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() == Type::getInt32Ty(Val->getContext()) &&
27 "Unexpected allocation size type!");
28 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
29 Offset = CI->getZExtValue();
31 return ConstantInt::get(Type::getInt32Ty(Val->getContext()), 0);
32 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
33 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
34 if (I->getOpcode() == Instruction::Shl) {
35 // This is a value scaled by '1 << the shift amt'.
36 Scale = 1U << RHS->getZExtValue();
38 return I->getOperand(0);
39 } else if (I->getOpcode() == Instruction::Mul) {
40 // This value is scaled by 'RHS'.
41 Scale = RHS->getZExtValue();
43 return I->getOperand(0);
44 } else if (I->getOpcode() == Instruction::Add) {
45 // We have X+C. Check to see if we really have (X*C2)+C1,
46 // where C1 is divisible by C2.
49 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
50 Offset += RHS->getZExtValue();
57 // Otherwise, we can't look past this.
63 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
64 /// try to eliminate the cast by moving the type information into the alloc.
65 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
67 // This requires TargetData to get the alloca alignment and size information.
70 const PointerType *PTy = cast<PointerType>(CI.getType());
72 BuilderTy AllocaBuilder(*Builder);
73 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
75 // Get the type really allocated and the type casted to.
76 const Type *AllocElTy = AI.getAllocatedType();
77 const Type *CastElTy = PTy->getElementType();
78 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
80 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
81 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
82 if (CastElTyAlign < AllocElTyAlign) return 0;
84 // If the allocation has multiple uses, only promote it if we are strictly
85 // increasing the alignment of the resultant allocation. If we keep it the
86 // same, we open the door to infinite loops of various kinds. (A reference
87 // from a dbg.declare doesn't count as a use for this purpose.)
88 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
89 CastElTyAlign == AllocElTyAlign) return 0;
91 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
92 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
93 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
95 // See if we can satisfy the modulus by pulling a scale out of the array
97 unsigned ArraySizeScale;
99 Value *NumElements = // See if the array size is a decomposable linear expr.
100 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
102 // If we can now satisfy the modulus, by using a non-1 scale, we really can
104 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
105 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
107 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
112 Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale);
113 // Insert before the alloca, not before the cast.
114 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
117 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
118 Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()),
120 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
123 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
124 New->setAlignment(AI.getAlignment());
127 // If the allocation has one real use plus a dbg.declare, just remove the
129 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
130 EraseInstFromFunction(*(Instruction*)DI);
132 // If the allocation has multiple real uses, insert a cast and change all
133 // things that used it to use the new cast. This will also hack on CI, but it
135 else if (!AI.hasOneUse()) {
136 // New is the allocation instruction, pointer typed. AI is the original
137 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
138 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
139 AI.replaceAllUsesWith(NewCast);
141 return ReplaceInstUsesWith(CI, New);
145 /// CanEvaluateInDifferentType - Return true if we can take the specified value
146 /// and return it as type Ty without inserting any new casts and without
147 /// changing the computed value. This is used by code that tries to decide
148 /// whether promoting or shrinking integer operations to wider or smaller types
149 /// will allow us to eliminate a truncate or extend.
151 /// This is a truncation operation if Ty is smaller than V->getType(), or an
152 /// extension operation if Ty is larger.
154 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
155 /// should return true if trunc(V) can be computed by computing V in the smaller
156 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
157 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
158 /// efficiently truncated.
160 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
161 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
162 /// the final result.
163 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
165 int &NumCastsRemoved){
166 // We can always evaluate constants in another type.
167 if (isa<Constant>(V))
170 Instruction *I = dyn_cast<Instruction>(V);
171 if (!I) return false;
173 const Type *OrigTy = V->getType();
175 // If this is an extension or truncate, we can often eliminate it.
176 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
177 // If this is a cast from the destination type, we can trivially eliminate
178 // it, and this will remove a cast overall.
179 if (I->getOperand(0)->getType() == Ty) {
180 // If the first operand is itself a cast, and is eliminable, do not count
181 // this as an eliminable cast. We would prefer to eliminate those two
183 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
189 // We can't extend or shrink something that has multiple uses: doing so would
190 // require duplicating the instruction in general, which isn't profitable.
191 if (!I->hasOneUse()) return false;
193 unsigned Opc = I->getOpcode();
195 case Instruction::Add:
196 case Instruction::Sub:
197 case Instruction::Mul:
198 case Instruction::And:
199 case Instruction::Or:
200 case Instruction::Xor:
201 // These operators can all arbitrarily be extended or truncated.
202 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
204 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
207 case Instruction::UDiv:
208 case Instruction::URem: {
209 // UDiv and URem can be truncated if all the truncated bits are zero.
210 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
211 uint32_t BitWidth = Ty->getScalarSizeInBits();
212 if (BitWidth < OrigBitWidth) {
213 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
214 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
215 MaskedValueIsZero(I->getOperand(1), Mask)) {
216 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
218 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
224 case Instruction::Shl:
225 // If we are truncating the result of this SHL, and if it's a shift of a
226 // constant amount, we can always perform a SHL in a smaller type.
227 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
228 uint32_t BitWidth = Ty->getScalarSizeInBits();
229 if (BitWidth < OrigTy->getScalarSizeInBits() &&
230 CI->getLimitedValue(BitWidth) < BitWidth)
231 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
235 case Instruction::LShr:
236 // If this is a truncate of a logical shr, we can truncate it to a smaller
237 // lshr iff we know that the bits we would otherwise be shifting in are
239 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
240 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
241 uint32_t BitWidth = Ty->getScalarSizeInBits();
242 if (BitWidth < OrigBitWidth &&
243 MaskedValueIsZero(I->getOperand(0),
244 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
245 CI->getLimitedValue(BitWidth) < BitWidth) {
246 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
251 case Instruction::ZExt:
252 case Instruction::SExt:
253 case Instruction::Trunc:
254 // If this is the same kind of case as our original (e.g. zext+zext), we
255 // can safely replace it. Note that replacing it does not reduce the number
256 // of casts in the input.
260 // sext (zext ty1), ty2 -> zext ty2
261 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
264 case Instruction::Select: {
265 SelectInst *SI = cast<SelectInst>(I);
266 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
268 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
271 case Instruction::PHI: {
272 // We can change a phi if we can change all operands.
273 PHINode *PN = cast<PHINode>(I);
274 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
275 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
281 // TODO: Can handle more cases here.
288 /// EvaluateInDifferentType - Given an expression that
289 /// CanEvaluateInDifferentType returns true for, actually insert the code to
290 /// evaluate the expression.
291 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
293 if (Constant *C = dyn_cast<Constant>(V))
294 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
296 // Otherwise, it must be an instruction.
297 Instruction *I = cast<Instruction>(V);
298 Instruction *Res = 0;
299 unsigned Opc = I->getOpcode();
301 case Instruction::Add:
302 case Instruction::Sub:
303 case Instruction::Mul:
304 case Instruction::And:
305 case Instruction::Or:
306 case Instruction::Xor:
307 case Instruction::AShr:
308 case Instruction::LShr:
309 case Instruction::Shl:
310 case Instruction::UDiv:
311 case Instruction::URem: {
312 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
313 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
314 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
317 case Instruction::Trunc:
318 case Instruction::ZExt:
319 case Instruction::SExt:
320 // If the source type of the cast is the type we're trying for then we can
321 // just return the source. There's no need to insert it because it is not
323 if (I->getOperand(0)->getType() == Ty)
324 return I->getOperand(0);
326 // Otherwise, must be the same type of cast, so just reinsert a new one.
327 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
329 case Instruction::Select: {
330 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
331 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
332 Res = SelectInst::Create(I->getOperand(0), True, False);
335 case Instruction::PHI: {
336 PHINode *OPN = cast<PHINode>(I);
337 PHINode *NPN = PHINode::Create(Ty);
338 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
339 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
340 NPN->addIncoming(V, OPN->getIncomingBlock(i));
346 // TODO: Can handle more cases here.
347 llvm_unreachable("Unreachable!");
352 return InsertNewInstBefore(Res, *I);
356 /// This function is a wrapper around CastInst::isEliminableCastPair. It
357 /// simply extracts arguments and returns what that function returns.
358 static Instruction::CastOps
359 isEliminableCastPair(
360 const CastInst *CI, ///< The first cast instruction
361 unsigned opcode, ///< The opcode of the second cast instruction
362 const Type *DstTy, ///< The target type for the second cast instruction
363 TargetData *TD ///< The target data for pointer size
366 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
367 const Type *MidTy = CI->getType(); // B from above
369 // Get the opcodes of the two Cast instructions
370 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
371 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
373 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
375 TD ? TD->getIntPtrType(CI->getContext()) : 0);
377 // We don't want to form an inttoptr or ptrtoint that converts to an integer
378 // type that differs from the pointer size.
379 if ((Res == Instruction::IntToPtr &&
380 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
381 (Res == Instruction::PtrToInt &&
382 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
385 return Instruction::CastOps(Res);
388 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
389 /// in any code being generated. It does not require codegen if V is simple
390 /// enough or if the cast can be folded into other casts.
391 bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V,
393 if (V->getType() == Ty || isa<Constant>(V)) return false;
395 // If this is another cast that can be eliminated, it isn't codegen either.
396 if (const CastInst *CI = dyn_cast<CastInst>(V))
397 if (isEliminableCastPair(CI, opcode, Ty, TD))
403 /// @brief Implement the transforms common to all CastInst visitors.
404 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
405 Value *Src = CI.getOperand(0);
407 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
409 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
410 if (Instruction::CastOps opc =
411 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
412 // The first cast (CSrc) is eliminable so we need to fix up or replace
413 // the second cast (CI). CSrc will then have a good chance of being dead.
414 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
418 // If we are casting a select then fold the cast into the select
419 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
420 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
423 // If we are casting a PHI then fold the cast into the PHI
424 if (isa<PHINode>(Src)) {
425 // We don't do this if this would create a PHI node with an illegal type if
426 // it is currently legal.
427 if (!isa<IntegerType>(Src->getType()) ||
428 !isa<IntegerType>(CI.getType()) ||
429 ShouldChangeType(CI.getType(), Src->getType()))
430 if (Instruction *NV = FoldOpIntoPhi(CI))
437 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
438 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
439 Value *Src = CI.getOperand(0);
441 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
442 // If casting the result of a getelementptr instruction with no offset, turn
443 // this into a cast of the original pointer!
444 if (GEP->hasAllZeroIndices()) {
445 // Changing the cast operand is usually not a good idea but it is safe
446 // here because the pointer operand is being replaced with another
447 // pointer operand so the opcode doesn't need to change.
449 CI.setOperand(0, GEP->getOperand(0));
453 // If the GEP has a single use, and the base pointer is a bitcast, and the
454 // GEP computes a constant offset, see if we can convert these three
455 // instructions into fewer. This typically happens with unions and other
456 // non-type-safe code.
457 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
458 if (GEP->hasAllConstantIndices()) {
459 // We are guaranteed to get a constant from EmitGEPOffset.
460 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
461 int64_t Offset = OffsetV->getSExtValue();
463 // Get the base pointer input of the bitcast, and the type it points to.
464 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
465 const Type *GEPIdxTy =
466 cast<PointerType>(OrigBase->getType())->getElementType();
467 SmallVector<Value*, 8> NewIndices;
468 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
469 // If we were able to index down into an element, create the GEP
470 // and bitcast the result. This eliminates one bitcast, potentially
472 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
473 Builder->CreateInBoundsGEP(OrigBase,
474 NewIndices.begin(), NewIndices.end()) :
475 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
478 if (isa<BitCastInst>(CI))
479 return new BitCastInst(NGEP, CI.getType());
480 assert(isa<PtrToIntInst>(CI));
481 return new PtrToIntInst(NGEP, CI.getType());
487 return commonCastTransforms(CI);
490 /// commonIntCastTransforms - This function implements the common transforms
491 /// for trunc, zext, and sext.
492 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
493 if (Instruction *Result = commonCastTransforms(CI))
496 Value *Src = CI.getOperand(0);
497 const Type *SrcTy = Src->getType();
498 const Type *DestTy = CI.getType();
499 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
500 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
502 // See if we can simplify any instructions used by the LHS whose sole
503 // purpose is to compute bits we don't care about.
504 if (SimplifyDemandedInstructionBits(CI))
507 // If the source isn't an instruction or has more than one use then we
508 // can't do anything more.
509 Instruction *SrcI = dyn_cast<Instruction>(Src);
510 if (!SrcI || !Src->hasOneUse())
513 // Attempt to propagate the cast into the instruction for int->int casts.
514 int NumCastsRemoved = 0;
515 // Only do this if the dest type is a simple type, don't convert the
516 // expression tree to something weird like i93 unless the source is also
518 if ((isa<VectorType>(DestTy) ||
519 ShouldChangeType(SrcI->getType(), DestTy)) &&
520 CanEvaluateInDifferentType(SrcI, DestTy,
521 CI.getOpcode(), NumCastsRemoved)) {
522 // If this cast is a truncate, evaluting in a different type always
523 // eliminates the cast, so it is always a win. If this is a zero-extension,
524 // we need to do an AND to maintain the clear top-part of the computation,
525 // so we require that the input have eliminated at least one cast. If this
526 // is a sign extension, we insert two new casts (to do the extension) so we
527 // require that two casts have been eliminated.
528 bool DoXForm = false;
529 bool JustReplace = false;
530 switch (CI.getOpcode()) {
532 // All the others use floating point so we shouldn't actually
533 // get here because of the check above.
534 llvm_unreachable("Unknown cast type");
535 case Instruction::Trunc:
538 case Instruction::ZExt: {
539 DoXForm = NumCastsRemoved >= 1;
542 // If it's unnecessary to issue an AND to clear the high bits, it's
543 // always profitable to do this xform.
544 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
545 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
546 if (MaskedValueIsZero(TryRes, Mask))
547 return ReplaceInstUsesWith(CI, TryRes);
549 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
550 if (TryI->use_empty())
551 EraseInstFromFunction(*TryI);
555 case Instruction::SExt: {
556 DoXForm = NumCastsRemoved >= 2;
557 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
558 // If we do not have to emit the truncate + sext pair, then it's always
559 // profitable to do this xform.
561 // It's not safe to eliminate the trunc + sext pair if one of the
562 // eliminated cast is a truncate. e.g.
563 // t2 = trunc i32 t1 to i16
564 // t3 = sext i16 t2 to i32
567 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
568 unsigned NumSignBits = ComputeNumSignBits(TryRes);
569 if (NumSignBits > (DestBitSize - SrcBitSize))
570 return ReplaceInstUsesWith(CI, TryRes);
572 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
573 if (TryI->use_empty())
574 EraseInstFromFunction(*TryI);
581 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
582 " to avoid cast: " << CI);
583 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
584 CI.getOpcode() == Instruction::SExt);
586 // Just replace this cast with the result.
587 return ReplaceInstUsesWith(CI, Res);
589 assert(Res->getType() == DestTy);
590 switch (CI.getOpcode()) {
591 default: llvm_unreachable("Unknown cast type!");
592 case Instruction::Trunc:
593 // Just replace this cast with the result.
594 return ReplaceInstUsesWith(CI, Res);
595 case Instruction::ZExt: {
596 assert(SrcBitSize < DestBitSize && "Not a zext?");
598 // If the high bits are already zero, just replace this cast with the
600 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
601 if (MaskedValueIsZero(Res, Mask))
602 return ReplaceInstUsesWith(CI, Res);
604 // We need to emit an AND to clear the high bits.
605 Constant *C = ConstantInt::get(CI.getContext(),
606 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
607 return BinaryOperator::CreateAnd(Res, C);
609 case Instruction::SExt: {
610 // If the high bits are already filled with sign bit, just replace this
611 // cast with the result.
612 unsigned NumSignBits = ComputeNumSignBits(Res);
613 if (NumSignBits > (DestBitSize - SrcBitSize))
614 return ReplaceInstUsesWith(CI, Res);
616 // We need to emit a cast to truncate, then a cast to sext.
617 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
623 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
624 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
626 switch (SrcI->getOpcode()) {
627 case Instruction::Add:
628 case Instruction::Mul:
629 case Instruction::And:
630 case Instruction::Or:
631 case Instruction::Xor:
632 // If we are discarding information, rewrite.
633 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
634 // Don't insert two casts unless at least one can be eliminated.
635 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy) ||
636 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy)) {
637 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
638 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
639 return BinaryOperator::Create(
640 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
644 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
645 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
646 SrcI->getOpcode() == Instruction::Xor &&
647 Op1 == ConstantInt::getTrue(CI.getContext()) &&
648 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
649 Value *New = Builder->CreateZExt(Op0, DestTy, Op0->getName());
650 return BinaryOperator::CreateXor(New,
651 ConstantInt::get(CI.getType(), 1));
655 case Instruction::Shl: {
656 // Canonicalize trunc inside shl, if we can.
657 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
658 if (CI && DestBitSize < SrcBitSize &&
659 CI->getLimitedValue(DestBitSize) < DestBitSize) {
660 Value *Op0c = Builder->CreateTrunc(Op0, DestTy, Op0->getName());
661 Value *Op1c = Builder->CreateTrunc(Op1, DestTy, Op1->getName());
662 return BinaryOperator::CreateShl(Op0c, Op1c);
671 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
672 if (Instruction *Result = commonIntCastTransforms(CI))
675 Value *Src = CI.getOperand(0);
676 const Type *Ty = CI.getType();
677 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
678 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
680 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
681 if (DestBitWidth == 1) {
682 Constant *One = ConstantInt::get(Src->getType(), 1);
683 Src = Builder->CreateAnd(Src, One, "tmp");
684 Value *Zero = Constant::getNullValue(Src->getType());
685 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
688 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
689 ConstantInt *ShAmtV = 0;
691 if (Src->hasOneUse() &&
692 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
693 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
695 // Get a mask for the bits shifting in.
696 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
697 if (MaskedValueIsZero(ShiftOp, Mask)) {
698 if (ShAmt >= DestBitWidth) // All zeros.
699 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
701 // Okay, we can shrink this. Truncate the input, then return a new
703 Value *V1 = Builder->CreateTrunc(ShiftOp, Ty, ShiftOp->getName());
704 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
705 return BinaryOperator::CreateLShr(V1, V2);
712 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
713 /// in order to eliminate the icmp.
714 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
716 // If we are just checking for a icmp eq of a single bit and zext'ing it
717 // to an integer, then shift the bit to the appropriate place and then
718 // cast to integer to avoid the comparison.
719 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
720 const APInt &Op1CV = Op1C->getValue();
722 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
723 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
724 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
725 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
726 if (!DoXform) return ICI;
728 Value *In = ICI->getOperand(0);
729 Value *Sh = ConstantInt::get(In->getType(),
730 In->getType()->getScalarSizeInBits()-1);
731 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
732 if (In->getType() != CI.getType())
733 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
735 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
736 Constant *One = ConstantInt::get(In->getType(), 1);
737 In = Builder->CreateXor(In, One, In->getName()+".not");
740 return ReplaceInstUsesWith(CI, In);
745 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
746 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
747 // zext (X == 1) to i32 --> X iff X has only the low bit set.
748 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
749 // zext (X != 0) to i32 --> X iff X has only the low bit set.
750 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
751 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
752 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
753 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
754 // This only works for EQ and NE
756 // If Op1C some other power of two, convert:
757 uint32_t BitWidth = Op1C->getType()->getBitWidth();
758 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
759 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
760 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
762 APInt KnownZeroMask(~KnownZero);
763 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
764 if (!DoXform) return ICI;
766 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
767 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
768 // (X&4) == 2 --> false
769 // (X&4) != 2 --> true
770 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
772 Res = ConstantExpr::getZExt(Res, CI.getType());
773 return ReplaceInstUsesWith(CI, Res);
776 uint32_t ShiftAmt = KnownZeroMask.logBase2();
777 Value *In = ICI->getOperand(0);
779 // Perform a logical shr by shiftamt.
780 // Insert the shift to put the result in the low bit.
781 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
782 In->getName()+".lobit");
785 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
786 Constant *One = ConstantInt::get(In->getType(), 1);
787 In = Builder->CreateXor(In, One, "tmp");
790 if (CI.getType() == In->getType())
791 return ReplaceInstUsesWith(CI, In);
793 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
798 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
799 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
800 // may lead to additional simplifications.
801 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
802 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
803 uint32_t BitWidth = ITy->getBitWidth();
804 Value *LHS = ICI->getOperand(0);
805 Value *RHS = ICI->getOperand(1);
807 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
808 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
809 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
810 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
811 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
813 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
814 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
815 APInt UnknownBit = ~KnownBits;
816 if (UnknownBit.countPopulation() == 1) {
817 if (!DoXform) return ICI;
819 Value *Result = Builder->CreateXor(LHS, RHS);
821 // Mask off any bits that are set and won't be shifted away.
822 if (KnownOneLHS.uge(UnknownBit))
823 Result = Builder->CreateAnd(Result,
824 ConstantInt::get(ITy, UnknownBit));
826 // Shift the bit we're testing down to the lsb.
827 Result = Builder->CreateLShr(
828 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
830 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
831 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
832 Result->takeName(ICI);
833 return ReplaceInstUsesWith(CI, Result);
842 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
843 // If one of the common conversion will work, do it.
844 if (Instruction *Result = commonIntCastTransforms(CI))
847 Value *Src = CI.getOperand(0);
849 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
850 // types and if the sizes are just right we can convert this into a logical
851 // 'and' which will be much cheaper than the pair of casts.
852 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
853 // Get the sizes of the types involved. We know that the intermediate type
854 // will be smaller than A or C, but don't know the relation between A and C.
855 Value *A = CSrc->getOperand(0);
856 unsigned SrcSize = A->getType()->getScalarSizeInBits();
857 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
858 unsigned DstSize = CI.getType()->getScalarSizeInBits();
859 // If we're actually extending zero bits, then if
860 // SrcSize < DstSize: zext(a & mask)
861 // SrcSize == DstSize: a & mask
862 // SrcSize > DstSize: trunc(a) & mask
863 if (SrcSize < DstSize) {
864 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
865 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
866 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
867 return new ZExtInst(And, CI.getType());
870 if (SrcSize == DstSize) {
871 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
872 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
875 if (SrcSize > DstSize) {
876 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
877 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
878 return BinaryOperator::CreateAnd(Trunc,
879 ConstantInt::get(Trunc->getType(),
884 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
885 return transformZExtICmp(ICI, CI);
887 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
888 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
889 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
890 // of the (zext icmp) will be transformed.
891 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
892 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
893 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
894 (transformZExtICmp(LHS, CI, false) ||
895 transformZExtICmp(RHS, CI, false))) {
896 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
897 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
898 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
902 // zext(trunc(t) & C) -> (t & zext(C)).
903 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
904 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
905 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
906 Value *TI0 = TI->getOperand(0);
907 if (TI0->getType() == CI.getType())
909 BinaryOperator::CreateAnd(TI0,
910 ConstantExpr::getZExt(C, CI.getType()));
913 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
914 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
915 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
916 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
917 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
918 And->getOperand(1) == C)
919 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
920 Value *TI0 = TI->getOperand(0);
921 if (TI0->getType() == CI.getType()) {
922 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
923 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
924 return BinaryOperator::CreateXor(NewAnd, ZC);
931 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
932 if (Instruction *I = commonIntCastTransforms(CI))
935 Value *Src = CI.getOperand(0);
937 // Canonicalize sign-extend from i1 to a select.
938 if (Src->getType() == Type::getInt1Ty(CI.getContext()))
939 return SelectInst::Create(Src,
940 Constant::getAllOnesValue(CI.getType()),
941 Constant::getNullValue(CI.getType()));
943 // See if the value being truncated is already sign extended. If so, just
944 // eliminate the trunc/sext pair.
945 if (Operator::getOpcode(Src) == Instruction::Trunc) {
946 Value *Op = cast<User>(Src)->getOperand(0);
947 unsigned OpBits = Op->getType()->getScalarSizeInBits();
948 unsigned MidBits = Src->getType()->getScalarSizeInBits();
949 unsigned DestBits = CI.getType()->getScalarSizeInBits();
950 unsigned NumSignBits = ComputeNumSignBits(Op);
952 if (OpBits == DestBits) {
953 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
954 // bits, it is already ready.
955 if (NumSignBits > DestBits-MidBits)
956 return ReplaceInstUsesWith(CI, Op);
957 } else if (OpBits < DestBits) {
958 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
959 // bits, just sext from i32.
960 if (NumSignBits > OpBits-MidBits)
961 return new SExtInst(Op, CI.getType(), "tmp");
963 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
964 // bits, just truncate to i32.
965 if (NumSignBits > OpBits-MidBits)
966 return new TruncInst(Op, CI.getType(), "tmp");
970 // If the input is a shl/ashr pair of a same constant, then this is a sign
971 // extension from a smaller value. If we could trust arbitrary bitwidth
972 // integers, we could turn this into a truncate to the smaller bit and then
973 // use a sext for the whole extension. Since we don't, look deeper and check
974 // for a truncate. If the source and dest are the same type, eliminate the
975 // trunc and extend and just do shifts. For example, turn:
976 // %a = trunc i32 %i to i8
978 // %c = ashr i8 %b, 6
979 // %d = sext i8 %c to i32
981 // %a = shl i32 %i, 30
982 // %d = ashr i32 %a, 30
984 ConstantInt *BA = 0, *CA = 0;
985 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
986 m_ConstantInt(CA))) &&
987 BA == CA && isa<TruncInst>(A)) {
988 Value *I = cast<TruncInst>(A)->getOperand(0);
989 if (I->getType() == CI.getType()) {
990 unsigned MidSize = Src->getType()->getScalarSizeInBits();
991 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
992 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
993 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
994 I = Builder->CreateShl(I, ShAmtV, CI.getName());
995 return BinaryOperator::CreateAShr(I, ShAmtV);
1003 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1004 /// in the specified FP type without changing its value.
1005 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1007 APFloat F = CFP->getValueAPF();
1008 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1010 return ConstantFP::get(CFP->getContext(), F);
1014 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1015 /// through it until we get the source value.
1016 static Value *LookThroughFPExtensions(Value *V) {
1017 if (Instruction *I = dyn_cast<Instruction>(V))
1018 if (I->getOpcode() == Instruction::FPExt)
1019 return LookThroughFPExtensions(I->getOperand(0));
1021 // If this value is a constant, return the constant in the smallest FP type
1022 // that can accurately represent it. This allows us to turn
1023 // (float)((double)X+2.0) into x+2.0f.
1024 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1025 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1026 return V; // No constant folding of this.
1027 // See if the value can be truncated to float and then reextended.
1028 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1030 if (CFP->getType() == Type::getDoubleTy(V->getContext()))
1031 return V; // Won't shrink.
1032 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1034 // Don't try to shrink to various long double types.
1040 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1041 if (Instruction *I = commonCastTransforms(CI))
1044 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1045 // smaller than the destination type, we can eliminate the truncate by doing
1046 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1047 // as many builtins (sqrt, etc).
1048 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1049 if (OpI && OpI->hasOneUse()) {
1050 switch (OpI->getOpcode()) {
1052 case Instruction::FAdd:
1053 case Instruction::FSub:
1054 case Instruction::FMul:
1055 case Instruction::FDiv:
1056 case Instruction::FRem:
1057 const Type *SrcTy = OpI->getType();
1058 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1059 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1060 if (LHSTrunc->getType() != SrcTy &&
1061 RHSTrunc->getType() != SrcTy) {
1062 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1063 // If the source types were both smaller than the destination type of
1064 // the cast, do this xform.
1065 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1066 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1067 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1068 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1069 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1078 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1079 return commonCastTransforms(CI);
1082 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1083 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1085 return commonCastTransforms(FI);
1087 // fptoui(uitofp(X)) --> X
1088 // fptoui(sitofp(X)) --> X
1089 // This is safe if the intermediate type has enough bits in its mantissa to
1090 // accurately represent all values of X. For example, do not do this with
1091 // i64->float->i64. This is also safe for sitofp case, because any negative
1092 // 'X' value would cause an undefined result for the fptoui.
1093 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1094 OpI->getOperand(0)->getType() == FI.getType() &&
1095 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1096 OpI->getType()->getFPMantissaWidth())
1097 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1099 return commonCastTransforms(FI);
1102 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1103 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1105 return commonCastTransforms(FI);
1107 // fptosi(sitofp(X)) --> X
1108 // fptosi(uitofp(X)) --> X
1109 // This is safe if the intermediate type has enough bits in its mantissa to
1110 // accurately represent all values of X. For example, do not do this with
1111 // i64->float->i64. This is also safe for sitofp case, because any negative
1112 // 'X' value would cause an undefined result for the fptoui.
1113 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1114 OpI->getOperand(0)->getType() == FI.getType() &&
1115 (int)FI.getType()->getScalarSizeInBits() <=
1116 OpI->getType()->getFPMantissaWidth())
1117 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1119 return commonCastTransforms(FI);
1122 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1123 return commonCastTransforms(CI);
1126 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1127 return commonCastTransforms(CI);
1130 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1131 // If the destination integer type is smaller than the intptr_t type for
1132 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
1133 // trunc to be exposed to other transforms. Don't do this for extending
1134 // ptrtoint's, because we don't know if the target sign or zero extends its
1137 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1138 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1139 TD->getIntPtrType(CI.getContext()),
1141 return new TruncInst(P, CI.getType());
1144 return commonPointerCastTransforms(CI);
1148 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1149 // If the source integer type is larger than the intptr_t type for
1150 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
1151 // allows the trunc to be exposed to other transforms. Don't do this for
1152 // extending inttoptr's, because we don't know if the target sign or zero
1153 // extends to pointers.
1154 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
1155 TD->getPointerSizeInBits()) {
1156 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1157 TD->getIntPtrType(CI.getContext()), "tmp");
1158 return new IntToPtrInst(P, CI.getType());
1161 if (Instruction *I = commonCastTransforms(CI))
1167 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1168 // If the operands are integer typed then apply the integer transforms,
1169 // otherwise just apply the common ones.
1170 Value *Src = CI.getOperand(0);
1171 const Type *SrcTy = Src->getType();
1172 const Type *DestTy = CI.getType();
1174 if (isa<PointerType>(SrcTy)) {
1175 if (Instruction *I = commonPointerCastTransforms(CI))
1178 if (Instruction *Result = commonCastTransforms(CI))
1183 // Get rid of casts from one type to the same type. These are useless and can
1184 // be replaced by the operand.
1185 if (DestTy == Src->getType())
1186 return ReplaceInstUsesWith(CI, Src);
1188 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1189 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1190 const Type *DstElTy = DstPTy->getElementType();
1191 const Type *SrcElTy = SrcPTy->getElementType();
1193 // If the address spaces don't match, don't eliminate the bitcast, which is
1194 // required for changing types.
1195 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1198 // If we are casting a alloca to a pointer to a type of the same
1199 // size, rewrite the allocation instruction to allocate the "right" type.
1200 // There is no need to modify malloc calls because it is their bitcast that
1201 // needs to be cleaned up.
1202 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1203 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1206 // If the source and destination are pointers, and this cast is equivalent
1207 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1208 // This can enhance SROA and other transforms that want type-safe pointers.
1209 Constant *ZeroUInt =
1210 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1211 unsigned NumZeros = 0;
1212 while (SrcElTy != DstElTy &&
1213 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
1214 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1215 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1219 // If we found a path from the src to dest, create the getelementptr now.
1220 if (SrcElTy == DstElTy) {
1221 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1222 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1223 ((Instruction*) NULL));
1227 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1228 if (DestVTy->getNumElements() == 1) {
1229 if (!isa<VectorType>(SrcTy)) {
1230 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1231 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1232 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1234 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1238 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1239 if (SrcVTy->getNumElements() == 1) {
1240 if (!isa<VectorType>(DestTy)) {
1242 Builder->CreateExtractElement(Src,
1243 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1244 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1249 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1250 if (SVI->hasOneUse()) {
1251 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
1252 // a bitconvert to a vector with the same # elts.
1253 if (isa<VectorType>(DestTy) &&
1254 cast<VectorType>(DestTy)->getNumElements() ==
1255 SVI->getType()->getNumElements() &&
1256 SVI->getType()->getNumElements() ==
1257 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1259 // If either of the operands is a cast from CI.getType(), then
1260 // evaluating the shuffle in the casted destination's type will allow
1261 // us to eliminate at least one cast.
1262 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
1263 Tmp->getOperand(0)->getType() == DestTy) ||
1264 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
1265 Tmp->getOperand(0)->getType() == DestTy)) {
1266 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1267 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1268 // Return a new shuffle vector. Use the same element ID's, as we
1269 // know the vector types match #elts.
1270 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));