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 "InstCombineInternal.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// Analyze 'Val', seeing if it is a simple linear expression.
26 /// If so, decompose it, returning some value X, such that Val is
29 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
31 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
32 Offset = CI->getZExtValue();
34 return ConstantInt::get(Val->getType(), 0);
37 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
38 // Cannot look past anything that might overflow.
39 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
40 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
46 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
47 if (I->getOpcode() == Instruction::Shl) {
48 // This is a value scaled by '1 << the shift amt'.
49 Scale = UINT64_C(1) << RHS->getZExtValue();
51 return I->getOperand(0);
54 if (I->getOpcode() == Instruction::Mul) {
55 // This value is scaled by 'RHS'.
56 Scale = RHS->getZExtValue();
58 return I->getOperand(0);
61 if (I->getOpcode() == Instruction::Add) {
62 // We have X+C. Check to see if we really have (X*C2)+C1,
63 // where C1 is divisible by C2.
66 decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
67 Offset += RHS->getZExtValue();
74 // Otherwise, we can't look past this.
80 /// If we find a cast of an allocation instruction, try to eliminate the cast by
81 /// moving the type information into the alloc.
82 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
84 PointerType *PTy = cast<PointerType>(CI.getType());
86 BuilderTy AllocaBuilder(*Builder);
87 AllocaBuilder.SetInsertPoint(&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 nullptr;
94 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy);
95 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy);
96 if (CastElTyAlign < AllocElTyAlign) return nullptr;
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 nullptr;
103 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy);
104 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy);
105 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
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 = DL.getTypeStoreSize(AllocElTy);
110 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy);
111 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
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 nullptr;
125 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
126 Value *Amt = nullptr;
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());
144 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
146 // If the allocation has multiple real uses, insert a cast and change all
147 // things that used it to use the new cast. This will also hack on CI, but it
149 if (!AI.hasOneUse()) {
150 // New is the allocation instruction, pointer typed. AI is the original
151 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
152 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
153 replaceInstUsesWith(AI, NewCast);
155 return replaceInstUsesWith(CI, New);
158 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
159 /// true for, actually 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 DL info.
165 if (Constant *FoldedC = ConstantFoldConstant(C, DL, &TLI))
170 // Otherwise, it must be an instruction.
171 Instruction *I = cast<Instruction>(V);
172 Instruction *Res = nullptr;
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) {
216 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
217 NPN->addIncoming(V, OPN->getIncomingBlock(i));
223 // TODO: Can handle more cases here.
224 llvm_unreachable("Unreachable!");
228 return InsertNewInstWith(Res, *I);
231 Instruction::CastOps InstCombiner::isEliminableCastPair(const CastInst *CI1,
232 const CastInst *CI2) {
233 Type *SrcTy = CI1->getSrcTy();
234 Type *MidTy = CI1->getDestTy();
235 Type *DstTy = CI2->getDestTy();
237 Instruction::CastOps firstOp = Instruction::CastOps(CI1->getOpcode());
238 Instruction::CastOps secondOp = Instruction::CastOps(CI2->getOpcode());
240 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
242 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
244 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
245 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
246 DstTy, SrcIntPtrTy, MidIntPtrTy,
249 // We don't want to form an inttoptr or ptrtoint that converts to an integer
250 // type that differs from the pointer size.
251 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
252 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
255 return Instruction::CastOps(Res);
258 /// @brief Implement the transforms common to all CastInst visitors.
259 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
260 Value *Src = CI.getOperand(0);
262 // Try to eliminate a cast of a cast.
263 if (auto *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
264 if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
265 // The first cast (CSrc) is eliminable so we need to fix up or replace
266 // the second cast (CI). CSrc will then have a good chance of being dead.
267 return CastInst::Create(NewOpc, CSrc->getOperand(0), CI.getType());
271 // If we are casting a select, then fold the cast into the select.
272 if (auto *SI = dyn_cast<SelectInst>(Src))
273 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
276 // If we are casting a PHI, then fold the cast into the PHI.
277 if (isa<PHINode>(Src)) {
278 // Don't do this if it would create a PHI node with an illegal type from a
280 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
281 ShouldChangeType(CI.getType(), Src->getType()))
282 if (Instruction *NV = FoldOpIntoPhi(CI))
289 /// Return true if we can evaluate the specified expression tree as type Ty
290 /// instead of its larger type, and arrive with the same value.
291 /// This is used by code that tries to eliminate truncates.
293 /// Ty will always be a type smaller than V. We should return true if trunc(V)
294 /// can be computed by computing V in the smaller type. If V is an instruction,
295 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
296 /// makes sense if x and y can be efficiently truncated.
298 /// This function works on both vectors and scalars.
300 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
302 // We can always evaluate constants in another type.
303 if (isa<Constant>(V))
306 Instruction *I = dyn_cast<Instruction>(V);
307 if (!I) return false;
309 Type *OrigTy = V->getType();
311 // If this is an extension from the dest type, we can eliminate it, even if it
312 // has multiple uses.
313 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
314 I->getOperand(0)->getType() == Ty)
317 // We can't extend or shrink something that has multiple uses: doing so would
318 // require duplicating the instruction in general, which isn't profitable.
319 if (!I->hasOneUse()) return false;
321 unsigned Opc = I->getOpcode();
323 case Instruction::Add:
324 case Instruction::Sub:
325 case Instruction::Mul:
326 case Instruction::And:
327 case Instruction::Or:
328 case Instruction::Xor:
329 // These operators can all arbitrarily be extended or truncated.
330 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
331 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
333 case Instruction::UDiv:
334 case Instruction::URem: {
335 // UDiv and URem can be truncated if all the truncated bits are zero.
336 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
337 uint32_t BitWidth = Ty->getScalarSizeInBits();
338 if (BitWidth < OrigBitWidth) {
339 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
340 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
341 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
342 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
343 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
348 case Instruction::Shl:
349 // If we are truncating the result of this SHL, and if it's a shift of a
350 // constant amount, we can always perform a SHL in a smaller type.
351 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
352 uint32_t BitWidth = Ty->getScalarSizeInBits();
353 if (CI->getLimitedValue(BitWidth) < BitWidth)
354 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
357 case Instruction::LShr:
358 // If this is a truncate of a logical shr, we can truncate it to a smaller
359 // lshr iff we know that the bits we would otherwise be shifting in are
361 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
362 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
363 uint32_t BitWidth = Ty->getScalarSizeInBits();
364 if (IC.MaskedValueIsZero(I->getOperand(0),
365 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
366 CI->getLimitedValue(BitWidth) < BitWidth) {
367 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
371 case Instruction::Trunc:
372 // trunc(trunc(x)) -> trunc(x)
374 case Instruction::ZExt:
375 case Instruction::SExt:
376 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
377 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
379 case Instruction::Select: {
380 SelectInst *SI = cast<SelectInst>(I);
381 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
382 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
384 case Instruction::PHI: {
385 // We can change a phi if we can change all operands. Note that we never
386 // get into trouble with cyclic PHIs here because we only consider
387 // instructions with a single use.
388 PHINode *PN = cast<PHINode>(I);
389 for (Value *IncValue : PN->incoming_values())
390 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
395 // TODO: Can handle more cases here.
