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/Analysis/ConstantFolding.h"
16 #include "llvm/DataLayout.h"
17 #include "llvm/Target/TargetLibraryInfo.h"
18 #include "llvm/Support/PatternMatch.h"
20 using namespace PatternMatch;
22 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
23 /// expression. If so, decompose it, returning some value X, such that Val is
26 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
28 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
29 Offset = CI->getZExtValue();
31 return ConstantInt::get(Val->getType(), 0);
34 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
35 // Cannot look past anything that might overflow.
36 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
37 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
43 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
44 if (I->getOpcode() == Instruction::Shl) {
45 // This is a value scaled by '1 << the shift amt'.
46 Scale = UINT64_C(1) << RHS->getZExtValue();
48 return I->getOperand(0);
51 if (I->getOpcode() == Instruction::Mul) {
52 // This value is scaled by 'RHS'.
53 Scale = RHS->getZExtValue();
55 return I->getOperand(0);
58 if (I->getOpcode() == Instruction::Add) {
59 // We have X+C. Check to see if we really have (X*C2)+C1,
60 // where C1 is divisible by C2.
63 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
64 Offset += RHS->getZExtValue();
71 // Otherwise, we can't look past this.
77 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
78 /// try to eliminate the cast by moving the type information into the alloc.
79 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
81 // This requires DataLayout to get the alloca alignment and size information.
84 PointerType *PTy = cast<PointerType>(CI.getType());
86 BuilderTy AllocaBuilder(*Builder);
87 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
89 // Get the type really allocated and the type casted to.
90 Type *AllocElTy = AI.getAllocatedType();
91 Type *CastElTy = PTy->getElementType();
92 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
94 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
95 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
96 if (CastElTyAlign < AllocElTyAlign) return 0;
98 // If the allocation has multiple uses, only promote it if we are strictly
99 // increasing the alignment of the resultant allocation. If we keep it the
100 // same, we open the door to infinite loops of various kinds.
101 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
103 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
104 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
105 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
107 // See if we can satisfy the modulus by pulling a scale out of the array
109 unsigned ArraySizeScale;
110 uint64_t ArrayOffset;
111 Value *NumElements = // See if the array size is a decomposable linear expr.
112 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
114 // If we can now satisfy the modulus, by using a non-1 scale, we really can
116 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
117 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
119 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
124 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
125 // Insert before the alloca, not before the cast.
126 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
129 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
130 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
132 Amt = AllocaBuilder.CreateAdd(Amt, Off);
135 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
136 New->setAlignment(AI.getAlignment());
139 // If the allocation has multiple real uses, insert a cast and change all
140 // things that used it to use the new cast. This will also hack on CI, but it
142 if (!AI.hasOneUse()) {
143 // New is the allocation instruction, pointer typed. AI is the original
144 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
145 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
146 ReplaceInstUsesWith(AI, NewCast);
148 return ReplaceInstUsesWith(CI, New);
151 /// EvaluateInDifferentType - Given an expression that
152 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
153 /// insert the code to evaluate the expression.
154 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
156 if (Constant *C = dyn_cast<Constant>(V)) {
157 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
158 // If we got a constantexpr back, try to simplify it with TD info.
159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
160 C = ConstantFoldConstantExpression(CE, TD, TLI);
164 // Otherwise, it must be an instruction.
165 Instruction *I = cast<Instruction>(V);
166 Instruction *Res = 0;
167 unsigned Opc = I->getOpcode();
169 case Instruction::Add:
170 case Instruction::Sub:
171 case Instruction::Mul:
172 case Instruction::And:
173 case Instruction::Or:
174 case Instruction::Xor:
175 case Instruction::AShr:
176 case Instruction::LShr:
177 case Instruction::Shl:
178 case Instruction::UDiv:
179 case Instruction::URem: {
180 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
181 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
182 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
185 case Instruction::Trunc:
186 case Instruction::ZExt:
187 case Instruction::SExt:
188 // If the source type of the cast is the type we're trying for then we can
189 // just return the source. There's no need to insert it because it is not
191 if (I->getOperand(0)->getType() == Ty)
192 return I->getOperand(0);
194 // Otherwise, must be the same type of cast, so just reinsert a new one.
195 // This also handles the case of zext(trunc(x)) -> zext(x).
196 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
197 Opc == Instruction::SExt);
199 case Instruction::Select: {
200 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
201 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
202 Res = SelectInst::Create(I->getOperand(0), True, False);
205 case Instruction::PHI: {
206 PHINode *OPN = cast<PHINode>(I);
207 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
208 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
209 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
210 NPN->addIncoming(V, OPN->getIncomingBlock(i));
216 // TODO: Can handle more cases here.
217 llvm_unreachable("Unreachable!");
221 return InsertNewInstWith(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 Type *DstTy, ///< The target type for the second cast instruction
232 DataLayout *TD ///< The target data for pointer size
235 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
236 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);
241 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
243 TD ? TD->getIntPtrType(DstTy) : 0);
245 // We don't want to form an inttoptr or ptrtoint that converts to an integer
246 // type that differs from the pointer size.
247 if ((Res == Instruction::IntToPtr &&
248 (!TD || SrcTy != TD->getIntPtrType(DstTy))) ||
249 (Res == Instruction::PtrToInt &&
250 (!TD || DstTy != TD->getIntPtrType(SrcTy))))
253 return Instruction::CastOps(Res);
256 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
257 /// results in any code being generated and is interesting to optimize out. If
258 /// the cast can be eliminated by some other simple transformation, we prefer
259 /// to do the simplification first.
260 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
262 // Noop casts and casts of constants should be eliminated trivially.
