• 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/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringMap.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/Config/config.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/Operator.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/MathExtras.h"
38 #include <cassert>
39 #include <cerrno>
40 #include <cfenv>
41 #include <cmath>
42 #include <limits>
43 
44 using namespace llvm;
45 
46 namespace {
47 
48 //===----------------------------------------------------------------------===//
49 // Constant Folding internal helper functions
50 //===----------------------------------------------------------------------===//
51 
52 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
53 /// This always returns a non-null constant, but it may be a
54 /// ConstantExpr if unfoldable.
FoldBitCast(Constant * C,Type * DestTy,const DataLayout & DL)55 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
56   // Catch the obvious splat cases.
57   if (C->isNullValue() && !DestTy->isX86_MMXTy())
58     return Constant::getNullValue(DestTy);
59   if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
60       !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
61     return Constant::getAllOnesValue(DestTy);
62 
63   // Handle a vector->integer cast.
64   if (auto *IT = dyn_cast<IntegerType>(DestTy)) {
65     auto *VTy = dyn_cast<VectorType>(C->getType());
66     if (!VTy)
67       return ConstantExpr::getBitCast(C, DestTy);
68 
69     unsigned NumSrcElts = VTy->getNumElements();
70     Type *SrcEltTy = VTy->getElementType();
71 
72     // If the vector is a vector of floating point, convert it to vector of int
73     // to simplify things.
74     if (SrcEltTy->isFloatingPointTy()) {
75       unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
76       Type *SrcIVTy =
77         VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
78       // Ask IR to do the conversion now that #elts line up.
79       C = ConstantExpr::getBitCast(C, SrcIVTy);
80     }
81 
82     // Now that we know that the input value is a vector of integers, just shift
83     // and insert them into our result.
84     unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
85     APInt Result(IT->getBitWidth(), 0);
86     for (unsigned i = 0; i != NumSrcElts; ++i) {
87       Constant *Element;
88       if (DL.isLittleEndian())
89         Element = C->getAggregateElement(NumSrcElts-i-1);
90       else
91         Element = C->getAggregateElement(i);
92 
93       auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
94       if (!ElementCI)
95         return ConstantExpr::getBitCast(C, DestTy);
96 
97       Result <<= BitShift;
98       Result |= ElementCI->getValue().zextOrSelf(IT->getBitWidth());
99     }
100 
101     return ConstantInt::get(IT, Result);
102   }
103 
104   // The code below only handles casts to vectors currently.
105   auto *DestVTy = dyn_cast<VectorType>(DestTy);
106   if (!DestVTy)
107     return ConstantExpr::getBitCast(C, DestTy);
108 
109   // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
110   // vector so the code below can handle it uniformly.
111   if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
112     Constant *Ops = C; // don't take the address of C!
113     return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
114   }
115 
116   // If this is a bitcast from constant vector -> vector, fold it.
117   if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
118     return ConstantExpr::getBitCast(C, DestTy);
119 
120   // If the element types match, IR can fold it.
121   unsigned NumDstElt = DestVTy->getNumElements();
122   unsigned NumSrcElt = C->getType()->getVectorNumElements();
123   if (NumDstElt == NumSrcElt)
124     return ConstantExpr::getBitCast(C, DestTy);
125 
126   Type *SrcEltTy = C->getType()->getVectorElementType();
127   Type *DstEltTy = DestVTy->getElementType();
128 
129   // Otherwise, we're changing the number of elements in a vector, which
130   // requires endianness information to do the right thing.  For example,
131   //    bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
132   // folds to (little endian):
133   //    <4 x i32> <i32 0, i32 0, i32 1, i32 0>
134   // and to (big endian):
135   //    <4 x i32> <i32 0, i32 0, i32 0, i32 1>
136 
137   // First thing is first.  We only want to think about integer here, so if
138   // we have something in FP form, recast it as integer.
139   if (DstEltTy->isFloatingPointTy()) {
140     // Fold to an vector of integers with same size as our FP type.
141     unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
142     Type *DestIVTy =
143       VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
144     // Recursively handle this integer conversion, if possible.
145     C = FoldBitCast(C, DestIVTy, DL);
146 
147     // Finally, IR can handle this now that #elts line up.
148     return ConstantExpr::getBitCast(C, DestTy);
149   }
150 
151   // Okay, we know the destination is integer, if the input is FP, convert
152   // it to integer first.
153   if (SrcEltTy->isFloatingPointTy()) {
154     unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
155     Type *SrcIVTy =
156       VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
157     // Ask IR to do the conversion now that #elts line up.
158     C = ConstantExpr::getBitCast(C, SrcIVTy);
159     // If IR wasn't able to fold it, bail out.
160     if (!isa<ConstantVector>(C) &&  // FIXME: Remove ConstantVector.
161         !isa<ConstantDataVector>(C))
162       return C;
163   }
164 
165   // Now we know that the input and output vectors are both integer vectors
166   // of the same size, and that their #elements is not the same.  Do the
167   // conversion here, which depends on whether the input or output has
168   // more elements.
169   bool isLittleEndian = DL.isLittleEndian();
170 
171   SmallVector<Constant*, 32> Result;
172   if (NumDstElt < NumSrcElt) {
173     // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
174     Constant *Zero = Constant::getNullValue(DstEltTy);
175     unsigned Ratio = NumSrcElt/NumDstElt;
176     unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
177     unsigned SrcElt = 0;
178     for (unsigned i = 0; i != NumDstElt; ++i) {
179       // Build each element of the result.
180       Constant *Elt = Zero;
181       unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
182       for (unsigned j = 0; j != Ratio; ++j) {
183         Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
184         if (!Src)  // Reject constantexpr elements.
185           return ConstantExpr::getBitCast(C, DestTy);
186 
187         // Zero extend the element to the right size.
188         Src = ConstantExpr::getZExt(Src, Elt->getType());
189 
190         // Shift it to the right place, depending on endianness.
191         Src = ConstantExpr::getShl(Src,
192                                    ConstantInt::get(Src->getType(), ShiftAmt));
193         ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
194 
195         // Mix it in.
196         Elt = ConstantExpr::getOr(Elt, Src);
197       }
198       Result.push_back(Elt);
199     }
200     return ConstantVector::get(Result);
201   }
202 
203   // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
204   unsigned Ratio = NumDstElt/NumSrcElt;
205   unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
206 
207   // Loop over each source value, expanding into multiple results.
208   for (unsigned i = 0; i != NumSrcElt; ++i) {
209     auto *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
210     if (!Src)  // Reject constantexpr elements.
211       return ConstantExpr::getBitCast(C, DestTy);
212 
213     unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
214     for (unsigned j = 0; j != Ratio; ++j) {
215       // Shift the piece of the value into the right place, depending on
216       // endianness.
217       Constant *Elt = ConstantExpr::getLShr(Src,
218                                   ConstantInt::get(Src->getType(), ShiftAmt));
219       ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
220 
221       // Truncate the element to an integer with the same pointer size and
222       // convert the element back to a pointer using a inttoptr.
223       if (DstEltTy->isPointerTy()) {
224         IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
225         Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
226         Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
227         continue;
228       }
229 
230       // Truncate and remember this piece.
231       Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
232     }
233   }
234 
235   return ConstantVector::get(Result);
236 }
237 
238 } // end anonymous namespace
239 
240 /// If this constant is a constant offset from a global, return the global and
241 /// the constant. Because of constantexprs, this function is recursive.
IsConstantOffsetFromGlobal(Constant * C,GlobalValue * & GV,APInt & Offset,const DataLayout & DL)242 bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
243                                       APInt &Offset, const DataLayout &DL) {
244   // Trivial case, constant is the global.
245   if ((GV = dyn_cast<GlobalValue>(C))) {
246     unsigned BitWidth = DL.getPointerTypeSizeInBits(GV->getType());
247     Offset = APInt(BitWidth, 0);
248     return true;
249   }
250 
251   // Otherwise, if this isn't a constant expr, bail out.
252   auto *CE = dyn_cast<ConstantExpr>(C);
253   if (!CE) return false;
254 
255   // Look through ptr->int and ptr->ptr casts.
256   if (CE->getOpcode() == Instruction::PtrToInt ||
257       CE->getOpcode() == Instruction::BitCast)
258     return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
259 
260   // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
261   auto *GEP = dyn_cast<GEPOperator>(CE);
262   if (!GEP)
263     return false;
264 
265   unsigned BitWidth = DL.getPointerTypeSizeInBits(GEP->getType());
266   APInt TmpOffset(BitWidth, 0);
267 
268   // If the base isn't a global+constant, we aren't either.
269   if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
270     return false;
271 
272   // Otherwise, add any offset that our operands provide.
273   if (!GEP->accumulateConstantOffset(DL, TmpOffset))
274     return false;
275 
276   Offset = TmpOffset;
277   return true;
278 }
279 
280 namespace {
281 
282 /// Recursive helper to read bits out of global. C is the constant being copied
283 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
284 /// results into and BytesLeft is the number of bytes left in
285 /// the CurPtr buffer. DL is the DataLayout.