402 /// Given a vector that is bitcast to an integer, optionally logically
403 /// right-shifted, and truncated, convert it to an extractelement.
404 /// Example (big endian):
405 /// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
407 /// extractelement <4 x i32> %X, 1
408 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc, InstCombiner &IC,
409 const DataLayout &DL) {
410 Value *TruncOp = Trunc.getOperand(0);
411 Type *DestType = Trunc.getType();
412 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
415 Value *VecInput = nullptr;
416 ConstantInt *ShiftVal = nullptr;
417 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
418 m_LShr(m_BitCast(m_Value(VecInput)),
419 m_ConstantInt(ShiftVal)))) ||
420 !isa<VectorType>(VecInput->getType()))
423 VectorType *VecType = cast<VectorType>(VecInput->getType());
424 unsigned VecWidth = VecType->getPrimitiveSizeInBits();
425 unsigned DestWidth = DestType->getPrimitiveSizeInBits();
426 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
428 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
431 // If the element type of the vector doesn't match the result type,
432 // bitcast it to a vector type that we can extract from.
433 unsigned NumVecElts = VecWidth / DestWidth;
434 if (VecType->getElementType() != DestType) {
435 VecType = VectorType::get(DestType, NumVecElts);
436 VecInput = IC.Builder->CreateBitCast(VecInput, VecType, "bc");
439 unsigned Elt = ShiftAmount / DestWidth;
440 if (DL.isBigEndian())
441 Elt = NumVecElts - 1 - Elt;
443 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
446 /// Try to narrow the width of bitwise logic instructions with constants.
447 Instruction *InstCombiner::shrinkBitwiseLogic(TruncInst &Trunc) {
448 Type *SrcTy = Trunc.getSrcTy();
449 Type *DestTy = Trunc.getType();
450 if (isa<IntegerType>(SrcTy) && !ShouldChangeType(SrcTy, DestTy))
453 BinaryOperator *LogicOp;
455 if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(LogicOp))) ||
456 !LogicOp->isBitwiseLogicOp() ||
457 !match(LogicOp->getOperand(1), m_Constant(C)))
460 // trunc (logic X, C) --> logic (trunc X, C')
461 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
462 Value *NarrowOp0 = Builder->CreateTrunc(LogicOp->getOperand(0), DestTy);
463 return BinaryOperator::Create(LogicOp->getOpcode(), NarrowOp0, NarrowC);
466 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
467 if (Instruction *Result = commonCastTransforms(CI))
470 // Test if the trunc is the user of a select which is part of a
471 // minimum or maximum operation. If so, don't do any more simplification.
472 // Even simplifying demanded bits can break the canonical form of a
475 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)))
476 if (matchSelectPattern(SI, LHS, RHS).Flavor != SPF_UNKNOWN)
479 // See if we can simplify any instructions used by the input whose sole
480 // purpose is to compute bits we don't care about.
481 if (SimplifyDemandedInstructionBits(CI))
484 Value *Src = CI.getOperand(0);
485 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
487 // Attempt to truncate the entire input expression tree to the destination
488 // type. Only do this if the dest type is a simple type, don't convert the
489 // expression tree to something weird like i93 unless the source is also
491 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
492 canEvaluateTruncated(Src, DestTy, *this, &CI)) {
494 // If this cast is a truncate, evaluting in a different type always
495 // eliminates the cast, so it is always a win.
496 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
497 " to avoid cast: " << CI << '\n');
498 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
499 assert(Res->getType() == DestTy);
500 return replaceInstUsesWith(CI, Res);
503 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
504 if (DestTy->getScalarSizeInBits() == 1) {
505 Constant *One = ConstantInt::get(SrcTy, 1);
506 Src = Builder->CreateAnd(Src, One);
507 Value *Zero = Constant::getNullValue(Src->getType());
508 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
511 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
512 Value *A = nullptr; ConstantInt *Cst = nullptr;
513 if (Src->hasOneUse() &&
514 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
515 // We have three types to worry about here, the type of A, the source of
516 // the truncate (MidSize), and the destination of the truncate. We know that
517 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
518 // between ASize and ResultSize.
519 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
521 // If the shift amount is larger than the size of A, then the result is
522 // known to be zero because all the input bits got shifted out.
523 if (Cst->getZExtValue() >= ASize)
524 return replaceInstUsesWith(CI, Constant::getNullValue(DestTy));
526 // Since we're doing an lshr and a zero extend, and know that the shift
527 // amount is smaller than ASize, it is always safe to do the shift in A's
528 // type, then zero extend or truncate to the result.
529 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
530 Shift->takeName(Src);
531 return CastInst::CreateIntegerCast(Shift, DestTy, false);
534 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type
536 // It works because bits coming from sign extension have the same value as
537 // the sign bit of the original value; performing ashr instead of lshr
538 // generates bits of the same value as the sign bit.
539 if (Src->hasOneUse() &&
540 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst))) &&
541 cast<Instruction>(Src)->getOperand(0)->hasOneUse()) {
542 const unsigned ASize = A->getType()->getPrimitiveSizeInBits();
543 // This optimization can be only performed when zero bits generated by
544 // the original lshr aren't pulled into the value after truncation, so we
545 // can only shift by values smaller than the size of destination type (in
547 if (Cst->getValue().ult(ASize)) {
548 Value *Shift = Builder->CreateAShr(A, Cst->getZExtValue());
549 Shift->takeName(Src);
550 return CastInst::CreateIntegerCast(Shift, CI.getType(), true);
554 if (Instruction *I = shrinkBitwiseLogic(CI))
557 if (Src->hasOneUse() && isa<IntegerType>(SrcTy) &&
558 ShouldChangeType(SrcTy, DestTy)) {
559 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
560 // dest type is native and cst < dest size.
561 if (match(Src, m_Shl(m_Value(A), m_ConstantInt(Cst))) &&
562 !match(A, m_Shr(m_Value(), m_Constant()))) {
563 // Skip shifts of shift by constants. It undoes a combine in
564 // FoldShiftByConstant and is the extend in reg pattern.
565 const unsigned DestSize = DestTy->getScalarSizeInBits();
566 if (Cst->getValue().ult(DestSize)) {
567 Value *NewTrunc = Builder->CreateTrunc(A, DestTy, A->getName() + ".tr");
569 return BinaryOperator::Create(
570 Instruction::Shl, NewTrunc,
571 ConstantInt::get(DestTy, Cst->getValue().trunc(DestSize)));
576 if (Instruction *I = foldVecTruncToExtElt(CI, *this, DL))
582 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, ZExtInst &CI,
584 // If we are just checking for a icmp eq of a single bit and zext'ing it
585 // to an integer, then shift the bit to the appropriate place and then
586 // cast to integer to avoid the comparison.
587 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
588 const APInt &Op1CV = Op1C->getValue();
590 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
591 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
592 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
593 (ICI->getPredicate() == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
594 if (!DoTransform) return ICI;
596 Value *In = ICI->getOperand(0);
597 Value *Sh = ConstantInt::get(In->getType(),
598 In->getType()->getScalarSizeInBits() - 1);
599 In = Builder->CreateLShr(In, Sh, In->getName() + ".lobit");
600 if (In->getType() != CI.getType())
601 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
603 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
604 Constant *One = ConstantInt::get(In->getType(), 1);
605 In = Builder->CreateXor(In, One, In->getName() + ".not");
608 return replaceInstUsesWith(CI, In);
611 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
612 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
613 // zext (X == 1) to i32 --> X iff X has only the low bit set.
