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 VMCore ConstantExpr simplifications,
13 // this file defines some additional folding routines that can make use of
14 // TargetData information. These functions cannot go in VMCore due to library
15 // dependency issues.
16 //
17 //===----------------------------------------------------------------------===//
18
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLibraryInfo.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/StringMap.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/FEnv.h"
36 #include <cerrno>
37 #include <cmath>
38 using namespace llvm;
39
40 //===----------------------------------------------------------------------===//
41 // Constant Folding internal helper functions
42 //===----------------------------------------------------------------------===//
43
44 /// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
45 /// TargetData. This always returns a non-null constant, but it may be a
46 /// ConstantExpr if unfoldable.
FoldBitCast(Constant * C,Type * DestTy,const TargetData & TD)47 static Constant *FoldBitCast(Constant *C, Type *DestTy,
48 const TargetData &TD) {
49 // Catch the obvious splat cases.
50 if (C->isNullValue() && !DestTy->isX86_MMXTy())
51 return Constant::getNullValue(DestTy);
52 if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
53 return Constant::getAllOnesValue(DestTy);
54
55 // Handle a vector->integer cast.
56 if (IntegerType *IT = dyn_cast<IntegerType>(DestTy)) {
57 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(C);
58 if (CDV == 0)
59 return ConstantExpr::getBitCast(C, DestTy);
60
61 unsigned NumSrcElts = CDV->getType()->getNumElements();
62
63 Type *SrcEltTy = CDV->getType()->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 VMCore to do the conversion now that #elts line up.
72 C = ConstantExpr::getBitCast(C, SrcIVTy);
73 CDV = cast<ConstantDataVector>(C);
74 }
75
76 // Now that we know that the input value is a vector of integers, just shift
77 // and insert them into our result.
78 unsigned BitShift = TD.getTypeAllocSizeInBits(SrcEltTy);
79 APInt Result(IT->getBitWidth(), 0);
80 for (unsigned i = 0; i != NumSrcElts; ++i) {
81 Result <<= BitShift;
82 if (TD.isLittleEndian())
83 Result |= CDV->getElementAsInteger(NumSrcElts-i-1);
84 else
85 Result |= CDV->getElementAsInteger(i);
86 }
87
88 return ConstantInt::get(IT, Result);
89 }
90
91 // The code below only handles casts to vectors currently.
92 VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
93 if (DestVTy == 0)
94 return ConstantExpr::getBitCast(C, DestTy);
95
96 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
97 // vector so the code below can handle it uniformly.
98 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
99 Constant *Ops = C; // don't take the address of C!
100 return FoldBitCast(ConstantVector::get(Ops), DestTy, TD);
101 }
102
103 // If this is a bitcast from constant vector -> vector, fold it.
104 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
105 return ConstantExpr::getBitCast(C, DestTy);
106
107 // If the element types match, VMCore can fold it.
108 unsigned NumDstElt = DestVTy->getNumElements();
109 unsigned NumSrcElt = C->getType()->getVectorNumElements();
110 if (NumDstElt == NumSrcElt)
111 return ConstantExpr::getBitCast(C, DestTy);
112
113 Type *SrcEltTy = C->getType()->getVectorElementType();
114 Type *DstEltTy = DestVTy->getElementType();
115
116 // Otherwise, we're changing the number of elements in a vector, which
117 // requires endianness information to do the right thing. For example,
118 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
119 // folds to (little endian):
120 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
121 // and to (big endian):
122 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
123
124 // First thing is first. We only want to think about integer here, so if
125 // we have something in FP form, recast it as integer.
126 if (DstEltTy->isFloatingPointTy()) {
127 // Fold to an vector of integers with same size as our FP type.
128 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
129 Type *DestIVTy =
130 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
131 // Recursively handle this integer conversion, if possible.
132 C = FoldBitCast(C, DestIVTy, TD);
133
134 // Finally, VMCore can handle this now that #elts line up.
135 return ConstantExpr::getBitCast(C, DestTy);
136 }
137
138 // Okay, we know the destination is integer, if the input is FP, convert
139 // it to integer first.
140 if (SrcEltTy->isFloatingPointTy()) {
141 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
142 Type *SrcIVTy =
143 VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
144 // Ask VMCore to do the conversion now that #elts line up.
145 C = ConstantExpr::getBitCast(C, SrcIVTy);
146 // If VMCore wasn't able to fold it, bail out.
147 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
148 !isa<ConstantDataVector>(C))
149 return C;
150 }
151
152 // Now we know that the input and output vectors are both integer vectors
153 // of the same size, and that their #elements is not the same. Do the
154 // conversion here, which depends on whether the input or output has
155 // more elements.
156 bool isLittleEndian = TD.isLittleEndian();
157
158 SmallVector<Constant*, 32> Result;
159 if (NumDstElt < NumSrcElt) {
160 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
161 Constant *Zero = Constant::getNullValue(DstEltTy);
162 unsigned Ratio = NumSrcElt/NumDstElt;
163 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
164 unsigned SrcElt = 0;
165 for (unsigned i = 0; i != NumDstElt; ++i) {
166 // Build each element of the result.
167 Constant *Elt = Zero;
168 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
169 for (unsigned j = 0; j != Ratio; ++j) {
170 Constant *Src =dyn_cast<ConstantInt>(C->getAggregateElement(SrcElt++));
171 if (!Src) // Reject constantexpr elements.
172 return ConstantExpr::getBitCast(C, DestTy);
173
174 // Zero extend the element to the right size.
175 Src = ConstantExpr::getZExt(Src, Elt->getType());
176
177 // Shift it to the right place, depending on endianness.
178 Src = ConstantExpr::getShl(Src,
179 ConstantInt::get(Src->getType(), ShiftAmt));
180 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
181
182 // Mix it in.
183 Elt = ConstantExpr::getOr(Elt, Src);
184 }
185 Result.push_back(Elt);
186 }
187 return ConstantVector::get(Result);
188 }
189
190 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
191 unsigned Ratio = NumDstElt/NumSrcElt;
192 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
193
194 // Loop over each source value, expanding into multiple results.
195 for (unsigned i = 0; i != NumSrcElt; ++i) {
196 Constant *Src = dyn_cast<ConstantInt>(C->getAggregateElement(i));
197 if (!Src) // Reject constantexpr elements.
198 return ConstantExpr::getBitCast(C, DestTy);
199
200 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
201 for (unsigned j = 0; j != Ratio; ++j) {
202 // Shift the piece of the value into the right place, depending on
203 // endianness.
204 Constant *Elt = ConstantExpr::getLShr(Src,
205 ConstantInt::get(Src->getType(), ShiftAmt));
206 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
207
208 // Truncate and remember this piece.
209 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
210 }
211 }
212
213 return ConstantVector::get(Result);
214 }
215
216
217 /// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
218 /// from a global, return the global and the constant. Because of
219 /// constantexprs, this function is recursive.