ReadDataFromGlobal(Constant * C,uint64_t ByteOffset,unsigned char * CurPtr,unsigned BytesLeft,const DataLayout & DL)286 bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
287                         unsigned BytesLeft, const DataLayout &DL) {
288   assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
289          "Out of range access");
290 
291   // If this element is zero or undefined, we can just return since *CurPtr is
292   // zero initialized.
293   if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
294     return true;
295 
296   if (auto *CI = dyn_cast<ConstantInt>(C)) {
297     if (CI->getBitWidth() > 64 ||
298         (CI->getBitWidth() & 7) != 0)
299       return false;
300 
301     uint64_t Val = CI->getZExtValue();
302     unsigned IntBytes = unsigned(CI->getBitWidth()/8);
303 
304     for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
305       int n = ByteOffset;
306       if (!DL.isLittleEndian())
307         n = IntBytes - n - 1;
308       CurPtr[i] = (unsigned char)(Val >> (n * 8));
309       ++ByteOffset;
310     }
311     return true;
312   }
313 
314   if (auto *CFP = dyn_cast<ConstantFP>(C)) {
315     if (CFP->getType()->isDoubleTy()) {
316       C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
317       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
318     }
319     if (CFP->getType()->isFloatTy()){
320       C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
321       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
322     }
323     if (CFP->getType()->isHalfTy()){
324       C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
325       return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
326     }
327     return false;
328   }
329 
330   if (auto *CS = dyn_cast<ConstantStruct>(C)) {
331     const StructLayout *SL = DL.getStructLayout(CS->getType());
332     unsigned Index = SL->getElementContainingOffset(ByteOffset);
333     uint64_t CurEltOffset = SL->getElementOffset(Index);
334     ByteOffset -= CurEltOffset;
335 
336     while (1) {
337       // If the element access is to the element itself and not to tail padding,
338       // read the bytes from the element.
339       uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
340 
341       if (ByteOffset < EltSize &&
342           !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
343                               BytesLeft, DL))
344         return false;
345 
346       ++Index;
347 
348       // Check to see if we read from the last struct element, if so we're done.
349       if (Index == CS->getType()->getNumElements())
350         return true;
351 
352       // If we read all of the bytes we needed from this element we're done.
353       uint64_t NextEltOffset = SL->getElementOffset(Index);
354 
355       if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
356         return true;
357 
358       // Move to the next element of the struct.
359       CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
360       BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
361       ByteOffset = 0;
362       CurEltOffset = NextEltOffset;
363     }
364     // not reached.
365   }
366 
367   if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
368       isa<ConstantDataSequential>(C)) {
369     Type *EltTy = C->getType()->getSequentialElementType();
370     uint64_t EltSize = DL.getTypeAllocSize(EltTy);
371     uint64_t Index = ByteOffset / EltSize;
372     uint64_t Offset = ByteOffset - Index * EltSize;
373     uint64_t NumElts;
374     if (auto *AT = dyn_cast<ArrayType>(C->getType()))
375       NumElts = AT->getNumElements();
376     else
377       NumElts = C->getType()->getVectorNumElements();
378 
379     for (; Index != NumElts; ++Index) {
380       if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
381                               BytesLeft, DL))
382         return false;
383 
384       uint64_t BytesWritten = EltSize - Offset;
385       assert(BytesWritten <= EltSize && "Not indexing into this element?");
386       if (BytesWritten >= BytesLeft)
387         return true;
388 
389       Offset = 0;
390       BytesLeft -= BytesWritten;
391       CurPtr += BytesWritten;
392     }
393     return true;
394   }
395 
396   if (auto *CE = dyn_cast<ConstantExpr>(C)) {
397     if (CE->getOpcode() == Instruction::IntToPtr &&
398         CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
399       return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
400                                 BytesLeft, DL);
401     }
402   }
403 
404   // Otherwise, unknown initializer type.
405   return false;
406 }
407 
FoldReinterpretLoadFromConstPtr(Constant * C,Type * LoadTy,const DataLayout & DL)408 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
409                                           const DataLayout &DL) {
410   auto *PTy = cast<PointerType>(C->getType());
411   auto *IntType = dyn_cast<IntegerType>(LoadTy);
412 
413   // If this isn't an integer load we can't fold it directly.
414   if (!IntType) {
415     unsigned AS = PTy->getAddressSpace();
416 
417     // If this is a float/double load, we can try folding it as an int32/64 load
418     // and then bitcast the result.  This can be useful for union cases.  Note
419     // that address spaces don't matter here since we're not going to result in
420     // an actual new load.
421     Type *MapTy;
422     if (LoadTy->isHalfTy())
423       MapTy = Type::getInt16Ty(C->getContext());
424     else if (LoadTy->isFloatTy())
425       MapTy = Type::getInt32Ty(C->getContext());
426     else if (LoadTy->isDoubleTy())
427       MapTy = Type::getInt64Ty(C->getContext());
428     else if (LoadTy->isVectorTy()) {
429       MapTy = PointerType::getIntNTy(C->getContext(),
430                                      DL.getTypeAllocSizeInBits(LoadTy));
431     } else
432       return nullptr;
433 
434     C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
435     if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL))
436       return FoldBitCast(Res, LoadTy, DL);
437     return nullptr;
438   }
439 
440   unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
441   if (BytesLoaded > 32 || BytesLoaded == 0)
442     return nullptr;
443 
444   GlobalValue *GVal;
445   APInt OffsetAI;
446   if (!IsConstantOffsetFromGlobal(C, GVal, OffsetAI, DL))
447     return nullptr;
448 
449   auto *GV = dyn_cast<GlobalVariable>(GVal);
450   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
451       !GV->getInitializer()->getType()->isSized())
452     return nullptr;
453 
454   int64_t Offset = OffsetAI.getSExtValue();
455   int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
456 
457   // If we're not accessing anything in this constant, the result is undefined.
458   if (Offset + BytesLoaded <= 0)
459     return UndefValue::get(IntType);
460 
461   // If we're not accessing anything in this constant, the result is undefined.
462   if (Offset >= InitializerSize)
463     return UndefValue::get(IntType);
464 
465   unsigned char RawBytes[32] = {0};
466   unsigned char *CurPtr = RawBytes;
467   unsigned BytesLeft = BytesLoaded;
468 
469   // If we're loading off the beginning of the global, some bytes may be valid.
470   if (Offset < 0) {
471     CurPtr += -Offset;
472     BytesLeft += Offset;
473     Offset = 0;
474   }
475 
476   if (!ReadDataFromGlobal(GV->getInitializer(), Offset, CurPtr, BytesLeft, DL))
477     return nullptr;
478 
479   APInt ResultVal = APInt(IntType->getBitWidth(), 0);
480   if (DL.isLittleEndian()) {
481     ResultVal = RawBytes[BytesLoaded - 1];
482     for (unsigned i = 1; i != BytesLoaded; ++i) {
483       ResultVal <<= 8;
484       ResultVal |= RawBytes[BytesLoaded - 1 - i];
485     }
486   } else {
487     ResultVal = RawBytes[0];
488     for (unsigned i = 1; i != BytesLoaded; ++i) {
489       ResultVal <<= 8;
490       ResultVal |= RawBytes[i];
491     }
492   }
493 
494   return ConstantInt::get(IntType->getContext(), ResultVal);
495 }
496 
ConstantFoldLoadThroughBitcast(ConstantExpr * CE,Type * DestTy,const DataLayout & DL)497 Constant *ConstantFoldLoadThroughBitcast(ConstantExpr *CE, Type *DestTy,
498                                          const DataLayout &DL) {
499   auto *SrcPtr = CE->getOperand(0);
500   auto *SrcPtrTy = dyn_cast<PointerType>(SrcPtr->getType());
501   if (!SrcPtrTy)
502     return nullptr;
503   Type *SrcTy = SrcPtrTy->getPointerElementType();
504 
505   Constant *C = ConstantFoldLoadFromConstPtr(SrcPtr, SrcTy, DL);
506   if (!C)
507     return nullptr;
508 
509   do {
510     Type *SrcTy = C->getType();
511 
512     // If the type sizes are the same and a cast is legal, just directly
513     // cast the constant.
514     if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
515       Instruction::CastOps Cast = Instruction::BitCast;
516       // If we are going from a pointer to int or vice versa, we spell the cast
517       // differently.
518       if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
519         Cast = Instruction::IntToPtr;
520       else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
521         Cast = Instruction::PtrToInt;
522 
523       if (CastInst::castIsValid(Cast, C, DestTy))
524         return ConstantExpr::getCast(Cast, C, DestTy);
525     }
526 
527     // If this isn't an aggregate type, there is nothing we can do to drill down
528     // and find a bitcastable constant.
529     if (!SrcTy->isAggregateType())
530       return nullptr;
531 
532     // We're simulating a load through a pointer that was bitcast to point to
533     // a different type, so we can try to walk down through the initial
534     // elements of an aggregate to see if some part of th e aggregate is
535     // castable to implement the "load" semantic model.
