ConstantFolding.cpp revision 360784 
1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
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 defines routines for folding instructions into constants.
10//
11// Also, to supplement the basic IR ConstantExpr simplifications,
12// this file defines some additional folding routines that can make use of
13// DataLayout information. These functions cannot go in IR due to library
14// dependency issues.
15//
16//===----------------------------------------------------------------------===//
17
18#include "llvm/Analysis/ConstantFolding.h"
19#include "llvm/ADT/APFloat.h"
20#include "llvm/ADT/APInt.h"
21#include "llvm/ADT/ArrayRef.h"
22#include "llvm/ADT/DenseMap.h"
23#include "llvm/ADT/STLExtras.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/StringRef.h"
26#include "llvm/Analysis/TargetLibraryInfo.h"
27#include "llvm/Analysis/ValueTracking.h"
28#include "llvm/Analysis/VectorUtils.h"
29#include "llvm/Config/config.h"
30#include "llvm/IR/Constant.h"
31#include "llvm/IR/Constants.h"
32#include "llvm/IR/DataLayout.h"
33#include "llvm/IR/DerivedTypes.h"
34#include "llvm/IR/Function.h"
35#include "llvm/IR/GlobalValue.h"
36#include "llvm/IR/GlobalVariable.h"
37#include "llvm/IR/InstrTypes.h"
38#include "llvm/IR/Instruction.h"
39#include "llvm/IR/Instructions.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/IntrinsicsX86.h"
42#include "llvm/IR/Operator.h"
43#include "llvm/IR/Type.h"
44#include "llvm/IR/Value.h"
45#include "llvm/Support/Casting.h"
46#include "llvm/Support/ErrorHandling.h"
47#include "llvm/Support/KnownBits.h"
48#include "llvm/Support/MathExtras.h"
49#include <cassert>
50#include <cerrno>
51#include <cfenv>
52#include <cmath>
53#include <cstddef>
54#include <cstdint>
55
56using namespace llvm;
57
58namespace {
59
60//===----------------------------------------------------------------------===//
61// Constant Folding internal helper functions
62//===----------------------------------------------------------------------===//
63
64static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
65                                        Constant *C, Type *SrcEltTy,
66                                        unsigned NumSrcElts,
67                                        const DataLayout &DL) {
68  // Now that we know that the input value is a vector of integers, just shift
69  // and insert them into our result.
70  unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
71  for (unsigned i = 0; i != NumSrcElts; ++i) {
72    Constant *Element;
73    if (DL.isLittleEndian())
74      Element = C->getAggregateElement(NumSrcElts - i - 1);
75    else
76      Element = C->getAggregateElement(i);
77
78    if (Element && isa<UndefValue>(Element)) {
79      Result <<= BitShift;
80      continue;
81    }
82
83    auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
84    if (!ElementCI)
85      return ConstantExpr::getBitCast(C, DestTy);
86
87    Result <<= BitShift;
88    Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth());
89  }
90
91  return nullptr;
92}
93
94/// Constant fold bitcast, symbolically evaluating it with DataLayout.
95/// This always returns a non-null constant, but it may be a
96/// ConstantExpr if unfoldable.
97Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
98  assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&
99         "Invalid constantexpr bitcast!");
100
101  // Catch the obvious splat cases.
102  if (C->isNullValue() && !DestTy->isX86_MMXTy())
103    return Constant::getNullValue(DestTy);
104  if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
105      !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
106    return Constant::getAllOnesValue(DestTy);
107
108  if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
109    // Handle a vector->scalar integer/fp cast.
110    if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
111      unsigned NumSrcElts = VTy->getNumElements();
112      Type *SrcEltTy = VTy->getElementType();
113
114      // If the vector is a vector of floating point, convert it to vector of int
115      // to simplify things.
116      if (SrcEltTy->isFloatingPointTy()) {
117        unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
118        Type *SrcIVTy =
119          VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
120        // Ask IR to do the conversion now that #elts line up.
121        C = ConstantExpr::getBitCast(C, SrcIVTy);
122      }
123
124      APInt Result(DL.getTypeSizeInBits(DestTy), 0);
125      if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
126                                                SrcEltTy, NumSrcElts, DL))
127        return CE;
128
129      if (isa<IntegerType>(DestTy))
130        return ConstantInt::get(DestTy, Result);
131
132      APFloat FP(DestTy->getFltSemantics(), Result);
133      return ConstantFP::get(DestTy->getContext(), FP);
134    }
135  }
136
137  // The code below only handles casts to vectors currently.
138  auto *DestVTy = dyn_cast<VectorType>(DestTy);
139  if (!DestVTy)
140    return ConstantExpr::getBitCast(C, DestTy);
141
142  // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
143  // vector so the code below can handle it uniformly.
144  if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
145    Constant *Ops = C; // don't take the address of C!
146    return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
147  }
148
149  // If this is a bitcast from constant vector -> vector, fold it.
150  if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
151    return ConstantExpr::getBitCast(C, DestTy);
152
153  // If the element types match, IR can fold it.
154  unsigned NumDstElt = DestVTy->getNumElements();
155  unsigned NumSrcElt = C->getType()->getVectorNumElements();
156  if (NumDstElt == NumSrcElt)
157    return ConstantExpr::getBitCast(C, DestTy);
158
159  Type *SrcEltTy = C->getType()->getVectorElementType();
160  Type *DstEltTy = DestVTy->getElementType();
161
162  // Otherwise, we're changing the number of elements in a vector, which
163  // requires endianness information to do the right thing.  For example,
164  //    bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
165  // folds to (little endian):
166  //    <4 x i32> <i32 0, i32 0, i32 1, i32 0>
167  // and to (big endian):
168  //    <4 x i32> <i32 0, i32 0, i32 0, i32 1>
169
170  // First thing is first.  We only want to think about integer here, so if
171  // we have something in FP form, recast it as integer.
172  if (DstEltTy->isFloatingPointTy()) {
173    // Fold to an vector of integers with same size as our FP type.
174    unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
175    Type *DestIVTy =
176      VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
177    // Recursively handle this integer conversion, if possible.
178    C = FoldBitCast(C, DestIVTy, DL);
179
180    // Finally, IR can handle this now that #elts line up.
181    return ConstantExpr::getBitCast(C, DestTy);
182  }
183
184  // Okay, we know the destination is integer, if the input is FP, convert
185  // it to integer first.
186  if (SrcEltTy->isFloatingPointTy()) {
187    unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
188    Type *SrcIVTy =
189      VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
190    // Ask IR to do the conversion now that #elts line up.
191    C = ConstantExpr::getBitCast(C, SrcIVTy);
192    // If IR wasn't able to fold it, bail out.
193    if (!isa<ConstantVector>(C) &&  // FIXME: Remove ConstantVector.
194        !isa<ConstantDataVector>(C))
195      return C;
196  }
197
198  // Now we know that the input and output vectors are both integer vectors
199  // of the same size, and that their #elements is not the same.  Do the
200  // conversion here, which depends on whether the input or output has
201  // more elements.
202  bool isLittleEndian = DL.isLittleEndian();
203
204  SmallVector<Constant*, 32> Result;
205  if (NumDstElt < NumSrcElt) {
206    // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
207    Constant *Zero = Constant::getNullValue(DstEltTy);
208    unsigned Ratio = NumSrcElt/NumDstElt;
209    unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
210    unsigned SrcElt = 0;
211    for (unsigned i = 0; i != NumDstElt; ++i) {
212      // Build each element of the result.
213      Constant *Elt = Zero;
214      unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
215      for (unsigned j = 0; j != Ratio; ++j) {
216        Constant *Src = C->getAggregateElement(SrcElt++);
217        if (Src && isa<UndefValue>(Src))
218          Src = Constant::getNullValue(C->getType()->getVectorElementType());
219        else
220          Src = dyn_cast_or_null<ConstantInt>(Src);
221        if (!Src)  // Reject constantexpr elements.
222          return ConstantExpr::getBitCast(C, DestTy);
223
224        // Zero extend the element to the right size.
225        Src = ConstantExpr::getZExt(Src, Elt->getType());
226
227        // Shift it to the right place, depending on endianness.
228        Src = ConstantExpr::getShl(Src,
229                                   ConstantInt::get(Src->getType(), ShiftAmt));
230        ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
231
232        // Mix it in.
233        Elt = ConstantExpr::getOr(Elt, Src);
234      }
235      Result.push_back(Elt);
236    }
237    return ConstantVector::get(Result);
238  }
239
240  // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
241  unsigned Ratio = NumDstElt/NumSrcElt;
242  unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
243
244  // Loop over each source value, expanding into multiple results.
245  for (unsigned i = 0; i != NumSrcElt; ++i) {
246    auto *Element = C->getAggregateElement(i);
247
248    if (!Element) // Reject constantexpr elements.
249      return ConstantExpr::getBitCast(C, DestTy);
250
251    if (isa<UndefValue>(Element)) {
252      // Correctly Propagate undef values.
253      Result.append(Ratio, UndefValue::get(DstEltTy));
254      continue;
255    }
256
257    auto *Src = dyn_cast<ConstantInt>(Element);
258    if (!Src)
259      return ConstantExpr::getBitCast(C, DestTy);
260
261    unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
262    for (unsigned j = 0; j != Ratio; ++j) {
263      // Shift the piece of the value into the right place, depending on
264      // endianness.
265      Constant *Elt = ConstantExpr::getLShr(Src,
266                                  ConstantInt::get(Src->getType(), ShiftAmt));
267      ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
268
269      // Truncate the element to an integer with the same pointer size and
270      // convert the element back to a pointer using a inttoptr.
271      if (DstEltTy->isPointerTy()) {
272        IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
273        Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
274        Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
275        continue;
276      }
277
278      // Truncate and remember this piece.
279      Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
280    }
281  }
282
283  return ConstantVector::get(Result);
284}
285
286} // end anonymous namespace
287
288/// If this constant is a constant offset from a global, return the global and
289/// the constant. Because of constantexprs, this function is recursive.
290bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
291                                      APInt &Offset, const DataLayout &DL) {
292  // Trivial case, constant is the global.
293  if ((GV = dyn_cast<GlobalValue>(C))) {
294    unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());
295    Offset = APInt(BitWidth, 0);
296    return true;
297  }
298
299  // Otherwise, if this isn't a constant expr, bail out.
300  auto *CE = dyn_cast<ConstantExpr>(C);
301  if (!CE) return false;
302
303  // Look through ptr->int and ptr->ptr casts.
304  if (CE->getOpcode() == Instruction::PtrToInt ||
305      CE->getOpcode() == Instruction::BitCast)
306    return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
307
308  // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
309  auto *GEP = dyn_cast<GEPOperator>(CE);
310  if (!GEP)
311    return false;
312
313  unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
314  APInt TmpOffset(BitWidth, 0);
315
316  // If the base isn't a global+constant, we aren't either.
317  if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
318    return false;
319
320  // Otherwise, add any offset that our operands provide.
321  if (!GEP->accumulateConstantOffset(DL, TmpOffset))
322    return false;
323
324  Offset = TmpOffset;
325  return true;
326}
327
328Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
329                                         const DataLayout &DL) {
330  do {
331    Type *SrcTy = C->getType();
332
333    // If the type sizes are the same and a cast is legal, just directly
334    // cast the constant.
335    if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
336      Instruction::CastOps Cast = Instruction::BitCast;
337      // If we are going from a pointer to int or vice versa, we spell the cast
338      // differently.
339      if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
340        Cast = Instruction::IntToPtr;
341      else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
342        Cast = Instruction::PtrToInt;
343
344      if (CastInst::castIsValid(Cast, C, DestTy))
345        return ConstantExpr::getCast(Cast, C, DestTy);
346    }
347
348    // If this isn't an aggregate type, there is nothing we can do to drill down
349    // and find a bitcastable constant.
350    if (!SrcTy->isAggregateType())
351      return nullptr;
352
353    // We're simulating a load through a pointer that was bitcast to point to
354    // a different type, so we can try to walk down through the initial
355    // elements of an aggregate to see if some part of the aggregate is
356    // castable to implement the "load" semantic model.
