Land the long talked about "type system rewrite" patch. This
[llvm/stm8.git] / lib / VMCore / ConstantFold.cpp
blob2c8f430176369fe4888f1f9cd804c4c2248bef28
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/Operator.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/ManagedStatic.h"
33 #include "llvm/Support/MathExtras.h"
34 #include <limits>
35 using namespace llvm;
37 //===----------------------------------------------------------------------===//
38 // ConstantFold*Instruction Implementations
39 //===----------------------------------------------------------------------===//
41 /// BitCastConstantVector - Convert the specified ConstantVector node to the
42 /// specified vector type. At this point, we know that the elements of the
43 /// input vector constant are all simple integer or FP values.
44 static Constant *BitCastConstantVector(ConstantVector *CV,
45 const VectorType *DstTy) {
47 if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48 if (CV->isNullValue()) return Constant::getNullValue(DstTy);
50 // If this cast changes element count then we can't handle it here:
51 // doing so requires endianness information. This should be handled by
52 // Analysis/ConstantFolding.cpp
53 unsigned NumElts = DstTy->getNumElements();
54 if (NumElts != CV->getNumOperands())
55 return 0;
57 // Check to verify that all elements of the input are simple.
58 for (unsigned i = 0; i != NumElts; ++i) {
59 if (!isa<ConstantInt>(CV->getOperand(i)) &&
60 !isa<ConstantFP>(CV->getOperand(i)))
61 return 0;
64 // Bitcast each element now.
65 std::vector<Constant*> Result;
66 const Type *DstEltTy = DstTy->getElementType();
67 for (unsigned i = 0; i != NumElts; ++i)
68 Result.push_back(ConstantExpr::getBitCast(CV->getOperand(i),
69 DstEltTy));
70 return ConstantVector::get(Result);
73 /// This function determines which opcode to use to fold two constant cast
74 /// expressions together. It uses CastInst::isEliminableCastPair to determine
75 /// the opcode. Consequently its just a wrapper around that function.
76 /// @brief Determine if it is valid to fold a cast of a cast
77 static unsigned
78 foldConstantCastPair(
79 unsigned opc, ///< opcode of the second cast constant expression
80 ConstantExpr *Op, ///< the first cast constant expression
81 const Type *DstTy ///< desintation type of the first cast
82 ) {
83 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
84 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
85 assert(CastInst::isCast(opc) && "Invalid cast opcode");
87 // The the types and opcodes for the two Cast constant expressions
88 const Type *SrcTy = Op->getOperand(0)->getType();
89 const Type *MidTy = Op->getType();
90 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
91 Instruction::CastOps secondOp = Instruction::CastOps(opc);
93 // Let CastInst::isEliminableCastPair do the heavy lifting.
94 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
95 Type::getInt64Ty(DstTy->getContext()));
98 static Constant *FoldBitCast(Constant *V, const Type *DestTy) {
99 const Type *SrcTy = V->getType();
100 if (SrcTy == DestTy)
101 return V; // no-op cast
103 // Check to see if we are casting a pointer to an aggregate to a pointer to
104 // the first element. If so, return the appropriate GEP instruction.
105 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
106 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy))
107 if (PTy->getAddressSpace() == DPTy->getAddressSpace()) {
108 SmallVector<Value*, 8> IdxList;
109 Value *Zero =
110 Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
111 IdxList.push_back(Zero);
112 const Type *ElTy = PTy->getElementType();
113 while (ElTy != DPTy->getElementType()) {
114 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
115 if (STy->getNumElements() == 0) break;
116 ElTy = STy->getElementType(0);
117 IdxList.push_back(Zero);
118 } else if (const SequentialType *STy =
119 dyn_cast<SequentialType>(ElTy)) {
120 if (ElTy->isPointerTy()) break; // Can't index into pointers!
121 ElTy = STy->getElementType();
122 IdxList.push_back(Zero);
123 } else {
124 break;
128 if (ElTy == DPTy->getElementType())
129 // This GEP is inbounds because all indices are zero.
130 return ConstantExpr::getInBoundsGetElementPtr(V, &IdxList[0],
131 IdxList.size());
134 // Handle casts from one vector constant to another. We know that the src
135 // and dest type have the same size (otherwise its an illegal cast).
136 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
137 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
138 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
139 "Not cast between same sized vectors!");
140 SrcTy = NULL;
141 // First, check for null. Undef is already handled.
142 if (isa<ConstantAggregateZero>(V))
143 return Constant::getNullValue(DestTy);
145 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
146 return BitCastConstantVector(CV, DestPTy);
149 // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
150 // This allows for other simplifications (although some of them
151 // can only be handled by Analysis/ConstantFolding.cpp).
152 if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
153 return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
156 // Finally, implement bitcast folding now. The code below doesn't handle
157 // bitcast right.
158 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
159 return ConstantPointerNull::get(cast<PointerType>(DestTy));
161 // Handle integral constant input.
162 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
163 if (DestTy->isIntegerTy())
164 // Integral -> Integral. This is a no-op because the bit widths must
165 // be the same. Consequently, we just fold to V.
166 return V;
168 if (DestTy->isFloatingPointTy())
169 return ConstantFP::get(DestTy->getContext(),
170 APFloat(CI->getValue(),
171 !DestTy->isPPC_FP128Ty()));
173 // Otherwise, can't fold this (vector?)
174 return 0;
177 // Handle ConstantFP input: FP -> Integral.
178 if (ConstantFP *FP = dyn_cast<ConstantFP>(V))
179 return ConstantInt::get(FP->getContext(),
180 FP->getValueAPF().bitcastToAPInt());
182 return 0;
186 /// ExtractConstantBytes - V is an integer constant which only has a subset of
187 /// its bytes used. The bytes used are indicated by ByteStart (which is the
188 /// first byte used, counting from the least significant byte) and ByteSize,
189 /// which is the number of bytes used.
191 /// This function analyzes the specified constant to see if the specified byte
192 /// range can be returned as a simplified constant. If so, the constant is
193 /// returned, otherwise null is returned.
194 ///
195 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
196 unsigned ByteSize) {
197 assert(C->getType()->isIntegerTy() &&
198 (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
199 "Non-byte sized integer input");
200 unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
201 assert(ByteSize && "Must be accessing some piece");
202 assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
203 assert(ByteSize != CSize && "Should not extract everything");
205 // Constant Integers are simple.
206 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
207 APInt V = CI->getValue();
208 if (ByteStart)
209 V = V.lshr(ByteStart*8);
210 V = V.trunc(ByteSize*8);
211 return ConstantInt::get(CI->getContext(), V);
214 // In the input is a constant expr, we might be able to recursively simplify.
215 // If not, we definitely can't do anything.
216 ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
217 if (CE == 0) return 0;
219 switch (CE->getOpcode()) {
220 default: return 0;
221 case Instruction::Or: {
222 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
223 if (RHS == 0)
224 return 0;
226 // X | -1 -> -1.
227 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
228 if (RHSC->isAllOnesValue())
229 return RHSC;
231 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
232 if (LHS == 0)
233 return 0;
234 return ConstantExpr::getOr(LHS, RHS);
236 case Instruction::And: {
237 Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
238 if (RHS == 0)
239 return 0;
241 // X & 0 -> 0.
242 if (RHS->isNullValue())
243 return RHS;
245 Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
246 if (LHS == 0)
247 return 0;
248 return ConstantExpr::getAnd(LHS, RHS);
250 case Instruction::LShr: {
251 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
252 if (Amt == 0)
253 return 0;
254 unsigned ShAmt = Amt->getZExtValue();
255 // Cannot analyze non-byte shifts.
256 if ((ShAmt & 7) != 0)
257 return 0;
258 ShAmt >>= 3;
260 // If the extract is known to be all zeros, return zero.
261 if (ByteStart >= CSize-ShAmt)
262 return Constant::getNullValue(IntegerType::get(CE->getContext(),
263 ByteSize*8));
264 // If the extract is known to be fully in the input, extract it.
265 if (ByteStart+ByteSize+ShAmt <= CSize)
266 return ExtractConstantBytes(CE->getOperand(0), ByteStart+ShAmt, ByteSize);
268 // TODO: Handle the 'partially zero' case.
269 return 0;
272 case Instruction::Shl: {
273 ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
274 if (Amt == 0)
275 return 0;
276 unsigned ShAmt = Amt->getZExtValue();
277 // Cannot analyze non-byte shifts.
278 if ((ShAmt & 7) != 0)
279 return 0;
280 ShAmt >>= 3;
282 // If the extract is known to be all zeros, return zero.
283 if (ByteStart+ByteSize <= ShAmt)
284 return Constant::getNullValue(IntegerType::get(CE->getContext(),
285 ByteSize*8));
286 // If the extract is known to be fully in the input, extract it.
287 if (ByteStart >= ShAmt)
288 return ExtractConstantBytes(CE->getOperand(0), ByteStart-ShAmt, ByteSize);
290 // TODO: Handle the 'partially zero' case.
291 return 0;
294 case Instruction::ZExt: {
295 unsigned SrcBitSize =
296 cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
298 // If extracting something that is completely zero, return 0.
299 if (ByteStart*8 >= SrcBitSize)
300 return Constant::getNullValue(IntegerType::get(CE->getContext(),
301 ByteSize*8));
303 // If exactly extracting the input, return it.
304 if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
305 return CE->getOperand(0);
307 // If extracting something completely in the input, if if the input is a
308 // multiple of 8 bits, recurse.
309 if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
310 return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
312 // Otherwise, if extracting a subset of the input, which is not multiple of
313 // 8 bits, do a shift and trunc to get the bits.
314 if ((ByteStart+ByteSize)*8 < SrcBitSize) {
315 assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
316 Constant *Res = CE->getOperand(0);
317 if (ByteStart)
318 Res = ConstantExpr::getLShr(Res,
319 ConstantInt::get(Res->getType(), ByteStart*8));
320 return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
321 ByteSize*8));
324 // TODO: Handle the 'partially zero' case.
