1 //===- InstCombineCasts.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for cast operations.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Target/TargetData.h"
16 #include "llvm/Support/PatternMatch.h"
18 using namespace PatternMatch;
20 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
21 /// expression. If so, decompose it, returning some value X, such that Val is
24 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
26 assert(Val->getType()->isInteger(32) && "Unexpected allocation size type!");
27 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
28 Offset = CI->getZExtValue();
30 return ConstantInt::get(Type::getInt32Ty(Val->getContext()), 0);
31 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
32 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
33 if (I->getOpcode() == Instruction::Shl) {
34 // This is a value scaled by '1 << the shift amt'.
35 Scale = 1U << RHS->getZExtValue();
37 return I->getOperand(0);
38 } else if (I->getOpcode() == Instruction::Mul) {
39 // This value is scaled by 'RHS'.
40 Scale = RHS->getZExtValue();
42 return I->getOperand(0);
43 } else if (I->getOpcode() == Instruction::Add) {
44 // We have X+C. Check to see if we really have (X*C2)+C1,
45 // where C1 is divisible by C2.
48 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
49 Offset += RHS->getZExtValue();
56 // Otherwise, we can't look past this.
62 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
63 /// try to eliminate the cast by moving the type information into the alloc.
64 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
66 // This requires TargetData to get the alloca alignment and size information.
69 const PointerType *PTy = cast<PointerType>(CI.getType());
71 BuilderTy AllocaBuilder(*Builder);
72 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
74 // Get the type really allocated and the type casted to.
75 const Type *AllocElTy = AI.getAllocatedType();
76 const Type *CastElTy = PTy->getElementType();
77 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
79 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
80 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
81 if (CastElTyAlign < AllocElTyAlign) return 0;
83 // If the allocation has multiple uses, only promote it if we are strictly
84 // increasing the alignment of the resultant allocation. If we keep it the
85 // same, we open the door to infinite loops of various kinds. (A reference
86 // from a dbg.declare doesn't count as a use for this purpose.)
87 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
88 CastElTyAlign == AllocElTyAlign) return 0;
90 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
91 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
92 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
94 // See if we can satisfy the modulus by pulling a scale out of the array
96 unsigned ArraySizeScale;
98 Value *NumElements = // See if the array size is a decomposable linear expr.
99 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
101 // If we can now satisfy the modulus, by using a non-1 scale, we really can
103 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
104 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
106 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
111 Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale);
112 // Insert before the alloca, not before the cast.
113 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
116 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
117 Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()),
119 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
122 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
123 New->setAlignment(AI.getAlignment());
126 // If the allocation has one real use plus a dbg.declare, just remove the
128 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
129 EraseInstFromFunction(*(Instruction*)DI);
131 // If the allocation has multiple real uses, insert a cast and change all
132 // things that used it to use the new cast. This will also hack on CI, but it
134 else if (!AI.hasOneUse()) {
135 // New is the allocation instruction, pointer typed. AI is the original
136 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
137 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
138 AI.replaceAllUsesWith(NewCast);
140 return ReplaceInstUsesWith(CI, New);
144 /// CanEvaluateInDifferentType - Return true if we can take the specified value
145 /// and return it as type Ty without inserting any new casts and without
146 /// changing the computed value. This is used by code that tries to decide
147 /// whether promoting or shrinking integer operations to wider or smaller types
148 /// will allow us to eliminate a truncate or extend.
150 /// This is a truncation operation if Ty is smaller than V->getType(), or an
151 /// extension operation if Ty is larger.
153 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
154 /// should return true if trunc(V) can be computed by computing V in the smaller
155 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
156 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
157 /// efficiently truncated.
159 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
160 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
161 /// the final result.
162 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
164 int &NumCastsRemoved){
165 // We can always evaluate constants in another type.
166 if (isa<Constant>(V))
169 Instruction *I = dyn_cast<Instruction>(V);
170 if (!I) return false;
172 const Type *OrigTy = V->getType();
174 // If this is an extension or truncate, we can often eliminate it.
175 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
176 // If this is a cast from the destination type, we can trivially eliminate
177 // it, and this will remove a cast overall.
178 if (I->getOperand(0)->getType() == Ty) {
179 // If the first operand is itself a cast, and is eliminable, do not count
180 // this as an eliminable cast. We would prefer to eliminate those two
182 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
188 // We can't extend or shrink something that has multiple uses: doing so would
189 // require duplicating the instruction in general, which isn't profitable.
190 if (!I->hasOneUse()) return false;
192 unsigned Opc = I->getOpcode();
194 case Instruction::Add:
195 case Instruction::Sub:
196 case Instruction::Mul:
197 case Instruction::And:
198 case Instruction::Or:
199 case Instruction::Xor:
200 // These operators can all arbitrarily be extended or truncated.
201 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
203 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
206 case Instruction::UDiv:
207 case Instruction::URem: {
208 // UDiv and URem can be truncated if all the truncated bits are zero.
209 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
210 uint32_t BitWidth = Ty->getScalarSizeInBits();
211 if (BitWidth < OrigBitWidth) {
212 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
213 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
214 MaskedValueIsZero(I->getOperand(1), Mask)) {
215 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
217 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
223 case Instruction::Shl:
224 // If we are truncating the result of this SHL, and if it's a shift of a
225 // constant amount, we can always perform a SHL in a smaller type.
226 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
227 uint32_t BitWidth = Ty->getScalarSizeInBits();
228 if (BitWidth < OrigTy->getScalarSizeInBits() &&
229 CI->getLimitedValue(BitWidth) < BitWidth)
230 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
234 case Instruction::LShr:
235 // If this is a truncate of a logical shr, we can truncate it to a smaller
236 // lshr iff we know that the bits we would otherwise be shifting in are
238 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
239 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
240 uint32_t BitWidth = Ty->getScalarSizeInBits();
241 if (BitWidth < OrigBitWidth &&
242 MaskedValueIsZero(I->getOperand(0),
243 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
244 CI->getLimitedValue(BitWidth) < BitWidth) {
245 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
250 case Instruction::ZExt:
251 case Instruction::SExt:
252 case Instruction::Trunc:
253 // If this is the same kind of case as our original (e.g. zext+zext), we
254 // can safely replace it. Note that replacing it does not reduce the number
255 // of casts in the input.
259 // sext (zext ty1), ty2 -> zext ty2
260 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
263 case Instruction::Select: {
264 SelectInst *SI = cast<SelectInst>(I);
265 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
267 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
270 case Instruction::PHI: {
271 // We can change a phi if we can change all operands.
272 PHINode *PN = cast<PHINode>(I);
273 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
274 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
280 // TODO: Can handle more cases here.
287 /// EvaluateInDifferentType - Given an expression that
288 /// CanEvaluateInDifferentType returns true for, actually insert the code to
289 /// evaluate the expression.
290 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
292 if (Constant *C = dyn_cast<Constant>(V))
293 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
295 // Otherwise, it must be an instruction.
296 Instruction *I = cast<Instruction>(V);
297 Instruction *Res = 0;
298 unsigned Opc = I->getOpcode();
300 case Instruction::Add:
301 case Instruction::Sub:
302 case Instruction::Mul:
303 case Instruction::And:
304 case Instruction::Or:
305 case Instruction::Xor:
306 case Instruction::AShr:
307 case Instruction::LShr:
308 case Instruction::Shl:
309 case Instruction::UDiv:
310 case Instruction::URem: {
311 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
312 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
313 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
316 case Instruction::Trunc:
317 case Instruction::ZExt:
318 case Instruction::SExt:
319 // If the source type of the cast is the type we're trying for then we can
320 // just return the source. There's no need to insert it because it is not
322 if (I->getOperand(0)->getType() == Ty)
323 return I->getOperand(0);
325 // Otherwise, must be the same type of cast, so just reinsert a new one.
326 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
328 case Instruction::Select: {
329 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
330 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
331 Res = SelectInst::Create(I->getOperand(0), True, False);
334 case Instruction::PHI: {
335 PHINode *OPN = cast<PHINode>(I);
336 PHINode *NPN = PHINode::Create(Ty);
337 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
338 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
339 NPN->addIncoming(V, OPN->getIncomingBlock(i));
345 // TODO: Can handle more cases here.
346 llvm_unreachable("Unreachable!");
351 return InsertNewInstBefore(Res, *I);
355 /// This function is a wrapper around CastInst::isEliminableCastPair. It
356 /// simply extracts arguments and returns what that function returns.
357 static Instruction::CastOps
358 isEliminableCastPair(
359 const CastInst *CI, ///< The first cast instruction
360 unsigned opcode, ///< The opcode of the second cast instruction
361 const Type *DstTy, ///< The target type for the second cast instruction
362 TargetData *TD ///< The target data for pointer size
365 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
366 const Type *MidTy = CI->getType(); // B from above
368 // Get the opcodes of the two Cast instructions
369 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
370 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
372 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
374 TD ? TD->getIntPtrType(CI->getContext()) : 0);
376 // We don't want to form an inttoptr or ptrtoint that converts to an integer
377 // type that differs from the pointer size.
378 if ((Res == Instruction::IntToPtr &&
379 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
380 (Res == Instruction::PtrToInt &&
381 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
384 return Instruction::CastOps(Res);
387 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
388 /// in any code being generated. It does not require codegen if V is simple
389 /// enough or if the cast can be folded into other casts.
390 bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V,
392 if (V->getType() == Ty || isa<Constant>(V)) return false;
394 // If this is another cast that can be eliminated, it isn't codegen either.
395 if (const CastInst *CI = dyn_cast<CastInst>(V))
396 if (isEliminableCastPair(CI, opcode, Ty, TD))
402 /// @brief Implement the transforms common to all CastInst visitors.
403 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
404 Value *Src = CI.getOperand(0);
406 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
408 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
409 if (Instruction::CastOps opc =
410 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
411 // The first cast (CSrc) is eliminable so we need to fix up or replace
412 // the second cast (CI). CSrc will then have a good chance of being dead.
413 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
417 // If we are casting a select then fold the cast into the select
418 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
419 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
422 // If we are casting a PHI then fold the cast into the PHI
423 if (isa<PHINode>(Src)) {
424 // We don't do this if this would create a PHI node with an illegal type if
425 // it is currently legal.
426 if (!isa<IntegerType>(Src->getType()) ||
427 !isa<IntegerType>(CI.getType()) ||
428 ShouldChangeType(CI.getType(), Src->getType()))
429 if (Instruction *NV = FoldOpIntoPhi(CI))
436 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
437 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
438 Value *Src = CI.getOperand(0);
440 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
441 // If casting the result of a getelementptr instruction with no offset, turn
442 // this into a cast of the original pointer!
443 if (GEP->hasAllZeroIndices()) {
444 // Changing the cast operand is usually not a good idea but it is safe
445 // here because the pointer operand is being replaced with another
446 // pointer operand so the opcode doesn't need to change.
448 CI.setOperand(0, GEP->getOperand(0));
452 // If the GEP has a single use, and the base pointer is a bitcast, and the
453 // GEP computes a constant offset, see if we can convert these three
454 // instructions into fewer. This typically happens with unions and other
455 // non-type-safe code.
456 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
457 if (GEP->hasAllConstantIndices()) {
458 // We are guaranteed to get a constant from EmitGEPOffset.
459 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
460 int64_t Offset = OffsetV->getSExtValue();
462 // Get the base pointer input of the bitcast, and the type it points to.
463 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
464 const Type *GEPIdxTy =
465 cast<PointerType>(OrigBase->getType())->getElementType();
466 SmallVector<Value*, 8> NewIndices;
467 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
468 // If we were able to index down into an element, create the GEP
469 // and bitcast the result. This eliminates one bitcast, potentially
471 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
472 Builder->CreateInBoundsGEP(OrigBase,
473 NewIndices.begin(), NewIndices.end()) :
474 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
477 if (isa<BitCastInst>(CI))
478 return new BitCastInst(NGEP, CI.getType());
479 assert(isa<PtrToIntInst>(CI));
480 return new PtrToIntInst(NGEP, CI.getType());
486 return commonCastTransforms(CI);
489 /// commonIntCastTransforms - This function implements the common transforms
490 /// for trunc, zext, and sext.
491 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
492 if (Instruction *Result = commonCastTransforms(CI))
495 Value *Src = CI.getOperand(0);
496 const Type *SrcTy = Src->getType();
497 const Type *DestTy = CI.getType();
498 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
499 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
501 // See if we can simplify any instructions used by the LHS whose sole
502 // purpose is to compute bits we don't care about.
503 if (SimplifyDemandedInstructionBits(CI))
506 // If the source isn't an instruction or has more than one use then we
507 // can't do anything more.
508 Instruction *SrcI = dyn_cast<Instruction>(Src);
509 if (!SrcI || !Src->hasOneUse())
512 // Attempt to propagate the cast into the instruction for int->int casts.
513 int NumCastsRemoved = 0;
514 // Only do this if the dest type is a simple type, don't convert the
515 // expression tree to something weird like i93 unless the source is also
517 if ((isa<VectorType>(DestTy) ||
518 ShouldChangeType(SrcI->getType(), DestTy)) &&
519 CanEvaluateInDifferentType(SrcI, DestTy,
520 CI.getOpcode(), NumCastsRemoved)) {
521 // If this cast is a truncate, evaluting in a different type always
522 // eliminates the cast, so it is always a win. If this is a zero-extension,
523 // we need to do an AND to maintain the clear top-part of the computation,
524 // so we require that the input have eliminated at least one cast. If this
525 // is a sign extension, we insert two new casts (to do the extension) so we
526 // require that two casts have been eliminated.
527 bool DoXForm = false;
528 bool JustReplace = false;
529 switch (CI.getOpcode()) {
531 // All the others use floating point so we shouldn't actually
532 // get here because of the check above.
533 llvm_unreachable("Unknown cast type");
534 case Instruction::Trunc:
537 case Instruction::ZExt: {
538 DoXForm = NumCastsRemoved >= 1;
541 // If it's unnecessary to issue an AND to clear the high bits, it's
542 // always profitable to do this xform.
543 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
544 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
545 if (MaskedValueIsZero(TryRes, Mask))
546 return ReplaceInstUsesWith(CI, TryRes);
548 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
549 if (TryI->use_empty())
550 EraseInstFromFunction(*TryI);
554 case Instruction::SExt: {
555 DoXForm = NumCastsRemoved >= 2;
556 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
557 // If we do not have to emit the truncate + sext pair, then it's always
558 // profitable to do this xform.
560 // It's not safe to eliminate the trunc + sext pair if one of the
561 // eliminated cast is a truncate. e.g.
562 // t2 = trunc i32 t1 to i16
563 // t3 = sext i16 t2 to i32
566 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
567 unsigned NumSignBits = ComputeNumSignBits(TryRes);
568 if (NumSignBits > (DestBitSize - SrcBitSize))
569 return ReplaceInstUsesWith(CI, TryRes);
571 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
572 if (TryI->use_empty())
573 EraseInstFromFunction(*TryI);
580 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
581 " to avoid cast: " << CI);
582 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
583 CI.getOpcode() == Instruction::SExt);
585 // Just replace this cast with the result.
586 return ReplaceInstUsesWith(CI, Res);
588 assert(Res->getType() == DestTy);
589 switch (CI.getOpcode()) {
590 default: llvm_unreachable("Unknown cast type!");
591 case Instruction::Trunc:
592 // Just replace this cast with the result.
593 return ReplaceInstUsesWith(CI, Res);
594 case Instruction::ZExt: {
595 assert(SrcBitSize < DestBitSize && "Not a zext?");
597 // If the high bits are already zero, just replace this cast with the
599 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
600 if (MaskedValueIsZero(Res, Mask))
601 return ReplaceInstUsesWith(CI, Res);
603 // We need to emit an AND to clear the high bits.
604 Constant *C = ConstantInt::get(CI.getContext(),
605 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
606 return BinaryOperator::CreateAnd(Res, C);
608 case Instruction::SExt: {
609 // If the high bits are already filled with sign bit, just replace this
610 // cast with the result.
611 unsigned NumSignBits = ComputeNumSignBits(Res);
612 if (NumSignBits > (DestBitSize - SrcBitSize))
613 return ReplaceInstUsesWith(CI, Res);
615 // We need to emit a cast to truncate, then a cast to sext.
616 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
622 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
623 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
625 switch (SrcI->getOpcode()) {
626 case Instruction::Add:
627 case Instruction::Mul:
628 case Instruction::And:
629 case Instruction::Or:
630 case Instruction::Xor:
631 // If we are discarding information, rewrite.
632 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
633 // Don't insert two casts unless at least one can be eliminated.
634 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy) ||
635 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy)) {
636 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
637 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
638 return BinaryOperator::Create(
639 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
643 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
644 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
645 SrcI->getOpcode() == Instruction::Xor &&
646 Op1 == ConstantInt::getTrue(CI.getContext()) &&
647 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
648 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
649 return BinaryOperator::CreateXor(New,
650 ConstantInt::get(CI.getType(), 1));
654 case Instruction::Shl: {
655 // Canonicalize trunc inside shl, if we can.
656 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
657 if (CI && DestBitSize < SrcBitSize &&
658 CI->getLimitedValue(DestBitSize) < DestBitSize) {
659 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
660 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
661 return BinaryOperator::CreateShl(Op0c, Op1c);
670 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
671 if (Instruction *Result = commonIntCastTransforms(CI))
674 Value *Src = CI.getOperand(0);
675 const Type *Ty = CI.getType();
676 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
677 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
679 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
680 if (DestBitWidth == 1) {
681 Constant *One = ConstantInt::get(Src->getType(), 1);
682 Src = Builder->CreateAnd(Src, One, "tmp");
683 Value *Zero = Constant::getNullValue(Src->getType());
684 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
687 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
688 ConstantInt *ShAmtV = 0;
690 if (Src->hasOneUse() &&
691 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
692 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
694 // Get a mask for the bits shifting in.
695 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
696 if (MaskedValueIsZero(ShiftOp, Mask)) {
697 if (ShAmt >= DestBitWidth) // All zeros.
698 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
700 // Okay, we can shrink this. Truncate the input, then return a new
702 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
703 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
704 return BinaryOperator::CreateLShr(V1, V2);
711 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
712 /// in order to eliminate the icmp.
713 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
715 // If we are just checking for a icmp eq of a single bit and zext'ing it
716 // to an integer, then shift the bit to the appropriate place and then
717 // cast to integer to avoid the comparison.
718 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
719 const APInt &Op1CV = Op1C->getValue();
721 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
722 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
723 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
724 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
725 if (!DoXform) return ICI;
727 Value *In = ICI->getOperand(0);
728 Value *Sh = ConstantInt::get(In->getType(),
729 In->getType()->getScalarSizeInBits()-1);
730 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
731 if (In->getType() != CI.getType())
732 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
734 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
735 Constant *One = ConstantInt::get(In->getType(), 1);
736 In = Builder->CreateXor(In, One, In->getName()+".not");
739 return ReplaceInstUsesWith(CI, In);
744 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
745 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
746 // zext (X == 1) to i32 --> X iff X has only the low bit set.
747 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
748 // zext (X != 0) to i32 --> X iff X has only the low bit set.
749 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
750 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
751 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
752 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
753 // This only works for EQ and NE
755 // If Op1C some other power of two, convert:
756 uint32_t BitWidth = Op1C->getType()->getBitWidth();
757 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
758 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
759 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
761 APInt KnownZeroMask(~KnownZero);
762 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
763 if (!DoXform) return ICI;
765 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
766 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
767 // (X&4) == 2 --> false
768 // (X&4) != 2 --> true
769 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
771 Res = ConstantExpr::getZExt(Res, CI.getType());
772 return ReplaceInstUsesWith(CI, Res);
775 uint32_t ShiftAmt = KnownZeroMask.logBase2();
776 Value *In = ICI->getOperand(0);
778 // Perform a logical shr by shiftamt.
779 // Insert the shift to put the result in the low bit.
780 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
781 In->getName()+".lobit");
784 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
785 Constant *One = ConstantInt::get(In->getType(), 1);
786 In = Builder->CreateXor(In, One, "tmp");
789 if (CI.getType() == In->getType())
790 return ReplaceInstUsesWith(CI, In);
792 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
797 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
798 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
799 // may lead to additional simplifications.
800 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
801 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
802 uint32_t BitWidth = ITy->getBitWidth();
803 Value *LHS = ICI->getOperand(0);
804 Value *RHS = ICI->getOperand(1);
806 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
807 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
808 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
809 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
810 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
812 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
813 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
814 APInt UnknownBit = ~KnownBits;
815 if (UnknownBit.countPopulation() == 1) {
816 if (!DoXform) return ICI;
818 Value *Result = Builder->CreateXor(LHS, RHS);
820 // Mask off any bits that are set and won't be shifted away.
821 if (KnownOneLHS.uge(UnknownBit))
822 Result = Builder->CreateAnd(Result,
823 ConstantInt::get(ITy, UnknownBit));
825 // Shift the bit we're testing down to the lsb.
826 Result = Builder->CreateLShr(
827 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
829 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
830 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
831 Result->takeName(ICI);
832 return ReplaceInstUsesWith(CI, Result);
841 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
842 // If one of the common conversion will work, do it.
843 if (Instruction *Result = commonIntCastTransforms(CI))
846 Value *Src = CI.getOperand(0);
848 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
849 // types and if the sizes are just right we can convert this into a logical
850 // 'and' which will be much cheaper than the pair of casts.
851 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
852 // Get the sizes of the types involved. We know that the intermediate type
853 // will be smaller than A or C, but don't know the relation between A and C.
854 Value *A = CSrc->getOperand(0);
855 unsigned SrcSize = A->getType()->getScalarSizeInBits();
856 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
857 unsigned DstSize = CI.getType()->getScalarSizeInBits();
858 // If we're actually extending zero bits, then if
859 // SrcSize < DstSize: zext(a & mask)
860 // SrcSize == DstSize: a & mask
861 // SrcSize > DstSize: trunc(a) & mask
862 if (SrcSize < DstSize) {
863 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
864 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
865 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
866 return new ZExtInst(And, CI.getType());
869 if (SrcSize == DstSize) {
870 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
871 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
874 if (SrcSize > DstSize) {
875 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
876 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
877 return BinaryOperator::CreateAnd(Trunc,
878 ConstantInt::get(Trunc->getType(),
883 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
884 return transformZExtICmp(ICI, CI);
886 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
887 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
888 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
889 // of the (zext icmp) will be transformed.
890 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
891 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
892 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
893 (transformZExtICmp(LHS, CI, false) ||
894 transformZExtICmp(RHS, CI, false))) {
895 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
896 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
897 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
901 // zext(trunc(t) & C) -> (t & zext(C)).
902 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
903 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
904 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
905 Value *TI0 = TI->getOperand(0);
906 if (TI0->getType() == CI.getType())
908 BinaryOperator::CreateAnd(TI0,
909 ConstantExpr::getZExt(C, CI.getType()));
912 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
913 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
914 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
915 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
916 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
917 And->getOperand(1) == C)
918 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
919 Value *TI0 = TI->getOperand(0);
920 if (TI0->getType() == CI.getType()) {
921 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
922 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
923 return BinaryOperator::CreateXor(NewAnd, ZC);
930 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
931 if (Instruction *I = commonIntCastTransforms(CI))
934 Value *Src = CI.getOperand(0);
936 // Canonicalize sign-extend from i1 to a select.
937 if (Src->getType()->isInteger(1))
938 return SelectInst::Create(Src,
939 Constant::getAllOnesValue(CI.getType()),
940 Constant::getNullValue(CI.getType()));
942 // See if the value being truncated is already sign extended. If so, just
943 // eliminate the trunc/sext pair.
944 if (Operator::getOpcode(Src) == Instruction::Trunc) {
945 Value *Op = cast<User>(Src)->getOperand(0);
946 unsigned OpBits = Op->getType()->getScalarSizeInBits();
947 unsigned MidBits = Src->getType()->getScalarSizeInBits();
948 unsigned DestBits = CI.getType()->getScalarSizeInBits();
949 unsigned NumSignBits = ComputeNumSignBits(Op);
951 if (OpBits == DestBits) {
952 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
953 // bits, it is already ready.
954 if (NumSignBits > DestBits-MidBits)
955 return ReplaceInstUsesWith(CI, Op);
956 } else if (OpBits < DestBits) {
957 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
958 // bits, just sext from i32.
959 if (NumSignBits > OpBits-MidBits)
960 return new SExtInst(Op, CI.getType(), "tmp");
962 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
963 // bits, just truncate to i32.
964 if (NumSignBits > OpBits-MidBits)
965 return new TruncInst(Op, CI.getType(), "tmp");
969 // If the input is a shl/ashr pair of a same constant, then this is a sign
970 // extension from a smaller value. If we could trust arbitrary bitwidth
971 // integers, we could turn this into a truncate to the smaller bit and then
972 // use a sext for the whole extension. Since we don't, look deeper and check
973 // for a truncate. If the source and dest are the same type, eliminate the
974 // trunc and extend and just do shifts. For example, turn:
975 // %a = trunc i32 %i to i8
977 // %c = ashr i8 %b, 6
978 // %d = sext i8 %c to i32
980 // %a = shl i32 %i, 30
981 // %d = ashr i32 %a, 30
983 ConstantInt *BA = 0, *CA = 0;
984 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
985 m_ConstantInt(CA))) &&
986 BA == CA && isa<TruncInst>(A)) {
987 Value *I = cast<TruncInst>(A)->getOperand(0);
988 if (I->getType() == CI.getType()) {
989 unsigned MidSize = Src->getType()->getScalarSizeInBits();
990 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
991 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
992 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
993 I = Builder->CreateShl(I, ShAmtV, CI.getName());
994 return BinaryOperator::CreateAShr(I, ShAmtV);
1002 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1003 /// in the specified FP type without changing its value.
1004 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1006 APFloat F = CFP->getValueAPF();
1007 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1009 return ConstantFP::get(CFP->getContext(), F);
1013 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1014 /// through it until we get the source value.
1015 static Value *LookThroughFPExtensions(Value *V) {
1016 if (Instruction *I = dyn_cast<Instruction>(V))
1017 if (I->getOpcode() == Instruction::FPExt)
1018 return LookThroughFPExtensions(I->getOperand(0));
1020 // If this value is a constant, return the constant in the smallest FP type
1021 // that can accurately represent it. This allows us to turn
1022 // (float)((double)X+2.0) into x+2.0f.
1023 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1024 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1025 return V; // No constant folding of this.
1026 // See if the value can be truncated to float and then reextended.
1027 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1029 if (CFP->getType()->isDoubleTy())
1030 return V; // Won't shrink.
1031 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1033 // Don't try to shrink to various long double types.
1039 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1040 if (Instruction *I = commonCastTransforms(CI))
1043 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1044 // smaller than the destination type, we can eliminate the truncate by doing
1045 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1046 // as many builtins (sqrt, etc).
1047 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1048 if (OpI && OpI->hasOneUse()) {
1049 switch (OpI->getOpcode()) {
1051 case Instruction::FAdd:
1052 case Instruction::FSub:
1053 case Instruction::FMul:
1054 case Instruction::FDiv:
1055 case Instruction::FRem:
1056 const Type *SrcTy = OpI->getType();
1057 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1058 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1059 if (LHSTrunc->getType() != SrcTy &&
1060 RHSTrunc->getType() != SrcTy) {
1061 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1062 // If the source types were both smaller than the destination type of
1063 // the cast, do this xform.
1064 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1065 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1066 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1067 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1068 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1077 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1078 return commonCastTransforms(CI);
1081 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1082 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1084 return commonCastTransforms(FI);
1086 // fptoui(uitofp(X)) --> X
1087 // fptoui(sitofp(X)) --> X
1088 // This is safe if the intermediate type has enough bits in its mantissa to
1089 // accurately represent all values of X. For example, do not do this with
1090 // i64->float->i64. This is also safe for sitofp case, because any negative
1091 // 'X' value would cause an undefined result for the fptoui.
1092 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1093 OpI->getOperand(0)->getType() == FI.getType() &&
1094 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1095 OpI->getType()->getFPMantissaWidth())
1096 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1098 return commonCastTransforms(FI);
1101 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1102 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1104 return commonCastTransforms(FI);
1106 // fptosi(sitofp(X)) --> X
1107 // fptosi(uitofp(X)) --> X
1108 // This is safe if the intermediate type has enough bits in its mantissa to
1109 // accurately represent all values of X. For example, do not do this with
1110 // i64->float->i64. This is also safe for sitofp case, because any negative
1111 // 'X' value would cause an undefined result for the fptoui.
1112 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1113 OpI->getOperand(0)->getType() == FI.getType() &&
1114 (int)FI.getType()->getScalarSizeInBits() <=
1115 OpI->getType()->getFPMantissaWidth())
1116 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1118 return commonCastTransforms(FI);
1121 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1122 return commonCastTransforms(CI);
1125 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1126 return commonCastTransforms(CI);
1129 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1130 // If the destination integer type is smaller than the intptr_t type for
1131 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
1132 // trunc to be exposed to other transforms. Don't do this for extending
1133 // ptrtoint's, because we don't know if the target sign or zero extends its
1136 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1137 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1138 TD->getIntPtrType(CI.getContext()),
1140 return new TruncInst(P, CI.getType());
1143 return commonPointerCastTransforms(CI);
1147 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1148 // If the source integer type is larger than the intptr_t type for
1149 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
1150 // allows the trunc to be exposed to other transforms. Don't do this for
1151 // extending inttoptr's, because we don't know if the target sign or zero
1152 // extends to pointers.
1153 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
1154 TD->getPointerSizeInBits()) {
1155 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1156 TD->getIntPtrType(CI.getContext()), "tmp");
1157 return new IntToPtrInst(P, CI.getType());
1160 if (Instruction *I = commonCastTransforms(CI))
1166 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1167 // If the operands are integer typed then apply the integer transforms,
1168 // otherwise just apply the common ones.
1169 Value *Src = CI.getOperand(0);
1170 const Type *SrcTy = Src->getType();
1171 const Type *DestTy = CI.getType();
1173 if (isa<PointerType>(SrcTy)) {
1174 if (Instruction *I = commonPointerCastTransforms(CI))
1177 if (Instruction *Result = commonCastTransforms(CI))
1182 // Get rid of casts from one type to the same type. These are useless and can
1183 // be replaced by the operand.
1184 if (DestTy == Src->getType())
1185 return ReplaceInstUsesWith(CI, Src);
1187 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1188 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1189 const Type *DstElTy = DstPTy->getElementType();
1190 const Type *SrcElTy = SrcPTy->getElementType();
1192 // If the address spaces don't match, don't eliminate the bitcast, which is
1193 // required for changing types.
1194 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1197 // If we are casting a alloca to a pointer to a type of the same
1198 // size, rewrite the allocation instruction to allocate the "right" type.
1199 // There is no need to modify malloc calls because it is their bitcast that
1200 // needs to be cleaned up.
1201 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1202 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1205 // If the source and destination are pointers, and this cast is equivalent
1206 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1207 // This can enhance SROA and other transforms that want type-safe pointers.
1208 Constant *ZeroUInt =
1209 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1210 unsigned NumZeros = 0;
1211 while (SrcElTy != DstElTy &&
1212 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
1213 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1214 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1218 // If we found a path from the src to dest, create the getelementptr now.
1219 if (SrcElTy == DstElTy) {
1220 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1221 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1222 ((Instruction*) NULL));
1226 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1227 if (DestVTy->getNumElements() == 1) {
1228 if (!isa<VectorType>(SrcTy)) {
1229 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1230 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1231 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1233 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1237 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1238 if (SrcVTy->getNumElements() == 1) {
1239 if (!isa<VectorType>(DestTy)) {
1241 Builder->CreateExtractElement(Src,
1242 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1243 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1248 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1249 if (SVI->hasOneUse()) {
1250 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
1251 // a bitconvert to a vector with the same # elts.
1252 if (isa<VectorType>(DestTy) &&
1253 cast<VectorType>(DestTy)->getNumElements() ==
1254 SVI->getType()->getNumElements() &&
1255 SVI->getType()->getNumElements() ==
1256 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1258 // If either of the operands is a cast from CI.getType(), then
1259 // evaluating the shuffle in the casted destination's type will allow
1260 // us to eliminate at least one cast.
1261 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
1262 Tmp->getOperand(0)->getType() == DestTy) ||
1263 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
1264 Tmp->getOperand(0)->getType() == DestTy)) {
1265 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1266 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1267 // Return a new shuffle vector. Use the same element ID's, as we
1268 // know the vector types match #elts.
1269 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));