LLVM 17.0.0git
InstCombineCalls.cpp
Go to the documentation of this file.
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"
21#include "llvm/ADT/Statistic.h"
26#include "llvm/Analysis/Loads.h"
30#include "llvm/IR/Attributes.h"
31#include "llvm/IR/BasicBlock.h"
32#include "llvm/IR/Constant.h"
33#include "llvm/IR/Constants.h"
34#include "llvm/IR/DataLayout.h"
35#include "llvm/IR/DebugInfo.h"
37#include "llvm/IR/Function.h"
39#include "llvm/IR/InlineAsm.h"
40#include "llvm/IR/InstrTypes.h"
41#include "llvm/IR/Instruction.h"
44#include "llvm/IR/Intrinsics.h"
45#include "llvm/IR/IntrinsicsAArch64.h"
46#include "llvm/IR/IntrinsicsAMDGPU.h"
47#include "llvm/IR/IntrinsicsARM.h"
48#include "llvm/IR/IntrinsicsHexagon.h"
49#include "llvm/IR/LLVMContext.h"
50#include "llvm/IR/Metadata.h"
52#include "llvm/IR/Statepoint.h"
53#include "llvm/IR/Type.h"
54#include "llvm/IR/User.h"
55#include "llvm/IR/Value.h"
56#include "llvm/IR/ValueHandle.h"
61#include "llvm/Support/Debug.h"
70#include <algorithm>
71#include <cassert>
72#include <cstdint>
73#include <optional>
74#include <utility>
75#include <vector>
76
77#define DEBUG_TYPE "instcombine"
79
80using namespace llvm;
81using namespace PatternMatch;
82
83STATISTIC(NumSimplified, "Number of library calls simplified");
84
86 "instcombine-guard-widening-window",
87 cl::init(3),
88 cl::desc("How wide an instruction window to bypass looking for "
89 "another guard"));
90
91namespace llvm {
92/// enable preservation of attributes in assume like:
93/// call void @llvm.assume(i1 true) [ "nonnull"(i32* %PTR) ]
95} // namespace llvm
96
97/// Return the specified type promoted as it would be to pass though a va_arg
98/// area.
100 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
101 if (ITy->getBitWidth() < 32)
102 return Type::getInt32Ty(Ty->getContext());
103 }
104 return Ty;
105}
106
107/// Recognize a memcpy/memmove from a trivially otherwise unused alloca.
108/// TODO: This should probably be integrated with visitAllocSites, but that
109/// requires a deeper change to allow either unread or unwritten objects.
111 auto *Src = MI->getRawSource();
112 while (isa<GetElementPtrInst>(Src) || isa<BitCastInst>(Src)) {
113 if (!Src->hasOneUse())
114 return false;
115 Src = cast<Instruction>(Src)->getOperand(0);
116 }
117 return isa<AllocaInst>(Src) && Src->hasOneUse();
118}
119
121 Align DstAlign = getKnownAlignment(MI->getRawDest(), DL, MI, &AC, &DT);
122 MaybeAlign CopyDstAlign = MI->getDestAlign();
123 if (!CopyDstAlign || *CopyDstAlign < DstAlign) {
124 MI->setDestAlignment(DstAlign);
125 return MI;
126 }
127
128 Align SrcAlign = getKnownAlignment(MI->getRawSource(), DL, MI, &AC, &DT);
129 MaybeAlign CopySrcAlign = MI->getSourceAlign();
130 if (!CopySrcAlign || *CopySrcAlign < SrcAlign) {
131 MI->setSourceAlignment(SrcAlign);
132 return MI;
133 }
134
135 // If we have a store to a location which is known constant, we can conclude
136 // that the store must be storing the constant value (else the memory
137 // wouldn't be constant), and this must be a noop.
138 if (!isModSet(AA->getModRefInfoMask(MI->getDest()))) {
139 // Set the size of the copy to 0, it will be deleted on the next iteration.
140 MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
141 return MI;
142 }
143
144 // If the source is provably undef, the memcpy/memmove doesn't do anything
145 // (unless the transfer is volatile).
146 if (hasUndefSource(MI) && !MI->isVolatile()) {
147 // Set the size of the copy to 0, it will be deleted on the next iteration.
148 MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
149 return MI;
150 }
151
152 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
153 // load/store.
154 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getLength());
155 if (!MemOpLength) return nullptr;
156
157 // Source and destination pointer types are always "i8*" for intrinsic. See
158 // if the size is something we can handle with a single primitive load/store.
159 // A single load+store correctly handles overlapping memory in the memmove
160 // case.
161 uint64_t Size = MemOpLength->getLimitedValue();
162 assert(Size && "0-sized memory transferring should be removed already.");
163
164 if (Size > 8 || (Size&(Size-1)))
165 return nullptr; // If not 1/2/4/8 bytes, exit.
166
167 // If it is an atomic and alignment is less than the size then we will
168 // introduce the unaligned memory access which will be later transformed
169 // into libcall in CodeGen. This is not evident performance gain so disable
170 // it now.
171 if (isa<AtomicMemTransferInst>(MI))
172 if (*CopyDstAlign < Size || *CopySrcAlign < Size)
173 return nullptr;
174
175 // Use an integer load+store unless we can find something better.
176 unsigned SrcAddrSp =
177 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
178 unsigned DstAddrSp =
179 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
180
181 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
182 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
183 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
184
185 // If the memcpy has metadata describing the members, see if we can get the
186 // TBAA tag describing our copy.
187 MDNode *CopyMD = nullptr;
188 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa)) {
189 CopyMD = M;
190 } else if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
191 if (M->getNumOperands() == 3 && M->getOperand(0) &&
192 mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
193 mdconst::extract<ConstantInt>(M->getOperand(0))->isZero() &&
194 M->getOperand(1) &&
195 mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
196 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
197 Size &&
198 M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
199 CopyMD = cast<MDNode>(M->getOperand(2));
200 }
201
202 Value *Src = Builder.CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
203 Value *Dest = Builder.CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
204 LoadInst *L = Builder.CreateLoad(IntType, Src);
205 // Alignment from the mem intrinsic will be better, so use it.
206 L->setAlignment(*CopySrcAlign);
207 if (CopyMD)
208 L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
209 MDNode *LoopMemParallelMD =
210 MI->getMetadata(LLVMContext::MD_mem_parallel_loop_access);
211 if (LoopMemParallelMD)
212 L->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
213 MDNode *AccessGroupMD = MI->getMetadata(LLVMContext::MD_access_group);
214 if (AccessGroupMD)
215 L->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
216
217 StoreInst *S = Builder.CreateStore(L, Dest);
218 // Alignment from the mem intrinsic will be better, so use it.
219 S->setAlignment(*CopyDstAlign);
220 if (CopyMD)
221 S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
222 if (LoopMemParallelMD)
223 S->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
224 if (AccessGroupMD)
225 S->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
226 S->copyMetadata(*MI, LLVMContext::MD_DIAssignID);
227
228 if (auto *MT = dyn_cast<MemTransferInst>(MI)) {
229 // non-atomics can be volatile
230 L->setVolatile(MT->isVolatile());
231 S->setVolatile(MT->isVolatile());
232 }
233 if (isa<AtomicMemTransferInst>(MI)) {
234 // atomics have to be unordered
237 }
238
239 // Set the size of the copy to 0, it will be deleted on the next iteration.
240 MI->setLength(Constant::getNullValue(MemOpLength->getType()));
241 return MI;
242}
243
245 const Align KnownAlignment =
246 getKnownAlignment(MI->getDest(), DL, MI, &AC, &DT);
247 MaybeAlign MemSetAlign = MI->getDestAlign();
248 if (!MemSetAlign || *MemSetAlign < KnownAlignment) {
249 MI->setDestAlignment(KnownAlignment);
250 return MI;
251 }
252
253 // If we have a store to a location which is known constant, we can conclude
254 // that the store must be storing the constant value (else the memory
255 // wouldn't be constant), and this must be a noop.
256 if (!isModSet(AA->getModRefInfoMask(MI->getDest()))) {
257 // Set the size of the copy to 0, it will be deleted on the next iteration.
258 MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
259 return MI;
260 }
261
262 // Remove memset with an undef value.
263 // FIXME: This is technically incorrect because it might overwrite a poison
264 // value. Change to PoisonValue once #52930 is resolved.
265 if (isa<UndefValue>(MI->getValue())) {
266 // Set the size of the copy to 0, it will be deleted on the next iteration.
267 MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
268 return MI;
269 }
270
271 // Extract the length and alignment and fill if they are constant.
272 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
273 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
274 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
275 return nullptr;
276 const uint64_t Len = LenC->getLimitedValue();
277 assert(Len && "0-sized memory setting should be removed already.");
278 const Align Alignment = MI->getDestAlign().valueOrOne();
279
280 // If it is an atomic and alignment is less than the size then we will
281 // introduce the unaligned memory access which will be later transformed
282 // into libcall in CodeGen. This is not evident performance gain so disable
283 // it now.
284 if (isa<AtomicMemSetInst>(MI))
285 if (Alignment < Len)
286 return nullptr;
287
288 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
289 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
290 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
291
292 Value *Dest = MI->getDest();
293 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
294 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
295 Dest = Builder.CreateBitCast(Dest, NewDstPtrTy);
296
297 // Extract the fill value and store.
298 const uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
299 Constant *FillVal = ConstantInt::get(ITy, Fill);
300 StoreInst *S = Builder.CreateStore(FillVal, Dest, MI->isVolatile());
301 S->copyMetadata(*MI, LLVMContext::MD_DIAssignID);
302 for (auto *DAI : at::getAssignmentMarkers(S)) {
303 if (any_of(DAI->location_ops(), [&](Value *V) { return V == FillC; }))
304 DAI->replaceVariableLocationOp(FillC, FillVal);
305 }
306
307 S->setAlignment(Alignment);
308 if (isa<AtomicMemSetInst>(MI))
310
311 // Set the size of the copy to 0, it will be deleted on the next iteration.
312 MI->setLength(Constant::getNullValue(LenC->getType()));
313 return MI;
314 }
315
316 return nullptr;
317}
318
319// TODO, Obvious Missing Transforms:
320// * Narrow width by halfs excluding zero/undef lanes
321Value *InstCombinerImpl::simplifyMaskedLoad(IntrinsicInst &II) {
322 Value *LoadPtr = II.getArgOperand(0);
323 const Align Alignment =
324 cast<ConstantInt>(II.getArgOperand(1))->getAlignValue();
325
326 // If the mask is all ones or undefs, this is a plain vector load of the 1st
327 // argument.
329 LoadInst *L = Builder.CreateAlignedLoad(II.getType(), LoadPtr, Alignment,
330 "unmaskedload");
331 L->copyMetadata(II);
332 return L;
333 }
334
335 // If we can unconditionally load from this address, replace with a
336 // load/select idiom. TODO: use DT for context sensitive query
337 if (isDereferenceablePointer(LoadPtr, II.getType(),
338 II.getModule()->getDataLayout(), &II, &AC)) {
339 LoadInst *LI = Builder.CreateAlignedLoad(II.getType(), LoadPtr, Alignment,
340 "unmaskedload");
341 LI->copyMetadata(II);
342 return Builder.CreateSelect(II.getArgOperand(2), LI, II.getArgOperand(3));
343 }
344
345 return nullptr;
346}
347
348// TODO, Obvious Missing Transforms:
349// * Single constant active lane -> store
350// * Narrow width by halfs excluding zero/undef lanes
351Instruction *InstCombinerImpl::simplifyMaskedStore(IntrinsicInst &II) {
352 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
353 if (!ConstMask)
354 return nullptr;
355
356 // If the mask is all zeros, this instruction does nothing.
357 if (ConstMask->isNullValue())
358 return eraseInstFromFunction(II);
359
360 // If the mask is all ones, this is a plain vector store of the 1st argument.
361 if (ConstMask->isAllOnesValue()) {
362 Value *StorePtr = II.getArgOperand(1);
363 Align Alignment = cast<ConstantInt>(II.getArgOperand(2))->getAlignValue();
364 StoreInst *S =
365 new StoreInst(II.getArgOperand(0), StorePtr, false, Alignment);
366 S->copyMetadata(II);
367 return S;
368 }
369
370 if (isa<ScalableVectorType>(ConstMask->getType()))
371 return nullptr;
372
373 // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
374 APInt DemandedElts = possiblyDemandedEltsInMask(ConstMask);
375 APInt UndefElts(DemandedElts.getBitWidth(), 0);
376 if (Value *V =
377 SimplifyDemandedVectorElts(II.getOperand(0), DemandedElts, UndefElts))
378 return replaceOperand(II, 0, V);
379
380 return nullptr;
381}
382
383// TODO, Obvious Missing Transforms:
384// * Single constant active lane load -> load
385// * Dereferenceable address & few lanes -> scalarize speculative load/selects
386// * Adjacent vector addresses -> masked.load
387// * Narrow width by halfs excluding zero/undef lanes
388// * Vector incrementing address -> vector masked load
389Instruction *InstCombinerImpl::simplifyMaskedGather(IntrinsicInst &II) {
390 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(2));
391 if (!ConstMask)
392 return nullptr;
393
394 // Vector splat address w/known mask -> scalar load
395 // Fold the gather to load the source vector first lane
396 // because it is reloading the same value each time
397 if (ConstMask->isAllOnesValue())
398 if (auto *SplatPtr = getSplatValue(II.getArgOperand(0))) {
399 auto *VecTy = cast<VectorType>(II.getType());
400 const Align Alignment =
401 cast<ConstantInt>(II.getArgOperand(1))->getAlignValue();
402 LoadInst *L = Builder.CreateAlignedLoad(VecTy->getElementType(), SplatPtr,
403 Alignment, "load.scalar");
404 Value *Shuf =
405 Builder.CreateVectorSplat(VecTy->getElementCount(), L, "broadcast");
406 return replaceInstUsesWith(II, cast<Instruction>(Shuf));
407 }
408
409 return nullptr;
410}
411
412// TODO, Obvious Missing Transforms:
413// * Single constant active lane -> store
414// * Adjacent vector addresses -> masked.store
415// * Narrow store width by halfs excluding zero/undef lanes
416// * Vector incrementing address -> vector masked store
417Instruction *InstCombinerImpl::simplifyMaskedScatter(IntrinsicInst &II) {
418 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
419 if (!ConstMask)
420 return nullptr;
421
422 // If the mask is all zeros, a scatter does nothing.
423 if (ConstMask->isNullValue())
424 return eraseInstFromFunction(II);
425
426 // Vector splat address -> scalar store
427 if (auto *SplatPtr = getSplatValue(II.getArgOperand(1))) {
428 // scatter(splat(value), splat(ptr), non-zero-mask) -> store value, ptr
429 if (auto *SplatValue = getSplatValue(II.getArgOperand(0))) {
430 Align Alignment = cast<ConstantInt>(II.getArgOperand(2))->getAlignValue();
431 StoreInst *S =
432 new StoreInst(SplatValue, SplatPtr, /*IsVolatile=*/false, Alignment);
433 S->copyMetadata(II);
434 return S;
435 }
436 // scatter(vector, splat(ptr), splat(true)) -> store extract(vector,
437 // lastlane), ptr
438 if (ConstMask->isAllOnesValue()) {
439 Align Alignment = cast<ConstantInt>(II.getArgOperand(2))->getAlignValue();
440 VectorType *WideLoadTy = cast<VectorType>(II.getArgOperand(1)->getType());
441 ElementCount VF = WideLoadTy->getElementCount();
442 Constant *EC =
444 Value *RunTimeVF = VF.isScalable() ? Builder.CreateVScale(EC) : EC;
445 Value *LastLane = Builder.CreateSub(RunTimeVF, Builder.getInt32(1));
446 Value *Extract =
448 StoreInst *S =
449 new StoreInst(Extract, SplatPtr, /*IsVolatile=*/false, Alignment);
450 S->copyMetadata(II);
451 return S;
452 }
453 }
454 if (isa<ScalableVectorType>(ConstMask->getType()))
455 return nullptr;
456
457 // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
458 APInt DemandedElts = possiblyDemandedEltsInMask(ConstMask);
459 APInt UndefElts(DemandedElts.getBitWidth(), 0);
460 if (Value *V =
461 SimplifyDemandedVectorElts(II.getOperand(0), DemandedElts, UndefElts))
462 return replaceOperand(II, 0, V);
463 if (Value *V =
464 SimplifyDemandedVectorElts(II.getOperand(1), DemandedElts, UndefElts))
465 return replaceOperand(II, 1, V);
466
467 return nullptr;
468}
469
470/// This function transforms launder.invariant.group and strip.invariant.group
471/// like:
472/// launder(launder(%x)) -> launder(%x) (the result is not the argument)
473/// launder(strip(%x)) -> launder(%x)
474/// strip(strip(%x)) -> strip(%x) (the result is not the argument)
475/// strip(launder(%x)) -> strip(%x)
476/// This is legal because it preserves the most recent information about
477/// the presence or absence of invariant.group.
479 InstCombinerImpl &IC) {
480 auto *Arg = II.getArgOperand(0);
481 auto *StrippedArg = Arg->stripPointerCasts();
482 auto *StrippedInvariantGroupsArg = StrippedArg;
483 while (auto *Intr = dyn_cast<IntrinsicInst>(StrippedInvariantGroupsArg)) {
484 if (Intr->getIntrinsicID() != Intrinsic::launder_invariant_group &&
485 Intr->getIntrinsicID() != Intrinsic::strip_invariant_group)
486 break;
487 StrippedInvariantGroupsArg = Intr->getArgOperand(0)->stripPointerCasts();
488 }
489 if (StrippedArg == StrippedInvariantGroupsArg)
490 return nullptr; // No launders/strips to remove.
491
492 Value *Result = nullptr;
493
494 if (II.getIntrinsicID() == Intrinsic::launder_invariant_group)
495 Result = IC.Builder.CreateLaunderInvariantGroup(StrippedInvariantGroupsArg);
496 else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group)
497 Result = IC.Builder.CreateStripInvariantGroup(StrippedInvariantGroupsArg);
498 else
500 "simplifyInvariantGroupIntrinsic only handles launder and strip");
501 if (Result->getType()->getPointerAddressSpace() !=
503 Result = IC.Builder.CreateAddrSpaceCast(Result, II.getType());
504 if (Result->getType() != II.getType())
505 Result = IC.Builder.CreateBitCast(Result, II.getType());
506
507 return cast<Instruction>(Result);
508}
509
511 assert((II.getIntrinsicID() == Intrinsic::cttz ||
512 II.getIntrinsicID() == Intrinsic::ctlz) &&
513 "Expected cttz or ctlz intrinsic");
514 bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz;
515 Value *Op0 = II.getArgOperand(0);
516 Value *Op1 = II.getArgOperand(1);
517 Value *X;
518 // ctlz(bitreverse(x)) -> cttz(x)
519 // cttz(bitreverse(x)) -> ctlz(x)
520 if (match(Op0, m_BitReverse(m_Value(X)))) {
521 Intrinsic::ID ID = IsTZ ? Intrinsic::ctlz : Intrinsic::cttz;
523 return CallInst::Create(F, {X, II.getArgOperand(1)});
524 }
525
526 if (II.getType()->isIntOrIntVectorTy(1)) {
527 // ctlz/cttz i1 Op0 --> not Op0
528 if (match(Op1, m_Zero()))
529 return BinaryOperator::CreateNot(Op0);
530 // If zero is poison, then the input can be assumed to be "true", so the
531 // instruction simplifies to "false".
532 assert(match(Op1, m_One()) && "Expected ctlz/cttz operand to be 0 or 1");
534 }
535
536 // If the operand is a select with constant arm(s), try to hoist ctlz/cttz.
537 if (auto *Sel = dyn_cast<SelectInst>(Op0))
538 if (Instruction *R = IC.FoldOpIntoSelect(II, Sel))
539 return R;
540
541 if (IsTZ) {
542 // cttz(-x) -> cttz(x)
543 if (match(Op0, m_Neg(m_Value(X))))
544 return IC.replaceOperand(II, 0, X);
545
546 // cttz(sext(x)) -> cttz(zext(x))
547 if (match(Op0, m_OneUse(m_SExt(m_Value(X))))) {
548 auto *Zext = IC.Builder.CreateZExt(X, II.getType());
549 auto *CttzZext =
550 IC.Builder.CreateBinaryIntrinsic(Intrinsic::cttz, Zext, Op1);
551 return IC.replaceInstUsesWith(II, CttzZext);
552 }
553
554 // Zext doesn't change the number of trailing zeros, so narrow:
555 // cttz(zext(x)) -> zext(cttz(x)) if the 'ZeroIsPoison' parameter is 'true'.
556 if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) && match(Op1, m_One())) {
557 auto *Cttz = IC.Builder.CreateBinaryIntrinsic(Intrinsic::cttz, X,
558 IC.Builder.getTrue());
559 auto *ZextCttz = IC.Builder.CreateZExt(Cttz, II.getType());
560 return IC.replaceInstUsesWith(II, ZextCttz);
561 }
562
563 // cttz(abs(x)) -> cttz(x)
564 // cttz(nabs(x)) -> cttz(x)
565 Value *Y;
567 if (SPF == SPF_ABS || SPF == SPF_NABS)
568 return IC.replaceOperand(II, 0, X);
569
570 if (match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(X))))
571 return IC.replaceOperand(II, 0, X);
572 }
573
574 KnownBits Known = IC.computeKnownBits(Op0, 0, &II);
575
576 // Create a mask for bits above (ctlz) or below (cttz) the first known one.
577 unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros()
578 : Known.countMaxLeadingZeros();
579 unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros()
580 : Known.countMinLeadingZeros();
581
582 // If all bits above (ctlz) or below (cttz) the first known one are known
583 // zero, this value is constant.
584 // FIXME: This should be in InstSimplify because we're replacing an
585 // instruction with a constant.
586 if (PossibleZeros == DefiniteZeros) {
587 auto *C = ConstantInt::get(Op0->getType(), DefiniteZeros);
588 return IC.replaceInstUsesWith(II, C);
589 }
590
591 // If the input to cttz/ctlz is known to be non-zero,
592 // then change the 'ZeroIsPoison' parameter to 'true'
593 // because we know the zero behavior can't affect the result.
594 if (!Known.One.isZero() ||
595 isKnownNonZero(Op0, IC.getDataLayout(), 0, &IC.getAssumptionCache(), &II,
596 &IC.getDominatorTree())) {
597 if (!match(II.getArgOperand(1), m_One()))
598 return IC.replaceOperand(II, 1, IC.Builder.getTrue());
599 }
600
601 // Add range metadata since known bits can't completely reflect what we know.
602 // TODO: Handle splat vectors.