325 return 0;
330 /// getFoldedSizeOf - Return a ConstantExpr with type DestTy for sizeof
331 /// on Ty, with any known factors factored out. If Folded is false,
332 /// return null if no factoring was possible, to avoid endlessly
333 /// bouncing an unfoldable expression back into the top-level folder.
335 static Constant *getFoldedSizeOf(const Type *Ty, const Type *DestTy,
336 bool Folded) {
337 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
338 Constant *N = ConstantInt::get(DestTy, ATy->getNumElements());
339 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
340 return ConstantExpr::getNUWMul(E, N);
343 if (const StructType *STy = dyn_cast<StructType>(Ty))
344 if (!STy->isPacked()) {
345 unsigned NumElems = STy->getNumElements();
346 // An empty struct has size zero.
347 if (NumElems == 0)
348 return ConstantExpr::getNullValue(DestTy);
349 // Check for a struct with all members having the same size.
350 Constant *MemberSize =
351 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
352 bool AllSame = true;
353 for (unsigned i = 1; i != NumElems; ++i)
354 if (MemberSize !=
355 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
356 AllSame = false;
357 break;
359 if (AllSame) {
360 Constant *N = ConstantInt::get(DestTy, NumElems);
361 return ConstantExpr::getNUWMul(MemberSize, N);
365 // Pointer size doesn't depend on the pointee type, so canonicalize them
366 // to an arbitrary pointee.
367 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
368 if (!PTy->getElementType()->isIntegerTy(1))
369 return
370 getFoldedSizeOf(PointerType::get(IntegerType::get(PTy->getContext(), 1),
371 PTy->getAddressSpace()),
372 DestTy, true);
374 // If there's no interesting folding happening, bail so that we don't create
375 // a constant that looks like it needs folding but really doesn't.
376 if (!Folded)
377 return 0;
379 // Base case: Get a regular sizeof expression.
380 Constant *C = ConstantExpr::getSizeOf(Ty);
381 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
382 DestTy, false),
383 C, DestTy);
384 return C;
387 /// getFoldedAlignOf - Return a ConstantExpr with type DestTy for alignof
388 /// on Ty, with any known factors factored out. If Folded is false,
389 /// return null if no factoring was possible, to avoid endlessly
390 /// bouncing an unfoldable expression back into the top-level folder.
392 static Constant *getFoldedAlignOf(const Type *Ty, const Type *DestTy,
393 bool Folded) {
394 // The alignment of an array is equal to the alignment of the
395 // array element. Note that this is not always true for vectors.
396 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
397 Constant *C = ConstantExpr::getAlignOf(ATy->getElementType());
398 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
399 DestTy,
400 false),
401 C, DestTy);
402 return C;
405 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
406 // Packed structs always have an alignment of 1.
407 if (STy->isPacked())
408 return ConstantInt::get(DestTy, 1);
410 // Otherwise, struct alignment is the maximum alignment of any member.
411 // Without target data, we can't compare much, but we can check to see
412 // if all the members have the same alignment.
413 unsigned NumElems = STy->getNumElements();
414 // An empty struct has minimal alignment.
415 if (NumElems == 0)
416 return ConstantInt::get(DestTy, 1);
417 // Check for a struct with all members having the same alignment.
418 Constant *MemberAlign =
419 getFoldedAlignOf(STy->getElementType(0), DestTy, true);
420 bool AllSame = true;
421 for (unsigned i = 1; i != NumElems; ++i)
422 if (MemberAlign != getFoldedAlignOf(STy->getElementType(i), DestTy, true)) {
423 AllSame = false;
424 break;
426 if (AllSame)
427 return MemberAlign;
430 // Pointer alignment doesn't depend on the pointee type, so canonicalize them
431 // to an arbitrary pointee.
432 if (const PointerType *PTy = dyn_cast<PointerType>(Ty))
433 if (!PTy->getElementType()->isIntegerTy(1))
434 return
435 getFoldedAlignOf(PointerType::get(IntegerType::get(PTy->getContext(),
437 PTy->getAddressSpace()),
438 DestTy, true);
440 // If there's no interesting folding happening, bail so that we don't create
441 // a constant that looks like it needs folding but really doesn't.
442 if (!Folded)
443 return 0;
445 // Base case: Get a regular alignof expression.
446 Constant *C = ConstantExpr::getAlignOf(Ty);
447 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
448 DestTy, false),
449 C, DestTy);
450 return C;
453 /// getFoldedOffsetOf - Return a ConstantExpr with type DestTy for offsetof
454 /// on Ty and FieldNo, with any known factors factored out. If Folded is false,
455 /// return null if no factoring was possible, to avoid endlessly
456 /// bouncing an unfoldable expression back into the top-level folder.
458 static Constant *getFoldedOffsetOf(const Type *Ty, Constant *FieldNo,
459 const Type *DestTy,
460 bool Folded) {
461 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
462 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo, false,
463 DestTy, false),
464 FieldNo, DestTy);
465 Constant *E = getFoldedSizeOf(ATy->getElementType(), DestTy, true);
466 return ConstantExpr::getNUWMul(E, N);
469 if (const StructType *STy = dyn_cast<StructType>(Ty))
470 if (!STy->isPacked()) {
471 unsigned NumElems = STy->getNumElements();
472 // An empty struct has no members.
473 if (NumElems == 0)
474 return 0;
475 // Check for a struct with all members having the same size.
476 Constant *MemberSize =
477 getFoldedSizeOf(STy->getElementType(0), DestTy, true);
478 bool AllSame = true;
479 for (unsigned i = 1; i != NumElems; ++i)
480 if (MemberSize !=
481 getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
482 AllSame = false;
483 break;
485 if (AllSame) {
486 Constant *N = ConstantExpr::getCast(CastInst::getCastOpcode(FieldNo,
487 false,
488 DestTy,
489 false),
490 FieldNo, DestTy);
491 return ConstantExpr::getNUWMul(MemberSize, N);
495 // If there's no interesting folding happening, bail so that we don't create
496 // a constant that looks like it needs folding but really doesn't.
497 if (!Folded)
498 return 0;
500 // Base case: Get a regular offsetof expression.
501 Constant *C = ConstantExpr::getOffsetOf(Ty, FieldNo);
502 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
503 DestTy, false),
504 C, DestTy);
505 return C;
508 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
509 const Type *DestTy) {
510 if (isa<UndefValue>(V)) {
511 // zext(undef) = 0, because the top bits will be zero.
512 // sext(undef) = 0, because the top bits will all be the same.
513 // [us]itofp(undef) = 0, because the result value is bounded.
514 if (opc == Instruction::ZExt || opc == Instruction::SExt ||
515 opc == Instruction::UIToFP || opc == Instruction::SIToFP)
516 return Constant::getNullValue(DestTy);
517 return UndefValue::get(DestTy);
520 // No compile-time operations on this type yet.
521 if (V->getType()->isPPC_FP128Ty() || DestTy->isPPC_FP128Ty())
522 return 0;
524 if (V->isNullValue() && !DestTy->isX86_MMXTy())
525 return Constant::getNullValue(DestTy);
527 // If the cast operand is a constant expression, there's a few things we can
528 // do to try to simplify it.
529 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
530 if (CE->isCast()) {
531 // Try hard to fold cast of cast because they are often eliminable.
532 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
533 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
534 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
535 // If all of the indexes in the GEP are null values, there is no pointer
536 // adjustment going on. We might as well cast the source pointer.
537 bool isAllNull = true;
538 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
539 if (!CE->getOperand(i)->isNullValue()) {
540 isAllNull = false;
541 break;
543 if (isAllNull)
544 // This is casting one pointer type to another, always BitCast
545 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
549 // If the cast operand is a constant vector, perform the cast by
550 // operating on each element. In the cast of bitcasts, the element
551 // count may be mismatched; don't attempt to handle that here.
552 if (ConstantVector *CV = dyn_cast<ConstantVector>(V))
553 if (DestTy->isVectorTy() &&
554 cast<VectorType>(DestTy)->getNumElements() ==
555 CV->getType()->getNumElements()) {
556 std::vector<Constant*> res;
557 const VectorType *DestVecTy = cast<VectorType>(DestTy);
558 const Type *DstEltTy = DestVecTy->getElementType();
559 for (unsigned i = 0, e = CV->getType()->getNumElements(); i != e; ++i)
560 res.push_back(ConstantExpr::getCast(opc,
561 CV->getOperand(i), DstEltTy));
562 return ConstantVector::get(res);
565 // We actually have to do a cast now. Perform the cast according to the
566 // opcode specified.
567 switch (opc) {
568 default:
569 llvm_unreachable("Failed to cast constant expression");
570 case Instruction::FPTrunc:
571 case Instruction::FPExt:
572 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
573 bool ignored;
574 APFloat Val = FPC->getValueAPF();
575 Val.convert(DestTy->isFloatTy() ? APFloat::IEEEsingle :
576 DestTy->isDoubleTy() ? APFloat::IEEEdouble :
577 DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended :
578 DestTy->isFP128Ty() ? APFloat::IEEEquad :
579 APFloat::Bogus,
580 APFloat::rmNearestTiesToEven, &ignored);
581 return ConstantFP::get(V->getContext(), Val);
583 return 0; // Can't fold.
584 case Instruction::FPToUI:
585 case Instruction::FPToSI:
586 if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
587 const APFloat &V = FPC->getValueAPF();
588 bool ignored;
589 uint64_t x[2];
590 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
591 (void) V.convertToInteger(x, DestBitWidth, opc==Instruction::FPToSI,
592 APFloat::rmTowardZero, &ignored);
593 APInt Val(DestBitWidth, 2, x);
594 return ConstantInt::get(FPC->getContext(), Val);
596 return 0; // Can't fold.
597 case Instruction::IntToPtr: //always treated as unsigned
598 if (V->isNullValue()) // Is it an integral null value?
599 return ConstantPointerNull::get(cast<PointerType>(DestTy));
600 return 0; // Other pointer types cannot be casted
601 case Instruction::PtrToInt: // always treated as unsigned
602 // Is it a null pointer value?
