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 // Cannot look past anything that might overflow.
34 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
35 if (OBI && !OBI->hasNoUnsignedWrap()) {
41 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
42 if (I->getOpcode() == Instruction::Shl) {
43 // This is a value scaled by '1 << the shift amt'.
44 Scale = UINT64_C(1) << RHS->getZExtValue();
46 return I->getOperand(0);
49 if (I->getOpcode() == Instruction::Mul) {
50 // This value is scaled by 'RHS'.
51 Scale = RHS->getZExtValue();
53 return I->getOperand(0);
56 if (I->getOpcode() == Instruction::Add) {
57 // We have X+C. Check to see if we really have (X*C2)+C1,
58 // where C1 is divisible by C2.
61 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
62 Offset += RHS->getZExtValue();
69 // Otherwise, we can't look past this.
75 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
76 /// try to eliminate the cast by moving the type information into the alloc.
77 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
79 // This requires TargetData to get the alloca alignment and size information.
82 PointerType *PTy = cast<PointerType>(CI.getType());
84 BuilderTy AllocaBuilder(*Builder);
85 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
87 // Get the type really allocated and the type casted to.
88 Type *AllocElTy = AI.getAllocatedType();
89 Type *CastElTy = PTy->getElementType();
90 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
92 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
93 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
94 if (CastElTyAlign < AllocElTyAlign) return 0;
96 // If the allocation has multiple uses, only promote it if we are strictly
97 // increasing the alignment of the resultant allocation. If we keep it the
98 // same, we open the door to infinite loops of various kinds.
99 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
101 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
102 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
103 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
105 // See if we can satisfy the modulus by pulling a scale out of the array
107 unsigned ArraySizeScale;
108 uint64_t ArrayOffset;
109 Value *NumElements = // See if the array size is a decomposable linear expr.
110 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
112 // If we can now satisfy the modulus, by using a non-1 scale, we really can
114 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
115 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
117 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
122 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
123 // Insert before the alloca, not before the cast.
124 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
127 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
128 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
130 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
133 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
134 New->setAlignment(AI.getAlignment());
137 // If the allocation has multiple real uses, insert a cast and change all
138 // things that used it to use the new cast. This will also hack on CI, but it
140 if (!AI.hasOneUse()) {
141 // New is the allocation instruction, pointer typed. AI is the original
142 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
143 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
144 ReplaceInstUsesWith(AI, NewCast);
146 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);
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!");
222 return InsertNewInstWith(Res, *I);
226 /// This function is a wrapper around CastInst::isEliminableCastPair. It
227 /// simply extracts arguments and returns what that function returns.
228 static Instruction::CastOps
229 isEliminableCastPair(
230 const CastInst *CI, ///< The first cast instruction
231 unsigned opcode, ///< The opcode of the second cast instruction
232 Type *DstTy, ///< The target type for the second cast instruction
233 TargetData *TD ///< The target data for pointer size
236 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
237 Type *MidTy = CI->getType(); // B from above
239 // Get the opcodes of the two Cast instructions
240 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
241 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
243 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
245 TD ? TD->getIntPtrType(CI->getContext()) : 0);
247 // We don't want to form an inttoptr or ptrtoint that converts to an integer
248 // type that differs from the pointer size.
249 if ((Res == Instruction::IntToPtr &&
250 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
251 (Res == Instruction::PtrToInt &&
252 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
255 return Instruction::CastOps(Res);
258 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
259 /// results in any code being generated and is interesting to optimize out. If
260 /// the cast can be eliminated by some other simple transformation, we prefer
261 /// to do the simplification first.
262 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
264 // Noop casts and casts of constants should be eliminated trivially.
265 if (V->getType() == Ty || isa<Constant>(V)) return false;
267 // If this is another cast that can be eliminated, we prefer to have it
269 if (const CastInst *CI = dyn_cast<CastInst>(V))
270 if (isEliminableCastPair(CI, opc, Ty, TD))
273 // If this is a vector sext from a compare, then we don't want to break the
274 // idiom where each element of the extended vector is either zero or all ones.
275 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
282 /// @brief Implement the transforms common to all CastInst visitors.
283 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
284 Value *Src = CI.getOperand(0);
286 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
288 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
289 if (Instruction::CastOps opc =
290 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
291 // The first cast (CSrc) is eliminable so we need to fix up or replace
292 // the second cast (CI). CSrc will then have a good chance of being dead.
293 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
297 // If we are casting a select then fold the cast into the select
298 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
299 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
302 // If we are casting a PHI then fold the cast into the PHI
303 if (isa<PHINode>(Src)) {
304 // We don't do this if this would create a PHI node with an illegal type if
305 // it is currently legal.
306 if (!Src->getType()->isIntegerTy() ||
307 !CI.getType()->isIntegerTy() ||
308 ShouldChangeType(CI.getType(), Src->getType()))
309 if (Instruction *NV = FoldOpIntoPhi(CI))
316 /// CanEvaluateTruncated - Return true if we can evaluate the specified
317 /// expression tree as type Ty instead of its larger type, and arrive with the
318 /// same value. This is used by code that tries to eliminate truncates.
320 /// Ty will always be a type smaller than V. We should return true if trunc(V)
321 /// can be computed by computing V in the smaller type. If V is an instruction,
322 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
323 /// makes sense if x and y can be efficiently truncated.
325 /// This function works on both vectors and scalars.
327 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
328 // We can always evaluate constants in another type.
329 if (isa<Constant>(V))
332 Instruction *I = dyn_cast<Instruction>(V);
333 if (!I) return false;
335 Type *OrigTy = V->getType();
337 // If this is an extension from the dest type, we can eliminate it, even if it
338 // has multiple uses.
