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/Target/TargetData.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()) {
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 TargetData 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 TargetData *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);
242 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
244 TD ? TD->getIntPtrType(CI->getContext()) : 0);
246 // We don't want to form an inttoptr or ptrtoint that converts to an integer
247 // type that differs from the pointer size.
248 if ((Res == Instruction::IntToPtr &&
249 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
250 (Res == Instruction::PtrToInt &&
251 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
254 return Instruction::CastOps(Res);
257 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
258 /// results in any code being generated and is interesting to optimize out. If
259 /// the cast can be eliminated by some other simple transformation, we prefer
260 /// to do the simplification first.
261 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
263 // Noop casts and casts of constants should be eliminated trivially.
264 if (V->getType() == Ty || isa<Constant>(V)) return false;
266 // If this is another cast that can be eliminated, we prefer to have it
268 if (const CastInst *CI = dyn_cast<CastInst>(V))
269 if (isEliminableCastPair(CI, opc, Ty, TD))
272 // If this is a vector sext from a compare, then we don't want to break the
273 // idiom where each element of the extended vector is either zero or all ones.
274 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
281 /// @brief Implement the transforms common to all CastInst visitors.
282 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
283 Value *Src = CI.getOperand(0);
285 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
287 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
288 if (Instruction::CastOps opc =
289 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
290 // The first cast (CSrc) is eliminable so we need to fix up or replace
291 // the second cast (CI). CSrc will then have a good chance of being dead.
292 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
296 // If we are casting a select then fold the cast into the select
297 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
298 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
301 // If we are casting a PHI then fold the cast into the PHI
302 if (isa<PHINode>(Src)) {
303 // We don't do this if this would create a PHI node with an illegal type if
304 // it is currently legal.
305 if (!Src->getType()->isIntegerTy() ||
306 !CI.getType()->isIntegerTy() ||
307 ShouldChangeType(CI.getType(), Src->getType()))
308 if (Instruction *NV = FoldOpIntoPhi(CI))
315 /// CanEvaluateTruncated - Return true if we can evaluate the specified
316 /// expression tree as type Ty instead of its larger type, and arrive with the
317 /// same value. This is used by code that tries to eliminate truncates.
319 /// Ty will always be a type smaller than V. We should return true if trunc(V)
320 /// can be computed by computing V in the smaller type. If V is an instruction,
321 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
322 /// makes sense if x and y can be efficiently truncated.
324 /// This function works on both vectors and scalars.
326 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
327 // We can always evaluate constants in another type.
328 if (isa<Constant>(V))
331 Instruction *I = dyn_cast<Instruction>(V);
332 if (!I) return false;
334 Type *OrigTy = V->getType();
336 // If this is an extension from the dest type, we can eliminate it, even if it
337 // has multiple uses.
338 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
339 I->getOperand(0)->getType() == Ty)
342 // We can't extend or shrink something that has multiple uses: doing so would
343 // require duplicating the instruction in general, which isn't profitable.
344 if (!I->hasOneUse()) return false;
346 unsigned Opc = I->getOpcode();
348 case Instruction::Add:
349 case Instruction::Sub:
350 case Instruction::Mul:
351 case Instruction::And:
352 case Instruction::Or:
353 case Instruction::Xor:
354 // These operators can all arbitrarily be extended or truncated.
355 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
356 CanEvaluateTruncated(I->getOperand(1), Ty);
358 case Instruction::UDiv:
359 case Instruction::URem: {
360 // UDiv and URem can be truncated if all the truncated bits are zero.
361 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
362 uint32_t BitWidth = Ty->getScalarSizeInBits();
363 if (BitWidth < OrigBitWidth) {
364 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
365 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
366 MaskedValueIsZero(I->getOperand(1), Mask)) {
367 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
368 CanEvaluateTruncated(I->getOperand(1), Ty);
373 case Instruction::Shl:
374 // If we are truncating the result of this SHL, and if it's a shift of a
375 // constant amount, we can always perform a SHL in a smaller type.
376 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
377 uint32_t BitWidth = Ty->getScalarSizeInBits();
378 if (CI->getLimitedValue(BitWidth) < BitWidth)
379 return CanEvaluateTruncated(I->getOperand(0), Ty);
382 case Instruction::LShr:
383 // If this is a truncate of a logical shr, we can truncate it to a smaller
384 // lshr iff we know that the bits we would otherwise be shifting in are
386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
387 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
388 uint32_t BitWidth = Ty->getScalarSizeInBits();
389 if (MaskedValueIsZero(I->getOperand(0),
390 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
391 CI->getLimitedValue(BitWidth) < BitWidth) {
392 return CanEvaluateTruncated(I->getOperand(0), Ty);
396 case Instruction::Trunc:
397 // trunc(trunc(x)) -> trunc(x)
399 case Instruction::ZExt:
400 case Instruction::SExt:
401 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
402 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
404 case Instruction::Select: {
405 SelectInst *SI = cast<SelectInst>(I);
406 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
407 CanEvaluateTruncated(SI->getFalseValue(), Ty);
409 case Instruction::PHI: {
410 // We can change a phi if we can change all operands. Note that we never
411 // get into trouble with cyclic PHIs here because we only consider
412 // instructions with a single use.