603 if (V->isNullValue())
604 return ConstantInt::get(DestTy, 0);
605 // If this is a sizeof-like expression, pull out multiplications by
606 // known factors to expose them to subsequent folding. If it's an
607 // alignof-like expression, factor out known factors.
608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
609 if (CE->getOpcode() == Instruction::GetElementPtr &&
610 CE->getOperand(0)->isNullValue()) {
611 const Type *Ty =
612 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
613 if (CE->getNumOperands() == 2) {
614 // Handle a sizeof-like expression.
615 Constant *Idx = CE->getOperand(1);
616 bool isOne = isa<ConstantInt>(Idx) && cast<ConstantInt>(Idx)->isOne();
617 if (Constant *C = getFoldedSizeOf(Ty, DestTy, !isOne)) {
618 Idx = ConstantExpr::getCast(CastInst::getCastOpcode(Idx, true,
619 DestTy, false),
620 Idx, DestTy);
621 return ConstantExpr::getMul(C, Idx);
623 } else if (CE->getNumOperands() == 3 &&
624 CE->getOperand(1)->isNullValue()) {
625 // Handle an alignof-like expression.
626 if (const StructType *STy = dyn_cast<StructType>(Ty))
627 if (!STy->isPacked()) {
628 ConstantInt *CI = cast<ConstantInt>(CE->getOperand(2));
629 if (CI->isOne() &&
630 STy->getNumElements() == 2 &&
631 STy->getElementType(0)->isIntegerTy(1)) {
632 return getFoldedAlignOf(STy->getElementType(1), DestTy, false);
635 // Handle an offsetof-like expression.
636 if (Ty->isStructTy() || Ty->isArrayTy()) {
637 if (Constant *C = getFoldedOffsetOf(Ty, CE->getOperand(2),
638 DestTy, false))
639 return C;
643 // Other pointer types cannot be casted
644 return 0;
645 case Instruction::UIToFP:
646 case Instruction::SIToFP:
647 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
648 APInt api = CI->getValue();
649 APFloat apf(APInt::getNullValue(DestTy->getPrimitiveSizeInBits()), true);
650 (void)apf.convertFromAPInt(api,
651 opc==Instruction::SIToFP,
652 APFloat::rmNearestTiesToEven);
653 return ConstantFP::get(V->getContext(), apf);
655 return 0;
656 case Instruction::ZExt:
657 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
658 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
659 return ConstantInt::get(V->getContext(),
660 CI->getValue().zext(BitWidth));
662 return 0;
663 case Instruction::SExt:
664 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
665 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
666 return ConstantInt::get(V->getContext(),
667 CI->getValue().sext(BitWidth));
669 return 0;
670 case Instruction::Trunc: {
671 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
672 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
673 return ConstantInt::get(V->getContext(),
674 CI->getValue().trunc(DestBitWidth));
677 // The input must be a constantexpr. See if we can simplify this based on
678 // the bytes we are demanding. Only do this if the source and dest are an
679 // even multiple of a byte.
680 if ((DestBitWidth & 7) == 0 &&
681 (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
682 if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
683 return Res;
685 return 0;
687 case Instruction::BitCast:
688 return FoldBitCast(V, DestTy);
692 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
693 Constant *V1, Constant *V2) {
694 if (ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
695 return CB->getZExtValue() ? V1 : V2;
697 // Check for zero aggregate and ConstantVector of zeros
698 if (Cond->isNullValue()) return V2;
700 if (ConstantVector* CondV = dyn_cast<ConstantVector>(Cond)) {
702 if (CondV->isAllOnesValue()) return V1;
704 const VectorType *VTy = cast<VectorType>(V1->getType());
705 ConstantVector *CP1 = dyn_cast<ConstantVector>(V1);
706 ConstantVector *CP2 = dyn_cast<ConstantVector>(V2);
708 if ((CP1 || isa<ConstantAggregateZero>(V1)) &&
709 (CP2 || isa<ConstantAggregateZero>(V2))) {
711 // Find the element type of the returned vector
712 const Type *EltTy = VTy->getElementType();
713 unsigned NumElem = VTy->getNumElements();
714 std::vector<Constant*> Res(NumElem);
716 bool Valid = true;
717 for (unsigned i = 0; i < NumElem; ++i) {
718 ConstantInt* c = dyn_cast<ConstantInt>(CondV->getOperand(i));
719 if (!c) {
720 Valid = false;
721 break;
723 Constant *C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
724 Constant *C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
725 Res[i] = c->getZExtValue() ? C1 : C2;
727 // If we were able to build the vector, return it
728 if (Valid) return ConstantVector::get(Res);
733 if (isa<UndefValue>(Cond)) {
734 if (isa<UndefValue>(V1)) return V1;
735 return V2;
737 if (isa<UndefValue>(V1)) return V2;
738 if (isa<UndefValue>(V2)) return V1;
739 if (V1 == V2) return V1;
741 if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
742 if (TrueVal->getOpcode() == Instruction::Select)
743 if (TrueVal->getOperand(0) == Cond)
744 return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
746 if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
747 if (FalseVal->getOpcode() == Instruction::Select)
748 if (FalseVal->getOperand(0) == Cond)
749 return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
752 return 0;
755 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
756 Constant *Idx) {
757 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
758 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
759 if (Val->isNullValue()) // ee(zero, x) -> zero
760 return Constant::getNullValue(
761 cast<VectorType>(Val->getType())->getElementType());
763 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
764 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
765 return CVal->getOperand(CIdx->getZExtValue());
766 } else if (isa<UndefValue>(Idx)) {
767 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
768 return CVal->getOperand(0);
771 return 0;
774 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
775 Constant *Elt,
776 Constant *Idx) {
777 ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
778 if (!CIdx) return 0;
779 APInt idxVal = CIdx->getValue();
780 if (isa<UndefValue>(Val)) {
781 // Insertion of scalar constant into vector undef
782 // Optimize away insertion of undef
783 if (isa<UndefValue>(Elt))
784 return Val;
785 // Otherwise break the aggregate undef into multiple undefs and do
786 // the insertion
787 unsigned numOps =
788 cast<VectorType>(Val->getType())->getNumElements();
789 std::vector<Constant*> Ops;
790 Ops.reserve(numOps);
791 for (unsigned i = 0; i < numOps; ++i) {
792 Constant *Op =
793 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
794 Ops.push_back(Op);
796 return ConstantVector::get(Ops);
798 if (isa<ConstantAggregateZero>(Val)) {
799 // Insertion of scalar constant into vector aggregate zero
800 // Optimize away insertion of zero
801 if (Elt->isNullValue())
802 return Val;
803 // Otherwise break the aggregate zero into multiple zeros and do
804 // the insertion
805 unsigned numOps =
806 cast<VectorType>(Val->getType())->getNumElements();
807 std::vector<Constant*> Ops;
808 Ops.reserve(numOps);
809 for (unsigned i = 0; i < numOps; ++i) {
810 Constant *Op =
811 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
812 Ops.push_back(Op);
814 return ConstantVector::get(Ops);
816 if (ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
817 // Insertion of scalar constant into vector constant
818 std::vector<Constant*> Ops;
819 Ops.reserve(CVal->getNumOperands());
820 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
821 Constant *Op =
822 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
823 Ops.push_back(Op);
825 return ConstantVector::get(Ops);
828 return 0;
831 /// GetVectorElement - If C is a ConstantVector, ConstantAggregateZero or Undef
832 /// return the specified element value. Otherwise return null.
833 static Constant *GetVectorElement(Constant *C, unsigned EltNo) {
834 if (ConstantVector *CV = dyn_cast<ConstantVector>(C))
835 return CV->getOperand(EltNo);
837 const Type *EltTy = cast<VectorType>(C->getType())->getElementType();
838 if (isa<ConstantAggregateZero>(C))
839 return Constant::getNullValue(EltTy);
840 if (isa<UndefValue>(C))
841 return UndefValue::get(EltTy);
842 return 0;
845 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1,
846 Constant *V2,
847 Constant *Mask) {
848 // Undefined shuffle mask -> undefined value.
849 if (isa<UndefValue>(Mask)) return UndefValue::get(V1->getType());
851 unsigned MaskNumElts = cast<VectorType>(Mask->getType())->getNumElements();
852 unsigned SrcNumElts = cast<VectorType>(V1->getType())->getNumElements();
853 const Type *EltTy = cast<VectorType>(V1->getType())->getElementType();
855 // Loop over the shuffle mask, evaluating each element.
856 SmallVector<Constant*, 32> Result;
857 for (unsigned i = 0; i != MaskNumElts; ++i) {
858 Constant *InElt = GetVectorElement(Mask, i);
859 if (InElt == 0) return 0;
861 if (isa<UndefValue>(InElt))
862 InElt = UndefValue::get(EltTy);
863 else if (ConstantInt *CI = dyn_cast<ConstantInt>(InElt)) {
864 unsigned Elt = CI->getZExtValue();
865 if (Elt >= SrcNumElts*2)
866 InElt = UndefValue::get(EltTy);
867 else if (Elt >= SrcNumElts)
868 InElt = GetVectorElement(V2, Elt - SrcNumElts);
869 else
870 InElt = GetVectorElement(V1, Elt);
871 if (InElt == 0) return 0;
872 } else {
873 // Unknown value.
874 return 0;
876 Result.push_back(InElt);
879 return ConstantVector::get(Result);
882 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
883 const unsigned *Idxs,
884 unsigned NumIdx) {
885 // Base case: no indices, so return the entire value.
886 if (NumIdx == 0)
887 return Agg;
889 if (isa<UndefValue>(Agg)) // ev(undef, x) -> undef
890 return UndefValue::get(ExtractValueInst::getIndexedType(Agg->getType(),
891 Idxs,
892 Idxs + NumIdx));
894 if (isa<ConstantAggregateZero>(Agg)) // ev(0, x) -> 0
895 return
896 Constant::getNullValue(ExtractValueInst::getIndexedType(Agg->getType(),
897 Idxs,
898 Idxs + NumIdx));
900 // Otherwise recurse.
