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/IR/DataLayout.h"
17 #include "llvm/Support/PatternMatch.h"
18 #include "llvm/Target/TargetLibraryInfo.h"
20 using namespace PatternMatch;
22 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
23 /// expression. If so, decompose it, returning some value X, such that Val is
26 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
28 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
29 Offset = CI->getZExtValue();
31 return ConstantInt::get(Val->getType(), 0);
34 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
35 // Cannot look past anything that might overflow.
36 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
37 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
43 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
44 if (I->getOpcode() == Instruction::Shl) {
45 // This is a value scaled by '1 << the shift amt'.
46 Scale = UINT64_C(1) << RHS->getZExtValue();
48 return I->getOperand(0);
51 if (I->getOpcode() == Instruction::Mul) {
52 // This value is scaled by 'RHS'.
53 Scale = RHS->getZExtValue();
55 return I->getOperand(0);
58 if (I->getOpcode() == Instruction::Add) {
59 // We have X+C. Check to see if we really have (X*C2)+C1,
60 // where C1 is divisible by C2.
63 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
64 Offset += RHS->getZExtValue();
71 // Otherwise, we can't look past this.
77 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
78 /// try to eliminate the cast by moving the type information into the alloc.
79 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
81 // This requires DataLayout to get the alloca alignment and size information.
84 PointerType *PTy = cast<PointerType>(CI.getType());
86 BuilderTy AllocaBuilder(*Builder);
87 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
89 // Get the type really allocated and the type casted to.
90 Type *AllocElTy = AI.getAllocatedType();
91 Type *CastElTy = PTy->getElementType();
92 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
94 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
95 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
96 if (CastElTyAlign < AllocElTyAlign) return 0;
98 // If the allocation has multiple uses, only promote it if we are strictly
99 // increasing the alignment of the resultant allocation. If we keep it the
100 // same, we open the door to infinite loops of various kinds.
101 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
103 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
104 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
105 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
107 // If the allocation has multiple uses, only promote it if we're not
108 // shrinking the amount of memory being allocated.
109 uint64_t AllocElTyStoreSize = TD->getTypeStoreSize(AllocElTy);
110 uint64_t CastElTyStoreSize = TD->getTypeStoreSize(CastElTy);
111 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return 0;
113 // See if we can satisfy the modulus by pulling a scale out of the array
115 unsigned ArraySizeScale;
116 uint64_t ArrayOffset;
117 Value *NumElements = // See if the array size is a decomposable linear expr.
118 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
120 // If we can now satisfy the modulus, by using a non-1 scale, we really can
122 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
123 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
125 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
130 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
131 // Insert before the alloca, not before the cast.
132 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
135 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
136 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
138 Amt = AllocaBuilder.CreateAdd(Amt, Off);
141 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
142 New->setAlignment(AI.getAlignment());
145 // If the allocation has multiple real uses, insert a cast and change all
146 // things that used it to use the new cast. This will also hack on CI, but it
148 if (!AI.hasOneUse()) {
149 // New is the allocation instruction, pointer typed. AI is the original
150 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
151 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
152 ReplaceInstUsesWith(AI, NewCast);
154 return ReplaceInstUsesWith(CI, New);
157 /// EvaluateInDifferentType - Given an expression that
158 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
159 /// insert the code to evaluate the expression.
160 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
162 if (Constant *C = dyn_cast<Constant>(V)) {
163 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
164 // If we got a constantexpr back, try to simplify it with TD info.
165 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
166 C = ConstantFoldConstantExpression(CE, TD, TLI);
170 // Otherwise, it must be an instruction.
171 Instruction *I = cast<Instruction>(V);
172 Instruction *Res = 0;
173 unsigned Opc = I->getOpcode();
175 case Instruction::Add:
176 case Instruction::Sub:
177 case Instruction::Mul:
178 case Instruction::And:
179 case Instruction::Or:
180 case Instruction::Xor:
181 case Instruction::AShr:
182 case Instruction::LShr:
183 case Instruction::Shl:
184 case Instruction::UDiv:
185 case Instruction::URem: {
186 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
187 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
188 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
191 case Instruction::Trunc:
192 case Instruction::ZExt:
193 case Instruction::SExt:
194 // If the source type of the cast is the type we're trying for then we can
195 // just return the source. There's no need to insert it because it is not
197 if (I->getOperand(0)->getType() == Ty)
198 return I->getOperand(0);
200 // Otherwise, must be the same type of cast, so just reinsert a new one.
201 // This also handles the case of zext(trunc(x)) -> zext(x).
202 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
203 Opc == Instruction::SExt);
205 case Instruction::Select: {
206 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
207 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
208 Res = SelectInst::Create(I->getOperand(0), True, False);
211 case Instruction::PHI: {
212 PHINode *OPN = cast<PHINode>(I);
213 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
214 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
215 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
216 NPN->addIncoming(V, OPN->getIncomingBlock(i));
222 // TODO: Can handle more cases here.
223 llvm_unreachable("Unreachable!");
227 return InsertNewInstWith(Res, *I);
231 /// This function is a wrapper around CastInst::isEliminableCastPair. It
232 /// simply extracts arguments and returns what that function returns.
233 static Instruction::CastOps
234 isEliminableCastPair(
235 const CastInst *CI, ///< The first cast instruction
236 unsigned opcode, ///< The opcode of the second cast instruction
237 Type *DstTy, ///< The target type for the second cast instruction
238 DataLayout *TD ///< The target data for pointer size
241 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
242 Type *MidTy = CI->getType(); // B from above
244 // Get the opcodes of the two Cast instructions
245 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
246 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
247 Type *SrcIntPtrTy = TD && SrcTy->isPtrOrPtrVectorTy() ?
248 TD->getIntPtrType(SrcTy) : 0;
249 Type *MidIntPtrTy = TD && MidTy->isPtrOrPtrVectorTy() ?
250 TD->getIntPtrType(MidTy) : 0;
251 Type *DstIntPtrTy = TD && DstTy->isPtrOrPtrVectorTy() ?
252 TD->getIntPtrType(DstTy) : 0;
253 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
254 DstTy, SrcIntPtrTy, MidIntPtrTy,
257 // We don't want to form an inttoptr or ptrtoint that converts to an integer
258 // type that differs from the pointer size.
259 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
260 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
263 return Instruction::CastOps(Res);
266 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
267 /// results in any code being generated and is interesting to optimize out. If
268 /// the cast can be eliminated by some other simple transformation, we prefer
269 /// to do the simplification first.
