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/ValueTracking.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/FEnv.h"
34 #include "llvm/Support/GetElementPtrTypeIterator.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Target/TargetLibraryInfo.h"
37 #include <cerrno>
38 #include <cmath>
39 using namespace llvm;
40
41 //===----------------------------------------------------------------------===//
42 // Constant Folding internal helper functions
43 //===----------------------------------------------------------------------===//
44
45 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
46 /// DataLayout. This always returns a non-null constant, but it may be a
47 /// ConstantExpr if unfoldable.
FoldBitCast(Constant * C,Type * DestTy,const DataLayout & TD)48 static Constant *FoldBitCast(Constant *C, Type *DestTy,
49 const DataLayout &TD) {
50 // Catch the obvious splat cases.
51 if (C->isNullValue() && !DestTy->isX86_MMXTy())
52 return Constant::getNullValue(DestTy);
53 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
54 return Constant::getAllOnesValue(DestTy);
55
56 // Handle a vector->integer cast.
57 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
58 VectorType *VTy = dyn_cast<VectorType>(C->getType());
59 if (VTy == 0)
60 return ConstantExpr::getBitCast(C, DestTy);
61
62 unsigned NumSrcElts = VTy->getNumElements();
63 Type *SrcEltTy = VTy->getElementType();
64
65 // If the vector is a vector of floating point, convert it to vector of int
66 // to simplify things.
67 if (SrcEltTy->isFloatingPointTy()) {
68 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
69 Type *SrcIVTy =
70 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
71 // Ask IR to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
73 }
74
75 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
76 if (CDV == 0)
77 return ConstantExpr::getBitCast(C, DestTy);
78
79 // Now that we know that the input value is a vector of integers, just shift
80 // and insert them into our result.
81 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
82 APInt Result(IT->getBitWidth(), 0);
83 for (unsigned i = 0; i != NumSrcElts; ++i) {
84 Result <<= BitShift;
85 if (TD.isLittleEndian())
86 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
87 else
88 Result |= CDV->getElementAsInteger(i);
89 }
90
91 return ConstantInt::get(IT, Result);
92 }
93
94 // The code below only handles casts to vectors currently.
95 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
96 if (DestVTy == 0)
97 return ConstantExpr::getBitCast(C, DestTy);
98
99 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
100 // vector so the code below can handle it uniformly.
101 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
102 Constant *Ops = C; // don't take the address of C!
103 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
104 }
105
106 // If this is a bitcast from constant vector -> vector, fold it.
107 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
108 return ConstantExpr::getBitCast(C, DestTy);
109
110 // If the element types match, IR can fold it.
111 unsigned NumDstElt = DestVTy->getNumElements();
112 unsigned NumSrcElt = C->getType()->getVectorNumElements();
113 if (NumDstElt == NumSrcElt)
114 return ConstantExpr::getBitCast(C, DestTy);
115
116 Type *SrcEltTy = C->getType()->getVectorElementType();
117 Type *DstEltTy = DestVTy->getElementType();
118
119 // Otherwise, we're changing the number of elements in a vector, which
120 // requires endianness information to do the right thing. For example,
121 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
122 // folds to (little endian):
123 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
124 // and to (big endian):
125 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
126
127 // First thing is first. We only want to think about integer here, so if
128 // we have something in FP form, recast it as integer.
129 if (DstEltTy->isFloatingPointTy()) {
130 // Fold to an vector of integers with same size as our FP type.
131 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
132 Type *DestIVTy =
133 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
134 // Recursively handle this integer conversion, if possible.
135 C = FoldBitCast(C, DestIVTy, TD);
136
137 // Finally, IR can handle this now that #elts line up.
138 return ConstantExpr::getBitCast(C, DestTy);
139 }
140
141 // Okay, we know the destination is integer, if the input is FP, convert
142 // it to integer first.
143 if (SrcEltTy->isFloatingPointTy()) {
144 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
145 Type *SrcIVTy =
146 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
147 // Ask IR to do the conversion now that #elts line up.
148 C = ConstantExpr::getBitCast(C, SrcIVTy);
149 // If IR wasn't able to fold it, bail out.
150 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
151 !isa<ConstantDataVector>(C))
152 return C;
153 }
154
155 // Now we know that the input and output vectors are both integer vectors
156 // of the same size, and that their #elements is not the same. Do the
157 // conversion here, which depends on whether the input or output has
158 // more elements.
159 bool isLittleEndian = TD.isLittleEndian();
160
161 SmallVector<Constant*, 32> Result;
162 if (NumDstElt < NumSrcElt) {
163 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
164 Constant *Zero = Constant::getNullValue(DstEltTy);
165 unsigned Ratio = NumSrcElt/NumDstElt;
166 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
167 unsigned SrcElt = 0;
168 for (unsigned i = 0; i != NumDstElt; ++i) {
169 // Build each element of the result.
170 Constant *Elt = Zero;
171 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
172 for (unsigned j = 0; j != Ratio; ++j) {
173 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
174 if (!Src) // Reject constantexpr elements.
175 return ConstantExpr::getBitCast(C, DestTy);
176
177 // Zero extend the element to the right size.
178 Src = ConstantExpr::getZExt(Src, Elt->getType());
179
180 // Shift it to the right place, depending on endianness.
181 Src = ConstantExpr::getShl(Src,
182 ConstantInt::get(Src->getType(), ShiftAmt));
183 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
184
185 // Mix it in.
186 Elt = ConstantExpr::getOr(Elt, Src);
187 }
188 Result.push_back(Elt);
189 }
190 return ConstantVector::get(Result);
191 }
192
193 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
194 unsigned Ratio = NumDstElt/NumSrcElt;
195 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
196
197 // Loop over each source value, expanding into multiple results.
198 for (unsigned i = 0; i != NumSrcElt; ++i) {
199 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
200 if (!Src) // Reject constantexpr elements.
201 return ConstantExpr::getBitCast(C, DestTy);
202
203 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
204 for (unsigned j = 0; j != Ratio; ++j) {
205 // Shift the piece of the value into the right place, depending on
206 // endianness.
207 Constant *Elt = ConstantExpr::getLShr(Src,
208 ConstantInt::get(Src->getType(), ShiftAmt));
209 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
210
211 // Truncate and remember this piece.
212 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
213 }
214 }
215
216 return ConstantVector::get(Result);
217 }
218
219
220 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
221 /// from a global, return the global and the constant. Because of
222 /// constantexprs, this function is recursive.