901 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg))
902 return ConstantFoldExtractValueInstruction(CS->getOperand(*Idxs),
903 Idxs+1, NumIdx-1);
905 if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg))
906 return ConstantFoldExtractValueInstruction(CA->getOperand(*Idxs),
907 Idxs+1, NumIdx-1);
908 ConstantVector *CV = cast<ConstantVector>(Agg);
909 return ConstantFoldExtractValueInstruction(CV->getOperand(*Idxs),
910 Idxs+1, NumIdx-1);
913 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
914 Constant *Val,
915 const unsigned *Idxs,
916 unsigned NumIdx) {
917 // Base case: no indices, so replace the entire value.
918 if (NumIdx == 0)
919 return Val;
921 if (isa<UndefValue>(Agg)) {
922 // Insertion of constant into aggregate undef
923 // Optimize away insertion of undef.
924 if (isa<UndefValue>(Val))
925 return Agg;
927 // Otherwise break the aggregate undef into multiple undefs and do
928 // the insertion.
929 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
930 unsigned numOps;
931 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
932 numOps = AR->getNumElements();
933 else
934 numOps = cast<StructType>(AggTy)->getNumElements();
936 std::vector<Constant*> Ops(numOps);
937 for (unsigned i = 0; i < numOps; ++i) {
938 const Type *MemberTy = AggTy->getTypeAtIndex(i);
939 Constant *Op =
940 (*Idxs == i) ?
941 ConstantFoldInsertValueInstruction(UndefValue::get(MemberTy),
942 Val, Idxs+1, NumIdx-1) :
943 UndefValue::get(MemberTy);
944 Ops[i] = Op;
947 if (const StructType* ST = dyn_cast<StructType>(AggTy))
948 return ConstantStruct::get(ST, Ops);
949 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
952 if (isa<ConstantAggregateZero>(Agg)) {
953 // Insertion of constant into aggregate zero
954 // Optimize away insertion of zero.
955 if (Val->isNullValue())
956 return Agg;
958 // Otherwise break the aggregate zero into multiple zeros and do
959 // the insertion.
960 const CompositeType *AggTy = cast<CompositeType>(Agg->getType());
961 unsigned numOps;
962 if (const ArrayType *AR = dyn_cast<ArrayType>(AggTy))
963 numOps = AR->getNumElements();
964 else
965 numOps = cast<StructType>(AggTy)->getNumElements();
967 std::vector<Constant*> Ops(numOps);
968 for (unsigned i = 0; i < numOps; ++i) {
969 const Type *MemberTy = AggTy->getTypeAtIndex(i);
970 Constant *Op =
971 (*Idxs == i) ?
972 ConstantFoldInsertValueInstruction(Constant::getNullValue(MemberTy),
973 Val, Idxs+1, NumIdx-1) :
974 Constant::getNullValue(MemberTy);
975 Ops[i] = Op;
978 if (const StructType *ST = dyn_cast<StructType>(AggTy))
979 return ConstantStruct::get(ST, Ops);
980 return ConstantArray::get(cast<ArrayType>(AggTy), Ops);
983 if (isa<ConstantStruct>(Agg) || isa<ConstantArray>(Agg)) {
984 // Insertion of constant into aggregate constant.
985 std::vector<Constant*> Ops(Agg->getNumOperands());
986 for (unsigned i = 0; i < Agg->getNumOperands(); ++i) {
987 Constant *Op = cast<Constant>(Agg->getOperand(i));
988 if (*Idxs == i)
989 Op = ConstantFoldInsertValueInstruction(Op, Val, Idxs+1, NumIdx-1);
990 Ops[i] = Op;
993 if (const StructType* ST = dyn_cast<StructType>(Agg->getType()))
994 return ConstantStruct::get(ST, Ops);
995 return ConstantArray::get(cast<ArrayType>(Agg->getType()), Ops);
998 return 0;
1002 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
1003 Constant *C1, Constant *C2) {
1004 // No compile-time operations on this type yet.
1005 if (C1->getType()->isPPC_FP128Ty())
1006 return 0;
1008 // Handle UndefValue up front.
1009 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1010 switch (Opcode) {
1011 case Instruction::Xor:
1012 if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
1013 // Handle undef ^ undef -> 0 special case. This is a common
1014 // idiom (misuse).
1015 return Constant::getNullValue(C1->getType());
1016 // Fallthrough
1017 case Instruction::Add:
1018 case Instruction::Sub:
1019 return UndefValue::get(C1->getType());
1020 case Instruction::And:
1021 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
1022 return C1;
1023 return Constant::getNullValue(C1->getType()); // undef & X -> 0
1024 case Instruction::Mul: {
1025 ConstantInt *CI;
1026 // X * undef -> undef if X is odd or undef
1027 if (((CI = dyn_cast<ConstantInt>(C1)) && CI->getValue()[0]) ||
1028 ((CI = dyn_cast<ConstantInt>(C2)) && CI->getValue()[0]) ||
1029 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1030 return UndefValue::get(C1->getType());
1032 // X * undef -> 0 otherwise
1033 return Constant::getNullValue(C1->getType());
1035 case Instruction::UDiv:
1036 case Instruction::SDiv:
1037 // undef / 1 -> undef
1038 if (Opcode == Instruction::UDiv || Opcode == Instruction::SDiv)
1039 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2))
1040 if (CI2->isOne())
1041 return C1;
1042 // FALL THROUGH
1043 case Instruction::URem:
1044 case Instruction::SRem:
1045 if (!isa<UndefValue>(C2)) // undef / X -> 0
1046 return Constant::getNullValue(C1->getType());
1047 return C2; // X / undef -> undef
1048 case Instruction::Or: // X | undef -> -1
1049 if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
1050 return C1;
1051 return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
1052 case Instruction::LShr:
1053 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
1054 return C1; // undef lshr undef -> undef
1055 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
1056 // undef lshr X -> 0
1057 case Instruction::AShr:
1058 if (!isa<UndefValue>(C2)) // undef ashr X --> all ones
1059 return Constant::getAllOnesValue(C1->getType());
1060 else if (isa<UndefValue>(C1))
1061 return C1; // undef ashr undef -> undef
1062 else
1063 return C1; // X ashr undef --> X
1064 case Instruction::Shl:
1065 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
1066 return C1; // undef shl undef -> undef
1067 // undef << X -> 0 or X << undef -> 0
1068 return Constant::getNullValue(C1->getType());
1072 // Handle simplifications when the RHS is a constant int.
1073 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1074 switch (Opcode) {
1075 case Instruction::Add:
1076 if (CI2->equalsInt(0)) return C1; // X + 0 == X
1077 break;
1078 case Instruction::Sub:
1079 if (CI2->equalsInt(0)) return C1; // X - 0 == X
1080 break;
1081 case Instruction::Mul:
1082 if (CI2->equalsInt(0)) return C2; // X * 0 == 0
1083 if (CI2->equalsInt(1))
1084 return C1; // X * 1 == X
1085 break;
1086 case Instruction::UDiv:
1087 case Instruction::SDiv:
1088 if (CI2->equalsInt(1))
1089 return C1; // X / 1 == X
1090 if (CI2->equalsInt(0))
1091 return UndefValue::get(CI2->getType()); // X / 0 == undef
1092 break;
1093 case Instruction::URem:
1094 case Instruction::SRem:
1095 if (CI2->equalsInt(1))
1096 return Constant::getNullValue(CI2->getType()); // X % 1 == 0
1097 if (CI2->equalsInt(0))
1098 return UndefValue::get(CI2->getType()); // X % 0 == undef
1099 break;
1100 case Instruction::And:
1101 if (CI2->isZero()) return C2; // X & 0 == 0
1102 if (CI2->isAllOnesValue())
1103 return C1; // X & -1 == X
1105 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1106 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1107 if (CE1->getOpcode() == Instruction::ZExt) {
1108 unsigned DstWidth = CI2->getType()->getBitWidth();
1109 unsigned SrcWidth =
1110 CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1111 APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1112 if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1113 return C1;
1116 // If and'ing the address of a global with a constant, fold it.
1117 if (CE1->getOpcode() == Instruction::PtrToInt &&
1118 isa<GlobalValue>(CE1->getOperand(0))) {
1119 GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1121 // Functions are at least 4-byte aligned.
1122 unsigned GVAlign = GV->getAlignment();
1123 if (isa<Function>(GV))
1124 GVAlign = std::max(GVAlign, 4U);
1126 if (GVAlign > 1) {
1127 unsigned DstWidth = CI2->getType()->getBitWidth();
1128 unsigned SrcWidth = std::min(DstWidth, Log2_32(GVAlign));
1129 APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1131 // If checking bits we know are clear, return zero.
1132 if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1133 return Constant::getNullValue(CI2->getType());
1137 break;
1138 case Instruction::Or:
1139 if (CI2->equalsInt(0)) return C1; // X | 0 == X
1140 if (CI2->isAllOnesValue())
1141 return C2; // X | -1 == -1
1142 break;
1143 case Instruction::Xor:
1144 if (CI2->equalsInt(0)) return C1; // X ^ 0 == X
1146 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1147 switch (CE1->getOpcode()) {
1148 default: break;
1149 case Instruction::ICmp:
1150 case Instruction::FCmp:
1151 // cmp pred ^ true -> cmp !pred
1152 assert(CI2->equalsInt(1));
1153 CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1154 pred = CmpInst::getInversePredicate(pred);
1155 return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1156 CE1->getOperand(1));
1159 break;
1160 case Instruction::AShr:
1161 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1162 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1163 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
1164 return ConstantExpr::getLShr(C1, C2);
1165 break;
1167 } else if (isa<ConstantInt>(C1)) {
1168 // If C1 is a ConstantInt and C2 is not, swap the operands.
1169 if (Instruction::isCommutative(Opcode))
1170 return ConstantExpr::get(Opcode, C2, C1);
1173 // At this point we know neither constant is an UndefValue.