614 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
615 // zext (X != 0) to i32 --> X iff X has only the low bit set.
616 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
617 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
618 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
619 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
620 // This only works for EQ and NE
622 // If Op1C some other power of two, convert:
623 uint32_t BitWidth = Op1C->getType()->getBitWidth();
624 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
625 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
627 APInt KnownZeroMask(~KnownZero);
628 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
629 if (!DoTransform) return ICI;
631 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
632 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
633 // (X&4) == 2 --> false
634 // (X&4) != 2 --> true
635 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
637 Res = ConstantExpr::getZExt(Res, CI.getType());
638 return replaceInstUsesWith(CI, Res);
641 uint32_t ShAmt = KnownZeroMask.logBase2();
642 Value *In = ICI->getOperand(0);
644 // Perform a logical shr by shiftamt.
645 // Insert the shift to put the result in the low bit.
646 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
647 In->getName() + ".lobit");
650 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
651 Constant *One = ConstantInt::get(In->getType(), 1);
652 In = Builder->CreateXor(In, One);
655 if (CI.getType() == In->getType())
656 return replaceInstUsesWith(CI, In);
658 Value *IntCast = Builder->CreateIntCast(In, CI.getType(), false);
659 return replaceInstUsesWith(CI, IntCast);
664 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
665 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
666 // may lead to additional simplifications.
667 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
668 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
669 uint32_t BitWidth = ITy->getBitWidth();
670 Value *LHS = ICI->getOperand(0);
671 Value *RHS = ICI->getOperand(1);
673 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
674 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
675 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
676 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
678 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
679 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
680 APInt UnknownBit = ~KnownBits;
681 if (UnknownBit.countPopulation() == 1) {
682 if (!DoTransform) return ICI;
684 Value *Result = Builder->CreateXor(LHS, RHS);
686 // Mask off any bits that are set and won't be shifted away.
687 if (KnownOneLHS.uge(UnknownBit))
688 Result = Builder->CreateAnd(Result,
689 ConstantInt::get(ITy, UnknownBit));
691 // Shift the bit we're testing down to the lsb.
692 Result = Builder->CreateLShr(
693 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
695 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
696 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
697 Result->takeName(ICI);
698 return replaceInstUsesWith(CI, Result);
707 /// Determine if the specified value can be computed in the specified wider type
708 /// and produce the same low bits. If not, return false.
710 /// If this function returns true, it can also return a non-zero number of bits
711 /// (in BitsToClear) which indicates that the value it computes is correct for
712 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
713 /// out. For example, to promote something like:
715 /// %B = trunc i64 %A to i32
716 /// %C = lshr i32 %B, 8
717 /// %E = zext i32 %C to i64
719 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
720 /// set to 8 to indicate that the promoted value needs to have bits 24-31
721 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
722 /// clear the top bits anyway, doing this has no extra cost.
724 /// This function works on both vectors and scalars.
725 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
726 InstCombiner &IC, Instruction *CxtI) {
728 if (isa<Constant>(V))
731 Instruction *I = dyn_cast<Instruction>(V);
732 if (!I) return false;
734 // If the input is a truncate from the destination type, we can trivially
736 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
739 // We can't extend or shrink something that has multiple uses: doing so would
740 // require duplicating the instruction in general, which isn't profitable.
741 if (!I->hasOneUse()) return false;
743 unsigned Opc = I->getOpcode(), Tmp;
745 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
746 case Instruction::SExt: // zext(sext(x)) -> sext(x).
747 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
749 case Instruction::And:
750 case Instruction::Or:
751 case Instruction::Xor:
752 case Instruction::Add:
753 case Instruction::Sub:
754 case Instruction::Mul:
755 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
756 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
758 // These can all be promoted if neither operand has 'bits to clear'.
759 if (BitsToClear == 0 && Tmp == 0)
762 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
763 // other side, BitsToClear is ok.
764 if (Tmp == 0 && I->isBitwiseLogicOp()) {
765 // We use MaskedValueIsZero here for generality, but the case we care
766 // about the most is constant RHS.
767 unsigned VSize = V->getType()->getScalarSizeInBits();
768 if (IC.MaskedValueIsZero(I->getOperand(1),
769 APInt::getHighBitsSet(VSize, BitsToClear),
774 // Otherwise, we don't know how to analyze this BitsToClear case yet.
777 case Instruction::Shl:
778 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
779 // upper bits we can reduce BitsToClear by the shift amount.
780 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
781 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
783 uint64_t ShiftAmt = Amt->getZExtValue();
784 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
788 case Instruction::LShr:
789 // We can promote lshr(x, cst) if we can promote x. This requires the
790 // ultimate 'and' to clear out the high zero bits we're clearing out though.
791 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
792 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
794 BitsToClear += Amt->getZExtValue();
795 if (BitsToClear > V->getType()->getScalarSizeInBits())
796 BitsToClear = V->getType()->getScalarSizeInBits();
799 // Cannot promote variable LSHR.
801 case Instruction::Select:
802 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
803 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
804 // TODO: If important, we could handle the case when the BitsToClear are
805 // known zero in the disagreeing side.
810 case Instruction::PHI: {
811 // We can change a phi if we can change all operands. Note that we never
812 // get into trouble with cyclic PHIs here because we only consider
813 // instructions with a single use.
814 PHINode *PN = cast<PHINode>(I);
815 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
817 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
818 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
819 // TODO: If important, we could handle the case when the BitsToClear
820 // are known zero in the disagreeing input.
826 // TODO: Can handle more cases here.
831 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
832 // If this zero extend is only used by a truncate, let the truncate be
833 // eliminated before we try to optimize this zext.
834 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
837 // If one of the common conversion will work, do it.
838 if (Instruction *Result = commonCastTransforms(CI))
841 // See if we can simplify any instructions used by the input whose sole
842 // purpose is to compute bits we don't care about.
843 if (SimplifyDemandedInstructionBits(CI))
846 Value *Src = CI.getOperand(0);
847 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
849 // Attempt to extend the entire input expression tree to the destination
850 // type. Only do this if the dest type is a simple type, don't convert the
851 // expression tree to something weird like i93 unless the source is also
853 unsigned BitsToClear;
854 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
855 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
856 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
857 "Unreasonable BitsToClear");
859 // Okay, we can transform this! Insert the new expression now.
860 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
861 " to avoid zero extend: " << CI << '\n');
862 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
863 assert(Res->getType() == DestTy);
865 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
866 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
868 // If the high bits are already filled with zeros, just replace this
869 // cast with the result.
870 if (MaskedValueIsZero(Res,
871 APInt::getHighBitsSet(DestBitSize,
872 DestBitSize-SrcBitsKept),
874 return replaceInstUsesWith(CI, Res);
876 // We need to emit an AND to clear the high bits.