339 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
340 I->getOperand(0)->getType() == Ty)
343 // We can't extend or shrink something that has multiple uses: doing so would
344 // require duplicating the instruction in general, which isn't profitable.
345 if (!I->hasOneUse()) return false;
347 unsigned Opc = I->getOpcode();
349 case Instruction::Add:
350 case Instruction::Sub:
351 case Instruction::Mul:
352 case Instruction::And:
353 case Instruction::Or:
354 case Instruction::Xor:
355 // These operators can all arbitrarily be extended or truncated.
356 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
357 CanEvaluateTruncated(I->getOperand(1), Ty);
359 case Instruction::UDiv:
360 case Instruction::URem: {
361 // UDiv and URem can be truncated if all the truncated bits are zero.
362 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
363 uint32_t BitWidth = Ty->getScalarSizeInBits();
364 if (BitWidth < OrigBitWidth) {
365 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
366 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
367 MaskedValueIsZero(I->getOperand(1), Mask)) {
368 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
369 CanEvaluateTruncated(I->getOperand(1), Ty);
374 case Instruction::Shl:
375 // If we are truncating the result of this SHL, and if it's a shift of a
376 // constant amount, we can always perform a SHL in a smaller type.
377 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
378 uint32_t BitWidth = Ty->getScalarSizeInBits();
379 if (CI->getLimitedValue(BitWidth) < BitWidth)
380 return CanEvaluateTruncated(I->getOperand(0), Ty);
383 case Instruction::LShr:
384 // If this is a truncate of a logical shr, we can truncate it to a smaller
385 // lshr iff we know that the bits we would otherwise be shifting in are
387 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
388 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
389 uint32_t BitWidth = Ty->getScalarSizeInBits();
390 if (MaskedValueIsZero(I->getOperand(0),
391 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
392 CI->getLimitedValue(BitWidth) < BitWidth) {
393 return CanEvaluateTruncated(I->getOperand(0), Ty);
397 case Instruction::Trunc:
398 // trunc(trunc(x)) -> trunc(x)
400 case Instruction::ZExt:
401 case Instruction::SExt:
402 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
403 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
405 case Instruction::Select: {
406 SelectInst *SI = cast<SelectInst>(I);
407 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
408 CanEvaluateTruncated(SI->getFalseValue(), Ty);
410 case Instruction::PHI: {
411 // We can change a phi if we can change all operands. Note that we never
412 // get into trouble with cyclic PHIs here because we only consider
413 // instructions with a single use.
414 PHINode *PN = cast<PHINode>(I);
415 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
416 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
421 // TODO: Can handle more cases here.
428 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
429 if (Instruction *Result = commonCastTransforms(CI))
432 // See if we can simplify any instructions used by the input whose sole
433 // purpose is to compute bits we don't care about.
434 if (SimplifyDemandedInstructionBits(CI))
437 Value *Src = CI.getOperand(0);
438 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
440 // Attempt to truncate the entire input expression tree to the destination
441 // type. Only do this if the dest type is a simple type, don't convert the
442 // expression tree to something weird like i93 unless the source is also
444 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
445 CanEvaluateTruncated(Src, DestTy)) {
447 // If this cast is a truncate, evaluting in a different type always
448 // eliminates the cast, so it is always a win.
449 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
450 " to avoid cast: " << CI << '\n');
451 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
452 assert(Res->getType() == DestTy);
453 return ReplaceInstUsesWith(CI, Res);
456 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
457 if (DestTy->getScalarSizeInBits() == 1) {
458 Constant *One = ConstantInt::get(Src->getType(), 1);
459 Src = Builder->CreateAnd(Src, One, "tmp");
460 Value *Zero = Constant::getNullValue(Src->getType());
461 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
464 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
465 Value *A = 0; ConstantInt *Cst = 0;
466 if (Src->hasOneUse() &&
467 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
468 // We have three types to worry about here, the type of A, the source of
469 // the truncate (MidSize), and the destination of the truncate. We know that
470 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
471 // between ASize and ResultSize.
472 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
474 // If the shift amount is larger than the size of A, then the result is
475 // known to be zero because all the input bits got shifted out.
476 if (Cst->getZExtValue() >= ASize)
477 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
479 // Since we're doing an lshr and a zero extend, and know that the shift
480 // amount is smaller than ASize, it is always safe to do the shift in A's
481 // type, then zero extend or truncate to the result.
482 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
483 Shift->takeName(Src);
484 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
487 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
488 // type isn't non-native.
489 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
490 ShouldChangeType(Src->getType(), CI.getType()) &&
491 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
492 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
493 return BinaryOperator::CreateAnd(NewTrunc,
494 ConstantExpr::getTrunc(Cst, CI.getType()));
500 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
501 /// in order to eliminate the icmp.
502 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
504 // If we are just checking for a icmp eq of a single bit and zext'ing it
505 // to an integer, then shift the bit to the appropriate place and then
506 // cast to integer to avoid the comparison.
507 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
508 const APInt &Op1CV = Op1C->getValue();
510 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
511 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
512 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
513 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
514 if (!DoXform) return ICI;
516 Value *In = ICI->getOperand(0);
517 Value *Sh = ConstantInt::get(In->getType(),
518 In->getType()->getScalarSizeInBits()-1);
519 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
520 if (In->getType() != CI.getType())
521 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
523 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
524 Constant *One = ConstantInt::get(In->getType(), 1);
525 In = Builder->CreateXor(In, One, In->getName()+".not");
528 return ReplaceInstUsesWith(CI, In);
533 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
534 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
535 // zext (X == 1) to i32 --> X iff X has only the low bit set.
