• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines routines for folding instructions into constants.
11 //
12 // Also, to supplement the basic IR ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // DataLayout information. These functions cannot go in IR due to library
15 // dependency issues.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/ADT/APFloat.h"
21 #include "llvm/ADT/APInt.h"
22 #include "llvm/ADT/ArrayRef.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/Analysis/ValueTracking.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/Operator.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/IR/Value.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/KnownBits.h"
46 #include "llvm/Support/MathExtras.h"
47 #include <cassert>
48 #include <cerrno>
49 #include <cfenv>
50 #include <cmath>
51 #include <cstddef>
52 #include <cstdint>
53 
54 using namespace llvm;
55 
56 namespace {
57 
58 //===----------------------------------------------------------------------===//
59 // Constant Folding internal helper functions
60 //===----------------------------------------------------------------------===//
61 
foldConstVectorToAPInt(APInt & Result,Type * DestTy,Constant * C,Type * SrcEltTy,unsigned NumSrcElts,const DataLayout & DL)62 static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
63                                         Constant *C, Type *SrcEltTy,
64                                         unsigned NumSrcElts,
65                                         const DataLayout &DL) {
66   // Now that we know that the input value is a vector of integers, just shift
67   // and insert them into our result.
68   unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
69   for (unsigned i = 0; i != NumSrcElts; ++i) {
70     Constant *Element;
71     if (DL.isLittleEndian())
72       Element = C->getAggregateElement(NumSrcElts - i - 1);
73     else
74       Element = C->getAggregateElement(i);
75 
76     if (Element && isa<UndefValue>(Element)) {
77       Result <<= BitShift;
78       continue;
79     }
80 
81     auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
82     if (!ElementCI)
83       return ConstantExpr::getBitCast(C, DestTy);
84 
85     Result <<= BitShift;
86     Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth());
87   }
88 
89   return nullptr;
90 }
91 
92 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
93 /// This always returns a non-null constant, but it may be a
94 /// ConstantExpr if unfoldable.
FoldBitCast(Constant * C,Type * DestTy,const DataLayout & DL)95 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
96   // Catch the obvious splat cases.
97   if (C->isNullValue() && !DestTy->isX86_MMXTy())
98     return Constant::getNullValue(DestTy);
99   if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
100       !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
101     return Constant::getAllOnesValue(DestTy);
102 
103   if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
104     // Handle a vector->scalar integer/fp cast.
105     if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
106       unsigned NumSrcElts = VTy->getNumElements();
107       Type *SrcEltTy = VTy->getElementType();
108 
109       // If the vector is a vector of floating point, convert it to vector of int
110       // to simplify things.
111       if (SrcEltTy->isFloatingPointTy()) {
112         unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
113         Type *SrcIVTy =
114           VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
115         // Ask IR to do the conversion now that #elts line up.
116         C = ConstantExpr::getBitCast(C, SrcIVTy);
117       }
118 
119       APInt Result(DL.getTypeSizeInBits(DestTy), 0);
120       if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
121                                                 SrcEltTy, NumSrcElts, DL))
122         return CE;
123 
124       if (isa<IntegerType>(DestTy))
125         return ConstantInt::get(DestTy, Result);
126 
127       APFloat FP(DestTy->getFltSemantics(), Result);
128       return ConstantFP::get(DestTy->getContext(), FP);
129     }
130   }
131 
132   // The code below only handles casts to vectors currently.
133   auto *DestVTy = dyn_cast<VectorType>(DestTy);
134   if (!DestVTy)
135     return ConstantExpr::getBitCast(C, DestTy);
136 
137   // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
138   // vector so the code below can handle it uniformly.
139   if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
140     Constant *Ops = C; // don't take the address of C!
141     return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
142   }
143 
144   // If this is a bitcast from constant vector -> vector, fold it.
145   if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
146     return ConstantExpr::getBitCast(C, DestTy);
147 
148   // If the element types match, IR can fold it.
149   unsigned NumDstElt = DestVTy->getNumElements();
150   unsigned NumSrcElt = C->getType()->getVectorNumElements();
151   if (NumDstElt == NumSrcElt)
152     return ConstantExpr::getBitCast(C, DestTy);
153 
154   Type *SrcEltTy = C->getType()->getVectorElementType();
155   Type *DstEltTy = DestVTy->getElementType();
156 
157   // Otherwise, we're changing the number of elements in a vector, which
158   // requires endianness information to do the right thing.  For example,
159   //    bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
160   // folds to (little endian):
161   //    <4 x i32> <i32 0, i32 0, i32 1, i32 0>
162   // and to (big endian):
163   //    <4 x i32> <i32 0, i32 0, i32 0, i32 1>
164 
165   // First thing is first.  We only want to think about integer here, so if
166   // we have something in FP form, recast it as integer.
167   if (DstEltTy->isFloatingPointTy()) {
168     // Fold to an vector of integers with same size as our FP type.
169     unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
170     Type *DestIVTy =
171       VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
172     // Recursively handle this integer conversion, if possible.
173     C = FoldBitCast(C, DestIVTy, DL);
174 
175     // Finally, IR can handle this now that #elts line up.
176     return ConstantExpr::getBitCast(C, DestTy);
177   }
178 
179   // Okay, we know the destination is integer, if the input is FP, convert
180   // it to integer first.
181   if (SrcEltTy->isFloatingPointTy()) {
182     unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
183     Type *SrcIVTy =
184       VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
185     // Ask IR to do the conversion now that #elts line up.
186     C = ConstantExpr::getBitCast(C, SrcIVTy);
187     // If IR wasn't able to fold it, bail out.
188     if (!isa<ConstantVector>(C) &&  // FIXME: Remove ConstantVector.
189         !isa<ConstantDataVector>(C))
190       return C;
191   }
192 
193   // Now we know that the input and output vectors are both integer vectors
194   // of the same size, and that their #elements is not the same.  Do the
195   // conversion here, which depends on whether the input or output has
196   // more elements.
197   bool isLittleEndian = DL.isLittleEndian();
198 
199   SmallVector<Constant*, 32> Result;
200   if (NumDstElt < NumSrcElt) {
201     // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
202     Constant *Zero = Constant::getNullValue(DstEltTy);
203     unsigned Ratio = NumSrcElt/NumDstElt;
204     unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
205     unsigned SrcElt = 0;
206     for (unsigned i = 0; i != NumDstElt; ++i) {
207       // Build each element of the result.
208       Constant *Elt = Zero;
209       unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
210       for (unsigned j = 0; j != Ratio; ++j) {
211         Constant *Src = C->getAggregateElement(SrcElt++);
212         if (Src && isa<UndefValue>(Src))
213           Src = Constant::getNullValue(C->getType()->getVectorElementType());
214         else
215           Src = dyn_cast_or_null<ConstantInt>(Src);
216         if (!Src)  // Reject constantexpr elements.
217           return ConstantExpr::getBitCast(C, DestTy);
218 
219         // Zero extend the element to the right size.
220         Src = ConstantExpr::getZExt(Src, Elt->getType());
221 
222         // Shift it to the right place, depending on endianness.
223         Src = ConstantExpr::getShl(Src,
224                                    ConstantInt::get(Src->getType(), ShiftAmt));
225         ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
226 
227         // Mix it in.
228         Elt = ConstantExpr::getOr(Elt, Src);
229       }
230       Result.push_back(Elt);
231     }
232     return ConstantVector::get(Result);
233   }
234 
235   // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
236   unsigned Ratio = NumDstElt/NumSrcElt;
237   unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
238 
239   // Loop over each source value, expanding into multiple results.
240   for (unsigned i = 0; i != NumSrcElt; ++i) {
241     auto *Element = C->getAggregateElement(i);
242 
243     if (!Element) // Reject constantexpr elements.
244       return ConstantExpr::getBitCast(C, DestTy);
245 
246     if (isa<UndefValue>(Element)) {
247       // Correctly Propagate undef values.
248       Result.append(Ratio, UndefValue::get(DstEltTy));
249       continue;
250     }
251 
252     auto *Src = dyn_cast<ConstantInt>(Element);
253     if (!Src)
254       return ConstantExpr::getBitCast(C, DestTy);
255 
256     unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
257     for (unsigned j = 0; j != Ratio; ++j) {
258       // Shift the piece of the value into the right place, depending on
259       // endianness.
260       Constant *Elt = ConstantExpr::getLShr(Src,
261                                   ConstantInt::get(Src->getType(), ShiftAmt));
262       ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
263 
264       // Truncate the element to an integer with the same pointer size and
265       // convert the element back to a pointer using a inttoptr.
266       if (DstEltTy->isPointerTy()) {
267         IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
268         Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
269         Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
270         continue;
271       }
272 
273       // Truncate and remember this piece.
274       Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
275     }
276   }
277 
278   return ConstantVector::get(Result);
279 }
280 
281 } // end anonymous namespace
282 
283 /// If this constant is a constant offset from a global, return the global and
284 /// the constant. Because of constantexprs, this function is recursive.
IsConstantOffsetFromGlobal(Constant * C,GlobalValue * & GV,APInt & Offset,const DataLayout & DL)285 bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
286                                       APInt &Offset, const DataLayout &DL) {
287   // Trivial case, constant is the global.
288   if ((GV = dyn_cast<GlobalValue>(C))) {
289     unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());
290     Offset = APInt(BitWidth, 0);
291     return true;
292   }
293 
294   // Otherwise, if this isn't a constant expr, bail out.
295   auto *CE = dyn_cast<ConstantExpr>(C);
296   if (!CE) return false;
297 
298   // Look through ptr->int and ptr->ptr casts.
299   if (CE->getOpcode() == Instruction::PtrToInt ||
300       CE->getOpcode() == Instruction::BitCast)
301     return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
302 
303   // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
304   auto *GEP = dyn_cast<GEPOperator>(CE);
305   if (!GEP)
306     return false;
307 
308   unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
309   APInt TmpOffset(BitWidth, 0);
310 
311   // If the base isn't a global+constant, we aren't either.
312   if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
313     return false;
314 
315   // Otherwise, add any offset that our operands provide.
316   if (!GEP->accumulateConstantOffset(DL, TmpOffset))
317     return false;
318 
319   Offset = TmpOffset;
320   return true;
321 }
322 
ConstantFoldLoadThroughBitcast(Constant * C,Type * DestTy,const DataLayout & DL)323 Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
324                                          const DataLayout &DL) {
325   do {
326     Type *SrcTy = C->getType();
327 
328     // If the type sizes are the same and a cast is legal, just directly
329     // cast the constant.
330     if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
331       Instruction::CastOps Cast = Instruction::BitCast;
332       // If we are going from a pointer to int or vice versa, we spell the cast
333       // differently.