357    if (SrcTy->isStructTy()) {
358      // Struct types might have leading zero-length elements like [0 x i32],
359      // which are certainly not what we are looking for, so skip them.
360      unsigned Elem = 0;
361      Constant *ElemC;
362      do {
363        ElemC = C->getAggregateElement(Elem++);
364      } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()) == 0);
365      C = ElemC;
366    } else {
367      C = C->getAggregateElement(0u);
368    }
369  } while (C);
370
371  return nullptr;
372}
373
374namespace {
375
376/// Recursive helper to read bits out of global. C is the constant being copied
377/// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
378/// results into and BytesLeft is the number of bytes left in
379/// the CurPtr buffer. DL is the DataLayout.
380bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
381                        unsigned BytesLeft, const DataLayout &DL) {
382  assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
383         "Out of range access");
384
385  // If this element is zero or undefined, we can just return since *CurPtr is
386  // zero initialized.
387  if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
388    return true;
389
390  if (auto *CI = dyn_cast<ConstantInt>(C)) {
391    if (CI->getBitWidth() > 64 ||
392        (CI->getBitWidth() & 7) != 0)
393      return false;
394
395    uint64_t Val = CI->getZExtValue();
396    unsigned IntBytes = unsigned(CI->getBitWidth()/8);
397
398    for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
399      int n = ByteOffset;
400      if (!DL.isLittleEndian())
401        n = IntBytes - n - 1;
402      CurPtr[i] = (unsigned char)(Val >> (n * 8));
403      ++ByteOffset;
404    }
405    return true;
406  }
407
408  if (auto *CFP = dyn_cast<ConstantFP>(C)) {
409    if (CFP->getType()->isDoubleTy()) {
410      C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
411      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
412    }
413    if (CFP->getType()->isFloatTy()){
414      C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
415      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
416    }
417    if (CFP->getType()->isHalfTy()){
418      C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
419      return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
420    }
421    return false;
422  }
423
424  if (auto *CS = dyn_cast<ConstantStruct>(C)) {
425    const StructLayout *SL = DL.getStructLayout(CS->getType());
426    unsigned Index = SL->getElementContainingOffset(ByteOffset);
427    uint64_t CurEltOffset = SL->getElementOffset(Index);
428    ByteOffset -= CurEltOffset;
429
430    while (true) {
431      // If the element access is to the element itself and not to tail padding,
432      // read the bytes from the element.
433      uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
434
435      if (ByteOffset < EltSize &&
436          !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
437                              BytesLeft, DL))
438        return false;
439
440      ++Index;
441
442      // Check to see if we read from the last struct element, if so we're done.
443      if (Index == CS->getType()->getNumElements())
444        return true;
445
446      // If we read all of the bytes we needed from this element we're done.
447      uint64_t NextEltOffset = SL->getElementOffset(Index);
448
449      if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
450        return true;
451
452      // Move to the next element of the struct.
453      CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
454      BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
455      ByteOffset = 0;
456      CurEltOffset = NextEltOffset;
457    }
458    // not reached.
459  }
460
461  if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
462      isa<ConstantDataSequential>(C)) {
463    Type *EltTy = C->getType()->getSequentialElementType();
464    uint64_t EltSize = DL.getTypeAllocSize(EltTy);
465    uint64_t Index = ByteOffset / EltSize;
466    uint64_t Offset = ByteOffset - Index * EltSize;
467    uint64_t NumElts;
468    if (auto *AT = dyn_cast<ArrayType>(C->getType()))
469      NumElts = AT->getNumElements();
470    else
471      NumElts = C->getType()->getVectorNumElements();
472
473    for (; Index != NumElts; ++Index) {
474      if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
475                              BytesLeft, DL))
476        return false;
477
478      uint64_t BytesWritten = EltSize - Offset;
479      assert(BytesWritten <= EltSize && "Not indexing into this element?");
480      if (BytesWritten >= BytesLeft)
481        return true;
482
483      Offset = 0;
484      BytesLeft -= BytesWritten;
485      CurPtr += BytesWritten;
486    }
487    return true;
488  }
489
490  if (auto *CE = dyn_cast<ConstantExpr>(C)) {
491    if (CE->getOpcode() == Instruction::IntToPtr &&
492        CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
493      return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
494                                BytesLeft, DL);
495    }
496  }
497
498  // Otherwise, unknown initializer type.
499  return false;
500}
501
502Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
503                                          const DataLayout &DL) {
504  auto *PTy = cast<PointerType>(C->getType());
505  auto *IntType = dyn_cast<IntegerType>(LoadTy);
506
507  // If this isn't an integer load we can't fold it directly.
508  if (!IntType) {
509    unsigned AS = PTy->getAddressSpace();
510
511    // If this is a float/double load, we can try folding it as an int32/64 load
512    // and then bitcast the result.  This can be useful for union cases.  Note
513    // that address spaces don't matter here since we're not going to result in
514    // an actual new load.
515    Type *MapTy;
516    if (LoadTy->isHalfTy())
517      MapTy = Type::getInt16Ty(C->getContext());
518    else if (LoadTy->isFloatTy())
519      MapTy = Type::getInt32Ty(C->getContext());
520    else if (LoadTy->isDoubleTy())
521      MapTy = Type::getInt64Ty(C->getContext());
522    else if (LoadTy->isVectorTy()) {
523      MapTy = PointerType::getIntNTy(C->getContext(),
524                                     DL.getTypeSizeInBits(LoadTy));
525    } else
526      return nullptr;
527
528    C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
529    if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL)) {
530      if (Res->isNullValue() && !LoadTy->isX86_MMXTy())
531        // Materializing a zero can be done trivially without a bitcast
532        return Constant::getNullValue(LoadTy);
533      Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;
534      Res = FoldBitCast(Res, CastTy, DL);
535      if (LoadTy->isPtrOrPtrVectorTy()) {
536        // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr
537        if (Res->isNullValue() && !LoadTy->isX86_MMXTy())
538          return Constant::getNullValue(LoadTy);
539        if (DL.isNonIntegralPointerType(LoadTy->getScalarType()))
540          // Be careful not to replace a load of an addrspace value with an inttoptr here
541          return nullptr;
542        Res = ConstantExpr::getCast(Instruction::IntToPtr, Res, LoadTy);
543      }
544      return Res;
545    }
546    return nullptr;
547  }
548
549  unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
550  if (BytesLoaded > 32 || BytesLoaded == 0)
551    return nullptr;
552
553  GlobalValue *GVal;
554  APInt OffsetAI;
555  if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL))
556    return nullptr;
557
558  auto *GV = dyn_cast<GlobalVariable>(GVal);
559  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
560      !GV->getInitializer()->getType()->isSized())
561    return nullptr;
562
563  int64_t Offset = OffsetAI.getSExtValue();
564  int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
565
566  // If we're not accessing anything in this constant, the result is undefined.
567  if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))
568    return UndefValue::get(IntType);
569
570  // If we're not accessing anything in this constant, the result is undefined.
571  if (Offset >= InitializerSize)
572    return UndefValue::get(IntType);
573
574  unsigned char RawBytes[32] = {0};
575  unsigned char *CurPtr = RawBytes;
576  unsigned BytesLeft = BytesLoaded;
577
578  // If we're loading off the beginning of the global, some bytes may be valid.
579  if (Offset < 0) {
580    CurPtr += -Offset;
581    BytesLeft += Offset;
582    Offset = 0;
583  }
584
585  if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL))
586    return nullptr;
587
588  APInt ResultVal = APInt(IntType->getBitWidth(), 0);
589  if (DL.isLittleEndian()) {
590    ResultVal = RawBytes[BytesLoaded - 1];
591    for (unsigned i = 1; i != BytesLoaded; ++i) {
592      ResultVal <<= 8;
593      ResultVal |= RawBytes[BytesLoaded - 1 - i];
594    }
595  } else {
596    ResultVal = RawBytes[0];
597    for (unsigned i = 1; i != BytesLoaded; ++i) {
598      ResultVal <<= 8;
599      ResultVal |= RawBytes[i];
600    }
601  }
602
603  return ConstantInt::get(IntType->getContext(), ResultVal);
604}
605
606Constant *ConstantFoldLoadThroughBitcastExpr(ConstantExpr *CE, Type *DestTy,
607                                             const DataLayout &DL) {
608  auto *SrcPtr = CE->getOperand(0);
609  auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType());
610  if (!SrcPtrTy)
611    return nullptr;
612  Type *SrcTy = SrcPtrTy->getPointerElementType();
613
614  Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL);
615  if (!C)
616    return nullptr;
617
618  return llvm::ConstantFoldLoadThroughBitcast(C, DestTy, DL);
619}
620
621} // end anonymous namespace
622
623Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
624                                             const DataLayout &DL) {
625  // First, try the easy cases:
626  if (auto *GV = dyn_cast<GlobalVariable>(C))
627    if (GV->isConstant() && GV->hasDefinitiveInitializer())
628      return GV->getInitializer();
629
630  if (auto *GA = dyn_cast<GlobalAlias>(C))
631    if (GA->getAliasee() && !GA->isInterposable())
632      return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL);
633
634  // If the loaded value isn't a constant expr, we can't handle it.
635  auto *CE = dyn_cast<ConstantExpr>(C);
636  if (!CE)
637    return nullptr;
638
639  if (CE->getOpcode() == Instruction::GetElementPtr) {
640    if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
641      if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
642        if (Constant *V =
643             ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
644          return V;
645      }
646    }
647  }
648
649  if (CE->getOpcode() == Instruction::BitCast)
650    if (Constant *LoadedC = ConstantFoldLoadThroughBitcastExpr(CE, Ty, DL))
651      return LoadedC;
652
653  // Instead of loading constant c string, use corresponding integer value
654  // directly if string length is small enough.
655  StringRef Str;
656  if (getConstantStringInfo(CE, Str) && !Str.empty()) {
657    size_t StrLen = Str.size();
658    unsigned NumBits = Ty->getPrimitiveSizeInBits();
659    // Replace load with immediate integer if the result is an integer or fp
660    // value.
661    if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
662        (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
663      APInt StrVal(NumBits, 0);
664      APInt SingleChar(NumBits, 0);
665      if (DL.isLittleEndian()) {
666        for (unsigned char C : reverse(Str.bytes())) {
667          SingleChar = static_cast<uint64_t>(C);
668          StrVal = (StrVal << 8) | SingleChar;
669        }
670      } else {
671        for (unsigned char C : Str.bytes()) {
672          SingleChar = static_cast<uint64_t>(C);
673          StrVal = (StrVal << 8) | SingleChar;
674        }
675        // Append NULL at the end.
676        SingleChar = 0;
677        StrVal = (StrVal << 8) | SingleChar;
678      }
679
680      Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
681      if (Ty->isFloatingPointTy())
682        Res = ConstantExpr::getBitCast(Res, Ty);
683      return Res;
684    }
685  }
686
687  // If this load comes from anywhere in a constant global, and if the global
688  // is all undef or zero, we know what it loads.
689  if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
690    if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
691      if (GV->getInitializer()->isNullValue())
692        return Constant::getNullValue(Ty);
693      if (isa<UndefValue>(GV->getInitializer()))
694        return UndefValue::get(Ty);
695    }
696  }
697
698  // Try hard to fold loads from bitcasted strange and non-type-safe things.
699  return FoldReinterpretLoadFromConstPtr(CE, Ty, DL);
700}
701
702namespace {
703
704Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) {
705  if (LI->isVolatile()) return nullptr;
706
707  if (auto *C = dyn_cast<Constant>(LI->getOperand(0)))
708    return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL);
709
710  return nullptr;
711}
712
713/// One of Op0/Op1 is a constant expression.
714/// Attempt to symbolically evaluate the result of a binary operator merging
715/// these together.  If target data info is available, it is provided as DL,
716/// otherwise DL is null.
717Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
718                                    const DataLayout &DL) {
719  // SROA
720
721  // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
722  // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
723  // bits.
724
725  if (Opc == Instruction::And) {
726    KnownBits Known0 = computeKnownBits(Op0, DL);
727    KnownBits Known1 = computeKnownBits(Op1, DL);
728    if ((Known1.One | Known0.Zero).isAllOnesValue()) {
729      // All the bits of Op0 that the 'and' could be masking are already zero.
730      return Op0;
731    }
732    if ((Known0.One | Known1.Zero).isAllOnesValue()) {
733      // All the bits of Op1 that the 'and' could be masking are already zero.
734      return Op1;
735    }
736
737    Known0.Zero |= Known1.Zero;
738    Known0.One &= Known1.One;
739    if (Known0.isConstant())
740      return ConstantInt::get(Op0->getType(), Known0.getConstant());
741  }
742
743  // If the constant expr is something like &A[123] - &A[4].f, fold this into a
744  // constant.  This happens frequently when iterating over a global array.
745  if (Opc == Instruction::Sub) {
746    GlobalValue *GV1, *GV2;
747    APInt Offs1, Offs2;
748
749    if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
750      if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
751        unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
752
753        // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
754        // PtrToInt may change the bitwidth so we have convert to the right size
755        // first.
756        return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
757                                                Offs2.zextOrTrunc(OpSize));
758      }
759  }
760
761  return nullptr;
762}
763
764/// If array indices are not pointer-sized integers, explicitly cast them so
765/// that they aren't implicitly casted by the getelementptr.
766Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
767                         Type *ResultTy, Optional<unsigned> InRangeIndex,
768                         const DataLayout &DL, const TargetLibraryInfo *TLI) {
769  Type *IntIdxTy = DL.getIndexType(ResultTy);
770  Type *IntIdxScalarTy = IntIdxTy->getScalarType();
771
772  bool Any = false;
773  SmallVector<Constant*, 32> NewIdxs;
774  for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
775    if ((i == 1 ||
776         !isa<StructType>(GetElementPtrInst::getIndexedType(
777             SrcElemTy, Ops.slice(1, i - 1)))) &&
778        Ops[i]->getType()->getScalarType() != IntIdxScalarTy) {
779      Any = true;
780      Type *NewType = Ops[i]->getType()->isVectorTy()
781                          ? IntIdxTy
782                          : IntIdxScalarTy;
783      NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
784                                                                      true,
785                                                                      NewType,
786                                                                      true),
787                                              Ops[i], NewType));
788    } else
789      NewIdxs.push_back(Ops[i]);
790  }
791
792  if (!Any)
793    return nullptr;
794
795  Constant *C = ConstantExpr::getGetElementPtr(
796      SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex);
797  if (Constant *Folded = ConstantFoldConstant(C, DL, TLI))
798    C = Folded;
799
800  return C;
801}
802
803/// Strip the pointer casts, but preserve the address space information.
804Constant *StripPtrCastKeepAS(Constant *Ptr, Type *&ElemTy) {
805  assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
806  auto *OldPtrTy = cast<PointerType>(Ptr->getType());
807  Ptr = cast<Constant>(Ptr->stripPointerCasts());
808  auto *NewPtrTy = cast<PointerType>(Ptr->getType());
809
810  ElemTy = NewPtrTy->getPointerElementType();
811
812  // Preserve the address space number of the pointer.
813  if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
814    NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace());
815    Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
816  }
817  return Ptr;
818}
819
820/// If we can symbolically evaluate the GEP constant expression, do so.
821Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
822                                  ArrayRef<Constant *> Ops,
823                                  const DataLayout &DL,
824                                  const TargetLibraryInfo *TLI) {
825  const GEPOperator *InnermostGEP = GEP;
826  bool InBounds = GEP->isInBounds();
827
828  Type *SrcElemTy = GEP->getSourceElementType();
829  Type *ResElemTy = GEP->getResultElementType();
830  Type *ResTy = GEP->getType();
831  if (!SrcElemTy->isSized())
832    return nullptr;
833
834  if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy,
835                                   GEP->getInRangeIndex(), DL, TLI))
836    return C;
837
838  Constant *Ptr = Ops[0];
839  if (!Ptr->getType()->isPointerTy())
840    return nullptr;
841
842  Type *IntIdxTy = DL.getIndexType(Ptr->getType());
843
844  // If this is a constant expr gep that is effectively computing an
845  // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
846  for (unsigned i = 1, e = Ops.size(); i != e; ++i)
847      if (!isa<ConstantInt>(Ops[i])) {
848
849        // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
850        // "inttoptr (sub (ptrtoint Ptr), V)"
851        if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
852          auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
853          assert((!CE || CE->getType() == IntIdxTy) &&
854                 "CastGEPIndices didn't canonicalize index types!");
855          if (CE && CE->getOpcode() == Instruction::Sub &&
856              CE->getOperand(0)->isNullValue()) {
857            Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
858            Res = ConstantExpr::getSub(Res, CE->getOperand(1));
859            Res = ConstantExpr::getIntToPtr(Res, ResTy);
860            if (auto *FoldedRes = ConstantFoldConstant(Res, DL, TLI))
861              Res = FoldedRes;
862            return Res;
863          }
864        }
865        return nullptr;
866      }
867
868  unsigned BitWidth = DL.getTypeSizeInBits(IntIdxTy);
869  APInt Offset =
870      APInt(BitWidth,
871            DL.getIndexedOffsetInType(
872                SrcElemTy,
873                makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
874  Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
875
876  // If this is a GEP of a GEP, fold it all into a single GEP.
877  while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
878    InnermostGEP = GEP;
879    InBounds &= GEP->isInBounds();
880
881    SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
882
883    // Do not try the incorporate the sub-GEP if some index is not a number.
884    bool AllConstantInt = true;
885    for (Value *NestedOp : NestedOps)
886      if (!isa<ConstantInt>(NestedOp)) {
887        AllConstantInt = false;
888        break;
889      }
890    if (!AllConstantInt)
891      break;
892
893    Ptr = cast<Constant>(GEP->getOperand(0));
894    SrcElemTy = GEP->getSourceElementType();
895    Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
896    Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
897  }
898
899  // If the base value for this address is a literal integer value, fold the
900  // getelementptr to the resulting integer value casted to the pointer type.
901  APInt BasePtr(BitWidth, 0);
902  if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
903    if (CE->getOpcode() == Instruction::IntToPtr) {
904      if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
905        BasePtr = Base->getValue().zextOrTrunc(BitWidth);
906    }
907  }
908
909  auto *PTy = cast<PointerType>(Ptr->getType());
910  if ((Ptr->isNullValue() || BasePtr != 0) &&
911      !DL.isNonIntegralPointerType(PTy)) {
912    Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
913    return ConstantExpr::getIntToPtr(C, ResTy);
914  }
915
916  // Otherwise form a regular getelementptr. Recompute the indices so that
917  // we eliminate over-indexing of the notional static type array bounds.
918  // This makes it easy to determine if the getelementptr is "inbounds".
919  // Also, this helps GlobalOpt do SROA on GlobalVariables.
920  Type *Ty = PTy;
921  SmallVector<Constant *, 32> NewIdxs;
922
923  do {
924    if (!Ty->isStructTy()) {
925      if (Ty->isPointerTy()) {
926        // The only pointer indexing we'll do is on the first index of the GEP.
927        if (!NewIdxs.empty())
928          break;
929
930        Ty = SrcElemTy;
931
932        // Only handle pointers to sized types, not pointers to functions.
933        if (!Ty->isSized())
934          return nullptr;
935      } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) {
936        Ty = ATy->getElementType();
937      } else {
938        // We've reached some non-indexable type.
939        break;
940      }
941
942      // Determine which element of the array the offset points into.
943      APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty));
944      if (ElemSize == 0) {
945        // The element size is 0. This may be [0 x Ty]*, so just use a zero
946        // index for this level and proceed to the next level to see if it can
947        // accommodate the offset.
948        NewIdxs.push_back(ConstantInt::get(IntIdxTy, 0));
949      } else {
950        // The element size is non-zero divide the offset by the element
951        // size (rounding down), to compute the index at this level.
952        bool Overflow;
953        APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow);
954        if (Overflow)
955          break;
956        Offset -= NewIdx * ElemSize;
957        NewIdxs.push_back(ConstantInt::get(IntIdxTy, NewIdx));
958      }
959    } else {
960      auto *STy = cast<StructType>(Ty);
961      // If we end up with an offset that isn't valid for this struct type, we
962      // can't re-form this GEP in a regular form, so bail out. The pointer
963      // operand likely went through casts that are necessary to make the GEP
964      // sensible.
965      const StructLayout &SL = *DL.getStructLayout(STy);
966      if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes()))
967        break;
968
969      // Determine which field of the struct the offset points into. The
970      // getZExtValue is fine as we've already ensured that the offset is
971      // within the range representable by the StructLayout API.
972      unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
973      NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
974                                         ElIdx));
975      Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
976      Ty = STy->getTypeAtIndex(ElIdx);
977    }
978  } while (Ty != ResElemTy);
979
980  // If we haven't used up the entire offset by descending the static
981  // type, then the offset is pointing into the middle of an indivisible
982  // member, so we can't simplify it.
983  if (Offset != 0)
984    return nullptr;
985
986  // Preserve the inrange index from the innermost GEP if possible. We must
987  // have calculated the same indices up to and including the inrange index.
988  Optional<unsigned> InRangeIndex;
989  if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex())
990    if (SrcElemTy == InnermostGEP->getSourceElementType() &&
991        NewIdxs.size() > *LastIRIndex) {
992      InRangeIndex = LastIRIndex;
993      for (unsigned I = 0; I <= *LastIRIndex; ++I)
994        if (NewIdxs[I] != InnermostGEP->getOperand(I + 1))
995          return nullptr;
996    }
997
998  // Create a GEP.
999  Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs,
1000                                               InBounds, InRangeIndex);
1001  assert(C->getType()->getPointerElementType() == Ty &&
1002         "Computed GetElementPtr has unexpected type!");
1003
1004  // If we ended up indexing a member with a type that doesn't match
1005  // the type of what the original indices indexed, add a cast.
1006  if (Ty != ResElemTy)
1007    C = FoldBitCast(C, ResTy, DL);
1008
1009  return C;
1010}
1011
1012/// Attempt to constant fold an instruction with the
1013/// specified opcode and operands.  If successful, the constant result is
1014/// returned, if not, null is returned.  Note that this function can fail when
1015/// attempting to fold instructions like loads and stores, which have no
1016/// constant expression form.
1017Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
1018                                       ArrayRef<Constant *> Ops,
1019                                       const DataLayout &DL,
1020                                       const TargetLibraryInfo *TLI) {
1021  Type *DestTy = InstOrCE->getType();
1022
1023  if (Instruction::isUnaryOp(Opcode))
1024    return ConstantFoldUnaryOpOperand(Opcode, Ops[0], DL);
1025
1026  if (Instruction::isBinaryOp(Opcode))
1027    return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
1028
1029  if (Instruction::isCast(Opcode))
1030    return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
1031
1032  if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
1033    if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1034      return C;
1035
1036    return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0],
1037                                          Ops.slice(1), GEP->isInBounds(),
1038                                          GEP->getInRangeIndex());
1039  }
1040
1041  if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))
1042    return CE->getWithOperands(Ops);
1043
1044  switch (Opcode) {
1045  default: return nullptr;
1046  case Instruction::ICmp:
1047  case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1048  case Instruction::Call:
1049    if (auto *F = dyn_cast<Function>(Ops.back())) {
1050      const auto *Call = cast<CallBase>(InstOrCE);
1051      if (canConstantFoldCallTo(Call, F))
1052        return ConstantFoldCall(Call, F, Ops.slice(0, Ops.size() - 1), TLI);
1053    }
1054    return nullptr;
1055  case Instruction::Select:
1056    return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1057  case Instruction::ExtractElement:
1058    return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1059  case Instruction::ExtractValue:
1060    return ConstantExpr::getExtractValue(
1061        Ops[0], cast<ExtractValueInst>(InstOrCE)->getIndices());
1062  case Instruction::InsertElement:
1063    return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1064  case Instruction::ShuffleVector:
1065    return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1066  }
1067}
1068
1069} // end anonymous namespace
1070
1071//===----------------------------------------------------------------------===//
1072// Constant Folding public APIs
1073//===----------------------------------------------------------------------===//
1074
1075namespace {
1076
1077Constant *
1078ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1079                         const TargetLibraryInfo *TLI,
1080                         SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1081  if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))
1082    return nullptr;
1083
1084  SmallVector<Constant *, 8> Ops;
1085  for (const Use &NewU : C->operands()) {
1086    auto *NewC = cast<Constant>(&NewU);
1087    // Recursively fold the ConstantExpr's operands. If we have already folded
1088    // a ConstantExpr, we don't have to process it again.