877 Constant *C = ConstantInt::get(Res->getType(),
878 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
879 return BinaryOperator::CreateAnd(Res, C);
882 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
883 // types and if the sizes are just right we can convert this into a logical
884 // 'and' which will be much cheaper than the pair of casts.
885 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
886 // TODO: Subsume this into EvaluateInDifferentType.
888 // Get the sizes of the types involved. We know that the intermediate type
889 // will be smaller than A or C, but don't know the relation between A and C.
890 Value *A = CSrc->getOperand(0);
891 unsigned SrcSize = A->getType()->getScalarSizeInBits();
892 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
893 unsigned DstSize = CI.getType()->getScalarSizeInBits();
894 // If we're actually extending zero bits, then if
895 // SrcSize < DstSize: zext(a & mask)
896 // SrcSize == DstSize: a & mask
897 // SrcSize > DstSize: trunc(a) & mask
898 if (SrcSize < DstSize) {
899 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
900 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
901 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
902 return new ZExtInst(And, CI.getType());
905 if (SrcSize == DstSize) {
906 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
907 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
910 if (SrcSize > DstSize) {
911 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
912 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
913 return BinaryOperator::CreateAnd(Trunc,
914 ConstantInt::get(Trunc->getType(),
919 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
920 return transformZExtICmp(ICI, CI);
922 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
923 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
924 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp) if at least one
925 // of the (zext icmp) can be eliminated. If so, immediately perform the
926 // according elimination.
927 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
928 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
929 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
930 (transformZExtICmp(LHS, CI, false) ||
931 transformZExtICmp(RHS, CI, false))) {
932 // zext (or icmp, icmp) -> or (zext icmp), (zext icmp)
933 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
934 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
935 BinaryOperator *Or = BinaryOperator::Create(Instruction::Or, LCast, RCast);
937 // Perform the elimination.
938 if (auto *LZExt = dyn_cast<ZExtInst>(LCast))
939 transformZExtICmp(LHS, *LZExt);
940 if (auto *RZExt = dyn_cast<ZExtInst>(RCast))
941 transformZExtICmp(RHS, *RZExt);
947 // zext(trunc(X) & C) -> (X & zext(C)).
951 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
952 X->getType() == CI.getType())
953 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
955 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
957 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
958 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
959 X->getType() == CI.getType()) {
960 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
961 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
967 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
968 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
969 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
970 ICmpInst::Predicate Pred = ICI->getPredicate();
972 // Don't bother if Op1 isn't of vector or integer type.
973 if (!Op1->getType()->isIntOrIntVectorTy())
976 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
977 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
978 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
979 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
980 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
982 Value *Sh = ConstantInt::get(Op0->getType(),
983 Op0->getType()->getScalarSizeInBits()-1);
984 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
985 if (In->getType() != CI.getType())
986 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
988 if (Pred == ICmpInst::ICMP_SGT)
989 In = Builder->CreateNot(In, In->getName()+".not");
990 return replaceInstUsesWith(CI, In);
994 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
995 // If we know that only one bit of the LHS of the icmp can be set and we
996 // have an equality comparison with zero or a power of 2, we can transform
997 // the icmp and sext into bitwise/integer operations.
998 if (ICI->hasOneUse() &&
999 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1000 unsigned BitWidth = Op1C->getType()->getBitWidth();
1001 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1002 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
1004 APInt KnownZeroMask(~KnownZero);
1005 if (KnownZeroMask.isPowerOf2()) {
1006 Value *In = ICI->getOperand(0);
1008 // If the icmp tests for a known zero bit we can constant fold it.
1009 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1010 Value *V = Pred == ICmpInst::ICMP_NE ?
1011 ConstantInt::getAllOnesValue(CI.getType()) :
1012 ConstantInt::getNullValue(CI.getType());
1013 return replaceInstUsesWith(CI, V);
1016 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1017 // sext ((x & 2^n) == 0) -> (x >> n) - 1
1018 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1019 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
1020 // Perform a right shift to place the desired bit in the LSB.
1022 In = Builder->CreateLShr(In,
1023 ConstantInt::get(In->getType(), ShiftAmt));
1025 // At this point "In" is either 1 or 0. Subtract 1 to turn
1026 // {1, 0} -> {0, -1}.
1027 In = Builder->CreateAdd(In,
1028 ConstantInt::getAllOnesValue(In->getType()),
1031 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
1032 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1033 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
1034 // Perform a left shift to place the desired bit in the MSB.
1036 In = Builder->CreateShl(In,
1037 ConstantInt::get(In->getType(), ShiftAmt));
1039 // Distribute the bit over the whole bit width.
1040 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
1041 BitWidth - 1), "sext");
1044 if (CI.getType() == In->getType())
1045 return replaceInstUsesWith(CI, In);
1046 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
1054 /// Return true if we can take the specified value and return it as type Ty
1055 /// without inserting any new casts and without changing the value of the common
1056 /// low bits. This is used by code that tries to promote integer operations to
1057 /// a wider types will allow us to eliminate the extension.
1059 /// This function works on both vectors and scalars.
1061 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1062 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1063 "Can't sign extend type to a smaller type");
1064 // If this is a constant, it can be trivially promoted.
1065 if (isa<Constant>(V))
1068 Instruction *I = dyn_cast<Instruction>(V);
1069 if (!I) return false;
1071 // If this is a truncate from the dest type, we can trivially eliminate it.
1072 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1075 // We can't extend or shrink something that has multiple uses: doing so would
1076 // require duplicating the instruction in general, which isn't profitable.
1077 if (!I->hasOneUse()) return false;
1079 switch (I->getOpcode()) {
1080 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1081 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1082 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1084 case Instruction::And:
1085 case Instruction::Or:
1086 case Instruction::Xor:
1087 case Instruction::Add:
1088 case Instruction::Sub:
1089 case Instruction::Mul:
1090 // These operators can all arbitrarily be extended if their inputs can.
1091 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1092 canEvaluateSExtd(I->getOperand(1), Ty);
1094 //case Instruction::Shl: TODO
1095 //case Instruction::LShr: TODO
1097 case Instruction::Select:
1098 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1099 canEvaluateSExtd(I->getOperand(2), Ty);
1101 case Instruction::PHI: {
1102 // We can change a phi if we can change all operands. Note that we never
1103 // get into trouble with cyclic PHIs here because we only consider
1104 // instructions with a single use.
1105 PHINode *PN = cast<PHINode>(I);
1106 for (Value *IncValue : PN->incoming_values())
1107 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1111 // TODO: Can handle more cases here.
1118 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1119 // If this sign extend is only used by a truncate, let the truncate be
1120 // eliminated before we try to optimize this sext.
1121 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1124 if (Instruction *I = commonCastTransforms(CI))
1127 // See if we can simplify any instructions used by the input whose sole
1128 // purpose is to compute bits we don't care about.
