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/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/PatternMatch.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
20 using namespace PatternMatch;
22 #define DEBUG_TYPE "instcombine"
24 /// Analyze 'Val', seeing if it is a simple linear expression.
25 /// If so, decompose it, returning some value X, such that Val is
28 static Value *decomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
30 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
31 Offset = CI->getZExtValue();
33 return ConstantInt::get(Val->getType(), 0);
36 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
37 // Cannot look past anything that might overflow.
38 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
39 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
45 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
46 if (I->getOpcode() == Instruction::Shl) {
47 // This is a value scaled by '1 << the shift amt'.
48 Scale = UINT64_C(1) << RHS->getZExtValue();
50 return I->getOperand(0);
53 if (I->getOpcode() == Instruction::Mul) {
54 // This value is scaled by 'RHS'.
55 Scale = RHS->getZExtValue();
57 return I->getOperand(0);
60 if (I->getOpcode() == Instruction::Add) {
61 // We have X+C. Check to see if we really have (X*C2)+C1,
62 // where C1 is divisible by C2.
65 decomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
66 Offset += RHS->getZExtValue();
73 // Otherwise, we can't look past this.
79 /// If we find a cast of an allocation instruction, try to eliminate the cast by
80 /// moving the type information into the alloc.
81 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
83 PointerType *PTy = cast<PointerType>(CI.getType());
85 BuilderTy AllocaBuilder(*Builder);
86 AllocaBuilder.SetInsertPoint(&AI);
88 // Get the type really allocated and the type casted to.
89 Type *AllocElTy = AI.getAllocatedType();
90 Type *CastElTy = PTy->getElementType();
91 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
93 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy);
94 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy);
95 if (CastElTyAlign < AllocElTyAlign) return nullptr;
97 // If the allocation has multiple uses, only promote it if we are strictly
98 // increasing the alignment of the resultant allocation. If we keep it the
99 // same, we open the door to infinite loops of various kinds.
100 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
102 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy);
103 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy);
104 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
106 // If the allocation has multiple uses, only promote it if we're not
107 // shrinking the amount of memory being allocated.
108 uint64_t AllocElTyStoreSize = DL.getTypeStoreSize(AllocElTy);
109 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy);
110 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
112 // See if we can satisfy the modulus by pulling a scale out of the array
114 unsigned ArraySizeScale;
115 uint64_t ArrayOffset;
116 Value *NumElements = // See if the array size is a decomposable linear expr.
117 decomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
119 // If we can now satisfy the modulus, by using a non-1 scale, we really can
121 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
122 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
124 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
125 Value *Amt = nullptr;
129 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
130 // Insert before the alloca, not before the cast.
131 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
134 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
135 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
137 Amt = AllocaBuilder.CreateAdd(Amt, Off);
140 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
141 New->setAlignment(AI.getAlignment());
143 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
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 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
158 /// true for, actually insert the code to evaluate the expression.
159 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
161 if (Constant *C = dyn_cast<Constant>(V)) {
162 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
163 // If we got a constantexpr back, try to simplify it with DL info.
164 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
165 C = ConstantFoldConstantExpression(CE, DL, TLI);
169 // Otherwise, it must be an instruction.
170 Instruction *I = cast<Instruction>(V);
171 Instruction *Res = nullptr;
172 unsigned Opc = I->getOpcode();
174 case Instruction::Add:
175 case Instruction::Sub:
176 case Instruction::Mul:
177 case Instruction::And:
178 case Instruction::Or:
179 case Instruction::Xor:
180 case Instruction::AShr:
181 case Instruction::LShr:
182 case Instruction::Shl:
183 case Instruction::UDiv:
184 case Instruction::URem: {
185 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
186 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
187 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
190 case Instruction::Trunc:
191 case Instruction::ZExt:
192 case Instruction::SExt:
193 // If the source type of the cast is the type we're trying for then we can
194 // just return the source. There's no need to insert it because it is not
196 if (I->getOperand(0)->getType() == Ty)
197 return I->getOperand(0);
199 // Otherwise, must be the same type of cast, so just reinsert a new one.