IsConstantOffsetFromGlobal(Constant * C,GlobalValue * & GV,APInt & Offset,const DataLayout & TD)223 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
224 APInt &Offset, const DataLayout &TD) {
225 // Trivial case, constant is the global.
226 if ((GV = dyn_cast<GlobalValue>(C))) {
227 Offset.clearAllBits();
228 return true;
229 }
230
231 // Otherwise, if this isn't a constant expr, bail out.
232 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
233 if (!CE) return false;
234
235 // Look through ptr->int and ptr->ptr casts.
236 if (CE->getOpcode() == Instruction::PtrToInt ||
237 CE->getOpcode() == Instruction::BitCast)
238 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
239
240 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
241 if (GEPOperator *GEP = dyn_cast<GEPOperator>(CE)) {
242 // If the base isn't a global+constant, we aren't either.
243 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
244 return false;
245
246 // Otherwise, add any offset that our operands provide.
247 return GEP->accumulateConstantOffset(TD, Offset);
248 }
249
250 return false;
251 }
252
253 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
254 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
255 /// pointer to copy results into and BytesLeft is the number of bytes left in
256 /// the CurPtr buffer. TD is the target data.
ReadDataFromGlobal(Constant * C,uint64_t ByteOffset,unsigned char * CurPtr,unsigned BytesLeft,const DataLayout & TD)257 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
258 unsigned char *CurPtr, unsigned BytesLeft,
259 const DataLayout &TD) {
260 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
261 "Out of range access");
262
263 // If this element is zero or undefined, we can just return since *CurPtr is
264 // zero initialized.
265 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
266 return true;
267
268 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
269 if (CI->getBitWidth() > 64 ||
270 (CI->getBitWidth() & 7) != 0)
271 return false;
272
273 uint64_t Val = CI->getZExtValue();
274 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
275
276 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
277 int n = ByteOffset;
278 if (!TD.isLittleEndian())
279 n = IntBytes - n - 1;
280 CurPtr[i] = (unsigned char)(Val >> (n * 8));
281 ++ByteOffset;
282 }
283 return true;
284 }
285
286 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
287 if (CFP->getType()->isDoubleTy()) {
288 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
289 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
290 }
291 if (CFP->getType()->isFloatTy()){
292 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
293 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
294 }
295 if (CFP->getType()->isHalfTy()){
296 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), TD);
297 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
298 }
299 return false;
300 }
301
302 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
303 const StructLayout *SL = TD.getStructLayout(CS->getType());
304 unsigned Index = SL->getElementContainingOffset(ByteOffset);
305 uint64_t CurEltOffset = SL->getElementOffset(Index);
306 ByteOffset -= CurEltOffset;
307
308 while (1) {
309 // If the element access is to the element itself and not to tail padding,
310 // read the bytes from the element.
311 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
312
313 if (ByteOffset < EltSize &&
314 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
315 BytesLeft, TD))
316 return false;
317
318 ++Index;
319
320 // Check to see if we read from the last struct element, if so we're done.
321 if (Index == CS->getType()->getNumElements())
322 return true;
323
324 // If we read all of the bytes we needed from this element we're done.
325 uint64_t NextEltOffset = SL->getElementOffset(Index);
326
327 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
328 return true;
329
330 // Move to the next element of the struct.
331 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
332 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
333 ByteOffset = 0;
334 CurEltOffset = NextEltOffset;
335 }
336 // not reached.
337 }
338
339 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
340 isa<ConstantDataSequential>(C)) {
341 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
342 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
343 uint64_t Index = ByteOffset / EltSize;
344 uint64_t Offset = ByteOffset - Index * EltSize;
345 uint64_t NumElts;
346 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
347 NumElts = AT->getNumElements();
348 else
349 NumElts = cast<VectorType>(C->getType())->getNumElements();
350
351 for (; Index != NumElts; ++Index) {
352 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
353 BytesLeft, TD))
354 return false;
355
356 uint64_t BytesWritten = EltSize - Offset;
357 assert(BytesWritten <= EltSize && "Not indexing into this element?");
358 if (BytesWritten >= BytesLeft)
359 return true;
360
361 Offset = 0;
362 BytesLeft -= BytesWritten;
363 CurPtr += BytesWritten;
364 }
365 return true;
366 }
367
368 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
369 if (CE->getOpcode() == Instruction::IntToPtr &&
370 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
371 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
372 BytesLeft, TD);
373 }
374
375 // Otherwise, unknown initializer type.
376 return false;
377 }
378
FoldReinterpretLoadFromConstPtr(Constant * C,const DataLayout & TD)379 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
380 const DataLayout &TD) {
381 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
382 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
383
384 // If this isn't an integer load we can't fold it directly.
385 if (!IntType) {
386 // If this is a float/double load, we can try folding it as an int32/64 load
387 // and then bitcast the result. This can be useful for union cases. Note
388 // that address spaces don't matter here since we're not going to result in
389 // an actual new load.
390 Type *MapTy;
391 if (LoadTy->isHalfTy())
392 MapTy = Type::getInt16PtrTy(C->getContext());
393 else if (LoadTy->isFloatTy())
394 MapTy = Type::getInt32PtrTy(C->getContext());
395 else if (LoadTy->isDoubleTy())
396 MapTy = Type::getInt64PtrTy(C->getContext());
397 else if (LoadTy->isVectorTy()) {
398 MapTy = IntegerType::get(C->getContext(),
399 TD.getTypeAllocSizeInBits(LoadTy));
400 MapTy = PointerType::getUnqual(MapTy);
401 } else
402 return 0;
403
404 C = FoldBitCast(C, MapTy, TD);
405 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
406 return FoldBitCast(Res, LoadTy, TD);
407 return 0;
408 }
409
410 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
411 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
412
413 GlobalValue *GVal;
414 APInt Offset(TD.getPointerSizeInBits(), 0);
415 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
416 return 0;
417
418 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
419 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
420 !GV->getInitializer()->getType()->isSized())
421 return 0;
422
423 // If we're loading off the beginning of the global, some bytes may be valid,
424 // but we don't try to handle this.
425 if (Offset.isNegative()) return 0;
426
427 // If we're not accessing anything in this constant, the result is undefined.