1174 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1175 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1176 using namespace APIntOps;
1177 const APInt &C1V = CI1->getValue();
1178 const APInt &C2V = CI2->getValue();
1179 switch (Opcode) {
1180 default:
1181 break;
1182 case Instruction::Add:
1183 return ConstantInt::get(CI1->getContext(), C1V + C2V);
1184 case Instruction::Sub:
1185 return ConstantInt::get(CI1->getContext(), C1V - C2V);
1186 case Instruction::Mul:
1187 return ConstantInt::get(CI1->getContext(), C1V * C2V);
1188 case Instruction::UDiv:
1189 assert(!CI2->isNullValue() && "Div by zero handled above");
1190 return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1191 case Instruction::SDiv:
1192 assert(!CI2->isNullValue() && "Div by zero handled above");
1193 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1194 return UndefValue::get(CI1->getType()); // MIN_INT / -1 -> undef
1195 return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1196 case Instruction::URem:
1197 assert(!CI2->isNullValue() && "Div by zero handled above");
1198 return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1199 case Instruction::SRem:
1200 assert(!CI2->isNullValue() && "Div by zero handled above");
1201 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
1202 return UndefValue::get(CI1->getType()); // MIN_INT % -1 -> undef
1203 return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1204 case Instruction::And:
1205 return ConstantInt::get(CI1->getContext(), C1V & C2V);
1206 case Instruction::Or:
1207 return ConstantInt::get(CI1->getContext(), C1V | C2V);
1208 case Instruction::Xor:
1209 return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1210 case Instruction::Shl: {
1211 uint32_t shiftAmt = C2V.getZExtValue();
1212 if (shiftAmt < C1V.getBitWidth())
1213 return ConstantInt::get(CI1->getContext(), C1V.shl(shiftAmt));
1214 else
1215 return UndefValue::get(C1->getType()); // too big shift is undef
1217 case Instruction::LShr: {
1218 uint32_t shiftAmt = C2V.getZExtValue();
1219 if (shiftAmt < C1V.getBitWidth())
1220 return ConstantInt::get(CI1->getContext(), C1V.lshr(shiftAmt));
1221 else
1222 return UndefValue::get(C1->getType()); // too big shift is undef
1224 case Instruction::AShr: {
1225 uint32_t shiftAmt = C2V.getZExtValue();
1226 if (shiftAmt < C1V.getBitWidth())
1227 return ConstantInt::get(CI1->getContext(), C1V.ashr(shiftAmt));
1228 else
1229 return UndefValue::get(C1->getType()); // too big shift is undef
1234 switch (Opcode) {
1235 case Instruction::SDiv:
1236 case Instruction::UDiv:
1237 case Instruction::URem:
1238 case Instruction::SRem:
1239 case Instruction::LShr:
1240 case Instruction::AShr:
1241 case Instruction::Shl:
1242 if (CI1->equalsInt(0)) return C1;
1243 break;
1244 default:
1245 break;
1247 } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1248 if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1249 APFloat C1V = CFP1->getValueAPF();
1250 APFloat C2V = CFP2->getValueAPF();
1251 APFloat C3V = C1V; // copy for modification
1252 switch (Opcode) {
1253 default:
1254 break;
1255 case Instruction::FAdd:
1256 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1257 return ConstantFP::get(C1->getContext(), C3V);
1258 case Instruction::FSub:
1259 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1260 return ConstantFP::get(C1->getContext(), C3V);
1261 case Instruction::FMul:
1262 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1263 return ConstantFP::get(C1->getContext(), C3V);
1264 case Instruction::FDiv:
1265 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1266 return ConstantFP::get(C1->getContext(), C3V);
1267 case Instruction::FRem:
1268 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
1269 return ConstantFP::get(C1->getContext(), C3V);
1272 } else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
1273 ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
1274 ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
1275 if ((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
1276 (CP2 != NULL || isa<ConstantAggregateZero>(C2))) {
1277 std::vector<Constant*> Res;
1278 const Type* EltTy = VTy->getElementType();
1279 Constant *C1 = 0;
1280 Constant *C2 = 0;
1281 switch (Opcode) {
1282 default:
1283 break;
1284 case Instruction::Add:
1285 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1286 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1287 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1288 Res.push_back(ConstantExpr::getAdd(C1, C2));
1290 return ConstantVector::get(Res);
1291 case Instruction::FAdd:
1292 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1293 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1294 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1295 Res.push_back(ConstantExpr::getFAdd(C1, C2));
1297 return ConstantVector::get(Res);
1298 case Instruction::Sub:
1299 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1300 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1301 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1302 Res.push_back(ConstantExpr::getSub(C1, C2));
1304 return ConstantVector::get(Res);
1305 case Instruction::FSub:
1306 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1307 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1308 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1309 Res.push_back(ConstantExpr::getFSub(C1, C2));
1311 return ConstantVector::get(Res);
1312 case Instruction::Mul:
1313 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1314 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1315 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1316 Res.push_back(ConstantExpr::getMul(C1, C2));
1318 return ConstantVector::get(Res);
1319 case Instruction::FMul:
1320 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1321 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1322 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1323 Res.push_back(ConstantExpr::getFMul(C1, C2));
1325 return ConstantVector::get(Res);
1326 case Instruction::UDiv:
1327 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1328 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1329 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1330 Res.push_back(ConstantExpr::getUDiv(C1, C2));
1332 return ConstantVector::get(Res);
1333 case Instruction::SDiv:
1334 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1335 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1336 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1337 Res.push_back(ConstantExpr::getSDiv(C1, C2));
1339 return ConstantVector::get(Res);
1340 case Instruction::FDiv:
1341 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1342 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1343 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1344 Res.push_back(ConstantExpr::getFDiv(C1, C2));
1346 return ConstantVector::get(Res);
1347 case Instruction::URem:
1348 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1349 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1350 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1351 Res.push_back(ConstantExpr::getURem(C1, C2));
1353 return ConstantVector::get(Res);
1354 case Instruction::SRem:
1355 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1356 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1357 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1358 Res.push_back(ConstantExpr::getSRem(C1, C2));
1360 return ConstantVector::get(Res);
1361 case Instruction::FRem:
1362 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1363 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1364 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1365 Res.push_back(ConstantExpr::getFRem(C1, C2));
1367 return ConstantVector::get(Res);
1368 case Instruction::And:
1369 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1370 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1371 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1372 Res.push_back(ConstantExpr::getAnd(C1, C2));
1374 return ConstantVector::get(Res);
1375 case Instruction::Or:
1376 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1377 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1378 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1379 Res.push_back(ConstantExpr::getOr(C1, C2));
1381 return ConstantVector::get(Res);
1382 case Instruction::Xor:
1383 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1384 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1385 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1386 Res.push_back(ConstantExpr::getXor(C1, C2));
1388 return ConstantVector::get(Res);
1389 case Instruction::LShr:
1390 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1391 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1392 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1393 Res.push_back(ConstantExpr::getLShr(C1, C2));
1395 return ConstantVector::get(Res);
1396 case Instruction::AShr:
1397 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1398 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1399 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1400 Res.push_back(ConstantExpr::getAShr(C1, C2));
1402 return ConstantVector::get(Res);
1403 case Instruction::Shl:
1404 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
1405 C1 = CP1 ? CP1->getOperand(i) : Constant::getNullValue(EltTy);
1406 C2 = CP2 ? CP2->getOperand(i) : Constant::getNullValue(EltTy);
1407 Res.push_back(ConstantExpr::getShl(C1, C2));
1409 return ConstantVector::get(Res);
1414 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1415 // There are many possible foldings we could do here. We should probably
1416 // at least fold add of a pointer with an integer into the appropriate
1417 // getelementptr. This will improve alias analysis a bit.
1419 // Given ((a + b) + c), if (b + c) folds to something interesting, return
1420 // (a + (b + c)).
1421 if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1422 Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1423 if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1424 return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1426 } else if (isa<ConstantExpr>(C2)) {
1427 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1428 // other way if possible.
1429 if (Instruction::isCommutative(Opcode))
1430 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1433 // i1 can be simplified in many cases.
1434 if (C1->getType()->isIntegerTy(1)) {
1435 switch (Opcode) {
1436 case Instruction::Add:
1437 case Instruction::Sub:
1438 return ConstantExpr::getXor(C1, C2);
1439 case Instruction::Mul:
1440 return ConstantExpr::getAnd(C1, C2);
1441 case Instruction::Shl:
1442 case Instruction::LShr:
1443 case Instruction::AShr:
1444 // We can assume that C2 == 0. If it were one the result would be
1445 // undefined because the shift value is as large as the bitwidth.
1446 return C1;
1447 case Instruction::SDiv:
1448 case Instruction::UDiv:
1449 // We can assume that C2 == 1. If it were zero the result would be
1450 // undefined through division by zero.
1451 return C1;
1452 case Instruction::URem:
1453 case Instruction::SRem:
1454 // We can assume that C2 == 1. If it were zero the result would be
1455 // undefined through division by zero.
1456 return ConstantInt::getFalse(C1->getContext());
1457 default:
1458 break;
1462 // We don't know how to fold this.
1463 return 0;
1466 /// isZeroSizedType - This type is zero sized if its an array or structure of
1467 /// zero sized types. The only leaf zero sized type is an empty structure.
1468 static bool isMaybeZeroSizedType(const Type *Ty) {
1469 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1470 if (STy->isOpaque()) return true; // Can't say.
1472 // If all of elements have zero size, this does too.
1473 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1474 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1475 return true;
1477 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1478 return isMaybeZeroSizedType(ATy->getElementType());
1480 return false;
1483 /// IdxCompare - Compare the two constants as though they were getelementptr
1484 /// indices. This allows coersion of the types to be the same thing.
1486 /// If the two constants are the "same" (after coersion), return 0. If the
1487 /// first is less than the second, return -1, if the second is less than the
1488 /// first, return 1. If the constants are not integral, return -2.
1490 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
1491 if (C1 == C2) return 0;
1493 // Ok, we found a different index. If they are not ConstantInt, we can't do
1494 // anything with them.
1495 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1496 return -2; // don't know!
1498 // Ok, we have two differing integer indices. Sign extend them to be the same
1499 // type. Long is always big enough, so we use it.
