IID == Intrinsic::vector_reduce_xor) {
3416 // Exclusive disjunction reduction over the vector with
3417 // (potentially-extended) i1 element type is actually a
3418 // (potentially-extended) arithmetic `add` reduction over the original
3419 // non-extended value:
3420 // vector_reduce_xor(?ext(<n x i1>))
3421 // -->
3422 // ?ext(vector_reduce_add(<n x i1>))
3423 Value *Arg = II->getArgOperand(i: 0);
3424 Value *Vect;
3425
3426 if (Value *NewOp =
3427 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3428 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3429 return II;
3430 }
3431
3432 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3433 if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType()))
3434 if (VTy->getElementType() == Builder.getInt1Ty()) {
3435 Value *Res = Builder.CreateAddReduce(Src: Vect);
3436 if (Arg != Vect)
3437 Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res,
3438 DestTy: II->getType());
3439 return replaceInstUsesWith(I&: CI, V: Res);
3440 }
3441 }
3442 }
3443 [[fallthrough]];
3444 }
3445 case Intrinsic::vector_reduce_mul: {
3446 if (IID == Intrinsic::vector_reduce_mul) {
3447 // Multiplicative reduction over the vector with (potentially-extended)
3448 // i1 element type is actually a (potentially zero-extended)
3449 // logical `and` reduction over the original non-extended value:
3450 // vector_reduce_mul(?ext(<n x i1>))
3451 // -->
3452 // zext(vector_reduce_and(<n x i1>))
3453 Value *Arg = II->getArgOperand(i: 0);
3454 Value *Vect;
3455
3456 if (Value *NewOp =
3457 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3458 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3459 return II;
3460 }
3461
3462 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3463 if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType()))
3464 if (VTy->getElementType() == Builder.getInt1Ty()) {
3465 Value *Res = Builder.CreateAndReduce(Src: Vect);
3466 if (Res->getType() != II->getType())
3467 Res = Builder.CreateZExt(V: Res, DestTy: II->getType());
3468 return replaceInstUsesWith(I&: CI, V: Res);
3469 }
3470 }
3471 }
3472 [[fallthrough]];
3473 }
3474 case Intrinsic::vector_reduce_umin:
3475 case Intrinsic::vector_reduce_umax: {
3476 if (IID == Intrinsic::vector_reduce_umin ||
3477 IID == Intrinsic::vector_reduce_umax) {
3478 // UMin/UMax reduction over the vector with (potentially-extended)
3479 // i1 element type is actually a (potentially-extended)
3480 // logical `and`/`or` reduction over the original non-extended value:
3481 // vector_reduce_u{min,max}(?ext(<n x i1>))
3482 // -->
3483 // ?ext(vector_reduce_{and,or}(<n x i1>))
3484 Value *Arg = II->getArgOperand(i: 0);
3485 Value *Vect;
3486
3487 if (Value *NewOp =
3488 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3489 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3490 return II;
3491 }
3492
3493 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3494 if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType()))
3495 if (VTy->getElementType() == Builder.getInt1Ty()) {
3496 Value *Res = IID == Intrinsic::vector_reduce_umin
3497 ? Builder.CreateAndReduce(Src: Vect)
3498 : Builder.CreateOrReduce(Src: Vect);
3499 if (Arg != Vect)
3500 Res = Builder.CreateCast(Op: cast<CastInst>(Val: Arg)->getOpcode(), V: Res,
3501 DestTy: II->getType());
3502 return replaceInstUsesWith(I&: CI, V: Res);
3503 }
3504 }
3505 }
3506 [[fallthrough]];
3507 }
3508 case Intrinsic::vector_reduce_smin:
3509 case Intrinsic::vector_reduce_smax: {
3510 if (IID == Intrinsic::vector_reduce_smin ||
3511 IID == Intrinsic::vector_reduce_smax) {
3512 // SMin/SMax reduction over the vector with (potentially-extended)
3513 // i1 element type is actually a (potentially-extended)
3514 // logical `and`/`or` reduction over the original non-extended value:
3515 // vector_reduce_s{min,max}(<n x i1>)
3516 // -->
3517 // vector_reduce_{or,and}(<n x i1>)
3518 // and
3519 // vector_reduce_s{min,max}(sext(<n x i1>))
3520 // -->
3521 // sext(vector_reduce_{or,and}(<n x i1>))
3522 // and
3523 // vector_reduce_s{min,max}(zext(<n x i1>))
3524 // -->
3525 // zext(vector_reduce_{and,or}(<n x i1>))
3526 Value *Arg = II->getArgOperand(i: 0);
3527 Value *Vect;
3528
3529 if (Value *NewOp =
3530 simplifyReductionOperand(Arg, /*CanReorderLanes=*/true)) {
3531 replaceUse(U&: II->getOperandUse(i: 0), NewValue: NewOp);
3532 return II;
3533 }
3534
3535 if (match(V: Arg, P: m_ZExtOrSExtOrSelf(Op: m_Value(V&: Vect)))) {
3536 if (auto *VTy = dyn_cast<VectorType>(Val: Vect->getType()))
3537 if (VTy->getElementType() == Builder.getInt1Ty()) {
3538 Instruction::CastOps ExtOpc = Instruction::CastOps::CastOpsEnd;
3539 if (Arg != Vect)
3540 ExtOpc = cast<CastInst>(Val: Arg)->getOpcode();
3541 Value *Res = ((IID == Intrinsic::vector_reduce_smin) ==
3542 (ExtOpc == Instruction::CastOps::ZExt))
3543 ? Builder.CreateAndReduce(Src: Vect)
3544 : Builder.CreateOrReduce(Src: Vect);
3545 if (Arg != Vect)
3546 Res = Builder.CreateCast(Op: ExtOpc, V: Res, DestTy: II->getType());
3547 return replaceInstUsesWith(I&: CI, V: Res);
3548 }
3549 }
3550 }
3551 [[fallthrough]];
3552 }
3553 case Intrinsic::vector_reduce_fmax:
3554 case Intrinsic::vector_reduce_fmin:
3555 case Intrinsic::vector_reduce_fadd:
3556 case Intrinsic::vector_reduce_fmul: {
3557 bool CanReorderLanes = (IID != Intrinsic::vector_reduce_fadd &&
3558 IID != Intrinsic::vector_reduce_fmul) ||
3559 II->hasAllowReassoc();
3560 const unsigned ArgIdx = (IID == Intrinsic::vector_reduce_fadd ||
3561 IID == Intrinsic::vector_reduce_fmul)
3562 ? 1
3563 : 0;
3564 Value *Arg = II->getArgOperand(i: ArgIdx);
3565 if (Value *NewOp = simplifyReductionOperand(Arg, CanReorderLanes)) {
3566 replaceUse(U&: II->getOperandUse(i: ArgIdx), NewValue: NewOp);
3567 return nullptr;
3568 }
3569 break;
3570 }
3571 case Intrinsic::is_fpclass: {
3572 if (Instruction *I = foldIntrinsicIsFPClass(II&: *II))
3573 return I;
3574 break;
3575 }
3576 case Intrinsic::threadlocal_address: {
3577 Align MinAlign = getKnownAlignment(V: II->getArgOperand(i: 0), DL, CxtI: II, AC: &AC, DT: &DT);
3578 MaybeAlign Align = II->getRetAlign();
3579 if (MinAlign > Align.valueOrOne()) {
3580 II->addRetAttr(Attr: Attribute::getWithAlignment(Context&: II->getContext(), Alignment: MinAlign));
3581 return II;
3582 }
3583 break;
3584 }
3585 default: {
3586 // Handle target specific intrinsics
3587 std::optional<Instruction *> V = targetInstCombineIntrinsic(II&: *II);
3588 if (V)
3589 return *V;
3590 break;
3591 }
3592 }
3593
3594 // Try to fold intrinsic into select operands. This is legal if:
3595 // * The intrinsic is speculatable.