603 auto *IT = dyn_cast<IntegerType>(Op0->getType());
604 if (IT && IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
605 Metadata *LowAndHigh[] = {
607 ConstantAsMetadata::get(ConstantInt::get(IT, PossibleZeros + 1))};
608 II.setMetadata(LLVMContext::MD_range,
610 return &II;
611 }
612
613 return nullptr;
614}
615
617 assert(II.getIntrinsicID() == Intrinsic::ctpop &&
618 "Expected ctpop intrinsic");
619 Type *Ty = II.getType();
620 unsigned BitWidth = Ty->getScalarSizeInBits();
621 Value *Op0 = II.getArgOperand(0);
622 Value *X, *Y;
623
624 // ctpop(bitreverse(x)) -> ctpop(x)
625 // ctpop(bswap(x)) -> ctpop(x)
626 if (match(Op0, m_BitReverse(m_Value(X))) || match(Op0, m_BSwap(m_Value(X))))
627 return IC.replaceOperand(II, 0, X);
628
629 // ctpop(rot(x)) -> ctpop(x)
630 if ((match(Op0, m_FShl(m_Value(X), m_Value(Y), m_Value())) ||
631 match(Op0, m_FShr(m_Value(X), m_Value(Y), m_Value()))) &&
632 X == Y)
633 return IC.replaceOperand(II, 0, X);
634
635 // ctpop(x | -x) -> bitwidth - cttz(x, false)
636 if (Op0->hasOneUse() &&
637 match(Op0, m_c_Or(m_Value(X), m_Neg(m_Deferred(X))))) {
638 Function *F =
639 Intrinsic::getDeclaration(II.getModule(), Intrinsic::cttz, Ty);
640 auto *Cttz = IC.Builder.CreateCall(F, {X, IC.Builder.getFalse()});
641 auto *Bw = ConstantInt::get(Ty, APInt(BitWidth, BitWidth));
642 return IC.replaceInstUsesWith(II, IC.Builder.CreateSub(Bw, Cttz));
643 }
644
645 // ctpop(~x & (x - 1)) -> cttz(x, false)
646 if (match(Op0,
648 Function *F =
649 Intrinsic::getDeclaration(II.getModule(), Intrinsic::cttz, Ty);
650 return CallInst::Create(F, {X, IC.Builder.getFalse()});
651 }
652
653 // Zext doesn't change the number of set bits, so narrow:
654 // ctpop (zext X) --> zext (ctpop X)
655 if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
656 Value *NarrowPop = IC.Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, X);
657 return CastInst::Create(Instruction::ZExt, NarrowPop, Ty);
658 }
659
660 // If the operand is a select with constant arm(s), try to hoist ctpop.
661 if (auto *Sel = dyn_cast<SelectInst>(Op0))
662 if (Instruction *R = IC.FoldOpIntoSelect(II, Sel))
663 return R;
664
665 KnownBits Known(BitWidth);
666 IC.computeKnownBits(Op0, Known, 0, &II);
667
668 // If all bits are zero except for exactly one fixed bit, then the result
669 // must be 0 or 1, and we can get that answer by shifting to LSB:
670 // ctpop (X & 32) --> (X & 32) >> 5
671 // TODO: Investigate removing this as its likely unnecessary given the below
672 // `isKnownToBeAPowerOfTwo` check.
673 if ((~Known.Zero).isPowerOf2())
674 return BinaryOperator::CreateLShr(
675 Op0, ConstantInt::get(Ty, (~Known.Zero).exactLogBase2()));
676
677 // More generally we can also handle non-constant power of 2 patterns such as
678 // shl/shr(Pow2, X), (X & -X), etc... by transforming:
679 // ctpop(Pow2OrZero) --> icmp ne X, 0
680 if (IC.isKnownToBeAPowerOfTwo(Op0, /* OrZero */ true))
681 return CastInst::Create(Instruction::ZExt,
684 Ty);
685
686 // FIXME: Try to simplify vectors of integers.
687 auto *IT = dyn_cast<IntegerType>(Ty);
688 if (!IT)
689 return nullptr;
690
691 // Add range metadata since known bits can't completely reflect what we know.
692 unsigned MinCount = Known.countMinPopulation();
693 unsigned MaxCount = Known.countMaxPopulation();
694 if (IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
695 Metadata *LowAndHigh[] = {
698 II.setMetadata(LLVMContext::MD_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.
711 InstCombiner::BuilderTy &Builder) {
712 // Bail out if the mask is not a constant.
713 auto *C = dyn_cast<Constant>(II.getArgOperand(1));
714 if (!C)
715 return nullptr;
716
717 auto *VecTy = cast<FixedVectorType>(II.getType());
718 unsigned NumElts = VecTy->getNumElements();
719
720 // Only perform this transformation for <8 x i8> vector types.
721 if (!VecTy->getElementType()->isIntegerTy(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(I);
728
729 if (!COp || !isa<ConstantInt>(COp))
730 return nullptr;
731
732 Indexes[I] = cast<ConstantInt>(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(0);
740 auto *V2 = Constant::getNullValue(V1->getType());
741 return Builder.CreateShuffleVector(V1, V2, 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
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>(&*BI)) {
774 if (I->isDebugOrPseudoInst() ||
775 I->getIntrinsicID() == EndI.getIntrinsicID())
776 continue;
777 if (IsStart(*I)) {
778 if (haveSameOperands(EndI, *I, EndI.arg_size())) {
780 IC.eraseInstFromFunction(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
794 removeTriviallyEmptyRange(I, *this, [](const IntrinsicInst &I) {
795 return I.getIntrinsicID() == Intrinsic::vastart ||
796 I.getIntrinsicID() == Intrinsic::vacopy;
797 });
798 return nullptr;
799}
800
802 assert(Call.arg_size() > 1 && "Need at least 2 args to swap");
803 Value *Arg0 = Call.getArgOperand(0), *Arg1 = Call.getArgOperand(1);
804 if (isa<Constant>(Arg0) && !isa<Constant>(Arg1)) {
805 Call.setArgOperand(0, Arg1);
806 Call.setArgOperand(1, 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.
815 Constant *Overflow) {
816 Constant *V[] = {PoisonValue::get(Result->getType()), Overflow};
817 StructType *ST = cast<StructType>(II->getType());
819 return InsertValueInst::Create(Struct, Result, 0);
820}
821
823InstCombinerImpl::foldIntrinsicWithOverflowCommon(IntrinsicInst *II) {
824 WithOverflowInst *WO = cast<WithOverflowInst>(II);
825 Value *OperationResult = nullptr;
826 Constant *OverflowResult = nullptr;
827 if (OptimizeOverflowCheck(WO->getBinaryOp(), WO->isSigned(), WO->getLHS(),
828 WO->getRHS(), *WO, OperationResult, OverflowResult))
829 return createOverflowTuple(WO, OperationResult, OverflowResult);
830 return nullptr;
831}
832
833static std::optional<bool> getKnownSign(Value *Op, Instruction *CxtI,
834 const DataLayout &DL,
835 AssumptionCache *AC,
836 DominatorTree *DT) {
837 KnownBits Known = computeKnownBits(Op, DL, 0, AC, CxtI, DT);
838 if (Known.isNonNegative())
839 return false;
840 if (Known.isNegative())
841 return true;
842
843 Value *X, *Y;
844 if (match(Op, m_NSWSub(m_Value(X), m_Value(Y))))
846
848 ICmpInst::ICMP_SLT, Op, Constant::getNullValue(Op->getType()), CxtI, DL);
849}
850
851/// Try to canonicalize min/max(X + C0, C1) as min/max(X, C1 - C0) + C0. This
852/// can trigger other combines.
854 InstCombiner::BuilderTy &Builder) {
855 Intrinsic::ID MinMaxID = II->getIntrinsicID();
856 assert((MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin ||
857 MinMaxID == Intrinsic::umax || MinMaxID == Intrinsic::umin) &&
858 "Expected a min or max intrinsic");
859
860 // TODO: Match vectors with undef elements, but undef may not propagate.
861 Value *Op0 = II->getArgOperand(0), *Op1 = II->getArgOperand(1);
862 Value *X;
863 const APInt *C0, *C1;
864 if (!match(Op0, m_OneUse(m_Add(m_Value(X), m_APInt(C0)))) ||
865 !match(Op1, m_APInt(C1)))
866 return nullptr;
867
868 // Check for necessary no-wrap and overflow constraints.
869 bool IsSigned = MinMaxID == Intrinsic::smax || MinMaxID == Intrinsic::smin;
870 auto *Add = cast<BinaryOperator>(Op0);
871 if ((IsSigned && !Add->hasNoSignedWrap()) ||
872 (!IsSigned && !Add->hasNoUnsignedWrap()))
873 return nullptr;
874
875 // If the constant difference overflows, then instsimplify should reduce the
876 // min/max to the add or C1.
877 bool Overflow;
878 APInt CDiff =
879 IsSigned ? C1->ssub_ov(*C0, Overflow) : C1->usub_ov(*C0, Overflow);
880 assert(!Overflow && "Expected simplify of min/max");
881
882 // min/max (add X, C0), C1 --> add (min/max X, C1 - C0), C0
883 // Note: the "mismatched" no-overflow setting does not propagate.
884 Constant *NewMinMaxC = ConstantInt::get(II->getType(), CDiff);
885 Value *NewMinMax = Builder.CreateBinaryIntrinsic(MinMaxID, X, NewMinMaxC);
886 return IsSigned ? BinaryOperator::CreateNSWAdd(NewMinMax, Add->getOperand(1))
887 : BinaryOperator::CreateNUWAdd(NewMinMax, Add->getOperand(1));
888}
889/// Match a sadd_sat or ssub_sat which is using min/max to clamp the value.
890Instruction *InstCombinerImpl::matchSAddSubSat(IntrinsicInst &MinMax1) {
891 Type *Ty = MinMax1.getType();
892
893 // We are looking for a tree of:
894 // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B))))
895 // Where the min and max could be reversed
896 Instruction *MinMax2;
898 const APInt *MinValue, *MaxValue;
899 if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) {
900 if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue))))
901 return nullptr;
902 } else if (match(&MinMax1,
903 m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) {
904 if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue))))
905 return nullptr;
906 } else
907 return nullptr;
908
909 // Check that the constants clamp a saturate, and that the new type would be
910 // sensible to convert to.
911 if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1)
912 return nullptr;
913 // In what bitwidth can this be treated as saturating arithmetics?
914 unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1;
915 // FIXME: This isn't quite right for vectors, but using the scalar type is a
916 // good first approximation for what should be done there.
917 if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth))
918 return nullptr;
919
920 // Also make sure that the inner min/max and the add/sub have one use.
921 if (!MinMax2->hasOneUse() || !AddSub->hasOneUse())
922 return nullptr;
923
924 // Create the new type (which can be a vector type)
925 Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth);
926
927 Intrinsic::ID IntrinsicID;
928 if (AddSub->getOpcode() == Instruction::Add)
929 IntrinsicID = Intrinsic::sadd_sat;
930 else if (AddSub->getOpcode() == Instruction::Sub)
931 IntrinsicID = Intrinsic::ssub_sat;
932 else
933 return nullptr;
934
935 // The two operands of the add/sub must be nsw-truncatable to the NewTy. This
936 // is usually achieved via a sext from a smaller type.
937 if (ComputeMaxSignificantBits(AddSub->getOperand(0), 0, AddSub) >
938 NewBitWidth ||
939 ComputeMaxSignificantBits(AddSub->getOperand(1), 0, AddSub) > NewBitWidth)
940 return nullptr;
941
942 // Finally create and return the sat intrinsic, truncated to the new type
943 Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy);
944 Value *AT = Builder.CreateTrunc(AddSub->getOperand(0), NewTy);
945 Value *BT = Builder.CreateTrunc(AddSub->getOperand(1), NewTy);
946 Value *Sat = Builder.CreateCall(F, {AT, BT});
947 return CastInst::Create(Instruction::SExt, Sat, Ty);
948}
949
950
951/// If we have a clamp pattern like max (min X, 42), 41 -- where the output
952/// can only be one of two possible constant values -- turn that into a select
953/// of constants.
955 InstCombiner::BuilderTy &Builder) {
956 Value *I0 = II->getArgOperand(0), *I1 = II->getArgOperand(1);
957 Value *X;
958 const APInt *C0, *C1;
959 if (!match(I1, m_APInt(C1)) || !I0->hasOneUse())
960 return nullptr;
961
963 switch (II->getIntrinsicID()) {
964 case Intrinsic::smax:
965 if (match(I0, m_SMin(m_Value(X), m_APInt(C0))) && *C0 == *C1 + 1)
966 Pred = ICmpInst::ICMP_SGT;
967 break;
968 case Intrinsic::smin:
969 if (match(I0, m_SMax(m_Value(X), m_APInt(C0))) && *C1 == *C0 + 1)
970 Pred = ICmpInst::ICMP_SLT;
971 break;
972 case Intrinsic::umax:
973 if (match(I0, m_UMin(m_Value(X), m_APInt(C0))) && *C0 == *C1 + 1)
974 Pred = ICmpInst::ICMP_UGT;
975 break;
976 case Intrinsic::umin:
977 if (match(I0, m_UMax(m_Value(X), m_APInt(C0))) && *C1 == *C0 + 1)
978 Pred = ICmpInst::ICMP_ULT;
979 break;
980 default:
981 llvm_unreachable("Expected min/max intrinsic");
982 }
983 if (Pred == CmpInst::BAD_ICMP_PREDICATE)
984 return nullptr;
985
986 // max (min X, 42), 41 --> X > 41 ? 42 : 41
987 // min (max X, 42), 43 --> X < 43 ? 42 : 43
988 Value *Cmp = Builder.CreateICmp(Pred, X, I1);
989 return SelectInst::Create(Cmp, ConstantInt::get(II->getType(), *C0), I1);
990}
991
992/// If this min/max has a constant operand and an operand that is a matching
993/// min/max with a constant operand, constant-fold the 2 constant operands.
995 Intrinsic::ID MinMaxID = II->getIntrinsicID();
996 auto *LHS = dyn_cast<IntrinsicInst>(II->getArgOperand(0));
997 if (!LHS || LHS->getIntrinsicID() != MinMaxID)
998 return nullptr;
999
1000 Constant *C0, *C1;
1001 if (!match(LHS->getArgOperand(1), m_ImmConstant(C0)) ||
1002 !match(II->getArgOperand(1), m_ImmConstant(C1)))
1003 return nullptr;
1004
1005 // max (max X, C0), C1 --> max X, (max C0, C1) --> max X, NewC
1007 Constant *CondC = ConstantExpr::getICmp(Pred, C0, C1);
1008 Constant *NewC = ConstantExpr::getSelect(CondC, C0, C1);
1009
1010 Module *Mod = II->getModule();
1012 return CallInst::Create(MinMax, {LHS->getArgOperand(0), NewC});
1013}
1014
1015/// If this min/max has a matching min/max operand with a constant, try to push
1016/// the constant operand into this instruction. This can enable more folds.
1017static Instruction *
1019 InstCombiner::BuilderTy &Builder) {
1020 // Match and capture a min/max operand candidate.
1021 Value *X, *Y;
1022 Constant *C;
1023 Instruction *Inner;
1025 m_Instruction(Inner),
1027 m_Value(Y))))
1028 return nullptr;
1029
1030 // The inner op must match. Check for constants to avoid infinite loops.
1031 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1032 auto *InnerMM = dyn_cast<IntrinsicInst>(Inner);
1033 if (!InnerMM || InnerMM->getIntrinsicID() != MinMaxID ||
1035 return nullptr;
1036
1037 // max (max X, C), Y --> max (max X, Y), C
1038 Function *MinMax =
1039 Intrinsic::getDeclaration(II->getModule(), MinMaxID, II->getType());
1040 Value *NewInner = Builder.CreateBinaryIntrinsic(MinMaxID, X, Y);
1041 NewInner->takeName(Inner);
1042 return CallInst::Create(MinMax, {NewInner, C});
1043}
1044
1045/// Reduce a sequence of min/max intrinsics with a common operand.
1047 // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
1048 auto *LHS = dyn_cast<IntrinsicInst>(II->getArgOperand(0));
1049 auto *RHS = dyn_cast<IntrinsicInst>(II->getArgOperand(1));
1050 Intrinsic::ID MinMaxID = II->getIntrinsicID();
1051 if (!LHS || !RHS || LHS->getIntrinsicID() != MinMaxID ||
1052 RHS->getIntrinsicID() != MinMaxID ||
1053 (!LHS->hasOneUse() && !RHS->hasOneUse()))
1054 return nullptr;
1055
1056 Value *A = LHS->getArgOperand(0);
1057 Value *B = LHS->getArgOperand(1);
1058 Value *C = RHS->getArgOperand(0);
1059 Value *D = RHS->getArgOperand(1);
1060
1061 // Look for a common operand.
1062 Value *MinMaxOp = nullptr;
1063 Value *ThirdOp = nullptr;
1064 if (LHS->hasOneUse()) {
1065 // If the LHS is only used in this chain and the RHS is used outside of it,
1066 // reuse the RHS min/max because that will eliminate the LHS.
1067 if (D == A || C == A) {
1068 // min(min(a, b), min(c, a)) --> min(min(c, a), b)
1069 // min(min(a, b), min(a, d)) --> min(min(a, d), b)
1070 MinMaxOp = RHS;
1071 ThirdOp = B;
1072 } else if (D == B || C == B) {
1073 // min(min(a, b), min(c, b)) --> min(min(c, b), a)
1074 // min(min(a, b), min(b, d)) --> min(min(b, d), a)
1075 MinMaxOp = RHS;
1076 ThirdOp = A;
1077 }
1078 } else {
1079 assert(RHS->hasOneUse() && "Expected one-use operand");
1080 // Reuse the LHS. This will eliminate the RHS.
1081 if (D == A || D == B) {
1082 // min(min(a, b), min(c, a)) --> min(min(a, b), c)
1083 // min(min(a, b), min(c, b)) --> min(min(a, b), c)
1084 MinMaxOp = LHS;
1085 ThirdOp = C;
1086 } else if (C == A || C == B) {
1087 // min(min(a, b), min(b, d)) --> min(min(a, b), d)
1088 // min(min(a, b), min(c, b)) --> min(min(a, b), d)
1089 MinMaxOp = LHS;
1090 ThirdOp = D;
1091 }
1092 }
1093
1094 if (!MinMaxOp || !ThirdOp)
1095 return nullptr;
1096
1097 Module *Mod = II->getModule();
1099 return CallInst::Create(MinMax, { MinMaxOp, ThirdOp });
1100}
1101
1102/// If all arguments of the intrinsic are unary shuffles with the same mask,
1103/// try to shuffle after the intrinsic.
1104static Instruction *
1106 InstCombiner::BuilderTy &Builder) {
1107 // TODO: This should be extended to handle other intrinsics like fshl, ctpop,
1108 // etc. Use llvm::isTriviallyVectorizable() and related to determine
1109 // which intrinsics are safe to shuffle?
1110 switch (II->getIntrinsicID()) {
1111 case Intrinsic::smax:
1112 case Intrinsic::smin:
1113 case Intrinsic::umax:
1114 case Intrinsic::umin:
1115 case Intrinsic::fma:
1116 case Intrinsic::fshl:
1117 case Intrinsic::fshr:
1118 break;
1119 default:
1120 return nullptr;
1121 }
1122
1123 Value *X;
1124 ArrayRef<int> Mask;
1125 if (!match(II->getArgOperand(0),
1126 m_Shuffle(m_Value(X), m_Undef(), m_Mask(Mask))))
1127 return nullptr;
1128
1129 // At least 1 operand must have 1 use because we are creating 2 instructions.
1130 if (none_of(II->args(), [](Value *V) { return V->hasOneUse(); }))
1131 return nullptr;
1132
1133 // See if all arguments are shuffled with the same mask.
1134 SmallVector<Value *, 4> NewArgs(II->arg_size());
1135 NewArgs[0] = X;
1136 Type *SrcTy = X->getType();
1137 for (unsigned i = 1, e = II->arg_size(); i != e; ++i) {
1138 if (!match(II->getArgOperand(i),
1139 m_Shuffle(m_Value(X), m_Undef(), m_SpecificMask(Mask))) ||
1140 X->getType() != SrcTy)
1141 return nullptr;
1142 NewArgs[i] = X;
1143 }
1144
1145 // intrinsic (shuf X, M), (shuf Y, M), ... --> shuf (intrinsic X, Y, ...), M
1146 Instruction *FPI = isa<FPMathOperator>(II) ? II : nullptr;
1147 Value *NewIntrinsic =
1148 Builder.CreateIntrinsic(II->getIntrinsicID(), SrcTy, NewArgs, FPI);
1149 return new ShuffleVectorInst(NewIntrinsic, Mask);
1150}
1151
1152/// CallInst simplification. This mostly only handles folding of intrinsic
1153/// instructions. For normal calls, it allows visitCallBase to do the heavy
1154/// lifting.
1156 // Don't try to simplify calls without uses. It will not do anything useful,
1157 // but will result in the following folds being skipped.
1158 if (!CI.use_empty())
1159 if (Value *V = simplifyCall(&CI, SQ.getWithInstruction(&CI)))
1160 return replaceInstUsesWith(CI, V);
1161
1162 if (Value *FreedOp = getFreedOperand(&CI, &TLI))
1163 return visitFree(CI, FreedOp);
1164
1165 // If the caller function (i.e. us, the function that contains this CallInst)
1166 // is nounwind, mark the call as nounwind, even if the callee isn't.
1167 if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) {
1168 CI.setDoesNotThrow();
1169 return &CI;
1170 }
1171
1172 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
1173 if (!II) return visitCallBase(CI);
1174
1175 // For atomic unordered mem intrinsics if len is not a positive or
1176 // not a multiple of element size then behavior is undefined.
1177 if (auto *AMI = dyn_cast<AtomicMemIntrinsic>(II))
1178 if (ConstantInt *NumBytes = dyn_cast<ConstantInt>(AMI->getLength()))
1179 if (NumBytes->getSExtValue() < 0 ||
1180 (NumBytes->getZExtValue() % AMI->getElementSizeInBytes() != 0)) {
1182 assert(AMI->getType()->isVoidTy() &&
1183 "non void atomic unordered mem intrinsic");
1184 return eraseInstFromFunction(*AMI);
1185 }
1186
1187 // Intrinsics cannot occur in an invoke or a callbr, so handle them here
1188 // instead of in visitCallBase.
1189 if (auto *MI = dyn_cast<AnyMemIntrinsic>(II)) {
1190 bool Changed = false;
1191
1192 // memmove/cpy/set of zero bytes is a noop.
1193 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
1194 if (NumBytes->isNullValue())
1195 return eraseInstFromFunction(CI);
1196 }
1197
1198 // No other transformations apply to volatile transfers.
1199 if (auto *M = dyn_cast<MemIntrinsic>(MI))
1200 if (M->isVolatile())
1201 return nullptr;
1202
1203 // If we have a memmove and the source operation is a constant global,
1204 // then the source and dest pointers can't alias, so we can change this
1205 // into a call to memcpy.
