LLVM  8.0.0svn
InstCombineCalls.cpp
Go to the documentation of this file.
1 //===- InstCombineCalls.cpp -----------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitCall and visitInvoke functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/Twine.h"
29 #include "llvm/IR/Attributes.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CallSite.h"
32 #include "llvm/IR/Constant.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/GlobalVariable.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instruction.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/IR/Statepoint.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/User.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/IR/ValueHandle.h"
52 #include "llvm/Support/Casting.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/KnownBits.h"
62 #include <algorithm>
63 #include <cassert>
64 #include <cstdint>
65 #include <cstring>
66 #include <utility>
67 #include <vector>
68 
69 using namespace llvm;
70 using namespace PatternMatch;
71 
72 #define DEBUG_TYPE "instcombine"
73 
74 STATISTIC(NumSimplified, "Number of library calls simplified");
75 
77  "instcombine-guard-widening-window",
78  cl::init(3),
79  cl::desc("How wide an instruction window to bypass looking for "
80  "another guard"));
81 
82 /// Return the specified type promoted as it would be to pass though a va_arg
83 /// area.
84 static Type *getPromotedType(Type *Ty) {
85  if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
86  if (ITy->getBitWidth() < 32)
87  return Type::getInt32Ty(Ty->getContext());
88  }
89  return Ty;
90 }
91 
92 /// Return a constant boolean vector that has true elements in all positions
93 /// where the input constant data vector has an element with the sign bit set.
96  IntegerType *BoolTy = Type::getInt1Ty(V->getContext());
97  for (unsigned I = 0, E = V->getNumElements(); I != E; ++I) {
98  Constant *Elt = V->getElementAsConstant(I);
99  assert((isa<ConstantInt>(Elt) || isa<ConstantFP>(Elt)) &&
100  "Unexpected constant data vector element type");
101  bool Sign = V->getElementType()->isIntegerTy()
102  ? cast<ConstantInt>(Elt)->isNegative()
103  : cast<ConstantFP>(Elt)->isNegative();
104  BoolVec.push_back(ConstantInt::get(BoolTy, Sign));
105  }
106  return ConstantVector::get(BoolVec);
107 }
108 
109 Instruction *InstCombiner::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) {
110  unsigned DstAlign = getKnownAlignment(MI->getRawDest(), DL, MI, &AC, &DT);
111  unsigned CopyDstAlign = MI->getDestAlignment();
112  if (CopyDstAlign < DstAlign){
113  MI->setDestAlignment(DstAlign);
114  return MI;
115  }
116 
117  unsigned SrcAlign = getKnownAlignment(MI->getRawSource(), DL, MI, &AC, &DT);
118  unsigned CopySrcAlign = MI->getSourceAlignment();
119  if (CopySrcAlign < SrcAlign) {
120  MI->setSourceAlignment(SrcAlign);
121  return MI;
122  }
123 
124  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
125  // load/store.
126  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getLength());
127  if (!MemOpLength) return nullptr;
128 
129  // Source and destination pointer types are always "i8*" for intrinsic. See
130  // if the size is something we can handle with a single primitive load/store.
131  // A single load+store correctly handles overlapping memory in the memmove
132  // case.
133  uint64_t Size = MemOpLength->getLimitedValue();
134  assert(Size && "0-sized memory transferring should be removed already.");
135 
136  if (Size > 8 || (Size&(Size-1)))
137  return nullptr; // If not 1/2/4/8 bytes, exit.
138 
139  // Use an integer load+store unless we can find something better.
140  unsigned SrcAddrSp =
141  cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
142  unsigned DstAddrSp =
143  cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
144 
145  IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
146  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
147  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
148 
149  // If the memcpy has metadata describing the members, see if we can get the
150  // TBAA tag describing our copy.
151  MDNode *CopyMD = nullptr;
152  if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa)) {
153  CopyMD = M;
154  } else if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
155  if (M->getNumOperands() == 3 && M->getOperand(0) &&
156  mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
157  mdconst::extract<ConstantInt>(M->getOperand(0))->isZero() &&
158  M->getOperand(1) &&
159  mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
160  mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
161  Size &&
162  M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
163  CopyMD = cast<MDNode>(M->getOperand(2));
164  }
165 
166  Value *Src = Builder.CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
167  Value *Dest = Builder.CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
168  LoadInst *L = Builder.CreateLoad(Src);
169  // Alignment from the mem intrinsic will be better, so use it.
170  L->setAlignment(CopySrcAlign);
171  if (CopyMD)
172  L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
173  MDNode *LoopMemParallelMD =
175  if (LoopMemParallelMD)
177 
178  StoreInst *S = Builder.CreateStore(L, Dest);
179  // Alignment from the mem intrinsic will be better, so use it.
180  S->setAlignment(CopyDstAlign);
181  if (CopyMD)
182  S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
183  if (LoopMemParallelMD)
185 
186  if (auto *MT = dyn_cast<MemTransferInst>(MI)) {
187  // non-atomics can be volatile
188  L->setVolatile(MT->isVolatile());
189  S->setVolatile(MT->isVolatile());
190  }
191  if (isa<AtomicMemTransferInst>(MI)) {
192  // atomics have to be unordered
195  }
196 
197  // Set the size of the copy to 0, it will be deleted on the next iteration.
198  MI->setLength(Constant::getNullValue(MemOpLength->getType()));
199  return MI;
200 }
201 
202 Instruction *InstCombiner::SimplifyAnyMemSet(AnyMemSetInst *MI) {
203  unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, &AC, &DT);
204  if (MI->getDestAlignment() < Alignment) {
205  MI->setDestAlignment(Alignment);
206  return MI;
207  }
208 
209  // Extract the length and alignment and fill if they are constant.
210  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
211  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
212  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
213  return nullptr;
214  uint64_t Len = LenC->getLimitedValue();
215  Alignment = MI->getDestAlignment();
216  assert(Len && "0-sized memory setting should be removed already.");
217 
218  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
219  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
220  Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
221 
222  Value *Dest = MI->getDest();
223  unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
224  Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
225  Dest = Builder.CreateBitCast(Dest, NewDstPtrTy);
226 
227  // Alignment 0 is identity for alignment 1 for memset, but not store.
228  if (Alignment == 0) Alignment = 1;
229 
230  // Extract the fill value and store.
231  uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
232  StoreInst *S = Builder.CreateStore(ConstantInt::get(ITy, Fill), Dest,
233  MI->isVolatile());
234  S->setAlignment(Alignment);
235  if (isa<AtomicMemSetInst>(MI))
237 
238  // Set the size of the copy to 0, it will be deleted on the next iteration.
239  MI->setLength(Constant::getNullValue(LenC->getType()));
240  return MI;
241  }
242 
243  return nullptr;
244 }
245 
247  InstCombiner::BuilderTy &Builder) {
248  bool IsAddition;
249 
250  switch (II.getIntrinsicID()) {
251  default: llvm_unreachable("Unexpected intrinsic!");
252  case Intrinsic::x86_sse2_padds_b:
253  case Intrinsic::x86_sse2_padds_w:
254  case Intrinsic::x86_avx2_padds_b:
255  case Intrinsic::x86_avx2_padds_w:
256  case Intrinsic::x86_avx512_padds_b_512:
257  case Intrinsic::x86_avx512_padds_w_512:
258  IsAddition = true;
259  break;
260  case Intrinsic::x86_sse2_psubs_b:
261  case Intrinsic::x86_sse2_psubs_w:
262  case Intrinsic::x86_avx2_psubs_b:
263  case Intrinsic::x86_avx2_psubs_w:
264  case Intrinsic::x86_avx512_psubs_b_512:
265  case Intrinsic::x86_avx512_psubs_w_512:
266  IsAddition = false;
267  break;
268  }
269 
270  auto *Arg0 = dyn_cast<Constant>(II.getOperand(0));
271  auto *Arg1 = dyn_cast<Constant>(II.getOperand(1));
272  auto VT = cast<VectorType>(II.getType());
273  auto SVT = VT->getElementType();
274  unsigned NumElems = VT->getNumElements();
275 
276  if (!Arg0 || !Arg1)
277  return nullptr;
278 
280 
281  APInt MaxValue = APInt::getSignedMaxValue(SVT->getIntegerBitWidth());
282  APInt MinValue = APInt::getSignedMinValue(SVT->getIntegerBitWidth());
283  for (unsigned i = 0; i < NumElems; ++i) {
284  auto *Elt0 = Arg0->getAggregateElement(i);
285  auto *Elt1 = Arg1->getAggregateElement(i);
286  if (isa<UndefValue>(Elt0) || isa<UndefValue>(Elt1)) {
287  Result.push_back(UndefValue::get(SVT));
288  continue;
289  }
290 
291  if (!isa<ConstantInt>(Elt0) || !isa<ConstantInt>(Elt1))
292  return nullptr;
293 
294  const APInt &Val0 = cast<ConstantInt>(Elt0)->getValue();
295  const APInt &Val1 = cast<ConstantInt>(Elt1)->getValue();
296  bool Overflow = false;
297  APInt ResultElem = IsAddition ? Val0.sadd_ov(Val1, Overflow)
298  : Val0.ssub_ov(Val1, Overflow);
299  if (Overflow)
300  ResultElem = Val0.isNegative() ? MinValue : MaxValue;
301  Result.push_back(Constant::getIntegerValue(SVT, ResultElem));
302  }
303 
304  return ConstantVector::get(Result);
305 }
306 
308  InstCombiner::BuilderTy &Builder) {
309  bool LogicalShift = false;
310  bool ShiftLeft = false;
311 
312  switch (II.getIntrinsicID()) {
313  default: llvm_unreachable("Unexpected intrinsic!");
314  case Intrinsic::x86_sse2_psra_d:
315  case Intrinsic::x86_sse2_psra_w:
316  case Intrinsic::x86_sse2_psrai_d:
317  case Intrinsic::x86_sse2_psrai_w:
318  case Intrinsic::x86_avx2_psra_d:
319  case Intrinsic::x86_avx2_psra_w:
320  case Intrinsic::x86_avx2_psrai_d:
321  case Intrinsic::x86_avx2_psrai_w:
322  case Intrinsic::x86_avx512_psra_q_128:
323  case Intrinsic::x86_avx512_psrai_q_128:
324  case Intrinsic::x86_avx512_psra_q_256:
325  case Intrinsic::x86_avx512_psrai_q_256:
326  case Intrinsic::x86_avx512_psra_d_512:
327  case Intrinsic::x86_avx512_psra_q_512:
328  case Intrinsic::x86_avx512_psra_w_512:
329  case Intrinsic::x86_avx512_psrai_d_512:
330  case Intrinsic::x86_avx512_psrai_q_512:
331  case Intrinsic::x86_avx512_psrai_w_512:
332  LogicalShift = false; ShiftLeft = false;
333  break;
334  case Intrinsic::x86_sse2_psrl_d:
335  case Intrinsic::x86_sse2_psrl_q:
336  case Intrinsic::x86_sse2_psrl_w:
337  case Intrinsic::x86_sse2_psrli_d:
338  case Intrinsic::x86_sse2_psrli_q:
339  case Intrinsic::x86_sse2_psrli_w:
340  case Intrinsic::x86_avx2_psrl_d:
341  case Intrinsic::x86_avx2_psrl_q:
342  case Intrinsic::x86_avx2_psrl_w:
343  case Intrinsic::x86_avx2_psrli_d:
344  case Intrinsic::x86_avx2_psrli_q:
345  case Intrinsic::x86_avx2_psrli_w:
346  case Intrinsic::x86_avx512_psrl_d_512:
347  case Intrinsic::x86_avx512_psrl_q_512:
348  case Intrinsic::x86_avx512_psrl_w_512:
349  case Intrinsic::x86_avx512_psrli_d_512:
350  case Intrinsic::x86_avx512_psrli_q_512:
351  case Intrinsic::x86_avx512_psrli_w_512:
352  LogicalShift = true; ShiftLeft = false;
353  break;
354  case Intrinsic::x86_sse2_psll_d:
355  case Intrinsic::x86_sse2_psll_q:
356  case Intrinsic::x86_sse2_psll_w:
357  case Intrinsic::x86_sse2_pslli_d:
358  case Intrinsic::x86_sse2_pslli_q:
359  case Intrinsic::x86_sse2_pslli_w:
360  case Intrinsic::x86_avx2_psll_d:
361  case Intrinsic::x86_avx2_psll_q:
362  case Intrinsic::x86_avx2_psll_w:
363  case Intrinsic::x86_avx2_pslli_d:
364  case Intrinsic::x86_avx2_pslli_q:
365  case Intrinsic::x86_avx2_pslli_w:
366  case Intrinsic::x86_avx512_psll_d_512:
367  case Intrinsic::x86_avx512_psll_q_512:
368  case Intrinsic::x86_avx512_psll_w_512:
369  case Intrinsic::x86_avx512_pslli_d_512:
370  case Intrinsic::x86_avx512_pslli_q_512:
371  case Intrinsic::x86_avx512_pslli_w_512:
372  LogicalShift = true; ShiftLeft = true;
373  break;
374  }
375  assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
376 
377  // Simplify if count is constant.
378  auto Arg1 = II.getArgOperand(1);
379  auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1);
380  auto CDV = dyn_cast<ConstantDataVector>(Arg1);
381  auto CInt = dyn_cast<ConstantInt>(Arg1);
382  if (!CAZ && !CDV && !CInt)
383  return nullptr;
384 
385  APInt Count(64, 0);
386  if (CDV) {
387  // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
388  // operand to compute the shift amount.
389  auto VT = cast<VectorType>(CDV->getType());
390  unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits();
391  assert((64 % BitWidth) == 0 && "Unexpected packed shift size");
392  unsigned NumSubElts = 64 / BitWidth;
393 
394  // Concatenate the sub-elements to create the 64-bit value.
395  for (unsigned i = 0; i != NumSubElts; ++i) {
396  unsigned SubEltIdx = (NumSubElts - 1) - i;
397  auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
398  Count <<= BitWidth;
399  Count |= SubElt->getValue().zextOrTrunc(64);
400  }
401  }
402  else if (CInt)
403  Count = CInt->getValue();
404 
405  auto Vec = II.getArgOperand(0);
406  auto VT = cast<VectorType>(Vec->getType());
407  auto SVT = VT->getElementType();
408  unsigned VWidth = VT->getNumElements();
409  unsigned BitWidth = SVT->getPrimitiveSizeInBits();
410 
411  // If shift-by-zero then just return the original value.
412  if (Count.isNullValue())
413  return Vec;
414 
415  // Handle cases when Shift >= BitWidth.
416  if (Count.uge(BitWidth)) {
417  // If LogicalShift - just return zero.
418  if (LogicalShift)
419  return ConstantAggregateZero::get(VT);
420 
421  // If ArithmeticShift - clamp Shift to (BitWidth - 1).
422  Count = APInt(64, BitWidth - 1);
423  }
424 
425  // Get a constant vector of the same type as the first operand.
426  auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
427  auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
428 
429  if (ShiftLeft)
430  return Builder.CreateShl(Vec, ShiftVec);
431 
432  if (LogicalShift)
433  return Builder.CreateLShr(Vec, ShiftVec);
434 
435  return Builder.CreateAShr(Vec, ShiftVec);
436 }
437 
438 // Attempt to simplify AVX2 per-element shift intrinsics to a generic IR shift.
439 // Unlike the generic IR shifts, the intrinsics have defined behaviour for out
440 // of range shift amounts (logical - set to zero, arithmetic - splat sign bit).
442  InstCombiner::BuilderTy &Builder) {
443  bool LogicalShift = false;
444  bool ShiftLeft = false;
445 
446  switch (II.getIntrinsicID()) {
447  default: llvm_unreachable("Unexpected intrinsic!");
448  case Intrinsic::x86_avx2_psrav_d:
449  case Intrinsic::x86_avx2_psrav_d_256:
450  case Intrinsic::x86_avx512_psrav_q_128:
451  case Intrinsic::x86_avx512_psrav_q_256:
452  case Intrinsic::x86_avx512_psrav_d_512:
453  case Intrinsic::x86_avx512_psrav_q_512:
454  case Intrinsic::x86_avx512_psrav_w_128:
455  case Intrinsic::x86_avx512_psrav_w_256:
456  case Intrinsic::x86_avx512_psrav_w_512:
457  LogicalShift = false;
458  ShiftLeft = false;
459  break;
460  case Intrinsic::x86_avx2_psrlv_d:
461  case Intrinsic::x86_avx2_psrlv_d_256:
462  case Intrinsic::x86_avx2_psrlv_q:
463  case Intrinsic::x86_avx2_psrlv_q_256:
464  case Intrinsic::x86_avx512_psrlv_d_512:
465  case Intrinsic::x86_avx512_psrlv_q_512:
466  case Intrinsic::x86_avx512_psrlv_w_128:
467  case Intrinsic::x86_avx512_psrlv_w_256:
468  case Intrinsic::x86_avx512_psrlv_w_512:
469  LogicalShift = true;
470  ShiftLeft = false;
471  break;
472  case Intrinsic::x86_avx2_psllv_d:
473  case Intrinsic::x86_avx2_psllv_d_256:
474  case Intrinsic::x86_avx2_psllv_q:
475  case Intrinsic::x86_avx2_psllv_q_256:
476  case Intrinsic::x86_avx512_psllv_d_512:
477  case Intrinsic::x86_avx512_psllv_q_512:
478  case Intrinsic::x86_avx512_psllv_w_128:
479  case Intrinsic::x86_avx512_psllv_w_256:
480  case Intrinsic::x86_avx512_psllv_w_512:
481  LogicalShift = true;
482  ShiftLeft = true;
483  break;
484  }
485  assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
486 
487  // Simplify if all shift amounts are constant/undef.
488  auto *CShift = dyn_cast<Constant>(II.getArgOperand(1));
489  if (!CShift)
490  return nullptr;
491 
492  auto Vec = II.getArgOperand(0);
493  auto VT = cast<VectorType>(II.getType());
494  auto SVT = VT->getVectorElementType();
495  int NumElts = VT->getNumElements();
496  int BitWidth = SVT->getIntegerBitWidth();
497 
498  // Collect each element's shift amount.
499  // We also collect special cases: UNDEF = -1, OUT-OF-RANGE = BitWidth.
500  bool AnyOutOfRange = false;
501  SmallVector<int, 8> ShiftAmts;
502  for (int I = 0; I < NumElts; ++I) {
503  auto *CElt = CShift->getAggregateElement(I);
504  if (CElt && isa<UndefValue>(CElt)) {
505  ShiftAmts.push_back(-1);
506  continue;
507  }
508 
509  auto *COp = dyn_cast_or_null<ConstantInt>(CElt);
510  if (!COp)
511  return nullptr;
512 
513  // Handle out of range shifts.
514  // If LogicalShift - set to BitWidth (special case).
515  // If ArithmeticShift - set to (BitWidth - 1) (sign splat).
516  APInt ShiftVal = COp->getValue();
517  if (ShiftVal.uge(BitWidth)) {
518  AnyOutOfRange = LogicalShift;
519  ShiftAmts.push_back(LogicalShift ? BitWidth : BitWidth - 1);
520  continue;
521  }
522 
523  ShiftAmts.push_back((int)ShiftVal.getZExtValue());
524  }
525 
526  // If all elements out of range or UNDEF, return vector of zeros/undefs.
527  // ArithmeticShift should only hit this if they are all UNDEF.
528  auto OutOfRange = [&](int Idx) { return (Idx < 0) || (BitWidth <= Idx); };
529  if (llvm::all_of(ShiftAmts, OutOfRange)) {
530  SmallVector<Constant *, 8> ConstantVec;
531  for (int Idx : ShiftAmts) {
532  if (Idx < 0) {
533  ConstantVec.push_back(UndefValue::get(SVT));
534  } else {
535  assert(LogicalShift && "Logical shift expected");
536  ConstantVec.push_back(ConstantInt::getNullValue(SVT));
537  }
538  }
539  return ConstantVector::get(ConstantVec);
540  }
541 
542  // We can't handle only some out of range values with generic logical shifts.
543  if (AnyOutOfRange)
544  return nullptr;
545 
546  // Build the shift amount constant vector.
547  SmallVector<Constant *, 8> ShiftVecAmts;
548  for (int Idx : ShiftAmts) {
549  if (Idx < 0)
550  ShiftVecAmts.push_back(UndefValue::get(SVT));
551  else
552  ShiftVecAmts.push_back(ConstantInt::get(SVT, Idx));
553  }
554  auto ShiftVec = ConstantVector::get(ShiftVecAmts);
555 
556  if (ShiftLeft)
557  return Builder.CreateShl(Vec, ShiftVec);
558 
559  if (LogicalShift)
560  return Builder.CreateLShr(Vec, ShiftVec);
561 
562  return Builder.CreateAShr(Vec, ShiftVec);
563 }
564 
565 static Value *simplifyX86pack(IntrinsicInst &II, bool IsSigned) {
566  Value *Arg0 = II.getArgOperand(0);
567  Value *Arg1 = II.getArgOperand(1);
568  Type *ResTy = II.getType();
569 
570  // Fast all undef handling.
571  if (isa<UndefValue>(Arg0) && isa<UndefValue>(Arg1))
572  return UndefValue::get(ResTy);
573 
574  Type *ArgTy = Arg0->getType();
575  unsigned NumLanes = ResTy->getPrimitiveSizeInBits() / 128;
576  unsigned NumDstElts = ResTy->getVectorNumElements();
577  unsigned NumSrcElts = ArgTy->getVectorNumElements();
578  assert(NumDstElts == (2 * NumSrcElts) && "Unexpected packing types");
579 
580  unsigned NumDstEltsPerLane = NumDstElts / NumLanes;
581  unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes;
582  unsigned DstScalarSizeInBits = ResTy->getScalarSizeInBits();
583  assert(ArgTy->getScalarSizeInBits() == (2 * DstScalarSizeInBits) &&
584  "Unexpected packing types");
585 
586  // Constant folding.
587  auto *Cst0 = dyn_cast<Constant>(Arg0);
588  auto *Cst1 = dyn_cast<Constant>(Arg1);
589  if (!Cst0 || !Cst1)
590  return nullptr;
591 
593  for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
594  for (unsigned Elt = 0; Elt != NumDstEltsPerLane; ++Elt) {
595  unsigned SrcIdx = Lane * NumSrcEltsPerLane + Elt % NumSrcEltsPerLane;
596  auto *Cst = (Elt >= NumSrcEltsPerLane) ? Cst1 : Cst0;
597  auto *COp = Cst->getAggregateElement(SrcIdx);
598  if (COp && isa<UndefValue>(COp)) {
599  Vals.push_back(UndefValue::get(ResTy->getScalarType()));
600  continue;
601  }
602 
603  auto *CInt = dyn_cast_or_null<ConstantInt>(COp);
604  if (!CInt)
605  return nullptr;
606 
607  APInt Val = CInt->getValue();
608  assert(Val.getBitWidth() == ArgTy->getScalarSizeInBits() &&
609  "Unexpected constant bitwidth");
610 
611  if (IsSigned) {
612  // PACKSS: Truncate signed value with signed saturation.
613  // Source values less than dst minint are saturated to minint.
614  // Source values greater than dst maxint are saturated to maxint.
615  if (Val.isSignedIntN(DstScalarSizeInBits))
616  Val = Val.trunc(DstScalarSizeInBits);
617  else if (Val.isNegative())
618  Val = APInt::getSignedMinValue(DstScalarSizeInBits);
619  else
620  Val = APInt::getSignedMaxValue(DstScalarSizeInBits);
621  } else {
622  // PACKUS: Truncate signed value with unsigned saturation.
623  // Source values less than zero are saturated to zero.
624  // Source values greater than dst maxuint are saturated to maxuint.
625  if (Val.isIntN(DstScalarSizeInBits))
626  Val = Val.trunc(DstScalarSizeInBits);
627  else if (Val.isNegative())
628  Val = APInt::getNullValue(DstScalarSizeInBits);
629  else
630  Val = APInt::getAllOnesValue(DstScalarSizeInBits);
631  }
632 
633  Vals.push_back(ConstantInt::get(ResTy->getScalarType(), Val));
634  }
635  }
636 
637  return ConstantVector::get(Vals);
638 }
639 
640 // Replace X86-specific intrinsics with generic floor-ceil where applicable.
642  InstCombiner::BuilderTy &Builder) {
643  ConstantInt *Arg = nullptr;
644  Intrinsic::ID IntrinsicID = II.getIntrinsicID();
645 
646  if (IntrinsicID == Intrinsic::x86_sse41_round_ss ||
647  IntrinsicID == Intrinsic::x86_sse41_round_sd)
648  Arg = dyn_cast<ConstantInt>(II.getArgOperand(2));
649  else if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss ||
650  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd)
651  Arg = dyn_cast<ConstantInt>(II.getArgOperand(4));
652  else
653  Arg = dyn_cast<ConstantInt>(II.getArgOperand(1));
654  if (!Arg)
655  return nullptr;
656  unsigned RoundControl = Arg->getZExtValue();
657 
658  Arg = nullptr;
659  unsigned SAE = 0;
660  if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_512 ||
661  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_512)
662  Arg = dyn_cast<ConstantInt>(II.getArgOperand(4));
663  else if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss ||
664  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd)
665  Arg = dyn_cast<ConstantInt>(II.getArgOperand(5));
666  else
667  SAE = 4;
668  if (!SAE) {
669  if (!Arg)
670  return nullptr;
671  SAE = Arg->getZExtValue();
672  }
673 
674  if (SAE != 4 || (RoundControl != 2 /*ceil*/ && RoundControl != 1 /*floor*/))
675  return nullptr;
676 
677  Value *Src, *Dst, *Mask;
678  bool IsScalar = false;
679  if (IntrinsicID == Intrinsic::x86_sse41_round_ss ||
680  IntrinsicID == Intrinsic::x86_sse41_round_sd ||
681  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss ||
682  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) {
683  IsScalar = true;
684  if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss ||
685  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) {
686  Mask = II.getArgOperand(3);
687  Value *Zero = Constant::getNullValue(Mask->getType());
688  Mask = Builder.CreateAnd(Mask, 1);
689  Mask = Builder.CreateICmp(ICmpInst::ICMP_NE, Mask, Zero);
690  Dst = II.getArgOperand(2);
691  } else
692  Dst = II.getArgOperand(0);
693  Src = Builder.CreateExtractElement(II.getArgOperand(1), (uint64_t)0);
694  } else {
695  Src = II.getArgOperand(0);
696  if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_128 ||
697  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_256 ||
698  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_512 ||
699  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_128 ||
700  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_256 ||
701  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_512) {
702  Dst = II.getArgOperand(2);
703  Mask = II.getArgOperand(3);
704  } else {
705  Dst = Src;
707  Builder.getIntNTy(Src->getType()->getVectorNumElements()));
708  }
709  }
710 
711  Intrinsic::ID ID = (RoundControl == 2) ? Intrinsic::ceil : Intrinsic::floor;
712  Value *Res = Builder.CreateUnaryIntrinsic(ID, Src, &II);
713  if (!IsScalar) {
714  if (auto *C = dyn_cast<Constant>(Mask))
715  if (C->isAllOnesValue())
716  return Res;
717  auto *MaskTy = VectorType::get(
718  Builder.getInt1Ty(), cast<IntegerType>(Mask->getType())->getBitWidth());
719  Mask = Builder.CreateBitCast(Mask, MaskTy);
720  unsigned Width = Src->getType()->getVectorNumElements();
721  if (MaskTy->getVectorNumElements() > Width) {
722  uint32_t Indices[4];
723  for (unsigned i = 0; i != Width; ++i)
724  Indices[i] = i;
725  Mask = Builder.CreateShuffleVector(Mask, Mask,
726  makeArrayRef(Indices, Width));
727  }
728  return Builder.CreateSelect(Mask, Res, Dst);
729  }
730  if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss ||
731  IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) {
732  Dst = Builder.CreateExtractElement(Dst, (uint64_t)0);
733  Res = Builder.CreateSelect(Mask, Res, Dst);
734  Dst = II.getArgOperand(0);
735  }
736  return Builder.CreateInsertElement(Dst, Res, (uint64_t)0);
737 }
738 
740  InstCombiner::BuilderTy &Builder) {
741  Value *Arg = II.getArgOperand(0);
742  Type *ResTy = II.getType();
743  Type *ArgTy = Arg->getType();
744 
745  // movmsk(undef) -> zero as we must ensure the upper bits are zero.