200 // This also handles the case of zext(trunc(x)) -> zext(x).
201 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
202 Opc == Instruction::SExt);
204 case Instruction::Select: {
205 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
206 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
207 Res = SelectInst::Create(I->getOperand(0), True, False);
210 case Instruction::PHI: {
211 PHINode *OPN = cast<PHINode>(I);
212 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
213 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
215 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(const CastInst *CI, ///< First cast instruction
235 unsigned opcode, ///< Opcode for the second cast
236 Type *DstTy, ///< Target type for the second cast
237 const DataLayout &DL) {
238 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
239 Type *MidTy = CI->getType(); // B from above
241 // Get the opcodes of the two Cast instructions
242 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
243 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
245 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
247 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
249 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
250 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
251 DstTy, SrcIntPtrTy, MidIntPtrTy,
254 // We don't want to form an inttoptr or ptrtoint that converts to an integer
255 // type that differs from the pointer size.
256 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
257 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
260 return Instruction::CastOps(Res);
263 /// Return true if the cast from "V to Ty" actually results in any code being
264 /// generated and is interesting to optimize out.
265 /// If the cast can be eliminated by some other simple transformation, we prefer
266 /// to do the simplification first.
267 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
269 // Noop casts and casts of constants should be eliminated trivially.
270 if (V->getType() == Ty || isa<Constant>(V)) return false;
272 // If this is another cast that can be eliminated, we prefer to have it
274 if (const CastInst *CI = dyn_cast<CastInst>(V))
275 if (isEliminableCastPair(CI, opc, Ty, DL))
278 // If this is a vector sext from a compare, then we don't want to break the
279 // idiom where each element of the extended vector is either zero or all ones.
280 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
287 /// @brief Implement the transforms common to all CastInst visitors.
288 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
289 Value *Src = CI.getOperand(0);
291 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
293 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
294 if (Instruction::CastOps opc =
295 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
296 // The first cast (CSrc) is eliminable so we need to fix up or replace
297 // the second cast (CI). CSrc will then have a good chance of being dead.
298 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
302 // If we are casting a select then fold the cast into the select
303 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
304 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
307 // If we are casting a PHI then fold the cast into the PHI
308 if (isa<PHINode>(Src)) {
309 // We don't do this if this would create a PHI node with an illegal type if
310 // it is currently legal.
311 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
312 ShouldChangeType(CI.getType(), Src->getType()))
313 if (Instruction *NV = FoldOpIntoPhi(CI))
320 /// Return true if we can evaluate the specified expression tree as type Ty
321 /// instead of its larger type, and arrive with the same value.
322 /// This is used by code that tries to eliminate truncates.
324 /// Ty will always be a type smaller than V. We should return true if trunc(V)
325 /// can be computed by computing V in the smaller type. If V is an instruction,
326 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
327 /// makes sense if x and y can be efficiently truncated.
329 /// This function works on both vectors and scalars.
331 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
333 // We can always evaluate constants in another type.
334 if (isa<Constant>(V))
337 Instruction *I = dyn_cast<Instruction>(V);
338 if (!I) return false;
340 Type *OrigTy = V->getType();
342 // If this is an extension from the dest type, we can eliminate it, even if it
343 // has multiple uses.
344 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
345 I->getOperand(0)->getType() == Ty)
348 // We can't extend or shrink something that has multiple uses: doing so would
349 // require duplicating the instruction in general, which isn't profitable.
350 if (!I->hasOneUse()) return false;
352 unsigned Opc = I->getOpcode();
354 case Instruction::Add:
355 case Instruction::Sub:
356 case Instruction::Mul:
357 case Instruction::And:
358 case Instruction::Or:
359 case Instruction::Xor:
360 // These operators can all arbitrarily be extended or truncated.