270 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
272 // Noop casts and casts of constants should be eliminated trivially.
273 if (V->getType() == Ty || isa<Constant>(V)) return false;
275 // If this is another cast that can be eliminated, we prefer to have it
277 if (const CastInst *CI = dyn_cast<CastInst>(V))
278 if (isEliminableCastPair(CI, opc, Ty, TD))
281 // If this is a vector sext from a compare, then we don't want to break the
282 // idiom where each element of the extended vector is either zero or all ones.
283 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
290 /// @brief Implement the transforms common to all CastInst visitors.
291 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
292 Value *Src = CI.getOperand(0);
294 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
296 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
297 if (Instruction::CastOps opc =
298 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
299 // The first cast (CSrc) is eliminable so we need to fix up or replace
300 // the second cast (CI). CSrc will then have a good chance of being dead.
301 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
305 // If we are casting a select then fold the cast into the select
306 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
307 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
310 // If we are casting a PHI then fold the cast into the PHI
311 if (isa<PHINode>(Src)) {
312 // We don't do this if this would create a PHI node with an illegal type if
313 // it is currently legal.
314 if (!Src->getType()->isIntegerTy() ||
315 !CI.getType()->isIntegerTy() ||
316 ShouldChangeType(CI.getType(), Src->getType()))
317 if (Instruction *NV = FoldOpIntoPhi(CI))
324 /// CanEvaluateTruncated - Return true if we can evaluate the specified
325 /// expression tree as type Ty instead of its larger type, and arrive with the
326 /// same value. This is used by code that tries to eliminate truncates.
328 /// Ty will always be a type smaller than V. We should return true if trunc(V)
329 /// can be computed by computing V in the smaller type. If V is an instruction,
330 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
331 /// makes sense if x and y can be efficiently truncated.
333 /// This function works on both vectors and scalars.
335 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
336 // We can always evaluate constants in another type.
337 if (isa<Constant>(V))
340 Instruction *I = dyn_cast<Instruction>(V);
341 if (!I) return false;
343 Type *OrigTy = V->getType();
345 // If this is an extension from the dest type, we can eliminate it, even if it
346 // has multiple uses.
347 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
348 I->getOperand(0)->getType() == Ty)
351 // We can't extend or shrink something that has multiple uses: doing so would
352 // require duplicating the instruction in general, which isn't profitable.
353 if (!I->hasOneUse()) return false;
355 unsigned Opc = I->getOpcode();
357 case Instruction::Add:
358 case Instruction::Sub:
359 case Instruction::Mul:
360 case Instruction::And:
361 case Instruction::Or:
362 case Instruction::Xor:
363 // These operators can all arbitrarily be extended or truncated.
364 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
365 CanEvaluateTruncated(I->getOperand(1), Ty);
367 case Instruction::UDiv:
368 case Instruction::URem: {
369 // UDiv and URem can be truncated if all the truncated bits are zero.
370 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
371 uint32_t BitWidth = Ty->getScalarSizeInBits();
372 if (BitWidth < OrigBitWidth) {
373 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
374 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
375 MaskedValueIsZero(I->getOperand(1), Mask)) {
376 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
377 CanEvaluateTruncated(I->getOperand(1), Ty);
382 case Instruction::Shl:
383 // If we are truncating the result of this SHL, and if it's a shift of a
384 // constant amount, we can always perform a SHL in a smaller type.
385 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
386 uint32_t BitWidth = Ty->getScalarSizeInBits();
387 if (CI->getLimitedValue(BitWidth) < BitWidth)
388 return CanEvaluateTruncated(I->getOperand(0), Ty);
391 case Instruction::LShr:
392 // If this is a truncate of a logical shr, we can truncate it to a smaller
393 // lshr iff we know that the bits we would otherwise be shifting in are
395 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
396 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
397 uint32_t BitWidth = Ty->getScalarSizeInBits();
398 if (MaskedValueIsZero(I->getOperand(0),
399 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
400 CI->getLimitedValue(BitWidth) < BitWidth) {
401 return CanEvaluateTruncated(I->getOperand(0), Ty);
405 case Instruction::Trunc:
406 // trunc(trunc(x)) -> trunc(x)
408 case Instruction::ZExt:
409 case Instruction::SExt:
410 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
411 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
413 case Instruction::Select: {
414 SelectInst *SI = cast<SelectInst>(I);
415 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
416 CanEvaluateTruncated(SI->getFalseValue(), Ty);
418 case Instruction::PHI: {
419 // We can change a phi if we can change all operands. Note that we never
420 // get into trouble with cyclic PHIs here because we only consider
421 // instructions with a single use.
422 PHINode *PN = cast<PHINode>(I);
423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
424 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
429 // TODO: Can handle more cases here.
436 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
437 if (Instruction *Result = commonCastTransforms(CI))
440 // See if we can simplify any instructions used by the input whose sole
441 // purpose is to compute bits we don't care about.
442 if (SimplifyDemandedInstructionBits(CI))
445 Value *Src = CI.getOperand(0);
446 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
448 // Attempt to truncate the entire input expression tree to the destination
449 // type. Only do this if the dest type is a simple type, don't convert the
450 // expression tree to something weird like i93 unless the source is also
452 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
453 CanEvaluateTruncated(Src, DestTy)) {
455 // If this cast is a truncate, evaluting in a different type always
456 // eliminates the cast, so it is always a win.
457 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
458 " to avoid cast: " << CI << '\n');
459 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
460 assert(Res->getType() == DestTy);
461 return ReplaceInstUsesWith(CI, Res);
464 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
465 if (DestTy->getScalarSizeInBits() == 1) {
466 Constant *One = ConstantInt::get(Src->getType(), 1);
467 Src = Builder->CreateAnd(Src, One);
468 Value *Zero = Constant::getNullValue(Src->getType());
469 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
472 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
473 Value *A = 0; ConstantInt *Cst = 0;
474 if (Src->hasOneUse() &&
475 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
476 // We have three types to worry about here, the type of A, the source of
477 // the truncate (MidSize), and the destination of the truncate. We know that
478 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
479 // between ASize and ResultSize.
480 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
482 // If the shift amount is larger than the size of A, then the result is
483 // known to be zero because all the input bits got shifted out.
484 if (Cst->getZExtValue() >= ASize)
485 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
487 // Since we're doing an lshr and a zero extend, and know that the shift
488 // amount is smaller than ASize, it is always safe to do the shift in A's
489 // type, then zero extend or truncate to the result.