1129 if (SimplifyDemandedInstructionBits(CI))
1132 Value *Src = CI.getOperand(0);
1133 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1135 // If we know that the value being extended is positive, we can use a zext
1137 bool KnownZero, KnownOne;
1138 ComputeSignBit(Src, KnownZero, KnownOne, 0, &CI);
1140 Value *ZExt = Builder->CreateZExt(Src, DestTy);
1141 return replaceInstUsesWith(CI, ZExt);
1144 // Attempt to extend the entire input expression tree to the destination
1145 // type. Only do this if the dest type is a simple type, don't convert the
1146 // expression tree to something weird like i93 unless the source is also
1148 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1149 canEvaluateSExtd(Src, DestTy)) {
1150 // Okay, we can transform this! Insert the new expression now.
1151 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1152 " to avoid sign extend: " << CI << '\n');
1153 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1154 assert(Res->getType() == DestTy);
1156 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1157 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1159 // If the high bits are already filled with sign bit, just replace this
1160 // cast with the result.
1161 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1162 return replaceInstUsesWith(CI, Res);
1164 // We need to emit a shl + ashr to do the sign extend.
1165 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1166 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1170 // If this input is a trunc from our destination, then turn sext(trunc(x))
1172 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1173 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1174 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1175 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1177 // We need to emit a shl + ashr to do the sign extend.
1178 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1179 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1180 return BinaryOperator::CreateAShr(Res, ShAmt);
1183 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1184 return transformSExtICmp(ICI, CI);
1186 // If the input is a shl/ashr pair of a same constant, then this is a sign
1187 // extension from a smaller value. If we could trust arbitrary bitwidth
1188 // integers, we could turn this into a truncate to the smaller bit and then
1189 // use a sext for the whole extension. Since we don't, look deeper and check
1190 // for a truncate. If the source and dest are the same type, eliminate the
1191 // trunc and extend and just do shifts. For example, turn:
1192 // %a = trunc i32 %i to i8
1193 // %b = shl i8 %a, 6
1194 // %c = ashr i8 %b, 6
1195 // %d = sext i8 %c to i32
1197 // %a = shl i32 %i, 30
1198 // %d = ashr i32 %a, 30
1200 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1201 ConstantInt *BA = nullptr, *CA = nullptr;
1202 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1203 m_ConstantInt(CA))) &&
1204 BA == CA && A->getType() == CI.getType()) {
1205 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1206 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1207 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1208 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1209 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1210 return BinaryOperator::CreateAShr(A, ShAmtV);
1217 /// Return a Constant* for the specified floating-point constant if it fits
1218 /// in the specified FP type without changing its value.
1219 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1221 APFloat F = CFP->getValueAPF();
1222 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1224 return ConstantFP::get(CFP->getContext(), F);
1228 /// If this is a floating-point extension instruction, look
1229 /// through it until we get the source value.
1230 static Value *lookThroughFPExtensions(Value *V) {
1231 if (Instruction *I = dyn_cast<Instruction>(V))
1232 if (I->getOpcode() == Instruction::FPExt)
1233 return lookThroughFPExtensions(I->getOperand(0));
1235 // If this value is a constant, return the constant in the smallest FP type
1236 // that can accurately represent it. This allows us to turn
1237 // (float)((double)X+2.0) into x+2.0f.
1238 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1239 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1240 return V; // No constant folding of this.
1241 // See if the value can be truncated to half and then reextended.
1242 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf()))
1244 // See if the value can be truncated to float and then reextended.
1245 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle()))
1247 if (CFP->getType()->isDoubleTy())
1248 return V; // Won't shrink.
1249 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble()))
1251 // Don't try to shrink to various long double types.
1257 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1258 if (Instruction *I = commonCastTransforms(CI))
1260 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1261 // simplify this expression to avoid one or more of the trunc/extend
1262 // operations if we can do so without changing the numerical results.
1264 // The exact manner in which the widths of the operands interact to limit
1265 // what we can and cannot do safely varies from operation to operation, and
1266 // is explained below in the various case statements.
1267 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1268 if (OpI && OpI->hasOneUse()) {
1269 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0));
1270 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1));
1271 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1272 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1273 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1274 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1275 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1276 switch (OpI->getOpcode()) {
1278 case Instruction::FAdd:
1279 case Instruction::FSub:
1280 // For addition and subtraction, the infinitely precise result can
1281 // essentially be arbitrarily wide; proving that double rounding
1282 // will not occur because the result of OpI is exact (as we will for
1283 // FMul, for example) is hopeless. However, we *can* nonetheless
1284 // frequently know that double rounding cannot occur (or that it is
1285 // innocuous) by taking advantage of the specific structure of
1286 // infinitely-precise results that admit double rounding.
1288 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1289 // to represent both sources, we can guarantee that the double
1290 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1291 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1292 // for proof of this fact).
1294 // Note: Figueroa does not consider the case where DstFormat !=
1295 // SrcFormat. It's possible (likely even!) that this analysis
1296 // could be tightened for those cases, but they are rare (the main
1297 // case of interest here is (float)((double)float + float)).
1298 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1299 if (LHSOrig->getType() != CI.getType())
1300 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1301 if (RHSOrig->getType() != CI.getType())
1302 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1304 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1305 RI->copyFastMathFlags(OpI);
1309 case Instruction::FMul:
1310 // For multiplication, the infinitely precise result has at most
1311 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1312 // that such a value can be exactly represented, then no double
1313 // rounding can possibly occur; we can safely perform the operation
1314 // in the destination format if it can represent both sources.
1315 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1316 if (LHSOrig->getType() != CI.getType())
1317 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1318 if (RHSOrig->getType() != CI.getType())
1319 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1321 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1322 RI->copyFastMathFlags(OpI);
1326 case Instruction::FDiv:
1327 // For division, we use again use the bound from Figueroa's
1328 // dissertation. I am entirely certain that this bound can be
1329 // tightened in the unbalanced operand case by an analysis based on
1330 // the diophantine rational approximation bound, but the well-known
1331 // condition used here is a good conservative first pass.
1332 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1333 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1334 if (LHSOrig->getType() != CI.getType())
1335 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1336 if (RHSOrig->getType() != CI.getType())
1337 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1339 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1340 RI->copyFastMathFlags(OpI);
1344 case Instruction::FRem:
1345 // Remainder is straightforward. Remainder is always exact, so the
1346 // type of OpI doesn't enter into things at all. We simply evaluate
1347 // in whichever source type is larger, then convert to the
1348 // destination type.
1349 if (SrcWidth == OpWidth)
1351 if (LHSWidth < SrcWidth)
1352 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1353 else if (RHSWidth <= SrcWidth)
1354 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1355 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1356 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1357 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1358 RI->copyFastMathFlags(OpI);
1359 return CastInst::CreateFPCast(ExactResult, CI.getType());
1363 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1364 if (BinaryOperator::isFNeg(OpI)) {
1365 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1367 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1368 RI->copyFastMathFlags(OpI);
1373 // (fptrunc (select cond, R1, Cst)) -->
1374 // (select cond, (fptrunc R1), (fptrunc Cst))
1376 // - but only if this isn't part of a min/max operation, else we'll
1377 // ruin min/max canonical form which is to have the select and
1378 // compare's operands be of the same type with no casts to look through.