361 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
362 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
364 case Instruction::UDiv:
365 case Instruction::URem: {
366 // UDiv and URem can be truncated if all the truncated bits are zero.
367 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
368 uint32_t BitWidth = Ty->getScalarSizeInBits();
369 if (BitWidth < OrigBitWidth) {
370 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
371 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
372 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
373 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
374 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
379 case Instruction::Shl:
380 // If we are truncating the result of this SHL, and if it's a shift of a
381 // constant amount, we can always perform a SHL in a smaller type.
382 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
383 uint32_t BitWidth = Ty->getScalarSizeInBits();
384 if (CI->getLimitedValue(BitWidth) < BitWidth)
385 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
388 case Instruction::LShr:
389 // If this is a truncate of a logical shr, we can truncate it to a smaller
390 // lshr iff we know that the bits we would otherwise be shifting in are
392 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
393 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
394 uint32_t BitWidth = Ty->getScalarSizeInBits();
395 if (IC.MaskedValueIsZero(I->getOperand(0),
396 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
397 CI->getLimitedValue(BitWidth) < BitWidth) {
398 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
402 case Instruction::Trunc:
403 // trunc(trunc(x)) -> trunc(x)
405 case Instruction::ZExt:
406 case Instruction::SExt:
407 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
408 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
410 case Instruction::Select: {
411 SelectInst *SI = cast<SelectInst>(I);
412 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
413 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
415 case Instruction::PHI: {
416 // We can change a phi if we can change all operands. Note that we never
417 // get into trouble with cyclic PHIs here because we only consider
418 // instructions with a single use.
419 PHINode *PN = cast<PHINode>(I);
420 for (Value *IncValue : PN->incoming_values())
421 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
426 // TODO: Can handle more cases here.
433 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
434 if (Instruction *Result = commonCastTransforms(CI))
437 // Test if the trunc is the user of a select which is part of a
438 // minimum or maximum operation. If so, don't do any more simplification.
439 // Even simplifying demanded bits can break the canonical form of a
442 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)))
443 if (matchSelectPattern(SI, LHS, RHS).Flavor != SPF_UNKNOWN)
446 // See if we can simplify any instructions used by the input whose sole
447 // purpose is to compute bits we don't care about.
448 if (SimplifyDemandedInstructionBits(CI))
451 Value *Src = CI.getOperand(0);
452 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
454 // Attempt to truncate the entire input expression tree to the destination
455 // type. Only do this if the dest type is a simple type, don't convert the
456 // expression tree to something weird like i93 unless the source is also
458 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
459 canEvaluateTruncated(Src, DestTy, *this, &CI)) {
461 // If this cast is a truncate, evaluting in a different type always
462 // eliminates the cast, so it is always a win.
463 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
464 " to avoid cast: " << CI << '\n');
465 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
466 assert(Res->getType() == DestTy);
467 return ReplaceInstUsesWith(CI, Res);
470 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
471 if (DestTy->getScalarSizeInBits() == 1) {
472 Constant *One = ConstantInt::get(Src->getType(), 1);
473 Src = Builder->CreateAnd(Src, One);
474 Value *Zero = Constant::getNullValue(Src->getType());
475 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
478 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
479 Value *A = nullptr; ConstantInt *Cst = nullptr;
480 if (Src->hasOneUse() &&
481 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
482 // We have three types to worry about here, the type of A, the source of
483 // the truncate (MidSize), and the destination of the truncate. We know that
484 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
485 // between ASize and ResultSize.
486 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
488 // If the shift amount is larger than the size of A, then the result is
489 // known to be zero because all the input bits got shifted out.
490 if (Cst->getZExtValue() >= ASize)
491 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
493 // Since we're doing an lshr and a zero extend, and know that the shift
494 // amount is smaller than ASize, it is always safe to do the shift in A's
495 // type, then zero extend or truncate to the result.