IsConstantOffsetFromGlobal(Constant * C,GlobalValue * & GV,int64_t & Offset,const TargetData & TD)220 static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
221 int64_t &Offset, const TargetData &TD) {
222 // Trivial case, constant is the global.
223 if ((GV = dyn_cast<GlobalValue>(C))) {
224 Offset = 0;
225 return true;
226 }
227
228 // Otherwise, if this isn't a constant expr, bail out.
229 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
230 if (!CE) return false;
231
232 // Look through ptr->int and ptr->ptr casts.
233 if (CE->getOpcode() == Instruction::PtrToInt ||
234 CE->getOpcode() == Instruction::BitCast)
235 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
236
237 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
238 if (CE->getOpcode() == Instruction::GetElementPtr) {
239 // Cannot compute this if the element type of the pointer is missing size
240 // info.
241 if (!cast<PointerType>(CE->getOperand(0)->getType())
242 ->getElementType()->isSized())
243 return false;
244
245 // If the base isn't a global+constant, we aren't either.
246 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
247 return false;
248
249 // Otherwise, add any offset that our operands provide.
250 gep_type_iterator GTI = gep_type_begin(CE);
251 for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
252 i != e; ++i, ++GTI) {
253 ConstantInt *CI = dyn_cast<ConstantInt>(*i);
254 if (!CI) return false; // Index isn't a simple constant?
255 if (CI->isZero()) continue; // Not adding anything.
256
257 if (StructType *ST = dyn_cast<StructType>(*GTI)) {
258 // N = N + Offset
259 Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
260 } else {
261 SequentialType *SQT = cast<SequentialType>(*GTI);
262 Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
263 }
264 }
265 return true;
266 }
267
268 return false;
269 }
270
271 /// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
272 /// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
273 /// pointer to copy results into and BytesLeft is the number of bytes left in
274 /// the CurPtr buffer. TD is the target data.
ReadDataFromGlobal(Constant * C,uint64_t ByteOffset,unsigned char * CurPtr,unsigned BytesLeft,const TargetData & TD)275 static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
276 unsigned char *CurPtr, unsigned BytesLeft,
277 const TargetData &TD) {
278 assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
279 "Out of range access");
280
281 // If this element is zero or undefined, we can just return since *CurPtr is
282 // zero initialized.
283 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
284 return true;
285
286 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
287 if (CI->getBitWidth() > 64 ||
288 (CI->getBitWidth() & 7) != 0)
289 return false;
290
291 uint64_t Val = CI->getZExtValue();
292 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
293
294 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
295 CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
296 ++ByteOffset;
297 }
298 return true;
299 }
300
301 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
302 if (CFP->getType()->isDoubleTy()) {
303 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
304 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
305 }
306 if (CFP->getType()->isFloatTy()){
307 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
308 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
309 }
310 return false;
311 }
312
313 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
314 const StructLayout *SL = TD.getStructLayout(CS->getType());
315 unsigned Index = SL->getElementContainingOffset(ByteOffset);
316 uint64_t CurEltOffset = SL->getElementOffset(Index);
317 ByteOffset -= CurEltOffset;
318
319 while (1) {
320 // If the element access is to the element itself and not to tail padding,
321 // read the bytes from the element.
322 uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
323
324 if (ByteOffset < EltSize &&
325 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
326 BytesLeft, TD))
327 return false;
328
329 ++Index;
330
331 // Check to see if we read from the last struct element, if so we're done.
332 if (Index == CS->getType()->getNumElements())
333 return true;
334
335 // If we read all of the bytes we needed from this element we're done.
336 uint64_t NextEltOffset = SL->getElementOffset(Index);
337
338 if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
339 return true;
340
341 // Move to the next element of the struct.
342 CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
343 BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
344 ByteOffset = 0;
345 CurEltOffset = NextEltOffset;
346 }
347 // not reached.
348 }
349
350 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
351 isa<ConstantDataSequential>(C)) {
352 Type *EltTy = cast<SequentialType>(C->getType())->getElementType();
353 uint64_t EltSize = TD.getTypeAllocSize(EltTy);
354 uint64_t Index = ByteOffset / EltSize;
355 uint64_t Offset = ByteOffset - Index * EltSize;
356 uint64_t NumElts;
357 if (ArrayType *AT = dyn_cast<ArrayType>(C->getType()))
358 NumElts = AT->getNumElements();
359 else
360 NumElts = cast<VectorType>(C->getType())->getNumElements();
361
362 for (; Index != NumElts; ++Index) {
363 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
364 BytesLeft, TD))
365 return false;
366
367 uint64_t BytesWritten = EltSize - Offset;
368 assert(BytesWritten <= EltSize && "Not indexing into this element?");
369 if (BytesWritten >= BytesLeft)
370 return true;
371
372 Offset = 0;
373 BytesLeft -= BytesWritten;
374 CurPtr += BytesWritten;
375 }
376 return true;
377 }
378
379 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
380 if (CE->getOpcode() == Instruction::IntToPtr &&
381 CE->getOperand(0)->getType() == TD.getIntPtrType(CE->getContext()))
382 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
383 BytesLeft, TD);
384 }
385
386 // Otherwise, unknown initializer type.
387 return false;
388 }
389
FoldReinterpretLoadFromConstPtr(Constant * C,const TargetData & TD)390 static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
391 const TargetData &TD) {
392 Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
393 IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
394
395 // If this isn't an integer load we can't fold it directly.
396 if (!IntType) {
397 // If this is a float/double load, we can try folding it as an int32/64 load
398 // and then bitcast the result. This can be useful for union cases. Note
399 // that address spaces don't matter here since we're not going to result in
400 // an actual new load.
401 Type *MapTy;
402 if (LoadTy->isFloatTy())
403 MapTy = Type::getInt32PtrTy(C->getContext());
404 else if (LoadTy->isDoubleTy())
405 MapTy = Type::getInt64PtrTy(C->getContext());
406 else if (LoadTy->isVectorTy()) {
407 MapTy = IntegerType::get(C->getContext(),
408 TD.getTypeAllocSizeInBits(LoadTy));
409 MapTy = PointerType::getUnqual(MapTy);
410 } else
411 return 0;
412
413 C = FoldBitCast(C, MapTy, TD);
414 if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
415 return FoldBitCast(Res, LoadTy, TD);
416 return 0;
417 }
418
419 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
420 if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
421
422 GlobalValue *GVal;
423 int64_t Offset;
424 if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
425 return 0;
426
427 GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
428 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
429 !GV->getInitializer()->getType()->isSized())
430 return 0;
431
432 // If we're loading off the beginning of the global, some bytes may be valid,
433 // but we don't try to handle this.
434 if (Offset < 0) return 0;
435
436 // If we're not accessing anything in this constant, the result is undefined.
