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/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringMap.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/Config/config.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include <cerrno>
38 #include <cmath>
39
40 #ifdef HAVE_FENV_H
41 #include <fenv.h>
42 #endif
43
44 using namespace llvm;
45
46 //===----------------------------------------------------------------------===//
47 // Constant Folding internal helper functions
48 //===----------------------------------------------------------------------===//
49
50 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
51 /// This always returns a non-null constant, but it may be a
52 /// ConstantExpr if unfoldable.
FoldBitCast(Constant * C,Type * DestTy,const DataLayout & DL)53 static Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
54 // Catch the obvious splat cases.
55 if (C->isNullValue() && !DestTy->isX86_MMXTy())
56 return Constant::getNullValue(DestTy);
57 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
58 !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
59 return Constant::getAllOnesValue(DestTy);
60
61 // Handle a vector->integer cast.
62 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
63 VectorType *VTy = dyn_cast<VectorType>(C->getType());
64 if (!VTy)
65 return ConstantExpr::getBitCast(C, DestTy);
66
67 unsigned NumSrcElts = VTy->getNumElements();
68 Type *SrcEltTy = VTy->getElementType();
69
70 // If the vector is a vector of floating point, convert it to vector of int
71 // to simplify things.
72 if (SrcEltTy->isFloatingPointTy()) {
73 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
74 Type *SrcIVTy =
75 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
76 // Ask IR to do the conversion now that #elts line up.
77 C = ConstantExpr::getBitCast(C, SrcIVTy);
78 }
79
80 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
81 if (!CDV)
82 return ConstantExpr::getBitCast(C, DestTy);
83
84 // Now that we know that the input value is a vector of integers, just shift
85 // and insert them into our result.
86 unsigned BitShift = DL.getTypeAllocSizeInBits(SrcEltTy);
87 APInt Result(IT->getBitWidth(), 0);
88 for (unsigned i = 0; i != NumSrcElts; ++i) {
89 Result <<= BitShift;
90 if (DL.isLittleEndian())
91 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
92 else
93 Result |= CDV->getElementAsInteger(i);
94 }
95
96 return ConstantInt::get(IT, Result);
97 }
98
99 // The code below only handles casts to vectors currently.
100 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
101 if (!DestVTy)
102 return ConstantExpr::getBitCast(C, DestTy);
103
104 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
105 // vector so the code below can handle it uniformly.
106 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
107 Constant *Ops = C; // don't take the address of C!
108 return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
109 }
110
111 // If this is a bitcast from constant vector -> vector, fold it.
112 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
113 return ConstantExpr::getBitCast(C, DestTy);
114
115 // If the element types match, IR can fold it.
116 unsigned NumDstElt = DestVTy->getNumElements();
117 unsigned NumSrcElt = C->getType()->getVectorNumElements();
118 if (NumDstElt == NumSrcElt)
119 return ConstantExpr::getBitCast(C, DestTy);
120
121 Type *SrcEltTy = C->getType()->getVectorElementType();
122 Type *DstEltTy = DestVTy->getElementType();
123
124 // Otherwise, we're changing the number of elements in a vector, which
125 // requires endianness information to do the right thing. For example,
126 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
127 // folds to (little endian):
128 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
129 // and to (big endian):
130 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
131
132 // First thing is first. We only want to think about integer here, so if
133 // we have something in FP form, recast it as integer.
134 if (DstEltTy->isFloatingPointTy()) {
135 // Fold to an vector of integers with same size as our FP type.
136 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
137 Type *DestIVTy =
138 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
139 // Recursively handle this integer conversion, if possible.
140 C = FoldBitCast(C, DestIVTy, DL);
141
142 // Finally, IR can handle this now that #elts line up.
143 return ConstantExpr::getBitCast(C, DestTy);
144 }
145
146 // Okay, we know the destination is integer, if the input is FP, convert
147 // it to integer first.
148 if (SrcEltTy->isFloatingPointTy()) {
149 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
150 Type *SrcIVTy =
151 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
152 // Ask IR to do the conversion now that #elts line up.
153 C = ConstantExpr::getBitCast(C, SrcIVTy);
154 // If IR wasn't able to fold it, bail out.
155 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
156 !isa<ConstantDataVector>(C))
157 return C;
158 }
159
160 // Now we know that the input and output vectors are both integer vectors
161 // of the same size, and that their #elements is not the same. Do the
162 // conversion here, which depends on whether the input or output has
163 // more elements.
164 bool isLittleEndian = DL.isLittleEndian();
165
166 SmallVector<Constant*, 32> Result;
167 if (NumDstElt < NumSrcElt) {
168 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
169 Constant *Zero = Constant::getNullValue(DstEltTy);
170 unsigned Ratio = NumSrcElt/NumDstElt;
171 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
172 unsigned SrcElt = 0;
173 for (unsigned i = 0; i != NumDstElt; ++i) {
174 // Build each element of the result.
175 Constant *Elt = Zero;
176 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
177 for (unsigned j = 0; j != Ratio; ++j) {
178 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
179 if (!Src) // Reject constantexpr elements.
180 return ConstantExpr::getBitCast(C, DestTy);
181
182 // Zero extend the element to the right size.
183 Src = ConstantExpr::getZExt(Src, Elt->getType());
184
185 // Shift it to the right place, depending on endianness.
186 Src = ConstantExpr::getShl(Src,
187 ConstantInt::get(Src->getType(), ShiftAmt));
188 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
189
190 // Mix it in.
191 Elt = ConstantExpr::getOr(Elt, Src);
192 }
193 Result.push_back(Elt);
194 }
195 return ConstantVector::get(Result);
196 }
197
198 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
199 unsigned Ratio = NumDstElt/NumSrcElt;
200 unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
201
202 // Loop over each source value, expanding into multiple results.
203 for (unsigned i = 0; i != NumSrcElt; ++i) {
204 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
205 if (!Src) // Reject constantexpr elements.
206 return ConstantExpr::getBitCast(C, DestTy);
207
208 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
209 for (unsigned j = 0; j != Ratio; ++j) {
210 // Shift the piece of the value into the right place, depending on
211 // endianness.
212 Constant *Elt = ConstantExpr::getLShr(Src,
213 ConstantInt::get(Src->getType(), ShiftAmt));
214 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
215
216 // Truncate the element to an integer with the same pointer size and
217 // convert the element back to a pointer using a inttoptr.
218 if (DstEltTy->isPointerTy()) {
219 IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
220 Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
221 Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
222 continue;
223 }
224
225 // Truncate and remember this piece.
226 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
227 }
228 }
229
230 return ConstantVector::get(Result);
231 }
232
233
234 /// If this constant is a constant offset from a global, return the global and
235 /// the constant. Because of constantexprs, this function is recursive.
IsConstantOffsetFromGlobal(Constant * C,GlobalValue * & GV,APInt & Offset,const DataLayout & DL)236 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
237 APInt &Offset, const DataLayout &DL) {
238 // Trivial case, constant is the global.
239 if ((GV = dyn_cast<GlobalValue>(C))) {
240 unsigned BitWidth = DL.getPointerTypeSizeInBits(GV->getType());
241 Offset = APInt(BitWidth, 0);
242 return true;
243 }
244
245 // Otherwise, if this isn't a constant expr, bail out.
246 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
247 if (!CE) return false;
248
249 // Look through ptr->int and ptr->ptr casts.
250 if (CE->getOpcode() == Instruction::PtrToInt ||
251 CE->getOpcode() == Instruction::BitCast ||
252 CE->getOpcode() == Instruction::AddrSpaceCast)
253 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL);
254
255 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
256 GEPOperator *GEP = dyn_cast<GEPOperator>(CE);
257 if (!GEP)
258 return false;
259
260 unsigned BitWidth = DL.getPointerTypeSizeInBits(GEP->getType());
261 APInt TmpOffset(BitWidth, 0);
262
263 // If the base isn't a global+constant, we aren't either.
