LLVM  10.0.0svn
InstCombineCalls.cpp
Go to the documentation of this file.
1 //===- InstCombineCalls.cpp -----------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitCall, visitInvoke, and visitCallBr functions.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APFloat.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/APSInt.h"
17 #include "llvm/ADT/ArrayRef.h"
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"
26 #include "llvm/Analysis/Loads.h"
30 #include "llvm/IR/Attributes.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/Constant.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/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"
63 #include <algorithm>
64 #include <cassert>
65 #include <cstdint>
66 #include <cstring>
67 #include <utility>
68 #include <vector>
69 
70 using namespace llvm;
71 using namespace PatternMatch;
72 
73 #define DEBUG_TYPE "instcombine"
74 
75 STATISTIC(NumSimplified, "Number of library calls simplified");
76 
78  "instcombine-guard-widening-window",
79  cl::init(3),
80  cl::desc("How wide an instruction window to bypass looking for "
81  "another guard"));
82 
83 /// Return the specified type promoted as it would be to pass though a va_arg
84 /// area.
85 static Type *getPromotedType(Type *Ty) {
86  if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
87  if (ITy->getBitWidth() < 32)
88  return Type::getInt32Ty(Ty->getContext());
89  }
90  return Ty;
91 }
92 
93 /// Return a constant boolean vector that has true elements in all positions
94 /// where the input constant data vector has an element with the sign bit set.
97  IntegerType *BoolTy = Type::getInt1Ty(V->getContext());
98  for (unsigned I = 0, E = V->getNumElements(); I != E; ++I) {
99  Constant *Elt = V->getElementAsConstant(I);
100  assert((isa<ConstantInt>(Elt) || isa<ConstantFP>(Elt)) &&
101  "Unexpected constant data vector element type");
102  bool Sign = V->getElementType()->isIntegerTy()
103  ? cast<ConstantInt>(Elt)->isNegative()
104  : cast<ConstantFP>(Elt)->isNegative();
105  BoolVec.push_back(ConstantInt::get(BoolTy, Sign));
106  }
107  return ConstantVector::get(BoolVec);
108 }
109 
110 Instruction *InstCombiner::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) {
111  unsigned DstAlign = getKnownAlignment(MI->getRawDest(), DL, MI, &AC, &DT);
112  unsigned CopyDstAlign = MI->getDestAlignment();
113  if (CopyDstAlign < DstAlign){
114  MI->setDestAlignment(DstAlign);
115  return MI;
116  }
117 
118  unsigned SrcAlign = getKnownAlignment(MI->getRawSource(), DL, MI, &AC, &DT);
119  unsigned CopySrcAlign = MI->getSourceAlignment();
120  if (CopySrcAlign < SrcAlign) {
121  MI->setSourceAlignment(SrcAlign);
122  return MI;
123  }
124 
125  // If we have a store to a location which is known constant, we can conclude
126  // that the store must be storing the constant value (else the memory
127  // wouldn't be constant), and this must be a noop.
128  if (AA->pointsToConstantMemory(MI->getDest())) {
129  // Set the size of the copy to 0, it will be deleted on the next iteration.
130  MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
131  return MI;
132  }
133 
134  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
135  // load/store.
136  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getLength());
137  if (!MemOpLength) return nullptr;
138 
139  // Source and destination pointer types are always "i8*" for intrinsic. See
140  // if the size is something we can handle with a single primitive load/store.
141  // A single load+store correctly handles overlapping memory in the memmove
142  // case.
143  uint64_t Size = MemOpLength->getLimitedValue();
144  assert(Size && "0-sized memory transferring should be removed already.");
145 
146  if (Size > 8 || (Size&(Size-1)))
147  return nullptr; // If not 1/2/4/8 bytes, exit.
148 
149  // If it is an atomic and alignment is less than the size then we will
150  // introduce the unaligned memory access which will be later transformed
151  // into libcall in CodeGen. This is not evident performance gain so disable
152  // it now.
153  if (isa<AtomicMemTransferInst>(MI))
154  if (CopyDstAlign < Size || CopySrcAlign < Size)
155  return nullptr;
156 
157  // Use an integer load+store unless we can find something better.
158  unsigned SrcAddrSp =
159  cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
160  unsigned DstAddrSp =
161  cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
162 
163  IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
164  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
165  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
166 
167  // If the memcpy has metadata describing the members, see if we can get the
168  // TBAA tag describing our copy.
169  MDNode *CopyMD = nullptr;
170  if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa)) {
171  CopyMD = M;
172  } else if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
173  if (M->getNumOperands() == 3 && M->getOperand(0) &&
174  mdconst::hasa<ConstantInt>(M->getOperand(0)) &&
175  mdconst::extract<ConstantInt>(M->getOperand(0))->isZero() &&
176  M->getOperand(1) &&
177  mdconst::hasa<ConstantInt>(M->getOperand(1)) &&
178  mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() ==
179  Size &&
180  M->getOperand(2) && isa<MDNode>(M->getOperand(2)))
181  CopyMD = cast<MDNode>(M->getOperand(2));
182  }
183 
184  Value *Src = Builder.CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
185  Value *Dest = Builder.CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
186  LoadInst *L = Builder.CreateLoad(IntType, Src);
187  // Alignment from the mem intrinsic will be better, so use it.
188  L->setAlignment(
189  MaybeAlign(CopySrcAlign)); // FIXME: Check if we can use Align instead.
190  if (CopyMD)
191  L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
192  MDNode *LoopMemParallelMD =
193  MI->getMetadata(LLVMContext::MD_mem_parallel_loop_access);
194  if (LoopMemParallelMD)
195  L->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
196  MDNode *AccessGroupMD = MI->getMetadata(LLVMContext::MD_access_group);
197  if (AccessGroupMD)
198  L->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
199 
200  StoreInst *S = Builder.CreateStore(L, Dest);
201  // Alignment from the mem intrinsic will be better, so use it.
202  S->setAlignment(
203  MaybeAlign(CopyDstAlign)); // FIXME: Check if we can use Align instead.
204  if (CopyMD)
205  S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
206  if (LoopMemParallelMD)
207  S->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
208  if (AccessGroupMD)
209  S->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
210 
211  if (auto *MT = dyn_cast<MemTransferInst>(MI)) {
212  // non-atomics can be volatile
213  L->setVolatile(MT->isVolatile());
214  S->setVolatile(MT->isVolatile());
215  }
216  if (isa<AtomicMemTransferInst>(MI)) {
217  // atomics have to be unordered
220  }
221 
222  // Set the size of the copy to 0, it will be deleted on the next iteration.
223  MI->setLength(Constant::getNullValue(MemOpLength->getType()));
224  return MI;
225 }
226 
227 Instruction *InstCombiner::SimplifyAnyMemSet(AnyMemSetInst *MI) {
228  const unsigned KnownAlignment =
229  getKnownAlignment(MI->getDest(), DL, MI, &AC, &DT);
230  if (MI->getDestAlignment() < KnownAlignment) {
231  MI->setDestAlignment(KnownAlignment);
232  return MI;
233  }
234 
235  // If we have a store to a location which is known constant, we can conclude
236  // that the store must be storing the constant value (else the memory
237  // wouldn't be constant), and this must be a noop.
238  if (AA->pointsToConstantMemory(MI->getDest())) {
239  // Set the size of the copy to 0, it will be deleted on the next iteration.
240  MI->setLength(Constant::getNullValue(MI->getLength()->getType()));
241  return MI;
242  }
243 
244  // Extract the length and alignment and fill if they are constant.
245  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
246  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
247  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
248  return nullptr;
249  const uint64_t Len = LenC->getLimitedValue();
250  assert(Len && "0-sized memory setting should be removed already.");
251  const Align Alignment = assumeAligned(MI->getDestAlignment());
252 
253  // If it is an atomic and alignment is less than the size then we will
254  // introduce the unaligned memory access which will be later transformed
255  // into libcall in CodeGen. This is not evident performance gain so disable
256  // it now.
257  if (isa<AtomicMemSetInst>(MI))
258  if (Alignment < Len)
259  return nullptr;
260 
261  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
262  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
263  Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8.
264 
265  Value *Dest = MI->getDest();
266  unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
267  Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
268  Dest = Builder.CreateBitCast(Dest, NewDstPtrTy);
269 
270  // Extract the fill value and store.
271  uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
272  StoreInst *S = Builder.CreateStore(ConstantInt::get(ITy, Fill), Dest,
273  MI->isVolatile());
274  S->setAlignment(Alignment);
275  if (isa<AtomicMemSetInst>(MI))
277 
278  // Set the size of the copy to 0, it will be deleted on the next iteration.
279  MI->setLength(Constant::getNullValue(LenC->getType()));
280  return MI;
281  }
282 
283  return nullptr;
284 }
285 
287  InstCombiner::BuilderTy &Builder) {
288  bool LogicalShift = false;
289  bool ShiftLeft = false;
290 
291  switch (II.getIntrinsicID()) {
292  default: llvm_unreachable("Unexpected intrinsic!");
293  case Intrinsic::x86_sse2_psra_d:
294  case Intrinsic::x86_sse2_psra_w:
295  case Intrinsic::x86_sse2_psrai_d:
296  case Intrinsic::x86_sse2_psrai_w:
297  case Intrinsic::x86_avx2_psra_d:
298  case Intrinsic::x86_avx2_psra_w:
299  case Intrinsic::x86_avx2_psrai_d:
300  case Intrinsic::x86_avx2_psrai_w:
301  case Intrinsic::x86_avx512_psra_q_128:
302  case Intrinsic::x86_avx512_psrai_q_128:
303  case Intrinsic::x86_avx512_psra_q_256:
304  case Intrinsic::x86_avx512_psrai_q_256:
305  case Intrinsic::x86_avx512_psra_d_512:
306  case Intrinsic::x86_avx512_psra_q_512:
307  case Intrinsic::x86_avx512_psra_w_512:
308  case Intrinsic::x86_avx512_psrai_d_512:
309  case Intrinsic::x86_avx512_psrai_q_512:
310  case Intrinsic::x86_avx512_psrai_w_512:
311  LogicalShift = false; ShiftLeft = false;
312  break;
313  case Intrinsic::x86_sse2_psrl_d:
314  case Intrinsic::x86_sse2_psrl_q:
315  case Intrinsic::x86_sse2_psrl_w:
316  case Intrinsic::x86_sse2_psrli_d:
317  case Intrinsic::x86_sse2_psrli_q:
318  case Intrinsic::x86_sse2_psrli_w:
319  case Intrinsic::x86_avx2_psrl_d:
320  case Intrinsic::x86_avx2_psrl_q:
321  case Intrinsic::x86_avx2_psrl_w:
322  case Intrinsic::x86_avx2_psrli_d:
323  case Intrinsic::x86_avx2_psrli_q:
324  case Intrinsic::x86_avx2_psrli_w:
325  case Intrinsic::x86_avx512_psrl_d_512:
326  case Intrinsic::x86_avx512_psrl_q_512:
327  case Intrinsic::x86_avx512_psrl_w_512:
328  case Intrinsic::x86_avx512_psrli_d_512:
329  case Intrinsic::x86_avx512_psrli_q_512:
330  case Intrinsic::x86_avx512_psrli_w_512:
331  LogicalShift = true; ShiftLeft = false;
332  break;
333  case Intrinsic::x86_sse2_psll_d:
334  case Intrinsic::x86_sse2_psll_q:
335  case Intrinsic::x86_sse2_psll_w:
336  case Intrinsic::x86_sse2_pslli_d:
337  case Intrinsic::x86_sse2_pslli_q:
338  case Intrinsic::x86_sse2_pslli_w:
339  case Intrinsic::x86_avx2_psll_d:
340  case Intrinsic::x86_avx2_psll_q:
341  case Intrinsic::x86_avx2_psll_w:
342  case Intrinsic::x86_avx2_pslli_d:
343  case Intrinsic::x86_avx2_pslli_q:
344  case Intrinsic::x86_avx2_pslli_w:
345  case Intrinsic::x86_avx512_psll_d_512:
346  case Intrinsic::x86_avx512_psll_q_512:
347  case Intrinsic::x86_avx512_psll_w_512:
348  case Intrinsic::x86_avx512_pslli_d_512:
349  case Intrinsic::x86_avx512_pslli_q_512:
350  case Intrinsic::x86_avx512_pslli_w_512:
351  LogicalShift = true; ShiftLeft = true;
352  break;
353  }
354  assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
355 
356  // Simplify if count is constant.
357  auto Arg1 = II.getArgOperand(1);
358  auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1);
359  auto CDV = dyn_cast<ConstantDataVector>(Arg1);
360  auto CInt = dyn_cast<ConstantInt>(Arg1);
361  if (!CAZ && !CDV && !CInt)
362  return nullptr;
363 
364  APInt Count(64, 0);
365  if (CDV) {
366  // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
367  // operand to compute the shift amount.
368  auto VT = cast<VectorType>(CDV->getType());
369  unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits();
370  assert((64 % BitWidth) == 0 && "Unexpected packed shift size");
371  unsigned NumSubElts = 64 / BitWidth;
372 
373  // Concatenate the sub-elements to create the 64-bit value.
374  for (unsigned i = 0; i != NumSubElts; ++i) {
375  unsigned SubEltIdx = (NumSubElts - 1) - i;
376  auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
377  Count <<= BitWidth;
378  Count |= SubElt->getValue().zextOrTrunc(64);
379  }
380  }
381  else if (CInt)
382  Count = CInt->getValue();
383 
384  auto Vec = II.getArgOperand(0);
385  auto VT = cast<VectorType>(Vec->getType());
386  auto SVT = VT->getElementType();
387  unsigned VWidth = VT->getNumElements();
388  unsigned BitWidth = SVT->getPrimitiveSizeInBits();
389 
390  // If shift-by-zero then just return the original value.
391  if (Count.isNullValue())
392  return Vec;
393 
394  // Handle cases when Shift >= BitWidth.
395  if (Count.uge(BitWidth)) {
396  // If LogicalShift - just return zero.
397  if (LogicalShift)
398  return ConstantAggregateZero::get(VT);
399 
400  // If ArithmeticShift - clamp Shift to (BitWidth - 1).
401  Count = APInt(64, BitWidth - 1);
402  }
403 
404  // Get a constant vector of the same type as the first operand.
405  auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
406  auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
407 
408  if (ShiftLeft)
409  return Builder.CreateShl(Vec, ShiftVec);
410 
411  if (LogicalShift)
412  return Builder.CreateLShr(Vec, ShiftVec);
413 
414  return Builder.CreateAShr(Vec, ShiftVec);
415 }
416 
417 // Attempt to simplify AVX2 per-element shift intrinsics to a generic IR shift.
418 // Unlike the generic IR shifts, the intrinsics have defined behaviour for out
419 // of range shift amounts (logical - set to zero, arithmetic - splat sign bit).
421  InstCombiner::BuilderTy &Builder) {
422  bool LogicalShift = false;
423  bool ShiftLeft = false;
424 
425  switch (II.getIntrinsicID()) {
426  default: llvm_unreachable("Unexpected intrinsic!");
427  case Intrinsic::x86_avx2_psrav_d:
428  case Intrinsic::x86_avx2_psrav_d_256:
429  case Intrinsic::x86_avx512_psrav_q_128:
430  case Intrinsic::x86_avx512_psrav_q_256:
431  case Intrinsic::x86_avx512_psrav_d_512:
432  case Intrinsic::x86_avx512_psrav_q_512:
433  case Intrinsic::x86_avx512_psrav_w_128:
434  case Intrinsic::x86_avx512_psrav_w_256:
435  case Intrinsic::x86_avx512_psrav_w_512:
436  LogicalShift = false;
437  ShiftLeft = false;
438  break;
439  case Intrinsic::x86_avx2_psrlv_d:
440  case Intrinsic::x86_avx2_psrlv_d_256:
441  case Intrinsic::x86_avx2_psrlv_q:
442  case Intrinsic::x86_avx2_psrlv_q_256:
443  case Intrinsic::x86_avx512_psrlv_d_512:
444  case Intrinsic::x86_avx512_psrlv_q_512:
445  case Intrinsic::x86_avx512_psrlv_w_128:
446  case Intrinsic::x86_avx512_psrlv_w_256:
447  case Intrinsic::x86_avx512_psrlv_w_512:
448  LogicalShift = true;
449  ShiftLeft = false;
450  break;
451  case Intrinsic::x86_avx2_psllv_d:
452  case Intrinsic::x86_avx2_psllv_d_256:
453  case Intrinsic::x86_avx2_psllv_q:
454  case Intrinsic::x86_avx2_psllv_q_256:
455  case Intrinsic::x86_avx512_psllv_d_512:
456  case Intrinsic::x86_avx512_psllv_q_512:
457  case Intrinsic::x86_avx512_psllv_w_128:
458  case Intrinsic::x86_avx512_psllv_w_256:
459  case Intrinsic::x86_avx512_psllv_w_512:
460  LogicalShift = true;
461  ShiftLeft = true;
462  break;
463  }
464  assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
465 
466  // Simplify if all shift amounts are constant/undef.
467  auto *CShift = dyn_cast<Constant>(II.getArgOperand(1));
468  if (!CShift)
469  return nullptr;
470 
471  auto Vec = II.getArgOperand(0);
472  auto VT = cast<VectorType>(II.getType());
473  auto SVT = VT->getVectorElementType();
474  int NumElts = VT->getNumElements();
475  int BitWidth = SVT->getIntegerBitWidth();
476 
477  // Collect each element's shift amount.
478  // We also collect special cases: UNDEF = -1, OUT-OF-RANGE = BitWidth.
479  bool AnyOutOfRange = false;
480  SmallVector<int, 8> ShiftAmts;
481  for (int I = 0; I < NumElts; ++I) {
482  auto *CElt = CShift->getAggregateElement(I);
483  if (CElt && isa<UndefValue>(CElt)) {
484  ShiftAmts.push_back(-1);
485  continue;
486  }
487 
488  auto *COp = dyn_cast_or_null<ConstantInt>(CElt);
489  if (!COp)
490  return nullptr;
491 
492  // Handle out of range shifts.
493  // If LogicalShift - set to BitWidth (special case).
494  // If ArithmeticShift - set to (BitWidth - 1) (sign splat).
495  APInt ShiftVal = COp->getValue();
496  if (ShiftVal.uge(BitWidth)) {
497  AnyOutOfRange = LogicalShift;
498  ShiftAmts.push_back(LogicalShift ? BitWidth : BitWidth - 1);
499  continue;
500  }
501 
502  ShiftAmts.push_back((int)ShiftVal.getZExtValue());
503  }
504 
505  // If all elements out of range or UNDEF, return vector of zeros/undefs.
506  // ArithmeticShift should only hit this if they are all UNDEF.
507  auto OutOfRange = [&](int Idx) { return (Idx < 0) || (BitWidth <= Idx); };
508  if (llvm::all_of(ShiftAmts, OutOfRange)) {
509  SmallVector<Constant *, 8> ConstantVec;
510  for (int Idx : ShiftAmts) {
511  if (Idx < 0) {
512  ConstantVec.push_back(UndefValue::get(SVT));
513  } else {
514  assert(LogicalShift && "Logical shift expected");
515  ConstantVec.push_back(ConstantInt::getNullValue(SVT));
516  }
517  }
518  return ConstantVector::get(ConstantVec);
519  }
520 
521  // We can't handle only some out of range values with generic logical shifts.
522  if (AnyOutOfRange)
523  return nullptr;
524 
525  // Build the shift amount constant vector.
526  SmallVector<Constant *, 8> ShiftVecAmts;
527  for (int Idx : ShiftAmts) {
528  if (Idx < 0)
529  ShiftVecAmts.push_back(UndefValue::get(SVT));
530  else
531  ShiftVecAmts.push_back(ConstantInt::get(SVT, Idx));
532  }
533  auto ShiftVec = ConstantVector::get(ShiftVecAmts);
534 
535  if (ShiftLeft)
536  return Builder.CreateShl(Vec, ShiftVec);
537 
538  if (LogicalShift)
539  return Builder.CreateLShr(Vec, ShiftVec);
540 
541  return Builder.CreateAShr(Vec, ShiftVec);
542 }
543 
545  InstCombiner::BuilderTy &Builder, bool IsSigned) {
546  Value *Arg0 = II.getArgOperand(0);
547  Value *Arg1 = II.getArgOperand(1);
548  Type *ResTy = II.getType();
549 
550  // Fast all undef handling.
551  if (isa<UndefValue>(Arg0) && isa<UndefValue>(Arg1))
552  return UndefValue::get(ResTy);
553 
554  Type *ArgTy = Arg0->getType();
555  unsigned NumLanes = ResTy->getPrimitiveSizeInBits() / 128;
556  unsigned NumSrcElts = ArgTy->getVectorNumElements();
557  assert(ResTy->getVectorNumElements() == (2 * NumSrcElts) &&
558  "Unexpected packing types");
559 
560  unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes;
561  unsigned DstScalarSizeInBits = ResTy->getScalarSizeInBits();
562  unsigned SrcScalarSizeInBits = ArgTy->getScalarSizeInBits();
563  assert(SrcScalarSizeInBits == (2 * DstScalarSizeInBits) &&
564  "Unexpected packing types");
565 
566  // Constant folding.
567  if (!isa<Constant>(Arg0) || !isa<Constant>(Arg1))
568  return nullptr;
569 
570  // Clamp Values - signed/unsigned both use signed clamp values, but they
571  // differ on the min/max values.
572  APInt MinValue, MaxValue;
573  if (IsSigned) {
574  // PACKSS: Truncate signed value with signed saturation.
575  // Source values less than dst minint are saturated to minint.
576  // Source values greater than dst maxint are saturated to maxint.
577  MinValue =
578  APInt::getSignedMinValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
579  MaxValue =
580  APInt::getSignedMaxValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
581  } else {
582  // PACKUS: Truncate signed value with unsigned saturation.
583  // Source values less than zero are saturated to zero.
584  // Source values greater than dst maxuint are saturated to maxuint.
585  MinValue = APInt::getNullValue(SrcScalarSizeInBits);
586  MaxValue = APInt::getLowBitsSet(SrcScalarSizeInBits, DstScalarSizeInBits);
587  }
588 
589  auto *MinC = Constant::getIntegerValue(ArgTy, MinValue);
590  auto *MaxC = Constant::getIntegerValue(ArgTy, MaxValue);
591  Arg0 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg0, MinC), MinC, Arg0);
592  Arg1 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg1, MinC), MinC, Arg1);
593  Arg0 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg0, MaxC), MaxC, Arg0);
594  Arg1 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg1, MaxC), MaxC, Arg1);
595 
596  // Shuffle clamped args together at the lane level.
597  SmallVector<unsigned, 32> PackMask;
598  for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
599  for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
600  PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane));
601  for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
602  PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane) + NumSrcElts);
603  }
604  auto *Shuffle = Builder.CreateShuffleVector(Arg0, Arg1, PackMask);
605 
606  // Truncate to dst size.
607  return Builder.CreateTrunc(Shuffle, ResTy);
608 }
609 
611  InstCombiner::BuilderTy &Builder) {
612  Value *Arg = II.getArgOperand(0);
613  Type *ResTy = II.getType();
614  Type *ArgTy = Arg->getType();
615 
616  // movmsk(undef) -> zero as we must ensure the upper bits are zero.
617  if (isa<UndefValue>(Arg))
618  return Constant::getNullValue(ResTy);
619 
620  // We can't easily peek through x86_mmx types.
621  if (!ArgTy->isVectorTy())
622  return nullptr;
623 
624  // Expand MOVMSK to compare/bitcast/zext:
625  // e.g. PMOVMSKB(v16i8 x):
626  // %cmp = icmp slt <16 x i8> %x, zeroinitializer
627  // %int = bitcast <16 x i1> %cmp to i16
628  // %res = zext i16 %int to i32
629  unsigned NumElts = ArgTy->getVectorNumElements();
630  Type *IntegerVecTy = VectorType::getInteger(cast<VectorType>(ArgTy));
631  Type *IntegerTy = Builder.getIntNTy(NumElts);
632 
633  Value *Res = Builder.CreateBitCast(Arg, IntegerVecTy);
634  Res = Builder.CreateICmpSLT(Res, Constant::getNullValue(IntegerVecTy));
635  Res = Builder.CreateBitCast(Res, IntegerTy);
636  Res = Builder.CreateZExtOrTrunc(Res, ResTy);
637  return Res;
638 }
639 
641  InstCombiner::BuilderTy &Builder) {
642  Value *CarryIn = II.getArgOperand(0);
643  Value *Op1 = II.getArgOperand(1);
644  Value *Op2 = II.getArgOperand(2);
645  Type *RetTy = II.getType();
646  Type *OpTy = Op1->getType();
647  assert(RetTy->getStructElementType(0)->isIntegerTy(8) &&
648  RetTy->getStructElementType(1) == OpTy && OpTy == Op2->getType() &&
649  "Unexpected types for x86 addcarry");
650 
651  // If carry-in is zero, this is just an unsigned add with overflow.
652  if (match(CarryIn, m_ZeroInt())) {
653  Value *UAdd = Builder.CreateIntrinsic(Intrinsic::uadd_with_overflow, OpTy,
654  { Op1, Op2 });
655  // The types have to be adjusted to match the x86 call types.
656  Value *UAddResult = Builder.CreateExtractValue(UAdd, 0);
657  Value *UAddOV = Builder.CreateZExt(Builder.CreateExtractValue(UAdd, 1),
658  Builder.getInt8Ty());
659  Value *Res = UndefValue::get(RetTy);
660  Res = Builder.CreateInsertValue(Res, UAddOV, 0);
661  return Builder.CreateInsertValue(Res, UAddResult, 1);
662  }
663 
664  return nullptr;
665 }
666 
668  InstCombiner::BuilderTy &Builder) {
669  auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2));
670  if (!CInt)
671  return nullptr;
672 
673  VectorType *VecTy = cast<VectorType>(II.getType());
674  assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
675 
676  // The immediate permute control byte looks like this:
677  // [3:0] - zero mask for each 32-bit lane
678  // [5:4] - select one 32-bit destination lane
679  // [7:6] - select one 32-bit source lane
680 
681  uint8_t Imm = CInt->getZExtValue();
682  uint8_t ZMask = Imm & 0xf;
683  uint8_t DestLane = (Imm >> 4) & 0x3;
684  uint8_t SourceLane = (Imm >> 6) & 0x3;
685 
687 
688  // If all zero mask bits are set, this was just a weird way to
689  // generate a zero vector.
690  if (ZMask == 0xf)
691  return ZeroVector;
692 
693  // Initialize by passing all of the first source bits through.
694  uint32_t ShuffleMask[4] = { 0, 1, 2, 3 };
695 
696  // We may replace the second operand with the zero vector.
697  Value *V1 = II.getArgOperand(1);
698 
699  if (ZMask) {
700  // If the zero mask is being used with a single input or the zero mask
701  // overrides the destination lane, this is a shuffle with the zero vector.
702  if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
703  (ZMask & (1 << DestLane))) {
704  V1 = ZeroVector;
705  // We may still move 32-bits of the first source vector from one lane
706  // to another.