437 if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
438 return UndefValue::get(IntType);
439
440 unsigned char RawBytes[32] = {0};
441 if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
442 BytesLoaded, TD))
443 return 0;
444
445 APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
446 for (unsigned i = 1; i != BytesLoaded; ++i) {
447 ResultVal <<= 8;
448 ResultVal |= RawBytes[BytesLoaded-1-i];
449 }
450
451 return ConstantInt::get(IntType->getContext(), ResultVal);
452 }
453
454 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
455 /// produce if it is constant and determinable. If this is not determinable,
456 /// return null.
ConstantFoldLoadFromConstPtr(Constant * C,const TargetData * TD)457 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
458 const TargetData *TD) {
459 // First, try the easy cases:
460 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
461 if (GV->isConstant() && GV->hasDefinitiveInitializer())
462 return GV->getInitializer();
463
464 // If the loaded value isn't a constant expr, we can't handle it.
465 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
466 if (!CE) return 0;
467
468 if (CE->getOpcode() == Instruction::GetElementPtr) {
469 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
470 if (GV->isConstant() && GV->hasDefinitiveInitializer())
471 if (Constant *V =
472 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
473 return V;
474 }
475
476 // Instead of loading constant c string, use corresponding integer value
477 // directly if string length is small enough.
478 StringRef Str;
479 if (TD && getConstantStringInfo(CE, Str) && !Str.empty()) {
480 unsigned StrLen = Str.size();
481 Type *Ty = cast<PointerType>(CE->getType())->getElementType();
482 unsigned NumBits = Ty->getPrimitiveSizeInBits();
483 // Replace load with immediate integer if the result is an integer or fp
484 // value.
485 if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
486 (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
487 APInt StrVal(NumBits, 0);
488 APInt SingleChar(NumBits, 0);
489 if (TD->isLittleEndian()) {
490 for (signed i = StrLen-1; i >= 0; i--) {
491 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
492 StrVal = (StrVal << 8) | SingleChar;
493 }
494 } else {
495 for (unsigned i = 0; i < StrLen; i++) {
496 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
497 StrVal = (StrVal << 8) | SingleChar;
498 }
499 // Append NULL at the end.
500 SingleChar = 0;
501 StrVal = (StrVal << 8) | SingleChar;
502 }
503
504 Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
505 if (Ty->isFloatingPointTy())
506 Res = ConstantExpr::getBitCast(Res, Ty);
507 return Res;
508 }
509 }
510
511 // If this load comes from anywhere in a constant global, and if the global
512 // is all undef or zero, we know what it loads.
513 if (GlobalVariable *GV =
514 dyn_cast<GlobalVariable>(GetUnderlyingObject(CE, TD))) {
515 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
516 Type *ResTy = cast<PointerType>(C->getType())->getElementType();
517 if (GV->getInitializer()->isNullValue())
518 return Constant::getNullValue(ResTy);
519 if (isa<UndefValue>(GV->getInitializer()))
520 return UndefValue::get(ResTy);
521 }
522 }
523
524 // Try hard to fold loads from bitcasted strange and non-type-safe things. We
525 // currently don't do any of this for big endian systems. It can be
526 // generalized in the future if someone is interested.
527 if (TD && TD->isLittleEndian())
528 return FoldReinterpretLoadFromConstPtr(CE, *TD);
529 return 0;
530 }
531
ConstantFoldLoadInst(const LoadInst * LI,const TargetData * TD)532 static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
533 if (LI->isVolatile()) return 0;
534
535 if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
536 return ConstantFoldLoadFromConstPtr(C, TD);
537
538 return 0;
539 }
540
541 /// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
542 /// Attempt to symbolically evaluate the result of a binary operator merging
543 /// these together. If target data info is available, it is provided as TD,
544 /// otherwise TD is null.
SymbolicallyEvaluateBinop(unsigned Opc,Constant * Op0,Constant * Op1,const TargetData * TD)545 static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
546 Constant *Op1, const TargetData *TD){
547 // SROA
548
549 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
550 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
551 // bits.
552
553
554 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
555 // constant. This happens frequently when iterating over a global array.
556 if (Opc == Instruction::Sub && TD) {
557 GlobalValue *GV1, *GV2;
558 int64_t Offs1, Offs2;
559
560 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
561 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
562 GV1 == GV2) {
563 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
564 return ConstantInt::get(Op0->getType(), Offs1-Offs2);
565 }
566 }
567
568 return 0;
569 }
570
571 /// CastGEPIndices - If array indices are not pointer-sized integers,
572 /// explicitly cast them so that they aren't implicitly casted by the
573 /// getelementptr.
CastGEPIndices(ArrayRef<Constant * > Ops,Type * ResultTy,const TargetData * TD,const TargetLibraryInfo * TLI)574 static Constant *CastGEPIndices(ArrayRef<Constant *> Ops,
575 Type *ResultTy, const TargetData *TD,
576 const TargetLibraryInfo *TLI) {
577 if (!TD) return 0;
578 Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
579
580 bool Any = false;
581 SmallVector<Constant*, 32> NewIdxs;
582 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
583 if ((i == 1 ||
584 !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
585 Ops.slice(1, i-1)))) &&
586 Ops[i]->getType() != IntPtrTy) {
587 Any = true;
588 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
589 true,
590 IntPtrTy,
591 true),
592 Ops[i], IntPtrTy));
593 } else
594 NewIdxs.push_back(Ops[i]);
595 }
596 if (!Any) return 0;
597
598 Constant *C =
599 ConstantExpr::getGetElementPtr(Ops[0], NewIdxs);
600 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
601 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
602 C = Folded;
603 return C;
604 }
605
606 /// Strip the pointer casts, but preserve the address space information.
StripPtrCastKeepAS(Constant * Ptr)607 static Constant* StripPtrCastKeepAS(Constant* Ptr) {
608 assert(Ptr->getType()->isPointerTy() && "Not a pointer type");
609 PointerType *OldPtrTy = cast<PointerType>(Ptr->getType());
610 Ptr = cast<Constant>(Ptr->stripPointerCasts());
611 PointerType *NewPtrTy = cast<PointerType>(Ptr->getType());
612
613 // Preserve the address space number of the pointer.
614 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
615 NewPtrTy = NewPtrTy->getElementType()->getPointerTo(
616 OldPtrTy->getAddressSpace());
617 Ptr = ConstantExpr::getBitCast(Ptr, NewPtrTy);
618 }
619 return Ptr;
620 }
621
622 /// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
623 /// constant expression, do so.