536 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
537 // zext (X != 0) to i32 --> X iff X has only the low bit set.
538 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
539 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
540 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
541 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
542 // This only works for EQ and NE
544 // If Op1C some other power of two, convert:
545 uint32_t BitWidth = Op1C->getType()->getBitWidth();
546 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
547 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
548 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
550 APInt KnownZeroMask(~KnownZero);
551 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
552 if (!DoXform) return ICI;
554 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
555 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
556 // (X&4) == 2 --> false
557 // (X&4) != 2 --> true
558 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
560 Res = ConstantExpr::getZExt(Res, CI.getType());
561 return ReplaceInstUsesWith(CI, Res);
564 uint32_t ShiftAmt = KnownZeroMask.logBase2();
565 Value *In = ICI->getOperand(0);
567 // Perform a logical shr by shiftamt.
568 // Insert the shift to put the result in the low bit.
569 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
570 In->getName()+".lobit");
573 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
574 Constant *One = ConstantInt::get(In->getType(), 1);
575 In = Builder->CreateXor(In, One, "tmp");
578 if (CI.getType() == In->getType())
579 return ReplaceInstUsesWith(CI, In);
580 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
585 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
586 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
587 // may lead to additional simplifications.
588 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
589 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
590 uint32_t BitWidth = ITy->getBitWidth();
591 Value *LHS = ICI->getOperand(0);
592 Value *RHS = ICI->getOperand(1);
594 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
595 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
596 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
597 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
598 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
600 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
601 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
602 APInt UnknownBit = ~KnownBits;
603 if (UnknownBit.countPopulation() == 1) {
604 if (!DoXform) return ICI;
606 Value *Result = Builder->CreateXor(LHS, RHS);
608 // Mask off any bits that are set and won't be shifted away.
609 if (KnownOneLHS.uge(UnknownBit))
610 Result = Builder->CreateAnd(Result,
611 ConstantInt::get(ITy, UnknownBit));
613 // Shift the bit we're testing down to the lsb.
614 Result = Builder->CreateLShr(
615 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
617 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
618 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
619 Result->takeName(ICI);
620 return ReplaceInstUsesWith(CI, Result);
629 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
630 /// specified wider type and produce the same low bits. If not, return false.
632 /// If this function returns true, it can also return a non-zero number of bits
633 /// (in BitsToClear) which indicates that the value it computes is correct for
634 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
635 /// out. For example, to promote something like:
637 /// %B = trunc i64 %A to i32
638 /// %C = lshr i32 %B, 8
639 /// %E = zext i32 %C to i64
641 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
642 /// set to 8 to indicate that the promoted value needs to have bits 24-31
643 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
644 /// clear the top bits anyway, doing this has no extra cost.
646 /// This function works on both vectors and scalars.
647 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
649 if (isa<Constant>(V))
652 Instruction *I = dyn_cast<Instruction>(V);
653 if (!I) return false;
655 // If the input is a truncate from the destination type, we can trivially
656 // eliminate it, even if it has multiple uses.
657 // FIXME: This is currently disabled until codegen can handle this without
658 // pessimizing code, PR5997.
659 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
662 // We can't extend or shrink something that has multiple uses: doing so would
663 // require duplicating the instruction in general, which isn't profitable.
664 if (!I->hasOneUse()) return false;
666 unsigned Opc = I->getOpcode(), Tmp;
668 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
669 case Instruction::SExt: // zext(sext(x)) -> sext(x).
670 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
672 case Instruction::And:
673 case Instruction::Or:
674 case Instruction::Xor:
675 case Instruction::Add:
676 case Instruction::Sub:
677 case Instruction::Mul:
678 case Instruction::Shl:
679 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
680 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
682 // These can all be promoted if neither operand has 'bits to clear'.
683 if (BitsToClear == 0 && Tmp == 0)
686 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
687 // other side, BitsToClear is ok.
689 (Opc == Instruction::And || Opc == Instruction::Or ||
690 Opc == Instruction::Xor)) {
691 // We use MaskedValueIsZero here for generality, but the case we care
692 // about the most is constant RHS.
693 unsigned VSize = V->getType()->getScalarSizeInBits();
694 if (MaskedValueIsZero(I->getOperand(1),
695 APInt::getHighBitsSet(VSize, BitsToClear)))
699 // Otherwise, we don't know how to analyze this BitsToClear case yet.
702 case Instruction::LShr:
703 // We can promote lshr(x, cst) if we can promote x. This requires the
704 // ultimate 'and' to clear out the high zero bits we're clearing out though.
705 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
706 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
708 BitsToClear += Amt->getZExtValue();
709 if (BitsToClear > V->getType()->getScalarSizeInBits())
710 BitsToClear = V->getType()->getScalarSizeInBits();
713 // Cannot promote variable LSHR.
715 case Instruction::Select:
716 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
717 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
718 // TODO: If important, we could handle the case when the BitsToClear are
719 // known zero in the disagreeing side.
724 case Instruction::PHI: {
725 // We can change a phi if we can change all operands. Note that we never
726 // get into trouble with cyclic PHIs here because we only consider
727 // instructions with a single use.
728 PHINode *PN = cast<PHINode>(I);
729 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
731 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
732 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
733 // TODO: If important, we could handle the case when the BitsToClear
734 // are known zero in the disagreeing input.
740 // TODO: Can handle more cases here.
745 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
746 // If this zero extend is only used by a truncate, let the truncate by
747 // eliminated before we try to optimize this zext.