707  ShuffleMask[DestLane] = SourceLane;
708  // The zero mask may override the previous insert operation.
709  for (unsigned i = 0; i < 4; ++i)
710  if ((ZMask >> i) & 0x1)
711  ShuffleMask[i] = i + 4;
712  } else {
713  // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
714  return nullptr;
715  }
716  } else {
717  // Replace the selected destination lane with the selected source lane.
718  ShuffleMask[DestLane] = SourceLane + 4;
719  }
720 
721  return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
722 }
723 
724 /// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding
725 /// or conversion to a shuffle vector.
727  ConstantInt *CILength, ConstantInt *CIIndex,
728  InstCombiner::BuilderTy &Builder) {
729  auto LowConstantHighUndef = [&](uint64_t Val) {
730  Type *IntTy64 = Type::getInt64Ty(II.getContext());
731  Constant *Args[] = {ConstantInt::get(IntTy64, Val),
732  UndefValue::get(IntTy64)};
733  return ConstantVector::get(Args);
734  };
735 
736  // See if we're dealing with constant values.
737  Constant *C0 = dyn_cast<Constant>(Op0);
738  ConstantInt *CI0 =
739  C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
740  : nullptr;
741 
742  // Attempt to constant fold.
743  if (CILength && CIIndex) {
744  // From AMD documentation: "The bit index and field length are each six
745  // bits in length other bits of the field are ignored."
746  APInt APIndex = CIIndex->getValue().zextOrTrunc(6);
747  APInt APLength = CILength->getValue().zextOrTrunc(6);
748 
749  unsigned Index = APIndex.getZExtValue();
750 
751  // From AMD documentation: "a value of zero in the field length is
752  // defined as length of 64".
753  unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
754 
755  // From AMD documentation: "If the sum of the bit index + length field
756  // is greater than 64, the results are undefined".
757  unsigned End = Index + Length;
758 
759  // Note that both field index and field length are 8-bit quantities.
760  // Since variables 'Index' and 'Length' are unsigned values
761  // obtained from zero-extending field index and field length
762  // respectively, their sum should never wrap around.
763  if (End > 64)
764  return UndefValue::get(II.getType());
765 
766  // If we are inserting whole bytes, we can convert this to a shuffle.
767  // Lowering can recognize EXTRQI shuffle masks.
768  if ((Length % 8) == 0 && (Index % 8) == 0) {
769  // Convert bit indices to byte indices.
770  Length /= 8;
771  Index /= 8;
772 
773  Type *IntTy8 = Type::getInt8Ty(II.getContext());
774  Type *IntTy32 = Type::getInt32Ty(II.getContext());
775  VectorType *ShufTy = VectorType::get(IntTy8, 16);
776 
777  SmallVector<Constant *, 16> ShuffleMask;
778  for (int i = 0; i != (int)Length; ++i)
779  ShuffleMask.push_back(
780  Constant::getIntegerValue(IntTy32, APInt(32, i + Index)));
781  for (int i = Length; i != 8; ++i)
782  ShuffleMask.push_back(
783  Constant::getIntegerValue(IntTy32, APInt(32, i + 16)));
784  for (int i = 8; i != 16; ++i)
785  ShuffleMask.push_back(UndefValue::get(IntTy32));
786 
787  Value *SV = Builder.CreateShuffleVector(
788  Builder.CreateBitCast(Op0, ShufTy),
789  ConstantAggregateZero::get(ShufTy), ConstantVector::get(ShuffleMask));
790  return Builder.CreateBitCast(SV, II.getType());
791  }
792 
793  // Constant Fold - shift Index'th bit to lowest position and mask off
794  // Length bits.
795  if (CI0) {
796  APInt Elt = CI0->getValue();
797  Elt.lshrInPlace(Index);
798  Elt = Elt.zextOrTrunc(Length);
799  return LowConstantHighUndef(Elt.getZExtValue());
800  }
801 
802  // If we were an EXTRQ call, we'll save registers if we convert to EXTRQI.
803  if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) {
804  Value *Args[] = {Op0, CILength, CIIndex};
805  Module *M = II.getModule();
806  Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi);
807  return Builder.CreateCall(F, Args);
808  }
809  }
810 
811  // Constant Fold - extraction from zero is always {zero, undef}.
812  if (CI0 && CI0->isZero())
813  return LowConstantHighUndef(0);
814 
815  return nullptr;
816 }
817 
818 /// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant
819 /// folding or conversion to a shuffle vector.
821  APInt APLength, APInt APIndex,
822  InstCombiner::BuilderTy &Builder) {
823  // From AMD documentation: "The bit index and field length are each six bits
824  // in length other bits of the field are ignored."
825  APIndex = APIndex.zextOrTrunc(6);
826  APLength = APLength.zextOrTrunc(6);
827 
828  // Attempt to constant fold.
829  unsigned Index = APIndex.getZExtValue();
830 
831  // From AMD documentation: "a value of zero in the field length is
832  // defined as length of 64".
833  unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
834 
835  // From AMD documentation: "If the sum of the bit index + length field
836  // is greater than 64, the results are undefined".
837  unsigned End = Index + Length;
838 
839  // Note that both field index and field length are 8-bit quantities.
840  // Since variables 'Index' and 'Length' are unsigned values
841  // obtained from zero-extending field index and field length
842  // respectively, their sum should never wrap around.
843  if (End > 64)
844  return UndefValue::get(II.getType());
845 
846  // If we are inserting whole bytes, we can convert this to a shuffle.
847  // Lowering can recognize INSERTQI shuffle masks.
848  if ((Length % 8) == 0 && (Index % 8) == 0) {
849  // Convert bit indices to byte indices.
850  Length /= 8;
851  Index /= 8;
852 
853  Type *IntTy8 = Type::getInt8Ty(II.getContext());
854  Type *IntTy32 = Type::getInt32Ty(II.getContext());
855  VectorType *ShufTy = VectorType::get(IntTy8, 16);
856 
857  SmallVector<Constant *, 16> ShuffleMask;
858  for (int i = 0; i != (int)Index; ++i)
859  ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i)));
860  for (int i = 0; i != (int)Length; ++i)
861  ShuffleMask.push_back(
862  Constant::getIntegerValue(IntTy32, APInt(32, i + 16)));
863  for (int i = Index + Length; i != 8; ++i)
864  ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i)));
865  for (int i = 8; i != 16; ++i)
866  ShuffleMask.push_back(UndefValue::get(IntTy32));
867 
868  Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy),
869  Builder.CreateBitCast(Op1, ShufTy),
870  ConstantVector::get(ShuffleMask));
871  return Builder.CreateBitCast(SV, II.getType());
872  }
873 
874  // See if we're dealing with constant values.
875  Constant *C0 = dyn_cast<Constant>(Op0);
876  Constant *C1 = dyn_cast<Constant>(Op1);
877  ConstantInt *CI00 =
878  C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
879  : nullptr;
880  ConstantInt *CI10 =
881  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
882  : nullptr;
883 
884  // Constant Fold - insert bottom Length bits starting at the Index'th bit.
885  if (CI00 && CI10) {
886  APInt V00 = CI00->getValue();
887  APInt V10 = CI10->getValue();
888  APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index);
889  V00 = V00 & ~Mask;
890  V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index);
891  APInt Val = V00 | V10;
892  Type *IntTy64 = Type::getInt64Ty(II.getContext());
893  Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()),
894  UndefValue::get(IntTy64)};
895  return ConstantVector::get(Args);
896  }
897 
898  // If we were an INSERTQ call, we'll save demanded elements if we convert to
899  // INSERTQI.
900  if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) {
901  Type *IntTy8 = Type::getInt8Ty(II.getContext());
902  Constant *CILength = ConstantInt::get(IntTy8, Length, false);
903  Constant *CIIndex = ConstantInt::get(IntTy8, Index, false);
904 
905  Value *Args[] = {Op0, Op1, CILength, CIIndex};
906  Module *M = II.getModule();
907  Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
908  return Builder.CreateCall(F, Args);
909  }
910 
911  return nullptr;
912 }
913 
914 /// Attempt to convert pshufb* to shufflevector if the mask is constant.
916  InstCombiner::BuilderTy &Builder) {
918  if (!V)
919  return nullptr;
920 
921  auto *VecTy = cast<VectorType>(II.getType());
922  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
923  unsigned NumElts = VecTy->getNumElements();
924  assert((NumElts == 16 || NumElts == 32 || NumElts == 64) &&
925  "Unexpected number of elements in shuffle mask!");
926 
927  // Construct a shuffle mask from constant integers or UNDEFs.
928  Constant *Indexes[64] = {nullptr};
929 
930  // Each byte in the shuffle control mask forms an index to permute the
931  // corresponding byte in the destination operand.
932  for (unsigned I = 0; I < NumElts; ++I) {
933  Constant *COp = V->getAggregateElement(I);
934  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
935  return nullptr;
936 
937  if (isa<UndefValue>(COp)) {
938  Indexes[I] = UndefValue::get(MaskEltTy);
939  continue;
940  }
941 
942  int8_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue();
943 
944  // If the most significant bit (bit[7]) of each byte of the shuffle
945  // control mask is set, then zero is written in the result byte.
946  // The zero vector is in the right-hand side of the resulting
947  // shufflevector.
948 
949  // The value of each index for the high 128-bit lane is the least
950  // significant 4 bits of the respective shuffle control byte.
951  Index = ((Index < 0) ? NumElts : Index & 0x0F) + (I & 0xF0);
952  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
953  }
954 
955  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts));
956  auto V1 = II.getArgOperand(0);
957  auto V2 = Constant::getNullValue(VecTy);
958  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
959 }
960 
961 /// Attempt to convert vpermilvar* to shufflevector if the mask is constant.
963  InstCombiner::BuilderTy &Builder) {
965  if (!V)
966  return nullptr;
967 
968  auto *VecTy = cast<VectorType>(II.getType());
969  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
970  unsigned NumElts = VecTy->getVectorNumElements();
971  bool IsPD = VecTy->getScalarType()->isDoubleTy();
972  unsigned NumLaneElts = IsPD ? 2 : 4;
973  assert(NumElts == 16 || NumElts == 8 || NumElts == 4 || NumElts == 2);
974 
975  // Construct a shuffle mask from constant integers or UNDEFs.
976  Constant *Indexes[16] = {nullptr};
977 
978  // The intrinsics only read one or two bits, clear the rest.
979  for (unsigned I = 0; I < NumElts; ++I) {
980  Constant *COp = V->getAggregateElement(I);
981  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
982  return nullptr;
983 
984  if (isa<UndefValue>(COp)) {
985  Indexes[I] = UndefValue::get(MaskEltTy);
986  continue;
987  }
988 
989  APInt Index = cast<ConstantInt>(COp)->getValue();
990  Index = Index.zextOrTrunc(32).getLoBits(2);
991 
992  // The PD variants uses bit 1 to select per-lane element index, so
993  // shift down to convert to generic shuffle mask index.
994  if (IsPD)
995  Index.lshrInPlace(1);
996 
997  // The _256 variants are a bit trickier since the mask bits always index
998  // into the corresponding 128 half. In order to convert to a generic
999  // shuffle, we have to make that explicit.
1000  Index += APInt(32, (I / NumLaneElts) * NumLaneElts);
1001 
1002  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
1003  }
1004 
1005  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts));
1006  auto V1 = II.getArgOperand(0);
1007  auto V2 = UndefValue::get(V1->getType());
1008  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1009 }
1010 
1011 /// Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
1013  InstCombiner::BuilderTy &Builder) {
1014  auto *V = dyn_cast<Constant>(II.getArgOperand(1));
1015  if (!V)
1016  return nullptr;
1017 
1018  auto *VecTy = cast<VectorType>(II.getType());
1019  auto *MaskEltTy = Type::getInt32Ty(II.getContext());
1020  unsigned Size = VecTy->getNumElements();
1021  assert((Size == 4 || Size == 8 || Size == 16 || Size == 32 || Size == 64) &&
1022  "Unexpected shuffle mask size");
1023 
1024  // Construct a shuffle mask from constant integers or UNDEFs.
1025  Constant *Indexes[64] = {nullptr};
1026 
1027  for (unsigned I = 0; I < Size; ++I) {
1028  Constant *COp = V->getAggregateElement(I);
1029  if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
1030  return nullptr;
1031 
1032  if (isa<UndefValue>(COp)) {
1033  Indexes[I] = UndefValue::get(MaskEltTy);
1034  continue;
1035  }
1036 
1037  uint32_t Index = cast<ConstantInt>(COp)->getZExtValue();
1038  Index &= Size - 1;
1039  Indexes[I] = ConstantInt::get(MaskEltTy, Index);
1040  }
1041 
1042  auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, Size));
1043  auto V1 = II.getArgOperand(0);
1044  auto V2 = UndefValue::get(VecTy);
1045  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1046 }
1047 
1048 // TODO, Obvious Missing Transforms:
1049 // * Narrow width by halfs excluding zero/undef lanes
1050 Value *InstCombiner::simplifyMaskedLoad(IntrinsicInst &II) {
1051  Value *LoadPtr = II.getArgOperand(0);
1052  unsigned Alignment = cast<ConstantInt>(II.getArgOperand(1))->getZExtValue();
1053 
1054  // If the mask is all ones or undefs, this is a plain vector load of the 1st
1055  // argument.
1057  return Builder.CreateAlignedLoad(II.getType(), LoadPtr, Alignment,
1058  "unmaskedload");
1059 
1060  // If we can unconditionally load from this address, replace with a
1061  // load/select idiom. TODO: use DT for context sensitive query
1063  LoadPtr, II.getType(), MaybeAlign(Alignment),
1064  II.getModule()->getDataLayout(), &II, nullptr)) {
1065  Value *LI = Builder.CreateAlignedLoad(II.getType(), LoadPtr, Alignment,
1066  "unmaskedload");
1067  return Builder.CreateSelect(II.getArgOperand(2), LI, II.getArgOperand(3));
1068  }
1069 
1070  return nullptr;
1071 }
1072 
1073 // TODO, Obvious Missing Transforms:
1074 // * Single constant active lane -> store
1075 // * Narrow width by halfs excluding zero/undef lanes
1076 Instruction *InstCombiner::simplifyMaskedStore(IntrinsicInst &II) {
1077  auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
1078  if (!ConstMask)
1079  return nullptr;
1080 
1081  // If the mask is all zeros, this instruction does nothing.
1082  if (ConstMask->isNullValue())
1083  return eraseInstFromFunction(II);
1084 
1085  // If the mask is all ones, this is a plain vector store of the 1st argument.
1086  if (ConstMask->isAllOnesValue()) {
1087  Value *StorePtr = II.getArgOperand(1);
1088  MaybeAlign Alignment(
1089  cast<ConstantInt>(II.getArgOperand(2))->getZExtValue());
1090  return new StoreInst(II.getArgOperand(0), StorePtr, false, Alignment);
1091  }
1092 
1093  // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
1094  APInt DemandedElts = possiblyDemandedEltsInMask(ConstMask);
1095  APInt UndefElts(DemandedElts.getBitWidth(), 0);
1096  if (Value *V = SimplifyDemandedVectorElts(II.getOperand(0),
1097  DemandedElts, UndefElts)) {
1098  II.setOperand(0, V);
1099  return &II;
1100  }
1101 
1102  return nullptr;
1103 }
1104 
1105 // TODO, Obvious Missing Transforms:
1106 // * Single constant active lane load -> load
1107 // * Dereferenceable address & few lanes -> scalarize speculative load/selects
1108 // * Adjacent vector addresses -> masked.load
1109 // * Narrow width by halfs excluding zero/undef lanes
1110 // * Vector splat address w/known mask -> scalar load
1111 // * Vector incrementing address -> vector masked load
1112 Instruction *InstCombiner::simplifyMaskedGather(IntrinsicInst &II) {
1113  return nullptr;
1114 }
1115 
1116 // TODO, Obvious Missing Transforms:
1117 // * Single constant active lane -> store
1118 // * Adjacent vector addresses -> masked.store
1119 // * Narrow store width by halfs excluding zero/undef lanes
1120 // * Vector splat address w/known mask -> scalar store
1121 // * Vector incrementing address -> vector masked store
1122 Instruction *InstCombiner::simplifyMaskedScatter(IntrinsicInst &II) {
1123  auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3));
1124  if (!ConstMask)
1125  return nullptr;
1126 
1127  // If the mask is all zeros, a scatter does nothing.
1128  if (ConstMask->isNullValue())
1129  return eraseInstFromFunction(II);
1130 
1131  // Use masked off lanes to simplify operands via SimplifyDemandedVectorElts
1132  APInt DemandedElts = possiblyDemandedEltsInMask(ConstMask);
1133  APInt UndefElts(DemandedElts.getBitWidth(), 0);
1134  if (Value *V = SimplifyDemandedVectorElts(II.getOperand(0),
1135  DemandedElts, UndefElts)) {
1136  II.setOperand(0, V);
1137  return &II;
1138  }
1139  if (Value *V = SimplifyDemandedVectorElts(II.getOperand(1),
1140  DemandedElts, UndefElts)) {
1141  II.setOperand(1, V);
1142  return &II;
1143  }
1144 
1145  return nullptr;
1146 }
1147 
1148 /// This function transforms launder.invariant.group and strip.invariant.group
1149 /// like:
1150 /// launder(launder(%x)) -> launder(%x) (the result is not the argument)
1151 /// launder(strip(%x)) -> launder(%x)
1152 /// strip(strip(%x)) -> strip(%x) (the result is not the argument)
1153 /// strip(launder(%x)) -> strip(%x)
1154 /// This is legal because it preserves the most recent information about
1155 /// the presence or absence of invariant.group.
1157  InstCombiner &IC) {
1158  auto *Arg = II.getArgOperand(0);
1159  auto *StrippedArg = Arg->stripPointerCasts();
1160  auto *StrippedInvariantGroupsArg = Arg->stripPointerCastsAndInvariantGroups();
1161  if (StrippedArg == StrippedInvariantGroupsArg)
1162  return nullptr; // No launders/strips to remove.
1163 
1164  Value *Result = nullptr;
1165 
1166  if (II.getIntrinsicID() == Intrinsic::launder_invariant_group)
1167  Result = IC.Builder.CreateLaunderInvariantGroup(StrippedInvariantGroupsArg);
1168  else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group)
1169  Result = IC.Builder.CreateStripInvariantGroup(StrippedInvariantGroupsArg);
1170  else
1172  "simplifyInvariantGroupIntrinsic only handles launder and strip");
1173  if (Result->getType()->getPointerAddressSpace() !=
1175  Result = IC.Builder.CreateAddrSpaceCast(Result, II.getType());
1176  if (Result->getType() != II.getType())
1177  Result = IC.Builder.CreateBitCast(Result, II.getType());
1178 
1179  return cast<Instruction>(Result);
1180 }
1181 
1183  assert((II.getIntrinsicID() == Intrinsic::cttz ||
1184  II.getIntrinsicID() == Intrinsic::ctlz) &&
1185  "Expected cttz or ctlz intrinsic");
1186  bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz;
1187  Value *Op0 = II.getArgOperand(0);
1188  Value *X;
1189  // ctlz(bitreverse(x)) -> cttz(x)
1190  // cttz(bitreverse(x)) -> ctlz(x)
1191  if (match(Op0, m_BitReverse(m_Value(X)))) {
1192  Intrinsic::ID ID = IsTZ ? Intrinsic::ctlz : Intrinsic::cttz;
1194  return CallInst::Create(F, {X, II.getArgOperand(1)});
1195  }
1196 
1197  if (IsTZ) {
1198  // cttz(-x) -> cttz(x)
1199  if (match(Op0, m_Neg(m_Value(X)))) {
1200  II.setOperand(0, X);
1201  return &II;
1202  }
1203 
1204  // cttz(abs(x)) -> cttz(x)
1205  // cttz(nabs(x)) -> cttz(x)
1206  Value *Y;
1208  if (SPF == SPF_ABS || SPF == SPF_NABS) {
1209  II.setOperand(0, X);
1210  return &II;
1211  }
1212  }
1213 
1214  KnownBits Known = IC.computeKnownBits(Op0, 0, &II);
1215 
1216  // Create a mask for bits above (ctlz) or below (cttz) the first known one.
1217  unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros()
1218  : Known.countMaxLeadingZeros();
1219  unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros()
1220  : Known.countMinLeadingZeros();
1221 
1222  // If all bits above (ctlz) or below (cttz) the first known one are known
1223  // zero, this value is constant.
1224  // FIXME: This should be in InstSimplify because we're replacing an
1225  // instruction with a constant.
1226  if (PossibleZeros == DefiniteZeros) {
1227  auto *C = ConstantInt::get(Op0->getType(), DefiniteZeros);
1228  return IC.replaceInstUsesWith(II, C);
1229  }
1230 
1231  // If the input to cttz/ctlz is known to be non-zero,
1232  // then change the 'ZeroIsUndef' parameter to 'true'
1233  // because we know the zero behavior can't affect the result.
1234  if (!Known.One.isNullValue() ||
1235  isKnownNonZero(Op0, IC.getDataLayout(), 0, &IC.getAssumptionCache(), &II,
1236  &IC.getDominatorTree())) {
1237  if (!match(II.getArgOperand(1), m_One())) {
1238  II.setOperand(1, IC.Builder.getTrue());
1239  return &II;
1240  }
1241  }
1242 
1243  // Add range metadata since known bits can't completely reflect what we know.
1244  // TODO: Handle splat vectors.
1245  auto *IT = dyn_cast<IntegerType>(Op0->getType());
1246  if (IT && IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
1247  Metadata *LowAndHigh[] = {
1248  ConstantAsMetadata::get(ConstantInt::get(IT, DefiniteZeros)),
1249  ConstantAsMetadata::get(ConstantInt::get(IT, PossibleZeros + 1))};
1250  II.setMetadata(LLVMContext::MD_range,
1252  return &II;
1253  }
1254 
1255  return nullptr;
1256 }
1257 
1259  assert(II.getIntrinsicID() == Intrinsic::ctpop &&
1260  "Expected ctpop intrinsic");
1261  Value *Op0 = II.getArgOperand(0);
1262  Value *X;
1263  // ctpop(bitreverse(x)) -> ctpop(x)
1264  // ctpop(bswap(x)) -> ctpop(x)
1265  if (match(Op0, m_BitReverse(m_Value(X))) || match(Op0, m_BSwap(m_Value(X)))) {
1266  II.setOperand(0, X);
1267  return &II;
1268  }
1269 
1270  // FIXME: Try to simplify vectors of integers.
1271  auto *IT = dyn_cast<IntegerType>(Op0->getType());
1272  if (!IT)
1273  return nullptr;
1274 
1275  unsigned BitWidth = IT->getBitWidth();
1276  KnownBits Known(BitWidth);
1277  IC.computeKnownBits(Op0, Known, 0, &II);
1278 
1279  unsigned MinCount = Known.countMinPopulation();
1280  unsigned MaxCount = Known.countMaxPopulation();
1281 
1282  // Add range metadata since known bits can't completely reflect what we know.
1283  if (IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) {
1284  Metadata *LowAndHigh[] = {
1286  ConstantAsMetadata::get(ConstantInt::get(IT, MaxCount + 1))};
1287  II.setMetadata(LLVMContext::MD_range,
1289  return &II;
1290  }
1291 
1292  return nullptr;
1293 }
1294 
1295 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
1296 // XMM register mask efficiently, we could transform all x86 masked intrinsics
1297 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
1299  Value *Ptr = II.getOperand(0);
1300  Value *Mask = II.getOperand(1);
1301  Constant *ZeroVec = Constant::getNullValue(II.getType());
1302 
1303  // Special case a zero mask since that's not a ConstantDataVector.
1304  // This masked load instruction creates a zero vector.
1305  if (isa<ConstantAggregateZero>(Mask))
1306  return IC.replaceInstUsesWith(II, ZeroVec);
1307 
1308  auto *ConstMask = dyn_cast<ConstantDataVector>(Mask);
1309  if (!ConstMask)
1310  return nullptr;
1311 
1312  // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic
1313  // to allow target-independent optimizations.
1314 
1315  // First, cast the x86 intrinsic scalar pointer to a vector pointer to match
1316  // the LLVM intrinsic definition for the pointer argument.
1317  unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
1318  PointerType *VecPtrTy = PointerType::get(II.getType(), AddrSpace);
1319  Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
1320 
1321  // Second, convert the x86 XMM integer vector mask to a vector of bools based
1322  // on each element's most significant bit (the sign bit).
1323  Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask);
1324 
1325  // The pass-through vector for an x86 masked load is a zero vector.
1326  CallInst *NewMaskedLoad =
1327  IC.Builder.CreateMaskedLoad(PtrCast, 1, BoolMask, ZeroVec);
1328  return IC.replaceInstUsesWith(II, NewMaskedLoad);
1329 }
1330 
1331 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
1332 // XMM register mask efficiently, we could transform all x86 masked intrinsics
1333 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
1335  Value *Ptr = II.getOperand(0);
1336  Value *Mask = II.getOperand(1);
1337  Value *Vec = II.getOperand(2);
1338 
1339  // Special case a zero mask since that's not a ConstantDataVector:
1340  // this masked store instruction does nothing.
1341  if (isa<ConstantAggregateZero>(Mask)) {
1342  IC.eraseInstFromFunction(II);
1343  return true;
1344  }
1345 
1346  // The SSE2 version is too weird (eg, unaligned but non-temporal) to do
1347  // anything else at this level.
1348  if (II.getIntrinsicID() == Intrinsic::x86_sse2_maskmov_dqu)
1349  return false;
1350 
1351  auto *ConstMask = dyn_cast<ConstantDataVector>(Mask);
1352  if (!ConstMask)
1353  return false;
1354 
1355  // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic
1356  // to allow target-independent optimizations.
1357 
1358  // First, cast the x86 intrinsic scalar pointer to a vector pointer to match
1359  // the LLVM intrinsic definition for the pointer argument.
1360  unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
1361  PointerType *VecPtrTy = PointerType::get(Vec->getType(), AddrSpace);
1362  Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
1363 
1364  // Second, convert the x86 XMM integer vector mask to a vector of bools based
1365  // on each element's most significant bit (the sign bit).
1366  Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask);
1367 
1368  IC.Builder.CreateMaskedStore(Vec, PtrCast, 1, BoolMask);
1369 
1370  // 'Replace uses' doesn't work for stores. Erase the original masked store.
1371  IC.eraseInstFromFunction(II);
1372  return true;
1373 }
1374 
1375 // Constant fold llvm.amdgcn.fmed3 intrinsics for standard inputs.
1376 //
1377 // A single NaN input is folded to minnum, so we rely on that folding for
1378 // handling NaNs.
1379 static APFloat fmed3AMDGCN(const APFloat &Src0, const APFloat &Src1,
1380  const APFloat &Src2) {
1381  APFloat Max3 = maxnum(maxnum(Src0, Src1), Src2);
1382 
1383  APFloat::cmpResult Cmp0 = Max3.compare(Src0);
1384  assert(Cmp0 != APFloat::cmpUnordered && "nans handled separately");
1385  if (Cmp0 == APFloat::cmpEqual)
1386  return maxnum(Src1, Src2);
1387 
1388  APFloat::cmpResult Cmp1 = Max3.compare(Src1);
1389  assert(Cmp1 != APFloat::cmpUnordered && "nans handled separately");
1390  if (Cmp1 == APFloat::cmpEqual)
1391  return maxnum(Src0, Src2);
1392 
1393  return maxnum(Src0, Src1);
1394 }
1395 
1396 /// Convert a table lookup to shufflevector if the mask is constant.