428 if (Offset.getZExtValue() >=
429 TD.getTypeAllocSize(GV->getInitializer()->getType()))
430 return UndefValue::get(IntType);
431
432 unsigned char RawBytes[32] = {0};
433 if (!ReadDataFromGlobal(GV->getInitializer(), Offset.getZExtValue(), RawBytes,
434 BytesLoaded, TD))
435 return 0;
436
437 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
438 if (TD.isLittleEndian()) {
439 ResultVal = RawBytes[BytesLoaded - 1];
440 for (unsigned i = 1; i != BytesLoaded; ++i) {
441 ResultVal <<= 8;
442 ResultVal |= RawBytes[BytesLoaded-1-i];
443 }
444 } else {
445 ResultVal = RawBytes[0];
446 for (unsigned i = 1; i != BytesLoaded; ++i) {
447 ResultVal <<= 8;
448 ResultVal |= RawBytes[i];
449 }
450 }
451
452 return ConstantInt::get(IntType->getContext(), ResultVal);
453 }
454
455 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
456 /// produce if it is constant and determinable. If this is not determinable,
457 /// return null.
ConstantFoldLoadFromConstPtr(Constant * C,const DataLayout * TD)458 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
459 const DataLayout *TD) {
460 // First, try the easy cases:
461 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
462 if (GV->isConstant() && GV->hasDefinitiveInitializer())
463 return GV->getInitializer();
464
465 // If the loaded value isn't a constant expr, we can't handle it.
466 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
467 if (!CE) return 0;
468
469 if (CE->getOpcode() == Instruction::GetElementPtr) {
470 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
471 if (GV->isConstant() && GV->hasDefinitiveInitializer())
472 if (Constant *V =
473 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
474 return V;
475 }
476
477 // Instead of loading constant c string, use corresponding integer value
478 // directly if string length is small enough.
479 StringRef Str;
480 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
481 unsigned StrLen = Str.size();
482 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
483 unsigned NumBits = Ty->getPrimitiveSizeInBits();
484 // Replace load with immediate integer if the result is an integer or fp
485 // value.
486 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
487 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
488 APInt StrVal(NumBits, 0);
489 APInt SingleChar(NumBits, 0);
490 if (TD->isLittleEndian()) {
491 for (signed i = StrLen-1; i >= 0; i--) {
492 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
493 StrVal = (StrVal << 8) | SingleChar;
494 }
495 } else {
496 for (unsigned i = 0; i < StrLen; i++) {
497 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
498 StrVal = (StrVal << 8) | SingleChar;
499 }
500 // Append NULL at the end.
501 SingleChar = 0;
502 StrVal = (StrVal << 8) | SingleChar;
503 }
504
505 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
506 if (Ty->isFloatingPointTy())
507 Res = ConstantExpr::getBitCast(Res, Ty);
508 return Res;
509 }
510 }
511
512 // If this load comes from anywhere in a constant global, and if the global
513 // is all undef or zero, we know what it loads.
514 if (GlobalVariable *GV =
515 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
516 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
517 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
518 if (GV->getInitializer()->isNullValue())
519 return Constant::getNullValue(ResTy);
520 if (isa<UndefValue>(GV->getInitializer()))
521 return UndefValue::get(ResTy);
522 }
523 }
524
525 // Try hard to fold loads from bitcasted strange and non-type-safe things.
526 if (TD)
527 return FoldReinterpretLoadFromConstPtr(CE, *TD);
528 return 0;
529 }
530
ConstantFoldLoadInst(const LoadInst * LI,const DataLayout * TD)531 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const DataLayout *TD){
532 if (LI->isVolatile()) return 0;
533
534 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
535 return ConstantFoldLoadFromConstPtr(C, TD);
536
537 return 0;
538 }
539
540 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
541 /// Attempt to symbolically evaluate the result of a binary operator merging
542 /// these together. If target data info is available, it is provided as DL,
543 /// otherwise DL is null.
SymbolicallyEvaluateBinop(unsigned Opc,Constant * Op0,Constant * Op1,const DataLayout * DL)544 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
545 Constant *Op1, const DataLayout *DL){
546 // SROA
547
548 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
549 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
550 // bits.
551
552
553 if (Opc == Instruction::And && DL) {
554 unsigned BitWidth = DL->getTypeSizeInBits(Op0->getType()->getScalarType());
555 APInt KnownZero0(BitWidth, 0), KnownOne0(BitWidth, 0);
556 APInt KnownZero1(BitWidth, 0), KnownOne1(BitWidth, 0);
557 ComputeMaskedBits(Op0, KnownZero0, KnownOne0, DL);
558 ComputeMaskedBits(Op1, KnownZero1, KnownOne1, DL);
559 if ((KnownOne1 | KnownZero0).isAllOnesValue()) {
560 // All the bits of Op0 that the 'and' could be masking are already zero.
561 return Op0;
562 }
563 if ((KnownOne0 | KnownZero1).isAllOnesValue()) {
564 // All the bits of Op1 that the 'and' could be masking are already zero.
565 return Op1;
566 }
567
568 APInt KnownZero = KnownZero0 | KnownZero1;
569 APInt KnownOne = KnownOne0 & KnownOne1;
570 if ((KnownZero | KnownOne).isAllOnesValue()) {
571 return ConstantInt::get(Op0->getType(), KnownOne);
572 }
573 }
574
575 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
576 // constant. This happens frequently when iterating over a global array.
577 if (Opc == Instruction::Sub && DL) {
578 GlobalValue *GV1, *GV2;
579 unsigned PtrSize = DL->getPointerSizeInBits();
580 unsigned OpSize = DL->getTypeSizeInBits(Op0->getType());
581 APInt Offs1(PtrSize, 0), Offs2(PtrSize, 0);
582
583 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *DL))
584 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *DL) &&
585 GV1 == GV2) {
586 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
587 // PtrToInt may change the bitwidth so we have convert to the right size
588 // first.
589 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
590 Offs2.zextOrTrunc(OpSize));
591 }
592 }
593
594 return 0;
595 }
596
597 /// CastGEPIndices - If array indices are not pointer-sized integers,
598 /// explicitly cast them so that they aren't implicitly casted by the
599 /// getelementptr.
CastGEPIndices(ArrayRef<Constant * > Ops,Type * ResultTy,const DataLayout * TD,const TargetLibraryInfo * TLI)600 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
601 Type *ResultTy, const DataLayout *TD,
602 const TargetLibraryInfo *TLI) {
603 if (!TD) return 0;
604 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
605
606 bool Any = false;
607 SmallVector<Constant*, 32> NewIdxs;
608 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
609 if ((i == 1 ||
610 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
611 Ops.slice(1, i-1)))) &&
612 Ops[i]->getType() != IntPtrTy) {
613 Any = true;
614 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
615 true,
616 IntPtrTy,
617 true),
618 Ops[i], IntPtrTy));
619 } else
620 NewIdxs.push_back(Ops[i]);
621 }
622 if (!Any) return 0;
623
624 Constant *C =
625 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
626 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
627 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
628 C = Folded;
629 return C;
630 }
631
632 /// Strip the pointer casts, but preserve the address space information.