1380 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1382 (isa<ConstantFP>(SI->getOperand(1)) ||
1383 isa<ConstantFP>(SI->getOperand(2))) &&
1384 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) {
1385 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1387 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1389 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1392 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1394 switch (II->getIntrinsicID()) {
1396 case Intrinsic::fabs: {
1397 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1398 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1400 Type *IntrinsicType[] = { CI.getType() };
1401 Function *Overload = Intrinsic::getDeclaration(
1402 CI.getModule(), II->getIntrinsicID(), IntrinsicType);
1404 SmallVector<OperandBundleDef, 1> OpBundles;
1405 II->getOperandBundlesAsDefs(OpBundles);
1407 Value *Args[] = { InnerTrunc };
1408 return CallInst::Create(Overload, Args, OpBundles, II->getName());
1416 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1417 return commonCastTransforms(CI);
1420 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1421 // This is safe if the intermediate type has enough bits in its mantissa to
1422 // accurately represent all values of X. For example, this won't work with
1423 // i64 -> float -> i64.
1424 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1425 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1427 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1429 Value *SrcI = OpI->getOperand(0);
1430 Type *FITy = FI.getType();
1431 Type *OpITy = OpI->getType();
1432 Type *SrcTy = SrcI->getType();
1433 bool IsInputSigned = isa<SIToFPInst>(OpI);
1434 bool IsOutputSigned = isa<FPToSIInst>(FI);
1436 // We can safely assume the conversion won't overflow the output range,
1437 // because (for example) (uint8_t)18293.f is undefined behavior.
1439 // Since we can assume the conversion won't overflow, our decision as to
1440 // whether the input will fit in the float should depend on the minimum
1441 // of the input range and output range.
1443 // This means this is also safe for a signed input and unsigned output, since
1444 // a negative input would lead to undefined behavior.
1445 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1446 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1447 int ActualSize = std::min(InputSize, OutputSize);
1449 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1450 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1451 if (IsInputSigned && IsOutputSigned)
1452 return new SExtInst(SrcI, FITy);
1453 return new ZExtInst(SrcI, FITy);
1455 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1456 return new TruncInst(SrcI, FITy);
1458 return replaceInstUsesWith(FI, SrcI);
1459 return new BitCastInst(SrcI, FITy);
1464 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1465 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1467 return commonCastTransforms(FI);
1469 if (Instruction *I = FoldItoFPtoI(FI))
1472 return commonCastTransforms(FI);
1475 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1476 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1478 return commonCastTransforms(FI);
1480 if (Instruction *I = FoldItoFPtoI(FI))
1483 return commonCastTransforms(FI);
1486 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1487 return commonCastTransforms(CI);
1490 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1491 return commonCastTransforms(CI);
1494 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1495 // If the source integer type is not the intptr_t type for this target, do a
1496 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1497 // cast to be exposed to other transforms.
1498 unsigned AS = CI.getAddressSpace();
1499 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1500 DL.getPointerSizeInBits(AS)) {
1501 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1502 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1503 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1505 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1506 return new IntToPtrInst(P, CI.getType());
1509 if (Instruction *I = commonCastTransforms(CI))
1515 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1516 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1517 Value *Src = CI.getOperand(0);
1519 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1520 // If casting the result of a getelementptr instruction with no offset, turn
1521 // this into a cast of the original pointer!
1522 if (GEP->hasAllZeroIndices() &&
1523 // If CI is an addrspacecast and GEP changes the poiner type, merging
1524 // GEP into CI would undo canonicalizing addrspacecast with different
1525 // pointer types, causing infinite loops.
1526 (!isa<AddrSpaceCastInst>(CI) ||
1527 GEP->getType() == GEP->getPointerOperand()->getType())) {
1528 // Changing the cast operand is usually not a good idea but it is safe
1529 // here because the pointer operand is being replaced with another
1530 // pointer operand so the opcode doesn't need to change.
1532 CI.setOperand(0, GEP->getOperand(0));
1537 return commonCastTransforms(CI);
1540 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1541 // If the destination integer type is not the intptr_t type for this target,
1542 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1543 // to be exposed to other transforms.
1545 Type *Ty = CI.getType();
1546 unsigned AS = CI.getPointerAddressSpace();
1548 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1549 return commonPointerCastTransforms(CI);
1551 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1552 if (Ty->isVectorTy()) // Handle vectors of pointers.
1553 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1555 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1556 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1559 /// This input value (which is known to have vector type) is being zero extended
1560 /// or truncated to the specified vector type.
1561 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
1563 /// The source and destination vector types may have different element types.
1564 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy,
1566 // We can only do this optimization if the output is a multiple of the input
1567 // element size, or the input is a multiple of the output element size.
1568 // Convert the input type to have the same element type as the output.
1569 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1571 if (SrcTy->getElementType() != DestTy->getElementType()) {
1572 // The input types don't need to be identical, but for now they must be the
1573 // same size. There is no specific reason we couldn't handle things like
1574 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1576 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1577 DestTy->getElementType()->getPrimitiveSizeInBits())
1580 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1581 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1584 // Now that the element types match, get the shuffle mask and RHS of the
1585 // shuffle to use, which depends on whether we're increasing or decreasing the
1586 // size of the input.
1587 SmallVector<uint32_t, 16> ShuffleMask;
1590 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1591 // If we're shrinking the number of elements, just shuffle in the low
1592 // elements from the input and use undef as the second shuffle input.
1593 V2 = UndefValue::get(SrcTy);
1594 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1595 ShuffleMask.push_back(i);
1598 // If we're increasing the number of elements, shuffle in all of the
1599 // elements from InVal and fill the rest of the result elements with zeros
1600 // from a constant zero.
1601 V2 = Constant::getNullValue(SrcTy);
1602 unsigned SrcElts = SrcTy->getNumElements();
1603 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1604 ShuffleMask.push_back(i);
1606 // The excess elements reference the first element of the zero input.
1607 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1608 ShuffleMask.push_back(SrcElts);
1611 return new ShuffleVectorInst(InVal, V2,
1612 ConstantDataVector::get(V2->getContext(),
1616 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1617 return Value % Ty->getPrimitiveSizeInBits() == 0;
1620 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1621 return Value / Ty->getPrimitiveSizeInBits();
1624 /// V is a value which is inserted into a vector of VecEltTy.
1625 /// Look through the value to see if we can decompose it into
1626 /// insertions into the vector. See the example in the comment for
1627 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1628 /// The type of V is always a non-zero multiple of VecEltTy's size.
1629 /// Shift is the number of bits between the lsb of V and the lsb of
1632 /// This returns false if the pattern can't be matched or true if it can,
1633 /// filling in Elements with the elements found here.
1634 static bool collectInsertionElements(Value *V, unsigned Shift,
1635 SmallVectorImpl<Value *> &Elements,
1636 Type *VecEltTy, bool isBigEndian) {
1637 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1638 "Shift should be a multiple of the element type size");
1640 // Undef values never contribute useful bits to the result.
1641 if (isa<UndefValue>(V)) return true;
1643 // If we got down to a value of the right type, we win, try inserting into the
1645 if (V->getType() == VecEltTy) {
1646 // Inserting null doesn't actually insert any elements.
1647 if (Constant *C = dyn_cast<Constant>(V))
1648 if (C->isNullValue())
1651 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1653 ElementIndex = Elements.size() - ElementIndex - 1;
1655 // Fail if multiple elements are inserted into this slot.