536     C = C->getAggregateElement(0u);
537   } while (C);
538 
539   return nullptr;
540 }
541 
542 } // end anonymous namespace
543 
ConstantFoldLoadFromConstPtr(Constant * C,Type * Ty,const DataLayout & DL)544 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
545                                              const DataLayout &DL) {
546   // First, try the easy cases:
547   if (auto *GV = dyn_cast<GlobalVariable>(C))
548     if (GV->isConstant() && GV->hasDefinitiveInitializer())
549       return GV->getInitializer();
550 
551   if (auto *GA = dyn_cast<GlobalAlias>(C))
552     if (GA->getAliasee() && !GA->isInterposable())
553       return ConstantFoldLoadFromConstPtr(GA->getAliasee(), Ty, DL);
554 
555   // If the loaded value isn't a constant expr, we can't handle it.
556   auto *CE = dyn_cast<ConstantExpr>(C);
557   if (!CE)
558     return nullptr;
559 
560   if (CE->getOpcode() == Instruction::GetElementPtr) {
561     if (auto *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
562       if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
563         if (Constant *V =
564              ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
565           return V;
566       }
567     }
568   }
569 
570   if (CE->getOpcode() == Instruction::BitCast)
571     if (Constant *LoadedC = ConstantFoldLoadThroughBitcast(CE, Ty, DL))
572       return LoadedC;
573 
574   // Instead of loading constant c string, use corresponding integer value
575   // directly if string length is small enough.
576   StringRef Str;
577   if (getConstantStringInfo(CE, Str) && !Str.empty()) {
578     size_t StrLen = Str.size();
579     unsigned NumBits = Ty->getPrimitiveSizeInBits();
580     // Replace load with immediate integer if the result is an integer or fp
581     // value.
582     if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
583         (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
584       APInt StrVal(NumBits, 0);
585       APInt SingleChar(NumBits, 0);
586       if (DL.isLittleEndian()) {
587         for (unsigned char C : reverse(Str.bytes())) {
588           SingleChar = static_cast<uint64_t>(C);
589           StrVal = (StrVal << 8) | SingleChar;
590         }
591       } else {
592         for (unsigned char C : Str.bytes()) {
593           SingleChar = static_cast<uint64_t>(C);
594           StrVal = (StrVal << 8) | SingleChar;
595         }
596         // Append NULL at the end.
597         SingleChar = 0;
598         StrVal = (StrVal << 8) | SingleChar;
599       }
600 
601       Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
602       if (Ty->isFloatingPointTy())
603         Res = ConstantExpr::getBitCast(Res, Ty);
604       return Res;
605     }
606   }
607 
608   // If this load comes from anywhere in a constant global, and if the global
609   // is all undef or zero, we know what it loads.
610   if (auto *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
611     if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
612       if (GV->getInitializer()->isNullValue())
613         return Constant::getNullValue(Ty);
614       if (isa<UndefValue>(GV->getInitializer()))
615         return UndefValue::get(Ty);
616     }
617   }
618 
619   // Try hard to fold loads from bitcasted strange and non-type-safe things.
620   return FoldReinterpretLoadFromConstPtr(CE, Ty, DL);
621 }
622 
623 namespace {
624 
ConstantFoldLoadInst(const LoadInst * LI,const DataLayout & DL)625 Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout &DL) {
626   if (LI->isVolatile()) return nullptr;
627 
628   if (auto *C = dyn_cast<Constant>(LI->getOperand(0)))
629     return ConstantFoldLoadFromConstPtr(C, LI->getType(), DL);
630 
631   return nullptr;
632 }
633 
634 /// One of Op0/Op1 is a constant expression.
635 /// Attempt to symbolically evaluate the result of a binary operator merging
636 /// these together.  If target data info is available, it is provided as DL,
637 /// otherwise DL is null.
SymbolicallyEvaluateBinop(unsigned Opc,Constant * Op0,Constant * Op1,const DataLayout & DL)638 Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
639                                     const DataLayout &DL) {
640   // SROA
641 
642   // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
643   // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
644   // bits.
645 
646   if (Opc == Instruction::And) {
647     unsigned BitWidth = DL.getTypeSizeInBits(Op0->getType()->getScalarType());
648     APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
649     APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
650     computeKnownBits(Op0, KnownZero0, KnownOne0, DL);
651     computeKnownBits(Op1, KnownZero1, KnownOne1, DL);
652     if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
653       // All the bits of Op0 that the 'and' could be masking are already zero.
654       return Op0;
655     }
656     if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
657       // All the bits of Op1 that the 'and' could be masking are already zero.
658       return Op1;
659     }
660 
661     APInt KnownZero = KnownZero0 | KnownZero1;
662     APInt KnownOne = KnownOne0 & KnownOne1;
663     if ((KnownZero | KnownOne).isAllOnesValue()) {
664       return ConstantInt::get(Op0->getType(), KnownOne);
665     }
666   }
667 
668   // If the constant expr is something like &A[123] - &A[4].f, fold this into a
669   // constant.  This happens frequently when iterating over a global array.
670   if (Opc == Instruction::Sub) {
671     GlobalValue *GV1, *GV2;
672     APInt Offs1, Offs2;
673 
674     if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
675       if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
676         unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
677 
678         // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
679         // PtrToInt may change the bitwidth so we have convert to the right size
680         // first.
681         return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
682                                                 Offs2.zextOrTrunc(OpSize));
683       }
684   }
685 
686   return nullptr;
687 }
688 
689 /// If array indices are not pointer-sized integers, explicitly cast them so
690 /// that they aren't implicitly casted by the getelementptr.
CastGEPIndices(Type * SrcElemTy,ArrayRef<Constant * > Ops,Type * ResultTy,const DataLayout & DL,const TargetLibraryInfo * TLI)691 Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
692                          Type *ResultTy, const DataLayout &DL,
693                          const TargetLibraryInfo *TLI) {
694   Type *IntPtrTy = DL.getIntPtrType(ResultTy);
695 
696   bool Any = false;
697   SmallVector<Constant*, 32> NewIdxs;
698   for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
699     if ((i == 1 ||
700          !isa<StructType>(GetElementPtrInst::getIndexedType(SrcElemTy,
701              Ops.slice(1, i - 1)))) &&
702         Ops[i]->getType() != IntPtrTy) {
703       Any = true;
704       NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
705                                                                       true,
706                                                                       IntPtrTy,
707                                                                       true),
708                                               Ops[i], IntPtrTy));
709     } else
710       NewIdxs.push_back(Ops[i]);
711   }
712 
713   if (!Any)
714     return nullptr;
715 
716   Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ops[0], NewIdxs);
717   if (auto *CE = dyn_cast<ConstantExpr>(C)) {
718     if (Constant *Folded = ConstantFoldConstantExpression(CE, DL, TLI))
719       C = Folded;
720   }
721 
722   return C;
723 }
724 
725 /// Strip the pointer casts, but preserve the address space information.
StripPtrCastKeepAS(Constant * Ptr,Type * & ElemTy)726 Constant* StripPtrCastKeepAS(Constant* Ptr, Type *&ElemTy) {
727   assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
728   auto *OldPtrTy = cast<PointerType>(Ptr->getType());
729   Ptr = Ptr->stripPointerCasts();
730   auto *NewPtrTy = cast<PointerType>(Ptr->getType());
731 
732   ElemTy = NewPtrTy->getPointerElementType();
733 
734   // Preserve the address space number of the pointer.
735   if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
736     NewPtrTy = ElemTy->getPointerTo(OldPtrTy->getAddressSpace());
737     Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
738   }
739   return Ptr;
740 }
741 
742 /// If we can symbolically evaluate the GEP constant expression, do so.
SymbolicallyEvaluateGEP(const GEPOperator * GEP,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)743 Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
744                                   ArrayRef<Constant *> Ops,
745                                   const DataLayout &DL,
746                                   const TargetLibraryInfo *TLI) {
747   Type *SrcElemTy = GEP->getSourceElementType();
748   Type *ResElemTy = GEP->getResultElementType();
749   Type *ResTy = GEP->getType();
750   if (!SrcElemTy->isSized())
751     return nullptr;
752 
753   if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy, DL, TLI))
754     return C;
755 
756   Constant *Ptr = Ops[0];
757   if (!Ptr->getType()->isPointerTy())
758     return nullptr;
759 
760   Type *IntPtrTy = DL.getIntPtrType(Ptr->getType());
761 
762   // If this is a constant expr gep that is effectively computing an
763   // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
764   for (unsigned i = 1, e = Ops.size(); i != e; ++i)
765     if (!isa<ConstantInt>(Ops[i])) {
766 
767       // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
768       // "inttoptr (sub (ptrtoint Ptr), V)"
769       if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
770         auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
771         assert((!CE || CE->getType() == IntPtrTy) &&
772                "CastGEPIndices didn't canonicalize index types!");
773         if (CE && CE->getOpcode() == Instruction::Sub &&
774             CE->getOperand(0)->isNullValue()) {
775           Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
776           Res = ConstantExpr::getSub(Res, CE->getOperand(1));
777           Res = ConstantExpr::getIntToPtr(Res, ResTy);
778           if (auto *ResCE = dyn_cast<ConstantExpr>(Res))
779             Res = ConstantFoldConstantExpression(ResCE, DL, TLI);
780           return Res;
781         }
782       }
783       return nullptr;
784     }
785 
786   unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy);
787   APInt Offset =
788       APInt(BitWidth,
789             DL.getIndexedOffsetInType(
790                 SrcElemTy,
791                 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
792   Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
793 
794   // If this is a GEP of a GEP, fold it all into a single GEP.
