SCCP doesn't actually preserve the CFG. It will delete and insert terminator
[llvm.git] / lib / VMCore / ConstantFold.cpp
blob3fea1910ffe153b1c8325f4a255d51e34384af61
1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
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 implements folding of constants for LLVM. This implements the
11 // (internal) ConstantFold.h interface, which is used by the
12 // ConstantExpr::get* methods to automatically fold constants when possible.
14 // The current constant folding implementation is implemented in two pieces: the
15 // pieces that don't need TargetData, and the pieces that do. This is to avoid
16 // a dependence in VMCore on Target.
18 //===----------------------------------------------------------------------===//
20 #include "ConstantFold.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Support/MathExtras.h"
33 #include <limits>
34 using namespace llvm;
36 //===----------------------------------------------------------------------===//
37 // ConstantFold*Instruction Implementations
38 //===----------------------------------------------------------------------===//
40 /// BitCastConstantVector - Convert the specified ConstantVector node to the
41 /// specified vector type. At this point, we know that the elements of the
42 /// input vector constant are all simple integer or FP values.
43 static Constant *BitCastConstantVector(ConstantVector *CV,
44 const VectorType *DstTy) {
45 // If this cast changes element count then we can't handle it here:
46 // doing so requires endianness information. This should be handled by
47 // Analysis/ConstantFolding.cpp
48 unsigned NumElts = DstTy->getNumElements();
49 if (NumElts != CV->getNumOperands())
50 return 0;
52 // Check to verify that all elements of the input are simple.
53 for (unsigned i = 0; i != NumElts; ++i) {
54 if (!isa<ConstantInt>(CV->getOperand(i)) &&
55 !isa<ConstantFP>(CV->getOperand(i)))
56 return 0;
59 // Bitcast each element now.
60 std::vector<Constant*> Result;
61 const Type *DstEltTy = DstTy->getElementType();
62 for (unsigned i = 0; i != NumElts; ++i)
63 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i),
64 DstEltTy));
65 return ConstantVector::get(Result);
68 /// This function determines which opcode to use to fold two constant cast
69 /// expressions together. It uses CastInst::isEliminableCastPair to determine
70 /// the opcode. Consequently its just a wrapper around that function.
71 /// @brief Determine if it is valid to fold a cast of a cast
72 static unsigned
73 foldConstantCastPair(
74 unsigned opc, ///< opcode of the second cast constant expression
75 ConstantExpr *Op, ///< the first cast constant expression
76 const Type *DstTy ///< desintation type of the first cast
77 ) {
78 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
79 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
80 assert(CastInst::isCast(opc) && "Invalid cast opcode");
82 // The the types and opcodes for the two Cast constant expressions
83 const Type *SrcTy = Op->getOperand(0)->getType();
84 const Type *MidTy = Op->getType();
85 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
86 Instruction::CastOps secondOp = Instruction::CastOps(opc);
88 // Let CastInst::isEliminableCastPair do the heavy lifting.
89 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
90 Type::getInt64Ty(DstTy->getContext()));
93 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
94 const Type *SrcTy = V->getType();
95 if (SrcTy == DestTy)
96 return V; // no-op cast
98 // Check to see if we are casting a pointer to an aggregate to a pointer to
99 // the first element. If so, return the appropriate GEP instruction.
100 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
101 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
102 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
103 SmallVector<Value*, 8> IdxList;
104 Value *Zero =
105 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
106 IdxList.push_back(Zero);
107 const Type *ElTy = PTy->getElementType();
108 while (ElTy != DPTy->getElementType()) {
109 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
110 if (STy->getNumElements() == 0) break;
111 ElTy = STy->getElementType(0);
112 IdxList.push_back(Zero);
113 } else if (const SequentialType *STy =
114 dyn_cast<SequentialType>(ElTy)) {
115 if (ElTy->isPointerTy()) break; // Can't index into pointers!
116 ElTy = STy->getElementType();
117 IdxList.push_back(Zero);
118 } else {
119 break;
123 if (ElTy == DPTy->getElementType())
124 // This GEP is inbounds because all indices are zero.
125 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0],
126 IdxList.size());
129 // Handle casts from one vector constant to another. We know that the src
130 // and dest type have the same size (otherwise its an illegal cast).
131 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
132 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
133 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
134 "Not cast between same sized vectors!");
135 SrcTy = NULL;
136 // First, check for null. Undef is already handled.
137 if (isa<ConstantAggregateZero>(V))
138 return Constant::getNullValue(DestTy);
140 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
141 return BitCastConstantVector(CV, DestPTy);
144 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
145 // This allows for other simplifications (although some of them
146 // can only be handled by Analysis/ConstantFolding.cpp).
147 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
148 return ConstantExpr::getBitCast(ConstantVector::get(&V, 1), DestPTy);
151 // Finally, implement bitcast folding now. The code below doesn't handle
152 // bitcast right.
153 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
154 return ConstantPointerNull::get(cast<PointerType>(DestTy));
156 // Handle integral constant input.
157 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
158 if (DestTy->isIntegerTy())
159 // Integral -> Integral. This is a no-op because the bit widths must
160 // be the same. Consequently, we just fold to V.
161 return V;
163 if (DestTy->isFloatingPointTy())
164 return ConstantFP::get(DestTy->getContext(),
165 APFloat(CI->getValue(),
166 !DestTy->isPPC_FP128Ty()));
168 // Otherwise, can't fold this (vector?)
169 return 0;
172 // Handle ConstantFP input: FP -> Integral.
173 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
174 return ConstantInt::get(FP->getContext(),
175 FP->getValueAPF().bitcastToAPInt());
177 return 0;
181 /// ExtractConstantBytes - V is an integer constant which only has a subset of
182 /// its bytes used. The bytes used are indicated by ByteStart (which is the
183 /// first byte used, counting from the least significant byte) and ByteSize,
184 /// which is the number of bytes used.
186 /// This function analyzes the specified constant to see if the specified byte
187 /// range can be returned as a simplified constant. If so, the constant is
188 /// returned, otherwise null is returned.
189 ///
190 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
191 unsigned ByteSize) {
192 assert(C->getType()->isIntegerTy() &&
193 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
194 "Non-byte sized integer input");
195 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
196 assert(ByteSize && "Must be accessing some piece");
197 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
198 assert(ByteSize != CSize && "Should not extract everything");
200 // Constant Integers are simple.
201 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
202 APInt V = CI->getValue();
203 if (ByteStart)
204 V = V.lshr(ByteStart*8);
205 V = V.trunc(ByteSize*8);
206 return ConstantInt::get(CI->getContext(), V);
209 // In the input is a constant expr, we might be able to recursively simplify.
210 // If not, we definitely can't do anything.
211 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
212 if (CE == 0) return 0;
214 switch (CE->getOpcode()) {
215 default: return 0;
216 case Instruction::Or: {
217 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
218 if (RHS == 0)
219 return 0;
221 // X | -1 -> -1.
222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
223 if (RHSC->isAllOnesValue())
224 return RHSC;
226 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
227 if (LHS == 0)
228 return 0;
229 return ConstantExpr::getOr(LHS, RHS);
231 case Instruction::And: {
232 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
233 if (RHS == 0)
234 return 0;
236 // X & 0 -> 0.
237 if (RHS->isNullValue())
238 return RHS;
240 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
241 if (LHS == 0)
242 return 0;
243 return ConstantExpr::getAnd(LHS, RHS);
245 case Instruction::LShr: {
246 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
247 if (Amt == 0)
248 return 0;
249 unsigned ShAmt = Amt->getZExtValue();
250 // Cannot analyze non-byte shifts.
251 if ((ShAmt & 7) != 0)
252 return 0;
253 ShAmt >>= 3;
255 // If the extract is known to be all zeros, return zero.
256 if (ByteStart >= CSize-ShAmt)
257 return Constant::getNullValue(IntegerType::get(CE->getContext(),
258 ByteSize*8));
259 // If the extract is known to be fully in the input, extract it.
260 if (ByteStart+ByteSize+ShAmt <= CSize)
261 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
263 // TODO: Handle the 'partially zero' case.
264 return 0;
267 case Instruction::Shl: {
268 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
269 if (Amt == 0)
270 return 0;
271 unsigned ShAmt = Amt->getZExtValue();
272 // Cannot analyze non-byte shifts.
273 if ((ShAmt & 7) != 0)
274 return 0;
275 ShAmt >>= 3;
277 // If the extract is known to be all zeros, return zero.