334       if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
335         Cast = Instruction::IntToPtr;
336       else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
337         Cast = Instruction::PtrToInt;
338 
339       if (CastInst::castIsValid(Cast, C, DestTy))
340         return ConstantExpr::getCast(Cast, C, DestTy);
341     }
342 
343     // If this isn't an aggregate type, there is nothing we can do to drill down
344     // and find a bitcastable constant.
345     if (!SrcTy->isAggregateType())
346       return nullptr;
347 
348     // We're simulating a load through a pointer that was bitcast to point to
349     // a different type, so we can try to walk down through the initial
350     // elements of an aggregate to see if some part of th e aggregate is
351     // castable to implement the "load" semantic model.
352     C = C->getAggregateElement(0u);
353   } while (C);
354 
355   return nullptr;
356 }
357 
358 namespace {
359 
360 /// Recursive helper to read bits out of global. C is the constant being copied
361 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
362 /// results into and BytesLeft is the number of bytes left in
363 /// the CurPtr buffer. DL is the DataLayout.
ReadDataFromGlobal(Constant * C,uint64_t ByteOffset,unsigned char * CurPtr,unsigned BytesLeft,const DataLayout & DL)364 bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
365                         unsigned BytesLeft, const DataLayout &DL) {
366   assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
367          "Out of range access");
368 
369   // If this element is zero or undefined, we can just return since *CurPtr is
370   // zero initialized.
371   if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
372     return true;
373 
374   if (auto *CI = dyn_cast<ConstantInt>(C)) {
375     if (CI->getBitWidth() > 64 ||
376         (CI->getBitWidth() & 7) != 0)
377       return false;
378 
379     uint64_t Val = CI->getZExtValue();
380     unsigned IntBytes = unsigned(CI->getBitWidth()/8);
381 
382     for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
383       int n = ByteOffset;
384       if (!DL.isLittleEndian())
385         n = IntBytes - n - 1;
386       CurPtr[i] = (unsigned char)(Val >> (n * 8));
387       ++ByteOffset;
388     }
389     return true;
390   }
391 
392   if (auto *CFP = dyn_cast<ConstantFP>(C)) {
393     if (CFP->getType()->isDoubleTy()) {
394       C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
395       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
396     }
397     if (CFP->getType()->isFloatTy()){
398       C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
399       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
400     }
401     if (CFP->getType()->isHalfTy()){
402       C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
403       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
404     }
405     return false;
406   }
407 
408   if (auto *CS = dyn_cast<ConstantStruct>(C)) {
409     const StructLayout *SL = DL.getStructLayout(CS->getType());
410     unsigned Index = SL->getElementContainingOffset(ByteOffset);
411     uint64_t CurEltOffset = SL->getElementOffset(Index);
412     ByteOffset -= CurEltOffset;
413 
414     while (true) {
415       // If the element access is to the element itself and not to tail padding,
416       // read the bytes from the element.
417       uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
418 
419       if (ByteOffset < EltSize &&
420           !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
421                               BytesLeft, DL))
422         return false;
423 
424       ++Index;
425 
426       // Check to see if we read from the last struct element, if so we're done.
427       if (Index == CS->getType()->getNumElements())
428         return true;
429 
430       // If we read all of the bytes we needed from this element we're done.
431       uint64_t NextEltOffset = SL->getElementOffset(Index);
432 
433       if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
434         return true;
435 
436       // Move to the next element of the struct.
437       CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
438       BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
439       ByteOffset = 0;
440       CurEltOffset = NextEltOffset;
441     }
442     // not reached.
443   }
444 
445   if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
446       isa<ConstantDataSequential>(C)) {
447     Type *EltTy = C->getType()->getSequentialElementType();
448     uint64_t EltSize = DL.getTypeAllocSize(EltTy);
449     uint64_t Index = ByteOffset / EltSize;
450     uint64_t Offset = ByteOffset - Index * EltSize;
451     uint64_t NumElts;
452     if (auto *AT = dyn_cast<ArrayType>(C->getType()))
453       NumElts = AT->getNumElements();
454     else
455       NumElts = C->getType()->getVectorNumElements();
456 
457     for (; Index != NumElts; ++Index) {
458       if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
459                               BytesLeft, DL))
460         return false;
461 
462       uint64_t BytesWritten = EltSize - Offset;
463       assert(BytesWritten <= EltSize && "Not indexing into this element?");
464       if (BytesWritten >= BytesLeft)
465         return true;
466 
467       Offset = 0;
468       BytesLeft -= BytesWritten;
469       CurPtr += BytesWritten;
470     }
471     return true;
472   }
473 
474   if (auto *CE = dyn_cast<ConstantExpr>(C)) {
475     if (CE->getOpcode() == Instruction::IntToPtr &&
476         CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
477       return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
478                                 BytesLeft, DL);
479     }
480   }
481 
482   // Otherwise, unknown initializer type.
483   return false;
484 }
485 
FoldReinterpretLoadFromConstPtr(Constant * C,Type * LoadTy,const DataLayout & DL)486 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
487                                           const DataLayout &DL) {
488   auto *PTy = cast<PointerType>(C->getType());
489   auto *IntType = dyn_cast<IntegerType>(LoadTy);
490 
491   // If this isn't an integer load we can't fold it directly.
492   if (!IntType) {
493     unsigned AS = PTy->getAddressSpace();
494 
495     // If this is a float/double load, we can try folding it as an int32/64 load
496     // and then bitcast the result.  This can be useful for union cases.  Note
497     // that address spaces don't matter here since we're not going to result in
498     // an actual new load.
499     Type *MapTy;
500     if (LoadTy->isHalfTy())
501       MapTy = Type::getInt16Ty(C->getContext());
502     else if (LoadTy->isFloatTy())
503       MapTy = Type::getInt32Ty(C->getContext());
504     else if (LoadTy->isDoubleTy())
505       MapTy = Type::getInt64Ty(C->getContext());
506     else if (LoadTy->isVectorTy()) {
507       MapTy = PointerType::getIntNTy(C->getContext(),
508                                      DL.getTypeAllocSizeInBits(LoadTy));
509     } else
510       return nullptr;
511 
512     C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
513     if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL))
514       return FoldBitCast(Res, LoadTy, DL);
515     return nullptr;
516   }
517 
518   unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
519   if (BytesLoaded > 32 || BytesLoaded == 0)
520     return nullptr;
521 
522   GlobalValue *GVal;
523   APInt OffsetAI;
524   if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL))
525     return nullptr;
526 
527   auto *GV = dyn_cast<GlobalVariable>(GVal);
528   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
529       !GV->getInitializer()->getType()->isSized())
530     return nullptr;
531 
532   int64_t Offset = OffsetAI.getSExtValue();
533   int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
534 
535   // If we're not accessing anything in this constant, the result is undefined.
536   if (Offset + BytesLoaded <= 0)
537     return UndefValue::get(IntType);
538 
539   // If we're not accessing anything in this constant, the result is undefined.
540   if (Offset >= InitializerSize)
541     return UndefValue::get(IntType);
542 
543   unsigned char RawBytes[32] = {0};
544   unsigned char *CurPtr = RawBytes;
545   unsigned BytesLeft = BytesLoaded;
546 
547   // If we're loading off the beginning of the global, some bytes may be valid.
548   if (Offset < 0) {
549     CurPtr += -Offset;
550     BytesLeft += Offset;
551     Offset = 0;
552   }
553 
554   if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL))
555     return nullptr;
556 
557   APInt ResultVal = APInt(IntType->getBitWidth(), 0);
558   if (DL.isLittleEndian()) {
559     ResultVal = RawBytes[BytesLoaded - 1];
560     for (unsigned i = 1; i != BytesLoaded; ++i) {
561       ResultVal <<= 8;
562       ResultVal |= RawBytes[BytesLoaded - 1 - i];
563     }
564   } else {
565     ResultVal = RawBytes[0];
566     for (unsigned i = 1; i != BytesLoaded; ++i) {
567       ResultVal <<= 8;
568       ResultVal |= RawBytes[i];
569     }
570   }
571 
572   return ConstantInt::get(IntType->getContext(), ResultVal);
573 }
574 
ConstantFoldLoadThroughBitcastExpr(ConstantExpr * CE,Type * DestTy,const DataLayout & DL)575 Constant *ConstantFoldLoadThroughBitcastExpr(ConstantExpr *CE, Type *DestTy,
576                                              const DataLayout &DL) {
577   auto *SrcPtr = CE->getOperand(0);
578   auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType());
579   if (!SrcPtrTy)
580     return nullptr;
581   Type *SrcTy = SrcPtrTy->getPointerElementType();
582 
583   Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL);
584   if (!C)
585     return nullptr;
586 
587   return llvm::ConstantFoldLoadThroughBitcast(C, DestTy, DL);
588 }
589 
590 } // end anonymous namespace
591 
ConstantFoldLoadFromConstPtr(Constant * C,Type * Ty,const DataLayout & DL)592 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
593                                              const DataLayout &DL) {
594   // First, try the easy cases:
595   if (auto *GV = dyn_cast<GlobalVariable>(C))
596     if (GV->isConstant() && GV->hasDefinitiveInitializer())
597       return GV->getInitializer();
598 
599   if (auto *GA = dyn_cast<GlobalAlias>(C))
600     if (GA->getAliasee() && !GA->isInterposable())
601       return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL);
602 
603   // If the loaded value isn't a constant expr, we can't handle it.
604   auto *CE = dyn_cast<ConstantExpr>(C);
605   if (!CE)
606     return nullptr;
607 
608   if (CE->getOpcode() == Instruction::GetElementPtr) {
609     if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
610       if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
611         if (Constant *V =
612              ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
613           return V;
614       }
615     }
616   }
617 
618   if (CE->getOpcode() == Instruction::BitCast)
619     if (Constant *LoadedC = ConstantFoldLoadThroughBitcastExpr(CE, Ty, DL))
620       return LoadedC;
621 
622   // Instead of loading constant c string, use corresponding integer value
623   // directly if string length is small enough.
624   StringRef Str;
625   if (getConstantStringInfo(CE, Str) && !Str.empty()) {
626     size_t StrLen = Str.size();
627     unsigned NumBits = Ty->getPrimitiveSizeInBits();
628     // Replace load with immediate integer if the result is an integer or fp
629     // value.