795   while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
796     SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
797 
798     // Do not try the incorporate the sub-GEP if some index is not a number.
799     bool AllConstantInt = true;
800     for (Value *NestedOp : NestedOps)
801       if (!isa<ConstantInt>(NestedOp)) {
802         AllConstantInt = false;
803         break;
804       }
805     if (!AllConstantInt)
806       break;
807 
808     Ptr = cast<Constant>(GEP->getOperand(0));
809     SrcElemTy = GEP->getSourceElementType();
810     Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
811     Ptr = StripPtrCastKeepAS(Ptr, SrcElemTy);
812   }
813 
814   // If the base value for this address is a literal integer value, fold the
815   // getelementptr to the resulting integer value casted to the pointer type.
816   APInt BasePtr(BitWidth, 0);
817   if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
818     if (CE->getOpcode() == Instruction::IntToPtr) {
819       if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
820         BasePtr = Base->getValue().zextOrTrunc(BitWidth);
821     }
822   }
823 
824   if (Ptr->isNullValue() || BasePtr != 0) {
825     Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
826     return ConstantExpr::getIntToPtr(C, ResTy);
827   }
828 
829   // Otherwise form a regular getelementptr. Recompute the indices so that
830   // we eliminate over-indexing of the notional static type array bounds.
831   // This makes it easy to determine if the getelementptr is "inbounds".
832   // Also, this helps GlobalOpt do SROA on GlobalVariables.
833   Type *Ty = Ptr->getType();
834   assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
835   SmallVector<Constant *, 32> NewIdxs;
836 
837   do {
838     if (!Ty->isStructTy()) {
839       if (Ty->isPointerTy()) {
840         // The only pointer indexing we'll do is on the first index of the GEP.
841         if (!NewIdxs.empty())
842           break;
843 
844         Ty = SrcElemTy;
845 
846         // Only handle pointers to sized types, not pointers to functions.
847         if (!Ty->isSized())
848           return nullptr;
849       } else if (auto *ATy = dyn_cast<SequentialType>(Ty)) {
850         Ty = ATy->getElementType();
851       } else {
852         // We've reached some non-indexable type.
853         break;
854       }
855 
856       // Determine which element of the array the offset points into.
857       APInt ElemSize(BitWidth, DL.getTypeAllocSize(Ty));
858       if (ElemSize == 0) {
859         // The element size is 0. This may be [0 x Ty]*, so just use a zero
860         // index for this level and proceed to the next level to see if it can
861         // accommodate the offset.
862         NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
863       } else {
864         // The element size is non-zero divide the offset by the element
865         // size (rounding down), to compute the index at this level.
866         bool Overflow;
867         APInt NewIdx = Offset.sdiv_ov(ElemSize, Overflow);
868         if (Overflow)
869           break;
870         Offset -= NewIdx * ElemSize;
871         NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
872       }
873     } else {
874       auto *STy = cast<StructType>(Ty);
875       // If we end up with an offset that isn't valid for this struct type, we
876       // can't re-form this GEP in a regular form, so bail out. The pointer
877       // operand likely went through casts that are necessary to make the GEP
878       // sensible.
879       const StructLayout &SL = *DL.getStructLayout(STy);
880       if (Offset.isNegative() || Offset.uge(SL.getSizeInBytes()))
881         break;
882 
883       // Determine which field of the struct the offset points into. The
884       // getZExtValue is fine as we've already ensured that the offset is
885       // within the range representable by the StructLayout API.
886       unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
887       NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
888                                          ElIdx));
889       Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
890       Ty = STy->getTypeAtIndex(ElIdx);
891     }
892   } while (Ty != ResElemTy);
893 
894   // If we haven't used up the entire offset by descending the static
895   // type, then the offset is pointing into the middle of an indivisible
896   // member, so we can't simplify it.
897   if (Offset != 0)
898     return nullptr;
899 
900   // Create a GEP.
901   Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs);
902   assert(C->getType()->getPointerElementType() == Ty &&
903          "Computed GetElementPtr has unexpected type!");
904 
905   // If we ended up indexing a member with a type that doesn't match
906   // the type of what the original indices indexed, add a cast.
907   if (Ty != ResElemTy)
908     C = FoldBitCast(C, ResTy, DL);
909 
910   return C;
911 }
912 
913 /// Attempt to constant fold an instruction with the
914 /// specified opcode and operands.  If successful, the constant result is
915 /// returned, if not, null is returned.  Note that this function can fail when
916 /// attempting to fold instructions like loads and stores, which have no
917 /// constant expression form.
918 ///
919 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
920 /// information, due to only being passed an opcode and operands. Constant
921 /// folding using this function strips this information.
922 ///
ConstantFoldInstOperandsImpl(const Value * InstOrCE,Type * DestTy,unsigned Opcode,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)923 Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, Type *DestTy,
924                                        unsigned Opcode,
925                                        ArrayRef<Constant *> Ops,
926                                        const DataLayout &DL,
927                                        const TargetLibraryInfo *TLI) {
928   // Handle easy binops first.
929   if (Instruction::isBinaryOp(Opcode))
930     return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
931 
932   if (Instruction::isCast(Opcode))
933     return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
934 
935   if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
936     if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
937       return C;
938 
939     return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(),
940                                           Ops[0], Ops.slice(1));
941   }
942 
943   switch (Opcode) {
944   default: return nullptr;
945   case Instruction::ICmp:
946   case Instruction::FCmp: llvm_unreachable("Invalid for compares");
947   case Instruction::Call:
948     if (auto *F = dyn_cast<Function>(Ops.back()))
949       if (canConstantFoldCallTo(F))
950         return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
951     return nullptr;
952   case Instruction::Select:
953     return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
954   case Instruction::ExtractElement:
955     return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
956   case Instruction::InsertElement:
957     return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
958   case Instruction::ShuffleVector:
959     return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
960   }
961 }
962 
963 } // end anonymous namespace
964 
965 //===----------------------------------------------------------------------===//
966 // Constant Folding public APIs
967 //===----------------------------------------------------------------------===//
968 
ConstantFoldInstruction(Instruction * I,const DataLayout & DL,const TargetLibraryInfo * TLI)969 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
970                                         const TargetLibraryInfo *TLI) {
971   // Handle PHI nodes quickly here...
972   if (auto *PN = dyn_cast<PHINode>(I)) {
973     Constant *CommonValue = nullptr;
974 
975     for (Value *Incoming : PN->incoming_values()) {
976       // If the incoming value is undef then skip it.  Note that while we could
977       // skip the value if it is equal to the phi node itself we choose not to
978       // because that would break the rule that constant folding only applies if
979       // all operands are constants.
980       if (isa<UndefValue>(Incoming))
981         continue;
982       // If the incoming value is not a constant, then give up.
983       auto *C = dyn_cast<Constant>(Incoming);
984       if (!C)
985         return nullptr;
986       // Fold the PHI's operands.
987       if (auto *NewC = dyn_cast<ConstantExpr>(C))
988         C = ConstantFoldConstantExpression(NewC, DL, TLI);
989       // If the incoming value is a different constant to
990       // the one we saw previously, then give up.
991       if (CommonValue && C != CommonValue)
992         return nullptr;
993       CommonValue = C;
994     }
995 
996 
997     // If we reach here, all incoming values are the same constant or undef.
998     return CommonValue ? CommonValue : UndefValue::get(PN->getType());
999   }
1000 
1001   // Scan the operand list, checking to see if they are all constants, if so,
1002   // hand off to ConstantFoldInstOperandsImpl.
1003   if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1004     return nullptr;
1005 
1006   SmallVector<Constant *, 8> Ops;
1007   for (const Use &OpU : I->operands()) {
1008     auto *Op = cast<Constant>(&OpU);
1009     // Fold the Instruction's operands.
