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. (A reference
91 // from a dbg.declare doesn't count as a use for this purpose.)
92 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
93 CastElTyAlign == AllocElTyAlign) return 0;
95 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
96 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
97 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
99 // See if we can satisfy the modulus by pulling a scale out of the array
101 unsigned ArraySizeScale;
102 uint64_t ArrayOffset;
103 Value *NumElements = // See if the array size is a decomposable linear expr.
104 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
106 // If we can now satisfy the modulus, by using a non-1 scale, we really can
108 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
109 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
111 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
116 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
117 // Insert before the alloca, not before the cast.
118 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
121 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
122 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
124 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
127 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
128 New->setAlignment(AI.getAlignment());
131 // If the allocation has one real use plus a dbg.declare, just remove the
133 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
134 EraseInstFromFunction(*(Instruction*)DI);
136 // If the allocation has multiple real uses, insert a cast and change all
137 // things that used it to use the new cast. This will also hack on CI, but it
139 else if (!AI.hasOneUse()) {
140 // New is the allocation instruction, pointer typed. AI is the original
141 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
142 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
143 AI.replaceAllUsesWith(NewCast);
145 return ReplaceInstUsesWith(CI, New);
150 /// EvaluateInDifferentType - Given an expression that
151 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
152 /// insert the code to evaluate the expression.
153 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
155 if (Constant *C = dyn_cast<Constant>(V)) {
156 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
157 // If we got a constantexpr back, try to simplify it with TD info.
158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
159 C = ConstantFoldConstantExpression(CE, TD);
163 // Otherwise, it must be an instruction.
164 Instruction *I = cast<Instruction>(V);
165 Instruction *Res = 0;
166 unsigned Opc = I->getOpcode();
168 case Instruction::Add:
169 case Instruction::Sub:
170 case Instruction::Mul:
171 case Instruction::And:
172 case Instruction::Or:
173 case Instruction::Xor:
174 case Instruction::AShr:
175 case Instruction::LShr:
176 case Instruction::Shl:
177 case Instruction::UDiv:
178 case Instruction::URem: {
179 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
180 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
181 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
184 case Instruction::Trunc:
185 case Instruction::ZExt:
186 case Instruction::SExt:
187 // If the source type of the cast is the type we're trying for then we can
188 // just return the source. There's no need to insert it because it is not
190 if (I->getOperand(0)->getType() == Ty)
191 return I->getOperand(0);
193 // Otherwise, must be the same type of cast, so just reinsert a new one.
194 // This also handles the case of zext(trunc(x)) -> zext(x).
195 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
196 Opc == Instruction::SExt);
198 case Instruction::Select: {
199 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
200 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
201 Res = SelectInst::Create(I->getOperand(0), True, False);
204 case Instruction::PHI: {
205 PHINode *OPN = cast<PHINode>(I);
206 PHINode *NPN = PHINode::Create(Ty);
207 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
208 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
209 NPN->addIncoming(V, OPN->getIncomingBlock(i));
215 // TODO: Can handle more cases here.
216 llvm_unreachable("Unreachable!");
221 return InsertNewInstBefore(Res, *I);
225 /// This function is a wrapper around CastInst::isEliminableCastPair. It
226 /// simply extracts arguments and returns what that function returns.
227 static Instruction::CastOps
228 isEliminableCastPair(
229 const CastInst *CI, ///< The first cast instruction
230 unsigned opcode, ///< The opcode of the second cast instruction
231 const Type *DstTy, ///< The target type for the second cast instruction
232 TargetData *TD ///< The target data for pointer size
235 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
236 const Type *MidTy = CI->getType(); // B from above
238 // Get the opcodes of the two Cast instructions
239 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
240 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
242 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
244 TD ? TD->getIntPtrType(CI->getContext()) : 0);
246 // We don't want to form an inttoptr or ptrtoint that converts to an integer
247 // type that differs from the pointer size.
248 if ((Res == Instruction::IntToPtr &&
249 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
250 (Res == Instruction::PtrToInt &&
251 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
254 return Instruction::CastOps(Res);
257 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
258 /// results in any code being generated and is interesting to optimize out. If
259 /// the cast can be eliminated by some other simple transformation, we prefer
260 /// to do the simplification first.
261 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
263 // Noop casts and casts of constants should be eliminated trivially.
264 if (V->getType() == Ty || isa<Constant>(V)) return false;
266 // If this is another cast that can be eliminated, we prefer to have it
268 if (const CastInst *CI = dyn_cast<CastInst>(V))
269 if (isEliminableCastPair(CI, opc, Ty, TD))
272 // If this is a vector sext from a compare, then we don't want to break the
273 // idiom where each element of the extended vector is either zero or all ones.
274 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
281 /// @brief Implement the transforms common to all CastInst visitors.
282 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
283 Value *Src = CI.getOperand(0);
285 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
287 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
288 if (Instruction::CastOps opc =
289 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
290 // The first cast (CSrc) is eliminable so we need to fix up or replace
291 // the second cast (CI). CSrc will then have a good chance of being dead.
292 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
296 // If we are casting a select then fold the cast into the select
297 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
298 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
301 // If we are casting a PHI then fold the cast into the PHI
302 if (isa<PHINode>(Src)) {
303 // We don't do this if this would create a PHI node with an illegal type if
304 // it is currently legal.
305 if (!Src->getType()->isIntegerTy() ||
306 !CI.getType()->isIntegerTy() ||
307 ShouldChangeType(CI.getType(), Src->getType()))
308 if (Instruction *NV = FoldOpIntoPhi(CI))
315 /// CanEvaluateTruncated - Return true if we can evaluate the specified
316 /// expression tree as type Ty instead of its larger type, and arrive with the
317 /// same value. This is used by code that tries to eliminate truncates.