1500 if (!C1->getType()->isIntegerTy(64))
1501 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
1503 if (!C2->getType()->isIntegerTy(64))
1504 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
1506 if (C1 == C2) return 0; // They are equal
1508 // If the type being indexed over is really just a zero sized type, there is
1509 // no pointer difference being made here.
1510 if (isMaybeZeroSizedType(ElTy))
1511 return -2; // dunno.
1513 // If they are really different, now that they are the same type, then we
1514 // found a difference!
1515 if (cast<ConstantInt>(C1)->getSExtValue() <
1516 cast<ConstantInt>(C2)->getSExtValue())
1517 return -1;
1518 else
1519 return 1;
1522 /// evaluateFCmpRelation - This function determines if there is anything we can
1523 /// decide about the two constants provided. This doesn't need to handle simple
1524 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
1525 /// If we can determine that the two constants have a particular relation to
1526 /// each other, we should return the corresponding FCmpInst predicate,
1527 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1528 /// ConstantFoldCompareInstruction.
1530 /// To simplify this code we canonicalize the relation so that the first
1531 /// operand is always the most "complex" of the two. We consider ConstantFP
1532 /// to be the simplest, and ConstantExprs to be the most complex.
1533 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1534 assert(V1->getType() == V2->getType() &&
1535 "Cannot compare values of different types!");
1537 // No compile-time operations on this type yet.
1538 if (V1->getType()->isPPC_FP128Ty())
1539 return FCmpInst::BAD_FCMP_PREDICATE;
1541 // Handle degenerate case quickly
1542 if (V1 == V2) return FCmpInst::FCMP_OEQ;
1544 if (!isa<ConstantExpr>(V1)) {
1545 if (!isa<ConstantExpr>(V2)) {
1546 // We distilled thisUse the standard constant folder for a few cases
1547 ConstantInt *R = 0;
1548 R = dyn_cast<ConstantInt>(
1549 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1550 if (R && !R->isZero())
1551 return FCmpInst::FCMP_OEQ;
1552 R = dyn_cast<ConstantInt>(
1553 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1554 if (R && !R->isZero())
1555 return FCmpInst::FCMP_OLT;
1556 R = dyn_cast<ConstantInt>(
1557 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1558 if (R && !R->isZero())
1559 return FCmpInst::FCMP_OGT;
1561 // Nothing more we can do
1562 return FCmpInst::BAD_FCMP_PREDICATE;
1565 // If the first operand is simple and second is ConstantExpr, swap operands.
1566 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1567 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1568 return FCmpInst::getSwappedPredicate(SwappedRelation);
1569 } else {
1570 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1571 // constantexpr or a simple constant.
1572 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1573 switch (CE1->getOpcode()) {
1574 case Instruction::FPTrunc:
1575 case Instruction::FPExt:
1576 case Instruction::UIToFP:
1577 case Instruction::SIToFP:
1578 // We might be able to do something with these but we don't right now.
1579 break;
1580 default:
1581 break;
1584 // There are MANY other foldings that we could perform here. They will
1585 // probably be added on demand, as they seem needed.
1586 return FCmpInst::BAD_FCMP_PREDICATE;
1589 /// evaluateICmpRelation - This function determines if there is anything we can
1590 /// decide about the two constants provided. This doesn't need to handle simple
1591 /// things like integer comparisons, but should instead handle ConstantExprs
1592 /// and GlobalValues. If we can determine that the two constants have a
1593 /// particular relation to each other, we should return the corresponding ICmp
1594 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
1596 /// To simplify this code we canonicalize the relation so that the first
1597 /// operand is always the most "complex" of the two. We consider simple
1598 /// constants (like ConstantInt) to be the simplest, followed by
1599 /// GlobalValues, followed by ConstantExpr's (the most complex).
1601 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1602 bool isSigned) {
1603 assert(V1->getType() == V2->getType() &&
1604 "Cannot compare different types of values!");
1605 if (V1 == V2) return ICmpInst::ICMP_EQ;
1607 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1608 !isa<BlockAddress>(V1)) {
1609 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1610 !isa<BlockAddress>(V2)) {
1611 // We distilled this down to a simple case, use the standard constant
1612 // folder.
1613 ConstantInt *R = 0;
1614 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1615 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1616 if (R && !R->isZero())
1617 return pred;
1618 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1619 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1620 if (R && !R->isZero())
1621 return pred;
1622 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1623 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1624 if (R && !R->isZero())
1625 return pred;
1627 // If we couldn't figure it out, bail.
1628 return ICmpInst::BAD_ICMP_PREDICATE;
1631 // If the first operand is simple, swap operands.
1632 ICmpInst::Predicate SwappedRelation =
1633 evaluateICmpRelation(V2, V1, isSigned);
1634 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1635 return ICmpInst::getSwappedPredicate(SwappedRelation);
1637 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1638 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1639 ICmpInst::Predicate SwappedRelation =
1640 evaluateICmpRelation(V2, V1, isSigned);
1641 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1642 return ICmpInst::getSwappedPredicate(SwappedRelation);
1643 return ICmpInst::BAD_ICMP_PREDICATE;
1646 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1647 // constant (which, since the types must match, means that it's a
1648 // ConstantPointerNull).
1649 if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1650 // Don't try to decide equality of aliases.
1651 if (!isa<GlobalAlias>(GV) && !isa<GlobalAlias>(GV2))
1652 if (!GV->hasExternalWeakLinkage() || !GV2->hasExternalWeakLinkage())
1653 return ICmpInst::ICMP_NE;
1654 } else if (isa<BlockAddress>(V2)) {
1655 return ICmpInst::ICMP_NE; // Globals never equal labels.
1656 } else {
1657 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1658 // GlobalVals can never be null unless they have external weak linkage.
1659 // We don't try to evaluate aliases here.
1660 if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV))
1661 return ICmpInst::ICMP_NE;
1663 } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1664 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
1665 ICmpInst::Predicate SwappedRelation =
1666 evaluateICmpRelation(V2, V1, isSigned);
1667 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1668 return ICmpInst::getSwappedPredicate(SwappedRelation);
1669 return ICmpInst::BAD_ICMP_PREDICATE;
1672 // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1673 // constant (which, since the types must match, means that it is a
1674 // ConstantPointerNull).
1675 if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1676 // Block address in another function can't equal this one, but block
1677 // addresses in the current function might be the same if blocks are
1678 // empty.
1679 if (BA2->getFunction() != BA->getFunction())
1680 return ICmpInst::ICMP_NE;
1681 } else {
1682 // Block addresses aren't null, don't equal the address of globals.
1683 assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1684 "Canonicalization guarantee!");
1685 return ICmpInst::ICMP_NE;
1687 } else {
1688 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
1689 // constantexpr, a global, block address, or a simple constant.
1690 ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1691 Constant *CE1Op0 = CE1->getOperand(0);
1693 switch (CE1->getOpcode()) {
1694 case Instruction::Trunc:
1695 case Instruction::FPTrunc:
1696 case Instruction::FPExt:
1697 case Instruction::FPToUI:
1698 case Instruction::FPToSI:
1699 break; // We can't evaluate floating point casts or truncations.
1701 case Instruction::UIToFP:
1702 case Instruction::SIToFP:
1703 case Instruction::BitCast:
1704 case Instruction::ZExt:
1705 case Instruction::SExt:
1706 // If the cast is not actually changing bits, and the second operand is a
1707 // null pointer, do the comparison with the pre-casted value.
1708 if (V2->isNullValue() &&
1709 (CE1->getType()->isPointerTy() || CE1->getType()->isIntegerTy())) {
1710 if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1711 if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1712 return evaluateICmpRelation(CE1Op0,
1713 Constant::getNullValue(CE1Op0->getType()),
1714 isSigned);
1716 break;
1718 case Instruction::GetElementPtr:
1719 // Ok, since this is a getelementptr, we know that the constant has a
1720 // pointer type. Check the various cases.
1721 if (isa<ConstantPointerNull>(V2)) {
1722 // If we are comparing a GEP to a null pointer, check to see if the base
1723 // of the GEP equals the null pointer.
1724 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1725 if (GV->hasExternalWeakLinkage())
1726 // Weak linkage GVals could be zero or not. We're comparing that
1727 // to null pointer so its greater-or-equal
1728 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1729 else
1730 // If its not weak linkage, the GVal must have a non-zero address
1731 // so the result is greater-than
1732 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1733 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1734 // If we are indexing from a null pointer, check to see if we have any
1735 // non-zero indices.
1736 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1737 if (!CE1->getOperand(i)->isNullValue())
1738 // Offsetting from null, must not be equal.
1739 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1740 // Only zero indexes from null, must still be zero.
1741 return ICmpInst::ICMP_EQ;
1743 // Otherwise, we can't really say if the first operand is null or not.
1744 } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1745 if (isa<ConstantPointerNull>(CE1Op0)) {
1746 if (GV2->hasExternalWeakLinkage())
1747 // Weak linkage GVals could be zero or not. We're comparing it to
1748 // a null pointer, so its less-or-equal
1749 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1750 else
1751 // If its not weak linkage, the GVal must have a non-zero address
1752 // so the result is less-than
1753 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1754 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1755 if (GV == GV2) {
1756 // If this is a getelementptr of the same global, then it must be
1757 // different. Because the types must match, the getelementptr could
1758 // only have at most one index, and because we fold getelementptr's
1759 // with a single zero index, it must be nonzero.
1760 assert(CE1->getNumOperands() == 2 &&
1761 !CE1->getOperand(1)->isNullValue() &&
1762 "Surprising getelementptr!");
1763 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1764 } else {
1765 // If they are different globals, we don't know what the value is,
1766 // but they can't be equal.
1767 return ICmpInst::ICMP_NE;
1770 } else {
1771 ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1772 Constant *CE2Op0 = CE2->getOperand(0);
1774 // There are MANY other foldings that we could perform here. They will
1775 // probably be added on demand, as they seem needed.
1776 switch (CE2->getOpcode()) {
1777 default: break;
1778 case Instruction::GetElementPtr:
1779 // By far the most common case to handle is when the base pointers are
1780 // obviously to the same or different globals.