3596 // * The select condition is not a vector, or the intrinsic does not
3597 // perform cross-lane operations.
3598 switch (IID) {
3599 case Intrinsic::ctlz:
3600 case Intrinsic::cttz:
3601 case Intrinsic::ctpop:
3602 case Intrinsic::umin:
3603 case Intrinsic::umax:
3604 case Intrinsic::smin:
3605 case Intrinsic::smax:
3606 case Intrinsic::usub_sat:
3607 case Intrinsic::uadd_sat:
3608 case Intrinsic::ssub_sat:
3609 case Intrinsic::sadd_sat:
3610 for (Value *Op : II->args())
3611 if (auto *Sel = dyn_cast<SelectInst>(Val: Op))
3612 if (Instruction *R = FoldOpIntoSelect(Op&: *II, SI: Sel))
3613 return R;
3614 [[fallthrough]];
3615 default:
3616 break;
3617 }
3618
3619 if (Instruction *Shuf = foldShuffledIntrinsicOperands(II, Builder))
3620 return Shuf;
3621
3622 // Some intrinsics (like experimental_gc_statepoint) can be used in invoke
3623 // context, so it is handled in visitCallBase and we should trigger it.
3624 return visitCallBase(Call&: *II);
3625}
3626
3627// Fence instruction simplification
3628Instruction *InstCombinerImpl::visitFenceInst(FenceInst &FI) {
3629 auto *NFI = dyn_cast<FenceInst>(Val: FI.getNextNonDebugInstruction());
3630 // This check is solely here to handle arbitrary target-dependent syncscopes.
3631 // TODO: Can remove if does not matter in practice.
3632 if (NFI && FI.isIdenticalTo(I: NFI))
3633 return eraseInstFromFunction(I&: FI);
3634
3635 // Returns true if FI1 is identical or stronger fence than FI2.
3636 auto isIdenticalOrStrongerFence = [](FenceInst *FI1, FenceInst *FI2) {
3637 auto FI1SyncScope = FI1->getSyncScopeID();
3638 // Consider same scope, where scope is global or single-thread.
3639 if (FI1SyncScope != FI2->getSyncScopeID() ||
3640 (FI1SyncScope != SyncScope::System &&
3641 FI1SyncScope != SyncScope::SingleThread))
3642 return false;
3643
3644 return isAtLeastOrStrongerThan(AO: FI1->getOrdering(), Other: FI2->getOrdering());
3645 };
3646 if (NFI && isIdenticalOrStrongerFence(NFI, &FI))
3647 return eraseInstFromFunction(I&: FI);
3648
3649 if (auto *PFI = dyn_cast_or_null<FenceInst>(Val: FI.getPrevNonDebugInstruction()))
3650 if (isIdenticalOrStrongerFence(PFI, &FI))
3651 return eraseInstFromFunction(I&: FI);
3652 return nullptr;
3653}
3654
3655// InvokeInst simplification
3656Instruction *InstCombinerImpl::visitInvokeInst(InvokeInst &II) {
3657 return visitCallBase(Call&: II);
3658}
3659
3660// CallBrInst simplification
3661Instruction *InstCombinerImpl::visitCallBrInst(CallBrInst &CBI) {
3662 return visitCallBase(Call&: CBI);
3663}
3664
3665Instruction *InstCombinerImpl::tryOptimizeCall(CallInst *CI) {
3666 if (!CI->getCalledFunction()) return nullptr;
3667
3668 // Skip optimizing notail and musttail calls so
3669 // LibCallSimplifier::optimizeCall doesn't have to preserve those invariants.
3670 // LibCallSimplifier::optimizeCall should try to preseve tail calls though.
3671 if (CI->isMustTailCall() || CI->isNoTailCall())
3672 return nullptr;
3673
3674 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
3675 replaceInstUsesWith(I&: *From, V: With);
3676 };
3677 auto InstCombineErase = [this](Instruction *I) {
3678 eraseInstFromFunction(I&: *I);
3679 };
3680 LibCallSimplifier Simplifier(DL, &TLI, &AC, ORE, BFI, PSI, InstCombineRAUW,
3681 InstCombineErase);
3682 if (Value *With = Simplifier.optimizeCall(CI, B&: Builder)) {
3683 ++NumSimplified;
3684 return CI->use_empty() ? CI : replaceInstUsesWith(I&: *CI, V: With);
3685 }
3686
3687 return nullptr;
3688}
3689
3690static IntrinsicInst *findInitTrampolineFromAlloca(Value *TrampMem) {
3691 // Strip off at most one level of pointer casts, looking for an alloca. This
3692 // is good enough in practice and simpler than handling any number of casts.