746  if (isa<UndefValue>(Arg))
747  return Constant::getNullValue(ResTy);
748 
749  // We can't easily peek through x86_mmx types.
750  if (!ArgTy->isVectorTy())
751  return nullptr;
752 
753  if (auto *C = dyn_cast<Constant>(Arg)) {
754  // Extract signbits of the vector input and pack into integer result.
755  APInt Result(ResTy->getPrimitiveSizeInBits(), 0);
756  for (unsigned I = 0, E = ArgTy->getVectorNumElements(); I != E; ++I) {
757  auto *COp = C->getAggregateElement(I);
758  if (!COp)
759  return nullptr;
760  if (isa<UndefValue>(COp))
761  continue;
762 
763  auto *CInt = dyn_cast<ConstantInt>(COp);
764  auto *CFp = dyn_cast<ConstantFP>(COp);
765  if (!CInt && !CFp)
766  return nullptr;
767 
768  if ((CInt && CInt->isNegative()) || (CFp && CFp->isNegative()))
769  Result.setBit(I);
770  }
771  return Constant::getIntegerValue(ResTy, Result);
772  }
773 
774  // Look for a sign-extended boolean source vector as the argument to this
775  // movmsk. If the argument is bitcast, look through that, but make sure the
776  // source of that bitcast is still a vector with the same number of elements.
777  // TODO: We can also convert a bitcast with wider elements, but that requires
778  // duplicating the bool source sign bits to match the number of elements
779  // expected by the movmsk call.
780  Arg = peekThroughBitcast(Arg);
781  Value *X;
782  if (Arg->getType()->isVectorTy() &&
783  Arg->getType()->getVectorNumElements() == ArgTy->getVectorNumElements() &&
784  match(Arg, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
785  // call iM movmsk(sext <N x i1> X) --> zext (bitcast <N x i1> X to iN) to iM
786  unsigned NumElts = X->getType()->getVectorNumElements();
787  Type *ScalarTy = Type::getIntNTy(Arg->getContext(), NumElts);
788  Value *BC = Builder.CreateBitCast(X, ScalarTy);
789  return Builder.CreateZExtOrTrunc(BC, ResTy);
790  }
791 
792  return nullptr;
793 }
794 
796  InstCombiner::BuilderTy &Builder) {
797  auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2));
798  if (!CInt)
799  return nullptr;
800 
801  VectorType *VecTy = cast<VectorType>(II.getType());
802  assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
803 
804  // The immediate permute control byte looks like this:
805  // [3:0] - zero mask for each 32-bit lane
806  // [5:4] - select one 32-bit destination lane
807  // [7:6] - select one 32-bit source lane
808 
809  uint8_t Imm = CInt->getZExtValue();
810  uint8_t ZMask = Imm & 0xf;
811  uint8_t DestLane = (Imm >> 4) & 0x3;
812  uint8_t SourceLane = (Imm >> 6) & 0x3;
813 
815 
816  // If all zero mask bits are set, this was just a weird way to
817  // generate a zero vector.
818  if (ZMask == 0xf)
819  return ZeroVector;
820 
821  // Initialize by passing all of the first source bits through.
822  uint32_t ShuffleMask[4] = { 0, 1, 2, 3 };
823 
824  // We may replace the second operand with the zero vector.
825  Value *V1 = II.getArgOperand(1);
826 
827  if (ZMask) {
828  // If the zero mask is being used with a single input or the zero mask
829  // overrides the destination lane, this is a shuffle with the zero vector.
830  if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
831  (ZMask & (1 << DestLane))) {
832  V1 = ZeroVector;
833  // We may still move 32-bits of the first source vector from one lane
834  // to another.
835  ShuffleMask[DestLane] = SourceLane;
836  // The zero mask may override the previous insert operation.
837  for (unsigned i = 0; i < 4; ++i)
838  if ((ZMask >> i) & 0x1)
839  ShuffleMask[i] = i + 4;
840  } else {
841  // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
842  return nullptr;
843  }
844  } else {
845  // Replace the selected destination lane with the selected source lane.
846  ShuffleMask[DestLane] = SourceLane + 4;
847  }
848 
849  return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
850 }
851 
852 /// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding
853 /// or conversion to a shuffle vector.
855  ConstantInt *CILength, ConstantInt *CIIndex,
856  InstCombiner::BuilderTy &Builder) {
857  auto LowConstantHighUndef = [&](uint64_t Val) {
858  Type *IntTy64 = Type::getInt64Ty(II.getContext());
859  Constant *Args[] = {ConstantInt::get(IntTy64, Val),
860  UndefValue::get(IntTy64)};
861  return ConstantVector::get(Args);
862  };
863 
864  // See if we're dealing with constant values.
865  Constant *C0 = dyn_cast<Constant>(Op0);
866  ConstantInt *CI0 =
867  C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
868  : nullptr;
869 
870  // Attempt to constant fold.
871  if (CILength && CIIndex) {
872  // From AMD documentation: "The bit index and field length are each six
873  // bits in length other bits of the field are ignored."
874  APInt APIndex = CIIndex->getValue().zextOrTrunc(6);
875  APInt APLength = CILength->getValue().zextOrTrunc(6);
876 
877  unsigned Index = APIndex.getZExtValue();
878 
879  // From AMD documentation: "a value of zero in the field length is
880  // defined as length of 64".
881  unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
882 
883  // From AMD documentation: "If the sum of the bit index + length field
884  // is greater than 64, the results are undefined".
885  unsigned End = Index + Length;
886 
887  // Note that both field index and field length are 8-bit quantities.
888  // Since variables 'Index' and 'Length' are unsigned values
889  // obtained from zero-extending field index and field length
890  // respectively, their sum should never wrap around.
891  if (End > 64)
892  return UndefValue::get(II.getType());
893 
894  // If we are inserting whole bytes, we can convert this to a shuffle.
895  // Lowering can recognize EXTRQI shuffle masks.
896  if ((Length % 8) == 0 && (Index % 8) == 0) {
897  // Convert bit indices to byte indices.
898  Length /= 8;
899  Index /= 8;
900 
901  Type *IntTy8 = Type::getInt8Ty(II.getContext());
902  Type *IntTy32 = Type::getInt32Ty(II.getContext());
903  VectorType *ShufTy = VectorType::get(IntTy8, 16);
904 
905  SmallVector<Constant *, 16> ShuffleMask;
906  for (int i = 0; i != (int)Length; ++i)
907  ShuffleMask.push_back(
908  Constant::getIntegerValue(IntTy32, APInt(32, i + Index)));
909  for (int i = Length; i != 8; ++i)
910  ShuffleMask.push_back(
911  Constant::getIntegerValue(IntTy32, APInt(32, i + 16)));
912  for (int i = 8; i != 16; ++i)
913  ShuffleMask.push_back(UndefValue::get(IntTy32));
914 
915  Value *SV = Builder.CreateShuffleVector(
916  Builder.CreateBitCast(Op0, ShufTy),
917  ConstantAggregateZero::get(ShufTy), ConstantVector::get(ShuffleMask));
918  return Builder.CreateBitCast(SV, II.getType());
919  }
920 
921  // Constant Fold - shift Index'th bit to lowest position and mask off
922  // Length bits.
923  if (CI0) {
924  APInt Elt = CI0->getValue();
925  Elt.lshrInPlace(Index);
926  Elt = Elt.zextOrTrunc(Length);
927  return LowConstantHighUndef(Elt.getZExtValue());
928  }
929 
930  // If we were an EXTRQ call, we'll save registers if we convert to EXTRQI.
931  if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) {
932  Value *Args[] = {Op0, CILength, CIIndex};
933  Module *M = II.getModule();
934  Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi);
935  return Builder.CreateCall(F, Args);
936  }
937  }
938 
939  // Constant Fold - extraction from zero is always {zero, undef}.
940  if (CI0 && CI0->isZero())
941  return LowConstantHighUndef(0);
942 
943  return nullptr;
944 }
945 
946 /// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant
947 /// folding or conversion to a shuffle vector.
949  APInt APLength, APInt APIndex,
950  InstCombiner::BuilderTy &Builder) {
951  // From AMD documentation: "The bit index and field length are each six bits
952  // in length other bits of the field are ignored."
953  APIndex = APIndex.zextOrTrunc(6);
954  APLength = APLength.zextOrTrunc(6);
955 
956  // Attempt to constant fold.
957  unsigned Index = APIndex.getZExtValue();
958 
959  // From AMD documentation: "a value of zero in the field length is
960  // defined as length of 64".
961  unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
962 
963  // From AMD documentation: "If the sum of the bit index + length field
964  // is greater than 64, the results are undefined".
965  unsigned End = Index + Length;
966 
967  // Note that both field index and field length are 8-bit quantities.
968  // Since variables 'Index' and 'Length' are unsigned values
969  // obtained from zero-extending field index and field length
970  // respectively, their sum should never wrap around.
971  if (End > 64)
972  return UndefValue::get(II.getType());
973 
974  // If we are inserting whole bytes, we can convert this to a shuffle.
975  // Lowering can recognize INSERTQI shuffle masks.
976  if ((Length % 8) == 0 && (Index % 8) == 0) {
977  // Convert bit indices to byte indices.
978  Length /= 8;
979  Index /= 8;
980 
981  Type *IntTy8 = Type::getInt8Ty(II.getContext());
982  Type *IntTy32 = Type::getInt32Ty(II.getContext());
983  VectorType *ShufTy = VectorType::get(IntTy8, 16);
984 
985  SmallVector<Constant *, 16> ShuffleMask;
986  for (int i = 0; i != (int)Index; ++i)
987  ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i)));
988  for (int i = 0; i != (int)Length; ++i)
989  ShuffleMask.push_back(
990  Constant::getIntegerValue(IntTy32, APInt(32, i + 16)));
991  for (int i = Index + Length; i != 8; ++i)
992  ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i)));
993  for (int i = 8; i != 16; ++i)
994  ShuffleMask.push_back(UndefValue::get(IntTy32));
995 
996  Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy),
997  Builder.CreateBitCast(Op1, ShufTy),
998  ConstantVector::get(ShuffleMask));
999  return Builder.CreateBitCast(SV, II.getType());
1000  }
1001 
1002  // See if we're dealing with constant values.
1003  Constant *C0 = dyn_cast<Constant>(Op0);
1004  Constant *C1 = dyn_cast<Constant>(Op1);
1005  ConstantInt *CI00 =
1006  C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
1007  : nullptr;
1008  ConstantInt *CI10 =
1009  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
1010  : nullptr;
1011 
1012  // Constant Fold - insert bottom Length bits starting at the Index'th bit.
1013  if (CI00 && CI10) {
1014  APInt V00 = CI00->getValue();
1015  APInt V10 = CI10->getValue();
1016  APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index);
1017  V00 = V00 & ~Mask;
1018  V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index);
1019  APInt Val = V00 | V10;
1020  Type *IntTy64 = Type::getInt64Ty(II.getContext());
1021  Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()),
1022  UndefValue::get(IntTy64)};
1023  return ConstantVector::get(Args);
1024  }
1025 
1026  // If we were an INSERTQ call, we'll save demanded elements if we convert to
1027  // INSERTQI.
1028  if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) {
1029  Type *IntTy8 = Type::getInt8Ty(II.getContext());
1030  Constant *CILength = ConstantInt::get(IntTy8, Length, false);
1031  Constant *CIIndex = ConstantInt::get(IntTy8, Index, false);
1032 
1033  Value *Args[] = {Op0, Op1, CILength, CIIndex};
1034  Module *M = II.getModule();
1035  Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1036  return Builder.CreateCall(F, Args);
1037  }
1038 
1039  return nullptr;
1040 }
1041 
1042 /// Attempt to convert pshufb* to shufflevector if the mask is constant.
1044  InstCombiner::BuilderTy &Builder) {
1045  Constant *V = dyn_cast<Constant>(II.getArgOperand(1));
1046  if (!V)
1047  return nullptr;
1048 
1049  auto *VecTy = cast<VectorType>(II.getType());
1050  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
1051  unsigned NumElts = VecTy->getNumElements();
1052  assert((NumElts == 16 || NumElts == 32 || NumElts == 64) &&
1053  "Unexpected number of elements in shuffle mask!");
1054 
1055  // Construct a shuffle mask from constant integers or UNDEFs.
1056  Constant *Indexes[64] = {nullptr};
1057 
1058  // Each byte in the shuffle control mask forms an index to permute the
1059  // corresponding byte in the destination operand.
1060  for (unsigned I = 0; I < NumElts; ++I) {
1061  Constant *COp = V->getAggregateElement(I);
1062  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
1063  return nullptr;
1064 
1065  if (isa<UndefValue>(COp)) {
1066  Indexes[I] = UndefValue::get(MaskEltTy);
1067  continue;
1068  }
1069 
1070  int8_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue();
1071 
1072  // If the most significant bit (bit[7]) of each byte of the shuffle
1073  // control mask is set, then zero is written in the result byte.
1074  // The zero vector is in the right-hand side of the resulting
1075  // shufflevector.
1076 
1077  // The value of each index for the high 128-bit lane is the least
1078  // significant 4 bits of the respective shuffle control byte.
1079  Index = ((Index < 0) ? NumElts : Index & 0x0F) + (I & 0xF0);
1080  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
1081  }
1082 
1083  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts));
1084  auto V1 = II.getArgOperand(0);
1085  auto V2 = Constant::getNullValue(VecTy);
1086  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1087 }
1088 
1089 /// Attempt to convert vpermilvar* to shufflevector if the mask is constant.
1091  InstCombiner::BuilderTy &Builder) {
1092  Constant *V = dyn_cast<Constant>(II.getArgOperand(1));
1093  if (!V)
1094  return nullptr;
1095 
1096  auto *VecTy = cast<VectorType>(II.getType());
1097  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
1098  unsigned NumElts = VecTy->getVectorNumElements();
1099  bool IsPD = VecTy->getScalarType()->isDoubleTy();
1100  unsigned NumLaneElts = IsPD ? 2 : 4;
1101  assert(NumElts == 16 || NumElts == 8 || NumElts == 4 || NumElts == 2);
1102 
1103  // Construct a shuffle mask from constant integers or UNDEFs.
1104  Constant *Indexes[16] = {nullptr};
1105 
1106  // The intrinsics only read one or two bits, clear the rest.
1107  for (unsigned I = 0; I < NumElts; ++I) {
1108  Constant *COp = V->getAggregateElement(I);
1109  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
1110  return nullptr;
1111 
1112  if (isa<UndefValue>(COp)) {
1113  Indexes[I] = UndefValue::get(MaskEltTy);
1114  continue;
1115  }
1116 
1117  APInt Index = cast<ConstantInt>(COp)->getValue();
1118  Index = Index.zextOrTrunc(32).getLoBits(2);
1119 
1120  // The PD variants uses bit 1 to select per-lane element index, so
1121  // shift down to convert to generic shuffle mask index.
1122  if (IsPD)
1123  Index.lshrInPlace(1);
1124 
1125  // The _256 variants are a bit trickier since the mask bits always index
1126  // into the corresponding 128 half. In order to convert to a generic
1127  // shuffle, we have to make that explicit.
1128  Index += APInt(32, (I / NumLaneElts) * NumLaneElts);
1129 
1130  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
1131  }
1132 
1133  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts));
1134  auto V1 = II.getArgOperand(0);
1135  auto V2 = UndefValue::get(V1->getType());
1136  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1137 }
1138 
1139 /// Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
1141  InstCombiner::BuilderTy &Builder) {
1142  auto *V = dyn_cast<Constant>(II.getArgOperand(1));
1143  if (!V)
1144  return nullptr;
1145 
1146  auto *VecTy = cast<VectorType>(II.getType());
1147  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
1148  unsigned Size = VecTy->getNumElements();
1149  assert((Size == 4 || Size == 8 || Size == 16 || Size == 32 || Size == 64) &&
1150  "Unexpected shuffle mask size");
1151 
1152  // Construct a shuffle mask from constant integers or UNDEFs.
1153  Constant *Indexes[64] = {nullptr};
1154 
1155  for (unsigned I = 0; I < Size; ++I) {
1156  Constant *COp = V->getAggregateElement(I);
1157  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
1158  return nullptr;
1159 
1160  if (isa<UndefValue>(COp)) {
1161  Indexes[I] = UndefValue::get(MaskEltTy);
1162  continue;
1163  }
1164 
1165  uint32_t Index = cast<ConstantInt>(COp)->getZExtValue();
1166  Index &= Size - 1;
1167  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
1168  }
1169 
1170  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, Size));
1171  auto V1 = II.getArgOperand(0);
1172  auto V2 = UndefValue::get(VecTy);
1173  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1174 }
1175 
1176 /// Decode XOP integer vector comparison intrinsics.
1178  InstCombiner::BuilderTy &Builder,
1179  bool IsSigned) {
1180  if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
1181  uint64_t Imm = CInt->getZExtValue() & 0x7;
1182  VectorType *VecTy = cast<VectorType>(II.getType());
1184 
1185  switch (Imm) {
1186  case 0x0:
1187  Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1188  break;
1189  case 0x1:
1190  Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1191  break;
1192  case 0x2:
1193  Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1194  break;
1195  case 0x3:
1196  Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1197  break;
1198  case 0x4:
1199  Pred = ICmpInst::ICMP_EQ; break;
1200  case 0x5:
1201  Pred = ICmpInst::ICMP_NE; break;
1202  case 0x6:
1203  return ConstantInt::getSigned(VecTy, 0); // FALSE
1204  case 0x7:
1205  return ConstantInt::getSigned(VecTy, -1); // TRUE
1206  }
1207 
1208  if (Value *Cmp = Builder.CreateICmp(Pred, II.getArgOperand(0),
1209  II.getArgOperand(1)))
1210  return Builder.CreateSExtOrTrunc(Cmp, VecTy);
1211  }
1212  return nullptr;
1213 }
1214 
1216  auto *ConstMask = dyn_cast<Constant>(Mask);
1217  if (!ConstMask)
1218  return false;
1219  if (ConstMask->isAllOnesValue() || isa<UndefValue>(ConstMask))
1220  return true;
1221  for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
1222  ++I) {
1223  if (auto *MaskElt = ConstMask->getAggregateElement(I))
1224  if (MaskElt->isAllOnesValue() || isa<UndefValue>(MaskElt))
1225  continue;
1226  return false;
1227  }
1228  return true;
1229 }
1230 
1232  InstCombiner::BuilderTy &Builder) {
1233  // If the mask is all ones or undefs, this is a plain vector load of the 1st
1234  // argument.
1235  if (maskIsAllOneOrUndef(II.getArgOperand(2))) {
1236  Value *LoadPtr = II.getArgOperand(0);
1237  unsigned Alignment = cast<ConstantInt>(II.getArgOperand(1))->getZExtValue();
1238  return Builder.CreateAlignedLoad(LoadPtr, Alignment, "unmaskedload");
1239  }
1240 
1241  return nullptr;
1242 }
1243 
1245  auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
1246  if (!ConstMask)
1247  return nullptr;
1248 
1249  // If the mask is all zeros, this instruction does nothing.
1250  if (ConstMask->isNullValue())
1251  return IC.eraseInstFromFunction(II);
1252 
1253  // If the mask is all ones, this is a plain vector store of the 1st argument.
1254  if (ConstMask->isAllOnesValue()) {
1255  Value *StorePtr = II.getArgOperand(1);
1256  unsigned Alignment = cast<ConstantInt>(II.getArgOperand(2))->getZExtValue();
1257  return new StoreInst(II.getArgOperand(0), StorePtr, false, Alignment);
1258  }
1259 
1260  return nullptr;
1261 }
1262 
1264  // If the mask is all zeros, return the "passthru" argument of the gather.
1265  auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(2));
1266  if (ConstMask && ConstMask->isNullValue())
1267  return IC.replaceInstUsesWith(II, II.getArgOperand(3));
1268 
1269  return nullptr;
1270 }
1271 
1272 /// This function transforms launder.invariant.group and strip.invariant.group
1273 /// like:
1274 /// launder(launder(%x)) -> launder(%x) (the result is not the argument)
1275 /// launder(strip(%x)) -> launder(%x)
1276 /// strip(strip(%x)) -> strip(%x) (the result is not the argument)
1277 /// strip(launder(%x)) -> strip(%x)
1278 /// This is legal because it preserves the most recent information about
1279 /// the presence or absence of invariant.group.
1281  InstCombiner &IC) {
1282  auto *Arg = II.getArgOperand(0);
1283  auto *StrippedArg = Arg->stripPointerCasts();
1284  auto *StrippedInvariantGroupsArg = Arg->stripPointerCastsAndInvariantGroups();
1285  if (StrippedArg == StrippedInvariantGroupsArg)
1286  return nullptr; // No launders/strips to remove.
1287 
1288  Value *Result = nullptr;
1289 
1290  if (II.getIntrinsicID() == Intrinsic::launder_invariant_group)
1291  Result = IC.Builder.CreateLaunderInvariantGroup(StrippedInvariantGroupsArg);
1292  else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group)
1293  Result = IC.Builder.CreateStripInvariantGroup(StrippedInvariantGroupsArg);
1294  else
1296  "simplifyInvariantGroupIntrinsic only handles launder and strip");
1297  if (Result->getType()->getPointerAddressSpace() !=
1299  Result = IC.Builder.CreateAddrSpaceCast(Result, II.getType());
1300  if (Result->getType() != II.getType())
1301  Result = IC.Builder.CreateBitCast(Result, II.getType());
1302 
1303  return cast<Instruction>(Result);
1304 }
1305 
1307  // If the mask is all zeros, a scatter does nothing.
1308  auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
1309  if (ConstMask && ConstMask->isNullValue())
1310  return IC.eraseInstFromFunction(II);
1311 
1312  return nullptr;
1313 }
1314 
1316  assert((II.getIntrinsicID() == Intrinsic::cttz ||
1317  II.getIntrinsicID() == Intrinsic::ctlz) &&
1318  "Expected cttz or ctlz intrinsic");
1319  Value *Op0 = II.getArgOperand(0);
1320 
1321  KnownBits Known = IC.computeKnownBits(Op0, 0, &II);
1322 
1323  // Create a mask for bits above (ctlz) or below (cttz) the first known one.
1324  bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz;
1325  unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros()
1326  : Known.countMaxLeadingZeros();
1327  unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros()
1328  : Known.countMinLeadingZeros();
1329 
1330  // If all bits above (ctlz) or below (cttz) the first known one are known
1331  // zero, this value is constant.
1332  // FIXME: This should be in InstSimplify because we're replacing an
1333  // instruction with a constant.
1334  if (PossibleZeros == DefiniteZeros) {
1335  auto *C = ConstantInt::get(Op0->getType(), DefiniteZeros);
1336  return IC.replaceInstUsesWith(II, C);
1337  }
1338 
1339  // If the input to cttz/ctlz is known to be non-zero,
1340  // then change the 'ZeroIsUndef' parameter to 'true'
1341  // because we know the zero behavior can't affect the result.
1342  if (!Known.One.isNullValue() ||
1343  isKnownNonZero(Op0, IC.getDataLayout(), 0, &IC.getAssumptionCache(), &II,
1344  &IC.getDominatorTree())) {
1345  if (!match(II.getArgOperand(1), m_One())) {
1346  II.setOperand(1, IC.Builder.getTrue());
1347  return &II;
1348  }
1349  }
1350 
1351  // Add range metadata since known bits can't completely reflect what we know.
1352  // TODO: Handle splat vectors.
1353  auto *IT = dyn_cast<IntegerType>(Op0->getType());
1354  if (IT && IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
1355  Metadata *LowAndHigh[] = {
1356  ConstantAsMetadata::get(ConstantInt::get(IT, DefiniteZeros)),
1357  ConstantAsMetadata::get(ConstantInt::get(IT, PossibleZeros + 1))};
1360  return &II;
1361  }
1362 
1363  return nullptr;
1364 }
1365 
1367  assert(II.getIntrinsicID() == Intrinsic::ctpop &&
1368  "Expected ctpop intrinsic");
1369  Value *Op0 = II.getArgOperand(0);
1370  // FIXME: Try to simplify vectors of integers.
1371  auto *IT = dyn_cast<IntegerType>(Op0->getType());
1372  if (!IT)
1373  return nullptr;
1374 
1375  unsigned BitWidth = IT->getBitWidth();
1376  KnownBits Known(BitWidth);
1377  IC.computeKnownBits(Op0, Known, 0, &II);
1378 
1379  unsigned MinCount = Known.countMinPopulation();
1380  unsigned MaxCount = Known.countMaxPopulation();
1381 
1382  // Add range metadata since known bits can't completely reflect what we know.
1383  if (IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
1384  Metadata *LowAndHigh[] = {
1386  ConstantAsMetadata::get(ConstantInt::get(IT, MaxCount + 1))};
1389  return &II;
1390  }
1391 
1392  return nullptr;
1393 }
1394 
1395 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
1396 // XMM register mask efficiently, we could transform all x86 masked intrinsics
1397 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
1399  Value *Ptr = II.getOperand(0);
1400  Value *Mask = II.getOperand(1);
1401  Constant *ZeroVec = Constant::getNullValue(II.getType());
1402 
1403  // Special case a zero mask since that's not a ConstantDataVector.
1404  // This masked load instruction creates a zero vector.
1405  if (isa<ConstantAggregateZero>(Mask))
1406  return IC.replaceInstUsesWith(II, ZeroVec);
1407 
1408  auto *ConstMask = dyn_cast<ConstantDataVector>(Mask);
1409  if (!ConstMask)
1410  return nullptr;
1411 
1412  // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic
1413  // to allow target-independent optimizations.
1414 
1415  // First, cast the x86 intrinsic scalar pointer to a vector pointer to match
1416  // the LLVM intrinsic definition for the pointer argument.
1417  unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
1418  PointerType *VecPtrTy = PointerType::get(II.getType(), AddrSpace);
1419  Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
1420 
1421  // Second, convert the x86 XMM integer vector mask to a vector of bools based
1422  // on each element's most significant bit (the sign bit).
1423  Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask);
1424 
1425  // The pass-through vector for an x86 masked load is a zero vector.