1397 /// This could benefit tbl1 if the mask is { 7,6,5,4,3,2,1,0 }, in
1398 /// which case we could lower the shufflevector with rev64 instructions
1399 /// as it's actually a byte reverse.
1401  InstCombiner::BuilderTy &Builder) {
1402  // Bail out if the mask is not a constant.
1403  auto *C = dyn_cast<Constant>(II.getArgOperand(1));
1404  if (!C)
1405  return nullptr;
1406 
1407  auto *VecTy = cast<VectorType>(II.getType());
1408  unsigned NumElts = VecTy->getNumElements();
1409 
1410  // Only perform this transformation for <8 x i8> vector types.
1411  if (!VecTy->getElementType()->isIntegerTy(8) || NumElts != 8)
1412  return nullptr;
1413 
1414  uint32_t Indexes[8];
1415 
1416  for (unsigned I = 0; I < NumElts; ++I) {
1417  Constant *COp = C->getAggregateElement(I);
1418 
1419  if (!COp || !isa<ConstantInt>(COp))
1420  return nullptr;
1421 
1422  Indexes[I] = cast<ConstantInt>(COp)->getLimitedValue();
1423 
1424  // Make sure the mask indices are in range.
1425  if (Indexes[I] >= NumElts)
1426  return nullptr;
1427  }
1428 
1429  auto *ShuffleMask = ConstantDataVector::get(II.getContext(),
1430  makeArrayRef(Indexes));
1431  auto *V1 = II.getArgOperand(0);
1432  auto *V2 = Constant::getNullValue(V1->getType());
1433  return Builder.CreateShuffleVector(V1, V2, ShuffleMask);
1434 }
1435 
1436 /// Convert a vector load intrinsic into a simple llvm load instruction.
1437 /// This is beneficial when the underlying object being addressed comes
1438 /// from a constant, since we get constant-folding for free.
1440  unsigned MemAlign,
1441  InstCombiner::BuilderTy &Builder) {
1442  auto *IntrAlign = dyn_cast<ConstantInt>(II.getArgOperand(1));
1443 
1444  if (!IntrAlign)
1445  return nullptr;
1446 
1447  unsigned Alignment = IntrAlign->getLimitedValue() < MemAlign ?
1448  MemAlign : IntrAlign->getLimitedValue();
1449 
1450  if (!isPowerOf2_32(Alignment))
1451  return nullptr;
1452 
1453  auto *BCastInst = Builder.CreateBitCast(II.getArgOperand(0),
1454  PointerType::get(II.getType(), 0));
1455  return Builder.CreateAlignedLoad(II.getType(), BCastInst, Alignment);
1456 }
1457 
1458 // Returns true iff the 2 intrinsics have the same operands, limiting the
1459 // comparison to the first NumOperands.
1460 static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E,
1461  unsigned NumOperands) {
1462  assert(I.getNumArgOperands() >= NumOperands && "Not enough operands");
1463  assert(E.getNumArgOperands() >= NumOperands && "Not enough operands");
1464  for (unsigned i = 0; i < NumOperands; i++)
1465  if (I.getArgOperand(i) != E.getArgOperand(i))
1466  return false;
1467  return true;
1468 }
1469 
1470 // Remove trivially empty start/end intrinsic ranges, i.e. a start
1471 // immediately followed by an end (ignoring debuginfo or other
1472 // start/end intrinsics in between). As this handles only the most trivial
1473 // cases, tracking the nesting level is not needed:
1474 //
1475 // call @llvm.foo.start(i1 0) ; &I
1476 // call @llvm.foo.start(i1 0)
1477 // call @llvm.foo.end(i1 0) ; This one will not be skipped: it will be removed
1478 // call @llvm.foo.end(i1 0)
1479 static bool removeTriviallyEmptyRange(IntrinsicInst &I, unsigned StartID,
1480  unsigned EndID, InstCombiner &IC) {
1481  assert(I.getIntrinsicID() == StartID &&
1482  "Start intrinsic does not have expected ID");
1483  BasicBlock::iterator BI(I), BE(I.getParent()->end());
1484  for (++BI; BI != BE; ++BI) {
1485  if (auto *E = dyn_cast<IntrinsicInst>(BI)) {
1486  if (isa<DbgInfoIntrinsic>(E) || E->getIntrinsicID() == StartID)
1487  continue;
1488  if (E->getIntrinsicID() == EndID &&
1489  haveSameOperands(I, *E, E->getNumArgOperands())) {
1490  IC.eraseInstFromFunction(*E);
1491  IC.eraseInstFromFunction(I);
1492  return true;
1493  }
1494  }
1495  break;
1496  }
1497 
1498  return false;
1499 }
1500 
1501 // Convert NVVM intrinsics to target-generic LLVM code where possible.
1503  // Each NVVM intrinsic we can simplify can be replaced with one of:
1504  //
1505  // * an LLVM intrinsic,
1506  // * an LLVM cast operation,
1507  // * an LLVM binary operation, or
1508  // * ad-hoc LLVM IR for the particular operation.
1509 
1510  // Some transformations are only valid when the module's
1511  // flush-denormals-to-zero (ftz) setting is true/false, whereas other
1512  // transformations are valid regardless of the module's ftz setting.
1513  enum FtzRequirementTy {
1514  FTZ_Any, // Any ftz setting is ok.
1515  FTZ_MustBeOn, // Transformation is valid only if ftz is on.
1516  FTZ_MustBeOff, // Transformation is valid only if ftz is off.
1517  };
1518  // Classes of NVVM intrinsics that can't be replaced one-to-one with a
1519  // target-generic intrinsic, cast op, or binary op but that we can nonetheless
1520  // simplify.
1521  enum SpecialCase {
1522  SPC_Reciprocal,
1523  };
1524 
1525  // SimplifyAction is a poor-man's variant (plus an additional flag) that
1526  // represents how to replace an NVVM intrinsic with target-generic LLVM IR.
1527  struct SimplifyAction {
1528  // Invariant: At most one of these Optionals has a value.
1532  Optional<SpecialCase> Special;
1533 
1534  FtzRequirementTy FtzRequirement = FTZ_Any;
1535 
1536  SimplifyAction() = default;
1537 
1538  SimplifyAction(Intrinsic::ID IID, FtzRequirementTy FtzReq)
1539  : IID(IID), FtzRequirement(FtzReq) {}
1540 
1541  // Cast operations don't have anything to do with FTZ, so we skip that
1542  // argument.
1543  SimplifyAction(Instruction::CastOps CastOp) : CastOp(CastOp) {}
1544 
1545  SimplifyAction(Instruction::BinaryOps BinaryOp, FtzRequirementTy FtzReq)
1546  : BinaryOp(BinaryOp), FtzRequirement(FtzReq) {}
1547 
1548  SimplifyAction(SpecialCase Special, FtzRequirementTy FtzReq)
1549  : Special(Special), FtzRequirement(FtzReq) {}
1550  };
1551 
1552  // Try to generate a SimplifyAction describing how to replace our
1553  // IntrinsicInstr with target-generic LLVM IR.
1554  const SimplifyAction Action = [II]() -> SimplifyAction {
1555  switch (II->getIntrinsicID()) {
1556  // NVVM intrinsics that map directly to LLVM intrinsics.
1557  case Intrinsic::nvvm_ceil_d:
1558  return {Intrinsic::ceil, FTZ_Any};
1559  case Intrinsic::nvvm_ceil_f:
1560  return {Intrinsic::ceil, FTZ_MustBeOff};
1561  case Intrinsic::nvvm_ceil_ftz_f:
1562  return {Intrinsic::ceil, FTZ_MustBeOn};
1563  case Intrinsic::nvvm_fabs_d:
1564  return {Intrinsic::fabs, FTZ_Any};
1565  case Intrinsic::nvvm_fabs_f:
1566  return {Intrinsic::fabs, FTZ_MustBeOff};
1567  case Intrinsic::nvvm_fabs_ftz_f:
1568  return {Intrinsic::fabs, FTZ_MustBeOn};
1569  case Intrinsic::nvvm_floor_d:
1570  return {Intrinsic::floor, FTZ_Any};
1571  case Intrinsic::nvvm_floor_f:
1572  return {Intrinsic::floor, FTZ_MustBeOff};
1573  case Intrinsic::nvvm_floor_ftz_f:
1574  return {Intrinsic::floor, FTZ_MustBeOn};
1575  case Intrinsic::nvvm_fma_rn_d:
1576  return {Intrinsic::fma, FTZ_Any};
1577  case Intrinsic::nvvm_fma_rn_f:
1578  return {Intrinsic::fma, FTZ_MustBeOff};
1579  case Intrinsic::nvvm_fma_rn_ftz_f:
1580  return {Intrinsic::fma, FTZ_MustBeOn};
1581  case Intrinsic::nvvm_fmax_d:
1582  return {Intrinsic::maxnum, FTZ_Any};
1583  case Intrinsic::nvvm_fmax_f:
1584  return {Intrinsic::maxnum, FTZ_MustBeOff};
1585  case Intrinsic::nvvm_fmax_ftz_f:
1586  return {Intrinsic::maxnum, FTZ_MustBeOn};
1587  case Intrinsic::nvvm_fmin_d:
1588  return {Intrinsic::minnum, FTZ_Any};
1589  case Intrinsic::nvvm_fmin_f:
1590  return {Intrinsic::minnum, FTZ_MustBeOff};
1591  case Intrinsic::nvvm_fmin_ftz_f:
1592  return {Intrinsic::minnum, FTZ_MustBeOn};
1593  case Intrinsic::nvvm_round_d:
1594  return {Intrinsic::round, FTZ_Any};
1595  case Intrinsic::nvvm_round_f:
1596  return {Intrinsic::round, FTZ_MustBeOff};
1597  case Intrinsic::nvvm_round_ftz_f:
1598  return {Intrinsic::round, FTZ_MustBeOn};
1599  case Intrinsic::nvvm_sqrt_rn_d:
1600  return {Intrinsic::sqrt, FTZ_Any};
1601  case Intrinsic::nvvm_sqrt_f:
1602  // nvvm_sqrt_f is a special case. For most intrinsics, foo_ftz_f is the
1603  // ftz version, and foo_f is the non-ftz version. But nvvm_sqrt_f adopts
1604  // the ftz-ness of the surrounding code. sqrt_rn_f and sqrt_rn_ftz_f are
1605  // the versions with explicit ftz-ness.
1606  return {Intrinsic::sqrt, FTZ_Any};
1607  case Intrinsic::nvvm_sqrt_rn_f:
1608  return {Intrinsic::sqrt, FTZ_MustBeOff};
1609  case Intrinsic::nvvm_sqrt_rn_ftz_f:
1610  return {Intrinsic::sqrt, FTZ_MustBeOn};
1611  case Intrinsic::nvvm_trunc_d:
1612  return {Intrinsic::trunc, FTZ_Any};
1613  case Intrinsic::nvvm_trunc_f:
1614  return {Intrinsic::trunc, FTZ_MustBeOff};
1615  case Intrinsic::nvvm_trunc_ftz_f:
1616  return {Intrinsic::trunc, FTZ_MustBeOn};
1617 
1618  // NVVM intrinsics that map to LLVM cast operations.
1619  //
1620  // Note that llvm's target-generic conversion operators correspond to the rz
1621  // (round to zero) versions of the nvvm conversion intrinsics, even though
1622  // most everything else here uses the rn (round to nearest even) nvvm ops.
1623  case Intrinsic::nvvm_d2i_rz:
1624  case Intrinsic::nvvm_f2i_rz:
1625  case Intrinsic::nvvm_d2ll_rz:
1626  case Intrinsic::nvvm_f2ll_rz:
1627  return {Instruction::FPToSI};
1628  case Intrinsic::nvvm_d2ui_rz:
1629  case Intrinsic::nvvm_f2ui_rz:
1630  case Intrinsic::nvvm_d2ull_rz:
1631  case Intrinsic::nvvm_f2ull_rz:
1632  return {Instruction::FPToUI};
1633  case Intrinsic::nvvm_i2d_rz:
1634  case Intrinsic::nvvm_i2f_rz:
1635  case Intrinsic::nvvm_ll2d_rz:
1636  case Intrinsic::nvvm_ll2f_rz:
1637  return {Instruction::SIToFP};
1638  case Intrinsic::nvvm_ui2d_rz:
1639  case Intrinsic::nvvm_ui2f_rz:
1640  case Intrinsic::nvvm_ull2d_rz:
1641  case Intrinsic::nvvm_ull2f_rz:
1642  return {Instruction::UIToFP};
1643 
1644  // NVVM intrinsics that map to LLVM binary ops.
1645  case Intrinsic::nvvm_add_rn_d:
1646  return {Instruction::FAdd, FTZ_Any};
1647  case Intrinsic::nvvm_add_rn_f:
1648  return {Instruction::FAdd, FTZ_MustBeOff};
1649  case Intrinsic::nvvm_add_rn_ftz_f:
1650  return {Instruction::FAdd, FTZ_MustBeOn};
1651  case Intrinsic::nvvm_mul_rn_d:
1652  return {Instruction::FMul, FTZ_Any};
1653  case Intrinsic::nvvm_mul_rn_f:
1654  return {Instruction::FMul, FTZ_MustBeOff};
1655  case Intrinsic::nvvm_mul_rn_ftz_f:
1656  return {Instruction::FMul, FTZ_MustBeOn};
1657  case Intrinsic::nvvm_div_rn_d:
1658  return {Instruction::FDiv, FTZ_Any};
1659  case Intrinsic::nvvm_div_rn_f:
1660  return {Instruction::FDiv, FTZ_MustBeOff};
1661  case Intrinsic::nvvm_div_rn_ftz_f:
1662  return {Instruction::FDiv, FTZ_MustBeOn};
1663 
1664  // The remainder of cases are NVVM intrinsics that map to LLVM idioms, but
1665  // need special handling.
1666  //
1667  // We seem to be missing intrinsics for rcp.approx.{ftz.}f32, which is just
1668  // as well.
1669  case Intrinsic::nvvm_rcp_rn_d:
1670  return {SPC_Reciprocal, FTZ_Any};
1671  case Intrinsic::nvvm_rcp_rn_f:
1672  return {SPC_Reciprocal, FTZ_MustBeOff};
1673  case Intrinsic::nvvm_rcp_rn_ftz_f:
1674  return {SPC_Reciprocal, FTZ_MustBeOn};
1675 
1676  // We do not currently simplify intrinsics that give an approximate answer.
1677  // These include:
1678  //
1679  // - nvvm_cos_approx_{f,ftz_f}
1680  // - nvvm_ex2_approx_{d,f,ftz_f}
1681  // - nvvm_lg2_approx_{d,f,ftz_f}
1682  // - nvvm_sin_approx_{f,ftz_f}
1683  // - nvvm_sqrt_approx_{f,ftz_f}
1684  // - nvvm_rsqrt_approx_{d,f,ftz_f}
1685  // - nvvm_div_approx_{ftz_d,ftz_f,f}
1686  // - nvvm_rcp_approx_ftz_d
1687  //
1688  // Ideally we'd encode them as e.g. "fast call @llvm.cos", where "fast"
1689  // means that fastmath is enabled in the intrinsic. Unfortunately only
1690  // binary operators (currently) have a fastmath bit in SelectionDAG, so this
1691  // information gets lost and we can't select on it.
1692  //
1693  // TODO: div and rcp are lowered to a binary op, so these we could in theory
1694  // lower them to "fast fdiv".
1695 
1696  default:
1697  return {};
1698  }
1699  }();
1700 
1701  // If Action.FtzRequirementTy is not satisfied by the module's ftz state, we
1702  // can bail out now. (Notice that in the case that IID is not an NVVM
1703  // intrinsic, we don't have to look up any module metadata, as
1704  // FtzRequirementTy will be FTZ_Any.)
1705  if (Action.FtzRequirement != FTZ_Any) {
1706  bool FtzEnabled =
1707  II->getFunction()->getFnAttribute("nvptx-f32ftz").getValueAsString() ==
1708  "true";
1709 
1710  if (FtzEnabled != (Action.FtzRequirement == FTZ_MustBeOn))
1711  return nullptr;
1712  }
1713 
1714  // Simplify to target-generic intrinsic.
1715  if (Action.IID) {
1717  // All the target-generic intrinsics currently of interest to us have one
1718  // type argument, equal to that of the nvvm intrinsic's argument.
1719  Type *Tys[] = {II->getArgOperand(0)->getType()};
1720  return CallInst::Create(
1721  Intrinsic::getDeclaration(II->getModule(), *Action.IID, Tys), Args);
1722  }
1723 
1724  // Simplify to target-generic binary op.
1725  if (Action.BinaryOp)
1726  return BinaryOperator::Create(*Action.BinaryOp, II->getArgOperand(0),
1727  II->getArgOperand(1), II->getName());
1728 
1729  // Simplify to target-generic cast op.
1730  if (Action.CastOp)
1731  return CastInst::Create(*Action.CastOp, II->getArgOperand(0), II->getType(),
1732  II->getName());
1733 
1734  // All that's left are the special cases.
1735  if (!Action.Special)
1736  return nullptr;
1737 
1738  switch (*Action.Special) {
1739  case SPC_Reciprocal:
1740  // Simplify reciprocal.
1741  return BinaryOperator::Create(
1742  Instruction::FDiv, ConstantFP::get(II->getArgOperand(0)->getType(), 1),
1743  II->getArgOperand(0), II->getName());
1744  }
1745  llvm_unreachable("All SpecialCase enumerators should be handled in switch.");
1746 }
1747 
1749  removeTriviallyEmptyRange(I, Intrinsic::vastart, Intrinsic::vaend, *this);
1750  return nullptr;
1751 }
1752 
1754  removeTriviallyEmptyRange(I, Intrinsic::vacopy, Intrinsic::vaend, *this);
1755  return nullptr;
1756 }
1757 
1759  assert(Call.getNumArgOperands() > 1 && "Need at least 2 args to swap");
1760  Value *Arg0 = Call.getArgOperand(0), *Arg1 = Call.getArgOperand(1);
1761  if (isa<Constant>(Arg0) && !isa<Constant>(Arg1)) {
1762  Call.setArgOperand(0, Arg1);
1763  Call.setArgOperand(1, Arg0);
1764  return &Call;
1765  }
1766  return nullptr;
1767 }
1768 
1769 Instruction *InstCombiner::foldIntrinsicWithOverflowCommon(IntrinsicInst *II) {
1770  WithOverflowInst *WO = cast<WithOverflowInst>(II);
1771  Value *OperationResult = nullptr;
1772  Constant *OverflowResult = nullptr;
1773  if (OptimizeOverflowCheck(WO->getBinaryOp(), WO->isSigned(), WO->getLHS(),
1774  WO->getRHS(), *WO, OperationResult, OverflowResult))
1775  return CreateOverflowTuple(WO, OperationResult, OverflowResult);
1776  return nullptr;
1777 }
1778 
1779 /// CallInst simplification. This mostly only handles folding of intrinsic
1780 /// instructions. For normal calls, it allows visitCallBase to do the heavy
1781 /// lifting.
1783  if (Value *V = SimplifyCall(&CI, SQ.getWithInstruction(&CI)))
1784  return replaceInstUsesWith(CI, V);
1785 
1786  if (isFreeCall(&CI, &TLI))
1787  return visitFree(CI);
1788 
1789  // If the caller function is nounwind, mark the call as nounwind, even if the
1790  // callee isn't.
1791  if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) {
1792  CI.setDoesNotThrow();
1793  return &CI;
1794  }
1795 
1796  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
1797  if (!II) return visitCallBase(CI);
1798 
1799  // Intrinsics cannot occur in an invoke or a callbr, so handle them here
1800  // instead of in visitCallBase.
1801  if (auto *MI = dyn_cast<AnyMemIntrinsic>(II)) {
1802  bool Changed = false;
1803 
1804  // memmove/cpy/set of zero bytes is a noop.
1805  if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
1806  if (NumBytes->isNullValue())
1807  return eraseInstFromFunction(CI);
1808 
1809  if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
1810  if (CI->getZExtValue() == 1) {
1811  // Replace the instruction with just byte operations. We would
1812  // transform other cases to loads/stores, but we don't know if
1813  // alignment is sufficient.
1814  }
1815  }
1816 
1817  // No other transformations apply to volatile transfers.
1818  if (auto *M = dyn_cast<MemIntrinsic>(MI))
1819  if (M->isVolatile())
1820  return nullptr;
1821 
1822  // If we have a memmove and the source operation is a constant global,
1823  // then the source and dest pointers can't alias, so we can change this
1824  // into a call to memcpy.
1825  if (auto *MMI = dyn_cast<AnyMemMoveInst>(MI)) {
1826  if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
1827  if (GVSrc->isConstant()) {
1828  Module *M = CI.getModule();
1829  Intrinsic::ID MemCpyID =
1830  isa<AtomicMemMoveInst>(MMI)
1831  ? Intrinsic::memcpy_element_unordered_atomic
1832  : Intrinsic::memcpy;
1833  Type *Tys[3] = { CI.getArgOperand(0)->getType(),
1834  CI.getArgOperand(1)->getType(),
1835  CI.getArgOperand(2)->getType() };
1836  CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
1837  Changed = true;
1838  }
1839  }
1840 
1841  if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1842  // memmove(x,x,size) -> noop.
1843  if (MTI->getSource() == MTI->getDest())
1844  return eraseInstFromFunction(CI);
1845  }
1846 
1847  // If we can determine a pointer alignment that is bigger than currently
1848  // set, update the alignment.
1849  if (auto *MTI = dyn_cast<AnyMemTransferInst>(MI)) {
1850  if (Instruction *I = SimplifyAnyMemTransfer(MTI))
1851  return I;
1852  } else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) {
1853  if (Instruction *I = SimplifyAnyMemSet(MSI))
1854  return I;
1855  }
1856 
1857  if (Changed) return II;
1858  }
1859 
1860  // For vector result intrinsics, use the generic demanded vector support.
1861  if (II->getType()->isVectorTy()) {
1862  auto VWidth = II->getType()->getVectorNumElements();
1863  APInt UndefElts(VWidth, 0);
1864  APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1865  if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) {
1866  if (V != II)
1867  return replaceInstUsesWith(*II, V);
1868  return II;
1869  }
1870  }
1871 
1872  if (Instruction *I = SimplifyNVVMIntrinsic(II, *this))
1873  return I;
1874 
1875  auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width,
1876  unsigned DemandedWidth) {
1877  APInt UndefElts(Width, 0);
1878  APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
1879  return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
1880  };
1881 
1882  Intrinsic::ID IID = II->getIntrinsicID();
1883  switch (IID) {
1884  default: break;
1885  case Intrinsic::objectsize:
1886  if (Value *V = lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/false))
1887  return replaceInstUsesWith(CI, V);
1888  return nullptr;
1889  case Intrinsic::bswap: {
1890  Value *IIOperand = II->getArgOperand(0);
1891  Value *X = nullptr;
1892 
1893  // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
1894  if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
1895  unsigned C = X->getType()->getPrimitiveSizeInBits() -
1896  IIOperand->getType()->getPrimitiveSizeInBits();
1897  Value *CV = ConstantInt::get(X->getType(), C);
1898  Value *V = Builder.CreateLShr(X, CV);
1899  return new TruncInst(V, IIOperand->getType());
1900  }
1901  break;
1902  }
1903  case Intrinsic::masked_load:
1904  if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(*II))
1905  return replaceInstUsesWith(CI, SimplifiedMaskedOp);
1906  break;
1907  case Intrinsic::masked_store:
1908  return simplifyMaskedStore(*II);
1909  case Intrinsic::masked_gather:
1910  return simplifyMaskedGather(*II);
1911  case Intrinsic::masked_scatter:
1912  return simplifyMaskedScatter(*II);
1913  case Intrinsic::launder_invariant_group:
1914  case Intrinsic::strip_invariant_group:
1915  if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(*II, *this))
1916  return replaceInstUsesWith(*II, SkippedBarrier);
1917  break;
1918  case Intrinsic::powi:
1919  if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
1920  // 0 and 1 are handled in instsimplify
1921 
1922  // powi(x, -1) -> 1/x
1923  if (Power->isMinusOne())
1924  return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
1925  II->getArgOperand(0));
1926  // powi(x, 2) -> x*x
1927  if (Power->equalsInt(2))
1928  return BinaryOperator::CreateFMul(II->getArgOperand(0),
1929  II->getArgOperand(0));
1930  }
1931  break;
1932 
1933  case Intrinsic::cttz:
1934  case Intrinsic::ctlz:
1935  if (auto *I = foldCttzCtlz(*II, *this))
1936  return I;
1937  break;
1938 
1939  case Intrinsic::ctpop:
1940  if (auto *I = foldCtpop(*II, *this))
1941  return I;
1942  break;
1943 
1944  case Intrinsic::fshl:
1945  case Intrinsic::fshr: {
1946  Value *Op0 = II->getArgOperand(0), *Op1 = II->getArgOperand(1);
1947  Type *Ty = II->getType();
1948  unsigned BitWidth = Ty->getScalarSizeInBits();
1949  Constant *ShAmtC;
1950  if (match(II->getArgOperand(2), m_Constant(ShAmtC)) &&
1951  !isa<ConstantExpr>(ShAmtC) && !ShAmtC->containsConstantExpression()) {
1952  // Canonicalize a shift amount constant operand to modulo the bit-width.
1953  Constant *WidthC = ConstantInt::get(Ty, BitWidth);
1954  Constant *ModuloC = ConstantExpr::getURem(ShAmtC, WidthC);
1955  if (ModuloC != ShAmtC) {
1956  II->setArgOperand(2, ModuloC);
1957  return II;
1958  }
1959  assert(ConstantExpr::getICmp(ICmpInst::ICMP_UGT, WidthC, ShAmtC) ==
1961  "Shift amount expected to be modulo bitwidth");
1962 
1963  // Canonicalize funnel shift right by constant to funnel shift left. This
1964  // is not entirely arbitrary. For historical reasons, the backend may
1965  // recognize rotate left patterns but miss rotate right patterns.