264 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL))
265 return false;
266
267 // Otherwise, add any offset that our operands provide.
268 if (!GEP->accumulateConstantOffset(DL, TmpOffset))
269 return false;
270
271 Offset = TmpOffset;
272 return true;
273 }
274
275 /// Recursive helper to read bits out of global. C is the constant being copied
276 /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
277 /// results into and BytesLeft is the number of bytes left in
278 /// the CurPtr buffer. DL is the DataLayout.
ReadDataFromGlobal(Constant * C,uint64_t ByteOffset,unsigned char * CurPtr,unsigned BytesLeft,const DataLayout & DL)279 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
280 unsigned char *CurPtr, unsigned BytesLeft,
281 const DataLayout &DL) {
282 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
283 "Out of range access");
284
285 // If this element is zero or undefined, we can just return since *CurPtr is
286 // zero initialized.
287 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
288 return true;
289
290 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
291 if (CI->getBitWidth() > 64 ||
292 (CI->getBitWidth() & 7) != 0)
293 return false;
294
295 uint64_t Val = CI->getZExtValue();
296 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
297
298 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
299 int n = ByteOffset;
300 if (!DL.isLittleEndian())
301 n = IntBytes - n - 1;
302 CurPtr[i] = (unsigned char)(Val >> (n * 8));
303 ++ByteOffset;
304 }
305 return true;
306 }
307
308 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
309 if (CFP->getType()->isDoubleTy()) {
310 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
311 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
312 }
313 if (CFP->getType()->isFloatTy()){
314 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
315 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
316 }
317 if (CFP->getType()->isHalfTy()){
318 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
319 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
320 }
321 return false;
322 }
323
324 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
325 const StructLayout *SL = DL.getStructLayout(CS->getType());
326 unsigned Index = SL->getElementContainingOffset(ByteOffset);
327 uint64_t CurEltOffset = SL->getElementOffset(Index);
328 ByteOffset -= CurEltOffset;
329
330 while (1) {
331 // If the element access is to the element itself and not to tail padding,
332 // read the bytes from the element.
333 uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
334
335 if (ByteOffset < EltSize &&
336 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
337 BytesLeft, DL))
338 return false;
339
340 ++Index;
341
342 // Check to see if we read from the last struct element, if so we're done.
343 if (Index == CS->getType()->getNumElements())
344 return true;
345
346 // If we read all of the bytes we needed from this element we're done.
347 uint64_t NextEltOffset = SL->getElementOffset(Index);
348
349 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
350 return true;
351
352 // Move to the next element of the struct.
353 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
354 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
355 ByteOffset = 0;
356 CurEltOffset = NextEltOffset;
357 }
358 // not reached.
359 }
360
361 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
362 isa<ConstantDataSequential>(C)) {
363 Type *EltTy = C->getType()->getSequentialElementType();
364 uint64_t EltSize = DL.getTypeAllocSize(EltTy);
365 uint64_t Index = ByteOffset / EltSize;
366 uint64_t Offset = ByteOffset - Index * EltSize;
367 uint64_t NumElts;
368 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
369 NumElts = AT->getNumElements();
370 else
371 NumElts = C->getType()->getVectorNumElements();
372
373 for (; Index != NumElts; ++Index) {
374 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
375 BytesLeft, DL))
376 return false;
377
378 uint64_t BytesWritten = EltSize - Offset;
379 assert(BytesWritten <= EltSize && "Not indexing into this element?");
380 if (BytesWritten >= BytesLeft)
381 return true;
382
383 Offset = 0;
384 BytesLeft -= BytesWritten;
385 CurPtr += BytesWritten;
386 }
387 return true;
388 }
389
390 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
391 if (CE->getOpcode() == Instruction::IntToPtr &&
392 CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
393 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
394 BytesLeft, DL);
395 }
396 }
397
398 // Otherwise, unknown initializer type.
399 return false;
400 }
401
FoldReinterpretLoadFromConstPtr(Constant * C,const DataLayout & DL)402 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
403 const DataLayout &DL) {
404 PointerType *PTy = cast<PointerType>(C->getType());
405 Type *LoadTy = PTy->getElementType();
406 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
407
408 // If this isn't an integer load we can't fold it directly.
409 if (!IntType) {
410 unsigned AS = PTy->getAddressSpace();
411
412 // If this is a float/double load, we can try folding it as an int32/64 load
413 // and then bitcast the result. This can be useful for union cases. Note
414 // that address spaces don't matter here since we're not going to result in
415 // an actual new load.
416 Type *MapTy;
417 if (LoadTy->isHalfTy())
418 MapTy = Type::getInt16PtrTy(C->getContext(), AS);
419 else if (LoadTy->isFloatTy())
420 MapTy = Type::getInt32PtrTy(C->getContext(), AS);
421 else if (LoadTy->isDoubleTy())
422 MapTy = Type::getInt64PtrTy(C->getContext(), AS);
423 else if (LoadTy->isVectorTy()) {
424 MapTy = PointerType::getIntNPtrTy(C->getContext(),
425 DL.getTypeAllocSizeInBits(LoadTy), AS);
426 } else
427 return nullptr;
428
429 C = FoldBitCast(C, MapTy, DL);
430 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, DL))
431 return FoldBitCast(Res, LoadTy, DL);
432 return nullptr;
433 }
434
435 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
436 if (BytesLoaded > 32 || BytesLoaded == 0)
437 return nullptr;
438
439 GlobalValue *GVal;
440 APInt Offset;
441 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, DL))
442 return nullptr;
443
444 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
445 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
446 !GV->getInitializer()->getType()->isSized())
447 return nullptr;
448
449 // If we're loading off the beginning of the global, some bytes may be valid,
450 // but we don't try to handle this.
451 if (Offset.isNegative())
452 return nullptr;
453
454 // If we're not accessing anything in this constant, the result is undefined.
455 if (Offset.getZExtValue() >=
456 DL.getTypeAllocSize(GV->getInitializer()->getType()))
457 return UndefValue::get(IntType);
458
459 unsigned char RawBytes[32] = {0};
460 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
461 BytesLoaded, DL))
462 return nullptr;
463
464 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
465 if (DL.isLittleEndian()) {
466 ResultVal = RawBytes[BytesLoaded - 1];
467 for (unsigned i = 1; i != BytesLoaded; ++i) {
468 ResultVal <<= 8;
469 ResultVal |= RawBytes[BytesLoaded - 1 - i];
470 }
471 } else {
472 ResultVal = RawBytes[0];
473 for (unsigned i = 1; i != BytesLoaded; ++i) {
474 ResultVal <<= 8;
475 ResultVal |= RawBytes[i];
476 }
477 }
478
479 return ConstantInt::get(IntType->getContext(), ResultVal);
480 }
481
ConstantFoldLoadThroughBitcast(ConstantExpr * CE,const DataLayout & DL)482 static Constant *ConstantFoldLoadThroughBitcast(ConstantExpr *CE,
483 const DataLayout &DL) {
484 auto *DestPtrTy = dyn_cast<PointerType>(CE->getType());
485 if (!DestPtrTy)
486 return nullptr;
487 Type *DestTy = DestPtrTy->getElementType();
488
489 Constant *C = ConstantFoldLoadFromConstPtr(CE->getOperand(0), DL);
490 if (!C)
491 return nullptr;
492
493 do {
494 Type *SrcTy = C->getType();
495
496 // If the type sizes are the same and a cast is legal, just directly
497 // cast the constant.