490 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
491 Shift->takeName(Src);
492 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
495 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
496 // type isn't non-native.
497 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
498 ShouldChangeType(Src->getType(), CI.getType()) &&
499 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
500 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
501 return BinaryOperator::CreateAnd(NewTrunc,
502 ConstantExpr::getTrunc(Cst, CI.getType()));
508 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
509 /// in order to eliminate the icmp.
510 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
512 // If we are just checking for a icmp eq of a single bit and zext'ing it
513 // to an integer, then shift the bit to the appropriate place and then
514 // cast to integer to avoid the comparison.
515 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
516 const APInt &Op1CV = Op1C->getValue();
518 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
519 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
520 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
521 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
522 if (!DoXform) return ICI;
524 Value *In = ICI->getOperand(0);
525 Value *Sh = ConstantInt::get(In->getType(),
526 In->getType()->getScalarSizeInBits()-1);
527 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
528 if (In->getType() != CI.getType())
529 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
531 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
532 Constant *One = ConstantInt::get(In->getType(), 1);
533 In = Builder->CreateXor(In, One, In->getName()+".not");
536 return ReplaceInstUsesWith(CI, In);
539 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
540 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
541 // zext (X == 1) to i32 --> X iff X has only the low bit set.
542 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
543 // zext (X != 0) to i32 --> X iff X has only the low bit set.
544 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
545 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
546 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
547 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
548 // This only works for EQ and NE
550 // If Op1C some other power of two, convert:
551 uint32_t BitWidth = Op1C->getType()->getBitWidth();
552 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
553 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne);
555 APInt KnownZeroMask(~KnownZero);
556 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
557 if (!DoXform) return ICI;
559 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
560 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
561 // (X&4) == 2 --> false
562 // (X&4) != 2 --> true
563 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
565 Res = ConstantExpr::getZExt(Res, CI.getType());
566 return ReplaceInstUsesWith(CI, Res);
569 uint32_t ShiftAmt = KnownZeroMask.logBase2();
570 Value *In = ICI->getOperand(0);
572 // Perform a logical shr by shiftamt.
573 // Insert the shift to put the result in the low bit.
574 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
575 In->getName()+".lobit");
578 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
579 Constant *One = ConstantInt::get(In->getType(), 1);
580 In = Builder->CreateXor(In, One);
583 if (CI.getType() == In->getType())
584 return ReplaceInstUsesWith(CI, In);
585 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
590 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
591 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
592 // may lead to additional simplifications.
593 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
594 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
595 uint32_t BitWidth = ITy->getBitWidth();
596 Value *LHS = ICI->getOperand(0);
597 Value *RHS = ICI->getOperand(1);
599 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
600 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
601 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS);
602 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS);
604 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
605 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
606 APInt UnknownBit = ~KnownBits;
607 if (UnknownBit.countPopulation() == 1) {
608 if (!DoXform) return ICI;
610 Value *Result = Builder->CreateXor(LHS, RHS);
612 // Mask off any bits that are set and won't be shifted away.
613 if (KnownOneLHS.uge(UnknownBit))
614 Result = Builder->CreateAnd(Result,
615 ConstantInt::get(ITy, UnknownBit));
617 // Shift the bit we're testing down to the lsb.
618 Result = Builder->CreateLShr(
619 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
621 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
622 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
623 Result->takeName(ICI);
624 return ReplaceInstUsesWith(CI, Result);
633 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
634 /// specified wider type and produce the same low bits. If not, return false.
636 /// If this function returns true, it can also return a non-zero number of bits
637 /// (in BitsToClear) which indicates that the value it computes is correct for
638 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
639 /// out. For example, to promote something like:
641 /// %B = trunc i64 %A to i32
642 /// %C = lshr i32 %B, 8
643 /// %E = zext i32 %C to i64
645 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
646 /// set to 8 to indicate that the promoted value needs to have bits 24-31
647 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
648 /// clear the top bits anyway, doing this has no extra cost.
650 /// This function works on both vectors and scalars.
651 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
653 if (isa<Constant>(V))
656 Instruction *I = dyn_cast<Instruction>(V);
657 if (!I) return false;
659 // If the input is a truncate from the destination type, we can trivially
661 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
664 // We can't extend or shrink something that has multiple uses: doing so would
665 // require duplicating the instruction in general, which isn't profitable.
666 if (!I->hasOneUse()) return false;
668 unsigned Opc = I->getOpcode(), Tmp;
670 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
671 case Instruction::SExt: // zext(sext(x)) -> sext(x).
672 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
674 case Instruction::And:
675 case Instruction::Or:
676 case Instruction::Xor:
677 case Instruction::Add:
678 case Instruction::Sub:
679 case Instruction::Mul:
680 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
681 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
683 // These can all be promoted if neither operand has 'bits to clear'.
684 if (BitsToClear == 0 && Tmp == 0)
687 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
688 // other side, BitsToClear is ok.
690 (Opc == Instruction::And || Opc == Instruction::Or ||
691 Opc == Instruction::Xor)) {
692 // We use MaskedValueIsZero here for generality, but the case we care
693 // about the most is constant RHS.
694 unsigned VSize = V->getType()->getScalarSizeInBits();
695 if (MaskedValueIsZero(I->getOperand(1),
696 APInt::getHighBitsSet(VSize, BitsToClear)))
700 // Otherwise, we don't know how to analyze this BitsToClear case yet.
703 case Instruction::Shl:
704 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
705 // upper bits we can reduce BitsToClear by the shift amount.
706 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
707 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
709 uint64_t ShiftAmt = Amt->getZExtValue();
710 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
714 case Instruction::LShr:
715 // We can promote lshr(x, cst) if we can promote x. This requires the
716 // ultimate 'and' to clear out the high zero bits we're clearing out though.
717 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
718 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
720 BitsToClear += Amt->getZExtValue();
721 if (BitsToClear > V->getType()->getScalarSizeInBits())
722 BitsToClear = V->getType()->getScalarSizeInBits();
725 // Cannot promote variable LSHR.
727 case Instruction::Select:
728 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
729 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
730 // TODO: If important, we could handle the case when the BitsToClear are
731 // known zero in the disagreeing side.
736 case Instruction::PHI: {
737 // We can change a phi if we can change all operands. Note that we never
738 // get into trouble with cyclic PHIs here because we only consider
739 // instructions with a single use.
740 PHINode *PN = cast<PHINode>(I);
741 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
743 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
744 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
745 // TODO: If important, we could handle the case when the BitsToClear
746 // are known zero in the disagreeing input.