278 if (ByteStart+ByteSize <= ShAmt)
279 return Constant::getNullValue(IntegerType::get(CE->getContext(),
280 ByteSize*8));
281 // If the extract is known to be fully in the input, extract it.
282 if (ByteStart >= ShAmt)
283 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
285 // TODO: Handle the 'partially zero' case.
286 return 0;
289 case Instruction::ZExt: {
290 unsigned SrcBitSize =
291 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
293 // If extracting something that is completely zero, return 0.
294 if (ByteStart*8 >= SrcBitSize)
295 return Constant::getNullValue(IntegerType::get(CE->getContext(),
296 ByteSize*8));
298 // If exactly extracting the input, return it.
299 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
300 return CE->getOperand(0);
302 // If extracting something completely in the input, if if the input is a
303 // multiple of 8 bits, recurse.
304 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
305 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
307 // Otherwise, if extracting a subset of the input, which is not multiple of
308 // 8 bits, do a shift and trunc to get the bits.
309 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
310 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
311 Constant *Res = CE->getOperand(0);
312 if (ByteStart)
313 Res = ConstantExpr::getLShr(Res,
314 ConstantInt::get(Res->getType(), ByteStart*8));
315 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
316 ByteSize*8));
319 // TODO: Handle the 'partially zero' case.
320 return 0;
325 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
326 /// on Ty, with any known factors factored out. If Folded is false,
327 /// return null if no factoring was possible, to avoid endlessly
328 /// bouncing an unfoldable expression back into the top-level folder.
330 static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy,
331 bool Folded) {
332 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
333 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
334 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
335 return ConstantExpr::getNUWMul(E, N);
338 if (const StructType *STy = dyn_cast<StructType>(Ty))
339 if (!STy->isPacked()) {
340 unsigned NumElems = STy->getNumElements();
341 // An empty struct has size zero.
342 if (NumElems == 0)
343 return ConstantExpr::getNullValue(DestTy);
344 // Check for a struct with all members having the same size.
345 Constant *MemberSize =
346 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
347 bool AllSame = true;
348 for (unsigned i = 1; i != NumElems; ++i)
349 if (MemberSize !=
350 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
351 AllSame = false;
352 break;
354 if (AllSame) {
355 Constant *N = ConstantInt::get(DestTy, NumElems);
356 return ConstantExpr::getNUWMul(MemberSize, N);
360 // Pointer size doesn't depend on the pointee type, so canonicalize them
361 // to an arbitrary pointee.
362 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
363 if (!PTy->getElementType()->isIntegerTy(1))
364 return
365 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
366 PTy->getAddressSpace()),
367 DestTy, true);
369 // If there's no interesting folding happening, bail so that we don't create
370 // a constant that looks like it needs folding but really doesn't.
371 if (!Folded)
372 return 0;
374 // Base case: Get a regular sizeof expression.
375 Constant *C = ConstantExpr::getSizeOf(Ty);
376 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
377 DestTy, false),
378 C, DestTy);
379 return C;
382 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
383 /// on Ty, with any known factors factored out. If Folded is false,
384 /// return null if no factoring was possible, to avoid endlessly
385 /// bouncing an unfoldable expression back into the top-level folder.
387 static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy,
388 bool Folded) {
389 // The alignment of an array is equal to the alignment of the
390 // array element. Note that this is not always true for vectors.
391 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
392 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
393 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
394 DestTy,
395 false),
396 C, DestTy);
397 return C;
400 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
401 // Packed structs always have an alignment of 1.
402 if (STy->isPacked())
403 return ConstantInt::get(DestTy, 1);
405 // Otherwise, struct alignment is the maximum alignment of any member.
406 // Without target data, we can't compare much, but we can check to see
407 // if all the members have the same alignment.
408 unsigned NumElems = STy->getNumElements();
409 // An empty struct has minimal alignment.
410 if (NumElems == 0)
411 return ConstantInt::get(DestTy, 1);
412 // Check for a struct with all members having the same alignment.
413 Constant *MemberAlign =
414 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
415 bool AllSame = true;
416 for (unsigned i = 1; i != NumElems; ++i)
417 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
418 AllSame = false;
419 break;
421 if (AllSame)
422 return MemberAlign;
425 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
426 // to an arbitrary pointee.
427 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
428 if (!PTy->getElementType()->isIntegerTy(1))
429 return
430 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
432 PTy->getAddressSpace()),
433 DestTy, true);
435 // If there's no interesting folding happening, bail so that we don't create
436 // a constant that looks like it needs folding but really doesn't.
437 if (!Folded)
438 return 0;
440 // Base case: Get a regular alignof expression.
441 Constant *C = ConstantExpr::getAlignOf(Ty);
442 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
443 DestTy, false),
444 C, DestTy);
445 return C;
448 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
449 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
450 /// return null if no factoring was possible, to avoid endlessly
451 /// bouncing an unfoldable expression back into the top-level folder.
453 static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo,
454 const Type *DestTy,
455 bool Folded) {
456 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
457 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
458 DestTy, false),
459 FieldNo, DestTy);
460 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
461 return ConstantExpr::getNUWMul(E, N);
464 if (const StructType *STy = dyn_cast<StructType>(Ty))
465 if (!STy->isPacked()) {
466 unsigned NumElems = STy->getNumElements();
467 // An empty struct has no members.
468 if (NumElems == 0)
469 return 0;
470 // Check for a struct with all members having the same size.
471 Constant *MemberSize =
472 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
473 bool AllSame = true;
474 for (unsigned i = 1; i != NumElems; ++i)
475 if (MemberSize !=
476 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
477 AllSame = false;
478 break;
480 if (AllSame) {
481 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
482 false,
483 DestTy,
484 false),
485 FieldNo, DestTy);
486 return ConstantExpr::getNUWMul(MemberSize, N);
490 // If there's no interesting folding happening, bail so that we don't create
491 // a constant that looks like it needs folding but really doesn't.
492 if (!Folded)
493 return 0;
495 // Base case: Get a regular offsetof expression.
496 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
497 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
498 DestTy, false),
499 C, DestTy);
500 return C;
503 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
504 const Type *DestTy) {
505 if (isa<UndefValue>(V)) {
506 // zext(undef) = 0, because the top bits will be zero.
507 // sext(undef) = 0, because the top bits will all be the same.
508 // [us]itofp(undef) = 0, because the result value is bounded.
509 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
510 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
511 return Constant::getNullValue(DestTy);
512 return UndefValue::get(DestTy);
515 // No compile-time operations on this type yet.
516 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
517 return 0;
519 if (V->isNullValue() && !DestTy->isX86_MMXTy())
520 return Constant::getNullValue(DestTy);
522 // If the cast operand is a constant expression, there's a few things we can
523 // do to try to simplify it.
524 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
525 if (CE->isCast()) {
526 // Try hard to fold cast of cast because they are often eliminable.
527 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
528 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
529 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
530 // If all of the indexes in the GEP are null values, there is no pointer
531 // adjustment going on. We might as well cast the source pointer.
532 bool isAllNull = true;
533 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
534 if (!CE->getOperand(i)->isNullValue()) {
535 isAllNull = false;
536 break;
538 if (isAllNull)
539 // This is casting one pointer type to another, always BitCast
540 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
544 // If the cast operand is a constant vector, perform the cast by
545 // operating on each element. In the cast of bitcasts, the element
546 // count may be mismatched; don't attempt to handle that here.
547 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
548 if (DestTy->isVectorTy() &&
549 cast<VectorType>(DestTy)->getNumElements() ==
550 CV->getType()->getNumElements()) {
551 std::vector<Constant*> res;
552 const VectorType *DestVecTy = cast<VectorType>(DestTy);
553 const Type *DstEltTy = DestVecTy->getElementType();
554 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
555 res.push_back(ConstantExpr::getCast(opc,
556 CV->getOperand(i), DstEltTy));
557 return ConstantVector::get(DestVecTy, res);
560 // We actually have to do a cast now. Perform the cast according to the
561 // opcode specified.
562 switch (opc) {
563 default:
564 llvm_unreachable("Failed to cast constant expression");
565 case Instruction::FPTrunc:
566 case Instruction::FPExt:
567 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
568 bool ignored;
569 APFloat Val = FPC->getValueAPF();
570 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
571 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
572 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
573 DestTy->isFP128Ty() ? APFloat::IEEEquad :
574 APFloat::Bogus,
575 APFloat::rmNearestTiesToEven, &ignored);
576 return ConstantFP::get(V->getContext(), Val);
578 return 0; // Can't fold.