1206 if (auto *MMI = dyn_cast<AnyMemMoveInst>(MI)) {
1207 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
1208 if (GVSrc->isConstant()) {
1209 Module *M = CI.getModule();
1210 Intrinsic::ID MemCpyID =
1211 isa<AtomicMemMoveInst>(MMI)
1212 ? Intrinsic::memcpy_element_unordered_atomic
1213 : Intrinsic::memcpy;
1214 Type *Tys[3] = { CI.getArgOperand(0)->getType(),
1215 CI.getArgOperand(1)->getType(),
1216 CI.getArgOperand(2)->getType() };
1217 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
1218 Changed = true;
1219 }
1220 }
1221
1222 if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1223 // memmove(x,x,size) -> noop.
1224 if (MTI->getSource() == MTI->getDest())
1225 return eraseInstFromFunction(CI);
1226 }
1227
1228 // If we can determine a pointer alignment that is bigger than currently
1229 // set, update the alignment.
1230 if (auto *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1232 return I;
1233 } else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) {
1234 if (Instruction *I = SimplifyAnyMemSet(MSI))
1235 return I;
1236 }
1237
1238 if (Changed) return II;
1239 }
1240
1241 // For fixed width vector result intrinsics, use the generic demanded vector
1242 // support.
1243 if (auto *IIFVTy = dyn_cast<FixedVectorType>(II->getType())) {
1244 auto VWidth = IIFVTy->getNumElements();
1245 APInt UndefElts(VWidth, 0);
1246 APInt AllOnesEltMask(APInt::getAllOnes(VWidth));
1247 if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) {
1248 if (V != II)
1249 return replaceInstUsesWith(*II, V);
1250 return II;
1251 }
1252 }
1253
1254 if (II->isCommutative()) {
1255 if (CallInst *NewCall = canonicalizeConstantArg0ToArg1(CI))
1256 return NewCall;
1257 }
1258
1259 // Unused constrained FP intrinsic calls may have declared side effect, which
1260 // prevents it from being removed. In some cases however the side effect is
1261 // actually absent. To detect this case, call SimplifyConstrainedFPCall. If it
1262 // returns a replacement, the call may be removed.
1263 if (CI.use_empty() && isa<ConstrainedFPIntrinsic>(CI)) {
1265 return eraseInstFromFunction(CI);
1266 }
1267
1268 Intrinsic::ID IID = II->getIntrinsicID();
1269 switch (IID) {
1270 case Intrinsic::objectsize:
1271 if (Value *V = lowerObjectSizeCall(II, DL, &TLI, AA, /*MustSucceed=*/false))
1272 return replaceInstUsesWith(CI, V);
1273 return nullptr;
1274 case Intrinsic::abs: {
1275 Value *IIOperand = II->getArgOperand(0);
1276 bool IntMinIsPoison = cast<Constant>(II->getArgOperand(1))->isOneValue();
1277
1278 // abs(-x) -> abs(x)
1279 // TODO: Copy nsw if it was present on the neg?
1280 Value *X;
1281 if (match(IIOperand, m_Neg(m_Value(X))))
1282 return replaceOperand(*II, 0, X);
1283 if (match(IIOperand, m_Select(m_Value(), m_Value(X), m_Neg(m_Deferred(X)))))
1284 return replaceOperand(*II, 0, X);
1285 if (match(IIOperand, m_Select(m_Value(), m_Neg(m_Value(X)), m_Deferred(X))))
1286 return replaceOperand(*II, 0, X);
1287
1288 if (std::optional<bool> Sign = getKnownSign(IIOperand, II, DL, &AC, &DT)) {
1289 // abs(x) -> x if x >= 0
1290 if (!*Sign)
1291 return replaceInstUsesWith(*II, IIOperand);
1292
1293 // abs(x) -> -x if x < 0
1294 if (IntMinIsPoison)
1295 return BinaryOperator::CreateNSWNeg(IIOperand);
1296 return BinaryOperator::CreateNeg(IIOperand);
1297 }
1298
1299 // abs (sext X) --> zext (abs X*)
1300 // Clear the IsIntMin (nsw) bit on the abs to allow narrowing.
1301 if (match(IIOperand, m_OneUse(m_SExt(m_Value(X))))) {
1302 Value *NarrowAbs =
1303 Builder.CreateBinaryIntrinsic(Intrinsic::abs, X, Builder.getFalse());
1304 return CastInst::Create(Instruction::ZExt, NarrowAbs, II->getType());
1305 }
1306
1307 // Match a complicated way to check if a number is odd/even:
1308 // abs (srem X, 2) --> and X, 1
1309 const APInt *C;
1310 if (match(IIOperand, m_SRem(m_Value(X), m_APInt(C))) && *C == 2)
1311 return BinaryOperator::CreateAnd(X, ConstantInt::get(II->getType(), 1));
1312
1313 break;
1314 }
1315 case Intrinsic::umin: {
1316 Value *I0 = II->getArgOperand(0), *I1 = II->getArgOperand(1);
1317 // umin(x, 1) == zext(x != 0)
1318 if (match(I1, m_One())) {
1319 assert(II->getType()->getScalarSizeInBits() != 1 &&
1320 "Expected simplify of umin with max constant");
1321 Value *Zero = Constant::getNullValue(I0->getType());
1322 Value *Cmp = Builder.CreateICmpNE(I0, Zero);
1323 return CastInst::Create(Instruction::ZExt, Cmp, II->getType());
1324 }
1325 [[fallthrough]];
1326 }
1327 case Intrinsic::umax: {
1328 Value *I0 = II->getArgOperand(0), *I1 = II->getArgOperand(1);
1329 Value *X, *Y;
1330 if (match(I0, m_ZExt(m_Value(X))) && match(I1, m_ZExt(m_Value(Y))) &&
1331 (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) {
1332 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(IID, X, Y);
1333 return CastInst::Create(Instruction::ZExt, NarrowMaxMin, II->getType());
1334 }
1335 Constant *C;
1336 if (match(I0, m_ZExt(m_Value(X))) && match(I1, m_Constant(C)) &&
1337 I0->hasOneUse()) {
1338 Constant *NarrowC = ConstantExpr::getTrunc(C, X->getType());
1339 if (ConstantExpr::getZExt(NarrowC, II->getType()) == C) {
1340 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(IID, X, NarrowC);
1341 return CastInst::Create(Instruction::ZExt, NarrowMaxMin, II->getType());
1342 }
1343 }
1344 // If both operands of unsigned min/max are sign-extended, it is still ok
1345 // to narrow the operation.
1346 [[fallthrough]];
1347 }
1348 case Intrinsic::smax:
1349 case Intrinsic::smin: {
1350 Value *I0 = II->getArgOperand(0), *I1 = II->getArgOperand(1);
1351 Value *X, *Y;
1352 if (match(I0, m_SExt(m_Value(X))) && match(I1, m_SExt(m_Value(Y))) &&
1353 (I0->hasOneUse() || I1->hasOneUse()) && X->getType() == Y->getType()) {
1354 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(IID, X, Y);
1355 return CastInst::Create(Instruction::SExt, NarrowMaxMin, II->getType());
1356 }
1357
1358 Constant *C;
1359 if (match(I0, m_SExt(m_Value(X))) && match(I1, m_Constant(C)) &&
1360 I0->hasOneUse()) {
1361 Constant *NarrowC = ConstantExpr::getTrunc(C, X->getType());
1362 if (ConstantExpr::getSExt(NarrowC, II->getType()) == C) {
1363 Value *NarrowMaxMin = Builder.CreateBinaryIntrinsic(IID, X, NarrowC);
1364 return CastInst::Create(Instruction::SExt, NarrowMaxMin, II->getType());
1365 }
1366 }
1367
1368 if (IID == Intrinsic::smax || IID == Intrinsic::smin) {
1369 // smax (neg nsw X), (neg nsw Y) --> neg nsw (smin X, Y)
1370 // smin (neg nsw X), (neg nsw Y) --> neg nsw (smax X, Y)
1371 // TODO: Canonicalize neg after min/max if I1 is constant.
1372 if (match(I0, m_NSWNeg(m_Value(X))) && match(I1, m_NSWNeg(m_Value(Y))) &&
1373 (I0->hasOneUse() || I1->hasOneUse())) {
1375 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, X, Y);
1376 return BinaryOperator::CreateNSWNeg(InvMaxMin);
1377 }
1378 }
1379
1380 // If we can eliminate ~A and Y is free to invert:
1381 // max ~A, Y --> ~(min A, ~Y)
1382 //
1383 // Examples:
1384 // max ~A, ~Y --> ~(min A, Y)
1385 // max ~A, C --> ~(min A, ~C)
1386 // max ~A, (max ~Y, ~Z) --> ~min( A, (min Y, Z))
1387 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * {
1388 Value *A;
1389 if (match(X, m_OneUse(m_Not(m_Value(A)))) &&
1390 !isFreeToInvert(A, A->hasOneUse()) &&
1391 isFreeToInvert(Y, Y->hasOneUse())) {
1392 Value *NotY = Builder.CreateNot(Y);
1394 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, A, NotY);
1395 return BinaryOperator::CreateNot(InvMaxMin);
1396 }
1397 return nullptr;
1398 };
1399
1400 if (Instruction *I = moveNotAfterMinMax(I0, I1))
1401 return I;
1402 if (Instruction *I = moveNotAfterMinMax(I1, I0))
1403 return I;
1404
1406 return I;
1407
1408 // smax(X, -X) --> abs(X)
1409 // smin(X, -X) --> -abs(X)
1410 // umax(X, -X) --> -abs(X)
1411 // umin(X, -X) --> abs(X)
1412 if (isKnownNegation(I0, I1)) {
1413 // We can choose either operand as the input to abs(), but if we can
1414 // eliminate the only use of a value, that's better for subsequent
1415 // transforms/analysis.
1416 if (I0->hasOneUse() && !I1->hasOneUse())
1417 std::swap(I0, I1);
1418
1419 // This is some variant of abs(). See if we can propagate 'nsw' to the abs
1420 // operation and potentially its negation.
1421 bool IntMinIsPoison = isKnownNegation(I0, I1, /* NeedNSW */ true);
1423 Intrinsic::abs, I0,
1424 ConstantInt::getBool(II->getContext(), IntMinIsPoison));
1425
1426 // We don't have a "nabs" intrinsic, so negate if needed based on the
1427 // max/min operation.
1428 if (IID == Intrinsic::smin || IID == Intrinsic::umax)
1429 Abs = Builder.CreateNeg(Abs, "nabs", /* NUW */ false, IntMinIsPoison);
1430 return replaceInstUsesWith(CI, Abs);
1431 }
1432
1433 if (Instruction *Sel = foldClampRangeOfTwo(II, Builder))
1434 return Sel;
1435
1436 if (Instruction *SAdd = matchSAddSubSat(*II))
1437 return SAdd;
1438
1439 if (match(I1, m_ImmConstant()))
1440 if (auto *Sel = dyn_cast<SelectInst>(I0))
1441 if (Instruction *R = FoldOpIntoSelect(*II, Sel))
1442 return R;
1443
1444 if (Instruction *NewMinMax = reassociateMinMaxWithConstants(II))
1445 return NewMinMax;
1446
1448 return R;
1449
1450 if (Instruction *NewMinMax = factorizeMinMaxTree(II))
1451 return NewMinMax;
1452
1453 break;
1454 }
1455 case Intrinsic::bitreverse: {
1456 // bitrev (zext i1 X to ?) --> X ? SignBitC : 0
1457 Value *X;
1458 if (match(II->getArgOperand(0), m_ZExt(m_Value(X))) &&
1459 X->getType()->isIntOrIntVectorTy(1)) {
1460 Type *Ty = II->getType();
1462 return SelectInst::Create(X, ConstantInt::get(Ty, SignBit),
1464 }
1465 break;
1466 }
1467 case Intrinsic::bswap: {
1468 Value *IIOperand = II->getArgOperand(0);
1469
1470 // Try to canonicalize bswap-of-logical-shift-by-8-bit-multiple as
1471 // inverse-shift-of-bswap:
1472 // bswap (shl X, Y) --> lshr (bswap X), Y
1473 // bswap (lshr X, Y) --> shl (bswap X), Y
1474 Value *X, *Y;
1475 if (match(IIOperand, m_OneUse(m_LogicalShift(m_Value(X), m_Value(Y))))) {
1476 // The transform allows undef vector elements, so try a constant match
1477 // first. If knownbits can handle that case, that clause could be removed.
1478 unsigned BitWidth = IIOperand->getType()->getScalarSizeInBits();
1479 const APInt *C;
1480 if ((match(Y, m_APIntAllowUndef(C)) && (*C & 7) == 0) ||
1482 Value *NewSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X);
1483 BinaryOperator::BinaryOps InverseShift =
1484 cast<BinaryOperator>(IIOperand)->getOpcode() == Instruction::Shl
1485 ? Instruction::LShr
1486 : Instruction::Shl;
1487 return BinaryOperator::Create(InverseShift, NewSwap, Y);
1488 }
1489 }
1490
1491 KnownBits Known = computeKnownBits(IIOperand, 0, II);
1492 uint64_t LZ = alignDown(Known.countMinLeadingZeros(), 8);
1493 uint64_t TZ = alignDown(Known.countMinTrailingZeros(), 8);
1494 unsigned BW = Known.getBitWidth();
1495
1496 // bswap(x) -> shift(x) if x has exactly one "active byte"
1497 if (BW - LZ - TZ == 8) {
1498 assert(LZ != TZ && "active byte cannot be in the middle");
1499 if (LZ > TZ) // -> shl(x) if the "active byte" is in the low part of x
1500 return BinaryOperator::CreateNUWShl(
1501 IIOperand, ConstantInt::get(IIOperand->getType(), LZ - TZ));
1502 // -> lshr(x) if the "active byte" is in the high part of x
1503 return BinaryOperator::CreateExactLShr(
1504 IIOperand, ConstantInt::get(IIOperand->getType(), TZ - LZ));
1505 }
1506
1507 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
1508 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
1509 unsigned C = X->getType()->getScalarSizeInBits() - BW;
1510 Value *CV = ConstantInt::get(X->getType(), C);
1511 Value *V = Builder.CreateLShr(X, CV);
1512 return new TruncInst(V, IIOperand->getType());
1513 }
1514 break;
1515 }
1516 case Intrinsic::masked_load:
1517 if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(*II))
1518 return replaceInstUsesWith(CI, SimplifiedMaskedOp);
1519 break;
1520 case Intrinsic::masked_store:
1521 return simplifyMaskedStore(*II);
1522 case Intrinsic::masked_gather:
1523 return simplifyMaskedGather(*II);
1524 case Intrinsic::masked_scatter:
1525 return simplifyMaskedScatter(*II);
1526 case Intrinsic::launder_invariant_group:
1527 case Intrinsic::strip_invariant_group:
1528 if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(*II, *this))
1529 return replaceInstUsesWith(*II, SkippedBarrier);
1530 break;
1531 case Intrinsic::powi:
1532 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
1533 // 0 and 1 are handled in instsimplify
1534 // powi(x, -1) -> 1/x
1535 if (Power->isMinusOne())
1537 II->getArgOperand(0), II);
1538 // powi(x, 2) -> x*x
1539 if (Power->equalsInt(2))
1541 II->getArgOperand(0), II);
1542
1543 if (!Power->getValue()[0]) {
1544 Value *X;
1545 // If power is even:
1546 // powi(-x, p) -> powi(x, p)
1547 // powi(fabs(x), p) -> powi(x, p)
1548 // powi(copysign(x, y), p) -> powi(x, p)
1549 if (match(II->getArgOperand(0), m_FNeg(m_Value(X))) ||
1550 match(II->getArgOperand(0), m_FAbs(m_Value(X))) ||
1551 match(II->getArgOperand(0),
1552 m_Intrinsic<Intrinsic::copysign>(m_Value(X), m_Value())))
1553 return replaceOperand(*II, 0, X);
1554 }
1555 }
1556 break;
1557
1558 case Intrinsic::cttz:
1559 case Intrinsic::ctlz:
1560 if (auto *I = foldCttzCtlz(*II, *this))
1561 return I;
1562 break;
1563
1564 case Intrinsic::ctpop:
1565 if (auto *I = foldCtpop(*II, *this))
1566 return I;
1567 break;
1568
1569 case Intrinsic::fshl:
1570 case Intrinsic::fshr: {
1571 Value *Op0 = II->getArgOperand(0), *Op1 = II->getArgOperand(1);
1572 Type *Ty = II->getType();
1573 unsigned BitWidth = Ty->getScalarSizeInBits();
1574 Constant *ShAmtC;
1575 if (match(II->getArgOperand(2), m_ImmConstant(ShAmtC))) {
1576 // Canonicalize a shift amount constant operand to modulo the bit-width.
1577 Constant *WidthC = ConstantInt::get(Ty, BitWidth);
1578 Constant *ModuloC =
1579 ConstantFoldBinaryOpOperands(Instruction::URem, ShAmtC, WidthC, DL);
1580 if (!ModuloC)
1581 return nullptr;
1582 if (ModuloC != ShAmtC)
1583 return replaceOperand(*II, 2, ModuloC);
1584
1587 "Shift amount expected to be modulo bitwidth");
1588
1589 // Canonicalize funnel shift right by constant to funnel shift left. This
1590 // is not entirely arbitrary. For historical reasons, the backend may
1591 // recognize rotate left patterns but miss rotate right patterns.
1592 if (IID == Intrinsic::fshr) {
1593 // fshr X, Y, C --> fshl X, Y, (BitWidth - C)
1594 Constant *LeftShiftC = ConstantExpr::getSub(WidthC, ShAmtC);
1595 Module *Mod = II->getModule();
1596 Function *Fshl = Intrinsic::getDeclaration(Mod, Intrinsic::fshl, Ty);
1597 return CallInst::Create(Fshl, { Op0, Op1, LeftShiftC });
1598 }
1599 assert(IID == Intrinsic::fshl &&
1600 "All funnel shifts by simple constants should go left");
1601
1602 // fshl(X, 0, C) --> shl X, C
1603 // fshl(X, undef, C) --> shl X, C
1604 if (match(Op1, m_ZeroInt()) || match(Op1, m_Undef()))
1605 return BinaryOperator::CreateShl(Op0, ShAmtC);
1606
1607 // fshl(0, X, C) --> lshr X, (BW-C)
1608 // fshl(undef, X, C) --> lshr X, (BW-C)
1609 if (match(Op0, m_ZeroInt()) || match(Op0, m_Undef()))
1610 return BinaryOperator::CreateLShr(Op1,
1611 ConstantExpr::getSub(WidthC, ShAmtC));
1612
1613 // fshl i16 X, X, 8 --> bswap i16 X (reduce to more-specific form)
1614 if (Op0 == Op1 && BitWidth == 16 && match(ShAmtC, m_SpecificInt(8))) {
1615 Module *Mod = II->getModule();
1616 Function *Bswap = Intrinsic::getDeclaration(Mod, Intrinsic::bswap, Ty);
1617 return CallInst::Create(Bswap, { Op0 });
1618 }
1619 }
1620
1621 // Left or right might be masked.
1623 return &CI;
1624
1625 // The shift amount (operand 2) of a funnel shift is modulo the bitwidth,
1626 // so only the low bits of the shift amount are demanded if the bitwidth is
1627 // a power-of-2.
1628 if (!isPowerOf2_32(BitWidth))
1629 break;
1631 KnownBits Op2Known(BitWidth);
1632 if (SimplifyDemandedBits(II, 2, Op2Demanded, Op2Known))
1633 return &CI;
1634 break;
1635 }
1636 case Intrinsic::uadd_with_overflow:
1637 case Intrinsic::sadd_with_overflow: {
1638 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
1639 return I;
1640
1641 // Given 2 constant operands whose sum does not overflow:
1642 // uaddo (X +nuw C0), C1 -> uaddo X, C0 + C1
1643 // saddo (X +nsw C0), C1 -> saddo X, C0 + C1
1644 Value *X;
1645 const APInt *C0, *C1;
1646 Value *Arg0 = II->getArgOperand(0);
1647 Value *Arg1 = II->getArgOperand(1);
1648 bool IsSigned = IID == Intrinsic::sadd_with_overflow;
1649 bool HasNWAdd = IsSigned ? match(Arg0, m_NSWAdd(m_Value(X), m_APInt(C0)))
1650 : match(Arg0, m_NUWAdd(m_Value(X), m_APInt(C0)));
1651 if (HasNWAdd && match(Arg1, m_APInt(C1))) {
1652 bool Overflow;
1653 APInt NewC =
1654 IsSigned ? C1->sadd_ov(*C0, Overflow) : C1->uadd_ov(*C0, Overflow);
1655 if (!Overflow)
1656 return replaceInstUsesWith(
1658 IID, X, ConstantInt::get(Arg1->getType(), NewC)));
1659 }
1660 break;
1661 }
1662
1663 case Intrinsic::umul_with_overflow:
1664 case Intrinsic::smul_with_overflow:
1665 case Intrinsic::usub_with_overflow:
1666 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
1667 return I;
1668 break;
1669
1670 case Intrinsic::ssub_with_overflow: {
1671 if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
1672 return I;
1673
1674 Constant *C;
1675 Value *Arg0 = II->getArgOperand(0);
1676 Value *Arg1 = II->getArgOperand(1);
1677 // Given a constant C that is not the minimum signed value
1678 // for an integer of a given bit width:
1679 //
1680 // ssubo X, C -> saddo X, -C
1681 if (match(Arg1, m_Constant(C)) && C->isNotMinSignedValue()) {
1682 Value *NegVal = ConstantExpr::getNeg(C);
1683 // Build a saddo call that is equivalent to the discovered
1684 // ssubo call.
1685 return replaceInstUsesWith(
1686 *II, Builder.CreateBinaryIntrinsic(Intrinsic::sadd_with_overflow,
1687 Arg0, NegVal));
1688 }
1689
1690 break;
1691 }
1692
1693 case Intrinsic::uadd_sat:
1694 case Intrinsic::sadd_sat:
1695 case Intrinsic::usub_sat:
1696 case Intrinsic::ssub_sat: {
1697 SaturatingInst *SI = cast<SaturatingInst>(II);
1698 Type *Ty = SI->getType();
1699 Value *Arg0 = SI->getLHS();
1700 Value *Arg1 = SI->getRHS();
1701
1702 // Make use of known overflow information.