263 if (V->getType() == Ty || isa<Constant>(V)) return false;
265 // If this is another cast that can be eliminated, we prefer to have it
267 if (const CastInst *CI = dyn_cast<CastInst>(V))
268 if (isEliminableCastPair(CI, opc, Ty, TD))
271 // If this is a vector sext from a compare, then we don't want to break the
272 // idiom where each element of the extended vector is either zero or all ones.
273 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
280 /// @brief Implement the transforms common to all CastInst visitors.
281 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
282 Value *Src = CI.getOperand(0);
284 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
286 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
287 if (Instruction::CastOps opc =
288 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
289 // The first cast (CSrc) is eliminable so we need to fix up or replace
290 // the second cast (CI). CSrc will then have a good chance of being dead.
291 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
295 // If we are casting a select then fold the cast into the select
296 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
297 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
300 // If we are casting a PHI then fold the cast into the PHI
301 if (isa<PHINode>(Src)) {
302 // We don't do this if this would create a PHI node with an illegal type if
303 // it is currently legal.
304 if (!Src->getType()->isIntegerTy() ||
305 !CI.getType()->isIntegerTy() ||
306 ShouldChangeType(CI.getType(), Src->getType()))
307 if (Instruction *NV = FoldOpIntoPhi(CI))
314 /// CanEvaluateTruncated - Return true if we can evaluate the specified
315 /// expression tree as type Ty instead of its larger type, and arrive with the
316 /// same value. This is used by code that tries to eliminate truncates.
318 /// Ty will always be a type smaller than V. We should return true if trunc(V)
319 /// can be computed by computing V in the smaller type. If V is an instruction,
320 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
321 /// makes sense if x and y can be efficiently truncated.
323 /// This function works on both vectors and scalars.
325 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
326 // We can always evaluate constants in another type.
327 if (isa<Constant>(V))
330 Instruction *I = dyn_cast<Instruction>(V);
331 if (!I) return false;
333 Type *OrigTy = V->getType();
335 // If this is an extension from the dest type, we can eliminate it, even if it
336 // has multiple uses.
337 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
338 I->getOperand(0)->getType() == Ty)
341 // We can't extend or shrink something that has multiple uses: doing so would
342 // require duplicating the instruction in general, which isn't profitable.
343 if (!I->hasOneUse()) return false;
345 unsigned Opc = I->getOpcode();
347 case Instruction::Add:
348 case Instruction::Sub:
349 case Instruction::Mul:
350 case Instruction::And:
351 case Instruction::Or:
352 case Instruction::Xor:
353 // These operators can all arbitrarily be extended or truncated.
354 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
355 CanEvaluateTruncated(I->getOperand(1), Ty);
357 case Instruction::UDiv:
358 case Instruction::URem: {
359 // UDiv and URem can be truncated if all the truncated bits are zero.
360 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
361 uint32_t BitWidth = Ty->getScalarSizeInBits();
362 if (BitWidth < OrigBitWidth) {
363 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
364 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
365 MaskedValueIsZero(I->getOperand(1), Mask)) {
366 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
367 CanEvaluateTruncated(I->getOperand(1), Ty);
372 case Instruction::Shl:
373 // If we are truncating the result of this SHL, and if it's a shift of a
374 // constant amount, we can always perform a SHL in a smaller type.
375 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
376 uint32_t BitWidth = Ty->getScalarSizeInBits();
377 if (CI->getLimitedValue(BitWidth) < BitWidth)
378 return CanEvaluateTruncated(I->getOperand(0), Ty);
381 case Instruction::LShr:
382 // If this is a truncate of a logical shr, we can truncate it to a smaller
383 // lshr iff we know that the bits we would otherwise be shifting in are
385 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
386 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
387 uint32_t BitWidth = Ty->getScalarSizeInBits();
388 if (MaskedValueIsZero(I->getOperand(0),
389 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
390 CI->getLimitedValue(BitWidth) < BitWidth) {
391 return CanEvaluateTruncated(I->getOperand(0), Ty);
395 case Instruction::Trunc:
396 // trunc(trunc(x)) -> trunc(x)
398 case Instruction::ZExt:
399 case Instruction::SExt:
400 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
401 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
403 case Instruction::Select: {
404 SelectInst *SI = cast<SelectInst>(I);
405 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
406 CanEvaluateTruncated(SI->getFalseValue(), Ty);
408 case Instruction::PHI: {
409 // We can change a phi if we can change all operands. Note that we never
410 // get into trouble with cyclic PHIs here because we only consider
411 // instructions with a single use.
412 PHINode *PN = cast<PHINode>(I);
413 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
414 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
419 // TODO: Can handle more cases here.
426 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
427 if (Instruction *Result = commonCastTransforms(CI))
430 // See if we can simplify any instructions used by the input whose sole
431 // purpose is to compute bits we don't care about.
432 if (SimplifyDemandedInstructionBits(CI))
435 Value *Src = CI.getOperand(0);
436 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
438 // Attempt to truncate the entire input expression tree to the destination
439 // type. Only do this if the dest type is a simple type, don't convert the
440 // expression tree to something weird like i93 unless the source is also
442 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
443 CanEvaluateTruncated(Src, DestTy)) {
445 // If this cast is a truncate, evaluting in a different type always
446 // eliminates the cast, so it is always a win.
447 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
448 " to avoid cast: " << CI << '\n');
449 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
450 assert(Res->getType() == DestTy);
451 return ReplaceInstUsesWith(CI, Res);
454 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
455 if (DestTy->getScalarSizeInBits() == 1) {
456 Constant *One = ConstantInt::get(Src->getType(), 1);
457 Src = Builder->CreateAnd(Src, One);
458 Value *Zero = Constant::getNullValue(Src->getType());
459 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
462 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
463 Value *A = 0; ConstantInt *Cst = 0;
464 if (Src->hasOneUse() &&
465 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
466 // We have three types to worry about here, the type of A, the source of
467 // the truncate (MidSize), and the destination of the truncate. We know that
468 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
469 // between ASize and ResultSize.