1426  CallInst *NewMaskedLoad =
1427  IC.Builder.CreateMaskedLoad(PtrCast, 1, BoolMask, ZeroVec);
1428  return IC.replaceInstUsesWith(II, NewMaskedLoad);
1429 }
1430 
1431 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
1432 // XMM register mask efficiently, we could transform all x86 masked intrinsics
1433 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
1435  Value *Ptr = II.getOperand(0);
1436  Value *Mask = II.getOperand(1);
1437  Value *Vec = II.getOperand(2);
1438 
1439  // Special case a zero mask since that's not a ConstantDataVector:
1440  // this masked store instruction does nothing.
1441  if (isa<ConstantAggregateZero>(Mask)) {
1442  IC.eraseInstFromFunction(II);
1443  return true;
1444  }
1445 
1446  // The SSE2 version is too weird (eg, unaligned but non-temporal) to do
1447  // anything else at this level.
1448  if (II.getIntrinsicID() == Intrinsic::x86_sse2_maskmov_dqu)
1449  return false;
1450 
1451  auto *ConstMask = dyn_cast<ConstantDataVector>(Mask);
1452  if (!ConstMask)
1453  return false;
1454 
1455  // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic
1456  // to allow target-independent optimizations.
1457 
1458  // First, cast the x86 intrinsic scalar pointer to a vector pointer to match
1459  // the LLVM intrinsic definition for the pointer argument.
1460  unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
1461  PointerType *VecPtrTy = PointerType::get(Vec->getType(), AddrSpace);
1462  Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
1463 
1464  // Second, convert the x86 XMM integer vector mask to a vector of bools based
1465  // on each element's most significant bit (the sign bit).
1466  Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask);
1467 
1468  IC.Builder.CreateMaskedStore(Vec, PtrCast, 1, BoolMask);
1469 
1470  // 'Replace uses' doesn't work for stores. Erase the original masked store.
1471  IC.eraseInstFromFunction(II);
1472  return true;
1473 }
1474 
1475 // Constant fold llvm.amdgcn.fmed3 intrinsics for standard inputs.
1476 //
1477 // A single NaN input is folded to minnum, so we rely on that folding for
1478 // handling NaNs.
1479 static APFloat fmed3AMDGCN(const APFloat &Src0, const APFloat &Src1,
1480  const APFloat &Src2) {
1481  APFloat Max3 = maxnum(maxnum(Src0, Src1), Src2);
1482 
1483  APFloat::cmpResult Cmp0 = Max3.compare(Src0);
1484  assert(Cmp0 != APFloat::cmpUnordered && "nans handled separately");
1485  if (Cmp0 == APFloat::cmpEqual)
1486  return maxnum(Src1, Src2);
1487 
1488  APFloat::cmpResult Cmp1 = Max3.compare(Src1);
1489  assert(Cmp1 != APFloat::cmpUnordered && "nans handled separately");
1490  if (Cmp1 == APFloat::cmpEqual)
1491  return maxnum(Src0, Src2);
1492 
1493  return maxnum(Src0, Src1);
1494 }
1495 
1496 /// Convert a table lookup to shufflevector if the mask is constant.
1497 /// This could benefit tbl1 if the mask is { 7,6,5,4,3,2,1,0 }, in
1498 /// which case we could lower the shufflevector with rev64 instructions
1499 /// as it's actually a byte reverse.
1501  InstCombiner::BuilderTy &Builder) {
1502  // Bail out if the mask is not a constant.
1503  auto *C = dyn_cast<Constant>(II.getArgOperand(1));
1504  if (!C)
1505  return nullptr;
1506 
1507  auto *VecTy = cast<VectorType>(II.getType());
1508  unsigned NumElts = VecTy->getNumElements();
1509 
1510  // Only perform this transformation for <8 x i8> vector types.
1511  if (!VecTy->getElementType()->isIntegerTy(8) || NumElts != 8)
1512  return nullptr;
1513 
1514  uint32_t Indexes[8];
1515 
1516  for (unsigned I = 0; I < NumElts; ++I) {
1517  Constant *COp = C->getAggregateElement(I);
1518 
1519  if (!COp || !isa<ConstantInt>(COp))
1520  return nullptr;
1521 
1522  Indexes[I] = cast<ConstantInt>(COp)->getLimitedValue();
1523 
1524  // Make sure the mask indices are in range.
1525  if (Indexes[I] >= NumElts)
1526  return nullptr;
1527  }
1528 
1529  auto *ShuffleMask = ConstantDataVector::get(II.getContext(),
1530  makeArrayRef(Indexes));
1531  auto *V1 = II.getArgOperand(0);
1532  auto *V2 = Constant::getNullValue(V1->getType());
1533  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1534 }
1535 
1536 /// Convert a vector load intrinsic into a simple llvm load instruction.
1537 /// This is beneficial when the underlying object being addressed comes
1538 /// from a constant, since we get constant-folding for free.
1540  unsigned MemAlign,
1541  InstCombiner::BuilderTy &Builder) {
1542  auto *IntrAlign = dyn_cast<ConstantInt>(II.getArgOperand(1));
1543 
1544  if (!IntrAlign)
1545  return nullptr;
1546 
1547  unsigned Alignment = IntrAlign->getLimitedValue() < MemAlign ?
1548  MemAlign : IntrAlign->getLimitedValue();
1549 
1550  if (!isPowerOf2_32(Alignment))
1551  return nullptr;
1552 
1553  auto *BCastInst = Builder.CreateBitCast(II.getArgOperand(0),
1554  PointerType::get(II.getType(), 0));
1555  return Builder.CreateAlignedLoad(BCastInst, Alignment);
1556 }
1557 
1558 // Returns true iff the 2 intrinsics have the same operands, limiting the
1559 // comparison to the first NumOperands.
1560 static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E,
1561  unsigned NumOperands) {
1562  assert(I.getNumArgOperands() >= NumOperands && "Not enough operands");
1563  assert(E.getNumArgOperands() >= NumOperands && "Not enough operands");
1564  for (unsigned i = 0; i < NumOperands; i++)
1565  if (I.getArgOperand(i) != E.getArgOperand(i))
1566  return false;
1567  return true;
1568 }
1569 
1570 // Remove trivially empty start/end intrinsic ranges, i.e. a start
1571 // immediately followed by an end (ignoring debuginfo or other
1572 // start/end intrinsics in between). As this handles only the most trivial
1573 // cases, tracking the nesting level is not needed:
1574 //
1575 // call @llvm.foo.start(i1 0) ; &I
1576 // call @llvm.foo.start(i1 0)
1577 // call @llvm.foo.end(i1 0) ; This one will not be skipped: it will be removed
1578 // call @llvm.foo.end(i1 0)
1579 static bool removeTriviallyEmptyRange(IntrinsicInst &I, unsigned StartID,
1580  unsigned EndID, InstCombiner &IC) {
1581  assert(I.getIntrinsicID() == StartID &&
1582  "Start intrinsic does not have expected ID");
1583  BasicBlock::iterator BI(I), BE(I.getParent()->end());
1584  for (++BI; BI != BE; ++BI) {
1585  if (auto *E = dyn_cast<IntrinsicInst>(BI)) {
1586  if (isa<DbgInfoIntrinsic>(E) || E->getIntrinsicID() == StartID)
1587  continue;
1588  if (E->getIntrinsicID() == EndID &&
1589  haveSameOperands(I, *E, E->getNumArgOperands())) {
1590  IC.eraseInstFromFunction(*E);
1591  IC.eraseInstFromFunction(I);
1592  return true;
1593  }
1594  }
1595  break;
1596  }
1597 
1598  return false;
1599 }
1600 
1601 // Convert NVVM intrinsics to target-generic LLVM code where possible.
1603  // Each NVVM intrinsic we can simplify can be replaced with one of:
1604  //
1605  // * an LLVM intrinsic,
1606  // * an LLVM cast operation,
1607  // * an LLVM binary operation, or
1608  // * ad-hoc LLVM IR for the particular operation.
1609 
1610  // Some transformations are only valid when the module's
1611  // flush-denormals-to-zero (ftz) setting is true/false, whereas other
1612  // transformations are valid regardless of the module's ftz setting.
1613  enum FtzRequirementTy {
1614  FTZ_Any, // Any ftz setting is ok.
1615  FTZ_MustBeOn, // Transformation is valid only if ftz is on.
1616  FTZ_MustBeOff, // Transformation is valid only if ftz is off.
1617  };
1618  // Classes of NVVM intrinsics that can't be replaced one-to-one with a
1619  // target-generic intrinsic, cast op, or binary op but that we can nonetheless
1620  // simplify.
1621  enum SpecialCase {
1622  SPC_Reciprocal,
1623  };
1624 
1625  // SimplifyAction is a poor-man's variant (plus an additional flag) that
1626  // represents how to replace an NVVM intrinsic with target-generic LLVM IR.
1627  struct SimplifyAction {
1628  // Invariant: At most one of these Optionals has a value.
1632  Optional<SpecialCase> Special;
1633 
1634  FtzRequirementTy FtzRequirement = FTZ_Any;
1635 
1636  SimplifyAction() = default;
1637 
1638  SimplifyAction(Intrinsic::ID IID, FtzRequirementTy FtzReq)
1639  : IID(IID), FtzRequirement(FtzReq) {}
1640 
1641  // Cast operations don't have anything to do with FTZ, so we skip that
1642  // argument.
1643  SimplifyAction(Instruction::CastOps CastOp) : CastOp(CastOp) {}
1644 
1645  SimplifyAction(Instruction::BinaryOps BinaryOp, FtzRequirementTy FtzReq)
1646  : BinaryOp(BinaryOp), FtzRequirement(FtzReq) {}
1647 
1648  SimplifyAction(SpecialCase Special, FtzRequirementTy FtzReq)
1649  : Special(Special), FtzRequirement(FtzReq) {}
1650  };
1651 
1652  // Try to generate a SimplifyAction describing how to replace our
1653  // IntrinsicInstr with target-generic LLVM IR.
1654  const SimplifyAction Action = [II]() -> SimplifyAction {
1655  switch (II->getIntrinsicID()) {
1656  // NVVM intrinsics that map directly to LLVM intrinsics.
1657  case Intrinsic::nvvm_ceil_d:
1658  return {Intrinsic::ceil, FTZ_Any};
1659  case Intrinsic::nvvm_ceil_f:
1660  return {Intrinsic::ceil, FTZ_MustBeOff};
1661  case Intrinsic::nvvm_ceil_ftz_f:
1662  return {Intrinsic::ceil, FTZ_MustBeOn};
1663  case Intrinsic::nvvm_fabs_d:
1664  return {Intrinsic::fabs, FTZ_Any};
1665  case Intrinsic::nvvm_fabs_f:
1666  return {Intrinsic::fabs, FTZ_MustBeOff};
1667  case Intrinsic::nvvm_fabs_ftz_f:
1668  return {Intrinsic::fabs, FTZ_MustBeOn};
1669  case Intrinsic::nvvm_floor_d:
1670  return {Intrinsic::floor, FTZ_Any};
1671  case Intrinsic::nvvm_floor_f:
1672  return {Intrinsic::floor, FTZ_MustBeOff};
1673  case Intrinsic::nvvm_floor_ftz_f:
1674  return {Intrinsic::floor, FTZ_MustBeOn};
1675  case Intrinsic::nvvm_fma_rn_d:
1676  return {Intrinsic::fma, FTZ_Any};
1677  case Intrinsic::nvvm_fma_rn_f:
1678  return {Intrinsic::fma, FTZ_MustBeOff};
1679  case Intrinsic::nvvm_fma_rn_ftz_f:
1680  return {Intrinsic::fma, FTZ_MustBeOn};
1681  case Intrinsic::nvvm_fmax_d:
1682  return {Intrinsic::maxnum, FTZ_Any};
1683  case Intrinsic::nvvm_fmax_f:
1684  return {Intrinsic::maxnum, FTZ_MustBeOff};
1685  case Intrinsic::nvvm_fmax_ftz_f:
1686  return {Intrinsic::maxnum, FTZ_MustBeOn};
1687  case Intrinsic::nvvm_fmin_d:
1688  return {Intrinsic::minnum, FTZ_Any};
1689  case Intrinsic::nvvm_fmin_f:
1690  return {Intrinsic::minnum, FTZ_MustBeOff};
1691  case Intrinsic::nvvm_fmin_ftz_f:
1692  return {Intrinsic::minnum, FTZ_MustBeOn};
1693  case Intrinsic::nvvm_round_d:
1694  return {Intrinsic::round, FTZ_Any};
1695  case Intrinsic::nvvm_round_f:
1696  return {Intrinsic::round, FTZ_MustBeOff};
1697  case Intrinsic::nvvm_round_ftz_f:
1698  return {Intrinsic::round, FTZ_MustBeOn};
1699  case Intrinsic::nvvm_sqrt_rn_d:
1700  return {Intrinsic::sqrt, FTZ_Any};
1701  case Intrinsic::nvvm_sqrt_f:
1702  // nvvm_sqrt_f is a special case. For most intrinsics, foo_ftz_f is the
1703  // ftz version, and foo_f is the non-ftz version. But nvvm_sqrt_f adopts
1704  // the ftz-ness of the surrounding code. sqrt_rn_f and sqrt_rn_ftz_f are
1705  // the versions with explicit ftz-ness.
1706  return {Intrinsic::sqrt, FTZ_Any};
1707  case Intrinsic::nvvm_sqrt_rn_f:
1708  return {Intrinsic::sqrt, FTZ_MustBeOff};
1709  case Intrinsic::nvvm_sqrt_rn_ftz_f:
1710  return {Intrinsic::sqrt, FTZ_MustBeOn};
1711  case Intrinsic::nvvm_trunc_d:
1712  return {Intrinsic::trunc, FTZ_Any};
1713  case Intrinsic::nvvm_trunc_f:
1714  return {Intrinsic::trunc, FTZ_MustBeOff};
1715  case Intrinsic::nvvm_trunc_ftz_f:
1716  return {Intrinsic::trunc, FTZ_MustBeOn};
1717 
1718  // NVVM intrinsics that map to LLVM cast operations.
1719  //
1720  // Note that llvm's target-generic conversion operators correspond to the rz
1721  // (round to zero) versions of the nvvm conversion intrinsics, even though
1722  // most everything else here uses the rn (round to nearest even) nvvm ops.
1723  case Intrinsic::nvvm_d2i_rz:
1724  case Intrinsic::nvvm_f2i_rz:
1725  case Intrinsic::nvvm_d2ll_rz:
1726  case Intrinsic::nvvm_f2ll_rz:
1727  return {Instruction::FPToSI};
1728  case Intrinsic::nvvm_d2ui_rz:
1729  case Intrinsic::nvvm_f2ui_rz:
1730  case Intrinsic::nvvm_d2ull_rz:
1731  case Intrinsic::nvvm_f2ull_rz:
1732  return {Instruction::FPToUI};
1733  case Intrinsic::nvvm_i2d_rz:
1734  case Intrinsic::nvvm_i2f_rz:
1735  case Intrinsic::nvvm_ll2d_rz:
1736  case Intrinsic::nvvm_ll2f_rz:
1737  return {Instruction::SIToFP};
1738  case Intrinsic::nvvm_ui2d_rz:
1739  case Intrinsic::nvvm_ui2f_rz:
1740  case Intrinsic::nvvm_ull2d_rz:
1741  case Intrinsic::nvvm_ull2f_rz:
1742  return {Instruction::UIToFP};
1743 
1744  // NVVM intrinsics that map to LLVM binary ops.
1745  case Intrinsic::nvvm_add_rn_d:
1746  return {Instruction::FAdd, FTZ_Any};
1747  case Intrinsic::nvvm_add_rn_f:
1748  return {Instruction::FAdd, FTZ_MustBeOff};
1749  case Intrinsic::nvvm_add_rn_ftz_f:
1750  return {Instruction::FAdd, FTZ_MustBeOn};
1751  case Intrinsic::nvvm_mul_rn_d:
1752  return {Instruction::FMul, FTZ_Any};
1753  case Intrinsic::nvvm_mul_rn_f:
1754  return {Instruction::FMul, FTZ_MustBeOff};
1755  case Intrinsic::nvvm_mul_rn_ftz_f:
1756  return {Instruction::FMul, FTZ_MustBeOn};
1757  case Intrinsic::nvvm_div_rn_d:
1758  return {Instruction::FDiv, FTZ_Any};
1759  case Intrinsic::nvvm_div_rn_f:
1760  return {Instruction::FDiv, FTZ_MustBeOff};
1761  case Intrinsic::nvvm_div_rn_ftz_f:
1762  return {Instruction::FDiv, FTZ_MustBeOn};
1763 
1764  // The remainder of cases are NVVM intrinsics that map to LLVM idioms, but
1765  // need special handling.
1766  //
1767  // We seem to be missing intrinsics for rcp.approx.{ftz.}f32, which is just
1768  // as well.
1769  case Intrinsic::nvvm_rcp_rn_d:
1770  return {SPC_Reciprocal, FTZ_Any};
1771  case Intrinsic::nvvm_rcp_rn_f:
1772  return {SPC_Reciprocal, FTZ_MustBeOff};
1773  case Intrinsic::nvvm_rcp_rn_ftz_f:
1774  return {SPC_Reciprocal, FTZ_MustBeOn};
1775 
1776  // We do not currently simplify intrinsics that give an approximate answer.
1777  // These include:
1778  //
1779  // - nvvm_cos_approx_{f,ftz_f}
1780  // - nvvm_ex2_approx_{d,f,ftz_f}
1781  // - nvvm_lg2_approx_{d,f,ftz_f}
1782  // - nvvm_sin_approx_{f,ftz_f}
1783  // - nvvm_sqrt_approx_{f,ftz_f}
1784  // - nvvm_rsqrt_approx_{d,f,ftz_f}
1785  // - nvvm_div_approx_{ftz_d,ftz_f,f}
1786  // - nvvm_rcp_approx_ftz_d
1787  //
1788  // Ideally we'd encode them as e.g. "fast call @llvm.cos", where "fast"
1789  // means that fastmath is enabled in the intrinsic. Unfortunately only
1790  // binary operators (currently) have a fastmath bit in SelectionDAG, so this
1791  // information gets lost and we can't select on it.
1792  //
1793  // TODO: div and rcp are lowered to a binary op, so these we could in theory
1794  // lower them to "fast fdiv".
1795 
1796  default:
1797  return {};
1798  }
1799  }();
1800 
1801  // If Action.FtzRequirementTy is not satisfied by the module's ftz state, we
1802  // can bail out now. (Notice that in the case that IID is not an NVVM
1803  // intrinsic, we don't have to look up any module metadata, as
1804  // FtzRequirementTy will be FTZ_Any.)
1805  if (Action.FtzRequirement != FTZ_Any) {
1806  bool FtzEnabled =
1807  II->getFunction()->getFnAttribute("nvptx-f32ftz").getValueAsString() ==
1808  "true";
1809 
1810  if (FtzEnabled != (Action.FtzRequirement == FTZ_MustBeOn))
1811  return nullptr;
1812  }
1813 
1814  // Simplify to target-generic intrinsic.
1815  if (Action.IID) {
1817  // All the target-generic intrinsics currently of interest to us have one
1818  // type argument, equal to that of the nvvm intrinsic's argument.
1819  Type *Tys[] = {II->getArgOperand(0)->getType()};
1820  return CallInst::Create(
1821  Intrinsic::getDeclaration(II->getModule(), *Action.IID, Tys), Args);
1822  }
1823 
1824  // Simplify to target-generic binary op.
1825  if (Action.BinaryOp)
1826  return BinaryOperator::Create(*Action.BinaryOp, II->getArgOperand(0),
1827  II->getArgOperand(1), II->getName());
1828 
1829  // Simplify to target-generic cast op.
1830  if (Action.CastOp)
1831  return CastInst::Create(*Action.CastOp, II->getArgOperand(0), II->getType(),
1832  II->getName());
1833 
1834  // All that's left are the special cases.
1835  if (!Action.Special)
1836  return nullptr;
1837 
1838  switch (*Action.Special) {
1839  case SPC_Reciprocal:
1840  // Simplify reciprocal.
1841  return BinaryOperator::Create(
1842  Instruction::FDiv, ConstantFP::get(II->getArgOperand(0)->getType(), 1),
1843  II->getArgOperand(0), II->getName());
1844  }
1845  llvm_unreachable("All SpecialCase enumerators should be handled in switch.");
1846 }
1847 
1849  removeTriviallyEmptyRange(I, Intrinsic::vastart, Intrinsic::vaend, *this);
1850  return nullptr;
1851 }
1852 
1854  removeTriviallyEmptyRange(I, Intrinsic::vacopy, Intrinsic::vaend, *this);
1855  return nullptr;
1856 }
1857 
1859  assert(Call.getNumArgOperands() > 1 && "Need at least 2 args to swap");
1860  Value *Arg0 = Call.getArgOperand(0), *Arg1 = Call.getArgOperand(1);
1861  if (isa<Constant>(Arg0) && !isa<Constant>(Arg1)) {
1862  Call.setArgOperand(0, Arg1);
1863  Call.setArgOperand(1, Arg0);
1864  return &Call;
1865  }
1866  return nullptr;
1867 }
1868 
1869 /// CallInst simplification. This mostly only handles folding of intrinsic
1870 /// instructions. For normal calls, it allows visitCallSite to do the heavy
1871 /// lifting.
1873  if (Value *V = SimplifyCall(&CI, SQ.getWithInstruction(&CI)))
1874  return replaceInstUsesWith(CI, V);
1875 
1876  if (isFreeCall(&CI, &TLI))
1877  return visitFree(CI);
1878 
1879  // If the caller function is nounwind, mark the call as nounwind, even if the
1880  // callee isn't.
1881  if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) {
1882  CI.setDoesNotThrow();
1883  return &CI;
1884  }
1885 
1886  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
1887  if (!II) return visitCallSite(&CI);
1888 
1889  // Intrinsics cannot occur in an invoke, so handle them here instead of in
1890  // visitCallSite.
1891  if (auto *MI = dyn_cast<AnyMemIntrinsic>(II)) {
1892  bool Changed = false;
1893 
1894  // memmove/cpy/set of zero bytes is a noop.
1895  if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
1896  if (NumBytes->isNullValue())
1897  return eraseInstFromFunction(CI);
1898 
1899  if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
1900  if (CI->getZExtValue() == 1) {
1901  // Replace the instruction with just byte operations. We would
1902  // transform other cases to loads/stores, but we don't know if
1903  // alignment is sufficient.
1904  }
1905  }
1906 
1907  // No other transformations apply to volatile transfers.
1908  if (auto *M = dyn_cast<MemIntrinsic>(MI))
1909  if (M->isVolatile())
1910  return nullptr;
1911 
1912  // If we have a memmove and the source operation is a constant global,
1913  // then the source and dest pointers can't alias, so we can change this
1914  // into a call to memcpy.
1915  if (auto *MMI = dyn_cast<AnyMemMoveInst>(MI)) {
1916  if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
1917  if (GVSrc->isConstant()) {
1918  Module *M = CI.getModule();
1919  Intrinsic::ID MemCpyID =
1920  isa<AtomicMemMoveInst>(MMI)
1921  ? Intrinsic::memcpy_element_unordered_atomic
1922  : Intrinsic::memcpy;
1923  Type *Tys[3] = { CI.getArgOperand(0)->getType(),
1924  CI.getArgOperand(1)->getType(),
1925  CI.getArgOperand(2)->getType() };
1926  CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
1927  Changed = true;
1928  }
1929  }
1930 
1931  if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1932  // memmove(x,x,size) -> noop.
1933  if (MTI->getSource() == MTI->getDest())
1934  return eraseInstFromFunction(CI);
1935  }
1936 
1937  // If we can determine a pointer alignment that is bigger than currently
1938  // set, update the alignment.
1939  if (auto *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1940  if (Instruction *I = SimplifyAnyMemTransfer(MTI))
1941  return I;
1942  } else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) {
1943  if (Instruction *I = SimplifyAnyMemSet(MSI))
1944  return I;
1945  }
1946 
1947  if (Changed) return II;
1948  }
1949 
1950  if (Instruction *I = SimplifyNVVMIntrinsic(II, *this))
1951  return I;
1952 
1953  auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width,
1954  unsigned DemandedWidth) {
1955  APInt UndefElts(Width, 0);
1956  APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
1957  return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
1958  };
1959 
1960  switch (II->getIntrinsicID()) {
1961  default: break;
1962  case Intrinsic::objectsize:
1963  if (ConstantInt *N =
1964  lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/false))
1965  return replaceInstUsesWith(CI, N);
1966  return nullptr;
1967  case Intrinsic::bswap: {
1968  Value *IIOperand = II->getArgOperand(0);
1969  Value *X = nullptr;
1970 
1971  // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
1972  if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
1973  unsigned C = X->getType()->getPrimitiveSizeInBits() -
1974  IIOperand->getType()->getPrimitiveSizeInBits();
1975  Value *CV = ConstantInt::get(X->getType(), C);
1976  Value *V = Builder.CreateLShr(X, CV);
1977  return new TruncInst(V, IIOperand->getType());
1978  }
1979  break;
1980  }
1981  case Intrinsic::masked_load:
1982  if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(*II, Builder))
1983  return replaceInstUsesWith(CI, SimplifiedMaskedOp);
1984  break;
1985  case Intrinsic::masked_store:
1986  return simplifyMaskedStore(*II, *this);
1987  case Intrinsic::masked_gather:
1988  return simplifyMaskedGather(*II, *this);
1989  case Intrinsic::masked_scatter:
1990  return simplifyMaskedScatter(*II, *this);
1991  case Intrinsic::launder_invariant_group:
1992  case Intrinsic::strip_invariant_group:
1993  if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(*II, *this))
1994  return replaceInstUsesWith(*II, SkippedBarrier);
1995  break;
1996  case Intrinsic::powi:
1997  if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
1998  // 0 and 1 are handled in instsimplify
1999 
2000  // powi(x, -1) -> 1/x
2001  if (Power->isMinusOne())
2002  return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
2003  II->getArgOperand(0));
2004  // powi(x, 2) -> x*x
2005  if (Power->equalsInt(2))
2006  return BinaryOperator::CreateFMul(II->getArgOperand(0),
2007  II->getArgOperand(0));
2008  }
2009  break;
2010 
2011  case Intrinsic::cttz:
2012  case Intrinsic::ctlz:
2013  if (auto *I = foldCttzCtlz(*II, *this))
2014  return I;
2015  break;
2016 
2017  case Intrinsic::ctpop:
2018  if (auto *I = foldCtpop(*II, *this))
2019  return I;
2020  break;
2021 
2022  case Intrinsic::fshl:
2023  case Intrinsic::fshr: {
2024  const APInt *SA;
2025  if (match(II->getArgOperand(2), m_APInt(SA))) {
2026  Value *Op0 = II->getArgOperand(0), *Op1 = II->getArgOperand(1);
2027  unsigned BitWidth = SA->getBitWidth();
2028  uint64_t ShiftAmt = SA->urem(BitWidth);
2029  assert(ShiftAmt != 0 && "SimplifyCall should have handled zero shift");
2030  // Normalize to funnel shift left.