579 case Instruction::FPToUI:
580 case Instruction::FPToSI:
581 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
582 const APFloat &V = FPC->getValueAPF();
583 bool ignored;
584 uint64_t x[2];
585 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
586 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
587 APFloat::rmTowardZero, &ignored);
588 APInt Val(DestBitWidth, 2, x);
589 return ConstantInt::get(FPC->getContext(), Val);
591 return 0; // Can't fold.
592 case Instruction::IntToPtr: //always treated as unsigned
593 if (V->isNullValue()) // Is it an integral null value?
594 return ConstantPointerNull::get(cast<PointerType>(DestTy));
595 return 0; // Other pointer types cannot be casted
596 case Instruction::PtrToInt: // always treated as unsigned
597 // Is it a null pointer value?
598 if (V->isNullValue())
599 return ConstantInt::get(DestTy, 0);
600 // If this is a sizeof-like expression, pull out multiplications by
601 // known factors to expose them to subsequent folding. If it's an
602 // alignof-like expression, factor out known factors.
603 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
604 if (CE->getOpcode() == Instruction::GetElementPtr &&
605 CE->getOperand(0)->isNullValue()) {
606 const Type *Ty =
607 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
608 if (CE->getNumOperands() == 2) {
609 // Handle a sizeof-like expression.
610 Constant *Idx = CE->getOperand(1);
611 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
612 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
613 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
614 DestTy, false),
615 Idx, DestTy);
616 return ConstantExpr::getMul(C, Idx);
618 } else if (CE->getNumOperands() == 3 &&
619 CE->getOperand(1)->isNullValue()) {
620 // Handle an alignof-like expression.
621 if (const StructType *STy = dyn_cast<StructType>(Ty))
622 if (!STy->isPacked()) {
623 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
624 if (CI->isOne() &&
625 STy->getNumElements() == 2 &&
626 STy->getElementType(0)->isIntegerTy(1)) {
627 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
630 // Handle an offsetof-like expression.
631 if (Ty->isStructTy() || Ty->isArrayTy()) {
632 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
633 DestTy, false))
634 return C;
638 // Other pointer types cannot be casted
639 return 0;
640 case Instruction::UIToFP:
641 case Instruction::SIToFP:
642 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
643 APInt api = CI->getValue();
644 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true);
645 (void)apf.convertFromAPInt(api,
646 opc==Instruction::SIToFP,
647 APFloat::rmNearestTiesToEven);
648 return ConstantFP::get(V->getContext(), apf);
650 return 0;
651 case Instruction::ZExt:
652 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
653 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
654 return ConstantInt::get(V->getContext(),
655 CI->getValue().zext(BitWidth));
657 return 0;
658 case Instruction::SExt:
659 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
660 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
661 return ConstantInt::get(V->getContext(),
662 CI->getValue().sext(BitWidth));
664 return 0;
665 case Instruction::Trunc: {
666 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
667 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
668 return ConstantInt::get(V->getContext(),
669 CI->getValue().trunc(DestBitWidth));
672 // The input must be a constantexpr. See if we can simplify this based on
673 // the bytes we are demanding. Only do this if the source and dest are an
674 // even multiple of a byte.
675 if ((DestBitWidth & 7) == 0 &&
676 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
677 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
678 return Res;
680 return 0;
682 case Instruction::BitCast:
683 return FoldBitCast(V, DestTy);
687 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
688 Constant *V1, Constant *V2) {
689 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
690 return CB->getZExtValue() ? V1 : V2;
692 if (isa<UndefValue>(V1)) return V2;
693 if (isa<UndefValue>(V2)) return V1;
694 if (isa<UndefValue>(Cond)) return V1;
695 if (V1 == V2) return V1;
696 return 0;
699 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
700 Constant *Idx) {
701 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
702 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
703 if (Val->isNullValue()) // ee(zero, x) -> zero
704 return Constant::getNullValue(
705 cast<VectorType>(Val->getType())->getElementType());
707 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
708 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
709 return CVal->getOperand(CIdx->getZExtValue());
710 } else if (isa<UndefValue>(Idx)) {
711 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
712 return CVal->getOperand(0);
715 return 0;
718 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
719 Constant *Elt,
720 Constant *Idx) {
721 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
722 if (!CIdx) return 0;
723 APInt idxVal = CIdx->getValue();
724 if (isa<UndefValue>(Val)) {
725 // Insertion of scalar constant into vector undef
726 // Optimize away insertion of undef
727 if (isa<UndefValue>(Elt))
728 return Val;
729 // Otherwise break the aggregate undef into multiple undefs and do
730 // the insertion
731 unsigned numOps =
732 cast<VectorType>(Val->getType())->getNumElements();
733 std::vector<Constant*> Ops;
734 Ops.reserve(numOps);
735 for (unsigned i = 0; i < numOps; ++i) {
736 Constant *Op =
737 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
738 Ops.push_back(Op);
740 return ConstantVector::get(Ops);
742 if (isa<ConstantAggregateZero>(Val)) {
743 // Insertion of scalar constant into vector aggregate zero
744 // Optimize away insertion of zero
745 if (Elt->isNullValue())
746 return Val;
747 // Otherwise break the aggregate zero into multiple zeros and do
748 // the insertion
749 unsigned numOps =
750 cast<VectorType>(Val->getType())->getNumElements();
751 std::vector<Constant*> Ops;
752 Ops.reserve(numOps);
753 for (unsigned i = 0; i < numOps; ++i) {
754 Constant *Op =
755 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
756 Ops.push_back(Op);
758 return ConstantVector::get(Ops);
760 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
761 // Insertion of scalar constant into vector constant
762 std::vector<Constant*> Ops;
763 Ops.reserve(CVal->getNumOperands());
764 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
765 Constant *Op =
766 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
767 Ops.push_back(Op);
769 return ConstantVector::get(Ops);
772 return 0;
775 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
776 /// return the specified element value. Otherwise return null.
777 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
778 if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
779 return CV->getOperand(EltNo);
781 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
782 if (isa<ConstantAggregateZero>(C))
783 return Constant::getNullValue(EltTy);
784 if (isa<UndefValue>(C))
785 return UndefValue::get(EltTy);
786 return 0;
789 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
790 Constant *V2,
791 Constant *Mask) {
792 // Undefined shuffle mask -> undefined value.
793 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
795 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
796 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
797 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
799 // Loop over the shuffle mask, evaluating each element.
800 SmallVector<Constant*, 32> Result;
801 for (unsigned i = 0; i != MaskNumElts; ++i) {
802 Constant *InElt = GetVectorElement(Mask, i);
803 if (InElt == 0) return 0;
805 if (isa<UndefValue>(InElt))
806 InElt = UndefValue::get(EltTy);
807 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
808 unsigned Elt = CI->getZExtValue();
809 if (Elt >= SrcNumElts*2)
810 InElt = UndefValue::get(EltTy);
811 else if (Elt >= SrcNumElts)
812 InElt = GetVectorElement(V2, Elt - SrcNumElts);
813 else
814 InElt = GetVectorElement(V1, Elt);
815 if (InElt == 0) return 0;
816 } else {
817 // Unknown value.
818 return 0;
820 Result.push_back(InElt);
823 return ConstantVector::get(&Result[0], Result.size());
826 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
827 const unsigned *Idxs,
828 unsigned NumIdx) {
829 // Base case: no indices, so return the entire value.
830 if (NumIdx == 0)
831 return Agg;
833 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
834 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
835 Idxs,
836 Idxs + NumIdx));
838 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
839 return
840 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
841 Idxs,
842 Idxs + NumIdx));
844 // Otherwise recurse.
845 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
846 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs),
847 Idxs+1, NumIdx-1);
849 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
850 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs),
851 Idxs+1, NumIdx-1);
852 ConstantVector *CV = cast<ConstantVector>(Agg);
853 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs),
854 Idxs+1, NumIdx-1);
857 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
858 Constant *Val,
859 const unsigned *Idxs,
860 unsigned NumIdx) {
861 // Base case: no indices, so replace the entire value.
862 if (NumIdx == 0)
863 return Val;
865 if (isa<UndefValue>(Agg)) {
866 // Insertion of constant into aggregate undef
867 // Optimize away insertion of undef.
868 if (isa<UndefValue>(Val))
869 return Agg;
871 // Otherwise break the aggregate undef into multiple undefs and do
872 // the insertion.