752 // TODO: Can handle more cases here.
757 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
758 // If this zero extend is only used by a truncate, let the truncate be
759 // eliminated before we try to optimize this zext.
760 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
763 // If one of the common conversion will work, do it.
764 if (Instruction *Result = commonCastTransforms(CI))
767 // See if we can simplify any instructions used by the input whose sole
768 // purpose is to compute bits we don't care about.
769 if (SimplifyDemandedInstructionBits(CI))
772 Value *Src = CI.getOperand(0);
773 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
775 // Attempt to extend the entire input expression tree to the destination
776 // type. Only do this if the dest type is a simple type, don't convert the
777 // expression tree to something weird like i93 unless the source is also
779 unsigned BitsToClear;
780 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
781 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
782 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
783 "Unreasonable BitsToClear");
785 // Okay, we can transform this! Insert the new expression now.
786 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
787 " to avoid zero extend: " << CI);
788 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
789 assert(Res->getType() == DestTy);
791 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
792 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
794 // If the high bits are already filled with zeros, just replace this
795 // cast with the result.
796 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
797 DestBitSize-SrcBitsKept)))
798 return ReplaceInstUsesWith(CI, Res);
800 // We need to emit an AND to clear the high bits.
801 Constant *C = ConstantInt::get(Res->getType(),
802 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
803 return BinaryOperator::CreateAnd(Res, C);
806 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
807 // types and if the sizes are just right we can convert this into a logical
808 // 'and' which will be much cheaper than the pair of casts.
809 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
810 // TODO: Subsume this into EvaluateInDifferentType.
812 // Get the sizes of the types involved. We know that the intermediate type
813 // will be smaller than A or C, but don't know the relation between A and C.
814 Value *A = CSrc->getOperand(0);
815 unsigned SrcSize = A->getType()->getScalarSizeInBits();
816 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
817 unsigned DstSize = CI.getType()->getScalarSizeInBits();
818 // If we're actually extending zero bits, then if
819 // SrcSize < DstSize: zext(a & mask)
820 // SrcSize == DstSize: a & mask
821 // SrcSize > DstSize: trunc(a) & mask
822 if (SrcSize < DstSize) {
823 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
824 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
825 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
826 return new ZExtInst(And, CI.getType());
829 if (SrcSize == DstSize) {
830 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
831 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
834 if (SrcSize > DstSize) {
835 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
836 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
837 return BinaryOperator::CreateAnd(Trunc,
838 ConstantInt::get(Trunc->getType(),
843 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
844 return transformZExtICmp(ICI, CI);
846 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
847 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
848 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
849 // of the (zext icmp) will be transformed.
850 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
851 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
852 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
853 (transformZExtICmp(LHS, CI, false) ||
854 transformZExtICmp(RHS, CI, false))) {
855 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
856 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
857 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
861 // zext(trunc(t) & C) -> (t & zext(C)).
862 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
863 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
864 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
865 Value *TI0 = TI->getOperand(0);
866 if (TI0->getType() == CI.getType())
868 BinaryOperator::CreateAnd(TI0,
869 ConstantExpr::getZExt(C, CI.getType()));
872 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
873 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
874 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
875 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
876 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
877 And->getOperand(1) == C)
878 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
879 Value *TI0 = TI->getOperand(0);
880 if (TI0->getType() == CI.getType()) {
881 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
882 Value *NewAnd = Builder->CreateAnd(TI0, ZC);
883 return BinaryOperator::CreateXor(NewAnd, ZC);
887 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
889 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
890 match(SrcI, m_Not(m_Value(X))) &&
891 (!X->hasOneUse() || !isa<CmpInst>(X))) {
892 Value *New = Builder->CreateZExt(X, CI.getType());
893 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
899 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
900 /// in order to eliminate the icmp.
901 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
902 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
903 ICmpInst::Predicate Pred = ICI->getPredicate();
905 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
906 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
907 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
908 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) ||
909 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
911 Value *Sh = ConstantInt::get(Op0->getType(),
912 Op0->getType()->getScalarSizeInBits()-1);
913 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
914 if (In->getType() != CI.getType())
915 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
917 if (Pred == ICmpInst::ICMP_SGT)
918 In = Builder->CreateNot(In, In->getName()+".not");
919 return ReplaceInstUsesWith(CI, In);
922 // If we know that only one bit of the LHS of the icmp can be set and we
923 // have an equality comparison with zero or a power of 2, we can transform
924 // the icmp and sext into bitwise/integer operations.
925 if (ICI->hasOneUse() &&
926 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
927 unsigned BitWidth = Op1C->getType()->getBitWidth();
928 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
929 ComputeMaskedBits(Op0, KnownZero, KnownOne);
931 APInt KnownZeroMask(~KnownZero);
932 if (KnownZeroMask.isPowerOf2()) {
933 Value *In = ICI->getOperand(0);
935 // If the icmp tests for a known zero bit we can constant fold it.
936 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
937 Value *V = Pred == ICmpInst::ICMP_NE ?
938 ConstantInt::getAllOnesValue(CI.getType()) :
939 ConstantInt::getNullValue(CI.getType());
940 return ReplaceInstUsesWith(CI, V);
943 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
944 // sext ((x & 2^n) == 0) -> (x >> n) - 1
945 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
946 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
947 // Perform a right shift to place the desired bit in the LSB.
949 In = Builder->CreateLShr(In,
950 ConstantInt::get(In->getType(), ShiftAmt));
952 // At this point "In" is either 1 or 0. Subtract 1 to turn
953 // {1, 0} -> {0, -1}.
954 In = Builder->CreateAdd(In,
955 ConstantInt::getAllOnesValue(In->getType()),
958 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
959 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
960 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
961 // Perform a left shift to place the desired bit in the MSB.
963 In = Builder->CreateShl(In,
964 ConstantInt::get(In->getType(), ShiftAmt));
966 // Distribute the bit over the whole bit width.
967 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
968 BitWidth - 1), "sext");
971 if (CI.getType() == In->getType())
972 return ReplaceInstUsesWith(CI, In);
973 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
978 // vector (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed.
979 if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) {
980 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) &&
981 Op0->getType() == CI.getType()) {
982 Type *EltTy = VTy->getElementType();
984 // splat the shift constant to a constant vector.
985 Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
986 Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit");
987 return ReplaceInstUsesWith(CI, In);
994 /// CanEvaluateSExtd - Return true if we can take the specified value
995 /// and return it as type Ty without inserting any new casts and without
996 /// changing the value of the common low bits. This is used by code that tries
997 /// to promote integer operations to a wider types will allow us to eliminate
1000 /// This function works on both vectors and scalars.