1703 OverflowResult OR = computeOverflow(SI->getBinaryOp(), SI->isSigned(),
1704 Arg0, Arg1, SI);
1705 switch (OR) {
1707 break;
1709 if (SI->isSigned())
1710 return BinaryOperator::CreateNSW(SI->getBinaryOp(), Arg0, Arg1);
1711 else
1712 return BinaryOperator::CreateNUW(SI->getBinaryOp(), Arg0, Arg1);
1714 unsigned BitWidth = Ty->getScalarSizeInBits();
1715 APInt Min = APSInt::getMinValue(BitWidth, !SI->isSigned());
1716 return replaceInstUsesWith(*SI, ConstantInt::get(Ty, Min));
1717 }
1719 unsigned BitWidth = Ty->getScalarSizeInBits();
1720 APInt Max = APSInt::getMaxValue(BitWidth, !SI->isSigned());
1721 return replaceInstUsesWith(*SI, ConstantInt::get(Ty, Max));
1722 }
1723 }
1724
1725 // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN
1726 Constant *C;
1727 if (IID == Intrinsic::ssub_sat && match(Arg1, m_Constant(C)) &&
1728 C->isNotMinSignedValue()) {
1729 Value *NegVal = ConstantExpr::getNeg(C);
1730 return replaceInstUsesWith(
1732 Intrinsic::sadd_sat, Arg0, NegVal));
1733 }
1734
1735 // sat(sat(X + Val2) + Val) -> sat(X + (Val+Val2))
1736 // sat(sat(X - Val2) - Val) -> sat(X - (Val+Val2))
1737 // if Val and Val2 have the same sign
1738 if (auto *Other = dyn_cast<IntrinsicInst>(Arg0)) {
1739 Value *X;
1740 const APInt *Val, *Val2;
1741 APInt NewVal;
1742 bool IsUnsigned =
1743 IID == Intrinsic::uadd_sat || IID == Intrinsic::usub_sat;
1744 if (Other->getIntrinsicID() == IID &&
1745 match(Arg1, m_APInt(Val)) &&
1746 match(Other->getArgOperand(0), m_Value(X)) &&
1747 match(Other->getArgOperand(1), m_APInt(Val2))) {
1748 if (IsUnsigned)
1749 NewVal = Val->uadd_sat(*Val2);
1750 else if (Val->isNonNegative() == Val2->isNonNegative()) {
1751 bool Overflow;
1752 NewVal = Val->sadd_ov(*Val2, Overflow);
1753 if (Overflow) {
1754 // Both adds together may add more than SignedMaxValue
1755 // without saturating the final result.
1756 break;
1757 }
1758 } else {
1759 // Cannot fold saturated addition with different signs.
1760 break;
1761 }
1762
1763 return replaceInstUsesWith(
1765 IID, X, ConstantInt::get(II->getType(), NewVal)));
1766 }
1767 }
1768 break;
1769 }
1770
1771 case Intrinsic::minnum:
1772 case Intrinsic::maxnum:
1773 case Intrinsic::minimum:
1774 case Intrinsic::maximum: {
1775 Value *Arg0 = II->getArgOperand(0);
1776 Value *Arg1 = II->getArgOperand(1);
1777 Value *X, *Y;
1778 if (match(Arg0, m_FNeg(m_Value(X))) && match(Arg1, m_FNeg(m_Value(Y))) &&
1779 (Arg0->hasOneUse() || Arg1->hasOneUse())) {
1780 // If both operands are negated, invert the call and negate the result:
1781 // min(-X, -Y) --> -(max(X, Y))
1782 // max(-X, -Y) --> -(min(X, Y))
1783 Intrinsic::ID NewIID;
1784 switch (IID) {
1785 case Intrinsic::maxnum:
1786 NewIID = Intrinsic::minnum;
1787 break;
1788 case Intrinsic::minnum:
1789 NewIID = Intrinsic::maxnum;
1790 break;
1791 case Intrinsic::maximum:
1792 NewIID = Intrinsic::minimum;
1793 break;
1794 case Intrinsic::minimum:
1795 NewIID = Intrinsic::maximum;
1796 break;
1797 default:
1798 llvm_unreachable("unexpected intrinsic ID");
1799 }
1800 Value *NewCall = Builder.CreateBinaryIntrinsic(NewIID, X, Y, II);
1801 Instruction *FNeg = UnaryOperator::CreateFNeg(NewCall);
1802 FNeg->copyIRFlags(II);
1803 return FNeg;
1804 }
1805
1806 // m(m(X, C2), C1) -> m(X, C)
1807 const APFloat *C1, *C2;
1808 if (auto *M = dyn_cast<IntrinsicInst>(Arg0)) {
1809 if (M->getIntrinsicID() == IID && match(Arg1, m_APFloat(C1)) &&
1810 ((match(M->getArgOperand(0), m_Value(X)) &&
1811 match(M->getArgOperand(1), m_APFloat(C2))) ||
1812 (match(M->getArgOperand(1), m_Value(X)) &&
1813 match(M->getArgOperand(0), m_APFloat(C2))))) {
1814 APFloat Res(0.0);
1815 switch (IID) {
1816 case Intrinsic::maxnum:
1817 Res = maxnum(*C1, *C2);
1818 break;
1819 case Intrinsic::minnum:
1820 Res = minnum(*C1, *C2);
1821 break;
1822 case Intrinsic::maximum:
1823 Res = maximum(*C1, *C2);
1824 break;
1825 case Intrinsic::minimum:
1826 Res = minimum(*C1, *C2);
1827 break;
1828 default:
1829 llvm_unreachable("unexpected intrinsic ID");
1830 }
1832 IID, X, ConstantFP::get(Arg0->getType(), Res), II);
1833 // TODO: Conservatively intersecting FMF. If Res == C2, the transform
1834 // was a simplification (so Arg0 and its original flags could
1835 // propagate?)
1836 NewCall->andIRFlags(M);
1837 return replaceInstUsesWith(*II, NewCall);
1838 }
1839 }
1840
1841 // m((fpext X), (fpext Y)) -> fpext (m(X, Y))
1842 if (match(Arg0, m_OneUse(m_FPExt(m_Value(X)))) &&
1843 match(Arg1, m_OneUse(m_FPExt(m_Value(Y)))) &&
1844 X->getType() == Y->getType()) {
1845 Value *NewCall =
1846 Builder.CreateBinaryIntrinsic(IID, X, Y, II, II->getName());
1847 return new FPExtInst(NewCall, II->getType());
1848 }
1849
1850 // max X, -X --> fabs X
1851 // min X, -X --> -(fabs X)
1852 // TODO: Remove one-use limitation? That is obviously better for max.
1853 // It would be an extra instruction for min (fnabs), but that is
1854 // still likely better for analysis and codegen.
1855 if ((match(Arg0, m_OneUse(m_FNeg(m_Value(X)))) && Arg1 == X) ||
1856 (match(Arg1, m_OneUse(m_FNeg(m_Value(X)))) && Arg0 == X)) {
1857 Value *R = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, II);
1858 if (IID == Intrinsic::minimum || IID == Intrinsic::minnum)
1859 R = Builder.CreateFNegFMF(R, II);
1860 return replaceInstUsesWith(*II, R);
1861 }
1862
1863 break;
1864 }
1865 case Intrinsic::matrix_multiply: {
1866 // Optimize negation in matrix multiplication.
1867
1868 // -A * -B -> A * B
1869 Value *A, *B;
1870 if (match(II->getArgOperand(0), m_FNeg(m_Value(A))) &&
1871 match(II->getArgOperand(1), m_FNeg(m_Value(B)))) {
1872 replaceOperand(*II, 0, A);
1873 replaceOperand(*II, 1, B);
1874 return II;
1875 }
1876
1877 Value *Op0 = II->getOperand(0);
1878 Value *Op1 = II->getOperand(1);
1879 Value *OpNotNeg, *NegatedOp;
1880 unsigned NegatedOpArg, OtherOpArg;
1881 if (match(Op0, m_FNeg(m_Value(OpNotNeg)))) {
1882 NegatedOp = Op0;
1883 NegatedOpArg = 0;
1884 OtherOpArg = 1;
1885 } else if (match(Op1, m_FNeg(m_Value(OpNotNeg)))) {
1886 NegatedOp = Op1;
1887 NegatedOpArg = 1;
1888 OtherOpArg = 0;
1889 } else
1890 // Multiplication doesn't have a negated operand.
1891 break;
1892
1893 // Only optimize if the negated operand has only one use.
1894 if (!NegatedOp->hasOneUse())
1895 break;
1896
1897 Value *OtherOp = II->getOperand(OtherOpArg);
1898 VectorType *RetTy = cast<VectorType>(II->getType());
1899 VectorType *NegatedOpTy = cast<VectorType>(NegatedOp->getType());
1900 VectorType *OtherOpTy = cast<VectorType>(OtherOp->getType());
1901 ElementCount NegatedCount = NegatedOpTy->getElementCount();
1902 ElementCount OtherCount = OtherOpTy->getElementCount();
1903 ElementCount RetCount = RetTy->getElementCount();
1904 // (-A) * B -> A * (-B), if it is cheaper to negate B and vice versa.
1905 if (ElementCount::isKnownGT(NegatedCount, OtherCount) &&
1906 ElementCount::isKnownLT(OtherCount, RetCount)) {
1907 Value *InverseOtherOp = Builder.CreateFNeg(OtherOp);
1908 replaceOperand(*II, NegatedOpArg, OpNotNeg);
1909 replaceOperand(*II, OtherOpArg, InverseOtherOp);
1910 return II;
1911 }
1912 // (-A) * B -> -(A * B), if it is cheaper to negate the result
1913 if (ElementCount::isKnownGT(NegatedCount, RetCount)) {
1914 SmallVector<Value *, 5> NewArgs(II->args());
1915 NewArgs[NegatedOpArg] = OpNotNeg;
1916 Instruction *NewMul =
1917 Builder.CreateIntrinsic(II->getType(), IID, NewArgs, II);
1918 return replaceInstUsesWith(*II, Builder.CreateFNegFMF(NewMul, II));
1919 }
1920 break;
1921 }
1922 case Intrinsic::fmuladd: {
1923 // Canonicalize fast fmuladd to the separate fmul + fadd.
1924 if (II->isFast()) {
1928 II->getArgOperand(1));
1930 Add->takeName(II);
1931 return replaceInstUsesWith(*II, Add);
1932 }
1933
1934 // Try to simplify the underlying FMul.
1935 if (Value *V = simplifyFMulInst(II->getArgOperand(0), II->getArgOperand(1),
1936 II->getFastMathFlags(),
1937 SQ.getWithInstruction(II))) {
1938 auto *FAdd = BinaryOperator::CreateFAdd(V, II->getArgOperand(2));
1939 FAdd->copyFastMathFlags(II);
1940 return FAdd;
1941 }
1942
1943 [[fallthrough]];
1944 }
1945 case Intrinsic::fma: {
1946 // fma fneg(x), fneg(y), z -> fma x, y, z
1947 Value *Src0 = II->getArgOperand(0);
1948 Value *Src1 = II->getArgOperand(1);
1949 Value *X, *Y;
1950 if (match(Src0, m_FNeg(m_Value(X))) && match(Src1, m_FNeg(m_Value(Y)))) {
1951 replaceOperand(*II, 0, X);
1952 replaceOperand(*II, 1, Y);
1953 return II;
1954 }
1955
1956 // fma fabs(x), fabs(x), z -> fma x, x, z
1957 if (match(Src0, m_FAbs(m_Value(X))) &&
1958 match(Src1, m_FAbs(m_Specific(X)))) {
1959 replaceOperand(*II, 0, X);
1960 replaceOperand(*II, 1, X);
1961 return II;
1962 }
1963
1964 // Try to simplify the underlying FMul. We can only apply simplifications
1965 // that do not require rounding.
1966 if (Value *V = simplifyFMAFMul(II->getArgOperand(0), II->getArgOperand(1),
1967 II->getFastMathFlags(),
1968 SQ.getWithInstruction(II))) {
1969 auto *FAdd = BinaryOperator::CreateFAdd(V, II->getArgOperand(2));
1970 FAdd->copyFastMathFlags(II);
1971 return FAdd;
1972 }
1973
1974 // fma x, y, 0 -> fmul x, y
1975 // This is always valid for -0.0, but requires nsz for +0.0 as
1976 // -0.0 + 0.0 = 0.0, which would not be the same as the fmul on its own.
1977 if (match(II->getArgOperand(2), m_NegZeroFP()) ||
1978 (match(II->getArgOperand(2), m_PosZeroFP()) &&
1980 return BinaryOperator::CreateFMulFMF(Src0, Src1, II);
1981
1982 break;
1983 }
1984 case Intrinsic::copysign: {
1985 Value *Mag = II->getArgOperand(0), *Sign = II->getArgOperand(1);
1986 if (SignBitMustBeZero(Sign, &TLI)) {
1987 // If we know that the sign argument is positive, reduce to FABS:
1988 // copysign Mag, +Sign --> fabs Mag
1989 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, Mag, II);
1990 return replaceInstUsesWith(*II, Fabs);
1991 }
1992 // TODO: There should be a ValueTracking sibling like SignBitMustBeOne.
1993 const APFloat *C;
1994 if (match(Sign, m_APFloat(C)) && C->isNegative()) {
1995 // If we know that the sign argument is negative, reduce to FNABS:
1996 // copysign Mag, -Sign --> fneg (fabs Mag)
1997 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, Mag, II);
1998 return replaceInstUsesWith(*II, Builder.CreateFNegFMF(Fabs, II));
1999 }
2000
2001 // Propagate sign argument through nested calls:
2002 // copysign Mag, (copysign ?, X) --> copysign Mag, X
2003 Value *X;
2004 if (match(Sign, m_Intrinsic<Intrinsic::copysign>(m_Value(), m_Value(X))))
2005 return replaceOperand(*II, 1, X);
2006
2007 // Peek through changes of magnitude's sign-bit. This call rewrites those:
2008 // copysign (fabs X), Sign --> copysign X, Sign
2009 // copysign (fneg X), Sign --> copysign X, Sign
2010 if (match(Mag, m_FAbs(m_Value(X))) || match(Mag, m_FNeg(m_Value(X))))
2011 return replaceOperand(*II, 0, X);
2012
2013 break;
2014 }
2015 case Intrinsic::fabs: {
2016 Value *Cond, *TVal, *FVal;
2017 if (match(II->getArgOperand(0),
2018 m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))) {
2019 // fabs (select Cond, TrueC, FalseC) --> select Cond, AbsT, AbsF
2020 if (isa<Constant>(TVal) && isa<Constant>(FVal)) {
2021 CallInst *AbsT = Builder.CreateCall(II->getCalledFunction(), {TVal});
2022 CallInst *AbsF = Builder.CreateCall(II->getCalledFunction(), {FVal});
2023 return SelectInst::Create(Cond, AbsT, AbsF);
2024 }
2025 // fabs (select Cond, -FVal, FVal) --> fabs FVal
2026 if (match(TVal, m_FNeg(m_Specific(FVal))))
2027 return replaceOperand(*II, 0, FVal);
2028 // fabs (select Cond, TVal, -TVal) --> fabs TVal
2029 if (match(FVal, m_FNeg(m_Specific(TVal))))
2030 return replaceOperand(*II, 0, TVal);
2031 }
2032
2033 Value *Magnitude, *Sign;
2034 if (match(II->getArgOperand(0),
2035 m_CopySign(m_Value(Magnitude), m_Value(Sign)))) {
2036 // fabs (copysign x, y) -> (fabs x)
2037 CallInst *AbsSign =
2038 Builder.CreateCall(II->getCalledFunction(), {Magnitude});
2039 AbsSign->copyFastMathFlags(II);
2040 return replaceInstUsesWith(*II, AbsSign);
2041 }
2042
2043 [[fallthrough]];
2044 }
2045 case Intrinsic::ceil:
2046 case Intrinsic::floor:
2047 case Intrinsic::round:
2048 case Intrinsic::roundeven:
2049 case Intrinsic::nearbyint:
2050 case Intrinsic::rint:
2051 case Intrinsic::trunc: {
2052 Value *ExtSrc;
2053 if (match(II->getArgOperand(0), m_OneUse(m_FPExt(m_Value(ExtSrc))))) {
2054 // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x)
2055 Value *NarrowII = Builder.CreateUnaryIntrinsic(IID, ExtSrc, II);
2056 return new FPExtInst(NarrowII, II->getType());
2057 }
2058 break;
2059 }
2060 case Intrinsic::cos:
2061 case Intrinsic::amdgcn_cos: {
2062 Value *X;
2063 Value *Src = II->getArgOperand(0);
2064 if (match(Src, m_FNeg(m_Value(X))) || match(Src, m_FAbs(m_Value(X)))) {
2065 // cos(-x) -> cos(x)
2066 // cos(fabs(x)) -> cos(x)
2067 return replaceOperand(*II, 0, X);
2068 }
2069 break;
2070 }
2071 case Intrinsic::sin: {
2072 Value *X;
2073 if (match(II->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X))))) {
2074 // sin(-x) --> -sin(x)
2075 Value *NewSin = Builder.CreateUnaryIntrinsic(Intrinsic::sin, X, II);
2076 Instruction *FNeg = UnaryOperator::CreateFNeg(NewSin);
2077 FNeg->copyFastMathFlags(II);
2078 return FNeg;
2079 }
2080 break;
2081 }
2082 case Intrinsic::ptrauth_auth:
2083 case Intrinsic::ptrauth_resign: {
2084 // (sign|resign) + (auth|resign) can be folded by omitting the middle
2085 // sign+auth component if the key and discriminator match.
2086 bool NeedSign = II->getIntrinsicID() == Intrinsic::ptrauth_resign;
2087 Value *Key = II->getArgOperand(1);
2088 Value *Disc = II->getArgOperand(2);
2089
2090 // AuthKey will be the key we need to end up authenticating against in
2091 // whatever we replace this sequence with.
2092 Value *AuthKey = nullptr, *AuthDisc = nullptr, *BasePtr;
2093 if (auto CI = dyn_cast<CallBase>(II->getArgOperand(0))) {
2094 BasePtr = CI->getArgOperand(0);
2095 if (CI->getIntrinsicID() == Intrinsic::ptrauth_sign) {
2096 if (CI->getArgOperand(1) != Key || CI->getArgOperand(2) != Disc)
2097 break;
2098 } else if (CI->getIntrinsicID() == Intrinsic::ptrauth_resign) {
2099 if (CI->getArgOperand(3) != Key || CI->getArgOperand(4) != Disc)
2100 break;
2101 AuthKey = CI->getArgOperand(1);
2102 AuthDisc = CI->getArgOperand(2);
2103 } else
2104 break;
2105 } else
2106 break;
2107
2108 unsigned NewIntrin;
2109 if (AuthKey && NeedSign) {
2110 // resign(0,1) + resign(1,2) = resign(0, 2)
2111 NewIntrin = Intrinsic::ptrauth_resign;
2112 } else if (AuthKey) {
2113 // resign(0,1) + auth(1) = auth(0)
2114 NewIntrin = Intrinsic::ptrauth_auth;
2115 } else if (NeedSign) {
2116 // sign(0) + resign(0, 1) = sign(1)
2117 NewIntrin = Intrinsic::ptrauth_sign;
2118 } else {
2119 // sign(0) + auth(0) = nop
2120 replaceInstUsesWith(*II, BasePtr);
2122 return nullptr;
2123 }
2124
2125 SmallVector<Value *, 4> CallArgs;
2126 CallArgs.push_back(BasePtr);
2127 if (AuthKey) {
2128 CallArgs.push_back(AuthKey);
2129 CallArgs.push_back(AuthDisc);
2130 }
2131
2132 if (NeedSign) {
2133 CallArgs.push_back(II->getArgOperand(3));
2134 CallArgs.push_back(II->getArgOperand(4));
2135 }
2136
2137 Function *NewFn = Intrinsic::getDeclaration(II->getModule(), NewIntrin);
2138 return CallInst::Create(NewFn, CallArgs);
2139 }
2140 case Intrinsic::arm_neon_vtbl1:
2141 case Intrinsic::aarch64_neon_tbl1:
2142 if (Value *V = simplifyNeonTbl1(*II, Builder))
2143 return replaceInstUsesWith(*II, V);
2144 break;
2145
2146 case Intrinsic::arm_neon_vmulls:
2147 case Intrinsic::arm_neon_vmullu:
2148 case Intrinsic::aarch64_neon_smull:
2149 case Intrinsic::aarch64_neon_umull: {
2150 Value *Arg0 = II->getArgOperand(0);
2151 Value *Arg1 = II->getArgOperand(1);
2152
2153 // Handle mul by zero first:
2154 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
2156 }
2157
2158 // Check for constant LHS & RHS - in this case we just simplify.
2159 bool Zext = (IID == Intrinsic::arm_neon_vmullu ||
2160 IID == Intrinsic::aarch64_neon_umull);
2161 VectorType *NewVT = cast<VectorType>(II->getType());
2162 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
2163 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
2164 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
2165 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
2166
2167 return replaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
2168 }
2169
2170 // Couldn't simplify - canonicalize constant to the RHS.
2171 std::swap(Arg0, Arg1);
2172 }
2173
2174 // Handle mul by one:
2175 if (Constant *CV1 = dyn_cast<Constant>(Arg1))
2176 if (ConstantInt *Splat =
2177 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
2178 if (Splat->isOne())
2179 return CastInst::CreateIntegerCast(Arg0, II->getType(),
2180 /*isSigned=*/!Zext);
2181
2182 break;
2183 }
2184 case Intrinsic::arm_neon_aesd:
2185 case Intrinsic::arm_neon_aese:
2186 case Intrinsic::aarch64_crypto_aesd:
2187 case Intrinsic::aarch64_crypto_aese: {
2188 Value *DataArg = II->getArgOperand(0);
2189 Value *KeyArg = II->getArgOperand(1);
2190
2191 // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR
2192 Value *Data, *Key;
2193 if (match(KeyArg, m_ZeroInt()) &&
2194 match(DataArg, m_Xor(m_Value(Data), m_Value(Key)))) {
2195 replaceOperand(*II, 0, Data);
2196 replaceOperand(*II, 1, Key);
2197 return II;
2198 }
2199 break;
2200 }
2201 case Intrinsic::hexagon_V6_vandvrt:
2202 case Intrinsic::hexagon_V6_vandvrt_128B: {
2203 // Simplify Q -> V -> Q conversion.
2204 if (auto Op0 = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
2205 Intrinsic::ID ID0 = Op0->getIntrinsicID();
2206 if (ID0 != Intrinsic::hexagon_V6_vandqrt &&
2207 ID0 != Intrinsic::hexagon_V6_vandqrt_128B)
2208 break;
2209 Value *Bytes = Op0->getArgOperand(1), *Mask = II->getArgOperand(1);
2210 uint64_t Bytes1 = computeKnownBits(Bytes, 0, Op0).One.getZExtValue();
2211 uint64_t Mask1 = computeKnownBits(Mask, 0, II).One.getZExtValue();
2212 // Check if every byte has common bits in Bytes and Mask.
2213 uint64_t C = Bytes1 & Mask1;
2214 if ((C & 0xFF) && (C & 0xFF00) && (C & 0xFF0000) && (C & 0xFF000000))
2215 return replaceInstUsesWith(*II, Op0->getArgOperand(0));
2216 }
2217 break;
2218 }
2219 case Intrinsic::stackrestore: {
2220 enum class ClassifyResult {
2221 None,
2222 Alloca,
2223 StackRestore,
2224 CallWithSideEffects,
2225 };
2226 auto Classify = [](const Instruction *I) {
2227 if (isa<AllocaInst>(I))
2228 return ClassifyResult::Alloca;
2229
2230 if (auto *CI = dyn_cast<CallInst>(I)) {
2231 if (auto *II = dyn_cast<IntrinsicInst>(CI)) {
2232 if (II->getIntrinsicID() == Intrinsic::stackrestore)
2233 return ClassifyResult::StackRestore;
2234
2235 if (II->mayHaveSideEffects())
2236 return ClassifyResult::CallWithSideEffects;
2237 } else {
2238 // Consider all non-intrinsic calls to be side effects
2239 return ClassifyResult::CallWithSideEffects;
2240 }
2241 }
2242
2243 return ClassifyResult::None;
2244 };
2245
2246 // If the stacksave and the stackrestore are in the same BB, and there is
2247 // no intervening call, alloca, or stackrestore of a different stacksave,
2248 // remove the restore. This can happen when variable allocas are DCE'd.