496 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
497 Shift->takeName(Src);
498 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
501 // Transform trunc(lshr (sext A), Cst) to ashr A, Cst to eliminate type
503 // It works because bits coming from sign extension have the same value as
504 // sign bit of the original value; performing ashr instead of lshr
505 // generates bits of the same value as the sign bit.
506 if (Src->hasOneUse() &&
507 match(Src, m_LShr(m_SExt(m_Value(A)), m_ConstantInt(Cst))) &&
508 cast<Instruction>(Src)->getOperand(0)->hasOneUse()) {
509 const unsigned ASize = A->getType()->getPrimitiveSizeInBits();
510 // This optimization can be only performed when zero bits generated by
511 // the original lshr aren't pulled into the value after truncation, so we
512 // can only shift by values smaller then the size of destination type (in
514 if (Cst->getValue().ult(ASize)) {
515 Value *Shift = Builder->CreateAShr(A, Cst->getZExtValue());
516 Shift->takeName(Src);
517 return CastInst::CreateIntegerCast(Shift, CI.getType(), true);
521 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
522 // type isn't non-native.
523 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
524 ShouldChangeType(Src->getType(), CI.getType()) &&
525 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
526 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
527 return BinaryOperator::CreateAnd(NewTrunc,
528 ConstantExpr::getTrunc(Cst, CI.getType()));
534 /// Transform (zext icmp) to bitwise / integer operations in order to eliminate
536 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
538 // If we are just checking for a icmp eq of a single bit and zext'ing it
539 // to an integer, then shift the bit to the appropriate place and then
540 // cast to integer to avoid the comparison.
541 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
542 const APInt &Op1CV = Op1C->getValue();
544 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
545 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
546 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
547 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
548 if (!DoXform) return ICI;
550 Value *In = ICI->getOperand(0);
551 Value *Sh = ConstantInt::get(In->getType(),
552 In->getType()->getScalarSizeInBits()-1);
553 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
554 if (In->getType() != CI.getType())
555 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
557 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
558 Constant *One = ConstantInt::get(In->getType(), 1);
559 In = Builder->CreateXor(In, One, In->getName()+".not");
562 return ReplaceInstUsesWith(CI, In);
565 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
566 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
567 // zext (X == 1) to i32 --> X iff X has only the low bit set.
568 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
569 // zext (X != 0) to i32 --> X iff X has only the low bit set.
570 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
571 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
572 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
573 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
574 // This only works for EQ and NE
576 // If Op1C some other power of two, convert:
577 uint32_t BitWidth = Op1C->getType()->getBitWidth();
578 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
579 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
581 APInt KnownZeroMask(~KnownZero);
582 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
583 if (!DoXform) return ICI;
585 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
586 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
587 // (X&4) == 2 --> false
588 // (X&4) != 2 --> true
589 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
591 Res = ConstantExpr::getZExt(Res, CI.getType());
592 return ReplaceInstUsesWith(CI, Res);
595 uint32_t ShiftAmt = KnownZeroMask.logBase2();
596 Value *In = ICI->getOperand(0);
598 // Perform a logical shr by shiftamt.
599 // Insert the shift to put the result in the low bit.
600 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
601 In->getName()+".lobit");
604 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
605 Constant *One = ConstantInt::get(In->getType(), 1);
606 In = Builder->CreateXor(In, One);
609 if (CI.getType() == In->getType())
610 return ReplaceInstUsesWith(CI, In);
611 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
616 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
617 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
618 // may lead to additional simplifications.
619 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
620 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
621 uint32_t BitWidth = ITy->getBitWidth();
622 Value *LHS = ICI->getOperand(0);
623 Value *RHS = ICI->getOperand(1);
625 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
626 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
627 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
628 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
630 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
631 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
632 APInt UnknownBit = ~KnownBits;
633 if (UnknownBit.countPopulation() == 1) {
634 if (!DoXform) return ICI;
636 Value *Result = Builder->CreateXor(LHS, RHS);
638 // Mask off any bits that are set and won't be shifted away.