StripPtrCastKeepAS(Constant * Ptr)633 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
634 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
635 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
636 Ptr = cast<Constant>(Ptr->stripPointerCasts());
637 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
638
639 // Preserve the address space number of the pointer.
640 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
641 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
642 OldPtrTy->getAddressSpace());
643 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy);
644 }
645 return Ptr;
646 }
647
648 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
649 /// constant expression, do so.
SymbolicallyEvaluateGEP(ArrayRef<Constant * > Ops,Type * ResultTy,const DataLayout * TD,const TargetLibraryInfo * TLI)650 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
651 Type *ResultTy, const DataLayout *TD,
652 const TargetLibraryInfo *TLI) {
653 Constant *Ptr = Ops[0];
654 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
655 !Ptr->getType()->isPointerTy())
656 return 0;
657
658 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
659
660 // If this is a constant expr gep that is effectively computing an
661 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
662 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
663 if (!isa<ConstantInt>(Ops[i])) {
664
665 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
666 // "inttoptr (sub (ptrtoint Ptr), V)"
667 if (Ops.size() == 2 &&
668 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
669 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
670 assert((CE == 0 || CE->getType() == IntPtrTy) &&
671 "CastGEPIndices didn't canonicalize index types!");
672 if (CE && CE->getOpcode() == Instruction::Sub &&
673 CE->getOperand(0)->isNullValue()) {
674 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
675 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
676 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
677 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
678 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
679 return Res;
680 }
681 }
682 return 0;
683 }
684
685 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
686 APInt Offset =
687 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
688 makeArrayRef((Value *const*)
689 Ops.data() + 1,
690 Ops.size() - 1)));
691 Ptr = StripPtrCastKeepAS(Ptr);
692
693 // If this is a GEP of a GEP, fold it all into a single GEP.
694 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
695 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
696
697 // Do not try the incorporate the sub-GEP if some index is not a number.
698 bool AllConstantInt = true;
699 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
700 if (!isa<ConstantInt>(NestedOps[i])) {
701 AllConstantInt = false;
702 break;
703 }
704 if (!AllConstantInt)
705 break;
706
707 Ptr = cast<Constant>(GEP->getOperand(0));
708 Offset += APInt(BitWidth,
709 TD->getIndexedOffset(Ptr->getType(), NestedOps));
710 Ptr = StripPtrCastKeepAS(Ptr);
711 }
712
713 // If the base value for this address is a literal integer value, fold the
714 // getelementptr to the resulting integer value casted to the pointer type.
715 APInt BasePtr(BitWidth, 0);
716 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
717 if (CE->getOpcode() == Instruction::IntToPtr)
718 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
719 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
720 if (Ptr->isNullValue() || BasePtr != 0) {
721 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
722 return ConstantExpr::getIntToPtr(C, ResultTy);
723 }
724
725 // Otherwise form a regular getelementptr. Recompute the indices so that
726 // we eliminate over-indexing of the notional static type array bounds.
727 // This makes it easy to determine if the getelementptr is "inbounds".
728 // Also, this helps GlobalOpt do SROA on GlobalVariables.
729 Type *Ty = Ptr->getType();
730 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
731 SmallVector<Constant*, 32> NewIdxs;
732 do {
733 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
734 if (ATy->isPointerTy()) {
735 // The only pointer indexing we'll do is on the first index of the GEP.
736 if (!NewIdxs.empty())
737 break;
738
739 // Only handle pointers to sized types, not pointers to functions.
740 if (!ATy->getElementType()->isSized())
741 return 0;
742 }
743
744 // Determine which element of the array the offset points into.
745 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
746 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
747 if (ElemSize == 0)
748 // The element size is 0. This may be [0 x Ty]*, so just use a zero
749 // index for this level and proceed to the next level to see if it can
750 // accommodate the offset.
751 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
752 else {
753 // The element size is non-zero divide the offset by the element
754 // size (rounding down), to compute the index at this level.
755 APInt NewIdx = Offset.udiv(ElemSize);
756 Offset -= NewIdx * ElemSize;
757 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
758 }
759 Ty = ATy->getElementType();
760 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
761 // If we end up with an offset that isn't valid for this struct type, we
762 // can't re-form this GEP in a regular form, so bail out. The pointer
763 // operand likely went through casts that are necessary to make the GEP
764 // sensible.
765 const StructLayout &SL = *TD->getStructLayout(STy);
766 if (Offset.uge(SL.getSizeInBytes()))
767 break;
768
769 // Determine which field of the struct the offset points into. The
770 // getZExtValue is fine as we've already ensured that the offset is
771 // within the range representable by the StructLayout API.
772 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
773 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
774 ElIdx));
775 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
776 Ty = STy->getTypeAtIndex(ElIdx);
777 } else {
778 // We've reached some non-indexable type.
779 break;
780 }
781 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
782
783 // If we haven't used up the entire offset by descending the static
784 // type, then the offset is pointing into the middle of an indivisible
785 // member, so we can't simplify it.
786 if (Offset != 0)
787 return 0;
788
789 // Create a GEP.
790 Constant *C =
791 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
792 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
793 "Computed GetElementPtr has unexpected type!");
794
795 // If we ended up indexing a member with a type that doesn't match
796 // the type of what the original indices indexed, add a cast.
797 if (Ty != cast<PointerType>(ResultTy)->getElementType())
798 C = FoldBitCast(C, ResultTy, *TD);
799
800 return C;
801 }
802
803
804
805 //===----------------------------------------------------------------------===//
806 // Constant Folding public APIs
807 //===----------------------------------------------------------------------===//
808
809 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
810 /// If successful, the constant result is returned, if not, null is returned.
811 /// Note that this fails if not all of the operands are constant. Otherwise,
812 /// this function can only fail when attempting to fold instructions like loads
813 /// and stores, which have no constant expression form.
ConstantFoldInstruction(Instruction * I,const DataLayout * TD,const TargetLibraryInfo * TLI)814 Constant *llvm::ConstantFoldInstruction(Instruction *I,
815 const DataLayout *TD,
816 const TargetLibraryInfo *TLI) {
817 // Handle PHI nodes quickly here...
818 if (PHINode *PN = dyn_cast<PHINode>(I)) {
819 Constant *CommonValue = 0;
820
821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
822 Value *Incoming = PN->getIncomingValue(i);
823 // If the incoming value is undef then skip it. Note that while we could
824 // skip the value if it is equal to the phi node itself we choose not to
825 // because that would break the rule that constant folding only applies if
826 // all operands are constants.