1002 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
1003 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1004 "Can't sign extend type to a smaller type");
1005 // If this is a constant, it can be trivially promoted.
1006 if (isa<Constant>(V))
1009 Instruction *I = dyn_cast<Instruction>(V);
1010 if (!I) return false;
1012 // If this is a truncate from the dest type, we can trivially eliminate it.
1013 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1016 // We can't extend or shrink something that has multiple uses: doing so would
1017 // require duplicating the instruction in general, which isn't profitable.
1018 if (!I->hasOneUse()) return false;
1020 switch (I->getOpcode()) {
1021 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1022 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1023 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1025 case Instruction::And:
1026 case Instruction::Or:
1027 case Instruction::Xor:
1028 case Instruction::Add:
1029 case Instruction::Sub:
1030 case Instruction::Mul:
1031 // These operators can all arbitrarily be extended if their inputs can.
1032 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1033 CanEvaluateSExtd(I->getOperand(1), Ty);
1035 //case Instruction::Shl: TODO
1036 //case Instruction::LShr: TODO
1038 case Instruction::Select:
1039 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1040 CanEvaluateSExtd(I->getOperand(2), Ty);
1042 case Instruction::PHI: {
1043 // We can change a phi if we can change all operands. Note that we never
1044 // get into trouble with cyclic PHIs here because we only consider
1045 // instructions with a single use.
1046 PHINode *PN = cast<PHINode>(I);
1047 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1048 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1052 // TODO: Can handle more cases here.
1059 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1060 // If this sign extend is only used by a truncate, let the truncate be
1061 // eliminated before we try to optimize this sext.
1062 if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
1065 if (Instruction *I = commonCastTransforms(CI))
1068 // See if we can simplify any instructions used by the input whose sole
1069 // purpose is to compute bits we don't care about.
1070 if (SimplifyDemandedInstructionBits(CI))
1073 Value *Src = CI.getOperand(0);
1074 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1076 // Attempt to extend the entire input expression tree to the destination
1077 // type. Only do this if the dest type is a simple type, don't convert the
1078 // expression tree to something weird like i93 unless the source is also
1080 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1081 CanEvaluateSExtd(Src, DestTy)) {
1082 // Okay, we can transform this! Insert the new expression now.
1083 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1084 " to avoid sign extend: " << CI);
1085 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1086 assert(Res->getType() == DestTy);
1088 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1089 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1091 // If the high bits are already filled with sign bit, just replace this
1092 // cast with the result.
1093 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1094 return ReplaceInstUsesWith(CI, Res);
1096 // We need to emit a shl + ashr to do the sign extend.
1097 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1098 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1102 // If this input is a trunc from our destination, then turn sext(trunc(x))
1104 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1105 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1106 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1107 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1109 // We need to emit a shl + ashr to do the sign extend.
1110 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1111 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1112 return BinaryOperator::CreateAShr(Res, ShAmt);
1115 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1116 return transformSExtICmp(ICI, CI);
1118 // If the input is a shl/ashr pair of a same constant, then this is a sign
1119 // extension from a smaller value. If we could trust arbitrary bitwidth
1120 // integers, we could turn this into a truncate to the smaller bit and then
1121 // use a sext for the whole extension. Since we don't, look deeper and check
1122 // for a truncate. If the source and dest are the same type, eliminate the
1123 // trunc and extend and just do shifts. For example, turn:
1124 // %a = trunc i32 %i to i8
1125 // %b = shl i8 %a, 6
1126 // %c = ashr i8 %b, 6
1127 // %d = sext i8 %c to i32
1129 // %a = shl i32 %i, 30
1130 // %d = ashr i32 %a, 30
1132 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1133 ConstantInt *BA = 0, *CA = 0;
1134 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1135 m_ConstantInt(CA))) &&
1136 BA == CA && A->getType() == CI.getType()) {
1137 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1138 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1139 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1140 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1141 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1142 return BinaryOperator::CreateAShr(A, ShAmtV);
1149 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1150 /// in the specified FP type without changing its value.
1151 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1153 APFloat F = CFP->getValueAPF();
1154 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1156 return ConstantFP::get(CFP->getContext(), F);
1160 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1161 /// through it until we get the source value.
1162 static Value *LookThroughFPExtensions(Value *V) {
1163 if (Instruction *I = dyn_cast<Instruction>(V))
1164 if (I->getOpcode() == Instruction::FPExt)
1165 return LookThroughFPExtensions(I->getOperand(0));
1167 // If this value is a constant, return the constant in the smallest FP type
1168 // that can accurately represent it. This allows us to turn
1169 // (float)((double)X+2.0) into x+2.0f.
1170 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1171 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1172 return V; // No constant folding of this.
1173 // See if the value can be truncated to half and then reextended.
1174 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1176 // See if the value can be truncated to float and then reextended.
1177 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1179 if (CFP->getType()->isDoubleTy())
1180 return V; // Won't shrink.
1181 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1183 // Don't try to shrink to various long double types.
1189 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1190 if (Instruction *I = commonCastTransforms(CI))
1192 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1193 // simpilify this expression to avoid one or more of the trunc/extend
1194 // operations if we can do so without changing the numerical results.
1196 // The exact manner in which the widths of the operands interact to limit
1197 // what we can and cannot do safely varies from operation to operation, and
1198 // is explained below in the various case statements.
1199 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1200 if (OpI && OpI->hasOneUse()) {
1201 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1202 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1203 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1204 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1205 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1206 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1207 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1208 switch (OpI->getOpcode()) {
1210 case Instruction::FAdd:
1211 case Instruction::FSub:
1212 // For addition and subtraction, the infinitely precise result can
1213 // essentially be arbitrarily wide; proving that double rounding
1214 // will not occur because the result of OpI is exact (as we will for
1215 // FMul, for example) is hopeless. However, we *can* nonetheless
1216 // frequently know that double rounding cannot occur (or that it is
1217 // innoculous) by taking advantage of the specific structure of
1218 // infinitely-precise results that admit double rounding.
1220 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficent
1221 // to represent both sources, we can guarantee that the double
1222 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1223 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1224 // for proof of this fact).
1226 // Note: Figueroa does not consider the case where DstFormat !=
1227 // SrcFormat. It's possible (likely even!) that this analysis
1228 // could be tightened for those cases, but they are rare (the main
1229 // case of interest here is (float)((double)float + float)).