639 if (KnownOneLHS.uge(UnknownBit))
640 Result = Builder->CreateAnd(Result,
641 ConstantInt::get(ITy, UnknownBit));
643 // Shift the bit we're testing down to the lsb.
644 Result = Builder->CreateLShr(
645 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
647 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
648 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
649 Result->takeName(ICI);
650 return ReplaceInstUsesWith(CI, Result);
659 /// Determine if the specified value can be computed in the specified wider type
660 /// and produce the same low bits. If not, return false.
662 /// If this function returns true, it can also return a non-zero number of bits
663 /// (in BitsToClear) which indicates that the value it computes is correct for
664 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
665 /// out. For example, to promote something like:
667 /// %B = trunc i64 %A to i32
668 /// %C = lshr i32 %B, 8
669 /// %E = zext i32 %C to i64
671 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
672 /// set to 8 to indicate that the promoted value needs to have bits 24-31
673 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
674 /// clear the top bits anyway, doing this has no extra cost.
676 /// This function works on both vectors and scalars.
677 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
678 InstCombiner &IC, Instruction *CxtI) {
680 if (isa<Constant>(V))
683 Instruction *I = dyn_cast<Instruction>(V);
684 if (!I) return false;
686 // If the input is a truncate from the destination type, we can trivially
688 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
691 // We can't extend or shrink something that has multiple uses: doing so would
692 // require duplicating the instruction in general, which isn't profitable.
693 if (!I->hasOneUse()) return false;
695 unsigned Opc = I->getOpcode(), Tmp;
697 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
698 case Instruction::SExt: // zext(sext(x)) -> sext(x).
699 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
701 case Instruction::And:
702 case Instruction::Or:
703 case Instruction::Xor:
704 case Instruction::Add:
705 case Instruction::Sub:
706 case Instruction::Mul:
707 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
708 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
710 // These can all be promoted if neither operand has 'bits to clear'.
711 if (BitsToClear == 0 && Tmp == 0)
714 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
715 // other side, BitsToClear is ok.
717 (Opc == Instruction::And || Opc == Instruction::Or ||
718 Opc == Instruction::Xor)) {
719 // We use MaskedValueIsZero here for generality, but the case we care
720 // about the most is constant RHS.
721 unsigned VSize = V->getType()->getScalarSizeInBits();
722 if (IC.MaskedValueIsZero(I->getOperand(1),
723 APInt::getHighBitsSet(VSize, BitsToClear),
728 // Otherwise, we don't know how to analyze this BitsToClear case yet.
731 case Instruction::Shl:
732 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
733 // upper bits we can reduce BitsToClear by the shift amount.
734 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
735 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
737 uint64_t ShiftAmt = Amt->getZExtValue();
738 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
742 case Instruction::LShr:
743 // We can promote lshr(x, cst) if we can promote x. This requires the
744 // ultimate 'and' to clear out the high zero bits we're clearing out though.
745 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
746 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
748 BitsToClear += Amt->getZExtValue();
749 if (BitsToClear > V->getType()->getScalarSizeInBits())
750 BitsToClear = V->getType()->getScalarSizeInBits();
753 // Cannot promote variable LSHR.
755 case Instruction::Select:
756 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
757 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
758 // TODO: If important, we could handle the case when the BitsToClear are
759 // known zero in the disagreeing side.
764 case Instruction::PHI: {
765 // We can change a phi if we can change all operands. Note that we never
766 // get into trouble with cyclic PHIs here because we only consider
767 // instructions with a single use.
768 PHINode *PN = cast<PHINode>(I);
769 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
771 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
772 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
773 // TODO: If important, we could handle the case when the BitsToClear
774 // are known zero in the disagreeing input.
780 // TODO: Can handle more cases here.
785 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
786 // If this zero extend is only used by a truncate, let the truncate be
787 // eliminated before we try to optimize this zext.
788 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
791 // If one of the common conversion will work, do it.