1010     if (auto *NewCE = dyn_cast<ConstantExpr>(Op))
1011       Op = ConstantFoldConstantExpression(NewCE, DL, TLI);
1012 
1013     Ops.push_back(Op);
1014   }
1015 
1016   if (const auto *CI = dyn_cast<CmpInst>(I))
1017     return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1018                                            DL, TLI);
1019 
1020   if (const auto *LI = dyn_cast<LoadInst>(I))
1021     return ConstantFoldLoadInst(LI, DL);
1022 
1023   if (auto *IVI = dyn_cast<InsertValueInst>(I)) {
1024     return ConstantExpr::getInsertValue(
1025                                 cast<Constant>(IVI->getAggregateOperand()),
1026                                 cast<Constant>(IVI->getInsertedValueOperand()),
1027                                 IVI->getIndices());
1028   }
1029 
1030   if (auto *EVI = dyn_cast<ExtractValueInst>(I)) {
1031     return ConstantExpr::getExtractValue(
1032                                     cast<Constant>(EVI->getAggregateOperand()),
1033                                     EVI->getIndices());
1034   }
1035 
1036   return ConstantFoldInstOperands(I, Ops, DL, TLI);
1037 }
1038 
1039 namespace {
1040 
1041 Constant *
ConstantFoldConstantExpressionImpl(const ConstantExpr * CE,const DataLayout & DL,const TargetLibraryInfo * TLI,SmallPtrSetImpl<ConstantExpr * > & FoldedOps)1042 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout &DL,
1043                                    const TargetLibraryInfo *TLI,
1044                                    SmallPtrSetImpl<ConstantExpr *> &FoldedOps) {
1045   SmallVector<Constant *, 8> Ops;
1046   for (const Use &NewU : CE->operands()) {
1047     auto *NewC = cast<Constant>(&NewU);
1048     // Recursively fold the ConstantExpr's operands. If we have already folded
1049     // a ConstantExpr, we don't have to process it again.
1050     if (auto *NewCE = dyn_cast<ConstantExpr>(NewC)) {
1051       if (FoldedOps.insert(NewCE).second)
1052         NewC = ConstantFoldConstantExpressionImpl(NewCE, DL, TLI, FoldedOps);
1053     }
1054     Ops.push_back(NewC);
1055   }
1056 
1057   if (CE->isCompare())
1058     return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1059                                            DL, TLI);
1060 
1061   return ConstantFoldInstOperandsImpl(CE, CE->getType(), CE->getOpcode(), Ops,
1062                                       DL, TLI);
1063 }
1064 
1065 } // end anonymous namespace
1066 
ConstantFoldConstantExpression(const ConstantExpr * CE,const DataLayout & DL,const TargetLibraryInfo * TLI)1067 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
1068                                                const DataLayout &DL,
1069                                                const TargetLibraryInfo *TLI) {
1070   SmallPtrSet<ConstantExpr *, 4> FoldedOps;
1071   return ConstantFoldConstantExpressionImpl(CE, DL, TLI, FoldedOps);
1072 }
1073 
ConstantFoldInstOperands(Instruction * I,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)1074 Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1075                                          ArrayRef<Constant *> Ops,
1076                                          const DataLayout &DL,
1077                                          const TargetLibraryInfo *TLI) {
1078   return ConstantFoldInstOperandsImpl(I, I->getType(), I->getOpcode(), Ops, DL,
1079                                       TLI);
1080 }
1081 
ConstantFoldInstOperands(unsigned Opcode,Type * DestTy,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)1082 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
1083                                          ArrayRef<Constant *> Ops,
1084                                          const DataLayout &DL,
1085                                          const TargetLibraryInfo *TLI) {
1086   assert(Opcode != Instruction::GetElementPtr && "Invalid for GEPs");
1087   return ConstantFoldInstOperandsImpl(nullptr, DestTy, Opcode, Ops, DL, TLI);
1088 }
1089 
ConstantFoldCompareInstOperands(unsigned Predicate,Constant * Ops0,Constant * Ops1,const DataLayout & DL,const TargetLibraryInfo * TLI)1090 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1091                                                 Constant *Ops0, Constant *Ops1,
1092                                                 const DataLayout &DL,
1093                                                 const TargetLibraryInfo *TLI) {
1094   // fold: icmp (inttoptr x), null         -> icmp x, 0
1095   // fold: icmp (ptrtoint x), 0            -> icmp x, null
1096   // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1097   // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1098   //
1099   // FIXME: The following comment is out of data and the DataLayout is here now.
1100   // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1101   // around to know if bit truncation is happening.
1102   if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1103     if (Ops1->isNullValue()) {
1104       if (CE0->getOpcode() == Instruction::IntToPtr) {
1105         Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1106         // Convert the integer value to the right size to ensure we get the
1107         // proper extension or truncation.
1108         Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1109                                                    IntPtrTy, false);
1110         Constant *Null = Constant::getNullValue(C->getType());
1111         return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1112       }
1113 
1114       // Only do this transformation if the int is intptrty in size, otherwise
1115       // there is a truncation or extension that we aren't modeling.
1116       if (CE0->getOpcode() == Instruction::PtrToInt) {
1117         Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1118         if (CE0->getType() == IntPtrTy) {
1119           Constant *C = CE0->getOperand(0);
1120           Constant *Null = Constant::getNullValue(C->getType());
1121           return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1122         }
1123       }
1124     }
1125 
1126     if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1127       if (CE0->getOpcode() == CE1->getOpcode()) {
1128         if (CE0->getOpcode() == Instruction::IntToPtr) {
1129           Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1130 
1131           // Convert the integer value to the right size to ensure we get the
1132           // proper extension or truncation.
1133           Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1134                                                       IntPtrTy, false);
1135           Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1136                                                       IntPtrTy, false);
1137           return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1138         }
1139 
1140         // Only do this transformation if the int is intptrty in size, otherwise
1141         // there is a truncation or extension that we aren't modeling.
1142         if (CE0->getOpcode() == Instruction::PtrToInt) {
1143           Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1144           if (CE0->getType() == IntPtrTy &&
1145               CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1146             return ConstantFoldCompareInstOperands(
1147                 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1148           }
1149         }
1150       }
1151     }
1152 
1153     // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1154     // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1155     if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1156         CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1157       Constant *LHS = ConstantFoldCompareInstOperands(
1158           Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1159       Constant *RHS = ConstantFoldCompareInstOperands(
1160           Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1161       unsigned OpC =
1162         Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1163       return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1164     }
1165   }
1166 
1167   return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1168 }
1169 
ConstantFoldBinaryOpOperands(unsigned Opcode,Constant * LHS,Constant * RHS,const DataLayout & DL)1170 Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1171                                              Constant *RHS,
1172                                              const DataLayout &DL) {
1173   assert(Instruction::isBinaryOp(Opcode));
1174   if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1175     if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1176       return C;
1177 
1178   return ConstantExpr::get(Opcode, LHS, RHS);
1179 }
1180 
ConstantFoldCastOperand(unsigned Opcode,Constant * C,Type * DestTy,const DataLayout & DL)1181 Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1182                                         Type *DestTy, const DataLayout &DL) {
1183   assert(Instruction::isCast(Opcode));
1184   switch (Opcode) {
1185   default:
1186     llvm_unreachable("Missing case");
1187   case Instruction::PtrToInt:
1188     // If the input is a inttoptr, eliminate the pair.  This requires knowing
1189     // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1190     if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1191       if (CE->getOpcode() == Instruction::IntToPtr) {
1192         Constant *Input = CE->getOperand(0);
1193         unsigned InWidth = Input->getType()->getScalarSizeInBits();
1194         unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1195         if (PtrWidth < InWidth) {
1196           Constant *Mask =
1197             ConstantInt::get(CE->getContext(),
1198                              APInt::getLowBitsSet(InWidth, PtrWidth));
1199           Input = ConstantExpr::getAnd(Input, Mask);
1200         }
1201         // Do a zext or trunc to get to the dest size.
1202         return ConstantExpr::getIntegerCast(Input, DestTy, false);
1203       }
1204     }
1205     return ConstantExpr::getCast(Opcode, C, DestTy);
1206   case Instruction::IntToPtr:
1207     // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1208     // the int size is >= the ptr size and the address spaces are the same.
1209     // This requires knowing the width of a pointer, so it can't be done in
1210     // ConstantExpr::getCast.
1211     if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1212       if (CE->getOpcode() == Instruction::PtrToInt) {
1213         Constant *SrcPtr = CE->getOperand(0);
1214         unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1215         unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1216 
1217         if (MidIntSize >= SrcPtrSize) {
1218           unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1219           if (SrcAS == DestTy->getPointerAddressSpace())
1220             return FoldBitCast(CE->getOperand(0), DestTy, DL);
1221         }
1222       }
1223     }
1224 
1225     return ConstantExpr::getCast(Opcode, C, DestTy);
1226   case Instruction::Trunc:
1227   case Instruction::ZExt:
1228   case Instruction::SExt:
1229   case Instruction::FPTrunc:
1230   case Instruction::FPExt:
1231   case Instruction::UIToFP:
1232   case Instruction::SIToFP:
1233   case Instruction::FPToUI:
1234   case Instruction::FPToSI:
1235   case Instruction::AddrSpaceCast:
1236       return ConstantExpr::getCast(Opcode, C, DestTy);
1237   case Instruction::BitCast:
1238     return FoldBitCast(C, DestTy, DL);
1239   }
1240 }
1241 
ConstantFoldLoadThroughGEPConstantExpr(Constant * C,ConstantExpr * CE)1242 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1243                                                        ConstantExpr *CE) {
1244   if (!CE->getOperand(1)->isNullValue())
1245     return nullptr;  // Do not allow stepping over the value!