827 if (isa<UndefValue>(Incoming))
828 continue;
829 // If the incoming value is not a constant, then give up.
830 Constant *C = dyn_cast<Constant>(Incoming);
831 if (!C)
832 return 0;
833 // Fold the PHI's operands.
834 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
835 C = ConstantFoldConstantExpression(NewC, TD, TLI);
836 // If the incoming value is a different constant to
837 // the one we saw previously, then give up.
838 if (CommonValue && C != CommonValue)
839 return 0;
840 CommonValue = C;
841 }
842
843
844 // If we reach here, all incoming values are the same constant or undef.
845 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
846 }
847
848 // Scan the operand list, checking to see if they are all constants, if so,
849 // hand off to ConstantFoldInstOperands.
850 SmallVector<Constant*, 8> Ops;
851 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
852 Constant *Op = dyn_cast<Constant>(*i);
853 if (!Op)
854 return 0; // All operands not constant!
855
856 // Fold the Instruction's operands.
857 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
858 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
859
860 Ops.push_back(Op);
861 }
862
863 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
864 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
865 TD, TLI);
866
867 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
868 return ConstantFoldLoadInst(LI, TD);
869
870 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
871 return ConstantExpr::getInsertValue(
872 cast<Constant>(IVI->getAggregateOperand()),
873 cast<Constant>(IVI->getInsertedValueOperand()),
874 IVI->getIndices());
875
876 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
877 return ConstantExpr::getExtractValue(
878 cast<Constant>(EVI->getAggregateOperand()),
879 EVI->getIndices());
880
881 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
882 }
883
884 static Constant *
ConstantFoldConstantExpressionImpl(const ConstantExpr * CE,const DataLayout * TD,const TargetLibraryInfo * TLI,SmallPtrSet<ConstantExpr *,4> & FoldedOps)885 ConstantFoldConstantExpressionImpl(const ConstantExpr *CE, const DataLayout *TD,
886 const TargetLibraryInfo *TLI,
887 SmallPtrSet<ConstantExpr *, 4> &FoldedOps) {
888 SmallVector<Constant *, 8> Ops;
889 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e;
890 ++i) {
891 Constant *NewC = cast<Constant>(*i);
892 // Recursively fold the ConstantExpr's operands. If we have already folded
893 // a ConstantExpr, we don't have to process it again.
894 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC)) {
895 if (FoldedOps.insert(NewCE))
896 NewC = ConstantFoldConstantExpressionImpl(NewCE, TD, TLI, FoldedOps);
897 }
898 Ops.push_back(NewC);
899 }
900
901 if (CE->isCompare())
902 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
903 TD, TLI);
904 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
905 }
906
907 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
908 /// using the specified DataLayout. If successful, the constant result is
909 /// result is returned, if not, null is returned.
ConstantFoldConstantExpression(const ConstantExpr * CE,const DataLayout * TD,const TargetLibraryInfo * TLI)910 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
911 const DataLayout *TD,
912 const TargetLibraryInfo *TLI) {
913 SmallPtrSet<ConstantExpr *, 4> FoldedOps;
914 return ConstantFoldConstantExpressionImpl(CE, TD, TLI, FoldedOps);
915 }
916
917 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
918 /// specified opcode and operands. If successful, the constant result is
919 /// returned, if not, null is returned. Note that this function can fail when
920 /// attempting to fold instructions like loads and stores, which have no
921 /// constant expression form.
922 ///
923 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
924 /// information, due to only being passed an opcode and operands. Constant
925 /// folding using this function strips this information.
926 ///
ConstantFoldInstOperands(unsigned Opcode,Type * DestTy,ArrayRef<Constant * > Ops,const DataLayout * TD,const TargetLibraryInfo * TLI)927 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
928 ArrayRef<Constant *> Ops,
929 const DataLayout *TD,
930 const TargetLibraryInfo *TLI) {
931 // Handle easy binops first.
932 if (Instruction::isBinaryOp(Opcode)) {
933 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
934 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
935 return C;
936
937 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
938 }
939
940 switch (Opcode) {
941 default: return 0;
942 case Instruction::ICmp:
943 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
944 case Instruction::Call:
945 if (Function *F = dyn_cast<Function>(Ops.back()))
946 if (canConstantFoldCallTo(F))
947 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
948 return 0;
949 case Instruction::PtrToInt:
950 // If the input is a inttoptr, eliminate the pair. This requires knowing
951 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
952 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
953 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
954 Constant *Input = CE->getOperand(0);
955 unsigned InWidth = Input->getType()->getScalarSizeInBits();
956 if (TD->getPointerSizeInBits() < InWidth) {
957 Constant *Mask =
958 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
959 TD->getPointerSizeInBits()));
960 Input = ConstantExpr::getAnd(Input, Mask);
961 }
962 // Do a zext or trunc to get to the dest size.
963 return ConstantExpr::getIntegerCast(Input, DestTy, false);
964 }
965 }
966 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
967 case Instruction::IntToPtr:
968 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
969 // the int size is >= the ptr size. This requires knowing the width of a
970 // pointer, so it can't be done in ConstantExpr::getCast.
971 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
972 if (TD &&
973 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
974 CE->getOpcode() == Instruction::PtrToInt)
975 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
976
977 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
978 case Instruction::Trunc:
979 case Instruction::ZExt:
980 case Instruction::SExt:
981 case Instruction::FPTrunc:
982 case Instruction::FPExt:
983 case Instruction::UIToFP:
984 case Instruction::SIToFP:
985 case Instruction::FPToUI:
986 case Instruction::FPToSI:
987 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
988 case Instruction::BitCast:
989 if (TD)
990 return FoldBitCast(Ops[0], DestTy, *TD);
991 return ConstantExpr::getBitCast(Ops[0], DestTy);
992 case Instruction::Select:
993 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
994 case Instruction::ExtractElement:
995 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
996 case Instruction::InsertElement:
997 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
998 case Instruction::ShuffleVector:
999 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
1000 case Instruction::GetElementPtr:
1001 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
1002 return C;
1003 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
1004 return C;
1005
1006 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
1007 }
1008 }
1009
1010 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
1011 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
1012 /// returns a constant expression of the specified operands.