630     if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
631         (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
632       APInt StrVal(NumBits, 0);
633       APInt SingleChar(NumBits, 0);
634       if (DL.isLittleEndian()) {
635         for (unsigned char C : reverse(Str.bytes())) {
636           SingleChar = static_cast<uint64_t>(C);
637           StrVal = (StrVal << 8) | SingleChar;
638         }
639       } else {
640         for (unsigned char C : Str.bytes()) {
641           SingleChar = static_cast<uint64_t>(C);
642           StrVal = (StrVal << 8) | SingleChar;
643         }
644         // Append NULL at the end.
645         SingleChar = 0;
646         StrVal = (StrVal << 8) | SingleChar;
647       }
648 
649       Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
650       if (Ty->isFloatingPointTy())
651         Res = ConstantExpr::getBitCast(Res, Ty);
652       return Res;
653     }
654   }
655 
656   // If this load comes from anywhere in a constant global, and if the global
657   // is all undef or zero, we know what it loads.
658   if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
659     if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
660       if (GV->getInitializer()->isNullValue())
661         return Constant::getNullValue(Ty);
662       if (isa<UndefValue>(GV->getInitializer()))
663         return UndefValue::get(Ty);
664     }
665   }
666 
667   // Try hard to fold loads from bitcasted strange and non-type-safe things.
668   return FoldReinterpretLoadFromConstPtr(CE, Ty, DL);
669 }
670 
671 namespace {
672 
ConstantFoldLoadInst(const LoadInst * LI,const DataLayout & DL)673 Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) {
674   if (LI->isVolatile()) return nullptr;
675 
676   if (auto *C = dyn_cast<Constant>(LI->getOperand(0)))
677     return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL);
678 
679   return nullptr;
680 }
681 
682 /// One of Op0/Op1 is a constant expression.
683 /// Attempt to symbolically evaluate the result of a binary operator merging
684 /// these together.  If target data info is available, it is provided as DL,
685 /// otherwise DL is null.
SymbolicallyEvaluateBinop(unsigned Opc,Constant * Op0,Constant * Op1,const DataLayout & DL)686 Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
687                                     const DataLayout &DL) {
688   // SROA
689 
690   // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
691   // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
692   // bits.
693 
694   if (Opc == Instruction::And) {
695     KnownBits Known0 = computeKnownBits(Op0, DL);
696     KnownBits Known1 = computeKnownBits(Op1, DL);
697     if ((Known1.One | Known0.Zero).isAllOnesValue()) {
698       // All the bits of Op0 that the 'and' could be masking are already zero.
699       return Op0;
700     }
701     if ((Known0.One | Known1.Zero).isAllOnesValue()) {
702       // All the bits of Op1 that the 'and' could be masking are already zero.
703       return Op1;
704     }
705 
706     Known0.Zero |= Known1.Zero;
707     Known0.One &= Known1.One;
708     if (Known0.isConstant())
709       return ConstantInt::get(Op0->getType(), Known0.getConstant());
710   }
711 
712   // If the constant expr is something like &A[123] - &A[4].f, fold this into a
713   // constant.  This happens frequently when iterating over a global array.
714   if (Opc == Instruction::Sub) {
715     GlobalValue *GV1, *GV2;
716     APInt Offs1, Offs2;
717 
718     if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
719       if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
720         unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
721 
722         // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
723         // PtrToInt may change the bitwidth so we have convert to the right size
724         // first.
725         return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
726                                                 Offs2.zextOrTrunc(OpSize));
727       }
728   }
729 
730   return nullptr;
731 }
732 
733 /// If array indices are not pointer-sized integers, explicitly cast them so
734 /// that they aren't implicitly casted by the getelementptr.
CastGEPIndices(Type * SrcElemTy,ArrayRef<Constant * > Ops,Type * ResultTy,Optional<unsigned> InRangeIndex,const DataLayout & DL,const TargetLibraryInfo * TLI)735 Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
736                          Type *ResultTy, Optional<unsigned> InRangeIndex,
737                          const DataLayout &DL, const TargetLibraryInfo *TLI) {
738   Type *IntPtrTy = DL.getIntPtrType(ResultTy);
739   Type *IntPtrScalarTy = IntPtrTy->getScalarType();
740 
741   bool Any = false;
742   SmallVector<Constant*, 32> NewIdxs;
743   for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
744     if ((i == 1 ||
745          !isa<StructType>(GetElementPtrInst::getIndexedType(
746              SrcElemTy, Ops.slice(1, i - 1)))) &&
747         Ops[i]->getType()->getScalarType() != IntPtrScalarTy) {
748       Any = true;
749       Type *NewType = Ops[i]->getType()->isVectorTy()
750                           ? IntPtrTy
751                           : IntPtrTy->getScalarType();
752       NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
753                                                                       true,
754                                                                       NewType,
755                                                                       true),
756                                               Ops[i], NewType));
757     } else
758       NewIdxs.push_back(Ops[i]);
759   }
760 
761   if (!Any)
762     return nullptr;
763 
764   Constant *C = ConstantExpr::getGetElementPtr(
765       SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex);
766   if (Constant *Folded = ConstantFoldConstant(C, DL, TLI))
767     C = Folded;
768 
769   return C;
770 }
771 
772 /// Strip the pointer casts, but preserve the address space information.
StripPtrCastKeepAS(Constant * Ptr,Type * & ElemTy)773 Constant* StripPtrCastKeepAS(Constant* Ptr, Type *&ElemTy) {
774   assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
775   auto *OldPtrTy = cast<PointerType>(Ptr->getType());
776   Ptr = Ptr->stripPointerCasts();
777   auto *NewPtrTy = cast<PointerType>(Ptr->getType());
778 
779   ElemTy = NewPtrTy->getPointerElementType();
780 
781   // Preserve the address space number of the pointer.
782   if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
783     NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace());
784     Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
785   }
786   return Ptr;
787 }
788 
789 /// If we can symbolically evaluate the GEP constant expression, do so.
SymbolicallyEvaluateGEP(const GEPOperator * GEP,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)790 Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
791                                   ArrayRef<Constant *> Ops,
792                                   const DataLayout &DL,
793                                   const TargetLibraryInfo *TLI) {
794   const GEPOperator *InnermostGEP = GEP;
795   bool InBounds = GEP->isInBounds();
796 
797   Type *SrcElemTy = GEP->getSourceElementType();
798   Type *ResElemTy = GEP->getResultElementType();
799   Type *ResTy = GEP->getType();
800   if (!SrcElemTy->isSized())
801     return nullptr;
802 
803   if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy,
804                                    GEP->getInRangeIndex(), DL, TLI))
805     return C;
806 
807   Constant *Ptr = Ops[0];
808   if (!Ptr->getType()->isPointerTy())
809     return nullptr;
810 
811   Type *IntPtrTy = DL.getIntPtrType(Ptr->getType());
812 
813   // If this is a constant expr gep that is effectively computing an
814   // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
815   for (unsigned i = 1, e = Ops.size(); i != e; ++i)
816       if (!isa<ConstantInt>(Ops[i])) {
817 
818         // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
819         // "inttoptr (sub (ptrtoint Ptr), V)"
820         if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
821           auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
822           assert((!CE || CE->getType() == IntPtrTy) &&
823                  "CastGEPIndices didn't canonicalize index types!");
824           if (CE && CE->getOpcode() == Instruction::Sub &&
825               CE->getOperand(0)->isNullValue()) {
826             Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
827             Res = ConstantExpr::getSub(Res, CE->getOperand(1));
828             Res = ConstantExpr::getIntToPtr(Res, ResTy);
829             if (auto *FoldedRes = ConstantFoldConstant(Res, DL, TLI))
830               Res = FoldedRes;
831             return Res;
832           }
833         }
834         return nullptr;
835       }
836 
837   unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy);
838   APInt Offset =
839       APInt(BitWidth,
840             DL.getIndexedOffsetInType(
841                 SrcElemTy,
842                 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
843   Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
844 
845   // If this is a GEP of a GEP, fold it all into a single GEP.
846   while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
847     InnermostGEP = GEP;
848     InBounds &= GEP->isInBounds();
849 
850     SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
851 
852     // Do not try the incorporate the sub-GEP if some index is not a number.
853     bool AllConstantInt = true;
854     for (Value *NestedOp : NestedOps)
855       if (!isa<ConstantInt>(NestedOp)) {
856         AllConstantInt = false;
857         break;
858       }
859     if (!AllConstantInt)
860       break;
861 
862     Ptr = cast<Constant>(GEP->getOperand(0));
863     SrcElemTy = GEP->getSourceElementType();
864     Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
865     Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
866   }
867 
868   // If the base value for this address is a literal integer value, fold the
869   // getelementptr to the resulting integer value casted to the pointer type.
870   APInt BasePtr(BitWidth, 0);
871   if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
872     if (CE->getOpcode() == Instruction::IntToPtr) {
873       if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
874         BasePtr = Base->getValue().zextOrTrunc(BitWidth);
875     }
876   }
877 
878   auto *PTy = cast<PointerType>(Ptr->getType());
879   if ((Ptr->isNullValue() || BasePtr != 0) &&
880       !DL.isNonIntegralPointerType(PTy)) {
881     Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
882     return ConstantExpr::getIntToPtr(C, ResTy);
883   }
884 
885   // Otherwise form a regular getelementptr. Recompute the indices so that
886   // we eliminate over-indexing of the notional static type array bounds.
887   // This makes it easy to determine if the getelementptr is "inbounds".
888   // Also, this helps GlobalOpt do SROA on GlobalVariables.
889   Type *Ty = PTy;
890   SmallVector<Constant *, 32> NewIdxs;
891 
892   do {
893     if (!Ty->isStructTy()) {
894       if (Ty->isPointerTy()) {
895         // The only pointer indexing we'll do is on the first index of the GEP.
896         if (!NewIdxs.empty())
897           break;
898 
899         Ty = SrcElemTy;
900 
901         // Only handle pointers to sized types, not pointers to functions.
902         if (!Ty->isSized())
903           return nullptr;
904       } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) {
905         Ty = ATy->getElementType();
906       } else {
907         // We've reached some non-indexable type.
908         break;
909       }
910 
911       // Determine which element of the array the offset points into.
912       APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty));
913       if (ElemSize == 0) {
914         // The element size is 0. This may be [0 x Ty]*, so just use a zero
915         // index for this level and proceed to the next level to see if it can
916         // accommodate the offset.
917         NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
918       } else {
919         // The element size is non-zero divide the offset by the element
920         // size (rounding down), to compute the index at this level.
921         bool Overflow;
922         APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow);
923         if (Overflow)
924           break;
925         Offset -= NewIdx * ElemSize;
926         NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
927       }
928     } else {
929       auto *STy = cast<StructType>(Ty);
930       // If we end up with an offset that isn't valid for this struct type, we
931       // can't re-form this GEP in a regular form, so bail out. The pointer
932       // operand likely went through casts that are necessary to make the GEP
933       // sensible.