748 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
751 // If one of the common conversion will work, do it.
752 if (Instruction *Result = commonCastTransforms(CI))
755 // See if we can simplify any instructions used by the input whose sole
756 // purpose is to compute bits we don't care about.
757 if (SimplifyDemandedInstructionBits(CI))
760 Value *Src = CI.getOperand(0);
761 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
763 // Attempt to extend the entire input expression tree to the destination
764 // type. Only do this if the dest type is a simple type, don't convert the
765 // expression tree to something weird like i93 unless the source is also
767 unsigned BitsToClear;
768 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
769 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
770 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
771 "Unreasonable BitsToClear");
773 // Okay, we can transform this! Insert the new expression now.
774 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
775 " to avoid zero extend: " << CI);
776 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
777 assert(Res->getType() == DestTy);
779 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
780 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
782 // If the high bits are already filled with zeros, just replace this
783 // cast with the result.
784 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
785 DestBitSize-SrcBitsKept)))
786 return ReplaceInstUsesWith(CI, Res);
788 // We need to emit an AND to clear the high bits.
789 Constant *C = ConstantInt::get(Res->getType(),
790 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
791 return BinaryOperator::CreateAnd(Res, C);
794 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
795 // types and if the sizes are just right we can convert this into a logical
796 // 'and' which will be much cheaper than the pair of casts.
797 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
798 // TODO: Subsume this into EvaluateInDifferentType.
800 // Get the sizes of the types involved. We know that the intermediate type
801 // will be smaller than A or C, but don't know the relation between A and C.
802 Value *A = CSrc->getOperand(0);
803 unsigned SrcSize = A->getType()->getScalarSizeInBits();
804 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
805 unsigned DstSize = CI.getType()->getScalarSizeInBits();
806 // If we're actually extending zero bits, then if
807 // SrcSize < DstSize: zext(a & mask)
808 // SrcSize == DstSize: a & mask
809 // SrcSize > DstSize: trunc(a) & mask
810 if (SrcSize < DstSize) {
811 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
812 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
813 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
814 return new ZExtInst(And, CI.getType());
817 if (SrcSize == DstSize) {
818 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
819 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
822 if (SrcSize > DstSize) {
823 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
824 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
825 return BinaryOperator::CreateAnd(Trunc,
826 ConstantInt::get(Trunc->getType(),
831 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
832 return transformZExtICmp(ICI, CI);
834 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
835 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
836 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
837 // of the (zext icmp) will be transformed.
838 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
839 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
840 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
841 (transformZExtICmp(LHS, CI, false) ||
842 transformZExtICmp(RHS, CI, false))) {
843 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
844 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
845 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
849 // zext(trunc(t) & C) -> (t & zext(C)).
850 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
851 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
852 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
853 Value *TI0 = TI->getOperand(0);
854 if (TI0->getType() == CI.getType())
856 BinaryOperator::CreateAnd(TI0,
857 ConstantExpr::getZExt(C, CI.getType()));
860 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
861 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
862 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
863 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
864 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
865 And->getOperand(1) == C)
866 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
867 Value *TI0 = TI->getOperand(0);
868 if (TI0->getType() == CI.getType()) {
869 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
870 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
871 return BinaryOperator::CreateXor(NewAnd, ZC);
875 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
877 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
878 match(SrcI, m_Not(m_Value(X))) &&
879 (!X->hasOneUse() || !isa<CmpInst>(X))) {
880 Value *New = Builder->CreateZExt(X, CI.getType());
881 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
887 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
888 /// in order to eliminate the icmp.
889 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
890 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
891 ICmpInst::Predicate Pred = ICI->getPredicate();
893 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
894 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
895 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
896 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) ||
897 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
899 Value *Sh = ConstantInt::get(Op0->getType(),
900 Op0->getType()->getScalarSizeInBits()-1);
901 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
902 if (In->getType() != CI.getType())
903 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/, "tmp");
905 if (Pred == ICmpInst::ICMP_SGT)
906 In = Builder->CreateNot(In, In->getName()+".not");
907 return ReplaceInstUsesWith(CI, In);
910 // If we know that only one bit of the LHS of the icmp can be set and we
911 // have an equality comparison with zero or a power of 2, we can transform
912 // the icmp and sext into bitwise/integer operations.
913 if (ICI->hasOneUse() &&
914 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
915 unsigned BitWidth = Op1C->getType()->getBitWidth();
916 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
917 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
918 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
920 APInt KnownZeroMask(~KnownZero);
921 if (KnownZeroMask.isPowerOf2()) {
922 Value *In = ICI->getOperand(0);
924 // If the icmp tests for a known zero bit we can constant fold it.
925 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
926 Value *V = Pred == ICmpInst::ICMP_NE ?
927 ConstantInt::getAllOnesValue(CI.getType()) :
928 ConstantInt::getNullValue(CI.getType());
929 return ReplaceInstUsesWith(CI, V);
932 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
933 // sext ((x & 2^n) == 0) -> (x >> n) - 1
934 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
935 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
936 // Perform a right shift to place the desired bit in the LSB.
938 In = Builder->CreateLShr(In,
939 ConstantInt::get(In->getType(), ShiftAmt));
941 // At this point "In" is either 1 or 0. Subtract 1 to turn
942 // {1, 0} -> {0, -1}.
943 In = Builder->CreateAdd(In,
944 ConstantInt::getAllOnesValue(In->getType()),
947 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
948 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
949 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
950 // Perform a left shift to place the desired bit in the MSB.