1089    if (isa<ConstantVector>(NewC) || isa<ConstantExpr>(NewC)) {
1090      auto It = FoldedOps.find(NewC);
1091      if (It == FoldedOps.end()) {
1092        if (auto *FoldedC =
1093                ConstantFoldConstantImpl(NewC, DL, TLI, FoldedOps)) {
1094          FoldedOps.insert({NewC, FoldedC});
1095          NewC = FoldedC;
1096        } else {
1097          FoldedOps.insert({NewC, NewC});
1098        }
1099      } else {
1100        NewC = It->second;
1101      }
1102    }
1103    Ops.push_back(NewC);
1104  }
1105
1106  if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1107    if (CE->isCompare())
1108      return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1109                                             DL, TLI);
1110
1111    return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI);
1112  }
1113
1114  assert(isa<ConstantVector>(C));
1115  return ConstantVector::get(Ops);
1116}
1117
1118} // end anonymous namespace
1119
1120Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
1121                                        const TargetLibraryInfo *TLI) {
1122  // Handle PHI nodes quickly here...
1123  if (auto *PN = dyn_cast<PHINode>(I)) {
1124    Constant *CommonValue = nullptr;
1125
1126    SmallDenseMap<Constant *, Constant *> FoldedOps;
1127    for (Value *Incoming : PN->incoming_values()) {
1128      // If the incoming value is undef then skip it.  Note that while we could
1129      // skip the value if it is equal to the phi node itself we choose not to
1130      // because that would break the rule that constant folding only applies if
1131      // all operands are constants.
1132      if (isa<UndefValue>(Incoming))
1133        continue;
1134      // If the incoming value is not a constant, then give up.
1135      auto *C = dyn_cast<Constant>(Incoming);
1136      if (!C)
1137        return nullptr;
1138      // Fold the PHI's operands.
1139      if (auto *FoldedC = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps))
1140        C = FoldedC;
1141      // If the incoming value is a different constant to
1142      // the one we saw previously, then give up.
1143      if (CommonValue && C != CommonValue)
1144        return nullptr;
1145      CommonValue = C;
1146    }
1147
1148    // If we reach here, all incoming values are the same constant or undef.
1149    return CommonValue ? CommonValue : UndefValue::get(PN->getType());
1150  }
1151
1152  // Scan the operand list, checking to see if they are all constants, if so,
1153  // hand off to ConstantFoldInstOperandsImpl.
1154  if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1155    return nullptr;
1156
1157  SmallDenseMap<Constant *, Constant *> FoldedOps;
1158  SmallVector<Constant *, 8> Ops;
1159  for (const Use &OpU : I->operands()) {
1160    auto *Op = cast<Constant>(&OpU);
1161    // Fold the Instruction's operands.
1162    if (auto *FoldedOp = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps))
1163      Op = FoldedOp;
1164
1165    Ops.push_back(Op);
1166  }
1167
1168  if (const auto *CI = dyn_cast<CmpInst>(I))
1169    return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1170                                           DL, TLI);
1171
1172  if (const auto *LI = dyn_cast<LoadInst>(I))
1173    return ConstantFoldLoadInst(LI, DL);
1174
1175  if (auto *IVI = dyn_cast<InsertValueInst>(I)) {
1176    return ConstantExpr::getInsertValue(
1177                                cast<Constant>(IVI->getAggregateOperand()),
1178                                cast<Constant>(IVI->getInsertedValueOperand()),
1179                                IVI->getIndices());
1180  }
1181
1182  if (auto *EVI = dyn_cast<ExtractValueInst>(I)) {
1183    return ConstantExpr::getExtractValue(
1184                                    cast<Constant>(EVI->getAggregateOperand()),
1185                                    EVI->getIndices());
1186  }
1187
1188  return ConstantFoldInstOperands(I, Ops, DL, TLI);
1189}
1190
1191Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1192                                     const TargetLibraryInfo *TLI) {
1193  SmallDenseMap<Constant *, Constant *> FoldedOps;
1194  return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1195}
1196
1197Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1198                                         ArrayRef<Constant *> Ops,
1199                                         const DataLayout &DL,
1200                                         const TargetLibraryInfo *TLI) {
1201  return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI);
1202}
1203
1204Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1205                                                Constant *Ops0, Constant *Ops1,
1206                                                const DataLayout &DL,
1207                                                const TargetLibraryInfo *TLI) {
1208  // fold: icmp (inttoptr x), null         -> icmp x, 0
1209  // fold: icmp null, (inttoptr x)         -> icmp 0, x
1210  // fold: icmp (ptrtoint x), 0            -> icmp x, null
1211  // fold: icmp 0, (ptrtoint x)            -> icmp null, x
1212  // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1213  // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1214  //
1215  // FIXME: The following comment is out of data and the DataLayout is here now.
1216  // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1217  // around to know if bit truncation is happening.
1218  if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1219    if (Ops1->isNullValue()) {
1220      if (CE0->getOpcode() == Instruction::IntToPtr) {
1221        Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1222        // Convert the integer value to the right size to ensure we get the
1223        // proper extension or truncation.
1224        Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1225                                                   IntPtrTy, false);
1226        Constant *Null = Constant::getNullValue(C->getType());
1227        return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1228      }
1229
1230      // Only do this transformation if the int is intptrty in size, otherwise
1231      // there is a truncation or extension that we aren't modeling.
1232      if (CE0->getOpcode() == Instruction::PtrToInt) {
1233        Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1234        if (CE0->getType() == IntPtrTy) {
1235          Constant *C = CE0->getOperand(0);
1236          Constant *Null = Constant::getNullValue(C->getType());
1237          return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1238        }
1239      }
1240    }
1241
1242    if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1243      if (CE0->getOpcode() == CE1->getOpcode()) {
1244        if (CE0->getOpcode() == Instruction::IntToPtr) {
1245          Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1246
1247          // Convert the integer value to the right size to ensure we get the
1248          // proper extension or truncation.
1249          Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1250                                                      IntPtrTy, false);
1251          Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1252                                                      IntPtrTy, false);
1253          return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1254        }
1255
1256        // Only do this transformation if the int is intptrty in size, otherwise
1257        // there is a truncation or extension that we aren't modeling.
1258        if (CE0->getOpcode() == Instruction::PtrToInt) {
1259          Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1260          if (CE0->getType() == IntPtrTy &&
1261              CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1262            return ConstantFoldCompareInstOperands(
1263                Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1264          }
1265        }
1266      }
1267    }
1268
1269    // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1270    // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1271    if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1272        CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1273      Constant *LHS = ConstantFoldCompareInstOperands(
1274          Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1275      Constant *RHS = ConstantFoldCompareInstOperands(
1276          Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1277      unsigned OpC =
1278        Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1279      return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1280    }
1281  } else if (isa<ConstantExpr>(Ops1)) {
1282    // If RHS is a constant expression, but the left side isn't, swap the
1283    // operands and try again.
1284    Predicate = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)Predicate);
1285    return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI);
1286  }
1287
1288  return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1289}
1290
1291Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op,
1292                                           const DataLayout &DL) {
1293  assert(Instruction::isUnaryOp(Opcode));
1294
1295  return ConstantExpr::get(Opcode, Op);
1296}
1297
1298Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1299                                             Constant *RHS,
1300                                             const DataLayout &DL) {
1301  assert(Instruction::isBinaryOp(Opcode));
1302  if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1303    if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1304      return C;
1305
1306  return ConstantExpr::get(Opcode, LHS, RHS);
1307}
1308
1309Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1310                                        Type *DestTy, const DataLayout &DL) {
1311  assert(Instruction::isCast(Opcode));
1312  switch (Opcode) {
1313  default:
1314    llvm_unreachable("Missing case");
1315  case Instruction::PtrToInt:
1316    // If the input is a inttoptr, eliminate the pair.  This requires knowing
1317    // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1318    if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1319      if (CE->getOpcode() == Instruction::IntToPtr) {
1320        Constant *Input = CE->getOperand(0);
1321        unsigned InWidth = Input->getType()->getScalarSizeInBits();
1322        unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1323        if (PtrWidth < InWidth) {
1324          Constant *Mask =
1325            ConstantInt::get(CE->getContext(),
1326                             APInt::getLowBitsSet(InWidth, PtrWidth));
1327          Input = ConstantExpr::getAnd(Input, Mask);
1328        }
1329        // Do a zext or trunc to get to the dest size.
1330        return ConstantExpr::getIntegerCast(Input, DestTy, false);
1331      }
1332    }
1333    return ConstantExpr::getCast(Opcode, C, DestTy);
1334  case Instruction::IntToPtr:
1335    // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1336    // the int size is >= the ptr size and the address spaces are the same.
1337    // This requires knowing the width of a pointer, so it can't be done in
1338    // ConstantExpr::getCast.
1339    if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1340      if (CE->getOpcode() == Instruction::PtrToInt) {
1341        Constant *SrcPtr = CE->getOperand(0);
1342        unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1343        unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1344
1345        if (MidIntSize >= SrcPtrSize) {
1346          unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1347          if (SrcAS == DestTy->getPointerAddressSpace())
1348            return FoldBitCast(CE->getOperand(0), DestTy, DL);
1349        }
1350      }
1351    }
1352
1353    return ConstantExpr::getCast(Opcode, C, DestTy);
1354  case Instruction::Trunc:
1355  case Instruction::ZExt:
1356  case Instruction::SExt:
1357  case Instruction::FPTrunc:
1358  case Instruction::FPExt:
1359  case Instruction::UIToFP:
1360  case Instruction::SIToFP:
1361  case Instruction::FPToUI:
1362  case Instruction::FPToSI:
1363  case Instruction::AddrSpaceCast:
1364      return ConstantExpr::getCast(Opcode, C, DestTy);
1365  case Instruction::BitCast:
1366    return FoldBitCast(C, DestTy, DL);
1367  }
1368}
1369
1370Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1371                                                       ConstantExpr *CE) {
1372  if (!CE->getOperand(1)->isNullValue())
1373    return nullptr;  // Do not allow stepping over the value!
1374
1375  // Loop over all of the operands, tracking down which value we are
1376  // addressing.
1377  for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1378    C = C->getAggregateElement(CE->getOperand(i));
1379    if (!C)
1380      return nullptr;
1381  }
1382  return C;
1383}
1384
1385Constant *
1386llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1387                                        ArrayRef<Constant *> Indices) {
1388  // Loop over all of the operands, tracking down which value we are
1389  // addressing.