3693 Value *Underlying = TrampMem->stripPointerCasts();
3694 if (Underlying != TrampMem &&
3695 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
3696 return nullptr;
3697 if (!isa<AllocaInst>(Val: Underlying))
3698 return nullptr;
3699
3700 IntrinsicInst *InitTrampoline = nullptr;
3701 for (User *U : TrampMem->users()) {
3702 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U);
3703 if (!II)
3704 return nullptr;
3705 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
3706 if (InitTrampoline)
3707 // More than one init_trampoline writes to this value. Give up.
3708 return nullptr;
3709 InitTrampoline = II;
3710 continue;
3711 }
3712 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
3713 // Allow any number of calls to adjust.trampoline.
3714 continue;
3715 return nullptr;
3716 }
3717
3718 // No call to init.trampoline found.
3719 if (!InitTrampoline)
3720 return nullptr;
3721
3722 // Check that the alloca is being used in the expected way.
3723 if (InitTrampoline->getOperand(i_nocapture: 0) != TrampMem)
3724 return nullptr;
3725
3726 return InitTrampoline;
3727}
3728
3729static IntrinsicInst *findInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
3730 Value *TrampMem) {
3731 // Visit all the previous instructions in the basic block, and try to find a
3732 // init.trampoline which has a direct path to the adjust.trampoline.
3733 for (BasicBlock::iterator I = AdjustTramp->getIterator(),
3734 E = AdjustTramp->getParent()->begin();
3735 I != E;) {
3736 Instruction *Inst = &*--I;
3737 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val&: I))
3738 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
3739 II->getOperand(i_nocapture: 0) == TrampMem)
3740 return II;
3741 if (Inst->mayWriteToMemory())
3742 return nullptr;
3743 }
3744 return nullptr;
3745}
3746
3747// Given a call to llvm.adjust.trampoline, find and return the corresponding
3748// call to llvm.init.trampoline if the call to the trampoline can be optimized
3749// to a direct call to a function. Otherwise return NULL.
3750static IntrinsicInst *findInitTrampoline(Value *Callee) {
3751 Callee = Callee->stripPointerCasts();
3752 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Val: Callee);
3753 if (!AdjustTramp ||
3754 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
3755 return nullptr;
3756
3757 Value *TrampMem = AdjustTramp->getOperand(i_nocapture: 0);
3758
3759 if (IntrinsicInst *IT = findInitTrampolineFromAlloca(TrampMem))
3760 return IT;
3761 if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem))
3762 return IT;
3763 return nullptr;
3764}
3765
3766bool InstCombinerImpl::annotateAnyAllocSite(CallBase &Call,
3767 const TargetLibraryInfo *TLI) {
3768 // Note: We only handle cases which can't be driven from generic attributes
3769 // here. So, for example, nonnull and noalias (which are common properties
3770 // of some allocation functions) are expected to be handled via annotation
3771 // of the respective allocator declaration with generic attributes.
3772 bool Changed = false;
3773
3774 if (!Call.getType()->isPointerTy())
3775 return Changed;
3776
3777 std::optional<APInt> Size = getAllocSize(CB: &Call, TLI);
3778 if (Size && *Size != 0) {
3779 // TODO: We really should just emit deref_or_null here and then
3780 // let the generic inference code combine that with nonnull.
3781 if (Call.hasRetAttr(Kind: Attribute::NonNull)) {
3782 Changed = !Call.hasRetAttr(Kind: Attribute::Dereferenceable);
3783 Call.addRetAttr(Attr: Attribute::getWithDereferenceableBytes(
3784 Context&: Call.getContext(), Bytes: Size->getLimitedValue()));
3785 } else {
3786 Changed = !Call.hasRetAttr(Kind: Attribute::DereferenceableOrNull);
3787 Call.addRetAttr(Attr: Attribute::getWithDereferenceableOrNullBytes(
3788 Context&: Call.getContext(), Bytes: Size->getLimitedValue()));
3789 }
3790 }
3791
3792 // Add alignment attribute if alignment is a power of two constant.
3793 Value *Alignment = getAllocAlignment(V: &Call, TLI);
3794 if (!Alignment)
3795 return Changed;
3796
3797 ConstantInt *AlignOpC = dyn_cast<ConstantInt>(Val: Alignment);
3798 if (AlignOpC && AlignOpC->getValue().ult(RHS: llvm::Value::MaximumAlignment)) {
3799 uint64_t AlignmentVal = AlignOpC->getZExtValue();
3800 if (llvm::isPowerOf2_64(Value: AlignmentVal)) {
3801 Align ExistingAlign = Call.getRetAlign().valueOrOne();
3802 Align NewAlign = Align(AlignmentVal);
3803 if (NewAlign > ExistingAlign) {
3804 Call.addRetAttr(
3805 Attr: Attribute::getWithAlignment(Context&: Call.getContext(), Alignment: NewAlign));
3806 Changed = true;
3807 }
3808 }
3809 }
3810 return Changed;
3811}
3812
3813/// Improvements for call, callbr and invoke instructions.
3814Instruction *InstCombinerImpl::visitCallBase(CallBase &Call) {
3815 bool Changed = annotateAnyAllocSite(Call, TLI: &TLI);
3816
3817 // Mark any parameters that are known to be non-null with the nonnull
3818 // attribute. This is helpful for inlining calls to functions with null
3819 // checks on their arguments.
3820 SmallVector<unsigned, 4> ArgNos;
3821 unsigned ArgNo = 0;
3822
3823 for (Value *V : Call.args()) {
3824 if (V->getType()->isPointerTy() &&
3825 !Call.paramHasAttr(ArgNo, Kind: Attribute::NonNull) &&
3826 isKnownNonZero(V, Q: getSimplifyQuery().getWithInstruction(I: &Call)))
3827 ArgNos.push_back(Elt: ArgNo);
3828 ArgNo++;
3829 }
3830
3831 assert(ArgNo == Call.arg_size() && "Call arguments not processed correctly.");
3832
3833 if (!ArgNos.empty()) {
3834 AttributeList AS = Call.getAttributes();
3835 LLVMContext &Ctx = Call.getContext();
3836 AS = AS.addParamAttribute(C&: Ctx, ArgNos,
3837 A: Attribute::get(Context&: Ctx, Kind: Attribute::NonNull));
3838 Call.setAttributes(AS);
3839 Changed = true;
3840 }
3841
3842 // If the callee is a pointer to a function, attempt to move any casts to the
3843 // arguments of the call/callbr/invoke.