1966  if (IID == Intrinsic::fshr) {
1967  // fshr X, Y, C --> fshl X, Y, (BitWidth - C)
1968  Constant *LeftShiftC = ConstantExpr::getSub(WidthC, ShAmtC);
1969  Module *Mod = II->getModule();
1970  Function *Fshl = Intrinsic::getDeclaration(Mod, Intrinsic::fshl, Ty);
1971  return CallInst::Create(Fshl, { Op0, Op1, LeftShiftC });
1972  }
1973  assert(IID == Intrinsic::fshl &&
1974  "All funnel shifts by simple constants should go left");
1975 
1976  // fshl(X, 0, C) --> shl X, C
1977  // fshl(X, undef, C) --> shl X, C
1978  if (match(Op1, m_ZeroInt()) || match(Op1, m_Undef()))
1979  return BinaryOperator::CreateShl(Op0, ShAmtC);
1980 
1981  // fshl(0, X, C) --> lshr X, (BW-C)
1982  // fshl(undef, X, C) --> lshr X, (BW-C)
1983  if (match(Op0, m_ZeroInt()) || match(Op0, m_Undef()))
1984  return BinaryOperator::CreateLShr(Op1,
1985  ConstantExpr::getSub(WidthC, ShAmtC));
1986 
1987  // fshl i16 X, X, 8 --> bswap i16 X (reduce to more-specific form)
1988  if (Op0 == Op1 && BitWidth == 16 && match(ShAmtC, m_SpecificInt(8))) {
1989  Module *Mod = II->getModule();
1990  Function *Bswap = Intrinsic::getDeclaration(Mod, Intrinsic::bswap, Ty);
1991  return CallInst::Create(Bswap, { Op0 });
1992  }
1993  }
1994 
1995  // Left or right might be masked.
1996  if (SimplifyDemandedInstructionBits(*II))
1997  return &CI;
1998 
1999  // The shift amount (operand 2) of a funnel shift is modulo the bitwidth,
2000  // so only the low bits of the shift amount are demanded if the bitwidth is
2001  // a power-of-2.
2002  if (!isPowerOf2_32(BitWidth))
2003  break;
2004  APInt Op2Demanded = APInt::getLowBitsSet(BitWidth, Log2_32_Ceil(BitWidth));
2005  KnownBits Op2Known(BitWidth);
2006  if (SimplifyDemandedBits(II, 2, Op2Demanded, Op2Known))
2007  return &CI;
2008  break;
2009  }
2010  case Intrinsic::uadd_with_overflow:
2011  case Intrinsic::sadd_with_overflow: {
2013  return I;
2014  if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2015  return I;
2016 
2017  // Given 2 constant operands whose sum does not overflow:
2018  // uaddo (X +nuw C0), C1 -> uaddo X, C0 + C1
2019  // saddo (X +nsw C0), C1 -> saddo X, C0 + C1
2020  Value *X;
2021  const APInt *C0, *C1;
2022  Value *Arg0 = II->getArgOperand(0);
2023  Value *Arg1 = II->getArgOperand(1);
2024  bool IsSigned = IID == Intrinsic::sadd_with_overflow;
2025  bool HasNWAdd = IsSigned ? match(Arg0, m_NSWAdd(m_Value(X), m_APInt(C0)))
2026  : match(Arg0, m_NUWAdd(m_Value(X), m_APInt(C0)));
2027  if (HasNWAdd && match(Arg1, m_APInt(C1))) {
2028  bool Overflow;
2029  APInt NewC =
2030  IsSigned ? C1->sadd_ov(*C0, Overflow) : C1->uadd_ov(*C0, Overflow);
2031  if (!Overflow)
2032  return replaceInstUsesWith(
2033  *II, Builder.CreateBinaryIntrinsic(
2034  IID, X, ConstantInt::get(Arg1->getType(), NewC)));
2035  }
2036  break;
2037  }
2038 
2039  case Intrinsic::umul_with_overflow:
2040  case Intrinsic::smul_with_overflow:
2042  return I;
2044 
2045  case Intrinsic::usub_with_overflow:
2046  if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2047  return I;
2048  break;
2049 
2050  case Intrinsic::ssub_with_overflow: {
2051  if (Instruction *I = foldIntrinsicWithOverflowCommon(II))
2052  return I;
2053 
2054  Constant *C;
2055  Value *Arg0 = II->getArgOperand(0);
2056  Value *Arg1 = II->getArgOperand(1);
2057  // Given a constant C that is not the minimum signed value
2058  // for an integer of a given bit width:
2059  //
2060  // ssubo X, C -> saddo X, -C
2061  if (match(Arg1, m_Constant(C)) && C->isNotMinSignedValue()) {
2062  Value *NegVal = ConstantExpr::getNeg(C);
2063  // Build a saddo call that is equivalent to the discovered
2064  // ssubo call.
2065  return replaceInstUsesWith(
2066  *II, Builder.CreateBinaryIntrinsic(Intrinsic::sadd_with_overflow,
2067  Arg0, NegVal));
2068  }
2069 
2070  break;
2071  }
2072 
2073  case Intrinsic::uadd_sat:
2074  case Intrinsic::sadd_sat:
2076  return I;
2078  case Intrinsic::usub_sat:
2079  case Intrinsic::ssub_sat: {
2080  SaturatingInst *SI = cast<SaturatingInst>(II);
2081  Type *Ty = SI->getType();
2082  Value *Arg0 = SI->getLHS();
2083  Value *Arg1 = SI->getRHS();
2084 
2085  // Make use of known overflow information.
2086  OverflowResult OR = computeOverflow(SI->getBinaryOp(), SI->isSigned(),
2087  Arg0, Arg1, SI);
2088  switch (OR) {
2090  break;
2092  if (SI->isSigned())
2093  return BinaryOperator::CreateNSW(SI->getBinaryOp(), Arg0, Arg1);
2094  else
2095  return BinaryOperator::CreateNUW(SI->getBinaryOp(), Arg0, Arg1);
2097  unsigned BitWidth = Ty->getScalarSizeInBits();
2098  APInt Min = APSInt::getMinValue(BitWidth, !SI->isSigned());
2099  return replaceInstUsesWith(*SI, ConstantInt::get(Ty, Min));
2100  }
2102  unsigned BitWidth = Ty->getScalarSizeInBits();
2103  APInt Max = APSInt::getMaxValue(BitWidth, !SI->isSigned());
2104  return replaceInstUsesWith(*SI, ConstantInt::get(Ty, Max));
2105  }
2106  }
2107 
2108  // ssub.sat(X, C) -> sadd.sat(X, -C) if C != MIN
2109  Constant *C;
2110  if (IID == Intrinsic::ssub_sat && match(Arg1, m_Constant(C)) &&
2111  C->isNotMinSignedValue()) {
2112  Value *NegVal = ConstantExpr::getNeg(C);
2113  return replaceInstUsesWith(
2114  *II, Builder.CreateBinaryIntrinsic(
2115  Intrinsic::sadd_sat, Arg0, NegVal));
2116  }
2117 
2118  // sat(sat(X + Val2) + Val) -> sat(X + (Val+Val2))
2119  // sat(sat(X - Val2) - Val) -> sat(X - (Val+Val2))
2120  // if Val and Val2 have the same sign
2121  if (auto *Other = dyn_cast<IntrinsicInst>(Arg0)) {
2122  Value *X;
2123  const APInt *Val, *Val2;
2124  APInt NewVal;
2125  bool IsUnsigned =
2126  IID == Intrinsic::uadd_sat || IID == Intrinsic::usub_sat;
2127  if (Other->getIntrinsicID() == IID &&
2128  match(Arg1, m_APInt(Val)) &&
2129  match(Other->getArgOperand(0), m_Value(X)) &&
2130  match(Other->getArgOperand(1), m_APInt(Val2))) {
2131  if (IsUnsigned)
2132  NewVal = Val->uadd_sat(*Val2);
2133  else if (Val->isNonNegative() == Val2->isNonNegative()) {
2134  bool Overflow;
2135  NewVal = Val->sadd_ov(*Val2, Overflow);
2136  if (Overflow) {
2137  // Both adds together may add more than SignedMaxValue
2138  // without saturating the final result.
2139  break;
2140  }
2141  } else {
2142  // Cannot fold saturated addition with different signs.
2143  break;
2144  }
2145 
2146  return replaceInstUsesWith(
2147  *II, Builder.CreateBinaryIntrinsic(
2148  IID, X, ConstantInt::get(II->getType(), NewVal)));
2149  }
2150  }
2151  break;
2152  }
2153 
2154  case Intrinsic::minnum:
2155  case Intrinsic::maxnum:
2156  case Intrinsic::minimum:
2157  case Intrinsic::maximum: {
2159  return I;
2160  Value *Arg0 = II->getArgOperand(0);
2161  Value *Arg1 = II->getArgOperand(1);
2162  Value *X, *Y;
2163  if (match(Arg0, m_FNeg(m_Value(X))) && match(Arg1, m_FNeg(m_Value(Y))) &&
2164  (Arg0->hasOneUse() || Arg1->hasOneUse())) {
2165  // If both operands are negated, invert the call and negate the result:
2166  // min(-X, -Y) --> -(max(X, Y))
2167  // max(-X, -Y) --> -(min(X, Y))
2168  Intrinsic::ID NewIID;
2169  switch (IID) {
2170  case Intrinsic::maxnum:
2171  NewIID = Intrinsic::minnum;
2172  break;
2173  case Intrinsic::minnum:
2174  NewIID = Intrinsic::maxnum;
2175  break;
2176  case Intrinsic::maximum:
2177  NewIID = Intrinsic::minimum;
2178  break;
2179  case Intrinsic::minimum:
2180  NewIID = Intrinsic::maximum;
2181  break;
2182  default:
2183  llvm_unreachable("unexpected intrinsic ID");
2184  }
2185  Value *NewCall = Builder.CreateBinaryIntrinsic(NewIID, X, Y, II);
2186  Instruction *FNeg = BinaryOperator::CreateFNeg(NewCall);
2187  FNeg->copyIRFlags(II);
2188  return FNeg;
2189  }
2190 
2191  // m(m(X, C2), C1) -> m(X, C)
2192  const APFloat *C1, *C2;
2193  if (auto *M = dyn_cast<IntrinsicInst>(Arg0)) {
2194  if (M->getIntrinsicID() == IID && match(Arg1, m_APFloat(C1)) &&
2195  ((match(M->getArgOperand(0), m_Value(X)) &&
2196  match(M->getArgOperand(1), m_APFloat(C2))) ||
2197  (match(M->getArgOperand(1), m_Value(X)) &&
2198  match(M->getArgOperand(0), m_APFloat(C2))))) {
2199  APFloat Res(0.0);
2200  switch (IID) {
2201  case Intrinsic::maxnum:
2202  Res = maxnum(*C1, *C2);
2203  break;
2204  case Intrinsic::minnum:
2205  Res = minnum(*C1, *C2);
2206  break;
2207  case Intrinsic::maximum:
2208  Res = maximum(*C1, *C2);
2209  break;
2210  case Intrinsic::minimum:
2211  Res = minimum(*C1, *C2);
2212  break;
2213  default:
2214  llvm_unreachable("unexpected intrinsic ID");
2215  }
2216  Instruction *NewCall = Builder.CreateBinaryIntrinsic(
2217  IID, X, ConstantFP::get(Arg0->getType(), Res));
2218  NewCall->copyIRFlags(II);
2219  return replaceInstUsesWith(*II, NewCall);
2220  }
2221  }
2222 
2223  break;
2224  }
2225  case Intrinsic::fmuladd: {
2226  // Canonicalize fast fmuladd to the separate fmul + fadd.
2227  if (II->isFast()) {
2228  BuilderTy::FastMathFlagGuard Guard(Builder);
2229  Builder.setFastMathFlags(II->getFastMathFlags());
2230  Value *Mul = Builder.CreateFMul(II->getArgOperand(0),
2231  II->getArgOperand(1));
2232  Value *Add = Builder.CreateFAdd(Mul, II->getArgOperand(2));
2233  Add->takeName(II);
2234  return replaceInstUsesWith(*II, Add);
2235  }
2236 
2237  // Try to simplify the underlying FMul.
2238  if (Value *V = SimplifyFMulInst(II->getArgOperand(0), II->getArgOperand(1),
2239  II->getFastMathFlags(),
2240  SQ.getWithInstruction(II))) {
2241  auto *FAdd = BinaryOperator::CreateFAdd(V, II->getArgOperand(2));
2242  FAdd->copyFastMathFlags(II);
2243  return FAdd;
2244  }
2245 
2247  }
2248  case Intrinsic::fma: {
2250  return I;
2251 
2252  // fma fneg(x), fneg(y), z -> fma x, y, z
2253  Value *Src0 = II->getArgOperand(0);
2254  Value *Src1 = II->getArgOperand(1);
2255  Value *X, *Y;
2256  if (match(Src0, m_FNeg(m_Value(X))) && match(Src1, m_FNeg(m_Value(Y)))) {
2257  II->setArgOperand(0, X);
2258  II->setArgOperand(1, Y);
2259  return II;
2260  }
2261 
2262  // fma fabs(x), fabs(x), z -> fma x, x, z
2263  if (match(Src0, m_FAbs(m_Value(X))) &&
2264  match(Src1, m_FAbs(m_Specific(X)))) {
2265  II->setArgOperand(0, X);
2266  II->setArgOperand(1, X);
2267  return II;
2268  }
2269 
2270  // Try to simplify the underlying FMul. We can only apply simplifications
2271  // that do not require rounding.
2272  if (Value *V = SimplifyFMAFMul(II->getArgOperand(0), II->getArgOperand(1),
2273  II->getFastMathFlags(),
2274  SQ.getWithInstruction(II))) {
2275  auto *FAdd = BinaryOperator::CreateFAdd(V, II->getArgOperand(2));
2276  FAdd->copyFastMathFlags(II);
2277  return FAdd;
2278  }
2279 
2280  break;
2281  }
2282  case Intrinsic::fabs: {
2283  Value *Cond;
2284  Constant *LHS, *RHS;
2285  if (match(II->getArgOperand(0),
2286  m_Select(m_Value(Cond), m_Constant(LHS), m_Constant(RHS)))) {
2287  CallInst *Call0 = Builder.CreateCall(II->getCalledFunction(), {LHS});
2288  CallInst *Call1 = Builder.CreateCall(II->getCalledFunction(), {RHS});
2289  return SelectInst::Create(Cond, Call0, Call1);
2290  }
2291 
2293  }
2294  case Intrinsic::ceil:
2295  case Intrinsic::floor:
2296  case Intrinsic::round:
2297  case Intrinsic::nearbyint:
2298  case Intrinsic::rint:
2299  case Intrinsic::trunc: {
2300  Value *ExtSrc;
2301  if (match(II->getArgOperand(0), m_OneUse(m_FPExt(m_Value(ExtSrc))))) {
2302  // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x)
2303  Value *NarrowII = Builder.CreateUnaryIntrinsic(IID, ExtSrc, II);
2304  return new FPExtInst(NarrowII, II->getType());
2305  }
2306  break;
2307  }
2308  case Intrinsic::cos:
2309  case Intrinsic::amdgcn_cos: {
2310  Value *X;
2311  Value *Src = II->getArgOperand(0);
2312  if (match(Src, m_FNeg(m_Value(X))) || match(Src, m_FAbs(m_Value(X)))) {
2313  // cos(-x) -> cos(x)
2314  // cos(fabs(x)) -> cos(x)
2315  II->setArgOperand(0, X);
2316  return II;
2317  }
2318  break;
2319  }
2320  case Intrinsic::sin: {
2321  Value *X;
2322  if (match(II->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X))))) {
2323  // sin(-x) --> -sin(x)
2324  Value *NewSin = Builder.CreateUnaryIntrinsic(Intrinsic::sin, X, II);
2325  Instruction *FNeg = BinaryOperator::CreateFNeg(NewSin);
2326  FNeg->copyFastMathFlags(II);
2327  return FNeg;
2328  }
2329  break;
2330  }
2331  case Intrinsic::ppc_altivec_lvx:
2332  case Intrinsic::ppc_altivec_lvxl:
2333  // Turn PPC lvx -> load if the pointer is known aligned.
2334  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC,
2335  &DT) >= 16) {
2336  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2337  PointerType::getUnqual(II->getType()));
2338  return new LoadInst(II->getType(), Ptr);
2339  }
2340  break;
2341  case Intrinsic::ppc_vsx_lxvw4x:
2342  case Intrinsic::ppc_vsx_lxvd2x: {
2343  // Turn PPC VSX loads into normal loads.
2344  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2345  PointerType::getUnqual(II->getType()));
2346  return new LoadInst(II->getType(), Ptr, Twine(""), false, Align::None());
2347  }
2348  case Intrinsic::ppc_altivec_stvx:
2349  case Intrinsic::ppc_altivec_stvxl:
2350  // Turn stvx -> store if the pointer is known aligned.
2351  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC,
2352  &DT) >= 16) {
2353  Type *OpPtrTy =
2354  PointerType::getUnqual(II->getArgOperand(0)->getType());
2355  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2356  return new StoreInst(II->getArgOperand(0), Ptr);
2357  }
2358  break;
2359  case Intrinsic::ppc_vsx_stxvw4x:
2360  case Intrinsic::ppc_vsx_stxvd2x: {
2361  // Turn PPC VSX stores into normal stores.
2362  Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType());
2363  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2364  return new StoreInst(II->getArgOperand(0), Ptr, false, Align::None());
2365  }
2366  case Intrinsic::ppc_qpx_qvlfs:
2367  // Turn PPC QPX qvlfs -> load if the pointer is known aligned.
2368  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC,
2369  &DT) >= 16) {
2370  Type *VTy = VectorType::get(Builder.getFloatTy(),
2371  II->getType()->getVectorNumElements());
2372  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2373  PointerType::getUnqual(VTy));
2374  Value *Load = Builder.CreateLoad(VTy, Ptr);
2375  return new FPExtInst(Load, II->getType());
2376  }
2377  break;
2378  case Intrinsic::ppc_qpx_qvlfd:
2379  // Turn PPC QPX qvlfd -> load if the pointer is known aligned.
2380  if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, &AC,
2381  &DT) >= 32) {
2382  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0),
2383  PointerType::getUnqual(II->getType()));
2384  return new LoadInst(II->getType(), Ptr);
2385  }
2386  break;
2387  case Intrinsic::ppc_qpx_qvstfs:
2388  // Turn PPC QPX qvstfs -> store if the pointer is known aligned.
2389  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC,
2390  &DT) >= 16) {
2391  Type *VTy = VectorType::get(Builder.getFloatTy(),
2392  II->getArgOperand(0)->getType()->getVectorNumElements());
2393  Value *TOp = Builder.CreateFPTrunc(II->getArgOperand(0), VTy);
2394  Type *OpPtrTy = PointerType::getUnqual(VTy);
2395  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2396  return new StoreInst(TOp, Ptr);
2397  }
2398  break;
2399  case Intrinsic::ppc_qpx_qvstfd:
2400  // Turn PPC QPX qvstfd -> store if the pointer is known aligned.
2401  if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, &AC,
2402  &DT) >= 32) {
2403  Type *OpPtrTy =
2404  PointerType::getUnqual(II->getArgOperand(0)->getType());
2405  Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy);
2406  return new StoreInst(II->getArgOperand(0), Ptr);
2407  }
2408  break;
2409 
2410  case Intrinsic::x86_bmi_bextr_32:
2411  case Intrinsic::x86_bmi_bextr_64:
2412  case Intrinsic::x86_tbm_bextri_u32:
2413  case Intrinsic::x86_tbm_bextri_u64:
2414  // If the RHS is a constant we can try some simplifications.
2415  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
2416  uint64_t Shift = C->getZExtValue();
2417  uint64_t Length = (Shift >> 8) & 0xff;
2418  Shift &= 0xff;
2419  unsigned BitWidth = II->getType()->getIntegerBitWidth();
2420  // If the length is 0 or the shift is out of range, replace with zero.
2421  if (Length == 0 || Shift >= BitWidth)
2422  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0));
2423  // If the LHS is also a constant, we can completely constant fold this.
2424  if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) {
2425  uint64_t Result = InC->getZExtValue() >> Shift;
2426  if (Length > BitWidth)
2427  Length = BitWidth;
2428  Result &= maskTrailingOnes<uint64_t>(Length);
2429  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result));
2430  }
2431  // TODO should we turn this into 'and' if shift is 0? Or 'shl' if we
2432  // are only masking bits that a shift already cleared?
2433  }
2434  break;
2435 
2436  case Intrinsic::x86_bmi_bzhi_32:
2437  case Intrinsic::x86_bmi_bzhi_64:
2438  // If the RHS is a constant we can try some simplifications.
2439  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
2440  uint64_t Index = C->getZExtValue() & 0xff;
2441  unsigned BitWidth = II->getType()->getIntegerBitWidth();
2442  if (Index >= BitWidth)
2443  return replaceInstUsesWith(CI, II->getArgOperand(0));
2444  if (Index == 0)
2445  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0));
2446  // If the LHS is also a constant, we can completely constant fold this.
2447  if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) {
2448  uint64_t Result = InC->getZExtValue();
2449  Result &= maskTrailingOnes<uint64_t>(Index);
2450  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result));
2451  }
2452  // TODO should we convert this to an AND if the RHS is constant?
2453  }
2454  break;
2455 
2456  case Intrinsic::x86_vcvtph2ps_128:
2457  case Intrinsic::x86_vcvtph2ps_256: {
2458  auto Arg = II->getArgOperand(0);
2459  auto ArgType = cast<VectorType>(Arg->getType());
2460  auto RetType = cast<VectorType>(II->getType());
2461  unsigned ArgWidth = ArgType->getNumElements();
2462  unsigned RetWidth = RetType->getNumElements();
2463  assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
2464  assert(ArgType->isIntOrIntVectorTy() &&
2465  ArgType->getScalarSizeInBits() == 16 &&
2466  "CVTPH2PS input type should be 16-bit integer vector");
2467  assert(RetType->getScalarType()->isFloatTy() &&
2468  "CVTPH2PS output type should be 32-bit float vector");
2469 
2470  // Constant folding: Convert to generic half to single conversion.
2471  if (isa<ConstantAggregateZero>(Arg))
2472  return replaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
2473 
2474  if (isa<ConstantDataVector>(Arg)) {
2475  auto VectorHalfAsShorts = Arg;
2476  if (RetWidth < ArgWidth) {
2477  SmallVector<uint32_t, 8> SubVecMask;
2478  for (unsigned i = 0; i != RetWidth; ++i)
2479  SubVecMask.push_back((int)i);
2480  VectorHalfAsShorts = Builder.CreateShuffleVector(
2481  Arg, UndefValue::get(ArgType), SubVecMask);
2482  }
2483 
2484  auto VectorHalfType =
2485  VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
2486  auto VectorHalfs =
2487  Builder.CreateBitCast(VectorHalfAsShorts, VectorHalfType);
2488  auto VectorFloats = Builder.CreateFPExt(VectorHalfs, RetType);
2489  return replaceInstUsesWith(*II, VectorFloats);
2490  }
2491 
2492  // We only use the lowest lanes of the argument.
2493  if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) {
2494  II->setArgOperand(0, V);
2495  return II;
2496  }
2497  break;
2498  }
2499 
2500  case Intrinsic::x86_sse_cvtss2si:
2501  case Intrinsic::x86_sse_cvtss2si64:
2502  case Intrinsic::x86_sse_cvttss2si:
2503  case Intrinsic::x86_sse_cvttss2si64:
2504  case Intrinsic::x86_sse2_cvtsd2si:
2505  case Intrinsic::x86_sse2_cvtsd2si64:
2506  case Intrinsic::x86_sse2_cvttsd2si:
2507  case Intrinsic::x86_sse2_cvttsd2si64:
2508  case Intrinsic::x86_avx512_vcvtss2si32:
2509  case Intrinsic::x86_avx512_vcvtss2si64:
2510  case Intrinsic::x86_avx512_vcvtss2usi32:
2511  case Intrinsic::x86_avx512_vcvtss2usi64:
2512  case Intrinsic::x86_avx512_vcvtsd2si32:
2513  case Intrinsic::x86_avx512_vcvtsd2si64:
2514  case Intrinsic::x86_avx512_vcvtsd2usi32:
2515  case Intrinsic::x86_avx512_vcvtsd2usi64:
2516  case Intrinsic::x86_avx512_cvttss2si:
2517  case Intrinsic::x86_avx512_cvttss2si64:
2518  case Intrinsic::x86_avx512_cvttss2usi:
2519  case Intrinsic::x86_avx512_cvttss2usi64:
2520  case Intrinsic::x86_avx512_cvttsd2si:
2521  case Intrinsic::x86_avx512_cvttsd2si64:
2522  case Intrinsic::x86_avx512_cvttsd2usi:
2523  case Intrinsic::x86_avx512_cvttsd2usi64: {
2524  // These intrinsics only demand the 0th element of their input vectors. If
2525  // we can simplify the input based on that, do so now.
2526  Value *Arg = II->getArgOperand(0);
2527  unsigned VWidth = Arg->getType()->getVectorNumElements();
2528  if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
2529  II->setArgOperand(0, V);
2530  return II;
2531  }
2532  break;
2533  }
2534 
2535  case Intrinsic::x86_mmx_pmovmskb:
2536  case Intrinsic::x86_sse_movmsk_ps:
2537  case Intrinsic::x86_sse2_movmsk_pd:
2538  case Intrinsic::x86_sse2_pmovmskb_128:
2539  case Intrinsic::x86_avx_movmsk_pd_256:
2540  case Intrinsic::x86_avx_movmsk_ps_256:
2541  case Intrinsic::x86_avx2_pmovmskb:
2542  if (Value *V = simplifyX86movmsk(*II, Builder))
2543  return replaceInstUsesWith(*II, V);
2544  break;
2545 
2546  case Intrinsic::x86_sse_comieq_ss:
2547  case Intrinsic::x86_sse_comige_ss:
2548  case Intrinsic::x86_sse_comigt_ss:
2549  case Intrinsic::x86_sse_comile_ss:
2550  case Intrinsic::x86_sse_comilt_ss:
2551  case Intrinsic::x86_sse_comineq_ss:
2552  case Intrinsic::x86_sse_ucomieq_ss:
2553  case Intrinsic::x86_sse_ucomige_ss:
2554  case Intrinsic::x86_sse_ucomigt_ss:
2555  case Intrinsic::x86_sse_ucomile_ss:
2556  case Intrinsic::x86_sse_ucomilt_ss:
2557  case Intrinsic::x86_sse_ucomineq_ss:
2558  case Intrinsic::x86_sse2_comieq_sd:
2559  case Intrinsic::x86_sse2_comige_sd:
2560  case Intrinsic::x86_sse2_comigt_sd:
2561  case Intrinsic::x86_sse2_comile_sd:
2562  case Intrinsic::x86_sse2_comilt_sd:
2563  case Intrinsic::x86_sse2_comineq_sd:
2564  case Intrinsic::x86_sse2_ucomieq_sd:
2565  case Intrinsic::x86_sse2_ucomige_sd:
2566  case Intrinsic::x86_sse2_ucomigt_sd:
2567  case Intrinsic::x86_sse2_ucomile_sd:
2568  case Intrinsic::x86_sse2_ucomilt_sd:
2569  case Intrinsic::x86_sse2_ucomineq_sd:
2570  case Intrinsic::x86_avx512_vcomi_ss:
2571  case Intrinsic::x86_avx512_vcomi_sd:
2572  case Intrinsic::x86_avx512_mask_cmp_ss:
2573  case Intrinsic::x86_avx512_mask_cmp_sd: {
2574  // These intrinsics only demand the 0th element of their input vectors. If
2575  // we can simplify the input based on that, do so now.