SymbolicallyEvaluateGEP(ArrayRef<Constant * > Ops,Type * ResultTy,const TargetData * TD,const TargetLibraryInfo * TLI)624 static Constant *SymbolicallyEvaluateGEP(ArrayRef<Constant *> Ops,
625 Type *ResultTy, const TargetData *TD,
626 const TargetLibraryInfo *TLI) {
627 Constant *Ptr = Ops[0];
628 if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized() ||
629 !Ptr->getType()->isPointerTy())
630 return 0;
631
632 Type *IntPtrTy = TD->getIntPtrType(Ptr->getContext());
633
634 // If this is a constant expr gep that is effectively computing an
635 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
636 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
637 if (!isa<ConstantInt>(Ops[i])) {
638
639 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
640 // "inttoptr (sub (ptrtoint Ptr), V)"
641 if (Ops.size() == 2 &&
642 cast<PointerType>(ResultTy)->getElementType()->isIntegerTy(8)) {
643 ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[1]);
644 assert((CE == 0 || CE->getType() == IntPtrTy) &&
645 "CastGEPIndices didn't canonicalize index types!");
646 if (CE && CE->getOpcode() == Instruction::Sub &&
647 CE->getOperand(0)->isNullValue()) {
648 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
649 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
650 Res = ConstantExpr::getIntToPtr(Res, ResultTy);
651 if (ConstantExpr *ResCE = dyn_cast<ConstantExpr>(Res))
652 Res = ConstantFoldConstantExpression(ResCE, TD, TLI);
653 return Res;
654 }
655 }
656 return 0;
657 }
658
659 unsigned BitWidth = TD->getTypeSizeInBits(IntPtrTy);
660 APInt Offset =
661 APInt(BitWidth, TD->getIndexedOffset(Ptr->getType(),
662 makeArrayRef((Value *const*)
663 Ops.data() + 1,
664 Ops.size() - 1)));
665 Ptr = StripPtrCastKeepAS(Ptr);
666
667 // If this is a GEP of a GEP, fold it all into a single GEP.
668 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
669 SmallVector<Value *, 4> NestedOps(GEP->op_begin()+1, GEP->op_end());
670
671 // Do not try the incorporate the sub-GEP if some index is not a number.
672 bool AllConstantInt = true;
673 for (unsigned i = 0, e = NestedOps.size(); i != e; ++i)
674 if (!isa<ConstantInt>(NestedOps[i])) {
675 AllConstantInt = false;
676 break;
677 }
678 if (!AllConstantInt)
679 break;
680
681 Ptr = cast<Constant>(GEP->getOperand(0));
682 Offset += APInt(BitWidth,
683 TD->getIndexedOffset(Ptr->getType(), NestedOps));
684 Ptr = StripPtrCastKeepAS(Ptr);
685 }
686
687 // If the base value for this address is a literal integer value, fold the
688 // getelementptr to the resulting integer value casted to the pointer type.
689 APInt BasePtr(BitWidth, 0);
690 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
691 if (CE->getOpcode() == Instruction::IntToPtr)
692 if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
693 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
694 if (Ptr->isNullValue() || BasePtr != 0) {
695 Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
696 return ConstantExpr::getIntToPtr(C, ResultTy);
697 }
698
699 // Otherwise form a regular getelementptr. Recompute the indices so that
700 // we eliminate over-indexing of the notional static type array bounds.
701 // This makes it easy to determine if the getelementptr is "inbounds".
702 // Also, this helps GlobalOpt do SROA on GlobalVariables.
703 Type *Ty = Ptr->getType();
704 assert(Ty->isPointerTy() && "Forming regular GEP of non-pointer type");
705 SmallVector<Constant*, 32> NewIdxs;
706 do {
707 if (SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
708 if (ATy->isPointerTy()) {
709 // The only pointer indexing we'll do is on the first index of the GEP.
710 if (!NewIdxs.empty())
711 break;
712
713 // Only handle pointers to sized types, not pointers to functions.
714 if (!ATy->getElementType()->isSized())
715 return 0;
716 }
717
718 // Determine which element of the array the offset points into.
719 APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
720 IntegerType *IntPtrTy = TD->getIntPtrType(Ty->getContext());
721 if (ElemSize == 0)
722 // The element size is 0. This may be [0 x Ty]*, so just use a zero
723 // index for this level and proceed to the next level to see if it can
724 // accommodate the offset.
725 NewIdxs.push_back(ConstantInt::get(IntPtrTy, 0));
726 else {
727 // The element size is non-zero divide the offset by the element
728 // size (rounding down), to compute the index at this level.
729 APInt NewIdx = Offset.udiv(ElemSize);
730 Offset -= NewIdx * ElemSize;
731 NewIdxs.push_back(ConstantInt::get(IntPtrTy, NewIdx));
732 }
733 Ty = ATy->getElementType();
734 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
735 // If we end up with an offset that isn't valid for this struct type, we
736 // can't re-form this GEP in a regular form, so bail out. The pointer
737 // operand likely went through casts that are necessary to make the GEP
738 // sensible.
739 const StructLayout &SL = *TD->getStructLayout(STy);
740 if (Offset.uge(SL.getSizeInBytes()))
741 break;
742
743 // Determine which field of the struct the offset points into. The
744 // getZExtValue is fine as we've already ensured that the offset is
745 // within the range representable by the StructLayout API.
746 unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
747 NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
748 ElIdx));
749 Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
750 Ty = STy->getTypeAtIndex(ElIdx);
751 } else {
752 // We've reached some non-indexable type.
753 break;
754 }
755 } while (Ty != cast<PointerType>(ResultTy)->getElementType());
756
757 // If we haven't used up the entire offset by descending the static
758 // type, then the offset is pointing into the middle of an indivisible
759 // member, so we can't simplify it.
760 if (Offset != 0)
761 return 0;
762
763 // Create a GEP.
764 Constant *C =
765 ConstantExpr::getGetElementPtr(Ptr, NewIdxs);
766 assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
767 "Computed GetElementPtr has unexpected type!");
768
769 // If we ended up indexing a member with a type that doesn't match
770 // the type of what the original indices indexed, add a cast.
771 if (Ty != cast<PointerType>(ResultTy)->getElementType())
772 C = FoldBitCast(C, ResultTy, *TD);
773
774 return C;
775 }
776
777
778
779 //===----------------------------------------------------------------------===//
780 // Constant Folding public APIs
781 //===----------------------------------------------------------------------===//
782
783 /// ConstantFoldInstruction - Try to constant fold the specified instruction.
784 /// If successful, the constant result is returned, if not, null is returned.
785 /// Note that this fails if not all of the operands are constant. Otherwise,
786 /// this function can only fail when attempting to fold instructions like loads
787 /// and stores, which have no constant expression form.
ConstantFoldInstruction(Instruction * I,const TargetData * TD,const TargetLibraryInfo * TLI)788 Constant *llvm::ConstantFoldInstruction(Instruction *I,
789 const TargetData *TD,
790 const TargetLibraryInfo *TLI) {
791 // Handle PHI nodes quickly here...
792 if (PHINode *PN = dyn_cast<PHINode>(I)) {
793 Constant *CommonValue = 0;
794
795 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
796 Value *Incoming = PN->getIncomingValue(i);
797 // If the incoming value is undef then skip it. Note that while we could
798 // skip the value if it is equal to the phi node itself we choose not to
799 // because that would break the rule that constant folding only applies if
800 // all operands are constants.