3844 Value *Callee = Call.getCalledOperand();
3845 Function *CalleeF = dyn_cast<Function>(Val: Callee);
3846 if ((!CalleeF || CalleeF->getFunctionType() != Call.getFunctionType()) &&
3847 transformConstExprCastCall(Call))
3848 return nullptr;
3849
3850 if (CalleeF) {
3851 // Remove the convergent attr on calls when the callee is not convergent.
3852 if (Call.isConvergent() && !CalleeF->isConvergent() &&
3853 !CalleeF->isIntrinsic()) {
3854 LLVM_DEBUG(dbgs() << "Removing convergent attr from instr " << Call
3855 << "\n");
3856 Call.setNotConvergent();
3857 return &Call;
3858 }
3859
3860 // If the call and callee calling conventions don't match, and neither one
3861 // of the calling conventions is compatible with C calling convention
3862 // this call must be unreachable, as the call is undefined.
3863 if ((CalleeF->getCallingConv() != Call.getCallingConv() &&
3864 !(CalleeF->getCallingConv() == llvm::CallingConv::C &&
3865 TargetLibraryInfoImpl::isCallingConvCCompatible(CI: &Call)) &&
3866 !(Call.getCallingConv() == llvm::CallingConv::C &&
3867 TargetLibraryInfoImpl::isCallingConvCCompatible(Callee: CalleeF))) &&
3868 // Only do this for calls to a function with a body. A prototype may
3869 // not actually end up matching the implementation's calling conv for a
3870 // variety of reasons (e.g. it may be written in assembly).
3871 !CalleeF->isDeclaration()) {
3872 Instruction *OldCall = &Call;
3873 CreateNonTerminatorUnreachable(InsertAt: OldCall);
3874 // If OldCall does not return void then replaceInstUsesWith poison.
3875 // This allows ValueHandlers and custom metadata to adjust itself.
3876 if (!OldCall->getType()->isVoidTy())
3877 replaceInstUsesWith(I&: *OldCall, V: PoisonValue::get(T: OldCall->getType()));
3878 if (isa<CallInst>(Val: OldCall))
3879 return eraseInstFromFunction(I&: *OldCall);
3880
3881 // We cannot remove an invoke or a callbr, because it would change thexi
3882 // CFG, just change the callee to a null pointer.
3883 cast<CallBase>(Val: OldCall)->setCalledFunction(
3884 FTy: CalleeF->getFunctionType(),
3885 Fn: Constant::getNullValue(Ty: CalleeF->getType()));
3886 return nullptr;
3887 }
3888 }
3889
3890 // Calling a null function pointer is undefined if a null address isn't
3891 // dereferenceable.
3892 if ((isa<ConstantPointerNull>(Val: Callee) &&
3893 !NullPointerIsDefined(F: Call.getFunction())) ||
3894 isa<UndefValue>(Val: Callee)) {
3895 // If Call does not return void then replaceInstUsesWith poison.
3896 // This allows ValueHandlers and custom metadata to adjust itself.
3897 if (!Call.getType()->isVoidTy())
3898 replaceInstUsesWith(I&: Call, V: PoisonValue::get(T: Call.getType()));
3899
3900 if (Call.isTerminator()) {
3901 // Can't remove an invoke or callbr because we cannot change the CFG.
3902 return nullptr;
3903 }
3904
3905 // This instruction is not reachable, just remove it.
3906 CreateNonTerminatorUnreachable(InsertAt: &Call);
3907 return eraseInstFromFunction(I&: Call);
3908 }
3909
3910 if (IntrinsicInst *II = findInitTrampoline(Callee))
3911 return transformCallThroughTrampoline(Call, Tramp&: *II);
3912
3913 if (isa<InlineAsm>(Val: Callee) && !Call.doesNotThrow()) {
3914 InlineAsm *IA = cast<InlineAsm>(Val: Callee);
3915 if (!IA->canThrow()) {
3916 // Normal inline asm calls cannot throw - mark them
3917 // 'nounwind'.
3918 Call.setDoesNotThrow();
3919 Changed = true;
3920 }
3921 }
3922
3923 // Try to optimize the call if possible, we require DataLayout for most of
3924 // this. None of these calls are seen as possibly dead so go ahead and
3925 // delete the instruction now.
3926 if (CallInst *CI = dyn_cast<CallInst>(Val: &Call)) {
3927 Instruction *I = tryOptimizeCall(CI);
3928 // If we changed something return the result, etc. Otherwise let
3929 // the fallthrough check.
3930 if (I) return eraseInstFromFunction(I&: *I);
3931 }
3932
3933 if (!Call.use_empty() && !Call.isMustTailCall())
3934 if (Value *ReturnedArg = Call.getReturnedArgOperand()) {
3935 Type *CallTy = Call.getType();
3936 Type *RetArgTy = ReturnedArg->getType();
3937 if (RetArgTy->canLosslesslyBitCastTo(Ty: CallTy))
3938 return replaceInstUsesWith(
3939 I&: Call, V: Builder.CreateBitOrPointerCast(V: ReturnedArg, DestTy: CallTy));
3940 }
3941
3942 // Drop unnecessary kcfi operand bundles from calls that were converted
3943 // into direct calls.
3944 auto Bundle = Call.getOperandBundle(ID: LLVMContext::OB_kcfi);
3945 if (Bundle && !Call.isIndirectCall()) {
3946 DEBUG_WITH_TYPE(DEBUG_TYPE "-kcfi", {
3947 if (CalleeF) {
3948 ConstantInt *FunctionType = nullptr;
3949 ConstantInt *ExpectedType = cast<ConstantInt>(Bundle->Inputs[0]);
3950
3951 if (MDNode *MD = CalleeF->getMetadata(LLVMContext::MD_kcfi_type))
3952 FunctionType = mdconst::extract<ConstantInt>(MD->getOperand(0));
3953
3954 if (FunctionType &&
3955 FunctionType->getZExtValue() != ExpectedType->getZExtValue())
3956 dbgs() << Call.getModule()->getName()
3957 << ": warning: kcfi: " << Call.getCaller()->getName()
3958 << ": call to " << CalleeF->getName()
3959 << " using a mismatching function pointer type\n";
3960 }
3961 });
3962
3963 return CallBase::removeOperandBundle(CB: &Call, ID: LLVMContext::OB_kcfi);
3964 }
3965
3966 if (isRemovableAlloc(V: &Call, TLI: &TLI))
3967 return visitAllocSite(FI&: Call);
3968
3969 // Handle intrinsics which can be used in both call and invoke context.