934       const StructLayout &SL = *DL.getStructLayout(STy);
935       if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes()))
936         break;
937 
938       // Determine which field of the struct the offset points into. The
939       // getZExtValue is fine as we've already ensured that the offset is
940       // within the range representable by the StructLayout API.
941       unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
942       NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
943                                          ElIdx));
944       Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
945       Ty = STy->getTypeAtIndex(ElIdx);
946     }
947   } while (Ty != ResElemTy);
948 
949   // If we haven't used up the entire offset by descending the static
950   // type, then the offset is pointing into the middle of an indivisible
951   // member, so we can't simplify it.
952   if (Offset != 0)
953     return nullptr;
954 
955   // Preserve the inrange index from the innermost GEP if possible. We must
956   // have calculated the same indices up to and including the inrange index.
957   Optional<unsigned> InRangeIndex;
958   if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex())
959     if (SrcElemTy == InnermostGEP->getSourceElementType() &&
960         NewIdxs.size() > *LastIRIndex) {
961       InRangeIndex = LastIRIndex;
962       for (unsigned I = 0; I <= *LastIRIndex; ++I)
963         if (NewIdxs[I] != InnermostGEP->getOperand(I + 1)) {
964           InRangeIndex = None;
965           break;
966         }
967     }
968 
969   // Create a GEP.
970   Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs,
971                                                InBounds, InRangeIndex);
972   assert(C->getType()->getPointerElementType() == Ty &&
973          "Computed GetElementPtr has unexpected type!");
974 
975   // If we ended up indexing a member with a type that doesn't match
976   // the type of what the original indices indexed, add a cast.
977   if (Ty != ResElemTy)
978     C = FoldBitCast(C, ResTy, DL);
979 
980   return C;
981 }
982 
983 /// Attempt to constant fold an instruction with the
984 /// specified opcode and operands.  If successful, the constant result is
985 /// returned, if not, null is returned.  Note that this function can fail when
986 /// attempting to fold instructions like loads and stores, which have no
987 /// constant expression form.
988 ///
989 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/inrange
990 /// etc information, due to only being passed an opcode and operands. Constant
991 /// folding using this function strips this information.
992 ///
ConstantFoldInstOperandsImpl(const Value * InstOrCE,unsigned Opcode,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)993 Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
994                                        ArrayRef<Constant *> Ops,
995                                        const DataLayout &DL,
996                                        const TargetLibraryInfo *TLI) {
997   Type *DestTy = InstOrCE->getType();
998 
999   // Handle easy binops first.
1000   if (Instruction::isBinaryOp(Opcode))
1001     return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
1002 
1003   if (Instruction::isCast(Opcode))
1004     return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
1005 
1006   if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
1007     if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1008       return C;
1009 
1010     return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0],
1011                                           Ops.slice(1), GEP->isInBounds(),
1012                                           GEP->getInRangeIndex());
1013   }
1014 
1015   if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))
1016     return CE->getWithOperands(Ops);
1017 
1018   switch (Opcode) {
1019   default: return nullptr;
1020   case Instruction::ICmp:
1021   case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1022   case Instruction::Call:
1023     if (auto *F = dyn_cast<Function>(Ops.back())) {
1024       ImmutableCallSite CS(cast<CallInst>(InstOrCE));
1025       if (canConstantFoldCallTo(CS, F))
1026         return ConstantFoldCall(CS, F, Ops.slice(0, Ops.size() - 1), TLI);
1027     }
1028     return nullptr;
1029   case Instruction::Select:
1030     return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1031   case Instruction::ExtractElement:
1032     return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1033   case Instruction::InsertElement:
1034     return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1035   case Instruction::ShuffleVector:
1036     return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1037   }
1038 }
1039 
1040 } // end anonymous namespace
1041 
1042 //===----------------------------------------------------------------------===//
1043 // Constant Folding public APIs
1044 //===----------------------------------------------------------------------===//
1045 
1046 namespace {
1047 
1048 Constant *
ConstantFoldConstantImpl(const Constant * C,const DataLayout & DL,const TargetLibraryInfo * TLI,SmallDenseMap<Constant *,Constant * > & FoldedOps)1049 ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1050                          const TargetLibraryInfo *TLI,
1051                          SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1052   if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))
1053     return nullptr;
1054 
1055   SmallVector<Constant *, 8> Ops;
1056   for (const Use &NewU : C->operands()) {
1057     auto *NewC = cast<Constant>(&NewU);
1058     // Recursively fold the ConstantExpr's operands. If we have already folded
1059     // a ConstantExpr, we don't have to process it again.
1060     if (isa<ConstantVector>(NewC) || isa<ConstantExpr>(NewC)) {
1061       auto It = FoldedOps.find(NewC);
1062       if (It == FoldedOps.end()) {
1063         if (auto *FoldedC =
1064                 ConstantFoldConstantImpl(NewC, DL, TLI, FoldedOps)) {
1065           FoldedOps.insert({NewC, FoldedC});
1066           NewC = FoldedC;
1067         } else {
1068           FoldedOps.insert({NewC, NewC});
1069         }
1070       } else {
1071         NewC = It->second;
1072       }
1073     }
1074     Ops.push_back(NewC);
1075   }
1076 
1077   if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1078     if (CE->isCompare())
1079       return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1080                                              DL, TLI);
1081 
1082     return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI);
1083   }
1084 
1085   assert(isa<ConstantVector>(C));
1086   return ConstantVector::get(Ops);
1087 }
1088 
1089 } // end anonymous namespace
1090 
ConstantFoldInstruction(Instruction * I,const DataLayout & DL,const TargetLibraryInfo * TLI)1091 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
1092                                         const TargetLibraryInfo *TLI) {
1093   // Handle PHI nodes quickly here...
1094   if (auto *PN = dyn_cast<PHINode>(I)) {
1095     Constant *CommonValue = nullptr;
1096 
1097     SmallDenseMap<Constant *, Constant *> FoldedOps;
1098     for (Value *Incoming : PN->incoming_values()) {
1099       // If the incoming value is undef then skip it.  Note that while we could
1100       // skip the value if it is equal to the phi node itself we choose not to
1101       // because that would break the rule that constant folding only applies if
1102       // all operands are constants.
1103       if (isa<UndefValue>(Incoming))
1104         continue;
1105       // If the incoming value is not a constant, then give up.
1106       auto *C = dyn_cast<Constant>(Incoming);
1107       if (!C)
1108         return nullptr;
1109       // Fold the PHI's operands.
1110       if (auto *FoldedC = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps))
1111         C = FoldedC;
1112       // If the incoming value is a different constant to
1113       // the one we saw previously, then give up.
1114       if (CommonValue && C != CommonValue)
1115         return nullptr;
1116       CommonValue = C;
1117     }
1118 
1119     // If we reach here, all incoming values are the same constant or undef.
1120     return CommonValue ? CommonValue : UndefValue::get(PN->getType());
1121   }
1122 
1123   // Scan the operand list, checking to see if they are all constants, if so,
1124   // hand off to ConstantFoldInstOperandsImpl.
1125   if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1126     return nullptr;
1127 
1128   SmallDenseMap<Constant *, Constant *> FoldedOps;
1129   SmallVector<Constant *, 8> Ops;
1130   for (const Use &OpU : I->operands()) {
1131     auto *Op = cast<Constant>(&OpU);
1132     // Fold the Instruction's operands.
1133     if (auto *FoldedOp = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps))
1134       Op = FoldedOp;
1135 
1136     Ops.push_back(Op);
1137   }
1138 
1139   if (const auto *CI = dyn_cast<CmpInst>(I))
1140     return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1141                                            DL, TLI);
1142 
1143   if (const auto *LI = dyn_cast<LoadInst>(I))
1144     return ConstantFoldLoadInst(LI, DL);
1145 
1146   if (auto *IVI = dyn_cast<InsertValueInst>(I)) {
1147     return ConstantExpr::getInsertValue(
1148                                 cast<Constant>(IVI->getAggregateOperand()),
1149                                 cast<Constant>(IVI->getInsertedValueOperand()),
1150                                 IVI->getIndices());
1151   }
1152 
1153   if (auto *EVI = dyn_cast<ExtractValueInst>(I)) {
1154     return ConstantExpr::getExtractValue(
1155                                     cast<Constant>(EVI->getAggregateOperand()),
1156                                     EVI->getIndices());
1157   }
1158 
1159   return ConstantFoldInstOperands(I, Ops, DL, TLI);
1160 }
1161 
ConstantFoldConstant(const Constant * C,const DataLayout & DL,const TargetLibraryInfo * TLI)1162 Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1163                                      const TargetLibraryInfo *TLI) {
1164   SmallDenseMap<Constant *, Constant *> FoldedOps;
1165   return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1166 }
1167 
ConstantFoldInstOperands(Instruction * I,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)1168 Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1169                                          ArrayRef<Constant *> Ops,
1170                                          const DataLayout &DL,
1171                                          const TargetLibraryInfo *TLI) {
1172   return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI);
1173 }
1174 
ConstantFoldCompareInstOperands(unsigned Predicate,Constant * Ops0,Constant * Ops1,const DataLayout & DL,const TargetLibraryInfo * TLI)1175 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1176                                                 Constant *Ops0, Constant *Ops1,
1177                                                 const DataLayout &DL,
1178                                                 const TargetLibraryInfo *TLI) {
1179   // fold: icmp (inttoptr x), null         -> icmp x, 0
1180   // fold: icmp null, (inttoptr x)         -> icmp 0, x
1181   // fold: icmp (ptrtoint x), 0            -> icmp x, null
1182   // fold: icmp 0, (ptrtoint x)            -> icmp null, x
1183   // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1184   // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1185   //
1186   // FIXME: The following comment is out of data and the DataLayout is here now.
1187   // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1188   // around to know if bit truncation is happening.
1189   if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1190     if (Ops1->isNullValue()) {
1191       if (CE0->getOpcode() == Instruction::IntToPtr) {
1192         Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1193         // Convert the integer value to the right size to ensure we get the
1194         // proper extension or truncation.
1195         Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1196                                                    IntPtrTy, false);
1197         Constant *Null = Constant::getNullValue(C->getType());
1198         return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1199       }
1200 
1201       // Only do this transformation if the int is intptrty in size, otherwise
1202       // there is a truncation or extension that we aren't modeling.