1656 if (Elements[ElementIndex])
1659 Elements[ElementIndex] = V;
1663 if (Constant *C = dyn_cast<Constant>(V)) {
1664 // Figure out the # elements this provides, and bitcast it or slice it up
1666 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1668 // If the constant is the size of a vector element, we just need to bitcast
1669 // it to the right type so it gets properly inserted.
1671 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1672 Shift, Elements, VecEltTy, isBigEndian);
1674 // Okay, this is a constant that covers multiple elements. Slice it up into
1675 // pieces and insert each element-sized piece into the vector.
1676 if (!isa<IntegerType>(C->getType()))
1677 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1678 C->getType()->getPrimitiveSizeInBits()));
1679 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1680 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1682 for (unsigned i = 0; i != NumElts; ++i) {
1683 unsigned ShiftI = Shift+i*ElementSize;
1684 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1686 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1687 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1694 if (!V->hasOneUse()) return false;
1696 Instruction *I = dyn_cast<Instruction>(V);
1697 if (!I) return false;
1698 switch (I->getOpcode()) {
1699 default: return false; // Unhandled case.
1700 case Instruction::BitCast:
1701 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1703 case Instruction::ZExt:
1704 if (!isMultipleOfTypeSize(
1705 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1708 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1710 case Instruction::Or:
1711 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1713 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1715 case Instruction::Shl: {
1716 // Must be shifting by a constant that is a multiple of the element size.
1717 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1718 if (!CI) return false;
1719 Shift += CI->getZExtValue();
1720 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1721 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1729 /// If the input is an 'or' instruction, we may be doing shifts and ors to
1730 /// assemble the elements of the vector manually.
1731 /// Try to rip the code out and replace it with insertelements. This is to
1732 /// optimize code like this:
1734 /// %tmp37 = bitcast float %inc to i32
1735 /// %tmp38 = zext i32 %tmp37 to i64
1736 /// %tmp31 = bitcast float %inc5 to i32
1737 /// %tmp32 = zext i32 %tmp31 to i64
1738 /// %tmp33 = shl i64 %tmp32, 32
1739 /// %ins35 = or i64 %tmp33, %tmp38
1740 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1742 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1743 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
1745 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1746 Value *IntInput = CI.getOperand(0);
1748 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1749 if (!collectInsertionElements(IntInput, 0, Elements,
1750 DestVecTy->getElementType(),
1751 IC.getDataLayout().isBigEndian()))
1754 // If we succeeded, we know that all of the element are specified by Elements
1755 // or are zero if Elements has a null entry. Recast this as a set of
1757 Value *Result = Constant::getNullValue(CI.getType());
1758 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1759 if (!Elements[i]) continue; // Unset element.
1761 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1762 IC.Builder->getInt32(i));
1768 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
1769 /// vector followed by extract element. The backend tends to handle bitcasts of
1770 /// vectors better than bitcasts of scalars because vector registers are
1771 /// usually not type-specific like scalar integer or scalar floating-point.
1772 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
1774 const DataLayout &DL) {
1775 // TODO: Create and use a pattern matcher for ExtractElementInst.
1776 auto *ExtElt = dyn_cast<ExtractElementInst>(BitCast.getOperand(0));
1777 if (!ExtElt || !ExtElt->hasOneUse())
1780 // The bitcast must be to a vectorizable type, otherwise we can't make a new
1781 // type to extract from.
1782 Type *DestType = BitCast.getType();
1783 if (!VectorType::isValidElementType(DestType))
1786 unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements();
1787 auto *NewVecType = VectorType::get(DestType, NumElts);
1788 auto *NewBC = IC.Builder->CreateBitCast(ExtElt->getVectorOperand(),
1790 return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand());
1793 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
1794 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
1795 InstCombiner::BuilderTy &Builder) {
1796 Type *DestTy = BitCast.getType();
1798 if (!DestTy->getScalarType()->isIntegerTy() ||
1799 !match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
1800 !BO->isBitwiseLogicOp())
1803 // FIXME: This transform is restricted to vector types to avoid backend
1804 // problems caused by creating potentially illegal operations. If a fix-up is
1805 // added to handle that situation, we can remove this check.
1806 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
1810 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
1811 X->getType() == DestTy && !isa<Constant>(X)) {
1812 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
1813 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
1814 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
1817 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
1818 X->getType() == DestTy && !isa<Constant>(X)) {
1819 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
1820 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
1821 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
1827 /// Change the type of a select if we can eliminate a bitcast.
1828 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
1829 InstCombiner::BuilderTy &Builder) {
1830 Value *Cond, *TVal, *FVal;
1831 if (!match(BitCast.getOperand(0),
1832 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
1835 // A vector select must maintain the same number of elements in its operands.
1836 Type *CondTy = Cond->getType();
1837 Type *DestTy = BitCast.getType();
1838 if (CondTy->isVectorTy()) {
1839 if (!DestTy->isVectorTy())
1841 if (DestTy->getVectorNumElements() != CondTy->getVectorNumElements())
1845 // FIXME: This transform is restricted from changing the select between
1846 // scalars and vectors to avoid backend problems caused by creating
1847 // potentially illegal operations. If a fix-up is added to handle that
1848 // situation, we can remove this check.
1849 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
1852 auto *Sel = cast<Instruction>(BitCast.getOperand(0));
1854 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1855 !isa<Constant>(X)) {
1856 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
1857 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
1858 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
1861 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
1862 !isa<Constant>(X)) {
1863 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
1864 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
1865 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
1871 /// Check if all users of CI are StoreInsts.
1872 static bool hasStoreUsersOnly(CastInst &CI) {
1873 for (User *U : CI.users()) {
1874 if (!isa<StoreInst>(U))
1880 /// This function handles following case
1886 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
1887 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
1888 Instruction *InstCombiner::optimizeBitCastFromPhi(CastInst &CI, PHINode *PN) {
1889 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
1890 if (hasStoreUsersOnly(CI))
1893 Value *Src = CI.getOperand(0);
1894 Type *SrcTy = Src->getType(); // Type B
1895 Type *DestTy = CI.getType(); // Type A
1897 SmallVector<PHINode *, 4> PhiWorklist;
1898 SmallSetVector<PHINode *, 4> OldPhiNodes;
1900 // Find all of the A->B casts and PHI nodes.
1901 // We need to inpect all related PHI nodes, but PHIs can be cyclic, so
1902 // OldPhiNodes is used to track all known PHI nodes, before adding a new
1903 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
1904 PhiWorklist.push_back(PN);
1905 OldPhiNodes.insert(PN);
1906 while (!PhiWorklist.empty()) {
1907 auto *OldPN = PhiWorklist.pop_back_val();
1908 for (Value *IncValue : OldPN->incoming_values()) {
1909 if (isa<Constant>(IncValue))
1912 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
1913 // If there is a sequence of one or more load instructions, each loaded
1914 // value is used as address of later load instruction, bitcast is
1915 // necessary to change the value type, don't optimize it. For
1916 // simplicity we give up if the load address comes from another load.