470 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
472 // If the shift amount is larger than the size of A, then the result is
473 // known to be zero because all the input bits got shifted out.
474 if (Cst->getZExtValue() >= ASize)
475 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
477 // Since we're doing an lshr and a zero extend, and know that the shift
478 // amount is smaller than ASize, it is always safe to do the shift in A's
479 // type, then zero extend or truncate to the result.
480 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
481 Shift->takeName(Src);
482 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
485 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
486 // type isn't non-native.
487 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
488 ShouldChangeType(Src->getType(), CI.getType()) &&
489 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
490 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
491 return BinaryOperator::CreateAnd(NewTrunc,
492 ConstantExpr::getTrunc(Cst, CI.getType()));
498 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
499 /// in order to eliminate the icmp.
500 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
502 // If we are just checking for a icmp eq of a single bit and zext'ing it
503 // to an integer, then shift the bit to the appropriate place and then
504 // cast to integer to avoid the comparison.
505 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
506 const APInt &Op1CV = Op1C->getValue();
508 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
509 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
510 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
511 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
512 if (!DoXform) return ICI;
514 Value *In = ICI->getOperand(0);
515 Value *Sh = ConstantInt::get(In->getType(),
516 In->getType()->getScalarSizeInBits()-1);
517 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
518 if (In->getType() != CI.getType())
519 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
521 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
522 Constant *One = ConstantInt::get(In->getType(), 1);
523 In = Builder->CreateXor(In, One, In->getName()+".not");
526 return ReplaceInstUsesWith(CI, In);
529 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
530 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
531 // zext (X == 1) to i32 --> X iff X has only the low bit set.
532 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
533 // zext (X != 0) to i32 --> X iff X has only the low bit set.
534 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
535 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
536 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
537 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
538 // This only works for EQ and NE
540 // If Op1C some other power of two, convert:
541 uint32_t BitWidth = Op1C->getType()->getBitWidth();
542 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
543 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne);
545 APInt KnownZeroMask(~KnownZero);
546 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
547 if (!DoXform) return ICI;
549 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
550 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
551 // (X&4) == 2 --> false
552 // (X&4) != 2 --> true
553 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
555 Res = ConstantExpr::getZExt(Res, CI.getType());
556 return ReplaceInstUsesWith(CI, Res);
559 uint32_t ShiftAmt = KnownZeroMask.logBase2();
560 Value *In = ICI->getOperand(0);
562 // Perform a logical shr by shiftamt.
563 // Insert the shift to put the result in the low bit.
564 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
565 In->getName()+".lobit");
568 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
569 Constant *One = ConstantInt::get(In->getType(), 1);
570 In = Builder->CreateXor(In, One);
573 if (CI.getType() == In->getType())
574 return ReplaceInstUsesWith(CI, In);
575 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
580 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
581 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
582 // may lead to additional simplifications.
583 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
584 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
585 uint32_t BitWidth = ITy->getBitWidth();
586 Value *LHS = ICI->getOperand(0);
587 Value *RHS = ICI->getOperand(1);
589 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
590 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
591 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS);
592 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS);
594 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
595 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
596 APInt UnknownBit = ~KnownBits;
597 if (UnknownBit.countPopulation() == 1) {
598 if (!DoXform) return ICI;
600 Value *Result = Builder->CreateXor(LHS, RHS);
602 // Mask off any bits that are set and won't be shifted away.
603 if (KnownOneLHS.uge(UnknownBit))
604 Result = Builder->CreateAnd(Result,
605 ConstantInt::get(ITy, UnknownBit));
607 // Shift the bit we're testing down to the lsb.
608 Result = Builder->CreateLShr(
609 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
611 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
612 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
613 Result->takeName(ICI);
614 return ReplaceInstUsesWith(CI, Result);
623 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
624 /// specified wider type and produce the same low bits. If not, return false.
626 /// If this function returns true, it can also return a non-zero number of bits
627 /// (in BitsToClear) which indicates that the value it computes is correct for
628 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
629 /// out. For example, to promote something like:
631 /// %B = trunc i64 %A to i32
632 /// %C = lshr i32 %B, 8
633 /// %E = zext i32 %C to i64
635 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
636 /// set to 8 to indicate that the promoted value needs to have bits 24-31
637 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
638 /// clear the top bits anyway, doing this has no extra cost.
640 /// This function works on both vectors and scalars.
641 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
643 if (isa<Constant>(V))
646 Instruction *I = dyn_cast<Instruction>(V);
647 if (!I) return false;
649 // If the input is a truncate from the destination type, we can trivially
651 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
654 // We can't extend or shrink something that has multiple uses: doing so would
655 // require duplicating the instruction in general, which isn't profitable.
656 if (!I->hasOneUse()) return false;
658 unsigned Opc = I->getOpcode(), Tmp;
660 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
661 case Instruction::SExt: // zext(sext(x)) -> sext(x).
662 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
664 case Instruction::And:
665 case Instruction::Or:
666 case Instruction::Xor:
667 case Instruction::Add:
668 case Instruction::Sub:
669 case Instruction::Mul:
670 case Instruction::Shl:
671 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
672 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
674 // These can all be promoted if neither operand has 'bits to clear'.
675 if (BitsToClear == 0 && Tmp == 0)
678 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
679 // other side, BitsToClear is ok.
681 (Opc == Instruction::And || Opc == Instruction::Or ||
682 Opc == Instruction::Xor)) {
683 // We use MaskedValueIsZero here for generality, but the case we care
684 // about the most is constant RHS.
685 unsigned VSize = V->getType()->getScalarSizeInBits();
686 if (MaskedValueIsZero(I->getOperand(1),
687 APInt::getHighBitsSet(VSize, BitsToClear)))
691 // Otherwise, we don't know how to analyze this BitsToClear case yet.