1246 
1247   // Loop over all of the operands, tracking down which value we are
1248   // addressing.
1249   for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1250     C = C->getAggregateElement(CE->getOperand(i));
1251     if (!C)
1252       return nullptr;
1253   }
1254   return C;
1255 }
1256 
1257 Constant *
ConstantFoldLoadThroughGEPIndices(Constant * C,ArrayRef<Constant * > Indices)1258 llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1259                                         ArrayRef<Constant *> Indices) {
1260   // Loop over all of the operands, tracking down which value we are
1261   // addressing.
1262   for (Constant *Index : Indices) {
1263     C = C->getAggregateElement(Index);
1264     if (!C)
1265       return nullptr;
1266   }
1267   return C;
1268 }
1269 
1270 //===----------------------------------------------------------------------===//
1271 //  Constant Folding for Calls
1272 //
1273 
canConstantFoldCallTo(const Function * F)1274 bool llvm::canConstantFoldCallTo(const Function *F) {
1275   switch (F->getIntrinsicID()) {
1276   case Intrinsic::fabs:
1277   case Intrinsic::minnum:
1278   case Intrinsic::maxnum:
1279   case Intrinsic::log:
1280   case Intrinsic::log2:
1281   case Intrinsic::log10:
1282   case Intrinsic::exp:
1283   case Intrinsic::exp2:
1284   case Intrinsic::floor:
1285   case Intrinsic::ceil:
1286   case Intrinsic::sqrt:
1287   case Intrinsic::sin:
1288   case Intrinsic::cos:
1289   case Intrinsic::trunc:
1290   case Intrinsic::rint:
1291   case Intrinsic::nearbyint:
1292   case Intrinsic::pow:
1293   case Intrinsic::powi:
1294   case Intrinsic::bswap:
1295   case Intrinsic::ctpop:
1296   case Intrinsic::ctlz:
1297   case Intrinsic::cttz:
1298   case Intrinsic::fma:
1299   case Intrinsic::fmuladd:
1300   case Intrinsic::copysign:
1301   case Intrinsic::round:
1302   case Intrinsic::masked_load:
1303   case Intrinsic::sadd_with_overflow:
1304   case Intrinsic::uadd_with_overflow:
1305   case Intrinsic::ssub_with_overflow:
1306   case Intrinsic::usub_with_overflow:
1307   case Intrinsic::smul_with_overflow:
1308   case Intrinsic::umul_with_overflow:
1309   case Intrinsic::convert_from_fp16:
1310   case Intrinsic::convert_to_fp16:
1311   case Intrinsic::bitreverse:
1312   case Intrinsic::x86_sse_cvtss2si:
1313   case Intrinsic::x86_sse_cvtss2si64:
1314   case Intrinsic::x86_sse_cvttss2si:
1315   case Intrinsic::x86_sse_cvttss2si64:
1316   case Intrinsic::x86_sse2_cvtsd2si:
1317   case Intrinsic::x86_sse2_cvtsd2si64:
1318   case Intrinsic::x86_sse2_cvttsd2si:
1319   case Intrinsic::x86_sse2_cvttsd2si64:
1320     return true;
1321   default:
1322     return false;
1323   case 0: break;
1324   }
1325 
1326   if (!F->hasName())
1327     return false;
1328   StringRef Name = F->getName();
1329 
1330   // In these cases, the check of the length is required.  We don't want to
1331   // return true for a name like "cos\0blah" which strcmp would return equal to
1332   // "cos", but has length 8.
1333   switch (Name[0]) {
1334   default:
1335     return false;
1336   case 'a':
1337     return Name == "acos" || Name == "asin" || Name == "atan" ||
1338            Name == "atan2" || Name == "acosf" || Name == "asinf" ||
1339            Name == "atanf" || Name == "atan2f";
1340   case 'c':
1341     return Name == "ceil" || Name == "cos" || Name == "cosh" ||
1342            Name == "ceilf" || Name == "cosf" || Name == "coshf";
1343   case 'e':
1344     return Name == "exp" || Name == "exp2" || Name == "expf" || Name == "exp2f";
1345   case 'f':
1346     return Name == "fabs" || Name == "floor" || Name == "fmod" ||
1347            Name == "fabsf" || Name == "floorf" || Name == "fmodf";
1348   case 'l':
1349     return Name == "log" || Name == "log10" || Name == "logf" ||
1350            Name == "log10f";
1351   case 'p':
1352     return Name == "pow" || Name == "powf";
1353   case 's':
1354     return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1355            Name == "sinf" || Name == "sinhf" || Name == "sqrtf";
1356   case 't':
1357     return Name == "tan" || Name == "tanh" || Name == "tanf" || Name == "tanhf";
1358   }
1359 }
1360 
1361 namespace {
1362 
GetConstantFoldFPValue(double V,Type * Ty)1363 Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1364   if (Ty->isHalfTy()) {
1365     APFloat APF(V);
1366     bool unused;
1367     APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1368     return ConstantFP::get(Ty->getContext(), APF);
1369   }
1370   if (Ty->isFloatTy())
1371     return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1372   if (Ty->isDoubleTy())
1373     return ConstantFP::get(Ty->getContext(), APFloat(V));
1374   llvm_unreachable("Can only constant fold half/float/double");
1375 }
1376 
1377 /// Clear the floating-point exception state.
llvm_fenv_clearexcept()1378 inline void llvm_fenv_clearexcept() {
1379 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1380   feclearexcept(FE_ALL_EXCEPT);
1381 #endif
1382   errno = 0;
1383 }
1384 
1385 /// Test if a floating-point exception was raised.
llvm_fenv_testexcept()1386 inline bool llvm_fenv_testexcept() {
1387   int errno_val = errno;
1388   if (errno_val == ERANGE || errno_val == EDOM)
1389     return true;
1390 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1391   if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1392     return true;
1393 #endif
1394   return false;
1395 }
1396 
ConstantFoldFP(double (* NativeFP)(double),double V,Type * Ty)1397 Constant *ConstantFoldFP(double (*NativeFP)(double), double V, Type *Ty) {
1398   llvm_fenv_clearexcept();
1399   V = NativeFP(V);
1400   if (llvm_fenv_testexcept()) {
1401     llvm_fenv_clearexcept();
1402     return nullptr;
1403   }
1404 
1405   return GetConstantFoldFPValue(V, Ty);
1406 }
1407 
ConstantFoldBinaryFP(double (* NativeFP)(double,double),double V,double W,Type * Ty)1408 Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), double V,
1409                                double W, Type *Ty) {
1410   llvm_fenv_clearexcept();
1411   V = NativeFP(V, W);
1412   if (llvm_fenv_testexcept()) {
1413     llvm_fenv_clearexcept();
1414     return nullptr;
1415   }
1416 
1417   return GetConstantFoldFPValue(V, Ty);
1418 }
1419 
1420 /// Attempt to fold an SSE floating point to integer conversion of a constant
1421 /// floating point. If roundTowardZero is false, the default IEEE rounding is
1422 /// used (toward nearest, ties to even). This matches the behavior of the
1423 /// non-truncating SSE instructions in the default rounding mode. The desired
1424 /// integer type Ty is used to select how many bits are available for the
1425 /// result. Returns null if the conversion cannot be performed, otherwise
1426 /// returns the Constant value resulting from the conversion.
ConstantFoldConvertToInt(const APFloat & Val,bool roundTowardZero,Type * Ty)1427 Constant *ConstantFoldConvertToInt(const APFloat &Val, bool roundTowardZero,
1428                                    Type *Ty) {
1429   // All of these conversion intrinsics form an integer of at most 64bits.