792 if (Instruction *Result = commonCastTransforms(CI))
795 // See if we can simplify any instructions used by the input whose sole
796 // purpose is to compute bits we don't care about.
797 if (SimplifyDemandedInstructionBits(CI))
800 Value *Src = CI.getOperand(0);
801 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
803 // Attempt to extend the entire input expression tree to the destination
804 // type. Only do this if the dest type is a simple type, don't convert the
805 // expression tree to something weird like i93 unless the source is also
807 unsigned BitsToClear;
808 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
809 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
810 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
811 "Unreasonable BitsToClear");
813 // Okay, we can transform this! Insert the new expression now.
814 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
815 " to avoid zero extend: " << CI);
816 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
817 assert(Res->getType() == DestTy);
819 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
820 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
822 // If the high bits are already filled with zeros, just replace this
823 // cast with the result.
824 if (MaskedValueIsZero(Res,
825 APInt::getHighBitsSet(DestBitSize,
826 DestBitSize-SrcBitsKept),
828 return ReplaceInstUsesWith(CI, Res);
830 // We need to emit an AND to clear the high bits.
831 Constant *C = ConstantInt::get(Res->getType(),
832 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
833 return BinaryOperator::CreateAnd(Res, C);
836 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
837 // types and if the sizes are just right we can convert this into a logical
838 // 'and' which will be much cheaper than the pair of casts.
839 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
840 // TODO: Subsume this into EvaluateInDifferentType.
842 // Get the sizes of the types involved. We know that the intermediate type
843 // will be smaller than A or C, but don't know the relation between A and C.
844 Value *A = CSrc->getOperand(0);
845 unsigned SrcSize = A->getType()->getScalarSizeInBits();
846 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
847 unsigned DstSize = CI.getType()->getScalarSizeInBits();
848 // If we're actually extending zero bits, then if
849 // SrcSize < DstSize: zext(a & mask)
850 // SrcSize == DstSize: a & mask
851 // SrcSize > DstSize: trunc(a) & mask
852 if (SrcSize < DstSize) {
853 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
854 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
855 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
856 return new ZExtInst(And, CI.getType());
859 if (SrcSize == DstSize) {
860 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
861 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
864 if (SrcSize > DstSize) {
865 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
866 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
867 return BinaryOperator::CreateAnd(Trunc,
868 ConstantInt::get(Trunc->getType(),
873 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
874 return transformZExtICmp(ICI, CI);
876 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
877 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
878 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
879 // of the (zext icmp) will be transformed.
880 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
881 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
882 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
883 (transformZExtICmp(LHS, CI, false) ||
884 transformZExtICmp(RHS, CI, false))) {
885 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
886 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
887 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
891 // zext(trunc(X) & C) -> (X & zext(C)).
895 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
896 X->getType() == CI.getType())
897 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
899 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
901 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
902 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
903 X->getType() == CI.getType()) {
904 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
905 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
908 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
909 if (SrcI && SrcI->hasOneUse() &&
910 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
911 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
912 Value *New = Builder->CreateZExt(X, CI.getType());
913 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
919 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
920 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
921 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
922 ICmpInst::Predicate Pred = ICI->getPredicate();
924 // Don't bother if Op1 isn't of vector or integer type.
925 if (!Op1->getType()->isIntOrIntVectorTy())
928 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
929 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
930 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
931 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
932 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
934 Value *Sh = ConstantInt::get(Op0->getType(),
935 Op0->getType()->getScalarSizeInBits()-1);
936 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
937 if (In->getType() != CI.getType())
938 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
940 if (Pred == ICmpInst::ICMP_SGT)
941 In = Builder->CreateNot(In, In->getName()+".not");
942 return ReplaceInstUsesWith(CI, In);
946 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
947 // If we know that only one bit of the LHS of the icmp can be set and we
948 // have an equality comparison with zero or a power of 2, we can transform
949 // the icmp and sext into bitwise/integer operations.