694 case Instruction::LShr:
695 // We can promote lshr(x, cst) if we can promote x. This requires the
696 // ultimate 'and' to clear out the high zero bits we're clearing out though.
697 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
698 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
700 BitsToClear += Amt->getZExtValue();
701 if (BitsToClear > V->getType()->getScalarSizeInBits())
702 BitsToClear = V->getType()->getScalarSizeInBits();
705 // Cannot promote variable LSHR.
707 case Instruction::Select:
708 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
709 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
710 // TODO: If important, we could handle the case when the BitsToClear are
711 // known zero in the disagreeing side.
716 case Instruction::PHI: {
717 // We can change a phi if we can change all operands. Note that we never
718 // get into trouble with cyclic PHIs here because we only consider
719 // instructions with a single use.
720 PHINode *PN = cast<PHINode>(I);
721 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
723 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
724 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
725 // TODO: If important, we could handle the case when the BitsToClear
726 // are known zero in the disagreeing input.
732 // TODO: Can handle more cases here.
737 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
738 // If this zero extend is only used by a truncate, let the truncate by
739 // eliminated before we try to optimize this zext.
740 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
743 // If one of the common conversion will work, do it.
744 if (Instruction *Result = commonCastTransforms(CI))
747 // See if we can simplify any instructions used by the input whose sole
748 // purpose is to compute bits we don't care about.
749 if (SimplifyDemandedInstructionBits(CI))
752 Value *Src = CI.getOperand(0);
753 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
755 // Attempt to extend the entire input expression tree to the destination
756 // type. Only do this if the dest type is a simple type, don't convert the
757 // expression tree to something weird like i93 unless the source is also
759 unsigned BitsToClear;
760 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
761 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
762 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
763 "Unreasonable BitsToClear");
765 // Okay, we can transform this! Insert the new expression now.
766 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
767 " to avoid zero extend: " << CI);
768 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
769 assert(Res->getType() == DestTy);
771 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
772 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
774 // If the high bits are already filled with zeros, just replace this
775 // cast with the result.
776 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
777 DestBitSize-SrcBitsKept)))
778 return ReplaceInstUsesWith(CI, Res);
780 // We need to emit an AND to clear the high bits.
781 Constant *C = ConstantInt::get(Res->getType(),
782 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
783 return BinaryOperator::CreateAnd(Res, C);
786 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
787 // types and if the sizes are just right we can convert this into a logical
788 // 'and' which will be much cheaper than the pair of casts.
789 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
790 // TODO: Subsume this into EvaluateInDifferentType.
792 // Get the sizes of the types involved. We know that the intermediate type
793 // will be smaller than A or C, but don't know the relation between A and C.
794 Value *A = CSrc->getOperand(0);
795 unsigned SrcSize = A->getType()->getScalarSizeInBits();
796 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
797 unsigned DstSize = CI.getType()->getScalarSizeInBits();
798 // If we're actually extending zero bits, then if
799 // SrcSize < DstSize: zext(a & mask)
800 // SrcSize == DstSize: a & mask
801 // SrcSize > DstSize: trunc(a) & mask
802 if (SrcSize < DstSize) {
803 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
804 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
805 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
806 return new ZExtInst(And, CI.getType());
809 if (SrcSize == DstSize) {
810 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
811 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
814 if (SrcSize > DstSize) {
815 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
816 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
817 return BinaryOperator::CreateAnd(Trunc,
818 ConstantInt::get(Trunc->getType(),
823 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
824 return transformZExtICmp(ICI, CI);
826 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
827 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
828 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
829 // of the (zext icmp) will be transformed.
830 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
831 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
832 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
833 (transformZExtICmp(LHS, CI, false) ||
834 transformZExtICmp(RHS, CI, false))) {
835 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
836 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
837 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
841 // zext(trunc(t) & C) -> (t & zext(C)).
842 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
843 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
844 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
845 Value *TI0 = TI->getOperand(0);
846 if (TI0->getType() == CI.getType())
848 BinaryOperator::CreateAnd(TI0,
849 ConstantExpr::getZExt(C, CI.getType()));
852 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
853 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
854 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
855 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
856 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
857 And->getOperand(1) == C)
858 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
859 Value *TI0 = TI->getOperand(0);
860 if (TI0->getType() == CI.getType()) {
861 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
862 Value *NewAnd = Builder->CreateAnd(TI0, ZC);
863 return BinaryOperator::CreateXor(NewAnd, ZC);
867 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
869 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
870 match(SrcI, m_Not(m_Value(X))) &&
871 (!X->hasOneUse() || !isa<CmpInst>(X))) {
872 Value *New = Builder->CreateZExt(X, CI.getType());
873 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
879 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
880 /// in order to eliminate the icmp.
881 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
882 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
883 ICmpInst::Predicate Pred = ICI->getPredicate();
885 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
886 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
887 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
888 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) ||
889 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
891 Value *Sh = ConstantInt::get(Op0->getType(),
892 Op0->getType()->getScalarSizeInBits()-1);
893 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
894 if (In->getType() != CI.getType())
895 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
897 if (Pred == ICmpInst::ICMP_SGT)
898 In = Builder->CreateNot(In, In->getName()+".not");
899 return ReplaceInstUsesWith(CI, In);
902 // If we know that only one bit of the LHS of the icmp can be set and we
903 // have an equality comparison with zero or a power of 2, we can transform
904 // the icmp and sext into bitwise/integer operations.
905 if (ICI->hasOneUse() &&
906 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
907 unsigned BitWidth = Op1C->getType()->getBitWidth();
908 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
909 ComputeMaskedBits(Op0, KnownZero, KnownOne);
911 APInt KnownZeroMask(~KnownZero);
912 if (KnownZeroMask.isPowerOf2()) {
913 Value *In = ICI->getOperand(0);
915 // If the icmp tests for a known zero bit we can constant fold it.