2576  bool MadeChange = false;
2577  Value *Arg0 = II->getArgOperand(0);
2578  Value *Arg1 = II->getArgOperand(1);
2579  unsigned VWidth = Arg0->getType()->getVectorNumElements();
2580  if (Value *V = SimplifyDemandedVectorEltsLow(Arg0, VWidth, 1)) {
2581  II->setArgOperand(0, V);
2582  MadeChange = true;
2583  }
2584  if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, 1)) {
2585  II->setArgOperand(1, V);
2586  MadeChange = true;
2587  }
2588  if (MadeChange)
2589  return II;
2590  break;
2591  }
2592  case Intrinsic::x86_avx512_cmp_pd_128:
2593  case Intrinsic::x86_avx512_cmp_pd_256:
2594  case Intrinsic::x86_avx512_cmp_pd_512:
2595  case Intrinsic::x86_avx512_cmp_ps_128:
2596  case Intrinsic::x86_avx512_cmp_ps_256:
2597  case Intrinsic::x86_avx512_cmp_ps_512: {
2598  // Folding cmp(sub(a,b),0) -> cmp(a,b) and cmp(0,sub(a,b)) -> cmp(b,a)
2599  Value *Arg0 = II->getArgOperand(0);
2600  Value *Arg1 = II->getArgOperand(1);
2601  bool Arg0IsZero = match(Arg0, m_PosZeroFP());
2602  if (Arg0IsZero)
2603  std::swap(Arg0, Arg1);
2604  Value *A, *B;
2605  // This fold requires only the NINF(not +/- inf) since inf minus
2606  // inf is nan.
2607  // NSZ(No Signed Zeros) is not needed because zeros of any sign are
2608  // equal for both compares.
2609  // NNAN is not needed because nans compare the same for both compares.
2610  // The compare intrinsic uses the above assumptions and therefore
2611  // doesn't require additional flags.
2612  if ((match(Arg0, m_OneUse(m_FSub(m_Value(A), m_Value(B)))) &&
2613  match(Arg1, m_PosZeroFP()) && isa<Instruction>(Arg0) &&
2614  cast<Instruction>(Arg0)->getFastMathFlags().noInfs())) {
2615  if (Arg0IsZero)
2616  std::swap(A, B);
2617  II->setArgOperand(0, A);
2618  II->setArgOperand(1, B);
2619  return II;
2620  }
2621  break;
2622  }
2623 
2624  case Intrinsic::x86_avx512_add_ps_512:
2625  case Intrinsic::x86_avx512_div_ps_512:
2626  case Intrinsic::x86_avx512_mul_ps_512:
2627  case Intrinsic::x86_avx512_sub_ps_512:
2628  case Intrinsic::x86_avx512_add_pd_512:
2629  case Intrinsic::x86_avx512_div_pd_512:
2630  case Intrinsic::x86_avx512_mul_pd_512:
2631  case Intrinsic::x86_avx512_sub_pd_512:
2632  // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2633  // IR operations.
2634  if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
2635  if (R->getValue() == 4) {
2636  Value *Arg0 = II->getArgOperand(0);
2637  Value *Arg1 = II->getArgOperand(1);
2638 
2639  Value *V;
2640  switch (IID) {
2641  default: llvm_unreachable("Case stmts out of sync!");
2642  case Intrinsic::x86_avx512_add_ps_512:
2643  case Intrinsic::x86_avx512_add_pd_512:
2644  V = Builder.CreateFAdd(Arg0, Arg1);
2645  break;
2646  case Intrinsic::x86_avx512_sub_ps_512:
2647  case Intrinsic::x86_avx512_sub_pd_512:
2648  V = Builder.CreateFSub(Arg0, Arg1);
2649  break;
2650  case Intrinsic::x86_avx512_mul_ps_512:
2651  case Intrinsic::x86_avx512_mul_pd_512:
2652  V = Builder.CreateFMul(Arg0, Arg1);
2653  break;
2654  case Intrinsic::x86_avx512_div_ps_512:
2655  case Intrinsic::x86_avx512_div_pd_512:
2656  V = Builder.CreateFDiv(Arg0, Arg1);
2657  break;
2658  }
2659 
2660  return replaceInstUsesWith(*II, V);
2661  }
2662  }
2663  break;
2664 
2665  case Intrinsic::x86_avx512_mask_add_ss_round:
2666  case Intrinsic::x86_avx512_mask_div_ss_round:
2667  case Intrinsic::x86_avx512_mask_mul_ss_round:
2668  case Intrinsic::x86_avx512_mask_sub_ss_round:
2669  case Intrinsic::x86_avx512_mask_add_sd_round:
2670  case Intrinsic::x86_avx512_mask_div_sd_round:
2671  case Intrinsic::x86_avx512_mask_mul_sd_round:
2672  case Intrinsic::x86_avx512_mask_sub_sd_round:
2673  // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2674  // IR operations.
2675  if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(4))) {
2676  if (R->getValue() == 4) {
2677  // Extract the element as scalars.
2678  Value *Arg0 = II->getArgOperand(0);
2679  Value *Arg1 = II->getArgOperand(1);
2680  Value *LHS = Builder.CreateExtractElement(Arg0, (uint64_t)0);
2681  Value *RHS = Builder.CreateExtractElement(Arg1, (uint64_t)0);
2682 
2683  Value *V;
2684  switch (IID) {
2685  default: llvm_unreachable("Case stmts out of sync!");
2686  case Intrinsic::x86_avx512_mask_add_ss_round:
2687  case Intrinsic::x86_avx512_mask_add_sd_round:
2688  V = Builder.CreateFAdd(LHS, RHS);
2689  break;
2690  case Intrinsic::x86_avx512_mask_sub_ss_round:
2691  case Intrinsic::x86_avx512_mask_sub_sd_round:
2692  V = Builder.CreateFSub(LHS, RHS);
2693  break;
2694  case Intrinsic::x86_avx512_mask_mul_ss_round:
2695  case Intrinsic::x86_avx512_mask_mul_sd_round:
2696  V = Builder.CreateFMul(LHS, RHS);
2697  break;
2698  case Intrinsic::x86_avx512_mask_div_ss_round:
2699  case Intrinsic::x86_avx512_mask_div_sd_round:
2700  V = Builder.CreateFDiv(LHS, RHS);
2701  break;
2702  }
2703 
2704  // Handle the masking aspect of the intrinsic.
2705  Value *Mask = II->getArgOperand(3);
2706  auto *C = dyn_cast<ConstantInt>(Mask);
2707  // We don't need a select if we know the mask bit is a 1.
2708  if (!C || !C->getValue()[0]) {
2709  // Cast the mask to an i1 vector and then extract the lowest element.
2710  auto *MaskTy = VectorType::get(Builder.getInt1Ty(),
2711  cast<IntegerType>(Mask->getType())->getBitWidth());
2712  Mask = Builder.CreateBitCast(Mask, MaskTy);
2713  Mask = Builder.CreateExtractElement(Mask, (uint64_t)0);
2714  // Extract the lowest element from the passthru operand.
2715  Value *Passthru = Builder.CreateExtractElement(II->getArgOperand(2),
2716  (uint64_t)0);
2717  V = Builder.CreateSelect(Mask, V, Passthru);
2718  }
2719 
2720  // Insert the result back into the original argument 0.
2721  V = Builder.CreateInsertElement(Arg0, V, (uint64_t)0);
2722 
2723  return replaceInstUsesWith(*II, V);
2724  }
2725  }
2726  break;
2727 
2728  // Constant fold ashr( <A x Bi>, Ci ).
2729  // Constant fold lshr( <A x Bi>, Ci ).
2730  // Constant fold shl( <A x Bi>, Ci ).
2731  case Intrinsic::x86_sse2_psrai_d:
2732  case Intrinsic::x86_sse2_psrai_w:
2733  case Intrinsic::x86_avx2_psrai_d:
2734  case Intrinsic::x86_avx2_psrai_w:
2735  case Intrinsic::x86_avx512_psrai_q_128:
2736  case Intrinsic::x86_avx512_psrai_q_256:
2737  case Intrinsic::x86_avx512_psrai_d_512:
2738  case Intrinsic::x86_avx512_psrai_q_512:
2739  case Intrinsic::x86_avx512_psrai_w_512:
2740  case Intrinsic::x86_sse2_psrli_d:
2741  case Intrinsic::x86_sse2_psrli_q:
2742  case Intrinsic::x86_sse2_psrli_w:
2743  case Intrinsic::x86_avx2_psrli_d:
2744  case Intrinsic::x86_avx2_psrli_q:
2745  case Intrinsic::x86_avx2_psrli_w:
2746  case Intrinsic::x86_avx512_psrli_d_512:
2747  case Intrinsic::x86_avx512_psrli_q_512:
2748  case Intrinsic::x86_avx512_psrli_w_512:
2749  case Intrinsic::x86_sse2_pslli_d:
2750  case Intrinsic::x86_sse2_pslli_q:
2751  case Intrinsic::x86_sse2_pslli_w:
2752  case Intrinsic::x86_avx2_pslli_d:
2753  case Intrinsic::x86_avx2_pslli_q:
2754  case Intrinsic::x86_avx2_pslli_w:
2755  case Intrinsic::x86_avx512_pslli_d_512:
2756  case Intrinsic::x86_avx512_pslli_q_512:
2757  case Intrinsic::x86_avx512_pslli_w_512:
2758  if (Value *V = simplifyX86immShift(*II, Builder))
2759  return replaceInstUsesWith(*II, V);
2760  break;
2761 
2762  case Intrinsic::x86_sse2_psra_d:
2763  case Intrinsic::x86_sse2_psra_w:
2764  case Intrinsic::x86_avx2_psra_d:
2765  case Intrinsic::x86_avx2_psra_w:
2766  case Intrinsic::x86_avx512_psra_q_128:
2767  case Intrinsic::x86_avx512_psra_q_256:
2768  case Intrinsic::x86_avx512_psra_d_512:
2769  case Intrinsic::x86_avx512_psra_q_512:
2770  case Intrinsic::x86_avx512_psra_w_512:
2771  case Intrinsic::x86_sse2_psrl_d:
2772  case Intrinsic::x86_sse2_psrl_q:
2773  case Intrinsic::x86_sse2_psrl_w:
2774  case Intrinsic::x86_avx2_psrl_d:
2775  case Intrinsic::x86_avx2_psrl_q:
2776  case Intrinsic::x86_avx2_psrl_w:
2777  case Intrinsic::x86_avx512_psrl_d_512:
2778  case Intrinsic::x86_avx512_psrl_q_512:
2779  case Intrinsic::x86_avx512_psrl_w_512:
2780  case Intrinsic::x86_sse2_psll_d:
2781  case Intrinsic::x86_sse2_psll_q:
2782  case Intrinsic::x86_sse2_psll_w:
2783  case Intrinsic::x86_avx2_psll_d:
2784  case Intrinsic::x86_avx2_psll_q:
2785  case Intrinsic::x86_avx2_psll_w:
2786  case Intrinsic::x86_avx512_psll_d_512:
2787  case Intrinsic::x86_avx512_psll_q_512:
2788  case Intrinsic::x86_avx512_psll_w_512: {
2789  if (Value *V = simplifyX86immShift(*II, Builder))
2790  return replaceInstUsesWith(*II, V);
2791 
2792  // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
2793  // operand to compute the shift amount.
2794  Value *Arg1 = II->getArgOperand(1);
2795  assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 &&
2796  "Unexpected packed shift size");
2797  unsigned VWidth = Arg1->getType()->getVectorNumElements();
2798 
2799  if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) {
2800  II->setArgOperand(1, V);
2801  return II;
2802  }
2803  break;
2804  }
2805 
2806  case Intrinsic::x86_avx2_psllv_d:
2807  case Intrinsic::x86_avx2_psllv_d_256:
2808  case Intrinsic::x86_avx2_psllv_q:
2809  case Intrinsic::x86_avx2_psllv_q_256:
2810  case Intrinsic::x86_avx512_psllv_d_512:
2811  case Intrinsic::x86_avx512_psllv_q_512:
2812  case Intrinsic::x86_avx512_psllv_w_128:
2813  case Intrinsic::x86_avx512_psllv_w_256:
2814  case Intrinsic::x86_avx512_psllv_w_512:
2815  case Intrinsic::x86_avx2_psrav_d:
2816  case Intrinsic::x86_avx2_psrav_d_256:
2817  case Intrinsic::x86_avx512_psrav_q_128:
2818  case Intrinsic::x86_avx512_psrav_q_256:
2819  case Intrinsic::x86_avx512_psrav_d_512:
2820  case Intrinsic::x86_avx512_psrav_q_512:
2821  case Intrinsic::x86_avx512_psrav_w_128:
2822  case Intrinsic::x86_avx512_psrav_w_256:
2823  case Intrinsic::x86_avx512_psrav_w_512:
2824  case Intrinsic::x86_avx2_psrlv_d:
2825  case Intrinsic::x86_avx2_psrlv_d_256:
2826  case Intrinsic::x86_avx2_psrlv_q:
2827  case Intrinsic::x86_avx2_psrlv_q_256:
2828  case Intrinsic::x86_avx512_psrlv_d_512:
2829  case Intrinsic::x86_avx512_psrlv_q_512:
2830  case Intrinsic::x86_avx512_psrlv_w_128:
2831  case Intrinsic::x86_avx512_psrlv_w_256:
2832  case Intrinsic::x86_avx512_psrlv_w_512:
2833  if (Value *V = simplifyX86varShift(*II, Builder))
2834  return replaceInstUsesWith(*II, V);
2835  break;
2836 
2837  case Intrinsic::x86_sse2_packssdw_128:
2838  case Intrinsic::x86_sse2_packsswb_128:
2839  case Intrinsic::x86_avx2_packssdw:
2840  case Intrinsic::x86_avx2_packsswb:
2841  case Intrinsic::x86_avx512_packssdw_512:
2842  case Intrinsic::x86_avx512_packsswb_512:
2843  if (Value *V = simplifyX86pack(*II, Builder, true))
2844  return replaceInstUsesWith(*II, V);
2845  break;
2846 
2847  case Intrinsic::x86_sse2_packuswb_128:
2848  case Intrinsic::x86_sse41_packusdw:
2849  case Intrinsic::x86_avx2_packusdw:
2850  case Intrinsic::x86_avx2_packuswb:
2851  case Intrinsic::x86_avx512_packusdw_512:
2852  case Intrinsic::x86_avx512_packuswb_512:
2853  if (Value *V = simplifyX86pack(*II, Builder, false))
2854  return replaceInstUsesWith(*II, V);
2855  break;
2856 
2857  case Intrinsic::x86_pclmulqdq:
2858  case Intrinsic::x86_pclmulqdq_256:
2859  case Intrinsic::x86_pclmulqdq_512: {
2860  if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
2861  unsigned Imm = C->getZExtValue();
2862 
2863  bool MadeChange = false;
2864  Value *Arg0 = II->getArgOperand(0);
2865  Value *Arg1 = II->getArgOperand(1);
2866  unsigned VWidth = Arg0->getType()->getVectorNumElements();
2867 
2868  APInt UndefElts1(VWidth, 0);
2869  APInt DemandedElts1 = APInt::getSplat(VWidth,
2870  APInt(2, (Imm & 0x01) ? 2 : 1));
2871  if (Value *V = SimplifyDemandedVectorElts(Arg0, DemandedElts1,
2872  UndefElts1)) {
2873  II->setArgOperand(0, V);
2874  MadeChange = true;
2875  }
2876 
2877  APInt UndefElts2(VWidth, 0);
2878  APInt DemandedElts2 = APInt::getSplat(VWidth,
2879  APInt(2, (Imm & 0x10) ? 2 : 1));
2880  if (Value *V = SimplifyDemandedVectorElts(Arg1, DemandedElts2,
2881  UndefElts2)) {
2882  II->setArgOperand(1, V);
2883  MadeChange = true;
2884  }
2885 
2886  // If either input elements are undef, the result is zero.
2887  if (DemandedElts1.isSubsetOf(UndefElts1) ||
2888  DemandedElts2.isSubsetOf(UndefElts2))
2889  return replaceInstUsesWith(*II,
2890  ConstantAggregateZero::get(II->getType()));
2891 
2892  if (MadeChange)
2893  return II;
2894  }
2895  break;
2896  }
2897 
2898  case Intrinsic::x86_sse41_insertps:
2899  if (Value *V = simplifyX86insertps(*II, Builder))
2900  return replaceInstUsesWith(*II, V);
2901  break;
2902 
2903  case Intrinsic::x86_sse4a_extrq: {
2904  Value *Op0 = II->getArgOperand(0);
2905  Value *Op1 = II->getArgOperand(1);
2906  unsigned VWidth0 = Op0->getType()->getVectorNumElements();
2907  unsigned VWidth1 = Op1->getType()->getVectorNumElements();
2908  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2909  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
2910  VWidth1 == 16 && "Unexpected operand sizes");
2911 
2912  // See if we're dealing with constant values.
2913  Constant *C1 = dyn_cast<Constant>(Op1);
2914  ConstantInt *CILength =
2915  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
2916  : nullptr;
2917  ConstantInt *CIIndex =
2918  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
2919  : nullptr;
2920 
2921  // Attempt to simplify to a constant, shuffle vector or EXTRQI call.
2922  if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder))
2923  return replaceInstUsesWith(*II, V);
2924 
2925  // EXTRQ only uses the lowest 64-bits of the first 128-bit vector
2926  // operands and the lowest 16-bits of the second.
2927  bool MadeChange = false;
2928  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
2929  II->setArgOperand(0, V);
2930  MadeChange = true;
2931  }
2932  if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
2933  II->setArgOperand(1, V);
2934  MadeChange = true;
2935  }
2936  if (MadeChange)
2937  return II;
2938  break;
2939  }
2940 
2941  case Intrinsic::x86_sse4a_extrqi: {
2942  // EXTRQI: Extract Length bits starting from Index. Zero pad the remaining
2943  // bits of the lower 64-bits. The upper 64-bits are undefined.
2944  Value *Op0 = II->getArgOperand(0);
2945  unsigned VWidth = Op0->getType()->getVectorNumElements();
2946  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
2947  "Unexpected operand size");
2948 
2949  // See if we're dealing with constant values.
2950  ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(1));
2951  ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(2));
2952 
2953  // Attempt to simplify to a constant or shuffle vector.
2954  if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder))
2955  return replaceInstUsesWith(*II, V);
2956 
2957  // EXTRQI only uses the lowest 64-bits of the first 128-bit vector
2958  // operand.
2959  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
2960  II->setArgOperand(0, V);
2961  return II;
2962  }
2963  break;
2964  }
2965 
2966  case Intrinsic::x86_sse4a_insertq: {
2967  Value *Op0 = II->getArgOperand(0);
2968  Value *Op1 = II->getArgOperand(1);
2969  unsigned VWidth = Op0->getType()->getVectorNumElements();
2970  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2971  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
2972  Op1->getType()->getVectorNumElements() == 2 &&
2973  "Unexpected operand size");
2974 
2975  // See if we're dealing with constant values.
2976  Constant *C1 = dyn_cast<Constant>(Op1);
2977  ConstantInt *CI11 =
2978  C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
2979  : nullptr;
2980 
2981  // Attempt to simplify to a constant, shuffle vector or INSERTQI call.
2982  if (CI11) {
2983  const APInt &V11 = CI11->getValue();
2984  APInt Len = V11.zextOrTrunc(6);
2985  APInt Idx = V11.lshr(8).zextOrTrunc(6);
2986  if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder))
2987  return replaceInstUsesWith(*II, V);
2988  }
2989 
2990  // INSERTQ only uses the lowest 64-bits of the first 128-bit vector
2991  // operand.
2992  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
2993  II->setArgOperand(0, V);
2994  return II;
2995  }
2996  break;
2997  }
2998 
2999  case Intrinsic::x86_sse4a_insertqi: {
3000  // INSERTQI: Extract lowest Length bits from lower half of second source and
3001  // insert over first source starting at Index bit. The upper 64-bits are
3002  // undefined.
3003  Value *Op0 = II->getArgOperand(0);
3004  Value *Op1 = II->getArgOperand(1);
3005  unsigned VWidth0 = Op0->getType()->getVectorNumElements();
3006  unsigned VWidth1 = Op1->getType()->getVectorNumElements();
3007  assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
3008  Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
3009  VWidth1 == 2 && "Unexpected operand sizes");
3010 
3011  // See if we're dealing with constant values.
3012  ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(2));
3013  ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3));
3014 
3015  // Attempt to simplify to a constant or shuffle vector.
3016  if (CILength && CIIndex) {
3017  APInt Len = CILength->getValue().zextOrTrunc(6);
3018  APInt Idx = CIIndex->getValue().zextOrTrunc(6);
3019  if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder))
3020  return replaceInstUsesWith(*II, V);
3021  }
3022 
3023  // INSERTQI only uses the lowest 64-bits of the first two 128-bit vector
3024  // operands.
3025  bool MadeChange = false;
3026  if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
3027  II->setArgOperand(0, V);
3028  MadeChange = true;
3029  }
3030  if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
3031  II->setArgOperand(1, V);
3032  MadeChange = true;
3033  }
3034  if (MadeChange)
3035  return II;
3036  break;
3037  }
3038 
3039  case Intrinsic::x86_sse41_pblendvb:
3040  case Intrinsic::x86_sse41_blendvps:
3041  case Intrinsic::x86_sse41_blendvpd:
3042  case Intrinsic::x86_avx_blendv_ps_256:
3043  case Intrinsic::x86_avx_blendv_pd_256:
3044  case Intrinsic::x86_avx2_pblendvb: {
3045  // fold (blend A, A, Mask) -> A
3046  Value *Op0 = II->getArgOperand(0);
3047  Value *Op1 = II->getArgOperand(1);
3048  Value *Mask = II->getArgOperand(2);
3049  if (Op0 == Op1)
3050  return replaceInstUsesWith(CI, Op0);
3051 
3052  // Zero Mask - select 1st argument.
3053  if (isa<ConstantAggregateZero>(Mask))
3054  return replaceInstUsesWith(CI, Op0);
3055 
3056  // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
3057  if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask)) {
3058  Constant *NewSelector = getNegativeIsTrueBoolVec(ConstantMask);
3059  return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
3060  }
3061 
3062  // Convert to a vector select if we can bypass casts and find a boolean
3063  // vector condition value.
3064  Value *BoolVec;
3065  Mask = peekThroughBitcast(Mask);
3066  if (match(Mask, m_SExt(m_Value(BoolVec))) &&
3067  BoolVec->getType()->isVectorTy() &&
3068  BoolVec->getType()->getScalarSizeInBits() == 1) {
3069  assert(Mask->getType()->getPrimitiveSizeInBits() ==
3070  II->getType()->getPrimitiveSizeInBits() &&
3071  "Not expecting mask and operands with different sizes");
3072 
3073  unsigned NumMaskElts = Mask->getType()->getVectorNumElements();
3074  unsigned NumOperandElts = II->getType()->getVectorNumElements();
3075  if (NumMaskElts == NumOperandElts)
3076  return SelectInst::Create(BoolVec, Op1, Op0);
3077 
3078  // If the mask has less elements than the operands, each mask bit maps to
3079  // multiple elements of the operands. Bitcast back and forth.
3080  if (NumMaskElts < NumOperandElts) {
3081  Value *CastOp0 = Builder.CreateBitCast(Op0, Mask->getType());
3082  Value *CastOp1 = Builder.CreateBitCast(Op1, Mask->getType());
3083  Value *Sel = Builder.CreateSelect(BoolVec, CastOp1, CastOp0);
3084  return new BitCastInst(Sel, II->getType());
3085  }
3086  }
3087 
3088  break;
3089  }
3090 
3091  case Intrinsic::x86_ssse3_pshuf_b_128:
3092  case Intrinsic::x86_avx2_pshuf_b:
3093  case Intrinsic::x86_avx512_pshuf_b_512:
3094  if (Value *V = simplifyX86pshufb(*II, Builder))
3095  return replaceInstUsesWith(*II, V);
3096  break;
3097 
3098  case Intrinsic::x86_avx_vpermilvar_ps:
3099  case Intrinsic::x86_avx_vpermilvar_ps_256:
3100  case Intrinsic::x86_avx512_vpermilvar_ps_512:
3101  case Intrinsic::x86_avx_vpermilvar_pd:
3102  case Intrinsic::x86_avx_vpermilvar_pd_256:
3103  case Intrinsic::x86_avx512_vpermilvar_pd_512:
3104  if (Value *V = simplifyX86vpermilvar(*II, Builder))
3105  return replaceInstUsesWith(*II, V);
3106  break;
3107 
3108  case Intrinsic::x86_avx2_permd:
3109  case Intrinsic::x86_avx2_permps:
3110  case Intrinsic::x86_avx512_permvar_df_256:
3111  case Intrinsic::x86_avx512_permvar_df_512:
3112  case Intrinsic::x86_avx512_permvar_di_256:
3113  case Intrinsic::x86_avx512_permvar_di_512:
3114  case Intrinsic::x86_avx512_permvar_hi_128:
3115  case Intrinsic::x86_avx512_permvar_hi_256:
3116  case Intrinsic::x86_avx512_permvar_hi_512:
3117  case Intrinsic::x86_avx512_permvar_qi_128:
3118  case Intrinsic::x86_avx512_permvar_qi_256:
3119  case Intrinsic::x86_avx512_permvar_qi_512:
3120  case Intrinsic::x86_avx512_permvar_sf_512:
3121  case Intrinsic::x86_avx512_permvar_si_512:
3122  if (Value *V = simplifyX86vpermv(*II, Builder))
3123  return replaceInstUsesWith(*II, V);
3124  break;
3125 
3126  case Intrinsic::x86_avx_maskload_ps:
3127  case Intrinsic::x86_avx_maskload_pd:
3128  case Intrinsic::x86_avx_maskload_ps_256:
3129  case Intrinsic::x86_avx_maskload_pd_256:
3130  case Intrinsic::x86_avx2_maskload_d:
3131  case Intrinsic::x86_avx2_maskload_q:
3132  case Intrinsic::x86_avx2_maskload_d_256:
3133  case Intrinsic::x86_avx2_maskload_q_256:
3134  if (Instruction *I = simplifyX86MaskedLoad(*II, *this))
3135  return I;
3136  break;
3137 
3138  case Intrinsic::x86_sse2_maskmov_dqu:
3139  case Intrinsic::x86_avx_maskstore_ps:
3140  case Intrinsic::x86_avx_maskstore_pd:
3141  case Intrinsic::x86_avx_maskstore_ps_256:
3142  case Intrinsic::x86_avx_maskstore_pd_256:
3143  case Intrinsic::x86_avx2_maskstore_d:
3144  case Intrinsic::x86_avx2_maskstore_q:
3145  case Intrinsic::x86_avx2_maskstore_d_256:
3146  case Intrinsic::x86_avx2_maskstore_q_256:
3147  if (simplifyX86MaskedStore(*II, *this))
3148  return nullptr;
3149  break;
3150 
3151  case Intrinsic::x86_addcarry_32:
3152  case Intrinsic::x86_addcarry_64:
3153  if (Value *V = simplifyX86addcarry(*II, Builder))
3154  return replaceInstUsesWith(*II, V);
3155  break;
3156 
3157  case Intrinsic::ppc_altivec_vperm:
3158  // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
3159  // Note that ppc_altivec_vperm has a big-endian bias, so when creating
3160  // a vectorshuffle for little endian, we must undo the transformation
3161  // performed on vec_perm in altivec.h. That is, we must complement
3162  // the permutation mask with respect to 31 and reverse the order of
3163  // V1 and V2.