1013 ///
ConstantFoldCompareInstOperands(unsigned Predicate,Constant * Ops0,Constant * Ops1,const DataLayout * TD,const TargetLibraryInfo * TLI)1014 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
1015 Constant *Ops0, Constant *Ops1,
1016 const DataLayout *TD,
1017 const TargetLibraryInfo *TLI) {
1018 // fold: icmp (inttoptr x), null -> icmp x, 0
1019 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1020 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1021 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1022 //
1023 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
1024 // around to know if bit truncation is happening.
1025 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1026 if (TD && Ops1->isNullValue()) {
1027 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1028 if (CE0->getOpcode() == Instruction::IntToPtr) {
1029 // Convert the integer value to the right size to ensure we get the
1030 // proper extension or truncation.
1031 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1032 IntPtrTy, false);
1033 Constant *Null = Constant::getNullValue(C->getType());
1034 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1035 }
1036
1037 // Only do this transformation if the int is intptrty in size, otherwise
1038 // there is a truncation or extension that we aren't modeling.
1039 if (CE0->getOpcode() == Instruction::PtrToInt &&
1040 CE0->getType() == IntPtrTy) {
1041 Constant *C = CE0->getOperand(0);
1042 Constant *Null = Constant::getNullValue(C->getType());
1043 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1044 }
1045 }
1046
1047 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1048 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1049 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1050
1051 if (CE0->getOpcode() == Instruction::IntToPtr) {
1052 // Convert the integer value to the right size to ensure we get the
1053 // proper extension or truncation.
1054 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1055 IntPtrTy, false);
1056 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1057 IntPtrTy, false);
1058 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1059 }
1060
1061 // Only do this transformation if the int is intptrty in size, otherwise
1062 // there is a truncation or extension that we aren't modeling.
1063 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1064 CE0->getType() == IntPtrTy &&
1065 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1066 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1067 CE1->getOperand(0), TD, TLI);
1068 }
1069 }
1070
1071 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1072 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1073 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1074 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1075 Constant *LHS =
1076 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1077 TD, TLI);
1078 Constant *RHS =
1079 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1080 TD, TLI);
1081 unsigned OpC =
1082 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1083 Constant *Ops[] = { LHS, RHS };
1084 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1085 }
1086 }
1087
1088 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1089 }
1090
1091
1092 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1093 /// getelementptr constantexpr, return the constant value being addressed by the
1094 /// constant expression, or null if something is funny and we can't decide.
ConstantFoldLoadThroughGEPConstantExpr(Constant * C,ConstantExpr * CE)1095 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1096 ConstantExpr *CE) {
1097 if (!CE->getOperand(1)->isNullValue())
1098 return 0; // Do not allow stepping over the value!
1099
1100 // Loop over all of the operands, tracking down which value we are
1101 // addressing.
1102 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1103 C = C->getAggregateElement(CE->getOperand(i));
1104 if (C == 0) return 0;
1105 }
1106 return C;
1107 }
1108
1109 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1110 /// indices (with an *implied* zero pointer index that is not in the list),
1111 /// return the constant value being addressed by a virtual load, or null if
1112 /// something is funny and we can't decide.
ConstantFoldLoadThroughGEPIndices(Constant * C,ArrayRef<Constant * > Indices)1113 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1114 ArrayRef<Constant*> Indices) {
1115 // Loop over all of the operands, tracking down which value we are
1116 // addressing.
1117 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1118 C = C->getAggregateElement(Indices[i]);
1119 if (C == 0) return 0;
1120 }
1121 return C;
1122 }
1123
1124
1125 //===----------------------------------------------------------------------===//
1126 // Constant Folding for Calls
1127 //
1128
1129 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1130 /// the specified function.
1131 bool
canConstantFoldCallTo(const Function * F)1132 llvm::canConstantFoldCallTo(const Function *F) {
1133 switch (F->getIntrinsicID()) {
1134 case Intrinsic::fabs:
1135 case Intrinsic::log:
1136 case Intrinsic::log2:
1137 case Intrinsic::log10:
1138 case Intrinsic::exp:
1139 case Intrinsic::exp2:
1140 case Intrinsic::floor:
1141 case Intrinsic::sqrt:
1142 case Intrinsic::pow:
1143 case Intrinsic::powi:
1144 case Intrinsic::bswap:
1145 case Intrinsic::ctpop:
1146 case Intrinsic::ctlz:
1147 case Intrinsic::cttz:
1148 case Intrinsic::sadd_with_overflow:
1149 case Intrinsic::uadd_with_overflow:
1150 case Intrinsic::ssub_with_overflow:
1151 case Intrinsic::usub_with_overflow:
1152 case Intrinsic::smul_with_overflow:
1153 case Intrinsic::umul_with_overflow:
1154 case Intrinsic::convert_from_fp16:
1155 case Intrinsic::convert_to_fp16:
1156 case Intrinsic::x86_sse_cvtss2si:
1157 case Intrinsic::x86_sse_cvtss2si64:
1158 case Intrinsic::x86_sse_cvttss2si:
1159 case Intrinsic::x86_sse_cvttss2si64:
1160 case Intrinsic::x86_sse2_cvtsd2si:
1161 case Intrinsic::x86_sse2_cvtsd2si64:
1162 case Intrinsic::x86_sse2_cvttsd2si:
1163 case Intrinsic::x86_sse2_cvttsd2si64:
1164 return true;
1165 default:
1166 return false;
1167 case 0: break;
1168 }
1169
1170 if (!F->hasName()) return false;
1171 StringRef Name = F->getName();
1172
1173 // In these cases, the check of the length is required. We don't want to
1174 // return true for a name like "cos\0blah" which strcmp would return equal to
1175 // "cos", but has length 8.
