Correct names.
[llvm-complete.git] / lib / VMCore / ConstantFold.cpp
blob917098d35137ce93926f5b51182ca762e7702c4e
1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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 // template-based folder for simple primitive constants like ConstantInt, and
16 // the special case hackery that we use to symbolically evaluate expressions
17 // that use ConstantExprs.
19 //===----------------------------------------------------------------------===//
21 #include "ConstantFold.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Function.h"
26 #include "llvm/GlobalAlias.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/ManagedStatic.h"
31 #include "llvm/Support/MathExtras.h"
32 #include <limits>
33 using namespace llvm;
35 //===----------------------------------------------------------------------===//
36 // ConstantFold*Instruction Implementations
37 //===----------------------------------------------------------------------===//
39 /// CastConstantVector - Convert the specified ConstantVector node to the
40 /// specified vector type. At this point, we know that the elements of the
41 /// input vector constant are all simple integer or FP values.
42 static Constant *CastConstantVector(ConstantVector *CV,
43 const VectorType *DstTy) {
44 unsigned SrcNumElts = CV->getType()->getNumElements();
45 unsigned DstNumElts = DstTy->getNumElements();
46 const Type *SrcEltTy = CV->getType()->getElementType();
47 const Type *DstEltTy = DstTy->getElementType();
49 // If both vectors have the same number of elements (thus, the elements
50 // are the same size), perform the conversion now.
51 if (SrcNumElts == DstNumElts) {
52 std::vector<Constant*> Result;
54 // If the src and dest elements are both integers, or both floats, we can
55 // just BitCast each element because the elements are the same size.
56 if ((SrcEltTy->isInteger() && DstEltTy->isInteger()) ||
57 (SrcEltTy->isFloatingPoint() && DstEltTy->isFloatingPoint())) {
58 for (unsigned i = 0; i != SrcNumElts; ++i)
59 Result.push_back(
60 ConstantExpr::getBitCast(CV->getOperand(i), DstEltTy));
61 return ConstantVector::get(Result);
64 // If this is an int-to-fp cast ..
65 if (SrcEltTy->isInteger()) {
66 // Ensure that it is int-to-fp cast
67 assert(DstEltTy->isFloatingPoint());
68 if (DstEltTy->getTypeID() == Type::DoubleTyID) {
69 for (unsigned i = 0; i != SrcNumElts; ++i) {
70 ConstantInt *CI = cast<ConstantInt>(CV->getOperand(i));
71 double V = CI->getValue().bitsToDouble();
72 Result.push_back(ConstantFP::get(Type::DoubleTy, APFloat(V)));
74 return ConstantVector::get(Result);
76 assert(DstEltTy == Type::FloatTy && "Unknown fp type!");
77 for (unsigned i = 0; i != SrcNumElts; ++i) {
78 ConstantInt *CI = cast<ConstantInt>(CV->getOperand(i));
79 float V = CI->getValue().bitsToFloat();
80 Result.push_back(ConstantFP::get(Type::FloatTy, APFloat(V)));
82 return ConstantVector::get(Result);
85 // Otherwise, this is an fp-to-int cast.
86 assert(SrcEltTy->isFloatingPoint() && DstEltTy->isInteger());
88 if (SrcEltTy->getTypeID() == Type::DoubleTyID) {
89 for (unsigned i = 0; i != SrcNumElts; ++i) {
90 uint64_t V = cast<ConstantFP>(CV->getOperand(i))->
91 getValueAPF().convertToAPInt().getZExtValue();
92 Constant *C = ConstantInt::get(Type::Int64Ty, V);
93 Result.push_back(ConstantExpr::getBitCast(C, DstEltTy ));
95 return ConstantVector::get(Result);
98 assert(SrcEltTy->getTypeID() == Type::FloatTyID);
99 for (unsigned i = 0; i != SrcNumElts; ++i) {
100 uint32_t V = (uint32_t)cast<ConstantFP>(CV->getOperand(i))->
101 getValueAPF().convertToAPInt().getZExtValue();
102 Constant *C = ConstantInt::get(Type::Int32Ty, V);
103 Result.push_back(ConstantExpr::getBitCast(C, DstEltTy));
105 return ConstantVector::get(Result);
108 // Otherwise, this is a cast that changes element count and size. Handle
109 // casts which shrink the elements here.
111 // FIXME: We need to know endianness to do this!
113 return 0;
116 /// This function determines which opcode to use to fold two constant cast
117 /// expressions together. It uses CastInst::isEliminableCastPair to determine
118 /// the opcode. Consequently its just a wrapper around that function.
119 /// @brief Determine if it is valid to fold a cast of a cast
120 static unsigned
121 foldConstantCastPair(
122 unsigned opc, ///< opcode of the second cast constant expression
123 const ConstantExpr*Op, ///< the first cast constant expression
124 const Type *DstTy ///< desintation type of the first cast
126 assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
127 assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
128 assert(CastInst::isCast(opc) && "Invalid cast opcode");
130 // The the types and opcodes for the two Cast constant expressions
131 const Type *SrcTy = Op->getOperand(0)->getType();
132 const Type *MidTy = Op->getType();
133 Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
134 Instruction::CastOps secondOp = Instruction::CastOps(opc);
136 // Let CastInst::isEliminableCastPair do the heavy lifting.
137 return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
138 Type::Int64Ty);
141 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, const Constant *V,
142 const Type *DestTy) {
143 const Type *SrcTy = V->getType();
145 if (isa<UndefValue>(V)) {
146 // zext(undef) = 0, because the top bits will be zero.
147 // sext(undef) = 0, because the top bits will all be the same.
148 if (opc == Instruction::ZExt || opc == Instruction::SExt)
149 return Constant::getNullValue(DestTy);
150 return UndefValue::get(DestTy);
153 // If the cast operand is a constant expression, there's a few things we can
154 // do to try to simplify it.
155 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
156 if (CE->isCast()) {
157 // Try hard to fold cast of cast because they are often eliminable.
158 if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
159 return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
160 } else if (CE->getOpcode() == Instruction::GetElementPtr) {
161 // If all of the indexes in the GEP are null values, there is no pointer
162 // adjustment going on. We might as well cast the source pointer.
163 bool isAllNull = true;
164 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
165 if (!CE->getOperand(i)->isNullValue()) {
166 isAllNull = false;
167 break;
169 if (isAllNull)
170 // This is casting one pointer type to another, always BitCast
171 return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
175 // We actually have to do a cast now. Perform the cast according to the
176 // opcode specified.
177 switch (opc) {
178 case Instruction::FPTrunc:
179 case Instruction::FPExt:
180 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
181 APFloat Val = FPC->getValueAPF();
182 Val.convert(DestTy == Type::FloatTy ? APFloat::IEEEsingle :
183 DestTy == Type::DoubleTy ? APFloat::IEEEdouble :
184 DestTy == Type::X86_FP80Ty ? APFloat::x87DoubleExtended :
185 DestTy == Type::FP128Ty ? APFloat::IEEEquad :
186 APFloat::Bogus,
187 APFloat::rmNearestTiesToEven);
188 return ConstantFP::get(DestTy, Val);
190 return 0; // Can't fold.
191 case Instruction::FPToUI:
192 case Instruction::FPToSI:
193 if (const ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
194 APFloat V = FPC->getValueAPF();
195 uint64_t x[2];
196 uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
197 APFloat::opStatus status = V.convertToInteger(x, DestBitWidth,
198 opc==Instruction::FPToSI,
199 APFloat::rmNearestTiesToEven);
200 if (status!=APFloat::opOK && status!=APFloat::opInexact)
201 return 0; // give up
202 APInt Val(DestBitWidth, 2, x);
203 return ConstantInt::get(Val);
205 return 0; // Can't fold.