2249 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
2250 if (SS->getIntrinsicID() == Intrinsic::stacksave &&
2251 SS->getParent() == II->getParent()) {
2252 BasicBlock::iterator BI(SS);
2253 bool CannotRemove = false;
2254 for (++BI; &*BI != II; ++BI) {
2255 switch (Classify(&*BI)) {
2256 case ClassifyResult::None:
2257 // So far so good, look at next instructions.
2258 break;
2259
2260 case ClassifyResult::StackRestore:
2261 // If we found an intervening stackrestore for a different
2262 // stacksave, we can't remove the stackrestore. Otherwise, continue.
2263 if (cast<IntrinsicInst>(*BI).getArgOperand(0) != SS)
2264 CannotRemove = true;
2265 break;
2266
2267 case ClassifyResult::Alloca:
2268 case ClassifyResult::CallWithSideEffects:
2269 // If we found an alloca, a non-intrinsic call, or an intrinsic
2270 // call with side effects, we can't remove the stackrestore.
2271 CannotRemove = true;
2272 break;
2273 }
2274 if (CannotRemove)
2275 break;
2276 }
2277
2278 if (!CannotRemove)
2279 return eraseInstFromFunction(CI);
2280 }
2281 }
2282
2283 // Scan down this block to see if there is another stack restore in the
2284 // same block without an intervening call/alloca.
2285 BasicBlock::iterator BI(II);
2286 Instruction *TI = II->getParent()->getTerminator();
2287 bool CannotRemove = false;
2288 for (++BI; &*BI != TI; ++BI) {
2289 switch (Classify(&*BI)) {
2290 case ClassifyResult::None:
2291 // So far so good, look at next instructions.
2292 break;
2293
2294 case ClassifyResult::StackRestore:
2295 // If there is a stackrestore below this one, remove this one.
2296 return eraseInstFromFunction(CI);
2297
2298 case ClassifyResult::Alloca:
2299 case ClassifyResult::CallWithSideEffects:
2300 // If we found an alloca, a non-intrinsic call, or an intrinsic call
2301 // with side effects (such as llvm.stacksave and llvm.read_register),
2302 // we can't remove the stack restore.
2303 CannotRemove = true;
2304 break;
2305 }
2306 if (CannotRemove)
2307 break;
2308 }
2309
2310 // If the stack restore is in a return, resume, or unwind block and if there
2311 // are no allocas or calls between the restore and the return, nuke the
2312 // restore.
2313 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
2314 return eraseInstFromFunction(CI);
2315 break;
2316 }
2317 case Intrinsic::lifetime_end:
2318 // Asan needs to poison memory to detect invalid access which is possible
2319 // even for empty lifetime range.
2320 if (II->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
2321 II->getFunction()->hasFnAttribute(Attribute::SanitizeMemory) ||
2322 II->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
2323 break;
2324
2325 if (removeTriviallyEmptyRange(*II, *this, [](const IntrinsicInst &I) {
2326 return I.getIntrinsicID() == Intrinsic::lifetime_start;
2327 }))
2328 return nullptr;
2329 break;
2330 case Intrinsic::assume: {
2331 Value *IIOperand = II->getArgOperand(0);
2333 II->getOperandBundlesAsDefs(OpBundles);
2334
2335 /// This will remove the boolean Condition from the assume given as
2336 /// argument and remove the assume if it becomes useless.
2337 /// always returns nullptr for use as a return values.
2338 auto RemoveConditionFromAssume = [&](Instruction *Assume) -> Instruction * {
2339 assert(isa<AssumeInst>(Assume));
2340 if (isAssumeWithEmptyBundle(*cast<AssumeInst>(II)))
2341 return eraseInstFromFunction(CI);
2343 return nullptr;
2344 };
2345 // Remove an assume if it is followed by an identical assume.
2346 // TODO: Do we need this? Unless there are conflicting assumptions, the
2347 // computeKnownBits(IIOperand) below here eliminates redundant assumes.
2349 if (match(Next, m_Intrinsic<Intrinsic::assume>(m_Specific(IIOperand))))
2350 return RemoveConditionFromAssume(Next);
2351
2352 // Canonicalize assume(a && b) -> assume(a); assume(b);
2353 // Note: New assumption intrinsics created here are registered by
2354 // the InstCombineIRInserter object.
2355 FunctionType *AssumeIntrinsicTy = II->getFunctionType();
2356 Value *AssumeIntrinsic = II->getCalledOperand();
2357 Value *A, *B;
2358 if (match(IIOperand, m_LogicalAnd(m_Value(A), m_Value(B)))) {
2359 Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic, A, OpBundles,
2360 II->getName());
2361 Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic, B, II->getName());
2362 return eraseInstFromFunction(*II);
2363 }
2364 // assume(!(a || b)) -> assume(!a); assume(!b);
2365 if (match(IIOperand, m_Not(m_LogicalOr(m_Value(A), m_Value(B))))) {
2366 Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic,
2367 Builder.CreateNot(A), OpBundles, II->getName());
2368 Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic,
2369 Builder.CreateNot(B), II->getName());
2370 return eraseInstFromFunction(*II);
2371 }
2372
2373 // assume( (load addr) != null ) -> add 'nonnull' metadata to load
2374 // (if assume is valid at the load)
2375 CmpInst::Predicate Pred;
2377 if (match(IIOperand, m_ICmp(Pred, m_Instruction(LHS), m_Zero())) &&
2378 Pred == ICmpInst::ICMP_NE && LHS->getOpcode() == Instruction::Load &&
2379 LHS->getType()->isPointerTy() &&
2381 MDNode *MD = MDNode::get(II->getContext(), std::nullopt);
2382 LHS->setMetadata(LLVMContext::MD_nonnull, MD);
2383 return RemoveConditionFromAssume(II);
2384
2385 // TODO: apply nonnull return attributes to calls and invokes
2386 // TODO: apply range metadata for range check patterns?
2387 }
2388
2389 // Convert nonnull assume like:
2390 // %A = icmp ne i32* %PTR, null
2391 // call void @llvm.assume(i1 %A)
2392 // into
2393 // call void @llvm.assume(i1 true) [ "nonnull"(i32* %PTR) ]
2395 match(IIOperand, m_Cmp(Pred, m_Value(A), m_Zero())) &&
2396 Pred == CmpInst::ICMP_NE && A->getType()->isPointerTy()) {
2397 if (auto *Replacement = buildAssumeFromKnowledge(
2398 {RetainedKnowledge{Attribute::NonNull, 0, A}}, Next, &AC, &DT)) {
2399
2400 Replacement->insertBefore(Next);
2401 AC.registerAssumption(Replacement);
2402 return RemoveConditionFromAssume(II);
2403 }
2404 }
2405
2406 // Convert alignment assume like:
2407 // %B = ptrtoint i32* %A to i64
2408 // %C = and i64 %B, Constant
2409 // %D = icmp eq i64 %C, 0
2410 // call void @llvm.assume(i1 %D)
2411 // into
2412 // call void @llvm.assume(i1 true) [ "align"(i32* [[A]], i64 Constant + 1)]
2413 uint64_t AlignMask;
2415 match(IIOperand,
2416 m_Cmp(Pred, m_And(m_Value(A), m_ConstantInt(AlignMask)),
2417 m_Zero())) &&
2418 Pred == CmpInst::ICMP_EQ) {
2419 if (isPowerOf2_64(AlignMask + 1)) {
2420 uint64_t Offset = 0;
2422 if (match(A, m_PtrToInt(m_Value(A)))) {
2423 /// Note: this doesn't preserve the offset information but merges
2424 /// offset and alignment.
2425 /// TODO: we can generate a GEP instead of merging the alignment with
2426 /// the offset.
2427 RetainedKnowledge RK{Attribute::Alignment,
2428 (unsigned)MinAlign(Offset, AlignMask + 1), A};
2429 if (auto *Replacement =
2430 buildAssumeFromKnowledge(RK, Next, &AC, &DT)) {
2431
2432 Replacement->insertAfter(II);
2433 AC.registerAssumption(Replacement);
2434 }
2435 return RemoveConditionFromAssume(II);
2436 }
2437 }
2438 }
2439
2440 /// Canonicalize Knowledge in operand bundles.
2442 for (unsigned Idx = 0; Idx < II->getNumOperandBundles(); Idx++) {
2443 auto &BOI = II->bundle_op_info_begin()[Idx];
2445 llvm::getKnowledgeFromBundle(cast<AssumeInst>(*II), BOI);
2446 if (BOI.End - BOI.Begin > 2)
2447 continue; // Prevent reducing knowledge in an align with offset since
2448 // extracting a RetainedKnowledge from them looses offset
2449 // information
2450 RetainedKnowledge CanonRK =
2451 llvm::simplifyRetainedKnowledge(cast<AssumeInst>(II), RK,
2453 &getDominatorTree());
2454 if (CanonRK == RK)
2455 continue;
2456 if (!CanonRK) {
2457 if (BOI.End - BOI.Begin > 0) {
2458 Worklist.pushValue(II->op_begin()[BOI.Begin]);
2459 Value::dropDroppableUse(II->op_begin()[BOI.Begin]);
2460 }
2461 continue;
2462 }
2463 assert(RK.AttrKind == CanonRK.AttrKind);
2464 if (BOI.End - BOI.Begin > 0)
2465 II->op_begin()[BOI.Begin].set(CanonRK.WasOn);
2466 if (BOI.End - BOI.Begin > 1)
2467 II->op_begin()[BOI.Begin + 1].set(ConstantInt::get(
2468 Type::getInt64Ty(II->getContext()), CanonRK.ArgValue));
2469 if (RK.WasOn)
2471 return II;
2472 }
2473 }
2474
2475 // If there is a dominating assume with the same condition as this one,
2476 // then this one is redundant, and should be removed.
2477 KnownBits Known(1);
2478 computeKnownBits(IIOperand, Known, 0, II);
2479 if (Known.isAllOnes() && isAssumeWithEmptyBundle(cast<AssumeInst>(*II)))
2480 return eraseInstFromFunction(*II);
2481
2482 // Update the cache of affected values for this assumption (we might be
2483 // here because we just simplified the condition).
2484 AC.updateAffectedValues(cast<AssumeInst>(II));
2485 break;
2486 }
2487 case Intrinsic::experimental_guard: {
2488 // Is this guard followed by another guard? We scan forward over a small
2489 // fixed window of instructions to handle common cases with conditions
2490 // computed between guards.
2491 Instruction *NextInst = II->getNextNonDebugInstruction();
2492 for (unsigned i = 0; i < GuardWideningWindow; i++) {
2493 // Note: Using context-free form to avoid compile time blow up
2494 if (!isSafeToSpeculativelyExecute(NextInst))
2495 break;
2496 NextInst = NextInst->getNextNonDebugInstruction();
2497 }
2498 Value *NextCond = nullptr;
2499 if (match(NextInst,
2500 m_Intrinsic<Intrinsic::experimental_guard>(m_Value(NextCond)))) {
2501 Value *CurrCond = II->getArgOperand(0);
2502
2503 // Remove a guard that it is immediately preceded by an identical guard.
2504 // Otherwise canonicalize guard(a); guard(b) -> guard(a & b).
2505 if (CurrCond != NextCond) {
2507 while (MoveI != NextInst) {
2508 auto *Temp = MoveI;
2509 MoveI = MoveI->getNextNonDebugInstruction();
2510 Temp->moveBefore(II);
2511 }
2512 replaceOperand(*II, 0, Builder.CreateAnd(CurrCond, NextCond));
2513 }
2514 eraseInstFromFunction(*NextInst);
2515 return II;
2516 }
2517 break;
2518 }
2519 case Intrinsic::vector_insert: {
2520 Value *Vec = II->getArgOperand(0);
2521 Value *SubVec = II->getArgOperand(1);
2522 Value *Idx = II->getArgOperand(2);
2523 auto *DstTy = dyn_cast<FixedVectorType>(II->getType());
2524 auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType());
2525 auto *SubVecTy = dyn_cast<FixedVectorType>(SubVec->getType());
2526
2527 // Only canonicalize if the destination vector, Vec, and SubVec are all
2528 // fixed vectors.
2529 if (DstTy && VecTy && SubVecTy) {
2530 unsigned DstNumElts = DstTy->getNumElements();
2531 unsigned VecNumElts = VecTy->getNumElements();
2532 unsigned SubVecNumElts = SubVecTy->getNumElements();
2533 unsigned IdxN = cast<ConstantInt>(Idx)->getZExtValue();
2534
2535 // An insert that entirely overwrites Vec with SubVec is a nop.
2536 if (VecNumElts == SubVecNumElts)
2537 return replaceInstUsesWith(CI, SubVec);
2538
2539 // Widen SubVec into a vector of the same width as Vec, since
2540 // shufflevector requires the two input vectors to be the same width.
2541 // Elements beyond the bounds of SubVec within the widened vector are
2542 // undefined.
2543 SmallVector<int, 8> WidenMask;
2544 unsigned i;
2545 for (i = 0; i != SubVecNumElts; ++i)
2546 WidenMask.push_back(i);
2547 for (; i != VecNumElts; ++i)
2548 WidenMask.push_back(UndefMaskElem);
2549
2550 Value *WidenShuffle = Builder.CreateShuffleVector(SubVec, WidenMask);
2551
2553 for (unsigned i = 0; i != IdxN; ++i)
2554 Mask.push_back(i);
2555 for (unsigned i = DstNumElts; i != DstNumElts + SubVecNumElts; ++i)
2556 Mask.push_back(i);
2557 for (unsigned i = IdxN + SubVecNumElts; i != DstNumElts; ++i)
2558 Mask.push_back(i);
2559
2560 Value *Shuffle = Builder.CreateShuffleVector(Vec, WidenShuffle, Mask);
2561 return replaceInstUsesWith(CI, Shuffle);
2562 }
2563 break;
2564 }
2565 case Intrinsic::vector_extract: {
2566 Value *Vec = II->getArgOperand(0);
2567 Value *Idx = II->getArgOperand(1);
2568
2569 Type *ReturnType = II->getType();
2570 // (extract_vector (insert_vector InsertTuple, InsertValue, InsertIdx),
2571 // ExtractIdx)
2572 unsigned ExtractIdx = cast<ConstantInt>(Idx)->getZExtValue();
2573 Value *InsertTuple, *InsertIdx, *InsertValue;
2574 if (match(Vec, m_Intrinsic<Intrinsic::vector_insert>(m_Value(InsertTuple),
2575 m_Value(InsertValue),
2576 m_Value(InsertIdx))) &&
2577 InsertValue->getType() == ReturnType) {
2578 unsigned Index = cast<ConstantInt>(InsertIdx)->getZExtValue();
2579 // Case where we get the same index right after setting it.
2580 // extract.vector(insert.vector(InsertTuple, InsertValue, Idx), Idx) -->
2581 // InsertValue
2582 if (ExtractIdx == Index)
2583 return replaceInstUsesWith(CI, InsertValue);
2584 // If we are getting a different index than what was set in the
2585 // insert.vector intrinsic. We can just set the input tuple to the one up
2586 // in the chain. extract.vector(insert.vector(InsertTuple, InsertValue,
2587 // InsertIndex), ExtractIndex)
2588 // --> extract.vector(InsertTuple, ExtractIndex)
2589 else
2590 return replaceOperand(CI, 0, InsertTuple);
2591 }
2592
2593 auto *DstTy = dyn_cast<FixedVectorType>(ReturnType);
2594 auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType());
2595
2596 // Only canonicalize if the the destination vector and Vec are fixed
2597 // vectors.
2598 if (DstTy && VecTy) {
2599 unsigned DstNumElts = DstTy->getNumElements();
2600 unsigned VecNumElts = VecTy->getNumElements();
2601 unsigned IdxN = cast<ConstantInt>(Idx)->getZExtValue();
2602
2603 // Extracting the entirety of Vec is a nop.
2604 if (VecNumElts == DstNumElts) {
2605 replaceInstUsesWith(CI, Vec);
2606 return eraseInstFromFunction(CI);
2607 }
2608
2610 for (unsigned i = 0; i != DstNumElts; ++i)
2611 Mask.push_back(IdxN + i);
2612
2614 return replaceInstUsesWith(CI, Shuffle);
2615 }
2616 break;
2617 }
2618 case Intrinsic::experimental_vector_reverse: {
2619 Value *BO0, *BO1, *X, *Y;
2620 Value *Vec = II->getArgOperand(0);
2621 if (match(Vec, m_OneUse(m_BinOp(m_Value(BO0), m_Value(BO1))))) {
2622 auto *OldBinOp = cast<BinaryOperator>(Vec);
2623 if (match(BO0, m_VecReverse(m_Value(X)))) {
2624 // rev(binop rev(X), rev(Y)) --> binop X, Y
2625 if (match(BO1, m_VecReverse(m_Value(Y))))
2626 return replaceInstUsesWith(CI,
2628 OldBinOp->getOpcode(), X, Y, OldBinOp,
2629 OldBinOp->getName(), II));
2630 // rev(binop rev(X), BO1Splat) --> binop X, BO1Splat
2631 if (isSplatValue(BO1))
2632 return replaceInstUsesWith(CI,
2634 OldBinOp->getOpcode(), X, BO1,
2635 OldBinOp, OldBinOp->getName(), II));
2636 }
2637 // rev(binop BO0Splat, rev(Y)) --> binop BO0Splat, Y
2638 if (match(BO1, m_VecReverse(m_Value(Y))) && isSplatValue(BO0))
2640 OldBinOp->getOpcode(), BO0, Y,
2641 OldBinOp, OldBinOp->getName(), II));
2642 }
2643 // rev(unop rev(X)) --> unop X
2644 if (match(Vec, m_OneUse(m_UnOp(m_VecReverse(m_Value(X)))))) {
2645 auto *OldUnOp = cast<UnaryOperator>(Vec);
2647 OldUnOp->getOpcode(), X, OldUnOp, OldUnOp->getName(), II);
2648 return replaceInstUsesWith(CI, NewUnOp);
2649 }
2650 break;
2651 }
2652 case Intrinsic::vector_reduce_or:
2653 case Intrinsic::vector_reduce_and: {
2654 // Canonicalize logical or/and reductions:
2655 // Or reduction for i1 is represented as:
2656 // %val = bitcast <ReduxWidth x i1> to iReduxWidth
2657 // %res = cmp ne iReduxWidth %val, 0
2658 // And reduction for i1 is represented as:
2659 // %val = bitcast <ReduxWidth x i1> to iReduxWidth
2660 // %res = cmp eq iReduxWidth %val, 11111
2661 Value *Arg = II->getArgOperand(0);
2662 Value *Vect;
2663 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2664 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2665 if (FTy->getElementType() == Builder.getInt1Ty()) {
2667 Vect, Builder.getIntNTy(FTy->getNumElements()));
2668 if (IID == Intrinsic::vector_reduce_and) {
2669 Res = Builder.CreateICmpEQ(
2671 } else {
2672 assert(IID == Intrinsic::vector_reduce_or &&
2673 "Expected or reduction.");
2674 Res = Builder.CreateIsNotNull(Res);
2675 }
2676 if (Arg != Vect)
2677 Res = Builder.CreateCast(cast<CastInst>(Arg)->getOpcode(), Res,
2678 II->getType());
2679 return replaceInstUsesWith(CI, Res);
2680 }
2681 }
2682 [[fallthrough]];
2683 }
2684 case Intrinsic::vector_reduce_add: {
2685 if (IID == Intrinsic::vector_reduce_add) {
2686 // Convert vector_reduce_add(ZExt(<n x i1>)) to
2687 // ZExtOrTrunc(ctpop(bitcast <n x i1> to in)).
2688 // Convert vector_reduce_add(SExt(<n x i1>)) to
2689 // -ZExtOrTrunc(ctpop(bitcast <n x i1> to in)).
2690 // Convert vector_reduce_add(<n x i1>) to
2691 // Trunc(ctpop(bitcast <n x i1> to in)).