1230 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1231 if (LHSOrig->getType() != CI.getType())
1232 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1233 if (RHSOrig->getType() != CI.getType())
1234 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1235 return BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1238 case Instruction::FMul:
1239 // For multiplication, the infinitely precise result has at most
1240 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1241 // that such a value can be exactly represented, then no double
1242 // rounding can possibly occur; we can safely perform the operation
1243 // in the destination format if it can represent both sources.
1244 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1245 if (LHSOrig->getType() != CI.getType())
1246 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1247 if (RHSOrig->getType() != CI.getType())
1248 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1249 return BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1252 case Instruction::FDiv:
1253 // For division, we use again use the bound from Figueroa's
1254 // dissertation. I am entirely certain that this bound can be
1255 // tightened in the unbalanced operand case by an analysis based on
1256 // the diophantine rational approximation bound, but the well-known
1257 // condition used here is a good conservative first pass.
1258 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1259 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1260 if (LHSOrig->getType() != CI.getType())
1261 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1262 if (RHSOrig->getType() != CI.getType())
1263 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1264 return BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1267 case Instruction::FRem:
1268 // Remainder is straightforward. Remainder is always exact, so the
1269 // type of OpI doesn't enter into things at all. We simply evaluate
1270 // in whichever source type is larger, then convert to the
1271 // destination type.
1272 if (LHSWidth < SrcWidth)
1273 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1274 else if (RHSWidth <= SrcWidth)
1275 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1276 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1277 return CastInst::CreateFPCast(ExactResult, CI.getType());
1280 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1281 if (BinaryOperator::isFNeg(OpI)) {
1282 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1284 return BinaryOperator::CreateFNeg(InnerTrunc);
1288 // (fptrunc (select cond, R1, Cst)) -->
1289 // (select cond, (fptrunc R1), (fptrunc Cst))
1290 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1292 (isa<ConstantFP>(SI->getOperand(1)) ||
1293 isa<ConstantFP>(SI->getOperand(2)))) {
1294 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1296 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1298 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1301 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1303 switch (II->getIntrinsicID()) {
1305 case Intrinsic::fabs: {
1306 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1307 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1309 Type *IntrinsicType[] = { CI.getType() };
1310 Function *Overload =
1311 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1312 II->getIntrinsicID(), IntrinsicType);
1314 Value *Args[] = { InnerTrunc };
1315 return CallInst::Create(Overload, Args, II->getName());
1320 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1321 // Note that we restrict this transformation based on
1322 // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because
1323 // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the
1324 // single-precision intrinsic can be expanded in the backend.
1325 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1326 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1327 (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) ||
1328 Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) &&
1329 Call->getNumArgOperands() == 1 &&
1330 Call->hasOneUse()) {
1331 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1332 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1333 CI.getType()->isFloatTy() &&
1334 Call->getType()->isDoubleTy() &&
1335 Arg->getType()->isDoubleTy() &&
1336 Arg->getOperand(0)->getType()->isFloatTy()) {
1337 Function *Callee = Call->getCalledFunction();
1338 Module *M = CI.getParent()->getParent()->getParent();
1339 Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ?
1340 Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) :
1341 M->getOrInsertFunction("sqrtf", Callee->getAttributes(),
1342 Builder->getFloatTy(), Builder->getFloatTy(),
1344 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1346 ret->setAttributes(Callee->getAttributes());
1349 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1350 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1351 EraseInstFromFunction(*Call);
1359 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1360 return commonCastTransforms(CI);
1363 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1364 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1366 return commonCastTransforms(FI);
1368 // fptoui(uitofp(X)) --> X
1369 // fptoui(sitofp(X)) --> X
1370 // This is safe if the intermediate type has enough bits in its mantissa to
1371 // accurately represent all values of X. For example, do not do this with
1372 // i64->float->i64. This is also safe for sitofp case, because any negative
1373 // 'X' value would cause an undefined result for the fptoui.
1374 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1375 OpI->getOperand(0)->getType() == FI.getType() &&
1376 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1377 OpI->getType()->getFPMantissaWidth())
1378 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1380 return commonCastTransforms(FI);
1383 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1384 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1386 return commonCastTransforms(FI);
1388 // fptosi(sitofp(X)) --> X
1389 // fptosi(uitofp(X)) --> X
1390 // This is safe if the intermediate type has enough bits in its mantissa to
1391 // accurately represent all values of X. For example, do not do this with
1392 // i64->float->i64. This is also safe for sitofp case, because any negative
1393 // 'X' value would cause an undefined result for the fptoui.
1394 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1395 OpI->getOperand(0)->getType() == FI.getType() &&
1396 (int)FI.getType()->getScalarSizeInBits() <=
1397 OpI->getType()->getFPMantissaWidth())
1398 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1400 return commonCastTransforms(FI);
1403 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1404 return commonCastTransforms(CI);
1407 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1408 return commonCastTransforms(CI);
1411 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1412 // If the source integer type is not the intptr_t type for this target, do a
1413 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1414 // cast to be exposed to other transforms.
1417 unsigned AS = CI.getAddressSpace();
1418 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1419 TD->getPointerSizeInBits(AS)) {
1420 Type *Ty = TD->getIntPtrType(CI.getContext(), AS);
1421 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1422 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1424 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1425 return new IntToPtrInst(P, CI.getType());
1429 if (Instruction *I = commonCastTransforms(CI))
1435 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1436 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1437 Value *Src = CI.getOperand(0);
1439 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1440 // If casting the result of a getelementptr instruction with no offset, turn
1441 // this into a cast of the original pointer!
1442 if (GEP->hasAllZeroIndices()) {
1443 // Changing the cast operand is usually not a good idea but it is safe
1444 // here because the pointer operand is being replaced with another
1445 // pointer operand so the opcode doesn't need to change.
1447 CI.setOperand(0, GEP->getOperand(0));
1452 return commonCastTransforms(CI);
1454 // If the GEP has a single use, and the base pointer is a bitcast, and the
1455 // GEP computes a constant offset, see if we can convert these three
1456 // instructions into fewer. This typically happens with unions and other
1457 // non-type-safe code.
1458 unsigned AS = GEP->getPointerAddressSpace();
1459 unsigned OffsetBits = TD->getPointerSizeInBits(AS);
1460 APInt Offset(OffsetBits, 0);
1461 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0));
1462 if (GEP->hasOneUse() &&
1464 GEP->accumulateConstantOffset(*TD, Offset)) {
1465 // Get the base pointer input of the bitcast, and the type it points to.