916 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
917 Value *V = Pred == ICmpInst::ICMP_NE ?
918 ConstantInt::getAllOnesValue(CI.getType()) :
919 ConstantInt::getNullValue(CI.getType());
920 return ReplaceInstUsesWith(CI, V);
923 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
924 // sext ((x & 2^n) == 0) -> (x >> n) - 1
925 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
926 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
927 // Perform a right shift to place the desired bit in the LSB.
929 In = Builder->CreateLShr(In,
930 ConstantInt::get(In->getType(), ShiftAmt));
932 // At this point "In" is either 1 or 0. Subtract 1 to turn
933 // {1, 0} -> {0, -1}.
934 In = Builder->CreateAdd(In,
935 ConstantInt::getAllOnesValue(In->getType()),
938 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
939 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
940 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
941 // Perform a left shift to place the desired bit in the MSB.
943 In = Builder->CreateShl(In,
944 ConstantInt::get(In->getType(), ShiftAmt));
946 // Distribute the bit over the whole bit width.
947 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
948 BitWidth - 1), "sext");
951 if (CI.getType() == In->getType())
952 return ReplaceInstUsesWith(CI, In);
953 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
958 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
959 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) {
960 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) &&
961 Op0->getType() == CI.getType()) {
962 Type *EltTy = VTy->getElementType();
964 // splat the shift constant to a constant vector.
965 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
966 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit");
967 return ReplaceInstUsesWith(CI, In);
974 /// CanEvaluateSExtd - Return true if we can take the specified value
975 /// and return it as type Ty without inserting any new casts and without
976 /// changing the value of the common low bits. This is used by code that tries
977 /// to promote integer operations to a wider types will allow us to eliminate
980 /// This function works on both vectors and scalars.
982 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
983 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
984 "Can't sign extend type to a smaller type");
985 // If this is a constant, it can be trivially promoted.
986 if (isa<Constant>(V))
989 Instruction *I = dyn_cast<Instruction>(V);
990 if (!I) return false;
992 // If this is a truncate from the dest type, we can trivially eliminate it.
993 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
996 // We can't extend or shrink something that has multiple uses: doing so would
997 // require duplicating the instruction in general, which isn't profitable.
998 if (!I->hasOneUse()) return false;
1000 switch (I->getOpcode()) {
1001 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1002 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1003 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1005 case Instruction::And:
1006 case Instruction::Or:
1007 case Instruction::Xor:
1008 case Instruction::Add:
1009 case Instruction::Sub:
1010 case Instruction::Mul:
1011 // These operators can all arbitrarily be extended if their inputs can.
1012 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1013 CanEvaluateSExtd(I->getOperand(1), Ty);
1015 //case Instruction::Shl: TODO
1016 //case Instruction::LShr: TODO
1018 case Instruction::Select:
1019 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1020 CanEvaluateSExtd(I->getOperand(2), Ty);
1022 case Instruction::PHI: {
1023 // We can change a phi if we can change all operands. Note that we never
1024 // get into trouble with cyclic PHIs here because we only consider
1025 // instructions with a single use.
1026 PHINode *PN = cast<PHINode>(I);
1027 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1028 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1032 // TODO: Can handle more cases here.
1039 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1040 // If this sign extend is only used by a truncate, let the truncate by
1041 // eliminated before we try to optimize this zext.
1042 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
1045 if (Instruction *I = commonCastTransforms(CI))
1048 // See if we can simplify any instructions used by the input whose sole
1049 // purpose is to compute bits we don't care about.
1050 if (SimplifyDemandedInstructionBits(CI))
1053 Value *Src = CI.getOperand(0);
1054 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1056 // Attempt to extend the entire input expression tree to the destination
1057 // type. Only do this if the dest type is a simple type, don't convert the
1058 // expression tree to something weird like i93 unless the source is also
1060 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1061 CanEvaluateSExtd(Src, DestTy)) {
1062 // Okay, we can transform this! Insert the new expression now.
1063 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1064 " to avoid sign extend: " << CI);
1065 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1066 assert(Res->getType() == DestTy);
1068 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1069 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1071 // If the high bits are already filled with sign bit, just replace this
1072 // cast with the result.
1073 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1074 return ReplaceInstUsesWith(CI, Res);
1076 // We need to emit a shl + ashr to do the sign extend.
1077 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1078 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1082 // If this input is a trunc from our destination, then turn sext(trunc(x))
1084 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1085 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1086 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1087 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1089 // We need to emit a shl + ashr to do the sign extend.
1090 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1091 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1092 return BinaryOperator::CreateAShr(Res, ShAmt);
1095 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1096 return transformSExtICmp(ICI, CI);
1098 // If the input is a shl/ashr pair of a same constant, then this is a sign
1099 // extension from a smaller value. If we could trust arbitrary bitwidth
1100 // integers, we could turn this into a truncate to the smaller bit and then
1101 // use a sext for the whole extension. Since we don't, look deeper and check
1102 // for a truncate. If the source and dest are the same type, eliminate the
1103 // trunc and extend and just do shifts. For example, turn:
1104 // %a = trunc i32 %i to i8
1105 // %b = shl i8 %a, 6
1106 // %c = ashr i8 %b, 6
1107 // %d = sext i8 %c to i32
1109 // %a = shl i32 %i, 30
1110 // %d = ashr i32 %a, 30
1112 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1113 ConstantInt *BA = 0, *CA = 0;
1114 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1115 m_ConstantInt(CA))) &&
1116 BA == CA && A->getType() == CI.getType()) {
1117 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1118 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1119 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1120 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1121 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1122 return BinaryOperator::CreateAShr(A, ShAmtV);
1129 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1130 /// in the specified FP type without changing its value.