1203       if (CE0->getOpcode() == Instruction::PtrToInt) {
1204         Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1205         if (CE0->getType() == IntPtrTy) {
1206           Constant *C = CE0->getOperand(0);
1207           Constant *Null = Constant::getNullValue(C->getType());
1208           return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1209         }
1210       }
1211     }
1212 
1213     if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1214       if (CE0->getOpcode() == CE1->getOpcode()) {
1215         if (CE0->getOpcode() == Instruction::IntToPtr) {
1216           Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1217 
1218           // Convert the integer value to the right size to ensure we get the
1219           // proper extension or truncation.
1220           Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1221                                                       IntPtrTy, false);
1222           Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1223                                                       IntPtrTy, false);
1224           return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1225         }
1226 
1227         // Only do this transformation if the int is intptrty in size, otherwise
1228         // there is a truncation or extension that we aren't modeling.
1229         if (CE0->getOpcode() == Instruction::PtrToInt) {
1230           Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1231           if (CE0->getType() == IntPtrTy &&
1232               CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1233             return ConstantFoldCompareInstOperands(
1234                 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1235           }
1236         }
1237       }
1238     }
1239 
1240     // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1241     // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1242     if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1243         CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1244       Constant *LHS = ConstantFoldCompareInstOperands(
1245           Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1246       Constant *RHS = ConstantFoldCompareInstOperands(
1247           Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1248       unsigned OpC =
1249         Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1250       return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1251     }
1252   } else if (isa<ConstantExpr>(Ops1)) {
1253     // If RHS is a constant expression, but the left side isn't, swap the
1254     // operands and try again.
1255     Predicate = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)Predicate);
1256     return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI);
1257   }
1258 
1259   return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1260 }
1261 
ConstantFoldBinaryOpOperands(unsigned Opcode,Constant * LHS,Constant * RHS,const DataLayout & DL)1262 Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1263                                              Constant *RHS,
1264                                              const DataLayout &DL) {
1265   assert(Instruction::isBinaryOp(Opcode));
1266   if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1267     if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1268       return C;
1269 
1270   return ConstantExpr::get(Opcode, LHS, RHS);
1271 }
1272 
ConstantFoldCastOperand(unsigned Opcode,Constant * C,Type * DestTy,const DataLayout & DL)1273 Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1274                                         Type *DestTy, const DataLayout &DL) {
1275   assert(Instruction::isCast(Opcode));
1276   switch (Opcode) {
1277   default:
1278     llvm_unreachable("Missing case");
1279   case Instruction::PtrToInt:
1280     // If the input is a inttoptr, eliminate the pair.  This requires knowing
1281     // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1282     if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1283       if (CE->getOpcode() == Instruction::IntToPtr) {
1284         Constant *Input = CE->getOperand(0);
1285         unsigned InWidth = Input->getType()->getScalarSizeInBits();
1286         unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1287         if (PtrWidth < InWidth) {
1288           Constant *Mask =
1289             ConstantInt::get(CE->getContext(),
1290                              APInt::getLowBitsSet(InWidth, PtrWidth));
1291           Input = ConstantExpr::getAnd(Input, Mask);
1292         }
1293         // Do a zext or trunc to get to the dest size.
1294         return ConstantExpr::getIntegerCast(Input, DestTy, false);
1295       }
1296     }
1297     return ConstantExpr::getCast(Opcode, C, DestTy);
1298   case Instruction::IntToPtr:
1299     // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1300     // the int size is >= the ptr size and the address spaces are the same.
1301     // This requires knowing the width of a pointer, so it can't be done in
1302     // ConstantExpr::getCast.
1303     if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1304       if (CE->getOpcode() == Instruction::PtrToInt) {
1305         Constant *SrcPtr = CE->getOperand(0);
1306         unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1307         unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1308 
1309         if (MidIntSize >= SrcPtrSize) {
1310           unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1311           if (SrcAS == DestTy->getPointerAddressSpace())
1312             return FoldBitCast(CE->getOperand(0), DestTy, DL);
1313         }
1314       }
1315     }
1316 
1317     return ConstantExpr::getCast(Opcode, C, DestTy);
1318   case Instruction::Trunc:
1319   case Instruction::ZExt:
1320   case Instruction::SExt:
1321   case Instruction::FPTrunc:
1322   case Instruction::FPExt:
1323   case Instruction::UIToFP:
1324   case Instruction::SIToFP:
1325   case Instruction::FPToUI:
1326   case Instruction::FPToSI:
1327   case Instruction::AddrSpaceCast:
1328       return ConstantExpr::getCast(Opcode, C, DestTy);
1329   case Instruction::BitCast:
1330     return FoldBitCast(C, DestTy, DL);
1331   }
1332 }
1333 
ConstantFoldLoadThroughGEPConstantExpr(Constant * C,ConstantExpr * CE)1334 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1335                                                        ConstantExpr *CE) {
1336   if (!CE->getOperand(1)->isNullValue())
1337     return nullptr;  // Do not allow stepping over the value!
1338 
1339   // Loop over all of the operands, tracking down which value we are
1340   // addressing.
1341   for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1342     C = C->getAggregateElement(CE->getOperand(i));
1343     if (!C)
1344       return nullptr;
1345   }
1346   return C;
1347 }
1348 
1349 Constant *
ConstantFoldLoadThroughGEPIndices(Constant * C,ArrayRef<Constant * > Indices)1350 llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1351                                         ArrayRef<Constant *> Indices) {
1352   // Loop over all of the operands, tracking down which value we are
1353   // addressing.
1354   for (Constant *Index : Indices) {
1355     C = C->getAggregateElement(Index);
1356     if (!C)
1357       return nullptr;
1358   }
1359   return C;
1360 }
1361 
1362 //===----------------------------------------------------------------------===//
1363 //  Constant Folding for Calls
1364 //
1365 
canConstantFoldCallTo(ImmutableCallSite CS,const Function * F)1366 bool llvm::canConstantFoldCallTo(ImmutableCallSite CS, const Function *F) {
1367   if (CS.isNoBuiltin() || CS.isStrictFP())
1368     return false;
1369   switch (F->getIntrinsicID()) {
1370   case Intrinsic::fabs:
1371   case Intrinsic::minnum:
1372   case Intrinsic::maxnum:
1373   case Intrinsic::log:
1374   case Intrinsic::log2:
1375   case Intrinsic::log10:
1376   case Intrinsic::exp:
1377   case Intrinsic::exp2:
1378   case Intrinsic::floor:
1379   case Intrinsic::ceil:
1380   case Intrinsic::sqrt:
1381   case Intrinsic::sin:
1382   case Intrinsic::cos:
1383   case Intrinsic::trunc:
1384   case Intrinsic::rint:
1385   case Intrinsic::nearbyint:
1386   case Intrinsic::pow:
1387   case Intrinsic::powi:
1388   case Intrinsic::bswap:
1389   case Intrinsic::ctpop:
1390   case Intrinsic::ctlz:
1391   case Intrinsic::cttz:
1392   case Intrinsic::fma:
1393   case Intrinsic::fmuladd:
1394   case Intrinsic::copysign:
1395   case Intrinsic::launder_invariant_group:
1396   case Intrinsic::strip_invariant_group:
1397   case Intrinsic::round:
1398   case Intrinsic::masked_load:
1399   case Intrinsic::sadd_with_overflow:
1400   case Intrinsic::uadd_with_overflow:
1401   case Intrinsic::ssub_with_overflow:
1402   case Intrinsic::usub_with_overflow:
1403   case Intrinsic::smul_with_overflow:
1404   case Intrinsic::umul_with_overflow:
1405   case Intrinsic::convert_from_fp16:
1406   case Intrinsic::convert_to_fp16:
1407   case Intrinsic::bitreverse:
1408   case Intrinsic::x86_sse_cvtss2si:
1409   case Intrinsic::x86_sse_cvtss2si64:
1410   case Intrinsic::x86_sse_cvttss2si:
1411   case Intrinsic::x86_sse_cvttss2si64:
1412   case Intrinsic::x86_sse2_cvtsd2si:
1413   case Intrinsic::x86_sse2_cvtsd2si64:
1414   case Intrinsic::x86_sse2_cvttsd2si:
1415   case Intrinsic::x86_sse2_cvttsd2si64:
1416     return true;
1417   default:
1418     return false;
1419   case Intrinsic::not_intrinsic: break;
1420   }
1421 
1422   if (!F->hasName())
1423     return false;
1424   StringRef Name = F->getName();
1425 
1426   // In these cases, the check of the length is required.  We don't want to
1427   // return true for a name like "cos\0blah" which strcmp would return equal to
1428   // "cos", but has length 8.
1429   switch (Name[0]) {
1430   default:
1431     return false;
1432   case 'a':
1433     return Name == "acos" || Name == "asin" || Name == "atan" ||
1434            Name == "atan2" || Name == "acosf" || Name == "asinf" ||
1435            Name == "atanf" || Name == "atan2f";
1436   case 'c':
1437     return Name == "ceil" || Name == "cos" || Name == "cosh" ||
1438            Name == "ceilf" || Name == "cosf" || Name == "coshf";
1439   case 'e':
1440     return Name == "exp" || Name == "exp2" || Name == "expf" || Name == "exp2f";
1441   case 'f':
1442     return Name == "fabs" || Name == "floor" || Name == "fmod" ||
1443            Name == "fabsf" || Name == "floorf" || Name == "fmodf";
1444   case 'l':
1445     return Name == "log" || Name == "log10" || Name == "logf" ||
1446            Name == "log10f";
1447   case 'p':
1448     return Name == "pow" || Name == "powf";
1449   case 'r':
1450     return Name == "round" || Name == "roundf";
1451   case 's':
1452     return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1453            Name == "sinf" || Name == "sinhf" || Name == "sqrtf";
1454   case 't':
1455     return Name == "tan" || Name == "tanh" || Name == "tanf" || Name == "tanhf";
1456   case '_':
1457 
1458     // Check for various function names that get used for the math functions
1459     // when the header files are preprocessed with the macro
1460     // __FINITE_MATH_ONLY__ enabled.
1461     // The '12' here is the length of the shortest name that can match.
1462     // We need to check the size before looking at Name[1] and Name[2]
1463     // so we may as well check a limit that will eliminate mismatches.