1176 switch (Name[0]) {
1177 default: return false;
1178 case 'a':
1179 return Name == "acos" || Name == "asin" || Name == "atan" || Name =="atan2";
1180 case 'c':
1181 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1182 case 'e':
1183 return Name == "exp" || Name == "exp2";
1184 case 'f':
1185 return Name == "fabs" || Name == "fmod" || Name == "floor";
1186 case 'l':
1187 return Name == "log" || Name == "log10";
1188 case 'p':
1189 return Name == "pow";
1190 case 's':
1191 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1192 Name == "sinf" || Name == "sqrtf";
1193 case 't':
1194 return Name == "tan" || Name == "tanh";
1195 }
1196 }
1197
ConstantFoldFP(double (* NativeFP)(double),double V,Type * Ty)1198 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1199 Type *Ty) {
1200 sys::llvm_fenv_clearexcept();
1201 V = NativeFP(V);
1202 if (sys::llvm_fenv_testexcept()) {
1203 sys::llvm_fenv_clearexcept();
1204 return 0;
1205 }
1206
1207 if (Ty->isHalfTy()) {
1208 APFloat APF(V);
1209 bool unused;
1210 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1211 return ConstantFP::get(Ty->getContext(), APF);
1212 }
1213 if (Ty->isFloatTy())
1214 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1215 if (Ty->isDoubleTy())
1216 return ConstantFP::get(Ty->getContext(), APFloat(V));
1217 llvm_unreachable("Can only constant fold half/float/double");
1218 }
1219
ConstantFoldBinaryFP(double (* NativeFP)(double,double),double V,double W,Type * Ty)1220 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1221 double V, double W, Type *Ty) {
1222 sys::llvm_fenv_clearexcept();
1223 V = NativeFP(V, W);
1224 if (sys::llvm_fenv_testexcept()) {
1225 sys::llvm_fenv_clearexcept();
1226 return 0;
1227 }
1228
1229 if (Ty->isHalfTy()) {
1230 APFloat APF(V);
1231 bool unused;
1232 APF.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &unused);
1233 return ConstantFP::get(Ty->getContext(), APF);
1234 }
1235 if (Ty->isFloatTy())
1236 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1237 if (Ty->isDoubleTy())
1238 return ConstantFP::get(Ty->getContext(), APFloat(V));
1239 llvm_unreachable("Can only constant fold half/float/double");
1240 }
1241
1242 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1243 /// conversion of a constant floating point. If roundTowardZero is false, the
1244 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1245 /// the behavior of the non-truncating SSE instructions in the default rounding
1246 /// mode. The desired integer type Ty is used to select how many bits are
1247 /// available for the result. Returns null if the conversion cannot be
1248 /// performed, otherwise returns the Constant value resulting from the
1249 /// conversion.
ConstantFoldConvertToInt(const APFloat & Val,bool roundTowardZero,Type * Ty)1250 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1251 bool roundTowardZero, Type *Ty) {
1252 // All of these conversion intrinsics form an integer of at most 64bits.
1253 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1254 assert(ResultWidth <= 64 &&
1255 "Can only constant fold conversions to 64 and 32 bit ints");
1256
1257 uint64_t UIntVal;
1258 bool isExact = false;
1259 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1260 : APFloat::rmNearestTiesToEven;
1261 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1262 /*isSigned=*/true, mode,
1263 &isExact);
1264 if (status != APFloat::opOK && status != APFloat::opInexact)
1265 return 0;
1266 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1267 }
1268
1269 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1270 /// with the specified arguments, returning null if unsuccessful.
1271 Constant *
ConstantFoldCall(Function * F,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1272 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1273 const TargetLibraryInfo *TLI) {
1274 if (!F->hasName()) return 0;
1275 StringRef Name = F->getName();
1276
1277 Type *Ty = F->getReturnType();
1278 if (Operands.size() == 1) {
1279 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1280 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1281 APFloat Val(Op->getValueAPF());
1282
1283 bool lost = false;
1284 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1285
1286 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1287 }
1288 if (!TLI)
1289 return 0;
1290
1291 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1292 return 0;
1293
1294 /// We only fold functions with finite arguments. Folding NaN and inf is
1295 /// likely to be aborted with an exception anyway, and some host libms
1296 /// have known errors raising exceptions.
1297 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1298 return 0;
1299
1300 /// Currently APFloat versions of these functions do not exist, so we use
1301 /// the host native double versions. Float versions are not called
1302 /// directly but for all these it is true (float)(f((double)arg)) ==
1303 /// f(arg). Long double not supported yet.
1304 double V;
1305 if (Ty->isFloatTy())
1306 V = Op->getValueAPF().convertToFloat();
1307 else if (Ty->isDoubleTy())
1308 V = Op->getValueAPF().convertToDouble();
1309 else {
1310 bool unused;
1311 APFloat APF = Op->getValueAPF();
1312 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1313 V = APF.convertToDouble();
1314 }
1315
1316 switch (F->getIntrinsicID()) {
1317 default: break;
1318 case Intrinsic::fabs:
1319 return ConstantFoldFP(fabs, V, Ty);
1320 #if HAVE_LOG2
1321 case Intrinsic::log2:
1322 return ConstantFoldFP(log2, V, Ty);
1323 #endif
1324 #if HAVE_LOG
1325 case Intrinsic::log:
1326 return ConstantFoldFP(log, V, Ty);
1327 #endif
1328 #if HAVE_LOG10
1329 case Intrinsic::log10:
1330 return ConstantFoldFP(log10, V, Ty);
1331 #endif
1332 #if HAVE_EXP
1333 case Intrinsic::exp:
1334 return ConstantFoldFP(exp, V, Ty);
1335 #endif
1336 #if HAVE_EXP2
1337 case Intrinsic::exp2:
1338 return ConstantFoldFP(exp2, V, Ty);
1339 #endif
1340 case Intrinsic::floor:
1341 return ConstantFoldFP(floor, V, Ty);
1342 }
1343
1344 switch (Name[0]) {
1345 case 'a':
1346 if (Name == "acos" && TLI->has(LibFunc::acos))
1347 return ConstantFoldFP(acos, V, Ty);
1348 else if (Name == "asin" && TLI->has(LibFunc::asin))
1349 return ConstantFoldFP(asin, V, Ty);
1350 else if (Name == "atan" && TLI->has(LibFunc::atan))
1351 return ConstantFoldFP(atan, V, Ty);
1352 break;
1353 case 'c':
1354 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1355 return ConstantFoldFP(ceil, V, Ty);
1356 else if (Name == "cos" && TLI->has(LibFunc::cos))
1357 return ConstantFoldFP(cos, V, Ty);
1358 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1359 return ConstantFoldFP(cosh, V, Ty);
1360 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1361 return ConstantFoldFP(cos, V, Ty);
1362 break;
1363 case 'e':
1364 if (Name == "exp" && TLI->has(LibFunc::exp))
1365 return ConstantFoldFP(exp, V, Ty);
1366
1367 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1368 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1369 // C99 library.