498 if (DL.getTypeSizeInBits(DestTy) == DL.getTypeSizeInBits(SrcTy)) {
499 Instruction::CastOps Cast = Instruction::BitCast;
500 // If we are going from a pointer to int or vice versa, we spell the cast
501 // differently.
502 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
503 Cast = Instruction::IntToPtr;
504 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
505 Cast = Instruction::PtrToInt;
506
507 if (CastInst::castIsValid(Cast, C, DestTy))
508 return ConstantExpr::getCast(Cast, C, DestTy);
509 }
510
511 // If this isn't an aggregate type, there is nothing we can do to drill down
512 // and find a bitcastable constant.
513 if (!SrcTy->isAggregateType())
514 return nullptr;
515
516 // We're simulating a load through a pointer that was bitcast to point to
517 // a different type, so we can try to walk down through the initial
518 // elements of an aggregate to see if some part of th e aggregate is
519 // castable to implement the "load" semantic model.
520 C = C->getAggregateElement(0u);
521 } while (C);
522
523 return nullptr;
524 }
525
526 /// Return the value that a load from C would produce if it is constant and
527 /// determinable. If this is not determinable, return null.
ConstantFoldLoadFromConstPtr(Constant * C,const DataLayout & DL)528 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
529 const DataLayout &DL) {
530 // First, try the easy cases:
531 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
532 if (GV->isConstant() && GV->hasDefinitiveInitializer())
533 return GV->getInitializer();
534
535 // If the loaded value isn't a constant expr, we can't handle it.
536 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
537 if (!CE)
538 return nullptr;
539
540 if (CE->getOpcode() == Instruction::GetElementPtr) {
541 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0))) {
542 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
543 if (Constant *V =
544 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
545 return V;
546 }
547 }
548 }
549
550 if (CE->getOpcode() == Instruction::BitCast)
551 if (Constant *LoadedC = ConstantFoldLoadThroughBitcast(CE, DL))
552 return LoadedC;
553
554 // Instead of loading constant c string, use corresponding integer value
555 // directly if string length is small enough.
556 StringRef Str;
557 if (getConstantStringInfo(CE, Str) && !Str.empty()) {
558 unsigned StrLen = Str.size();
559 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
560 unsigned NumBits = Ty->getPrimitiveSizeInBits();
561 // Replace load with immediate integer if the result is an integer or fp
562 // value.
563 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
564 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
565 APInt StrVal(NumBits, 0);
566 APInt SingleChar(NumBits, 0);
567 if (DL.isLittleEndian()) {
568 for (signed i = StrLen-1; i >= 0; i--) {
569 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
570 StrVal = (StrVal << 8) | SingleChar;
571 }
572 } else {
573 for (unsigned i = 0; i < StrLen; i++) {
574 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
575 StrVal = (StrVal << 8) | SingleChar;
576 }
577 // Append NULL at the end.
578 SingleChar = 0;
579 StrVal = (StrVal << 8) | SingleChar;
580 }
581
582 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
583 if (Ty->isFloatingPointTy())
584 Res = ConstantExpr::getBitCast(Res, Ty);
585 return Res;
586 }
587 }
588
589 // If this load comes from anywhere in a constant global, and if the global
590 // is all undef or zero, we know what it loads.
591 if (GlobalVariable *GV =
592 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, DL))) {
593 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
594 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
595 if (GV->getInitializer()->isNullValue())
596 return Constant::getNullValue(ResTy);
597 if (isa<UndefValue>(GV->getInitializer()))
598 return UndefValue::get(ResTy);
599 }
600 }
601
602 // Try hard to fold loads from bitcasted strange and non-type-safe things.
603 return FoldReinterpretLoadFromConstPtr(CE, DL);
604 }
605
ConstantFoldLoadInst(const LoadInst * LI,const DataLayout & DL)606 static Constant *ConstantFoldLoadInst(const LoadInst *LI,
607 const DataLayout &DL) {
608 if (LI->isVolatile()) return nullptr;
609
610 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
611 return ConstantFoldLoadFromConstPtr(C, DL);
612
613 return nullptr;
614 }
615
616 /// One of Op0/Op1 is a constant expression.
617 /// Attempt to symbolically evaluate the result of a binary operator merging
618 /// these together. If target data info is available, it is provided as DL,
619 /// otherwise DL is null.
SymbolicallyEvaluateBinop(unsigned Opc,Constant * Op0,Constant * Op1,const DataLayout & DL)620 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
621 Constant *Op1,
622 const DataLayout &DL) {
623 // SROA
624
625 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
626 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
627 // bits.
628
629 if (Opc == Instruction::And) {
630 unsigned BitWidth = DL.getTypeSizeInBits(Op0->getType()->getScalarType());
631 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
632 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
633 computeKnownBits(Op0, KnownZero0, KnownOne0, DL);
634 computeKnownBits(Op1, KnownZero1, KnownOne1, DL);
635 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
636 // All the bits of Op0 that the 'and' could be masking are already zero.
637 return Op0;
638 }
639 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
640 // All the bits of Op1 that the 'and' could be masking are already zero.
641 return Op1;
642 }
643
644 APInt KnownZero = KnownZero0 | KnownZero1;
645 APInt KnownOne = KnownOne0 & KnownOne1;
646 if ((KnownZero | KnownOne).isAllOnesValue()) {
647 return ConstantInt::get(Op0->getType(), KnownOne);
648 }
649 }
650
651 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
652 // constant. This happens frequently when iterating over a global array.
653 if (Opc == Instruction::Sub) {
654 GlobalValue *GV1, *GV2;
655 APInt Offs1, Offs2;
656
657 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
658 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
659 unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
660
661 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
662 // PtrToInt may change the bitwidth so we have convert to the right size
663 // first.
664 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
665 Offs2.zextOrTrunc(OpSize));
666 }
667 }
668
669 return nullptr;
670 }
671
672 /// If array indices are not pointer-sized integers, explicitly cast them so
673 /// that they aren't implicitly casted by the getelementptr.
CastGEPIndices(Type * SrcTy,ArrayRef<Constant * > Ops,Type * ResultTy,const DataLayout & DL,const TargetLibraryInfo * TLI)674 static Constant *CastGEPIndices(Type *SrcTy, ArrayRef<Constant *> Ops,
675 Type *ResultTy, const DataLayout &DL,
676 const TargetLibraryInfo *TLI) {
677 Type *IntPtrTy = DL.getIntPtrType(ResultTy);
678
679 bool Any = false;
680 SmallVector<Constant*, 32> NewIdxs;
681 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
682 if ((i == 1 ||
683 !isa<StructType>(GetElementPtrInst::getIndexedType(
684 cast<PointerType>(Ops[0]->getType()->getScalarType())
685 ->getElementType(),
686 Ops.slice(1, i - 1)))) &&
687 Ops[i]->getType() != IntPtrTy) {
688 Any = true;
689 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
690 true,
691 IntPtrTy,
692 true),
693 Ops[i], IntPtrTy));
694 } else
695 NewIdxs.push_back(Ops[i]);
696 }
697
698 if (!Any)
699 return nullptr;
700
701 Constant *C = ConstantExpr::getGetElementPtr(SrcTy, Ops[0], NewIdxs);
702 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
703 if (Constant *Folded = ConstantFoldConstantExpression(CE, DL, TLI))
704 C = Folded;
705 }
706
707 return C;
708 }
709
710 /// Strip the pointer casts, but preserve the address space information.
StripPtrCastKeepAS(Constant * Ptr)711 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
712 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
713 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
714 Ptr = Ptr->stripPointerCasts();
715 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
716
717 // Preserve the address space number of the pointer.