2031  if (II->getIntrinsicID() == Intrinsic::fshr)
2032  ShiftAmt = BitWidth - ShiftAmt;
2033 
2034  // fshl(X, 0, C) -> shl X, C
2035  // fshl(X, undef, C) -> shl X, C
2036  if (match(Op1, m_Zero()) || match(Op1, m_Undef()))
2037  return BinaryOperator::CreateShl(
2038  Op0, ConstantInt::get(II->getType(), ShiftAmt));
2039 
2040  // fshl(0, X, C) -> lshr X, (BW-C)
2041  // fshl(undef, X, C) -> lshr X, (BW-C)
2042  if (match(Op0, m_Zero()) || match(Op0, m_Undef()))
2043  return BinaryOperator::CreateLShr(
2044  Op1, ConstantInt::get(II->getType(), BitWidth - ShiftAmt));
2045  }
2046 
2047  // The shift amount (operand 2) of a funnel shift is modulo the bitwidth,
2048  // so only the low bits of the shift amount are demanded if the bitwidth is
2049  // a power-of-2.
2050  unsigned BitWidth = II->getType()->getScalarSizeInBits();
2051  if (!isPowerOf2_32(BitWidth))
2052  break;
2053  APInt Op2Demanded = APInt::getLowBitsSet(BitWidth, Log2_32_Ceil(BitWidth));
2054  KnownBits Op2Known(BitWidth);
2055  if (SimplifyDemandedBits(II, 2, Op2Demanded, Op2Known))
2056  return &CI;
2057  break;
2058  }
2059  case Intrinsic::uadd_with_overflow:
2060  case Intrinsic::sadd_with_overflow:
2061  case Intrinsic::umul_with_overflow:
2062  case Intrinsic::smul_with_overflow:
2064  return I;
2066 
2067  case Intrinsic::usub_with_overflow:
2068  case Intrinsic::ssub_with_overflow: {
2069  OverflowCheckFlavor OCF =
2071  assert(OCF != OCF_INVALID && "unexpected!");
2072 
2073  Value *OperationResult = nullptr;
2074  Constant *OverflowResult = nullptr;
2075  if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1),
2076  *II, OperationResult, OverflowResult))
2077  return CreateOverflowTuple(II, OperationResult, OverflowResult);
2078 
2079  break;
2080  }
2081 
2082  case Intrinsic::uadd_sat:
2083  case Intrinsic::sadd_sat:
2085  return I;
2087  case Intrinsic::usub_sat:
2088  case Intrinsic::ssub_sat: {
2089  Value *Arg0 = II->getArgOperand(0);
2090  Value *Arg1 = II->getArgOperand(1);
2091  Intrinsic::ID IID = II->getIntrinsicID();
2092 
2093  // Make use of known overflow information.
2095  switch (IID) {
2096  default:
2097  llvm_unreachable("Unexpected intrinsic!");
2098  case Intrinsic::uadd_sat:
2099  OR = computeOverflowForUnsignedAdd(Arg0, Arg1, II);
2101  return BinaryOperator::CreateNUWAdd(Arg0, Arg1);
2103  return replaceInstUsesWith(*II,
2105  break;
2106  case Intrinsic::usub_sat:
2107  OR = computeOverflowForUnsignedSub(Arg0, Arg1, II);
2109  return BinaryOperator::CreateNUWSub(Arg0, Arg1);
2111  return replaceInstUsesWith(*II,
2113  break;
2114  case Intrinsic::sadd_sat:
2115  if (willNotOverflowSignedAdd(Arg0, Arg1, *II))
2116  return BinaryOperator::CreateNSWAdd(Arg0, Arg1);
2117  break;
2118  case Intrinsic::ssub_sat:
2119  if (willNotOverflowSignedSub(Arg0, Arg1, *II))
2120  return BinaryOperator::CreateNSWSub(Arg0, Arg1);
2121  break;
2122  }
2123 
2124  // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN
2125  Constant *C;
2126  if (IID == Intrinsic::ssub_sat && match(Arg1, m_Constant(C)) &&
2127  C->isNotMinSignedValue()) {
2128  Value *NegVal = ConstantExpr::getNeg(C);
2129  return replaceInstUsesWith(
2130  *II, Builder.CreateBinaryIntrinsic(
2131  Intrinsic::sadd_sat, Arg0, NegVal));
2132  }
2133 
2134  // sat(sat(X + Val2) + Val) -> sat(X + (Val+Val2))
2135  // sat(sat(X - Val2) - Val) -> sat(X - (Val+Val2))
2136  // if Val and Val2 have the same sign
2137  if (auto *Other = dyn_cast<IntrinsicInst>(Arg0)) {
2138  Value *X;
2139  const APInt *Val, *Val2;
2140  APInt NewVal;
2141  bool IsUnsigned =
2142  IID == Intrinsic::uadd_sat || IID == Intrinsic::usub_sat;
2143  if (Other->getIntrinsicID() == II->getIntrinsicID() &&
2144  match(Arg1, m_APInt(Val)) &&
2145  match(Other->getArgOperand(0), m_Value(X)) &&
2146  match(Other->getArgOperand(1), m_APInt(Val2))) {
2147  if (IsUnsigned)
2148  NewVal = Val->uadd_sat(*Val2);
2149  else if (Val->isNonNegative() == Val2->isNonNegative()) {
2150  bool Overflow;
2151  NewVal = Val->sadd_ov(*Val2, Overflow);
2152  if (Overflow) {
2153  // Both adds together may add more than SignedMaxValue
2154  // without saturating the final result.
2155  break;
2156  }
2157  } else {
2158  // Cannot fold saturated addition with different signs.
2159  break;
2160  }
2161 
2162  return replaceInstUsesWith(
2163  *II, Builder.CreateBinaryIntrinsic(
2164  IID, X, ConstantInt::get(II->getType(), NewVal)));
2165  }
2166  }
2167  break;
2168  }
2169 
2170  case Intrinsic::minnum:
2171  case Intrinsic::maxnum:
2172  case Intrinsic::minimum:
2173  case Intrinsic::maximum: {
2175  return I;
2176  Value *Arg0 = II->getArgOperand(0);
2177  Value *Arg1 = II->getArgOperand(1);
2178  Intrinsic::ID IID = II->getIntrinsicID();
2179  Value *X, *Y;
2180  if (match(Arg0, m_FNeg(m_Value(X))) && match(Arg1, m_FNeg(m_Value(Y))) &&
2181  (Arg0->hasOneUse() || Arg1->hasOneUse())) {
2182  // If both operands are negated, invert the call and negate the result:
2183  // min(-X, -Y) --> -(max(X, Y))
2184  // max(-X, -Y) --> -(min(X, Y))
2185  Intrinsic::ID NewIID;
2186  switch (IID) {
2187  case Intrinsic::maxnum:
2188  NewIID = Intrinsic::minnum;
2189  break;
2190  case Intrinsic::minnum:
2191  NewIID = Intrinsic::maxnum;
2192  break;
2193  case Intrinsic::maximum:
2194  NewIID = Intrinsic::minimum;
2195  break;
2196  case Intrinsic::minimum:
2197  NewIID = Intrinsic::maximum;
2198  break;
2199  default:
2200  llvm_unreachable("unexpected intrinsic ID");
2201  }
2202  Value *NewCall = Builder.CreateBinaryIntrinsic(NewIID, X, Y, II);
2203  Instruction *FNeg = BinaryOperator::CreateFNeg(NewCall);
2204  FNeg->copyIRFlags(II);
2205  return FNeg;
2206  }
2207 
2208  // m(m(X, C2), C1) -> m(X, C)
2209  const APFloat *C1, *C2;
2210  if (auto *M = dyn_cast<IntrinsicInst>(Arg0)) {
2211  if (M->getIntrinsicID() == IID && match(Arg1, m_APFloat(C1)) &&
2212  ((match(M->getArgOperand(0), m_Value(X)) &&
2213  match(M->getArgOperand(1), m_APFloat(C2))) ||
2214  (match(M->getArgOperand(1), m_Value(X)) &&
2215  match(M->getArgOperand(0), m_APFloat(C2))))) {
2216  APFloat Res(0.0);
2217  switch (IID) {
2218  case Intrinsic::maxnum:
2219  Res = maxnum(*C1, *C2);
2220  break;
2221  case Intrinsic::minnum:
2222  Res = minnum(*C1, *C2);
2223  break;
2224  case Intrinsic::maximum:
2225  Res = maximum(*C1, *C2);
2226  break;
2227  case Intrinsic::minimum:
2228  Res = minimum(*C1, *C2);
2229  break;
2230  default:
2231  llvm_unreachable("unexpected intrinsic ID");
2232  }
2233  Instruction *NewCall = Builder.CreateBinaryIntrinsic(
2234  IID, X, ConstantFP::get(Arg0->getType(), Res));
2235  NewCall->copyIRFlags(II);
2236  return replaceInstUsesWith(*II, NewCall);
2237  }
2238  }
2239 
2240  break;
2241  }
2242  case Intrinsic::fmuladd: {
2243  // Canonicalize fast fmuladd to the separate fmul + fadd.
2244  if (II->isFast()) {
2245  BuilderTy::FastMathFlagGuard Guard(Builder);
2246  Builder.setFastMathFlags(II->getFastMathFlags());
2247  Value *Mul = Builder.CreateFMul(II->getArgOperand(0),
2248  II->getArgOperand(1));
2249  Value *Add = Builder.CreateFAdd(Mul, II->getArgOperand(2));
2250  Add->takeName(II);
2251  return replaceInstUsesWith(*II, Add);
2252  }
2253 
2255  }
2256  case Intrinsic::fma: {
2258  return I;
2259 
2260  // fma fneg(x), fneg(y), z -> fma x, y, z
2261  Value *Src0 = II->getArgOperand(0);
2262  Value *Src1 = II->getArgOperand(1);
2263  Value *X, *Y;
2264  if (match(Src0, m_FNeg(m_Value(X))) && match(Src1, m_FNeg(m_Value(Y)))) {
2265  II->setArgOperand(0, X);
2266  II->setArgOperand(1, Y);
2267  return II;
2268  }
2269 
2270  // fma fabs(x), fabs(x), z -> fma x, x, z
2271  if (match(Src0, m_FAbs(m_Value(X))) &&
2272  match(Src1, m_FAbs(m_Specific(X)))) {
2273  II->setArgOperand(0, X);
2274  II->setArgOperand(1, X);
2275  return II;
2276  }
2277 
2278  // fma x, 1, z -> fadd x, z
2279  if (match(Src1, m_FPOne())) {
2280  auto *FAdd = BinaryOperator::CreateFAdd(Src0, II->getArgOperand(2));
2281  FAdd->copyFastMathFlags(II);
2282  return FAdd;
2283  }
2284 
2285  break;
2286  }
2287  case Intrinsic::fabs: {
2288  Value *Cond;
2289  Constant *LHS, *RHS;
2290  if (match(II->getArgOperand(0),
2291  m_Select(m_Value(Cond), m_Constant(LHS), m_Constant(RHS)))) {
2292  CallInst *Call0 = Builder.CreateCall(II->getCalledFunction(), {LHS});
2293  CallInst *Call1 = Builder.CreateCall(II->getCalledFunction(), {RHS});
2294  return SelectInst::Create(Cond, Call0, Call1);
2295  }
2296 
2298  }
2299  case Intrinsic::ceil:
2300  case Intrinsic::floor:
2301  case Intrinsic::round:
2302  case Intrinsic::nearbyint:
2303  case Intrinsic::rint:
2304  case Intrinsic::trunc: {
2305  Value *ExtSrc;
2306  if (match(II->getArgOperand(0), m_OneUse(m_FPExt(m_Value(ExtSrc))))) {
2307  // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x)
2308  Value *NarrowII =
2309  Builder.CreateUnaryIntrinsic(II->getIntrinsicID(), ExtSrc, II);
2310  return new FPExtInst(NarrowII, II->getType());
2311  }
2312  break;
2313  }
2314  case Intrinsic::cos:
2315  case Intrinsic::amdgcn_cos: {
2316  Value *X;
2317  Value *Src = II->getArgOperand(0);
2318  if (match(Src, m_FNeg(m_Value(X))) || match(Src, m_FAbs(m_Value(X)))) {
2319  // cos(-x) -> cos(x)
2320  // cos(fabs(x)) -> cos(x)
2321  II->setArgOperand(0, X);
2322  return II;
2323  }
2324  break;
2325  }
2326  case Intrinsic::sin: {
2327  Value *X;
2328  if (match(II->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X))))) {
2329  // sin(-x) --> -sin(x)
2330  Value *NewSin = Builder.CreateUnaryIntrinsic(Intrinsic::sin, X, II);
2331  Instruction *FNeg = BinaryOperator::CreateFNeg(NewSin);
2332  FNeg->copyFastMathFlags(II);
2333  return FNeg;
2334  }
2335  break;
2336  }
2337  case Intrinsic::ppc_altivec_lvx:
2338  case Intrinsic::ppc_altivec_lvxl:
2339  // Turn PPC lvx -> load if the pointer is known aligned.
2340  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC,
2341  &DT) >= 16) {
2342  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2343  PointerType::getUnqual(II->getType()));
2344  return new LoadInst(Ptr);
2345  }
2346  break;
2347  case Intrinsic::ppc_vsx_lxvw4x:
2348  case Intrinsic::ppc_vsx_lxvd2x: {
2349  // Turn PPC VSX loads into normal loads.
2350  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2351  PointerType::getUnqual(II->getType()));
2352  return new LoadInst(Ptr, Twine(""), false, 1);
2353  }
2354  case Intrinsic::ppc_altivec_stvx:
2355  case Intrinsic::ppc_altivec_stvxl:
2356  // Turn stvx -> store if the pointer is known aligned.
2357  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC,
2358  &DT) >= 16) {
2359  Type *OpPtrTy =
2360  PointerType::getUnqual(II->getArgOperand(0)->getType());
2361  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2362  return new StoreInst(II->getArgOperand(0), Ptr);
2363  }
2364  break;
2365  case Intrinsic::ppc_vsx_stxvw4x:
2366  case Intrinsic::ppc_vsx_stxvd2x: {
2367  // Turn PPC VSX stores into normal stores.
2368  Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
2369  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2370  return new StoreInst(II->getArgOperand(0), Ptr, false, 1);
2371  }
2372  case Intrinsic::ppc_qpx_qvlfs:
2373  // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
2374  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC,
2375  &DT) >= 16) {
2376  Type *VTy = VectorType::get(Builder.getFloatTy(),
2377  II->getType()->getVectorNumElements());
2378  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2379  PointerType::getUnqual(VTy));
2380  Value *Load = Builder.CreateLoad(Ptr);
2381  return new FPExtInst(Load, II->getType());
2382  }
2383  break;
2384  case Intrinsic::ppc_qpx_qvlfd:
2385  // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
2386  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, &AC,
2387  &DT) >= 32) {
2388  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2389  PointerType::getUnqual(II->getType()));
2390  return new LoadInst(Ptr);
2391  }
2392  break;
2393  case Intrinsic::ppc_qpx_qvstfs:
2394  // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
2395  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC,
2396  &DT) >= 16) {
2397  Type *VTy = VectorType::get(Builder.getFloatTy(),
2398  II->getArgOperand(0)->getType()->getVectorNumElements());
2399  Value *TOp = Builder.CreateFPTrunc(II->getArgOperand(0), VTy);
2400  Type *OpPtrTy = PointerType::getUnqual(VTy);
2401  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2402  return new StoreInst(TOp, Ptr);
2403  }
2404  break;
2405  case Intrinsic::ppc_qpx_qvstfd:
2406  // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
2407  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, &AC,
2408  &DT) >= 32) {
2409  Type *OpPtrTy =
2410  PointerType::getUnqual(II->getArgOperand(0)->getType());
2411  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2412  return new StoreInst(II->getArgOperand(0), Ptr);
2413  }
2414  break;
2415 
2416  case Intrinsic::x86_bmi_bextr_32:
2417  case Intrinsic::x86_bmi_bextr_64:
2418  case Intrinsic::x86_tbm_bextri_u32:
2419  case Intrinsic::x86_tbm_bextri_u64:
2420  // If the RHS is a constant we can try some simplifications.
2421  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
2422  uint64_t Shift = C->getZExtValue();
2423  uint64_t Length = (Shift >> 8) & 0xff;
2424  Shift &= 0xff;
2425  unsigned BitWidth = II->getType()->getIntegerBitWidth();
2426  // If the length is 0 or the shift is out of range, replace with zero.
2427  if (Length == 0 || Shift >= BitWidth)
2428  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0));
2429  // If the LHS is also a constant, we can completely constant fold this.
2430  if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) {
2431  uint64_t Result = InC->getZExtValue() >> Shift;
2432  if (Length > BitWidth)
2433  Length = BitWidth;
2434  Result &= maskTrailingOnes<uint64_t>(Length);
2435  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result));
2436  }
2437  // TODO should we turn this into 'and' if shift is 0? Or 'shl' if we
2438  // are only masking bits that a shift already cleared?
2439  }
2440  break;
2441 
2442  case Intrinsic::x86_bmi_bzhi_32:
2443  case Intrinsic::x86_bmi_bzhi_64:
2444  // If the RHS is a constant we can try some simplifications.
2445  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
2446  uint64_t Index = C->getZExtValue() & 0xff;
2447  unsigned BitWidth = II->getType()->getIntegerBitWidth();
2448  if (Index >= BitWidth)
2449  return replaceInstUsesWith(CI, II->getArgOperand(0));
2450  if (Index == 0)
2451  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0));
2452  // If the LHS is also a constant, we can completely constant fold this.
2453  if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) {
2454  uint64_t Result = InC->getZExtValue();
2455  Result &= maskTrailingOnes<uint64_t>(Index);
2456  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result));
2457  }
2458  // TODO should we convert this to an AND if the RHS is constant?
2459  }
2460  break;
2461 
2462  case Intrinsic::x86_vcvtph2ps_128:
2463  case Intrinsic::x86_vcvtph2ps_256: {
2464  auto Arg = II->getArgOperand(0);
2465  auto ArgType = cast<VectorType>(Arg->getType());
2466  auto RetType = cast<VectorType>(II->getType());
2467  unsigned ArgWidth = ArgType->getNumElements();
2468  unsigned RetWidth = RetType->getNumElements();
2469  assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
2470  assert(ArgType->isIntOrIntVectorTy() &&
2471  ArgType->getScalarSizeInBits() == 16 &&
2472  "CVTPH2PS input type should be 16-bit integer vector");
2473  assert(RetType->getScalarType()->isFloatTy() &&
2474  "CVTPH2PS output type should be 32-bit float vector");
2475 
2476  // Constant folding: Convert to generic half to single conversion.
2477  if (isa<ConstantAggregateZero>(Arg))
2478  return replaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
2479 
2480  if (isa<ConstantDataVector>(Arg)) {
2481  auto VectorHalfAsShorts = Arg;
2482  if (RetWidth < ArgWidth) {
2483  SmallVector<uint32_t, 8> SubVecMask;
2484  for (unsigned i = 0; i != RetWidth; ++i)
2485  SubVecMask.push_back((int)i);
2486  VectorHalfAsShorts = Builder.CreateShuffleVector(
2487  Arg, UndefValue::get(ArgType), SubVecMask);
2488  }
2489 
2490  auto VectorHalfType =
2491  VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
2492  auto VectorHalfs =
2493  Builder.CreateBitCast(VectorHalfAsShorts, VectorHalfType);
2494  auto VectorFloats = Builder.CreateFPExt(VectorHalfs, RetType);
2495  return replaceInstUsesWith(*II, VectorFloats);
2496  }
2497 
2498  // We only use the lowest lanes of the argument.
2499  if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) {
2500  II->setArgOperand(0, V);
2501  return II;
2502  }
2503  break;
2504  }
2505 
2506  case Intrinsic::x86_sse_cvtss2si:
2507  case Intrinsic::x86_sse_cvtss2si64:
2508  case Intrinsic::x86_sse_cvttss2si:
2509  case Intrinsic::x86_sse_cvttss2si64:
2510  case Intrinsic::x86_sse2_cvtsd2si:
2511  case Intrinsic::x86_sse2_cvtsd2si64:
2512  case Intrinsic::x86_sse2_cvttsd2si:
2513  case Intrinsic::x86_sse2_cvttsd2si64:
2514  case Intrinsic::x86_avx512_vcvtss2si32:
2515  case Intrinsic::x86_avx512_vcvtss2si64:
2516  case Intrinsic::x86_avx512_vcvtss2usi32:
2517  case Intrinsic::x86_avx512_vcvtss2usi64:
2518  case Intrinsic::x86_avx512_vcvtsd2si32:
2519  case Intrinsic::x86_avx512_vcvtsd2si64:
2520  case Intrinsic::x86_avx512_vcvtsd2usi32:
2521  case Intrinsic::x86_avx512_vcvtsd2usi64:
2522  case Intrinsic::x86_avx512_cvttss2si:
2523  case Intrinsic::x86_avx512_cvttss2si64:
2524  case Intrinsic::x86_avx512_cvttss2usi:
2525  case Intrinsic::x86_avx512_cvttss2usi64:
2526  case Intrinsic::x86_avx512_cvttsd2si:
2527  case Intrinsic::x86_avx512_cvttsd2si64:
2528  case Intrinsic::x86_avx512_cvttsd2usi:
2529  case Intrinsic::x86_avx512_cvttsd2usi64: {
2530  // These intrinsics only demand the 0th element of their input vectors. If
2531  // we can simplify the input based on that, do so now.
2532  Value *Arg = II->getArgOperand(0);
2533  unsigned VWidth = Arg->getType()->getVectorNumElements();
2534  if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
2535  II->setArgOperand(0, V);
2536  return II;
2537  }
2538  break;
2539  }
2540 
2541  case Intrinsic::x86_sse41_round_ps:
2542  case Intrinsic::x86_sse41_round_pd:
2543  case Intrinsic::x86_avx_round_ps_256:
2544  case Intrinsic::x86_avx_round_pd_256:
2545  case Intrinsic::x86_avx512_mask_rndscale_ps_128:
2546  case Intrinsic::x86_avx512_mask_rndscale_ps_256:
2547  case Intrinsic::x86_avx512_mask_rndscale_ps_512:
2548  case Intrinsic::x86_avx512_mask_rndscale_pd_128:
2549  case Intrinsic::x86_avx512_mask_rndscale_pd_256:
2550  case Intrinsic::x86_avx512_mask_rndscale_pd_512:
2551  case Intrinsic::x86_avx512_mask_rndscale_ss:
2552  case Intrinsic::x86_avx512_mask_rndscale_sd:
2553  if (Value *V = simplifyX86round(*II, Builder))
2554  return replaceInstUsesWith(*II, V);
2555  break;
2556 
2557  case Intrinsic::x86_mmx_pmovmskb:
2558  case Intrinsic::x86_sse_movmsk_ps:
2559  case Intrinsic::x86_sse2_movmsk_pd:
2560  case Intrinsic::x86_sse2_pmovmskb_128:
2561  case Intrinsic::x86_avx_movmsk_pd_256:
2562  case Intrinsic::x86_avx_movmsk_ps_256:
2563  case Intrinsic::x86_avx2_pmovmskb:
2564  if (Value *V = simplifyX86movmsk(*II, Builder))
2565  return replaceInstUsesWith(*II, V);
2566  break;
2567 
2568  case Intrinsic::x86_sse_comieq_ss:
2569  case Intrinsic::x86_sse_comige_ss:
2570  case Intrinsic::x86_sse_comigt_ss:
2571  case Intrinsic::x86_sse_comile_ss:
2572  case Intrinsic::x86_sse_comilt_ss:
2573  case Intrinsic::x86_sse_comineq_ss:
2574  case Intrinsic::x86_sse_ucomieq_ss:
2575  case Intrinsic::x86_sse_ucomige_ss:
2576  case Intrinsic::x86_sse_ucomigt_ss:
2577  case Intrinsic::x86_sse_ucomile_ss:
2578  case Intrinsic::x86_sse_ucomilt_ss:
2579  case Intrinsic::x86_sse_ucomineq_ss:
2580  case Intrinsic::x86_sse2_comieq_sd:
2581  case Intrinsic::x86_sse2_comige_sd:
2582  case Intrinsic::x86_sse2_comigt_sd:
2583  case Intrinsic::x86_sse2_comile_sd:
2584  case Intrinsic::x86_sse2_comilt_sd:
2585  case Intrinsic::x86_sse2_comineq_sd:
2586  case Intrinsic::x86_sse2_ucomieq_sd:
2587  case Intrinsic::x86_sse2_ucomige_sd:
2588  case Intrinsic::x86_sse2_ucomigt_sd:
2589  case Intrinsic::x86_sse2_ucomile_sd:
2590  case Intrinsic::x86_sse2_ucomilt_sd:
2591  case Intrinsic::x86_sse2_ucomineq_sd:
2592  case Intrinsic::x86_avx512_vcomi_ss:
2593  case Intrinsic::x86_avx512_vcomi_sd:
2594  case Intrinsic::x86_avx512_mask_cmp_ss:
2595  case Intrinsic::x86_avx512_mask_cmp_sd: {
2596  // These intrinsics only demand the 0th element of their input vectors. If
2597  // we can simplify the input based on that, do so now.
2598  bool MadeChange = false;
2599  Value *Arg0 = II->getArgOperand(0);
2600  Value *Arg1 = II->getArgOperand(1);
2601  unsigned VWidth = Arg0->getType()->getVectorNumElements();
2602  if (Value *V = SimplifyDemandedVectorEltsLow(Arg0, VWidth, 1)) {
2603  II->setArgOperand(0, V);
2604  MadeChange = true;
2605  }
2606  if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, 1)) {
2607  II->setArgOperand(1, V);
2608  MadeChange = true;
2609  }
2610  if (MadeChange)
2611  return II;
2612  break;
2613  }
2614  case Intrinsic::x86_avx512_cmp_pd_128:
2615  case Intrinsic::x86_avx512_cmp_pd_256:
2616  case Intrinsic::x86_avx512_cmp_pd_512:
2617  case Intrinsic::x86_avx512_cmp_ps_128:
2618  case Intrinsic::x86_avx512_cmp_ps_256:
2619  case Intrinsic::x86_avx512_cmp_ps_512: {
2620  // Folding cmp(sub(a,b),0) -> cmp(a,b) and cmp(0,sub(a,b)) -> cmp(b,a)
2621  Value *Arg0 = II->getArgOperand(0);
2622  Value *Arg1 = II->getArgOperand(1);
2623  bool Arg0IsZero = match(Arg0, m_PosZeroFP());
2624  if (Arg0IsZero)
2625  std::swap(Arg0, Arg1);
2626  Value *A, *B;
2627  // This fold requires only the NINF(not +/- inf) since inf minus
2628  // inf is nan.
2629  // NSZ(No Signed Zeros) is not needed because zeros of any sign are
2630  // equal for both compares.
2631  // NNAN is not needed because nans compare the same for both compares.
2632  // The compare intrinsic uses the above assumptions and therefore
2633  // doesn't require additional flags.
2634  if ((match(Arg0, m_OneUse(m_FSub(m_Value(A), m_Value(B)))) &&
2635  match(Arg1, m_PosZeroFP()) && isa<Instruction>(Arg0) &&
2636  cast<Instruction>(Arg0)->getFastMathFlags().noInfs())) {
2637  if (Arg0IsZero)
2638  std::swap(A, B);
2639  II->setArgOperand(0, A);
2640  II->setArgOperand(1, B);
2641  return II;
2642  }
2643  break;
2644  }
2645 
2646  case Intrinsic::x86_avx512_add_ps_512:
2647  case Intrinsic::x86_avx512_div_ps_512:
2648  case Intrinsic::x86_avx512_mul_ps_512:
2649  case Intrinsic::x86_avx512_sub_ps_512:
2650  case Intrinsic::x86_avx512_add_pd_512:
2651  case Intrinsic::x86_avx512_div_pd_512:
2652  case Intrinsic::x86_avx512_mul_pd_512:
2653  case Intrinsic::x86_avx512_sub_pd_512:
2654  // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2655  // IR operations.
2656  if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
2657  if (R->getValue() == 4) {
2658  Value *Arg0 = II->getArgOperand(0);
2659  Value *Arg1 = II->getArgOperand(1);
2660 
2661  Value *V;
2662  switch (II->getIntrinsicID()) {
2663  default: llvm_unreachable("Case stmts out of sync!");
2664  case Intrinsic::x86_avx512_add_ps_512:
2665  case Intrinsic::x86_avx512_add_pd_512:
2666  V = Builder.CreateFAdd(Arg0, Arg1);
2667  break;
2668  case Intrinsic::x86_avx512_sub_ps_512:
2669  case Intrinsic::x86_avx512_sub_pd_512:
2670  V = Builder.CreateFSub(Arg0, Arg1);
2671  break;
2672  case Intrinsic::x86_avx512_mul_ps_512:
2673  case Intrinsic::x86_avx512_mul_pd_512:
2674  V = Builder.CreateFMul(Arg0, Arg1);
2675  break;
2676  case Intrinsic::x86_avx512_div_ps_512:
2677  case Intrinsic::x86_avx512_div_pd_512:
2678  V = Builder.CreateFDiv(Arg0, Arg1);
2679  break;
2680  }
2681 
2682  return replaceInstUsesWith(*II, V);
2683  }
2684  }
2685  break;
2686 
2687  case Intrinsic::x86_avx512_mask_add_ss_round:
2688  case Intrinsic::x86_avx512_mask_div_ss_round:
2689  case Intrinsic::x86_avx512_mask_mul_ss_round:
2690  case Intrinsic::x86_avx512_mask_sub_ss_round:
2691  case Intrinsic::x86_avx512_mask_add_sd_round:
2692  case Intrinsic::x86_avx512_mask_div_sd_round:
2693  case Intrinsic::x86_avx512_mask_mul_sd_round:
2694  case Intrinsic::x86_avx512_mask_sub_sd_round:
2695  // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2696  // IR operations.
2697  if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(4))) {
2698  if (R->getValue() == 4) {
2699  // Extract the element as scalars.
2700  Value *Arg0 = II->getArgOperand(0);
2701  Value *Arg1 = II->getArgOperand(1);
2702  Value *LHS = Builder.CreateExtractElement(Arg0, (uint64_t)0);
2703  Value *RHS = Builder.CreateExtractElement(Arg1, (uint64_t)0);
2704 
2705  Value *V;
2706  switch (II->getIntrinsicID()) {
2707  default: llvm_unreachable("Case stmts out of sync!");
2708  case Intrinsic::x86_avx512_mask_add_ss_round:
2709  case Intrinsic::x86_avx512_mask_add_sd_round:
2710  V = Builder.CreateFAdd(LHS, RHS);
2711  break;
2712  case Intrinsic::x86_avx512_mask_sub_ss_round:
2713  case Intrinsic::x86_avx512_mask_sub_sd_round:
2714  V = Builder.CreateFSub(LHS, RHS);
2715  break;
2716  case Intrinsic::x86_avx512_mask_mul_ss_round:
2717  case Intrinsic::x86_avx512_mask_mul_sd_round:
2718  V = Builder.CreateFMul(LHS, RHS);
2719  break;
2720  case Intrinsic::x86_avx512_mask_div_ss_round:
2721  case Intrinsic::x86_avx512_mask_div_sd_round:
2722  V = Builder.CreateFDiv(LHS, RHS);
2723  break;
2724  }
2725 
2726  // Handle the masking aspect of the intrinsic.
2727  Value *Mask = II->getArgOperand(3);
2728  auto *C = dyn_cast<ConstantInt>(Mask);
2729  // We don't need a select if we know the mask bit is a 1.
2730  if (!C || !C->getValue()[0]) {
2731  // Cast the mask to an i1 vector and then extract the lowest element.
2732  auto *MaskTy = VectorType::get(Builder.getInt1Ty(),
2733  cast<IntegerType>(Mask->getType())->getBitWidth());
2734  Mask = Builder.CreateBitCast(Mask, MaskTy);
2735  Mask = Builder.CreateExtractElement(Mask, (uint64_t)0);
2736  // Extract the lowest element from the passthru operand.
2737  Value *Passthru = Builder.CreateExtractElement(II->getArgOperand(2),
2738  (uint64_t)0);
2739  V = Builder.CreateSelect(Mask, V, Passthru);
2740  }
2741 
2742  // Insert the result back into the original argument 0.
2743  V = Builder.CreateInsertElement(Arg0, V, (uint64_t)0);
2744 
2745  return replaceInstUsesWith(*II, V);
2746  }
2747  }
2749 
2750  // X86 scalar intrinsics simplified with SimplifyDemandedVectorElts.
2751  case Intrinsic::x86_avx512_mask_max_ss_round:
2752  case Intrinsic::x86_avx512_mask_min_ss_round:
2753  case Intrinsic::x86_avx512_mask_max_sd_round:
2754  case Intrinsic::x86_avx512_mask_min_sd_round:
2755  case Intrinsic::x86_sse_cmp_ss:
2756  case Intrinsic::x86_sse_min_ss:
2757  case Intrinsic::x86_sse_max_ss:
2758  case Intrinsic::x86_sse2_cmp_sd:
2759  case Intrinsic::x86_sse2_min_sd:
2760  case Intrinsic::x86_sse2_max_sd:
2761  case Intrinsic::x86_xop_vfrcz_ss:
2762  case Intrinsic::x86_xop_vfrcz_sd: {
2763  unsigned VWidth = II->getType()->getVectorNumElements();
2764  APInt UndefElts(VWidth, 0);
2765  APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
2766  if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) {
2767  if (V != II)
2768  return replaceInstUsesWith(*II, V);
2769  return II;
2770  }
2771  break;
2772  }
2773  case Intrinsic::x86_sse41_round_ss:
2774  case Intrinsic::x86_sse41_round_sd: {
2775  unsigned VWidth = II->getType()->getVectorNumElements();
2776  APInt UndefElts(VWidth, 0);
2777  APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
2778  if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) {
2779  if (V != II)
2780  return replaceInstUsesWith(*II, V);
2781  return II;
2782  } else if (Value *V = simplifyX86round(*II, Builder))
2783  return replaceInstUsesWith(*II, V);
2784  break;
2785  }
2786 
2787  // Constant fold add/sub with saturation intrinsics.
2788  case Intrinsic::x86_sse2_padds_b:
2789  case Intrinsic::x86_sse2_padds_w:
2790  case Intrinsic::x86_sse2_psubs_b:
2791  case Intrinsic::x86_sse2_psubs_w:
2792  case Intrinsic::x86_avx2_padds_b:
2793  case Intrinsic::x86_avx2_padds_w:
2794  case Intrinsic::x86_avx2_psubs_b:
2795  case Intrinsic::x86_avx2_psubs_w:
2796  case Intrinsic::x86_avx512_padds_b_512:
2797  case Intrinsic::x86_avx512_padds_w_512:
2798  case Intrinsic::x86_avx512_psubs_b_512:
2799  case Intrinsic::x86_avx512_psubs_w_512:
2800  if (Value *V = simplifyX86AddsSubs(*II, Builder))
2801  return replaceInstUsesWith(*II, V);
2802  break;
2803 
2804  // Constant fold ashr( <A x Bi>, Ci ).
2805  // Constant fold lshr( <A x Bi>, Ci ).
2806  // Constant fold shl( <A x Bi>, Ci ).
2807  case Intrinsic::x86_sse2_psrai_d:
2808  case Intrinsic::x86_sse2_psrai_w:
2809  case Intrinsic::x86_avx2_psrai_d:
2810  case Intrinsic::x86_avx2_psrai_w:
2811  case Intrinsic::x86_avx512_psrai_q_128:
2812  case Intrinsic::x86_avx512_psrai_q_256:
2813  case Intrinsic::x86_avx512_psrai_d_512:
2814  case Intrinsic::x86_avx512_psrai_q_512:
2815  case Intrinsic::x86_avx512_psrai_w_512:
2816  case Intrinsic::x86_sse2_psrli_d:
2817  case Intrinsic::x86_sse2_psrli_q:
2818  case Intrinsic::x86_sse2_psrli_w:
2819  case Intrinsic::x86_avx2_psrli_d:
2820  case Intrinsic::x86_avx2_psrli_q:
2821  case Intrinsic::x86_avx2_psrli_w:
2822  case Intrinsic::x86_avx512_psrli_d_512:
2823  case Intrinsic::x86_avx512_psrli_q_512:
2824  case Intrinsic::x86_avx512_psrli_w_512:
2825  case Intrinsic::x86_sse2_pslli_d:
2826  case Intrinsic::x86_sse2_pslli_q:
2827  case Intrinsic::x86_sse2_pslli_w:
2828  case Intrinsic::x86_avx2_pslli_d:
2829  case Intrinsic::x86_avx2_pslli_q:
2830  case Intrinsic::x86_avx2_pslli_w:
2831  case Intrinsic::x86_avx512_pslli_d_512:
2832  case Intrinsic::x86_avx512_pslli_q_512:
2833  case Intrinsic::x86_avx512_pslli_w_512:
2834  if (Value *V = simplifyX86immShift(*II, Builder))
2835  return replaceInstUsesWith(*II, V);
2836  break;
2837 
2838  case Intrinsic::x86_sse2_psra_d:
2839  case Intrinsic::x86_sse2_psra_w:
2840  case Intrinsic::x86_avx2_psra_d:
2841  case Intrinsic::x86_avx2_psra_w:
2842  case Intrinsic::x86_avx512_psra_q_128:
2843  case Intrinsic::x86_avx512_psra_q_256:
2844  case Intrinsic::x86_avx512_psra_d_512:
2845  case Intrinsic::x86_avx512_psra_q_512:
2846  case Intrinsic::x86_avx512_psra_w_512:
2847  case Intrinsic::x86_sse2_psrl_d:
2848  case Intrinsic::x86_sse2_psrl_q:
2849  case Intrinsic::x86_sse2_psrl_w:
2850  case Intrinsic::x86_avx2_psrl_d:
2851  case Intrinsic::x86_avx2_psrl_q:
2852  case Intrinsic::x86_avx2_psrl_w:
2853  case Intrinsic::x86_avx512_psrl_d_512:
2854  case Intrinsic::x86_avx512_psrl_q_512:
2855  case Intrinsic::x86_avx512_psrl_w_512:
2856  case Intrinsic::x86_sse2_psll_d:
2857  case Intrinsic::x86_sse2_psll_q:
2858  case Intrinsic::x86_sse2_psll_w:
2859  case Intrinsic::x86_avx2_psll_d:
2860  case Intrinsic::x86_avx2_psll_q:
2861  case Intrinsic::x86_avx2_psll_w:
2862  case Intrinsic::x86_avx512_psll_d_512:
2863  case Intrinsic::x86_avx512_psll_q_512:
2864  case Intrinsic::x86_avx512_psll_w_512: {
2865  if (Value *V = simplifyX86immShift(*II, Builder))
2866  return replaceInstUsesWith(*II, V);
2867 
2868  // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
2869  // operand to compute the shift amount.
2870  Value *Arg1 = II->getArgOperand(1);
2871  assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 &&
2872  "Unexpected packed shift size");
2873  unsigned VWidth = Arg1->getType()->getVectorNumElements();
2874 
2875  if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) {
2876  II->setArgOperand(1, V);
2877  return II;
2878  }
2879  break;
2880  }
2881 
2882  case Intrinsic::x86_avx2_psllv_d:
2883  case Intrinsic::x86_avx2_psllv_d_256:
2884  case Intrinsic::x86_avx2_psllv_q:
2885  case Intrinsic::x86_avx2_psllv_q_256:
2886  case Intrinsic::x86_avx512_psllv_d_512:
2887  case Intrinsic::x86_avx512_psllv_q_512:
2888  case Intrinsic::x86_avx512_psllv_w_128:
2889  case Intrinsic::x86_avx512_psllv_w_256:
2890  case Intrinsic::x86_avx512_psllv_w_512:
2891  case Intrinsic::x86_avx2_psrav_d:
2892  case Intrinsic::x86_avx2_psrav_d_256:
2893  case Intrinsic::x86_avx512_psrav_q_128:
2894  case Intrinsic::x86_avx512_psrav_q_256:
2895  case Intrinsic::x86_avx512_psrav_d_512:
2896  case Intrinsic::x86_avx512_psrav_q_512:
2897  case Intrinsic::x86_avx512_psrav_w_128:
2898  case Intrinsic::x86_avx512_psrav_w_256:
2899  case Intrinsic::x86_avx512_psrav_w_512:
2900  case Intrinsic::x86_avx2_psrlv_d:
2901  case Intrinsic::x86_avx2_psrlv_d_256:
2902  case Intrinsic::x86_avx2_psrlv_q:
2903  case Intrinsic::x86_avx2_psrlv_q_256:
2904  case Intrinsic::x86_avx512_psrlv_d_512:
2905  case Intrinsic::x86_avx512_psrlv_q_512:
2906  case Intrinsic::x86_avx512_psrlv_w_128:
2907  case Intrinsic::x86_avx512_psrlv_w_256:
2908  case Intrinsic::x86_avx512_psrlv_w_512:
2909  if (Value *V = simplifyX86varShift(*II, Builder))
2910  return replaceInstUsesWith(*II, V);
2911  break;
2912 
2913  case Intrinsic::x86_sse2_packssdw_128:
2914  case Intrinsic::x86_sse2_packsswb_128:
2915  case Intrinsic::x86_avx2_packssdw:
2916  case Intrinsic::x86_avx2_packsswb:
2917  case Intrinsic::x86_avx512_packssdw_512:
2918  case Intrinsic::x86_avx512_packsswb_512:
2919  if (Value *V = simplifyX86pack(*II, true))
2920  return replaceInstUsesWith(*II, V);
2921  break;
2922 
2923  case Intrinsic::x86_sse2_packuswb_128:
2924  case Intrinsic::x86_sse41_packusdw:
2925  case Intrinsic::x86_avx2_packusdw:
2926  case Intrinsic::x86_avx2_packuswb:
2927  case Intrinsic::x86_avx512_packusdw_512:
2928  case Intrinsic::x86_avx512_packuswb_512:
2929  if (Value *V = simplifyX86pack(*II, false))
2930  return replaceInstUsesWith(*II, V);
2931  break;
2932 
2933  case Intrinsic::x86_pclmulqdq:
2934  case Intrinsic::x86_pclmulqdq_256:
2935  case Intrinsic::x86_pclmulqdq_512: {
2936  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
2937  unsigned Imm = C->getZExtValue();
2938 
2939  bool MadeChange = false;
2940  Value *Arg0 = II->getArgOperand(0);
2941  Value *Arg1 = II->getArgOperand(1);
2942  unsigned VWidth = Arg0->getType()->getVectorNumElements();
2943 
2944  APInt UndefElts1(VWidth, 0);
2945  APInt DemandedElts1 = APInt::getSplat(VWidth,
2946  APInt(2, (Imm & 0x01) ? 2 : 1));
2947  if (Value *V = SimplifyDemandedVectorElts(Arg0, DemandedElts1,
2948  UndefElts1)) {
2949  II->setArgOperand(0, V);
2950  MadeChange = true;
2951  }
2952 
2953  APInt UndefElts2(VWidth, 0);
2954  APInt DemandedElts2 = APInt::getSplat(VWidth,
2955  APInt(2, (Imm & 0x10) ? 2 : 1));
2956  if (Value *V = SimplifyDemandedVectorElts(Arg1, DemandedElts2,
2957  UndefElts2)) {
2958  II->setArgOperand(1, V);
2959  MadeChange = true;
2960  }
2961 
2962  // If either input elements are undef, the result is zero.
2963  if (DemandedElts1.isSubsetOf(UndefElts1) ||
2964  DemandedElts2.isSubsetOf(UndefElts2))
2965  return replaceInstUsesWith(*II,
2966  ConstantAggregateZero::get(II->getType()));
2967 
2968  if (MadeChange)
2969  return II;
2970  }
2971  break;
2972  }
2973 
2974  case Intrinsic::x86_sse41_insertps:
2975  if (Value *V = simplifyX86insertps(*II, Builder))
2976  return replaceInstUsesWith(*II, V);
2977  break;
2978 
2979  case Intrinsic::x86_sse4a_extrq: {
2980  Value *Op0 = II->getArgOperand(0);
2981  Value *Op1 = II->getArgOperand(1);
2982  unsigned VWidth0 = Op0->getType()->getVectorNumElements();
2983  unsigned VWidth1 = Op1->getType()->getVectorNumElements();
2984  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2985  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
2986  VWidth1 == 16 && "Unexpected operand sizes");
2987 
2988  // See if we're dealing with constant values.
2989  Constant *C1 = dyn_cast<Constant>(Op1);
2990  ConstantInt *CILength =
2991  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
2992  : nullptr;
2993  ConstantInt *CIIndex =
2994  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
2995  : nullptr;
2996 
2997  // Attempt to simplify to a constant, shuffle vector or EXTRQI call.
2998  if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder))
2999  return replaceInstUsesWith(*II, V);
3000 
3001  // EXTRQ only uses the lowest 64-bits of the first 128-bit vector
3002  // operands and the lowest 16-bits of the second.
3003  bool MadeChange = false;
3004  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
3005  II->setArgOperand(0, V);
3006  MadeChange = true;
3007  }
3008  if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
3009  II->setArgOperand(1, V);
3010  MadeChange = true;
3011  }
3012  if (MadeChange)
3013  return II;
3014  break;
3015  }
3016 
3017  case Intrinsic::x86_sse4a_extrqi: {
3018  // EXTRQI: Extract Length bits starting from Index. Zero pad the remaining
3019  // bits of the lower 64-bits. The upper 64-bits are undefined.
3020  Value *Op0 = II->getArgOperand(0);
3021  unsigned VWidth = Op0->getType()->getVectorNumElements();
3022  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
3023  "Unexpected operand size");
3024 
3025  // See if we're dealing with constant values.
3026  ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(1));
3027  ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(2));
3028 
3029  // Attempt to simplify to a constant or shuffle vector.
3030  if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder))
3031  return replaceInstUsesWith(*II, V);
3032 
3033  // EXTRQI only uses the lowest 64-bits of the first 128-bit vector
3034  // operand.
3035  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
3036  II->setArgOperand(0, V);
3037  return II;
3038  }
3039  break;
3040  }
3041 
3042  case Intrinsic::x86_sse4a_insertq: {
3043  Value *Op0 = II->getArgOperand(0);
3044  Value *Op1 = II->getArgOperand(1);
3045  unsigned VWidth = Op0->getType()->getVectorNumElements();
3046  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
3047  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
3048  Op1->getType()->getVectorNumElements() == 2 &&
3049  "Unexpected operand size");
3050 
3051  // See if we're dealing with constant values.
3052  Constant *C1 = dyn_cast<Constant>(Op1);
3053  ConstantInt *CI11 =
3054  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
3055  : nullptr;
3056 
3057  // Attempt to simplify to a constant, shuffle vector or INSERTQI call.
3058  if (CI11) {
3059  const APInt &V11 = CI11->getValue();
3060  APInt Len = V11.zextOrTrunc(6);
3061  APInt Idx = V11.lshr(8).zextOrTrunc(6);
3062  if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder))
3063  return replaceInstUsesWith(*II, V);
3064  }
3065 
3066  // INSERTQ only uses the lowest 64-bits of the first 128-bit vector
3067  // operand.
3068  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
3069  II->setArgOperand(0, V);
3070  return II;
3071  }
3072  break;
3073  }
3074 
3075  case Intrinsic::x86_sse4a_insertqi: {
3076  // INSERTQI: Extract lowest Length bits from lower half of second source and
3077  // insert over first source starting at Index bit. The upper 64-bits are
3078  // undefined.
3079  Value *Op0 = II->getArgOperand(0);
3080  Value *Op1 = II->getArgOperand(1);
3081  unsigned VWidth0 = Op0->getType()->getVectorNumElements();
3082  unsigned VWidth1 = Op1->getType()->getVectorNumElements();
3083  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
3084  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
3085  VWidth1 == 2 && "Unexpected operand sizes");
3086 
3087  // See if we're dealing with constant values.
3088  ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(2));
3089  ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3));
3090 
3091  // Attempt to simplify to a constant or shuffle vector.
3092  if (CILength && CIIndex) {
3093  APInt Len = CILength->getValue().zextOrTrunc(6);
3094  APInt Idx = CIIndex->getValue().zextOrTrunc(6);
3095  if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder))
3096  return replaceInstUsesWith(*II, V);
3097  }
3098 
3099  // INSERTQI only uses the lowest 64-bits of the first two 128-bit vector
3100  // operands.
3101  bool MadeChange = false;
3102  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
3103  II->setArgOperand(0, V);
3104  MadeChange = true;
3105  }
3106  if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
3107  II->setArgOperand(1, V);
3108  MadeChange = true;
3109  }
3110  if (MadeChange)
3111  return II;
3112  break;
3113  }
3114 
3115  case Intrinsic::x86_sse41_pblendvb:
3116  case Intrinsic::x86_sse41_blendvps:
3117  case Intrinsic::x86_sse41_blendvpd:
3118  case Intrinsic::x86_avx_blendv_ps_256:
3119  case Intrinsic::x86_avx_blendv_pd_256:
3120  case Intrinsic::x86_avx2_pblendvb: {
3121  // fold (blend A, A, Mask) -> A
3122  Value *Op0 = II->getArgOperand(0);
3123  Value *Op1 = II->getArgOperand(1);
3124  Value *Mask = II->getArgOperand(2);
3125  if (Op0 == Op1)
3126  return replaceInstUsesWith(CI, Op0);
3127 
3128  // Zero Mask - select 1st argument.
3129  if (isa<ConstantAggregateZero>(Mask))
3130  return replaceInstUsesWith(CI, Op0);
3131 
3132  // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
3133  if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask)) {
3134  Constant *NewSelector = getNegativeIsTrueBoolVec(ConstantMask);
3135  return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
3136  }
3137 
3138  // Convert to a vector select if we can bypass casts and find a boolean
3139  // vector condition value.
3140  Value *BoolVec;
3141  Mask = peekThroughBitcast(Mask);
3142  if (match(Mask, m_SExt(m_Value(BoolVec))) &&
3143  BoolVec->getType()->isVectorTy() &&
3144  BoolVec->getType()->getScalarSizeInBits() == 1) {
3145  assert(Mask->getType()->getPrimitiveSizeInBits() ==
3146  II->getType()->getPrimitiveSizeInBits() &&
3147  "Not expecting mask and operands with different sizes");
3148 
3149  unsigned NumMaskElts = Mask->getType()->getVectorNumElements();
3150  unsigned NumOperandElts = II->getType()->getVectorNumElements();
3151  if (NumMaskElts == NumOperandElts)
3152  return SelectInst::Create(BoolVec, Op1, Op0);
3153 
3154  // If the mask has less elements than the operands, each mask bit maps to
3155  // multiple elements of the operands. Bitcast back and forth.
3156  if (NumMaskElts < NumOperandElts) {
3157  Value *CastOp0 = Builder.CreateBitCast(Op0, Mask->getType());
3158  Value *CastOp1 = Builder.CreateBitCast(Op1, Mask->getType());
3159  Value *Sel = Builder.CreateSelect(BoolVec, CastOp1, CastOp0);
3160  return new BitCastInst(Sel, II->getType());
3161  }
3162  }
3163 
3164  break;
3165  }
3166 
3167  case Intrinsic::x86_ssse3_pshuf_b_128:
3168  case Intrinsic::x86_avx2_pshuf_b:
3169  case Intrinsic::x86_avx512_pshuf_b_512:
3170  if (Value *V = simplifyX86pshufb(*II, Builder))
3171  return replaceInstUsesWith(*II, V);
3172  break;
3173 
3174  case Intrinsic::x86_avx_vpermilvar_ps:
3175  case Intrinsic::x86_avx_vpermilvar_ps_256:
3176  case Intrinsic::x86_avx512_vpermilvar_ps_512:
3177  case Intrinsic::x86_avx_vpermilvar_pd:
3178  case Intrinsic::x86_avx_vpermilvar_pd_256:
3179  case Intrinsic::x86_avx512_vpermilvar_pd_512:
3180  if (Value *V = simplifyX86vpermilvar(*II, Builder))
3181  return replaceInstUsesWith(*II, V);
3182  break;
3183 
3184  case Intrinsic::x86_avx2_permd:
3185  case Intrinsic::x86_avx2_permps:
3186  case Intrinsic::x86_avx512_permvar_df_256:
3187  case Intrinsic::x86_avx512_permvar_df_512:
3188  case Intrinsic::x86_avx512_permvar_di_256:
3189  case Intrinsic::x86_avx512_permvar_di_512:
3190  case Intrinsic::x86_avx512_permvar_hi_128:
3191  case Intrinsic::x86_avx512_permvar_hi_256:
3192  case Intrinsic::x86_avx512_permvar_hi_512:
3193  case Intrinsic::x86_avx512_permvar_qi_128:
3194  case Intrinsic::x86_avx512_permvar_qi_256:
3195  case Intrinsic::x86_avx512_permvar_qi_512:
3196  case Intrinsic::x86_avx512_permvar_sf_512:
3197  case Intrinsic::x86_avx512_permvar_si_512:
3198  if (Value *V = simplifyX86vpermv(*II, Builder))
3199  return replaceInstUsesWith(*II, V);
3200  break;
3201 
3202  case Intrinsic::x86_avx_maskload_ps:
3203  case Intrinsic::x86_avx_maskload_pd:
3204  case Intrinsic::x86_avx_maskload_ps_256:
3205  case Intrinsic::x86_avx_maskload_pd_256:
3206  case Intrinsic::x86_avx2_maskload_d:
3207  case Intrinsic::x86_avx2_maskload_q:
3208  case Intrinsic::x86_avx2_maskload_d_256:
3209  case Intrinsic::x86_avx2_maskload_q_256:
3210  if (Instruction *I = simplifyX86MaskedLoad(*II, *this))
3211  return I;
3212  break;
3213 
3214  case Intrinsic::x86_sse2_maskmov_dqu:
3215  case Intrinsic::x86_avx_maskstore_ps:
3216  case Intrinsic::x86_avx_maskstore_pd:
3217  case Intrinsic::x86_avx_maskstore_ps_256:
3218  case Intrinsic::x86_avx_maskstore_pd_256:
3219  case Intrinsic::x86_avx2_maskstore_d:
3220  case Intrinsic::x86_avx2_maskstore_q:
3221  case Intrinsic::x86_avx2_maskstore_d_256:
3222  case Intrinsic::x86_avx2_maskstore_q_256:
3223  if (simplifyX86MaskedStore(*II, *this))
3224  return nullptr;
3225  break;
3226 
3227  case Intrinsic::x86_xop_vpcomb:
3228  case Intrinsic::x86_xop_vpcomd:
3229  case Intrinsic::x86_xop_vpcomq:
3230  case Intrinsic::x86_xop_vpcomw:
3231  if (Value *V = simplifyX86vpcom(*II, Builder, true))
3232  return replaceInstUsesWith(*II, V);
3233  break;
3234 
3235  case Intrinsic::x86_xop_vpcomub:
3236  case Intrinsic::x86_xop_vpcomud:
3237  case Intrinsic::x86_xop_vpcomuq:
3238  case Intrinsic::x86_xop_vpcomuw:
3239  if (Value *V = simplifyX86vpcom(*II, Builder, false))
3240  return replaceInstUsesWith(*II, V);
3241  break;
3242 
3243  case Intrinsic::ppc_altivec_vperm:
3244  // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
3245  // Note that ppc_altivec_vperm has a big-endian bias, so when creating
3246  // a vectorshuffle for little endian, we must undo the transformation
3247  // performed on vec_perm in altivec.h. That is, we must complement
3248  // the permutation mask with respect to 31 and reverse the order of
3249  // V1 and V2.
3250  if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
3251  assert(Mask->getType()->getVectorNumElements() == 16 &&
3252  "Bad type for intrinsic!");
3253 
3254  // Check that all of the elements are integer constants or undefs.
3255  bool AllEltsOk = true;
3256  for (unsigned i = 0; i != 16; ++i) {
3257  Constant *Elt = Mask->getAggregateElement(i);
3258  if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
3259  AllEltsOk = false;
3260  break;
3261  }
3262  }
3263 
3264  if (AllEltsOk) {
3265  // Cast the input vectors to byte vectors.
3266  Value *Op0 = Builder.CreateBitCast(II->getArgOperand(0),
3267  Mask->getType());
3268  Value *Op1 = Builder.CreateBitCast(II->getArgOperand(1),
3269  Mask->getType());
3270  Value *Result = UndefValue::get(Op0->getType());
3271 
3272  // Only extract each element once.
3273  Value *ExtractedElts[32];
3274  memset(ExtractedElts, 0, sizeof(ExtractedElts));
3275 
3276  for (unsigned i = 0; i != 16; ++i) {
3277  if (isa<UndefValue>(Mask->getAggregateElement(i)))
3278  continue;
3279  unsigned Idx =
3280  cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
3281  Idx &= 31; // Match the hardware behavior.
3282  if (DL.isLittleEndian())
3283  Idx = 31 - Idx;
3284 
3285  if (!ExtractedElts[Idx]) {
3286  Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
3287  Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
3288  ExtractedElts[Idx] =
3289  Builder.CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
3290  Builder.getInt32(Idx&15));
3291  }
3292 
3293  // Insert this value into the result vector.
3294  Result = Builder.CreateInsertElement(Result, ExtractedElts[Idx],
3295  Builder.getInt32(i));
3296  }
3297  return CastInst::Create(Instruction::BitCast, Result, CI.getType());
3298  }
3299  }
3300  break;
3301 
3302  case Intrinsic::arm_neon_vld1: {
3303  unsigned MemAlign = getKnownAlignment(II->getArgOperand(0),
3304  DL, II, &AC, &DT);
3305  if (Value *V = simplifyNeonVld1(*II, MemAlign, Builder))
3306  return replaceInstUsesWith(*II, V);
3307  break;
3308  }
3309 
3310  case Intrinsic::arm_neon_vld2:
3311  case Intrinsic::arm_neon_vld3:
3312  case Intrinsic::arm_neon_vld4:
3313  case Intrinsic::arm_neon_vld2lane:
3314  case Intrinsic::arm_neon_vld3lane:
3315  case Intrinsic::arm_neon_vld4lane:
3316  case Intrinsic::arm_neon_vst1:
3317  case Intrinsic::arm_neon_vst2:
3318  case Intrinsic::arm_neon_vst3:
3319  case Intrinsic::arm_neon_vst4:
3320  case Intrinsic::arm_neon_vst2lane:
3321  case Intrinsic::arm_neon_vst3lane:
3322  case Intrinsic::arm_neon_vst4lane: {
3323  unsigned MemAlign =
3324  getKnownAlignment(II->getArgOperand(0), DL, II, &AC, &DT);
3325  unsigned AlignArg = II->getNumArgOperands() - 1;
3326  ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
3327  if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
3328  II->setArgOperand(AlignArg,
3329  ConstantInt::get(Type::getInt32Ty(II->getContext()),
3330  MemAlign, false));
3331  return II;
3332  }
3333  break;
3334  }
3335 
3336  case Intrinsic::arm_neon_vtbl1:
3337  case Intrinsic::aarch64_neon_tbl1:
3338  if (Value *V = simplifyNeonTbl1(*II, Builder))
3339  return replaceInstUsesWith(*II, V);
3340  break;
3341 
3342  case Intrinsic::arm_neon_vmulls:
3343  case Intrinsic::arm_neon_vmullu:
3344  case Intrinsic::aarch64_neon_smull:
3345  case Intrinsic::aarch64_neon_umull: {
3346  Value *Arg0 = II->getArgOperand(0);
3347  Value *Arg1 = II->getArgOperand(1);
3348 
3349  // Handle mul by zero first:
3350  if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
3351  return replaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
3352  }
3353 
3354  // Check for constant LHS & RHS - in this case we just simplify.
3355  bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
3356  II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
3357  VectorType *NewVT = cast<VectorType>(II->getType());
3358  if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
3359  if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
3360  CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
3361  CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
3362 
3363  return replaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
3364  }
3365 
3366  // Couldn't simplify - canonicalize constant to the RHS.
3367  std::swap(Arg0, Arg1);
3368  }
3369 
3370  // Handle mul by one:
3371  if (Constant *CV1 = dyn_cast<Constant>(Arg1))
3372  if (ConstantInt *Splat =
3373  dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
3374  if (Splat->isOne())
3375  return CastInst::CreateIntegerCast(Arg0, II->getType(),
3376  /*isSigned=*/!Zext);
3377 
3378  break;
3379  }
3380  case Intrinsic::arm_neon_aesd:
3381  case Intrinsic::arm_neon_aese:
3382  case Intrinsic::aarch64_crypto_aesd:
3383  case Intrinsic::aarch64_crypto_aese: {
3384  Value *DataArg = II->getArgOperand(0);
3385  Value *KeyArg = II->getArgOperand(1);
3386 
3387  // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR
3388  Value *Data, *Key;
3389  if (match(KeyArg, m_ZeroInt()) &&
3390  match(DataArg, m_Xor(m_Value(Data), m_Value(Key)))) {
3391  II->setArgOperand(0, Data);
3392  II->setArgOperand(1, Key);
3393  return II;
3394  }
3395  break;
3396  }
3397  case Intrinsic::amdgcn_rcp: {
3398  Value *Src = II->getArgOperand(0);
3399 
3400  // TODO: Move to ConstantFolding/InstSimplify?
3401  if (isa<UndefValue>(Src))
3402  return replaceInstUsesWith(CI, Src);
3403 
3404  if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) {
3405  const APFloat &ArgVal = C->getValueAPF();
3406  APFloat Val(ArgVal.getSemantics(), 1.0);
3407  APFloat::opStatus Status = Val.divide(ArgVal,
3409  // Only do this if it was exact and therefore not dependent on the
3410  // rounding mode.
3411  if (Status == APFloat::opOK)
3412  return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
3413  }
3414 
3415  break;
3416  }
3417  case Intrinsic::amdgcn_rsq: {
3418  Value *Src = II->getArgOperand(0);
3419 
3420  // TODO: Move to ConstantFolding/InstSimplify?
3421  if (isa<UndefValue>(Src))
3422  return replaceInstUsesWith(CI, Src);
3423  break;
3424  }
3425  case Intrinsic::amdgcn_frexp_mant:
3426  case Intrinsic::amdgcn_frexp_exp: {
3427  Value *Src = II->getArgOperand(0);
3428  if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) {
3429  int Exp;
3430  APFloat Significand = frexp(C->getValueAPF(), Exp,
3432 
3433  if (II->getIntrinsicID() == Intrinsic::amdgcn_frexp_mant) {
3434  return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(),
3435  Significand));
3436  }
3437 
3438  // Match instruction special case behavior.
3439  if (Exp == APFloat::IEK_NaN || Exp == APFloat::IEK_Inf)
3440  Exp = 0;
3441 
3442  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Exp));
3443  }
3444 
3445  if (isa<UndefValue>(Src))
3446  return replaceInstUsesWith(CI, UndefValue::get(II->getType()));
3447 
3448  break;
3449  }
3450  case Intrinsic::amdgcn_class: {
3451  enum {
3452  S_NAN = 1 << 0, // Signaling NaN
3453  Q_NAN = 1 << 1, // Quiet NaN
3454  N_INFINITY = 1 << 2, // Negative infinity
3455  N_NORMAL = 1 << 3, // Negative normal
3456  N_SUBNORMAL = 1 << 4, // Negative subnormal
3457  N_ZERO = 1 << 5, // Negative zero
3458  P_ZERO = 1 << 6, // Positive zero
3459  P_SUBNORMAL = 1 << 7, // Positive subnormal
3460  P_NORMAL = 1 << 8, // Positive normal
3461  P_INFINITY = 1 << 9 // Positive infinity
3462  };
3463 
3464  const uint32_t FullMask = S_NAN | Q_NAN | N_INFINITY | N_NORMAL |
3466 
3467  Value *Src0 = II->getArgOperand(0);
3468  Value *Src1 = II->getArgOperand(1);
3469  const ConstantInt *CMask = dyn_cast<ConstantInt>(Src1);
3470  if (!CMask) {
3471  if (isa<UndefValue>(Src0))
3472  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3473 
3474  if (isa<UndefValue>(Src1))
3475  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false));
3476  break;
3477  }
3478 
3479  uint32_t Mask = CMask->getZExtValue();
3480 
3481  // If all tests are made, it doesn't matter what the value is.
3482  if ((Mask & FullMask) == FullMask)
3483  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), true));
3484 
3485  if ((Mask & FullMask) == 0)
3486  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false));
3487 
3488  if (Mask == (S_NAN | Q_NAN)) {
3489  // Equivalent of isnan. Replace with standard fcmp.
3490  Value *FCmp = Builder.CreateFCmpUNO(Src0, Src0);
3491  FCmp->takeName(II);
3492  return replaceInstUsesWith(*II, FCmp);
3493  }
3494 
3495  if (Mask == (N_ZERO | P_ZERO)) {
3496  // Equivalent of == 0.
3497  Value *FCmp = Builder.CreateFCmpOEQ(
3498  Src0, ConstantFP::get(Src0->getType(), 0.0));
3499 
3500  FCmp->takeName(II);
3501  return replaceInstUsesWith(*II, FCmp);
3502  }
3503 
3504  // fp_class (nnan x), qnan|snan|other -> fp_class (nnan x), other
3505  if (((Mask & S_NAN) || (Mask & Q_NAN)) && isKnownNeverNaN(Src0, &TLI)) {
3506  II->setArgOperand(1, ConstantInt::get(Src1->getType(),
3507  Mask & ~(S_NAN | Q_NAN)));
3508  return II;
3509  }
3510 
3511  const ConstantFP *CVal = dyn_cast<ConstantFP>(Src0);
3512  if (!CVal) {
3513  if (isa<UndefValue>(Src0))
3514  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3515 
3516  // Clamp mask to used bits
3517  if ((Mask & FullMask) != Mask) {
3518  CallInst *NewCall = Builder.CreateCall(II->getCalledFunction(),
3519  { Src0, ConstantInt::get(Src1->getType(), Mask & FullMask) }
3520  );
3521 
3522  NewCall->takeName(II);
3523  return replaceInstUsesWith(*II, NewCall);
3524  }
3525 
3526  break;
3527  }
3528 
3529  const APFloat &Val = CVal->getValueAPF();
3530 
3531  bool Result =
3532  ((Mask & S_NAN) && Val.isNaN() && Val.isSignaling()) ||
3533  ((Mask & Q_NAN) && Val.isNaN() && !Val.isSignaling()) ||
3534  ((Mask & N_INFINITY) && Val.isInfinity() && Val.isNegative()) ||
3535  ((Mask & N_NORMAL) && Val.isNormal() && Val.isNegative()) ||
3536  ((Mask & N_SUBNORMAL) && Val.isDenormal() && Val.isNegative()) ||
3537  ((Mask & N_ZERO) && Val.isZero() && Val.isNegative()) ||
3538  ((Mask & P_ZERO) && Val.isZero() && !Val.isNegative()) ||
3539  ((Mask & P_SUBNORMAL) && Val.isDenormal() && !Val.isNegative()) ||
3540  ((Mask & P_NORMAL) && Val.isNormal() && !Val.isNegative()) ||
3541  ((Mask & P_INFINITY) && Val.isInfinity() && !Val.isNegative());
3542 
3543  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), Result));
3544  }
3545  case Intrinsic::amdgcn_cvt_pkrtz: {
3546  Value *Src0 = II->getArgOperand(0);
3547  Value *Src1 = II->getArgOperand(1);
3548  if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) {
3549  if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) {
3550  const fltSemantics &HalfSem
3551  = II->getType()->getScalarType()->getFltSemantics();
3552  bool LosesInfo;
3553  APFloat Val0 = C0->getValueAPF();
3554  APFloat Val1 = C1->getValueAPF();
3555  Val0.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo);
3556  Val1.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo);
3557 
3558  Constant *Folded = ConstantVector::get({
3559  ConstantFP::get(II->getContext(), Val0),
3560  ConstantFP::get(II->getContext(), Val1) });
3561  return replaceInstUsesWith(*II, Folded);
3562  }
3563  }
3564 
3565  if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1))
3566  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3567 
3568  break;
3569  }
3570  case Intrinsic::amdgcn_cvt_pknorm_i16:
3571  case Intrinsic::amdgcn_cvt_pknorm_u16:
3572  case Intrinsic::amdgcn_cvt_pk_i16:
3573  case Intrinsic::amdgcn_cvt_pk_u16: {
3574  Value *Src0 = II->getArgOperand(0);
3575  Value *Src1 = II->getArgOperand(1);
3576 
3577  if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1))
3578  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3579 
3580  break;
3581  }
3582  case Intrinsic::amdgcn_ubfe:
3583  case Intrinsic::amdgcn_sbfe: {
3584  // Decompose simple cases into standard shifts.
3585  Value *Src = II->getArgOperand(0);
3586  if (isa<UndefValue>(Src))
3587  return replaceInstUsesWith(*II, Src);
3588 
3589  unsigned Width;
3590  Type *Ty = II->getType();
3591  unsigned IntSize = Ty->getIntegerBitWidth();
3592 
3593  ConstantInt *CWidth = dyn_cast<ConstantInt>(II->getArgOperand(2));
3594  if (CWidth) {
3595  Width = CWidth->getZExtValue();
3596  if ((Width & (IntSize - 1)) == 0)
3597  return replaceInstUsesWith(*II, ConstantInt::getNullValue(Ty));
3598 
3599  if (Width >= IntSize) {
3600  // Hardware ignores high bits, so remove those.
3601  II->setArgOperand(2, ConstantInt::get(CWidth->getType(),
3602  Width & (IntSize - 1)));
3603  return II;
3604  }
3605  }
3606 
3607  unsigned Offset;
3608  ConstantInt *COffset = dyn_cast<ConstantInt>(II->getArgOperand(1));
3609  if (COffset) {
3610  Offset = COffset->getZExtValue();
3611  if (Offset >= IntSize) {
3612  II->setArgOperand(1, ConstantInt::get(COffset->getType(),
3613  Offset & (IntSize - 1)));
3614  return II;
3615  }
3616  }
3617 
3618  bool Signed = II->getIntrinsicID() == Intrinsic::amdgcn_sbfe;
3619 
3620  if (!CWidth || !COffset)
3621  break;
3622 
3623  // The case of Width == 0 is handled above, which makes this tranformation
3624  // safe. If Width == 0, then the ashr and lshr instructions become poison
3625  // value since the shift amount would be equal to the bit size.
3626  assert(Width != 0);
3627 
3628  // TODO: This allows folding to undef when the hardware has specific
3629  // behavior?
3630  if (Offset + Width < IntSize) {
3631  Value *Shl = Builder.CreateShl(Src, IntSize - Offset - Width);
3632  Value *RightShift = Signed ? Builder.CreateAShr(Shl, IntSize - Width)
3633  : Builder.CreateLShr(Shl, IntSize - Width);
3634  RightShift->takeName(II);
3635  return replaceInstUsesWith(*II, RightShift);
3636  }
3637 
3638  Value *RightShift = Signed ? Builder.CreateAShr(Src, Offset)
3639  : Builder.CreateLShr(Src, Offset);
3640 
3641  RightShift->takeName(II);
3642  return replaceInstUsesWith(*II, RightShift);
3643  }
3644  case Intrinsic::amdgcn_exp:
3645  case Intrinsic::amdgcn_exp_compr: {
3646  ConstantInt *En = dyn_cast<ConstantInt>(II->getArgOperand(1));
3647  if (!En) // Illegal.
3648  break;
3649 
3650  unsigned EnBits = En->getZExtValue();
3651  if (EnBits == 0xf)
3652  break; // All inputs enabled.
3653 
3654  bool IsCompr = II->getIntrinsicID() == Intrinsic::amdgcn_exp_compr;
3655  bool Changed = false;
3656  for (int I = 0; I < (IsCompr ? 2 : 4); ++I) {
3657  if ((!IsCompr && (EnBits & (1 << I)) == 0) ||
3658  (IsCompr && ((EnBits & (0x3 << (2 * I))) == 0))) {
3659  Value *Src = II->getArgOperand(I + 2);
3660  if (!isa<UndefValue>(Src)) {
3661  II->setArgOperand(I + 2, UndefValue::get(Src->getType()));
3662  Changed = true;
3663  }
3664  }
3665  }
3666 
3667  if (Changed)
3668  return II;
3669 
3670  break;
3671  }
3672  case Intrinsic::amdgcn_fmed3: {
3673  // Note this does not preserve proper sNaN behavior if IEEE-mode is enabled
3674  // for the shader.
3675 
3676  Value *Src0 = II->getArgOperand(0);
3677  Value *Src1 = II->getArgOperand(1);
3678  Value *Src2 = II->getArgOperand(2);
3679 
3680  // Checking for NaN before canonicalization provides better fidelity when
3681  // mapping other operations onto fmed3 since the order of operands is
3682  // unchanged.
3683  CallInst *NewCall = nullptr;
3684  if (match(Src0, m_NaN()) || isa<UndefValue>(Src0)) {
3685  NewCall = Builder.CreateMinNum(Src1, Src2);
3686  } else if (match(Src1, m_NaN()) || isa<UndefValue>(Src1)) {
3687  NewCall = Builder.CreateMinNum(Src0, Src2);
3688  } else if (match(Src2, m_NaN()) || isa<UndefValue>(Src2)) {
3689  NewCall = Builder.CreateMaxNum(Src0, Src1);
3690  }
3691 
3692  if (NewCall) {
3693  NewCall->copyFastMathFlags(II);
3694  NewCall->takeName(II);
3695  return replaceInstUsesWith(*II, NewCall);
3696  }
3697 
3698  bool Swap = false;
3699  // Canonicalize constants to RHS operands.
3700  //
3701  // fmed3(c0, x, c1) -> fmed3(x, c0, c1)
3702  if (isa<Constant>(Src0) && !isa<Constant>(Src1)) {
3703  std::swap(Src0, Src1);
3704  Swap = true;
3705  }
3706 
3707  if (isa<Constant>(Src1) && !isa<Constant>(Src2)) {
3708  std::swap(Src1, Src2);
3709  Swap = true;
3710  }
3711 
3712  if (isa<Constant>(Src0) && !isa<Constant>(Src1)) {
3713  std::swap(Src0, Src1);
3714  Swap = true;
3715  }
3716 
3717  if (Swap) {
3718  II->setArgOperand(0, Src0);
3719  II->setArgOperand(1, Src1);
3720  II->setArgOperand(2, Src2);
3721  return II;
3722  }
3723 
3724  if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) {
3725  if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) {
3726  if (const ConstantFP *C2 = dyn_cast<ConstantFP>(Src2)) {
3727  APFloat Result = fmed3AMDGCN(C0->getValueAPF(), C1->getValueAPF(),
3728  C2->getValueAPF());
3729  return replaceInstUsesWith(*II,
3730  ConstantFP::get(Builder.getContext(), Result));
3731  }
3732  }
3733  }
3734 
3735  break;
3736  }
3737  case Intrinsic::amdgcn_icmp:
3738  case Intrinsic::amdgcn_fcmp: {
3739  const ConstantInt *CC = dyn_cast<ConstantInt>(II->getArgOperand(2));
3740  if (!CC)
3741  break;
3742 
3743  // Guard against invalid arguments.
3744  int64_t CCVal = CC->getZExtValue();
3745  bool IsInteger = II->getIntrinsicID() == Intrinsic::amdgcn_icmp;
3746  if ((IsInteger && (CCVal < CmpInst::FIRST_ICMP_PREDICATE ||
3747  CCVal > CmpInst::LAST_ICMP_PREDICATE)) ||
3748  (!IsInteger && (CCVal < CmpInst::FIRST_FCMP_PREDICATE ||
3749  CCVal > CmpInst::LAST_FCMP_PREDICATE)))
3750  break;
3751 
3752  Value *Src0 = II->getArgOperand(0);
3753  Value *Src1 = II->getArgOperand(1);
3754 
3755  if (auto *CSrc0 = dyn_cast<Constant>(Src0)) {
3756  if (auto *CSrc1 = dyn_cast<Constant>(Src1)) {
3757  Constant *CCmp = ConstantExpr::getCompare(CCVal, CSrc0, CSrc1);
3758  if (CCmp->isNullValue()) {
3759  return replaceInstUsesWith(
3760  *II, ConstantExpr::getSExt(CCmp, II->getType()));
3761  }
3762 
3763  // The result of V_ICMP/V_FCMP assembly instructions (which this
3764  // intrinsic exposes) is one bit per thread, masked with the EXEC
3765  // register (which contains the bitmask of live threads). So a
3766  // comparison that always returns true is the same as a read of the
3767  // EXEC register.
3769  II->getModule(), Intrinsic::read_register, II->getType());
3770  Metadata *MDArgs[] = {MDString::get(II->getContext(), "exec")};
3771  MDNode *MD = MDNode::get(II->getContext(), MDArgs);
3772  Value *Args[] = {MetadataAsValue::get(II->getContext(), MD)};
3773  CallInst *NewCall = Builder.CreateCall(NewF, Args);
3776  NewCall->takeName(II);
3777  return replaceInstUsesWith(*II, NewCall);
3778  }
3779 
3780  // Canonicalize constants to RHS.
3781  CmpInst::Predicate SwapPred
3782  = CmpInst::getSwappedPredicate(static_cast<CmpInst::Predicate>(CCVal));
3783  II->setArgOperand(0, Src1);
3784  II->setArgOperand(1, Src0);
3785  II->setArgOperand(2, ConstantInt::get(CC->getType(),
3786  static_cast<int>(SwapPred)));
3787  return II;
3788  }
3789 
3790  if (CCVal != CmpInst::ICMP_EQ && CCVal != CmpInst::ICMP_NE)
3791  break;
3792 
3793  // Canonicalize compare eq with true value to compare != 0
3794  // llvm.amdgcn.icmp(zext (i1 x), 1, eq)
3795  // -> llvm.amdgcn.icmp(zext (i1 x), 0, ne)
3796  // llvm.amdgcn.icmp(sext (i1 x), -1, eq)
3797  // -> llvm.amdgcn.icmp(sext (i1 x), 0, ne)
3798  Value *ExtSrc;
3799  if (CCVal == CmpInst::ICMP_EQ &&
3800  ((match(Src1, m_One()) && match(Src0, m_ZExt(m_Value(ExtSrc)))) ||
3801  (match(Src1, m_AllOnes()) && match(Src0, m_SExt(m_Value(ExtSrc))))) &&
3802  ExtSrc->getType()->isIntegerTy(1)) {
3803  II->setArgOperand(1, ConstantInt::getNullValue(Src1->getType()));
3804  II->setArgOperand(2, ConstantInt::get(CC->getType(), CmpInst::ICMP_NE));
3805  return II;
3806  }
3807 
3808  CmpInst::Predicate SrcPred;
3809  Value *SrcLHS;
3810  Value *SrcRHS;
3811 
3812  // Fold compare eq/ne with 0 from a compare result as the predicate to the
3813  // intrinsic. The typical use is a wave vote function in the library, which
3814  // will be fed from a user code condition compared with 0. Fold in the
3815  // redundant compare.
3816 
3817  // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, ne)
3818  // -> llvm.amdgcn.[if]cmp(a, b, pred)
3819  //
3820  // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, eq)
3821  // -> llvm.amdgcn.[if]cmp(a, b, inv pred)
3822  if (match(Src1, m_Zero()) &&
3823  match(Src0,
3824  m_ZExtOrSExt(m_Cmp(SrcPred, m_Value(SrcLHS), m_Value(SrcRHS))))) {
3825  if (CCVal == CmpInst::ICMP_EQ)
3826  SrcPred = CmpInst::getInversePredicate(SrcPred);
3827 
3828  Intrinsic::ID NewIID = CmpInst::isFPPredicate(SrcPred) ?
3829  Intrinsic::amdgcn_fcmp : Intrinsic::amdgcn_icmp;
3830 
3831  Type *Ty = SrcLHS->getType();
3832  if (auto *CmpType = dyn_cast<IntegerType>(Ty)) {
3833  // Promote to next legal integer type.
3834  unsigned Width = CmpType->getBitWidth();
3835  unsigned NewWidth = Width;
3836  if (Width <= 16)
3837  NewWidth = 16;
3838  else if (Width <= 32)
3839  NewWidth = 32;
3840  else if (Width <= 64)
3841  NewWidth = 64;
3842  else if (Width > 64)
3843  break; // Can't handle this.
3844 
3845  if (Width != NewWidth) {
3846  IntegerType *CmpTy = Builder.getIntNTy(NewWidth);
3847  if (CmpInst::isSigned(SrcPred)) {
3848  SrcLHS = Builder.CreateSExt(SrcLHS, CmpTy);
3849  SrcRHS = Builder.CreateSExt(SrcRHS, CmpTy);
3850  } else {
3851  SrcLHS = Builder.CreateZExt(SrcLHS, CmpTy);
3852  SrcRHS = Builder.CreateZExt(SrcRHS, CmpTy);
3853  }
3854  }
3855  } else if (!Ty->isFloatTy() && !Ty->isDoubleTy() && !Ty->isHalfTy())
3856  break;
3857 
3858  Value *NewF = Intrinsic::getDeclaration(II->getModule(), NewIID,
3859  SrcLHS->getType());
3860  Value *Args[] = { SrcLHS, SrcRHS,
3861  ConstantInt::get(CC->getType(), SrcPred) };
3862  CallInst *NewCall = Builder.CreateCall(NewF, Args);
3863  NewCall->takeName(II);
3864  return replaceInstUsesWith(*II, NewCall);
3865  }
3866 
3867  break;
3868  }
3869  case Intrinsic::amdgcn_wqm_vote: {
3870  // wqm_vote is identity when the argument is constant.