873 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
874 unsigned numOps;
875 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
876 numOps = AR->getNumElements();
877 else
878 numOps = cast<StructType>(AggTy)->getNumElements();
880 std::vector<Constant*> Ops(numOps);
881 for (unsigned i = 0; i < numOps; ++i) {
882 const Type *MemberTy = AggTy->getTypeAtIndex(i);
883 Constant *Op =
884 (*Idxs == i) ?
885 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
886 Val, Idxs+1, NumIdx-1) :
887 UndefValue::get(MemberTy);
888 Ops[i] = Op;
891 if (const StructType* ST = dyn_cast<StructType>(AggTy))
892 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
893 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
896 if (isa<ConstantAggregateZero>(Agg)) {
897 // Insertion of constant into aggregate zero
898 // Optimize away insertion of zero.
899 if (Val->isNullValue())
900 return Agg;
902 // Otherwise break the aggregate zero into multiple zeros and do
903 // the insertion.
904 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
905 unsigned numOps;
906 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
907 numOps = AR->getNumElements();
908 else
909 numOps = cast<StructType>(AggTy)->getNumElements();
911 std::vector<Constant*> Ops(numOps);
912 for (unsigned i = 0; i < numOps; ++i) {
913 const Type *MemberTy = AggTy->getTypeAtIndex(i);
914 Constant *Op =
915 (*Idxs == i) ?
916 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
917 Val, Idxs+1, NumIdx-1) :
918 Constant::getNullValue(MemberTy);
919 Ops[i] = Op;
922 if (const StructType *ST = dyn_cast<StructType>(AggTy))
923 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
924 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
927 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
928 // Insertion of constant into aggregate constant.
929 std::vector<Constant*> Ops(Agg->getNumOperands());
930 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
931 Constant *Op = cast<Constant>(Agg->getOperand(i));
932 if (*Idxs == i)
933 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1);
934 Ops[i] = Op;
937 if (const StructType* ST = dyn_cast<StructType>(Agg->getType()))
938 return ConstantStruct::get(ST->getContext(), Ops, ST->isPacked());
939 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
942 return 0;
946 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
947 Constant *C1, Constant *C2) {
948 // No compile-time operations on this type yet.
949 if (C1->getType()->isPPC_FP128Ty())
950 return 0;
952 // Handle UndefValue up front.
953 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
954 switch (Opcode) {
955 case Instruction::Xor:
956 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
957 // Handle undef ^ undef -> 0 special case. This is a common
958 // idiom (misuse).
959 return Constant::getNullValue(C1->getType());
960 // Fallthrough
961 case Instruction::Add:
962 case Instruction::Sub:
963 return UndefValue::get(C1->getType());
964 case Instruction::Mul:
965 case Instruction::And:
966 return Constant::getNullValue(C1->getType());
967 case Instruction::UDiv:
968 case Instruction::SDiv:
969 case Instruction::URem:
970 case Instruction::SRem:
971 if (!isa<UndefValue>(C2)) // undef / X -> 0
972 return Constant::getNullValue(C1->getType());
973 return C2; // X / undef -> undef
974 case Instruction::Or: // X | undef -> -1
975 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
976 return Constant::getAllOnesValue(PTy);
977 return Constant::getAllOnesValue(C1->getType());
978 case Instruction::LShr:
979 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
980 return C1; // undef lshr undef -> undef
981 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
982 // undef lshr X -> 0
983 case Instruction::AShr:
984 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
985 return Constant::getAllOnesValue(C1->getType());
986 else if (isa<UndefValue>(C1))
987 return C1; // undef ashr undef -> undef
988 else
989 return C1; // X ashr undef --> X
990 case Instruction::Shl:
991 // undef << X -> 0 or X << undef -> 0
992 return Constant::getNullValue(C1->getType());
996 // Handle simplifications when the RHS is a constant int.
997 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
998 switch (Opcode) {
999 case Instruction::Add:
1000 if (CI2->equalsInt(0)) return C1; // X + 0 == X
1001 break;
1002 case Instruction::Sub:
1003 if (CI2->equalsInt(0)) return C1; // X - 0 == X
1004 break;
1005 case Instruction::Mul:
1006 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
1007 if (CI2->equalsInt(1))
1008 return C1; // X * 1 == X
1009 break;
1010 case Instruction::UDiv:
1011 case Instruction::SDiv:
1012 if (CI2->equalsInt(1))
1013 return C1; // X / 1 == X
1014 if (CI2->equalsInt(0))
1015 return UndefValue::get(CI2->getType()); // X / 0 == undef
1016 break;
1017 case Instruction::URem:
1018 case Instruction::SRem:
1019 if (CI2->equalsInt(1))
1020 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1021 if (CI2->equalsInt(0))
1022 return UndefValue::get(CI2->getType()); // X % 0 == undef
1023 break;
1024 case Instruction::And:
1025 if (CI2->isZero()) return C2; // X & 0 == 0
1026 if (CI2->isAllOnesValue())
1027 return C1; // X & -1 == X
1029 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1030 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1031 if (CE1->getOpcode() == Instruction::ZExt) {
1032 unsigned DstWidth = CI2->getType()->getBitWidth();
1033 unsigned SrcWidth =
1034 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1035 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1036 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1037 return C1;
1040 // If and'ing the address of a global with a constant, fold it.
1041 if (CE1->getOpcode() == Instruction::PtrToInt &&
1042 isa<GlobalValue>(CE1->getOperand(0))) {
1043 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1045 // Functions are at least 4-byte aligned.
1046 unsigned GVAlign = GV->getAlignment();
1047 if (isa<Function>(GV))
1048 GVAlign = std::max(GVAlign, 4U);
1050 if (GVAlign > 1) {
1051 unsigned DstWidth = CI2->getType()->getBitWidth();
1052 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1053 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1055 // If checking bits we know are clear, return zero.
1056 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1057 return Constant::getNullValue(CI2->getType());
1061 break;
1062 case Instruction::Or:
1063 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1064 if (CI2->isAllOnesValue())
1065 return C2; // X | -1 == -1
1066 break;
1067 case Instruction::Xor:
1068 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1070 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1071 switch (CE1->getOpcode()) {
1072 default: break;
1073 case Instruction::ICmp:
1074 case Instruction::FCmp:
1075 // cmp pred ^ true -> cmp !pred
1076 assert(CI2->equalsInt(1));
1077 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1078 pred = CmpInst::getInversePredicate(pred);
1079 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1080 CE1->getOperand(1));
1083 break;
1084 case Instruction::AShr:
1085 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1086 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1087 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1088 return ConstantExpr::getLShr(C1, C2);
1089 break;
1091 } else if (isa<ConstantInt>(C1)) {
1092 // If C1 is a ConstantInt and C2 is not, swap the operands.
1093 if (Instruction::isCommutative(Opcode))
1094 return ConstantExpr::get(Opcode, C2, C1);
1097 // At this point we know neither constant is an UndefValue.