952 In = Builder->CreateShl(In,
953 ConstantInt::get(In->getType(), ShiftAmt));
955 // Distribute the bit over the whole bit width.
956 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
957 BitWidth - 1), "sext");
960 if (CI.getType() == In->getType())
961 return ReplaceInstUsesWith(CI, In);
962 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
967 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
968 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) {
969 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) &&
970 Op0->getType() == CI.getType()) {
971 Type *EltTy = VTy->getElementType();
973 // splat the shift constant to a constant vector.
974 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
975 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit");
976 return ReplaceInstUsesWith(CI, In);
983 /// CanEvaluateSExtd - Return true if we can take the specified value
984 /// and return it as type Ty without inserting any new casts and without
985 /// changing the value of the common low bits. This is used by code that tries
986 /// to promote integer operations to a wider types will allow us to eliminate
989 /// This function works on both vectors and scalars.
991 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
992 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
993 "Can't sign extend type to a smaller type");
994 // If this is a constant, it can be trivially promoted.
995 if (isa<Constant>(V))
998 Instruction *I = dyn_cast<Instruction>(V);
999 if (!I) return false;
1001 // If this is a truncate from the dest type, we can trivially eliminate it,
1002 // even if it has multiple uses.
1003 // FIXME: This is currently disabled until codegen can handle this without
1004 // pessimizing code, PR5997.
1005 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1008 // We can't extend or shrink something that has multiple uses: doing so would
1009 // require duplicating the instruction in general, which isn't profitable.
1010 if (!I->hasOneUse()) return false;
1012 switch (I->getOpcode()) {
1013 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1014 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1015 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1017 case Instruction::And:
1018 case Instruction::Or:
1019 case Instruction::Xor:
1020 case Instruction::Add:
1021 case Instruction::Sub:
1022 case Instruction::Mul:
1023 // These operators can all arbitrarily be extended if their inputs can.
1024 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1025 CanEvaluateSExtd(I->getOperand(1), Ty);
1027 //case Instruction::Shl: TODO
1028 //case Instruction::LShr: TODO
1030 case Instruction::Select:
1031 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1032 CanEvaluateSExtd(I->getOperand(2), Ty);
1034 case Instruction::PHI: {
1035 // We can change a phi if we can change all operands. Note that we never
1036 // get into trouble with cyclic PHIs here because we only consider
1037 // instructions with a single use.
1038 PHINode *PN = cast<PHINode>(I);
1039 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1040 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1044 // TODO: Can handle more cases here.
1051 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1052 // If this sign extend is only used by a truncate, let the truncate by
1053 // eliminated before we try to optimize this zext.
1054 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
1057 if (Instruction *I = commonCastTransforms(CI))
1060 // See if we can simplify any instructions used by the input whose sole
1061 // purpose is to compute bits we don't care about.
1062 if (SimplifyDemandedInstructionBits(CI))
1065 Value *Src = CI.getOperand(0);
1066 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1068 // Attempt to extend the entire input expression tree to the destination
1069 // type. Only do this if the dest type is a simple type, don't convert the
1070 // expression tree to something weird like i93 unless the source is also
1072 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1073 CanEvaluateSExtd(Src, DestTy)) {
1074 // Okay, we can transform this! Insert the new expression now.
1075 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1076 " to avoid sign extend: " << CI);
1077 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1078 assert(Res->getType() == DestTy);
1080 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1081 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1083 // If the high bits are already filled with sign bit, just replace this
1084 // cast with the result.
1085 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1086 return ReplaceInstUsesWith(CI, Res);
1088 // We need to emit a shl + ashr to do the sign extend.
1089 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1090 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1094 // If this input is a trunc from our destination, then turn sext(trunc(x))
1096 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1097 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1098 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1099 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1101 // We need to emit a shl + ashr to do the sign extend.
1102 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1103 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1104 return BinaryOperator::CreateAShr(Res, ShAmt);
1107 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1108 return transformSExtICmp(ICI, CI);
1110 // If the input is a shl/ashr pair of a same constant, then this is a sign
1111 // extension from a smaller value. If we could trust arbitrary bitwidth
1112 // integers, we could turn this into a truncate to the smaller bit and then
1113 // use a sext for the whole extension. Since we don't, look deeper and check
1114 // for a truncate. If the source and dest are the same type, eliminate the
1115 // trunc and extend and just do shifts. For example, turn:
1116 // %a = trunc i32 %i to i8
1117 // %b = shl i8 %a, 6
1118 // %c = ashr i8 %b, 6
1119 // %d = sext i8 %c to i32
1121 // %a = shl i32 %i, 30
1122 // %d = ashr i32 %a, 30
1124 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1125 ConstantInt *BA = 0, *CA = 0;
1126 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1127 m_ConstantInt(CA))) &&
1128 BA == CA && A->getType() == CI.getType()) {
1129 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1130 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1131 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1132 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1133 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1134 return BinaryOperator::CreateAShr(A, ShAmtV);
1141 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1142 /// in the specified FP type without changing its value.
1143 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1145 APFloat F = CFP->getValueAPF();
1146 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1148 return ConstantFP::get(CFP->getContext(), F);
1152 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1153 /// through it until we get the source value.
1154 static Value *LookThroughFPExtensions(Value *V) {
1155 if (Instruction *I = dyn_cast<Instruction>(V))
1156 if (I->getOpcode() == Instruction::FPExt)
1157 return LookThroughFPExtensions(I->getOperand(0));
1159 // If this value is a constant, return the constant in the smallest FP type
1160 // that can accurately represent it. This allows us to turn
1161 // (float)((double)X+2.0) into x+2.0f.