2692 Value *Arg = II->getArgOperand(0);
2693 Value *Vect;
2694 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2695 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2696 if (FTy->getElementType() == Builder.getInt1Ty()) {
2698 Vect, Builder.getIntNTy(FTy->getNumElements()));
2699 Value *Res = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, V);
2700 if (Res->getType() != II->getType())
2701 Res = Builder.CreateZExtOrTrunc(Res, II->getType());
2702 if (Arg != Vect &&
2703 cast<Instruction>(Arg)->getOpcode() == Instruction::SExt)
2704 Res = Builder.CreateNeg(Res);
2705 return replaceInstUsesWith(CI, Res);
2706 }
2707 }
2708 }
2709 [[fallthrough]];
2710 }
2711 case Intrinsic::vector_reduce_xor: {
2712 if (IID == Intrinsic::vector_reduce_xor) {
2713 // Exclusive disjunction reduction over the vector with
2714 // (potentially-extended) i1 element type is actually a
2715 // (potentially-extended) arithmetic `add` reduction over the original
2716 // non-extended value:
2717 // vector_reduce_xor(?ext(<n x i1>))
2718 // -->
2719 // ?ext(vector_reduce_add(<n x i1>))
2720 Value *Arg = II->getArgOperand(0);
2721 Value *Vect;
2722 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2723 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2724 if (FTy->getElementType() == Builder.getInt1Ty()) {
2725 Value *Res = Builder.CreateAddReduce(Vect);
2726 if (Arg != Vect)
2727 Res = Builder.CreateCast(cast<CastInst>(Arg)->getOpcode(), Res,
2728 II->getType());
2729 return replaceInstUsesWith(CI, Res);
2730 }
2731 }
2732 }
2733 [[fallthrough]];
2734 }
2735 case Intrinsic::vector_reduce_mul: {
2736 if (IID == Intrinsic::vector_reduce_mul) {
2737 // Multiplicative reduction over the vector with (potentially-extended)
2738 // i1 element type is actually a (potentially zero-extended)
2739 // logical `and` reduction over the original non-extended value:
2740 // vector_reduce_mul(?ext(<n x i1>))
2741 // -->
2742 // zext(vector_reduce_and(<n x i1>))
2743 Value *Arg = II->getArgOperand(0);
2744 Value *Vect;
2745 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2746 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2747 if (FTy->getElementType() == Builder.getInt1Ty()) {
2748 Value *Res = Builder.CreateAndReduce(Vect);
2749 if (Res->getType() != II->getType())
2750 Res = Builder.CreateZExt(Res, II->getType());
2751 return replaceInstUsesWith(CI, Res);
2752 }
2753 }
2754 }
2755 [[fallthrough]];
2756 }
2757 case Intrinsic::vector_reduce_umin:
2758 case Intrinsic::vector_reduce_umax: {
2759 if (IID == Intrinsic::vector_reduce_umin ||
2760 IID == Intrinsic::vector_reduce_umax) {
2761 // UMin/UMax reduction over the vector with (potentially-extended)
2762 // i1 element type is actually a (potentially-extended)
2763 // logical `and`/`or` reduction over the original non-extended value:
2764 // vector_reduce_u{min,max}(?ext(<n x i1>))
2765 // -->
2766 // ?ext(vector_reduce_{and,or}(<n x i1>))
2767 Value *Arg = II->getArgOperand(0);
2768 Value *Vect;
2769 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2770 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2771 if (FTy->getElementType() == Builder.getInt1Ty()) {
2772 Value *Res = IID == Intrinsic::vector_reduce_umin
2773 ? Builder.CreateAndReduce(Vect)
2774 : Builder.CreateOrReduce(Vect);
2775 if (Arg != Vect)
2776 Res = Builder.CreateCast(cast<CastInst>(Arg)->getOpcode(), Res,
2777 II->getType());
2778 return replaceInstUsesWith(CI, Res);
2779 }
2780 }
2781 }
2782 [[fallthrough]];
2783 }
2784 case Intrinsic::vector_reduce_smin:
2785 case Intrinsic::vector_reduce_smax: {
2786 if (IID == Intrinsic::vector_reduce_smin ||
2787 IID == Intrinsic::vector_reduce_smax) {
2788 // SMin/SMax reduction over the vector with (potentially-extended)
2789 // i1 element type is actually a (potentially-extended)
2790 // logical `and`/`or` reduction over the original non-extended value:
2791 // vector_reduce_s{min,max}(<n x i1>)
2792 // -->
2793 // vector_reduce_{or,and}(<n x i1>)
2794 // and
2795 // vector_reduce_s{min,max}(sext(<n x i1>))
2796 // -->
2797 // sext(vector_reduce_{or,and}(<n x i1>))
2798 // and
2799 // vector_reduce_s{min,max}(zext(<n x i1>))
2800 // -->
2801 // zext(vector_reduce_{and,or}(<n x i1>))
2802 Value *Arg = II->getArgOperand(0);
2803 Value *Vect;
2804 if (match(Arg, m_ZExtOrSExtOrSelf(m_Value(Vect)))) {
2805 if (auto *FTy = dyn_cast<FixedVectorType>(Vect->getType()))
2806 if (FTy->getElementType() == Builder.getInt1Ty()) {
2807 Instruction::CastOps ExtOpc = Instruction::CastOps::CastOpsEnd;
2808 if (Arg != Vect)
2809 ExtOpc = cast<CastInst>(Arg)->getOpcode();
2810 Value *Res = ((IID == Intrinsic::vector_reduce_smin) ==
2811 (ExtOpc == Instruction::CastOps::ZExt))
2812 ? Builder.CreateAndReduce(Vect)
2813 : Builder.CreateOrReduce(Vect);
2814 if (Arg != Vect)
2815 Res = Builder.CreateCast(ExtOpc, Res, II->getType());
2816 return replaceInstUsesWith(CI, Res);
2817 }
2818 }
2819 }
2820 [[fallthrough]];
2821 }
2822 case Intrinsic::vector_reduce_fmax:
2823 case Intrinsic::vector_reduce_fmin:
2824 case Intrinsic::vector_reduce_fadd:
2825 case Intrinsic::vector_reduce_fmul: {
2826 bool CanBeReassociated = (IID != Intrinsic::vector_reduce_fadd &&
2827 IID != Intrinsic::vector_reduce_fmul) ||
2828 II->hasAllowReassoc();
2829 const unsigned ArgIdx = (IID == Intrinsic::vector_reduce_fadd ||
2830 IID == Intrinsic::vector_reduce_fmul)
2831 ? 1
2832 : 0;
2833 Value *Arg = II->getArgOperand(ArgIdx);
2834 Value *V;
2835 ArrayRef<int> Mask;
2836 if (!isa<FixedVectorType>(Arg->getType()) || !CanBeReassociated ||
2837 !match(Arg, m_Shuffle(m_Value(V), m_Undef(), m_Mask(Mask))) ||
2838 !cast<ShuffleVectorInst>(Arg)->isSingleSource())
2839 break;
2840 int Sz = Mask.size();
2841 SmallBitVector UsedIndices(Sz);
2842 for (int Idx : Mask) {
2843 if (Idx == UndefMaskElem || UsedIndices.test(Idx))
2844 break;
2845 UsedIndices.set(Idx);
2846 }
2847 // Can remove shuffle iff just shuffled elements, no repeats, undefs, or
2848 // other changes.
2849 if (UsedIndices.all()) {
2850 replaceUse(II->getOperandUse(ArgIdx), V);
2851 return nullptr;
2852 }
2853 break;
2854 }
2855 default: {
2856 // Handle target specific intrinsics
2857 std::optional<Instruction *> V = targetInstCombineIntrinsic(*II);
2858 if (V)
2859 return *V;
2860 break;
2861 }
2862 }
2863
2865 return Shuf;
2866
2867 // Some intrinsics (like experimental_gc_statepoint) can be used in invoke
2868 // context, so it is handled in visitCallBase and we should trigger it.
2869 return visitCallBase(*II);
2870}
2871
2872// Fence instruction simplification
2874 auto *NFI = dyn_cast<FenceInst>(FI.getNextNonDebugInstruction());
2875 // This check is solely here to handle arbitrary target-dependent syncscopes.
2876 // TODO: Can remove if does not matter in practice.
2877 if (NFI && FI.isIdenticalTo(NFI))
2878 return eraseInstFromFunction(FI);
2879
2880 // Returns true if FI1 is identical or stronger fence than FI2.
2881 auto isIdenticalOrStrongerFence = [](FenceInst *FI1, FenceInst *FI2) {
2882 auto FI1SyncScope = FI1->getSyncScopeID();
2883 // Consider same scope, where scope is global or single-thread.
2884 if (FI1SyncScope != FI2->getSyncScopeID() ||
2885 (FI1SyncScope != SyncScope::System &&
2886 FI1SyncScope != SyncScope::SingleThread))
2887 return false;
2888
2889 return isAtLeastOrStrongerThan(FI1->getOrdering(), FI2->getOrdering());
2890 };
2891 if (NFI && isIdenticalOrStrongerFence(NFI, &FI))
2892 return eraseInstFromFunction(FI);
2893
2894 if (auto *PFI = dyn_cast_or_null<FenceInst>(FI.getPrevNonDebugInstruction()))
2895 if (isIdenticalOrStrongerFence(PFI, &FI))
2896 return eraseInstFromFunction(FI);
2897 return nullptr;
2898}
2899
2900// InvokeInst simplification
2902 return visitCallBase(II);
2903}
2904
2905// CallBrInst simplification
2907 return visitCallBase(CBI);
2908}
2909
2910/// If this cast does not affect the value passed through the varargs area, we
2911/// can eliminate the use of the cast.
2913 const DataLayout &DL,
2914 const CastInst *const CI,
2915 const int ix) {
2916 if (!CI->isLosslessCast())
2917 return false;
2918
2919 // If this is a GC intrinsic, avoid munging types. We need types for
2920 // statepoint reconstruction in SelectionDAG.
2921 // TODO: This is probably something which should be expanded to all
2922 // intrinsics since the entire point of intrinsics is that
2923 // they are understandable by the optimizer.
2924 if (isa<GCStatepointInst>(Call) || isa<GCRelocateInst>(Call) ||
2925 isa<GCResultInst>(Call))
2926 return false;
2927
2928 // Opaque pointers are compatible with any byval types.
2929 PointerType *SrcTy = cast<PointerType>(CI->getOperand(0)->getType());
2930 if (SrcTy->isOpaque())
2931 return true;
2932
2933 // The size of ByVal or InAlloca arguments is derived from the type, so we
2934 // can't change to a type with a different size. If the size were
2935 // passed explicitly we could avoid this check.
2936 if (!Call.isPassPointeeByValueArgument(ix))
2937 return true;
2938
2939 // The transform currently only handles type replacement for byval, not other
2940 // type-carrying attributes.
2941 if (!Call.isByValArgument(ix))
2942 return false;
2943
2944 Type *SrcElemTy = SrcTy->getNonOpaquePointerElementType();
2945 Type *DstElemTy = Call.getParamByValType(ix);
2946 if (!SrcElemTy->isSized() || !DstElemTy->isSized())
2947 return false;
2948 if (DL.getTypeAllocSize(SrcElemTy) != DL.getTypeAllocSize(DstElemTy))
2949 return false;
2950 return true;
2951}
2952
2953Instruction *InstCombinerImpl::tryOptimizeCall(CallInst *CI) {
2954 if (!CI->getCalledFunction()) return nullptr;
2955
2956 // Skip optimizing notail and musttail calls so
2957 // LibCallSimplifier::optimizeCall doesn't have to preserve those invariants.
2958 // LibCallSimplifier::optimizeCall should try to preseve tail calls though.
2959 if (CI->isMustTailCall() || CI->isNoTailCall())
2960 return nullptr;
2961
2962 auto InstCombineRAUW = [this](Instruction *From, Value *With) {
2963 replaceInstUsesWith(*From, With);
2964 };
2965 auto InstCombineErase = [this](Instruction *I) {
2967 };
2968 LibCallSimplifier Simplifier(DL, &TLI, ORE, BFI, PSI, InstCombineRAUW,
2969 InstCombineErase);
2970 if (Value *With = Simplifier.optimizeCall(CI, Builder)) {
2971 ++NumSimplified;
2972 return CI->use_empty() ? CI : replaceInstUsesWith(*CI, With);
2973 }
2974
2975 return nullptr;
2976}
2977
2979 // Strip off at most one level of pointer casts, looking for an alloca. This
2980 // is good enough in practice and simpler than handling any number of casts.
2981 Value *Underlying = TrampMem->stripPointerCasts();
2982 if (Underlying != TrampMem &&
2983 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
2984 return nullptr;
2985 if (!isa<AllocaInst>(Underlying))
2986 return nullptr;
2987
2988 IntrinsicInst *InitTrampoline = nullptr;
2989 for (User *U : TrampMem->users()) {
2990 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
2991 if (!II)
2992 return nullptr;
2993 if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
2994 if (InitTrampoline)
2995 // More than one init_trampoline writes to this value. Give up.
2996 return nullptr;
2997 InitTrampoline = II;
2998 continue;
2999 }
3000 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
3001 // Allow any number of calls to adjust.trampoline.
3002 continue;
3003 return nullptr;
3004 }
3005
3006 // No call to init.trampoline found.
3007 if (!InitTrampoline)
3008 return nullptr;
3009
3010 // Check that the alloca is being used in the expected way.
3011 if (InitTrampoline->getOperand(0) != TrampMem)
3012 return nullptr;
3013
3014 return InitTrampoline;
3015}
3016
3018 Value *TrampMem) {
3019 // Visit all the previous instructions in the basic block, and try to find a
3020 // init.trampoline which has a direct path to the adjust.trampoline.
3021 for (BasicBlock::iterator I = AdjustTramp->getIterator(),
3022 E = AdjustTramp->getParent()->begin();
3023 I != E;) {
3024 Instruction *Inst = &*--I;
3025 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
3026 if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
3027 II->getOperand(0) == TrampMem)
3028 return II;
3029 if (Inst->mayWriteToMemory())
3030 return nullptr;
3031 }
3032 return nullptr;
3033}
3034
3035// Given a call to llvm.adjust.trampoline, find and return the corresponding
3036// call to llvm.init.trampoline if the call to the trampoline can be optimized
3037// to a direct call to a function. Otherwise return NULL.
3039 Callee = Callee->stripPointerCasts();
3040 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
3041 if (!AdjustTramp ||
3042 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
3043 return nullptr;
3044
3045 Value *TrampMem = AdjustTramp->getOperand(0);
3046
3048 return IT;
3049 if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem))
3050 return IT;
3051 return nullptr;
3052}
3053
3054bool InstCombinerImpl::annotateAnyAllocSite(CallBase &Call,
3055 const TargetLibraryInfo *TLI) {
3056 // Note: We only handle cases which can't be driven from generic attributes
3057 // here. So, for example, nonnull and noalias (which are common properties
3058 // of some allocation functions) are expected to be handled via annotation
3059 // of the respective allocator declaration with generic attributes.
3060 bool Changed = false;
3061
3062 if (!Call.getType()->isPointerTy())
3063 return Changed;
3064
3065 std::optional<APInt> Size = getAllocSize(&Call, TLI);
3066 if (Size && *Size != 0) {
3067 // TODO: We really should just emit deref_or_null here and then
3068 // let the generic inference code combine that with nonnull.
3069 if (Call.hasRetAttr(Attribute::NonNull)) {
3070 Changed = !Call.hasRetAttr(Attribute::Dereferenceable);
3072 Call.getContext(), Size->getLimitedValue()));
3073 } else {
3074 Changed = !Call.hasRetAttr(Attribute::DereferenceableOrNull);
3076 Call.getContext(), Size->getLimitedValue()));
3077 }
3078 }
3079
3080 // Add alignment attribute if alignment is a power of two constant.
3081 Value *Alignment = getAllocAlignment(&Call, TLI);
3082 if (!Alignment)
3083 return Changed;
3084
3085 ConstantInt *AlignOpC = dyn_cast<ConstantInt>(Alignment);
3086 if (AlignOpC && AlignOpC->getValue().ult(llvm::Value::MaximumAlignment)) {
3087 uint64_t AlignmentVal = AlignOpC->getZExtValue();
3088 if (llvm::isPowerOf2_64(AlignmentVal)) {
3089 Align ExistingAlign = Call.getRetAlign().valueOrOne();
3090 Align NewAlign = Align(AlignmentVal);
3091 if (NewAlign > ExistingAlign) {
3092 Call.addRetAttr(
3093 Attribute::getWithAlignment(Call.getContext(), NewAlign));
3094 Changed = true;
3095 }
3096 }
3097 }
3098 return Changed;
3099}
3100
3101/// Improvements for call, callbr and invoke instructions.
3102Instruction *InstCombinerImpl::visitCallBase(CallBase &Call) {
3103 bool Changed = annotateAnyAllocSite(Call, &TLI);
3104
3105 // Mark any parameters that are known to be non-null with the nonnull
3106 // attribute. This is helpful for inlining calls to functions with null
3107 // checks on their arguments.
3109 unsigned ArgNo = 0;
3110
3111 for (Value *V : Call.args()) {
3112 if (V->getType()->isPointerTy() &&
3113 !Call.paramHasAttr(ArgNo, Attribute::NonNull) &&
3114 isKnownNonZero(V, DL, 0, &AC, &Call, &DT))
3115 ArgNos.push_back(ArgNo);
3116 ArgNo++;
3117 }
3118
3119 assert(ArgNo == Call.arg_size() && "Call arguments not processed correctly.");
3120
3121 if (!ArgNos.empty()) {
3122 AttributeList AS = Call.getAttributes();
3123 LLVMContext &Ctx = Call.getContext();
3124 AS = AS.addParamAttribute(Ctx, ArgNos,
3125 Attribute::get(Ctx, Attribute::NonNull));
3126 Call.setAttributes(AS);
3127 Changed = true;
3128 }
3129
3130 // If the callee is a pointer to a function, attempt to move any casts to the
3131 // arguments of the call/callbr/invoke.
3132 Value *Callee = Call.getCalledOperand();
3133 Function *CalleeF = dyn_cast<Function>(Callee);
3134 if ((!CalleeF || CalleeF->getFunctionType() != Call.getFunctionType()) &&
3135 transformConstExprCastCall(Call))
3136 return nullptr;
3137
3138 if (CalleeF) {
3139 // Remove the convergent attr on calls when the callee is not convergent.
3140 if (Call.isConvergent() && !CalleeF->isConvergent() &&
3141 !CalleeF->isIntrinsic()) {
3142 LLVM_DEBUG(dbgs() << "Removing convergent attr from instr " << Call
3143 << "\n");
3144 Call.setNotConvergent();
3145 return &Call;
3146 }
3147
3148 // If the call and callee calling conventions don't match, and neither one
3149 // of the calling conventions is compatible with C calling convention
3150 // this call must be unreachable, as the call is undefined.
3151 if ((CalleeF->getCallingConv() != Call.getCallingConv() &&
3152 !(CalleeF->getCallingConv() == llvm::CallingConv::C &&
3154 !(Call.getCallingConv() == llvm::CallingConv::C &&
3156 // Only do this for calls to a function with a body. A prototype may
3157 // not actually end up matching the implementation's calling conv for a
3158 // variety of reasons (e.g. it may be written in assembly).
3159 !CalleeF->isDeclaration()) {
3160 Instruction *OldCall = &Call;
3162 // If OldCall does not return void then replaceInstUsesWith poison.
3163 // This allows ValueHandlers and custom metadata to adjust itself.
3164 if (!OldCall->getType()->isVoidTy())
3165 replaceInstUsesWith(*OldCall, PoisonValue::get(OldCall->getType()));
3166 if (isa<CallInst>(OldCall))
3167 return eraseInstFromFunction(*OldCall);
3168
3169 // We cannot remove an invoke or a callbr, because it would change thexi
3170 // CFG, just change the callee to a null pointer.
3171 cast<CallBase>(OldCall)->setCalledFunction(
3172 CalleeF->getFunctionType(),
3173 Constant::getNullValue(CalleeF->getType()));
3174 return nullptr;
3175 }
3176 }
3177
3178 // Calling a null function pointer is undefined if a null address isn't
3179 // dereferenceable.
3180 if ((isa<ConstantPointerNull>(Callee) &&
3181 !NullPointerIsDefined(Call.getFunction())) ||
3182 isa<UndefValue>(Callee)) {
3183 // If Call does not return void then replaceInstUsesWith poison.
3184 // This allows ValueHandlers and custom metadata to adjust itself.
3185 if (!Call.getType()->isVoidTy())
3186 replaceInstUsesWith(Call, PoisonValue::get(Call.getType()));
3187
3188 if (Call.isTerminator()) {
3189 // Can't remove an invoke or callbr because we cannot change the CFG.
3190 return nullptr;
3191 }
3192
3193 // This instruction is not reachable, just remove it.
3195 return eraseInstFromFunction(Call);
3196 }
3197
3199 return transformCallThroughTrampoline(Call, *II);
3200
3201 // TODO: Drop this transform once opaque pointer transition is done.
3202 FunctionType *FTy = Call.getFunctionType();
3203 if (FTy->isVarArg()) {
3204 int ix = FTy->getNumParams();
3205 // See if we can optimize any arguments passed through the varargs area of
3206 // the call.
3207 for (auto I = Call.arg_begin() + FTy->getNumParams(), E = Call.arg_end();
3208 I != E; ++I, ++ix) {
3209 CastInst *CI = dyn_cast<CastInst>(*I);
3210 if (CI && isSafeToEliminateVarargsCast(Call, DL, CI, ix)) {
3211 replaceUse(*I, CI->getOperand(0));
3212
3213 // Update the byval type to match the pointer type.
3214 // Not necessary for opaque pointers.
3215 PointerType *NewTy = cast<PointerType>(CI->getOperand(0)->getType());
3216 if (!NewTy->isOpaque() && Call.isByValArgument(ix)) {
3217 Call.removeParamAttr(ix, Attribute::ByVal);
3218 Call.addParamAttr(ix, Attribute::getWithByValType(
3219 Call.getContext(),
3220 NewTy->getNonOpaquePointerElementType()));
3221 }
3222 Changed = true;
3223 }
3224 }
3225 }
3226
3227 if (isa<InlineAsm>(Callee) && !Call.doesNotThrow()) {
3228 InlineAsm *IA = cast<InlineAsm>(Callee);
3229 if (!IA->canThrow()) {
3230 // Normal inline asm calls cannot throw - mark them
3231 // 'nounwind'.
3232 Call.setDoesNotThrow();
3233 Changed = true;
3234 }
3235 }
3236
3237 // Try to optimize the call if possible, we require DataLayout for most of
3238 // this. None of these calls are seen as possibly dead so go ahead and
3239 // delete the instruction now.
3240 if (CallInst *CI = dyn_cast<CallInst>(&Call)) {
3241 Instruction *I = tryOptimizeCall(CI);
3242 // If we changed something return the result, etc. Otherwise let
3243 // the fallthrough check.
3244 if (I) return eraseInstFromFunction(*I);
3245 }
3246
3247 if (!Call.use_empty() && !Call.isMustTailCall())
3248 if (Value *ReturnedArg = Call.getReturnedArgOperand()) {
3249 Type *CallTy = Call.getType();
3250 Type *RetArgTy = ReturnedArg->getType();
3251 if (RetArgTy->canLosslesslyBitCastTo(CallTy))
3252 return replaceInstUsesWith(
3253 Call, Builder.CreateBitOrPointerCast(ReturnedArg, CallTy));
3254 }
3255
3256 // Drop unnecessary kcfi operand bundles from calls that were converted
3257 // into direct calls.
3258 auto Bundle = Call.getOperandBundle(LLVMContext::OB_kcfi);
3259 if (Bundle && !Call.isIndirectCall()) {
3260 DEBUG_WITH_TYPE(DEBUG_TYPE "-kcfi", {
3261 if (CalleeF) {
3262 ConstantInt *FunctionType = nullptr;
3263 ConstantInt *ExpectedType = cast<ConstantInt>(Bundle->Inputs[0]);
3264
3265 if (MDNode *MD = CalleeF->getMetadata(LLVMContext::MD_kcfi_type))
3266 FunctionType = mdconst::extract<ConstantInt>(MD->getOperand(0));
3267
3268 if (FunctionType &&
3269 FunctionType->getZExtValue() != ExpectedType->getZExtValue())
3270 dbgs() << Call.getModule()->getName()
3271 << ": warning: kcfi: " << Call.getCaller()->getName()
3272 << ": call to " << CalleeF->getName()
3273 << " using a mismatching function pointer type\n";
3274 }
3275 });
3276
3278 }
3279
3280 if (isRemovableAlloc(&Call, &TLI))
3281 return visitAllocSite(Call);
3282
3283 // Handle intrinsics which can be used in both call and invoke context.
3284 switch (Call.getIntrinsicID()) {
3285 case Intrinsic::experimental_gc_statepoint: {
3286 GCStatepointInst &GCSP = *cast<GCStatepointInst>(&Call);
3287 SmallPtrSet<Value *, 32> LiveGcValues;
3288 for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) {
3289 GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc);
3290
3291 // Remove the relocation if unused.
3292 if (GCR.use_empty()) {
3294 continue;
3295 }
3296
3297 Value *DerivedPtr = GCR.getDerivedPtr();
3298 Value *BasePtr = GCR.getBasePtr();
3299
3300 // Undef is undef, even after relocation.