1098 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1099 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1100 using namespace APIntOps;
1101 const APInt &C1V = CI1->getValue();
1102 const APInt &C2V = CI2->getValue();
1103 switch (Opcode) {
1104 default:
1105 break;
1106 case Instruction::Add:
1107 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1108 case Instruction::Sub:
1109 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1110 case Instruction::Mul:
1111 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1112 case Instruction::UDiv:
1113 assert(!CI2->isNullValue() && "Div by zero handled above");
1114 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1115 case Instruction::SDiv:
1116 assert(!CI2->isNullValue() && "Div by zero handled above");
1117 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1118 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1119 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1120 case Instruction::URem:
1121 assert(!CI2->isNullValue() && "Div by zero handled above");
1122 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1123 case Instruction::SRem:
1124 assert(!CI2->isNullValue() && "Div by zero handled above");
1125 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1126 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1127 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1128 case Instruction::And:
1129 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1130 case Instruction::Or:
1131 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1132 case Instruction::Xor:
1133 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1134 case Instruction::Shl: {
1135 uint32_t shiftAmt = C2V.getZExtValue();
1136 if (shiftAmt < C1V.getBitWidth())
1137 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1138 else
1139 return UndefValue::get(C1->getType()); // too big shift is undef
1141 case Instruction::LShr: {
1142 uint32_t shiftAmt = C2V.getZExtValue();
1143 if (shiftAmt < C1V.getBitWidth())
1144 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1145 else
1146 return UndefValue::get(C1->getType()); // too big shift is undef
1148 case Instruction::AShr: {
1149 uint32_t shiftAmt = C2V.getZExtValue();
1150 if (shiftAmt < C1V.getBitWidth())
1151 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1152 else
1153 return UndefValue::get(C1->getType()); // too big shift is undef
1158 switch (Opcode) {
1159 case Instruction::SDiv:
1160 case Instruction::UDiv:
1161 case Instruction::URem:
1162 case Instruction::SRem:
1163 case Instruction::LShr:
1164 case Instruction::AShr:
1165 case Instruction::Shl:
1166 if (CI1->equalsInt(0)) return C1;
1167 break;
1168 default:
1169 break;
1171 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1172 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1173 APFloat C1V = CFP1->getValueAPF();
1174 APFloat C2V = CFP2->getValueAPF();
1175 APFloat C3V = C1V; // copy for modification
1176 switch (Opcode) {
1177 default:
1178 break;
1179 case Instruction::FAdd:
1180 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1181 return ConstantFP::get(C1->getContext(), C3V);
1182 case Instruction::FSub:
1183 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1184 return ConstantFP::get(C1->getContext(), C3V);
1185 case Instruction::FMul:
1186 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1187 return ConstantFP::get(C1->getContext(), C3V);
1188 case Instruction::FDiv:
1189 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1190 return ConstantFP::get(C1->getContext(), C3V);
1191 case Instruction::FRem:
1192 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1193 return ConstantFP::get(C1->getContext(), C3V);
1196 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1197 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1198 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1199 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1200 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1201 std::vector<Constant*> Res;
1202 const Type* EltTy = VTy->getElementType();
1203 Constant *C1 = 0;
1204 Constant *C2 = 0;
1205 switch (Opcode) {
1206 default:
1207 break;
1208 case Instruction::Add:
1209 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1210 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1211 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1212 Res.push_back(ConstantExpr::getAdd(C1, C2));
1214 return ConstantVector::get(Res);
1215 case Instruction::FAdd:
1216 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1217 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1218 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1219 Res.push_back(ConstantExpr::getFAdd(C1, C2));
1221 return ConstantVector::get(Res);
1222 case Instruction::Sub:
1223 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1224 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1225 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1226 Res.push_back(ConstantExpr::getSub(C1, C2));
1228 return ConstantVector::get(Res);
1229 case Instruction::FSub:
1230 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1231 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1232 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1233 Res.push_back(ConstantExpr::getFSub(C1, C2));
1235 return ConstantVector::get(Res);
1236 case Instruction::Mul:
1237 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1238 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1239 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1240 Res.push_back(ConstantExpr::getMul(C1, C2));
1242 return ConstantVector::get(Res);
1243 case Instruction::FMul:
1244 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1245 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1246 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1247 Res.push_back(ConstantExpr::getFMul(C1, C2));
1249 return ConstantVector::get(Res);
1250 case Instruction::UDiv:
1251 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1252 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1253 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1254 Res.push_back(ConstantExpr::getUDiv(C1, C2));
1256 return ConstantVector::get(Res);
1257 case Instruction::SDiv:
1258 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1259 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1260 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1261 Res.push_back(ConstantExpr::getSDiv(C1, C2));
1263 return ConstantVector::get(Res);
1264 case Instruction::FDiv:
1265 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1266 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1267 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1268 Res.push_back(ConstantExpr::getFDiv(C1, C2));
1270 return ConstantVector::get(Res);
1271 case Instruction::URem:
1272 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1273 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1274 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1275 Res.push_back(ConstantExpr::getURem(C1, C2));
1277 return ConstantVector::get(Res);
1278 case Instruction::SRem:
1279 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1280 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1281 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1282 Res.push_back(ConstantExpr::getSRem(C1, C2));
1284 return ConstantVector::get(Res);
1285 case Instruction::FRem:
1286 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1287 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1288 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1289 Res.push_back(ConstantExpr::getFRem(C1, C2));
1291 return ConstantVector::get(Res);
1292 case Instruction::And:
1293 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1294 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1295 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1296 Res.push_back(ConstantExpr::getAnd(C1, C2));
1298 return ConstantVector::get(Res);
1299 case Instruction::Or:
1300 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1301 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1302 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1303 Res.push_back(ConstantExpr::getOr(C1, C2));
1305 return ConstantVector::get(Res);
1306 case Instruction::Xor:
1307 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1308 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1309 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1310 Res.push_back(ConstantExpr::getXor(C1, C2));
1312 return ConstantVector::get(Res);
1313 case Instruction::LShr:
1314 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1315 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1316 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1317 Res.push_back(ConstantExpr::getLShr(C1, C2));
1319 return ConstantVector::get(Res);
1320 case Instruction::AShr:
1321 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1322 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1323 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1324 Res.push_back(ConstantExpr::getAShr(C1, C2));
1326 return ConstantVector::get(Res);
1327 case Instruction::Shl:
1328 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1329 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1330 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1331 Res.push_back(ConstantExpr::getShl(C1, C2));
1333 return ConstantVector::get(Res);
1338 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1339 // There are many possible foldings we could do here. We should probably
1340 // at least fold add of a pointer with an integer into the appropriate
1341 // getelementptr. This will improve alias analysis a bit.
1343 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1344 // (a + (b + c)).
1345 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1346 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1347 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1348 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1350 } else if (isa<ConstantExpr>(C2)) {
1351 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1352 // other way if possible.
1353 if (Instruction::isCommutative(Opcode))
1354 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1357 // i1 can be simplified in many cases.
1358 if (C1->getType()->isIntegerTy(1)) {
1359 switch (Opcode) {
1360 case Instruction::Add:
1361 case Instruction::Sub:
1362 return ConstantExpr::getXor(C1, C2);
1363 case Instruction::Mul:
1364 return ConstantExpr::getAnd(C1, C2);
1365 case Instruction::Shl:
1366 case Instruction::LShr:
1367 case Instruction::AShr:
1368 // We can assume that C2 == 0. If it were one the result would be
1369 // undefined because the shift value is as large as the bitwidth.
1370 return C1;
1371 case Instruction::SDiv:
1372 case Instruction::UDiv:
1373 // We can assume that C2 == 1. If it were zero the result would be
1374 // undefined through division by zero.
1375 return C1;
1376 case Instruction::URem:
1377 case Instruction::SRem:
1378 // We can assume that C2 == 1. If it were zero the result would be
1379 // undefined through division by zero.
1380 return ConstantInt::getFalse(C1->getContext());
1381 default:
1382 break;
1386 // We don't know how to fold this.
1387 return 0;
1390 /// isZeroSizedType - This type is zero sized if its an array or structure of
1391 /// zero sized types. The only leaf zero sized type is an empty structure.
1392 static bool isMaybeZeroSizedType(const Type *Ty) {
1393 if (Ty->isOpaqueTy()) return true; // Can't say.
1394 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1396 // If all of elements have zero size, this does too.
1397 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1398 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1399 return true;
1401 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1402 return isMaybeZeroSizedType(ATy->getElementType());
1404 return false;
1407 /// IdxCompare - Compare the two constants as though they were getelementptr
1408 /// indices. This allows coersion of the types to be the same thing.
1410 /// If the two constants are the "same" (after coersion), return 0. If the
1411 /// first is less than the second, return -1, if the second is less than the
1412 /// first, return 1. If the constants are not integral, return -2.
1414 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
1415 if (C1 == C2) return 0;
1417 // Ok, we found a different index. If they are not ConstantInt, we can't do
1418 // anything with them.
1419 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1420 return -2; // don't know!
1422 // Ok, we have two differing integer indices. Sign extend them to be the same
1423 // type. Long is always big enough, so we use it.
1424 if (!C1->getType()->isIntegerTy(64))
1425 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1427 if (!C2->getType()->isIntegerTy(64))
1428 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1430 if (C1 == C2) return 0; // They are equal
1432 // If the type being indexed over is really just a zero sized type, there is
1433 // no pointer difference being made here.
1434 if (isMaybeZeroSizedType(ElTy))
1435 return -2; // dunno.
1437 // If they are really different, now that they are the same type, then we
1438 // found a difference!
1439 if (cast<ConstantInt>(C1)->getSExtValue() <
1440 cast<ConstantInt>(C2)->getSExtValue())
1441 return -1;
1442 else
1443 return 1;
1446 /// evaluateFCmpRelation - This function determines if there is anything we can
1447 /// decide about the two constants provided. This doesn't need to handle simple
1448 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1449 /// If we can determine that the two constants have a particular relation to
1450 /// each other, we should return the corresponding FCmpInst predicate,
1451 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1452 /// ConstantFoldCompareInstruction.