1466 Value *OrigBase = BCI->getOperand(0);
1467 SmallVector<Value*, 8> NewIndices;
1468 if (FindElementAtOffset(OrigBase->getType(),
1469 Offset.getSExtValue(),
1471 // If we were able to index down into an element, create the GEP
1472 // and bitcast the result. This eliminates one bitcast, potentially
1474 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1475 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1476 Builder->CreateGEP(OrigBase, NewIndices);
1477 NGEP->takeName(GEP);
1479 if (isa<BitCastInst>(CI))
1480 return new BitCastInst(NGEP, CI.getType());
1481 assert(isa<PtrToIntInst>(CI));
1482 return new PtrToIntInst(NGEP, CI.getType());
1487 return commonCastTransforms(CI);
1490 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1491 // If the destination integer type is not the intptr_t type for this target,
1492 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1493 // to be exposed to other transforms.
1496 return commonPointerCastTransforms(CI);
1498 Type *Ty = CI.getType();
1499 unsigned AS = CI.getPointerAddressSpace();
1501 if (Ty->getScalarSizeInBits() == TD->getPointerSizeInBits(AS))
1502 return commonPointerCastTransforms(CI);
1504 Type *PtrTy = TD->getIntPtrType(CI.getContext(), AS);
1505 if (Ty->isVectorTy()) // Handle vectors of pointers.
1506 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1508 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1509 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1512 /// OptimizeVectorResize - This input value (which is known to have vector type)
1513 /// is being zero extended or truncated to the specified vector type. Try to
1514 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1516 /// The source and destination vector types may have different element types.
1517 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1519 // We can only do this optimization if the output is a multiple of the input
1520 // element size, or the input is a multiple of the output element size.
1521 // Convert the input type to have the same element type as the output.
1522 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1524 if (SrcTy->getElementType() != DestTy->getElementType()) {
1525 // The input types don't need to be identical, but for now they must be the
1526 // same size. There is no specific reason we couldn't handle things like
1527 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1529 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1530 DestTy->getElementType()->getPrimitiveSizeInBits())
1533 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1534 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1537 // Now that the element types match, get the shuffle mask and RHS of the
1538 // shuffle to use, which depends on whether we're increasing or decreasing the
1539 // size of the input.
1540 SmallVector<uint32_t, 16> ShuffleMask;
1543 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1544 // If we're shrinking the number of elements, just shuffle in the low
1545 // elements from the input and use undef as the second shuffle input.
1546 V2 = UndefValue::get(SrcTy);
1547 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1548 ShuffleMask.push_back(i);
1551 // If we're increasing the number of elements, shuffle in all of the
1552 // elements from InVal and fill the rest of the result elements with zeros
1553 // from a constant zero.
1554 V2 = Constant::getNullValue(SrcTy);
1555 unsigned SrcElts = SrcTy->getNumElements();
1556 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1557 ShuffleMask.push_back(i);
1559 // The excess elements reference the first element of the zero input.
1560 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1561 ShuffleMask.push_back(SrcElts);
1564 return new ShuffleVectorInst(InVal, V2,
1565 ConstantDataVector::get(V2->getContext(),
1569 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1570 return Value % Ty->getPrimitiveSizeInBits() == 0;
1573 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1574 return Value / Ty->getPrimitiveSizeInBits();
1577 /// CollectInsertionElements - V is a value which is inserted into a vector of
1578 /// VecEltTy. Look through the value to see if we can decompose it into
1579 /// insertions into the vector. See the example in the comment for
1580 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1581 /// The type of V is always a non-zero multiple of VecEltTy's size.
1582 /// Shift is the number of bits between the lsb of V and the lsb of
1585 /// This returns false if the pattern can't be matched or true if it can,
1586 /// filling in Elements with the elements found here.
1587 static bool CollectInsertionElements(Value *V, unsigned Shift,
1588 SmallVectorImpl<Value*> &Elements,
1589 Type *VecEltTy, InstCombiner &IC) {
1590 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1591 "Shift should be a multiple of the element type size");
1593 // Undef values never contribute useful bits to the result.
1594 if (isa<UndefValue>(V)) return true;
1596 // If we got down to a value of the right type, we win, try inserting into the
1598 if (V->getType() == VecEltTy) {
1599 // Inserting null doesn't actually insert any elements.
1600 if (Constant *C = dyn_cast<Constant>(V))
1601 if (C->isNullValue())
1604 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1605 if (IC.getDataLayout()->isBigEndian())
1606 ElementIndex = Elements.size() - ElementIndex - 1;
1608 // Fail if multiple elements are inserted into this slot.
1609 if (Elements[ElementIndex] != 0)
1612 Elements[ElementIndex] = V;
1616 if (Constant *C = dyn_cast<Constant>(V)) {
1617 // Figure out the # elements this provides, and bitcast it or slice it up
1619 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1621 // If the constant is the size of a vector element, we just need to bitcast
1622 // it to the right type so it gets properly inserted.
1624 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1625 Shift, Elements, VecEltTy, IC);
1627 // Okay, this is a constant that covers multiple elements. Slice it up into
1628 // pieces and insert each element-sized piece into the vector.
1629 if (!isa<IntegerType>(C->getType()))
1630 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1631 C->getType()->getPrimitiveSizeInBits()));
1632 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1633 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1635 for (unsigned i = 0; i != NumElts; ++i) {
1636 unsigned ShiftI = Shift+i*ElementSize;
1637 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1639 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1640 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC))
1646 if (!V->hasOneUse()) return false;
1648 Instruction *I = dyn_cast<Instruction>(V);
1649 if (I == 0) return false;
1650 switch (I->getOpcode()) {
1651 default: return false; // Unhandled case.
1652 case Instruction::BitCast:
1653 return CollectInsertionElements(I->getOperand(0), Shift,
1654 Elements, VecEltTy, IC);
1655 case Instruction::ZExt:
1656 if (!isMultipleOfTypeSize(
1657 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1660 return CollectInsertionElements(I->getOperand(0), Shift,
1661 Elements, VecEltTy, IC);
1662 case Instruction::Or:
1663 return CollectInsertionElements(I->getOperand(0), Shift,
1664 Elements, VecEltTy, IC) &&
1665 CollectInsertionElements(I->getOperand(1), Shift,
1666 Elements, VecEltTy, IC);
1667 case Instruction::Shl: {
1668 // Must be shifting by a constant that is a multiple of the element size.