1464     if (Name.size() < 12 || Name[1] != '_')
1465       return false;
1466     switch (Name[2]) {
1467     default:
1468       return false;
1469     case 'a':
1470       return Name == "__acos_finite" || Name == "__acosf_finite" ||
1471              Name == "__asin_finite" || Name == "__asinf_finite" ||
1472              Name == "__atan2_finite" || Name == "__atan2f_finite";
1473     case 'c':
1474       return Name == "__cosh_finite" || Name == "__coshf_finite";
1475     case 'e':
1476       return Name == "__exp_finite" || Name == "__expf_finite" ||
1477              Name == "__exp2_finite" || Name == "__exp2f_finite";
1478     case 'l':
1479       return Name == "__log_finite" || Name == "__logf_finite" ||
1480              Name == "__log10_finite" || Name == "__log10f_finite";
1481     case 'p':
1482       return Name == "__pow_finite" || Name == "__powf_finite";
1483     case 's':
1484       return Name == "__sinh_finite" || Name == "__sinhf_finite";
1485     }
1486   }
1487 }
1488 
1489 namespace {
1490 
GetConstantFoldFPValue(double V,Type * Ty)1491 Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1492   if (Ty->isHalfTy()) {
1493     APFloat APF(V);
1494     bool unused;
1495     APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &unused);
1496     return ConstantFP::get(Ty->getContext(), APF);
1497   }
1498   if (Ty->isFloatTy())
1499     return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1500   if (Ty->isDoubleTy())
1501     return ConstantFP::get(Ty->getContext(), APFloat(V));
1502   llvm_unreachable("Can only constant fold half/float/double");
1503 }
1504 
1505 /// Clear the floating-point exception state.
llvm_fenv_clearexcept()1506 inline void llvm_fenv_clearexcept() {
1507 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1508   feclearexcept(FE_ALL_EXCEPT);
1509 #endif
1510   errno = 0;
1511 }
1512 
1513 /// Test if a floating-point exception was raised.
llvm_fenv_testexcept()1514 inline bool llvm_fenv_testexcept() {
1515   int errno_val = errno;
1516   if (errno_val == ERANGE || errno_val == EDOM)
1517     return true;
1518 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1519   if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1520     return true;
1521 #endif
1522   return false;
1523 }
1524 
ConstantFoldFP(double (* NativeFP)(double),double V,Type * Ty)1525 Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) {
1526   llvm_fenv_clearexcept();
1527   V = NativeFP(V);
1528   if (llvm_fenv_testexcept()) {
1529     llvm_fenv_clearexcept();
1530     return nullptr;
1531   }
1532 
1533   return GetConstantFoldFPValue(V, Ty);
1534 }
1535 
ConstantFoldBinaryFP(double (* NativeFP)(double,double),double V,double W,Type * Ty)1536 Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
1537                                double W, Type *Ty) {
1538   llvm_fenv_clearexcept();
1539   V = NativeFP(V, W);
1540   if (llvm_fenv_testexcept()) {
1541     llvm_fenv_clearexcept();
1542     return nullptr;
1543   }
1544 
1545   return GetConstantFoldFPValue(V, Ty);
1546 }
1547 
1548 /// Attempt to fold an SSE floating point to integer conversion of a constant
1549 /// floating point. If roundTowardZero is false, the default IEEE rounding is
1550 /// used (toward nearest, ties to even). This matches the behavior of the
1551 /// non-truncating SSE instructions in the default rounding mode. The desired
1552 /// integer type Ty is used to select how many bits are available for the
1553 /// result. Returns null if the conversion cannot be performed, otherwise
1554 /// returns the Constant value resulting from the conversion.
ConstantFoldSSEConvertToInt(const APFloat & Val,bool roundTowardZero,Type * Ty)1555 Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
1556                                       Type *Ty) {
1557   // All of these conversion intrinsics form an integer of at most 64bits.
1558   unsigned ResultWidth = Ty->getIntegerBitWidth();
1559   assert(ResultWidth <= 64 &&
1560          "Can only constant fold conversions to 64 and 32 bit ints");
1561 
1562   uint64_t UIntVal;
1563   bool isExact = false;
1564   APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1565                                               : APFloat::rmNearestTiesToEven;
1566   APFloat::opStatus status =
1567       Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth,
1568                            /*isSigned=*/true, mode, &isExact);
1569   if (status != APFloat::opOK &&
1570       (!roundTowardZero || status != APFloat::opInexact))
1571     return nullptr;
1572   return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1573 }
1574 
getValueAsDouble(ConstantFP * Op)1575 double getValueAsDouble(ConstantFP *Op) {
1576   Type *Ty = Op->getType();
1577 
1578   if (Ty->isFloatTy())
1579     return Op->getValueAPF().convertToFloat();
1580 
1581   if (Ty->isDoubleTy())
1582     return Op->getValueAPF().convertToDouble();
1583 
1584   bool unused;
1585   APFloat APF = Op->getValueAPF();
1586   APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
1587   return APF.convertToDouble();
1588 }
1589 
ConstantFoldScalarCall(StringRef Name,unsigned IntrinsicID,Type * Ty,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI,ImmutableCallSite CS)1590 Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
1591                                  ArrayRef<Constant *> Operands,
1592                                  const TargetLibraryInfo *TLI,
1593                                  ImmutableCallSite CS) {
1594   if (Operands.size() == 1) {
1595     if (isa<UndefValue>(Operands[0])) {
1596       // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN
1597       if (IntrinsicID == Intrinsic::cos)
1598         return Constant::getNullValue(Ty);
1599       if (IntrinsicID == Intrinsic::bswap ||
1600           IntrinsicID == Intrinsic::bitreverse ||
1601           IntrinsicID == Intrinsic::launder_invariant_group ||
1602           IntrinsicID == Intrinsic::strip_invariant_group)
1603         return Operands[0];
1604     }
1605 
1606     if (isa<ConstantPointerNull>(Operands[0])) {
1607       // launder(null) == null == strip(null) iff in addrspace 0
1608       if (IntrinsicID == Intrinsic::launder_invariant_group ||
1609           IntrinsicID == Intrinsic::strip_invariant_group) {
1610         // If instruction is not yet put in a basic block (e.g. when cloning
1611         // a function during inlining), CS caller may not be available.
1612         // So check CS's BB first before querying CS.getCaller.
1613         const Function *Caller = CS.getParent() ? CS.getCaller() : nullptr;
1614         if (Caller &&
1615             !NullPointerIsDefined(
1616                 Caller, Operands[0]->getType()->getPointerAddressSpace())) {
1617           return Operands[0];
1618         }
1619         return nullptr;
1620       }
1621     }
1622 
1623     if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1624       if (IntrinsicID == Intrinsic::convert_to_fp16) {
1625         APFloat Val(Op->getValueAPF());
1626 
1627         bool lost = false;
1628         Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);
1629 
1630         return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1631       }
1632 
1633       if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1634         return nullptr;
1635 
1636       if (IntrinsicID == Intrinsic::round) {
1637         APFloat V = Op->getValueAPF();
1638         V.roundToIntegral(APFloat::rmNearestTiesToAway);
1639         return ConstantFP::get(Ty->getContext(), V);
1640       }
1641 
1642       if (IntrinsicID == Intrinsic::floor) {
1643         APFloat V = Op->getValueAPF();
1644         V.roundToIntegral(APFloat::rmTowardNegative);
1645         return ConstantFP::get(Ty->getContext(), V);
1646       }
1647 
1648       if (IntrinsicID == Intrinsic::ceil) {
1649         APFloat V = Op->getValueAPF();
1650         V.roundToIntegral(APFloat::rmTowardPositive);
1651         return ConstantFP::get(Ty->getContext(), V);
1652       }
1653 
1654       if (IntrinsicID == Intrinsic::trunc) {
1655         APFloat V = Op->getValueAPF();
1656         V.roundToIntegral(APFloat::rmTowardZero);
1657         return ConstantFP::get(Ty->getContext(), V);
1658       }
1659 
1660       if (IntrinsicID == Intrinsic::rint) {
1661         APFloat V = Op->getValueAPF();
1662         V.roundToIntegral(APFloat::rmNearestTiesToEven);
1663         return ConstantFP::get(Ty->getContext(), V);
1664       }
1665 
1666       if (IntrinsicID == Intrinsic::nearbyint) {
1667         APFloat V = Op->getValueAPF();
1668         V.roundToIntegral(APFloat::rmNearestTiesToEven);
1669         return ConstantFP::get(Ty->getContext(), V);
1670       }
1671 
1672       /// We only fold functions with finite arguments. Folding NaN and inf is
1673       /// likely to be aborted with an exception anyway, and some host libms
1674       /// have known errors raising exceptions.
1675       if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1676         return nullptr;
1677 
1678       /// Currently APFloat versions of these functions do not exist, so we use
1679       /// the host native double versions.  Float versions are not called
1680       /// directly but for all these it is true (float)(f((double)arg)) ==
1681       /// f(arg).  Long double not supported yet.
1682       double V = getValueAsDouble(Op);
1683 
1684       switch (IntrinsicID) {
1685         default: break;
1686         case Intrinsic::fabs:
1687           return ConstantFoldFP(fabs, V, Ty);
1688         case Intrinsic::log2:
1689           return ConstantFoldFP(Log2, V, Ty);
1690         case Intrinsic::log:
1691           return ConstantFoldFP(log, V, Ty);
1692         case Intrinsic::log10:
1693           return ConstantFoldFP(log10, V, Ty);
1694         case Intrinsic::exp:
1695           return ConstantFoldFP(exp, V, Ty);
1696         case Intrinsic::exp2:
1697           return ConstantFoldFP(exp2, V, Ty);
1698         case Intrinsic::sin:
1699           return ConstantFoldFP(sin, V, Ty);
1700         case Intrinsic::cos:
1701           return ConstantFoldFP(cos, V, Ty);
1702         case Intrinsic::sqrt:
1703           return ConstantFoldFP(sqrt, V, Ty);
1704       }
1705 
1706       if (!TLI)
1707         return nullptr;
1708 
1709       char NameKeyChar = Name[0];
1710       if (Name[0] == '_' && Name.size() > 2 && Name[1] == '_')
1711         NameKeyChar = Name[2];
1712 
1713       switch (NameKeyChar) {
1714       case 'a':
1715         if ((Name == "acos" && TLI->has(LibFunc_acos)) ||
1716             (Name == "acosf" && TLI->has(LibFunc_acosf)) ||
1717             (Name == "__acos_finite" && TLI->has(LibFunc_acos_finite)) ||
1718             (Name == "__acosf_finite" && TLI->has(LibFunc_acosf_finite)))
1719           return ConstantFoldFP(acos, V, Ty);
1720         else if ((Name == "asin" && TLI->has(LibFunc_asin)) ||
1721                  (Name == "asinf" && TLI->has(LibFunc_asinf)) ||
1722                  (Name == "__asin_finite" && TLI->has(LibFunc_asin_finite)) ||
1723                  (Name == "__asinf_finite" && TLI->has(LibFunc_asinf_finite)))
1724           return ConstantFoldFP(asin, V, Ty);
1725         else if ((Name == "atan" && TLI->has(LibFunc_atan)) ||
1726                  (Name == "atanf" && TLI->has(LibFunc_atanf)))
1727           return ConstantFoldFP(atan, V, Ty);
1728         break;
1729       case 'c':
1730         if ((Name == "ceil" && TLI->has(LibFunc_ceil)) ||
1731             (Name == "ceilf" && TLI->has(LibFunc_ceilf)))
1732           return ConstantFoldFP(ceil, V, Ty);
1733         else if ((Name == "cos" && TLI->has(LibFunc_cos)) ||
1734                  (Name == "cosf" && TLI->has(LibFunc_cosf)))
1735           return ConstantFoldFP(cos, V, Ty);
1736         else if ((Name == "cosh" && TLI->has(LibFunc_cosh)) ||
1737                  (Name == "coshf" && TLI->has(LibFunc_coshf)) ||
1738                  (Name == "__cosh_finite" && TLI->has(LibFunc_cosh_finite)) ||
1739                  (Name == "__coshf_finite" && TLI->has(LibFunc_coshf_finite)))
1740           return ConstantFoldFP(cosh, V, Ty);
1741         break;
1742       case 'e':
1743         if ((Name == "exp" && TLI->has(LibFunc_exp)) ||
1744             (Name == "expf" && TLI->has(LibFunc_expf)) ||
1745             (Name == "__exp_finite" && TLI->has(LibFunc_exp_finite)) ||
1746             (Name == "__expf_finite" && TLI->has(LibFunc_expf_finite)))
1747           return ConstantFoldFP(exp, V, Ty);
1748         if ((Name == "exp2" && TLI->has(LibFunc_exp2)) ||
1749             (Name == "exp2f" && TLI->has(LibFunc_exp2f)) ||
1750             (Name == "__exp2_finite" && TLI->has(LibFunc_exp2_finite)) ||
1751             (Name == "__exp2f_finite" && TLI->has(LibFunc_exp2f_finite)))
1752           // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1753           // C99 library.