1162 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1163 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1164 return V; // No constant folding of this.
1165 // See if the value can be truncated to float and then reextended.
1166 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1168 if (CFP->getType()->isDoubleTy())
1169 return V; // Won't shrink.
1170 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1172 // Don't try to shrink to various long double types.
1178 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1179 if (Instruction *I = commonCastTransforms(CI))
1182 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1183 // smaller than the destination type, we can eliminate the truncate by doing
1184 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1185 // as many builtins (sqrt, etc).
1186 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1187 if (OpI && OpI->hasOneUse()) {
1188 switch (OpI->getOpcode()) {
1190 case Instruction::FAdd:
1191 case Instruction::FSub:
1192 case Instruction::FMul:
1193 case Instruction::FDiv:
1194 case Instruction::FRem:
1195 Type *SrcTy = OpI->getType();
1196 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1197 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1198 if (LHSTrunc->getType() != SrcTy &&
1199 RHSTrunc->getType() != SrcTy) {
1200 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1201 // If the source types were both smaller than the destination type of
1202 // the cast, do this xform.
1203 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1204 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1205 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1206 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1207 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1214 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1215 // NOTE: This should be disabled by -fno-builtin-sqrt if we ever support it.
1216 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1217 if (Call && Call->getCalledFunction() &&
1218 Call->getCalledFunction()->getName() == "sqrt" &&
1219 Call->getNumArgOperands() == 1 &&
1220 Call->hasOneUse()) {
1221 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1222 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1223 CI.getType()->isFloatTy() &&
1224 Call->getType()->isDoubleTy() &&
1225 Arg->getType()->isDoubleTy() &&
1226 Arg->getOperand(0)->getType()->isFloatTy()) {
1227 Function *Callee = Call->getCalledFunction();
1228 Module *M = CI.getParent()->getParent()->getParent();
1229 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1230 Callee->getAttributes(),
1231 Builder->getFloatTy(),
1232 Builder->getFloatTy(),
1234 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1236 ret->setAttributes(Callee->getAttributes());
1239 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1240 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1241 EraseInstFromFunction(*Call);
1249 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1250 return commonCastTransforms(CI);
1253 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1254 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1256 return commonCastTransforms(FI);
1258 // fptoui(uitofp(X)) --> X
1259 // fptoui(sitofp(X)) --> X
1260 // This is safe if the intermediate type has enough bits in its mantissa to
1261 // accurately represent all values of X. For example, do not do this with
1262 // i64->float->i64. This is also safe for sitofp case, because any negative
1263 // 'X' value would cause an undefined result for the fptoui.
1264 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1265 OpI->getOperand(0)->getType() == FI.getType() &&
1266 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1267 OpI->getType()->getFPMantissaWidth())
1268 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1270 return commonCastTransforms(FI);
1273 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1274 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1276 return commonCastTransforms(FI);
1278 // fptosi(sitofp(X)) --> X
1279 // fptosi(uitofp(X)) --> X
1280 // This is safe if the intermediate type has enough bits in its mantissa to
1281 // accurately represent all values of X. For example, do not do this with
1282 // i64->float->i64. This is also safe for sitofp case, because any negative
1283 // 'X' value would cause an undefined result for the fptoui.
1284 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1285 OpI->getOperand(0)->getType() == FI.getType() &&
1286 (int)FI.getType()->getScalarSizeInBits() <=
1287 OpI->getType()->getFPMantissaWidth())
1288 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1290 return commonCastTransforms(FI);
1293 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1294 return commonCastTransforms(CI);
1297 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1298 return commonCastTransforms(CI);
1301 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1302 // If the source integer type is not the intptr_t type for this target, do a
1303 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1304 // cast to be exposed to other transforms.
1306 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
1307 TD->getPointerSizeInBits()) {
1308 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1309 TD->getIntPtrType(CI.getContext()), "tmp");
1310 return new IntToPtrInst(P, CI.getType());
1312 if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
1313 TD->getPointerSizeInBits()) {
1314 Value *P = Builder->CreateZExt(CI.getOperand(0),
1315 TD->getIntPtrType(CI.getContext()), "tmp");
1316 return new IntToPtrInst(P, CI.getType());
1320 if (Instruction *I = commonCastTransforms(CI))
1326 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1327 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1328 Value *Src = CI.getOperand(0);
1330 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1331 // If casting the result of a getelementptr instruction with no offset, turn
1332 // this into a cast of the original pointer!
1333 if (GEP->hasAllZeroIndices()) {
1334 // Changing the cast operand is usually not a good idea but it is safe
1335 // here because the pointer operand is being replaced with another
1336 // pointer operand so the opcode doesn't need to change.
1338 CI.setOperand(0, GEP->getOperand(0));
1342 // If the GEP has a single use, and the base pointer is a bitcast, and the
1343 // GEP computes a constant offset, see if we can convert these three
1344 // instructions into fewer. This typically happens with unions and other
1345 // non-type-safe code.
1346 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1347 GEP->hasAllConstantIndices()) {
1348 // We are guaranteed to get a constant from EmitGEPOffset.
1349 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
1350 int64_t Offset = OffsetV->getSExtValue();
1352 // Get the base pointer input of the bitcast, and the type it points to.