1430   unsigned ResultWidth = Ty->getIntegerBitWidth();
1431   assert(ResultWidth <= 64 &&
1432          "Can only constant fold conversions to 64 and 32 bit ints");
1433 
1434   uint64_t UIntVal;
1435   bool isExact = false;
1436   APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1437                                               : APFloat::rmNearestTiesToEven;
1438   APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1439                                                   /*isSigned=*/true, mode,
1440                                                   &isExact);
1441   if (status != APFloat::opOK && status != APFloat::opInexact)
1442     return nullptr;
1443   return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1444 }
1445 
getValueAsDouble(ConstantFP * Op)1446 double getValueAsDouble(ConstantFP *Op) {
1447   Type *Ty = Op->getType();
1448 
1449   if (Ty->isFloatTy())
1450     return Op->getValueAPF().convertToFloat();
1451 
1452   if (Ty->isDoubleTy())
1453     return Op->getValueAPF().convertToDouble();
1454 
1455   bool unused;
1456   APFloat APF = Op->getValueAPF();
1457   APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1458   return APF.convertToDouble();
1459 }
1460 
ConstantFoldScalarCall(StringRef Name,unsigned IntrinsicID,Type * Ty,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1461 Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID, Type *Ty,
1462                                  ArrayRef<Constant *> Operands,
1463                                  const TargetLibraryInfo *TLI) {
1464   if (Operands.size() == 1) {
1465     if (isa<UndefValue>(Operands[0])) {
1466       // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN
1467       if (IntrinsicID == Intrinsic::cos)
1468         return Constant::getNullValue(Ty);
1469     }
1470     if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1471       if (IntrinsicID == Intrinsic::convert_to_fp16) {
1472         APFloat Val(Op->getValueAPF());
1473 
1474         bool lost = false;
1475         Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1476 
1477         return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1478       }
1479 
1480       if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1481         return nullptr;
1482 
1483       if (IntrinsicID == Intrinsic::round) {
1484         APFloat V = Op->getValueAPF();
1485         V.roundToIntegral(APFloat::rmNearestTiesToAway);
1486         return ConstantFP::get(Ty->getContext(), V);
1487       }
1488 
1489       if (IntrinsicID == Intrinsic::floor) {
1490         APFloat V = Op->getValueAPF();
1491         V.roundToIntegral(APFloat::rmTowardNegative);
1492         return ConstantFP::get(Ty->getContext(), V);
1493       }
1494 
1495       if (IntrinsicID == Intrinsic::ceil) {
1496         APFloat V = Op->getValueAPF();
1497         V.roundToIntegral(APFloat::rmTowardPositive);
1498         return ConstantFP::get(Ty->getContext(), V);
1499       }
1500 
1501       if (IntrinsicID == Intrinsic::trunc) {
1502         APFloat V = Op->getValueAPF();
1503         V.roundToIntegral(APFloat::rmTowardZero);
1504         return ConstantFP::get(Ty->getContext(), V);
1505       }
1506 
1507       if (IntrinsicID == Intrinsic::rint) {
1508         APFloat V = Op->getValueAPF();
1509         V.roundToIntegral(APFloat::rmNearestTiesToEven);
1510         return ConstantFP::get(Ty->getContext(), V);
1511       }
1512 
1513       if (IntrinsicID == Intrinsic::nearbyint) {
1514         APFloat V = Op->getValueAPF();
1515         V.roundToIntegral(APFloat::rmNearestTiesToEven);
1516         return ConstantFP::get(Ty->getContext(), V);
1517       }
1518 
1519       /// We only fold functions with finite arguments. Folding NaN and inf is
1520       /// likely to be aborted with an exception anyway, and some host libms
1521       /// have known errors raising exceptions.
1522       if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1523         return nullptr;
1524 
1525       /// Currently APFloat versions of these functions do not exist, so we use
1526       /// the host native double versions.  Float versions are not called
1527       /// directly but for all these it is true (float)(f((double)arg)) ==
1528       /// f(arg).  Long double not supported yet.
1529       double V = getValueAsDouble(Op);
1530 
1531       switch (IntrinsicID) {
1532         default: break;
1533         case Intrinsic::fabs:
1534           return ConstantFoldFP(fabs, V, Ty);
1535         case Intrinsic::log2:
1536           return ConstantFoldFP(Log2, V, Ty);
1537         case Intrinsic::log:
1538           return ConstantFoldFP(log, V, Ty);
1539         case Intrinsic::log10:
1540           return ConstantFoldFP(log10, V, Ty);
1541         case Intrinsic::exp:
1542           return ConstantFoldFP(exp, V, Ty);
1543         case Intrinsic::exp2:
1544           return ConstantFoldFP(exp2, V, Ty);
1545         case Intrinsic::sin:
1546           return ConstantFoldFP(sin, V, Ty);
1547         case Intrinsic::cos:
1548           return ConstantFoldFP(cos, V, Ty);
1549       }
1550 
1551       if (!TLI)
1552         return nullptr;
1553 
1554       switch (Name[0]) {
1555       case 'a':
1556         if ((Name == "acos" && TLI->has(LibFunc::acos)) ||
1557             (Name == "acosf" && TLI->has(LibFunc::acosf)))
1558           return ConstantFoldFP(acos, V, Ty);
1559         else if ((Name == "asin" && TLI->has(LibFunc::asin)) ||
1560                  (Name == "asinf" && TLI->has(LibFunc::asinf)))
1561           return ConstantFoldFP(asin, V, Ty);
1562         else if ((Name == "atan" && TLI->has(LibFunc::atan)) ||
1563                  (Name == "atanf" && TLI->has(LibFunc::atanf)))
1564           return ConstantFoldFP(atan, V, Ty);
1565         break;
1566       case 'c':
1567         if ((Name == "ceil" && TLI->has(LibFunc::ceil)) ||
1568             (Name == "ceilf" && TLI->has(LibFunc::ceilf)))
1569           return ConstantFoldFP(ceil, V, Ty);
1570         else if ((Name == "cos" && TLI->has(LibFunc::cos)) ||
1571                  (Name == "cosf" && TLI->has(LibFunc::cosf)))
1572           return ConstantFoldFP(cos, V, Ty);
1573         else if ((Name == "cosh" && TLI->has(LibFunc::cosh)) ||
1574                  (Name == "coshf" && TLI->has(LibFunc::coshf)))
1575           return ConstantFoldFP(cosh, V, Ty);
1576         break;
1577       case 'e':
1578         if ((Name == "exp" && TLI->has(LibFunc::exp)) ||
1579             (Name == "expf" && TLI->has(LibFunc::expf)))
1580           return ConstantFoldFP(exp, V, Ty);
1581         if ((Name == "exp2" && TLI->has(LibFunc::exp2)) ||
1582             (Name == "exp2f" && TLI->has(LibFunc::exp2f)))
1583           // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1584           // C99 library.
1585           return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1586         break;
1587       case 'f':
1588         if ((Name == "fabs" && TLI->has(LibFunc::fabs)) ||
1589             (Name == "fabsf" && TLI->has(LibFunc::fabsf)))
1590           return ConstantFoldFP(fabs, V, Ty);
1591         else if ((Name == "floor" && TLI->has(LibFunc::floor)) ||
1592                  (Name == "floorf" && TLI->has(LibFunc::floorf)))
1593           return ConstantFoldFP(floor, V, Ty);
1594         break;
1595       case 'l':
1596         if ((Name == "log" && V > 0 && TLI->has(LibFunc::log)) ||
1597             (Name == "logf" && V > 0 && TLI->has(LibFunc::logf)))
1598           return ConstantFoldFP(log, V, Ty);
1599         else if ((Name == "log10" && V > 0 && TLI->has(LibFunc::log10)) ||
1600                  (Name == "log10f" && V > 0 && TLI->has(LibFunc::log10f)))
1601           return ConstantFoldFP(log10, V, Ty);
1602         else if (IntrinsicID == Intrinsic::sqrt &&
1603                  (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1604           if (V >= -0.0)
1605             return ConstantFoldFP(sqrt, V, Ty);
1606           else {
1607             // Unlike the sqrt definitions in C/C++, POSIX, and IEEE-754 - which
1608             // all guarantee or favor returning NaN - the square root of a
1609             // negative number is not defined for the LLVM sqrt intrinsic.
1610             // This is because the intrinsic should only be emitted in place of
1611             // libm's sqrt function when using "no-nans-fp-math".
1612             return UndefValue::get(Ty);
1613           }
1614         }
1615         break;
1616       case 's':
1617         if ((Name == "sin" && TLI->has(LibFunc::sin)) ||
1618             (Name == "sinf" && TLI->has(LibFunc::sinf)))
1619           return ConstantFoldFP(sin, V, Ty);
1620         else if ((Name == "sinh" && TLI->has(LibFunc::sinh)) ||
1621                  (Name == "sinhf" && TLI->has(LibFunc::sinhf)))
1622           return ConstantFoldFP(sinh, V, Ty);
1623         else if ((Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt)) ||
1624                  (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf)))
1625           return ConstantFoldFP(sqrt, V, Ty);
1626         break;
1627       case 't':
1628         if ((Name == "tan" && TLI->has(LibFunc::tan)) ||
1629             (Name == "tanf" && TLI->has(LibFunc::tanf)))
1630           return ConstantFoldFP(tan, V, Ty);
1631         else if ((Name == "tanh" && TLI->has(LibFunc::tanh)) ||
1632                  (Name == "tanhf" && TLI->has(LibFunc::tanhf)))
1633           return ConstantFoldFP(tanh, V, Ty);
1634         break;
1635       default:
1636         break;
1637       }
1638       return nullptr;
1639     }
1640 
1641     if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
1642       switch (IntrinsicID) {
1643       case Intrinsic::bswap:
1644         return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1645       case Intrinsic::ctpop:
1646         return ConstantInt::get(Ty, Op->getValue().countPopulation());
1647       case Intrinsic::bitreverse:
1648         return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
1649       case Intrinsic::convert_from_fp16: {
1650         APFloat Val(APFloat::IEEEhalf, Op->getValue());
1651 
1652         bool lost = false;
1653         APFloat::opStatus status = Val.convert(
1654             Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
1655 
1656         // Conversion is always precise.
1657         (void)status;
1658         assert(status == APFloat::opOK && !lost &&
1659                "Precision lost during fp16 constfolding");
1660 
1661         return ConstantFP::get(Ty->getContext(), Val);
1662       }
1663       default:
1664         return nullptr;
1665       }
1666     }
1667 
1668     // Support ConstantVector in case we have an Undef in the top.