1370 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1371 }
1372 break;
1373 case 'f':
1374 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1375 return ConstantFoldFP(fabs, V, Ty);
1376 else if (Name == "floor" && TLI->has(LibFunc::floor))
1377 return ConstantFoldFP(floor, V, Ty);
1378 break;
1379 case 'l':
1380 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1381 return ConstantFoldFP(log, V, Ty);
1382 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1383 return ConstantFoldFP(log10, V, Ty);
1384 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1385 (Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())) {
1386 if (V >= -0.0)
1387 return ConstantFoldFP(sqrt, V, Ty);
1388 else // Undefined
1389 return Constant::getNullValue(Ty);
1390 }
1391 break;
1392 case 's':
1393 if (Name == "sin" && TLI->has(LibFunc::sin))
1394 return ConstantFoldFP(sin, V, Ty);
1395 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1396 return ConstantFoldFP(sinh, V, Ty);
1397 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1398 return ConstantFoldFP(sqrt, V, Ty);
1399 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1400 return ConstantFoldFP(sqrt, V, Ty);
1401 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1402 return ConstantFoldFP(sin, V, Ty);
1403 break;
1404 case 't':
1405 if (Name == "tan" && TLI->has(LibFunc::tan))
1406 return ConstantFoldFP(tan, V, Ty);
1407 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1408 return ConstantFoldFP(tanh, V, Ty);
1409 break;
1410 default:
1411 break;
1412 }
1413 return 0;
1414 }
1415
1416 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1417 switch (F->getIntrinsicID()) {
1418 case Intrinsic::bswap:
1419 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1420 case Intrinsic::ctpop:
1421 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1422 case Intrinsic::convert_from_fp16: {
1423 APFloat Val(APFloat::IEEEhalf, Op->getValue());
1424
1425 bool lost = false;
1426 APFloat::opStatus status =
1427 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1428
1429 // Conversion is always precise.
1430 (void)status;
1431 assert(status == APFloat::opOK && !lost &&
1432 "Precision lost during fp16 constfolding");
1433
1434 return ConstantFP::get(F->getContext(), Val);
1435 }
1436 default:
1437 return 0;
1438 }
1439 }
1440
1441 // Support ConstantVector in case we have an Undef in the top.
1442 if (isa<ConstantVector>(Operands[0]) ||
1443 isa<ConstantDataVector>(Operands[0])) {
1444 Constant *Op = cast<Constant>(Operands[0]);
1445 switch (F->getIntrinsicID()) {
1446 default: break;
1447 case Intrinsic::x86_sse_cvtss2si:
1448 case Intrinsic::x86_sse_cvtss2si64:
1449 case Intrinsic::x86_sse2_cvtsd2si:
1450 case Intrinsic::x86_sse2_cvtsd2si64:
1451 if (ConstantFP *FPOp =
1452 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1453 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1454 /*roundTowardZero=*/false, Ty);
1455 case Intrinsic::x86_sse_cvttss2si:
1456 case Intrinsic::x86_sse_cvttss2si64:
1457 case Intrinsic::x86_sse2_cvttsd2si:
1458 case Intrinsic::x86_sse2_cvttsd2si64:
1459 if (ConstantFP *FPOp =
1460 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1461 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1462 /*roundTowardZero=*/true, Ty);
1463 }
1464 }
1465
1466 if (isa<UndefValue>(Operands[0])) {
1467 if (F->getIntrinsicID() == Intrinsic::bswap)
1468 return Operands[0];
1469 return 0;
1470 }
1471
1472 return 0;
1473 }
1474
1475 if (Operands.size() == 2) {
1476 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1477 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1478 return 0;
1479 double Op1V;
1480 if (Ty->isFloatTy())
1481 Op1V = Op1->getValueAPF().convertToFloat();
1482 else if (Ty->isDoubleTy())
1483 Op1V = Op1->getValueAPF().convertToDouble();
1484 else {
1485 bool unused;
1486 APFloat APF = Op1->getValueAPF();
1487 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1488 Op1V = APF.convertToDouble();
1489 }
1490
1491 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1492 if (Op2->getType() != Op1->getType())
1493 return 0;
1494
1495 double Op2V;
1496 if (Ty->isFloatTy())
1497 Op2V = Op2->getValueAPF().convertToFloat();
1498 else if (Ty->isDoubleTy())
1499 Op2V = Op2->getValueAPF().convertToDouble();
1500 else {
1501 bool unused;
1502 APFloat APF = Op2->getValueAPF();
1503 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &unused);
1504 Op2V = APF.convertToDouble();
1505 }
1506
1507 if (F->getIntrinsicID() == Intrinsic::pow) {
1508 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1509 }
1510 if (!TLI)
1511 return 0;
1512 if (Name == "pow" && TLI->has(LibFunc::pow))
1513 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1514 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1515 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1516 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1517 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1518 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1519 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isHalfTy())
1520 return ConstantFP::get(F->getContext(),
1521 APFloat((float)std::pow((float)Op1V,
1522 (int)Op2C->getZExtValue())));
1523 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1524 return ConstantFP::get(F->getContext(),
1525 APFloat((float)std::pow((float)Op1V,
1526 (int)Op2C->getZExtValue())));
1527 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1528 return ConstantFP::get(F->getContext(),
1529 APFloat((double)std::pow((double)Op1V,
1530 (int)Op2C->getZExtValue())));
1531 }
1532 return 0;
1533 }
1534
1535 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1536 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1537 switch (F->getIntrinsicID()) {
1538 default: break;
1539 case Intrinsic::sadd_with_overflow:
1540 case Intrinsic::uadd_with_overflow:
1541 case Intrinsic::ssub_with_overflow:
1542 case Intrinsic::usub_with_overflow:
1543 case Intrinsic::smul_with_overflow:
1544 case Intrinsic::umul_with_overflow: {
1545 APInt Res;
1546 bool Overflow;
1547 switch (F->getIntrinsicID()) {
1548 default: llvm_unreachable("Invalid case");
1549 case Intrinsic::sadd_with_overflow:
1550 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1551 break;
1552 case Intrinsic::uadd_with_overflow:
1553 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1554 break;
1555 case Intrinsic::ssub_with_overflow:
1556 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1557 break;
1558 case Intrinsic::usub_with_overflow:
1559 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1560 break;
1561 case Intrinsic::smul_with_overflow:
1562 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1563 break;
1564 case Intrinsic::umul_with_overflow:
1565 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1566 break;
1567 }
1568 Constant *Ops[] = {
1569 ConstantInt::get(F->getContext(), Res),
1570 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1571 };
1572 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1573 }
1574 case Intrinsic::cttz:
1575 if (Op2->isOne() && Op1->isZero()) // cttz(0, 1) is undef.
1576 return UndefValue::get(Ty);
1577 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1578 case Intrinsic::ctlz:
1579 if (Op2->isOne() && Op1->isZero()) // ctlz(0, 1) is undef.
1580 return UndefValue::get(Ty);
1581 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1582 }
1583 }
1584
1585 return 0;
1586 }
1587 return 0;
1588 }
1589 return 0;
1590 }
1591