718 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
719 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
720 OldPtrTy->getAddressSpace());
721 Ptr = ConstantExpr::getPointerCast(Ptr, NewPtrTy);
722 }
723 return Ptr;
724 }
725
726 /// If we can symbolically evaluate the GEP constant expression, do so.
SymbolicallyEvaluateGEP(Type * SrcTy,ArrayRef<Constant * > Ops,Type * ResultTy,const DataLayout & DL,const TargetLibraryInfo * TLI)727 static Constant *SymbolicallyEvaluateGEP(Type *SrcTy, ArrayRef<Constant *> Ops,
728 Type *ResultTy, const DataLayout &DL,
729 const TargetLibraryInfo *TLI) {
730 Constant *Ptr = Ops[0];
731 if (!Ptr->getType()->getPointerElementType()->isSized() ||
732 !Ptr->getType()->isPointerTy())
733 return nullptr;
734
735 Type *IntPtrTy = DL.getIntPtrType(Ptr->getType());
736 Type *ResultElementTy = ResultTy->getPointerElementType();
737
738 // If this is a constant expr gep that is effectively computing an
739 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
740 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
741 if (!isa<ConstantInt>(Ops[i])) {
742
743 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
744 // "inttoptr (sub (ptrtoint Ptr), V)"
745 if (Ops.size() == 2 && ResultElementTy->isIntegerTy(8)) {
746 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
747 assert((!CE || CE->getType() == IntPtrTy) &&
748 "CastGEPIndices didn't canonicalize index types!");
749 if (CE && CE->getOpcode() == Instruction::Sub &&
750 CE->getOperand(0)->isNullValue()) {
751 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
752 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
753 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
754 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
755 Res = ConstantFoldConstantExpression(ResCE, DL, TLI);
756 return Res;
757 }
758 }
759 return nullptr;
760 }
761
762 unsigned BitWidth = DL.getTypeSizeInBits(IntPtrTy);
763 APInt Offset =
764 APInt(BitWidth,
765 DL.getIndexedOffset(
766 Ptr->getType(),
767 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
768 Ptr = StripPtrCastKeepAS(Ptr);
769
770 // If this is a GEP of a GEP, fold it all into a single GEP.
771 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
772 SmallVector<Value *, 4> NestedOps(GEP->op_begin() + 1, GEP->op_end());
773
774 // Do not try the incorporate the sub-GEP if some index is not a number.
775 bool AllConstantInt = true;
776 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
777 if (!isa<ConstantInt>(NestedOps[i])) {
778 AllConstantInt = false;
779 break;
780 }
781 if (!AllConstantInt)
782 break;
783
784 Ptr = cast<Constant>(GEP->getOperand(0));
785 Offset += APInt(BitWidth, DL.getIndexedOffset(Ptr->getType(), NestedOps));
786 Ptr = StripPtrCastKeepAS(Ptr);
787 }
788
789 // If the base value for this address is a literal integer value, fold the
790 // getelementptr to the resulting integer value casted to the pointer type.
791 APInt BasePtr(BitWidth, 0);
792 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
793 if (CE->getOpcode() == Instruction::IntToPtr) {
794 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
795 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
796 }
797 }
798
799 if (Ptr->isNullValue() || BasePtr != 0) {
800 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
801 return ConstantExpr::getIntToPtr(C, ResultTy);
802 }
803
804 // Otherwise form a regular getelementptr. Recompute the indices so that
805 // we eliminate over-indexing of the notional static type array bounds.
806 // This makes it easy to determine if the getelementptr is "inbounds".
807 // Also, this helps GlobalOpt do SROA on GlobalVariables.
808 Type *Ty = Ptr->getType();
809 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
810 SmallVector<Constant *, 32> NewIdxs;
811
812 do {
813 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
814 if (ATy->isPointerTy()) {
815 // The only pointer indexing we'll do is on the first index of the GEP.
816 if (!NewIdxs.empty())
817 break;
818
819 // Only handle pointers to sized types, not pointers to functions.
820 if (!ATy->getElementType()->isSized())
821 return nullptr;
822 }
823
824 // Determine which element of the array the offset points into.
825 APInt ElemSize(BitWidth, DL.getTypeAllocSize(ATy->getElementType()));
826 if (ElemSize == 0)
827 // The element size is 0. This may be [0 x Ty]*, so just use a zero
828 // index for this level and proceed to the next level to see if it can
829 // accommodate the offset.
830 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
831 else {
832 // The element size is non-zero divide the offset by the element
833 // size (rounding down), to compute the index at this level.
834 APInt NewIdx = Offset.udiv(ElemSize);
835 Offset -= NewIdx * ElemSize;
836 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
837 }
838 Ty = ATy->getElementType();
839 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
840 // If we end up with an offset that isn't valid for this struct type, we
841 // can't re-form this GEP in a regular form, so bail out. The pointer
842 // operand likely went through casts that are necessary to make the GEP
843 // sensible.
844 const StructLayout &SL = *DL.getStructLayout(STy);
845 if (Offset.uge(SL.getSizeInBytes()))
846 break;
847
848 // Determine which field of the struct the offset points into. The
849 // getZExtValue is fine as we've already ensured that the offset is
850 // within the range representable by the StructLayout API.
851 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
852 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
853 ElIdx));
854 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
855 Ty = STy->getTypeAtIndex(ElIdx);
856 } else {
857 // We've reached some non-indexable type.
858 break;
859 }
860 } while (Ty != ResultElementTy);
861
862 // If we haven't used up the entire offset by descending the static
863 // type, then the offset is pointing into the middle of an indivisible
864 // member, so we can't simplify it.
865 if (Offset != 0)
866 return nullptr;
867
868 // Create a GEP.
869 Constant *C = ConstantExpr::getGetElementPtr(SrcTy, Ptr, NewIdxs);
870 assert(C->getType()->getPointerElementType() == Ty &&
871 "Computed GetElementPtr has unexpected type!");
872
873 // If we ended up indexing a member with a type that doesn't match
874 // the type of what the original indices indexed, add a cast.
875 if (Ty != ResultElementTy)
876 C = FoldBitCast(C, ResultTy, DL);
877
878 return C;
879 }
880
881
882
883 //===----------------------------------------------------------------------===//
884 // Constant Folding public APIs
885 //===----------------------------------------------------------------------===//
886
887 /// Try to constant fold the specified instruction.
888 /// If successful, the constant result is returned, if not, null is returned.
889 /// Note that this fails if not all of the operands are constant. Otherwise,
890 /// this function can only fail when attempting to fold instructions like loads
891 /// and stores, which have no constant expression form.
ConstantFoldInstruction(Instruction * I,const DataLayout & DL,const TargetLibraryInfo * TLI)892 Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
893 const TargetLibraryInfo *TLI) {
894 // Handle PHI nodes quickly here...
895 if (PHINode *PN = dyn_cast<PHINode>(I)) {
896 Constant *CommonValue = nullptr;
897
898 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
899 Value *Incoming = PN->getIncomingValue(i);
900 // If the incoming value is undef then skip it. Note that while we could
901 // skip the value if it is equal to the phi node itself we choose not to
902 // because that would break the rule that constant folding only applies if
903 // all operands are constants.
904 if (isa<UndefValue>(Incoming))
905 continue;
906 // If the incoming value is not a constant, then give up.
907 Constant *C = dyn_cast<Constant>(Incoming);
908 if (!C)
909 return nullptr;
910 // Fold the PHI's operands.
911 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
912 C = ConstantFoldConstantExpression(NewC, DL, TLI);
913 // If the incoming value is a different constant to
914 // the one we saw previously, then give up.
915 if (CommonValue && C != CommonValue)
916 return nullptr;
917 CommonValue = C;
918 }
919
920
921 // If we reach here, all incoming values are the same constant or undef.