1669     if (isa<ConstantVector>(Operands[0]) ||
1670         isa<ConstantDataVector>(Operands[0])) {
1671       auto *Op = cast<Constant>(Operands[0]);
1672       switch (IntrinsicID) {
1673       default: break;
1674       case Intrinsic::x86_sse_cvtss2si:
1675       case Intrinsic::x86_sse_cvtss2si64:
1676       case Intrinsic::x86_sse2_cvtsd2si:
1677       case Intrinsic::x86_sse2_cvtsd2si64:
1678         if (ConstantFP *FPOp =
1679               dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1680           return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1681                                           /*roundTowardZero=*/false, Ty);
1682       case Intrinsic::x86_sse_cvttss2si:
1683       case Intrinsic::x86_sse_cvttss2si64:
1684       case Intrinsic::x86_sse2_cvttsd2si:
1685       case Intrinsic::x86_sse2_cvttsd2si64:
1686         if (ConstantFP *FPOp =
1687               dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1688           return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1689                                           /*roundTowardZero=*/true, Ty);
1690       }
1691     }
1692 
1693     if (isa<UndefValue>(Operands[0])) {
1694       if (IntrinsicID == Intrinsic::bswap)
1695         return Operands[0];
1696       return nullptr;
1697     }
1698 
1699     return nullptr;
1700   }
1701 
1702   if (Operands.size() == 2) {
1703     if (auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1704       if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1705         return nullptr;
1706       double Op1V = getValueAsDouble(Op1);
1707 
1708       if (auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1709         if (Op2->getType() != Op1->getType())
1710           return nullptr;
1711 
1712         double Op2V = getValueAsDouble(Op2);
1713         if (IntrinsicID == Intrinsic::pow) {
1714           return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1715         }
1716         if (IntrinsicID == Intrinsic::copysign) {
1717           APFloat V1 = Op1->getValueAPF();
1718           const APFloat &V2 = Op2->getValueAPF();
1719           V1.copySign(V2);
1720           return ConstantFP::get(Ty->getContext(), V1);
1721         }
1722 
1723         if (IntrinsicID == Intrinsic::minnum) {
1724           const APFloat &C1 = Op1->getValueAPF();
1725           const APFloat &C2 = Op2->getValueAPF();
1726           return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
1727         }
1728 
1729         if (IntrinsicID == Intrinsic::maxnum) {
1730           const APFloat &C1 = Op1->getValueAPF();
1731           const APFloat &C2 = Op2->getValueAPF();
1732           return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
1733         }
1734 
1735         if (!TLI)
1736           return nullptr;
1737         if ((Name == "pow" && TLI->has(LibFunc::pow)) ||
1738             (Name == "powf" && TLI->has(LibFunc::powf)))
1739           return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1740         if ((Name == "fmod" && TLI->has(LibFunc::fmod)) ||
1741             (Name == "fmodf" && TLI->has(LibFunc::fmodf)))
1742           return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1743         if ((Name == "atan2" && TLI->has(LibFunc::atan2)) ||
1744             (Name == "atan2f" && TLI->has(LibFunc::atan2f)))
1745           return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1746       } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1747         if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1748           return ConstantFP::get(Ty->getContext(),
1749                                  APFloat((float)std::pow((float)Op1V,
1750                                                  (int)Op2C->getZExtValue())));
1751         if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1752           return ConstantFP::get(Ty->getContext(),
1753                                  APFloat((float)std::pow((float)Op1V,
1754                                                  (int)Op2C->getZExtValue())));
1755         if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1756           return ConstantFP::get(Ty->getContext(),
1757                                  APFloat((double)std::pow((double)Op1V,
1758                                                    (int)Op2C->getZExtValue())));
1759       }
1760       return nullptr;
1761     }
1762 
1763     if (auto *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1764       if (auto *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1765         switch (IntrinsicID) {
1766         default: break;
1767         case Intrinsic::sadd_with_overflow:
1768         case Intrinsic::uadd_with_overflow:
1769         case Intrinsic::ssub_with_overflow:
1770         case Intrinsic::usub_with_overflow:
1771         case Intrinsic::smul_with_overflow:
1772         case Intrinsic::umul_with_overflow: {
1773           APInt Res;
1774           bool Overflow;
1775           switch (IntrinsicID) {
1776           default: llvm_unreachable("Invalid case");
1777           case Intrinsic::sadd_with_overflow:
1778             Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1779             break;
1780           case Intrinsic::uadd_with_overflow:
1781             Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1782             break;
1783           case Intrinsic::ssub_with_overflow:
1784             Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1785             break;
1786           case Intrinsic::usub_with_overflow:
1787             Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1788             break;
1789           case Intrinsic::smul_with_overflow:
1790             Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1791             break;
1792           case Intrinsic::umul_with_overflow:
1793             Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1794             break;
1795           }
1796           Constant *Ops[] = {
1797             ConstantInt::get(Ty->getContext(), Res),
1798             ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1799           };
1800           return ConstantStruct::get(cast<StructType>(Ty), Ops);
1801         }
1802         case Intrinsic::cttz:
1803           if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1804             return UndefValue::get(Ty);
1805           return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1806         case Intrinsic::ctlz:
1807           if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1808             return UndefValue::get(Ty);
1809           return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1810         }
1811       }
1812 
1813       return nullptr;
1814     }
1815     return nullptr;
1816   }
1817 
1818   if (Operands.size() != 3)
1819     return nullptr;
1820 
1821   if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1822     if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1823       if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1824         switch (IntrinsicID) {
1825         default: break;
1826         case Intrinsic::fma:
1827         case Intrinsic::fmuladd: {
1828           APFloat V = Op1->getValueAPF();
1829           APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
1830                                                    Op3->getValueAPF(),
1831                                                    APFloat::rmNearestTiesToEven);
1832           if (s != APFloat::opInvalidOp)
1833             return ConstantFP::get(Ty->getContext(), V);
1834 
1835           return nullptr;
1836         }
1837         }
1838       }
1839     }
1840   }
1841 
1842   return nullptr;
1843 }
1844 
ConstantFoldVectorCall(StringRef Name,unsigned IntrinsicID,VectorType * VTy,ArrayRef<Constant * > Operands,const DataLayout & DL,const TargetLibraryInfo * TLI)1845 Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
1846                                  VectorType *VTy, ArrayRef<Constant *> Operands,
1847                                  const DataLayout &DL,
1848                                  const TargetLibraryInfo *TLI) {
1849   SmallVector<Constant *, 4> Result(VTy->getNumElements());
1850   SmallVector<Constant *, 4> Lane(Operands.size());
1851   Type *Ty = VTy->getElementType();
1852 
1853   if (IntrinsicID == Intrinsic::masked_load) {
1854     auto *SrcPtr = Operands[0];
1855     auto *Mask = Operands[2];
1856     auto *Passthru = Operands[3];
1857 
1858     Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, VTy, DL);
1859 
1860     SmallVector<Constant *, 32> NewElements;
1861     for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1862       auto *MaskElt = Mask->getAggregateElement(I);
1863       if (!MaskElt)
1864         break;
1865       auto *PassthruElt = Passthru->getAggregateElement(I);
1866       auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
1867       if (isa<UndefValue>(MaskElt)) {
1868         if (PassthruElt)
1869           NewElements.push_back(PassthruElt);
1870         else if (VecElt)
1871           NewElements.push_back(VecElt);
1872         else
1873           return nullptr;
1874       }
1875       if (MaskElt->isNullValue()) {
1876         if (!PassthruElt)
1877           return nullptr;
1878         NewElements.push_back(PassthruElt);
1879       } else if (MaskElt->isOneValue()) {
1880         if (!VecElt)
1881           return nullptr;
1882         NewElements.push_back(VecElt);
1883       } else {
1884         return nullptr;
1885       }
1886     }
1887     if (NewElements.size() != VTy->getNumElements())
1888       return nullptr;
1889     return ConstantVector::get(NewElements);
1890   }
1891 
1892   for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1893     // Gather a column of constants.
1894     for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
1895       Constant *Agg = Operands[J]->getAggregateElement(I);
1896       if (!Agg)
1897         return nullptr;
1898 
1899       Lane[J] = Agg;
1900     }
1901 
1902     // Use the regular scalar folding to simplify this column.
1903     Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI);
1904     if (!Folded)
1905       return nullptr;
1906     Result[I] = Folded;
1907   }
1908 
1909   return ConstantVector::get(Result);
1910 }
1911 
1912 } // end anonymous namespace
1913 
1914 Constant *
ConstantFoldCall(Function * F,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1915 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1916                        const TargetLibraryInfo *TLI) {
1917   if (!F->hasName())
1918     return nullptr;
1919   StringRef Name = F->getName();
1920 
1921   Type *Ty = F->getReturnType();
1922 
1923   if (auto *VTy = dyn_cast<VectorType>(Ty))
1924     return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
1925                                   F->getParent()->getDataLayout(), TLI);
1926 
1927   return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);
1928 }
1929