206 case Instruction::IntToPtr: //always treated as unsigned
207 if (V->isNullValue()) // Is it an integral null value?
208 return ConstantPointerNull::get(cast<PointerType>(DestTy));
209 return 0; // Other pointer types cannot be casted
210 case Instruction::PtrToInt: // always treated as unsigned
211 if (V->isNullValue()) // is it a null pointer value?
212 return ConstantInt::get(DestTy, 0);
213 return 0; // Other pointer types cannot be casted
214 case Instruction::UIToFP:
215 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
216 double d = CI->getValue().roundToDouble();
217 if (DestTy==Type::FloatTy)
218 return ConstantFP::get(DestTy, APFloat((float)d));
219 else if (DestTy==Type::DoubleTy)
220 return ConstantFP::get(DestTy, APFloat(d));
221 else
222 return 0; // FIXME do this for long double
224 return 0;
225 case Instruction::SIToFP:
226 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
227 double d = CI->getValue().signedRoundToDouble();
228 if (DestTy==Type::FloatTy)
229 return ConstantFP::get(DestTy, APFloat((float)d));
230 else if (DestTy==Type::DoubleTy)
231 return ConstantFP::get(DestTy, APFloat(d));
232 else
233 return 0; // FIXME do this for long double
235 return 0;
236 case Instruction::ZExt:
237 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
238 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
239 APInt Result(CI->getValue());
240 Result.zext(BitWidth);
241 return ConstantInt::get(Result);
243 return 0;
244 case Instruction::SExt:
245 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
246 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
247 APInt Result(CI->getValue());
248 Result.sext(BitWidth);
249 return ConstantInt::get(Result);
251 return 0;
252 case Instruction::Trunc:
253 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
254 uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
255 APInt Result(CI->getValue());
256 Result.trunc(BitWidth);
257 return ConstantInt::get(Result);
259 return 0;
260 case Instruction::BitCast:
261 if (SrcTy == DestTy)
262 return (Constant*)V; // no-op cast
264 // Check to see if we are casting a pointer to an aggregate to a pointer to
265 // the first element. If so, return the appropriate GEP instruction.
266 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
267 if (const PointerType *DPTy = dyn_cast<PointerType>(DestTy)) {
268 SmallVector<Value*, 8> IdxList;
269 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
270 const Type *ElTy = PTy->getElementType();
271 while (ElTy != DPTy->getElementType()) {
272 if (const StructType *STy = dyn_cast<StructType>(ElTy)) {
273 if (STy->getNumElements() == 0) break;
274 ElTy = STy->getElementType(0);
275 IdxList.push_back(Constant::getNullValue(Type::Int32Ty));
276 } else if (const SequentialType *STy =
277 dyn_cast<SequentialType>(ElTy)) {
278 if (isa<PointerType>(ElTy)) break; // Can't index into pointers!
279 ElTy = STy->getElementType();
280 IdxList.push_back(IdxList[0]);
281 } else {
282 break;
286 if (ElTy == DPTy->getElementType())
287 return ConstantExpr::getGetElementPtr(
288 const_cast<Constant*>(V), &IdxList[0], IdxList.size());
291 // Handle casts from one vector constant to another. We know that the src
292 // and dest type have the same size (otherwise its an illegal cast).
293 if (const VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
294 if (const VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
295 assert(DestPTy->getBitWidth() == SrcTy->getBitWidth() &&
296 "Not cast between same sized vectors!");
297 // First, check for null and undef
298 if (isa<ConstantAggregateZero>(V))
299 return Constant::getNullValue(DestTy);
300 if (isa<UndefValue>(V))
301 return UndefValue::get(DestTy);
303 if (const ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
304 // This is a cast from a ConstantVector of one type to a
305 // ConstantVector of another type. Check to see if all elements of
306 // the input are simple.
307 bool AllSimpleConstants = true;
308 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
309 if (!isa<ConstantInt>(CV->getOperand(i)) &&
310 !isa<ConstantFP>(CV->getOperand(i))) {
311 AllSimpleConstants = false;
312 break;
316 // If all of the elements are simple constants, we can fold this.
317 if (AllSimpleConstants)
318 return CastConstantVector(const_cast<ConstantVector*>(CV), DestPTy);
323 // Finally, implement bitcast folding now. The code below doesn't handle
324 // bitcast right.
325 if (isa<ConstantPointerNull>(V)) // ptr->ptr cast.
326 return ConstantPointerNull::get(cast<PointerType>(DestTy));
328 // Handle integral constant input.
329 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
330 if (DestTy->isInteger())
331 // Integral -> Integral. This is a no-op because the bit widths must
332 // be the same. Consequently, we just fold to V.
333 return const_cast<Constant*>(V);
335 if (DestTy->isFloatingPoint()) {
336 assert((DestTy == Type::DoubleTy || DestTy == Type::FloatTy) &&
337 "Unknown FP type!");
338 return ConstantFP::get(DestTy, APFloat(CI->getValue()));
340 // Otherwise, can't fold this (vector?)
341 return 0;
344 // Handle ConstantFP input.
345 if (const ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
346 // FP -> Integral.
347 if (DestTy == Type::Int32Ty) {
348 return ConstantInt::get(FP->getValueAPF().convertToAPInt());
349 } else {
350 assert(DestTy == Type::Int64Ty && "only support f32/f64 for now!");
351 return ConstantInt::get(FP->getValueAPF().convertToAPInt());
354 return 0;
355 default:
356 assert(!"Invalid CE CastInst opcode");
357 break;
360 assert(0 && "Failed to cast constant expression");
361 return 0;
364 Constant *llvm::ConstantFoldSelectInstruction(const Constant *Cond,
365 const Constant *V1,
366 const Constant *V2) {
367 if (const ConstantInt *CB = dyn_cast<ConstantInt>(Cond))
368 return const_cast<Constant*>(CB->getZExtValue() ? V1 : V2);
370 if (isa<UndefValue>(V1)) return const_cast<Constant*>(V2);
371 if (isa<UndefValue>(V2)) return const_cast<Constant*>(V1);
372 if (isa<UndefValue>(Cond)) return const_cast<Constant*>(V1);
373 if (V1 == V2) return const_cast<Constant*>(V1);
374 return 0;
377 Constant *llvm::ConstantFoldExtractElementInstruction(const Constant *Val,
378 const Constant *Idx) {
379 if (isa<UndefValue>(Val)) // ee(undef, x) -> undef
380 return UndefValue::get(cast<VectorType>(Val->getType())->getElementType());
381 if (Val->isNullValue()) // ee(zero, x) -> zero
382 return Constant::getNullValue(
383 cast<VectorType>(Val->getType())->getElementType());
385 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
386 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx)) {
387 return const_cast<Constant*>(CVal->getOperand(CIdx->getZExtValue()));
388 } else if (isa<UndefValue>(Idx)) {
389 // ee({w,x,y,z}, undef) -> w (an arbitrary value).