1754           return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1755         break;
1756       case 'f':
1757         if ((Name == "fabs" && TLI->has(LibFunc_fabs)) ||
1758             (Name == "fabsf" && TLI->has(LibFunc_fabsf)))
1759           return ConstantFoldFP(fabs, V, Ty);
1760         else if ((Name == "floor" && TLI->has(LibFunc_floor)) ||
1761                  (Name == "floorf" && TLI->has(LibFunc_floorf)))
1762           return ConstantFoldFP(floor, V, Ty);
1763         break;
1764       case 'l':
1765         if ((Name == "log" && V > 0 && TLI->has(LibFunc_log)) ||
1766             (Name == "logf" && V > 0 && TLI->has(LibFunc_logf)) ||
1767             (Name == "__log_finite" && V > 0 &&
1768               TLI->has(LibFunc_log_finite)) ||
1769             (Name == "__logf_finite" && V > 0 &&
1770               TLI->has(LibFunc_logf_finite)))
1771           return ConstantFoldFP(log, V, Ty);
1772         else if ((Name == "log10" && V > 0 && TLI->has(LibFunc_log10)) ||
1773                  (Name == "log10f" && V > 0 && TLI->has(LibFunc_log10f)) ||
1774                  (Name == "__log10_finite" && V > 0 &&
1775                    TLI->has(LibFunc_log10_finite)) ||
1776                  (Name == "__log10f_finite" && V > 0 &&
1777                    TLI->has(LibFunc_log10f_finite)))
1778           return ConstantFoldFP(log10, V, Ty);
1779         break;
1780       case 'r':
1781         if ((Name == "round" && TLI->has(LibFunc_round)) ||
1782             (Name == "roundf" && TLI->has(LibFunc_roundf)))
1783           return ConstantFoldFP(round, V, Ty);
1784         break;
1785       case 's':
1786         if ((Name == "sin" && TLI->has(LibFunc_sin)) ||
1787             (Name == "sinf" && TLI->has(LibFunc_sinf)))
1788           return ConstantFoldFP(sin, V, Ty);
1789         else if ((Name == "sinh" && TLI->has(LibFunc_sinh)) ||
1790                  (Name == "sinhf" && TLI->has(LibFunc_sinhf)) ||
1791                  (Name == "__sinh_finite" && TLI->has(LibFunc_sinh_finite)) ||
1792                  (Name == "__sinhf_finite" && TLI->has(LibFunc_sinhf_finite)))
1793           return ConstantFoldFP(sinh, V, Ty);
1794         else if ((Name == "sqrt" && V >= 0 && TLI->has(LibFunc_sqrt)) ||
1795                  (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc_sqrtf)))
1796           return ConstantFoldFP(sqrt, V, Ty);
1797         break;
1798       case 't':
1799         if ((Name == "tan" && TLI->has(LibFunc_tan)) ||
1800             (Name == "tanf" && TLI->has(LibFunc_tanf)))
1801           return ConstantFoldFP(tan, V, Ty);
1802         else if ((Name == "tanh" && TLI->has(LibFunc_tanh)) ||
1803                  (Name == "tanhf" && TLI->has(LibFunc_tanhf)))
1804           return ConstantFoldFP(tanh, V, Ty);
1805         break;
1806       default:
1807         break;
1808       }
1809       return nullptr;
1810     }
1811 
1812     if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
1813       switch (IntrinsicID) {
1814       case Intrinsic::bswap:
1815         return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1816       case Intrinsic::ctpop:
1817         return ConstantInt::get(Ty, Op->getValue().countPopulation());
1818       case Intrinsic::bitreverse:
1819         return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
1820       case Intrinsic::convert_from_fp16: {
1821         APFloat Val(APFloat::IEEEhalf(), Op->getValue());
1822 
1823         bool lost = false;
1824         APFloat::opStatus status = Val.convert(
1825             Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
1826 
1827         // Conversion is always precise.
1828         (void)status;
1829         assert(status == APFloat::opOK && !lost &&
1830                "Precision lost during fp16 constfolding");
1831 
1832         return ConstantFP::get(Ty->getContext(), Val);
1833       }
1834       default:
1835         return nullptr;
1836       }
1837     }
1838 
1839     // Support ConstantVector in case we have an Undef in the top.
1840     if (isa<ConstantVector>(Operands[0]) ||
1841         isa<ConstantDataVector>(Operands[0])) {
1842       auto *Op = cast<Constant>(Operands[0]);
1843       switch (IntrinsicID) {
1844       default: break;
1845       case Intrinsic::x86_sse_cvtss2si:
1846       case Intrinsic::x86_sse_cvtss2si64:
1847       case Intrinsic::x86_sse2_cvtsd2si:
1848       case Intrinsic::x86_sse2_cvtsd2si64:
1849         if (ConstantFP *FPOp =
1850                 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1851           return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1852                                              /*roundTowardZero=*/false, Ty);
1853         break;
1854       case Intrinsic::x86_sse_cvttss2si:
1855       case Intrinsic::x86_sse_cvttss2si64:
1856       case Intrinsic::x86_sse2_cvttsd2si:
1857       case Intrinsic::x86_sse2_cvttsd2si64:
1858         if (ConstantFP *FPOp =
1859                 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1860           return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
1861                                              /*roundTowardZero=*/true, Ty);
1862         break;
1863       }
1864     }
1865 
1866     return nullptr;
1867   }
1868 
1869   if (Operands.size() == 2) {
1870     if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1871       if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1872         return nullptr;
1873       double Op1V = getValueAsDouble(Op1);
1874 
1875       if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1876         if (Op2->getType() != Op1->getType())
1877           return nullptr;
1878 
1879         double Op2V = getValueAsDouble(Op2);
1880         if (IntrinsicID == Intrinsic::pow) {
1881           return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1882         }
1883         if (IntrinsicID == Intrinsic::copysign) {
1884           APFloat V1 = Op1->getValueAPF();
1885           const APFloat &V2 = Op2->getValueAPF();
1886           V1.copySign(V2);
1887           return ConstantFP::get(Ty->getContext(), V1);
1888         }
1889 
1890         if (IntrinsicID == Intrinsic::minnum) {
1891           const APFloat &C1 = Op1->getValueAPF();
1892           const APFloat &C2 = Op2->getValueAPF();
1893           return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
1894         }
1895 
1896         if (IntrinsicID == Intrinsic::maxnum) {
1897           const APFloat &C1 = Op1->getValueAPF();
1898           const APFloat &C2 = Op2->getValueAPF();
1899           return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
1900         }
1901 
1902         if (!TLI)
1903           return nullptr;
1904         if ((Name == "pow" && TLI->has(LibFunc_pow)) ||
1905             (Name == "powf" && TLI->has(LibFunc_powf)) ||
1906             (Name == "__pow_finite" && TLI->has(LibFunc_pow_finite)) ||
1907             (Name == "__powf_finite" && TLI->has(LibFunc_powf_finite)))
1908           return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1909         if ((Name == "fmod" && TLI->has(LibFunc_fmod)) ||
1910             (Name == "fmodf" && TLI->has(LibFunc_fmodf)))
1911           return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1912         if ((Name == "atan2" && TLI->has(LibFunc_atan2)) ||
1913             (Name == "atan2f" && TLI->has(LibFunc_atan2f)) ||
1914             (Name == "__atan2_finite" && TLI->has(LibFunc_atan2_finite)) ||
1915             (Name == "__atan2f_finite" && TLI->has(LibFunc_atan2f_finite)))
1916           return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1917       } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1918         if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1919           return ConstantFP::get(Ty->getContext(),
1920                                  APFloat((float)std::pow((float)Op1V,
1921                                                  (int)Op2C->getZExtValue())));
1922         if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1923           return ConstantFP::get(Ty->getContext(),
1924                                  APFloat((float)std::pow((float)Op1V,
1925                                                  (int)Op2C->getZExtValue())));
1926         if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1927           return ConstantFP::get(Ty->getContext(),
1928                                  APFloat((double)std::pow((double)Op1V,
1929                                                    (int)Op2C->getZExtValue())));
1930       }
1931       return nullptr;
1932     }
1933 
1934     if (auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1935       if (auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1936         switch (IntrinsicID) {
1937         default: break;
1938         case Intrinsic::sadd_with_overflow:
1939         case Intrinsic::uadd_with_overflow:
1940         case Intrinsic::ssub_with_overflow:
1941         case Intrinsic::usub_with_overflow:
1942         case Intrinsic::smul_with_overflow:
1943         case Intrinsic::umul_with_overflow: {
1944           APInt Res;
1945           bool Overflow;
1946           switch (IntrinsicID) {
1947           default: llvm_unreachable("Invalid case");
1948           case Intrinsic::sadd_with_overflow:
1949             Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1950             break;
1951           case Intrinsic::uadd_with_overflow:
1952             Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1953             break;
1954           case Intrinsic::ssub_with_overflow:
1955             Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1956             break;
1957           case Intrinsic::usub_with_overflow:
1958             Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1959             break;
1960           case Intrinsic::smul_with_overflow:
1961             Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1962             break;
1963           case Intrinsic::umul_with_overflow:
1964             Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1965             break;
1966           }
1967           Constant *Ops[] = {
1968             ConstantInt::get(Ty->getContext(), Res),
1969             ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1970           };
1971           return ConstantStruct::get(cast<StructType>(Ty), Ops);
1972         }
1973         case Intrinsic::cttz:
1974           if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1975             return UndefValue::get(Ty);
1976           return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1977         case Intrinsic::ctlz:
1978           if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1979             return UndefValue::get(Ty);