1669 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1670 if (CI == 0) return false;
1671 Shift += CI->getZExtValue();
1672 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1673 return CollectInsertionElements(I->getOperand(0), Shift,
1674 Elements, VecEltTy, IC);
1681 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1682 /// may be doing shifts and ors to assemble the elements of the vector manually.
1683 /// Try to rip the code out and replace it with insertelements. This is to
1684 /// optimize code like this:
1686 /// %tmp37 = bitcast float %inc to i32
1687 /// %tmp38 = zext i32 %tmp37 to i64
1688 /// %tmp31 = bitcast float %inc5 to i32
1689 /// %tmp32 = zext i32 %tmp31 to i64
1690 /// %tmp33 = shl i64 %tmp32, 32
1691 /// %ins35 = or i64 %tmp33, %tmp38
1692 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1694 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1695 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1697 // We need to know the target byte order to perform this optimization.
1698 if (!IC.getDataLayout()) return 0;
1700 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1701 Value *IntInput = CI.getOperand(0);
1703 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1704 if (!CollectInsertionElements(IntInput, 0, Elements,
1705 DestVecTy->getElementType(), IC))
1708 // If we succeeded, we know that all of the element are specified by Elements
1709 // or are zero if Elements has a null entry. Recast this as a set of
1711 Value *Result = Constant::getNullValue(CI.getType());
1712 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1713 if (Elements[i] == 0) continue; // Unset element.
1715 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1716 IC.Builder->getInt32(i));
1723 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1724 /// bitcast. The various long double bitcasts can't get in here.
1725 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1726 // We need to know the target byte order to perform this optimization.
1727 if (!IC.getDataLayout()) return 0;
1729 Value *Src = CI.getOperand(0);
1730 Type *DestTy = CI.getType();
1732 // If this is a bitcast from int to float, check to see if the int is an
1733 // extraction from a vector.
1734 Value *VecInput = 0;
1735 // bitcast(trunc(bitcast(somevector)))
1736 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1737 isa<VectorType>(VecInput->getType())) {
1738 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1739 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1741 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1742 // If the element type of the vector doesn't match the result type,
1743 // bitcast it to be a vector type we can extract from.
1744 if (VecTy->getElementType() != DestTy) {
1745 VecTy = VectorType::get(DestTy,
1746 VecTy->getPrimitiveSizeInBits() / DestWidth);
1747 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1751 if (IC.getDataLayout()->isBigEndian())
1752 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1753 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1757 // bitcast(trunc(lshr(bitcast(somevector), cst))
1758 ConstantInt *ShAmt = 0;
1759 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1760 m_ConstantInt(ShAmt)))) &&
1761 isa<VectorType>(VecInput->getType())) {
1762 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1763 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1764 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1765 ShAmt->getZExtValue() % DestWidth == 0) {
1766 // If the element type of the vector doesn't match the result type,
1767 // bitcast it to be a vector type we can extract from.
1768 if (VecTy->getElementType() != DestTy) {
1769 VecTy = VectorType::get(DestTy,
1770 VecTy->getPrimitiveSizeInBits() / DestWidth);
1771 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1774 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1775 if (IC.getDataLayout()->isBigEndian())
1776 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1777 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1783 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1784 // If the operands are integer typed then apply the integer transforms,
1785 // otherwise just apply the common ones.
1786 Value *Src = CI.getOperand(0);
1787 Type *SrcTy = Src->getType();
1788 Type *DestTy = CI.getType();
1790 // Get rid of casts from one type to the same type. These are useless and can
1791 // be replaced by the operand.
1792 if (DestTy == Src->getType())
1793 return ReplaceInstUsesWith(CI, Src);
1795 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1796 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1797 Type *DstElTy = DstPTy->getElementType();
1798 Type *SrcElTy = SrcPTy->getElementType();
1800 // If the address spaces don't match, don't eliminate the bitcast, which is
1801 // required for changing types.
1802 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1805 // If we are casting a alloca to a pointer to a type of the same
1806 // size, rewrite the allocation instruction to allocate the "right" type.
1807 // There is no need to modify malloc calls because it is their bitcast that
1808 // needs to be cleaned up.
1809 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1810 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1813 // If the source and destination are pointers, and this cast is equivalent
1814 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1815 // This can enhance SROA and other transforms that want type-safe pointers.
1816 Constant *ZeroUInt =
1817 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1818 unsigned NumZeros = 0;
1819 while (SrcElTy != DstElTy &&
1820 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1821 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1822 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1826 // If we found a path from the src to dest, create the getelementptr now.
1827 if (SrcElTy == DstElTy) {
1828 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1829 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1833 // Try to optimize int -> float bitcasts.
1834 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1835 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1838 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1839 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1840 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1841 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1842 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1843 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1846 if (isa<IntegerType>(SrcTy)) {
1847 // If this is a cast from an integer to vector, check to see if the input
1848 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1849 // the casts with a shuffle and (potentially) a bitcast.
1850 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1851 CastInst *SrcCast = cast<CastInst>(Src);
1852 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1853 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1854 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1855 cast<VectorType>(DestTy), *this))
1859 // If the input is an 'or' instruction, we may be doing shifts and ors to
1860 // assemble the elements of the vector manually. Try to rip the code out
1861 // and replace it with insertelements.
1862 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1863 return ReplaceInstUsesWith(CI, V);
1867 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1868 if (SrcVTy->getNumElements() == 1) {
1869 // If our destination is not a vector, then make this a straight
1870 // scalar-scalar cast.
1871 if (!DestTy->isVectorTy()) {
1873 Builder->CreateExtractElement(Src,
1874 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1875 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1878 // Otherwise, see if our source is an insert. If so, then use the scalar
1879 // component directly.
1880 if (InsertElementInst *IEI =
1881 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1882 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1887 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1888 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1889 // a bitcast to a vector with the same # elts.
1890 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1891 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1892 SVI->getType()->getNumElements() ==
1893 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1895 // If either of the operands is a cast from CI.getType(), then
1896 // evaluating the shuffle in the casted destination's type will allow
1897 // us to eliminate at least one cast.
1898 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1899 Tmp->getOperand(0)->getType() == DestTy) ||
1900 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1901 Tmp->getOperand(0)->getType() == DestTy)) {
1902 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1903 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1904 // Return a new shuffle vector. Use the same element ID's, as we
1905 // know the vector types match #elts.
1906 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1911 if (SrcTy->isPointerTy())
1912 return commonPointerCastTransforms(CI);
1913 return commonCastTransforms(CI);
1916 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1917 return commonCastTransforms(CI);