413 PHINode *PN = cast<PHINode>(I);
414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
415 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
420 // TODO: Can handle more cases here.
427 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
428 if (Instruction *Result = commonCastTransforms(CI))
431 // See if we can simplify any instructions used by the input whose sole
432 // purpose is to compute bits we don't care about.
433 if (SimplifyDemandedInstructionBits(CI))
436 Value *Src = CI.getOperand(0);
437 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
439 // Attempt to truncate the entire input expression tree to the destination
440 // type. Only do this if the dest type is a simple type, don't convert the
441 // expression tree to something weird like i93 unless the source is also
443 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
444 CanEvaluateTruncated(Src, DestTy)) {
446 // If this cast is a truncate, evaluting in a different type always
447 // eliminates the cast, so it is always a win.
448 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
449 " to avoid cast: " << CI << '\n');
450 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
451 assert(Res->getType() == DestTy);
452 return ReplaceInstUsesWith(CI, Res);
455 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
456 if (DestTy->getScalarSizeInBits() == 1) {
457 Constant *One = ConstantInt::get(Src->getType(), 1);
458 Src = Builder->CreateAnd(Src, One);
459 Value *Zero = Constant::getNullValue(Src->getType());
460 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
463 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
464 Value *A = 0; ConstantInt *Cst = 0;
465 if (Src->hasOneUse() &&
466 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
467 // We have three types to worry about here, the type of A, the source of
468 // the truncate (MidSize), and the destination of the truncate. We know that
469 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
470 // between ASize and ResultSize.
471 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
473 // If the shift amount is larger than the size of A, then the result is
474 // known to be zero because all the input bits got shifted out.
475 if (Cst->getZExtValue() >= ASize)
476 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
478 // Since we're doing an lshr and a zero extend, and know that the shift
479 // amount is smaller than ASize, it is always safe to do the shift in A's
480 // type, then zero extend or truncate to the result.
481 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
482 Shift->takeName(Src);
483 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
486 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
487 // type isn't non-native.
488 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
489 ShouldChangeType(Src->getType(), CI.getType()) &&
490 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
491 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
492 return BinaryOperator::CreateAnd(NewTrunc,
493 ConstantExpr::getTrunc(Cst, CI.getType()));
499 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
500 /// in order to eliminate the icmp.
501 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
503 // If we are just checking for a icmp eq of a single bit and zext'ing it
504 // to an integer, then shift the bit to the appropriate place and then
505 // cast to integer to avoid the comparison.
506 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
507 const APInt &Op1CV = Op1C->getValue();
509 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
510 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
511 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
512 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
513 if (!DoXform) return ICI;
515 Value *In = ICI->getOperand(0);
516 Value *Sh = ConstantInt::get(In->getType(),
517 In->getType()->getScalarSizeInBits()-1);
518 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
519 if (In->getType() != CI.getType())
520 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
522 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
523 Constant *One = ConstantInt::get(In->getType(), 1);
524 In = Builder->CreateXor(In, One, In->getName()+".not");
527 return ReplaceInstUsesWith(CI, In);
530 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
531 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
532 // zext (X == 1) to i32 --> X iff X has only the low bit set.
533 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
534 // zext (X != 0) to i32 --> X iff X has only the low bit set.
535 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
536 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
537 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
538 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
539 // This only works for EQ and NE
541 // If Op1C some other power of two, convert:
542 uint32_t BitWidth = Op1C->getType()->getBitWidth();
543 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
544 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
545 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
547 APInt KnownZeroMask(~KnownZero);
548 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
549 if (!DoXform) return ICI;
551 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
552 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
553 // (X&4) == 2 --> false
554 // (X&4) != 2 --> true
555 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
557 Res = ConstantExpr::getZExt(Res, CI.getType());
558 return ReplaceInstUsesWith(CI, Res);
561 uint32_t ShiftAmt = KnownZeroMask.logBase2();
562 Value *In = ICI->getOperand(0);
564 // Perform a logical shr by shiftamt.
565 // Insert the shift to put the result in the low bit.
566 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
567 In->getName()+".lobit");
570 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
571 Constant *One = ConstantInt::get(In->getType(), 1);
572 In = Builder->CreateXor(In, One);
575 if (CI.getType() == In->getType())
576 return ReplaceInstUsesWith(CI, In);
577 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
582 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
583 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
584 // may lead to additional simplifications.
585 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
586 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
587 uint32_t BitWidth = ITy->getBitWidth();
588 Value *LHS = ICI->getOperand(0);
589 Value *RHS = ICI->getOperand(1);
591 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
592 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
593 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
594 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
595 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
597 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
598 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
599 APInt UnknownBit = ~KnownBits;
600 if (UnknownBit.countPopulation() == 1) {
601 if (!DoXform) return ICI;
603 Value *Result = Builder->CreateXor(LHS, RHS);
605 // Mask off any bits that are set and won't be shifted away.
606 if (KnownOneLHS.uge(UnknownBit))
607 Result = Builder->CreateAnd(Result,
608 ConstantInt::get(ITy, UnknownBit));
610 // Shift the bit we're testing down to the lsb.
611 Result = Builder->CreateLShr(
612 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
614 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
615 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
616 Result->takeName(ICI);
617 return ReplaceInstUsesWith(CI, Result);
626 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
627 /// specified wider type and produce the same low bits. If not, return false.
629 /// If this function returns true, it can also return a non-zero number of bits
630 /// (in BitsToClear) which indicates that the value it computes is correct for
631 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
632 /// out. For example, to promote something like:
634 /// %B = trunc i64 %A to i32
635 /// %C = lshr i32 %B, 8
636 /// %E = zext i32 %C to i64
638 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
639 /// set to 8 to indicate that the promoted value needs to have bits 24-31
640 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
641 /// clear the top bits anyway, doing this has no extra cost.
643 /// This function works on both vectors and scalars.
644 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
646 if (isa<Constant>(V))
649 Instruction *I = dyn_cast<Instruction>(V);
650 if (!I) return false;
652 // If the input is a truncate from the destination type, we can trivially
653 // eliminate it, even if it has multiple uses.
654 // FIXME: This is currently disabled until codegen can handle this without
655 // pessimizing code, PR5997.
656 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
659 // We can't extend or shrink something that has multiple uses: doing so would
660 // require duplicating the instruction in general, which isn't profitable.
661 if (!I->hasOneUse()) return false;
663 unsigned Opc = I->getOpcode(), Tmp;
665 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
666 case Instruction::SExt: // zext(sext(x)) -> sext(x).
667 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
669 case Instruction::And:
670 case Instruction::Or:
671 case Instruction::Xor:
672 case Instruction::Add:
673 case Instruction::Sub:
674 case Instruction::Mul:
675 case Instruction::Shl:
676 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
677 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
679 // These can all be promoted if neither operand has 'bits to clear'.
680 if (BitsToClear == 0 && Tmp == 0)
683 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
684 // other side, BitsToClear is ok.
686 (Opc == Instruction::And || Opc == Instruction::Or ||
687 Opc == Instruction::Xor)) {
688 // We use MaskedValueIsZero here for generality, but the case we care
689 // about the most is constant RHS.
690 unsigned VSize = V->getType()->getScalarSizeInBits();
691 if (MaskedValueIsZero(I->getOperand(1),
692 APInt::getHighBitsSet(VSize, BitsToClear)))
696 // Otherwise, we don't know how to analyze this BitsToClear case yet.
699 case Instruction::LShr:
700 // We can promote lshr(x, cst) if we can promote x. This requires the
701 // ultimate 'and' to clear out the high zero bits we're clearing out though.
702 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
703 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
705 BitsToClear += Amt->getZExtValue();
706 if (BitsToClear > V->getType()->getScalarSizeInBits())
707 BitsToClear = V->getType()->getScalarSizeInBits();
710 // Cannot promote variable LSHR.
712 case Instruction::Select:
713 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
714 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
715 // TODO: If important, we could handle the case when the BitsToClear are
716 // known zero in the disagreeing side.
721 case Instruction::PHI: {
722 // We can change a phi if we can change all operands. Note that we never
723 // get into trouble with cyclic PHIs here because we only consider
724 // instructions with a single use.
725 PHINode *PN = cast<PHINode>(I);
726 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
728 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
729 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
730 // TODO: If important, we could handle the case when the BitsToClear
731 // are known zero in the disagreeing input.
737 // TODO: Can handle more cases here.
742 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
743 // If this zero extend is only used by a truncate, let the truncate by
744 // eliminated before we try to optimize this zext.
745 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
748 // If one of the common conversion will work, do it.
749 if (Instruction *Result = commonCastTransforms(CI))
752 // See if we can simplify any instructions used by the input whose sole
753 // purpose is to compute bits we don't care about.
754 if (SimplifyDemandedInstructionBits(CI))
757 Value *Src = CI.getOperand(0);
758 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
760 // Attempt to extend the entire input expression tree to the destination
761 // type. Only do this if the dest type is a simple type, don't convert the
762 // expression tree to something weird like i93 unless the source is also
764 unsigned BitsToClear;
765 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
766 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
767 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
768 "Unreasonable BitsToClear");
770 // Okay, we can transform this! Insert the new expression now.
771 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
772 " to avoid zero extend: " << CI);
773 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
774 assert(Res->getType() == DestTy);
776 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
777 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
779 // If the high bits are already filled with zeros, just replace this
780 // cast with the result.
781 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
782 DestBitSize-SrcBitsKept)))
783 return ReplaceInstUsesWith(CI, Res);
785 // We need to emit an AND to clear the high bits.
786 Constant *C = ConstantInt::get(Res->getType(),
787 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
788 return BinaryOperator::CreateAnd(Res, C);
791 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
792 // types and if the sizes are just right we can convert this into a logical
793 // 'and' which will be much cheaper than the pair of casts.
794 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
795 // TODO: Subsume this into EvaluateInDifferentType.
797 // Get the sizes of the types involved. We know that the intermediate type
798 // will be smaller than A or C, but don't know the relation between A and C.
799 Value *A = CSrc->getOperand(0);
800 unsigned SrcSize = A->getType()->getScalarSizeInBits();
801 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
802 unsigned DstSize = CI.getType()->getScalarSizeInBits();
803 // If we're actually extending zero bits, then if
804 // SrcSize < DstSize: zext(a & mask)
805 // SrcSize == DstSize: a & mask
806 // SrcSize > DstSize: trunc(a) & mask
807 if (SrcSize < DstSize) {
808 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
809 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
810 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
811 return new ZExtInst(And, CI.getType());
814 if (SrcSize == DstSize) {
815 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
816 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
819 if (SrcSize > DstSize) {
820 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
821 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
822 return BinaryOperator::CreateAnd(Trunc,
823 ConstantInt::get(Trunc->getType(),
828 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
829 return transformZExtICmp(ICI, CI);
831 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
832 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
833 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
834 // of the (zext icmp) will be transformed.
835 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
836 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
837 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
838 (transformZExtICmp(LHS, CI, false) ||
839 transformZExtICmp(RHS, CI, false))) {
840 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
841 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
842 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
846 // zext(trunc(t) & C) -> (t & zext(C)).
847 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
848 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
849 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
850 Value *TI0 = TI->getOperand(0);
851 if (TI0->getType() == CI.getType())
853 BinaryOperator::CreateAnd(TI0,
854 ConstantExpr::getZExt(C, CI.getType()));
857 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
858 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
859 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
860 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
861 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
862 And->getOperand(1) == C)
863 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
864 Value *TI0 = TI->getOperand(0);
865 if (TI0->getType() == CI.getType()) {
866 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
867 Value *NewAnd = Builder->CreateAnd(TI0, ZC);
868 return BinaryOperator::CreateXor(NewAnd, ZC);
872 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
874 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
875 match(SrcI, m_Not(m_Value(X))) &&
876 (!X->hasOneUse() || !isa<CmpInst>(X))) {
877 Value *New = Builder->CreateZExt(X, CI.getType());
878 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
884 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
885 /// in order to eliminate the icmp.
886 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
887 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
888 ICmpInst::Predicate Pred = ICI->getPredicate();
890 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
891 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
892 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
893 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) ||
894 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
896 Value *Sh = ConstantInt::get(Op0->getType(),
897 Op0->getType()->getScalarSizeInBits()-1);
898 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
899 if (In->getType() != CI.getType())
900 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
902 if (Pred == ICmpInst::ICMP_SGT)
903 In = Builder->CreateNot(In, In->getName()+".not");
904 return ReplaceInstUsesWith(CI, In);
907 // If we know that only one bit of the LHS of the icmp can be set and we
908 // have an equality comparison with zero or a power of 2, we can transform
909 // the icmp and sext into bitwise/integer operations.
910 if (ICI->hasOneUse() &&
911 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
912 unsigned BitWidth = Op1C->getType()->getBitWidth();
913 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
914 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
915 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
917 APInt KnownZeroMask(~KnownZero);
918 if (KnownZeroMask.isPowerOf2()) {
919 Value *In = ICI->getOperand(0);
921 // If the icmp tests for a known zero bit we can constant fold it.
922 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
923 Value *V = Pred == ICmpInst::ICMP_NE ?
924 ConstantInt::getAllOnesValue(CI.getType()) :
925 ConstantInt::getNullValue(CI.getType());
926 return ReplaceInstUsesWith(CI, V);
929 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
930 // sext ((x & 2^n) == 0) -> (x >> n) - 1
931 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
932 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
933 // Perform a right shift to place the desired bit in the LSB.
935 In = Builder->CreateLShr(In,
936 ConstantInt::get(In->getType(), ShiftAmt));
938 // At this point "In" is either 1 or 0. Subtract 1 to turn
939 // {1, 0} -> {0, -1}.
940 In = Builder->CreateAdd(In,
941 ConstantInt::getAllOnesValue(In->getType()),
944 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
945 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
946 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
947 // Perform a left shift to place the desired bit in the MSB.
949 In = Builder->CreateShl(In,
950 ConstantInt::get(In->getType(), ShiftAmt));
952 // Distribute the bit over the whole bit width.
953 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
954 BitWidth - 1), "sext");
957 if (CI.getType() == In->getType())
958 return ReplaceInstUsesWith(CI, In);
959 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
964 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
965 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) {
966 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) &&
967 Op0->getType() == CI.getType()) {
968 Type *EltTy = VTy->getElementType();
970 // splat the shift constant to a constant vector.
971 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
972 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit");
973 return ReplaceInstUsesWith(CI, In);
980 /// CanEvaluateSExtd - Return true if we can take the specified value
981 /// and return it as type Ty without inserting any new casts and without
982 /// changing the value of the common low bits. This is used by code that tries
983 /// to promote integer operations to a wider types will allow us to eliminate
986 /// This function works on both vectors and scalars.
988 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
989 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
990 "Can't sign extend type to a smaller type");
991 // If this is a constant, it can be trivially promoted.
992 if (isa<Constant>(V))
995 Instruction *I = dyn_cast<Instruction>(V);
996 if (!I) return false;
998 // If this is a truncate from the dest type, we can trivially eliminate it,
999 // even if it has multiple uses.
1000 // FIXME: This is currently disabled until codegen can handle this without
1001 // pessimizing code, PR5997.
1002 if (0 && isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1005 // We can't extend or shrink something that has multiple uses: doing so would
1006 // require duplicating the instruction in general, which isn't profitable.
1007 if (!I->hasOneUse()) return false;
1009 switch (I->getOpcode()) {
1010 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1011 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1012 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1014 case Instruction::And:
1015 case Instruction::Or:
1016 case Instruction::Xor:
1017 case Instruction::Add:
1018 case Instruction::Sub:
1019 case Instruction::Mul:
1020 // These operators can all arbitrarily be extended if their inputs can.
1021 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1022 CanEvaluateSExtd(I->getOperand(1), Ty);
1024 //case Instruction::Shl: TODO
1025 //case Instruction::LShr: TODO
1027 case Instruction::Select:
1028 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1029 CanEvaluateSExtd(I->getOperand(2), Ty);
1031 case Instruction::PHI: {
1032 // We can change a phi if we can change all operands. Note that we never
1033 // get into trouble with cyclic PHIs here because we only consider
1034 // instructions with a single use.
1035 PHINode *PN = cast<PHINode>(I);
1036 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1037 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1041 // TODO: Can handle more cases here.
1048 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1049 // If this sign extend is only used by a truncate, let the truncate by
1050 // eliminated before we try to optimize this zext.
1051 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
1054 if (Instruction *I = commonCastTransforms(CI))
1057 // See if we can simplify any instructions used by the input whose sole
1058 // purpose is to compute bits we don't care about.
1059 if (SimplifyDemandedInstructionBits(CI))
1062 Value *Src = CI.getOperand(0);
1063 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1065 // Attempt to extend the entire input expression tree to the destination
1066 // type. Only do this if the dest type is a simple type, don't convert the
1067 // expression tree to something weird like i93 unless the source is also
1069 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1070 CanEvaluateSExtd(Src, DestTy)) {
1071 // Okay, we can transform this! Insert the new expression now.
1072 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1073 " to avoid sign extend: " << CI);
1074 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1075 assert(Res->getType() == DestTy);
1077 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1078 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1080 // If the high bits are already filled with sign bit, just replace this
1081 // cast with the result.
1082 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1083 return ReplaceInstUsesWith(CI, Res);
1085 // We need to emit a shl + ashr to do the sign extend.
1086 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1087 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1091 // If this input is a trunc from our destination, then turn sext(trunc(x))
1093 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1094 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1095 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1096 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1098 // We need to emit a shl + ashr to do the sign extend.
1099 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1100 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1101 return BinaryOperator::CreateAShr(Res, ShAmt);
1104 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1105 return transformSExtICmp(ICI, CI);
1107 // If the input is a shl/ashr pair of a same constant, then this is a sign
1108 // extension from a smaller value. If we could trust arbitrary bitwidth
1109 // integers, we could turn this into a truncate to the smaller bit and then
1110 // use a sext for the whole extension. Since we don't, look deeper and check
1111 // for a truncate. If the source and dest are the same type, eliminate the
1112 // trunc and extend and just do shifts. For example, turn:
1113 // %a = trunc i32 %i to i8
1114 // %b = shl i8 %a, 6
1115 // %c = ashr i8 %b, 6
1116 // %d = sext i8 %c to i32
1118 // %a = shl i32 %i, 30
1119 // %d = ashr i32 %a, 30
1121 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1122 ConstantInt *BA = 0, *CA = 0;
1123 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1124 m_ConstantInt(CA))) &&
1125 BA == CA && A->getType() == CI.getType()) {
1126 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1127 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1128 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1129 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1130 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1131 return BinaryOperator::CreateAShr(A, ShAmtV);
1138 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1139 /// in the specified FP type without changing its value.
1140 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1142 APFloat F = CFP->getValueAPF();
1143 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1145 return ConstantFP::get(CFP->getContext(), F);
1149 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1150 /// through it until we get the source value.
1151 static Value *LookThroughFPExtensions(Value *V) {
1152 if (Instruction *I = dyn_cast<Instruction>(V))
1153 if (I->getOpcode() == Instruction::FPExt)
1154 return LookThroughFPExtensions(I->getOperand(0));
1156 // If this value is a constant, return the constant in the smallest FP type
1157 // that can accurately represent it. This allows us to turn
1158 // (float)((double)X+2.0) into x+2.0f.
1159 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1160 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1161 return V; // No constant folding of this.
1162 // See if the value can be truncated to half and then reextended.
1163 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
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 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1216 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1217 Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) &&
1218 Call->getNumArgOperands() == 1 &&
1219 Call->hasOneUse()) {
1220 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1221 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1222 CI.getType()->isFloatTy() &&
1223 Call->getType()->isDoubleTy() &&
1224 Arg->getType()->isDoubleTy() &&
1225 Arg->getOperand(0)->getType()->isFloatTy()) {
1226 Function *Callee = Call->getCalledFunction();
1227 Module *M = CI.getParent()->getParent()->getParent();
1228 Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1229 Callee->getAttributes(),
1230 Builder->getFloatTy(),
1231 Builder->getFloatTy(),
1233 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1235 ret->setAttributes(Callee->getAttributes());
1238 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1239 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1240 EraseInstFromFunction(*Call);
1248 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1249 return commonCastTransforms(CI);
1252 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1253 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1255 return commonCastTransforms(FI);
1257 // fptoui(uitofp(X)) --> X
1258 // fptoui(sitofp(X)) --> X
1259 // This is safe if the intermediate type has enough bits in its mantissa to
1260 // accurately represent all values of X. For example, do not do this with
1261 // i64->float->i64. This is also safe for sitofp case, because any negative
1262 // 'X' value would cause an undefined result for the fptoui.
1263 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1264 OpI->getOperand(0)->getType() == FI.getType() &&
1265 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1266 OpI->getType()->getFPMantissaWidth())
1267 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1269 return commonCastTransforms(FI);
1272 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1273 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1275 return commonCastTransforms(FI);
1277 // fptosi(sitofp(X)) --> X
1278 // fptosi(uitofp(X)) --> X
1279 // This is safe if the intermediate type has enough bits in its mantissa to
1280 // accurately represent all values of X. For example, do not do this with
1281 // i64->float->i64. This is also safe for sitofp case, because any negative
1282 // 'X' value would cause an undefined result for the fptoui.
1283 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1284 OpI->getOperand(0)->getType() == FI.getType() &&
1285 (int)FI.getType()->getScalarSizeInBits() <=
1286 OpI->getType()->getFPMantissaWidth())
1287 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1289 return commonCastTransforms(FI);
1292 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1293 return commonCastTransforms(CI);
1296 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1297 return commonCastTransforms(CI);
1300 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1301 // If the source integer type is not the intptr_t type for this target, do a
1302 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1303 // cast to be exposed to other transforms.
1305 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
1306 TD->getPointerSizeInBits()) {
1307 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1308 TD->getIntPtrType(CI.getContext()));
1309 return new IntToPtrInst(P, CI.getType());
1311 if (CI.getOperand(0)->getType()->getScalarSizeInBits() <
1312 TD->getPointerSizeInBits()) {
1313 Value *P = Builder->CreateZExt(CI.getOperand(0),
1314 TD->getIntPtrType(CI.getContext()));
1315 return new IntToPtrInst(P, CI.getType());
1319 if (Instruction *I = commonCastTransforms(CI))
1325 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1326 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1327 Value *Src = CI.getOperand(0);
1329 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1330 // If casting the result of a getelementptr instruction with no offset, turn
1331 // this into a cast of the original pointer!
1332 if (GEP->hasAllZeroIndices()) {
1333 // Changing the cast operand is usually not a good idea but it is safe
1334 // here because the pointer operand is being replaced with another
1335 // pointer operand so the opcode doesn't need to change.
1337 CI.setOperand(0, GEP->getOperand(0));
1341 // If the GEP has a single use, and the base pointer is a bitcast, and the
1342 // GEP computes a constant offset, see if we can convert these three
1343 // instructions into fewer. This typically happens with unions and other
1344 // non-type-safe code.
1345 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1346 GEP->hasAllConstantIndices()) {
1347 // We are guaranteed to get a constant from EmitGEPOffset.
1348 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
1349 int64_t Offset = OffsetV->getSExtValue();
1351 // Get the base pointer input of the bitcast, and the type it points to.
1352 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1354 cast<PointerType>(OrigBase->getType())->getElementType();
1355 SmallVector<Value*, 8> NewIndices;
1356 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
1357 // If we were able to index down into an element, create the GEP
1358 // and bitcast the result. This eliminates one bitcast, potentially
1360 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1361 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1362 Builder->CreateGEP(OrigBase, NewIndices);
1363 NGEP->takeName(GEP);
1365 if (isa<BitCastInst>(CI))
1366 return new BitCastInst(NGEP, CI.getType());
1367 assert(isa<PtrToIntInst>(CI));
1368 return new PtrToIntInst(NGEP, CI.getType());
1373 return commonCastTransforms(CI);
1376 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1377 // If the destination integer type is not the intptr_t type for this target,
1378 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1379 // to be exposed to other transforms.
1381 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1382 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1383 TD->getIntPtrType(CI.getContext()));
1384 return new TruncInst(P, CI.getType());
1386 if (CI.getType()->getScalarSizeInBits() > TD->getPointerSizeInBits()) {
1387 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1388 TD->getIntPtrType(CI.getContext()));
1389 return new ZExtInst(P, CI.getType());
1393 return commonPointerCastTransforms(CI);
1396 /// OptimizeVectorResize - This input value (which is known to have vector type)
1397 /// is being zero extended or truncated to the specified vector type. Try to
1398 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1400 /// The source and destination vector types may have different element types.
1401 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1403 // We can only do this optimization if the output is a multiple of the input
1404 // element size, or the input is a multiple of the output element size.
1405 // Convert the input type to have the same element type as the output.
1406 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1408 if (SrcTy->getElementType() != DestTy->getElementType()) {
1409 // The input types don't need to be identical, but for now they must be the
1410 // same size. There is no specific reason we couldn't handle things like
1411 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1413 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1414 DestTy->getElementType()->getPrimitiveSizeInBits())
1417 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1418 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1421 // Now that the element types match, get the shuffle mask and RHS of the
1422 // shuffle to use, which depends on whether we're increasing or decreasing the
1423 // size of the input.
1424 SmallVector<uint32_t, 16> ShuffleMask;
1427 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1428 // If we're shrinking the number of elements, just shuffle in the low
1429 // elements from the input and use undef as the second shuffle input.
1430 V2 = UndefValue::get(SrcTy);
1431 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1432 ShuffleMask.push_back(i);
1435 // If we're increasing the number of elements, shuffle in all of the
1436 // elements from InVal and fill the rest of the result elements with zeros
1437 // from a constant zero.
1438 V2 = Constant::getNullValue(SrcTy);
1439 unsigned SrcElts = SrcTy->getNumElements();
1440 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1441 ShuffleMask.push_back(i);
1443 // The excess elements reference the first element of the zero input.
1444 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1445 ShuffleMask.push_back(SrcElts);
1448 return new ShuffleVectorInst(InVal, V2,
1449 ConstantDataVector::get(V2->getContext(),
1453 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1454 return Value % Ty->getPrimitiveSizeInBits() == 0;
1457 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1458 return Value / Ty->getPrimitiveSizeInBits();
1461 /// CollectInsertionElements - V is a value which is inserted into a vector of
1462 /// VecEltTy. Look through the value to see if we can decompose it into
1463 /// insertions into the vector. See the example in the comment for
1464 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1465 /// The type of V is always a non-zero multiple of VecEltTy's size.
1467 /// This returns false if the pattern can't be matched or true if it can,
1468 /// filling in Elements with the elements found here.
1469 static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1470 SmallVectorImpl<Value*> &Elements,
1472 // Undef values never contribute useful bits to the result.
1473 if (isa<UndefValue>(V)) return true;
1475 // If we got down to a value of the right type, we win, try inserting into the
1477 if (V->getType() == VecEltTy) {
1478 // Inserting null doesn't actually insert any elements.
1479 if (Constant *C = dyn_cast<Constant>(V))
1480 if (C->isNullValue())
1483 // Fail if multiple elements are inserted into this slot.
1484 if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1487 Elements[ElementIndex] = V;
1491 if (Constant *C = dyn_cast<Constant>(V)) {
1492 // Figure out the # elements this provides, and bitcast it or slice it up
1494 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1496 // If the constant is the size of a vector element, we just need to bitcast
1497 // it to the right type so it gets properly inserted.
1499 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1500 ElementIndex, Elements, VecEltTy);
1502 // Okay, this is a constant that covers multiple elements. Slice it up into
1503 // pieces and insert each element-sized piece into the vector.
1504 if (!isa<IntegerType>(C->getType()))
1505 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1506 C->getType()->getPrimitiveSizeInBits()));
1507 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1508 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1510 for (unsigned i = 0; i != NumElts; ++i) {
1511 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1513 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1514 if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1520 if (!V->hasOneUse()) return false;
1522 Instruction *I = dyn_cast<Instruction>(V);
1523 if (I == 0) return false;
1524 switch (I->getOpcode()) {
1525 default: return false; // Unhandled case.
1526 case Instruction::BitCast:
1527 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1528 Elements, VecEltTy);
1529 case Instruction::ZExt:
1530 if (!isMultipleOfTypeSize(
1531 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1534 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1535 Elements, VecEltTy);
1536 case Instruction::Or:
1537 return CollectInsertionElements(I->getOperand(0), ElementIndex,
1538 Elements, VecEltTy) &&
1539 CollectInsertionElements(I->getOperand(1), ElementIndex,
1540 Elements, VecEltTy);
1541 case Instruction::Shl: {
1542 // Must be shifting by a constant that is a multiple of the element size.
1543 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1544 if (CI == 0) return false;
1545 if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1546 unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1548 return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1549 Elements, VecEltTy);
1556 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1557 /// may be doing shifts and ors to assemble the elements of the vector manually.
1558 /// Try to rip the code out and replace it with insertelements. This is to
1559 /// optimize code like this:
1561 /// %tmp37 = bitcast float %inc to i32
1562 /// %tmp38 = zext i32 %tmp37 to i64
1563 /// %tmp31 = bitcast float %inc5 to i32
1564 /// %tmp32 = zext i32 %tmp31 to i64
1565 /// %tmp33 = shl i64 %tmp32, 32
1566 /// %ins35 = or i64 %tmp33, %tmp38
1567 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1569 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1570 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1572 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1573 Value *IntInput = CI.getOperand(0);
1575 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1576 if (!CollectInsertionElements(IntInput, 0, Elements,
1577 DestVecTy->getElementType()))
1580 // If we succeeded, we know that all of the element are specified by Elements
1581 // or are zero if Elements has a null entry. Recast this as a set of
1583 Value *Result = Constant::getNullValue(CI.getType());
1584 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1585 if (Elements[i] == 0) continue; // Unset element.
1587 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1588 IC.Builder->getInt32(i));
1595 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1596 /// bitcast. The various long double bitcasts can't get in here.
1597 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1598 Value *Src = CI.getOperand(0);
1599 Type *DestTy = CI.getType();
1601 // If this is a bitcast from int to float, check to see if the int is an
1602 // extraction from a vector.
1603 Value *VecInput = 0;
1604 // bitcast(trunc(bitcast(somevector)))
1605 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1606 isa<VectorType>(VecInput->getType())) {
1607 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1608 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1610 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1611 // If the element type of the vector doesn't match the result type,
1612 // bitcast it to be a vector type we can extract from.
1613 if (VecTy->getElementType() != DestTy) {
1614 VecTy = VectorType::get(DestTy,
1615 VecTy->getPrimitiveSizeInBits() / DestWidth);
1616 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1619 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(0));
1623 // bitcast(trunc(lshr(bitcast(somevector), cst))
1624 ConstantInt *ShAmt = 0;
1625 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1626 m_ConstantInt(ShAmt)))) &&
1627 isa<VectorType>(VecInput->getType())) {
1628 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1629 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1630 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1631 ShAmt->getZExtValue() % DestWidth == 0) {
1632 // If the element type of the vector doesn't match the result type,
1633 // bitcast it to be a vector type we can extract from.
1634 if (VecTy->getElementType() != DestTy) {
1635 VecTy = VectorType::get(DestTy,
1636 VecTy->getPrimitiveSizeInBits() / DestWidth);
1637 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1640 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1641 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1647 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1648 // If the operands are integer typed then apply the integer transforms,
1649 // otherwise just apply the common ones.
1650 Value *Src = CI.getOperand(0);
1651 Type *SrcTy = Src->getType();
1652 Type *DestTy = CI.getType();
1654 // Get rid of casts from one type to the same type. These are useless and can
1655 // be replaced by the operand.
1656 if (DestTy == Src->getType())
1657 return ReplaceInstUsesWith(CI, Src);
1659 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1660 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1661 Type *DstElTy = DstPTy->getElementType();
1662 Type *SrcElTy = SrcPTy->getElementType();
1664 // If the address spaces don't match, don't eliminate the bitcast, which is
1665 // required for changing types.
1666 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1669 // If we are casting a alloca to a pointer to a type of the same
1670 // size, rewrite the allocation instruction to allocate the "right" type.
1671 // There is no need to modify malloc calls because it is their bitcast that
1672 // needs to be cleaned up.
1673 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1674 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1677 // If the source and destination are pointers, and this cast is equivalent
1678 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1679 // This can enhance SROA and other transforms that want type-safe pointers.
1680 Constant *ZeroUInt =
1681 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1682 unsigned NumZeros = 0;
1683 while (SrcElTy != DstElTy &&
1684 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1685 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1686 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1690 // If we found a path from the src to dest, create the getelementptr now.
1691 if (SrcElTy == DstElTy) {
1692 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1693 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1697 // Try to optimize int -> float bitcasts.
1698 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1699 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1702 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1703 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1704 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1705 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1706 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1707 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1710 if (isa<IntegerType>(SrcTy)) {
1711 // If this is a cast from an integer to vector, check to see if the input
1712 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1713 // the casts with a shuffle and (potentially) a bitcast.
1714 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1715 CastInst *SrcCast = cast<CastInst>(Src);
1716 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1717 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1718 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1719 cast<VectorType>(DestTy), *this))
1723 // If the input is an 'or' instruction, we may be doing shifts and ors to
1724 // assemble the elements of the vector manually. Try to rip the code out
1725 // and replace it with insertelements.
1726 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1727 return ReplaceInstUsesWith(CI, V);
1731 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1732 if (SrcVTy->getNumElements() == 1 && !DestTy->isVectorTy()) {
1734 Builder->CreateExtractElement(Src,
1735 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1736 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1740 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1741 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1742 // a bitcast to a vector with the same # elts.
1743 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1744 cast<VectorType>(DestTy)->getNumElements() ==
1745 SVI->getType()->getNumElements() &&
1746 SVI->getType()->getNumElements() ==
1747 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1749 // If either of the operands is a cast from CI.getType(), then
1750 // evaluating the shuffle in the casted destination's type will allow
1751 // us to eliminate at least one cast.
1752 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1753 Tmp->getOperand(0)->getType() == DestTy) ||
1754 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1755 Tmp->getOperand(0)->getType() == DestTy)) {
1756 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1757 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1758 // Return a new shuffle vector. Use the same element ID's, as we
1759 // know the vector types match #elts.
1760 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1765 if (SrcTy->isPointerTy())
1766 return commonPointerCastTransforms(CI);
1767 return commonCastTransforms(CI);