801 if (isa<UndefValue>(Incoming))
802 continue;
803 // If the incoming value is not a constant, then give up.
804 Constant *C = dyn_cast<Constant>(Incoming);
805 if (!C)
806 return 0;
807 // Fold the PHI's operands.
808 if (ConstantExpr *NewC = dyn_cast<ConstantExpr>(C))
809 C = ConstantFoldConstantExpression(NewC, TD, TLI);
810 // If the incoming value is a different constant to
811 // the one we saw previously, then give up.
812 if (CommonValue && C != CommonValue)
813 return 0;
814 CommonValue = C;
815 }
816
817
818 // If we reach here, all incoming values are the same constant or undef.
819 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
820 }
821
822 // Scan the operand list, checking to see if they are all constants, if so,
823 // hand off to ConstantFoldInstOperands.
824 SmallVector<Constant*, 8> Ops;
825 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
826 Constant *Op = dyn_cast<Constant>(*i);
827 if (!Op)
828 return 0; // All operands not constant!
829
830 // Fold the Instruction's operands.
831 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(Op))
832 Op = ConstantFoldConstantExpression(NewCE, TD, TLI);
833
834 Ops.push_back(Op);
835 }
836
837 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
838 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
839 TD, TLI);
840
841 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
842 return ConstantFoldLoadInst(LI, TD);
843
844 if (InsertValueInst *IVI = dyn_cast<InsertValueInst>(I))
845 return ConstantExpr::getInsertValue(
846 cast<Constant>(IVI->getAggregateOperand()),
847 cast<Constant>(IVI->getInsertedValueOperand()),
848 IVI->getIndices());
849
850 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I))
851 return ConstantExpr::getExtractValue(
852 cast<Constant>(EVI->getAggregateOperand()),
853 EVI->getIndices());
854
855 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Ops, TD, TLI);
856 }
857
858 /// ConstantFoldConstantExpression - Attempt to fold the constant expression
859 /// using the specified TargetData. If successful, the constant result is
860 /// result is returned, if not, null is returned.
ConstantFoldConstantExpression(const ConstantExpr * CE,const TargetData * TD,const TargetLibraryInfo * TLI)861 Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
862 const TargetData *TD,
863 const TargetLibraryInfo *TLI) {
864 SmallVector<Constant*, 8> Ops;
865 for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end();
866 i != e; ++i) {
867 Constant *NewC = cast<Constant>(*i);
868 // Recursively fold the ConstantExpr's operands.
869 if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
870 NewC = ConstantFoldConstantExpression(NewCE, TD, TLI);
871 Ops.push_back(NewC);
872 }
873
874 if (CE->isCompare())
875 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
876 TD, TLI);
877 return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(), Ops, TD, TLI);
878 }
879
880 /// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
881 /// specified opcode and operands. If successful, the constant result is
882 /// returned, if not, null is returned. Note that this function can fail when
883 /// attempting to fold instructions like loads and stores, which have no
884 /// constant expression form.
885 ///
886 /// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
887 /// information, due to only being passed an opcode and operands. Constant
888 /// folding using this function strips this information.
889 ///
ConstantFoldInstOperands(unsigned Opcode,Type * DestTy,ArrayRef<Constant * > Ops,const TargetData * TD,const TargetLibraryInfo * TLI)890 Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, Type *DestTy,
891 ArrayRef<Constant *> Ops,
892 const TargetData *TD,
893 const TargetLibraryInfo *TLI) {
894 // Handle easy binops first.
895 if (Instruction::isBinaryOp(Opcode)) {
896 if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
897 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
898 return C;
899
900 return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
901 }
902
903 switch (Opcode) {
904 default: return 0;
905 case Instruction::ICmp:
906 case Instruction::FCmp: llvm_unreachable("Invalid for compares");
907 case Instruction::Call:
908 if (Function *F = dyn_cast<Function>(Ops.back()))
909 if (canConstantFoldCallTo(F))
910 return ConstantFoldCall(F, Ops.slice(0, Ops.size() - 1), TLI);
911 return 0;
912 case Instruction::PtrToInt:
913 // If the input is a inttoptr, eliminate the pair. This requires knowing
914 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
915 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
916 if (TD && CE->getOpcode() == Instruction::IntToPtr) {
917 Constant *Input = CE->getOperand(0);
918 unsigned InWidth = Input->getType()->getScalarSizeInBits();
919 if (TD->getPointerSizeInBits() < InWidth) {
920 Constant *Mask =
921 ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
922 TD->getPointerSizeInBits()));
923 Input = ConstantExpr::getAnd(Input, Mask);
924 }
925 // Do a zext or trunc to get to the dest size.
926 return ConstantExpr::getIntegerCast(Input, DestTy, false);
927 }
928 }
929 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
930 case Instruction::IntToPtr:
931 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
932 // the int size is >= the ptr size. This requires knowing the width of a
933 // pointer, so it can't be done in ConstantExpr::getCast.
934 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0]))
935 if (TD &&
936 TD->getPointerSizeInBits() <= CE->getType()->getScalarSizeInBits() &&
937 CE->getOpcode() == Instruction::PtrToInt)
938 return FoldBitCast(CE->getOperand(0), DestTy, *TD);
939
940 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
941 case Instruction::Trunc:
942 case Instruction::ZExt:
943 case Instruction::SExt:
944 case Instruction::FPTrunc:
945 case Instruction::FPExt:
946 case Instruction::UIToFP:
947 case Instruction::SIToFP:
948 case Instruction::FPToUI:
949 case Instruction::FPToSI:
950 return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
951 case Instruction::BitCast:
952 if (TD)
953 return FoldBitCast(Ops[0], DestTy, *TD);
954 return ConstantExpr::getBitCast(Ops[0], DestTy);
955 case Instruction::Select:
956 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
957 case Instruction::ExtractElement:
958 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
959 case Instruction::InsertElement:
960 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
961 case Instruction::ShuffleVector:
962 return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
963 case Instruction::GetElementPtr:
964 if (Constant *C = CastGEPIndices(Ops, DestTy, TD, TLI))
965 return C;
966 if (Constant *C = SymbolicallyEvaluateGEP(Ops, DestTy, TD, TLI))
967 return C;
968
969 return ConstantExpr::getGetElementPtr(Ops[0], Ops.slice(1));
970 }
971 }
972
973 /// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
974 /// instruction (icmp/fcmp) with the specified operands. If it fails, it
975 /// returns a constant expression of the specified operands.