1131 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1133 APFloat F = CFP->getValueAPF();
1134 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1136 return ConstantFP::get(CFP->getContext(), F);
1140 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1141 /// through it until we get the source value.
1142 static Value *LookThroughFPExtensions(Value *V) {
1143 if (Instruction *I = dyn_cast<Instruction>(V))
1144 if (I->getOpcode() == Instruction::FPExt)
1145 return LookThroughFPExtensions(I->getOperand(0));
1147 // If this value is a constant, return the constant in the smallest FP type
1148 // that can accurately represent it. This allows us to turn
1149 // (float)((double)X+2.0) into x+2.0f.
1150 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1151 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1152 return V; // No constant folding of this.
1153 // See if the value can be truncated to half and then reextended.
1154 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1156 // See if the value can be truncated to float and then reextended.
1157 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1159 if (CFP->getType()->isDoubleTy())
1160 return V; // Won't shrink.
1161 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1163 // Don't try to shrink to various long double types.
1169 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1170 if (Instruction *I = commonCastTransforms(CI))
1173 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1174 // smaller than the destination type, we can eliminate the truncate by doing
1175 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1176 // as many builtins (sqrt, etc).
1177 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1178 if (OpI && OpI->hasOneUse()) {
1179 switch (OpI->getOpcode()) {
1181 case Instruction::FAdd:
1182 case Instruction::FSub:
1183 case Instruction::FMul:
1184 case Instruction::FDiv:
1185 case Instruction::FRem:
1186 Type *SrcTy = OpI->getType();
1187 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1188 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1189 if (LHSTrunc->getType() != SrcTy &&
1190 RHSTrunc->getType() != SrcTy) {
1191 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1192 // If the source types were both smaller than the destination type of
1193 // the cast, do this xform.
1194 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1195 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1196 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1197 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1198 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1205 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1206 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1207 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1208 Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) &&
1209 Call->getNumArgOperands() == 1 &&
1210 Call->hasOneUse()) {
1211 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1212 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1213 CI.getType()->isFloatTy() &&
1214 Call->getType()->isDoubleTy() &&
1215 Arg->getType()->isDoubleTy() &&
1216 Arg->getOperand(0)->getType()->isFloatTy()) {
1217 Function *Callee = Call->getCalledFunction();
1218 Module *M = CI.getParent()->getParent()->getParent();
1219 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1220 Callee->getAttributes(),
1221 Builder->getFloatTy(),
1222 Builder->getFloatTy(),
1224 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1226 ret->setAttributes(Callee->getAttributes());
1229 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1230 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1231 EraseInstFromFunction(*Call);
1239 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1240 return commonCastTransforms(CI);
1243 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1244 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1246 return commonCastTransforms(FI);
1248 // fptoui(uitofp(X)) --> X
1249 // fptoui(sitofp(X)) --> X
1250 // This is safe if the intermediate type has enough bits in its mantissa to
1251 // accurately represent all values of X. For example, do not do this with
1252 // i64->float->i64. This is also safe for sitofp case, because any negative
1253 // 'X' value would cause an undefined result for the fptoui.
1254 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1255 OpI->getOperand(0)->getType() == FI.getType() &&
1256 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1257 OpI->getType()->getFPMantissaWidth())
1258 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1260 return commonCastTransforms(FI);
1263 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1264 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1266 return commonCastTransforms(FI);
1268 // fptosi(sitofp(X)) --> X
1269 // fptosi(uitofp(X)) --> X
1270 // This is safe if the intermediate type has enough bits in its mantissa to
1271 // accurately represent all values of X. For example, do not do this with
1272 // i64->float->i64. This is also safe for sitofp case, because any negative
1273 // 'X' value would cause an undefined result for the fptoui.
1274 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1275 OpI->getOperand(0)->getType() == FI.getType() &&
1276 (int)FI.getType()->getScalarSizeInBits() <=
1277 OpI->getType()->getFPMantissaWidth())
1278 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1280 return commonCastTransforms(FI);
1283 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1284 return commonCastTransforms(CI);
1287 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1288 return commonCastTransforms(CI);
1291 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1292 // If the source integer type is not the intptr_t type for this target, do a
1293 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1294 // cast to be exposed to other transforms.
1295 unsigned AS = CI.getAddressSpace();
1297 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
1298 TD->getPointerSizeInBits(AS)) {
1299 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1300 TD->getIntPtrType(CI.getType()));
1301 return new IntToPtrInst(P, CI.getType());
1303 if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
1304 TD->getPointerSizeInBits(AS)) {
1305 Value *P = Builder->CreateZExt(CI.getOperand(0),
1306 TD->getIntPtrType(CI.getType()));
1307 return new IntToPtrInst(P, CI.getType());
1311 if (Instruction *I = commonCastTransforms(CI))
1317 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1318 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1319 Value *Src = CI.getOperand(0);
1321 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1322 // If casting the result of a getelementptr instruction with no offset, turn
1323 // this into a cast of the original pointer!
1324 if (GEP->hasAllZeroIndices()) {
1325 // Changing the cast operand is usually not a good idea but it is safe
1326 // here because the pointer operand is being replaced with another
1327 // pointer operand so the opcode doesn't need to change.
1329 CI.setOperand(0, GEP->getOperand(0));
1333 // If the GEP has a single use, and the base pointer is a bitcast, and the
1334 // GEP computes a constant offset, see if we can convert these three
1335 // instructions into fewer. This typically happens with unions and other
1336 // non-type-safe code.
1337 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1338 GEP->hasAllConstantIndices()) {
1339 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
1340 int64_t Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops);
1342 // Get the base pointer input of the bitcast, and the type it points to.