1454 /// To simplify this code we canonicalize the relation so that the first
1455 /// operand is always the most "complex" of the two. We consider ConstantFP
1456 /// to be the simplest, and ConstantExprs to be the most complex.
1457 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1458 assert(V1->getType() == V2->getType() &&
1459 "Cannot compare values of different types!");
1461 // No compile-time operations on this type yet.
1462 if (V1->getType()->isPPC_FP128Ty())
1463 return FCmpInst::BAD_FCMP_PREDICATE;
1465 // Handle degenerate case quickly
1466 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1468 if (!isa<ConstantExpr>(V1)) {
1469 if (!isa<ConstantExpr>(V2)) {
1470 // We distilled thisUse the standard constant folder for a few cases
1471 ConstantInt *R = 0;
1472 R = dyn_cast<ConstantInt>(
1473 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1474 if (R && !R->isZero())
1475 return FCmpInst::FCMP_OEQ;
1476 R = dyn_cast<ConstantInt>(
1477 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1478 if (R && !R->isZero())
1479 return FCmpInst::FCMP_OLT;
1480 R = dyn_cast<ConstantInt>(
1481 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1482 if (R && !R->isZero())
1483 return FCmpInst::FCMP_OGT;
1485 // Nothing more we can do
1486 return FCmpInst::BAD_FCMP_PREDICATE;
1489 // If the first operand is simple and second is ConstantExpr, swap operands.
1490 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1491 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1492 return FCmpInst::getSwappedPredicate(SwappedRelation);
1493 } else {
1494 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1495 // constantexpr or a simple constant.
1496 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1497 switch (CE1->getOpcode()) {
1498 case Instruction::FPTrunc:
1499 case Instruction::FPExt:
1500 case Instruction::UIToFP:
1501 case Instruction::SIToFP:
1502 // We might be able to do something with these but we don't right now.
1503 break;
1504 default:
1505 break;
1508 // There are MANY other foldings that we could perform here. They will
1509 // probably be added on demand, as they seem needed.
1510 return FCmpInst::BAD_FCMP_PREDICATE;
1513 /// evaluateICmpRelation - This function determines if there is anything we can
1514 /// decide about the two constants provided. This doesn't need to handle simple
1515 /// things like integer comparisons, but should instead handle ConstantExprs
1516 /// and GlobalValues. If we can determine that the two constants have a
1517 /// particular relation to each other, we should return the corresponding ICmp
1518 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1520 /// To simplify this code we canonicalize the relation so that the first
1521 /// operand is always the most "complex" of the two. We consider simple
1522 /// constants (like ConstantInt) to be the simplest, followed by
1523 /// GlobalValues, followed by ConstantExpr's (the most complex).
1525 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1526 bool isSigned) {
1527 assert(V1->getType() == V2->getType() &&
1528 "Cannot compare different types of values!");
1529 if (V1 == V2) return ICmpInst::ICMP_EQ;
1531 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1532 !isa<BlockAddress>(V1)) {
1533 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1534 !isa<BlockAddress>(V2)) {
1535 // We distilled this down to a simple case, use the standard constant
1536 // folder.
1537 ConstantInt *R = 0;
1538 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1539 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1540 if (R && !R->isZero())
1541 return pred;
1542 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1543 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1544 if (R && !R->isZero())
1545 return pred;
1546 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1547 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1548 if (R && !R->isZero())
1549 return pred;
1551 // If we couldn't figure it out, bail.
1552 return ICmpInst::BAD_ICMP_PREDICATE;
1555 // If the first operand is simple, swap operands.
1556 ICmpInst::Predicate SwappedRelation =
1557 evaluateICmpRelation(V2, V1, isSigned);
1558 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1559 return ICmpInst::getSwappedPredicate(SwappedRelation);
1561 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1562 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1563 ICmpInst::Predicate SwappedRelation =
1564 evaluateICmpRelation(V2, V1, isSigned);
1565 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1566 return ICmpInst::getSwappedPredicate(SwappedRelation);
1567 return ICmpInst::BAD_ICMP_PREDICATE;
1570 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1571 // constant (which, since the types must match, means that it's a
1572 // ConstantPointerNull).
1573 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1574 // Don't try to decide equality of aliases.
1575 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1576 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1577 return ICmpInst::ICMP_NE;
1578 } else if (isa<BlockAddress>(V2)) {
1579 return ICmpInst::ICMP_NE; // Globals never equal labels.
1580 } else {
1581 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1582 // GlobalVals can never be null unless they have external weak linkage.
1583 // We don't try to evaluate aliases here.
1584 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1585 return ICmpInst::ICMP_NE;
1587 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1588 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1589 ICmpInst::Predicate SwappedRelation =
1590 evaluateICmpRelation(V2, V1, isSigned);
1591 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1592 return ICmpInst::getSwappedPredicate(SwappedRelation);
1593 return ICmpInst::BAD_ICMP_PREDICATE;
1596 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1597 // constant (which, since the types must match, means that it is a
1598 // ConstantPointerNull).
1599 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1600 // Block address in another function can't equal this one, but block
1601 // addresses in the current function might be the same if blocks are
1602 // empty.
1603 if (BA2->getFunction() != BA->getFunction())
1604 return ICmpInst::ICMP_NE;
1605 } else {
1606 // Block addresses aren't null, don't equal the address of globals.
1607 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1608 "Canonicalization guarantee!");
1609 return ICmpInst::ICMP_NE;
1611 } else {
1612 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1613 // constantexpr, a global, block address, or a simple constant.
1614 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1615 Constant *CE1Op0 = CE1->getOperand(0);
1617 switch (CE1->getOpcode()) {
1618 case Instruction::Trunc:
1619 case Instruction::FPTrunc:
1620 case Instruction::FPExt:
1621 case Instruction::FPToUI:
1622 case Instruction::FPToSI:
1623 break; // We can't evaluate floating point casts or truncations.
1625 case Instruction::UIToFP:
1626 case Instruction::SIToFP:
1627 case Instruction::BitCast:
1628 case Instruction::ZExt:
1629 case Instruction::SExt:
1630 // If the cast is not actually changing bits, and the second operand is a
1631 // null pointer, do the comparison with the pre-casted value.
1632 if (V2->isNullValue() &&
1633 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1634 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1635 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1636 return evaluateICmpRelation(CE1Op0,
1637 Constant::getNullValue(CE1Op0->getType()),
1638 isSigned);
1640 break;
1642 case Instruction::GetElementPtr:
1643 // Ok, since this is a getelementptr, we know that the constant has a
1644 // pointer type. Check the various cases.
1645 if (isa<ConstantPointerNull>(V2)) {
1646 // If we are comparing a GEP to a null pointer, check to see if the base
1647 // of the GEP equals the null pointer.
1648 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1649 if (GV->hasExternalWeakLinkage())
1650 // Weak linkage GVals could be zero or not. We're comparing that
1651 // to null pointer so its greater-or-equal
1652 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1653 else
1654 // If its not weak linkage, the GVal must have a non-zero address
1655 // so the result is greater-than
1656 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1657 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1658 // If we are indexing from a null pointer, check to see if we have any
1659 // non-zero indices.
1660 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1661 if (!CE1->getOperand(i)->isNullValue())
1662 // Offsetting from null, must not be equal.
1663 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1664 // Only zero indexes from null, must still be zero.
1665 return ICmpInst::ICMP_EQ;
1667 // Otherwise, we can't really say if the first operand is null or not.
1668 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1669 if (isa<ConstantPointerNull>(CE1Op0)) {
1670 if (GV2->hasExternalWeakLinkage())
1671 // Weak linkage GVals could be zero or not. We're comparing it to
1672 // a null pointer, so its less-or-equal
1673 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1674 else
1675 // If its not weak linkage, the GVal must have a non-zero address
1676 // so the result is less-than
1677 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1678 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1679 if (GV == GV2) {
1680 // If this is a getelementptr of the same global, then it must be
1681 // different. Because the types must match, the getelementptr could
1682 // only have at most one index, and because we fold getelementptr's
1683 // with a single zero index, it must be nonzero.
1684 assert(CE1->getNumOperands() == 2 &&
1685 !CE1->getOperand(1)->isNullValue() &&
1686 "Suprising getelementptr!");
1687 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1688 } else {
1689 // If they are different globals, we don't know what the value is,
1690 // but they can't be equal.