976 ///
ConstantFoldCompareInstOperands(unsigned Predicate,Constant * Ops0,Constant * Ops1,const TargetData * TD,const TargetLibraryInfo * TLI)977 Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
978 Constant *Ops0, Constant *Ops1,
979 const TargetData *TD,
980 const TargetLibraryInfo *TLI) {
981 // fold: icmp (inttoptr x), null -> icmp x, 0
982 // fold: icmp (ptrtoint x), 0 -> icmp x, null
983 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
984 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
985 //
986 // ConstantExpr::getCompare cannot do this, because it doesn't have TD
987 // around to know if bit truncation is happening.
988 if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
989 if (TD && Ops1->isNullValue()) {
990 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
991 if (CE0->getOpcode() == Instruction::IntToPtr) {
992 // Convert the integer value to the right size to ensure we get the
993 // proper extension or truncation.
994 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
995 IntPtrTy, false);
996 Constant *Null = Constant::getNullValue(C->getType());
997 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
998 }
999
1000 // Only do this transformation if the int is intptrty in size, otherwise
1001 // there is a truncation or extension that we aren't modeling.
1002 if (CE0->getOpcode() == Instruction::PtrToInt &&
1003 CE0->getType() == IntPtrTy) {
1004 Constant *C = CE0->getOperand(0);
1005 Constant *Null = Constant::getNullValue(C->getType());
1006 return ConstantFoldCompareInstOperands(Predicate, C, Null, TD, TLI);
1007 }
1008 }
1009
1010 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1011 if (TD && CE0->getOpcode() == CE1->getOpcode()) {
1012 Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
1013
1014 if (CE0->getOpcode() == Instruction::IntToPtr) {
1015 // Convert the integer value to the right size to ensure we get the
1016 // proper extension or truncation.
1017 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1018 IntPtrTy, false);
1019 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1020 IntPtrTy, false);
1021 return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD, TLI);
1022 }
1023
1024 // Only do this transformation if the int is intptrty in size, otherwise
1025 // there is a truncation or extension that we aren't modeling.
1026 if ((CE0->getOpcode() == Instruction::PtrToInt &&
1027 CE0->getType() == IntPtrTy &&
1028 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
1029 return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
1030 CE1->getOperand(0), TD, TLI);
1031 }
1032 }
1033
1034 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1035 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1036 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1037 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1038 Constant *LHS =
1039 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,
1040 TD, TLI);
1041 Constant *RHS =
1042 ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,
1043 TD, TLI);
1044 unsigned OpC =
1045 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1046 Constant *Ops[] = { LHS, RHS };
1047 return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, TD, TLI);
1048 }
1049 }
1050
1051 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1052 }
1053
1054
1055 /// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
1056 /// getelementptr constantexpr, return the constant value being addressed by the
1057 /// constant expression, or null if something is funny and we can't decide.
ConstantFoldLoadThroughGEPConstantExpr(Constant * C,ConstantExpr * CE)1058 Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
1059 ConstantExpr *CE) {
1060 if (!CE->getOperand(1)->isNullValue())
1061 return 0; // Do not allow stepping over the value!
1062
1063 // Loop over all of the operands, tracking down which value we are
1064 // addressing.
1065 for (unsigned i = 2, e = CE->getNumOperands(); i != e; ++i) {
1066 C = C->getAggregateElement(CE->getOperand(i));
1067 if (C == 0) return 0;
1068 }
1069 return C;
1070 }
1071
1072 /// ConstantFoldLoadThroughGEPIndices - Given a constant and getelementptr
1073 /// indices (with an *implied* zero pointer index that is not in the list),
1074 /// return the constant value being addressed by a virtual load, or null if
1075 /// something is funny and we can't decide.
ConstantFoldLoadThroughGEPIndices(Constant * C,ArrayRef<Constant * > Indices)1076 Constant *llvm::ConstantFoldLoadThroughGEPIndices(Constant *C,
1077 ArrayRef<Constant*> Indices) {
1078 // Loop over all of the operands, tracking down which value we are
1079 // addressing.
1080 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1081 C = C->getAggregateElement(Indices[i]);
1082 if (C == 0) return 0;
1083 }
1084 return C;
1085 }
1086
1087
1088 //===----------------------------------------------------------------------===//
1089 // Constant Folding for Calls
1090 //
1091
1092 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
1093 /// the specified function.
1094 bool
canConstantFoldCallTo(const Function * F)1095 llvm::canConstantFoldCallTo(const Function *F) {
1096 switch (F->getIntrinsicID()) {
1097 case Intrinsic::sqrt:
1098 case Intrinsic::pow:
1099 case Intrinsic::powi:
1100 case Intrinsic::bswap:
1101 case Intrinsic::ctpop:
1102 case Intrinsic::ctlz:
1103 case Intrinsic::cttz:
1104 case Intrinsic::sadd_with_overflow:
1105 case Intrinsic::uadd_with_overflow:
1106 case Intrinsic::ssub_with_overflow:
1107 case Intrinsic::usub_with_overflow:
1108 case Intrinsic::smul_with_overflow:
1109 case Intrinsic::umul_with_overflow:
1110 case Intrinsic::convert_from_fp16:
1111 case Intrinsic::convert_to_fp16:
1112 case Intrinsic::x86_sse_cvtss2si:
1113 case Intrinsic::x86_sse_cvtss2si64:
1114 case Intrinsic::x86_sse_cvttss2si:
1115 case Intrinsic::x86_sse_cvttss2si64:
1116 case Intrinsic::x86_sse2_cvtsd2si:
1117 case Intrinsic::x86_sse2_cvtsd2si64:
1118 case Intrinsic::x86_sse2_cvttsd2si:
1119 case Intrinsic::x86_sse2_cvttsd2si64:
1120 return true;
1121 default:
1122 return false;
1123 case 0: break;
1124 }
1125
1126 if (!F->hasName()) return false;
1127 StringRef Name = F->getName();
1128
1129 // In these cases, the check of the length is required. We don't want to
1130 // return true for a name like "cos\0blah" which strcmp would return equal to
1131 // "cos", but has length 8.