3871  if (!isa<Constant>(II->getArgOperand(0)))
3872  break;
3873 
3874  return replaceInstUsesWith(*II, II->getArgOperand(0));
3875  }
3876  case Intrinsic::amdgcn_kill: {
3877  const ConstantInt *C = dyn_cast<ConstantInt>(II->getArgOperand(0));
3878  if (!C || !C->getZExtValue())
3879  break;
3880 
3881  // amdgcn.kill(i1 1) is a no-op
3882  return eraseInstFromFunction(CI);
3883  }
3884  case Intrinsic::amdgcn_update_dpp: {
3885  Value *Old = II->getArgOperand(0);
3886 
3887  auto BC = dyn_cast<ConstantInt>(II->getArgOperand(5));
3888  auto RM = dyn_cast<ConstantInt>(II->getArgOperand(3));
3889  auto BM = dyn_cast<ConstantInt>(II->getArgOperand(4));
3890  if (!BC || !RM || !BM ||
3891  BC->isZeroValue() ||
3892  RM->getZExtValue() != 0xF ||
3893  BM->getZExtValue() != 0xF ||
3894  isa<UndefValue>(Old))
3895  break;
3896 
3897  // If bound_ctrl = 1, row mask = bank mask = 0xf we can omit old value.
3898  II->setOperand(0, UndefValue::get(Old->getType()));
3899  return II;
3900  }
3901  case Intrinsic::stackrestore: {
3902  // If the save is right next to the restore, remove the restore. This can
3903  // happen when variable allocas are DCE'd.
3904  if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
3905  if (SS->getIntrinsicID() == Intrinsic::stacksave) {
3906  // Skip over debug info.
3907  if (SS->getNextNonDebugInstruction() == II) {
3908  return eraseInstFromFunction(CI);
3909  }
3910  }
3911  }
3912 
3913  // Scan down this block to see if there is another stack restore in the
3914  // same block without an intervening call/alloca.
3915  BasicBlock::iterator BI(II);
3916  Instruction *TI = II->getParent()->getTerminator();
3917  bool CannotRemove = false;
3918  for (++BI; &*BI != TI; ++BI) {
3919  if (isa<AllocaInst>(BI)) {
3920  CannotRemove = true;
3921  break;
3922  }
3923  if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
3924  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
3925  // If there is a stackrestore below this one, remove this one.
3926  if (II->getIntrinsicID() == Intrinsic::stackrestore)
3927  return eraseInstFromFunction(CI);
3928 
3929  // Bail if we cross over an intrinsic with side effects, such as
3930  // llvm.stacksave, llvm.read_register, or llvm.setjmp.
3931  if (II->mayHaveSideEffects()) {
3932  CannotRemove = true;
3933  break;
3934  }
3935  } else {
3936  // If we found a non-intrinsic call, we can't remove the stack
3937  // restore.
3938  CannotRemove = true;
3939  break;
3940  }
3941  }
3942  }
3943 
3944  // If the stack restore is in a return, resume, or unwind block and if there
3945  // are no allocas or calls between the restore and the return, nuke the
3946  // restore.
3947  if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
3948  return eraseInstFromFunction(CI);
3949  break;
3950  }
3951  case Intrinsic::lifetime_start:
3952  // Asan needs to poison memory to detect invalid access which is possible
3953  // even for empty lifetime range.
3954  if (II->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3955  II->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3956  break;
3957 
3958  if (removeTriviallyEmptyRange(*II, Intrinsic::lifetime_start,
3959  Intrinsic::lifetime_end, *this))
3960  return nullptr;
3961  break;
3962  case Intrinsic::assume: {
3963  Value *IIOperand = II->getArgOperand(0);
3964  // Remove an assume if it is followed by an identical assume.
3965  // TODO: Do we need this? Unless there are conflicting assumptions, the
3966  // computeKnownBits(IIOperand) below here eliminates redundant assumes.
3968  if (match(Next, m_Intrinsic<Intrinsic::assume>(m_Specific(IIOperand))))
3969  return eraseInstFromFunction(CI);
3970 
3971  // Canonicalize assume(a && b) -> assume(a); assume(b);
3972  // Note: New assumption intrinsics created here are registered by
3973  // the InstCombineIRInserter object.
3974  Value *AssumeIntrinsic = II->getCalledValue(), *A, *B;
3975  if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
3976  Builder.CreateCall(AssumeIntrinsic, A, II->getName());
3977  Builder.CreateCall(AssumeIntrinsic, B, II->getName());
3978  return eraseInstFromFunction(*II);
3979  }
3980  // assume(!(a || b)) -> assume(!a); assume(!b);
3981  if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
3982  Builder.CreateCall(AssumeIntrinsic, Builder.CreateNot(A), II->getName());
3983  Builder.CreateCall(AssumeIntrinsic, Builder.CreateNot(B), II->getName());
3984  return eraseInstFromFunction(*II);
3985  }
3986 
3987  // assume( (load addr) != null ) -> add 'nonnull' metadata to load
3988  // (if assume is valid at the load)
3989  CmpInst::Predicate Pred;
3990  Instruction *LHS;
3991  if (match(IIOperand, m_ICmp(Pred, m_Instruction(LHS), m_Zero())) &&
3992  Pred == ICmpInst::ICMP_NE && LHS->getOpcode() == Instruction::Load &&
3993  LHS->getType()->isPointerTy() &&
3994  isValidAssumeForContext(II, LHS, &DT)) {
3995  MDNode *MD = MDNode::get(II->getContext(), None);
3997  return eraseInstFromFunction(*II);
3998 
3999  // TODO: apply nonnull return attributes to calls and invokes
4000  // TODO: apply range metadata for range check patterns?
4001  }
4002 
4003  // If there is a dominating assume with the same condition as this one,
4004  // then this one is redundant, and should be removed.
4005  KnownBits Known(1);
4006  computeKnownBits(IIOperand, Known, 0, II);
4007  if (Known.isAllOnes())
4008  return eraseInstFromFunction(*II);
4009 
4010  // Update the cache of affected values for this assumption (we might be
4011  // here because we just simplified the condition).
4012  AC.updateAffectedValues(II);
4013  break;
4014  }
4015  case Intrinsic::experimental_gc_relocate: {
4016  // Translate facts known about a pointer before relocating into
4017  // facts about the relocate value, while being careful to
4018  // preserve relocation semantics.
4019  Value *DerivedPtr = cast<GCRelocateInst>(II)->getDerivedPtr();
4020 
4021  // Remove the relocation if unused, note that this check is required
4022  // to prevent the cases below from looping forever.
4023  if (II->use_empty())
4024  return eraseInstFromFunction(*II);
4025 
4026  // Undef is undef, even after relocation.
4027  // TODO: provide a hook for this in GCStrategy. This is clearly legal for
4028  // most practical collectors, but there was discussion in the review thread
4029  // about whether it was legal for all possible collectors.
4030  if (isa<UndefValue>(DerivedPtr))
4031  // Use undef of gc_relocate's type to replace it.
4032  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
4033 
4034  if (auto *PT = dyn_cast<PointerType>(II->getType())) {
4035  // The relocation of null will be null for most any collector.
4036  // TODO: provide a hook for this in GCStrategy. There might be some
4037  // weird collector this property does not hold for.
4038  if (isa<ConstantPointerNull>(DerivedPtr))
4039  // Use null-pointer of gc_relocate's type to replace it.
4040  return replaceInstUsesWith(*II, ConstantPointerNull::get(PT));
4041 
4042  // isKnownNonNull -> nonnull attribute
4043  if (!II->hasRetAttr(Attribute::NonNull) &&
4044  isKnownNonZero(DerivedPtr, DL, 0, &AC, II, &DT)) {
4045  II->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull);
4046  return II;
4047  }
4048  }
4049 
4050  // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
4051  // Canonicalize on the type from the uses to the defs
4052 
4053  // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
4054  break;
4055  }
4056 
4057  case Intrinsic::experimental_guard: {
4058  // Is this guard followed by another guard? We scan forward over a small
4059  // fixed window of instructions to handle common cases with conditions
4060  // computed between guards.
4061  Instruction *NextInst = II->getNextNode();
4062  for (unsigned i = 0; i < GuardWideningWindow; i++) {
4063  // Note: Using context-free form to avoid compile time blow up
4064  if (!isSafeToSpeculativelyExecute(NextInst))
4065  break;
4066  NextInst = NextInst->getNextNode();
4067  }
4068  Value *NextCond = nullptr;
4069  if (match(NextInst,
4070  m_Intrinsic<Intrinsic::experimental_guard>(m_Value(NextCond)))) {
4071  Value *CurrCond = II->getArgOperand(0);
4072 
4073  // Remove a guard that it is immediately preceded by an identical guard.
4074  if (CurrCond == NextCond)
4075  return eraseInstFromFunction(*NextInst);
4076 
4077  // Otherwise canonicalize guard(a); guard(b) -> guard(a & b).
4078  Instruction* MoveI = II->getNextNode();
4079  while (MoveI != NextInst) {
4080  auto *Temp = MoveI;
4081  MoveI = MoveI->getNextNode();
4082  Temp->moveBefore(II);
4083  }
4084  II->setArgOperand(0, Builder.CreateAnd(CurrCond, NextCond));
4085  return eraseInstFromFunction(*NextInst);
4086  }
4087  break;
4088  }
4089  }
4090  return visitCallSite(II);
4091 }
4092 
4093 // Fence instruction simplification
4095  // Remove identical consecutive fences.
4097  if (auto *NFI = dyn_cast<FenceInst>(Next))
4098  if (FI.isIdenticalTo(NFI))
4099  return eraseInstFromFunction(FI);
4100  return nullptr;
4101 }
4102 
4103 // InvokeInst simplification
4105  return visitCallSite(&II);
4106 }
4107 
4108 /// If this cast does not affect the value passed through the varargs area, we
4109 /// can eliminate the use of the cast.
4111  const DataLayout &DL,
4112  const CastInst *const CI,
4113  const int ix) {
4114  if (!CI->isLosslessCast())
4115  return false;
4116 
4117  // If this is a GC intrinsic, avoid munging types. We need types for
4118  // statepoint reconstruction in SelectionDAG.
4119  // TODO: This is probably something which should be expanded to all
4120  // intrinsics since the entire point of intrinsics is that
4121  // they are understandable by the optimizer.
4122  if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS))
4123  return false;
4124 
4125  // The size of ByVal or InAlloca arguments is derived from the type, so we
4126  // can't change to a type with a different size. If the size were
4127  // passed explicitly we could avoid this check.
4128  if (!CS.isByValOrInAllocaArgument(ix))
4129  return true;
4130 
4131  Type* SrcTy =
4132  cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
4133  Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
4134  if (!SrcTy->isSized() || !DstTy->isSized())
4135  return false;
4136  if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
4137  return false;
4138  return true;
4139 }
4140 
4141 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
4142  if (!CI->getCalledFunction()) return nullptr;
4143 
4144  auto InstCombineRAUW = [this](Instruction *From, Value *With) {
4145  replaceInstUsesWith(*From, With);
4146  };
4147  auto InstCombineErase = [this](Instruction *I) {
4148  eraseInstFromFunction(*I);
4149  };
4150  LibCallSimplifier Simplifier(DL, &TLI, ORE, InstCombineRAUW,
4151  InstCombineErase);
4152  if (Value *With = Simplifier.optimizeCall(CI)) {
4153  ++NumSimplified;
4154  return CI->use_empty() ? CI : replaceInstUsesWith(*CI, With);
4155  }
4156 
4157  return nullptr;
4158 }
4159 
4161  // Strip off at most one level of pointer casts, looking for an alloca. This
4162  // is good enough in practice and simpler than handling any number of casts.
4163  Value *Underlying = TrampMem->stripPointerCasts();
4164  if (Underlying != TrampMem &&
4165  (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
4166  return nullptr;
4167  if (!isa<AllocaInst>(Underlying))
4168  return nullptr;
4169 
4170  IntrinsicInst *InitTrampoline = nullptr;
4171  for (User *U : TrampMem->users()) {
4173  if (!II)
4174  return nullptr;
4175  if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
4176  if (InitTrampoline)
4177  // More than one init_trampoline writes to this value. Give up.
4178  return nullptr;
4179  InitTrampoline = II;
4180  continue;
4181  }
4182  if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
4183  // Allow any number of calls to adjust.trampoline.
4184  continue;
4185  return nullptr;
4186  }
4187 
4188  // No call to init.trampoline found.
4189  if (!InitTrampoline)
4190  return nullptr;
4191 
4192  // Check that the alloca is being used in the expected way.
4193  if (InitTrampoline->getOperand(0) != TrampMem)
4194  return nullptr;
4195 
4196  return InitTrampoline;
4197 }
4198 
4200  Value *TrampMem) {
4201  // Visit all the previous instructions in the basic block, and try to find a
4202  // init.trampoline which has a direct path to the adjust.trampoline.
4203  for (BasicBlock::iterator I = AdjustTramp->getIterator(),
4204  E = AdjustTramp->getParent()->begin();
4205  I != E;) {
4206  Instruction *Inst = &*--I;
4207  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
4208  if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
4209  II->getOperand(0) == TrampMem)
4210  return II;
4211  if (Inst->mayWriteToMemory())
4212  return nullptr;
4213  }
4214  return nullptr;
4215 }
4216 
4217 // Given a call to llvm.adjust.trampoline, find and return the corresponding
4218 // call to llvm.init.trampoline if the call to the trampoline can be optimized
4219 // to a direct call to a function. Otherwise return NULL.
4221  Callee = Callee->stripPointerCasts();
4222  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
4223  if (!AdjustTramp ||
4224  AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
4225  return nullptr;
4226 
4227  Value *TrampMem = AdjustTramp->getOperand(0);
4228 
4230  return IT;
4231  if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem))
4232  return IT;
4233  return nullptr;
4234 }
4235 
4236 /// Improvements for call and invoke instructions.
4237 Instruction *InstCombiner::visitCallSite(CallSite CS) {
4238  if (isAllocLikeFn(CS.getInstruction(), &TLI))
4239  return visitAllocSite(*CS.getInstruction());
4240 
4241  bool Changed = false;
4242 
4243  // Mark any parameters that are known to be non-null with the nonnull
4244  // attribute. This is helpful for inlining calls to functions with null
4245  // checks on their arguments.
4246  SmallVector<unsigned, 4> ArgNos;
4247  unsigned ArgNo = 0;
4248 
4249  for (Value *V : CS.args()) {
4250  if (V->getType()->isPointerTy() &&
4251  !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
4252  isKnownNonZero(V, DL, 0, &AC, CS.getInstruction(), &DT))
4253  ArgNos.push_back(ArgNo);
4254  ArgNo++;
4255  }
4256 
4257  assert(ArgNo == CS.arg_size() && "sanity check");
4258 
4259  if (!ArgNos.empty()) {
4260  AttributeList AS = CS.getAttributes();
4261  LLVMContext &Ctx = CS.getInstruction()->getContext();
4262  AS = AS.addParamAttribute(Ctx, ArgNos,
4263  Attribute::get(Ctx, Attribute::NonNull));
4264  CS.setAttributes(AS);
4265  Changed = true;
4266  }
4267 
4268  // If the callee is a pointer to a function, attempt to move any casts to the
4269  // arguments of the call/invoke.
4270  Value *Callee = CS.getCalledValue();
4271  if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
4272  return nullptr;
4273 
4274  if (Function *CalleeF = dyn_cast<Function>(Callee)) {
4275  // Remove the convergent attr on calls when the callee is not convergent.
4276  if (CS.isConvergent() && !CalleeF->isConvergent() &&
4277  !CalleeF->isIntrinsic()) {
4278  LLVM_DEBUG(dbgs() << "Removing convergent attr from instr "
4279  << CS.getInstruction() << "\n");
4280  CS.setNotConvergent();
4281  return CS.getInstruction();
4282  }
4283 
4284  // If the call and callee calling conventions don't match, this call must
4285  // be unreachable, as the call is undefined.
4286  if (CalleeF->getCallingConv() != CS.getCallingConv() &&
4287  // Only do this for calls to a function with a body. A prototype may
4288  // not actually end up matching the implementation's calling conv for a
4289  // variety of reasons (e.g. it may be written in assembly).
4290  !CalleeF->isDeclaration()) {
4291  Instruction *OldCall = CS.getInstruction();
4292  new StoreInst(ConstantInt::getTrue(Callee->getContext()),
4294  OldCall);
4295  // If OldCall does not return void then replaceAllUsesWith undef.
4296  // This allows ValueHandlers and custom metadata to adjust itself.
4297  if (!OldCall->getType()->isVoidTy())
4298  replaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
4299  if (isa<CallInst>(OldCall))
4300  return eraseInstFromFunction(*OldCall);
4301 
4302  // We cannot remove an invoke, because it would change the CFG, just
4303  // change the callee to a null pointer.
4304  cast<InvokeInst>(OldCall)->setCalledFunction(
4305  Constant::getNullValue(CalleeF->getType()));
4306  return nullptr;
4307  }
4308  }
4309 
4310  if ((isa<ConstantPointerNull>(Callee) &&
4312  isa<UndefValue>(Callee)) {
4313  // If CS does not return void then replaceAllUsesWith undef.
4314  // This allows ValueHandlers and custom metadata to adjust itself.
4315  if (!CS.getInstruction()->getType()->isVoidTy())
4316  replaceInstUsesWith(*CS.getInstruction(),
4318 
4319  if (isa<InvokeInst>(CS.getInstruction())) {
4320  // Can't remove an invoke because we cannot change the CFG.
4321  return nullptr;
4322  }
4323 
4324  // This instruction is not reachable, just remove it. We insert a store to
4325  // undef so that we know that this code is not reachable, despite the fact
4326  // that we can't modify the CFG here.
4327  new StoreInst(ConstantInt::getTrue(Callee->getContext()),
4329  CS.getInstruction());
4330 
4331  return eraseInstFromFunction(*CS.getInstruction());
4332  }
4333 
4334  if (IntrinsicInst *II = findInitTrampoline(Callee))
4335  return transformCallThroughTrampoline(CS, II);
4336 
4337  PointerType *PTy = cast<PointerType>(Callee->getType());
4338  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4339  if (FTy->isVarArg()) {
4340  int ix = FTy->getNumParams();
4341  // See if we can optimize any arguments passed through the varargs area of
4342  // the call.
4343  for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
4344  E = CS.arg_end(); I != E; ++I, ++ix) {
4345  CastInst *CI = dyn_cast<CastInst>(*I);
4346  if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) {
4347  *I = CI->getOperand(0);
4348  Changed = true;
4349  }
4350  }
4351  }
4352 
4353  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
4354  // Inline asm calls cannot throw - mark them 'nounwind'.
4355  CS.setDoesNotThrow();
4356  Changed = true;
4357  }
4358 
4359  // Try to optimize the call if possible, we require DataLayout for most of
4360  // this. None of these calls are seen as possibly dead so go ahead and
4361  // delete the instruction now.
4362  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
4363  Instruction *I = tryOptimizeCall(CI);
4364  // If we changed something return the result, etc. Otherwise let
4365  // the fallthrough check.
4366  if (I) return eraseInstFromFunction(*I);
4367  }
4368 
4369  return Changed ? CS.getInstruction() : nullptr;
4370 }
4371 
4372 /// If the callee is a constexpr cast of a function, attempt to move the cast to
4373 /// the arguments of the call/invoke.
4374 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
4376  if (!Callee)
4377  return false;
4378 
4379  // If this is a call to a thunk function, don't remove the cast. Thunks are
4380  // used to transparently forward all incoming parameters and outgoing return
4381  // values, so it's important to leave the cast in place.
4382  if (Callee->hasFnAttribute("thunk"))
4383  return false;
4384 
4385  // If this is a musttail call, the callee's prototype must match the caller's
4386  // prototype with the exception of pointee types. The code below doesn't
4387  // implement that, so we can't do this transform.
4388  // TODO: Do the transform if it only requires adding pointer casts.
4389  if (CS.isMustTailCall())
4390  return false;
4391 
4392  Instruction *Caller = CS.getInstruction();
4393  const AttributeList &CallerPAL = CS.getAttributes();
4394 
4395  // Okay, this is a cast from a function to a different type. Unless doing so
4396  // would cause a type conversion of one of our arguments, change this call to
4397  // be a direct call with arguments casted to the appropriate types.
4398  FunctionType *FT = Callee->getFunctionType();
4399  Type *OldRetTy = Caller->getType();
4400  Type *NewRetTy = FT->getReturnType();
4401 
4402  // Check to see if we are changing the return type...
4403  if (OldRetTy != NewRetTy) {
4404 
4405  if (NewRetTy->isStructTy())
4406  return false; // TODO: Handle multiple return values.
4407 
4408  if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
4409  if (Callee->isDeclaration())
4410  return false; // Cannot transform this return value.
4411 
4412  if (!Caller->use_empty() &&
4413  // void -> non-void is handled specially
4414  !NewRetTy->isVoidTy())
4415  return false; // Cannot transform this return value.
4416  }
4417 
4418  if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
4419  AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex);
4420  if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
4421  return false; // Attribute not compatible with transformed value.
4422  }
4423 
4424  // If the callsite is an invoke instruction, and the return value is used by
4425  // a PHI node in a successor, we cannot change the return type of the call
4426  // because there is no place to put the cast instruction (without breaking
4427  // the critical edge). Bail out in this case.
4428  if (!Caller->use_empty())
4429  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
4430  for (User *U : II->users())
4431  if (PHINode *PN = dyn_cast<PHINode>(U))
4432  if (PN->getParent() == II->getNormalDest() ||
4433  PN->getParent() == II->getUnwindDest())
4434  return false;
4435  }
4436 
4437  unsigned NumActualArgs = CS.arg_size();
4438  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
4439 
4440  // Prevent us turning:
4441  // declare void @takes_i32_inalloca(i32* inalloca)
4442  // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
4443  //
4444  // into:
4445  // call void @takes_i32_inalloca(i32* null)
4446  //
4447  // Similarly, avoid folding away bitcasts of byval calls.
4448  if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
4449  Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
4450  return false;
4451 
4453  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
4454  Type *ParamTy = FT->getParamType(i);
4455  Type *ActTy = (*AI)->getType();
4456 
4457  if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
4458  return false; // Cannot transform this parameter value.
4459 
4460  if (AttrBuilder(CallerPAL.getParamAttributes(i))
4461  .overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
4462  return false; // Attribute not compatible with transformed value.
4463 
4464  if (CS.isInAllocaArgument(i))
4465  return false; // Cannot transform to and from inalloca.
4466 
4467  // If the parameter is passed as a byval argument, then we have to have a
4468  // sized type and the sized type has to have the same size as the old type.
4469  if (ParamTy != ActTy && CallerPAL.hasParamAttribute(i, Attribute::ByVal)) {
4470  PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
4471  if (!ParamPTy || !ParamPTy->getElementType()->isSized())
4472  return false;
4473 
4474  Type *CurElTy = ActTy->getPointerElementType();
4475  if (DL.getTypeAllocSize(CurElTy) !=
4476  DL.getTypeAllocSize(ParamPTy->getElementType()))
4477  return false;
4478  }
4479  }
4480 
4481  if (Callee->isDeclaration()) {
4482  // Do not delete arguments unless we have a function body.
4483  if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
4484  return false;
4485 
4486  // If the callee is just a declaration, don't change the varargsness of the
4487  // call. We don't want to introduce a varargs call where one doesn't
4488  // already exist.
4489  PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
4490  if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
4491  return false;
4492 
4493  // If both the callee and the cast type are varargs, we still have to make
4494  // sure the number of fixed parameters are the same or we have the same
4495  // ABI issues as if we introduce a varargs call.
4496  if (FT->isVarArg() &&
4497  cast<FunctionType>(APTy->getElementType())->isVarArg() &&
4498  FT->getNumParams() !=
4499  cast<FunctionType>(APTy->getElementType())->getNumParams())
4500  return false;
4501  }
4502 
4503  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
4504  !CallerPAL.isEmpty()) {
4505  // In this case we have more arguments than the new function type, but we
4506  // won't be dropping them. Check that these extra arguments have attributes
4507  // that are compatible with being a vararg call argument.
4508  unsigned SRetIdx;
4509  if (CallerPAL.hasAttrSomewhere(Attribute::StructRet, &SRetIdx) &&
4510  SRetIdx > FT->getNumParams())
4511  return false;
4512  }
4513 
4514  // Okay, we decided that this is a safe thing to do: go ahead and start
4515  // inserting cast instructions as necessary.
4518  Args.reserve(NumActualArgs);
4519  ArgAttrs.reserve(NumActualArgs);
4520 
4521  // Get any return attributes.
4522  AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex);
4523 
4524  // If the return value is not being used, the type may not be compatible
4525  // with the existing attributes. Wipe out any problematic attributes.
4526  RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
4527 
4528  AI = CS.arg_begin();
4529  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
4530  Type *ParamTy = FT->getParamType(i);
4531 
4532  Value *NewArg = *AI;
4533  if ((*AI)->getType() != ParamTy)
4534  NewArg = Builder.CreateBitOrPointerCast(*AI, ParamTy);
4535  Args.push_back(NewArg);
4536 
4537  // Add any parameter attributes.
4538  ArgAttrs.push_back(CallerPAL.getParamAttributes(i));
4539  }
4540 
4541  // If the function takes more arguments than the call was taking, add them
4542  // now.
4543  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) {
4545  ArgAttrs.push_back(AttributeSet());
4546  }
4547 
4548  // If we are removing arguments to the function, emit an obnoxious warning.
4549  if (FT->getNumParams() < NumActualArgs) {
4550  // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
4551  if (FT->isVarArg()) {
4552  // Add all of the arguments in their promoted form to the arg list.
4553  for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
4554  Type *PTy = getPromotedType((*AI)->getType());
4555  Value *NewArg = *AI;
4556  if (PTy != (*AI)->getType()) {
4557  // Must promote to pass through va_arg area!
4558  Instruction::CastOps opcode =
4559  CastInst::getCastOpcode(*AI, false, PTy, false);
4560  NewArg = Builder.CreateCast(opcode, *AI, PTy);
4561  }
4562  Args.push_back(NewArg);
4563 
4564  // Add any parameter attributes.