1343 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1345 cast<PointerType>(OrigBase->getType())->getElementType();
1346 SmallVector<Value*, 8> NewIndices;
1347 Type *IntPtrTy = TD->getIntPtrType(OrigBase->getType());
1348 if (FindElementAtOffset(GEPIdxTy, Offset, IntPtrTy, NewIndices)) {
1349 // If we were able to index down into an element, create the GEP
1350 // and bitcast the result. This eliminates one bitcast, potentially
1352 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1353 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1354 Builder->CreateGEP(OrigBase, NewIndices);
1355 NGEP->takeName(GEP);
1357 if (isa<BitCastInst>(CI))
1358 return new BitCastInst(NGEP, CI.getType());
1359 assert(isa<PtrToIntInst>(CI));
1360 return new PtrToIntInst(NGEP, CI.getType());
1365 return commonCastTransforms(CI);
1368 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1369 // If the destination integer type is not the intptr_t type for this target,
1370 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1371 // to be exposed to other transforms.
1372 unsigned AS = CI.getPointerAddressSpace();
1374 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits(AS)) {
1375 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1376 TD->getIntPtrType(CI.getContext(), AS));
1377 return new TruncInst(P, CI.getType());
1379 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits(AS)) {
1380 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1381 TD->getIntPtrType(CI.getContext(), AS));
1382 return new ZExtInst(P, CI.getType());
1386 return commonPointerCastTransforms(CI);
1389 /// OptimizeVectorResize - This input value (which is known to have vector type)
1390 /// is being zero extended or truncated to the specified vector type. Try to
1391 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1393 /// The source and destination vector types may have different element types.
1394 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1396 // We can only do this optimization if the output is a multiple of the input
1397 // element size, or the input is a multiple of the output element size.
1398 // Convert the input type to have the same element type as the output.
1399 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1401 if (SrcTy->getElementType() != DestTy->getElementType()) {
1402 // The input types don't need to be identical, but for now they must be the
1403 // same size. There is no specific reason we couldn't handle things like
1404 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1406 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1407 DestTy->getElementType()->getPrimitiveSizeInBits())
1410 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1411 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1414 // Now that the element types match, get the shuffle mask and RHS of the
1415 // shuffle to use, which depends on whether we're increasing or decreasing the
1416 // size of the input.
1417 SmallVector<uint32_t, 16> ShuffleMask;
1420 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1421 // If we're shrinking the number of elements, just shuffle in the low
1422 // elements from the input and use undef as the second shuffle input.
1423 V2 = UndefValue::get(SrcTy);
1424 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1425 ShuffleMask.push_back(i);
1428 // If we're increasing the number of elements, shuffle in all of the
1429 // elements from InVal and fill the rest of the result elements with zeros
1430 // from a constant zero.
1431 V2 = Constant::getNullValue(SrcTy);
1432 unsigned SrcElts = SrcTy->getNumElements();
1433 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1434 ShuffleMask.push_back(i);
1436 // The excess elements reference the first element of the zero input.
1437 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1438 ShuffleMask.push_back(SrcElts);
1441 return new ShuffleVectorInst(InVal, V2,
1442 ConstantDataVector::get(V2->getContext(),
1446 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1447 return Value % Ty->getPrimitiveSizeInBits() == 0;
1450 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1451 return Value / Ty->getPrimitiveSizeInBits();
1454 /// CollectInsertionElements - V is a value which is inserted into a vector of
1455 /// VecEltTy. Look through the value to see if we can decompose it into
1456 /// insertions into the vector. See the example in the comment for
1457 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1458 /// The type of V is always a non-zero multiple of VecEltTy's size.
1460 /// This returns false if the pattern can't be matched or true if it can,
1461 /// filling in Elements with the elements found here.
1462 static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1463 SmallVectorImpl<Value*> &Elements,
1465 // Undef values never contribute useful bits to the result.
1466 if (isa<UndefValue>(V)) return true;
1468 // If we got down to a value of the right type, we win, try inserting into the
1470 if (V->getType() == VecEltTy) {
1471 // Inserting null doesn't actually insert any elements.
1472 if (Constant *C = dyn_cast<Constant>(V))
1473 if (C->isNullValue())
1476 // Fail if multiple elements are inserted into this slot.
1477 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1480 Elements[ElementIndex] = V;
1484 if (Constant *C = dyn_cast<Constant>(V)) {
1485 // Figure out the # elements this provides, and bitcast it or slice it up
1487 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1489 // If the constant is the size of a vector element, we just need to bitcast
1490 // it to the right type so it gets properly inserted.
1492 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1493 ElementIndex, Elements, VecEltTy);
1495 // Okay, this is a constant that covers multiple elements. Slice it up into
1496 // pieces and insert each element-sized piece into the vector.
1497 if (!isa<IntegerType>(C->getType()))
1498 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1499 C->getType()->getPrimitiveSizeInBits()));
1500 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1501 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1503 for (unsigned i = 0; i != NumElts; ++i) {
1504 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1506 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1507 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1513 if (!V->hasOneUse()) return false;
1515 Instruction *I = dyn_cast<Instruction>(V);
1516 if (I == 0) return false;
1517 switch (I->getOpcode()) {
1518 default: return false; // Unhandled case.
1519 case Instruction::BitCast:
1520 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1521 Elements, VecEltTy);
1522 case Instruction::ZExt:
1523 if (!isMultipleOfTypeSize(
1524 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1527 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1528 Elements, VecEltTy);
1529 case Instruction::Or:
1530 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1531 Elements, VecEltTy) &&
1532 CollectInsertionElements(I->getOperand(1), ElementIndex,
1533 Elements, VecEltTy);
1534 case Instruction::Shl: {
1535 // Must be shifting by a constant that is a multiple of the element size.