1980           return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1981         }
1982       }
1983 
1984       return nullptr;
1985     }
1986     return nullptr;
1987   }
1988 
1989   if (Operands.size() != 3)
1990     return nullptr;
1991 
1992   if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1993     if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1994       if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1995         switch (IntrinsicID) {
1996         default: break;
1997         case Intrinsic::fma:
1998         case Intrinsic::fmuladd: {
1999           APFloat V = Op1->getValueAPF();
2000           APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
2001                                                    Op3->getValueAPF(),
2002                                                    APFloat::rmNearestTiesToEven);
2003           if (s != APFloat::opInvalidOp)
2004             return ConstantFP::get(Ty->getContext(), V);
2005 
2006           return nullptr;
2007         }
2008         }
2009       }
2010     }
2011   }
2012 
2013   return nullptr;
2014 }
2015 
ConstantFoldVectorCall(StringRef Name,unsigned IntrinsicID,VectorType * VTy,ArrayRef<Constant * > Operands,const DataLayout & DL,const TargetLibraryInfo * TLI,ImmutableCallSite CS)2016 Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
2017                                  VectorType *VTy, ArrayRef<Constant *> Operands,
2018                                  const DataLayout &DL,
2019                                  const TargetLibraryInfo *TLI,
2020                                  ImmutableCallSite CS) {
2021   SmallVector<Constant *, 4> Result(VTy->getNumElements());
2022   SmallVector<Constant *, 4> Lane(Operands.size());
2023   Type *Ty = VTy->getElementType();
2024 
2025   if (IntrinsicID == Intrinsic::masked_load) {
2026     auto *SrcPtr = Operands[0];
2027     auto *Mask = Operands[2];
2028     auto *Passthru = Operands[3];
2029 
2030     Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL);
2031 
2032     SmallVector<Constant *, 32> NewElements;
2033     for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2034       auto *MaskElt = Mask->getAggregateElement(I);
2035       if (!MaskElt)
2036         break;
2037       auto *PassthruElt = Passthru->getAggregateElement(I);
2038       auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
2039       if (isa<UndefValue>(MaskElt)) {
2040         if (PassthruElt)
2041           NewElements.push_back(PassthruElt);
2042         else if (VecElt)
2043           NewElements.push_back(VecElt);
2044         else
2045           return nullptr;
2046       }
2047       if (MaskElt->isNullValue()) {
2048         if (!PassthruElt)
2049           return nullptr;
2050         NewElements.push_back(PassthruElt);
2051       } else if (MaskElt->isOneValue()) {
2052         if (!VecElt)
2053           return nullptr;
2054         NewElements.push_back(VecElt);
2055       } else {
2056         return nullptr;
2057       }
2058     }
2059     if (NewElements.size() != VTy->getNumElements())
2060       return nullptr;
2061     return ConstantVector::get(NewElements);
2062   }
2063 
2064   for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
2065     // Gather a column of constants.
2066     for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
2067       // These intrinsics use a scalar type for their second argument.
2068       if (J == 1 &&
2069           (IntrinsicID == Intrinsic::cttz || IntrinsicID == Intrinsic::ctlz ||
2070            IntrinsicID == Intrinsic::powi)) {
2071         Lane[J] = Operands[J];
2072         continue;
2073       }
2074 
2075       Constant *Agg = Operands[J]->getAggregateElement(I);
2076       if (!Agg)
2077         return nullptr;
2078 
2079       Lane[J] = Agg;
2080     }
2081 
2082     // Use the regular scalar folding to simplify this column.
2083     Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, CS);
2084     if (!Folded)
2085       return nullptr;
2086     Result[I] = Folded;
2087   }
2088 
2089   return ConstantVector::get(Result);
2090 }
2091 
2092 } // end anonymous namespace
2093 
2094 Constant *
ConstantFoldCall(ImmutableCallSite CS,Function * F,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)2095 llvm::ConstantFoldCall(ImmutableCallSite CS, Function *F,
2096                        ArrayRef<Constant *> Operands,
2097                        const TargetLibraryInfo *TLI) {
2098   if (CS.isNoBuiltin() || CS.isStrictFP())
2099     return nullptr;
2100   if (!F->hasName())
2101     return nullptr;
2102   StringRef Name = F->getName();
2103 
2104   Type *Ty = F->getReturnType();
2105 
2106   if (auto *VTy = dyn_cast<VectorType>(Ty))
2107     return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
2108                                   F->getParent()->getDataLayout(), TLI, CS);
2109 
2110   return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI, CS);
2111 }
2112 
isMathLibCallNoop(CallSite CS,const TargetLibraryInfo * TLI)2113 bool llvm::isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI) {
2114   // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
2115   // (and to some extent ConstantFoldScalarCall).
2116   if (CS.isNoBuiltin() || CS.isStrictFP())
2117     return false;
2118   Function *F = CS.getCalledFunction();
2119   if (!F)
2120     return false;
2121 
2122   LibFunc Func;
2123   if (!TLI || !TLI->getLibFunc(*F, Func))
2124     return false;
2125 
2126   if (CS.getNumArgOperands() == 1) {
2127     if (ConstantFP *OpC = dyn_cast<ConstantFP>(CS.getArgOperand(0))) {
2128       const APFloat &Op = OpC->getValueAPF();
2129       switch (Func) {
2130       case LibFunc_logl:
2131       case LibFunc_log:
2132       case LibFunc_logf:
2133       case LibFunc_log2l:
2134       case LibFunc_log2:
2135       case LibFunc_log2f:
2136       case LibFunc_log10l:
2137       case LibFunc_log10:
2138       case LibFunc_log10f:
2139         return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
2140 
2141       case LibFunc_expl:
2142       case LibFunc_exp:
2143       case LibFunc_expf:
2144         // FIXME: These boundaries are slightly conservative.
2145         if (OpC->getType()->isDoubleTy())
2146           return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan &&
2147                  Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan;
2148         if (OpC->getType()->isFloatTy())
2149           return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan &&
2150                  Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan;
2151         break;
2152 
2153       case LibFunc_exp2l:
2154       case LibFunc_exp2:
2155       case LibFunc_exp2f:
2156         // FIXME: These boundaries are slightly conservative.
2157         if (OpC->getType()->isDoubleTy())
2158           return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan &&
2159                  Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan;
2160         if (OpC->getType()->isFloatTy())
2161           return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan &&
2162                  Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan;
2163         break;
2164 
2165       case LibFunc_sinl:
2166       case LibFunc_sin:
2167       case LibFunc_sinf:
2168       case LibFunc_cosl:
2169       case LibFunc_cos:
2170       case LibFunc_cosf:
2171         return !Op.isInfinity();
2172 
2173       case LibFunc_tanl:
2174       case LibFunc_tan:
2175       case LibFunc_tanf: {
2176         // FIXME: Stop using the host math library.
2177         // FIXME: The computation isn't done in the right precision.
2178         Type *Ty = OpC->getType();
2179         if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2180           double OpV = getValueAsDouble(OpC);
2181           return ConstantFoldFP(tan, OpV, Ty) != nullptr;
2182         }
2183         break;
2184       }
2185 
2186       case LibFunc_asinl:
2187       case LibFunc_asin:
2188       case LibFunc_asinf:
2189       case LibFunc_acosl:
2190       case LibFunc_acos:
2191       case LibFunc_acosf:
2192         return Op.compare(APFloat(Op.getSemantics(), "-1")) !=
2193                    APFloat::cmpLessThan &&
2194                Op.compare(APFloat(Op.getSemantics(), "1")) !=
2195                    APFloat::cmpGreaterThan;
2196 
2197       case LibFunc_sinh:
2198       case LibFunc_cosh:
2199       case LibFunc_sinhf:
2200       case LibFunc_coshf:
2201       case LibFunc_sinhl:
2202       case LibFunc_coshl:
2203         // FIXME: These boundaries are slightly conservative.
2204         if (OpC->getType()->isDoubleTy())
2205           return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan &&
2206                  Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan;
2207         if (OpC->getType()->isFloatTy())
2208           return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan &&
2209                  Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan;
2210         break;
2211 
2212       case LibFunc_sqrtl:
2213       case LibFunc_sqrt:
2214       case LibFunc_sqrtf:
2215         return Op.isNaN() || Op.isZero() || !Op.isNegative();
2216 
2217       // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
2218       // maybe others?
2219       default:
2220         break;
2221       }
2222     }
2223   }
2224 
2225   if (CS.getNumArgOperands() == 2) {
2226     ConstantFP *Op0C = dyn_cast<ConstantFP>(CS.getArgOperand(0));
2227     ConstantFP *Op1C = dyn_cast<ConstantFP>(CS.getArgOperand(1));
2228     if (Op0C && Op1C) {
2229       const APFloat &Op0 = Op0C->getValueAPF();
2230       const APFloat &Op1 = Op1C->getValueAPF();
2231 
2232       switch (Func) {
2233       case LibFunc_powl:
2234       case LibFunc_pow:
2235       case LibFunc_powf: {
2236         // FIXME: Stop using the host math library.
2237         // FIXME: The computation isn't done in the right precision.
2238         Type *Ty = Op0C->getType();
2239         if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
2240           if (Ty == Op1C->getType()) {
2241             double Op0V = getValueAsDouble(Op0C);
2242             double Op1V = getValueAsDouble(Op1C);
2243             return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr;
2244           }
2245         }
2246         break;
2247       }
2248 
2249       case LibFunc_fmodl:
2250       case LibFunc_fmod:
2251       case LibFunc_fmodf:
2252         return Op0.isNaN() || Op1.isNaN() ||
2253                (!Op0.isInfinity() && !Op1.isZero());
2254 
2255       default:
2256         break;
2257       }
2258     }
2259   }
2260 
2261   return false;
2262 }
2263