1691 return ICmpInst::ICMP_NE;
1694 } else {
1695 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1696 Constant *CE2Op0 = CE2->getOperand(0);
1698 // There are MANY other foldings that we could perform here. They will
1699 // probably be added on demand, as they seem needed.
1700 switch (CE2->getOpcode()) {
1701 default: break;
1702 case Instruction::GetElementPtr:
1703 // By far the most common case to handle is when the base pointers are
1704 // obviously to the same or different globals.
1705 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1706 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1707 return ICmpInst::ICMP_NE;
1708 // Ok, we know that both getelementptr instructions are based on the
1709 // same global. From this, we can precisely determine the relative
1710 // ordering of the resultant pointers.
1711 unsigned i = 1;
1713 // The logic below assumes that the result of the comparison
1714 // can be determined by finding the first index that differs.
1715 // This doesn't work if there is over-indexing in any
1716 // subsequent indices, so check for that case first.
1717 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1718 !CE2->isGEPWithNoNotionalOverIndexing())
1719 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1721 // Compare all of the operands the GEP's have in common.
1722 gep_type_iterator GTI = gep_type_begin(CE1);
1723 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1724 ++i, ++GTI)
1725 switch (IdxCompare(CE1->getOperand(i),
1726 CE2->getOperand(i), GTI.getIndexedType())) {
1727 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1728 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1729 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1732 // Ok, we ran out of things they have in common. If any leftovers
1733 // are non-zero then we have a difference, otherwise we are equal.
1734 for (; i < CE1->getNumOperands(); ++i)
1735 if (!CE1->getOperand(i)->isNullValue()) {
1736 if (isa<ConstantInt>(CE1->getOperand(i)))
1737 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1738 else
1739 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1742 for (; i < CE2->getNumOperands(); ++i)
1743 if (!CE2->getOperand(i)->isNullValue()) {
1744 if (isa<ConstantInt>(CE2->getOperand(i)))
1745 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1746 else
1747 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1749 return ICmpInst::ICMP_EQ;
1753 default:
1754 break;
1758 return ICmpInst::BAD_ICMP_PREDICATE;
1761 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1762 Constant *C1, Constant *C2) {
1763 const Type *ResultTy;
1764 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1765 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1766 VT->getNumElements());
1767 else
1768 ResultTy = Type::getInt1Ty(C1->getContext());
1770 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1771 if (pred == FCmpInst::FCMP_FALSE)
1772 return Constant::getNullValue(ResultTy);
1774 if (pred == FCmpInst::FCMP_TRUE)
1775 return Constant::getAllOnesValue(ResultTy);
1777 // Handle some degenerate cases first
1778 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1779 // For EQ and NE, we can always pick a value for the undef to make the
1780 // predicate pass or fail, so we can return undef.
1781 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)))
1782 return UndefValue::get(ResultTy);
1783 // Otherwise, pick the same value as the non-undef operand, and fold
1784 // it to true or false.
1785 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1788 // No compile-time operations on this type yet.
1789 if (C1->getType()->isPPC_FP128Ty())
1790 return 0;
1792 // icmp eq/ne(null,GV) -> false/true
1793 if (C1->isNullValue()) {
1794 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1795 // Don't try to evaluate aliases. External weak GV can be null.
1796 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1797 if (pred == ICmpInst::ICMP_EQ)
1798 return ConstantInt::getFalse(C1->getContext());
1799 else if (pred == ICmpInst::ICMP_NE)
1800 return ConstantInt::getTrue(C1->getContext());
1802 // icmp eq/ne(GV,null) -> false/true
1803 } else if (C2->isNullValue()) {
1804 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1805 // Don't try to evaluate aliases. External weak GV can be null.
1806 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1807 if (pred == ICmpInst::ICMP_EQ)
1808 return ConstantInt::getFalse(C1->getContext());
1809 else if (pred == ICmpInst::ICMP_NE)
1810 return ConstantInt::getTrue(C1->getContext());
1814 // If the comparison is a comparison between two i1's, simplify it.
1815 if (C1->getType()->isIntegerTy(1)) {
1816 switch(pred) {
1817 case ICmpInst::ICMP_EQ:
1818 if (isa<ConstantInt>(C2))
1819 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1820 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1821 case ICmpInst::ICMP_NE:
1822 return ConstantExpr::getXor(C1, C2);
1823 default:
1824 break;
1828 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1829 APInt V1 = cast<ConstantInt>(C1)->getValue();
1830 APInt V2 = cast<ConstantInt>(C2)->getValue();
1831 switch (pred) {
1832 default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1833 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1834 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1835 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1836 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1837 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1838 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1839 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1840 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1841 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1842 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1844 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1845 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1846 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1847 APFloat::cmpResult R = C1V.compare(C2V);
1848 switch (pred) {
1849 default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1850 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1851 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1852 case FCmpInst::FCMP_UNO:
1853 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1854 case FCmpInst::FCMP_ORD:
1855 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1856 case FCmpInst::FCMP_UEQ:
1857 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1858 R==APFloat::cmpEqual);
1859 case FCmpInst::FCMP_OEQ:
1860 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1861 case FCmpInst::FCMP_UNE:
1862 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1863 case FCmpInst::FCMP_ONE:
1864 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1865 R==APFloat::cmpGreaterThan);
1866 case FCmpInst::FCMP_ULT:
1867 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1868 R==APFloat::cmpLessThan);
1869 case FCmpInst::FCMP_OLT:
1870 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1871 case FCmpInst::FCMP_UGT:
1872 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1873 R==APFloat::cmpGreaterThan);
1874 case FCmpInst::FCMP_OGT:
1875 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1876 case FCmpInst::FCMP_ULE:
1877 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1878 case FCmpInst::FCMP_OLE:
1879 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1880 R==APFloat::cmpEqual);
1881 case FCmpInst::FCMP_UGE:
1882 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1883 case FCmpInst::FCMP_OGE:
1884 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1885 R==APFloat::cmpEqual);
1887 } else if (C1->getType()->isVectorTy()) {
1888 SmallVector<Constant*, 16> C1Elts, C2Elts;
1889 C1->getVectorElements(C1Elts);
1890 C2->getVectorElements(C2Elts);
1891 if (C1Elts.empty() || C2Elts.empty())
1892 return 0;
1894 // If we can constant fold the comparison of each element, constant fold
1895 // the whole vector comparison.
1896 SmallVector<Constant*, 4> ResElts;
1897 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i) {
1898 // Compare the elements, producing an i1 result or constant expr.
1899 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1901 return ConstantVector::get(&ResElts[0], ResElts.size());
1904 if (C1->getType()->isFloatingPointTy()) {
1905 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1906 switch (evaluateFCmpRelation(C1, C2)) {
1907 default: llvm_unreachable("Unknown relation!");
1908 case FCmpInst::FCMP_UNO:
1909 case FCmpInst::FCMP_ORD:
1910 case FCmpInst::FCMP_UEQ:
1911 case FCmpInst::FCMP_UNE:
1912 case FCmpInst::FCMP_ULT:
1913 case FCmpInst::FCMP_UGT:
1914 case FCmpInst::FCMP_ULE:
1915 case FCmpInst::FCMP_UGE:
1916 case FCmpInst::FCMP_TRUE:
1917 case FCmpInst::FCMP_FALSE:
1918 case FCmpInst::BAD_FCMP_PREDICATE:
1919 break; // Couldn't determine anything about these constants.
1920 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1921 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1922 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1923 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1924 break;
1925 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1926 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1927 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1928 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1929 break;
1930 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1931 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1932 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1933 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1934 break;
1935 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1936 // We can only partially decide this relation.
1937 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1938 Result = 0;
1939 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1940 Result = 1;
1941 break;
1942 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1943 // We can only partially decide this relation.
1944 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1945 Result = 0;
1946 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1947 Result = 1;
1948 break;
1949 case ICmpInst::ICMP_NE: // We know that C1 != C2
1950 // We can only partially decide this relation.
1951 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1952 Result = 0;
1953 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1954 Result = 1;
1955 break;
1958 // If we evaluated the result, return it now.
1959 if (Result != -1)
1960 return ConstantInt::get(ResultTy, Result);
1962 } else {
1963 // Evaluate the relation between the two constants, per the predicate.
1964 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1965 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1966 default: llvm_unreachable("Unknown relational!");
1967 case ICmpInst::BAD_ICMP_PREDICATE:
1968 break; // Couldn't determine anything about these constants.