1917 Value *Addr = LI->getOperand(0);
1918 if (Addr == &CI || isa<LoadInst>(Addr))
1920 if (LI->hasOneUse() && LI->isSimple())
1922 // If a LoadInst has more than one use, changing the type of loaded
1923 // value may create another bitcast.
1927 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
1928 if (OldPhiNodes.insert(PNode))
1929 PhiWorklist.push_back(PNode);
1933 auto *BCI = dyn_cast<BitCastInst>(IncValue);
1934 // We can't handle other instructions.
1938 // Verify it's a A->B cast.
1939 Type *TyA = BCI->getOperand(0)->getType();
1940 Type *TyB = BCI->getType();
1941 if (TyA != DestTy || TyB != SrcTy)
1946 // For each old PHI node, create a corresponding new PHI node with a type A.
1947 SmallDenseMap<PHINode *, PHINode *> NewPNodes;
1948 for (auto *OldPN : OldPhiNodes) {
1949 Builder->SetInsertPoint(OldPN);
1950 PHINode *NewPN = Builder->CreatePHI(DestTy, OldPN->getNumOperands());
1951 NewPNodes[OldPN] = NewPN;
1954 // Fill in the operands of new PHI nodes.
1955 for (auto *OldPN : OldPhiNodes) {
1956 PHINode *NewPN = NewPNodes[OldPN];
1957 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
1958 Value *V = OldPN->getOperand(j);
1959 Value *NewV = nullptr;
1960 if (auto *C = dyn_cast<Constant>(V)) {
1961 NewV = ConstantExpr::getBitCast(C, DestTy);
1962 } else if (auto *LI = dyn_cast<LoadInst>(V)) {
1963 Builder->SetInsertPoint(LI->getNextNode());
1964 NewV = Builder->CreateBitCast(LI, DestTy);
1966 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
1967 NewV = BCI->getOperand(0);
1968 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
1969 NewV = NewPNodes[PrevPN];
1972 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
1976 // If there is a store with type B, change it to type A.
1977 for (User *U : PN->users()) {
1978 auto *SI = dyn_cast<StoreInst>(U);
1979 if (SI && SI->isSimple() && SI->getOperand(0) == PN) {
1980 Builder->SetInsertPoint(SI);
1982 cast<BitCastInst>(Builder->CreateBitCast(NewPNodes[PN], SrcTy));
1983 SI->setOperand(0, NewBC);
1985 assert(hasStoreUsersOnly(*NewBC));
1989 return replaceInstUsesWith(CI, NewPNodes[PN]);
1992 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1993 // If the operands are integer typed then apply the integer transforms,
1994 // otherwise just apply the common ones.
1995 Value *Src = CI.getOperand(0);
1996 Type *SrcTy = Src->getType();
1997 Type *DestTy = CI.getType();
1999 // Get rid of casts from one type to the same type. These are useless and can
2000 // be replaced by the operand.
2001 if (DestTy == Src->getType())
2002 return replaceInstUsesWith(CI, Src);
2004 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
2005 PointerType *SrcPTy = cast<PointerType>(SrcTy);
2006 Type *DstElTy = DstPTy->getElementType();
2007 Type *SrcElTy = SrcPTy->getElementType();
2009 // If we are casting a alloca to a pointer to a type of the same
2010 // size, rewrite the allocation instruction to allocate the "right" type.
2011 // There is no need to modify malloc calls because it is their bitcast that
2012 // needs to be cleaned up.
2013 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
2014 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
2017 // When the type pointed to is not sized the cast cannot be
2018 // turned into a gep.
2020 cast<PointerType>(Src->getType()->getScalarType())->getElementType();
2021 if (!PointeeType->isSized())
2024 // If the source and destination are pointers, and this cast is equivalent
2025 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
2026 // This can enhance SROA and other transforms that want type-safe pointers.
2027 unsigned NumZeros = 0;
2028 while (SrcElTy != DstElTy &&
2029 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
2030 SrcElTy->getNumContainedTypes() /* not "{}" */) {
2031 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
2035 // If we found a path from the src to dest, create the getelementptr now.
2036 if (SrcElTy == DstElTy) {
2037 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0));
2038 return GetElementPtrInst::CreateInBounds(Src, Idxs);
2042 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
2043 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
2044 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
2045 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
2046 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2047 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
2050 if (isa<IntegerType>(SrcTy)) {
2051 // If this is a cast from an integer to vector, check to see if the input
2052 // is a trunc or zext of a bitcast from vector. If so, we can replace all
2053 // the casts with a shuffle and (potentially) a bitcast.
2054 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2055 CastInst *SrcCast = cast<CastInst>(Src);
2056 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2057 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2058 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0),
2059 cast<VectorType>(DestTy), *this))
2063 // If the input is an 'or' instruction, we may be doing shifts and ors to
2064 // assemble the elements of the vector manually. Try to rip the code out
2065 // and replace it with insertelements.
2066 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2067 return replaceInstUsesWith(CI, V);
2071 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
2072 if (SrcVTy->getNumElements() == 1) {
2073 // If our destination is not a vector, then make this a straight
2074 // scalar-scalar cast.
2075 if (!DestTy->isVectorTy()) {
2077 Builder->CreateExtractElement(Src,
2078 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2079 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2082 // Otherwise, see if our source is an insert. If so, then use the scalar
2083 // component directly.
2084 if (InsertElementInst *IEI =
2085 dyn_cast<InsertElementInst>(CI.getOperand(0)))
2086 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
2091 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
2092 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
2093 // a bitcast to a vector with the same # elts.
2094 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
2095 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
2096 SVI->getType()->getNumElements() ==
2097 SVI->getOperand(0)->getType()->getVectorNumElements()) {
2099 // If either of the operands is a cast from CI.getType(), then
2100 // evaluating the shuffle in the casted destination's type will allow
2101 // us to eliminate at least one cast.
2102 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
2103 Tmp->getOperand(0)->getType() == DestTy) ||
2104 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
2105 Tmp->getOperand(0)->getType() == DestTy)) {
2106 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
2107 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
2108 // Return a new shuffle vector. Use the same element ID's, as we
2109 // know the vector types match #elts.
2110 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
2115 // Handle the A->B->A cast, and there is an intervening PHI node.
2116 if (PHINode *PN = dyn_cast<PHINode>(Src))
2117 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2120 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this, DL))
2123 if (Instruction *I = foldBitCastBitwiseLogic(CI, *Builder))
2126 if (Instruction *I = foldBitCastSelect(CI, *Builder))
2129 if (SrcTy->isPointerTy())
2130 return commonPointerCastTransforms(CI);
2131 return commonCastTransforms(CI);
2134 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2135 // If the destination pointer element type is not the same as the source's
2136 // first do a bitcast to the destination type, and then the addrspacecast.
2137 // This allows the cast to be exposed to other transforms.
2138 Value *Src = CI.getOperand(0);
2139 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
2140 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
2142 Type *DestElemTy = DestTy->getElementType();
2143 if (SrcTy->getElementType() != DestElemTy) {
2144 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
2145 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
2146 // Handle vectors of pointers.
2147 MidTy = VectorType::get(MidTy, VT->getNumElements());
2150 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
2151 return new AddrSpaceCastInst(NewBitCast, CI.getType());
2154 return commonPointerCastTransforms(CI);