319 /// Ty will always be a type smaller than V. We should return true if trunc(V)
320 /// can be computed by computing V in the smaller type. If V is an instruction,
321 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
322 /// makes sense if x and y can be efficiently truncated.
324 /// This function works on both vectors and scalars.
326 static bool CanEvaluateTruncated(Value *V, const Type *Ty) {
327 // We can always evaluate constants in another type.
328 if (isa<Constant>(V))
331 Instruction *I = dyn_cast<Instruction>(V);
332 if (!I) return false;
334 const Type *OrigTy = V->getType();
336 // If this is an extension from the dest type, we can eliminate it, even if it
337 // has multiple uses.
338 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
339 I->getOperand(0)->getType() == Ty)
342 // We can't extend or shrink something that has multiple uses: doing so would
343 // require duplicating the instruction in general, which isn't profitable.
344 if (!I->hasOneUse()) return false;
346 unsigned Opc = I->getOpcode();
348 case Instruction::Add:
349 case Instruction::Sub:
350 case Instruction::Mul:
351 case Instruction::And:
352 case Instruction::Or:
353 case Instruction::Xor:
354 // These operators can all arbitrarily be extended or truncated.
355 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
356 CanEvaluateTruncated(I->getOperand(1), Ty);
358 case Instruction::UDiv:
359 case Instruction::URem: {
360 // UDiv and URem can be truncated if all the truncated bits are zero.
361 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
362 uint32_t BitWidth = Ty->getScalarSizeInBits();
363 if (BitWidth < OrigBitWidth) {
364 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
365 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
366 MaskedValueIsZero(I->getOperand(1), Mask)) {
367 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
368 CanEvaluateTruncated(I->getOperand(1), Ty);
373 case Instruction::Shl:
374 // If we are truncating the result of this SHL, and if it's a shift of a
375 // constant amount, we can always perform a SHL in a smaller type.
376 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
377 uint32_t BitWidth = Ty->getScalarSizeInBits();
378 if (CI->getLimitedValue(BitWidth) < BitWidth)
379 return CanEvaluateTruncated(I->getOperand(0), Ty);
382 case Instruction::LShr:
383 // If this is a truncate of a logical shr, we can truncate it to a smaller
384 // lshr iff we know that the bits we would otherwise be shifting in are
386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
387 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
388 uint32_t BitWidth = Ty->getScalarSizeInBits();
389 if (MaskedValueIsZero(I->getOperand(0),
390 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
391 CI->getLimitedValue(BitWidth) < BitWidth) {
392 return CanEvaluateTruncated(I->getOperand(0), Ty);
396 case Instruction::Trunc:
397 // trunc(trunc(x)) -> trunc(x)
399 case Instruction::ZExt:
400 case Instruction::SExt:
401 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
402 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
404 case Instruction::Select: {
405 SelectInst *SI = cast<SelectInst>(I);
406 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
407 CanEvaluateTruncated(SI->getFalseValue(), Ty);
409 case Instruction::PHI: {
410 // We can change a phi if we can change all operands. Note that we never
411 // get into trouble with cyclic PHIs here because we only consider
412 // instructions with a single use.
413 PHINode *PN = cast<PHINode>(I);
414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
415 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
420 // TODO: Can handle more cases here.
427 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
428 if (Instruction *Result = commonCastTransforms(CI))
431 // See if we can simplify any instructions used by the input whose sole
432 // purpose is to compute bits we don't care about.
433 if (SimplifyDemandedInstructionBits(CI))
436 Value *Src = CI.getOperand(0);
437 const Type *DestTy = CI.getType(), *SrcTy = Src->getType();
439 // Attempt to truncate the entire input expression tree to the destination
440 // type. Only do this if the dest type is a simple type, don't convert the
441 // expression tree to something weird like i93 unless the source is also
443 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
444 CanEvaluateTruncated(Src, DestTy)) {
446 // If this cast is a truncate, evaluting in a different type always
447 // eliminates the cast, so it is always a win.
448 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
449 " to avoid cast: " << CI << '\n');
450 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
451 assert(Res->getType() == DestTy);
452 return ReplaceInstUsesWith(CI, Res);
455 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
456 if (DestTy->getScalarSizeInBits() == 1) {
457 Constant *One = ConstantInt::get(Src->getType(), 1);
458 Src = Builder->CreateAnd(Src, One, "tmp");
459 Value *Zero = Constant::getNullValue(Src->getType());
460 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
463 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
464 Value *A = 0; ConstantInt *Cst = 0;
465 if (Src->hasOneUse() &&
466 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
467 // We have three types to worry about here, the type of A, the source of
468 // the truncate (MidSize), and the destination of the truncate. We know that
469 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
470 // between ASize and ResultSize.
471 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
473 // If the shift amount is larger than the size of A, then the result is
474 // known to be zero because all the input bits got shifted out.
475 if (Cst->getZExtValue() >= ASize)
476 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
478 // Since we're doing an lshr and a zero extend, and know that the shift
479 // amount is smaller than ASize, it is always safe to do the shift in A's
480 // type, then zero extend or truncate to the result.
481 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
482 Shift->takeName(Src);
483 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
486 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
487 // type isn't non-native.
488 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
489 ShouldChangeType(Src->getType(), CI.getType()) &&
490 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
491 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
492 return BinaryOperator::CreateAnd(NewTrunc,
493 ConstantExpr::getTrunc(Cst, CI.getType()));
499 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
500 /// in order to eliminate the icmp.
501 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
503 // If we are just checking for a icmp eq of a single bit and zext'ing it
504 // to an integer, then shift the bit to the appropriate place and then
505 // cast to integer to avoid the comparison.