1969 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1970 // If we know the constants are equal, we can decide the result of this
1971 // computation precisely.
1972 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1973 break;
1974 case ICmpInst::ICMP_ULT:
1975 switch (pred) {
1976 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1977 Result = 1; break;
1978 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1979 Result = 0; break;
1981 break;
1982 case ICmpInst::ICMP_SLT:
1983 switch (pred) {
1984 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1985 Result = 1; break;
1986 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1987 Result = 0; break;
1989 break;
1990 case ICmpInst::ICMP_UGT:
1991 switch (pred) {
1992 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1993 Result = 1; break;
1994 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1995 Result = 0; break;
1997 break;
1998 case ICmpInst::ICMP_SGT:
1999 switch (pred) {
2000 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2001 Result = 1; break;
2002 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2003 Result = 0; break;
2005 break;
2006 case ICmpInst::ICMP_ULE:
2007 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2008 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2009 break;
2010 case ICmpInst::ICMP_SLE:
2011 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2012 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2013 break;
2014 case ICmpInst::ICMP_UGE:
2015 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2016 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2017 break;
2018 case ICmpInst::ICMP_SGE:
2019 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2020 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2021 break;
2022 case ICmpInst::ICMP_NE:
2023 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2024 if (pred == ICmpInst::ICMP_NE) Result = 1;
2025 break;
2028 // If we evaluated the result, return it now.
2029 if (Result != -1)
2030 return ConstantInt::get(ResultTy, Result);
2032 // If the right hand side is a bitcast, try using its inverse to simplify
2033 // it by moving it to the left hand side. We can't do this if it would turn
2034 // a vector compare into a scalar compare or visa versa.
2035 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2036 Constant *CE2Op0 = CE2->getOperand(0);
2037 if (CE2->getOpcode() == Instruction::BitCast &&
2038 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
2039 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2040 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2044 // If the left hand side is an extension, try eliminating it.
2045 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2046 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
2047 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
2048 Constant *CE1Op0 = CE1->getOperand(0);
2049 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2050 if (CE1Inverse == CE1Op0) {
2051 // Check whether we can safely truncate the right hand side.
2052 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2053 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2054 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2060 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2061 (C1->isNullValue() && !C2->isNullValue())) {
2062 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2063 // other way if possible.
2064 // Also, if C1 is null and C2 isn't, flip them around.
2065 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2066 return ConstantExpr::getICmp(pred, C2, C1);
2069 return 0;
2072 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2073 /// is "inbounds".
2074 template<typename IndexTy>
2075 static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) {
2076 // No indices means nothing that could be out of bounds.
2077 if (NumIdx == 0) return true;
2079 // If the first index is zero, it's in bounds.
2080 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2082 // If the first index is one and all the rest are zero, it's in bounds,
2083 // by the one-past-the-end rule.
2084 if (!cast<ConstantInt>(Idxs[0])->isOne())
2085 return false;
2086 for (unsigned i = 1, e = NumIdx; i != e; ++i)
2087 if (!cast<Constant>(Idxs[i])->isNullValue())
2088 return false;
2089 return true;
2092 template<typename IndexTy>
2093 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
2094 bool inBounds,
2095 IndexTy const *Idxs,
2096 unsigned NumIdx) {
2097 Constant *Idx0 = cast<Constant>(Idxs[0]);
2098 if (NumIdx == 0 ||
2099 (NumIdx == 1 && Idx0->isNullValue()))
2100 return C;
2102 if (isa<UndefValue>(C)) {
2103 const PointerType *Ptr = cast<PointerType>(C->getType());
2104 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, Idxs+NumIdx);
2105 assert(Ty != 0 && "Invalid indices for GEP!");
2106 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2109 if (C->isNullValue()) {
2110 bool isNull = true;
2111 for (unsigned i = 0, e = NumIdx; i != e; ++i)
2112 if (!cast<Constant>(Idxs[i])->isNullValue()) {
2113 isNull = false;
2114 break;
2116 if (isNull) {
2117 const PointerType *Ptr = cast<PointerType>(C->getType());
2118 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs,
2119 Idxs+NumIdx);
2120 assert(Ty != 0 && "Invalid indices for GEP!");
2121 return ConstantPointerNull::get(
2122 PointerType::get(Ty,Ptr->getAddressSpace()));
2126 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2127 // Combine Indices - If the source pointer to this getelementptr instruction
2128 // is a getelementptr instruction, combine the indices of the two
2129 // getelementptr instructions into a single instruction.
2131 if (CE->getOpcode() == Instruction::GetElementPtr) {
2132 const Type *LastTy = 0;
2133 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2134 I != E; ++I)
2135 LastTy = *I;
2137 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) {
2138 SmallVector<Value*, 16> NewIndices;
2139 NewIndices.reserve(NumIdx + CE->getNumOperands());
2140 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2141 NewIndices.push_back(CE->getOperand(i));
2143 // Add the last index of the source with the first index of the new GEP.
2144 // Make sure to handle the case when they are actually different types.
2145 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2146 // Otherwise it must be an array.
2147 if (!Idx0->isNullValue()) {
2148 const Type *IdxTy = Combined->getType();
2149 if (IdxTy != Idx0->getType()) {
2150 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2151 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2152 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2153 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2154 } else {
2155 Combined =
2156 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2160 NewIndices.push_back(Combined);
2161 NewIndices.append(Idxs+1, Idxs+NumIdx);
2162 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ?
2163 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0),
2164 &NewIndices[0],
2165 NewIndices.size()) :
2166 ConstantExpr::getGetElementPtr(CE->getOperand(0),
2167 &NewIndices[0],
2168 NewIndices.size());
2172 // Implement folding of:
2173 // int* getelementptr ([2 x int]* bitcast ([3 x int]* %X to [2 x int]*),
2174 // long 0, long 0)
2175 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
2177 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
2178 if (const PointerType *SPT =
2179 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2180 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2181 if (const ArrayType *CAT =
2182 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2183 if (CAT->getElementType() == SAT->getElementType())
2184 return inBounds ?
2185 ConstantExpr::getInBoundsGetElementPtr(
2186 (Constant*)CE->getOperand(0), Idxs, NumIdx) :
2187 ConstantExpr::getGetElementPtr(
2188 (Constant*)CE->getOperand(0), Idxs, NumIdx);
2192 // Check to see if any array indices are not within the corresponding
2193 // notional array bounds. If so, try to determine if they can be factored
2194 // out into preceding dimensions.
2195 bool Unknown = false;
2196 SmallVector<Constant *, 8> NewIdxs;
2197 const Type *Ty = C->getType();
2198 const Type *Prev = 0;
2199 for (unsigned i = 0; i != NumIdx;
2200 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2201 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2202 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2203 if (ATy->getNumElements() <= INT64_MAX &&
2204 ATy->getNumElements() != 0 &&
2205 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2206 if (isa<SequentialType>(Prev)) {
2207 // It's out of range, but we can factor it into the prior
2208 // dimension.
2209 NewIdxs.resize(NumIdx);
2210 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2211 ATy->getNumElements());
2212 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2214 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2215 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2217 // Before adding, extend both operands to i64 to avoid
2218 // overflow trouble.
2219 if (!PrevIdx->getType()->isIntegerTy(64))
2220 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2221 Type::getInt64Ty(Div->getContext()));
2222 if (!Div->getType()->isIntegerTy(64))
2223 Div = ConstantExpr::getSExt(Div,
2224 Type::getInt64Ty(Div->getContext()));
2226 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2227 } else {
2228 // It's out of range, but the prior dimension is a struct
2229 // so we can't do anything about it.
2230 Unknown = true;
2233 } else {
2234 // We don't know if it's in range or not.
2235 Unknown = true;
2239 // If we did any factoring, start over with the adjusted indices.
2240 if (!NewIdxs.empty()) {
2241 for (unsigned i = 0; i != NumIdx; ++i)
2242 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2243 return inBounds ?
2244 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(),
2245 NewIdxs.size()) :
2246 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size());
2249 // If all indices are known integers and normalized, we can do a simple
2250 // check for the "inbounds" property.
2251 if (!Unknown && !inBounds &&
2252 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx))
2253 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx);
2255 return 0;
2258 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2259 bool inBounds,
2260 Constant* const *Idxs,
2261 unsigned NumIdx) {
2262 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);
2265 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2266 bool inBounds,
2267 Value* const *Idxs,
2268 unsigned NumIdx) {
2269 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);