1390  for (Constant *Index : Indices) {
1391    C = C->getAggregateElement(Index);
1392    if (!C)
1393      return nullptr;
1394  }
1395  return C;
1396}
1397
1398//===----------------------------------------------------------------------===//
1399//  Constant Folding for Calls
1400//
1401
1402bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
1403  if (Call->isNoBuiltin() || Call->isStrictFP())
1404    return false;
1405  switch (F->getIntrinsicID()) {
1406  case Intrinsic::fabs:
1407  case Intrinsic::minnum:
1408  case Intrinsic::maxnum:
1409  case Intrinsic::minimum:
1410  case Intrinsic::maximum:
1411  case Intrinsic::log:
1412  case Intrinsic::log2:
1413  case Intrinsic::log10:
1414  case Intrinsic::exp:
1415  case Intrinsic::exp2:
1416  case Intrinsic::floor:
1417  case Intrinsic::ceil:
1418  case Intrinsic::sqrt:
1419  case Intrinsic::sin:
1420  case Intrinsic::cos:
1421  case Intrinsic::trunc:
1422  case Intrinsic::rint:
1423  case Intrinsic::nearbyint:
1424  case Intrinsic::pow:
1425  case Intrinsic::powi:
1426  case Intrinsic::bswap:
1427  case Intrinsic::ctpop:
1428  case Intrinsic::ctlz:
1429  case Intrinsic::cttz:
1430  case Intrinsic::fshl:
1431  case Intrinsic::fshr:
1432  case Intrinsic::fma:
1433  case Intrinsic::fmuladd:
1434  case Intrinsic::copysign:
1435  case Intrinsic::launder_invariant_group:
1436  case Intrinsic::strip_invariant_group:
1437  case Intrinsic::round:
1438  case Intrinsic::masked_load:
1439  case Intrinsic::sadd_with_overflow:
1440  case Intrinsic::uadd_with_overflow:
1441  case Intrinsic::ssub_with_overflow:
1442  case Intrinsic::usub_with_overflow:
1443  case Intrinsic::smul_with_overflow:
1444  case Intrinsic::umul_with_overflow:
1445  case Intrinsic::sadd_sat:
1446  case Intrinsic::uadd_sat:
1447  case Intrinsic::ssub_sat:
1448  case Intrinsic::usub_sat:
1449  case Intrinsic::smul_fix:
1450  case Intrinsic::smul_fix_sat:
1451  case Intrinsic::convert_from_fp16:
1452  case Intrinsic::convert_to_fp16:
1453  case Intrinsic::bitreverse:
1454  case Intrinsic::x86_sse_cvtss2si:
1455  case Intrinsic::x86_sse_cvtss2si64:
1456  case Intrinsic::x86_sse_cvttss2si:
1457  case Intrinsic::x86_sse_cvttss2si64:
1458  case Intrinsic::x86_sse2_cvtsd2si:
1459  case Intrinsic::x86_sse2_cvtsd2si64:
1460  case Intrinsic::x86_sse2_cvttsd2si:
1461  case Intrinsic::x86_sse2_cvttsd2si64:
1462  case Intrinsic::x86_avx512_vcvtss2si32:
1463  case Intrinsic::x86_avx512_vcvtss2si64:
1464  case Intrinsic::x86_avx512_cvttss2si:
1465  case Intrinsic::x86_avx512_cvttss2si64:
1466  case Intrinsic::x86_avx512_vcvtsd2si32:
1467  case Intrinsic::x86_avx512_vcvtsd2si64:
1468  case Intrinsic::x86_avx512_cvttsd2si:
1469  case Intrinsic::x86_avx512_cvttsd2si64:
1470  case Intrinsic::x86_avx512_vcvtss2usi32:
1471  case Intrinsic::x86_avx512_vcvtss2usi64:
1472  case Intrinsic::x86_avx512_cvttss2usi:
1473  case Intrinsic::x86_avx512_cvttss2usi64:
1474  case Intrinsic::x86_avx512_vcvtsd2usi32:
1475  case Intrinsic::x86_avx512_vcvtsd2usi64:
1476  case Intrinsic::x86_avx512_cvttsd2usi:
1477  case Intrinsic::x86_avx512_cvttsd2usi64:
1478  case Intrinsic::is_constant:
1479    return true;
1480  default:
1481    return false;
1482  case Intrinsic::not_intrinsic: break;
1483  }
1484
1485  if (!F->hasName())
1486    return false;
1487
1488  // In these cases, the check of the length is required.  We don't want to
1489  // return true for a name like "cos\0blah" which strcmp would return equal to
1490  // "cos", but has length 8.
1491  StringRef Name = F->getName();
1492  switch (Name[0]) {
1493  default:
1494    return false;
1495  case 'a':
1496    return Name == "acos" || Name == "acosf" ||
1497           Name == "asin" || Name == "asinf" ||
1498           Name == "atan" || Name == "atanf" ||
1499           Name == "atan2" || Name == "atan2f";
1500  case 'c':
1501    return Name == "ceil" || Name == "ceilf" ||
1502           Name == "cos" || Name == "cosf" ||
1503           Name == "cosh" || Name == "coshf";
1504  case 'e':
1505    return Name == "exp" || Name == "expf" ||
1506           Name == "exp2" || Name == "exp2f";
1507  case 'f':
1508    return Name == "fabs" || Name == "fabsf" ||
1509           Name == "floor" || Name == "floorf" ||
1510           Name == "fmod" || Name == "fmodf";
1511  case 'l':
1512    return Name == "log" || Name == "logf" ||
1513           Name == "log2" || Name == "log2f" ||
1514           Name == "log10" || Name == "log10f";
1515  case 'n':
1516    return Name == "nearbyint" || Name == "nearbyintf";
1517  case 'p':
1518    return Name == "pow" || Name == "powf";
1519  case 'r':
1520    return Name == "rint" || Name == "rintf" ||
1521           Name == "round" || Name == "roundf";
1522  case 's':
1523    return Name == "sin" || Name == "sinf" ||
1524           Name == "sinh" || Name == "sinhf" ||
1525           Name == "sqrt" || Name == "sqrtf";
1526  case 't':
1527    return Name == "tan" || Name == "tanf" ||
1528           Name == "tanh" || Name == "tanhf" ||
1529           Name == "trunc" || Name == "truncf";
1530  case '_':
1531    // Check for various function names that get used for the math functions
1532    // when the header files are preprocessed with the macro
1533    // __FINITE_MATH_ONLY__ enabled.
1534    // The '12' here is the length of the shortest name that can match.
1535    // We need to check the size before looking at Name[1] and Name[2]
1536    // so we may as well check a limit that will eliminate mismatches.
1537    if (Name.size() < 12 || Name[1] != '_')
1538      return false;
1539    switch (Name[2]) {
1540    default:
1541      return false;
1542    case 'a':
1543      return Name == "__acos_finite" || Name == "__acosf_finite" ||
1544             Name == "__asin_finite" || Name == "__asinf_finite" ||
1545             Name == "__atan2_finite" || Name == "__atan2f_finite";
1546    case 'c':
1547      return Name == "__cosh_finite" || Name == "__coshf_finite";
1548    case 'e':
1549      return Name == "__exp_finite" || Name == "__expf_finite" ||
1550             Name == "__exp2_finite" || Name == "__exp2f_finite";
1551    case 'l':
1552      return Name == "__log_finite" || Name == "__logf_finite" ||
1553             Name == "__log10_finite" || Name == "__log10f_finite";
1554    case 'p':
1555      return Name == "__pow_finite" || Name == "__powf_finite";
1556    case 's':
1557      return Name == "__sinh_finite" || Name == "__sinhf_finite";
1558    }
1559  }
1560}
1561
1562namespace {
1563
1564Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1565  if (Ty->isHalfTy() || Ty->isFloatTy()) {
1566    APFloat APF(V);
1567    bool unused;
1568    APF.convert(Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &unused);
1569    return ConstantFP::get(Ty->getContext(), APF);
1570  }
1571  if (Ty->isDoubleTy())
1572    return ConstantFP::get(Ty->getContext(), APFloat(V));
1573  llvm_unreachable("Can only constant fold half/float/double");
1574}
1575
1576/// Clear the floating-point exception state.
1577inline void llvm_fenv_clearexcept() {
1578#if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1579  feclearexcept(FE_ALL_EXCEPT);
1580#endif
1581  errno = 0;
1582}
1583
1584/// Test if a floating-point exception was raised.
1585inline bool llvm_fenv_testexcept() {
1586  int errno_val = errno;
1587  if (errno_val == ERANGE || errno_val == EDOM)
1588    return true;
1589#if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1590  if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1591    return true;
1592#endif
1593  return false;
1594}
1595
1596Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) {
1597  llvm_fenv_clearexcept();
1598  V = NativeFP(V);
1599  if (llvm_fenv_testexcept()) {
1600    llvm_fenv_clearexcept();
1601    return nullptr;
1602  }
1603
1604  return GetConstantFoldFPValue(V, Ty);
1605}
1606
1607Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
1608                               double W, Type *Ty) {
1609  llvm_fenv_clearexcept();
1610  V = NativeFP(V, W);
1611  if (llvm_fenv_testexcept()) {
1612    llvm_fenv_clearexcept();
1613    return nullptr;
1614  }
1615
1616  return GetConstantFoldFPValue(V, Ty);
1617}
1618
1619/// Attempt to fold an SSE floating point to integer conversion of a constant
1620/// floating point. If roundTowardZero is false, the default IEEE rounding is
1621/// used (toward nearest, ties to even). This matches the behavior of the
1622/// non-truncating SSE instructions in the default rounding mode. The desired
1623/// integer type Ty is used to select how many bits are available for the
1624/// result. Returns null if the conversion cannot be performed, otherwise
1625/// returns the Constant value resulting from the conversion.
1626Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
1627                                      Type *Ty, bool IsSigned) {
1628  // All of these conversion intrinsics form an integer of at most 64bits.
1629  unsigned ResultWidth = Ty->getIntegerBitWidth();
1630  assert(ResultWidth <= 64 &&
1631         "Can only constant fold conversions to 64 and 32 bit ints");
1632
1633  uint64_t UIntVal;
1634  bool isExact = false;
1635  APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1636                                              : APFloat::rmNearestTiesToEven;
1637  APFloat::opStatus status =
1638      Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth,
1639                           IsSigned, mode, &isExact);
1640  if (status != APFloat::opOK &&
1641      (!roundTowardZero || status != APFloat::opInexact))
1642    return nullptr;
1643  return ConstantInt::get(Ty, UIntVal, IsSigned);
1644}
1645
1646double getValueAsDouble(ConstantFP *Op) {
1647  Type *Ty = Op->getType();
1648
1649  if (Ty->isFloatTy())
1650    return Op->getValueAPF().convertToFloat();
1651
1652  if (Ty->isDoubleTy())
1653    return Op->getValueAPF().convertToDouble();
1654
1655  bool unused;
1656  APFloat APF = Op->getValueAPF();
1657  APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
1658  return APF.convertToDouble();
1659}
1660
1661static bool isManifestConstant(const Constant *c) {
1662  if (isa<ConstantData>(c)) {
1663    return true;
1664  } else if (isa<ConstantAggregate>(c) || isa<ConstantExpr>(c)) {
1665    for (const Value *subc : c->operand_values()) {
1666      if (!isManifestConstant(cast<Constant>(subc)))
1667        return false;
1668    }
1669    return true;
1670  }
1671  return false;
1672}
1673
1674static bool getConstIntOrUndef(Value *Op, const APInt *&C) {
1675  if (auto *CI = dyn_cast<ConstantInt>(Op)) {
1676    C = &CI->getValue();
1677    return true;
1678  }
1679  if (isa<UndefValue>(Op)) {
1680    C = nullptr;
1681    return true;
1682  }
1683  return false;
1684}
1685
1686static Constant *ConstantFoldScalarCall1(StringRef Name,
1687                                         Intrinsic::ID IntrinsicID,
1688                                         Type *Ty,
1689                                         ArrayRef<Constant *> Operands,
1690                                         const TargetLibraryInfo *TLI,
1691                                         const CallBase *Call) {
1692  assert(Operands.size() == 1 && "Wrong number of operands.");
1693
1694  if (IntrinsicID == Intrinsic::is_constant) {
1695    // We know we have a "Constant" argument. But we want to only
1696    // return true for manifest constants, not those that depend on
1697    // constants with unknowable values, e.g. GlobalValue or BlockAddress.
1698    if (isManifestConstant(Operands[0]))
1699      return ConstantInt::getTrue(Ty->getContext());
1700    return nullptr;
1701  }
1702  if (isa<UndefValue>(Operands[0])) {
1703    // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.