3164  if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
3165  assert(Mask->getType()->getVectorNumElements() == 16 &&
3166  "Bad type for intrinsic!");
3167 
3168  // Check that all of the elements are integer constants or undefs.
3169  bool AllEltsOk = true;
3170  for (unsigned i = 0; i != 16; ++i) {
3171  Constant *Elt = Mask->getAggregateElement(i);
3172  if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
3173  AllEltsOk = false;
3174  break;
3175  }
3176  }
3177 
3178  if (AllEltsOk) {
3179  // Cast the input vectors to byte vectors.
3180  Value *Op0 = Builder.CreateBitCast(II->getArgOperand(0),
3181  Mask->getType());
3182  Value *Op1 = Builder.CreateBitCast(II->getArgOperand(1),
3183  Mask->getType());
3184  Value *Result = UndefValue::get(Op0->getType());
3185 
3186  // Only extract each element once.
3187  Value *ExtractedElts[32];
3188  memset(ExtractedElts, 0, sizeof(ExtractedElts));
3189 
3190  for (unsigned i = 0; i != 16; ++i) {
3191  if (isa<UndefValue>(Mask->getAggregateElement(i)))
3192  continue;
3193  unsigned Idx =
3194  cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
3195  Idx &= 31; // Match the hardware behavior.
3196  if (DL.isLittleEndian())
3197  Idx = 31 - Idx;
3198 
3199  if (!ExtractedElts[Idx]) {
3200  Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0;
3201  Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1;
3202  ExtractedElts[Idx] =
3203  Builder.CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
3204  Builder.getInt32(Idx&15));
3205  }
3206 
3207  // Insert this value into the result vector.
3208  Result = Builder.CreateInsertElement(Result, ExtractedElts[Idx],
3209  Builder.getInt32(i));
3210  }
3211  return CastInst::Create(Instruction::BitCast, Result, CI.getType());
3212  }
3213  }
3214  break;
3215 
3216  case Intrinsic::arm_neon_vld1: {
3217  unsigned MemAlign = getKnownAlignment(II->getArgOperand(0),
3218  DL, II, &AC, &DT);
3219  if (Value *V = simplifyNeonVld1(*II, MemAlign, Builder))
3220  return replaceInstUsesWith(*II, V);
3221  break;
3222  }
3223 
3224  case Intrinsic::arm_neon_vld2:
3225  case Intrinsic::arm_neon_vld3:
3226  case Intrinsic::arm_neon_vld4:
3227  case Intrinsic::arm_neon_vld2lane:
3228  case Intrinsic::arm_neon_vld3lane:
3229  case Intrinsic::arm_neon_vld4lane:
3230  case Intrinsic::arm_neon_vst1:
3231  case Intrinsic::arm_neon_vst2:
3232  case Intrinsic::arm_neon_vst3:
3233  case Intrinsic::arm_neon_vst4:
3234  case Intrinsic::arm_neon_vst2lane:
3235  case Intrinsic::arm_neon_vst3lane:
3236  case Intrinsic::arm_neon_vst4lane: {
3237  unsigned MemAlign =
3238  getKnownAlignment(II->getArgOperand(0), DL, II, &AC, &DT);
3239  unsigned AlignArg = II->getNumArgOperands() - 1;
3240  ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
3241  if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
3242  II->setArgOperand(AlignArg,
3243  ConstantInt::get(Type::getInt32Ty(II->getContext()),
3244  MemAlign, false));
3245  return II;
3246  }
3247  break;
3248  }
3249 
3250  case Intrinsic::arm_neon_vtbl1:
3251  case Intrinsic::aarch64_neon_tbl1:
3252  if (Value *V = simplifyNeonTbl1(*II, Builder))
3253  return replaceInstUsesWith(*II, V);
3254  break;
3255 
3256  case Intrinsic::arm_neon_vmulls:
3257  case Intrinsic::arm_neon_vmullu:
3258  case Intrinsic::aarch64_neon_smull:
3259  case Intrinsic::aarch64_neon_umull: {
3260  Value *Arg0 = II->getArgOperand(0);
3261  Value *Arg1 = II->getArgOperand(1);
3262 
3263  // Handle mul by zero first:
3264  if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
3265  return replaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
3266  }
3267 
3268  // Check for constant LHS & RHS - in this case we just simplify.
3269  bool Zext = (IID == Intrinsic::arm_neon_vmullu ||
3270  IID == Intrinsic::aarch64_neon_umull);
3271  VectorType *NewVT = cast<VectorType>(II->getType());
3272  if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
3273  if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
3274  CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
3275  CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
3276 
3277  return replaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
3278  }
3279 
3280  // Couldn't simplify - canonicalize constant to the RHS.
3281  std::swap(Arg0, Arg1);
3282  }
3283 
3284  // Handle mul by one:
3285  if (Constant *CV1 = dyn_cast<Constant>(Arg1))
3286  if (ConstantInt *Splat =
3287  dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
3288  if (Splat->isOne())
3289  return CastInst::CreateIntegerCast(Arg0, II->getType(),
3290  /*isSigned=*/!Zext);
3291 
3292  break;
3293  }
3294  case Intrinsic::arm_neon_aesd:
3295  case Intrinsic::arm_neon_aese:
3296  case Intrinsic::aarch64_crypto_aesd:
3297  case Intrinsic::aarch64_crypto_aese: {
3298  Value *DataArg = II->getArgOperand(0);
3299  Value *KeyArg = II->getArgOperand(1);
3300 
3301  // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR
3302  Value *Data, *Key;
3303  if (match(KeyArg, m_ZeroInt()) &&
3304  match(DataArg, m_Xor(m_Value(Data), m_Value(Key)))) {
3305  II->setArgOperand(0, Data);
3306  II->setArgOperand(1, Key);
3307  return II;
3308  }
3309  break;
3310  }
3311  case Intrinsic::amdgcn_rcp: {
3312  Value *Src = II->getArgOperand(0);
3313 
3314  // TODO: Move to ConstantFolding/InstSimplify?
3315  if (isa<UndefValue>(Src))
3316  return replaceInstUsesWith(CI, Src);
3317 
3318  if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) {
3319  const APFloat &ArgVal = C->getValueAPF();
3320  APFloat Val(ArgVal.getSemantics(), 1.0);
3321  APFloat::opStatus Status = Val.divide(ArgVal,
3323  // Only do this if it was exact and therefore not dependent on the
3324  // rounding mode.
3325  if (Status == APFloat::opOK)
3326  return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
3327  }
3328 
3329  break;
3330  }
3331  case Intrinsic::amdgcn_rsq: {
3332  Value *Src = II->getArgOperand(0);
3333 
3334  // TODO: Move to ConstantFolding/InstSimplify?
3335  if (isa<UndefValue>(Src))
3336  return replaceInstUsesWith(CI, Src);
3337  break;
3338  }
3339  case Intrinsic::amdgcn_frexp_mant:
3340  case Intrinsic::amdgcn_frexp_exp: {
3341  Value *Src = II->getArgOperand(0);
3342  if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) {
3343  int Exp;
3344  APFloat Significand = frexp(C->getValueAPF(), Exp,
3346 
3347  if (IID == Intrinsic::amdgcn_frexp_mant) {
3348  return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(),
3349  Significand));
3350  }
3351 
3352  // Match instruction special case behavior.
3353  if (Exp == APFloat::IEK_NaN || Exp == APFloat::IEK_Inf)
3354  Exp = 0;
3355 
3356  return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Exp));
3357  }
3358 
3359  if (isa<UndefValue>(Src))
3360  return replaceInstUsesWith(CI, UndefValue::get(II->getType()));
3361 
3362  break;
3363  }
3364  case Intrinsic::amdgcn_class: {
3365  enum {
3366  S_NAN = 1 << 0, // Signaling NaN
3367  Q_NAN = 1 << 1, // Quiet NaN
3368  N_INFINITY = 1 << 2, // Negative infinity
3369  N_NORMAL = 1 << 3, // Negative normal
3370  N_SUBNORMAL = 1 << 4, // Negative subnormal
3371  N_ZERO = 1 << 5, // Negative zero
3372  P_ZERO = 1 << 6, // Positive zero
3373  P_SUBNORMAL = 1 << 7, // Positive subnormal
3374  P_NORMAL = 1 << 8, // Positive normal
3375  P_INFINITY = 1 << 9 // Positive infinity
3376  };
3377 
3378  const uint32_t FullMask = S_NAN | Q_NAN | N_INFINITY | N_NORMAL |
3380 
3381  Value *Src0 = II->getArgOperand(0);
3382  Value *Src1 = II->getArgOperand(1);
3383  const ConstantInt *CMask = dyn_cast<ConstantInt>(Src1);
3384  if (!CMask) {
3385  if (isa<UndefValue>(Src0))
3386  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3387 
3388  if (isa<UndefValue>(Src1))
3389  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false));
3390  break;
3391  }
3392 
3393  uint32_t Mask = CMask->getZExtValue();
3394 
3395  // If all tests are made, it doesn't matter what the value is.
3396  if ((Mask & FullMask) == FullMask)
3397  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), true));
3398 
3399  if ((Mask & FullMask) == 0)
3400  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false));
3401 
3402  if (Mask == (S_NAN | Q_NAN)) {
3403  // Equivalent of isnan. Replace with standard fcmp.
3404  Value *FCmp = Builder.CreateFCmpUNO(Src0, Src0);
3405  FCmp->takeName(II);
3406  return replaceInstUsesWith(*II, FCmp);
3407  }
3408 
3409  if (Mask == (N_ZERO | P_ZERO)) {
3410  // Equivalent of == 0.
3411  Value *FCmp = Builder.CreateFCmpOEQ(
3412  Src0, ConstantFP::get(Src0->getType(), 0.0));
3413 
3414  FCmp->takeName(II);
3415  return replaceInstUsesWith(*II, FCmp);
3416  }
3417 
3418  // fp_class (nnan x), qnan|snan|other -> fp_class (nnan x), other
3419  if (((Mask & S_NAN) || (Mask & Q_NAN)) && isKnownNeverNaN(Src0, &TLI)) {
3420  II->setArgOperand(1, ConstantInt::get(Src1->getType(),
3421  Mask & ~(S_NAN | Q_NAN)));
3422  return II;
3423  }
3424 
3425  const ConstantFP *CVal = dyn_cast<ConstantFP>(Src0);
3426  if (!CVal) {
3427  if (isa<UndefValue>(Src0))
3428  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3429 
3430  // Clamp mask to used bits
3431  if ((Mask & FullMask) != Mask) {
3432  CallInst *NewCall = Builder.CreateCall(II->getCalledFunction(),
3433  { Src0, ConstantInt::get(Src1->getType(), Mask & FullMask) }
3434  );
3435 
3436  NewCall->takeName(II);
3437  return replaceInstUsesWith(*II, NewCall);
3438  }
3439 
3440  break;
3441  }
3442 
3443  const APFloat &Val = CVal->getValueAPF();
3444 
3445  bool Result =
3446  ((Mask & S_NAN) && Val.isNaN() && Val.isSignaling()) ||
3447  ((Mask & Q_NAN) && Val.isNaN() && !Val.isSignaling()) ||
3448  ((Mask & N_INFINITY) && Val.isInfinity() && Val.isNegative()) ||
3449  ((Mask & N_NORMAL) && Val.isNormal() && Val.isNegative()) ||
3450  ((Mask & N_SUBNORMAL) && Val.isDenormal() && Val.isNegative()) ||
3451  ((Mask & N_ZERO) && Val.isZero() && Val.isNegative()) ||
3452  ((Mask & P_ZERO) && Val.isZero() && !Val.isNegative()) ||
3453  ((Mask & P_SUBNORMAL) && Val.isDenormal() && !Val.isNegative()) ||
3454  ((Mask & P_NORMAL) && Val.isNormal() && !Val.isNegative()) ||
3455  ((Mask & P_INFINITY) && Val.isInfinity() && !Val.isNegative());
3456 
3457  return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), Result));
3458  }
3459  case Intrinsic::amdgcn_cvt_pkrtz: {
3460  Value *Src0 = II->getArgOperand(0);
3461  Value *Src1 = II->getArgOperand(1);
3462  if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) {
3463  if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) {
3464  const fltSemantics &HalfSem
3465  = II->getType()->getScalarType()->getFltSemantics();
3466  bool LosesInfo;
3467  APFloat Val0 = C0->getValueAPF();
3468  APFloat Val1 = C1->getValueAPF();
3469  Val0.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo);
3470  Val1.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo);
3471 
3472  Constant *Folded = ConstantVector::get({
3473  ConstantFP::get(II->getContext(), Val0),
3474  ConstantFP::get(II->getContext(), Val1) });
3475  return replaceInstUsesWith(*II, Folded);
3476  }
3477  }
3478 
3479  if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1))
3480  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3481 
3482  break;
3483  }
3484  case Intrinsic::amdgcn_cvt_pknorm_i16:
3485  case Intrinsic::amdgcn_cvt_pknorm_u16:
3486  case Intrinsic::amdgcn_cvt_pk_i16:
3487  case Intrinsic::amdgcn_cvt_pk_u16: {
3488  Value *Src0 = II->getArgOperand(0);
3489  Value *Src1 = II->getArgOperand(1);
3490 
3491  if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1))
3492  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3493 
3494  break;
3495  }
3496  case Intrinsic::amdgcn_ubfe:
3497  case Intrinsic::amdgcn_sbfe: {
3498  // Decompose simple cases into standard shifts.
3499  Value *Src = II->getArgOperand(0);
3500  if (isa<UndefValue>(Src))
3501  return replaceInstUsesWith(*II, Src);
3502 
3503  unsigned Width;
3504  Type *Ty = II->getType();
3505  unsigned IntSize = Ty->getIntegerBitWidth();
3506 
3507  ConstantInt *CWidth = dyn_cast<ConstantInt>(II->getArgOperand(2));
3508  if (CWidth) {
3509  Width = CWidth->getZExtValue();
3510  if ((Width & (IntSize - 1)) == 0)
3511  return replaceInstUsesWith(*II, ConstantInt::getNullValue(Ty));
3512 
3513  if (Width >= IntSize) {
3514  // Hardware ignores high bits, so remove those.
3515  II->setArgOperand(2, ConstantInt::get(CWidth->getType(),
3516  Width & (IntSize - 1)));
3517  return II;
3518  }
3519  }
3520 
3521  unsigned Offset;
3522  ConstantInt *COffset = dyn_cast<ConstantInt>(II->getArgOperand(1));
3523  if (COffset) {
3524  Offset = COffset->getZExtValue();
3525  if (Offset >= IntSize) {
3526  II->setArgOperand(1, ConstantInt::get(COffset->getType(),
3527  Offset & (IntSize - 1)));
3528  return II;
3529  }
3530  }
3531 
3532  bool Signed = IID == Intrinsic::amdgcn_sbfe;
3533 
3534  if (!CWidth || !COffset)
3535  break;
3536 
3537  // The case of Width == 0 is handled above, which makes this tranformation
3538  // safe. If Width == 0, then the ashr and lshr instructions become poison
3539  // value since the shift amount would be equal to the bit size.
3540  assert(Width != 0);
3541 
3542  // TODO: This allows folding to undef when the hardware has specific
3543  // behavior?
3544  if (Offset + Width < IntSize) {
3545  Value *Shl = Builder.CreateShl(Src, IntSize - Offset - Width);
3546  Value *RightShift = Signed ? Builder.CreateAShr(Shl, IntSize - Width)
3547  : Builder.CreateLShr(Shl, IntSize - Width);
3548  RightShift->takeName(II);
3549  return replaceInstUsesWith(*II, RightShift);
3550  }
3551 
3552  Value *RightShift = Signed ? Builder.CreateAShr(Src, Offset)
3553  : Builder.CreateLShr(Src, Offset);
3554 
3555  RightShift->takeName(II);
3556  return replaceInstUsesWith(*II, RightShift);
3557  }
3558  case Intrinsic::amdgcn_exp:
3559  case Intrinsic::amdgcn_exp_compr: {
3560  ConstantInt *En = cast<ConstantInt>(II->getArgOperand(1));
3561  unsigned EnBits = En->getZExtValue();
3562  if (EnBits == 0xf)
3563  break; // All inputs enabled.
3564 
3565  bool IsCompr = IID == Intrinsic::amdgcn_exp_compr;
3566  bool Changed = false;
3567  for (int I = 0; I < (IsCompr ? 2 : 4); ++I) {
3568  if ((!IsCompr && (EnBits & (1 << I)) == 0) ||
3569  (IsCompr && ((EnBits & (0x3 << (2 * I))) == 0))) {
3570  Value *Src = II->getArgOperand(I + 2);
3571  if (!isa<UndefValue>(Src)) {
3572  II->setArgOperand(I + 2, UndefValue::get(Src->getType()));
3573  Changed = true;
3574  }
3575  }
3576  }
3577 
3578  if (Changed)
3579  return II;
3580 
3581  break;
3582  }
3583  case Intrinsic::amdgcn_fmed3: {
3584  // Note this does not preserve proper sNaN behavior if IEEE-mode is enabled
3585  // for the shader.
3586 
3587  Value *Src0 = II->getArgOperand(0);
3588  Value *Src1 = II->getArgOperand(1);
3589  Value *Src2 = II->getArgOperand(2);
3590 
3591  // Checking for NaN before canonicalization provides better fidelity when
3592  // mapping other operations onto fmed3 since the order of operands is
3593  // unchanged.
3594  CallInst *NewCall = nullptr;
3595  if (match(Src0, m_NaN()) || isa<UndefValue>(Src0)) {
3596  NewCall = Builder.CreateMinNum(Src1, Src2);
3597  } else if (match(Src1, m_NaN()) || isa<UndefValue>(Src1)) {
3598  NewCall = Builder.CreateMinNum(Src0, Src2);
3599  } else if (match(Src2, m_NaN()) || isa<UndefValue>(Src2)) {
3600  NewCall = Builder.CreateMaxNum(Src0, Src1);
3601  }
3602 
3603  if (NewCall) {
3604  NewCall->copyFastMathFlags(II);
3605  NewCall->takeName(II);
3606  return replaceInstUsesWith(*II, NewCall);
3607  }
3608 
3609  bool Swap = false;
3610  // Canonicalize constants to RHS operands.
3611  //
3612  // fmed3(c0, x, c1) -> fmed3(x, c0, c1)
3613  if (isa<Constant>(Src0) && !isa<Constant>(Src1)) {
3614  std::swap(Src0, Src1);
3615  Swap = true;
3616  }
3617 
3618  if (isa<Constant>(Src1) && !isa<Constant>(Src2)) {
3619  std::swap(Src1, Src2);
3620  Swap = true;
3621  }
3622 
3623  if (isa<Constant>(Src0) && !isa<Constant>(Src1)) {
3624  std::swap(Src0, Src1);
3625  Swap = true;
3626  }
3627 
3628  if (Swap) {
3629  II->setArgOperand(0, Src0);
3630  II->setArgOperand(1, Src1);
3631  II->setArgOperand(2, Src2);
3632  return II;
3633  }
3634 
3635  if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) {
3636  if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) {
3637  if (const ConstantFP *C2 = dyn_cast<ConstantFP>(Src2)) {
3638  APFloat Result = fmed3AMDGCN(C0->getValueAPF(), C1->getValueAPF(),
3639  C2->getValueAPF());
3640  return replaceInstUsesWith(*II,
3641  ConstantFP::get(Builder.getContext(), Result));
3642  }
3643  }
3644  }
3645 
3646  break;
3647  }
3648  case Intrinsic::amdgcn_icmp:
3649  case Intrinsic::amdgcn_fcmp: {
3650  const ConstantInt *CC = cast<ConstantInt>(II->getArgOperand(2));
3651  // Guard against invalid arguments.
3652  int64_t CCVal = CC->getZExtValue();
3653  bool IsInteger = IID == Intrinsic::amdgcn_icmp;
3654  if ((IsInteger && (CCVal < CmpInst::FIRST_ICMP_PREDICATE ||
3655  CCVal > CmpInst::LAST_ICMP_PREDICATE)) ||
3656  (!IsInteger && (CCVal < CmpInst::FIRST_FCMP_PREDICATE ||
3657  CCVal > CmpInst::LAST_FCMP_PREDICATE)))
3658  break;
3659 
3660  Value *Src0 = II->getArgOperand(0);
3661  Value *Src1 = II->getArgOperand(1);
3662 
3663  if (auto *CSrc0 = dyn_cast<Constant>(Src0)) {
3664  if (auto *CSrc1 = dyn_cast<Constant>(Src1)) {
3665  Constant *CCmp = ConstantExpr::getCompare(CCVal, CSrc0, CSrc1);
3666  if (CCmp->isNullValue()) {
3667  return replaceInstUsesWith(
3668  *II, ConstantExpr::getSExt(CCmp, II->getType()));
3669  }
3670 
3671  // The result of V_ICMP/V_FCMP assembly instructions (which this
3672  // intrinsic exposes) is one bit per thread, masked with the EXEC
3673  // register (which contains the bitmask of live threads). So a
3674  // comparison that always returns true is the same as a read of the
3675  // EXEC register.
3677  II->getModule(), Intrinsic::read_register, II->getType());
3678  Metadata *MDArgs[] = {MDString::get(II->getContext(), "exec")};
3679  MDNode *MD = MDNode::get(II->getContext(), MDArgs);
3680  Value *Args[] = {MetadataAsValue::get(II->getContext(), MD)};
3681  CallInst *NewCall = Builder.CreateCall(NewF, Args);
3684  NewCall->takeName(II);
3685  return replaceInstUsesWith(*II, NewCall);
3686  }
3687 
3688  // Canonicalize constants to RHS.
3689  CmpInst::Predicate SwapPred
3690  = CmpInst::getSwappedPredicate(static_cast<CmpInst::Predicate>(CCVal));
3691  II->setArgOperand(0, Src1);
3692  II->setArgOperand(1, Src0);
3693  II->setArgOperand(2, ConstantInt::get(CC->getType(),
3694  static_cast<int>(SwapPred)));
3695  return II;
3696  }
3697 
3698  if (CCVal != CmpInst::ICMP_EQ && CCVal != CmpInst::ICMP_NE)
3699  break;
3700 
3701  // Canonicalize compare eq with true value to compare != 0
3702  // llvm.amdgcn.icmp(zext (i1 x), 1, eq)
3703  // -> llvm.amdgcn.icmp(zext (i1 x), 0, ne)
3704  // llvm.amdgcn.icmp(sext (i1 x), -1, eq)
3705  // -> llvm.amdgcn.icmp(sext (i1 x), 0, ne)
3706  Value *ExtSrc;
3707  if (CCVal == CmpInst::ICMP_EQ &&
3708  ((match(Src1, m_One()) && match(Src0, m_ZExt(m_Value(ExtSrc)))) ||
3709  (match(Src1, m_AllOnes()) && match(Src0, m_SExt(m_Value(ExtSrc))))) &&
3710  ExtSrc->getType()->isIntegerTy(1)) {
3711  II->setArgOperand(1, ConstantInt::getNullValue(Src1->getType()));
3712  II->setArgOperand(2, ConstantInt::get(CC->getType(), CmpInst::ICMP_NE));
3713  return II;
3714  }
3715 
3716  CmpInst::Predicate SrcPred;
3717  Value *SrcLHS;
3718  Value *SrcRHS;
3719 
3720  // Fold compare eq/ne with 0 from a compare result as the predicate to the
3721  // intrinsic. The typical use is a wave vote function in the library, which
3722  // will be fed from a user code condition compared with 0. Fold in the
3723  // redundant compare.
3724 
3725  // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, ne)
3726  // -> llvm.amdgcn.[if]cmp(a, b, pred)
3727  //
3728  // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, eq)
3729  // -> llvm.amdgcn.[if]cmp(a, b, inv pred)
3730  if (match(Src1, m_Zero()) &&
3731  match(Src0,
3732  m_ZExtOrSExt(m_Cmp(SrcPred, m_Value(SrcLHS), m_Value(SrcRHS))))) {
3733  if (CCVal == CmpInst::ICMP_EQ)
3734  SrcPred = CmpInst::getInversePredicate(SrcPred);
3735 
3736  Intrinsic::ID NewIID = CmpInst::isFPPredicate(SrcPred) ?
3737  Intrinsic::amdgcn_fcmp : Intrinsic::amdgcn_icmp;
3738 
3739  Type *Ty = SrcLHS->getType();
3740  if (auto *CmpType = dyn_cast<IntegerType>(Ty)) {
3741  // Promote to next legal integer type.
3742  unsigned Width = CmpType->getBitWidth();
3743  unsigned NewWidth = Width;
3744 
3745  // Don't do anything for i1 comparisons.
3746  if (Width == 1)
3747  break;
3748 
3749  if (Width <= 16)
3750  NewWidth = 16;
3751  else if (Width <= 32)
3752  NewWidth = 32;
3753  else if (Width <= 64)
3754  NewWidth = 64;
3755  else if (Width > 64)
3756  break; // Can't handle this.
3757 
3758  if (Width != NewWidth) {
3759  IntegerType *CmpTy = Builder.getIntNTy(NewWidth);
3760  if (CmpInst::isSigned(SrcPred)) {
3761  SrcLHS = Builder.CreateSExt(SrcLHS, CmpTy);
3762  SrcRHS = Builder.CreateSExt(SrcRHS, CmpTy);
3763  } else {
3764  SrcLHS = Builder.CreateZExt(SrcLHS, CmpTy);
3765  SrcRHS = Builder.CreateZExt(SrcRHS, CmpTy);
3766  }
3767  }
3768  } else if (!Ty->isFloatTy() && !Ty->isDoubleTy() && !Ty->isHalfTy())
3769  break;
3770 
3771  Function *NewF =
3772  Intrinsic::getDeclaration(II->getModule(), NewIID,
3773  { II->getType(),
3774  SrcLHS->getType() });
3775  Value *Args[] = { SrcLHS, SrcRHS,
3776  ConstantInt::get(CC->getType(), SrcPred) };
3777  CallInst *NewCall = Builder.CreateCall(NewF, Args);
3778  NewCall->takeName(II);
3779  return replaceInstUsesWith(*II, NewCall);
3780  }
3781 
3782  break;
3783  }
3784  case Intrinsic::amdgcn_wqm_vote: {
3785  // wqm_vote is identity when the argument is constant.