390 return const_cast<Constant*>(CVal->getOperand(0));
393 return 0;
396 Constant *llvm::ConstantFoldInsertElementInstruction(const Constant *Val,
397 const Constant *Elt,
398 const Constant *Idx) {
399 const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
400 if (!CIdx) return 0;
401 APInt idxVal = CIdx->getValue();
402 if (isa<UndefValue>(Val)) {
403 // Insertion of scalar constant into vector undef
404 // Optimize away insertion of undef
405 if (isa<UndefValue>(Elt))
406 return const_cast<Constant*>(Val);
407 // Otherwise break the aggregate undef into multiple undefs and do
408 // the insertion
409 unsigned numOps =
410 cast<VectorType>(Val->getType())->getNumElements();
411 std::vector<Constant*> Ops;
412 Ops.reserve(numOps);
413 for (unsigned i = 0; i < numOps; ++i) {
414 const Constant *Op =
415 (idxVal == i) ? Elt : UndefValue::get(Elt->getType());
416 Ops.push_back(const_cast<Constant*>(Op));
418 return ConstantVector::get(Ops);
420 if (isa<ConstantAggregateZero>(Val)) {
421 // Insertion of scalar constant into vector aggregate zero
422 // Optimize away insertion of zero
423 if (Elt->isNullValue())
424 return const_cast<Constant*>(Val);
425 // Otherwise break the aggregate zero into multiple zeros and do
426 // the insertion
427 unsigned numOps =
428 cast<VectorType>(Val->getType())->getNumElements();
429 std::vector<Constant*> Ops;
430 Ops.reserve(numOps);
431 for (unsigned i = 0; i < numOps; ++i) {
432 const Constant *Op =
433 (idxVal == i) ? Elt : Constant::getNullValue(Elt->getType());
434 Ops.push_back(const_cast<Constant*>(Op));
436 return ConstantVector::get(Ops);
438 if (const ConstantVector *CVal = dyn_cast<ConstantVector>(Val)) {
439 // Insertion of scalar constant into vector constant
440 std::vector<Constant*> Ops;
441 Ops.reserve(CVal->getNumOperands());
442 for (unsigned i = 0; i < CVal->getNumOperands(); ++i) {
443 const Constant *Op =
444 (idxVal == i) ? Elt : cast<Constant>(CVal->getOperand(i));
445 Ops.push_back(const_cast<Constant*>(Op));
447 return ConstantVector::get(Ops);
449 return 0;
452 Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1,
453 const Constant *V2,
454 const Constant *Mask) {
455 // TODO:
456 return 0;
459 /// EvalVectorOp - Given two vector constants and a function pointer, apply the
460 /// function pointer to each element pair, producing a new ConstantVector
461 /// constant.
462 static Constant *EvalVectorOp(const ConstantVector *V1,
463 const ConstantVector *V2,
464 Constant *(*FP)(Constant*, Constant*)) {
465 std::vector<Constant*> Res;
466 for (unsigned i = 0, e = V1->getNumOperands(); i != e; ++i)
467 Res.push_back(FP(const_cast<Constant*>(V1->getOperand(i)),
468 const_cast<Constant*>(V2->getOperand(i))));
469 return ConstantVector::get(Res);
472 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
473 const Constant *C1,
474 const Constant *C2) {
475 // Handle UndefValue up front
476 if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
477 switch (Opcode) {
478 case Instruction::Add:
479 case Instruction::Sub:
480 case Instruction::Xor:
481 return UndefValue::get(C1->getType());
482 case Instruction::Mul:
483 case Instruction::And:
484 return Constant::getNullValue(C1->getType());
485 case Instruction::UDiv:
486 case Instruction::SDiv:
487 case Instruction::FDiv:
488 case Instruction::URem:
489 case Instruction::SRem:
490 case Instruction::FRem:
491 if (!isa<UndefValue>(C2)) // undef / X -> 0
492 return Constant::getNullValue(C1->getType());
493 return const_cast<Constant*>(C2); // X / undef -> undef
494 case Instruction::Or: // X | undef -> -1
495 if (const VectorType *PTy = dyn_cast<VectorType>(C1->getType()))
496 return ConstantVector::getAllOnesValue(PTy);
497 return ConstantInt::getAllOnesValue(C1->getType());
498 case Instruction::LShr:
499 if (isa<UndefValue>(C2) && isa<UndefValue>(C1))
500 return const_cast<Constant*>(C1); // undef lshr undef -> undef
501 return Constant::getNullValue(C1->getType()); // X lshr undef -> 0
502 // undef lshr X -> 0
503 case Instruction::AShr:
504 if (!isa<UndefValue>(C2))
505 return const_cast<Constant*>(C1); // undef ashr X --> undef
506 else if (isa<UndefValue>(C1))
507 return const_cast<Constant*>(C1); // undef ashr undef -> undef
508 else
509 return const_cast<Constant*>(C1); // X ashr undef --> X
510 case Instruction::Shl:
511 // undef << X -> 0 or X << undef -> 0
512 return Constant::getNullValue(C1->getType());
516 if (const ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
517 if (isa<ConstantExpr>(C2)) {
518 // There are many possible foldings we could do here. We should probably
519 // at least fold add of a pointer with an integer into the appropriate
520 // getelementptr. This will improve alias analysis a bit.
521 } else {
522 // Just implement a couple of simple identities.
523 switch (Opcode) {
524 case Instruction::Add:
525 if (C2->isNullValue()) return const_cast<Constant*>(C1); // X + 0 == X
526 break;
527 case Instruction::Sub:
528 if (C2->isNullValue()) return const_cast<Constant*>(C1); // X - 0 == X
529 break;
530 case Instruction::Mul:
531 if (C2->isNullValue()) return const_cast<Constant*>(C2); // X * 0 == 0
532 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
533 if (CI->equalsInt(1))
534 return const_cast<Constant*>(C1); // X * 1 == X
535 break;
536 case Instruction::UDiv:
537 case Instruction::SDiv:
538 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
539 if (CI->equalsInt(1))
540 return const_cast<Constant*>(C1); // X / 1 == X
541 break;
542 case Instruction::URem:
543 case Instruction::SRem:
544 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
545 if (CI->equalsInt(1))
546 return Constant::getNullValue(CI->getType()); // X % 1 == 0
547 break;
548 case Instruction::And:
549 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2)) {
550 if (CI->isZero()) return const_cast<Constant*>(C2); // X & 0 == 0
551 if (CI->isAllOnesValue())
552 return const_cast<Constant*>(C1); // X & -1 == X
554 // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
555 if (CE1->getOpcode() == Instruction::ZExt) {
556 APInt PossiblySetBits
557 = cast<IntegerType>(CE1->getOperand(0)->getType())->getMask();
558 PossiblySetBits.zext(C1->getType()->getPrimitiveSizeInBits());
559 if ((PossiblySetBits & CI->getValue()) == PossiblySetBits)
560 return const_cast<Constant*>(C1);
563 if (CE1->isCast() && isa<GlobalValue>(CE1->getOperand(0))) {
564 GlobalValue *CPR = cast<GlobalValue>(CE1->getOperand(0));
566 // Functions are at least 4-byte aligned. If and'ing the address of a
567 // function with a constant < 4, fold it to zero.
568 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
569 if (CI->getValue().ult(APInt(CI->getType()->getBitWidth(),4)) &&
570 isa<Function>(CPR))
571 return Constant::getNullValue(CI->getType());
573 break;
574 case Instruction::Or:
575 if (C2->isNullValue()) return const_cast<Constant*>(C1); // X | 0 == X
576 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C2))
577 if (CI->isAllOnesValue())
578 return const_cast<Constant*>(C2); // X | -1 == -1
579 break;
580 case Instruction::Xor:
581 if (C2->isNullValue()) return const_cast<Constant*>(C1); // X ^ 0 == X
582 break;
583 case Instruction::AShr:
584 // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
585 if (CE1->getOpcode() == Instruction::ZExt) // Top bits known zero.
586 return ConstantExpr::getLShr(const_cast<Constant*>(C1),
587 const_cast<Constant*>(C2));
588 break;
591 } else if (isa<ConstantExpr>(C2)) {
592 // If C2 is a constant expr and C1 isn't, flop them around and fold the
593 // other way if possible.
594 switch (Opcode) {
595 case Instruction::Add:
596 case Instruction::Mul:
597 case Instruction::And:
598 case Instruction::Or:
599 case Instruction::Xor:
600 // No change of opcode required.
601 return ConstantFoldBinaryInstruction(Opcode, C2, C1);
603 case Instruction::Shl:
604 case Instruction::LShr:
605 case Instruction::AShr:
606 case Instruction::Sub:
607 case Instruction::SDiv:
608 case Instruction::UDiv:
609 case Instruction::FDiv:
610 case Instruction::URem:
611 case Instruction::SRem:
612 case Instruction::FRem:
613 default: // These instructions cannot be flopped around.
614 return 0;
618 // At this point we know neither constant is an UndefValue nor a ConstantExpr
619 // so look at directly computing the value.
620 if (const ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
621 if (const ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
622 using namespace APIntOps;
623 APInt C1V = CI1->getValue();
624 APInt C2V = CI2->getValue();
625 switch (Opcode) {
626 default:
627 break;
628 case Instruction::Add:
629 return ConstantInt::get(C1V + C2V);
630 case Instruction::Sub:
631 return ConstantInt::get(C1V - C2V);
632 case Instruction::Mul:
633 return ConstantInt::get(C1V * C2V);
634 case Instruction::UDiv:
635 if (CI2->isNullValue())
636 return 0; // X / 0 -> can't fold
637 return ConstantInt::get(C1V.udiv(C2V));
638 case Instruction::SDiv:
639 if (CI2->isNullValue())
640 return 0; // X / 0 -> can't fold
641 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
642 return 0; // MIN_INT / -1 -> overflow
643 return ConstantInt::get(C1V.sdiv(C2V));
644 case Instruction::URem:
645 if (C2->isNullValue())
646 return 0; // X / 0 -> can't fold
647 return ConstantInt::get(C1V.urem(C2V));
648 case Instruction::SRem:
649 if (CI2->isNullValue())
650 return 0; // X % 0 -> can't fold
651 if (C2V.isAllOnesValue() && C1V.isMinSignedValue())
652 return 0; // MIN_INT % -1 -> overflow
653 return ConstantInt::get(C1V.srem(C2V));
654 case Instruction::And:
655 return ConstantInt::get(C1V & C2V);
656 case Instruction::Or:
657 return ConstantInt::get(C1V | C2V);
658 case Instruction::Xor:
659 return ConstantInt::get(C1V ^ C2V);
660 case Instruction::Shl:
661 if (uint32_t shiftAmt = C2V.getZExtValue())
662 if (shiftAmt < C1V.getBitWidth())
663 return ConstantInt::get(C1V.shl(shiftAmt));
664 else
665 return UndefValue::get(C1->getType()); // too big shift is undef
666 return const_cast<ConstantInt*>(CI1); // Zero shift is identity
667 case Instruction::LShr:
668 if (uint32_t shiftAmt = C2V.getZExtValue())
669 if (shiftAmt < C1V.getBitWidth())
670 return ConstantInt::get(C1V.lshr(shiftAmt));
671 else
672 return UndefValue::get(C1->getType()); // too big shift is undef
673 return const_cast<ConstantInt*>(CI1); // Zero shift is identity
674 case Instruction::AShr:
675 if (uint32_t shiftAmt = C2V.getZExtValue())
676 if (shiftAmt < C1V.getBitWidth())
677 return ConstantInt::get(C1V.ashr(shiftAmt));
678 else
679 return UndefValue::get(C1->getType()); // too big shift is undef
680 return const_cast<ConstantInt*>(CI1); // Zero shift is identity
683 } else if (const ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
684 if (const ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
685 APFloat C1V = CFP1->getValueAPF();
686 APFloat C2V = CFP2->getValueAPF();
687 APFloat C3V = C1V; // copy for modification
688 bool isDouble = CFP1->getType()==Type::DoubleTy;
689 switch (Opcode) {
690 default:
691 break;
692 case Instruction::Add:
693 (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
694 return ConstantFP::get(CFP1->getType(), C3V);
695 case Instruction::Sub:
696 (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
697 return ConstantFP::get(CFP1->getType(), C3V);
698 case Instruction::Mul:
699 (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
700 return ConstantFP::get(CFP1->getType(), C3V);
701 case Instruction::FDiv:
702 // FIXME better to look at the return code
703 if (C2V.isZero())
704 if (C1V.isZero())
705 // IEEE 754, Section 7.1, #4
706 return ConstantFP::get(CFP1->getType(), isDouble ?
707 APFloat(std::numeric_limits<double>::quiet_NaN()) :
708 APFloat(std::numeric_limits<float>::quiet_NaN()));
709 else if (C2V.isNegZero() || C1V.isNegative())
710 // IEEE 754, Section 7.2, negative infinity case
711 return ConstantFP::get(CFP1->getType(), isDouble ?
712 APFloat(-std::numeric_limits<double>::infinity()) :
713 APFloat(-std::numeric_limits<float>::infinity()));
714 else
715 // IEEE 754, Section 7.2, positive infinity case
716 return ConstantFP::get(CFP1->getType(), isDouble ?
717 APFloat(std::numeric_limits<double>::infinity()) :
718 APFloat(std::numeric_limits<float>::infinity()));
719 (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
720 return ConstantFP::get(CFP1->getType(), C3V);
721 case Instruction::FRem:
722 if (C2V.isZero())
723 // IEEE 754, Section 7.1, #5
724 return ConstantFP::get(CFP1->getType(), isDouble ?
725 APFloat(std::numeric_limits<double>::quiet_NaN()) :
726 APFloat(std::numeric_limits<float>::quiet_NaN()));
727 (void)C3V.mod(C2V, APFloat::rmNearestTiesToEven);
728 return ConstantFP::get(CFP1->getType(), C3V);
731 } else if (const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1)) {
732 if (const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2)) {
733 switch (Opcode) {
734 default:
735 break;
736 case Instruction::Add:
737 return EvalVectorOp(CP1, CP2, ConstantExpr::getAdd);
738 case Instruction::Sub:
739 return EvalVectorOp(CP1, CP2, ConstantExpr::getSub);
740 case Instruction::Mul:
741 return EvalVectorOp(CP1, CP2, ConstantExpr::getMul);
742 case Instruction::UDiv:
743 return EvalVectorOp(CP1, CP2, ConstantExpr::getUDiv);
744 case Instruction::SDiv:
745 return EvalVectorOp(CP1, CP2, ConstantExpr::getSDiv);
746 case Instruction::FDiv:
747 return EvalVectorOp(CP1, CP2, ConstantExpr::getFDiv);
748 case Instruction::URem:
749 return EvalVectorOp(CP1, CP2, ConstantExpr::getURem);
750 case Instruction::SRem:
751 return EvalVectorOp(CP1, CP2, ConstantExpr::getSRem);
752 case Instruction::FRem:
753 return EvalVectorOp(CP1, CP2, ConstantExpr::getFRem);
754 case Instruction::And:
755 return EvalVectorOp(CP1, CP2, ConstantExpr::getAnd);
756 case Instruction::Or:
757 return EvalVectorOp(CP1, CP2, ConstantExpr::getOr);
758 case Instruction::Xor:
759 return EvalVectorOp(CP1, CP2, ConstantExpr::getXor);
764 // We don't know how to fold this
765 return 0;
768 /// isZeroSizedType - This type is zero sized if its an array or structure of
769 /// zero sized types. The only leaf zero sized type is an empty structure.
770 static bool isMaybeZeroSizedType(const Type *Ty) {
771 if (isa<OpaqueType>(Ty)) return true; // Can't say.
772 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
774 // If all of elements have zero size, this does too.
775 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
776 if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
777 return true;
779 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
780 return isMaybeZeroSizedType(ATy->getElementType());
782 return false;
785 /// IdxCompare - Compare the two constants as though they were getelementptr
786 /// indices. This allows coersion of the types to be the same thing.
788 /// If the two constants are the "same" (after coersion), return 0. If the
789 /// first is less than the second, return -1, if the second is less than the
790 /// first, return 1. If the constants are not integral, return -2.
792 static int IdxCompare(Constant *C1, Constant *C2, const Type *ElTy) {
793 if (C1 == C2) return 0;
795 // Ok, we found a different index. If they are not ConstantInt, we can't do
796 // anything with them.
797 if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
798 return -2; // don't know!
800 // Ok, we have two differing integer indices. Sign extend them to be the same
801 // type. Long is always big enough, so we use it.
802 if (C1->getType() != Type::Int64Ty)
803 C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
805 if (C2->getType() != Type::Int64Ty)
806 C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
808 if (C1 == C2) return 0; // They are equal
810 // If the type being indexed over is really just a zero sized type, there is
811 // no pointer difference being made here.
812 if (isMaybeZeroSizedType(ElTy))
813 return -2; // dunno.
815 // If they are really different, now that they are the same type, then we
816 // found a difference!
817 if (cast<ConstantInt>(C1)->getSExtValue() <
818 cast<ConstantInt>(C2)->getSExtValue())
819 return -1;
820 else
821 return 1;
824 /// evaluateFCmpRelation - This function determines if there is anything we can
825 /// decide about the two constants provided. This doesn't need to handle simple
826 /// things like ConstantFP comparisons, but should instead handle ConstantExprs.
827 /// If we can determine that the two constants have a particular relation to
828 /// each other, we should return the corresponding FCmpInst predicate,
829 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
830 /// ConstantFoldCompareInstruction.
832 /// To simplify this code we canonicalize the relation so that the first
833 /// operand is always the most "complex" of the two. We consider ConstantFP
834 /// to be the simplest, and ConstantExprs to be the most complex.
835 static FCmpInst::Predicate evaluateFCmpRelation(const Constant *V1,
836 const Constant *V2) {
837 assert(V1->getType() == V2->getType() &&
838 "Cannot compare values of different types!");
839 // Handle degenerate case quickly
840 if (V1 == V2) return FCmpInst::FCMP_OEQ;
842 if (!isa<ConstantExpr>(V1)) {
843 if (!isa<ConstantExpr>(V2)) {
844 // We distilled thisUse the standard constant folder for a few cases
845 ConstantInt *R = 0;
846 Constant *C1 = const_cast<Constant*>(V1);
847 Constant *C2 = const_cast<Constant*>(V2);
848 R = dyn_cast<ConstantInt>(
849 ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, C1, C2));
850 if (R && !R->isZero())
851 return FCmpInst::FCMP_OEQ;
852 R = dyn_cast<ConstantInt>(
853 ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, C1, C2));
854 if (R && !R->isZero())
855 return FCmpInst::FCMP_OLT;
856 R = dyn_cast<ConstantInt>(
857 ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, C1, C2));
858 if (R && !R->isZero())
859 return FCmpInst::FCMP_OGT;
861 // Nothing more we can do
862 return FCmpInst::BAD_FCMP_PREDICATE;
865 // If the first operand is simple and second is ConstantExpr, swap operands.
866 FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
867 if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
868 return FCmpInst::getSwappedPredicate(SwappedRelation);
869 } else {
870 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
871 // constantexpr or a simple constant.
872 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
873 switch (CE1->getOpcode()) {
874 case Instruction::FPTrunc:
875 case Instruction::FPExt:
876 case Instruction::UIToFP:
877 case Instruction::SIToFP:
878 // We might be able to do something with these but we don't right now.
879 break;
880 default:
881 break;
884 // There are MANY other foldings that we could perform here. They will
885 // probably be added on demand, as they seem needed.
886 return FCmpInst::BAD_FCMP_PREDICATE;
889 /// evaluateICmpRelation - This function determines if there is anything we can
890 /// decide about the two constants provided. This doesn't need to handle simple
891 /// things like integer comparisons, but should instead handle ConstantExprs
892 /// and GlobalValues. If we can determine that the two constants have a
893 /// particular relation to each other, we should return the corresponding ICmp
894 /// predicate, otherwise return ICmpInst::BAD_ICMP_PREDICATE.
896 /// To simplify this code we canonicalize the relation so that the first
897 /// operand is always the most "complex" of the two. We consider simple
898 /// constants (like ConstantInt) to be the simplest, followed by
899 /// GlobalValues, followed by ConstantExpr's (the most complex).
901 static ICmpInst::Predicate evaluateICmpRelation(const Constant *V1,
902 const Constant *V2,
903 bool isSigned) {
904 assert(V1->getType() == V2->getType() &&
905 "Cannot compare different types of values!");
906 if (V1 == V2) return ICmpInst::ICMP_EQ;
908 if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1)) {
909 if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2)) {
910 // We distilled this down to a simple case, use the standard constant
911 // folder.
912 ConstantInt *R = 0;
913 Constant *C1 = const_cast<Constant*>(V1);
914 Constant *C2 = const_cast<Constant*>(V2);
915 ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
916 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
917 if (R && !R->isZero())
918 return pred;
919 pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
920 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
921 if (R && !R->isZero())
922 return pred;
923 pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
924 R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, C1, C2));
925 if (R && !R->isZero())
926 return pred;
928 // If we couldn't figure it out, bail.
929 return ICmpInst::BAD_ICMP_PREDICATE;
932 // If the first operand is simple, swap operands.
933 ICmpInst::Predicate SwappedRelation =
934 evaluateICmpRelation(V2, V1, isSigned);
935 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
936 return ICmpInst::getSwappedPredicate(SwappedRelation);
938 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(V1)) {
939 if (isa<ConstantExpr>(V2)) { // Swap as necessary.
940 ICmpInst::Predicate SwappedRelation =
941 evaluateICmpRelation(V2, V1, isSigned);
942 if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
943 return ICmpInst::getSwappedPredicate(SwappedRelation);
944 else
945 return ICmpInst::BAD_ICMP_PREDICATE;
948 // Now we know that the RHS is a GlobalValue or simple constant,
949 // which (since the types must match) means that it's a ConstantPointerNull.
950 if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
951 // Don't try to decide equality of aliases.
952 if (!isa<GlobalAlias>(CPR1) && !isa<GlobalAlias>(CPR2))
953 if (!CPR1->hasExternalWeakLinkage() || !CPR2->hasExternalWeakLinkage())
954 return ICmpInst::ICMP_NE;
955 } else {
956 assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
957 // GlobalVals can never be null. Don't try to evaluate aliases.
958 if (!CPR1->hasExternalWeakLinkage() && !isa<GlobalAlias>(CPR1))
959 return ICmpInst::ICMP_NE;
961 } else {
962 // Ok, the LHS is known to be a constantexpr. The RHS can be any of a
963 // constantexpr, a CPR, or a simple constant.
964 const ConstantExpr *CE1 = cast<ConstantExpr>(V1);
965 const Constant *CE1Op0 = CE1->getOperand(0);
967 switch (CE1->getOpcode()) {
968 case Instruction::Trunc:
969 case Instruction::FPTrunc:
970 case Instruction::FPExt:
971 case Instruction::FPToUI:
972 case Instruction::FPToSI:
973 break; // We can't evaluate floating point casts or truncations.
975 case Instruction::UIToFP:
976 case Instruction::SIToFP:
977 case Instruction::IntToPtr:
978 case Instruction::BitCast:
979 case Instruction::ZExt:
980 case Instruction::SExt:
981 case Instruction::PtrToInt:
982 // If the cast is not actually changing bits, and the second operand is a
983 // null pointer, do the comparison with the pre-casted value.
984 if (V2->isNullValue() &&
985 (isa<PointerType>(CE1->getType()) || CE1->getType()->isInteger())) {
986 bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false :
987 (CE1->getOpcode() == Instruction::SExt ? true :
988 (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned));
989 return evaluateICmpRelation(
990 CE1Op0, Constant::getNullValue(CE1Op0->getType()), sgnd);
993 // If the dest type is a pointer type, and the RHS is a constantexpr cast
994 // from the same type as the src of the LHS, evaluate the inputs. This is
995 // important for things like "icmp eq (cast 4 to int*), (cast 5 to int*)",
996 // which happens a lot in compilers with tagged integers.
997 if (const ConstantExpr *CE2 = dyn_cast<ConstantExpr>(V2))
998 if (CE2->isCast() && isa<PointerType>(CE1->getType()) &&
999 CE1->getOperand(0)->getType() == CE2->getOperand(0)->getType() &&
1000 CE1->getOperand(0)->getType()->isInteger()) {
1001 bool sgnd = CE1->getOpcode() == Instruction::ZExt ? false :
1002 (CE1->getOpcode() == Instruction::SExt ? true :
1003 (CE1->getOpcode() == Instruction::PtrToInt ? false : isSigned));
1004 return evaluateICmpRelation(CE1->getOperand(0), CE2->getOperand(0),
1005 sgnd);
1007 break;
1009 case Instruction::GetElementPtr:
1010 // Ok, since this is a getelementptr, we know that the constant has a
1011 // pointer type. Check the various cases.
1012 if (isa<ConstantPointerNull>(V2)) {
1013 // If we are comparing a GEP to a null pointer, check to see if the base
1014 // of the GEP equals the null pointer.
1015 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1016 if (GV->hasExternalWeakLinkage())
1017 // Weak linkage GVals could be zero or not. We're comparing that
1018 // to null pointer so its greater-or-equal
1019 return isSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1020 else
1021 // If its not weak linkage, the GVal must have a non-zero address
1022 // so the result is greater-than
1023 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1024 } else if (isa<ConstantPointerNull>(CE1Op0)) {
1025 // If we are indexing from a null pointer, check to see if we have any
1026 // non-zero indices.
1027 for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1028 if (!CE1->getOperand(i)->isNullValue())
1029 // Offsetting from null, must not be equal.
1030 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1031 // Only zero indexes from null, must still be zero.
1032 return ICmpInst::ICMP_EQ;
1034 // Otherwise, we can't really say if the first operand is null or not.
1035 } else if (const GlobalValue *CPR2 = dyn_cast<GlobalValue>(V2)) {
1036 if (isa<ConstantPointerNull>(CE1Op0)) {
1037 if (CPR2->hasExternalWeakLinkage())
1038 // Weak linkage GVals could be zero or not. We're comparing it to
1039 // a null pointer, so its less-or-equal
1040 return isSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1041 else
1042 // If its not weak linkage, the GVal must have a non-zero address
1043 // so the result is less-than
1044 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1045 } else if (const GlobalValue *CPR1 = dyn_cast<GlobalValue>(CE1Op0)) {
1046 if (CPR1 == CPR2) {
1047 // If this is a getelementptr of the same global, then it must be
1048 // different. Because the types must match, the getelementptr could
1049 // only have at most one index, and because we fold getelementptr's
1050 // with a single zero index, it must be nonzero.
1051 assert(CE1->getNumOperands() == 2 &&
1052 !CE1->getOperand(1)->isNullValue() &&
1053 "Suprising getelementptr!");
1054 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1055 } else {
1056 // If they are different globals, we don't know what the value is,
1057 // but they can't be equal.
1058 return ICmpInst::ICMP_NE;
1061 } else {
1062 const ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1063 const Constant *CE2Op0 = CE2->getOperand(0);
1065 // There are MANY other foldings that we could perform here. They will
1066 // probably be added on demand, as they seem needed.
1067 switch (CE2->getOpcode()) {
1068 default: break;
1069 case Instruction::GetElementPtr:
1070 // By far the most common case to handle is when the base pointers are
1071 // obviously to the same or different globals.
1072 if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1073 if (CE1Op0 != CE2Op0) // Don't know relative ordering, but not equal
1074 return ICmpInst::ICMP_NE;
1075 // Ok, we know that both getelementptr instructions are based on the
1076 // same global. From this, we can precisely determine the relative
1077 // ordering of the resultant pointers.
1078 unsigned i = 1;
1080 // Compare all of the operands the GEP's have in common.
1081 gep_type_iterator GTI = gep_type_begin(CE1);
1082 for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1083 ++i, ++GTI)
1084 switch (IdxCompare(CE1->getOperand(i), CE2->getOperand(i),
1085 GTI.getIndexedType())) {
1086 case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1087 case 1: return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1088 case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1091 // Ok, we ran out of things they have in common. If any leftovers
1092 // are non-zero then we have a difference, otherwise we are equal.
1093 for (; i < CE1->getNumOperands(); ++i)
1094 if (!CE1->getOperand(i)->isNullValue())
1095 if (isa<ConstantInt>(CE1->getOperand(i)))
1096 return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1097 else
1098 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1100 for (; i < CE2->getNumOperands(); ++i)
1101 if (!CE2->getOperand(i)->isNullValue())
1102 if (isa<ConstantInt>(CE2->getOperand(i)))
1103 return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1104 else
1105 return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1106 return ICmpInst::ICMP_EQ;
1110 default:
1111 break;
1115 return ICmpInst::BAD_ICMP_PREDICATE;
1118 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1119 const Constant *C1,
1120 const Constant *C2) {
1122 // Handle some degenerate cases first
1123 if (isa<UndefValue>(C1) || isa<UndefValue>(C2))
1124 return UndefValue::get(Type::Int1Ty);
1126 // icmp eq/ne(null,GV) -> false/true
1127 if (C1->isNullValue()) {
1128 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1129 // Don't try to evaluate aliases. External weak GV can be null.
1130 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage())
1131 if (pred == ICmpInst::ICMP_EQ)
1132 return ConstantInt::getFalse();
1133 else if (pred == ICmpInst::ICMP_NE)
1134 return ConstantInt::getTrue();
1135 // icmp eq/ne(GV,null) -> false/true
1136 } else if (C2->isNullValue()) {
1137 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1))
1138 // Don't try to evaluate aliases. External weak GV can be null.
1139 if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage())
1140 if (pred == ICmpInst::ICMP_EQ)
1141 return ConstantInt::getFalse();
1142 else if (pred == ICmpInst::ICMP_NE)
1143 return ConstantInt::getTrue();
1146 if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1147 APInt V1 = cast<ConstantInt>(C1)->getValue();
1148 APInt V2 = cast<ConstantInt>(C2)->getValue();
1149 switch (pred) {
1150 default: assert(0 && "Invalid ICmp Predicate"); return 0;
1151 case ICmpInst::ICMP_EQ: return ConstantInt::get(Type::Int1Ty, V1 == V2);
1152 case ICmpInst::ICMP_NE: return ConstantInt::get(Type::Int1Ty, V1 != V2);
1153 case ICmpInst::ICMP_SLT:return ConstantInt::get(Type::Int1Ty, V1.slt(V2));
1154 case ICmpInst::ICMP_SGT:return ConstantInt::get(Type::Int1Ty, V1.sgt(V2));
1155 case ICmpInst::ICMP_SLE:return ConstantInt::get(Type::Int1Ty, V1.sle(V2));
1156 case ICmpInst::ICMP_SGE:return ConstantInt::get(Type::Int1Ty, V1.sge(V2));
1157 case ICmpInst::ICMP_ULT:return ConstantInt::get(Type::Int1Ty, V1.ult(V2));
1158 case ICmpInst::ICMP_UGT:return ConstantInt::get(Type::Int1Ty, V1.ugt(V2));
1159 case ICmpInst::ICMP_ULE:return ConstantInt::get(Type::Int1Ty, V1.ule(V2));
1160 case ICmpInst::ICMP_UGE:return ConstantInt::get(Type::Int1Ty, V1.uge(V2));
1162 } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1163 APFloat C1V = cast<ConstantFP>(C1)->getValueAPF();
1164 APFloat C2V = cast<ConstantFP>(C2)->getValueAPF();
1165 APFloat::cmpResult R = C1V.compare(C2V);
1166 switch (pred) {
1167 default: assert(0 && "Invalid FCmp Predicate"); return 0;
1168 case FCmpInst::FCMP_FALSE: return ConstantInt::getFalse();
1169 case FCmpInst::FCMP_TRUE: return ConstantInt::getTrue();
1170 case FCmpInst::FCMP_UNO:
1171 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered);
1172 case FCmpInst::FCMP_ORD:
1173 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpUnordered);
1174 case FCmpInst::FCMP_UEQ:
1175 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1176 R==APFloat::cmpEqual);
1177 case FCmpInst::FCMP_OEQ:
1178 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpEqual);
1179 case FCmpInst::FCMP_UNE:
1180 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpEqual);
1181 case FCmpInst::FCMP_ONE:
1182 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1183 R==APFloat::cmpGreaterThan);
1184 case FCmpInst::FCMP_ULT:
1185 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1186 R==APFloat::cmpLessThan);
1187 case FCmpInst::FCMP_OLT:
1188 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan);
1189 case FCmpInst::FCMP_UGT:
1190 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpUnordered ||
1191 R==APFloat::cmpGreaterThan);
1192 case FCmpInst::FCMP_OGT:
1193 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan);
1194 case FCmpInst::FCMP_ULE:
1195 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpGreaterThan);
1196 case FCmpInst::FCMP_OLE:
1197 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpLessThan ||
1198 R==APFloat::cmpEqual);
1199 case FCmpInst::FCMP_UGE:
1200 return ConstantInt::get(Type::Int1Ty, R!=APFloat::cmpLessThan);
1201 case FCmpInst::FCMP_OGE:
1202 return ConstantInt::get(Type::Int1Ty, R==APFloat::cmpGreaterThan ||
1203 R==APFloat::cmpEqual);
1205 } else if (const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1)) {
1206 if (const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2)) {
1207 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ) {
1208 for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) {
1209 Constant *C= ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ,
1210 const_cast<Constant*>(CP1->getOperand(i)),
1211 const_cast<Constant*>(CP2->getOperand(i)));
1212 if (ConstantInt *CB = dyn_cast<ConstantInt>(C))
1213 return CB;
1215 // Otherwise, could not decide from any element pairs.
1216 return 0;
1217 } else if (pred == ICmpInst::ICMP_EQ) {
1218 for (unsigned i = 0, e = CP1->getNumOperands(); i != e; ++i) {
1219 Constant *C = ConstantExpr::getICmp(ICmpInst::ICMP_EQ,
1220 const_cast<Constant*>(CP1->getOperand(i)),
1221 const_cast<Constant*>(CP2->getOperand(i)));
1222 if (ConstantInt *CB = dyn_cast<ConstantInt>(C))
1223 return CB;
1225 // Otherwise, could not decide from any element pairs.
1226 return 0;
1231 if (C1->getType()->isFloatingPoint()) {
1232 switch (evaluateFCmpRelation(C1, C2)) {
1233 default: assert(0 && "Unknown relation!");
1234 case FCmpInst::FCMP_UNO:
1235 case FCmpInst::FCMP_ORD:
1236 case FCmpInst::FCMP_UEQ:
1237 case FCmpInst::FCMP_UNE:
1238 case FCmpInst::FCMP_ULT:
1239 case FCmpInst::FCMP_UGT:
1240 case FCmpInst::FCMP_ULE:
1241 case FCmpInst::FCMP_UGE:
1242 case FCmpInst::FCMP_TRUE:
1243 case FCmpInst::FCMP_FALSE:
1244 case FCmpInst::BAD_FCMP_PREDICATE:
1245 break; // Couldn't determine anything about these constants.
1246 case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1247 return ConstantInt::get(Type::Int1Ty,
1248 pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1249 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1250 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1251 case FCmpInst::FCMP_OLT: // We know that C1 < C2
1252 return ConstantInt::get(Type::Int1Ty,
1253 pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1254 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1255 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1256 case FCmpInst::FCMP_OGT: // We know that C1 > C2
1257 return ConstantInt::get(Type::Int1Ty,
1258 pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1259 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1260 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1261 case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1262 // We can only partially decide this relation.
1263 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1264 return ConstantInt::getFalse();
1265 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1266 return ConstantInt::getTrue();
1267 break;
1268 case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1269 // We can only partially decide this relation.
1270 if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1271 return ConstantInt::getFalse();
1272 if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1273 return ConstantInt::getTrue();
1274 break;
1275 case ICmpInst::ICMP_NE: // We know that C1 != C2
1276 // We can only partially decide this relation.
1277 if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1278 return ConstantInt::getFalse();
1279 if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1280 return ConstantInt::getTrue();
1281 break;
1283 } else {
1284 // Evaluate the relation between the two constants, per the predicate.
1285 switch (evaluateICmpRelation(C1, C2, CmpInst::isSigned(pred))) {
1286 default: assert(0 && "Unknown relational!");
1287 case ICmpInst::BAD_ICMP_PREDICATE:
1288 break; // Couldn't determine anything about these constants.
1289 case ICmpInst::ICMP_EQ: // We know the constants are equal!
1290 // If we know the constants are equal, we can decide the result of this
1291 // computation precisely.
1292 return ConstantInt::get(Type::Int1Ty,
1293 pred == ICmpInst::ICMP_EQ ||
1294 pred == ICmpInst::ICMP_ULE ||
1295 pred == ICmpInst::ICMP_SLE ||
1296 pred == ICmpInst::ICMP_UGE ||
1297 pred == ICmpInst::ICMP_SGE);
1298 case ICmpInst::ICMP_ULT:
1299 // If we know that C1 < C2, we can decide the result of this computation
1300 // precisely.
1301 return ConstantInt::get(Type::Int1Ty,
1302 pred == ICmpInst::ICMP_ULT ||
1303 pred == ICmpInst::ICMP_NE ||
1304 pred == ICmpInst::ICMP_ULE);
1305 case ICmpInst::ICMP_SLT:
1306 // If we know that C1 < C2, we can decide the result of this computation
1307 // precisely.
1308 return ConstantInt::get(Type::Int1Ty,
1309 pred == ICmpInst::ICMP_SLT ||
1310 pred == ICmpInst::ICMP_NE ||
1311 pred == ICmpInst::ICMP_SLE);
1312 case ICmpInst::ICMP_UGT:
1313 // If we know that C1 > C2, we can decide the result of this computation
1314 // precisely.
1315 return ConstantInt::get(Type::Int1Ty,
1316 pred == ICmpInst::ICMP_UGT ||
1317 pred == ICmpInst::ICMP_NE ||
1318 pred == ICmpInst::ICMP_UGE);
1319 case ICmpInst::ICMP_SGT:
1320 // If we know that C1 > C2, we can decide the result of this computation
1321 // precisely.
1322 return ConstantInt::get(Type::Int1Ty,
1323 pred == ICmpInst::ICMP_SGT ||
1324 pred == ICmpInst::ICMP_NE ||
1325 pred == ICmpInst::ICMP_SGE);
1326 case ICmpInst::ICMP_ULE:
1327 // If we know that C1 <= C2, we can only partially decide this relation.
1328 if (pred == ICmpInst::ICMP_UGT) return ConstantInt::getFalse();
1329 if (pred == ICmpInst::ICMP_ULT) return ConstantInt::getTrue();
1330 break;
1331 case ICmpInst::ICMP_SLE:
1332 // If we know that C1 <= C2, we can only partially decide this relation.
1333 if (pred == ICmpInst::ICMP_SGT) return ConstantInt::getFalse();
1334 if (pred == ICmpInst::ICMP_SLT) return ConstantInt::getTrue();
1335 break;
1337 case ICmpInst::ICMP_UGE:
1338 // If we know that C1 >= C2, we can only partially decide this relation.
1339 if (pred == ICmpInst::ICMP_ULT) return ConstantInt::getFalse();
1340 if (pred == ICmpInst::ICMP_UGT) return ConstantInt::getTrue();
1341 break;
1342 case ICmpInst::ICMP_SGE:
1343 // If we know that C1 >= C2, we can only partially decide this relation.
1344 if (pred == ICmpInst::ICMP_SLT) return ConstantInt::getFalse();
1345 if (pred == ICmpInst::ICMP_SGT) return ConstantInt::getTrue();
1346 break;
1348 case ICmpInst::ICMP_NE:
1349 // If we know that C1 != C2, we can only partially decide this relation.
1350 if (pred == ICmpInst::ICMP_EQ) return ConstantInt::getFalse();
1351 if (pred == ICmpInst::ICMP_NE) return ConstantInt::getTrue();
1352 break;
1355 if (!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) {
1356 // If C2 is a constant expr and C1 isn't, flop them around and fold the
1357 // other way if possible.
1358 switch (pred) {
1359 case ICmpInst::ICMP_EQ:
1360 case ICmpInst::ICMP_NE:
1361 // No change of predicate required.
1362 return ConstantFoldCompareInstruction(pred, C2, C1);
1364 case ICmpInst::ICMP_ULT:
1365 case ICmpInst::ICMP_SLT:
1366 case ICmpInst::ICMP_UGT:
1367 case ICmpInst::ICMP_SGT:
1368 case ICmpInst::ICMP_ULE:
1369 case ICmpInst::ICMP_SLE:
1370 case ICmpInst::ICMP_UGE:
1371 case ICmpInst::ICMP_SGE:
1372 // Change the predicate as necessary to swap the operands.
1373 pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
1374 return ConstantFoldCompareInstruction(pred, C2, C1);
1376 default: // These predicates cannot be flopped around.
1377 break;
1381 return 0;
1384 Constant *llvm::ConstantFoldGetElementPtr(const Constant *C,
1385 Constant* const *Idxs,
1386 unsigned NumIdx) {
1387 if (NumIdx == 0 ||
1388 (NumIdx == 1 && Idxs[0]->isNullValue()))
1389 return const_cast<Constant*>(C);
1391 if (isa<UndefValue>(C)) {
1392 const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(),
1393 (Value **)Idxs,
1394 (Value **)Idxs+NumIdx,
1395 true);
1396 assert(Ty != 0 && "Invalid indices for GEP!");
1397 return UndefValue::get(PointerType::get(Ty));
1400 Constant *Idx0 = Idxs[0];
1401 if (C->isNullValue()) {
1402 bool isNull = true;
1403 for (unsigned i = 0, e = NumIdx; i != e; ++i)
1404 if (!Idxs[i]->isNullValue()) {
1405 isNull = false;
1406 break;
1408 if (isNull) {
1409 const Type *Ty = GetElementPtrInst::getIndexedType(C->getType(),
1410 (Value**)Idxs,
1411 (Value**)Idxs+NumIdx,
1412 true);
1413 assert(Ty != 0 && "Invalid indices for GEP!");
1414 return ConstantPointerNull::get(PointerType::get(Ty));
1418 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(const_cast<Constant*>(C))) {
1419 // Combine Indices - If the source pointer to this getelementptr instruction
1420 // is a getelementptr instruction, combine the indices of the two
1421 // getelementptr instructions into a single instruction.
1423 if (CE->getOpcode() == Instruction::GetElementPtr) {
1424 const Type *LastTy = 0;
1425 for (gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
1426 I != E; ++I)
1427 LastTy = *I;
1429 if ((LastTy && isa<ArrayType>(LastTy)) || Idx0->isNullValue()) {
1430 SmallVector<Value*, 16> NewIndices;
1431 NewIndices.reserve(NumIdx + CE->getNumOperands());
1432 for (unsigned i = 1, e = CE->getNumOperands()-1; i != e; ++i)
1433 NewIndices.push_back(CE->getOperand(i));
1435 // Add the last index of the source with the first index of the new GEP.
1436 // Make sure to handle the case when they are actually different types.
1437 Constant *Combined = CE->getOperand(CE->getNumOperands()-1);
1438 // Otherwise it must be an array.
1439 if (!Idx0->isNullValue()) {
1440 const Type *IdxTy = Combined->getType();
1441 if (IdxTy != Idx0->getType()) {
1442 Constant *C1 = ConstantExpr::getSExtOrBitCast(Idx0, Type::Int64Ty);
1443 Constant *C2 = ConstantExpr::getSExtOrBitCast(Combined,
1444 Type::Int64Ty);
1445 Combined = ConstantExpr::get(Instruction::Add, C1, C2);
1446 } else {
1447 Combined =
1448 ConstantExpr::get(Instruction::Add, Idx0, Combined);
1452 NewIndices.push_back(Combined);
1453 NewIndices.insert(NewIndices.end(), Idxs+1, Idxs+NumIdx);
1454 return ConstantExpr::getGetElementPtr(CE->getOperand(0), &NewIndices[0],
1455 NewIndices.size());
1459 // Implement folding of:
1460 // int* getelementptr ([2 x int]* cast ([3 x int]* %X to [2 x int]*),
1461 // long 0, long 0)
1462 // To: int* getelementptr ([3 x int]* %X, long 0, long 0)
1464 if (CE->isCast() && NumIdx > 1 && Idx0->isNullValue()) {
1465 if (const PointerType *SPT =
1466 dyn_cast<PointerType>(CE->getOperand(0)->getType()))
1467 if (const ArrayType *SAT = dyn_cast<ArrayType>(SPT->getElementType()))
1468 if (const ArrayType *CAT =
1469 dyn_cast<ArrayType>(cast<PointerType>(C->getType())->getElementType()))
1470 if (CAT->getElementType() == SAT->getElementType())
1471 return ConstantExpr::getGetElementPtr(
1472 (Constant*)CE->getOperand(0), Idxs, NumIdx);
1475 // Fold: getelementptr (i8* inttoptr (i64 1 to i8*), i32 -1)
1476 // Into: inttoptr (i64 0 to i8*)
1477 // This happens with pointers to member functions in C++.
1478 if (CE->getOpcode() == Instruction::IntToPtr && NumIdx == 1 &&
1479 isa<ConstantInt>(CE->getOperand(0)) && isa<ConstantInt>(Idxs[0]) &&
1480 cast<PointerType>(CE->getType())->getElementType() == Type::Int8Ty) {
1481 Constant *Base = CE->getOperand(0);
1482 Constant *Offset = Idxs[0];
1484 // Convert the smaller integer to the larger type.
1485 if (Offset->getType()->getPrimitiveSizeInBits() <
1486 Base->getType()->getPrimitiveSizeInBits())
1487 Offset = ConstantExpr::getSExt(Offset, Base->getType());
1488 else if (Base->getType()->getPrimitiveSizeInBits() <
1489 Offset->getType()->getPrimitiveSizeInBits())
1490 Base = ConstantExpr::getZExt(Base, Base->getType());
1492 Base = ConstantExpr::getAdd(Base, Offset);
1493 return ConstantExpr::getIntToPtr(Base, CE->getType());
1496 return 0;