1353 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1355 cast<PointerType>(OrigBase->getType())->getElementType();
1356 SmallVector<Value*, 8> NewIndices;
1357 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
1358 // If we were able to index down into an element, create the GEP
1359 // and bitcast the result. This eliminates one bitcast, potentially
1361 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1362 Builder->CreateInBoundsGEP(OrigBase,
1363 NewIndices.begin(), NewIndices.end()) :
1364 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
1365 NGEP->takeName(GEP);
1367 if (isa<BitCastInst>(CI))
1368 return new BitCastInst(NGEP, CI.getType());
1369 assert(isa<PtrToIntInst>(CI));
1370 return new PtrToIntInst(NGEP, CI.getType());
1375 return commonCastTransforms(CI);
1378 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1379 // If the destination integer type is not the intptr_t type for this target,
1380 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1381 // to be exposed to other transforms.
1383 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1384 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1385 TD->getIntPtrType(CI.getContext()),
1387 return new TruncInst(P, CI.getType());
1389 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) {
1390 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1391 TD->getIntPtrType(CI.getContext()),
1393 return new ZExtInst(P, CI.getType());
1397 return commonPointerCastTransforms(CI);
1400 /// OptimizeVectorResize - This input value (which is known to have vector type)
1401 /// is being zero extended or truncated to the specified vector type. Try to
1402 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1404 /// The source and destination vector types may have different element types.
1405 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1407 // We can only do this optimization if the output is a multiple of the input
1408 // element size, or the input is a multiple of the output element size.
1409 // Convert the input type to have the same element type as the output.
1410 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1412 if (SrcTy->getElementType() != DestTy->getElementType()) {
1413 // The input types don't need to be identical, but for now they must be the
1414 // same size. There is no specific reason we couldn't handle things like
1415 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1417 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1418 DestTy->getElementType()->getPrimitiveSizeInBits())
1421 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1422 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1425 // Now that the element types match, get the shuffle mask and RHS of the
1426 // shuffle to use, which depends on whether we're increasing or decreasing the
1427 // size of the input.
1428 SmallVector<Constant*, 16> ShuffleMask;
1430 IntegerType *Int32Ty = Type::getInt32Ty(SrcTy->getContext());
1432 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1433 // If we're shrinking the number of elements, just shuffle in the low
1434 // elements from the input and use undef as the second shuffle input.
1435 V2 = UndefValue::get(SrcTy);
1436 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1437 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1440 // If we're increasing the number of elements, shuffle in all of the
1441 // elements from InVal and fill the rest of the result elements with zeros
1442 // from a constant zero.
1443 V2 = Constant::getNullValue(SrcTy);
1444 unsigned SrcElts = SrcTy->getNumElements();
1445 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1446 ShuffleMask.push_back(ConstantInt::get(Int32Ty, i));
1448 // The excess elements reference the first element of the zero input.
1449 ShuffleMask.append(DestTy->getNumElements()-SrcElts,
1450 ConstantInt::get(Int32Ty, SrcElts));
1453 return new ShuffleVectorInst(InVal, V2, ConstantVector::get(ShuffleMask));
1456 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1457 return Value % Ty->getPrimitiveSizeInBits() == 0;
1460 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1461 return Value / Ty->getPrimitiveSizeInBits();
1464 /// CollectInsertionElements - V is a value which is inserted into a vector of
1465 /// VecEltTy. Look through the value to see if we can decompose it into
1466 /// insertions into the vector. See the example in the comment for
1467 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1468 /// The type of V is always a non-zero multiple of VecEltTy's size.
1470 /// This returns false if the pattern can't be matched or true if it can,
1471 /// filling in Elements with the elements found here.
1472 static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1473 SmallVectorImpl<Value*> &Elements,
1475 // Undef values never contribute useful bits to the result.
1476 if (isa<UndefValue>(V)) return true;
1478 // If we got down to a value of the right type, we win, try inserting into the
1480 if (V->getType() == VecEltTy) {
1481 // Inserting null doesn't actually insert any elements.
1482 if (Constant *C = dyn_cast<Constant>(V))
1483 if (C->isNullValue())
1486 // Fail if multiple elements are inserted into this slot.
1487 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1490 Elements[ElementIndex] = V;
1494 if (Constant *C = dyn_cast<Constant>(V)) {
1495 // Figure out the # elements this provides, and bitcast it or slice it up
1497 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1499 // If the constant is the size of a vector element, we just need to bitcast
1500 // it to the right type so it gets properly inserted.
1502 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1503 ElementIndex, Elements, VecEltTy);
1505 // Okay, this is a constant that covers multiple elements. Slice it up into
1506 // pieces and insert each element-sized piece into the vector.
1507 if (!isa<IntegerType>(C->getType()))
1508 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1509 C->getType()->getPrimitiveSizeInBits()));
1510 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1511 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1513 for (unsigned i = 0; i != NumElts; ++i) {
1514 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1516 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1517 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1523 if (!V->hasOneUse()) return false;
1525 Instruction *I = dyn_cast<Instruction>(V);
1526 if (I == 0) return false;
1527 switch (I->getOpcode()) {
1528 default: return false; // Unhandled case.
1529 case Instruction::BitCast:
1530 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1531 Elements, VecEltTy);
1532 case Instruction::ZExt:
1533 if (!isMultipleOfTypeSize(
1534 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1537 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1538 Elements, VecEltTy);
1539 case Instruction::Or:
1540 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1541 Elements, VecEltTy) &&
1542 CollectInsertionElements(I->getOperand(1), ElementIndex,
1543 Elements, VecEltTy);
1544 case Instruction::Shl: {
1545 // Must be shifting by a constant that is a multiple of the element size.
1546 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1547 if (CI == 0) return false;
1548 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1549 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1551 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1552 Elements, VecEltTy);
1559 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1560 /// may be doing shifts and ors to assemble the elements of the vector manually.