1536 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1537 if (CI == 0) return false;
1538 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1539 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1541 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1542 Elements, VecEltTy);
1549 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1550 /// may be doing shifts and ors to assemble the elements of the vector manually.
1551 /// Try to rip the code out and replace it with insertelements. This is to
1552 /// optimize code like this:
1554 /// %tmp37 = bitcast float %inc to i32
1555 /// %tmp38 = zext i32 %tmp37 to i64
1556 /// %tmp31 = bitcast float %inc5 to i32
1557 /// %tmp32 = zext i32 %tmp31 to i64
1558 /// %tmp33 = shl i64 %tmp32, 32
1559 /// %ins35 = or i64 %tmp33, %tmp38
1560 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1562 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1563 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1565 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1566 Value *IntInput = CI.getOperand(0);
1568 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1569 if (!CollectInsertionElements(IntInput, 0, Elements,
1570 DestVecTy->getElementType()))
1573 // If we succeeded, we know that all of the element are specified by Elements
1574 // or are zero if Elements has a null entry. Recast this as a set of
1576 Value *Result = Constant::getNullValue(CI.getType());
1577 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1578 if (Elements[i] == 0) continue; // Unset element.
1580 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1581 IC.Builder->getInt32(i));
1588 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1589 /// bitcast. The various long double bitcasts can't get in here.
1590 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1591 Value *Src = CI.getOperand(0);
1592 Type *DestTy = CI.getType();
1594 // If this is a bitcast from int to float, check to see if the int is an
1595 // extraction from a vector.
1596 Value *VecInput = 0;
1597 // bitcast(trunc(bitcast(somevector)))
1598 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1599 isa<VectorType>(VecInput->getType())) {
1600 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1601 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1603 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1604 // If the element type of the vector doesn't match the result type,
1605 // bitcast it to be a vector type we can extract from.
1606 if (VecTy->getElementType() != DestTy) {
1607 VecTy = VectorType::get(DestTy,
1608 VecTy->getPrimitiveSizeInBits() / DestWidth);
1609 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1612 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
1616 // bitcast(trunc(lshr(bitcast(somevector), cst))
1617 ConstantInt *ShAmt = 0;
1618 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1619 m_ConstantInt(ShAmt)))) &&
1620 isa<VectorType>(VecInput->getType())) {
1621 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1622 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1623 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1624 ShAmt->getZExtValue() % DestWidth == 0) {
1625 // If the element type of the vector doesn't match the result type,
1626 // bitcast it to be a vector type we can extract from.
1627 if (VecTy->getElementType() != DestTy) {
1628 VecTy = VectorType::get(DestTy,
1629 VecTy->getPrimitiveSizeInBits() / DestWidth);
1630 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1633 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1634 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1640 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1641 // If the operands are integer typed then apply the integer transforms,
1642 // otherwise just apply the common ones.
1643 Value *Src = CI.getOperand(0);
1644 Type *SrcTy = Src->getType();
1645 Type *DestTy = CI.getType();
1647 // Get rid of casts from one type to the same type. These are useless and can
1648 // be replaced by the operand.
1649 if (DestTy == Src->getType())
1650 return ReplaceInstUsesWith(CI, Src);
1652 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1653 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1654 Type *DstElTy = DstPTy->getElementType();
1655 Type *SrcElTy = SrcPTy->getElementType();
1657 // If the address spaces don't match, don't eliminate the bitcast, which is
1658 // required for changing types.
1659 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1662 // If we are casting a alloca to a pointer to a type of the same
1663 // size, rewrite the allocation instruction to allocate the "right" type.
1664 // There is no need to modify malloc calls because it is their bitcast that
1665 // needs to be cleaned up.
1666 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1667 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1670 // If the source and destination are pointers, and this cast is equivalent
1671 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1672 // This can enhance SROA and other transforms that want type-safe pointers.
1673 Constant *ZeroUInt =
1674 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1675 unsigned NumZeros = 0;
1676 while (SrcElTy != DstElTy &&
1677 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1678 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1679 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1683 // If we found a path from the src to dest, create the getelementptr now.
1684 if (SrcElTy == DstElTy) {
1685 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1686 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1690 // Try to optimize int -> float bitcasts.
1691 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1692 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1695 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1696 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1697 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1698 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1699 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1700 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1703 if (isa<IntegerType>(SrcTy)) {
1704 // If this is a cast from an integer to vector, check to see if the input
1705 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1706 // the casts with a shuffle and (potentially) a bitcast.
1707 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1708 CastInst *SrcCast = cast<CastInst>(Src);
1709 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1710 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1711 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1712 cast<VectorType>(DestTy), *this))
1716 // If the input is an 'or' instruction, we may be doing shifts and ors to
1717 // assemble the elements of the vector manually. Try to rip the code out
1718 // and replace it with insertelements.
1719 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1720 return ReplaceInstUsesWith(CI, V);
1724 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1725 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
1727 Builder->CreateExtractElement(Src,
1728 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1729 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1733 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1734 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1735 // a bitcast to a vector with the same # elts.
1736 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1737 cast<VectorType>(DestTy)->getNumElements() ==
1738 SVI->getType()->getNumElements() &&
1739 SVI->getType()->getNumElements() ==
1740 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1742 // If either of the operands is a cast from CI.getType(), then
1743 // evaluating the shuffle in the casted destination's type will allow
1744 // us to eliminate at least one cast.
1745 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1746 Tmp->getOperand(0)->getType() == DestTy) ||
1747 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1748 Tmp->getOperand(0)->getType() == DestTy)) {
1749 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1750 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1751 // Return a new shuffle vector. Use the same element ID's, as we
1752 // know the vector types match #elts.
1753 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1758 if (SrcTy->isPointerTy())
1759 return commonPointerCastTransforms(CI);
1760 return commonCastTransforms(CI);