922 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
923 }
924
925 // Scan the operand list, checking to see if they are all constants, if so,
926 // hand off to ConstantFoldInstOperands.
927 SmallVector<Constant*, 8> Ops;
928 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
929 Constant *Op = dyn_cast<Constant>(*i);
930 if (!Op)
931 return nullptr; // All operands not constant!
932
933 // Fold the Instruction's operands.
934 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
935 Op = ConstantFoldConstantExpression(NewCE, DL, TLI);
936
937 Ops.push_back(Op);
938 }
939
940 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
941 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
942 DL, TLI);
943
944 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
945 return ConstantFoldLoadInst(LI, DL);
946
947 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I)) {
948 return ConstantExpr::getInsertValue(
949 cast<Constant>(IVI->getAggregateOperand()),
950 cast<Constant>(IVI->getInsertedValueOperand()),
951 IVI->getIndices());
952 }
953
954 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I)) {
955 return ConstantExpr::getExtractValue(
956 cast<Constant>(EVI->getAggregateOperand()),
957 EVI->getIndices());
958 }
959
960 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, DL, TLI);
961 }
962
963 static Constant *
ConstantFoldConstantExpressionImpl(const ConstantExpr * CE,const DataLayout & DL,const TargetLibraryInfo * TLI,SmallPtrSetImpl<ConstantExpr * > & FoldedOps)964 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout &DL,
965 const TargetLibraryInfo *TLI,
966 SmallPtrSetImpl<ConstantExpr *> &FoldedOps) {
967 SmallVector<Constant *, 8> Ops;
968 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e;
969 ++i) {
970 Constant *NewC = cast<Constant>(*i);
971 // Recursively fold the ConstantExpr's operands. If we have already folded
972 // a ConstantExpr, we don't have to process it again.
973 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) {
974 if (FoldedOps.insert(NewCE).second)
975 NewC = ConstantFoldConstantExpressionImpl(NewCE, DL, TLI, FoldedOps);
976 }
977 Ops.push_back(NewC);
978 }
979
980 if (CE->isCompare())
981 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
982 DL, TLI);
983 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, DL, TLI);
984 }
985
986 /// Attempt to fold the constant expression
987 /// using the specified DataLayout. If successful, the constant result is
988 /// result is returned, if not, null is returned.
ConstantFoldConstantExpression(const ConstantExpr * CE,const DataLayout & DL,const TargetLibraryInfo * TLI)989 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
990 const DataLayout &DL,
991 const TargetLibraryInfo *TLI) {
992 SmallPtrSet<ConstantExpr *, 4> FoldedOps;
993 return ConstantFoldConstantExpressionImpl(CE, DL, TLI, FoldedOps);
994 }
995
996 /// Attempt to constant fold an instruction with the
997 /// specified opcode and operands. If successful, the constant result is
998 /// returned, if not, null is returned. Note that this function can fail when
999 /// attempting to fold instructions like loads and stores, which have no
1000 /// constant expression form.
1001 ///
1002 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
1003 /// information, due to only being passed an opcode and operands. Constant
1004 /// folding using this function strips this information.
1005 ///
ConstantFoldInstOperands(unsigned Opcode,Type * DestTy,ArrayRef<Constant * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI)1006 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
1007 ArrayRef<Constant *> Ops,
1008 const DataLayout &DL,
1009 const TargetLibraryInfo *TLI) {
1010 // Handle easy binops first.
1011 if (Instruction::isBinaryOp(Opcode)) {
1012 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1])) {
1013 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], DL))
1014 return C;
1015 }
1016
1017 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
1018 }
1019
1020 switch (Opcode) {
1021 default: return nullptr;
1022 case Instruction::ICmp:
1023 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
1024 case Instruction::Call:
1025 if (Function *F = dyn_cast<Function>(Ops.back()))
1026 if (canConstantFoldCallTo(F))
1027 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
1028 return nullptr;
1029 case Instruction::PtrToInt:
1030 // If the input is a inttoptr, eliminate the pair. This requires knowing
1031 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1032 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1033 if (CE->getOpcode() == Instruction::IntToPtr) {
1034 Constant *Input = CE->getOperand(0);
1035 unsigned InWidth = Input->getType()->getScalarSizeInBits();
1036 unsigned PtrWidth = DL.getPointerTypeSizeInBits(CE->getType());
1037 if (PtrWidth < InWidth) {
1038 Constant *Mask =
1039 ConstantInt::get(CE->getContext(),
1040 APInt::getLowBitsSet(InWidth, PtrWidth));
1041 Input = ConstantExpr::getAnd(Input, Mask);
1042 }
1043 // Do a zext or trunc to get to the dest size.
1044 return ConstantExpr::getIntegerCast(Input, DestTy, false);
1045 }
1046 }
1047 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1048 case Instruction::IntToPtr:
1049 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1050 // the int size is >= the ptr size and the address spaces are the same.
1051 // This requires knowing the width of a pointer, so it can't be done in
1052 // ConstantExpr::getCast.
1053 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
1054 if (CE->getOpcode() == Instruction::PtrToInt) {
1055 Constant *SrcPtr = CE->getOperand(0);
1056 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1057 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1058
1059 if (MidIntSize >= SrcPtrSize) {
1060 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1061 if (SrcAS == DestTy->getPointerAddressSpace())
1062 return FoldBitCast(CE->getOperand(0), DestTy, DL);
1063 }
1064 }
1065 }
1066
1067 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1068 case Instruction::Trunc:
1069 case Instruction::ZExt:
1070 case Instruction::SExt:
1071 case Instruction::FPTrunc:
1072 case Instruction::FPExt:
1073 case Instruction::UIToFP:
1074 case Instruction::SIToFP:
1075 case Instruction::FPToUI:
1076 case Instruction::FPToSI:
1077 case Instruction::AddrSpaceCast:
1078 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
1079 case Instruction::BitCast:
1080 return FoldBitCast(Ops[0], DestTy, DL);
1081 case Instruction::Select:
1082 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1083 case Instruction::ExtractElement:
1084 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1085 case Instruction::InsertElement:
1086 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1087 case Instruction::ShuffleVector:
1088 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1089 case Instruction::GetElementPtr: {
1090 Type *SrcTy = nullptr;
1091 if (Constant *C = CastGEPIndices(SrcTy, Ops, DestTy, DL, TLI))
1092 return C;
1093 if (Constant *C = SymbolicallyEvaluateGEP(SrcTy, Ops, DestTy, DL, TLI))
1094 return C;
1095
1096 return ConstantExpr::getGetElementPtr(SrcTy, Ops[0], Ops.slice(1));
1097 }
1098 }
1099 }
1100
1101 /// Attempt to constant fold a compare
1102 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
1103 /// returns a constant expression of the specified operands.
ConstantFoldCompareInstOperands(unsigned Predicate,Constant * Ops0,Constant * Ops1,const DataLayout & DL,const TargetLibraryInfo * TLI)1104 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1105 Constant *Ops0, Constant *Ops1,
1106 const DataLayout &DL,
1107 const TargetLibraryInfo *TLI) {
1108 // fold: icmp (inttoptr x), null -> icmp x, 0
1109 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1110 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1111 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1112 //
1113 // FIXME: The following comment is out of data and the DataLayout is here now.
1114 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1115 // around to know if bit truncation is happening.
1116 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1117 if (Ops1->isNullValue()) {
1118 if (CE0->getOpcode() == Instruction::IntToPtr) {
1119 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1120 // Convert the integer value to the right size to ensure we get the
1121 // proper extension or truncation.