1132 switch (Name[0]) {
1133 default: return false;
1134 case 'a':
1135 return Name == "acos" || Name == "asin" ||
1136 Name == "atan" || Name == "atan2";
1137 case 'c':
1138 return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
1139 case 'e':
1140 return Name == "exp" || Name == "exp2";
1141 case 'f':
1142 return Name == "fabs" || Name == "fmod" || Name == "floor";
1143 case 'l':
1144 return Name == "log" || Name == "log10";
1145 case 'p':
1146 return Name == "pow";
1147 case 's':
1148 return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
1149 Name == "sinf" || Name == "sqrtf";
1150 case 't':
1151 return Name == "tan" || Name == "tanh";
1152 }
1153 }
1154
ConstantFoldFP(double (* NativeFP)(double),double V,Type * Ty)1155 static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
1156 Type *Ty) {
1157 sys::llvm_fenv_clearexcept();
1158 V = NativeFP(V);
1159 if (sys::llvm_fenv_testexcept()) {
1160 sys::llvm_fenv_clearexcept();
1161 return 0;
1162 }
1163
1164 if (Ty->isFloatTy())
1165 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1166 if (Ty->isDoubleTy())
1167 return ConstantFP::get(Ty->getContext(), APFloat(V));
1168 llvm_unreachable("Can only constant fold float/double");
1169 }
1170
ConstantFoldBinaryFP(double (* NativeFP)(double,double),double V,double W,Type * Ty)1171 static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1172 double V, double W, Type *Ty) {
1173 sys::llvm_fenv_clearexcept();
1174 V = NativeFP(V, W);
1175 if (sys::llvm_fenv_testexcept()) {
1176 sys::llvm_fenv_clearexcept();
1177 return 0;
1178 }
1179
1180 if (Ty->isFloatTy())
1181 return ConstantFP::get(Ty->getContext(), APFloat((float)V));
1182 if (Ty->isDoubleTy())
1183 return ConstantFP::get(Ty->getContext(), APFloat(V));
1184 llvm_unreachable("Can only constant fold float/double");
1185 }
1186
1187 /// ConstantFoldConvertToInt - Attempt to an SSE floating point to integer
1188 /// conversion of a constant floating point. If roundTowardZero is false, the
1189 /// default IEEE rounding is used (toward nearest, ties to even). This matches
1190 /// the behavior of the non-truncating SSE instructions in the default rounding
1191 /// mode. The desired integer type Ty is used to select how many bits are
1192 /// available for the result. Returns null if the conversion cannot be
1193 /// performed, otherwise returns the Constant value resulting from the
1194 /// conversion.
ConstantFoldConvertToInt(const APFloat & Val,bool roundTowardZero,Type * Ty)1195 static Constant *ConstantFoldConvertToInt(const APFloat &Val,
1196 bool roundTowardZero, Type *Ty) {
1197 // All of these conversion intrinsics form an integer of at most 64bits.
1198 unsigned ResultWidth = cast<IntegerType>(Ty)->getBitWidth();
1199 assert(ResultWidth <= 64 &&
1200 "Can only constant fold conversions to 64 and 32 bit ints");
1201
1202 uint64_t UIntVal;
1203 bool isExact = false;
1204 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1205 : APFloat::rmNearestTiesToEven;
1206 APFloat::opStatus status = Val.convertToInteger(&UIntVal, ResultWidth,
1207 /*isSigned=*/true, mode,
1208 &isExact);
1209 if (status != APFloat::opOK && status != APFloat::opInexact)
1210 return 0;
1211 return ConstantInt::get(Ty, UIntVal, /*isSigned=*/true);
1212 }
1213
1214 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
1215 /// with the specified arguments, returning null if unsuccessful.
1216 Constant *
ConstantFoldCall(Function * F,ArrayRef<Constant * > Operands,const TargetLibraryInfo * TLI)1217 llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
1218 const TargetLibraryInfo *TLI) {
1219 if (!F->hasName()) return 0;
1220 StringRef Name = F->getName();
1221
1222 Type *Ty = F->getReturnType();
1223 if (Operands.size() == 1) {
1224 if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
1225 if (F->getIntrinsicID() == Intrinsic::convert_to_fp16) {
1226 APFloat Val(Op->getValueAPF());
1227
1228 bool lost = false;
1229 Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
1230
1231 return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
1232 }
1233 if (!TLI)
1234 return 0;
1235
1236 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1237 return 0;
1238
1239 /// We only fold functions with finite arguments. Folding NaN and inf is
1240 /// likely to be aborted with an exception anyway, and some host libms
1241 /// have known errors raising exceptions.
1242 if (Op->getValueAPF().isNaN() || Op->getValueAPF().isInfinity())
1243 return 0;
1244
1245 /// Currently APFloat versions of these functions do not exist, so we use
1246 /// the host native double versions. Float versions are not called
1247 /// directly but for all these it is true (float)(f((double)arg)) ==
1248 /// f(arg). Long double not supported yet.
1249 double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
1250 Op->getValueAPF().convertToDouble();
1251 switch (Name[0]) {
1252 case 'a':
1253 if (Name == "acos" && TLI->has(LibFunc::acos))
1254 return ConstantFoldFP(acos, V, Ty);
1255 else if (Name == "asin" && TLI->has(LibFunc::asin))
1256 return ConstantFoldFP(asin, V, Ty);
1257 else if (Name == "atan" && TLI->has(LibFunc::atan))
1258 return ConstantFoldFP(atan, V, Ty);
1259 break;
1260 case 'c':
1261 if (Name == "ceil" && TLI->has(LibFunc::ceil))
1262 return ConstantFoldFP(ceil, V, Ty);
1263 else if (Name == "cos" && TLI->has(LibFunc::cos))
1264 return ConstantFoldFP(cos, V, Ty);
1265 else if (Name == "cosh" && TLI->has(LibFunc::cosh))
1266 return ConstantFoldFP(cosh, V, Ty);
1267 else if (Name == "cosf" && TLI->has(LibFunc::cosf))
1268 return ConstantFoldFP(cos, V, Ty);
1269 break;
1270 case 'e':
1271 if (Name == "exp" && TLI->has(LibFunc::exp))
1272 return ConstantFoldFP(exp, V, Ty);
1273
1274 if (Name == "exp2" && TLI->has(LibFunc::exp2)) {
1275 // Constant fold exp2(x) as pow(2,x) in case the host doesn't have a
1276 // C99 library.
1277 return ConstantFoldBinaryFP(pow, 2.0, V, Ty);
1278 }
1279 break;
1280 case 'f':
1281 if (Name == "fabs" && TLI->has(LibFunc::fabs))
1282 return ConstantFoldFP(fabs, V, Ty);
1283 else if (Name == "floor" && TLI->has(LibFunc::floor))
1284 return ConstantFoldFP(floor, V, Ty);
1285 break;
1286 case 'l':
1287 if (Name == "log" && V > 0 && TLI->has(LibFunc::log))
1288 return ConstantFoldFP(log, V, Ty);
1289 else if (Name == "log10" && V > 0 && TLI->has(LibFunc::log10))
1290 return ConstantFoldFP(log10, V, Ty);
1291 else if (F->getIntrinsicID() == Intrinsic::sqrt &&
1292 (Ty->isFloatTy() || Ty->isDoubleTy())) {
1293 if (V >= -0.0)
1294 return ConstantFoldFP(sqrt, V, Ty);
1295 else // Undefined
1296 return Constant::getNullValue(Ty);
1297 }
1298 break;
1299 case 's':
1300 if (Name == "sin" && TLI->has(LibFunc::sin))
1301 return ConstantFoldFP(sin, V, Ty);
1302 else if (Name == "sinh" && TLI->has(LibFunc::sinh))
1303 return ConstantFoldFP(sinh, V, Ty);
1304 else if (Name == "sqrt" && V >= 0 && TLI->has(LibFunc::sqrt))
1305 return ConstantFoldFP(sqrt, V, Ty);
1306 else if (Name == "sqrtf" && V >= 0 && TLI->has(LibFunc::sqrtf))
1307 return ConstantFoldFP(sqrt, V, Ty);
1308 else if (Name == "sinf" && TLI->has(LibFunc::sinf))
1309 return ConstantFoldFP(sin, V, Ty);
1310 break;
1311 case 't':
1312 if (Name == "tan" && TLI->has(LibFunc::tan))
1313 return ConstantFoldFP(tan, V, Ty);
1314 else if (Name == "tanh" && TLI->has(LibFunc::tanh))
1315 return ConstantFoldFP(tanh, V, Ty);
1316 break;
1317 default:
1318 break;
1319 }
1320 return 0;
1321 }
1322
1323 if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
1324 switch (F->getIntrinsicID()) {
1325 case Intrinsic::bswap:
1326 return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
1327 case Intrinsic::ctpop:
1328 return ConstantInt::get(Ty, Op->getValue().countPopulation());
1329 case Intrinsic::convert_from_fp16: {
1330 APFloat Val(Op->getValue());
1331
1332 bool lost = false;
1333 APFloat::opStatus status =
1334 Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
1335
1336 // Conversion is always precise.