1704    // ctpop() is between 0 and bitwidth, pick 0 for undef.
1705    if (IntrinsicID == Intrinsic::cos ||
1706        IntrinsicID == Intrinsic::ctpop)
1707      return Constant::getNullValue(Ty);
1708    if (IntrinsicID == Intrinsic::bswap ||
1709        IntrinsicID == Intrinsic::bitreverse ||
1710        IntrinsicID == Intrinsic::launder_invariant_group ||
1711        IntrinsicID == Intrinsic::strip_invariant_group)
1712      return Operands[0];
1713  }
1714
1715  if (isa<ConstantPointerNull>(Operands[0])) {
1716    // launder(null) == null == strip(null) iff in addrspace 0
1717    if (IntrinsicID == Intrinsic::launder_invariant_group ||
1718        IntrinsicID == Intrinsic::strip_invariant_group) {
1719      // If instruction is not yet put in a basic block (e.g. when cloning
1720      // a function during inlining), Call's caller may not be available.
1721      // So check Call's BB first before querying Call->getCaller.
1722      const Function *Caller =
1723          Call->getParent() ? Call->getCaller() : nullptr;
1724      if (Caller &&
1725          !NullPointerIsDefined(
1726              Caller, Operands[0]->getType()->getPointerAddressSpace())) {
1727        return Operands[0];
1728      }
1729      return nullptr;
1730    }
1731  }
1732
1733  if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1734    if (IntrinsicID == Intrinsic::convert_to_fp16) {
1735      APFloat Val(Op->getValueAPF());
1736
1737      bool lost = false;
1738      Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);
1739
1740      return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1741    }
1742
1743    if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1744      return nullptr;
1745
1746    // Use internal versions of these intrinsics.
1747    APFloat U = Op->getValueAPF();
1748
1749    if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint) {
1750      U.roundToIntegral(APFloat::rmNearestTiesToEven);
1751      return ConstantFP::get(Ty->getContext(), U);
1752    }
1753
1754    if (IntrinsicID == Intrinsic::round) {
1755      U.roundToIntegral(APFloat::rmNearestTiesToAway);
1756      return ConstantFP::get(Ty->getContext(), U);
1757    }
1758
1759    if (IntrinsicID == Intrinsic::ceil) {
1760      U.roundToIntegral(APFloat::rmTowardPositive);
1761      return ConstantFP::get(Ty->getContext(), U);
1762    }
1763
1764    if (IntrinsicID == Intrinsic::floor) {
1765      U.roundToIntegral(APFloat::rmTowardNegative);
1766      return ConstantFP::get(Ty->getContext(), U);
1767    }
1768
1769    if (IntrinsicID == Intrinsic::trunc) {
1770      U.roundToIntegral(APFloat::rmTowardZero);
1771      return ConstantFP::get(Ty->getContext(), U);
1772    }
1773
1774    if (IntrinsicID == Intrinsic::fabs) {
1775      U.clearSign();
1776      return ConstantFP::get(Ty->getContext(), U);
1777    }
1778
1779    /// We only fold functions with finite arguments. Folding NaN and inf is
1780    /// likely to be aborted with an exception anyway, and some host libms
1781    /// have known errors raising exceptions.
1782    if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1783      return nullptr;
1784
1785    /// Currently APFloat versions of these functions do not exist, so we use
1786    /// the host native double versions.  Float versions are not called
1787    /// directly but for all these it is true (float)(f((double)arg)) ==
1788    /// f(arg).  Long double not supported yet.
1789    double V = getValueAsDouble(Op);
1790
1791    switch (IntrinsicID) {
1792      default: break;
1793      case Intrinsic::log:
1794        return ConstantFoldFP(log, V, Ty);
1795      case Intrinsic::log2:
1796        // TODO: What about hosts that lack a C99 library?
1797        return ConstantFoldFP(Log2, V, Ty);
1798      case Intrinsic::log10:
1799        // TODO: What about hosts that lack a C99 library?
1800        return ConstantFoldFP(log10, V, Ty);
1801      case Intrinsic::exp:
1802        return ConstantFoldFP(exp, V, Ty);
1803      case Intrinsic::exp2:
1804        // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
1805        return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1806      case Intrinsic::sin:
1807        return ConstantFoldFP(sin, V, Ty);
1808      case Intrinsic::cos:
1809        return ConstantFoldFP(cos, V, Ty);
1810      case Intrinsic::sqrt:
1811        return ConstantFoldFP(sqrt, V, Ty);
1812    }
1813
1814    if (!TLI)
1815      return nullptr;
1816
1817    LibFunc Func = NotLibFunc;
1818    TLI->getLibFunc(Name, Func);
1819    switch (Func) {
1820    default:
1821      break;
1822    case LibFunc_acos:
1823    case LibFunc_acosf:
1824    case LibFunc_acos_finite:
1825    case LibFunc_acosf_finite:
1826      if (TLI->has(Func))
1827        return ConstantFoldFP(acos, V, Ty);
1828      break;
1829    case LibFunc_asin:
1830    case LibFunc_asinf:
1831    case LibFunc_asin_finite:
1832    case LibFunc_asinf_finite:
1833      if (TLI->has(Func))
1834        return ConstantFoldFP(asin, V, Ty);
1835      break;
1836    case LibFunc_atan:
1837    case LibFunc_atanf:
1838      if (TLI->has(Func))
1839        return ConstantFoldFP(atan, V, Ty);
1840      break;
1841    case LibFunc_ceil:
1842    case LibFunc_ceilf:
1843      if (TLI->has(Func)) {
1844        U.roundToIntegral(APFloat::rmTowardPositive);
1845        return ConstantFP::get(Ty->getContext(), U);
1846      }
1847      break;
1848    case LibFunc_cos:
1849    case LibFunc_cosf:
1850      if (TLI->has(Func))
1851        return ConstantFoldFP(cos, V, Ty);
1852      break;
1853    case LibFunc_cosh:
1854    case LibFunc_coshf:
1855    case LibFunc_cosh_finite:
1856    case LibFunc_coshf_finite:
1857      if (TLI->has(Func))
1858        return ConstantFoldFP(cosh, V, Ty);
1859      break;
1860    case LibFunc_exp:
1861    case LibFunc_expf:
1862    case LibFunc_exp_finite:
1863    case LibFunc_expf_finite:
1864      if (TLI->has(Func))
1865        return ConstantFoldFP(exp, V, Ty);
1866      break;
1867    case LibFunc_exp2:
1868    case LibFunc_exp2f:
1869    case LibFunc_exp2_finite:
1870    case LibFunc_exp2f_finite:
1871      if (TLI->has(Func))
1872        // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
1873        return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1874      break;
1875    case LibFunc_fabs:
1876    case LibFunc_fabsf:
1877      if (TLI->has(Func)) {
1878        U.clearSign();
1879        return ConstantFP::get(Ty->getContext(), U);
1880      }
1881      break;
1882    case LibFunc_floor:
1883    case LibFunc_floorf:
1884      if (TLI->has(Func)) {
1885        U.roundToIntegral(APFloat::rmTowardNegative);
1886        return ConstantFP::get(Ty->getContext(), U);
1887      }
1888      break;
1889    case LibFunc_log:
1890    case LibFunc_logf:
1891    case LibFunc_log_finite:
1892    case LibFunc_logf_finite:
1893      if (V > 0.0 && TLI->has(Func))
1894        return ConstantFoldFP(log, V, Ty);
1895      break;
1896    case LibFunc_log2:
1897    case LibFunc_log2f:
1898    case LibFunc_log2_finite:
1899    case LibFunc_log2f_finite:
1900      if (V > 0.0 && TLI->has(Func))
1901        // TODO: What about hosts that lack a C99 library?
1902        return ConstantFoldFP(Log2, V, Ty);
1903      break;
1904    case LibFunc_log10:
1905    case LibFunc_log10f:
1906    case LibFunc_log10_finite:
1907    case LibFunc_log10f_finite:
1908      if (V > 0.0 && TLI->has(Func))
1909        // TODO: What about hosts that lack a C99 library?
1910        return ConstantFoldFP(log10, V, Ty);
1911      break;
1912    case LibFunc_nearbyint:
1913    case LibFunc_nearbyintf:
1914    case LibFunc_rint:
1915    case LibFunc_rintf:
1916      if (TLI->has(Func)) {
1917        U.roundToIntegral(APFloat::rmNearestTiesToEven);
1918        return ConstantFP::get(Ty->getContext(), U);
1919      }
1920      break;
1921    case LibFunc_round:
1922    case LibFunc_roundf:
1923      if (TLI->has(Func)) {
1924        U.roundToIntegral(APFloat::rmNearestTiesToAway);
1925        return ConstantFP::get(Ty->getContext(), U);
1926      }
1927      break;
1928    case LibFunc_sin:
1929    case LibFunc_sinf:
1930      if (TLI->has(Func))
1931        return ConstantFoldFP(sin, V, Ty);
1932      break;
1933    case LibFunc_sinh:
1934    case LibFunc_sinhf:
1935    case LibFunc_sinh_finite:
1936    case LibFunc_sinhf_finite:
1937      if (TLI->has(Func))
1938        return ConstantFoldFP(sinh, V, Ty);
1939      break;
1940    case LibFunc_sqrt:
1941    case LibFunc_sqrtf:
1942      if (V >= 0.0 && TLI->has(Func))
1943        return ConstantFoldFP(sqrt, V, Ty);
1944      break;
1945    case LibFunc_tan:
1946    case LibFunc_tanf:
1947      if (TLI->has(Func))
1948        return ConstantFoldFP(tan, V, Ty);
1949      break;
1950    case LibFunc_tanh:
1951    case LibFunc_tanhf:
1952      if (TLI->has(Func))
1953        return ConstantFoldFP(tanh, V, Ty);
1954      break;
1955    case LibFunc_trunc:
1956    case LibFunc_truncf:
1957      if (TLI->has(Func)) {
1958        U.roundToIntegral(APFloat::rmTowardZero);
1959        return ConstantFP::get(Ty->getContext(), U);
1960      }
1961      break;
1962    }
1963    return nullptr;
1964  }
1965
1966  if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
1967    switch (IntrinsicID) {
1968    case Intrinsic::bswap:
1969      return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1970    case Intrinsic::ctpop:
1971      return ConstantInt::get(Ty, Op->getValue().countPopulation());
1972    case Intrinsic::bitreverse:
1973      return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
1974    case Intrinsic::convert_from_fp16: {
1975      APFloat Val(APFloat::IEEEhalf(), Op->getValue());
1976
1977      bool lost = false;
1978      APFloat::opStatus status = Val.convert(
1979          Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
1980
1981      // Conversion is always precise.
1982      (void)status;
1983      assert(status == APFloat::opOK && !lost &&
1984             "Precision lost during fp16 constfolding");
1985
1986      return ConstantFP::get(Ty->getContext(), Val);
1987    }
1988    default:
1989      return nullptr;
1990    }
1991  }
1992
1993  // Support ConstantVector in case we have an Undef in the top.