3301 if (isa<UndefValue>(DerivedPtr) || isa<UndefValue>(BasePtr)) {
3304 continue;
3305 }
3306
3307 if (auto *PT = dyn_cast<PointerType>(GCR.getType())) {
3308 // The relocation of null will be null for most any collector.
3309 // TODO: provide a hook for this in GCStrategy. There might be some
3310 // weird collector this property does not hold for.
3311 if (isa<ConstantPointerNull>(DerivedPtr)) {
3312 // Use null-pointer of gc_relocate's type to replace it.
3315 continue;
3316 }
3317
3318 // isKnownNonNull -> nonnull attribute
3319 if (!GCR.hasRetAttr(Attribute::NonNull) &&
3320 isKnownNonZero(DerivedPtr, DL, 0, &AC, &Call, &DT)) {
3321 GCR.addRetAttr(Attribute::NonNull);
3322 // We discovered new fact, re-check users.
3324 }
3325 }
3326
3327 // If we have two copies of the same pointer in the statepoint argument
3328 // list, canonicalize to one. This may let us common gc.relocates.
3329 if (GCR.getBasePtr() == GCR.getDerivedPtr() &&
3330 GCR.getBasePtrIndex() != GCR.getDerivedPtrIndex()) {
3331 auto *OpIntTy = GCR.getOperand(2)->getType();
3332 GCR.setOperand(2, ConstantInt::get(OpIntTy, GCR.getBasePtrIndex()));
3333 }
3334
3335 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
3336 // Canonicalize on the type from the uses to the defs
3337
3338 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
3339 LiveGcValues.insert(BasePtr);
3340 LiveGcValues.insert(DerivedPtr);
3341 }
3342 std::optional<OperandBundleUse> Bundle =
3344 unsigned NumOfGCLives = LiveGcValues.size();
3345 if (!Bundle || NumOfGCLives == Bundle->Inputs.size())
3346 break;
3347 // We can reduce the size of gc live bundle.
3349 std::vector<Value *> NewLiveGc;
3350 for (Value *V : Bundle->Inputs) {
3351 if (Val2Idx.count(V))
3352 continue;
3353 if (LiveGcValues.count(V)) {
3354 Val2Idx[V] = NewLiveGc.size();
3355 NewLiveGc.push_back(V);
3356 } else
3357 Val2Idx[V] = NumOfGCLives;
3358 }
3359 // Update all gc.relocates
3360 for (const GCRelocateInst *Reloc : GCSP.getGCRelocates()) {
3361 GCRelocateInst &GCR = *const_cast<GCRelocateInst *>(Reloc);
3362 Value *BasePtr = GCR.getBasePtr();
3363 assert(Val2Idx.count(BasePtr) && Val2Idx[BasePtr] != NumOfGCLives &&
3364 "Missed live gc for base pointer");
3365 auto *OpIntTy1 = GCR.getOperand(1)->getType();
3366 GCR.setOperand(1, ConstantInt::get(OpIntTy1, Val2Idx[BasePtr]));
3367 Value *DerivedPtr = GCR.getDerivedPtr();
3368 assert(Val2Idx.count(DerivedPtr) && Val2Idx[DerivedPtr] != NumOfGCLives &&
3369 "Missed live gc for derived pointer");
3370 auto *OpIntTy2 = GCR.getOperand(2)->getType();
3371 GCR.setOperand(2, ConstantInt::get(OpIntTy2, Val2Idx[DerivedPtr]));
3372 }
3373 // Create new statepoint instruction.
3374 OperandBundleDef NewBundle("gc-live", NewLiveGc);
3375 return CallBase::Create(&Call, NewBundle);
3376 }
3377 default: { break; }
3378 }
3379
3380 return Changed ? &Call : nullptr;
3381}
3382
3383/// If the callee is a constexpr cast of a function, attempt to move the cast to
3384/// the arguments of the call/callbr/invoke.
3385bool InstCombinerImpl::transformConstExprCastCall(CallBase &Call) {
3386 auto *Callee =
3387 dyn_cast<Function>(Call.getCalledOperand()->stripPointerCasts());
3388 if (!Callee)
3389 return false;
3390
3391 // If this is a call to a thunk function, don't remove the cast. Thunks are
3392 // used to transparently forward all incoming parameters and outgoing return
3393 // values, so it's important to leave the cast in place.
3394 if (Callee->hasFnAttribute("thunk"))
3395 return false;
3396
3397 // If this is a musttail call, the callee's prototype must match the caller's
3398 // prototype with the exception of pointee types. The code below doesn't
3399 // implement that, so we can't do this transform.
3400 // TODO: Do the transform if it only requires adding pointer casts.
3401 if (Call.isMustTailCall())
3402 return false;
3403
3405 const AttributeList &CallerPAL = Call.getAttributes();
3406
3407 // Okay, this is a cast from a function to a different type. Unless doing so
3408 // would cause a type conversion of one of our arguments, change this call to
3409 // be a direct call with arguments casted to the appropriate types.
3411 Type *OldRetTy = Caller->getType();
3412 Type *NewRetTy = FT->getReturnType();
3413
3414 // Check to see if we are changing the return type...
3415 if (OldRetTy != NewRetTy) {
3416
3417 if (NewRetTy->isStructTy())
3418 return false; // TODO: Handle multiple return values.
3419
3420 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
3421 if (Callee->isDeclaration())
3422 return false; // Cannot transform this return value.
3423
3424 if (!Caller->use_empty() &&
3425 // void -> non-void is handled specially
3426 !NewRetTy->isVoidTy())
3427 return false; // Cannot transform this return value.
3428 }
3429
3430 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
3431 AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs());
3432 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
3433 return false; // Attribute not compatible with transformed value.
3434 }
3435
3436 // If the callbase is an invoke/callbr instruction, and the return value is
3437 // used by a PHI node in a successor, we cannot change the return type of
3438 // the call because there is no place to put the cast instruction (without
3439 // breaking the critical edge). Bail out in this case.
3440 if (!Caller->use_empty()) {
3441 BasicBlock *PhisNotSupportedBlock = nullptr;
3442 if (auto *II = dyn_cast<InvokeInst>(Caller))
3443 PhisNotSupportedBlock = II->getNormalDest();
3444 if (auto *CB = dyn_cast<CallBrInst>(Caller))
3445 PhisNotSupportedBlock = CB->getDefaultDest();
3446 if (PhisNotSupportedBlock)
3447 for (User *U : Caller->users())
3448 if (PHINode *PN = dyn_cast<PHINode>(U))
3449 if (PN->getParent() == PhisNotSupportedBlock)
3450 return false;
3451 }
3452 }
3453
3454 unsigned NumActualArgs = Call.arg_size();
3455 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
3456
3457 // Prevent us turning:
3458 // declare void @takes_i32_inalloca(i32* inalloca)
3459 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
3460 //
3461 // into:
3462 // call void @takes_i32_inalloca(i32* null)
3463 //
3464 // Similarly, avoid folding away bitcasts of byval calls.
3465 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
3466 Callee->getAttributes().hasAttrSomewhere(Attribute::Preallocated))
3467 return false;
3468
3469 auto AI = Call.arg_begin();
3470 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
3471 Type *ParamTy = FT->getParamType(i);
3472 Type *ActTy = (*AI)->getType();
3473
3474 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
3475 return false; // Cannot transform this parameter value.
3476
3477 // Check if there are any incompatible attributes we cannot drop safely.
3478 if (AttrBuilder(FT->getContext(), CallerPAL.getParamAttrs(i))
3481 return false; // Attribute not compatible with transformed value.
3482
3483 if (Call.isInAllocaArgument(i) ||
3484 CallerPAL.hasParamAttr(i, Attribute::Preallocated))
3485 return false; // Cannot transform to and from inalloca/preallocated.
3486
3487 if (CallerPAL.hasParamAttr(i, Attribute::SwiftError))
3488 return false;
3489
3490 if (CallerPAL.hasParamAttr(i, Attribute::ByVal) !=
3491 Callee->getAttributes().hasParamAttr(i, Attribute::ByVal))
3492 return false; // Cannot transform to or from byval.
3493
3494 // If the parameter is passed as a byval argument, then we have to have a
3495 // sized type and the sized type has to have the same size as the old type.
3496 if (ParamTy != ActTy && CallerPAL.hasParamAttr(i, Attribute::ByVal)) {
3497 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
3498 if (!ParamPTy)
3499 return false;
3500
3501 if (!ParamPTy->isOpaque()) {
3502 Type *ParamElTy = ParamPTy->getNonOpaquePointerElementType();
3503 if (!ParamElTy->isSized())
3504 return false;
3505
3506 Type *CurElTy = Call.getParamByValType(i);
3507 if (DL.getTypeAllocSize(CurElTy) != DL.getTypeAllocSize(ParamElTy))
3508 return false;
3509 }
3510 }
3511 }
3512
3513 if (Callee->isDeclaration()) {
3514 // Do not delete arguments unless we have a function body.
3515 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
3516 return false;
3517
3518 // If the callee is just a declaration, don't change the varargsness of the
3519 // call. We don't want to introduce a varargs call where one doesn't
3520 // already exist.
3521 if (FT->isVarArg() != Call.getFunctionType()->isVarArg())
3522 return false;
3523
3524 // If both the callee and the cast type are varargs, we still have to make
3525 // sure the number of fixed parameters are the same or we have the same
3526 // ABI issues as if we introduce a varargs call.
3527 if (FT->isVarArg() && Call.getFunctionType()->isVarArg() &&
3528 FT->getNumParams() != Call.getFunctionType()->getNumParams())
3529 return false;
3530 }
3531
3532 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
3533 !CallerPAL.isEmpty()) {
3534 // In this case we have more arguments than the new function type, but we
3535 // won't be dropping them. Check that these extra arguments have attributes
3536 // that are compatible with being a vararg call argument.
3537 unsigned SRetIdx;
3538 if (CallerPAL.hasAttrSomewhere(Attribute::StructRet, &SRetIdx) &&
3539 SRetIdx - AttributeList::FirstArgIndex >= FT->getNumParams())
3540 return false;
3541 }
3542
3543 // Okay, we decided that this is a safe thing to do: go ahead and start
3544 // inserting cast instructions as necessary.
3547 Args.reserve(NumActualArgs);
3548 ArgAttrs.reserve(NumActualArgs);
3549
3550 // Get any return attributes.
3551 AttrBuilder RAttrs(FT->getContext(), CallerPAL.getRetAttrs());
3552
3553 // If the return value is not being used, the type may not be compatible
3554 // with the existing attributes. Wipe out any problematic attributes.
3555 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
3556
3557 LLVMContext &Ctx = Call.getContext();
3558 AI = Call.arg_begin();
3559 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
3560 Type *ParamTy = FT->getParamType(i);
3561
3562 Value *NewArg = *AI;
3563 if ((*AI)->getType() != ParamTy)
3564 NewArg = Builder.CreateBitOrPointerCast(*AI, ParamTy);
3565 Args.push_back(NewArg);
3566
3567 // Add any parameter attributes except the ones incompatible with the new
3568 // type. Note that we made sure all incompatible ones are safe to drop.
3571 if (CallerPAL.hasParamAttr(i, Attribute::ByVal) &&
3572 !ParamTy->isOpaquePointerTy()) {
3573 AttrBuilder AB(Ctx, CallerPAL.getParamAttrs(i).removeAttributes(
3574 Ctx, IncompatibleAttrs));
3575 AB.addByValAttr(ParamTy->getNonOpaquePointerElementType());
3576 ArgAttrs.push_back(AttributeSet::get(Ctx, AB));
3577 } else {
3578 ArgAttrs.push_back(
3579 CallerPAL.getParamAttrs(i).removeAttributes(Ctx, IncompatibleAttrs));
3580 }
3581 }
3582
3583 // If the function takes more arguments than the call was taking, add them
3584 // now.
3585 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) {
3586 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
3587 ArgAttrs.push_back(AttributeSet());
3588 }
3589
3590 // If we are removing arguments to the function, emit an obnoxious warning.
3591 if (FT->getNumParams() < NumActualArgs) {
3592 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
3593 if (FT->isVarArg()) {
3594 // Add all of the arguments in their promoted form to the arg list.
3595 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
3596 Type *PTy = getPromotedType((*AI)->getType());
3597 Value *NewArg = *AI;
3598 if (PTy != (*AI)->getType()) {
3599 // Must promote to pass through va_arg area!
3600 Instruction::CastOps opcode =
3601 CastInst::getCastOpcode(*AI, false, PTy, false);
3602 NewArg = Builder.CreateCast(opcode, *AI, PTy);
3603 }
3604 Args.push_back(NewArg);
3605
3606 // Add any parameter attributes.
3607 ArgAttrs.push_back(CallerPAL.getParamAttrs(i));
3608 }
3609 }
3610 }
3611
3612 AttributeSet FnAttrs = CallerPAL.getFnAttrs();
3613
3614 if (NewRetTy->isVoidTy())
3615 Caller->setName(""); // Void type should not have a name.
3616
3617 assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) &&
3618 "missing argument attributes");
3619 AttributeList NewCallerPAL = AttributeList::get(
3620 Ctx, FnAttrs, AttributeSet::get(Ctx, RAttrs), ArgAttrs);
3621
3623 Call.getOperandBundlesAsDefs(OpBundles);
3624
3625 CallBase *NewCall;
3626 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
3627 NewCall = Builder.CreateInvoke(Callee, II->getNormalDest(),
3628 II->getUnwindDest(), Args, OpBundles);
3629 } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(Caller)) {
3630 NewCall = Builder.CreateCallBr(Callee, CBI->getDefaultDest(),
3631 CBI->getIndirectDests(), Args, OpBundles);
3632 } else {
3633 NewCall = Builder.CreateCall(Callee, Args, OpBundles);
3634 cast<CallInst>(NewCall)->setTailCallKind(
3635 cast<CallInst>(Caller)->getTailCallKind());
3636 }
3637 NewCall->takeName(Caller);
3638 NewCall->setCallingConv(Call.getCallingConv());
3639 NewCall->setAttributes(NewCallerPAL);
3640
3641 // Preserve prof metadata if any.
3642 NewCall->copyMetadata(*Caller, {LLVMContext::MD_prof});
3643
3644 // Insert a cast of the return type as necessary.
3645 Instruction *NC = NewCall;
3646 Value *NV = NC;
3647 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
3648 if (!NV->getType()->isVoidTy()) {
3650 NC->setDebugLoc(Caller->getDebugLoc());
3651
3652 Instruction *InsertPt = NewCall->getInsertionPointAfterDef();
3653 assert(InsertPt && "No place to insert cast");
3654 InsertNewInstBefore(NC, *InsertPt);
3656 } else {
3657 NV = PoisonValue::get(Caller->getType());
3658 }
3659 }
3660
3661 if (!Caller->use_empty())
3662 replaceInstUsesWith(*Caller, NV);
3663 else if (Caller->hasValueHandle()) {
3664 if (OldRetTy == NV->getType())
3666 else
3667 // We cannot call ValueIsRAUWd with a different type, and the
3668 // actual tracked value will disappear.
3670 }
3671
3672 eraseInstFromFunction(*Caller);
3673 return true;
3674}
3675
3676/// Turn a call to a function created by init_trampoline / adjust_trampoline
3677/// intrinsic pair into a direct call to the underlying function.
3679InstCombinerImpl::transformCallThroughTrampoline(CallBase &Call,
3680 IntrinsicInst &Tramp) {
3681 Value *Callee = Call.getCalledOperand();
3682 Type *CalleeTy = Callee->getType();
3683 FunctionType *FTy = Call.getFunctionType();
3684 AttributeList Attrs = Call.getAttributes();
3685
3686 // If the call already has the 'nest' attribute somewhere then give up -
3687 // otherwise 'nest' would occur twice after splicing in the chain.
3688 if (Attrs.hasAttrSomewhere(Attribute::Nest))
3689 return nullptr;
3690
3691 Function *NestF = cast<Function>(Tramp.getArgOperand(1)->stripPointerCasts());
3692 FunctionType *NestFTy = NestF->getFunctionType();
3693
3694 AttributeList NestAttrs = NestF->getAttributes();
3695 if (!NestAttrs.isEmpty()) {
3696 unsigned NestArgNo = 0;
3697 Type *NestTy = nullptr;
3698 AttributeSet NestAttr;
3699
3700 // Look for a parameter marked with the 'nest' attribute.
3701 for (FunctionType::param_iterator I = NestFTy->param_begin(),
3702 E = NestFTy->param_end();
3703 I != E; ++NestArgNo, ++I) {
3704 AttributeSet AS = NestAttrs.getParamAttrs(NestArgNo);
3705 if (AS.hasAttribute(Attribute::Nest)) {
3706 // Record the parameter type and any other attributes.
3707 NestTy = *I;
3708 NestAttr = AS;
3709 break;
3710 }
3711 }
3712
3713 if (NestTy) {
3714 std::vector<Value*> NewArgs;
3715 std::vector<AttributeSet> NewArgAttrs;
3716 NewArgs.reserve(Call.arg_size() + 1);
3717 NewArgAttrs.reserve(Call.arg_size());
3718
3719 // Insert the nest argument into the call argument list, which may
3720 // mean appending it. Likewise for attributes.
3721
3722 {
3723 unsigned ArgNo = 0;
3724 auto I = Call.arg_begin(), E = Call.arg_end();
3725 do {
3726 if (ArgNo == NestArgNo) {
3727 // Add the chain argument and attributes.
3728 Value *NestVal = Tramp.getArgOperand(2);
3729 if (NestVal->getType() != NestTy)
3730 NestVal = Builder.CreateBitCast(NestVal, NestTy, "nest");
3731 NewArgs.push_back(NestVal);
3732 NewArgAttrs.push_back(NestAttr);
3733 }
3734
3735 if (I == E)
3736 break;
3737
3738 // Add the original argument and attributes.
3739 NewArgs.push_back(*I);
3740 NewArgAttrs.push_back(Attrs.getParamAttrs(ArgNo));
3741
3742 ++ArgNo;
3743 ++I;
3744 } while (true);
3745 }
3746
3747 // The trampoline may have been bitcast to a bogus type (FTy).
3748 // Handle this by synthesizing a new function type, equal to FTy
3749 // with the chain parameter inserted.
3750
3751 std::vector<Type*> NewTypes;
3752 NewTypes.reserve(FTy->getNumParams()+1);
3753
3754 // Insert the chain's type into the list of parameter types, which may
3755 // mean appending it.
3756 {
3757 unsigned ArgNo = 0;
3758 FunctionType::param_iterator I = FTy->param_begin(),
3759 E = FTy->param_end();
3760
3761 do {
3762 if (ArgNo == NestArgNo)
3763 // Add the chain's type.
3764 NewTypes.push_back(NestTy);
3765
3766 if (I == E)
3767 break;
3768
3769 // Add the original type.
3770 NewTypes.push_back(*I);
3771
3772 ++ArgNo;
3773 ++I;
3774 } while (true);
3775 }
3776
3777 // Replace the trampoline call with a direct call. Let the generic
3778 // code sort out any function type mismatches.
3779 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
3780 FTy->isVarArg());
3781 Constant *NewCallee =
3782 NestF->getType() == PointerType::getUnqual(NewFTy) ?
3783 NestF : ConstantExpr::getBitCast(NestF,
3784 PointerType::getUnqual(NewFTy));
3785 AttributeList NewPAL =
3786 AttributeList::get(FTy->getContext(), Attrs.getFnAttrs(),
3787 Attrs.getRetAttrs(), NewArgAttrs);
3788
3790 Call.getOperandBundlesAsDefs(OpBundles);
3791
3792 Instruction *NewCaller;
3793 if (InvokeInst *II = dyn_cast<InvokeInst>(&Call)) {
3794 NewCaller = InvokeInst::Create(NewFTy, NewCallee,
3795 II->getNormalDest(), II->getUnwindDest(),
3796 NewArgs, OpBundles);
3797 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
3798 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
3799 } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(&Call)) {
3800 NewCaller =
3801 CallBrInst::Create(NewFTy, NewCallee, CBI->getDefaultDest(),
3802 CBI->getIndirectDests(), NewArgs, OpBundles);
3803 cast<CallBrInst>(NewCaller)->setCallingConv(CBI->getCallingConv());
3804 cast<CallBrInst>(NewCaller)->setAttributes(NewPAL);
3805 } else {
3806 NewCaller = CallInst::Create(NewFTy, NewCallee, NewArgs, OpBundles);
3807 cast<CallInst>(NewCaller)->setTailCallKind(
3808 cast<CallInst>(Call).getTailCallKind());
3809 cast<CallInst>(NewCaller)->setCallingConv(
3810 cast<CallInst>(Call).getCallingConv());
3811 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
3812 }
3813 NewCaller->setDebugLoc(Call.getDebugLoc());
3814
3815 return NewCaller;
3816 }
3817 }
3818
3819 // Replace the trampoline call with a direct call. Since there is no 'nest'
3820 // parameter, there is no need to adjust the argument list. Let the generic
3821 // code sort out any function type mismatches.
3822 Constant *NewCallee = ConstantExpr::getBitCast(NestF, CalleeTy);
3823 Call.setCalledFunction(FTy, NewCallee);
3824 return &Call;
3825}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
unsigned Intr
amdgpu Simplify well known AMD library false FunctionCallee Callee
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
This file declares a class to represent arbitrary precision floating point values and provide a varie...
This file implements a class to represent arbitrary precision integral constant values and operations...
This file implements the APSInt class, which is a simple class that represents an arbitrary sized int...
static cl::opt< ITMode > IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT), cl::values(clEnumValN(DefaultIT, "arm-default-it", "Generate any type of IT block"), clEnumValN(RestrictedIT, "arm-restrict-it", "Disallow complex IT blocks")))
assume Assume Builder
Atomic ordering constants.
This file contains the simple types necessary to represent the attributes associated with functions a...
SmallVector< MachineOperand, 4 > Cond
BlockVerifier::State From
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
return RetTy
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
#define LLVM_DEBUG(X)
Definition: Debug.h:101
#define DEBUG_WITH_TYPE(TYPE, X)
DEBUG_WITH_TYPE macro - This macro should be used by passes to emit debug information.