1781 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1782 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1783 return ICmpInst::ICMP_NE;
1784 // Ok, we know that both getelementptr instructions are based on the
1785 // same global. From this, we can precisely determine the relative
1786 // ordering of the resultant pointers.
1787 unsigned i = 1;
1789 // The logic below assumes that the result of the comparison
1790 // can be determined by finding the first index that differs.
1791 // This doesn't work if there is over-indexing in any
1792 // subsequent indices, so check for that case first.
1793 if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1794 !CE2->isGEPWithNoNotionalOverIndexing())
1795 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1797 // Compare all of the operands the GEP's have in common.
1798 gep_type_iterator GTI = gep_type_begin(CE1);
1799 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1800 ++i, ++GTI)
1801 switch (IdxCompare(CE1->getOperand(i),
1802 CE2->getOperand(i), GTI.getIndexedType())) {
1803 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1804 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1805 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1808 // Ok, we ran out of things they have in common. If any leftovers
1809 // are non-zero then we have a difference, otherwise we are equal.
1810 for (; i < CE1->getNumOperands(); ++i)
1811 if (!CE1->getOperand(i)->isNullValue()) {
1812 if (isa<ConstantInt>(CE1->getOperand(i)))
1813 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1814 else
1815 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1818 for (; i < CE2->getNumOperands(); ++i)
1819 if (!CE2->getOperand(i)->isNullValue()) {
1820 if (isa<ConstantInt>(CE2->getOperand(i)))
1821 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1822 else
1823 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1825 return ICmpInst::ICMP_EQ;
1829 default:
1830 break;
1834 return ICmpInst::BAD_ICMP_PREDICATE;
1837 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1838 Constant *C1, Constant *C2) {
1839 const Type *ResultTy;
1840 if (const VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1841 ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1842 VT->getNumElements());
1843 else
1844 ResultTy = Type::getInt1Ty(C1->getContext());
1846 // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1847 if (pred == FCmpInst::FCMP_FALSE)
1848 return Constant::getNullValue(ResultTy);
1850 if (pred == FCmpInst::FCMP_TRUE)
1851 return Constant::getAllOnesValue(ResultTy);
1853 // Handle some degenerate cases first
1854 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1855 // For EQ and NE, we can always pick a value for the undef to make the
1856 // predicate pass or fail, so we can return undef.
1857 // Also, if both operands are undef, we can return undef.
1858 if (ICmpInst::isEquality(ICmpInst::Predicate(pred)) ||
1859 (isa<UndefValue>(C1) && isa<UndefValue>(C2)))
1860 return UndefValue::get(ResultTy);
1861 // Otherwise, pick the same value as the non-undef operand, and fold
1862 // it to true or false.
1863 return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(pred));
1866 // No compile-time operations on this type yet.
1867 if (C1->getType()->isPPC_FP128Ty())
1868 return 0;
1870 // icmp eq/ne(null,GV) -> false/true
1871 if (C1->isNullValue()) {
1872 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1873 // Don't try to evaluate aliases. External weak GV can be null.
1874 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1875 if (pred == ICmpInst::ICMP_EQ)
1876 return ConstantInt::getFalse(C1->getContext());
1877 else if (pred == ICmpInst::ICMP_NE)
1878 return ConstantInt::getTrue(C1->getContext());
1880 // icmp eq/ne(GV,null) -> false/true
1881 } else if (C2->isNullValue()) {
1882 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1883 // Don't try to evaluate aliases. External weak GV can be null.
1884 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage()) {
1885 if (pred == ICmpInst::ICMP_EQ)
1886 return ConstantInt::getFalse(C1->getContext());
1887 else if (pred == ICmpInst::ICMP_NE)
1888 return ConstantInt::getTrue(C1->getContext());
1892 // If the comparison is a comparison between two i1's, simplify it.
1893 if (C1->getType()->isIntegerTy(1)) {
1894 switch(pred) {
1895 case ICmpInst::ICMP_EQ:
1896 if (isa<ConstantInt>(C2))
1897 return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1898 return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1899 case ICmpInst::ICMP_NE:
1900 return ConstantExpr::getXor(C1, C2);
1901 default:
1902 break;
1906 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1907 APInt V1 = cast<ConstantInt>(C1)->getValue();
1908 APInt V2 = cast<ConstantInt>(C2)->getValue();
1909 switch (pred) {
1910 default: llvm_unreachable("Invalid ICmp Predicate"); return 0;
1911 case ICmpInst::ICMP_EQ: return ConstantInt::get(ResultTy, V1 == V2);
1912 case ICmpInst::ICMP_NE: return ConstantInt::get(ResultTy, V1 != V2);
1913 case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1914 case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1915 case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1916 case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1917 case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1918 case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1919 case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1920 case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1922 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1923 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1924 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1925 APFloat::cmpResult R = C1V.compare(C2V);
1926 switch (pred) {
1927 default: llvm_unreachable("Invalid FCmp Predicate"); return 0;
1928 case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1929 case FCmpInst::FCMP_TRUE: return Constant::getAllOnesValue(ResultTy);
1930 case FCmpInst::FCMP_UNO:
1931 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1932 case FCmpInst::FCMP_ORD:
1933 return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1934 case FCmpInst::FCMP_UEQ:
1935 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1936 R==APFloat::cmpEqual);
1937 case FCmpInst::FCMP_OEQ:
1938 return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1939 case FCmpInst::FCMP_UNE:
1940 return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1941 case FCmpInst::FCMP_ONE:
1942 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1943 R==APFloat::cmpGreaterThan);
1944 case FCmpInst::FCMP_ULT:
1945 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1946 R==APFloat::cmpLessThan);
1947 case FCmpInst::FCMP_OLT:
1948 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1949 case FCmpInst::FCMP_UGT:
1950 return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1951 R==APFloat::cmpGreaterThan);
1952 case FCmpInst::FCMP_OGT:
1953 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1954 case FCmpInst::FCMP_ULE:
1955 return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1956 case FCmpInst::FCMP_OLE:
1957 return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1958 R==APFloat::cmpEqual);
1959 case FCmpInst::FCMP_UGE:
1960 return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1961 case FCmpInst::FCMP_OGE:
1962 return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1963 R==APFloat::cmpEqual);
1965 } else if (C1->getType()->isVectorTy()) {
1966 SmallVector<Constant*, 16> C1Elts, C2Elts;
1967 C1->getVectorElements(C1Elts);
1968 C2->getVectorElements(C2Elts);
1969 if (C1Elts.empty() || C2Elts.empty())
1970 return 0;
1972 // If we can constant fold the comparison of each element, constant fold
1973 // the whole vector comparison.
1974 SmallVector<Constant*, 4> ResElts;
1975 // Compare the elements, producing an i1 result or constant expr.
1976 for (unsigned i = 0, e = C1Elts.size(); i != e; ++i)
1977 ResElts.push_back(ConstantExpr::getCompare(pred, C1Elts[i], C2Elts[i]));
1979 return ConstantVector::get(ResElts);
1982 if (C1->getType()->isFloatingPointTy()) {
1983 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
1984 switch (evaluateFCmpRelation(C1, C2)) {
1985 default: llvm_unreachable("Unknown relation!");
1986 case FCmpInst::FCMP_UNO:
1987 case FCmpInst::FCMP_ORD:
1988 case FCmpInst::FCMP_UEQ:
1989 case FCmpInst::FCMP_UNE:
1990 case FCmpInst::FCMP_ULT:
1991 case FCmpInst::FCMP_UGT:
1992 case FCmpInst::FCMP_ULE:
1993 case FCmpInst::FCMP_UGE:
1994 case FCmpInst::FCMP_TRUE:
1995 case FCmpInst::FCMP_FALSE:
1996 case FCmpInst::BAD_FCMP_PREDICATE:
1997 break; // Couldn't determine anything about these constants.
1998 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1999 Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
2000 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
2001 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2002 break;
2003 case FCmpInst::FCMP_OLT: // We know that C1 < C2
2004 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2005 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
2006 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
2007 break;
2008 case FCmpInst::FCMP_OGT: // We know that C1 > C2
2009 Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
2010 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
2011 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
2012 break;
2013 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
2014 // We can only partially decide this relation.
2015 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2016 Result = 0;
2017 else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2018 Result = 1;
2019 break;
2020 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
2021 // We can only partially decide this relation.
2022 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
2023 Result = 0;
2024 else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
2025 Result = 1;
2026 break;
2027 case FCmpInst::FCMP_ONE: // We know that C1 != C2
2028 // We can only partially decide this relation.
2029 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
2030 Result = 0;
2031 else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
2032 Result = 1;
2033 break;
2036 // If we evaluated the result, return it now.
2037 if (Result != -1)
2038 return ConstantInt::get(ResultTy, Result);
2040 } else {
2041 // Evaluate the relation between the two constants, per the predicate.
2042 int Result = -1; // -1 = unknown, 0 = known false, 1 = known true.
2043 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
2044 default: llvm_unreachable("Unknown relational!");
2045 case ICmpInst::BAD_ICMP_PREDICATE:
2046 break; // Couldn't determine anything about these constants.
2047 case ICmpInst::ICMP_EQ: // We know the constants are equal!
2048 // If we know the constants are equal, we can decide the result of this
2049 // computation precisely.
2050 Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
2051 break;
2052 case ICmpInst::ICMP_ULT:
2053 switch (pred) {
2054 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
2055 Result = 1; break;
2056 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
2057 Result = 0; break;
2059 break;
2060 case ICmpInst::ICMP_SLT:
2061 switch (pred) {
2062 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
2063 Result = 1; break;
2064 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
2065 Result = 0; break;
2067 break;
2068 case ICmpInst::ICMP_UGT:
2069 switch (pred) {
2070 case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
2071 Result = 1; break;
2072 case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
2073 Result = 0; break;
2075 break;
2076 case ICmpInst::ICMP_SGT:
2077 switch (pred) {
2078 case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
2079 Result = 1; break;
2080 case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
2081 Result = 0; break;
2083 break;
2084 case ICmpInst::ICMP_ULE:
2085 if (pred == ICmpInst::ICMP_UGT) Result = 0;
2086 if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
2087 break;
2088 case ICmpInst::ICMP_SLE:
2089 if (pred == ICmpInst::ICMP_SGT) Result = 0;
2090 if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2091 break;
2092 case ICmpInst::ICMP_UGE:
2093 if (pred == ICmpInst::ICMP_ULT) Result = 0;
2094 if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2095 break;
2096 case ICmpInst::ICMP_SGE:
2097 if (pred == ICmpInst::ICMP_SLT) Result = 0;
2098 if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2099 break;
2100 case ICmpInst::ICMP_NE:
2101 if (pred == ICmpInst::ICMP_EQ) Result = 0;
2102 if (pred == ICmpInst::ICMP_NE) Result = 1;
2103 break;
2106 // If we evaluated the result, return it now.