950 if (ICI->hasOneUse() &&
951 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
952 unsigned BitWidth = Op1C->getType()->getBitWidth();
953 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
954 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
956 APInt KnownZeroMask(~KnownZero);
957 if (KnownZeroMask.isPowerOf2()) {
958 Value *In = ICI->getOperand(0);
960 // If the icmp tests for a known zero bit we can constant fold it.
961 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
962 Value *V = Pred == ICmpInst::ICMP_NE ?
963 ConstantInt::getAllOnesValue(CI.getType()) :
964 ConstantInt::getNullValue(CI.getType());
965 return ReplaceInstUsesWith(CI, V);
968 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
969 // sext ((x & 2^n) == 0) -> (x >> n) - 1
970 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
971 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
972 // Perform a right shift to place the desired bit in the LSB.
974 In = Builder->CreateLShr(In,
975 ConstantInt::get(In->getType(), ShiftAmt));
977 // At this point "In" is either 1 or 0. Subtract 1 to turn
978 // {1, 0} -> {0, -1}.
979 In = Builder->CreateAdd(In,
980 ConstantInt::getAllOnesValue(In->getType()),
983 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
984 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
985 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
986 // Perform a left shift to place the desired bit in the MSB.
988 In = Builder->CreateShl(In,
989 ConstantInt::get(In->getType(), ShiftAmt));
991 // Distribute the bit over the whole bit width.
992 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
993 BitWidth - 1), "sext");
996 if (CI.getType() == In->getType())
997 return ReplaceInstUsesWith(CI, In);
998 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
1006 /// Return true if we can take the specified value and return it as type Ty
1007 /// without inserting any new casts and without changing the value of the common
1008 /// low bits. This is used by code that tries to promote integer operations to
1009 /// a wider types will allow us to eliminate the extension.
1011 /// This function works on both vectors and scalars.
1013 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1014 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1015 "Can't sign extend type to a smaller type");
1016 // If this is a constant, it can be trivially promoted.
1017 if (isa<Constant>(V))
1020 Instruction *I = dyn_cast<Instruction>(V);
1021 if (!I) return false;
1023 // If this is a truncate from the dest type, we can trivially eliminate it.
1024 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1027 // We can't extend or shrink something that has multiple uses: doing so would
1028 // require duplicating the instruction in general, which isn't profitable.
1029 if (!I->hasOneUse()) return false;
1031 switch (I->getOpcode()) {
1032 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1033 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1034 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1036 case Instruction::And:
1037 case Instruction::Or:
1038 case Instruction::Xor:
1039 case Instruction::Add:
1040 case Instruction::Sub:
1041 case Instruction::Mul:
1042 // These operators can all arbitrarily be extended if their inputs can.
1043 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1044 canEvaluateSExtd(I->getOperand(1), Ty);
1046 //case Instruction::Shl: TODO
1047 //case Instruction::LShr: TODO
1049 case Instruction::Select:
1050 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1051 canEvaluateSExtd(I->getOperand(2), Ty);
1053 case Instruction::PHI: {
1054 // We can change a phi if we can change all operands. Note that we never
1055 // get into trouble with cyclic PHIs here because we only consider
1056 // instructions with a single use.
1057 PHINode *PN = cast<PHINode>(I);
1058 for (Value *IncValue : PN->incoming_values())
1059 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1063 // TODO: Can handle more cases here.
1070 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1071 // If this sign extend is only used by a truncate, let the truncate be
1072 // eliminated before we try to optimize this sext.
1073 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1076 if (Instruction *I = commonCastTransforms(CI))
1079 // See if we can simplify any instructions used by the input whose sole
1080 // purpose is to compute bits we don't care about.