1994  if (isa<ConstantVector>(Operands[0]) ||
1995      isa<ConstantDataVector>(Operands[0])) {
1996    auto *Op = cast<Constant>(Operands[0]);
1997    switch (IntrinsicID) {
1998    default: break;
1999    case Intrinsic::x86_sse_cvtss2si:
2000    case Intrinsic::x86_sse_cvtss2si64:
2001    case Intrinsic::x86_sse2_cvtsd2si:
2002    case Intrinsic::x86_sse2_cvtsd2si64:
2003      if (ConstantFP *FPOp =
2004              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2005        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2006                                           /*roundTowardZero=*/false, Ty,
2007                                           /*IsSigned*/true);
2008      break;
2009    case Intrinsic::x86_sse_cvttss2si:
2010    case Intrinsic::x86_sse_cvttss2si64:
2011    case Intrinsic::x86_sse2_cvttsd2si:
2012    case Intrinsic::x86_sse2_cvttsd2si64:
2013      if (ConstantFP *FPOp =
2014              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2015        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2016                                           /*roundTowardZero=*/true, Ty,
2017                                           /*IsSigned*/true);
2018      break;
2019    }
2020  }
2021
2022  return nullptr;
2023}
2024
2025static Constant *ConstantFoldScalarCall2(StringRef Name,
2026                                         Intrinsic::ID IntrinsicID,
2027                                         Type *Ty,
2028                                         ArrayRef<Constant *> Operands,
2029                                         const TargetLibraryInfo *TLI,
2030                                         const CallBase *Call) {
2031  assert(Operands.size() == 2 && "Wrong number of operands.");
2032
2033  if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
2034    if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
2035      return nullptr;
2036    double Op1V = getValueAsDouble(Op1);
2037
2038    if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
2039      if (Op2->getType() != Op1->getType())
2040        return nullptr;
2041
2042      double Op2V = getValueAsDouble(Op2);
2043      if (IntrinsicID == Intrinsic::pow) {
2044        return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
2045      }
2046      if (IntrinsicID == Intrinsic::copysign) {
2047        APFloat V1 = Op1->getValueAPF();
2048        const APFloat &V2 = Op2->getValueAPF();
2049        V1.copySign(V2);
2050        return ConstantFP::get(Ty->getContext(), V1);
2051      }
2052
2053      if (IntrinsicID == Intrinsic::minnum) {
2054        const APFloat &C1 = Op1->getValueAPF();
2055        const APFloat &C2 = Op2->getValueAPF();
2056        return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
2057      }
2058
2059      if (IntrinsicID == Intrinsic::maxnum) {
2060        const APFloat &C1 = Op1->getValueAPF();
2061        const APFloat &C2 = Op2->getValueAPF();
2062        return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
2063      }
2064
2065      if (IntrinsicID == Intrinsic::minimum) {
2066        const APFloat &C1 = Op1->getValueAPF();
2067        const APFloat &C2 = Op2->getValueAPF();
2068        return ConstantFP::get(Ty->getContext(), minimum(C1, C2));
2069      }
2070
2071      if (IntrinsicID == Intrinsic::maximum) {
2072        const APFloat &C1 = Op1->getValueAPF();
2073        const APFloat &C2 = Op2->getValueAPF();
2074        return ConstantFP::get(Ty->getContext(), maximum(C1, C2));
2075      }
2076
2077      if (!TLI)
2078        return nullptr;
2079
2080      LibFunc Func = NotLibFunc;
2081      TLI->getLibFunc(Name, Func);
2082      switch (Func) {
2083      default:
2084        break;
2085      case LibFunc_pow:
2086      case LibFunc_powf:
2087      case LibFunc_pow_finite:
2088      case LibFunc_powf_finite:
2089        if (TLI->has(Func))
2090          return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
2091        break;
2092      case LibFunc_fmod:
2093      case LibFunc_fmodf:
2094        if (TLI->has(Func)) {
2095          APFloat V = Op1->getValueAPF();
2096          if (APFloat::opStatus::opOK == V.mod(Op2->getValueAPF()))
2097            return ConstantFP::get(Ty->getContext(), V);
2098        }
2099        break;
2100      case LibFunc_atan2:
2101      case LibFunc_atan2f:
2102      case LibFunc_atan2_finite:
2103      case LibFunc_atan2f_finite:
2104        if (TLI->has(Func))
2105          return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
2106        break;
2107      }
2108    } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
2109      if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
2110        return ConstantFP::get(Ty->getContext(),
2111                               APFloat((float)std::pow((float)Op1V,
2112                                               (int)Op2C->getZExtValue())));
2113      if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
2114        return ConstantFP::get(Ty->getContext(),
2115                               APFloat((float)std::pow((float)Op1V,
2116                                               (int)Op2C->getZExtValue())));
2117      if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
2118        return ConstantFP::get(Ty->getContext(),
2119                               APFloat((double)std::pow((double)Op1V,
2120                                                 (int)Op2C->getZExtValue())));
2121    }
2122    return nullptr;
2123  }
2124
2125  if (Operands[0]->getType()->isIntegerTy() &&
2126      Operands[1]->getType()->isIntegerTy()) {
2127    const APInt *C0, *C1;
2128    if (!getConstIntOrUndef(Operands[0], C0) ||
2129        !getConstIntOrUndef(Operands[1], C1))
2130      return nullptr;
2131
2132    switch (IntrinsicID) {
2133    default: break;
2134    case Intrinsic::usub_with_overflow:
2135    case Intrinsic::ssub_with_overflow:
2136    case Intrinsic::uadd_with_overflow:
2137    case Intrinsic::sadd_with_overflow:
2138      // X - undef -> { undef, false }
2139      // undef - X -> { undef, false }
2140      // X + undef -> { undef, false }
2141      // undef + x -> { undef, false }
2142      if (!C0 || !C1) {
2143        return ConstantStruct::get(
2144            cast<StructType>(Ty),
2145            {UndefValue::get(Ty->getStructElementType(0)),
2146             Constant::getNullValue(Ty->getStructElementType(1))});
2147      }
2148      LLVM_FALLTHROUGH;
2149    case Intrinsic::smul_with_overflow:
2150    case Intrinsic::umul_with_overflow: {
2151      // undef * X -> { 0, false }
2152      // X * undef -> { 0, false }
2153      if (!C0 || !C1)
2154        return Constant::getNullValue(Ty);
2155
2156      APInt Res;
2157      bool Overflow;
2158      switch (IntrinsicID) {
2159      default: llvm_unreachable("Invalid case");
2160      case Intrinsic::sadd_with_overflow:
2161        Res = C0->sadd_ov(*C1, Overflow);
2162        break;
2163      case Intrinsic::uadd_with_overflow:
2164        Res = C0->uadd_ov(*C1, Overflow);
2165        break;
2166      case Intrinsic::ssub_with_overflow:
2167        Res = C0->ssub_ov(*C1, Overflow);
2168        break;
2169      case Intrinsic::usub_with_overflow:
2170        Res = C0->usub_ov(*C1, Overflow);
2171        break;
2172      case Intrinsic::smul_with_overflow:
2173        Res = C0->smul_ov(*C1, Overflow);
2174        break;
2175      case Intrinsic::umul_with_overflow:
2176        Res = C0->umul_ov(*C1, Overflow);
2177        break;
2178      }
2179      Constant *Ops[] = {
2180        ConstantInt::get(Ty->getContext(), Res),
2181        ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
2182      };
2183      return ConstantStruct::get(cast<StructType>(Ty), Ops);
2184    }
2185    case Intrinsic::uadd_sat:
2186    case Intrinsic::sadd_sat:
2187      if (!C0 && !C1)
2188        return UndefValue::get(Ty);
2189      if (!C0 || !C1)
2190        return Constant::getAllOnesValue(Ty);
2191      if (IntrinsicID == Intrinsic::uadd_sat)
2192        return ConstantInt::get(Ty, C0->uadd_sat(*C1));
2193      else
2194        return ConstantInt::get(Ty, C0->sadd_sat(*C1));
2195    case Intrinsic::usub_sat:
2196    case Intrinsic::ssub_sat:
2197      if (!C0 && !C1)
2198        return UndefValue::get(Ty);
2199      if (!C0 || !C1)
2200        return Constant::getNullValue(Ty);
2201      if (IntrinsicID == Intrinsic::usub_sat)
2202        return ConstantInt::get(Ty, C0->usub_sat(*C1));
2203      else
2204        return ConstantInt::get(Ty, C0->ssub_sat(*C1));
2205    case Intrinsic::cttz:
2206    case Intrinsic::ctlz:
2207      assert(C1 && "Must be constant int");
2208
2209      // cttz(0, 1) and ctlz(0, 1) are undef.
2210      if (C1->isOneValue() && (!C0 || C0->isNullValue()))
2211        return UndefValue::get(Ty);
2212      if (!C0)
2213        return Constant::getNullValue(Ty);
2214      if (IntrinsicID == Intrinsic::cttz)
2215        return ConstantInt::get(Ty, C0->countTrailingZeros());
2216      else
2217        return ConstantInt::get(Ty, C0->countLeadingZeros());
2218    }
2219
2220    return nullptr;
2221  }
2222
2223  // Support ConstantVector in case we have an Undef in the top.
2224  if ((isa<ConstantVector>(Operands[0]) ||
2225       isa<ConstantDataVector>(Operands[0])) &&
2226      // Check for default rounding mode.
2227      // FIXME: Support other rounding modes?
2228      isa<ConstantInt>(Operands[1]) &&
2229      cast<ConstantInt>(Operands[1])->getValue() == 4) {
2230    auto *Op = cast<Constant>(Operands[0]);
2231    switch (IntrinsicID) {
2232    default: break;
2233    case Intrinsic::x86_avx512_vcvtss2si32:
2234    case Intrinsic::x86_avx512_vcvtss2si64:
2235    case Intrinsic::x86_avx512_vcvtsd2si32:
2236    case Intrinsic::x86_avx512_vcvtsd2si64:
2237      if (ConstantFP *FPOp =
2238              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2239        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2240                                           /*roundTowardZero=*/false, Ty,
2241                                           /*IsSigned*/true);
2242      break;
2243    case Intrinsic::x86_avx512_vcvtss2usi32:
2244    case Intrinsic::x86_avx512_vcvtss2usi64:
2245    case Intrinsic::x86_avx512_vcvtsd2usi32:
2246    case Intrinsic::x86_avx512_vcvtsd2usi64:
2247      if (ConstantFP *FPOp =
2248              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2249        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2250                                           /*roundTowardZero=*/false, Ty,
2251                                           /*IsSigned*/false);
2252      break;
2253    case Intrinsic::x86_avx512_cvttss2si:
2254    case Intrinsic::x86_avx512_cvttss2si64:
2255    case Intrinsic::x86_avx512_cvttsd2si:
2256    case Intrinsic::x86_avx512_cvttsd2si64:
2257      if (ConstantFP *FPOp =
2258              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2259        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2260                                           /*roundTowardZero=*/true, Ty,
2261                                           /*IsSigned*/true);
2262      break;
2263    case Intrinsic::x86_avx512_cvttss2usi:
2264    case Intrinsic::x86_avx512_cvttss2usi64:
2265    case Intrinsic::x86_avx512_cvttsd2usi:
2266    case Intrinsic::x86_avx512_cvttsd2usi64:
2267      if (ConstantFP *FPOp =
2268              dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2269        return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2270                                           /*roundTowardZero=*/true, Ty,
2271                                           /*IsSigned*/false);
2272      break;
2273    }
2274  }
2275  return nullptr;
2276}
2277
2278static Constant *ConstantFoldScalarCall3(StringRef Name,
2279                                         Intrinsic::ID IntrinsicID,
2280                                         Type *Ty,
2281                                         ArrayRef<Constant *> Operands,
2282                                         const TargetLibraryInfo *TLI,
2283                                         const CallBase *Call) {
2284  assert(Operands.size() == 3 && "Wrong number of operands.");
2285
2286  if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
2287    if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
2288      if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
2289        switch (IntrinsicID) {
2290        default: break;
2291        case Intrinsic::fma:
2292        case Intrinsic::fmuladd: {
2293          APFloat V = Op1->getValueAPF();
2294          V.fusedMultiplyAdd(Op2->getValueAPF(), Op3->getValueAPF(),
2295                             APFloat::rmNearestTiesToEven);
2296          return ConstantFP::get(Ty->getContext(), V);
2297        }
2298        }
2299      }
2300    }
2301  }
2302
2303  if (const auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
2304    if (const auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
2305      if (const auto *Op3 = dyn_cast<ConstantInt>(Operands[2])) {
2306        switch (IntrinsicID) {
2307        default: break;
2308        case Intrinsic::smul_fix:
2309        case Intrinsic::smul_fix_sat: {
2310          // This code performs rounding towards negative infinity in case the
2311          // result cannot be represented exactly for the given scale. Targets
2312          // that do care about rounding should use a target hook for specifying
2313          // how rounding should be done, and provide their own folding to be
2314          // consistent with rounding. This is the same approach as used by
2315          // DAGTypeLegalizer::ExpandIntRes_MULFIX.