1122 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1123 IntPtrTy, false);
1124 Constant *Null = Constant::getNullValue(C->getType());
1125 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1126 }
1127
1128 // Only do this transformation if the int is intptrty in size, otherwise
1129 // there is a truncation or extension that we aren't modeling.
1130 if (CE0->getOpcode() == Instruction::PtrToInt) {
1131 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1132 if (CE0->getType() == IntPtrTy) {
1133 Constant *C = CE0->getOperand(0);
1134 Constant *Null = Constant::getNullValue(C->getType());
1135 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1136 }
1137 }
1138 }
1139
1140 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1141 if (CE0->getOpcode() == CE1->getOpcode()) {
1142 if (CE0->getOpcode() == Instruction::IntToPtr) {
1143 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1144
1145 // Convert the integer value to the right size to ensure we get the
1146 // proper extension or truncation.
1147 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1148 IntPtrTy, false);
1149 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1150 IntPtrTy, false);
1151 return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1152 }
1153
1154 // Only do this transformation if the int is intptrty in size, otherwise
1155 // there is a truncation or extension that we aren't modeling.
1156 if (CE0->getOpcode() == Instruction::PtrToInt) {
1157 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1158 if (CE0->getType() == IntPtrTy &&
1159 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1160 return ConstantFoldCompareInstOperands(
1161 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1162 }
1163 }
1164 }
1165 }
1166
1167 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1168 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1169 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1170 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1171 Constant *LHS = ConstantFoldCompareInstOperands(
1172 Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1173 Constant *RHS = ConstantFoldCompareInstOperands(
1174 Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1175 unsigned OpC =
1176 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1177 Constant *Ops[] = { LHS, RHS };
1178 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, DL, TLI);
1179 }
1180 }
1181
1182 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1183 }
1184
1185
1186 /// Given a constant and a getelementptr constantexpr, return the constant value
1187 /// being addressed by the constant expression, or null if something is funny
1188 /// and we can't decide.
ConstantFoldLoadThroughGEPConstantExpr(Constant * C,ConstantExpr * CE)1189 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1190 ConstantExpr *CE) {
1191 if (!CE->getOperand(1)->isNullValue())
1192 return nullptr; // Do not allow stepping over the value!
1193
1194 // Loop over all of the operands, tracking down which value we are
1195 // addressing.
1196 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1197 C = C->getAggregateElement(CE->getOperand(i));
1198 if (!C)
1199 return nullptr;
1200 }
1201 return C;
1202 }
1203
1204 /// Given a constant and getelementptr indices (with an *implied* zero pointer
1205 /// index that is not in the list), return the constant value being addressed by
1206 /// a virtual load, or null if something is funny and we can't decide.
ConstantFoldLoadThroughGEPIndices(Constant * C,ArrayRef<Constant * > Indices)1207 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1208 ArrayRef<Constant*> Indices) {
1209 // Loop over all of the operands, tracking down which value we are
1210 // addressing.
1211 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1212 C = C->getAggregateElement(Indices[i]);
1213 if (!C)
1214 return nullptr;
1215 }
1216 return C;
1217 }
1218
1219
1220 //===----------------------------------------------------------------------===//
1221 // Constant Folding for Calls
1222 //
1223
1224 /// Return true if it's even possible to fold a call to the specified function.
canConstantFoldCallTo(const Function * F)1225 bool llvm::canConstantFoldCallTo(const Function *F) {
1226 switch (F->getIntrinsicID()) {
1227 case Intrinsic::fabs:
1228 case Intrinsic::minnum:
1229 case Intrinsic::maxnum:
1230 case Intrinsic::log:
1231 case Intrinsic::log2:
1232 case Intrinsic::log10:
1233 case Intrinsic::exp:
1234 case Intrinsic::exp2:
1235 case Intrinsic::floor:
1236 case Intrinsic::ceil:
1237 case Intrinsic::sqrt:
1238 case Intrinsic::pow:
1239 case Intrinsic::powi:
1240 case Intrinsic::bswap:
1241 case Intrinsic::ctpop:
1242 case Intrinsic::ctlz:
1243 case Intrinsic::cttz:
1244 case Intrinsic::fma:
1245 case Intrinsic::fmuladd:
1246 case Intrinsic::copysign:
1247 case Intrinsic::round:
1248 case Intrinsic::sadd_with_overflow:
1249 case Intrinsic::uadd_with_overflow:
1250 case Intrinsic::ssub_with_overflow:
1251 case Intrinsic::usub_with_overflow:
1252 case Intrinsic::smul_with_overflow:
1253 case Intrinsic::umul_with_overflow:
1254 case Intrinsic::convert_from_fp16:
1255 case Intrinsic::convert_to_fp16:
1256 case Intrinsic::x86_sse_cvtss2si:
1257 case Intrinsic::x86_sse_cvtss2si64:
1258 case Intrinsic::x86_sse_cvttss2si:
1259 case Intrinsic::x86_sse_cvttss2si64:
1260 case Intrinsic::x86_sse2_cvtsd2si:
1261 case Intrinsic::x86_sse2_cvtsd2si64:
1262 case Intrinsic::x86_sse2_cvttsd2si:
1263 case Intrinsic::x86_sse2_cvttsd2si64:
1264 return true;
1265 default:
1266 return false;
1267 case 0: break;
1268 }
1269
1270 if (!F->hasName())
1271 return false;
1272 StringRef Name = F->getName();
1273
1274 // In these cases, the check of the length is required. We don't want to
1275 // return true for a name like "cos\0blah" which strcmp would return equal to
1276 // "cos", but has length 8.
1277 switch (Name[0]) {
1278 default: return false;
1279 case 'a':
1280 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2";
1281 case 'c':
1282 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1283 case 'e':
1284 return Name == "exp" || Name == "exp2";
1285 case 'f':
1286 return Name == "fabs" || Name == "fmod" || Name == "floor";
1287 case 'l':
1288 return Name == "log" || Name == "log10";
1289 case 'p':
1290 return Name == "pow";
1291 case 's':
1292 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1293 Name == "sinf" || Name == "sqrtf";
1294 case 't':
1295 return Name == "tan" || Name == "tanh";
1296 }
1297 }
1298
GetConstantFoldFPValue(double V,Type * Ty)1299 static Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1300 if (Ty->isHalfTy()) {
1301 APFloat APF(V);
1302 bool unused;
1303 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1304 return ConstantFP::get(Ty->getContext(), APF);
1305 }
1306 if (Ty->isFloatTy())
1307 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1308 if (Ty->isDoubleTy())
1309 return ConstantFP::get(Ty->getContext(), APFloat(V));
1310 llvm_unreachable("Can only constant fold half/float/double");
1311
1312 }
1313
1314 namespace {
1315 /// Clear the floating-point exception state.
llvm_fenv_clearexcept()1316 static inline void llvm_fenv_clearexcept() {
1317 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT
1318 feclearexcept(FE_ALL_EXCEPT);
1319 #endif
1320 errno = 0;
1321 }
1322
1323 /// Test if a floating-point exception was raised.
llvm_fenv_testexcept()1324 static inline bool llvm_fenv_testexcept() {
1325 int errno_val = errno;
1326 if (errno_val == ERANGE || errno_val == EDOM)
1327 return true;
1328 #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
1329 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
1330 return true;
1331 #endif
1332 return false;
1333 }
1334 } // End namespace
1335
ConstantFoldFP(double (* NativeFP)(double),double V,Type * Ty)1336 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1337 Type *Ty) {
1338 llvm_fenv_clearexcept();
1339 V = NativeFP(V);
1340 if (llvm_fenv_testexcept()) {
1341 llvm_fenv_clearexcept();
1342 return nullptr;
1343 }
1344
1345 return GetConstantFoldFPValue(V, Ty);
1346 }
1347
ConstantFoldBinaryFP(double (* NativeFP)(double,double),double V,double W,Type * Ty)1348 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1349 double V, double W, Type *Ty) {
1350 llvm_fenv_clearexcept();
1351 V = NativeFP(V, W);
1352 if (llvm_fenv_testexcept()) {
1353 llvm_fenv_clearexcept();
1354 return nullptr;
1355 }
1356
1357 return GetConstantFoldFPValue(V, Ty);
1358 }
1359
1360 /// Attempt to fold an SSE floating point to integer conversion of a constant
1361 /// floating point. If roundTowardZero is false, the default IEEE rounding is
1362 /// used (toward nearest, ties to even). This matches the behavior of the
1363 /// non-truncating SSE instructions in the default rounding mode. The desired
1364 /// integer type Ty is used to select how many bits are available for the
1365 /// result. Returns null if the conversion cannot be performed, otherwise
1366 /// returns the Constant value resulting from the conversion.