4565  ArgAttrs.push_back(CallerPAL.getParamAttributes(i));
4566  }
4567  }
4568  }
4569 
4570  AttributeSet FnAttrs = CallerPAL.getFnAttributes();
4571 
4572  if (NewRetTy->isVoidTy())
4573  Caller->setName(""); // Void type should not have a name.
4574 
4575  assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) &&
4576  "missing argument attributes");
4577  LLVMContext &Ctx = Callee->getContext();
4578  AttributeList NewCallerPAL = AttributeList::get(
4579  Ctx, FnAttrs, AttributeSet::get(Ctx, RAttrs), ArgAttrs);
4580 
4582  CS.getOperandBundlesAsDefs(OpBundles);
4583 
4584  CallSite NewCS;
4585  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4586  NewCS = Builder.CreateInvoke(Callee, II->getNormalDest(),
4587  II->getUnwindDest(), Args, OpBundles);
4588  } else {
4589  NewCS = Builder.CreateCall(Callee, Args, OpBundles);
4590  cast<CallInst>(NewCS.getInstruction())
4591  ->setTailCallKind(cast<CallInst>(Caller)->getTailCallKind());
4592  }
4593  NewCS->takeName(Caller);
4594  NewCS.setCallingConv(CS.getCallingConv());
4595  NewCS.setAttributes(NewCallerPAL);
4596 
4597  // Preserve the weight metadata for the new call instruction. The metadata
4598  // is used by SamplePGO to check callsite's hotness.
4599  uint64_t W;
4600  if (Caller->extractProfTotalWeight(W))
4601  NewCS->setProfWeight(W);
4602 
4603  // Insert a cast of the return type as necessary.
4604  Instruction *NC = NewCS.getInstruction();
4605  Value *NV = NC;
4606  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
4607  if (!NV->getType()->isVoidTy()) {
4608  NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
4609  NC->setDebugLoc(Caller->getDebugLoc());
4610 
4611  // If this is an invoke instruction, we should insert it after the first
4612  // non-phi, instruction in the normal successor block.
4613  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4614  BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
4615  InsertNewInstBefore(NC, *I);
4616  } else {
4617  // Otherwise, it's a call, just insert cast right after the call.
4618  InsertNewInstBefore(NC, *Caller);
4619  }
4620  Worklist.AddUsersToWorkList(*Caller);
4621  } else {
4622  NV = UndefValue::get(Caller->getType());
4623  }
4624  }
4625 
4626  if (!Caller->use_empty())
4627  replaceInstUsesWith(*Caller, NV);
4628  else if (Caller->hasValueHandle()) {
4629  if (OldRetTy == NV->getType())
4630  ValueHandleBase::ValueIsRAUWd(Caller, NV);
4631  else
4632  // We cannot call ValueIsRAUWd with a different type, and the
4633  // actual tracked value will disappear.
4635  }
4636 
4637  eraseInstFromFunction(*Caller);
4638  return true;
4639 }
4640 
4641 /// Turn a call to a function created by init_trampoline / adjust_trampoline
4642 /// intrinsic pair into a direct call to the underlying function.
4643 Instruction *
4644 InstCombiner::transformCallThroughTrampoline(CallSite CS,
4645  IntrinsicInst *Tramp) {
4646  Value *Callee = CS.getCalledValue();
4647  PointerType *PTy = cast<PointerType>(Callee->getType());
4648  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4650 
4651  // If the call already has the 'nest' attribute somewhere then give up -
4652  // otherwise 'nest' would occur twice after splicing in the chain.
4653  if (Attrs.hasAttrSomewhere(Attribute::Nest))
4654  return nullptr;
4655 
4656  assert(Tramp &&
4657  "transformCallThroughTrampoline called with incorrect CallSite.");
4658 
4659  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
4660  FunctionType *NestFTy = cast<FunctionType>(NestF->getValueType());
4661 
4662  AttributeList NestAttrs = NestF->getAttributes();
4663  if (!NestAttrs.isEmpty()) {
4664  unsigned NestArgNo = 0;
4665  Type *NestTy = nullptr;
4666  AttributeSet NestAttr;
4667 
4668  // Look for a parameter marked with the 'nest' attribute.
4669  for (FunctionType::param_iterator I = NestFTy->param_begin(),
4670  E = NestFTy->param_end();
4671  I != E; ++NestArgNo, ++I) {
4672  AttributeSet AS = NestAttrs.getParamAttributes(NestArgNo);
4673  if (AS.hasAttribute(Attribute::Nest)) {
4674  // Record the parameter type and any other attributes.
4675  NestTy = *I;
4676  NestAttr = AS;
4677  break;
4678  }
4679  }
4680 
4681  if (NestTy) {
4682  Instruction *Caller = CS.getInstruction();
4683  std::vector<Value*> NewArgs;
4684  std::vector<AttributeSet> NewArgAttrs;
4685  NewArgs.reserve(CS.arg_size() + 1);
4686  NewArgAttrs.reserve(CS.arg_size());
4687 
4688  // Insert the nest argument into the call argument list, which may
4689  // mean appending it. Likewise for attributes.
4690 
4691  {
4692  unsigned ArgNo = 0;
4693  CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
4694  do {
4695  if (ArgNo == NestArgNo) {
4696  // Add the chain argument and attributes.
4697  Value *NestVal = Tramp->getArgOperand(2);
4698  if (NestVal->getType() != NestTy)
4699  NestVal = Builder.CreateBitCast(NestVal, NestTy, "nest");
4700  NewArgs.push_back(NestVal);
4701  NewArgAttrs.push_back(NestAttr);
4702  }
4703 
4704  if (I == E)
4705  break;
4706 
4707  // Add the original argument and attributes.
4708  NewArgs.push_back(*I);
4709  NewArgAttrs.push_back(Attrs.getParamAttributes(ArgNo));
4710 
4711  ++ArgNo;
4712  ++I;
4713  } while (true);
4714  }
4715 
4716  // The trampoline may have been bitcast to a bogus type (FTy).
4717  // Handle this by synthesizing a new function type, equal to FTy
4718  // with the chain parameter inserted.
4719 
4720  std::vector<Type*> NewTypes;
4721  NewTypes.reserve(FTy->getNumParams()+1);
4722 
4723  // Insert the chain's type into the list of parameter types, which may
4724  // mean appending it.
4725  {
4726  unsigned ArgNo = 0;
4727  FunctionType::param_iterator I = FTy->param_begin(),
4728  E = FTy->param_end();
4729 
4730  do {
4731  if (ArgNo == NestArgNo)
4732  // Add the chain's type.
4733  NewTypes.push_back(NestTy);
4734 
4735  if (I == E)
4736  break;
4737 
4738  // Add the original type.
4739  NewTypes.push_back(*I);
4740 
4741  ++ArgNo;
4742  ++I;
4743  } while (true);
4744  }
4745 
4746  // Replace the trampoline call with a direct call. Let the generic
4747  // code sort out any function type mismatches.
4748  FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
4749  FTy->isVarArg());
4750  Constant *NewCallee =
4751  NestF->getType() == PointerType::getUnqual(NewFTy) ?
4752  NestF : ConstantExpr::getBitCast(NestF,
4753  PointerType::getUnqual(NewFTy));
4754  AttributeList NewPAL =
4755  AttributeList::get(FTy->getContext(), Attrs.getFnAttributes(),
4756  Attrs.getRetAttributes(), NewArgAttrs);
4757 
4759  CS.getOperandBundlesAsDefs(OpBundles);
4760 
4761  Instruction *NewCaller;
4762  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4763  NewCaller = InvokeInst::Create(NewCallee,
4764  II->getNormalDest(), II->getUnwindDest(),
4765  NewArgs, OpBundles);
4766  cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
4767  cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
4768  } else {
4769  NewCaller = CallInst::Create(NewCallee, NewArgs, OpBundles);
4770  cast<CallInst>(NewCaller)->setTailCallKind(
4771  cast<CallInst>(Caller)->getTailCallKind());
4772  cast<CallInst>(NewCaller)->setCallingConv(
4773  cast<CallInst>(Caller)->getCallingConv());
4774  cast<CallInst>(NewCaller)->setAttributes(NewPAL);
4775  }
4776  NewCaller->setDebugLoc(Caller->getDebugLoc());
4777 
4778  return NewCaller;
4779  }
4780  }
4781 
4782  // Replace the trampoline call with a direct call. Since there is no 'nest'
4783  // parameter, there is no need to adjust the argument list. Let the generic
4784  // code sort out any function type mismatches.
4785  Constant *NewCallee =
4786  NestF->getType() == PTy ? NestF :
4787  ConstantExpr::getBitCast(NestF, PTy);
4788  CS.setCalledFunction(NewCallee);
4789  return CS.getInstruction();
4790 }
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
bool isFPPredicate() const
Definition: InstrTypes.h:734
const NoneType None
Definition: None.h:24
A vector constant whose element type is a simple 1/2/4/8-byte integer or float/double, and whose elements are just simple data values (i.e.
Definition: Constants.h:762
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:750
uint64_t CallInst * C
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, OptimizationRemarkEmitter *ORE=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
User::op_iterator arg_iterator
The type of iterator to use when looping over actual arguments at this call site. ...
Definition: CallSite.h:213
LibCallSimplifier - This class implements a collection of optimizations that replace well formed call...
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
unsigned Log2_32_Ceil(uint32_t Value)
Return the ceil log base 2 of the specified value, 32 if the value is zero.
Definition: MathExtras.h:552
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, const Instruction *CxtI) const
void copyFastMathFlags(FastMathFlags FMF)
Convenience function for transferring all fast-math flag values to this instruction, which must be an operator which supports these flags.
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
static void ValueIsDeleted(Value *V)
Definition: Value.cpp:832
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1858
class_match< UndefValue > m_Undef()
Match an arbitrary undef constant.
Definition: PatternMatch.h:88
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
bool isZero() const
Definition: APFloat.h:1143
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:173
class_match< CmpInst > m_Cmp()
Matches any compare instruction and ignore it.
Definition: PatternMatch.h:80
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1563
unsigned getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, const DataLayout &DL, const Instruction *CxtI=nullptr, AssumptionCache *AC=nullptr, const DominatorTree *DT=nullptr)
Try to ensure that the alignment of V is at least PrefAlign bytes.
Definition: Local.cpp:1184
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
static Value * simplifyX86immShift(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
static APInt getAllOnesValue(unsigned numBits)
Get the all-ones value.
Definition: APInt.h:562
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:377
DiagnosticInfoOptimizationBase::Argument NV
unsigned arg_size() const
Definition: CallSite.h:219
CallingConv::ID getCallingConv() const
Get the calling convention of the call.
Definition: CallSite.h:312
Atomic ordering constants.
Value * CreateAddrSpaceCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1646
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1578
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:289
This class represents lattice values for constants.
Definition: AllocatorList.h:24
Type * getParamType(unsigned i) const
Parameter type accessors.
Definition: DerivedTypes.h:135
Constant * getElementAsConstant(unsigned i) const
Return a Constant for a specified index&#39;s element.
Definition: Constants.cpp:2762
unsigned countMinPopulation() const
Returns the number of bits known to be one.
Definition: KnownBits.h:186
bool isInAllocaArgument(unsigned ArgNo) const
Determine whether this argument is passed in an alloca.
Definition: CallSite.h:603
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:65
Instruction * visitCallInst(CallInst &CI)
CallInst simplification.
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:265
#define LLVM_FALLTHROUGH
Definition: Compiler.h:86
An instruction for ordering other memory operations.
Definition: Instructions.h:444
static MDString * get(LLVMContext &Context, StringRef Str)
Definition: Metadata.cpp:454
Instruction * visitVACopyInst(VACopyInst &I)
static Instruction * simplifyInvariantGroupIntrinsic(IntrinsicInst &II, InstCombiner &IC)
This function transforms launder.invariant.group and strip.invariant.group like: launder(launder(x)) ...
static ConstantAggregateZero * get(Type *Ty)
Definition: Constants.cpp:1332
APInt uadd_sat(const APInt &RHS) const
Definition: APInt.cpp:1960
static bool simplifyX86MaskedStore(IntrinsicInst &II, InstCombiner &IC)
This class represents a function call, abstracting a target machine&#39;s calling convention.
m_Intrinsic_Ty< Opnd0 >::Ty m_FAbs(const Opnd0 &Op0)
This file contains the declarations for metadata subclasses.
static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet)
Get a value with low bits set.
Definition: APInt.h:648
void setOrdering(AtomicOrdering Ordering)
Sets the ordering constraint of this load instruction.
Definition: Instructions.h:243
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:91
static PointerType * get(Type *ElementType, unsigned AddressSpace)
This constructs a pointer to an object of the specified type in a numbered address space...
Definition: Type.cpp:630
iterator_range< IterTy > args() const
Definition: CallSite.h:215
static uint64_t round(uint64_t Acc, uint64_t Input)
Definition: xxhash.cpp:57
m_Intrinsic_Ty< Opnd0 >::Ty m_BSwap(const Opnd0 &Op0)
bool hasValueHandle() const
Return true if there is a value handle associated with this value.
Definition: Value.h:486
unsigned less or equal
Definition: InstrTypes.h:668
bool mayWriteToMemory() const
Return true if this instruction may modify memory.
unsigned less than
Definition: InstrTypes.h:667
bool isSubsetOf(const APInt &RHS) const
This operation checks that all bits set in this APInt are also set in RHS.
Definition: APInt.h:1329
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *InsertBefore=nullptr, Instruction *MDFrom=nullptr)
static Instruction * foldCttzCtlz(IntrinsicInst &II, InstCombiner &IC)
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:705
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
static Value * simplifyX86AddsSubs(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1180
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:811
bool isValidAssumeForContext(const Instruction *I, const Instruction *CxtI, const DominatorTree *DT=nullptr)
Return true if it is valid to use the assumptions provided by an assume intrinsic, I, at the point in the control-flow identified by the context instruction, CxtI.
STATISTIC(NumFunctions, "Total number of functions")
void setArgOperand(unsigned i, Value *v)
Definition: InstrTypes.h:1084
Metadata node.
Definition: Metadata.h:864
F(f)
static CallInst * Create(Value *Func, ArrayRef< Value *> Args, ArrayRef< OperandBundleDef > Bundles=None, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
const fltSemantics & getSemantics() const
Definition: APFloat.h:1155
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
BinaryOp_match< LHS, RHS, Instruction::FSub > m_FSub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:661
An instruction for reading from memory.
Definition: Instructions.h:168
static IntegerType * getInt64Ty(LLVMContext &C)
Definition: Type.cpp:177
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:876
static Constant * getCompare(unsigned short pred, Constant *C1, Constant *C2, bool OnlyIfReduced=false)
Return an ICmp or FCmp comparison operator constant expression.
Definition: Constants.cpp:1956
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.cpp:138
unsigned countMaxTrailingZeros() const
Returns the maximum number of trailing zero bits possible.
Definition: KnownBits.h:166
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:230
static OverflowCheckFlavor IntrinsicIDToOverflowCheckFlavor(unsigned ID)
Returns the OverflowCheckFlavor corresponding to a overflow_with_op intrinsic.
LLVM_READONLY APFloat maximum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2018 maximum semantics.
Definition: APFloat.h:1262
void reserve(size_type N)
Definition: SmallVector.h:376
void addAttribute(unsigned i, Attribute::AttrKind Kind)
adds the attribute to the list of attributes.
Definition: InstrTypes.h:1170
Value * getLength() const
void copyIRFlags(const Value *V, bool IncludeWrapFlags=true)
Convenience method to copy supported exact, fast-math, and (optionally) wrapping flags from V to this...
static Instruction * simplifyMaskedStore(IntrinsicInst &II, InstCombiner &IC)
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:365
Instruction * visitVAStartInst(VAStartInst &I)
static APInt getSignedMaxValue(unsigned numBits)
Gets maximum signed value of APInt for a specific bit width.
Definition: APInt.h:535
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1509
Value * CreateLaunderInvariantGroup(Value *Ptr)
Create a launder.invariant.group intrinsic call.
Definition: IRBuilder.h:2034
bool isGCRelocate(ImmutableCallSite CS)
Definition: Statepoint.cpp:43
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
const CallInst * isFreeCall(const Value *I, const TargetLibraryInfo *TLI)
isFreeCall - Returns non-null if the value is a call to the builtin free()
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:265
unsigned countMinTrailingZeros() const
Returns the minimum number of trailing zero bits.
Definition: KnownBits.h:136
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...
Value * getDest() const
This is just like getRawDest, but it strips off any cast instructions (including addrspacecast) that ...
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:269
bool isIdenticalTo(const Instruction *I) const
Return true if the specified instruction is exactly identical to the current one. ...
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1079
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
opStatus divide(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:968
static Instruction * SimplifyNVVMIntrinsic(IntrinsicInst *II, InstCombiner &IC)
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
Instruction * visitInvokeInst(InvokeInst &II)
static Constant * getIntegerCast(Constant *C, Type *Ty, bool isSigned)
Create a ZExt, Bitcast or Trunc for integer -> integer casts.
Definition: Constants.cpp:1613
bool isSigned() const
Definition: InstrTypes.h:812
APInt getLoBits(unsigned numBits) const
Compute an APInt containing numBits lowbits from this APInt.
Definition: APInt.cpp:516
static APFloat fmed3AMDGCN(const APFloat &Src0, const APFloat &Src1, const APFloat &Src2)
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
Definition: PatternMatch.h:762
Type * getPointerElementType() const
Definition: Type.h:376
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:741
OverflowCheckFlavor
Specific patterns of overflow check idioms that we match.
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
static Value * simplifyX86movmsk(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:349
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:369
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:451
APInt shl(unsigned shiftAmt) const
Left-shift function.
Definition: APInt.h:993
static Value * simplifyNeonTbl1(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Convert a table lookup to shufflevector if the mask is constant.
IterTy arg_end() const
Definition: CallSite.h:575
OverflowResult computeOverflowForUnsignedAdd(const Value *LHS, const Value *RHS, const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, bool UseInstrInfo=true)
Instruction * eraseInstFromFunction(Instruction &I)
Combiner aware instruction erasure.
CastClass_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:743
The core instruction combiner logic.
static bool isSafeToEliminateVarargsCast(const CallSite CS, 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...
void setCalledFunction(Value *Fn)
Sets the function called, including updating the function type.
Definition: InstrTypes.h:1119
This file contains the simple types necessary to represent the attributes associated with functions a...
LLVM_READONLY APFloat minimum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2018 minimum semantics.
Definition: APFloat.h:1249
InstrTy * getInstruction() const
Definition: CallSite.h:92
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1651
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:286
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
void lshrInPlace(unsigned ShiftAmt)
Logical right-shift this APInt by ShiftAmt in place.
Definition: APInt.h:978
ELFYAML::ELF_STO Other
Definition: ELFYAML.cpp:783
This file implements a class to represent arbitrary precision integral constant values and operations...
All zero aggregate value.
Definition: Constants.h:341
static Value * simplifyX86vpermv(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
Metadata * LowAndHigh[]
ValTy * getCalledValue() const
Return the pointer to function that is being called.
Definition: CallSite.h:100
static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E, unsigned NumOperands)
DominatorTree & getDominatorTree() const
unsigned countMaxPopulation() const
Returns the maximum number of bits that could be one.
Definition: KnownBits.h:191
Key
PAL metadata keys.
bool doesNotThrow() const
Determine if the call cannot unwind.
Definition: InstrTypes.h:1410
bool isNullValue() const
Return true if this is the value that would be returned by getNullValue.
Definition: Constants.cpp:85
Class to represent function types.
Definition: DerivedTypes.h:103
static Value * peekThroughBitcast(Value *V, bool OneUseOnly=false)
Return the source operand of a potentially bitcasted value while optionally checking if it has one us...
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1641
bool isInfinity() const
Definition: APFloat.h:1144
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
Value * CreateSExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a SExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1593
cstfp_pred_ty< is_nan > m_NaN()
Match an arbitrary NaN constant.
Definition: PatternMatch.h:428
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:181
This represents the llvm.va_start intrinsic.
CastClass_match< OpTy, Instruction::FPExt > m_FPExt(const OpTy &Op)
Matches FPExt.
opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, bool *losesInfo)
Definition: APFloat.cpp:4444
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
AttributeSet getParamAttributes(unsigned ArgNo) const
The attributes for the argument or parameter at the given index are returned.
bool isVarArg() const
Definition: DerivedTypes.h:123
This class represents a no-op cast from one type to another.
bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const
Return true if the call or the callee has the given attribute.
Definition: CallSite.h:377
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:221
Value * CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name="")
Return a vector value that contains.
Definition: IRBuilder.h:2088
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
cstfp_pred_ty< is_pos_zero_fp > m_PosZeroFP()
Match a floating-point positive zero.
Definition: PatternMatch.h:446
iterator_range< User::op_iterator > arg_operands()
Iteration adapter for range-for loops.
Definition: InstrTypes.h:1070
AttrBuilder & remove(const AttrBuilder &B)
Remove the attributes from the builder.
static Value * simplifyX86pack(IntrinsicInst &II, bool IsSigned)
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:224
cmpResult
IEEE-754R 5.11: Floating Point Comparison Relations.
Definition: APFloat.h:166
An instruction for storing to memory.
Definition: Instructions.h:310
bool extractProfTotalWeight(uint64_t &TotalVal) const
Retrieve total raw weight values of a branch.
Definition: Metadata.cpp:1340
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
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:734
static void ValueIsRAUWd(Value *Old, Value *New)
Definition: Value.cpp:885
static Value * simplifyX86vpcom(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder, bool IsSigned)
Decode XOP integer vector comparison intrinsics.
constexpr char Attrs[]
Key for Kernel::Metadata::mAttrs.
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:292
static ConstantAsMetadata * get(Constant *C)
Definition: Metadata.h:410
amdgpu Simplify well known AMD library false Value * Callee
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1021
This class represents a truncation of integer types.
Type * getElementType() const
Return the element type of the array/vector.
Definition: Constants.cpp:2422
Value * getOperand(unsigned i) const
Definition: User.h:170
Class to represent pointers.
Definition: DerivedTypes.h:467
bool hasAttribute(Attribute::AttrKind Kind) const
Return true if the attribute exists in this set.
Definition: Attributes.cpp:578
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:335
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:304
const DataLayout & getDataLayout() const
static MetadataAsValue * get(LLVMContext &Context, Metadata *MD)
Definition: Metadata.cpp:106
static Constant * getBitCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1773
bool isVoidTy() const
Return true if this is &#39;void&#39;.
Definition: Type.h:141
bool isFloatTy() const
Return true if this is &#39;float&#39;, a 32-bit IEEE fp type.
Definition: Type.h:147
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...
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:63
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1166
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:364
void setAttributes(AttributeList PAL)
Set the parameter attributes of the call.
Definition: CallSite.h:333
Instruction * visitFenceInst(FenceInst &FI)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:419
static Instruction * simplifyMaskedScatter(IntrinsicInst &II, InstCombiner &IC)
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:149
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static AttributeSet get(LLVMContext &C, const AttrBuilder &B)
Definition: Attributes.cpp:513
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:177
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:308
bool isNegative() const
Definition: APFloat.h:1147
static ConstantPointerNull * get(PointerType *T)
Static factory methods - Return objects of the specified value.
Definition: Constants.cpp:1401
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1104
APInt urem(const APInt &RHS) const
Unsigned remainder operation.
Definition: APInt.cpp:1613
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:429
ConstantInt * lowerObjectSizeCall(IntrinsicInst *ObjectSize, const DataLayout &DL, const TargetLibraryInfo *TLI, bool MustSucceed)
Try to turn a call to @llvm.objectsize into an integer value of the given Type.
LLVM_NODISCARD AttributeList addParamAttribute(LLVMContext &C, unsigned ArgNo, Attribute::AttrKind Kind) const
Add an argument attribute to the list.
Definition: Attributes.h:403
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
Definition: PatternMatch.h:756
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:69
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:287
bool isNaN() const
Definition: APFloat.h:1145
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:1913
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:224
static ManagedStatic< OptionRegistry > OR
Definition: Options.cpp:31
unsigned getNumParams() const
Return the number of fixed parameters this function type requires.
Definition: DerivedTypes.h:139
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
APInt ssub_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1888
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:310
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:503
const Instruction * getNextNonDebugInstruction() const
Return a pointer to the next non-debug instruction in the same basic block as &#39;this&#39;, or nullptr if no such instruction exists.
This file declares a class to represent arbitrary precision floating point values and provide a varie...
bool isFast() const
Determine whether all fast-math-flags are set.
std::underlying_type< E >::type Underlying(E Val)
Check that Val is in range for E, and return Val cast to E&#39;s underlying type.
Definition: BitmaskEnum.h:91
static IntrinsicInst * findInitTrampolineFromBB(IntrinsicInst *AdjustTramp, Value *TrampMem)
bool isHalfTy() const
Return true if this is &#39;half&#39;, a 16-bit IEEE fp type.
Definition: Type.h:144
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:642
match_combine_or< CastClass_match< OpTy, Instruction::ZExt >, CastClass_match< OpTy, Instruction::SExt > > m_ZExtOrSExt(const OpTy &Op)
bool isAllOnes() const
Returns true if value is all one bits.
Definition: KnownBits.h:78
void setCallingConv(CallingConv::ID CC)
Set the calling convention of the call.
Definition: CallSite.h:316
bool isGCResult(ImmutableCallSite CS)
Definition: Statepoint.cpp:53
This class represents any memset intrinsic.
static FunctionType * get(Type *Result, ArrayRef< Type *> Params, bool isVarArg)
This static method is the primary way of constructing a FunctionType.
Definition: Type.cpp:297
self_iterator getIterator()
Definition: ilist_node.h:82
Class to represent integer types.
Definition: DerivedTypes.h:40
IntegerType * getIntNTy(unsigned N)
Fetch the type representing an N-bit integer.
Definition: IRBuilder.h:360
Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")
Definition: IRBuilder.h:1933
bool isIntN(unsigned N) const
Check if this APInt has an N-bits unsigned integer value.
Definition: APInt.h:450
void setNotConvergent()
Definition: CallSite.h:527
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:60
void setAlignment(unsigned Align)
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:319
static Value * simplifyX86varShift(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1415
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:530
size_t size() const
Definition: SmallVector.h:53
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1226
LLVM_READONLY APFloat maxnum(const APFloat &A, const APFloat &B)
Implements IEEE maxNum semantics.
Definition: APFloat.h:1238
static InvokeInst * Create(Value *Func, BasicBlock *IfNormal, BasicBlock *IfException, ArrayRef< Value *> Args, const Twine &NameStr, Instruction *InsertBefore=nullptr)
APInt lshr(unsigned shiftAmt) const
Logical right-shift function.
Definition: APInt.h:971
signed greater than
Definition: InstrTypes.h:669
static Constant * getIntegerValue(Type *Ty, const APInt &V)
Return the value for an integer or pointer constant, or a vector thereof, with the given scalar value...
Definition: Constants.cpp:302
static Value * simplifyX86extrq(IntrinsicInst &II, Value *Op0, ConstantInt *CILength, ConstantInt *CIIndex, InstCombiner::BuilderTy &Builder)
Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding or conversion to a shuffle...
const APFloat & getValueAPF() const
Definition: Constants.h:303
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:51
bool doesNotThrow() const
Determine if the function cannot unwind.
Definition: Function.h:520
static BinaryOperator * CreateFNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
static Type * getHalfTy(LLVMContext &C)
Definition: Type.cpp:163
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:240
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.
Iterator for intrusive lists based on ilist_node.
unsigned countMaxLeadingZeros() const
Returns the maximum number of leading zero bits possible.
Definition: KnownBits.h:176
bool hasParamAttribute(unsigned ArgNo, Attribute::AttrKind Kind) const
Equivalent to hasAttribute(ArgNo + FirstArgIndex, Kind).
static PointerType * getInt1PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:216