2316          APInt Lhs = Op1->getValue();
2317          APInt Rhs = Op2->getValue();
2318          unsigned Scale = Op3->getValue().getZExtValue();
2319          unsigned Width = Lhs.getBitWidth();
2320          assert(Scale < Width && "Illegal scale.");
2321          unsigned ExtendedWidth = Width * 2;
2322          APInt Product = (Lhs.sextOrSelf(ExtendedWidth) *
2323                           Rhs.sextOrSelf(ExtendedWidth)).ashr(Scale);
2324          if (IntrinsicID == Intrinsic::smul_fix_sat) {
2325            APInt MaxValue =
2326              APInt::getSignedMaxValue(Width).sextOrSelf(ExtendedWidth);
2327            APInt MinValue =
2328              APInt::getSignedMinValue(Width).sextOrSelf(ExtendedWidth);
2329            Product = APIntOps::smin(Product, MaxValue);
2330            Product = APIntOps::smax(Product, MinValue);
2331          }
2332          return ConstantInt::get(Ty->getContext(),
2333                                  Product.sextOrTrunc(Width));
2334        }
2335        }
2336      }
2337    }
2338  }
2339
2340  if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
2341    const APInt *C0, *C1, *C2;
2342    if (!getConstIntOrUndef(Operands[0], C0) ||
2343        !getConstIntOrUndef(Operands[1], C1) ||
2344        !getConstIntOrUndef(Operands[2], C2))
2345      return nullptr;
2346
2347    bool IsRight = IntrinsicID == Intrinsic::fshr;
2348    if (!C2)
2349      return Operands[IsRight ? 1 : 0];
2350    if (!C0 && !C1)
2351      return UndefValue::get(Ty);
2352
2353    // The shift amount is interpreted as modulo the bitwidth. If the shift
2354    // amount is effectively 0, avoid UB due to oversized inverse shift below.
2355    unsigned BitWidth = C2->getBitWidth();
2356    unsigned ShAmt = C2->urem(BitWidth);
2357    if (!ShAmt)
2358      return Operands[IsRight ? 1 : 0];
2359
2360    // (C0 << ShlAmt) | (C1 >> LshrAmt)
2361    unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;
2362    unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;
2363    if (!C0)
2364      return ConstantInt::get(Ty, C1->lshr(LshrAmt));
2365    if (!C1)
2366      return ConstantInt::get(Ty, C0->shl(ShlAmt));
2367    return ConstantInt::get(Ty, C0->shl(ShlAmt) | C1->lshr(LshrAmt));
2368  }
2369
2370  return nullptr;
2371}
2372
2373static Constant *ConstantFoldScalarCall(StringRef Name,
2374                                        Intrinsic::ID IntrinsicID,
2375                                        Type *Ty,
2376                                        ArrayRef<Constant *> Operands,
2377                                        const TargetLibraryInfo *TLI,
2378                                        const CallBase *Call) {
2379  if (Operands.size() == 1)
2380    return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call);
2381
2382  if (Operands.size() == 2)
2383    return ConstantFoldScalarCall2(Name, IntrinsicID, Ty, Operands, TLI, Call);
2384
2385  if (Operands.size() == 3)
2386    return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call);
2387
2388  return nullptr;
2389}
2390
2391static Constant *ConstantFoldVectorCall(StringRef Name,
2392                                        Intrinsic::ID IntrinsicID,
2393                                        VectorType *VTy,
2394                                        ArrayRef<Constant *> Operands,
2395                                        const DataLayout &DL,
2396                                        const TargetLibraryInfo *TLI,
2397                                        const CallBase *Call) {
2398  SmallVector<Constant *, 4> Result(VTy->getNumElements());
2399  SmallVector<Constant *, 4> Lane(Operands.size());
2400  Type *Ty = VTy->getElementType();
2401
2402  if (IntrinsicID == Intrinsic::masked_load) {
2403    auto *SrcPtr = Operands[0];
2404    auto *Mask = Operands[2];
2405    auto *Passthru = Operands[3];
2406
2407    Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL);
2408
2409    SmallVector<Constant *, 32> NewElements;
2410    for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2411      auto *MaskElt = Mask->getAggregateElement(I);
2412      if (!MaskElt)
2413        break;
2414      auto *PassthruElt = Passthru->getAggregateElement(I);
2415      auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
2416      if (isa<UndefValue>(MaskElt)) {
2417        if (PassthruElt)
2418          NewElements.push_back(PassthruElt);
2419        else if (VecElt)
2420          NewElements.push_back(VecElt);
2421        else
2422          return nullptr;
2423      }
2424      if (MaskElt->isNullValue()) {
2425        if (!PassthruElt)
2426          return nullptr;
2427        NewElements.push_back(PassthruElt);
2428      } else if (MaskElt->isOneValue()) {
2429        if (!VecElt)
2430          return nullptr;
2431        NewElements.push_back(VecElt);
2432      } else {
2433        return nullptr;
2434      }
2435    }
2436    if (NewElements.size() != VTy->getNumElements())
2437      return nullptr;
2438    return ConstantVector::get(NewElements);
2439  }
2440
2441  for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2442    // Gather a column of constants.
2443    for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
2444      // Some intrinsics use a scalar type for certain arguments.
2445      if (hasVectorInstrinsicScalarOpd(IntrinsicID, J)) {
2446        Lane[J] = Operands[J];
2447        continue;
2448      }
2449
2450      Constant *Agg = Operands[J]->getAggregateElement(I);
2451      if (!Agg)
2452        return nullptr;
2453
2454      Lane[J] = Agg;
2455    }
2456
2457    // Use the regular scalar folding to simplify this column.
2458    Constant *Folded =
2459        ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, Call);
2460    if (!Folded)
2461      return nullptr;
2462    Result[I] = Folded;
2463  }
2464
2465  return ConstantVector::get(Result);
2466}
2467
2468} // end anonymous namespace
2469
2470Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F,
2471                                 ArrayRef<Constant *> Operands,
2472                                 const TargetLibraryInfo *TLI) {
2473  if (Call->isNoBuiltin() || Call->isStrictFP())
2474    return nullptr;
2475  if (!F->hasName())
2476    return nullptr;
2477  StringRef Name = F->getName();
2478
2479  Type *Ty = F->getReturnType();
2480
2481  if (auto *VTy = dyn_cast<VectorType>(Ty))
2482    return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
2483                                  F->getParent()->getDataLayout(), TLI, Call);
2484
2485  return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI,
2486                                Call);
2487}
2488
2489bool llvm::isMathLibCallNoop(const CallBase *Call,
2490                             const TargetLibraryInfo *TLI) {
2491  // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
2492  // (and to some extent ConstantFoldScalarCall).
2493  if (Call->isNoBuiltin() || Call->isStrictFP())
2494    return false;
2495  Function *F = Call->getCalledFunction();
2496  if (!F)
2497    return false;
2498
2499  LibFunc Func;
2500  if (!TLI || !TLI->getLibFunc(*F, Func))
2501    return false;
2502
2503  if (Call->getNumArgOperands() == 1) {
2504    if (ConstantFP *OpC = dyn_cast<ConstantFP>(Call->getArgOperand(0))) {
2505      const APFloat &Op = OpC->getValueAPF();
2506      switch (Func) {
2507      case LibFunc_logl:
2508      case LibFunc_log:
2509      case LibFunc_logf:
2510      case LibFunc_log2l:
2511      case LibFunc_log2:
2512      case LibFunc_log2f:
2513      case LibFunc_log10l:
2514      case LibFunc_log10:
2515      case LibFunc_log10f:
2516        return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
2517
2518      case LibFunc_expl:
2519      case LibFunc_exp:
2520      case LibFunc_expf:
2521        // FIXME: These boundaries are slightly conservative.
2522        if (OpC->getType()->isDoubleTy())
2523          return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan &&
2524                 Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan;
2525        if (OpC->getType()->isFloatTy())
2526          return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan &&
2527                 Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan;
2528        break;
2529
2530      case LibFunc_exp2l:
2531      case LibFunc_exp2:
2532      case LibFunc_exp2f:
2533        // FIXME: These boundaries are slightly conservative.
2534        if (OpC->getType()->isDoubleTy())
2535          return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan &&
2536                 Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan;
2537        if (OpC->getType()->isFloatTy())
2538          return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan &&
2539                 Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan;
2540        break;
2541
2542      case LibFunc_sinl:
2543      case LibFunc_sin:
2544      case LibFunc_sinf:
2545      case LibFunc_cosl:
2546      case LibFunc_cos:
2547      case LibFunc_cosf:
2548        return !Op.isInfinity();
2549
2550      case LibFunc_tanl:
2551      case LibFunc_tan:
2552      case LibFunc_tanf: {
2553        // FIXME: Stop using the host math library.
2554        // FIXME: The computation isn't done in the right precision.
2555        Type *Ty = OpC->getType();
2556        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2557          double OpV = getValueAsDouble(OpC);
2558          return ConstantFoldFP(tan, OpV, Ty) != nullptr;
2559        }
2560        break;
2561      }
2562
2563      case LibFunc_asinl:
2564      case LibFunc_asin:
2565      case LibFunc_asinf:
2566      case LibFunc_acosl:
2567      case LibFunc_acos:
2568      case LibFunc_acosf:
2569        return Op.compare(APFloat(Op.getSemantics(), "-1")) !=
2570                   APFloat::cmpLessThan &&
2571               Op.compare(APFloat(Op.getSemantics(), "1")) !=
2572                   APFloat::cmpGreaterThan;
2573
2574      case LibFunc_sinh:
2575      case LibFunc_cosh:
2576      case LibFunc_sinhf:
2577      case LibFunc_coshf:
2578      case LibFunc_sinhl:
2579      case LibFunc_coshl:
2580        // FIXME: These boundaries are slightly conservative.
2581        if (OpC->getType()->isDoubleTy())
2582          return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan &&
2583                 Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan;
2584        if (OpC->getType()->isFloatTy())
2585          return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan &&
2586                 Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan;
2587        break;
2588
2589      case LibFunc_sqrtl:
2590      case LibFunc_sqrt:
2591      case LibFunc_sqrtf:
2592        return Op.isNaN() || Op.isZero() || !Op.isNegative();
2593
2594      // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
2595      // maybe others?
2596      default:
2597        break;
2598      }
2599    }
2600  }
2601
2602  if (Call->getNumArgOperands() == 2) {
2603    ConstantFP *Op0C = dyn_cast<ConstantFP>(Call->getArgOperand(0));
2604    ConstantFP *Op1C = dyn_cast<ConstantFP>(Call->getArgOperand(1));
2605    if (Op0C && Op1C) {
2606      const APFloat &Op0 = Op0C->getValueAPF();
2607      const APFloat &Op1 = Op1C->getValueAPF();
2608
2609      switch (Func) {
2610      case LibFunc_powl:
2611      case LibFunc_pow:
2612      case LibFunc_powf: {
2613        // FIXME: Stop using the host math library.
2614        // FIXME: The computation isn't done in the right precision.
2615        Type *Ty = Op0C->getType();
2616        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2617          if (Ty == Op1C->getType()) {
2618            double Op0V = getValueAsDouble(Op0C);
2619            double Op1V = getValueAsDouble(Op1C);
2620            return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr;
2621          }
2622        }
2623        break;
2624      }
2625
2626      case LibFunc_fmodl:
2627      case LibFunc_fmod:
2628      case LibFunc_fmodf:
2629        return Op0.isNaN() || Op1.isNaN() ||
2630               (!Op0.isInfinity() && !Op1.isZero());
2631
2632      default:
2633        break;
2634      }
2635    }
2636  }
2637
2638  return false;
2639}
2640