1561 /// Try to rip the code out and replace it with insertelements. This is to
1562 /// optimize code like this:
1564 /// %tmp37 = bitcast float %inc to i32
1565 /// %tmp38 = zext i32 %tmp37 to i64
1566 /// %tmp31 = bitcast float %inc5 to i32
1567 /// %tmp32 = zext i32 %tmp31 to i64
1568 /// %tmp33 = shl i64 %tmp32, 32
1569 /// %ins35 = or i64 %tmp33, %tmp38
1570 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1572 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1573 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1575 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1576 Value *IntInput = CI.getOperand(0);
1578 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1579 if (!CollectInsertionElements(IntInput, 0, Elements,
1580 DestVecTy->getElementType()))
1583 // If we succeeded, we know that all of the element are specified by Elements
1584 // or are zero if Elements has a null entry. Recast this as a set of
1586 Value *Result = Constant::getNullValue(CI.getType());
1587 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1588 if (Elements[i] == 0) continue; // Unset element.
1590 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1591 IC.Builder->getInt32(i));
1598 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1599 /// bitcast. The various long double bitcasts can't get in here.
1600 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1601 Value *Src = CI.getOperand(0);
1602 Type *DestTy = CI.getType();
1604 // If this is a bitcast from int to float, check to see if the int is an
1605 // extraction from a vector.
1606 Value *VecInput = 0;
1607 // bitcast(trunc(bitcast(somevector)))
1608 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1609 isa<VectorType>(VecInput->getType())) {
1610 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1611 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1613 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1614 // If the element type of the vector doesn't match the result type,
1615 // bitcast it to be a vector type we can extract from.
1616 if (VecTy->getElementType() != DestTy) {
1617 VecTy = VectorType::get(DestTy,
1618 VecTy->getPrimitiveSizeInBits() / DestWidth);
1619 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1622 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
1626 // bitcast(trunc(lshr(bitcast(somevector), cst))
1627 ConstantInt *ShAmt = 0;
1628 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1629 m_ConstantInt(ShAmt)))) &&
1630 isa<VectorType>(VecInput->getType())) {
1631 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1632 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1633 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1634 ShAmt->getZExtValue() % DestWidth == 0) {
1635 // If the element type of the vector doesn't match the result type,
1636 // bitcast it to be a vector type we can extract from.
1637 if (VecTy->getElementType() != DestTy) {
1638 VecTy = VectorType::get(DestTy,
1639 VecTy->getPrimitiveSizeInBits() / DestWidth);
1640 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1643 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1644 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1650 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1651 // If the operands are integer typed then apply the integer transforms,
1652 // otherwise just apply the common ones.
1653 Value *Src = CI.getOperand(0);
1654 Type *SrcTy = Src->getType();
1655 Type *DestTy = CI.getType();
1657 // Get rid of casts from one type to the same type. These are useless and can
1658 // be replaced by the operand.
1659 if (DestTy == Src->getType())
1660 return ReplaceInstUsesWith(CI, Src);
1662 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1663 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1664 Type *DstElTy = DstPTy->getElementType();
1665 Type *SrcElTy = SrcPTy->getElementType();
1667 // If the address spaces don't match, don't eliminate the bitcast, which is
1668 // required for changing types.
1669 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1672 // If we are casting a alloca to a pointer to a type of the same
1673 // size, rewrite the allocation instruction to allocate the "right" type.
1674 // There is no need to modify malloc calls because it is their bitcast that
1675 // needs to be cleaned up.
1676 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1677 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1680 // If the source and destination are pointers, and this cast is equivalent
1681 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1682 // This can enhance SROA and other transforms that want type-safe pointers.
1683 Constant *ZeroUInt =
1684 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1685 unsigned NumZeros = 0;
1686 while (SrcElTy != DstElTy &&
1687 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1688 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1689 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1693 // If we found a path from the src to dest, create the getelementptr now.
1694 if (SrcElTy == DstElTy) {
1695 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1696 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end());
1700 // Try to optimize int -> float bitcasts.
1701 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1702 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1705 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1706 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1707 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1708 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1709 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1710 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1713 if (isa<IntegerType>(SrcTy)) {
1714 // If this is a cast from an integer to vector, check to see if the input
1715 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1716 // the casts with a shuffle and (potentially) a bitcast.
1717 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1718 CastInst *SrcCast = cast<CastInst>(Src);
1719 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1720 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1721 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1722 cast<VectorType>(DestTy), *this))
1726 // If the input is an 'or' instruction, we may be doing shifts and ors to
1727 // assemble the elements of the vector manually. Try to rip the code out
1728 // and replace it with insertelements.
1729 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1730 return ReplaceInstUsesWith(CI, V);
1734 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1735 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
1737 Builder->CreateExtractElement(Src,
1738 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1739 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1743 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1744 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1745 // a bitcast to a vector with the same # elts.
1746 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1747 cast<VectorType>(DestTy)->getNumElements() ==
1748 SVI->getType()->getNumElements() &&
1749 SVI->getType()->getNumElements() ==
1750 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1752 // If either of the operands is a cast from CI.getType(), then
1753 // evaluating the shuffle in the casted destination's type will allow
1754 // us to eliminate at least one cast.
1755 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1756 Tmp->getOperand(0)->getType() == DestTy) ||
1757 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1758 Tmp->getOperand(0)->getType() == DestTy)) {
1759 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1760 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1761 // Return a new shuffle vector. Use the same element ID's, as we
1762 // know the vector types match #elts.
1763 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1768 if (SrcTy->isPointerTy())
1769 return commonPointerCastTransforms(CI);
1770 return commonCastTransforms(CI);