1337 (void)status;
1338 assert(status == APFloat::opOK && !lost &&
1339 "Precision lost during fp16 constfolding");
1340
1341 return ConstantFP::get(F->getContext(), Val);
1342 }
1343 default:
1344 return 0;
1345 }
1346 }
1347
1348 // Support ConstantVector in case we have an Undef in the top.
1349 if (isa<ConstantVector>(Operands[0]) ||
1350 isa<ConstantDataVector>(Operands[0])) {
1351 Constant *Op = cast<Constant>(Operands[0]);
1352 switch (F->getIntrinsicID()) {
1353 default: break;
1354 case Intrinsic::x86_sse_cvtss2si:
1355 case Intrinsic::x86_sse_cvtss2si64:
1356 case Intrinsic::x86_sse2_cvtsd2si:
1357 case Intrinsic::x86_sse2_cvtsd2si64:
1358 if (ConstantFP *FPOp =
1359 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1360 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1361 /*roundTowardZero=*/false, Ty);
1362 case Intrinsic::x86_sse_cvttss2si:
1363 case Intrinsic::x86_sse_cvttss2si64:
1364 case Intrinsic::x86_sse2_cvttsd2si:
1365 case Intrinsic::x86_sse2_cvttsd2si64:
1366 if (ConstantFP *FPOp =
1367 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
1368 return ConstantFoldConvertToInt(FPOp->getValueAPF(),
1369 /*roundTowardZero=*/true, Ty);
1370 }
1371 }
1372
1373 if (isa<UndefValue>(Operands[0])) {
1374 if (F->getIntrinsicID() == Intrinsic::bswap)
1375 return Operands[0];
1376 return 0;
1377 }
1378
1379 return 0;
1380 }
1381
1382 if (Operands.size() == 2) {
1383 if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
1384 if (!Ty->isFloatTy() && !Ty->isDoubleTy())
1385 return 0;
1386 double Op1V = Ty->isFloatTy() ?
1387 (double)Op1->getValueAPF().convertToFloat() :
1388 Op1->getValueAPF().convertToDouble();
1389 if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
1390 if (Op2->getType() != Op1->getType())
1391 return 0;
1392
1393 double Op2V = Ty->isFloatTy() ?
1394 (double)Op2->getValueAPF().convertToFloat():
1395 Op2->getValueAPF().convertToDouble();
1396
1397 if (F->getIntrinsicID() == Intrinsic::pow) {
1398 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1399 }
1400 if (!TLI)
1401 return 0;
1402 if (Name == "pow" && TLI->has(LibFunc::pow))
1403 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
1404 if (Name == "fmod" && TLI->has(LibFunc::fmod))
1405 return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
1406 if (Name == "atan2" && TLI->has(LibFunc::atan2))
1407 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
1408 } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
1409 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isFloatTy())
1410 return ConstantFP::get(F->getContext(),
1411 APFloat((float)std::pow((float)Op1V,
1412 (int)Op2C->getZExtValue())));
1413 if (F->getIntrinsicID() == Intrinsic::powi && Ty->isDoubleTy())
1414 return ConstantFP::get(F->getContext(),
1415 APFloat((double)std::pow((double)Op1V,
1416 (int)Op2C->getZExtValue())));
1417 }
1418 return 0;
1419 }
1420
1421 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
1422 if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
1423 switch (F->getIntrinsicID()) {
1424 default: break;
1425 case Intrinsic::sadd_with_overflow:
1426 case Intrinsic::uadd_with_overflow:
1427 case Intrinsic::ssub_with_overflow:
1428 case Intrinsic::usub_with_overflow:
1429 case Intrinsic::smul_with_overflow:
1430 case Intrinsic::umul_with_overflow: {
1431 APInt Res;
1432 bool Overflow;
1433 switch (F->getIntrinsicID()) {
1434 default: llvm_unreachable("Invalid case");
1435 case Intrinsic::sadd_with_overflow:
1436 Res = Op1->getValue().sadd_ov(Op2->getValue(), Overflow);
1437 break;
1438 case Intrinsic::uadd_with_overflow:
1439 Res = Op1->getValue().uadd_ov(Op2->getValue(), Overflow);
1440 break;
1441 case Intrinsic::ssub_with_overflow:
1442 Res = Op1->getValue().ssub_ov(Op2->getValue(), Overflow);
1443 break;
1444 case Intrinsic::usub_with_overflow:
1445 Res = Op1->getValue().usub_ov(Op2->getValue(), Overflow);
1446 break;
1447 case Intrinsic::smul_with_overflow:
1448 Res = Op1->getValue().smul_ov(Op2->getValue(), Overflow);
1449 break;
1450 case Intrinsic::umul_with_overflow:
1451 Res = Op1->getValue().umul_ov(Op2->getValue(), Overflow);
1452 break;
1453 }
1454 Constant *Ops[] = {
1455 ConstantInt::get(F->getContext(), Res),
1456 ConstantInt::get(Type::getInt1Ty(F->getContext()), Overflow)
1457 };
1458 return ConstantStruct::get(cast<StructType>(F->getReturnType()), Ops);
1459 }
1460 case Intrinsic::cttz:
1461 // FIXME: This should check for Op2 == 1, and become unreachable if
1462 // Op1 == 0.
1463 return ConstantInt::get(Ty, Op1->getValue().countTrailingZeros());
1464 case Intrinsic::ctlz:
1465 // FIXME: This should check for Op2 == 1, and become unreachable if
1466 // Op1 == 0.
1467 return ConstantInt::get(Ty, Op1->getValue().countLeadingZeros());
1468 }
1469 }
1470
1471 return 0;
1472 }
1473 return 0;
1474 }
1475 return 0;
1476 }
1477