3970 switch (Call.getIntrinsicID()) {
3971 case Intrinsic::experimental_gc_statepoint: {
3972 GCStatepointInst &GCSP = *cast<GCStatepointInst>(Val: &Call);
3973 SmallPtrSet<Value *, 32> LiveGcValues;
3974 for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) {
3975 GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc);
3976
3977 // Remove the relocation if unused.
3978 if (GCR.use_empty()) {
3979 eraseInstFromFunction(I&: GCR);
3980 continue;
3981 }
3982
3983 Value *DerivedPtr = GCR.getDerivedPtr();
3984 Value *BasePtr = GCR.getBasePtr();
3985
3986 // Undef is undef, even after relocation.
3987 if (isa<UndefValue>(Val: DerivedPtr) || isa<UndefValue>(Val: BasePtr)) {
3988 replaceInstUsesWith(I&: GCR, V: UndefValue::get(T: GCR.getType()));
3989 eraseInstFromFunction(I&: GCR);
3990 continue;
3991 }
3992
3993 if (auto *PT = dyn_cast<PointerType>(Val: GCR.getType())) {
3994 // The relocation of null will be null for most any collector.
3995 // TODO: provide a hook for this in GCStrategy. There might be some
3996 // weird collector this property does not hold for.
3997 if (isa<ConstantPointerNull>(Val: DerivedPtr)) {
3998 // Use null-pointer of gc_relocate's type to replace it.
3999 replaceInstUsesWith(I&: GCR, V: ConstantPointerNull::get(T: PT));
4000 eraseInstFromFunction(I&: GCR);
4001 continue;
4002 }
4003
4004 // isKnownNonNull -> nonnull attribute
4005 if (!GCR.hasRetAttr(Kind: Attribute::NonNull) &&
4006 isKnownNonZero(V: DerivedPtr,
4007 Q: getSimplifyQuery().getWithInstruction(I: &Call))) {
4008 GCR.addRetAttr(Kind: Attribute::NonNull);
4009 // We discovered new fact, re-check users.
4010 Worklist.pushUsersToWorkList(I&: GCR);
4011 }
4012 }
4013
4014 // If we have two copies of the same pointer in the statepoint argument
4015 // list, canonicalize to one. This may let us common gc.relocates.
4016 if (GCR.getBasePtr() == GCR.getDerivedPtr() &&
4017 GCR.getBasePtrIndex() != GCR.getDerivedPtrIndex()) {
4018 auto *OpIntTy = GCR.getOperand(i_nocapture: 2)->getType();
4019 GCR.setOperand(i_nocapture: 2, Val_nocapture: ConstantInt::get(Ty: OpIntTy, V: GCR.getBasePtrIndex()));
4020 }
4021
4022 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
4023 // Canonicalize on the type from the uses to the defs
4024
4025 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
4026 LiveGcValues.insert(Ptr: BasePtr);
4027 LiveGcValues.insert(Ptr: DerivedPtr);
4028 }
4029 std::optional<OperandBundleUse> Bundle =
4030 GCSP.getOperandBundle(ID: LLVMContext::OB_gc_live);
4031 unsigned NumOfGCLives = LiveGcValues.size();
4032 if (!Bundle || NumOfGCLives == Bundle->Inputs.size())
4033 break;
4034 // We can reduce the size of gc live bundle.
4035 DenseMap<Value *, unsigned> Val2Idx;
4036 std::vector<Value *> NewLiveGc;
4037 for (Value *V : Bundle->Inputs) {
4038 if (Val2Idx.count(Val: V))
4039 continue;
4040 if (LiveGcValues.count(Ptr: V)) {
4041 Val2Idx[V] = NewLiveGc.size();
4042 NewLiveGc.push_back(x: V);
4043 } else
4044 Val2Idx[V] = NumOfGCLives;
4045 }
4046 // Update all gc.relocates
4047 for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) {
4048 GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc);
4049 Value *BasePtr = GCR.getBasePtr();
4050 assert(Val2Idx.count(BasePtr) && Val2Idx[BasePtr] != NumOfGCLives &&
4051 "Missed live gc for base pointer");
4052 auto *OpIntTy1 = GCR.getOperand(i_nocapture: 1)->getType();
4053 GCR.setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: OpIntTy1, V: Val2Idx[BasePtr]));
4054 Value *DerivedPtr = GCR.getDerivedPtr();
4055 assert(Val2Idx.count(DerivedPtr) && Val2Idx[DerivedPtr] != NumOfGCLives &&
4056 "Missed live gc for derived pointer");
4057 auto *OpIntTy2 = GCR.getOperand(i_nocapture: 2)->getType();
4058 GCR.setOperand(i_nocapture: 2, Val_nocapture: ConstantInt::get(Ty: OpIntTy2, V: Val2Idx[DerivedPtr]));
4059 }
4060 // Create new statepoint instruction.
4061 OperandBundleDef NewBundle("gc-live", NewLiveGc);
4062 return CallBase::Create(CB: &Call, Bundle: NewBundle);
4063 }
4064 default: { break; }
4065 }
4066
4067 return Changed ? &Call : nullptr;
4068}
4069
4070/// If the callee is a constexpr cast of a function, attempt to move the cast to
4071/// the arguments of the call/invoke.
4072/// CallBrInst is not supported.
4073bool InstCombinerImpl::transformConstExprCastCall(CallBase &Call) {
4074 auto *Callee =
4075 dyn_cast<Function>(Val: Call.getCalledOperand()->stripPointerCasts());
4076 if (!Callee)
4077 return false;
4078
4079 assert(!isa<CallBrInst>(Call) &&
4080 "CallBr's don't have a single point after a def to insert at");
4081
4082 // If this is a call to a thunk function, don't remove the cast. Thunks are
4083 // used to transparently forward all incoming parameters and outgoing return
4084 // values, so it's important to leave the cast in place.