506 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
507 const APInt &Op1CV = Op1C->getValue();
509 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
510 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
511 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
512 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
513 if (!DoXform) return ICI;
515 Value *In = ICI->getOperand(0);
516 Value *Sh = ConstantInt::get(In->getType(),
517 In->getType()->getScalarSizeInBits()-1);
518 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
519 if (In->getType() != CI.getType())
520 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
522 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
523 Constant *One = ConstantInt::get(In->getType(), 1);
524 In = Builder->CreateXor(In, One, In->getName()+".not");
527 return ReplaceInstUsesWith(CI, In);
532 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
533 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
534 // zext (X == 1) to i32 --> X iff X has only the low bit set.
535 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
536 // zext (X != 0) to i32 --> X iff X has only the low bit set.
537 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
538 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
539 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
540 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
541 // This only works for EQ and NE
543 // If Op1C some other power of two, convert:
544 uint32_t BitWidth = Op1C->getType()->getBitWidth();
545 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
546 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
547 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
549 APInt KnownZeroMask(~KnownZero);
550 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
551 if (!DoXform) return ICI;
553 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
554 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
555 // (X&4) == 2 --> false
556 // (X&4) != 2 --> true
557 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
559 Res = ConstantExpr::getZExt(Res, CI.getType());
560 return ReplaceInstUsesWith(CI, Res);
563 uint32_t ShiftAmt = KnownZeroMask.logBase2();
564 Value *In = ICI->getOperand(0);
566 // Perform a logical shr by shiftamt.
567 // Insert the shift to put the result in the low bit.
568 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
569 In->getName()+".lobit");
572 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
573 Constant *One = ConstantInt::get(In->getType(), 1);
574 In = Builder->CreateXor(In, One, "tmp");
577 if (CI.getType() == In->getType())
578 return ReplaceInstUsesWith(CI, In);
579 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
584 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
585 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
586 // may lead to additional simplifications.
587 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
588 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
589 uint32_t BitWidth = ITy->getBitWidth();
590 Value *LHS = ICI->getOperand(0);
591 Value *RHS = ICI->getOperand(1);
593 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
594 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
595 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
596 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
597 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
599 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
600 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
601 APInt UnknownBit = ~KnownBits;
602 if (UnknownBit.countPopulation() == 1) {
603 if (!DoXform) return ICI;
605 Value *Result = Builder->CreateXor(LHS, RHS);
607 // Mask off any bits that are set and won't be shifted away.
608 if (KnownOneLHS.uge(UnknownBit))
609 Result = Builder->CreateAnd(Result,
610 ConstantInt::get(ITy, UnknownBit));
612 // Shift the bit we're testing down to the lsb.
613 Result = Builder->CreateLShr(
614 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
616 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
617 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
618 Result->takeName(ICI);
619 return ReplaceInstUsesWith(CI, Result);
628 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
629 /// specified wider type and produce the same low bits. If not, return false.
631 /// If this function returns true, it can also return a non-zero number of bits
632 /// (in BitsToClear) which indicates that the value it computes is correct for
633 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
634 /// out. For example, to promote something like:
636 /// %B = trunc i64 %A to i32
637 /// %C = lshr i32 %B, 8
638 /// %E = zext i32 %C to i64
640 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
641 /// set to 8 to indicate that the promoted value needs to have bits 24-31
642 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
643 /// clear the top bits anyway, doing this has no extra cost.
645 /// This function works on both vectors and scalars.
646 static bool CanEvaluateZExtd(Value *V, const Type *Ty, unsigned &BitsToClear) {
648 if (isa<Constant>(V))
651 Instruction *I = dyn_cast<Instruction>(V);
652 if (!I) return false;
654 // If the input is a truncate from the destination type, we can trivially
655 // eliminate it, even if it has multiple uses.
656 // FIXME: This is currently disabled until codegen can handle this without
657 // pessimizing code, PR5997.
658 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
661 // We can't extend or shrink something that has multiple uses: doing so would
662 // require duplicating the instruction in general, which isn't profitable.
663 if (!I->hasOneUse()) return false;
665 unsigned Opc = I->getOpcode(), Tmp;
667 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
668 case Instruction::SExt: // zext(sext(x)) -> sext(x).
669 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
671 case Instruction::And:
672 case Instruction::Or:
673 case Instruction::Xor:
674 case Instruction::Add:
675 case Instruction::Sub:
676 case Instruction::Mul:
677 case Instruction::Shl:
678 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
679 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
681 // These can all be promoted if neither operand has 'bits to clear'.
682 if (BitsToClear == 0 && Tmp == 0)
685 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
686 // other side, BitsToClear is ok.
688 (Opc == Instruction::And || Opc == Instruction::Or ||
689 Opc == Instruction::Xor)) {
690 // We use MaskedValueIsZero here for generality, but the case we care
691 // about the most is constant RHS.
692 unsigned VSize = V->getType()->getScalarSizeInBits();
693 if (MaskedValueIsZero(I->getOperand(1),
694 APInt::getHighBitsSet(VSize, BitsToClear)))
698 // Otherwise, we don't know how to analyze this BitsToClear case yet.
701 case Instruction::LShr:
702 // We can promote lshr(x, cst) if we can promote x. This requires the
703 // ultimate 'and' to clear out the high zero bits we're clearing out though.
704 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
705 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
707 BitsToClear += Amt->getZExtValue();
708 if (BitsToClear > V->getType()->getScalarSizeInBits())
709 BitsToClear = V->getType()->getScalarSizeInBits();
712 // Cannot promote variable LSHR.