3786  if (!isa<Constant>(II->getArgOperand(0)))
3787  break;
3788 
3789  return replaceInstUsesWith(*II, II->getArgOperand(0));
3790  }
3791  case Intrinsic::amdgcn_kill: {
3792  const ConstantInt *C = dyn_cast<ConstantInt>(II->getArgOperand(0));
3793  if (!C || !C->getZExtValue())
3794  break;
3795 
3796  // amdgcn.kill(i1 1) is a no-op
3797  return eraseInstFromFunction(CI);
3798  }
3799  case Intrinsic::amdgcn_update_dpp: {
3800  Value *Old = II->getArgOperand(0);
3801 
3802  auto BC = cast<ConstantInt>(II->getArgOperand(5));
3803  auto RM = cast<ConstantInt>(II->getArgOperand(3));
3804  auto BM = cast<ConstantInt>(II->getArgOperand(4));
3805  if (BC->isZeroValue() ||
3806  RM->getZExtValue() != 0xF ||
3807  BM->getZExtValue() != 0xF ||
3808  isa<UndefValue>(Old))
3809  break;
3810 
3811  // If bound_ctrl = 1, row mask = bank mask = 0xf we can omit old value.
3812  II->setOperand(0, UndefValue::get(Old->getType()));
3813  return II;
3814  }
3815  case Intrinsic::amdgcn_readfirstlane:
3816  case Intrinsic::amdgcn_readlane: {
3817  // A constant value is trivially uniform.
3818  if (Constant *C = dyn_cast<Constant>(II->getArgOperand(0)))
3819  return replaceInstUsesWith(*II, C);
3820 
3821  // The rest of these may not be safe if the exec may not be the same between
3822  // the def and use.
3823  Value *Src = II->getArgOperand(0);
3824  Instruction *SrcInst = dyn_cast<Instruction>(Src);
3825  if (SrcInst && SrcInst->getParent() != II->getParent())
3826  break;
3827 
3828  // readfirstlane (readfirstlane x) -> readfirstlane x
3829  // readlane (readfirstlane x), y -> readfirstlane x
3830  if (match(Src, m_Intrinsic<Intrinsic::amdgcn_readfirstlane>()))
3831  return replaceInstUsesWith(*II, Src);
3832 
3833  if (IID == Intrinsic::amdgcn_readfirstlane) {
3834  // readfirstlane (readlane x, y) -> readlane x, y
3835  if (match(Src, m_Intrinsic<Intrinsic::amdgcn_readlane>()))
3836  return replaceInstUsesWith(*II, Src);
3837  } else {
3838  // readlane (readlane x, y), y -> readlane x, y
3839  if (match(Src, m_Intrinsic<Intrinsic::amdgcn_readlane>(
3840  m_Value(), m_Specific(II->getArgOperand(1)))))
3841  return replaceInstUsesWith(*II, Src);
3842  }
3843 
3844  break;
3845  }
3846  case Intrinsic::stackrestore: {
3847  // If the save is right next to the restore, remove the restore. This can
3848  // happen when variable allocas are DCE'd.
3849  if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
3850  if (SS->getIntrinsicID() == Intrinsic::stacksave) {
3851  // Skip over debug info.
3852  if (SS->getNextNonDebugInstruction() == II) {
3853  return eraseInstFromFunction(CI);
3854  }
3855  }
3856  }
3857 
3858  // Scan down this block to see if there is another stack restore in the
3859  // same block without an intervening call/alloca.
3860  BasicBlock::iterator BI(II);
3861  Instruction *TI = II->getParent()->getTerminator();
3862  bool CannotRemove = false;
3863  for (++BI; &*BI != TI; ++BI) {
3864  if (isa<AllocaInst>(BI)) {
3865  CannotRemove = true;
3866  break;
3867  }
3868  if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
3869  if (auto *II2 = dyn_cast<IntrinsicInst>(BCI)) {
3870  // If there is a stackrestore below this one, remove this one.
3871  if (II2->getIntrinsicID() == Intrinsic::stackrestore)
3872  return eraseInstFromFunction(CI);
3873 
3874  // Bail if we cross over an intrinsic with side effects, such as
3875  // llvm.stacksave, llvm.read_register, or llvm.setjmp.
3876  if (II2->mayHaveSideEffects()) {
3877  CannotRemove = true;
3878  break;
3879  }
3880  } else {
3881  // If we found a non-intrinsic call, we can't remove the stack
3882  // restore.
3883  CannotRemove = true;
3884  break;
3885  }
3886  }
3887  }
3888 
3889  // If the stack restore is in a return, resume, or unwind block and if there
3890  // are no allocas or calls between the restore and the return, nuke the
3891  // restore.
3892  if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
3893  return eraseInstFromFunction(CI);
3894  break;
3895  }
3896  case Intrinsic::lifetime_start:
3897  // Asan needs to poison memory to detect invalid access which is possible
3898  // even for empty lifetime range.
3899  if (II->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3900  II->getFunction()->hasFnAttribute(Attribute::SanitizeMemory) ||
3901  II->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3902  break;
3903 
3904  if (removeTriviallyEmptyRange(*II, Intrinsic::lifetime_start,
3905  Intrinsic::lifetime_end, *this))
3906  return nullptr;
3907  break;
3908  case Intrinsic::assume: {
3909  Value *IIOperand = II->getArgOperand(0);
3910  // Remove an assume if it is followed by an identical assume.
3911  // TODO: Do we need this? Unless there are conflicting assumptions, the
3912  // computeKnownBits(IIOperand) below here eliminates redundant assumes.
3914  if (match(Next, m_Intrinsic<Intrinsic::assume>(m_Specific(IIOperand))))
3915  return eraseInstFromFunction(CI);
3916 
3917  // Canonicalize assume(a && b) -> assume(a); assume(b);
3918  // Note: New assumption intrinsics created here are registered by
3919  // the InstCombineIRInserter object.
3920  FunctionType *AssumeIntrinsicTy = II->getFunctionType();
3921  Value *AssumeIntrinsic = II->getCalledValue();
3922  Value *A, *B;
3923  if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) {
3924  Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic, A, II->getName());
3925  Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic, B, II->getName());
3926  return eraseInstFromFunction(*II);
3927  }
3928  // assume(!(a || b)) -> assume(!a); assume(!b);
3929  if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) {
3930  Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic,
3931  Builder.CreateNot(A), II->getName());
3932  Builder.CreateCall(AssumeIntrinsicTy, AssumeIntrinsic,
3933  Builder.CreateNot(B), II->getName());
3934  return eraseInstFromFunction(*II);
3935  }
3936 
3937  // assume( (load addr) != null ) -> add 'nonnull' metadata to load
3938  // (if assume is valid at the load)
3939  CmpInst::Predicate Pred;
3940  Instruction *LHS;
3941  if (match(IIOperand, m_ICmp(Pred, m_Instruction(LHS), m_Zero())) &&
3942  Pred == ICmpInst::ICMP_NE && LHS->getOpcode() == Instruction::Load &&
3943  LHS->getType()->isPointerTy() &&
3944  isValidAssumeForContext(II, LHS, &DT)) {
3945  MDNode *MD = MDNode::get(II->getContext(), None);
3946  LHS->setMetadata(LLVMContext::MD_nonnull, MD);
3947  return eraseInstFromFunction(*II);
3948 
3949  // TODO: apply nonnull return attributes to calls and invokes
3950  // TODO: apply range metadata for range check patterns?
3951  }
3952 
3953  // If there is a dominating assume with the same condition as this one,
3954  // then this one is redundant, and should be removed.
3955  KnownBits Known(1);
3956  computeKnownBits(IIOperand, Known, 0, II);
3957  if (Known.isAllOnes())
3958  return eraseInstFromFunction(*II);
3959 
3960  // Update the cache of affected values for this assumption (we might be
3961  // here because we just simplified the condition).
3962  AC.updateAffectedValues(II);
3963  break;
3964  }
3965  case Intrinsic::experimental_gc_relocate: {
3966  auto &GCR = *cast<GCRelocateInst>(II);
3967 
3968  // If we have two copies of the same pointer in the statepoint argument
3969  // list, canonicalize to one. This may let us common gc.relocates.
3970  if (GCR.getBasePtr() == GCR.getDerivedPtr() &&
3971  GCR.getBasePtrIndex() != GCR.getDerivedPtrIndex()) {
3972  auto *OpIntTy = GCR.getOperand(2)->getType();
3973  II->setOperand(2, ConstantInt::get(OpIntTy, GCR.getBasePtrIndex()));
3974  return II;
3975  }
3976 
3977  // Translate facts known about a pointer before relocating into
3978  // facts about the relocate value, while being careful to
3979  // preserve relocation semantics.
3980  Value *DerivedPtr = GCR.getDerivedPtr();
3981 
3982  // Remove the relocation if unused, note that this check is required
3983  // to prevent the cases below from looping forever.
3984  if (II->use_empty())
3985  return eraseInstFromFunction(*II);
3986 
3987  // Undef is undef, even after relocation.
3988  // TODO: provide a hook for this in GCStrategy. This is clearly legal for
3989  // most practical collectors, but there was discussion in the review thread
3990  // about whether it was legal for all possible collectors.
3991  if (isa<UndefValue>(DerivedPtr))
3992  // Use undef of gc_relocate's type to replace it.
3993  return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
3994 
3995  if (auto *PT = dyn_cast<PointerType>(II->getType())) {
3996  // The relocation of null will be null for most any collector.
3997  // TODO: provide a hook for this in GCStrategy. There might be some
3998  // weird collector this property does not hold for.
3999  if (isa<ConstantPointerNull>(DerivedPtr))
4000  // Use null-pointer of gc_relocate's type to replace it.
4001  return replaceInstUsesWith(*II, ConstantPointerNull::get(PT));
4002 
4003  // isKnownNonNull -> nonnull attribute
4004  if (!II->hasRetAttr(Attribute::NonNull) &&
4005  isKnownNonZero(DerivedPtr, DL, 0, &AC, II, &DT)) {
4006  II->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull);
4007  return II;
4008  }
4009  }
4010 
4011  // TODO: bitcast(relocate(p)) -> relocate(bitcast(p))
4012  // Canonicalize on the type from the uses to the defs
4013 
4014  // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...)
4015  break;
4016  }
4017 
4018  case Intrinsic::experimental_guard: {
4019  // Is this guard followed by another guard? We scan forward over a small
4020  // fixed window of instructions to handle common cases with conditions
4021  // computed between guards.
4022  Instruction *NextInst = II->getNextNode();
4023  for (unsigned i = 0; i < GuardWideningWindow; i++) {
4024  // Note: Using context-free form to avoid compile time blow up
4025  if (!isSafeToSpeculativelyExecute(NextInst))
4026  break;
4027  NextInst = NextInst->getNextNode();
4028  }
4029  Value *NextCond = nullptr;
4030  if (match(NextInst,
4031  m_Intrinsic<Intrinsic::experimental_guard>(m_Value(NextCond)))) {
4032  Value *CurrCond = II->getArgOperand(0);
4033 
4034  // Remove a guard that it is immediately preceded by an identical guard.
4035  if (CurrCond == NextCond)
4036  return eraseInstFromFunction(*NextInst);
4037 
4038  // Otherwise canonicalize guard(a); guard(b) -> guard(a & b).
4039  Instruction* MoveI = II->getNextNode();
4040  while (MoveI != NextInst) {
4041  auto *Temp = MoveI;
4042  MoveI = MoveI->getNextNode();
4043  Temp->moveBefore(II);
4044  }
4045  II->setArgOperand(0, Builder.CreateAnd(CurrCond, NextCond));
4046  return eraseInstFromFunction(*NextInst);
4047  }
4048  break;
4049  }
4050  }
4051  return visitCallBase(*II);
4052 }
4053 
4054 // Fence instruction simplification
4056  // Remove identical consecutive fences.
4058  if (auto *NFI = dyn_cast<FenceInst>(Next))
4059  if (FI.isIdenticalTo(NFI))
4060  return eraseInstFromFunction(FI);
4061  return nullptr;
4062 }
4063 
4064 // InvokeInst simplification
4066  return visitCallBase(II);
4067 }
4068 
4069 // CallBrInst simplification
4071  return visitCallBase(CBI);
4072 }
4073 
4074 /// If this cast does not affect the value passed through the varargs area, we
4075 /// can eliminate the use of the cast.
4076 static bool isSafeToEliminateVarargsCast(const CallBase &Call,
4077  const DataLayout &DL,
4078  const CastInst *const CI,
4079  const int ix) {
4080  if (!CI->isLosslessCast())
4081  return false;
4082 
4083  // If this is a GC intrinsic, avoid munging types. We need types for
4084  // statepoint reconstruction in SelectionDAG.
4085  // TODO: This is probably something which should be expanded to all
4086  // intrinsics since the entire point of intrinsics is that
4087  // they are understandable by the optimizer.
4088  if (isStatepoint(&Call) || isGCRelocate(&Call) || isGCResult(&Call))
4089  return false;
4090 
4091  // The size of ByVal or InAlloca arguments is derived from the type, so we
4092  // can't change to a type with a different size. If the size were
4093  // passed explicitly we could avoid this check.
4094  if (!Call.isByValOrInAllocaArgument(ix))
4095  return true;
4096 
4097  Type* SrcTy =
4098  cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
4099  Type *DstTy = Call.isByValArgument(ix)
4100  ? Call.getParamByValType(ix)
4101  : cast<PointerType>(CI->getType())->getElementType();
4102  if (!SrcTy->isSized() || !DstTy->isSized())
4103  return false;
4104  if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy))
4105  return false;
4106  return true;
4107 }
4108 
4109 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) {
4110  if (!CI->getCalledFunction()) return nullptr;
4111 
4112  auto InstCombineRAUW = [this](Instruction *From, Value *With) {
4113  replaceInstUsesWith(*From, With);
4114  };
4115  auto InstCombineErase = [this](Instruction *I) {
4116  eraseInstFromFunction(*I);
4117  };
4118  LibCallSimplifier Simplifier(DL, &TLI, ORE, BFI, PSI, InstCombineRAUW,
4119  InstCombineErase);
4120  if (Value *With = Simplifier.optimizeCall(CI)) {
4121  ++NumSimplified;
4122  return CI->use_empty() ? CI : replaceInstUsesWith(*CI, With);
4123  }
4124 
4125  return nullptr;
4126 }
4127 
4129  // Strip off at most one level of pointer casts, looking for an alloca. This
4130  // is good enough in practice and simpler than handling any number of casts.
4131  Value *Underlying = TrampMem->stripPointerCasts();
4132  if (Underlying != TrampMem &&
4133  (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
4134  return nullptr;
4135  if (!isa<AllocaInst>(Underlying))
4136  return nullptr;
4137 
4138  IntrinsicInst *InitTrampoline = nullptr;
4139  for (User *U : TrampMem->users()) {
4141  if (!II)
4142  return nullptr;
4143  if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
4144  if (InitTrampoline)
4145  // More than one init_trampoline writes to this value. Give up.
4146  return nullptr;
4147  InitTrampoline = II;
4148  continue;
4149  }
4150  if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
4151  // Allow any number of calls to adjust.trampoline.
4152  continue;
4153  return nullptr;
4154  }
4155 
4156  // No call to init.trampoline found.
4157  if (!InitTrampoline)
4158  return nullptr;
4159 
4160  // Check that the alloca is being used in the expected way.
4161  if (InitTrampoline->getOperand(0) != TrampMem)
4162  return nullptr;
4163 
4164  return InitTrampoline;
4165 }
4166 
4168  Value *TrampMem) {
4169  // Visit all the previous instructions in the basic block, and try to find a
4170  // init.trampoline which has a direct path to the adjust.trampoline.
4171  for (BasicBlock::iterator I = AdjustTramp->getIterator(),
4172  E = AdjustTramp->getParent()->begin();
4173  I != E;) {
4174  Instruction *Inst = &*--I;
4175  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
4176  if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
4177  II->getOperand(0) == TrampMem)
4178  return II;
4179  if (Inst->mayWriteToMemory())
4180  return nullptr;
4181  }
4182  return nullptr;
4183 }
4184 
4185 // Given a call to llvm.adjust.trampoline, find and return the corresponding
4186 // call to llvm.init.trampoline if the call to the trampoline can be optimized
4187 // to a direct call to a function. Otherwise return NULL.
4189  Callee = Callee->stripPointerCasts();
4190  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
4191  if (!AdjustTramp ||
4192  AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
4193  return nullptr;
4194 
4195  Value *TrampMem = AdjustTramp->getOperand(0);
4196 
4198  return IT;
4199  if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem))
4200  return IT;
4201  return nullptr;
4202 }
4203 
4204 static void annotateAnyAllocSite(CallBase &Call, const TargetLibraryInfo *TLI) {
4205  unsigned NumArgs = Call.getNumArgOperands();
4206  ConstantInt *Op0C = dyn_cast<ConstantInt>(Call.getOperand(0));
4207  ConstantInt *Op1C =
4208  (NumArgs == 1) ? nullptr : dyn_cast<ConstantInt>(Call.getOperand(1));
4209  // Bail out if the allocation size is zero.
4210  if ((Op0C && Op0C->isNullValue()) || (Op1C && Op1C->isNullValue()))
4211  return;
4212 
4213  if (isMallocLikeFn(&Call, TLI) && Op0C) {
4214  if (isOpNewLikeFn(&Call, TLI))
4217  Call.getContext(), Op0C->getZExtValue()));
4218  else
4221  Call.getContext(), Op0C->getZExtValue()));
4222  } else if (isReallocLikeFn(&Call, TLI) && Op1C) {
4225  Call.getContext(), Op1C->getZExtValue()));
4226  } else if (isCallocLikeFn(&Call, TLI) && Op0C && Op1C) {
4227  bool Overflow;
4228  const APInt &N = Op0C->getValue();
4229  APInt Size = N.umul_ov(Op1C->getValue(), Overflow);
4230  if (!Overflow)
4233  Call.getContext(), Size.getZExtValue()));
4234  } else if (isStrdupLikeFn(&Call, TLI)) {
4235  uint64_t Len = GetStringLength(Call.getOperand(0));
4236  if (Len) {
4237  // strdup
4238  if (NumArgs == 1)
4241  Call.getContext(), Len));
4242  // strndup
4243  else if (NumArgs == 2 && Op1C)
4244  Call.addAttribute(
4247  Call.getContext(), std::min(Len, Op1C->getZExtValue() + 1)));
4248  }
4249  }
4250 }
4251 
4252 /// Improvements for call, callbr and invoke instructions.
4253 Instruction *InstCombiner::visitCallBase(CallBase &Call) {
4254  if (isAllocationFn(&Call, &TLI))
4255  annotateAnyAllocSite(Call, &TLI);
4256 
4257  bool Changed = false;
4258 
4259  // Mark any parameters that are known to be non-null with the nonnull
4260  // attribute. This is helpful for inlining calls to functions with null
4261  // checks on their arguments.
4262  SmallVector<unsigned, 4> ArgNos;
4263  unsigned ArgNo = 0;
4264 
4265  for (Value *V : Call.args()) {
4266  if (V->getType()->isPointerTy() &&
4267  !Call.paramHasAttr(ArgNo, Attribute::NonNull) &&
4268  isKnownNonZero(V, DL, 0, &AC, &Call, &DT))
4269  ArgNos.push_back(ArgNo);
4270  ArgNo++;
4271  }
4272 
4273  assert(ArgNo == Call.arg_size() && "sanity check");
4274 
4275  if (!ArgNos.empty()) {
4276  AttributeList AS = Call.getAttributes();
4277  LLVMContext &Ctx = Call.getContext();
4278  AS = AS.addParamAttribute(Ctx, ArgNos,
4279  Attribute::get(Ctx, Attribute::NonNull));
4280  Call.setAttributes(AS);
4281  Changed = true;
4282  }
4283 
4284  // If the callee is a pointer to a function, attempt to move any casts to the
4285  // arguments of the call/callbr/invoke.
4286  Value *Callee = Call.getCalledValue();
4287  if (!isa<Function>(Callee) && transformConstExprCastCall(Call))
4288  return nullptr;
4289 
4290  if (Function *CalleeF = dyn_cast<Function>(Callee)) {
4291  // Remove the convergent attr on calls when the callee is not convergent.
4292  if (Call.isConvergent() && !CalleeF->isConvergent() &&
4293  !CalleeF->isIntrinsic()) {
4294  LLVM_DEBUG(dbgs() << "Removing convergent attr from instr " << Call
4295  << "\n");
4296  Call.setNotConvergent();
4297  return &Call;
4298  }
4299 
4300  // If the call and callee calling conventions don't match, this call must
4301  // be unreachable, as the call is undefined.
4302  if (CalleeF->getCallingConv() != Call.getCallingConv() &&
4303  // Only do this for calls to a function with a body. A prototype may
4304  // not actually end up matching the implementation's calling conv for a
4305  // variety of reasons (e.g. it may be written in assembly).
4306  !CalleeF->isDeclaration()) {
4307  Instruction *OldCall = &Call;
4308  CreateNonTerminatorUnreachable(OldCall);
4309  // If OldCall does not return void then replaceAllUsesWith undef.
4310  // This allows ValueHandlers and custom metadata to adjust itself.
4311  if (!OldCall->getType()->isVoidTy())
4312  replaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
4313  if (isa<CallInst>(OldCall))
4314  return eraseInstFromFunction(*OldCall);
4315 
4316  // We cannot remove an invoke or a callbr, because it would change thexi
4317  // CFG, just change the callee to a null pointer.
4318  cast<CallBase>(OldCall)->setCalledFunction(
4319  CalleeF->getFunctionType(),
4320  Constant::getNullValue(CalleeF->getType()));
4321  return nullptr;
4322  }
4323  }
4324 
4325  if ((isa<ConstantPointerNull>(Callee) &&
4326  !NullPointerIsDefined(Call.getFunction())) ||
4327  isa<UndefValue>(Callee)) {
4328  // If Call does not return void then replaceAllUsesWith undef.
4329  // This allows ValueHandlers and custom metadata to adjust itself.
4330  if (!Call.getType()->isVoidTy())
4331  replaceInstUsesWith(Call, UndefValue::get(Call.getType()));
4332 
4333  if (Call.isTerminator()) {
4334  // Can't remove an invoke or callbr because we cannot change the CFG.
4335  return nullptr;
4336  }
4337 
4338  // This instruction is not reachable, just remove it.
4339  CreateNonTerminatorUnreachable(&Call);
4340  return eraseInstFromFunction(Call);
4341  }
4342 
4343  if (IntrinsicInst *II = findInitTrampoline(Callee))
4344  return transformCallThroughTrampoline(Call, *II);
4345 
4346  PointerType *PTy = cast<PointerType>(Callee->getType());
4347  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
4348  if (FTy->isVarArg()) {
4349  int ix = FTy->getNumParams();
4350  // See if we can optimize any arguments passed through the varargs area of
4351  // the call.
4352  for (auto I = Call.arg_begin() + FTy->getNumParams(), E = Call.arg_end();
4353  I != E; ++I, ++ix) {
4354  CastInst *CI = dyn_cast<CastInst>(*I);
4355  if (CI && isSafeToEliminateVarargsCast(Call, DL, CI, ix)) {
4356  *I = CI->getOperand(0);
4357 
4358  // Update the byval type to match the argument type.
4359  if (Call.isByValArgument(ix)) {
4360  Call.removeParamAttr(ix, Attribute::ByVal);
4361  Call.addParamAttr(
4363  Call.getContext(),
4364  CI->getOperand(0)->getType()->getPointerElementType()));
4365  }
4366  Changed = true;
4367  }
4368  }
4369  }
4370 
4371  if (isa<InlineAsm>(Callee) && !Call.doesNotThrow()) {
4372  // Inline asm calls cannot throw - mark them 'nounwind'.
4373  Call.setDoesNotThrow();
4374  Changed = true;
4375  }
4376 
4377  // Try to optimize the call if possible, we require DataLayout for most of
4378  // this. None of these calls are seen as possibly dead so go ahead and
4379  // delete the instruction now.
4380  if (CallInst *CI = dyn_cast<CallInst>(&Call)) {
4381  Instruction *I = tryOptimizeCall(CI);
4382  // If we changed something return the result, etc. Otherwise let
4383  // the fallthrough check.
4384  if (I) return eraseInstFromFunction(*I);
4385  }
4386 
4387  if (isAllocLikeFn(&Call, &TLI))
4388  return visitAllocSite(Call);
4389 
4390  return Changed ? &Call : nullptr;
4391 }
4392 
4393 /// If the callee is a constexpr cast of a function, attempt to move the cast to
4394 /// the arguments of the call/callbr/invoke.
4395 bool InstCombiner::transformConstExprCastCall(CallBase &Call) {
4397  if (!Callee)
4398  return false;
4399 
4400  // If this is a call to a thunk function, don't remove the cast. Thunks are
4401  // used to transparently forward all incoming parameters and outgoing return
4402  // values, so it's important to leave the cast in place.
4403  if (Callee->hasFnAttribute("thunk"))
4404  return false;
4405 
4406  // If this is a musttail call, the callee's prototype must match the caller's
4407  // prototype with the exception of pointee types. The code below doesn't
4408  // implement that, so we can't do this transform.
4409  // TODO: Do the transform if it only requires adding pointer casts.
4410  if (Call.isMustTailCall())
4411  return false;
4412 
4413  Instruction *Caller = &Call;
4414  const AttributeList &CallerPAL = Call.getAttributes();
4415 
4416  // Okay, this is a cast from a function to a different type. Unless doing so
4417  // would cause a type conversion of one of our arguments, change this call to
4418  // be a direct call with arguments casted to the appropriate types.
4420  Type *OldRetTy = Caller->getType();
4421  Type *NewRetTy = FT->getReturnType();
4422 
4423  // Check to see if we are changing the return type...
4424  if (OldRetTy != NewRetTy) {
4425 
4426  if (NewRetTy->isStructTy())
4427  return false; // TODO: Handle multiple return values.
4428 
4429  if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) {
4430  if (Callee->isDeclaration())
4431  return false; // Cannot transform this return value.
4432 
4433  if (!Caller->use_empty() &&
4434  // void -> non-void is handled specially
4435  !NewRetTy->isVoidTy())
4436  return false; // Cannot transform this return value.
4437  }
4438 
4439  if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
4440  AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex);
4441  if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy)))
4442  return false; // Attribute not compatible with transformed value.
4443  }
4444 
4445  // If the callbase is an invoke/callbr instruction, and the return value is
4446  // used by a PHI node in a successor, we cannot change the return type of
4447  // the call because there is no place to put the cast instruction (without
4448  // breaking the critical edge). Bail out in this case.
4449  if (!Caller->use_empty()) {
4450  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
4451  for (User *U : II->users())
4452  if (PHINode *PN = dyn_cast<PHINode>(U))
4453  if (PN->getParent() == II->getNormalDest() ||
4454  PN->getParent() == II->getUnwindDest())
4455  return false;
4456  // FIXME: Be conservative for callbr to avoid a quadratic search.
4457  if (isa<CallBrInst>(Caller))
4458  return false;
4459  }
4460  }
4461 
4462  unsigned NumActualArgs = Call.arg_size();
4463  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
4464 
4465  // Prevent us turning:
4466  // declare void @takes_i32_inalloca(i32* inalloca)
4467  // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0)
4468  //
4469  // into:
4470  // call void @takes_i32_inalloca(i32* null)
4471  //
4472  // Similarly, avoid folding away bitcasts of byval calls.