4085 if (Callee->hasFnAttribute(Kind: "thunk"))
4086 return false;
4087
4088 // If this is a call to a naked function, the assembly might be
4089 // using an argument, or otherwise rely on the frame layout,
4090 // the function prototype will mismatch.
4091 if (Callee->hasFnAttribute(Kind: Attribute::Naked))
4092 return false;
4093
4094 // If this is a musttail call, the callee's prototype must match the caller's
4095 // prototype with the exception of pointee types. The code below doesn't
4096 // implement that, so we can't do this transform.
4097 // TODO: Do the transform if it only requires adding pointer casts.
4098 if (Call.isMustTailCall())
4099 return false;
4100
4101 Instruction *Caller = &Call;
4102 const AttributeList &CallerPAL = Call.getAttributes();
4103
4104 // Okay, this is a cast from a function to a different type. Unless doing so
4105 // would cause a type conversion of one of our arguments, change this call to
4106 // be a direct call with arguments casted to the appropriate types.
4107 FunctionType *FT = Callee->getFunctionType();
4108 Type *OldRetTy = Caller->getType();
4109 Type *NewRetTy = FT->getReturnType();
4110
4111 // Check to see if we are changing the return type...
4112 if (OldRetTy != NewRetTy) {
4113
4114 if (NewRetTy->isStructTy())
4115 return false; // TODO: Handle multiple return values.
4116
4117 if (!CastInst::isBitOrNoopPointerCastable(SrcTy: NewRetTy, DestTy: OldRetTy, DL)) {
4118 if (Callee->isDeclaration())
4119 return false; // Cannot transform this return value.
4120
4121 if (!Caller->use_empty() &&
4122 // void -> non-void is handled specially
4123 !NewRetTy->isVoidTy())
4124 return false; // Cannot transform this return value.
4125 }
4126
4127 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
4128 AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs());
4129 if (RAttrs.overlaps(AM: AttributeFuncs::typeIncompatible(Ty: NewRetTy)))
4130 return false; // Attribute not compatible with transformed value.
4131 }
4132
4133 // If the callbase is an invoke instruction, and the return value is
4134 // used by a PHI node in a successor, we cannot change the return type of
4135 // the call because there is no place to put the cast instruction (without
4136 // breaking the critical edge). Bail out in this case.
4137 if (!Caller->use_empty()) {
4138 BasicBlock *PhisNotSupportedBlock = nullptr;
4139 if (auto *II = dyn_cast<InvokeInst>(Val: Caller))
4140 PhisNotSupportedBlock = II->getNormalDest();
4141 if (PhisNotSupportedBlock)
4142 for (User *U : Caller->users())
4143 if (PHINode *PN = dyn_cast<PHINode>(Val: U))
4144 if (PN->getParent() == PhisNotSupportedBlock)
4145 return false;
4146 }
4147 }
4148
4149 unsigned NumActualArgs = Call.arg_size();
4150 unsigned NumCommonArgs = std::min(a: FT->getNumParams(), b: NumActualArgs);
4151
4152 // Prevent us turning:
4153 // declare void @takes_i32_inalloca(i32* inalloca)
4154 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
4155 //
4156 // into:
4157 // call void @takes_i32_inalloca(i32* null)
4158 //
4159 // Similarly, avoid folding away bitcasts of byval calls.
4160 if (Callee->getAttributes().hasAttrSomewhere(Kind: Attribute::InAlloca) ||
4161 Callee->getAttributes().hasAttrSomewhere(Kind: Attribute::Preallocated))
4162 return false;
4163
4164 auto AI = Call.arg_begin();
4165 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
4166 Type *ParamTy = FT->getParamType(i);
4167 Type *ActTy = (*AI)->getType();
4168
4169 if (!CastInst::isBitOrNoopPointerCastable(SrcTy: ActTy, DestTy: ParamTy, DL))
4170 return false; // Cannot transform this parameter value.
4171
4172 // Check if there are any incompatible attributes we cannot drop safely.
4173 if (AttrBuilder(FT->getContext(), CallerPAL.getParamAttrs(ArgNo: i))
4174 .overlaps(AM: AttributeFuncs::typeIncompatible(
4175 Ty: ParamTy, ASK: AttributeFuncs::ASK_UNSAFE_TO_DROP)))
4176 return false; // Attribute not compatible with transformed value.
4177
4178 if (Call.isInAllocaArgument(ArgNo: i) ||
4179 CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::Preallocated))
4180 return false; // Cannot transform to and from inalloca/preallocated.
4181
4182 if (CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::SwiftError))
4183 return false;
4184
4185 if (CallerPAL.hasParamAttr(ArgNo: i, Kind: Attribute::ByVal) !=
4186 Callee->getAttributes().hasParamAttr(ArgNo: i, Kind: Attribute::ByVal))
4187 return false; // Cannot transform to or from byval.
4188 }
4189
4190 if (Callee->isDeclaration()) {
4191 // Do not delete arguments unless we have a function body.
4192 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
4193 return false;
4194
4195 // If the callee is just a declaration, don't change the varargsness of the
4196 // call. We don't want to introduce a varargs call where one doesn't
4197 // already exist.
4198 if (FT->isVarArg() != Call.getFunctionType()->isVarArg())
4199 return false;
4200
4201 // If both the callee and the cast type are varargs, we still have to make
4202 // sure the number of fixed parameters are the same or we have the same
4203 // ABI issues as if we introduce a varargs call.
4204 if (FT->isVarArg() && Call.getFunctionType()->isVarArg() &&
4205 FT->getNumParams() != Call.getFunctionType()->getNumParams())
4206 return false;
4207 }
4208
4209 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
4210 !CallerPAL.isEmpty()) {
4211 // In this case we have more arguments than the new function type, but we
4212 // won't be dropping them. Check that these extra arguments have attributes
4213 // that are compatible with being a vararg call argument.
4214 unsigned SRetIdx;
4215 if (CallerPAL.hasAttrSomewhere(Kind: Attribute::StructRet, Index: &SRetIdx) &&
4216 SRetIdx - AttributeList::FirstArgIndex >= FT->getNumParams())
4217 return false;
4218 }
4219
4220 // Okay, we decided that this is a safe thing to do: go ahead and start
4221 // inserting cast instructions as necessary.
4222 SmallVector<Value *, 8> Args;
4223 SmallVector<AttributeSet, 8> ArgAttrs;
4224 Args.reserve(N: NumActualArgs);
4225 ArgAttrs.reserve(N: NumActualArgs);
4226
4227 // Get any return attributes.
4228 AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs());
4229
4230 // If the return value is not being used, the type may not be compatible
4231 // with the existing attributes. Wipe out any problematic attributes.
4232 RAttrs.remove(AM: AttributeFuncs::typeIncompatible(Ty: NewRetTy));
4233
4234 LLVMContext &Ctx = Call.getContext();
4235 AI = Call.arg_begin();
4236 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
4237 Type *ParamTy = FT->getParamType(i);
4238
4239 Value *NewArg = *AI;
4240 if ((*AI)->getType() != ParamTy)
4241 NewArg = Builder.CreateBitOrPointerCast(V: *AI, DestTy: ParamTy);
4242 Args.push_back(Elt: NewArg);
4243
4244 // Add any parameter attributes except the ones incompatible with the new
4245 // type. Note that we made sure all incompatible ones are safe to drop.
4246 AttributeMask IncompatibleAttrs = AttributeFuncs::typeIncompatible(
4247 Ty: ParamTy, ASK: AttributeFuncs::ASK_SAFE_TO_DROP);
4248 ArgAttrs.push_back(
4249 Elt: CallerPAL.getParamAttrs(ArgNo: i).removeAttributes(C&: Ctx, AttrsToRemove: IncompatibleAttrs));
4250 }
4251
4252 // If the function takes more arguments than the call was taking, add them
4253 // now.
4254 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) {
4255 Args.push_back(Elt: Constant::getNullValue(Ty: FT->getParamType(i)));
4256 ArgAttrs.push_back(Elt: AttributeSet());
4257 }
4258
4259 // If we are removing arguments to the function, emit an obnoxious warning.
4260 if (FT->getNumParams() < NumActualArgs) {
4261 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
4262 if (FT->isVarArg()) {
4263 // Add all of the arguments in their promoted form to the arg list.
4264 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
4265 Type *PTy = getPromotedType(Ty: (*AI)->getType());
4266 Value *NewArg = *AI;
4267 if (PTy != (*AI)->getType()) {
4268 // Must promote to pass through va_arg area!
4269 Instruction::CastOps opcode =
4270 CastInst::getCastOpcode(Val: *AI, SrcIsSigned: false, Ty: PTy, DstIsSigned: false);
4271 NewArg = Builder.CreateCast(Op: opcode, V: *AI, DestTy: PTy);
4272 }
4273 Args.push_back(Elt: NewArg);
4274
4275 // Add any parameter attributes.
4276 ArgAttrs.push_back(Elt: CallerPAL.getParamAttrs(ArgNo: i));
4277 }
4278 }
4279 }
4280
4281 AttributeSet FnAttrs = CallerPAL.getFnAttrs();
4282
4283 if (NewRetTy->isVoidTy())
4284 Caller->setName(""); // Void type should not have a name.
4285
4286 assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) &&
4287 "missing argument attributes");
4288 AttributeList NewCallerPAL = AttributeList::get(
4289 C&: Ctx, FnAttrs, RetAttrs: AttributeSet::get(C&: Ctx, B: RAttrs), ArgAttrs);
4290
4291 SmallVector<OperandBundleDef, 1> OpBundles;
4292 Call.getOperandBundlesAsDefs(Defs&: OpBundles);
4293
4294 CallBase *NewCall;
4295 if (InvokeInst *II = dyn_cast<InvokeInst>(Val: Caller)) {
4296 NewCall = Builder.CreateInvoke(Callee, NormalDest: II->getNormalDest(),
4297 UnwindDest: II->getUnwindDest(), Args, OpBundles);
4298 } else {
4299 NewCall = Builder.CreateCall(Callee, Args, OpBundles);
4300 cast<CallInst>(Val: NewCall)->setTailCallKind(
4301 cast<CallInst>(Val: Caller)->getTailCallKind());
4302 }
4303 NewCall->takeName(V: Caller);
4304 NewCall->setCallingConv(Call.getCallingConv());
4305 NewCall->setAttributes(NewCallerPAL);
4306
4307 // Preserve prof metadata if any.
4308 NewCall->copyMetadata(SrcInst: *Caller, WL: {LLVMContext::MD_prof});
4309
4310 // Insert a cast of the return type as necessary.
4311 Instruction *NC = NewCall;
4312 Value *NV = NC;
4313 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
4314 if (!NV->getType()->isVoidTy()) {
4315 NV = NC = CastInst::CreateBitOrPointerCast(S: NC, Ty: OldRetTy);
4316 NC->setDebugLoc(Caller->getDebugLoc());
4317
4318 auto OptInsertPt = NewCall->getInsertionPointAfterDef();
4319 assert(OptInsertPt && "No place to insert cast");
4320 InsertNewInstBefore(New: NC, Old: *OptInsertPt);
4321 Worklist.pushUsersToWorkList(I&: *Caller);
4322 } else {
4323 NV = PoisonValue::get(T: Caller->getType());
4324 }
4325 }
4326
4327 if (!Caller->use_empty())
4328 replaceInstUsesWith(I&: *Caller, V: NV);
4329 else if (Caller->hasValueHandle()) {
4330 if (OldRetTy == NV->getType())
4331 ValueHandleBase::ValueIsRAUWd(Old: Caller, New: NV);
4332 else
4333 // We cannot call ValueIsRAUWd with a different type, and the
4334 // actual tracked value will disappear.
4335 ValueHandleBase::ValueIsDeleted(V: Caller);
4336 }
4337
4338 eraseInstFromFunction(I&: *Caller);
4339 return true;
4340}
4341
4342/// Turn a call to a function created by init_trampoline / adjust_trampoline
4343/// intrinsic pair into a direct call to the underlying function.
4344Instruction *
4345InstCombinerImpl::transformCallThroughTrampoline(CallBase &Call,
4346 IntrinsicInst &Tramp) {
4347 FunctionType *FTy = Call.getFunctionType();
4348 AttributeList Attrs = Call.getAttributes();
4349
4350 // If the call already has the 'nest' attribute somewhere then give up -
4351 // otherwise 'nest' would occur twice after splicing in the chain.