714 case Instruction::Select:
715 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
716 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
717 // TODO: If important, we could handle the case when the BitsToClear are
718 // known zero in the disagreeing side.
723 case Instruction::PHI: {
724 // We can change a phi if we can change all operands. Note that we never
725 // get into trouble with cyclic PHIs here because we only consider
726 // instructions with a single use.
727 PHINode *PN = cast<PHINode>(I);
728 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
730 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
731 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
732 // TODO: If important, we could handle the case when the BitsToClear
733 // are known zero in the disagreeing input.
739 // TODO: Can handle more cases here.
744 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
745 // If this zero extend is only used by a truncate, let the truncate by
746 // eliminated before we try to optimize this zext.
747 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
750 // If one of the common conversion will work, do it.
751 if (Instruction *Result = commonCastTransforms(CI))
754 // See if we can simplify any instructions used by the input whose sole
755 // purpose is to compute bits we don't care about.
756 if (SimplifyDemandedInstructionBits(CI))
759 Value *Src = CI.getOperand(0);
760 const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
762 // Attempt to extend the entire input expression tree to the destination
763 // type. Only do this if the dest type is a simple type, don't convert the
764 // expression tree to something weird like i93 unless the source is also
766 unsigned BitsToClear;
767 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
768 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
769 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
770 "Unreasonable BitsToClear");
772 // Okay, we can transform this! Insert the new expression now.
773 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
774 " to avoid zero extend: " << CI);
775 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
776 assert(Res->getType() == DestTy);
778 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
779 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
781 // If the high bits are already filled with zeros, just replace this
782 // cast with the result.
783 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
784 DestBitSize-SrcBitsKept)))
785 return ReplaceInstUsesWith(CI, Res);
787 // We need to emit an AND to clear the high bits.
788 Constant *C = ConstantInt::get(Res->getType(),
789 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
790 return BinaryOperator::CreateAnd(Res, C);
793 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
794 // types and if the sizes are just right we can convert this into a logical
795 // 'and' which will be much cheaper than the pair of casts.
796 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
797 // TODO: Subsume this into EvaluateInDifferentType.
799 // Get the sizes of the types involved. We know that the intermediate type
800 // will be smaller than A or C, but don't know the relation between A and C.
801 Value *A = CSrc->getOperand(0);
802 unsigned SrcSize = A->getType()->getScalarSizeInBits();
803 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
804 unsigned DstSize = CI.getType()->getScalarSizeInBits();
805 // If we're actually extending zero bits, then if
806 // SrcSize < DstSize: zext(a & mask)
807 // SrcSize == DstSize: a & mask
808 // SrcSize > DstSize: trunc(a) & mask
809 if (SrcSize < DstSize) {
810 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
811 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
812 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
813 return new ZExtInst(And, CI.getType());
816 if (SrcSize == DstSize) {
817 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
818 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
821 if (SrcSize > DstSize) {
822 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
823 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
824 return BinaryOperator::CreateAnd(Trunc,
825 ConstantInt::get(Trunc->getType(),
830 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
831 return transformZExtICmp(ICI, CI);
833 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
834 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
835 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
836 // of the (zext icmp) will be transformed.
837 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
838 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
839 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
840 (transformZExtICmp(LHS, CI, false) ||
841 transformZExtICmp(RHS, CI, false))) {
842 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
843 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
844 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
848 // zext(trunc(t) & C) -> (t & zext(C)).
849 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
850 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
851 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
852 Value *TI0 = TI->getOperand(0);
853 if (TI0->getType() == CI.getType())
855 BinaryOperator::CreateAnd(TI0,
856 ConstantExpr::getZExt(C, CI.getType()));
859 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
860 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
861 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
862 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
863 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
864 And->getOperand(1) == C)
865 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
866 Value *TI0 = TI->getOperand(0);
867 if (TI0->getType() == CI.getType()) {
868 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
869 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
870 return BinaryOperator::CreateXor(NewAnd, ZC);
874 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
876 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
877 match(SrcI, m_Not(m_Value(X))) &&
878 (!X->hasOneUse() || !isa<CmpInst>(X))) {
879 Value *New = Builder->CreateZExt(X, CI.getType());
880 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
886 /// CanEvaluateSExtd - Return true if we can take the specified value
887 /// and return it as type Ty without inserting any new casts and without
888 /// changing the value of the common low bits. This is used by code that tries
889 /// to promote integer operations to a wider types will allow us to eliminate
892 /// This function works on both vectors and scalars.
894 static bool CanEvaluateSExtd(Value *V, const Type *Ty) {
895 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
896 "Can't sign extend type to a smaller type");
897 // If this is a constant, it can be trivially promoted.
898 if (isa<Constant>(V))
901 Instruction *I = dyn_cast<Instruction>(V);
902 if (!I) return false;
904 // If this is a truncate from the dest type, we can trivially eliminate it,
905 // even if it has multiple uses.
906 // FIXME: This is currently disabled until codegen can handle this without
907 // pessimizing code, PR5997.
908 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
911 // We can't extend or shrink something that has multiple uses: doing so would
912 // require duplicating the instruction in general, which isn't profitable.
913 if (!I->hasOneUse()) return false;
915 switch (I->getOpcode()) {
916 case Instruction::SExt: // sext(sext(x)) -> sext(x)
917 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
918 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
920 case Instruction::And:
921 case Instruction::Or:
922 case Instruction::Xor:
923 case Instruction::Add:
924 case Instruction::Sub:
925 case Instruction::Mul:
926 // These operators can all arbitrarily be extended if their inputs can.
