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);
33 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
34 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
35 if (I->getOpcode() == Instruction::Shl) {
36 // This is a value scaled by '1 << the shift amt'.
37 Scale = 1U << RHS->getZExtValue();
39 return I->getOperand(0);
42 if (I->getOpcode() == Instruction::Mul) {
43 // This value is scaled by 'RHS'.
44 Scale = RHS->getZExtValue();
46 return I->getOperand(0);
49 if (I->getOpcode() == Instruction::Add) {
50 // We have X+C. Check to see if we really have (X*C2)+C1,
51 // where C1 is divisible by C2.
54 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
55 Offset += RHS->getZExtValue();
62 // Otherwise, we can't look past this.
68 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
69 /// try to eliminate the cast by moving the type information into the alloc.
70 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
72 // This requires TargetData to get the alloca alignment and size information.
75 const PointerType *PTy = cast<PointerType>(CI.getType());
77 BuilderTy AllocaBuilder(*Builder);
78 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
80 // Get the type really allocated and the type casted to.
81 const Type *AllocElTy = AI.getAllocatedType();
82 const Type *CastElTy = PTy->getElementType();
83 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
85 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
86 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
87 if (CastElTyAlign < AllocElTyAlign) return 0;
89 // If the allocation has multiple uses, only promote it if we are strictly
90 // increasing the alignment of the resultant allocation. If we keep it the
91 // same, we open the door to infinite loops of various kinds. (A reference
92 // from a dbg.declare doesn't count as a use for this purpose.)
93 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
94 CastElTyAlign == AllocElTyAlign) return 0;
96 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
97 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
98 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
100 // See if we can satisfy the modulus by pulling a scale out of the array
102 unsigned ArraySizeScale;
104 Value *NumElements = // See if the array size is a decomposable linear expr.
105 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
107 // If we can now satisfy the modulus, by using a non-1 scale, we really can
109 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
110 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
112 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
117 Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale);
118 // Insert before the alloca, not before the cast.
119 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
122 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
123 Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()),
125 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
128 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
129 New->setAlignment(AI.getAlignment());
132 // If the allocation has one real use plus a dbg.declare, just remove the
134 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
135 EraseInstFromFunction(*(Instruction*)DI);
137 // If the allocation has multiple real uses, insert a cast and change all
138 // things that used it to use the new cast. This will also hack on CI, but it
140 else if (!AI.hasOneUse()) {
141 // New is the allocation instruction, pointer typed. AI is the original
142 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
143 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
144 AI.replaceAllUsesWith(NewCast);
146 return ReplaceInstUsesWith(CI, New);
150 /// CanEvaluateInDifferentType - Return true if we can take the specified value
151 /// and return it as type Ty without inserting any new casts and without
152 /// changing the computed value. This is used by code that tries to decide
153 /// whether promoting or shrinking integer operations to wider or smaller types
154 /// will allow us to eliminate a truncate or extend.
156 /// This is a truncation operation if Ty is smaller than V->getType(), or an
157 /// extension operation if Ty is larger.
159 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
160 /// should return true if trunc(V) can be computed by computing V in the smaller
161 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
162 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
163 /// efficiently truncated.
165 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
166 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
167 /// the final result.
168 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
170 int &NumCastsRemoved){
171 // We can always evaluate constants in another type.
172 if (isa<Constant>(V))
175 Instruction *I = dyn_cast<Instruction>(V);
176 if (!I) return false;
178 const Type *OrigTy = V->getType();
180 // If this is an extension or truncate, we can often eliminate it.
181 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
182 // If this is a cast from the destination type, we can trivially eliminate
183 // it, and this will remove a cast overall.
184 if (I->getOperand(0)->getType() == Ty) {
185 // If the first operand is itself a cast, and is eliminable, do not count
186 // this as an eliminable cast. We would prefer to eliminate those two
188 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
194 // We can't extend or shrink something that has multiple uses: doing so would
195 // require duplicating the instruction in general, which isn't profitable.
196 if (!I->hasOneUse()) return false;
198 unsigned Opc = I->getOpcode();
200 case Instruction::Add:
201 case Instruction::Sub:
202 case Instruction::Mul:
203 case Instruction::And:
204 case Instruction::Or:
205 case Instruction::Xor:
206 // These operators can all arbitrarily be extended or truncated.
207 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
209 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
212 case Instruction::UDiv:
213 case Instruction::URem: {
214 // UDiv and URem can be truncated if all the truncated bits are zero.
215 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
216 uint32_t BitWidth = Ty->getScalarSizeInBits();
217 if (BitWidth < OrigBitWidth) {
218 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
219 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
220 MaskedValueIsZero(I->getOperand(1), Mask)) {
221 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
223 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
229 case Instruction::Shl:
230 // If we are truncating the result of this SHL, and if it's a shift of a
231 // constant amount, we can always perform a SHL in a smaller type.
232 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
233 uint32_t BitWidth = Ty->getScalarSizeInBits();
234 if (BitWidth < OrigTy->getScalarSizeInBits() &&
235 CI->getLimitedValue(BitWidth) < BitWidth)
236 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
240 case Instruction::LShr:
241 // If this is a truncate of a logical shr, we can truncate it to a smaller
242 // lshr iff we know that the bits we would otherwise be shifting in are
244 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
245 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
246 uint32_t BitWidth = Ty->getScalarSizeInBits();
247 if (BitWidth < OrigBitWidth &&
248 MaskedValueIsZero(I->getOperand(0),
249 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
250 CI->getLimitedValue(BitWidth) < BitWidth) {
251 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
256 case Instruction::ZExt:
257 case Instruction::SExt:
258 case Instruction::Trunc:
259 // If this is the same kind of case as our original (e.g. zext+zext), we
260 // can safely replace it. Note that replacing it does not reduce the number
261 // of casts in the input.
265 // sext (zext ty1), ty2 -> zext ty2
266 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
269 case Instruction::Select: {
270 SelectInst *SI = cast<SelectInst>(I);
271 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
273 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
276 case Instruction::PHI: {
277 // We can change a phi if we can change all operands.
278 PHINode *PN = cast<PHINode>(I);
279 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
280 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
286 // TODO: Can handle more cases here.
293 /// EvaluateInDifferentType - Given an expression that
294 /// CanEvaluateInDifferentType returns true for, actually insert the code to
295 /// evaluate the expression.
296 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
298 if (Constant *C = dyn_cast<Constant>(V))
299 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
301 // Otherwise, it must be an instruction.
302 Instruction *I = cast<Instruction>(V);
303 Instruction *Res = 0;
304 unsigned Opc = I->getOpcode();
306 case Instruction::Add:
307 case Instruction::Sub:
308 case Instruction::Mul:
309 case Instruction::And:
310 case Instruction::Or:
311 case Instruction::Xor:
312 case Instruction::AShr:
313 case Instruction::LShr:
314 case Instruction::Shl:
315 case Instruction::UDiv:
316 case Instruction::URem: {
317 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
318 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
319 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
322 case Instruction::Trunc:
323 case Instruction::ZExt:
324 case Instruction::SExt:
325 // If the source type of the cast is the type we're trying for then we can
326 // just return the source. There's no need to insert it because it is not
328 if (I->getOperand(0)->getType() == Ty)
329 return I->getOperand(0);
331 // Otherwise, must be the same type of cast, so just reinsert a new one.
332 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
334 case Instruction::Select: {
335 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
336 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
337 Res = SelectInst::Create(I->getOperand(0), True, False);
340 case Instruction::PHI: {
341 PHINode *OPN = cast<PHINode>(I);
342 PHINode *NPN = PHINode::Create(Ty);
343 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
344 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
345 NPN->addIncoming(V, OPN->getIncomingBlock(i));
351 // TODO: Can handle more cases here.
352 llvm_unreachable("Unreachable!");
357 return InsertNewInstBefore(Res, *I);
361 /// This function is a wrapper around CastInst::isEliminableCastPair. It
362 /// simply extracts arguments and returns what that function returns.
363 static Instruction::CastOps
364 isEliminableCastPair(
365 const CastInst *CI, ///< The first cast instruction
366 unsigned opcode, ///< The opcode of the second cast instruction
367 const Type *DstTy, ///< The target type for the second cast instruction
368 TargetData *TD ///< The target data for pointer size
371 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
372 const Type *MidTy = CI->getType(); // B from above
374 // Get the opcodes of the two Cast instructions
375 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
376 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
378 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
380 TD ? TD->getIntPtrType(CI->getContext()) : 0);
382 // We don't want to form an inttoptr or ptrtoint that converts to an integer
383 // type that differs from the pointer size.
384 if ((Res == Instruction::IntToPtr &&
385 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
386 (Res == Instruction::PtrToInt &&
387 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
390 return Instruction::CastOps(Res);
393 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
394 /// in any code being generated. It does not require codegen if V is simple
395 /// enough or if the cast can be folded into other casts.
396 bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V,
398 if (V->getType() == Ty || isa<Constant>(V)) return false;
400 // If this is another cast that can be eliminated, it isn't codegen either.
401 if (const CastInst *CI = dyn_cast<CastInst>(V))
402 if (isEliminableCastPair(CI, opcode, Ty, TD))
408 /// @brief Implement the transforms common to all CastInst visitors.
409 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
410 Value *Src = CI.getOperand(0);
412 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
414 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
415 if (Instruction::CastOps opc =
416 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
417 // The first cast (CSrc) is eliminable so we need to fix up or replace
418 // the second cast (CI). CSrc will then have a good chance of being dead.
419 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
423 // If we are casting a select then fold the cast into the select
424 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
425 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
428 // If we are casting a PHI then fold the cast into the PHI
429 if (isa<PHINode>(Src)) {
430 // We don't do this if this would create a PHI node with an illegal type if
431 // it is currently legal.
432 if (!isa<IntegerType>(Src->getType()) ||
433 !isa<IntegerType>(CI.getType()) ||
434 ShouldChangeType(CI.getType(), Src->getType()))
435 if (Instruction *NV = FoldOpIntoPhi(CI))
442 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
443 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
444 Value *Src = CI.getOperand(0);
446 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
447 // If casting the result of a getelementptr instruction with no offset, turn
448 // this into a cast of the original pointer!
449 if (GEP->hasAllZeroIndices()) {
450 // Changing the cast operand is usually not a good idea but it is safe
451 // here because the pointer operand is being replaced with another
452 // pointer operand so the opcode doesn't need to change.
454 CI.setOperand(0, GEP->getOperand(0));
458 // If the GEP has a single use, and the base pointer is a bitcast, and the
459 // GEP computes a constant offset, see if we can convert these three
460 // instructions into fewer. This typically happens with unions and other
461 // non-type-safe code.
462 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
463 if (GEP->hasAllConstantIndices()) {
464 // We are guaranteed to get a constant from EmitGEPOffset.
465 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
466 int64_t Offset = OffsetV->getSExtValue();
468 // Get the base pointer input of the bitcast, and the type it points to.
469 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
470 const Type *GEPIdxTy =
471 cast<PointerType>(OrigBase->getType())->getElementType();
472 SmallVector<Value*, 8> NewIndices;
473 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
474 // If we were able to index down into an element, create the GEP
475 // and bitcast the result. This eliminates one bitcast, potentially
477 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
478 Builder->CreateInBoundsGEP(OrigBase,
479 NewIndices.begin(), NewIndices.end()) :
480 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
483 if (isa<BitCastInst>(CI))
484 return new BitCastInst(NGEP, CI.getType());
485 assert(isa<PtrToIntInst>(CI));
486 return new PtrToIntInst(NGEP, CI.getType());
492 return commonCastTransforms(CI);
495 /// commonIntCastTransforms - This function implements the common transforms
496 /// for trunc, zext, and sext.
497 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
498 if (Instruction *Result = commonCastTransforms(CI))
501 Value *Src = CI.getOperand(0);
502 const Type *SrcTy = Src->getType();
503 const Type *DestTy = CI.getType();
504 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
505 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
507 // See if we can simplify any instructions used by the LHS whose sole
508 // purpose is to compute bits we don't care about.
509 if (SimplifyDemandedInstructionBits(CI))
512 // If the source isn't an instruction or has more than one use then we
513 // can't do anything more.
514 Instruction *SrcI = dyn_cast<Instruction>(Src);
515 if (!SrcI || !Src->hasOneUse())
518 // Attempt to propagate the cast into the instruction for int->int casts.
519 int NumCastsRemoved = 0;
520 // Only do this if the dest type is a simple type, don't convert the
521 // expression tree to something weird like i93 unless the source is also
523 if ((isa<VectorType>(DestTy) ||
524 ShouldChangeType(SrcI->getType(), DestTy)) &&
525 CanEvaluateInDifferentType(SrcI, DestTy,
526 CI.getOpcode(), NumCastsRemoved)) {
527 // If this cast is a truncate, evaluting in a different type always
528 // eliminates the cast, so it is always a win. If this is a zero-extension,
529 // we need to do an AND to maintain the clear top-part of the computation,
530 // so we require that the input have eliminated at least one cast. If this
531 // is a sign extension, we insert two new casts (to do the extension) so we
532 // require that two casts have been eliminated.
533 bool DoXForm = false;
534 bool JustReplace = false;
535 switch (CI.getOpcode()) {
537 // All the others use floating point so we shouldn't actually
538 // get here because of the check above.
539 llvm_unreachable("Unknown cast type");
540 case Instruction::Trunc:
543 case Instruction::ZExt: {
544 DoXForm = NumCastsRemoved >= 1;
547 // If it's unnecessary to issue an AND to clear the high bits, it's
548 // always profitable to do this xform.
549 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
550 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
551 if (MaskedValueIsZero(TryRes, Mask))
552 return ReplaceInstUsesWith(CI, TryRes);
554 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
555 if (TryI->use_empty())
556 EraseInstFromFunction(*TryI);
560 case Instruction::SExt: {
561 DoXForm = NumCastsRemoved >= 2;
562 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
563 // If we do not have to emit the truncate + sext pair, then it's always
564 // profitable to do this xform.
566 // It's not safe to eliminate the trunc + sext pair if one of the
567 // eliminated cast is a truncate. e.g.
568 // t2 = trunc i32 t1 to i16
569 // t3 = sext i16 t2 to i32
572 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
573 unsigned NumSignBits = ComputeNumSignBits(TryRes);
574 if (NumSignBits > (DestBitSize - SrcBitSize))
575 return ReplaceInstUsesWith(CI, TryRes);
577 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
578 if (TryI->use_empty())
579 EraseInstFromFunction(*TryI);
586 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
587 " to avoid cast: " << CI);
588 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
589 CI.getOpcode() == Instruction::SExt);
591 // Just replace this cast with the result.
592 return ReplaceInstUsesWith(CI, Res);
594 assert(Res->getType() == DestTy);
595 switch (CI.getOpcode()) {
596 default: llvm_unreachable("Unknown cast type!");
597 case Instruction::Trunc:
598 // Just replace this cast with the result.
599 return ReplaceInstUsesWith(CI, Res);
600 case Instruction::ZExt: {
601 assert(SrcBitSize < DestBitSize && "Not a zext?");
603 // If the high bits are already zero, just replace this cast with the
605 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
606 if (MaskedValueIsZero(Res, Mask))
607 return ReplaceInstUsesWith(CI, Res);
609 // We need to emit an AND to clear the high bits.
610 Constant *C = ConstantInt::get(CI.getContext(),
611 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
612 return BinaryOperator::CreateAnd(Res, C);
614 case Instruction::SExt: {
615 // If the high bits are already filled with sign bit, just replace this
616 // cast with the result.
617 unsigned NumSignBits = ComputeNumSignBits(Res);
618 if (NumSignBits > (DestBitSize - SrcBitSize))
619 return ReplaceInstUsesWith(CI, Res);
621 // We need to emit a cast to truncate, then a cast to sext.
622 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
631 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
632 if (Instruction *Result = commonIntCastTransforms(CI))
635 Value *Src = CI.getOperand(0);
636 const Type *DestTy = CI.getType();
637 uint32_t DestBitWidth = DestTy->getScalarSizeInBits();
638 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
640 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
641 if (DestBitWidth == 1) {
642 Constant *One = ConstantInt::get(Src->getType(), 1);
643 Src = Builder->CreateAnd(Src, One, "tmp");
644 Value *Zero = Constant::getNullValue(Src->getType());
645 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
651 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
652 /// in order to eliminate the icmp.
653 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
655 // If we are just checking for a icmp eq of a single bit and zext'ing it
656 // to an integer, then shift the bit to the appropriate place and then
657 // cast to integer to avoid the comparison.
658 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
659 const APInt &Op1CV = Op1C->getValue();
661 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
662 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
663 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
664 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
665 if (!DoXform) return ICI;
667 Value *In = ICI->getOperand(0);
668 Value *Sh = ConstantInt::get(In->getType(),
669 In->getType()->getScalarSizeInBits()-1);
670 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
671 if (In->getType() != CI.getType())
672 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
674 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
675 Constant *One = ConstantInt::get(In->getType(), 1);
676 In = Builder->CreateXor(In, One, In->getName()+".not");
679 return ReplaceInstUsesWith(CI, In);
684 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
685 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
686 // zext (X == 1) to i32 --> X iff X has only the low bit set.
687 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
688 // zext (X != 0) to i32 --> X iff X has only the low bit set.
689 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
690 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
691 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
692 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
693 // This only works for EQ and NE
695 // If Op1C some other power of two, convert:
696 uint32_t BitWidth = Op1C->getType()->getBitWidth();
697 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
698 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
699 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
701 APInt KnownZeroMask(~KnownZero);
702 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
703 if (!DoXform) return ICI;
705 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
706 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
707 // (X&4) == 2 --> false
708 // (X&4) != 2 --> true
709 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
711 Res = ConstantExpr::getZExt(Res, CI.getType());
712 return ReplaceInstUsesWith(CI, Res);
715 uint32_t ShiftAmt = KnownZeroMask.logBase2();
716 Value *In = ICI->getOperand(0);
718 // Perform a logical shr by shiftamt.
719 // Insert the shift to put the result in the low bit.
720 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
721 In->getName()+".lobit");
724 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
725 Constant *One = ConstantInt::get(In->getType(), 1);
726 In = Builder->CreateXor(In, One, "tmp");
729 if (CI.getType() == In->getType())
730 return ReplaceInstUsesWith(CI, In);
732 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
737 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
738 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
739 // may lead to additional simplifications.
740 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
741 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
742 uint32_t BitWidth = ITy->getBitWidth();
743 Value *LHS = ICI->getOperand(0);
744 Value *RHS = ICI->getOperand(1);
746 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
747 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
748 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
749 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
750 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
752 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
753 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
754 APInt UnknownBit = ~KnownBits;
755 if (UnknownBit.countPopulation() == 1) {
756 if (!DoXform) return ICI;
758 Value *Result = Builder->CreateXor(LHS, RHS);
760 // Mask off any bits that are set and won't be shifted away.
761 if (KnownOneLHS.uge(UnknownBit))
762 Result = Builder->CreateAnd(Result,
763 ConstantInt::get(ITy, UnknownBit));
765 // Shift the bit we're testing down to the lsb.
766 Result = Builder->CreateLShr(
767 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
769 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
770 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
771 Result->takeName(ICI);
772 return ReplaceInstUsesWith(CI, Result);
781 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
782 // If one of the common conversion will work, do it.
783 if (Instruction *Result = commonIntCastTransforms(CI))
786 Value *Src = CI.getOperand(0);
788 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
789 // types and if the sizes are just right we can convert this into a logical
790 // 'and' which will be much cheaper than the pair of casts.
791 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
792 // Get the sizes of the types involved. We know that the intermediate type
793 // will be smaller than A or C, but don't know the relation between A and C.
794 Value *A = CSrc->getOperand(0);
795 unsigned SrcSize = A->getType()->getScalarSizeInBits();
796 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
797 unsigned DstSize = CI.getType()->getScalarSizeInBits();
798 // If we're actually extending zero bits, then if
799 // SrcSize < DstSize: zext(a & mask)
800 // SrcSize == DstSize: a & mask
801 // SrcSize > DstSize: trunc(a) & mask
802 if (SrcSize < DstSize) {
803 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
804 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
805 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
806 return new ZExtInst(And, CI.getType());
809 if (SrcSize == DstSize) {
810 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
811 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
814 if (SrcSize > DstSize) {
815 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
816 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
817 return BinaryOperator::CreateAnd(Trunc,
818 ConstantInt::get(Trunc->getType(),
823 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
824 return transformZExtICmp(ICI, CI);
826 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
827 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
828 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
829 // of the (zext icmp) will be transformed.
830 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
831 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
832 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
833 (transformZExtICmp(LHS, CI, false) ||
834 transformZExtICmp(RHS, CI, false))) {
835 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
836 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
837 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
841 // zext(trunc(t) & C) -> (t & zext(C)).
842 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
843 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
844 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
845 Value *TI0 = TI->getOperand(0);
846 if (TI0->getType() == CI.getType())
848 BinaryOperator::CreateAnd(TI0,
849 ConstantExpr::getZExt(C, CI.getType()));
852 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
853 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
854 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
855 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
856 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
857 And->getOperand(1) == C)
858 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
859 Value *TI0 = TI->getOperand(0);
860 if (TI0->getType() == CI.getType()) {
861 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
862 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
863 return BinaryOperator::CreateXor(NewAnd, ZC);
867 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
869 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) &&
870 match(SrcI, m_Not(m_Value(X))) &&
871 (!X->hasOneUse() || !isa<CmpInst>(X))) {
872 Value *New = Builder->CreateZExt(X, CI.getType());
873 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
879 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
880 if (Instruction *I = commonIntCastTransforms(CI))
883 Value *Src = CI.getOperand(0);
885 // Canonicalize sign-extend from i1 to a select.
886 if (Src->getType()->isInteger(1))
887 return SelectInst::Create(Src,
888 Constant::getAllOnesValue(CI.getType()),
889 Constant::getNullValue(CI.getType()));
891 // See if the value being truncated is already sign extended. If so, just
892 // eliminate the trunc/sext pair.
893 if (Operator::getOpcode(Src) == Instruction::Trunc) {
894 Value *Op = cast<User>(Src)->getOperand(0);
895 unsigned OpBits = Op->getType()->getScalarSizeInBits();
896 unsigned MidBits = Src->getType()->getScalarSizeInBits();
897 unsigned DestBits = CI.getType()->getScalarSizeInBits();
898 unsigned NumSignBits = ComputeNumSignBits(Op);
900 if (OpBits == DestBits) {
901 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
902 // bits, it is already ready.
903 if (NumSignBits > DestBits-MidBits)
904 return ReplaceInstUsesWith(CI, Op);
905 } else if (OpBits < DestBits) {
906 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
907 // bits, just sext from i32.
908 if (NumSignBits > OpBits-MidBits)
909 return new SExtInst(Op, CI.getType(), "tmp");
911 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
912 // bits, just truncate to i32.
913 if (NumSignBits > OpBits-MidBits)
914 return new TruncInst(Op, CI.getType(), "tmp");
918 // If the input is a shl/ashr pair of a same constant, then this is a sign
919 // extension from a smaller value. If we could trust arbitrary bitwidth
920 // integers, we could turn this into a truncate to the smaller bit and then
921 // use a sext for the whole extension. Since we don't, look deeper and check
922 // for a truncate. If the source and dest are the same type, eliminate the
923 // trunc and extend and just do shifts. For example, turn:
924 // %a = trunc i32 %i to i8
926 // %c = ashr i8 %b, 6
927 // %d = sext i8 %c to i32
929 // %a = shl i32 %i, 30
930 // %d = ashr i32 %a, 30
932 ConstantInt *BA = 0, *CA = 0;
933 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
934 m_ConstantInt(CA))) &&
935 BA == CA && isa<TruncInst>(A)) {
936 Value *I = cast<TruncInst>(A)->getOperand(0);
937 if (I->getType() == CI.getType()) {
938 unsigned MidSize = Src->getType()->getScalarSizeInBits();
939 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
940 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
941 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
942 I = Builder->CreateShl(I, ShAmtV, CI.getName());
943 return BinaryOperator::CreateAShr(I, ShAmtV);
951 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
952 /// in the specified FP type without changing its value.
953 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
955 APFloat F = CFP->getValueAPF();
956 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
958 return ConstantFP::get(CFP->getContext(), F);
962 /// LookThroughFPExtensions - If this is an fp extension instruction, look
963 /// through it until we get the source value.
964 static Value *LookThroughFPExtensions(Value *V) {
965 if (Instruction *I = dyn_cast<Instruction>(V))
966 if (I->getOpcode() == Instruction::FPExt)
967 return LookThroughFPExtensions(I->getOperand(0));
969 // If this value is a constant, return the constant in the smallest FP type
970 // that can accurately represent it. This allows us to turn
971 // (float)((double)X+2.0) into x+2.0f.
972 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
973 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
974 return V; // No constant folding of this.
975 // See if the value can be truncated to float and then reextended.
976 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
978 if (CFP->getType()->isDoubleTy())
979 return V; // Won't shrink.
980 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
982 // Don't try to shrink to various long double types.
988 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
989 if (Instruction *I = commonCastTransforms(CI))
992 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
993 // smaller than the destination type, we can eliminate the truncate by doing
994 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
995 // as many builtins (sqrt, etc).
996 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
997 if (OpI && OpI->hasOneUse()) {
998 switch (OpI->getOpcode()) {
1000 case Instruction::FAdd:
1001 case Instruction::FSub:
1002 case Instruction::FMul:
1003 case Instruction::FDiv:
1004 case Instruction::FRem:
1005 const Type *SrcTy = OpI->getType();
1006 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1007 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1008 if (LHSTrunc->getType() != SrcTy &&
1009 RHSTrunc->getType() != SrcTy) {
1010 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1011 // If the source types were both smaller than the destination type of
1012 // the cast, do this xform.
1013 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1014 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1015 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1016 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1017 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1026 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1027 return commonCastTransforms(CI);
1030 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1031 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1033 return commonCastTransforms(FI);
1035 // fptoui(uitofp(X)) --> X
1036 // fptoui(sitofp(X)) --> X
1037 // This is safe if the intermediate type has enough bits in its mantissa to
1038 // accurately represent all values of X. For example, do not do this with
1039 // i64->float->i64. This is also safe for sitofp case, because any negative
1040 // 'X' value would cause an undefined result for the fptoui.
1041 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1042 OpI->getOperand(0)->getType() == FI.getType() &&
1043 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1044 OpI->getType()->getFPMantissaWidth())
1045 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1047 return commonCastTransforms(FI);
1050 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1051 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1053 return commonCastTransforms(FI);
1055 // fptosi(sitofp(X)) --> X
1056 // fptosi(uitofp(X)) --> X
1057 // This is safe if the intermediate type has enough bits in its mantissa to
1058 // accurately represent all values of X. For example, do not do this with
1059 // i64->float->i64. This is also safe for sitofp case, because any negative
1060 // 'X' value would cause an undefined result for the fptoui.
1061 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1062 OpI->getOperand(0)->getType() == FI.getType() &&
1063 (int)FI.getType()->getScalarSizeInBits() <=
1064 OpI->getType()->getFPMantissaWidth())
1065 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1067 return commonCastTransforms(FI);
1070 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1071 return commonCastTransforms(CI);
1074 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1075 return commonCastTransforms(CI);
1078 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1079 // If the destination integer type is smaller than the intptr_t type for
1080 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
1081 // trunc to be exposed to other transforms. Don't do this for extending
1082 // ptrtoint's, because we don't know if the target sign or zero extends its
1085 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1086 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1087 TD->getIntPtrType(CI.getContext()),
1089 return new TruncInst(P, CI.getType());
1092 return commonPointerCastTransforms(CI);
1096 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1097 // If the source integer type is larger than the intptr_t type for
1098 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
1099 // allows the trunc to be exposed to other transforms. Don't do this for
1100 // extending inttoptr's, because we don't know if the target sign or zero
1101 // extends to pointers.
1102 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
1103 TD->getPointerSizeInBits()) {
1104 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1105 TD->getIntPtrType(CI.getContext()), "tmp");
1106 return new IntToPtrInst(P, CI.getType());
1109 if (Instruction *I = commonCastTransforms(CI))
1115 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1116 // If the operands are integer typed then apply the integer transforms,
1117 // otherwise just apply the common ones.
1118 Value *Src = CI.getOperand(0);
1119 const Type *SrcTy = Src->getType();
1120 const Type *DestTy = CI.getType();
1122 if (isa<PointerType>(SrcTy)) {
1123 if (Instruction *I = commonPointerCastTransforms(CI))
1126 if (Instruction *Result = commonCastTransforms(CI))
1131 // Get rid of casts from one type to the same type. These are useless and can
1132 // be replaced by the operand.
1133 if (DestTy == Src->getType())
1134 return ReplaceInstUsesWith(CI, Src);
1136 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1137 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1138 const Type *DstElTy = DstPTy->getElementType();
1139 const Type *SrcElTy = SrcPTy->getElementType();
1141 // If the address spaces don't match, don't eliminate the bitcast, which is
1142 // required for changing types.
1143 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1146 // If we are casting a alloca to a pointer to a type of the same
1147 // size, rewrite the allocation instruction to allocate the "right" type.
1148 // There is no need to modify malloc calls because it is their bitcast that
1149 // needs to be cleaned up.
1150 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1151 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1154 // If the source and destination are pointers, and this cast is equivalent
1155 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1156 // This can enhance SROA and other transforms that want type-safe pointers.
1157 Constant *ZeroUInt =
1158 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1159 unsigned NumZeros = 0;
1160 while (SrcElTy != DstElTy &&
1161 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
1162 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1163 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1167 // If we found a path from the src to dest, create the getelementptr now.
1168 if (SrcElTy == DstElTy) {
1169 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1170 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1171 ((Instruction*) NULL));
1175 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1176 if (DestVTy->getNumElements() == 1) {
1177 if (!isa<VectorType>(SrcTy)) {
1178 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1179 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1180 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1182 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1186 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1187 if (SrcVTy->getNumElements() == 1) {
1188 if (!isa<VectorType>(DestTy)) {
1190 Builder->CreateExtractElement(Src,
1191 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1192 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1197 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1198 if (SVI->hasOneUse()) {
1199 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
1200 // a bitconvert to a vector with the same # elts.
1201 if (isa<VectorType>(DestTy) &&
1202 cast<VectorType>(DestTy)->getNumElements() ==
1203 SVI->getType()->getNumElements() &&
1204 SVI->getType()->getNumElements() ==
1205 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1207 // If either of the operands is a cast from CI.getType(), then
1208 // evaluating the shuffle in the casted destination's type will allow
1209 // us to eliminate at least one cast.
1210 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
1211 Tmp->getOperand(0)->getType() == DestTy) ||
1212 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
1213 Tmp->getOperand(0)->getType() == DestTy)) {
1214 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1215 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1216 // Return a new shuffle vector. Use the same element ID's, as we
1217 // know the vector types match #elts.
1218 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));