4352 if (Attrs.hasAttrSomewhere(Kind: Attribute::Nest))
4353 return nullptr;
4354
4355 Function *NestF = cast<Function>(Val: Tramp.getArgOperand(i: 1)->stripPointerCasts());
4356 FunctionType *NestFTy = NestF->getFunctionType();
4357
4358 AttributeList NestAttrs = NestF->getAttributes();
4359 if (!NestAttrs.isEmpty()) {
4360 unsigned NestArgNo = 0;
4361 Type *NestTy = nullptr;
4362 AttributeSet NestAttr;
4363
4364 // Look for a parameter marked with the 'nest' attribute.
4365 for (FunctionType::param_iterator I = NestFTy->param_begin(),
4366 E = NestFTy->param_end();
4367 I != E; ++NestArgNo, ++I) {
4368 AttributeSet AS = NestAttrs.getParamAttrs(ArgNo: NestArgNo);
4369 if (AS.hasAttribute(Kind: Attribute::Nest)) {
4370 // Record the parameter type and any other attributes.
4371 NestTy = *I;
4372 NestAttr = AS;
4373 break;
4374 }
4375 }
4376
4377 if (NestTy) {
4378 std::vector<Value*> NewArgs;
4379 std::vector<AttributeSet> NewArgAttrs;
4380 NewArgs.reserve(n: Call.arg_size() + 1);
4381 NewArgAttrs.reserve(n: Call.arg_size());
4382
4383 // Insert the nest argument into the call argument list, which may
4384 // mean appending it. Likewise for attributes.
4385
4386 {
4387 unsigned ArgNo = 0;
4388 auto I = Call.arg_begin(), E = Call.arg_end();
4389 do {
4390 if (ArgNo == NestArgNo) {
4391 // Add the chain argument and attributes.
4392 Value *NestVal = Tramp.getArgOperand(i: 2);
4393 if (NestVal->getType() != NestTy)
4394 NestVal = Builder.CreateBitCast(V: NestVal, DestTy: NestTy, Name: "nest");
4395 NewArgs.push_back(x: NestVal);
4396 NewArgAttrs.push_back(x: NestAttr);
4397 }
4398
4399 if (I == E)
4400 break;
4401
4402 // Add the original argument and attributes.
4403 NewArgs.push_back(x: *I);
4404 NewArgAttrs.push_back(x: Attrs.getParamAttrs(ArgNo));
4405
4406 ++ArgNo;
4407 ++I;
4408 } while (true);
4409 }
4410
4411 // The trampoline may have been bitcast to a bogus type (FTy).
4412 // Handle this by synthesizing a new function type, equal to FTy
4413 // with the chain parameter inserted.
4414
4415 std::vector<Type*> NewTypes;
4416 NewTypes.reserve(n: FTy->getNumParams()+1);
4417
4418 // Insert the chain's type into the list of parameter types, which may
4419 // mean appending it.
4420 {
4421 unsigned ArgNo = 0;
4422 FunctionType::param_iterator I = FTy->param_begin(),
4423 E = FTy->param_end();
4424
4425 do {
4426 if (ArgNo == NestArgNo)
4427 // Add the chain's type.
4428 NewTypes.push_back(x: NestTy);
4429
4430 if (I == E)
4431 break;
4432
4433 // Add the original type.
4434 NewTypes.push_back(x: *I);
4435
4436 ++ArgNo;
4437 ++I;
4438 } while (true);
4439 }
4440
4441 // Replace the trampoline call with a direct call. Let the generic
4442 // code sort out any function type mismatches.
4443 FunctionType *NewFTy =
4444 FunctionType::get(Result: FTy->getReturnType(), Params: NewTypes, isVarArg: FTy->isVarArg());
4445 AttributeList NewPAL =
4446 AttributeList::get(C&: FTy->getContext(), FnAttrs: Attrs.getFnAttrs(),
4447 RetAttrs: Attrs.getRetAttrs(), ArgAttrs: NewArgAttrs);
4448
4449 SmallVector<OperandBundleDef, 1> OpBundles;
4450 Call.getOperandBundlesAsDefs(Defs&: OpBundles);
4451
4452 Instruction *NewCaller;
4453 if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &Call)) {
4454 NewCaller = InvokeInst::Create(Ty: NewFTy, Func: NestF, IfNormal: II->getNormalDest(),
4455 IfException: II->getUnwindDest(), Args: NewArgs, Bundles: OpBundles);
4456 cast<InvokeInst>(Val: NewCaller)->setCallingConv(II->getCallingConv());
4457 cast<InvokeInst>(Val: NewCaller)->setAttributes(NewPAL);
4458 } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(Val: &Call)) {
4459 NewCaller =
4460 CallBrInst::Create(Ty: NewFTy, Func: NestF, DefaultDest: CBI->getDefaultDest(),
4461 IndirectDests: CBI->getIndirectDests(), Args: NewArgs, Bundles: OpBundles);
4462 cast<CallBrInst>(Val: NewCaller)->setCallingConv(CBI->getCallingConv());
4463 cast<CallBrInst>(Val: NewCaller)->setAttributes(NewPAL);
4464 } else {
4465 NewCaller = CallInst::Create(Ty: NewFTy, Func: NestF, Args: NewArgs, Bundles: OpBundles);
4466 cast<CallInst>(Val: NewCaller)->setTailCallKind(
4467 cast<CallInst>(Val&: Call).getTailCallKind());
4468 cast<CallInst>(Val: NewCaller)->setCallingConv(
4469 cast<CallInst>(Val&: Call).getCallingConv());
4470 cast<CallInst>(Val: NewCaller)->setAttributes(NewPAL);
4471 }
4472 NewCaller->setDebugLoc(Call.getDebugLoc());
4473
4474 return NewCaller;
4475 }
4476 }
4477
4478 // Replace the trampoline call with a direct call. Since there is no 'nest'
4479 // parameter, there is no need to adjust the argument list. Let the generic
4480 // code sort out any function type mismatches.
4481 Call.setCalledFunction(FTy, Fn: NestF);
4482 return &Call;
4483}
4484