ConstantFoldConvertToInt(const APFloat & Val,bool roundTowardZero,Type * Ty)1367 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1368 bool roundTowardZero, Type *Ty) {
1369 // All of these conversion intrinsics form an integer of at most 64bits.
1370 unsigned ResultWidth = Ty->getIntegerBitWidth();
1371 assert(ResultWidth <= 64 &&
1372 "Can only constant fold conversions to 64 and 32 bit ints");
1373
1374 uint64_t UIntVal;
1375 bool isExact = false;
1376 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1377 : APFloat::rmNearestTiesToEven;
1378 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1379 /*isSigned=*/true, mode,
1380 &isExact);
1381 if (status != APFloat::opOK && status != APFloat::opInexact)
1382 return nullptr;
1383 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1384 }
1385
getValueAsDouble(ConstantFP * Op)1386 static double getValueAsDouble(ConstantFP *Op) {
1387 Type *Ty = Op->getType();
1388
1389 if (Ty->isFloatTy())
1390 return Op->getValueAPF().convertToFloat();
1391
1392 if (Ty->isDoubleTy())
1393 return Op->getValueAPF().convertToDouble();
1394
1395 bool unused;
1396 APFloat APF = Op->getValueAPF();
1397 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1398 return APF.convertToDouble();
1399 }
1400
ConstantFoldScalarCall(StringRef Name,unsigned IntrinsicID,Type * Ty,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1401 static Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID,
1402 Type *Ty, ArrayRef<Constant *> Operands,
1403 const TargetLibraryInfo *TLI) {
1404 if (Operands.size() == 1) {
1405 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1406 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1407 APFloat Val(Op->getValueAPF());
1408
1409 bool lost = false;
1410 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1411
1412 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1413 }
1414
1415 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1416 return nullptr;
1417
1418 if (IntrinsicID == Intrinsic::round) {
1419 APFloat V = Op->getValueAPF();
1420 V.roundToIntegral(APFloat::rmNearestTiesToAway);
1421 return ConstantFP::get(Ty->getContext(), V);
1422 }
1423
1424 /// We only fold functions with finite arguments. Folding NaN and inf is
1425 /// likely to be aborted with an exception anyway, and some host libms
1426 /// have known errors raising exceptions.
1427 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1428 return nullptr;
1429
1430 /// Currently APFloat versions of these functions do not exist, so we use
1431 /// the host native double versions. Float versions are not called
1432 /// directly but for all these it is true (float)(f((double)arg)) ==
1433 /// f(arg). Long double not supported yet.
1434 double V = getValueAsDouble(Op);
1435
1436 switch (IntrinsicID) {
1437 default: break;
1438 case Intrinsic::fabs:
1439 return ConstantFoldFP(fabs, V, Ty);
1440 case Intrinsic::log2:
1441 return ConstantFoldFP(log2, V, Ty);
1442 case Intrinsic::log:
1443 return ConstantFoldFP(log, V, Ty);
1444 case Intrinsic::log10:
1445 return ConstantFoldFP(log10, V, Ty);
1446 case Intrinsic::exp:
1447 return ConstantFoldFP(exp, V, Ty);
1448 case Intrinsic::exp2:
1449 return ConstantFoldFP(exp2, V, Ty);
1450 case Intrinsic::floor:
1451 return ConstantFoldFP(floor, V, Ty);
1452 case Intrinsic::ceil:
1453 return ConstantFoldFP(ceil, V, Ty);
1454 }
1455
1456 if (!TLI)
1457 return nullptr;
1458
1459 switch (Name[0]) {
1460 case 'a':
1461 if (Name == "acos" && TLI->has(LibFunc::acos))
1462 return ConstantFoldFP(acos, V, Ty);
1463 else if (Name == "asin" && TLI->has(LibFunc::asin))
1464 return ConstantFoldFP(asin, V, Ty);
1465 else if (Name == "atan" && TLI->has(LibFunc::atan))
1466 return ConstantFoldFP(atan, V, Ty);
1467 break;
1468 case 'c':
1469 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1470 return ConstantFoldFP(ceil, V, Ty);
1471 else if (Name == "cos" && TLI->has(LibFunc::cos))
1472 return ConstantFoldFP(cos, V, Ty);
1473 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1474 return ConstantFoldFP(cosh, V, Ty);
1475 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1476 return ConstantFoldFP(cos, V, Ty);
1477 break;
1478 case 'e':
1479 if (Name == "exp" && TLI->has(LibFunc::exp))
1480 return ConstantFoldFP(exp, V, Ty);
1481
1482 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1483 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1484 // C99 library.
1485 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1486 }
1487 break;
1488 case 'f':
1489 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1490 return ConstantFoldFP(fabs, V, Ty);
1491 else if (Name == "floor" && TLI->has(LibFunc::floor))
1492 return ConstantFoldFP(floor, V, Ty);
1493 break;
1494 case 'l':
1495 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1496 return ConstantFoldFP(log, V, Ty);
1497 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1498 return ConstantFoldFP(log10, V, Ty);
1499 else if (IntrinsicID == Intrinsic::sqrt &&
1500 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1501 if (V >= -0.0)
1502 return ConstantFoldFP(sqrt, V, Ty);
1503 else {
1504 // Unlike the sqrt definitions in C/C++, POSIX, and IEEE-754 - which
1505 // all guarantee or favor returning NaN - the square root of a
1506 // negative number is not defined for the LLVM sqrt intrinsic.
1507 // This is because the intrinsic should only be emitted in place of
1508 // libm's sqrt function when using "no-nans-fp-math".
1509 return UndefValue::get(Ty);
1510 }
1511 }
1512 break;
1513 case 's':
1514 if (Name == "sin" && TLI->has(LibFunc::sin))
1515 return ConstantFoldFP(sin, V, Ty);
1516 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1517 return ConstantFoldFP(sinh, V, Ty);
1518 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1519 return ConstantFoldFP(sqrt, V, Ty);
1520 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1521 return ConstantFoldFP(sqrt, V, Ty);
1522 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1523 return ConstantFoldFP(sin, V, Ty);
1524 break;
1525 case 't':
1526 if (Name == "tan" && TLI->has(LibFunc::tan))
1527 return ConstantFoldFP(tan, V, Ty);
1528 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1529 return ConstantFoldFP(tanh, V, Ty);
1530 break;
1531 default:
1532 break;
1533 }
1534 return nullptr;
1535 }
1536
1537 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1538 switch (IntrinsicID) {
1539 case Intrinsic::bswap:
1540 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
1541 case Intrinsic::ctpop:
1542 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1543 case Intrinsic::convert_from_fp16: {
1544 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1545
1546 bool lost = false;
1547 APFloat::opStatus status =
1548 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1549
1550 // Conversion is always precise.