4473  if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) ||
4474  Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal))
4475  return false;
4476 
4477  auto AI = Call.arg_begin();
4478  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
4479  Type *ParamTy = FT->getParamType(i);
4480  Type *ActTy = (*AI)->getType();
4481 
4482  if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL))
4483  return false; // Cannot transform this parameter value.
4484 
4485  if (AttrBuilder(CallerPAL.getParamAttributes(i))
4486  .overlaps(AttributeFuncs::typeIncompatible(ParamTy)))
4487  return false; // Attribute not compatible with transformed value.
4488 
4489  if (Call.isInAllocaArgument(i))
4490  return false; // Cannot transform to and from inalloca.
4491 
4492  // If the parameter is passed as a byval argument, then we have to have a
4493  // sized type and the sized type has to have the same size as the old type.
4494  if (ParamTy != ActTy && CallerPAL.hasParamAttribute(i, Attribute::ByVal)) {
4495  PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
4496  if (!ParamPTy || !ParamPTy->getElementType()->isSized())
4497  return false;
4498 
4499  Type *CurElTy = Call.getParamByValType(i);
4500  if (DL.getTypeAllocSize(CurElTy) !=
4501  DL.getTypeAllocSize(ParamPTy->getElementType()))
4502  return false;
4503  }
4504  }
4505 
4506  if (Callee->isDeclaration()) {
4507  // Do not delete arguments unless we have a function body.
4508  if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
4509  return false;
4510 
4511  // If the callee is just a declaration, don't change the varargsness of the
4512  // call. We don't want to introduce a varargs call where one doesn't
4513  // already exist.
4514  PointerType *APTy = cast<PointerType>(Call.getCalledValue()->getType());
4515  if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
4516  return false;
4517 
4518  // If both the callee and the cast type are varargs, we still have to make
4519  // sure the number of fixed parameters are the same or we have the same
4520  // ABI issues as if we introduce a varargs call.
4521  if (FT->isVarArg() &&
4522  cast<FunctionType>(APTy->getElementType())->isVarArg() &&
4523  FT->getNumParams() !=
4524  cast<FunctionType>(APTy->getElementType())->getNumParams())
4525  return false;
4526  }
4527 
4528  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
4529  !CallerPAL.isEmpty()) {
4530  // In this case we have more arguments than the new function type, but we
4531  // won't be dropping them. Check that these extra arguments have attributes
4532  // that are compatible with being a vararg call argument.
4533  unsigned SRetIdx;
4534  if (CallerPAL.hasAttrSomewhere(Attribute::StructRet, &SRetIdx) &&
4535  SRetIdx > FT->getNumParams())
4536  return false;
4537  }
4538 
4539  // Okay, we decided that this is a safe thing to do: go ahead and start
4540  // inserting cast instructions as necessary.
4543  Args.reserve(NumActualArgs);
4544  ArgAttrs.reserve(NumActualArgs);
4545 
4546  // Get any return attributes.
4547  AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex);
4548 
4549  // If the return value is not being used, the type may not be compatible
4550  // with the existing attributes. Wipe out any problematic attributes.
4551  RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy));
4552 
4553  LLVMContext &Ctx = Call.getContext();
4554  AI = Call.arg_begin();
4555  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
4556  Type *ParamTy = FT->getParamType(i);
4557 
4558  Value *NewArg = *AI;
4559  if ((*AI)->getType() != ParamTy)
4560  NewArg = Builder.CreateBitOrPointerCast(*AI, ParamTy);
4561  Args.push_back(NewArg);
4562 
4563  // Add any parameter attributes.
4564  if (CallerPAL.hasParamAttribute(i, Attribute::ByVal)) {
4565  AttrBuilder AB(CallerPAL.getParamAttributes(i));
4566  AB.addByValAttr(NewArg->getType()->getPointerElementType());
4567  ArgAttrs.push_back(AttributeSet::get(Ctx, AB));
4568  } else
4569  ArgAttrs.push_back(CallerPAL.getParamAttributes(i));
4570  }
4571 
4572  // If the function takes more arguments than the call was taking, add them
4573  // now.
4574  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) {
4576  ArgAttrs.push_back(AttributeSet());
4577  }
4578 
4579  // If we are removing arguments to the function, emit an obnoxious warning.
4580  if (FT->getNumParams() < NumActualArgs) {
4581  // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
4582  if (FT->isVarArg()) {
4583  // Add all of the arguments in their promoted form to the arg list.
4584  for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
4585  Type *PTy = getPromotedType((*AI)->getType());
4586  Value *NewArg = *AI;
4587  if (PTy != (*AI)->getType()) {
4588  // Must promote to pass through va_arg area!
4589  Instruction::CastOps opcode =
4590  CastInst::getCastOpcode(*AI, false, PTy, false);
4591  NewArg = Builder.CreateCast(opcode, *AI, PTy);
4592  }
4593  Args.push_back(NewArg);
4594 
4595  // Add any parameter attributes.
4596  ArgAttrs.push_back(CallerPAL.getParamAttributes(i));
4597  }
4598  }
4599  }
4600 
4601  AttributeSet FnAttrs = CallerPAL.getFnAttributes();
4602 
4603  if (NewRetTy->isVoidTy())
4604  Caller->setName(""); // Void type should not have a name.
4605 
4606  assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) &&
4607  "missing argument attributes");
4608  AttributeList NewCallerPAL = AttributeList::get(
4609  Ctx, FnAttrs, AttributeSet::get(Ctx, RAttrs), ArgAttrs);
4610 
4612  Call.getOperandBundlesAsDefs(OpBundles);
4613 
4614  CallBase *NewCall;
4615  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4616  NewCall = Builder.CreateInvoke(Callee, II->getNormalDest(),
4617  II->getUnwindDest(), Args, OpBundles);
4618  } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(Caller)) {
4619  NewCall = Builder.CreateCallBr(Callee, CBI->getDefaultDest(),
4620  CBI->getIndirectDests(), Args, OpBundles);
4621  } else {
4622  NewCall = Builder.CreateCall(Callee, Args, OpBundles);
4623  cast<CallInst>(NewCall)->setTailCallKind(
4624  cast<CallInst>(Caller)->getTailCallKind());
4625  }
4626  NewCall->takeName(Caller);
4627  NewCall->setCallingConv(Call.getCallingConv());
4628  NewCall->setAttributes(NewCallerPAL);
4629 
4630  // Preserve the weight metadata for the new call instruction. The metadata
4631  // is used by SamplePGO to check callsite's hotness.
4632  uint64_t W;
4633  if (Caller->extractProfTotalWeight(W))
4634  NewCall->setProfWeight(W);
4635 
4636  // Insert a cast of the return type as necessary.
4637  Instruction *NC = NewCall;
4638  Value *NV = NC;
4639  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
4640  if (!NV->getType()->isVoidTy()) {
4641  NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy);
4642  NC->setDebugLoc(Caller->getDebugLoc());
4643 
4644  // If this is an invoke/callbr instruction, we should insert it after the
4645  // first non-phi instruction in the normal successor block.
4646  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
4647  BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
4648  InsertNewInstBefore(NC, *I);
4649  } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(Caller)) {
4650  BasicBlock::iterator I = CBI->getDefaultDest()->getFirstInsertionPt();
4651  InsertNewInstBefore(NC, *I);
4652  } else {
4653  // Otherwise, it's a call, just insert cast right after the call.
4654  InsertNewInstBefore(NC, *Caller);
4655  }
4656  Worklist.AddUsersToWorkList(*Caller);
4657  } else {
4658  NV = UndefValue::get(Caller->getType());
4659  }
4660  }
4661 
4662  if (!Caller->use_empty())
4663  replaceInstUsesWith(*Caller, NV);
4664  else if (Caller->hasValueHandle()) {
4665  if (OldRetTy == NV->getType())
4666  ValueHandleBase::ValueIsRAUWd(Caller, NV);
4667  else
4668  // We cannot call ValueIsRAUWd with a different type, and the
4669  // actual tracked value will disappear.
4671  }
4672 
4673  eraseInstFromFunction(*Caller);
4674  return true;
4675 }
4676 
4677 /// Turn a call to a function created by init_trampoline / adjust_trampoline
4678 /// intrinsic pair into a direct call to the underlying function.
4679 Instruction *
4680 InstCombiner::transformCallThroughTrampoline(CallBase &Call,
4681  IntrinsicInst &Tramp) {
4682  Value *Callee = Call.getCalledValue();
4683  Type *CalleeTy = Callee->getType();
4684  FunctionType *FTy = Call.getFunctionType();
4686 
4687  // If the call already has the 'nest' attribute somewhere then give up -
4688  // otherwise 'nest' would occur twice after splicing in the chain.
4689  if (Attrs.hasAttrSomewhere(Attribute::Nest))
4690  return nullptr;
4691 
4692  Function *NestF = cast<Function>(Tramp.getArgOperand(1)->stripPointerCasts());
4693  FunctionType *NestFTy = NestF->getFunctionType();
4694 
4695  AttributeList NestAttrs = NestF->getAttributes();
4696  if (!NestAttrs.isEmpty()) {
4697  unsigned NestArgNo = 0;
4698  Type *NestTy = nullptr;
4699  AttributeSet NestAttr;
4700 
4701  // Look for a parameter marked with the 'nest' attribute.
4702  for (FunctionType::param_iterator I = NestFTy->param_begin(),
4703  E = NestFTy->param_end();
4704  I != E; ++NestArgNo, ++I) {
4705  AttributeSet AS = NestAttrs.getParamAttributes(NestArgNo);
4706  if (AS.hasAttribute(Attribute::Nest)) {
4707  // Record the parameter type and any other attributes.
4708  NestTy = *I;
4709  NestAttr = AS;
4710  break;
4711  }
4712  }
4713 
4714  if (NestTy) {
4715  std::vector<Value*> NewArgs;
4716  std::vector<AttributeSet> NewArgAttrs;
4717  NewArgs.reserve(Call.arg_size() + 1);
4718  NewArgAttrs.reserve(Call.arg_size());
4719 
4720  // Insert the nest argument into the call argument list, which may
4721  // mean appending it. Likewise for attributes.
4722 
4723  {
4724  unsigned ArgNo = 0;
4725  auto I = Call.arg_begin(), E = Call.arg_end();
4726  do {
4727  if (ArgNo == NestArgNo) {
4728  // Add the chain argument and attributes.
4729  Value *NestVal = Tramp.getArgOperand(2);
4730  if (NestVal->getType() != NestTy)
4731  NestVal = Builder.CreateBitCast(NestVal, NestTy, "nest");
4732  NewArgs.push_back(NestVal);
4733  NewArgAttrs.push_back(NestAttr);
4734  }
4735 
4736  if (I == E)
4737  break;
4738 
4739  // Add the original argument and attributes.
4740  NewArgs.push_back(*I);
4741  NewArgAttrs.push_back(Attrs.getParamAttributes(ArgNo));
4742 
4743  ++ArgNo;
4744  ++I;
4745  } while (true);
4746  }
4747 
4748  // The trampoline may have been bitcast to a bogus type (FTy).
4749  // Handle this by synthesizing a new function type, equal to FTy
4750  // with the chain parameter inserted.
4751 
4752  std::vector<Type*> NewTypes;
4753  NewTypes.reserve(FTy->getNumParams()+1);
4754 
4755  // Insert the chain's type into the list of parameter types, which may
4756  // mean appending it.
4757  {
4758  unsigned ArgNo = 0;
4760  E = FTy->param_end();
4761 
4762  do {
4763  if (ArgNo == NestArgNo)
4764  // Add the chain's type.
4765  NewTypes.push_back(NestTy);
4766 
4767  if (I == E)
4768  break;
4769 
4770  // Add the original type.
4771  NewTypes.push_back(*I);
4772 
4773  ++ArgNo;
4774  ++I;
4775  } while (true);
4776  }
4777 
4778  // Replace the trampoline call with a direct call. Let the generic
4779  // code sort out any function type mismatches.
4780  FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
4781  FTy->isVarArg());
4782  Constant *NewCallee =
4783  NestF->getType() == PointerType::getUnqual(NewFTy) ?
4784  NestF : ConstantExpr::getBitCast(NestF,
4785  PointerType::getUnqual(NewFTy));
4786  AttributeList NewPAL =
4788  Attrs.getRetAttributes(), NewArgAttrs);
4789 
4791  Call.getOperandBundlesAsDefs(OpBundles);
4792 
4793  Instruction *NewCaller;
4794  if (InvokeInst *II = dyn_cast<InvokeInst>(&Call)) {
4795  NewCaller = InvokeInst::Create(NewFTy, NewCallee,
4796  II->getNormalDest(), II->getUnwindDest(),
4797  NewArgs, OpBundles);
4798  cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
4799  cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
4800  } else if (CallBrInst *CBI = dyn_cast<CallBrInst>(&Call)) {
4801  NewCaller =
4802  CallBrInst::Create(NewFTy, NewCallee, CBI->getDefaultDest(),
4803  CBI->getIndirectDests(), NewArgs, OpBundles);
4804  cast<CallBrInst>(NewCaller)->setCallingConv(CBI->getCallingConv());
4805  cast<CallBrInst>(NewCaller)->setAttributes(NewPAL);
4806  } else {
4807  NewCaller = CallInst::Create(NewFTy, NewCallee, NewArgs, OpBundles);
4808  cast<CallInst>(NewCaller)->setTailCallKind(
4809  cast<CallInst>(Call).getTailCallKind());
4810  cast<CallInst>(NewCaller)->setCallingConv(
4811  cast<CallInst>(Call).getCallingConv());
4812  cast<CallInst>(NewCaller)->setAttributes(NewPAL);
4813  }
4814  NewCaller->setDebugLoc(Call.getDebugLoc());
4815 
4816  return NewCaller;
4817  }
4818  }
4819 
4820  // Replace the trampoline call with a direct call. Since there is no 'nest'
4821  // parameter, there is no need to adjust the argument list. Let the generic
4822  // code sort out any function type mismatches.
4823  Constant *NewCallee = ConstantExpr::getBitCast(NestF, CalleeTy);
4824  Call.setCalledFunction(FTy, NewCallee);
4825  return &Call;
4826 }
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
bool isFPPredicate() const
Definition: InstrTypes.h:824
const NoneType None
Definition: None.h:23
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:761
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
Definition: PatternMatch.h:874
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...
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:171
bool isAllocationFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates or reallocates memory (eith...
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:598
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:70
static void ValueIsDeleted(Value *V)
Definition: Value.cpp:870
class_match< UndefValue > m_Undef()
Match an arbitrary undef constant.
Definition: PatternMatch.h:86
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
bool isZero() const
Definition: APFloat.h:1158
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:177
class_match< CmpInst > m_Cmp()
Matches any compare instruction and ignore it.
Definition: PatternMatch.h:78
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1571
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:1181
APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:888
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:561
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:403
DiagnosticInfoOptimizationBase::Argument NV
Atomic ordering constants.
Value * CreateAddrSpaceCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1968
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
Definition: IRBuilder.h:1887
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:288
static Type * makeCmpResultType(Type *opnd_type)
Create a result type for fcmp/icmp.
Definition: InstrTypes.h:975
This class represents lattice values for constants.
Definition: AllocatorList.h:23
Type * getParamType(unsigned i) const
Parameter type accessors.
Definition: DerivedTypes.h:140
Constant * getElementAsConstant(unsigned i) const
Return a Constant for a specified index&#39;s element.
Definition: Constants.cpp:2773
unsigned countMinPopulation() const
Returns the number of bits known to be one.
Definition: KnownBits.h:196
static Value * simplifyX86pack(IntrinsicInst &II, InstCombiner::BuilderTy &Builder, bool IsSigned)
Represents an op.with.overflow intrinsic.
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:66
m_Intrinsic_Ty< Opnd0 >::Ty m_BitReverse(const Opnd0 &Op0)
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
LoadInst * CreateAlignedLoad(Type *Ty, Value *Ptr, unsigned Align, const char *Name)
Provided to resolve &#39;CreateAlignedLoad(Ptr, Align, "...")&#39; correctly, instead of converting the strin...
Definition: IRBuilder.h:1611
bool isConvergent() const
Determine if the invoke is convergent.
Definition: InstrTypes.h:1701
static Attribute getWithDereferenceableBytes(LLVMContext &Context, uint64_t Bytes)
Definition: Attributes.cpp:155
An instruction for ordering other memory operations.
Definition: Instructions.h:460
static MDString * get(LLVMContext &Context, StringRef Str)
Definition: Metadata.cpp:453
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
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:1363
void setProfWeight(uint64_t W)
Sets the branch_weights metadata to W for CallInst.
Value * CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2133
APInt uadd_sat(const APInt &RHS) const
Definition: APInt.cpp:2023
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:647
void setOrdering(AtomicOrdering Ordering)
Sets the ordering constraint of this load instruction.
Definition: Instructions.h:258
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:89
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:637
static uint64_t round(uint64_t Acc, uint64_t Input)
Definition: xxhash.cpp:57
Optional< std::vector< StOtherPiece > > Other
Definition: ELFYAML.cpp:953
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:506
bool mayWriteToMemory() const
Return true if this instruction may modify memory.
bool isSubsetOf(const APInt &RHS) const
This operation checks that all bits set in this APInt are also set in RHS.
Definition: APInt.h:1328
bool isDereferenceableAndAlignedPointer(const Value *V, Type *Ty, MaybeAlign Alignment, const DataLayout &DL, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Returns true if V is always a dereferenceable pointer with alignment greater or equal than requested...
Definition: Loads.cpp:138
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)
Always overflows in the direction of signed/unsigned min value.
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:743
bool isTerminator() const
Definition: Instruction.h:128
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
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:1165
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:1246
Metadata node.
Definition: Metadata.h:863
Base class for all callable instructions (InvokeInst and CallInst) Holds everything related to callin...
Definition: InstrTypes.h:1100
F(f)
Type * getStructElementType(unsigned N) const
Definition: DerivedTypes.h:370
User::op_iterator arg_end()
Return the iterator pointing to the end of the argument list.
Definition: InstrTypes.h:1212
const fltSemantics & getSemantics() const
Definition: APFloat.h:1170
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:635
param_iterator param_end() const
Definition: DerivedTypes.h:134
BinaryOp_match< LHS, RHS, Instruction::FSub > m_FSub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:777
An instruction for reading from memory.
Definition: Instructions.h:169
bool isReallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that reallocates memory (e...
static IntegerType * getInt64Ty(LLVMContext &C)
Definition: Type.cpp:181
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:930
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:1968
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2261
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:144
unsigned countMaxTrailingZeros() const
Returns the maximum number of trailing zero bits possible.
Definition: KnownBits.h:176
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:230
LLVM_READONLY APFloat maximum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2018 maximum semantics.
Definition: APFloat.h:1277
void reserve(size_type N)
Definition: SmallVector.h:369
void addAttribute(unsigned i, Attribute::AttrKind Kind)
adds the attribute to the list of attributes.
Definition: InstrTypes.h:1383
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...
void setAlignment(MaybeAlign Align)
static CallBrInst * Create(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest, ArrayRef< BasicBlock *> IndirectDests, ArrayRef< Value *> Args, const Twine &NameStr, Instruction *InsertBefore=nullptr)
cst_pred_ty< is_zero_int > m_ZeroInt()
Match an integer 0 or a vector with all elements equal to 0.
Definition: PatternMatch.h:391
Instruction * visitVAStartInst(VAStartInst &I)
static APInt getSignedMaxValue(unsigned numBits)
Gets maximum signed value of APInt for a specific bit width.
Definition: APInt.h:534
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1517
Value * CreateLaunderInvariantGroup(Value *Ptr)
Create a launder.invariant.group intrinsic call.
Definition: IRBuilder.h:2410
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:289
unsigned countMinTrailingZeros() const
Returns the minimum number of trailing zero bits.
Definition: KnownBits.h:146
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:273
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:1241
static BinaryOperator * CreateNSW(BinaryOps Opc, Value *V1, Value *V2, const Twine &Name="")
Definition: InstrTypes.h:289
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
opStatus divide(const APFloat &RHS, roundingMode RM)
Definition: APFloat.h:983
static Instruction * SimplifyNVVMIntrinsic(IntrinsicInst *II, InstCombiner &IC)
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:47
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:1644
bool isSigned() const
Definition: InstrTypes.h:902
APInt getLoBits(unsigned numBits) const
Compute an APInt containing numBits lowbits from this APInt.
Definition: APInt.cpp:570
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:886
Type * getPointerElementType() const
Definition: Type.h:381
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:831
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:80
static Value * simplifyX86movmsk(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Absolute value.
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:439
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:368
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:450
APInt shl(unsigned shiftAmt) const
Left-shift function.
Definition: APInt.h:992
static Value * simplifyNeonTbl1(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Convert a table lookup to shufflevector if the mask is constant.
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
static Attribute getWithDereferenceableOrNullBytes(LLVMContext &Context, uint64_t Bytes)
Definition: Attributes.cpp:161
The core instruction combiner logic.
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:1264
AttributeSet getRetAttributes() const
The attributes for the ret value are returned.
static Constant * getSExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1682
AttrBuilder & addByValAttr(Type *Ty)
This turns a byval type into the form used internally in Attribute.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:285
Value * CreateICmpSGT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2125
uint64_t getNumElements() const
For scalable vectors, this will return the minimum number of elements in the vector.
Definition: DerivedTypes.h:398
void lshrInPlace(unsigned ShiftAmt)
Logical right-shift this APInt by ShiftAmt in place.
Definition: APInt.h:977
bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const
Determine whether the argument or parameter has the given attribute.
This file implements a class to represent arbitrary precision integral constant values and operations...
All zero aggregate value.
Definition: Constants.h:340
static Value * simplifyX86vpermv(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
Metadata * LowAndHigh[]
static Value * simplifyX86addcarry(const IntrinsicInst &II, InstCombiner::BuilderTy &Builder)
Value * 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.
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:201
Key
PAL metadata keys.
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWAdd(const LHS &L, const RHS &R)
Definition: PatternMatch.h:970
bool doesNotThrow() const
Determine if the call cannot unwind.
Definition: InstrTypes.h:1689
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:108
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:1963
bool isInfinity() const
Definition: APFloat.h:1159
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:246
cstfp_pred_ty< is_nan > m_NaN()
Match an arbitrary NaN constant.
Definition: PatternMatch.h:506
apfloat_match m_APFloat(const APFloat *&Res)
Match a ConstantFP or splatted ConstantVector, binding the specified pointer to the contained APFloat...
Definition: PatternMatch.h:198
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:4483
May or may not overflow.
bool isStatepoint(const CallBase *Call)
Definition: Statepoint.cpp:20
void setCalledFunction(Function *Fn)
Sets the function called, including updating the function type.
Definition: InstrTypes.h:1323
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:128
This class represents a no-op cast from one type to another.
bool isOpNewLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates memory and throws if an all...
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:244
Value * CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name="")
Return a vector value that contains.
Definition: IRBuilder.h:2464
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:137
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
bool maskIsAllOneOrUndef(Value *Mask)
Given a mask vector of the form <Y x="" i1>="">, Return true if all of the elements of this predicate...
cstfp_pred_ty< is_pos_zero_fp > m_PosZeroFP()
Match a floating-point positive zero.
Definition: PatternMatch.h:524
iterator_range< User::op_iterator > arg_operands()
Definition: InstrTypes.h:1233
AttrBuilder & remove(const AttrBuilder &B)
Remove the attributes from the builder.
AttributeList getAttributes() const
Return the attribute list for this Function.
Definition: Function.h:223
cmpResult
IEEE-754R 5.11: Floating Point Comparison Relations.
Definition: APFloat.h:176
An instruction for storing to memory.
Definition: Instructions.h:325
bool extractProfTotalWeight(uint64_t &TotalVal) const
Retrieve total raw weight values of a branch.
Definition: Metadata.cpp:1336
Value * getRHS() const
static void ValueIsRAUWd(Value *Old, Value *New)
Definition: Value.cpp:923
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1877
constexpr char Attrs[]
Key for Kernel::Metadata::mAttrs.
uint64_t GetStringLength(const Value *V, unsigned CharSize=8)
If we can compute the length of the string pointed to by the specified pointer, return &#39;len+1&#39;...
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
static ConstantAsMetadata * get(Constant *C)
Definition: Metadata.h:409
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:1093
This class represents a truncation of integer types.
Type * getElementType() const
Return the element type of the array/vector.
Definition: Constants.cpp:2433
Value * getOperand(unsigned i) const
Definition: User.h:169
Class to represent pointers.
Definition: DerivedTypes.h:579
bool hasAttribute(Attribute::AttrKind Kind) const
Return true if the attribute exists in this set.
Definition: Attributes.cpp:654
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:359
const DataLayout & getDataLayout() const
static MetadataAsValue * get(LLVMContext &Context, Metadata *MD)
Definition: Metadata.cpp:105
static Constant * getBitCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1804
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:61
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1165
Instruction * visitFenceInst(FenceInst &FI)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:432
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:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static AttributeSet get(LLVMContext &C, const AttrBuilder &B)
Definition: Attributes.cpp:589
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:194
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:328
bool isNegative() const
Definition: APFloat.h:1162
static ConstantPointerNull * get(PointerType *T)
Static factory methods - Return objects of the specified value.
Definition: Constants.cpp:1432
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1248
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:465
unsigned arg_size() const
Definition: InstrTypes.h:1229
Value * getCalledValue() const
Definition: InstrTypes.h:1280
TypeSize getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
LLVM_NODISCARD AttributeList addParamAttribute(LLVMContext &C, unsigned ArgNo, Attribute::AttrKind Kind) const
Add an argument attribute to the list.
Definition: Attributes.h:413
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:880
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:64
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:323
Value * SimplifyFMAFMul(Value *LHS, Value *RHS, FastMathFlags FMF, const SimplifyQuery &Q)
Given operands for the multiplication of a FMA, fold the result or return null.
bool isNaN() const
Definition: APFloat.h:1160
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
void setAlignment(MaybeAlign Align)
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:2287
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:224
bool isSigned() const
Whether the intrinsic is signed or unsigned.
static BinaryOperator * CreateNUW(BinaryOps Opc, Value *V1, Value *V2, const Twine &Name="")
Definition: InstrTypes.h:308
TypeSize getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:486
FunctionType * getFunctionType() const
Definition: InstrTypes.h:1144
static ManagedStatic< OptionRegistry > OR
Definition: Options.cpp:30
unsigned getNumParams() const
Return the number of fixed parameters this function type requires.
Definition: DerivedTypes.h:144
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:263
specific_intval m_SpecificInt(APInt V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:664
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:336
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:587
Instruction * visitCallBrInst(CallBrInst &CBI)
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.
param_iterator param_begin() const
Definition: DerivedTypes.h:133
This file declares a class to represent arbitrary precision floating point values and provide a varie...
static VectorType * getInteger(VectorType *VTy)
This static method gets a VectorType with the same number of elements as the input type...
Definition: DerivedTypes.h:468
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:90
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:732
constexpr double e
Definition: MathExtras.h:57
void setCallingConv(CallingConv::ID CC)
Definition: InstrTypes.h:1348
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:77
This class represents any memset intrinsic.
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
static FunctionType * get(Type *Result, ArrayRef< Type *> Params, bool isVarArg)
This static method is the primary way of constructing a FunctionType.
Definition: