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 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
27 Offset = CI->getZExtValue();
29 return ConstantInt::get(Val->getType(), 0);
32 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 = UINT64_C(1) << RHS->getZExtValue();
38 return I->getOperand(0);
41 if (I->getOpcode() == Instruction::Mul) {
42 // This value is scaled by 'RHS'.
43 Scale = RHS->getZExtValue();
45 return I->getOperand(0);
48 if (I->getOpcode() == Instruction::Add) {
49 // We have X+C. Check to see if we really have (X*C2)+C1,
50 // where C1 is divisible by C2.
53 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
54 Offset += RHS->getZExtValue();
61 // Otherwise, we can't look past this.
67 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
68 /// try to eliminate the cast by moving the type information into the alloc.
69 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
71 // This requires TargetData to get the alloca alignment and size information.
74 const PointerType *PTy = cast<PointerType>(CI.getType());
76 BuilderTy AllocaBuilder(*Builder);
77 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
79 // Get the type really allocated and the type casted to.
80 const Type *AllocElTy = AI.getAllocatedType();
81 const Type *CastElTy = PTy->getElementType();
82 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
84 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
85 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
86 if (CastElTyAlign < AllocElTyAlign) return 0;
88 // If the allocation has multiple uses, only promote it if we are strictly
89 // increasing the alignment of the resultant allocation. If we keep it the
90 // same, we open the door to infinite loops of various kinds.
91 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
93 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
94 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
95 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
97 // See if we can satisfy the modulus by pulling a scale out of the array
99 unsigned ArraySizeScale;
100 uint64_t ArrayOffset;
101 Value *NumElements = // See if the array size is a decomposable linear expr.
102 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
104 // If we can now satisfy the modulus, by using a non-1 scale, we really can
106 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
107 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
109 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
114 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
115 // Insert before the alloca, not before the cast.
116 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
119 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
120 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
122 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
125 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
126 New->setAlignment(AI.getAlignment());
129 // If the allocation has multiple real uses, insert a cast and change all
130 // things that used it to use the new cast. This will also hack on CI, but it
132 if (!AI.hasOneUse()) {
133 // New is the allocation instruction, pointer typed. AI is the original
134 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
135 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
136 AI.replaceAllUsesWith(NewCast);
138 return ReplaceInstUsesWith(CI, New);
143 /// EvaluateInDifferentType - Given an expression that
144 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
145 /// insert the code to evaluate the expression.
146 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
148 if (Constant *C = dyn_cast<Constant>(V)) {
149 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
150 // If we got a constantexpr back, try to simplify it with TD info.
151 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
152 C = ConstantFoldConstantExpression(CE, TD);
156 // Otherwise, it must be an instruction.
157 Instruction *I = cast<Instruction>(V);
158 Instruction *Res = 0;
159 unsigned Opc = I->getOpcode();
161 case Instruction::Add:
162 case Instruction::Sub:
163 case Instruction::Mul:
164 case Instruction::And:
165 case Instruction::Or:
166 case Instruction::Xor:
167 case Instruction::AShr:
168 case Instruction::LShr:
169 case Instruction::Shl:
170 case Instruction::UDiv:
171 case Instruction::URem: {
172 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
173 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
174 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
177 case Instruction::Trunc:
178 case Instruction::ZExt:
179 case Instruction::SExt:
180 // If the source type of the cast is the type we're trying for then we can
181 // just return the source. There's no need to insert it because it is not
183 if (I->getOperand(0)->getType() == Ty)
184 return I->getOperand(0);
186 // Otherwise, must be the same type of cast, so just reinsert a new one.
187 // This also handles the case of zext(trunc(x)) -> zext(x).
188 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
189 Opc == Instruction::SExt);
191 case Instruction::Select: {
192 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
193 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
194 Res = SelectInst::Create(I->getOperand(0), True, False);
197 case Instruction::PHI: {
198 PHINode *OPN = cast<PHINode>(I);
199 PHINode *NPN = PHINode::Create(Ty);
200 NPN->reserveOperandSpace(OPN->getNumIncomingValues());
201 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
202 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
203 NPN->addIncoming(V, OPN->getIncomingBlock(i));
209 // TODO: Can handle more cases here.
210 llvm_unreachable("Unreachable!");
215 return InsertNewInstBefore(Res, *I);
219 /// This function is a wrapper around CastInst::isEliminableCastPair. It
220 /// simply extracts arguments and returns what that function returns.
221 static Instruction::CastOps
222 isEliminableCastPair(
223 const CastInst *CI, ///< The first cast instruction
224 unsigned opcode, ///< The opcode of the second cast instruction
225 const Type *DstTy, ///< The target type for the second cast instruction
226 TargetData *TD ///< The target data for pointer size
229 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
230 const Type *MidTy = CI->getType(); // B from above
232 // Get the opcodes of the two Cast instructions
233 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
234 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
236 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
238 TD ? TD->getIntPtrType(CI->getContext()) : 0);
240 // We don't want to form an inttoptr or ptrtoint that converts to an integer
241 // type that differs from the pointer size.
242 if ((Res == Instruction::IntToPtr &&
243 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
244 (Res == Instruction::PtrToInt &&
245 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
248 return Instruction::CastOps(Res);
251 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
252 /// results in any code being generated and is interesting to optimize out. If
253 /// the cast can be eliminated by some other simple transformation, we prefer
254 /// to do the simplification first.
255 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
257 // Noop casts and casts of constants should be eliminated trivially.
258 if (V->getType() == Ty || isa<Constant>(V)) return false;
260 // If this is another cast that can be eliminated, we prefer to have it
262 if (const CastInst *CI = dyn_cast<CastInst>(V))
263 if (isEliminableCastPair(CI, opc, Ty, TD))
266 // If this is a vector sext from a compare, then we don't want to break the
267 // idiom where each element of the extended vector is either zero or all ones.
268 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
275 /// @brief Implement the transforms common to all CastInst visitors.
276 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
277 Value *Src = CI.getOperand(0);
279 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
281 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
282 if (Instruction::CastOps opc =
283 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
284 // The first cast (CSrc) is eliminable so we need to fix up or replace
285 // the second cast (CI). CSrc will then have a good chance of being dead.
286 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
290 // If we are casting a select then fold the cast into the select
291 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
292 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
295 // If we are casting a PHI then fold the cast into the PHI
296 if (isa<PHINode>(Src)) {
297 // We don't do this if this would create a PHI node with an illegal type if
298 // it is currently legal.
299 if (!Src->getType()->isIntegerTy() ||
300 !CI.getType()->isIntegerTy() ||
301 ShouldChangeType(CI.getType(), Src->getType()))
302 if (Instruction *NV = FoldOpIntoPhi(CI))
309 /// CanEvaluateTruncated - Return true if we can evaluate the specified
310 /// expression tree as type Ty instead of its larger type, and arrive with the
311 /// same value. This is used by code that tries to eliminate truncates.
313 /// Ty will always be a type smaller than V. We should return true if trunc(V)
314 /// can be computed by computing V in the smaller type. If V is an instruction,
315 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
316 /// makes sense if x and y can be efficiently truncated.
318 /// This function works on both vectors and scalars.
320 static bool CanEvaluateTruncated(Value *V, const Type *Ty) {
321 // We can always evaluate constants in another type.
322 if (isa<Constant>(V))
325 Instruction *I = dyn_cast<Instruction>(V);
326 if (!I) return false;
328 const Type *OrigTy = V->getType();
330 // If this is an extension from the dest type, we can eliminate it, even if it
331 // has multiple uses.
332 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
333 I->getOperand(0)->getType() == Ty)
336 // We can't extend or shrink something that has multiple uses: doing so would
337 // require duplicating the instruction in general, which isn't profitable.
338 if (!I->hasOneUse()) return false;
340 unsigned Opc = I->getOpcode();
342 case Instruction::Add:
343 case Instruction::Sub:
344 case Instruction::Mul:
345 case Instruction::And:
346 case Instruction::Or:
347 case Instruction::Xor:
348 // These operators can all arbitrarily be extended or truncated.
349 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
350 CanEvaluateTruncated(I->getOperand(1), Ty);
352 case Instruction::UDiv:
353 case Instruction::URem: {
354 // UDiv and URem can be truncated if all the truncated bits are zero.
355 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
356 uint32_t BitWidth = Ty->getScalarSizeInBits();
357 if (BitWidth < OrigBitWidth) {
358 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
359 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
360 MaskedValueIsZero(I->getOperand(1), Mask)) {
361 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
362 CanEvaluateTruncated(I->getOperand(1), Ty);
367 case Instruction::Shl:
368 // If we are truncating the result of this SHL, and if it's a shift of a
369 // constant amount, we can always perform a SHL in a smaller type.
370 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
371 uint32_t BitWidth = Ty->getScalarSizeInBits();
372 if (CI->getLimitedValue(BitWidth) < BitWidth)
373 return CanEvaluateTruncated(I->getOperand(0), Ty);
376 case Instruction::LShr:
377 // If this is a truncate of a logical shr, we can truncate it to a smaller
378 // lshr iff we know that the bits we would otherwise be shifting in are
380 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
381 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
382 uint32_t BitWidth = Ty->getScalarSizeInBits();
383 if (MaskedValueIsZero(I->getOperand(0),
384 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
385 CI->getLimitedValue(BitWidth) < BitWidth) {
386 return CanEvaluateTruncated(I->getOperand(0), Ty);
390 case Instruction::Trunc:
391 // trunc(trunc(x)) -> trunc(x)
393 case Instruction::ZExt:
394 case Instruction::SExt:
395 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
396 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
398 case Instruction::Select: {
399 SelectInst *SI = cast<SelectInst>(I);
400 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
401 CanEvaluateTruncated(SI->getFalseValue(), Ty);
403 case Instruction::PHI: {
404 // We can change a phi if we can change all operands. Note that we never
405 // get into trouble with cyclic PHIs here because we only consider
406 // instructions with a single use.
407 PHINode *PN = cast<PHINode>(I);
408 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
409 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
414 // TODO: Can handle more cases here.
421 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
422 if (Instruction *Result = commonCastTransforms(CI))
425 // See if we can simplify any instructions used by the input whose sole
426 // purpose is to compute bits we don't care about.
427 if (SimplifyDemandedInstructionBits(CI))
430 Value *Src = CI.getOperand(0);
431 const Type *DestTy = CI.getType(), *SrcTy = Src->getType();
433 // Attempt to truncate the entire input expression tree to the destination
434 // type. Only do this if the dest type is a simple type, don't convert the
435 // expression tree to something weird like i93 unless the source is also
437 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
438 CanEvaluateTruncated(Src, DestTy)) {
440 // If this cast is a truncate, evaluting in a different type always
441 // eliminates the cast, so it is always a win.
442 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
443 " to avoid cast: " << CI << '\n');
444 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
445 assert(Res->getType() == DestTy);
446 return ReplaceInstUsesWith(CI, Res);
449 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
450 if (DestTy->getScalarSizeInBits() == 1) {
451 Constant *One = ConstantInt::get(Src->getType(), 1);
452 Src = Builder->CreateAnd(Src, One, "tmp");
453 Value *Zero = Constant::getNullValue(Src->getType());
454 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
457 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
458 Value *A = 0; ConstantInt *Cst = 0;
459 if (Src->hasOneUse() &&
460 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
461 // We have three types to worry about here, the type of A, the source of
462 // the truncate (MidSize), and the destination of the truncate. We know that
463 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
464 // between ASize and ResultSize.
465 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
467 // If the shift amount is larger than the size of A, then the result is
468 // known to be zero because all the input bits got shifted out.
469 if (Cst->getZExtValue() >= ASize)
470 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
472 // Since we're doing an lshr and a zero extend, and know that the shift
473 // amount is smaller than ASize, it is always safe to do the shift in A's
474 // type, then zero extend or truncate to the result.
475 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
476 Shift->takeName(Src);
477 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
480 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
481 // type isn't non-native.
482 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
483 ShouldChangeType(Src->getType(), CI.getType()) &&
484 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
485 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
486 return BinaryOperator::CreateAnd(NewTrunc,
487 ConstantExpr::getTrunc(Cst, CI.getType()));
493 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
494 /// in order to eliminate the icmp.
495 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
497 // If we are just checking for a icmp eq of a single bit and zext'ing it
498 // to an integer, then shift the bit to the appropriate place and then
499 // cast to integer to avoid the comparison.
500 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
501 const APInt &Op1CV = Op1C->getValue();
503 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
504 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
505 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
506 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
507 if (!DoXform) return ICI;
509 Value *In = ICI->getOperand(0);
510 Value *Sh = ConstantInt::get(In->getType(),
511 In->getType()->getScalarSizeInBits()-1);
512 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
513 if (In->getType() != CI.getType())
514 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
516 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
517 Constant *One = ConstantInt::get(In->getType(), 1);
518 In = Builder->CreateXor(In, One, In->getName()+".not");
521 return ReplaceInstUsesWith(CI, In);
526 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
527 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
528 // zext (X == 1) to i32 --> X iff X has only the low bit set.
529 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
530 // zext (X != 0) to i32 --> X iff X has only the low bit set.
531 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
532 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
533 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
534 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
535 // This only works for EQ and NE
537 // If Op1C some other power of two, convert:
538 uint32_t BitWidth = Op1C->getType()->getBitWidth();
539 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
540 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
541 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
543 APInt KnownZeroMask(~KnownZero);
544 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
545 if (!DoXform) return ICI;
547 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
548 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
549 // (X&4) == 2 --> false
550 // (X&4) != 2 --> true
551 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
553 Res = ConstantExpr::getZExt(Res, CI.getType());
554 return ReplaceInstUsesWith(CI, Res);
557 uint32_t ShiftAmt = KnownZeroMask.logBase2();
558 Value *In = ICI->getOperand(0);
560 // Perform a logical shr by shiftamt.
561 // Insert the shift to put the result in the low bit.
562 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
563 In->getName()+".lobit");
566 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
567 Constant *One = ConstantInt::get(In->getType(), 1);
568 In = Builder->CreateXor(In, One, "tmp");
571 if (CI.getType() == In->getType())
572 return ReplaceInstUsesWith(CI, In);
573 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
578 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
579 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
580 // may lead to additional simplifications.
581 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
582 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
583 uint32_t BitWidth = ITy->getBitWidth();
584 Value *LHS = ICI->getOperand(0);
585 Value *RHS = ICI->getOperand(1);
587 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
588 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
589 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
590 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
591 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
593 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
594 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
595 APInt UnknownBit = ~KnownBits;
596 if (UnknownBit.countPopulation() == 1) {
597 if (!DoXform) return ICI;
599 Value *Result = Builder->CreateXor(LHS, RHS);
601 // Mask off any bits that are set and won't be shifted away.
602 if (KnownOneLHS.uge(UnknownBit))
603 Result = Builder->CreateAnd(Result,
604 ConstantInt::get(ITy, UnknownBit));
606 // Shift the bit we're testing down to the lsb.
607 Result = Builder->CreateLShr(
608 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
610 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
611 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
612 Result->takeName(ICI);
613 return ReplaceInstUsesWith(CI, Result);
622 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
623 /// specified wider type and produce the same low bits. If not, return false.
625 /// If this function returns true, it can also return a non-zero number of bits
626 /// (in BitsToClear) which indicates that the value it computes is correct for
627 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
628 /// out. For example, to promote something like:
630 /// %B = trunc i64 %A to i32
631 /// %C = lshr i32 %B, 8
632 /// %E = zext i32 %C to i64
634 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
635 /// set to 8 to indicate that the promoted value needs to have bits 24-31
636 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
637 /// clear the top bits anyway, doing this has no extra cost.
639 /// This function works on both vectors and scalars.
640 static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) {
642 if (isa<Constant>(V))
645 Instruction *I = dyn_cast<Instruction>(V);
646 if (!I) return false;
648 // If the input is a truncate from the destination type, we can trivially
649 // eliminate it, even if it has multiple uses.
650 // FIXME: This is currently disabled until codegen can handle this without
651 // pessimizing code, PR5997.
652 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
655 // We can't extend or shrink something that has multiple uses: doing so would
656 // require duplicating the instruction in general, which isn't profitable.
657 if (!I->hasOneUse()) return false;
659 unsigned Opc = I->getOpcode(), Tmp;
661 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
662 case Instruction::SExt: // zext(sext(x)) -> sext(x).
663 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
665 case Instruction::And:
666 case Instruction::Or:
667 case Instruction::Xor:
668 case Instruction::Add:
669 case Instruction::Sub:
670 case Instruction::Mul:
671 case Instruction::Shl:
672 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
673 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
675 // These can all be promoted if neither operand has 'bits to clear'.
676 if (BitsToClear == 0 && Tmp == 0)
679 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
680 // other side, BitsToClear is ok.
682 (Opc == Instruction::And || Opc == Instruction::Or ||
683 Opc == Instruction::Xor)) {
684 // We use MaskedValueIsZero here for generality, but the case we care
685 // about the most is constant RHS.
686 unsigned VSize = V->getType()->getScalarSizeInBits();
687 if (MaskedValueIsZero(I->getOperand(1),
688 APInt::getHighBitsSet(VSize, BitsToClear)))
692 // Otherwise, we don't know how to analyze this BitsToClear case yet.
695 case Instruction::LShr:
696 // We can promote lshr(x, cst) if we can promote x. This requires the
697 // ultimate 'and' to clear out the high zero bits we're clearing out though.
698 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
699 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
701 BitsToClear += Amt->getZExtValue();
702 if (BitsToClear > V->getType()->getScalarSizeInBits())
703 BitsToClear = V->getType()->getScalarSizeInBits();
706 // Cannot promote variable LSHR.
708 case Instruction::Select:
709 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
710 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
711 // TODO: If important, we could handle the case when the BitsToClear are
712 // known zero in the disagreeing side.
717 case Instruction::PHI: {
718 // We can change a phi if we can change all operands. Note that we never
719 // get into trouble with cyclic PHIs here because we only consider
720 // instructions with a single use.
721 PHINode *PN = cast<PHINode>(I);
722 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
724 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
725 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
726 // TODO: If important, we could handle the case when the BitsToClear
727 // are known zero in the disagreeing input.
733 // TODO: Can handle more cases here.
738 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
739 // If this zero extend is only used by a truncate, let the truncate by
740 // eliminated before we try to optimize this zext.
741 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
744 // If one of the common conversion will work, do it.
745 if (Instruction *Result = commonCastTransforms(CI))
748 // See if we can simplify any instructions used by the input whose sole
749 // purpose is to compute bits we don't care about.
750 if (SimplifyDemandedInstructionBits(CI))
753 Value *Src = CI.getOperand(0);
754 const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
756 // Attempt to extend the entire input expression tree to the destination
757 // type. Only do this if the dest type is a simple type, don't convert the
758 // expression tree to something weird like i93 unless the source is also
760 unsigned BitsToClear;
761 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
762 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
763 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
764 "Unreasonable BitsToClear");
766 // Okay, we can transform this! Insert the new expression now.
767 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
768 " to avoid zero extend: " << CI);
769 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
770 assert(Res->getType() == DestTy);
772 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
773 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
775 // If the high bits are already filled with zeros, just replace this
776 // cast with the result.
777 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
778 DestBitSize-SrcBitsKept)))
779 return ReplaceInstUsesWith(CI, Res);
781 // We need to emit an AND to clear the high bits.
782 Constant *C = ConstantInt::get(Res->getType(),
783 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
784 return BinaryOperator::CreateAnd(Res, C);
787 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
788 // types and if the sizes are just right we can convert this into a logical
789 // 'and' which will be much cheaper than the pair of casts.
790 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
791 // TODO: Subsume this into EvaluateInDifferentType.
793 // Get the sizes of the types involved. We know that the intermediate type
794 // will be smaller than A or C, but don't know the relation between A and C.
795 Value *A = CSrc->getOperand(0);
796 unsigned SrcSize = A->getType()->getScalarSizeInBits();
797 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
798 unsigned DstSize = CI.getType()->getScalarSizeInBits();
799 // If we're actually extending zero bits, then if
800 // SrcSize < DstSize: zext(a & mask)
801 // SrcSize == DstSize: a & mask
802 // SrcSize > DstSize: trunc(a) & mask
803 if (SrcSize < DstSize) {
804 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
805 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
806 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
807 return new ZExtInst(And, CI.getType());
810 if (SrcSize == DstSize) {
811 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
812 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
815 if (SrcSize > DstSize) {
816 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
817 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
818 return BinaryOperator::CreateAnd(Trunc,
819 ConstantInt::get(Trunc->getType(),
824 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
825 return transformZExtICmp(ICI, CI);
827 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
828 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
829 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
830 // of the (zext icmp) will be transformed.
831 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
832 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
833 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
834 (transformZExtICmp(LHS, CI, false) ||
835 transformZExtICmp(RHS, CI, false))) {
836 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
837 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
838 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
842 // zext(trunc(t) & C) -> (t & zext(C)).
843 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
844 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
845 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
846 Value *TI0 = TI->getOperand(0);
847 if (TI0->getType() == CI.getType())
849 BinaryOperator::CreateAnd(TI0,
850 ConstantExpr::getZExt(C, CI.getType()));
853 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
854 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
855 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
856 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
857 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
858 And->getOperand(1) == C)
859 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
860 Value *TI0 = TI->getOperand(0);
861 if (TI0->getType() == CI.getType()) {
862 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
863 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
864 return BinaryOperator::CreateXor(NewAnd, ZC);
868 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
870 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
871 match(SrcI, m_Not(m_Value(X))) &&
872 (!X->hasOneUse() || !isa<CmpInst>(X))) {
873 Value *New = Builder->CreateZExt(X, CI.getType());
874 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
880 /// CanEvaluateSExtd - Return true if we can take the specified value
881 /// and return it as type Ty without inserting any new casts and without
882 /// changing the value of the common low bits. This is used by code that tries
883 /// to promote integer operations to a wider types will allow us to eliminate
886 /// This function works on both vectors and scalars.
888 static bool CanEvaluateSExtd(Value *V, const Type *Ty) {
889 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
890 "Can't sign extend type to a smaller type");
891 // If this is a constant, it can be trivially promoted.
892 if (isa<Constant>(V))
895 Instruction *I = dyn_cast<Instruction>(V);
896 if (!I) return false;
898 // If this is a truncate from the dest type, we can trivially eliminate it,
899 // even if it has multiple uses.
900 // FIXME: This is currently disabled until codegen can handle this without
901 // pessimizing code, PR5997.
902 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
905 // We can't extend or shrink something that has multiple uses: doing so would
906 // require duplicating the instruction in general, which isn't profitable.
907 if (!I->hasOneUse()) return false;
909 switch (I->getOpcode()) {
910 case Instruction::SExt: // sext(sext(x)) -> sext(x)
911 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
912 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
914 case Instruction::And:
915 case Instruction::Or:
916 case Instruction::Xor:
917 case Instruction::Add:
918 case Instruction::Sub:
919 case Instruction::Mul:
920 // These operators can all arbitrarily be extended if their inputs can.
921 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
922 CanEvaluateSExtd(I->getOperand(1), Ty);
924 //case Instruction::Shl: TODO
925 //case Instruction::LShr: TODO
927 case Instruction::Select:
928 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
929 CanEvaluateSExtd(I->getOperand(2), Ty);
931 case Instruction::PHI: {
932 // We can change a phi if we can change all operands. Note that we never
933 // get into trouble with cyclic PHIs here because we only consider
934 // instructions with a single use.
935 PHINode *PN = cast<PHINode>(I);
936 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
937 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
941 // TODO: Can handle more cases here.
948 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
949 // If this sign extend is only used by a truncate, let the truncate by
950 // eliminated before we try to optimize this zext.
951 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
954 if (Instruction *I = commonCastTransforms(CI))
957 // See if we can simplify any instructions used by the input whose sole
958 // purpose is to compute bits we don't care about.
959 if (SimplifyDemandedInstructionBits(CI))
962 Value *Src = CI.getOperand(0);
963 const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
965 // Attempt to extend the entire input expression tree to the destination
966 // type. Only do this if the dest type is a simple type, don't convert the
967 // expression tree to something weird like i93 unless the source is also
969 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
970 CanEvaluateSExtd(Src, DestTy)) {
971 // Okay, we can transform this! Insert the new expression now.
972 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
973 " to avoid sign extend: " << CI);
974 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
975 assert(Res->getType() == DestTy);
977 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
978 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
980 // If the high bits are already filled with sign bit, just replace this
981 // cast with the result.
982 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
983 return ReplaceInstUsesWith(CI, Res);
985 // We need to emit a shl + ashr to do the sign extend.
986 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
987 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
991 // If this input is a trunc from our destination, then turn sext(trunc(x))
993 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
994 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
995 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
996 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
998 // We need to emit a shl + ashr to do the sign extend.
999 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1000 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1001 return BinaryOperator::CreateAShr(Res, ShAmt);
1005 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
1006 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
1008 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS;
1009 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) {
1010 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
1011 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
1012 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) ||
1013 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) {
1014 Value *Sh = ConstantInt::get(CmpLHS->getType(),
1015 CmpLHS->getType()->getScalarSizeInBits()-1);
1016 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit");
1017 if (In->getType() != CI.getType())
1018 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp");
1020 if (Pred == ICmpInst::ICMP_SGT)
1021 In = Builder->CreateNot(In, In->getName()+".not");
1022 return ReplaceInstUsesWith(CI, In);
1027 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
1028 if (const VectorType *VTy = dyn_cast<VectorType>(DestTy)) {
1029 ICmpInst::Predicate Pred; Value *CmpLHS;
1030 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_Zero()))) {
1031 if (Pred == ICmpInst::ICMP_SLT && CmpLHS->getType() == DestTy) {
1032 const Type *EltTy = VTy->getElementType();
1034 // splat the shift constant to a constant vector.
1035 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
1036 Value *In = Builder->CreateAShr(CmpLHS, VSh,CmpLHS->getName()+".lobit");
1037 return ReplaceInstUsesWith(CI, In);
1042 // If the input is a shl/ashr pair of a same constant, then this is a sign
1043 // extension from a smaller value. If we could trust arbitrary bitwidth
1044 // integers, we could turn this into a truncate to the smaller bit and then
1045 // use a sext for the whole extension. Since we don't, look deeper and check
1046 // for a truncate. If the source and dest are the same type, eliminate the
1047 // trunc and extend and just do shifts. For example, turn:
1048 // %a = trunc i32 %i to i8
1049 // %b = shl i8 %a, 6
1050 // %c = ashr i8 %b, 6
1051 // %d = sext i8 %c to i32
1053 // %a = shl i32 %i, 30
1054 // %d = ashr i32 %a, 30
1056 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1057 ConstantInt *BA = 0, *CA = 0;
1058 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1059 m_ConstantInt(CA))) &&
1060 BA == CA && A->getType() == CI.getType()) {
1061 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1062 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1063 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1064 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1065 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1066 return BinaryOperator::CreateAShr(A, ShAmtV);
1073 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1074 /// in the specified FP type without changing its value.
1075 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1077 APFloat F = CFP->getValueAPF();
1078 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1080 return ConstantFP::get(CFP->getContext(), F);
1084 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1085 /// through it until we get the source value.
1086 static Value *LookThroughFPExtensions(Value *V) {
1087 if (Instruction *I = dyn_cast<Instruction>(V))
1088 if (I->getOpcode() == Instruction::FPExt)
1089 return LookThroughFPExtensions(I->getOperand(0));
1091 // If this value is a constant, return the constant in the smallest FP type
1092 // that can accurately represent it. This allows us to turn
1093 // (float)((double)X+2.0) into x+2.0f.
1094 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1095 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1096 return V; // No constant folding of this.
1097 // See if the value can be truncated to float and then reextended.
1098 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1100 if (CFP->getType()->isDoubleTy())
1101 return V; // Won't shrink.
1102 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1104 // Don't try to shrink to various long double types.
1110 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1111 if (Instruction *I = commonCastTransforms(CI))
1114 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1115 // smaller than the destination type, we can eliminate the truncate by doing
1116 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1117 // as many builtins (sqrt, etc).
1118 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1119 if (OpI && OpI->hasOneUse()) {
1120 switch (OpI->getOpcode()) {
1122 case Instruction::FAdd:
1123 case Instruction::FSub:
1124 case Instruction::FMul:
1125 case Instruction::FDiv:
1126 case Instruction::FRem:
1127 const Type *SrcTy = OpI->getType();
1128 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1129 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1130 if (LHSTrunc->getType() != SrcTy &&
1131 RHSTrunc->getType() != SrcTy) {
1132 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1133 // If the source types were both smaller than the destination type of
1134 // the cast, do this xform.
1135 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1136 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1137 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1138 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1139 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1146 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1147 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it.
1148 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1149 if (Call && Call->getCalledFunction() &&
1150 Call->getCalledFunction()->getName() == "sqrt" &&
1151 Call->getNumArgOperands() == 1) {
1152 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1153 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1154 CI.getType()->isFloatTy() &&
1155 Call->getType()->isDoubleTy() &&
1156 Arg->getType()->isDoubleTy() &&
1157 Arg->getOperand(0)->getType()->isFloatTy()) {
1158 Function *Callee = Call->getCalledFunction();
1159 Module *M = CI.getParent()->getParent()->getParent();
1160 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1161 Callee->getAttributes(),
1162 Builder->getFloatTy(),
1163 Builder->getFloatTy(),
1165 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1167 ret->setAttributes(Callee->getAttributes());
1170 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1171 Call->replaceAllUsesWith(UndefValue::get(Call->getType()));
1172 EraseInstFromFunction(*Call);
1180 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1181 return commonCastTransforms(CI);
1184 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1185 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1187 return commonCastTransforms(FI);
1189 // fptoui(uitofp(X)) --> X
1190 // fptoui(sitofp(X)) --> X
1191 // This is safe if the intermediate type has enough bits in its mantissa to
1192 // accurately represent all values of X. For example, do not do this with
1193 // i64->float->i64. This is also safe for sitofp case, because any negative
1194 // 'X' value would cause an undefined result for the fptoui.
1195 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1196 OpI->getOperand(0)->getType() == FI.getType() &&
1197 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1198 OpI->getType()->getFPMantissaWidth())
1199 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1201 return commonCastTransforms(FI);
1204 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1205 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1207 return commonCastTransforms(FI);
1209 // fptosi(sitofp(X)) --> X
1210 // fptosi(uitofp(X)) --> X
1211 // This is safe if the intermediate type has enough bits in its mantissa to
1212 // accurately represent all values of X. For example, do not do this with
1213 // i64->float->i64. This is also safe for sitofp case, because any negative
1214 // 'X' value would cause an undefined result for the fptoui.
1215 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1216 OpI->getOperand(0)->getType() == FI.getType() &&
1217 (int)FI.getType()->getScalarSizeInBits() <=
1218 OpI->getType()->getFPMantissaWidth())
1219 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1221 return commonCastTransforms(FI);
1224 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1225 return commonCastTransforms(CI);
1228 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1229 return commonCastTransforms(CI);
1232 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1233 // If the source integer type is not the intptr_t type for this target, do a
1234 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1235 // cast to be exposed to other transforms.
1237 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
1238 TD->getPointerSizeInBits()) {
1239 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1240 TD->getIntPtrType(CI.getContext()), "tmp");
1241 return new IntToPtrInst(P, CI.getType());
1243 if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
1244 TD->getPointerSizeInBits()) {
1245 Value *P = Builder->CreateZExt(CI.getOperand(0),
1246 TD->getIntPtrType(CI.getContext()), "tmp");
1247 return new IntToPtrInst(P, CI.getType());
1251 if (Instruction *I = commonCastTransforms(CI))
1257 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1258 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1259 Value *Src = CI.getOperand(0);
1261 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1262 // If casting the result of a getelementptr instruction with no offset, turn
1263 // this into a cast of the original pointer!
1264 if (GEP->hasAllZeroIndices()) {
1265 // Changing the cast operand is usually not a good idea but it is safe
1266 // here because the pointer operand is being replaced with another
1267 // pointer operand so the opcode doesn't need to change.
1269 CI.setOperand(0, GEP->getOperand(0));
1273 // If the GEP has a single use, and the base pointer is a bitcast, and the
1274 // GEP computes a constant offset, see if we can convert these three
1275 // instructions into fewer. This typically happens with unions and other
1276 // non-type-safe code.
1277 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1278 GEP->hasAllConstantIndices()) {
1279 // We are guaranteed to get a constant from EmitGEPOffset.
1280 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
1281 int64_t Offset = OffsetV->getSExtValue();
1283 // Get the base pointer input of the bitcast, and the type it points to.
1284 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1285 const Type *GEPIdxTy =
1286 cast<PointerType>(OrigBase->getType())->getElementType();
1287 SmallVector<Value*, 8> NewIndices;
1288 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
1289 // If we were able to index down into an element, create the GEP
1290 // and bitcast the result. This eliminates one bitcast, potentially
1292 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1293 Builder->CreateInBoundsGEP(OrigBase,
1294 NewIndices.begin(), NewIndices.end()) :
1295 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
1296 NGEP->takeName(GEP);
1298 if (isa<BitCastInst>(CI))
1299 return new BitCastInst(NGEP, CI.getType());
1300 assert(isa<PtrToIntInst>(CI));
1301 return new PtrToIntInst(NGEP, CI.getType());
1306 return commonCastTransforms(CI);
1309 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1310 // If the destination integer type is not the intptr_t type for this target,
1311 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1312 // to be exposed to other transforms.
1314 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1315 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1316 TD->getIntPtrType(CI.getContext()),
1318 return new TruncInst(P, CI.getType());
1320 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) {
1321 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1322 TD->getIntPtrType(CI.getContext()),
1324 return new ZExtInst(P, CI.getType());
1328 return commonPointerCastTransforms(CI);
1331 /// OptimizeVectorResize - This input value (which is known to have vector type)
1332 /// is being zero extended or truncated to the specified vector type. Try to
1333 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1335 /// The source and destination vector types may have different element types.
1336 static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy,
1338 // We can only do this optimization if the output is a multiple of the input
1339 // element size, or the input is a multiple of the output element size.
1340 // Convert the input type to have the same element type as the output.
1341 const VectorType *SrcTy = cast<VectorType>(InVal->getType());
1343 if (SrcTy->getElementType() != DestTy->getElementType()) {
1344 // The input types don't need to be identical, but for now they must be the
1345 // same size. There is no specific reason we couldn't handle things like
1346 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1348 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1349 DestTy->getElementType()->getPrimitiveSizeInBits())
1352 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1353 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1356 // Now that the element types match, get the shuffle mask and RHS of the
1357 // shuffle to use, which depends on whether we're increasing or decreasing the
1358 // size of the input.
1359 SmallVector<Constant*, 16> ShuffleMask;
1361 const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext());
1363 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1364 // If we're shrinking the number of elements, just shuffle in the low
1365 // elements from the input and use undef as the second shuffle input.
1366 V2 = UndefValue::get(SrcTy);
1367 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1368 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1371 // If we're increasing the number of elements, shuffle in all of the
1372 // elements from InVal and fill the rest of the result elements with zeros
1373 // from a constant zero.
1374 V2 = Constant::getNullValue(SrcTy);
1375 unsigned SrcElts = SrcTy->getNumElements();
1376 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1377 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1379 // The excess elements reference the first element of the zero input.
1380 ShuffleMask.append(DestTy->getNumElements()-SrcElts,
1381 ConstantInt::get(Int32Ty, SrcElts));
1384 return new ShuffleVectorInst(InVal, V2, ConstantVector::get(ShuffleMask));
1387 static bool isMultipleOfTypeSize(unsigned Value, const Type *Ty) {
1388 return Value % Ty->getPrimitiveSizeInBits() == 0;
1391 static unsigned getTypeSizeIndex(unsigned Value, const Type *Ty) {
1392 return Value / Ty->getPrimitiveSizeInBits();
1395 /// CollectInsertionElements - V is a value which is inserted into a vector of
1396 /// VecEltTy. Look through the value to see if we can decompose it into
1397 /// insertions into the vector. See the example in the comment for
1398 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1399 /// The type of V is always a non-zero multiple of VecEltTy's size.
1401 /// This returns false if the pattern can't be matched or true if it can,
1402 /// filling in Elements with the elements found here.
1403 static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1404 SmallVectorImpl<Value*> &Elements,
1405 const Type *VecEltTy) {
1406 // Undef values never contribute useful bits to the result.
1407 if (isa<UndefValue>(V)) return true;
1409 // If we got down to a value of the right type, we win, try inserting into the
1411 if (V->getType() == VecEltTy) {
1412 // Inserting null doesn't actually insert any elements.
1413 if (Constant *C = dyn_cast<Constant>(V))
1414 if (C->isNullValue())
1417 // Fail if multiple elements are inserted into this slot.
1418 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1421 Elements[ElementIndex] = V;
1425 if (Constant *C = dyn_cast<Constant>(V)) {
1426 // Figure out the # elements this provides, and bitcast it or slice it up
1428 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1430 // If the constant is the size of a vector element, we just need to bitcast
1431 // it to the right type so it gets properly inserted.
1433 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1434 ElementIndex, Elements, VecEltTy);
1436 // Okay, this is a constant that covers multiple elements. Slice it up into
1437 // pieces and insert each element-sized piece into the vector.
1438 if (!isa<IntegerType>(C->getType()))
1439 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1440 C->getType()->getPrimitiveSizeInBits()));
1441 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1442 const Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1444 for (unsigned i = 0; i != NumElts; ++i) {
1445 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1447 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1448 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1454 if (!V->hasOneUse()) return false;
1456 Instruction *I = dyn_cast<Instruction>(V);
1457 if (I == 0) return false;
1458 switch (I->getOpcode()) {
1459 default: return false; // Unhandled case.
1460 case Instruction::BitCast:
1461 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1462 Elements, VecEltTy);
1463 case Instruction::ZExt:
1464 if (!isMultipleOfTypeSize(
1465 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1468 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1469 Elements, VecEltTy);
1470 case Instruction::Or:
1471 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1472 Elements, VecEltTy) &&
1473 CollectInsertionElements(I->getOperand(1), ElementIndex,
1474 Elements, VecEltTy);
1475 case Instruction::Shl: {
1476 // Must be shifting by a constant that is a multiple of the element size.
1477 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1478 if (CI == 0) return false;
1479 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1480 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1482 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1483 Elements, VecEltTy);
1490 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1491 /// may be doing shifts and ors to assemble the elements of the vector manually.
1492 /// Try to rip the code out and replace it with insertelements. This is to
1493 /// optimize code like this:
1495 /// %tmp37 = bitcast float %inc to i32
1496 /// %tmp38 = zext i32 %tmp37 to i64
1497 /// %tmp31 = bitcast float %inc5 to i32
1498 /// %tmp32 = zext i32 %tmp31 to i64
1499 /// %tmp33 = shl i64 %tmp32, 32
1500 /// %ins35 = or i64 %tmp33, %tmp38
1501 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1503 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1504 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1506 const VectorType *DestVecTy = cast<VectorType>(CI.getType());
1507 Value *IntInput = CI.getOperand(0);
1509 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1510 if (!CollectInsertionElements(IntInput, 0, Elements,
1511 DestVecTy->getElementType()))
1514 // If we succeeded, we know that all of the element are specified by Elements
1515 // or are zero if Elements has a null entry. Recast this as a set of
1517 Value *Result = Constant::getNullValue(CI.getType());
1518 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1519 if (Elements[i] == 0) continue; // Unset element.
1521 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1522 IC.Builder->getInt32(i));
1529 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1530 /// bitcast. The various long double bitcasts can't get in here.
1531 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1532 Value *Src = CI.getOperand(0);
1533 const Type *DestTy = CI.getType();
1535 // If this is a bitcast from int to float, check to see if the int is an
1536 // extraction from a vector.
1537 Value *VecInput = 0;
1538 // bitcast(trunc(bitcast(somevector)))
1539 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1540 isa<VectorType>(VecInput->getType())) {
1541 const VectorType *VecTy = cast<VectorType>(VecInput->getType());
1542 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1544 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1545 // If the element type of the vector doesn't match the result type,
1546 // bitcast it to be a vector type we can extract from.
1547 if (VecTy->getElementType() != DestTy) {
1548 VecTy = VectorType::get(DestTy,
1549 VecTy->getPrimitiveSizeInBits() / DestWidth);
1550 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1553 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
1557 // bitcast(trunc(lshr(bitcast(somevector), cst))
1558 ConstantInt *ShAmt = 0;
1559 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1560 m_ConstantInt(ShAmt)))) &&
1561 isa<VectorType>(VecInput->getType())) {
1562 const VectorType *VecTy = cast<VectorType>(VecInput->getType());
1563 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1564 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1565 ShAmt->getZExtValue() % DestWidth == 0) {
1566 // If the element type of the vector doesn't match the result type,
1567 // bitcast it to be a vector type we can extract from.
1568 if (VecTy->getElementType() != DestTy) {
1569 VecTy = VectorType::get(DestTy,
1570 VecTy->getPrimitiveSizeInBits() / DestWidth);
1571 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1574 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1575 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1581 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1582 // If the operands are integer typed then apply the integer transforms,
1583 // otherwise just apply the common ones.
1584 Value *Src = CI.getOperand(0);
1585 const Type *SrcTy = Src->getType();
1586 const Type *DestTy = CI.getType();
1588 // Get rid of casts from one type to the same type. These are useless and can
1589 // be replaced by the operand.
1590 if (DestTy == Src->getType())
1591 return ReplaceInstUsesWith(CI, Src);
1593 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1594 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1595 const Type *DstElTy = DstPTy->getElementType();
1596 const Type *SrcElTy = SrcPTy->getElementType();
1598 // If the address spaces don't match, don't eliminate the bitcast, which is
1599 // required for changing types.
1600 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1603 // If we are casting a alloca to a pointer to a type of the same
1604 // size, rewrite the allocation instruction to allocate the "right" type.
1605 // There is no need to modify malloc calls because it is their bitcast that
1606 // needs to be cleaned up.
1607 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1608 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1611 // If the source and destination are pointers, and this cast is equivalent
1612 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1613 // This can enhance SROA and other transforms that want type-safe pointers.
1614 Constant *ZeroUInt =
1615 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1616 unsigned NumZeros = 0;
1617 while (SrcElTy != DstElTy &&
1618 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1619 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1620 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1624 // If we found a path from the src to dest, create the getelementptr now.
1625 if (SrcElTy == DstElTy) {
1626 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1627 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1628 ((Instruction*)NULL));
1632 // Try to optimize int -> float bitcasts.
1633 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1634 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1637 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1638 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1639 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1640 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1641 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1642 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1645 if (isa<IntegerType>(SrcTy)) {
1646 // If this is a cast from an integer to vector, check to see if the input
1647 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1648 // the casts with a shuffle and (potentially) a bitcast.
1649 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1650 CastInst *SrcCast = cast<CastInst>(Src);
1651 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1652 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1653 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1654 cast<VectorType>(DestTy), *this))
1658 // If the input is an 'or' instruction, we may be doing shifts and ors to
1659 // assemble the elements of the vector manually. Try to rip the code out
1660 // and replace it with insertelements.
1661 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1662 return ReplaceInstUsesWith(CI, V);
1666 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1667 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
1669 Builder->CreateExtractElement(Src,
1670 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1671 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1675 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1676 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1677 // a bitcast to a vector with the same # elts.
1678 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1679 cast<VectorType>(DestTy)->getNumElements() ==
1680 SVI->getType()->getNumElements() &&
1681 SVI->getType()->getNumElements() ==
1682 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1684 // If either of the operands is a cast from CI.getType(), then
1685 // evaluating the shuffle in the casted destination's type will allow
1686 // us to eliminate at least one cast.
1687 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1688 Tmp->getOperand(0)->getType() == DestTy) ||
1689 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1690 Tmp->getOperand(0)->getType() == DestTy)) {
1691 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1692 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1693 // Return a new shuffle vector. Use the same element ID's, as we
1694 // know the vector types match #elts.
1695 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1700 if (SrcTy->isPointerTy())
1701 return commonPointerCastTransforms(CI);
1702 return commonCastTransforms(CI);