1551 (void)status;
1552 assert(status == APFloat::opOK && !lost &&
1553 "Precision lost during fp16 constfolding");
1554
1555 return ConstantFP::get(Ty->getContext(), Val);
1556 }
1557 default:
1558 return nullptr;
1559 }
1560 }
1561
1562 // Support ConstantVector in case we have an Undef in the top.
1563 if (isa<ConstantVector>(Operands[0]) ||
1564 isa<ConstantDataVector>(Operands[0])) {
1565 Constant *Op = cast<Constant>(Operands[0]);
1566 switch (IntrinsicID) {
1567 default: break;
1568 case Intrinsic::x86_sse_cvtss2si:
1569 case Intrinsic::x86_sse_cvtss2si64:
1570 case Intrinsic::x86_sse2_cvtsd2si:
1571 case Intrinsic::x86_sse2_cvtsd2si64:
1572 if (ConstantFP *FPOp =
1573 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1574 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1575 /*roundTowardZero=*/false, Ty);
1576 case Intrinsic::x86_sse_cvttss2si:
1577 case Intrinsic::x86_sse_cvttss2si64:
1578 case Intrinsic::x86_sse2_cvttsd2si:
1579 case Intrinsic::x86_sse2_cvttsd2si64:
1580 if (ConstantFP *FPOp =
1581 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1582 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1583 /*roundTowardZero=*/true, Ty);
1584 }
1585 }
1586
1587 if (isa<UndefValue>(Operands[0])) {
1588 if (IntrinsicID == Intrinsic::bswap)
1589 return Operands[0];
1590 return nullptr;
1591 }
1592
1593 return nullptr;
1594 }
1595
1596 if (Operands.size() == 2) {
1597 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1598 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1599 return nullptr;
1600 double Op1V = getValueAsDouble(Op1);
1601
1602 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1603 if (Op2->getType() != Op1->getType())
1604 return nullptr;
1605
1606 double Op2V = getValueAsDouble(Op2);
1607 if (IntrinsicID == Intrinsic::pow) {
1608 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1609 }
1610 if (IntrinsicID == Intrinsic::copysign) {
1611 APFloat V1 = Op1->getValueAPF();
1612 APFloat V2 = Op2->getValueAPF();
1613 V1.copySign(V2);
1614 return ConstantFP::get(Ty->getContext(), V1);
1615 }
1616
1617 if (IntrinsicID == Intrinsic::minnum) {
1618 const APFloat &C1 = Op1->getValueAPF();
1619 const APFloat &C2 = Op2->getValueAPF();
1620 return ConstantFP::get(Ty->getContext(), minnum(C1, C2));
1621 }
1622
1623 if (IntrinsicID == Intrinsic::maxnum) {
1624 const APFloat &C1 = Op1->getValueAPF();
1625 const APFloat &C2 = Op2->getValueAPF();
1626 return ConstantFP::get(Ty->getContext(), maxnum(C1, C2));
1627 }
1628
1629 if (!TLI)
1630 return nullptr;
1631 if (Name == "pow" && TLI->has(LibFunc::pow))
1632 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1633 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1634 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1635 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1636 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1637 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1638 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
1639 return ConstantFP::get(Ty->getContext(),
1640 APFloat((float)std::pow((float)Op1V,
1641 (int)Op2C->getZExtValue())));
1642 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
1643 return ConstantFP::get(Ty->getContext(),
1644 APFloat((float)std::pow((float)Op1V,
1645 (int)Op2C->getZExtValue())));
1646 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
1647 return ConstantFP::get(Ty->getContext(),
1648 APFloat((double)std::pow((double)Op1V,
1649 (int)Op2C->getZExtValue())));
1650 }
1651 return nullptr;
1652 }
1653
1654 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1655 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1656 switch (IntrinsicID) {
1657 default: break;
1658 case Intrinsic::sadd_with_overflow:
1659 case Intrinsic::uadd_with_overflow:
1660 case Intrinsic::ssub_with_overflow:
1661 case Intrinsic::usub_with_overflow:
1662 case Intrinsic::smul_with_overflow:
1663 case Intrinsic::umul_with_overflow: {
1664 APInt Res;
1665 bool Overflow;
1666 switch (IntrinsicID) {
1667 default: llvm_unreachable("Invalid case");
1668 case Intrinsic::sadd_with_overflow:
1669 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1670 break;
1671 case Intrinsic::uadd_with_overflow:
1672 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1673 break;
1674 case Intrinsic::ssub_with_overflow:
1675 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1676 break;
1677 case Intrinsic::usub_with_overflow:
1678 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1679 break;
1680 case Intrinsic::smul_with_overflow:
1681 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1682 break;
1683 case Intrinsic::umul_with_overflow:
1684 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1685 break;
1686 }
1687 Constant *Ops[] = {
1688 ConstantInt::get(Ty->getContext(), Res),
1689 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
1690 };
1691 return ConstantStruct::get(cast<StructType>(Ty), Ops);
1692 }
1693 case Intrinsic::cttz:
1694 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1695 return UndefValue::get(Ty);
1696 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1697 case Intrinsic::ctlz:
1698 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1699 return UndefValue::get(Ty);
1700 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1701 }
1702 }
1703
1704 return nullptr;
1705 }
1706 return nullptr;
1707 }
1708
1709 if (Operands.size() != 3)
1710 return nullptr;
1711
1712 if (const ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1713 if (const ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1714 if (const ConstantFP *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
1715 switch (IntrinsicID) {
1716 default: break;
1717 case Intrinsic::fma:
1718 case Intrinsic::fmuladd: {
1719 APFloat V = Op1->getValueAPF();
1720 APFloat::opStatus s = V.fusedMultiplyAdd(Op2->getValueAPF(),
1721 Op3->getValueAPF(),
1722 APFloat::rmNearestTiesToEven);
1723 if (s != APFloat::opInvalidOp)
1724 return ConstantFP::get(Ty->getContext(), V);
1725
1726 return nullptr;
1727 }
1728 }
1729 }
1730 }
1731 }
1732
1733 return nullptr;
1734 }
1735
ConstantFoldVectorCall(StringRef Name,unsigned IntrinsicID,VectorType * VTy,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1736 static Constant *ConstantFoldVectorCall(StringRef Name, unsigned IntrinsicID,
1737 VectorType *VTy,
1738 ArrayRef<Constant *> Operands,
1739 const TargetLibraryInfo *TLI) {
1740 SmallVector<Constant *, 4> Result(VTy->getNumElements());
1741 SmallVector<Constant *, 4> Lane(Operands.size());
1742 Type *Ty = VTy->getElementType();
1743
1744 for (unsigned I = 0, E = VTy->getNumElements(); I != E; ++I) {
1745 // Gather a column of constants.
1746 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
1747 Constant *Agg = Operands[J]->getAggregateElement(I);
1748 if (!Agg)
1749 return nullptr;
1750
1751 Lane[J] = Agg;
1752 }
1753
1754 // Use the regular scalar folding to simplify this column.
1755 Constant *Folded = ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI);
1756 if (!Folded)
1757 return nullptr;
1758 Result[I] = Folded;
1759 }
1760
1761 return ConstantVector::get(Result);
1762 }
1763
1764 /// Attempt to constant fold a call to the specified function
1765 /// with the specified arguments, returning null if unsuccessful.
1766 Constant *
ConstantFoldCall(Function * F,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1767 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1768 const TargetLibraryInfo *TLI) {
1769 if (!F->hasName())
1770 return nullptr;
1771 StringRef Name = F->getName();
1772
1773 Type *Ty = F->getReturnType();
1774
1775 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
1776 return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands, TLI);
1777
1778 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);
1779 }
1780