2107 if (Result != -1)
2108 return ConstantInt::get(ResultTy, Result);
2110 // If the right hand side is a bitcast, try using its inverse to simplify
2111 // it by moving it to the left hand side. We can't do this if it would turn
2112 // a vector compare into a scalar compare or visa versa.
2113 if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2114 Constant *CE2Op0 = CE2->getOperand(0);
2115 if (CE2->getOpcode() == Instruction::BitCast &&
2116 CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy()) {
2117 Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2118 return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2122 // If the left hand side is an extension, try eliminating it.
2123 if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2124 if ((CE1->getOpcode() == Instruction::SExt && ICmpInst::isSigned(pred)) ||
2125 (CE1->getOpcode() == Instruction::ZExt && !ICmpInst::isSigned(pred))){
2126 Constant *CE1Op0 = CE1->getOperand(0);
2127 Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2128 if (CE1Inverse == CE1Op0) {
2129 // Check whether we can safely truncate the right hand side.
2130 Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2131 if (ConstantExpr::getZExt(C2Inverse, C2->getType()) == C2) {
2132 return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2138 if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2139 (C1->isNullValue() && !C2->isNullValue())) {
2140 // If C2 is a constant expr and C1 isn't, flip them around and fold the
2141 // other way if possible.
2142 // Also, if C1 is null and C2 isn't, flip them around.
2143 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2144 return ConstantExpr::getICmp(pred, C2, C1);
2147 return 0;
2150 /// isInBoundsIndices - Test whether the given sequence of *normalized* indices
2151 /// is "inbounds".
2152 template<typename IndexTy>
2153 static bool isInBoundsIndices(IndexTy const *Idxs, size_t NumIdx) {
2154 // No indices means nothing that could be out of bounds.
2155 if (NumIdx == 0) return true;
2157 // If the first index is zero, it's in bounds.
2158 if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2160 // If the first index is one and all the rest are zero, it's in bounds,
2161 // by the one-past-the-end rule.
2162 if (!cast<ConstantInt>(Idxs[0])->isOne())
2163 return false;
2164 for (unsigned i = 1, e = NumIdx; i != e; ++i)
2165 if (!cast<Constant>(Idxs[i])->isNullValue())
2166 return false;
2167 return true;
2170 template<typename IndexTy>
2171 static Constant *ConstantFoldGetElementPtrImpl(Constant *C,
2172 bool inBounds,
2173 IndexTy const *Idxs,
2174 unsigned NumIdx) {
2175 Constant *Idx0 = cast<Constant>(Idxs[0]);
2176 if (NumIdx == 0 ||
2177 (NumIdx == 1 && Idx0->isNullValue()))
2178 return C;
2180 if (isa<UndefValue>(C)) {
2181 const PointerType *Ptr = cast<PointerType>(C->getType());
2182 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs, Idxs+NumIdx);
2183 assert(Ty != 0 && "Invalid indices for GEP!");
2184 return UndefValue::get(PointerType::get(Ty, Ptr->getAddressSpace()));
2187 if (C->isNullValue()) {
2188 bool isNull = true;
2189 for (unsigned i = 0, e = NumIdx; i != e; ++i)
2190 if (!cast<Constant>(Idxs[i])->isNullValue()) {
2191 isNull = false;
2192 break;
2194 if (isNull) {
2195 const PointerType *Ptr = cast<PointerType>(C->getType());
2196 const Type *Ty = GetElementPtrInst::getIndexedType(Ptr, Idxs,
2197 Idxs+NumIdx);
2198 assert(Ty != 0 && "Invalid indices for GEP!");
2199 return ConstantPointerNull::get(PointerType::get(Ty,
2200 Ptr->getAddressSpace()));
2204 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2205 // Combine Indices - If the source pointer to this getelementptr instruction
2206 // is a getelementptr instruction, combine the indices of the two
2207 // getelementptr instructions into a single instruction.
2209 if (CE->getOpcode() == Instruction::GetElementPtr) {
2210 const Type *LastTy = 0;
2211 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
2212 I != E; ++I)
2213 LastTy = *I;
2215 if ((LastTy && LastTy->isArrayTy()) || Idx0->isNullValue()) {
2216 SmallVector<Value*, 16> NewIndices;
2217 NewIndices.reserve(NumIdx + CE->getNumOperands());
2218 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
2219 NewIndices.push_back(CE->getOperand(i));
2221 // Add the last index of the source with the first index of the new GEP.
2222 // Make sure to handle the case when they are actually different types.
2223 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
2224 // Otherwise it must be an array.
2225 if (!Idx0->isNullValue()) {
2226 const Type *IdxTy = Combined->getType();
2227 if (IdxTy != Idx0->getType()) {
2228 const Type *Int64Ty = Type::getInt64Ty(IdxTy->getContext());
2229 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Int64Ty);
2230 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined, Int64Ty);
2231 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
2232 } else {
2233 Combined =
2234 ConstantExpr::get(Instruction::Add, Idx0, Combined);
2238 NewIndices.push_back(Combined);
2239 NewIndices.append(Idxs+1, Idxs+NumIdx);
2240 return (inBounds && cast<GEPOperator>(CE)->isInBounds()) ?
2241 ConstantExpr::getInBoundsGetElementPtr(CE->getOperand(0),
2242 &NewIndices[0],
2243 NewIndices.size()) :
2244 ConstantExpr::getGetElementPtr(CE->getOperand(0),
2245 &NewIndices[0],
2246 NewIndices.size());
2250 // Implement folding of:
2251 // i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2252 // i64 0, i64 0)
2253 // To: i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2255 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
2256 if (const PointerType *SPT =
2257 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
2258 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
2259 if (const ArrayType *CAT =
2260 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
2261 if (CAT->getElementType() == SAT->getElementType())
2262 return inBounds ?
2263 ConstantExpr::getInBoundsGetElementPtr(
2264 (Constant*)CE->getOperand(0), Idxs, NumIdx) :
2265 ConstantExpr::getGetElementPtr(
2266 (Constant*)CE->getOperand(0), Idxs, NumIdx);
2270 // Check to see if any array indices are not within the corresponding
2271 // notional array bounds. If so, try to determine if they can be factored
2272 // out into preceding dimensions.
2273 bool Unknown = false;
2274 SmallVector<Constant *, 8> NewIdxs;
2275 const Type *Ty = C->getType();
2276 const Type *Prev = 0;
2277 for (unsigned i = 0; i != NumIdx;
2278 Prev = Ty, Ty = cast<CompositeType>(Ty)->getTypeAtIndex(Idxs[i]), ++i) {
2279 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2280 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
2281 if (ATy->getNumElements() <= INT64_MAX &&
2282 ATy->getNumElements() != 0 &&
2283 CI->getSExtValue() >= (int64_t)ATy->getNumElements()) {
2284 if (isa<SequentialType>(Prev)) {
2285 // It's out of range, but we can factor it into the prior
2286 // dimension.
2287 NewIdxs.resize(NumIdx);
2288 ConstantInt *Factor = ConstantInt::get(CI->getType(),
2289 ATy->getNumElements());
2290 NewIdxs[i] = ConstantExpr::getSRem(CI, Factor);
2292 Constant *PrevIdx = cast<Constant>(Idxs[i-1]);
2293 Constant *Div = ConstantExpr::getSDiv(CI, Factor);
2295 // Before adding, extend both operands to i64 to avoid
2296 // overflow trouble.
2297 if (!PrevIdx->getType()->isIntegerTy(64))
2298 PrevIdx = ConstantExpr::getSExt(PrevIdx,
2299 Type::getInt64Ty(Div->getContext()));
2300 if (!Div->getType()->isIntegerTy(64))
2301 Div = ConstantExpr::getSExt(Div,
2302 Type::getInt64Ty(Div->getContext()));
2304 NewIdxs[i-1] = ConstantExpr::getAdd(PrevIdx, Div);
2305 } else {
2306 // It's out of range, but the prior dimension is a struct
2307 // so we can't do anything about it.
2308 Unknown = true;
2311 } else {
2312 // We don't know if it's in range or not.
2313 Unknown = true;
2317 // If we did any factoring, start over with the adjusted indices.
2318 if (!NewIdxs.empty()) {
2319 for (unsigned i = 0; i != NumIdx; ++i)
2320 if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2321 return inBounds ?
2322 ConstantExpr::getInBoundsGetElementPtr(C, NewIdxs.data(),
2323 NewIdxs.size()) :
2324 ConstantExpr::getGetElementPtr(C, NewIdxs.data(), NewIdxs.size());
2327 // If all indices are known integers and normalized, we can do a simple
2328 // check for the "inbounds" property.
2329 if (!Unknown && !inBounds &&
2330 isa<GlobalVariable>(C) && isInBoundsIndices(Idxs, NumIdx))
2331 return ConstantExpr::getInBoundsGetElementPtr(C, Idxs, NumIdx);
2333 return 0;
2336 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2337 bool inBounds,
2338 Constant* const *Idxs,
2339 unsigned NumIdx) {
2340 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);
2343 Constant *llvm::ConstantFoldGetElementPtr(Constant *C,
2344 bool inBounds,
2345 Value* const *Idxs,
2346 unsigned NumIdx) {
2347 return ConstantFoldGetElementPtrImpl(C, inBounds, Idxs, NumIdx);