Definition: Debug.h:64
uint64_t Size
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
#define DEBUG_TYPE
IRTranslator LLVM IR MI
static Type * getPromotedType(Type *Ty)
Return the specified type promoted as it would be to pass though a va_arg area.
static Instruction * createOverflowTuple(IntrinsicInst *II, Value *Result, Constant *Overflow)
Creates a result tuple for an overflow intrinsic II with a given Result and a constant Overflow value...
static bool isSafeToEliminateVarargsCast(const CallBase &Call, const DataLayout &DL, const CastInst *const CI, const int ix)
If this cast does not affect the value passed through the varargs area, we can eliminate the use of t...
static IntrinsicInst * findInitTrampolineFromAlloca(Value *TrampMem)
static bool removeTriviallyEmptyRange(IntrinsicInst &EndI, InstCombinerImpl &IC, std::function< bool(const IntrinsicInst &)> IsStart)
static Instruction * reassociateMinMaxWithConstantInOperand(IntrinsicInst *II, InstCombiner::BuilderTy &Builder)
If this min/max has a matching min/max operand with a constant, try to push the constant operand into...
static Instruction * moveAddAfterMinMax(IntrinsicInst *II, InstCombiner::BuilderTy &Builder)
Try to canonicalize min/max(X + C0, C1) as min/max(X, C1 - C0) + C0.
static Instruction * simplifyInvariantGroupIntrinsic(IntrinsicInst &II, InstCombinerImpl &IC)
This function transforms launder.invariant.group and strip.invariant.group like: launder(launder(x)) ...
static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E, unsigned NumOperands)
static cl::opt< unsigned > GuardWideningWindow("instcombine-guard-widening-window", cl::init(3), cl::desc("How wide an instruction window to bypass looking for " "another guard"))
static bool hasUndefSource(AnyMemTransferInst *MI)
Recognize a memcpy/memmove from a trivially otherwise unused alloca.
static Instruction * foldShuffledIntrinsicOperands(IntrinsicInst *II, InstCombiner::BuilderTy &Builder)
If all arguments of the intrinsic are unary shuffles with the same mask, try to shuffle after the int...
static Instruction * factorizeMinMaxTree(IntrinsicInst *II)
Reduce a sequence of min/max intrinsics with a common operand.
static Value * simplifyNeonTbl1(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Convert a table lookup to shufflevector if the mask is constant.
static Instruction * foldClampRangeOfTwo(IntrinsicInst *II, InstCombiner::BuilderTy &Builder)
If we have a clamp pattern like max (min X, 42), 41 – where the output can only be one of two possibl...
static IntrinsicInst * findInitTrampolineFromBB(IntrinsicInst *AdjustTramp, Value *TrampMem)
static Instruction * reassociateMinMaxWithConstants(IntrinsicInst *II)
If this min/max has a constant operand and an operand that is a matching min/max with a constant oper...
static Instruction * foldCtpop(IntrinsicInst &II, InstCombinerImpl &IC)
static Instruction * foldCttzCtlz(IntrinsicInst &II, InstCombinerImpl &IC)
static IntrinsicInst * findInitTrampoline(Value *Callee)
static std::optional< bool > getKnownSign(Value *Op, Instruction *CxtI, const DataLayout &DL, AssumptionCache *AC, DominatorTree *DT)
static CallInst * canonicalizeConstantArg0ToArg1(CallInst &Call)
This file provides internal interfaces used to implement the InstCombine.
This file provides the interface for the instcombine pass implementation.
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
This file contains the declarations for metadata subclasses.
Metadata * LowAndHigh[]
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
@ SI
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file implements the SmallBitVector class.
This file defines the SmallVector class.
This file defines the 'Statistic' class, which is designed to be an easy way to expose various metric...
#define STATISTIC(VARNAME, DESC)
Definition: Statistic.h:167
@ Struct
static std::optional< unsigned > getOpcode(ArrayRef< VPValue * > Values)
Returns the opcode of Values or ~0 if they do not all agree.
Definition: VPlanSLP.cpp:191
Value * RHS
Value * LHS
ModRefInfo getModRefInfoMask(const MemoryLocation &Loc, bool IgnoreLocals=false)
Returns a bitmask that should be unconditionally applied to the ModRef info of a memory location.
Class for arbitrary precision integers.
Definition: APInt.h:75
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition: APInt.h:214
static APInt getSignMask(unsigned BitWidth)
Get the SignMask for a specific bit width.
Definition: APInt.h:209
APInt usub_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1946
bool isZero() const
Determine if this value is zero, i.e. all bits are clear.
Definition: APInt.h:366
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1439
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1089
APInt sadd_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1926
APInt uadd_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1933
APInt uadd_sat(const APInt &RHS) const
Definition: APInt.cpp:2019
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:317
static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet)
Constructs an APInt value that has the bottom loBitsSet bits set.
Definition: APInt.h:289
APInt ssub_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1939
static APSInt getMinValue(uint32_t numBits, bool Unsigned)
Return the APSInt representing the minimum integer value with the given bit width and signedness.
Definition: APSInt.h:311
static APSInt getMaxValue(uint32_t numBits, bool Unsigned)
Return the APSInt representing the maximum integer value with the given bit width and signedness.
Definition: APSInt.h:303
This class represents any memset intrinsic.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
A cache of @llvm.assume calls within a function.
void updateAffectedValues(CondGuardInst *CI)
Update the cache of values being affected by this assumption (i.e.
void registerAssumption(CondGuardInst *CI)
Add an @llvm.assume intrinsic to this function's cache.
bool overlaps(const AttributeMask &AM) const
Return true if the builder has any attribute that's in the specified builder.
AttributeSet getFnAttrs() const
The function attributes are returned.
static AttributeList get(LLVMContext &C, ArrayRef< std::pair< unsigned, Attribute > > Attrs)
Create an AttributeList with the specified parameters in it.
bool isEmpty() const
Return true if there are no attributes.
Definition: Attributes.h:942
AttributeSet getRetAttrs() const
The attributes for the ret value are returned.
bool hasAttrSomewhere(Attribute::AttrKind Kind, unsigned *Index=nullptr) const
Return true if the specified attribute is set for at least one parameter or for the return value.
bool hasParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) const
Return true if the attribute exists for the given argument.
Definition: Attributes.h:764
AttributeSet getParamAttrs(unsigned ArgNo) const
The attributes for the argument or parameter at the given index are returned.
AttributeList addParamAttribute(LLVMContext &C, unsigned ArgNo, Attribute::AttrKind Kind) const
Add an argument attribute to the list.
Definition: Attributes.h:570
bool hasAttribute(Attribute::AttrKind Kind) const
Return true if the attribute exists in this set.
Definition: Attributes.cpp:760
AttributeSet removeAttributes(LLVMContext &C, const AttributeMask &AttrsToRemove) const
Remove the specified attributes from this set.
Definition: Attributes.cpp:745
static AttributeSet get(LLVMContext &C, const AttrBuilder &B)
Definition: Attributes.cpp:693
static Attribute get(LLVMContext &Context, AttrKind Kind, uint64_t Val=0)
Return a uniquified Attribute object.
Definition: Attributes.cpp:91
static Attribute getWithDereferenceableBytes(LLVMContext &Context, uint64_t Bytes)
Definition: Attributes.cpp:177
static Attribute getWithDereferenceableOrNullBytes(LLVMContext &Context, uint64_t Bytes)
Definition: Attributes.cpp:183
static Attribute getWithByValType(LLVMContext &Context, Type *Ty)
Definition: Attributes.cpp:189
static Attribute getWithAlignment(LLVMContext &Context, Align Alignment)
Return a uniquified Attribute object that has the specific alignment set.
Definition: Attributes.cpp:167
LLVM Basic Block Representation.
Definition: BasicBlock.h:56
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:314
InstListType::reverse_iterator reverse_iterator
Definition: BasicBlock.h:89
reverse_iterator rend()
Definition: BasicBlock.h:321
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:87
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.h:127
Value * getRHS() const
bool isSigned() const
Whether the intrinsic is signed or unsigned.
Instruction::BinaryOps getBinaryOp() const
Returns the binary operation underlying the intrinsic.
Value * getLHS() const
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
static BinaryOperator * CreateFDivFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:271
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
static BinaryOperator * CreateNSW(BinaryOps Opc, Value *V1, Value *V2, const Twine &Name="")
Definition: InstrTypes.h:282
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
static BinaryOperator * CreateNSWNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
static BinaryOperator * CreateWithCopiedFlags(BinaryOps Opc, Value *V1, Value *V2, Instruction *CopyO, const Twine &Name="", Instruction *InsertBefore=nullptr)
Definition: InstrTypes.h:248
static BinaryOperator * CreateFMulFMF(Value *V1, Value *V2, Instruction *FMFSource, const Twine &Name="")
Definition: InstrTypes.h:266
static BinaryOperator * CreateNUW(BinaryOps Opc, Value *V1, Value *V2, const Twine &Name="")
Definition: InstrTypes.h:301
Base class for all callable instructions (InvokeInst and CallInst) Holds everything related to callin...
Definition: InstrTypes.h:1184
void setCallingConv(CallingConv::ID CC)
Definition: InstrTypes.h:1469
bundle_op_iterator bundle_op_info_begin()
Return the start of the list of BundleOpInfo instances associated with this OperandBundleUser.
Definition: InstrTypes.h:2197
void setDoesNotThrow()
Definition: InstrTypes.h:1909
void getOperandBundlesAsDefs(SmallVectorImpl< OperandBundleDef > &Defs) const
Return the list of operand bundles attached to this instruction as a vector of OperandBundleDefs.
std::optional< OperandBundleUse > getOperandBundle(StringRef Name) const
Return an operand bundle by name, if present.
Definition: InstrTypes.h:2036
Function * getCalledFunction() const
Returns the function called, or null if this is an indirect function invocation or the function signa...
Definition: InstrTypes.h:1406
bool hasRetAttr(Attribute::AttrKind Kind) const
Determine whether the return value has the given attribute.
Definition: InstrTypes.h:1605
unsigned getNumOperandBundles() const
Return the number of operand bundles associated with this User.
Definition: InstrTypes.h:1949
CallingConv::ID getCallingConv() const
Definition: InstrTypes.h:1465
static CallBase * Create(CallBase *CB, ArrayRef< OperandBundleDef > Bundles, Instruction *InsertPt=nullptr)
Create a clone of CB with a different set of operand bundles and insert it before InsertPt.
static CallBase * removeOperandBundle(CallBase *CB, uint32_t ID, Instruction *InsertPt=nullptr)
Create a clone of CB with operand bundle ID removed.
Value * getCalledOperand() const
Definition: InstrTypes.h:1399
void setAttributes(AttributeList A)
Set the parameter attributes for this call.
Definition: InstrTypes.h:1488
bool doesNotThrow() const
Determine if the call cannot unwind.
Definition: InstrTypes.h:1908
void addRetAttr(Attribute::AttrKind Kind)
Adds the attribute to the return value.
Definition: InstrTypes.h:1526
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1351
FunctionType * getFunctionType() const
Definition: InstrTypes.h:1264
Intrinsic::ID getIntrinsicID() const
Returns the intrinsic ID of the intrinsic called or Intrinsic::not_intrinsic if the called function i...
iterator_range< User::op_iterator > args()
Iteration adapter for range-for loops.
Definition: InstrTypes.h:1342
unsigned arg_size() const
Definition: InstrTypes.h:1349
bool hasOperandBundles() const
Return true if this User has any operand bundles.
Definition: InstrTypes.h:1954
void setCalledFunction(Function *Fn)
Sets the function called, including updating the function type.
Definition: InstrTypes.h:1445
CallBr instruction, tracking function calls that may not return control but instead transfer it to a ...
static CallBrInst * Create(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest, ArrayRef< BasicBlock * > IndirectDests, ArrayRef< Value * > Args, const Twine &NameStr, Instruction *InsertBefore=nullptr)
This class represents a function call, abstracting a target machine's calling convention.
bool isNoTailCall() const
void setTailCallKind(TailCallKind TCK)
bool isMustTailCall() const
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:428
static Instruction::CastOps getCastOpcode(const Value *Val, bool SrcIsSigned, Type *Ty, bool DstIsSigned)
Returns the opcode necessary to cast Val into Ty using usual casting rules.
static CastInst * CreateIntegerCast(Value *S, Type *Ty, bool isSigned, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a ZExt, BitCast, or Trunc for int -> int casts.
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
bool isLosslessCast() const
A lossless cast is one that does not alter the basic value.
static bool isBitOrNoopPointerCastable(Type *SrcTy, Type *DestTy, const DataLayout &DL)
Check whether a bitcast, inttoptr, or ptrtoint cast between these types is valid and a no-op.
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:1054
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:718
@ ICMP_SLT
signed less than
Definition: InstrTypes.h:747
@ ICMP_UGT
unsigned greater than
Definition: InstrTypes.h:741
@ ICMP_SGT
signed greater than
Definition: InstrTypes.h:745
@ ICMP_ULT
unsigned less than
Definition: InstrTypes.h:743
@ ICMP_EQ
equal
Definition: InstrTypes.h:739
@ ICMP_NE
not equal
Definition: InstrTypes.h:740
static ConstantAggregateZero * get(Type *Ty)
Definition: Constants.cpp:1595
static ConstantAsMetadata * get(Constant *C)
Definition: Metadata.h:419
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2664
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2119
static Constant * getICmp(unsigned short pred, Constant *LHS, Constant *RHS, bool OnlyIfReduced=false)
get* - Return some common constants without having to specify the full Instruction::OPCODE identifier...
Definition: Constants.cpp:2523
static Constant * getSelect(Constant *C, Constant *V1, Constant *V2, Type *OnlyIfReducedTy=nullptr)
Select constant expr.
Definition: Constants.cpp:2441
static Constant * getIntegerCast(Constant *C, Type *Ty, bool IsSigned)
Create a ZExt, Bitcast or Trunc for integer -> integer casts.
Definition: Constants.cpp:2067
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2105
static Constant * getMul(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2671
static Constant * getBitCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2229
static Constant * getTrunc(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2091
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2644
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:934
This is the shared class of boolean and integer constants.
Definition: Constants.h:78
uint64_t getLimitedValue(uint64_t Limit=~0ULL) const
getLimitedValue - If the value is smaller than the specified limit, return it, otherwise return the l...
Definition: Constants.h:243
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:835
static Constant * get(Type *Ty, uint64_t V, bool IsSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:887
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:141
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:132
static ConstantInt * getBool(LLVMContext &Context, bool V)
Definition: Constants.cpp:849
static ConstantPointerNull * get(PointerType *T)
Static factory methods - Return objects of the specified value.
Definition: Constants.cpp:1707
static Constant * get(StructType *T, ArrayRef< Constant * > V)
Definition: Constants.cpp:1314
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:403
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:356
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:114
TypeSize getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:507
unsigned size() const
Definition: DenseMap.h:99
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:145
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:166
This class represents an extension of floating point types.
bool noSignedZeros() const
Definition: FMF.h:69
An instruction for ordering other memory operations.
Definition: Instructions.h:436
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this fence instruction.
Definition: Instructions.h:472
AtomicOrdering getOrdering() const
Returns the ordering constraint of this fence instruction.
Definition: Instructions.h:461
FunctionType * getFunctionType()
Definition: DerivedTypes.h:182
Class to represent function types.
Definition: DerivedTypes.h:103
Type::subtype_iterator param_iterator
Definition: DerivedTypes.h:126
static FunctionType * get(Type *Result, ArrayRef< Type * > Params, bool isVarArg)
This static method is the primary way of constructing a FunctionType.
bool isConvergent() const
Determine if the call is convergent.
Definition: Function.h:551
FunctionType * getFunctionType() const
Returns the FunctionType for me.
Definition: Function.h:174
CallingConv::ID getCallingConv() const
getCallingConv()/setCallingConv(CC) - These method get and set the calling convention of this functio...
Definition: Function.h:237
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:313
bool doesNotThrow() const
Determine if the function cannot unwind.
Definition: Function.h:535
bool isIntrinsic() const
isIntrinsic - Returns true if the function's name starts with "llvm.".
Definition: Function.h:209
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.cpp:640
Represents calls to the gc.relocate intrinsic.
Value * getBasePtr() const
unsigned getBasePtrIndex() const
The index into the associate statepoint's argument list which contains the base pointer of the pointe...
Value * getDerivedPtr() const
unsigned getDerivedPtrIndex() const
The index into the associate statepoint's argument list which contains the pointer whose relocation t...
Represents a gc.statepoint intrinsic call.
Definition: Statepoint.h:61
std::vector< const GCRelocateInst * > getGCRelocates() const
Get list of all gc reloactes linked to this statepoint May contain several relocations for the same b...
Definition: Statepoint.h:206
MDNode * getMetadata(unsigned KindID) const
Get the current metadata attachments for the given kind, if any.
Definition: Metadata.cpp:1288
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:275
PointerType * getType() const
Global values are always pointers.
Definition: GlobalValue.h:290
CallInst * CreateUnaryIntrinsic(Intrinsic::ID ID, Value *V, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 1 operand which is mangled on its type.
Definition: IRBuilder.cpp:948
CallBrInst * CreateCallBr(FunctionType *Ty, Value *Callee, BasicBlock *DefaultDest, ArrayRef< BasicBlock * > IndirectDests, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="")
Create a callbr instruction.
Definition: IRBuilder.h:1112
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1912
Value * CreateVScale(Constant *Scaling, const Twine &Name="")
Create a call to llvm.vscale, multiplied by Scaling.
Definition: IRBuilder.cpp:97
Value * CreateLaunderInvariantGroup(Value *Ptr)
Create a launder.invariant.group intrinsic call.
Definition: IRBuilder.cpp:1156
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1631
IntegerType * getInt1Ty()
Fetch the type representing a single bit.
Definition: IRBuilder.h:497
Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")
Definition: IRBuilder.h:2341
IntegerType * getIntNTy(unsigned N)
Fetch the type representing an N-bit integer.
Definition: IRBuilder.h:525
LoadInst * CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, const char *Name)
Definition: IRBuilder.h:1722
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1926
Value * CreateFAdd(Value *L, Value *R, const Twine &Name="", MDNode *FPMD=nullptr)
Definition: IRBuilder.h:1448
CallInst * CreateAndReduce(Value *Src)
Create a vector int AND reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:445
Value * CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name="")
Return a vector value that contains.
Definition: IRBuilder.cpp:1249
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:452
CallInst * CreateIntrinsic(Intrinsic::ID ID, ArrayRef< Type * > Types, ArrayRef< Value * > Args, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with Args, mangled using Types.
Definition: IRBuilder.cpp:965
Value * CreateFNegFMF(Value *V, Instruction *FMFSource, const Twine &Name="")
Copy fast-math-flags from an instruction rather than using the builder's default FMF.
Definition: IRBuilder.h:1655
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.cpp:1126
InvokeInst * CreateInvoke(FunctionType *Ty, Value *Callee, BasicBlock *NormalDest, BasicBlock *UnwindDest, ArrayRef< Value * > Args, ArrayRef< OperandBundleDef > OpBundles, const Twine &Name="")
Create an invoke instruction.
Definition: IRBuilder.h:1073
CallInst * CreateAddReduce(Value *Src)
Create a vector int add reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:437
Value * CreateLShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1352
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:512
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:297
Value * CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2126
CallInst * CreateOrReduce(Value *Src)
Create a vector int OR reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:449
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1916
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Definition: IRBuilder.h:472
Value * CreateBitOrPointerCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2086
Value * CreateNot(Value *V, const Twine &Name="")
Definition: IRBuilder.h:1664
Value * CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2122
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1259
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2008
LoadInst * CreateLoad(Type *Ty, Value *Ptr, const char *Name)
Provided to resolve 'CreateLoad(Ty, Ptr, "...")' correctly, instead of converting the string to 'bool...
Definition: IRBuilder.h:1705
CallInst * CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 2 operands which is mangled on the first type.
Definition: IRBuilder.cpp:956
Value * CreateShuffleVector(Value *V1, Value *V2, Value *Mask, const Twine &Name="")
Definition: IRBuilder.h:2375
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1390
StoreInst * CreateStore(Value *Val, Value *Ptr, bool isVolatile=false)
Definition: IRBuilder.h:1718
ConstantInt * getFalse()
Get the constant value for i1 false.
Definition: IRBuilder.h:457
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2430
Value * CreateCast(Instruction::CastOps Op, Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2045
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2293
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2232
Value * CreateFMul(Value *L, Value *R, const Twine &Name="", MDNode *FPMD=nullptr)
Definition: IRBuilder.h:1502
Value * CreateFNeg(Value *V, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1645
Value * CreateAddrSpaceCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:2013
Value * CreateStripInvariantGroup(Value *Ptr)
Create a strip.invariant.group intrinsic call.
Definition: IRBuilder.cpp:1180
static InsertValueInst * Create(Value *Agg, Value *Val, ArrayRef< unsigned > Idxs, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Instruction * FoldOpIntoSelect(Instruction &Op, SelectInst *SI, bool FoldWithMultiUse=false)
Given an instruction with a select as one operand and a constant as the other operand,...
bool SimplifyDemandedBits(Instruction *I, unsigned Op, const APInt &DemandedMask, KnownBits &Known, unsigned Depth=0) override
This form of SimplifyDemandedBits simplifies the specified instruction operand if possible,...
Value * SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, APInt &UndefElts, unsigned Depth=0, bool AllowMultipleUsers=false) override
The specified value produces a vector with any number of elements.
Instruction * SimplifyAnyMemSet(AnyMemSetInst *MI)
Instruction * visitFree(CallInst &FI, Value *FreedOp)
Instruction * visitCallBrInst(CallBrInst &CBI)
Instruction * eraseInstFromFunction(Instruction &I) override
Combiner aware instruction erasure.
Instruction * visitFenceInst(FenceInst &FI)
Instruction * visitInvokeInst(InvokeInst &II)
bool SimplifyDemandedInstructionBits(Instruction &Inst)
Tries to simplify operands to an integer instruction based on its demanded bits.
void CreateNonTerminatorUnreachable(Instruction *InsertAt)
Create and insert the idiom we use to indicate a block is unreachable without having to rewrite the C...
Instruction * visitVAEndInst(VAEndInst &I)
Instruction * visitAllocSite(Instruction &FI)
Instruction * SimplifyAnyMemTransfer(AnyMemTransferInst *MI)
OverflowResult computeOverflow(Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS, Instruction *CxtI) const
Instruction * visitCallInst(CallInst &CI)
CallInst simplification.
const DataLayout & getDataLayout() const
Definition: InstCombiner.h:372
DominatorTree & getDominatorTree() const
Definition: InstCombiner.h:371
BlockFrequencyInfo * BFI
Definition: InstCombiner.h:76
TargetLibraryInfo & TLI
Definition: InstCombiner.h:71
bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero=false, unsigned Depth=0, const Instruction *CxtI=nullptr)
Definition: InstCombiner.h:471
AAResults * AA
Definition: InstCombiner.h:67
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:418
static bool isFreeToInvert(Value *V, bool WillInvertAllUses)
Return true if the specified value is free to invert (apply ~ to).
Definition: InstCombiner.h:235
void replaceUse(Use &U, Value *NewValue)
Replace use and add the previously used value to the worklist.
Definition: InstCombiner.h:449
const SimplifyQuery SQ
Definition: InstCombiner.h:74
InstructionWorklist & Worklist
A worklist of the instructions that need to be simplified.
Definition: InstCombiner.h:62
const DataLayout & DL
Definition: InstCombiner.h:73
Instruction * InsertNewInstBefore(Instruction *New, Instruction &Old)
Inserts an instruction New before instruction Old.
Definition: InstCombiner.h:397
std::optional< Instruction * > targetInstCombineIntrinsic(IntrinsicInst &II)
AssumptionCache & AC
Definition: InstCombiner.h:70
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:442
DominatorTree & DT
Definition: InstCombiner.h:72
ProfileSummaryInfo * PSI
Definition: InstCombiner.h:77