927 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
928 CanEvaluateSExtd(I->getOperand(1), Ty);
930 //case Instruction::Shl: TODO
931 //case Instruction::LShr: TODO
933 case Instruction::Select:
934 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
935 CanEvaluateSExtd(I->getOperand(2), Ty);
937 case Instruction::PHI: {
938 // We can change a phi if we can change all operands. Note that we never
939 // get into trouble with cyclic PHIs here because we only consider
940 // instructions with a single use.
941 PHINode *PN = cast<PHINode>(I);
942 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
943 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
947 // TODO: Can handle more cases here.
954 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
955 // If this sign extend is only used by a truncate, let the truncate by
956 // eliminated before we try to optimize this zext.
957 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
960 if (Instruction *I = commonCastTransforms(CI))
963 // See if we can simplify any instructions used by the input whose sole
964 // purpose is to compute bits we don't care about.
965 if (SimplifyDemandedInstructionBits(CI))
968 Value *Src = CI.getOperand(0);
969 const Type *SrcTy = Src->getType(), *DestTy = CI.getType();
971 // Attempt to extend the entire input expression tree to the destination
972 // type. Only do this if the dest type is a simple type, don't convert the
973 // expression tree to something weird like i93 unless the source is also
975 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
976 CanEvaluateSExtd(Src, DestTy)) {
977 // Okay, we can transform this! Insert the new expression now.
978 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
979 " to avoid sign extend: " << CI);
980 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
981 assert(Res->getType() == DestTy);
983 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
984 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
986 // If the high bits are already filled with sign bit, just replace this
987 // cast with the result.
988 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
989 return ReplaceInstUsesWith(CI, Res);
991 // We need to emit a shl + ashr to do the sign extend.
992 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
993 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
997 // If this input is a trunc from our destination, then turn sext(trunc(x))
999 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1000 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1001 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1002 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1004 // We need to emit a shl + ashr to do the sign extend.
1005 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1006 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1007 return BinaryOperator::CreateAShr(Res, ShAmt);
1011 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
1012 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
1014 ICmpInst::Predicate Pred; Value *CmpLHS; ConstantInt *CmpRHS;
1015 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_ConstantInt(CmpRHS)))) {
1016 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
1017 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
1018 if ((Pred == ICmpInst::ICMP_SLT && CmpRHS->isZero()) ||
1019 (Pred == ICmpInst::ICMP_SGT && CmpRHS->isAllOnesValue())) {
1020 Value *Sh = ConstantInt::get(CmpLHS->getType(),
1021 CmpLHS->getType()->getScalarSizeInBits()-1);
1022 Value *In = Builder->CreateAShr(CmpLHS, Sh, CmpLHS->getName()+".lobit");
1023 if (In->getType() != CI.getType())
1024 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp");
1026 if (Pred == ICmpInst::ICMP_SGT)
1027 In = Builder->CreateNot(In, In->getName()+".not");
1028 return ReplaceInstUsesWith(CI, In);
1033 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
1034 if (const VectorType *VTy = dyn_cast<VectorType>(DestTy)) {
1035 ICmpInst::Predicate Pred; Value *CmpLHS;
1036 if (match(Src, m_ICmp(Pred, m_Value(CmpLHS), m_Zero()))) {
1037 if (Pred == ICmpInst::ICMP_SLT && CmpLHS->getType() == DestTy) {
1038 const Type *EltTy = VTy->getElementType();
1040 // splat the shift constant to a cosntant vector
1041 Constant *Sh = ConstantInt::get(EltTy, EltTy->getScalarSizeInBits()-1);
1042 std::vector<Constant *> Elts(VTy->getNumElements(), Sh);
1043 Constant *VSh = ConstantVector::get(Elts);
1045 Value *In = Builder->CreateAShr(CmpLHS, VSh,CmpLHS->getName()+".lobit");
1046 return ReplaceInstUsesWith(CI, In);
1051 // If the input is a shl/ashr pair of a same constant, then this is a sign
1052 // extension from a smaller value. If we could trust arbitrary bitwidth
1053 // integers, we could turn this into a truncate to the smaller bit and then
1054 // use a sext for the whole extension. Since we don't, look deeper and check
1055 // for a truncate. If the source and dest are the same type, eliminate the
1056 // trunc and extend and just do shifts. For example, turn:
1057 // %a = trunc i32 %i to i8
1058 // %b = shl i8 %a, 6
1059 // %c = ashr i8 %b, 6
1060 // %d = sext i8 %c to i32
1062 // %a = shl i32 %i, 30
1063 // %d = ashr i32 %a, 30
1065 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1066 ConstantInt *BA = 0, *CA = 0;
1067 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1068 m_ConstantInt(CA))) &&
1069 BA == CA && A->getType() == CI.getType()) {
1070 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1071 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1072 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1073 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1074 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1075 return BinaryOperator::CreateAShr(A, ShAmtV);
1082 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1083 /// in the specified FP type without changing its value.
1084 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1086 APFloat F = CFP->getValueAPF();
1087 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1089 return ConstantFP::get(CFP->getContext(), F);
1093 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1094 /// through it until we get the source value.
1095 static Value *LookThroughFPExtensions(Value *V) {
1096 if (Instruction *I = dyn_cast<Instruction>(V))
1097 if (I->getOpcode() == Instruction::FPExt)
1098 return LookThroughFPExtensions(I->getOperand(0));
1100 // If this value is a constant, return the constant in the smallest FP type
1101 // that can accurately represent it. This allows us to turn
1102 // (float)((double)X+2.0) into x+2.0f.
1103 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1104 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1105 return V; // No constant folding of this.
1106 // See if the value can be truncated to float and then reextended.
1107 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1109 if (CFP->getType()->isDoubleTy())
1110 return V; // Won't shrink.
1111 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1113 // Don't try to shrink to various long double types.
1119 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1120 if (Instruction *I = commonCastTransforms(CI))
1123 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1124 // smaller than the destination type, we can eliminate the truncate by doing
1125 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1126 // as many builtins (sqrt, etc).
1127 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1128 if (OpI && OpI->hasOneUse()) {
1129 switch (OpI->getOpcode()) {
1131 case Instruction::FAdd:
1132 case Instruction::FSub:
1133 case Instruction::FMul:
1134 case Instruction::FDiv:
1135 case Instruction::FRem:
1136 const Type *SrcTy = OpI->getType();
1137 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1138 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1139 if (LHSTrunc->getType() != SrcTy &&
1140 RHSTrunc->getType() != SrcTy) {
1141 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1142 // If the source types were both smaller than the destination type of
1143 // the cast, do this xform.
1144 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1145 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1146 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1147 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1148 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1155 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1156 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it.
1157 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1158 if (Call && Call->getCalledFunction() &&
1159 Call->getCalledFunction()->getName() == "sqrt" &&
1160 Call->getNumArgOperands() == 1) {
1161 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1162 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1163 CI.getType()->isFloatTy() &&
1164 Call->getType()->isDoubleTy() &&
1165 Arg->getType()->isDoubleTy() &&
1166 Arg->getOperand(0)->getType()->isFloatTy()) {
1167 Function *Callee = Call->getCalledFunction();
1168 Module *M = CI.getParent()->getParent()->getParent();
1169 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1170 Callee->getAttributes(),
1171 Builder->getFloatTy(),
1172 Builder->getFloatTy(),
1174 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1176 ret->setAttributes(Callee->getAttributes());
1179 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1180 Call->replaceAllUsesWith(UndefValue::get(Call->getType()));
1181 EraseInstFromFunction(*Call);
1189 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1190 return commonCastTransforms(CI);
1193 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1194 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1196 return commonCastTransforms(FI);
1198 // fptoui(uitofp(X)) --> X
1199 // fptoui(sitofp(X)) --> X
1200 // This is safe if the intermediate type has enough bits in its mantissa to
1201 // accurately represent all values of X. For example, do not do this with
1202 // i64->float->i64. This is also safe for sitofp case, because any negative
1203 // 'X' value would cause an undefined result for the fptoui.
1204 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1205 OpI->getOperand(0)->getType() == FI.getType() &&
1206 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1207 OpI->getType()->getFPMantissaWidth())
1208 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1210 return commonCastTransforms(FI);
1213 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1214 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1216 return commonCastTransforms(FI);
1218 // fptosi(sitofp(X)) --> X
1219 // fptosi(uitofp(X)) --> X
1220 // This is safe if the intermediate type has enough bits in its mantissa to
1221 // accurately represent all values of X. For example, do not do this with
1222 // i64->float->i64. This is also safe for sitofp case, because any negative
1223 // 'X' value would cause an undefined result for the fptoui.
1224 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1225 OpI->getOperand(0)->getType() == FI.getType() &&
1226 (int)FI.getType()->getScalarSizeInBits() <=
1227 OpI->getType()->getFPMantissaWidth())
1228 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1230 return commonCastTransforms(FI);
1233 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1234 return commonCastTransforms(CI);
1237 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1238 return commonCastTransforms(CI);
1241 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1242 // If the source integer type is not the intptr_t type for this target, do a
1243 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1244 // cast to be exposed to other transforms.
1246 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
1247 TD->getPointerSizeInBits()) {
1248 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1249 TD->getIntPtrType(CI.getContext()), "tmp");
1250 return new IntToPtrInst(P, CI.getType());
1252 if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
1253 TD->getPointerSizeInBits()) {
1254 Value *P = Builder->CreateZExt(CI.getOperand(0),
1255 TD->getIntPtrType(CI.getContext()), "tmp");
1256 return new IntToPtrInst(P, CI.getType());
1260 if (Instruction *I = commonCastTransforms(CI))
1266 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1267 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1268 Value *Src = CI.getOperand(0);
1270 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1271 // If casting the result of a getelementptr instruction with no offset, turn
1272 // this into a cast of the original pointer!
1273 if (GEP->hasAllZeroIndices()) {
1274 // Changing the cast operand is usually not a good idea but it is safe
1275 // here because the pointer operand is being replaced with another
1276 // pointer operand so the opcode doesn't need to change.
1278 CI.setOperand(0, GEP->getOperand(0));
1282 // If the GEP has a single use, and the base pointer is a bitcast, and the
1283 // GEP computes a constant offset, see if we can convert these three
1284 // instructions into fewer. This typically happens with unions and other
1285 // non-type-safe code.
1286 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1287 GEP->hasAllConstantIndices()) {
1288 // We are guaranteed to get a constant from EmitGEPOffset.
1289 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
1290 int64_t Offset = OffsetV->getSExtValue();
1292 // Get the base pointer input of the bitcast, and the type it points to.
1293 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1294 const Type *GEPIdxTy =
1295 cast<PointerType>(OrigBase->getType())->getElementType();
1296 SmallVector<Value*, 8> NewIndices;
1297 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
1298 // If we were able to index down into an element, create the GEP
1299 // and bitcast the result. This eliminates one bitcast, potentially
1301 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1302 Builder->CreateInBoundsGEP(OrigBase,
1303 NewIndices.begin(), NewIndices.end()) :
1304 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
1305 NGEP->takeName(GEP);
1307 if (isa<BitCastInst>(CI))
1308 return new BitCastInst(NGEP, CI.getType());
1309 assert(isa<PtrToIntInst>(CI));
1310 return new PtrToIntInst(NGEP, CI.getType());
1315 return commonCastTransforms(CI);
1318 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1319 // If the destination integer type is not the intptr_t type for this target,
1320 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1321 // to be exposed to other transforms.
1323 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1324 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1325 TD->getIntPtrType(CI.getContext()),
1327 return new TruncInst(P, CI.getType());
1329 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) {
1330 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1331 TD->getIntPtrType(CI.getContext()),
1333 return new ZExtInst(P, CI.getType());
1337 return commonPointerCastTransforms(CI);
1340 /// OptimizeVectorResize - This input value (which is known to have vector type)
1341 /// is being zero extended or truncated to the specified vector type. Try to
1342 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1344 /// The source and destination vector types may have different element types.
1345 static Instruction *OptimizeVectorResize(Value *InVal, const VectorType *DestTy,
1347 // We can only do this optimization if the output is a multiple of the input
1348 // element size, or the input is a multiple of the output element size.
1349 // Convert the input type to have the same element type as the output.
1350 const VectorType *SrcTy = cast<VectorType>(InVal->getType());
1352 if (SrcTy->getElementType() != DestTy->getElementType()) {
1353 // The input types don't need to be identical, but for now they must be the
1354 // same size. There is no specific reason we couldn't handle things like
1355 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1357 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1358 DestTy->getElementType()->getPrimitiveSizeInBits())
1361 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1362 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1365 // Now that the element types match, get the shuffle mask and RHS of the
1366 // shuffle to use, which depends on whether we're increasing or decreasing the
1367 // size of the input.
1368 SmallVector<Constant*, 16> ShuffleMask;
1370 const IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext());
1372 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1373 // If we're shrinking the number of elements, just shuffle in the low
1374 // elements from the input and use undef as the second shuffle input.
1375 V2 = UndefValue::get(SrcTy);
1376 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1377 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1380 // If we're increasing the number of elements, shuffle in all of the
1381 // elements from InVal and fill the rest of the result elements with zeros
1382 // from a constant zero.
1383 V2 = Constant::getNullValue(SrcTy);
1384 unsigned SrcElts = SrcTy->getNumElements();
1385 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1386 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1388 // The excess elements reference the first element of the zero input.
1389 ShuffleMask.append(DestTy->getNumElements()-SrcElts,
1390 ConstantInt::get(Int32Ty, SrcElts));
1393 Constant *Mask = ConstantVector::get(ShuffleMask.data(), ShuffleMask.size());
1394 return new ShuffleVectorInst(InVal, V2, Mask);
1397 static bool isMultipleOfTypeSize(unsigned Value, const Type *Ty) {
1398 return Value % Ty->getPrimitiveSizeInBits() == 0;
1401 static unsigned getTypeSizeIndex(unsigned Value, const Type *Ty) {
1402 return Value / Ty->getPrimitiveSizeInBits();
1405 /// CollectInsertionElements - V is a value which is inserted into a vector of
1406 /// VecEltTy. Look through the value to see if we can decompose it into
1407 /// insertions into the vector. See the example in the comment for
1408 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1409 /// The type of V is always a non-zero multiple of VecEltTy's size.
1411 /// This returns false if the pattern can't be matched or true if it can,
1412 /// filling in Elements with the elements found here.
1413 static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1414 SmallVectorImpl<Value*> &Elements,
1415 const Type *VecEltTy) {
1416 // Undef values never contribute useful bits to the result.
1417 if (isa<UndefValue>(V)) return true;
1419 // If we got down to a value of the right type, we win, try inserting into the
1421 if (V->getType() == VecEltTy) {
1422 // Inserting null doesn't actually insert any elements.
1423 if (Constant *C = dyn_cast<Constant>(V))
1424 if (C->isNullValue())
1427 // Fail if multiple elements are inserted into this slot.
1428 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1431 Elements[ElementIndex] = V;
1435 if (Constant *C = dyn_cast<Constant>(V)) {
1436 // Figure out the # elements this provides, and bitcast it or slice it up
1438 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1440 // If the constant is the size of a vector element, we just need to bitcast
1441 // it to the right type so it gets properly inserted.
1443 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1444 ElementIndex, Elements, VecEltTy);
1446 // Okay, this is a constant that covers multiple elements. Slice it up into
1447 // pieces and insert each element-sized piece into the vector.
1448 if (!isa<IntegerType>(C->getType()))
1449 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1450 C->getType()->getPrimitiveSizeInBits()));
1451 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1452 const Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1454 for (unsigned i = 0; i != NumElts; ++i) {
1455 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1457 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1458 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1464 if (!V->hasOneUse()) return false;
1466 Instruction *I = dyn_cast<Instruction>(V);
1467 if (I == 0) return false;
1468 switch (I->getOpcode()) {
1469 default: return false; // Unhandled case.
1470 case Instruction::BitCast:
1471 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1472 Elements, VecEltTy);
1473 case Instruction::ZExt:
1474 if (!isMultipleOfTypeSize(
1475 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1478 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1479 Elements, VecEltTy);
1480 case Instruction::Or:
1481 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1482 Elements, VecEltTy) &&
1483 CollectInsertionElements(I->getOperand(1), ElementIndex,
1484 Elements, VecEltTy);
1485 case Instruction::Shl: {
1486 // Must be shifting by a constant that is a multiple of the element size.
1487 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1488 if (CI == 0) return false;
1489 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1490 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1492 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1493 Elements, VecEltTy);
1500 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1501 /// may be doing shifts and ors to assemble the elements of the vector manually.
1502 /// Try to rip the code out and replace it with insertelements. This is to
1503 /// optimize code like this:
1505 /// %tmp37 = bitcast float %inc to i32
1506 /// %tmp38 = zext i32 %tmp37 to i64
1507 /// %tmp31 = bitcast float %inc5 to i32
1508 /// %tmp32 = zext i32 %tmp31 to i64
1509 /// %tmp33 = shl i64 %tmp32, 32
1510 /// %ins35 = or i64 %tmp33, %tmp38
1511 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1513 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1514 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1516 const VectorType *DestVecTy = cast<VectorType>(CI.getType());
1517 Value *IntInput = CI.getOperand(0);
1519 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1520 if (!CollectInsertionElements(IntInput, 0, Elements,
1521 DestVecTy->getElementType()))
1524 // If we succeeded, we know that all of the element are specified by Elements
1525 // or are zero if Elements has a null entry. Recast this as a set of
1527 Value *Result = Constant::getNullValue(CI.getType());
1528 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1529 if (Elements[i] == 0) continue; // Unset element.
1531 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1532 IC.Builder->getInt32(i));
1539 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1540 /// bitcast. The various long double bitcasts can't get in here.
1541 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1542 Value *Src = CI.getOperand(0);
1543 const Type *DestTy = CI.getType();
1545 // If this is a bitcast from int to float, check to see if the int is an
1546 // extraction from a vector.
1547 Value *VecInput = 0;
1548 // bitcast(trunc(bitcast(somevector)))
1549 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1550 isa<VectorType>(VecInput->getType())) {
1551 const VectorType *VecTy = cast<VectorType>(VecInput->getType());
1552 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1554 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1555 // If the element type of the vector doesn't match the result type,
1556 // bitcast it to be a vector type we can extract from.
1557 if (VecTy->getElementType() != DestTy) {
1558 VecTy = VectorType::get(DestTy,
1559 VecTy->getPrimitiveSizeInBits() / DestWidth);
1560 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1563 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
1567 // bitcast(trunc(lshr(bitcast(somevector), cst))
1568 ConstantInt *ShAmt = 0;
1569 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1570 m_ConstantInt(ShAmt)))) &&
1571 isa<VectorType>(VecInput->getType())) {
1572 const VectorType *VecTy = cast<VectorType>(VecInput->getType());
1573 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1574 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1575 ShAmt->getZExtValue() % DestWidth == 0) {
1576 // If the element type of the vector doesn't match the result type,
1577 // bitcast it to be a vector type we can extract from.
1578 if (VecTy->getElementType() != DestTy) {
1579 VecTy = VectorType::get(DestTy,
1580 VecTy->getPrimitiveSizeInBits() / DestWidth);
1581 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1584 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1585 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1591 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1592 // If the operands are integer typed then apply the integer transforms,
1593 // otherwise just apply the common ones.
1594 Value *Src = CI.getOperand(0);
1595 const Type *SrcTy = Src->getType();
1596 const Type *DestTy = CI.getType();
1598 // Get rid of casts from one type to the same type. These are useless and can
1599 // be replaced by the operand.
1600 if (DestTy == Src->getType())
1601 return ReplaceInstUsesWith(CI, Src);
1603 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1604 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1605 const Type *DstElTy = DstPTy->getElementType();
1606 const Type *SrcElTy = SrcPTy->getElementType();
1608 // If the address spaces don't match, don't eliminate the bitcast, which is
1609 // required for changing types.
1610 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1613 // If we are casting a alloca to a pointer to a type of the same
1614 // size, rewrite the allocation instruction to allocate the "right" type.
1615 // There is no need to modify malloc calls because it is their bitcast that
1616 // needs to be cleaned up.
1617 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1618 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1621 // If the source and destination are pointers, and this cast is equivalent
1622 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1623 // This can enhance SROA and other transforms that want type-safe pointers.
1624 Constant *ZeroUInt =
1625 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1626 unsigned NumZeros = 0;
1627 while (SrcElTy != DstElTy &&
1628 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1629 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1630 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1634 // If we found a path from the src to dest, create the getelementptr now.
1635 if (SrcElTy == DstElTy) {
1636 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1637 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1638 ((Instruction*)NULL));
1642 // Try to optimize int -> float bitcasts.
1643 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1644 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1647 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1648 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1649 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1650 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1651 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1652 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1655 if (isa<IntegerType>(SrcTy)) {
1656 // If this is a cast from an integer to vector, check to see if the input
1657 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1658 // the casts with a shuffle and (potentially) a bitcast.
1659 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1660 CastInst *SrcCast = cast<CastInst>(Src);
1661 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1662 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1663 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1664 cast<VectorType>(DestTy), *this))
1668 // If the input is an 'or' instruction, we may be doing shifts and ors to
1669 // assemble the elements of the vector manually. Try to rip the code out
1670 // and replace it with insertelements.
1671 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1672 return ReplaceInstUsesWith(CI, V);
1676 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1677 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
1679 Builder->CreateExtractElement(Src,
1680 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1681 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1685 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1686 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1687 // a bitcast to a vector with the same # elts.
1688 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1689 cast<VectorType>(DestTy)->getNumElements() ==
1690 SVI->getType()->getNumElements() &&
1691 SVI->getType()->getNumElements() ==
1692 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1694 // If either of the operands is a cast from CI.getType(), then
1695 // evaluating the shuffle in the casted destination's type will allow
1696 // us to eliminate at least one cast.
1697 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1698 Tmp->getOperand(0)->getType() == DestTy) ||
1699 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1700 Tmp->getOperand(0)->getType() == DestTy)) {
1701 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1702 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1703 // Return a new shuffle vector. Use the same element ID's, as we
1704 // know the vector types match #elts.
1705 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1710 if (SrcTy->isPointerTy())
1711 return commonPointerCastTransforms(CI);
1712 return commonCastTransforms(CI);