1081 if (SimplifyDemandedInstructionBits(CI))
1084 Value *Src = CI.getOperand(0);
1085 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1087 // If we know that the value being extended is positive, we can use a zext
1089 bool KnownZero, KnownOne;
1090 ComputeSignBit(Src, KnownZero, KnownOne, 0, &CI);
1092 Value *ZExt = Builder->CreateZExt(Src, DestTy);
1093 return ReplaceInstUsesWith(CI, ZExt);
1096 // Attempt to extend the entire input expression tree to the destination
1097 // type. Only do this if the dest type is a simple type, don't convert the
1098 // expression tree to something weird like i93 unless the source is also
1100 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1101 canEvaluateSExtd(Src, DestTy)) {
1102 // Okay, we can transform this! Insert the new expression now.
1103 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1104 " to avoid sign extend: " << CI);
1105 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1106 assert(Res->getType() == DestTy);
1108 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1109 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1111 // If the high bits are already filled with sign bit, just replace this
1112 // cast with the result.
1113 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1114 return ReplaceInstUsesWith(CI, Res);
1116 // We need to emit a shl + ashr to do the sign extend.
1117 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1118 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1122 // If this input is a trunc from our destination, then turn sext(trunc(x))
1124 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1125 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1126 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1127 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1129 // We need to emit a shl + ashr to do the sign extend.
1130 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1131 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1132 return BinaryOperator::CreateAShr(Res, ShAmt);
1135 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1136 return transformSExtICmp(ICI, CI);
1138 // If the input is a shl/ashr pair of a same constant, then this is a sign
1139 // extension from a smaller value. If we could trust arbitrary bitwidth
1140 // integers, we could turn this into a truncate to the smaller bit and then
1141 // use a sext for the whole extension. Since we don't, look deeper and check
1142 // for a truncate. If the source and dest are the same type, eliminate the
1143 // trunc and extend and just do shifts. For example, turn:
1144 // %a = trunc i32 %i to i8
1145 // %b = shl i8 %a, 6
1146 // %c = ashr i8 %b, 6
1147 // %d = sext i8 %c to i32
1149 // %a = shl i32 %i, 30
1150 // %d = ashr i32 %a, 30
1152 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1153 ConstantInt *BA = nullptr, *CA = nullptr;
1154 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1155 m_ConstantInt(CA))) &&
1156 BA == CA && A->getType() == CI.getType()) {
1157 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1158 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1159 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1160 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1161 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1162 return BinaryOperator::CreateAShr(A, ShAmtV);
1169 /// Return a Constant* for the specified floating-point constant if it fits
1170 /// in the specified FP type without changing its value.
1171 static Constant *fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1173 APFloat F = CFP->getValueAPF();
1174 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1176 return ConstantFP::get(CFP->getContext(), F);
1180 /// If this is a floating-point extension instruction, look
1181 /// through it until we get the source value.
1182 static Value *lookThroughFPExtensions(Value *V) {
1183 if (Instruction *I = dyn_cast<Instruction>(V))
1184 if (I->getOpcode() == Instruction::FPExt)
1185 return lookThroughFPExtensions(I->getOperand(0));
1187 // If this value is a constant, return the constant in the smallest FP type
1188 // that can accurately represent it. This allows us to turn
1189 // (float)((double)X+2.0) into x+2.0f.
1190 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1191 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1192 return V; // No constant folding of this.
1193 // See if the value can be truncated to half and then reextended.
1194 if (Value *V = fitsInFPType(CFP, APFloat::IEEEhalf))
1196 // See if the value can be truncated to float and then reextended.
1197 if (Value *V = fitsInFPType(CFP, APFloat::IEEEsingle))
1199 if (CFP->getType()->isDoubleTy())
1200 return V; // Won't shrink.
1201 if (Value *V = fitsInFPType(CFP, APFloat::IEEEdouble))
1203 // Don't try to shrink to various long double types.
1209 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1210 if (Instruction *I = commonCastTransforms(CI))
1212 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1213 // simpilify this expression to avoid one or more of the trunc/extend
1214 // operations if we can do so without changing the numerical results.
1216 // The exact manner in which the widths of the operands interact to limit
1217 // what we can and cannot do safely varies from operation to operation, and
1218 // is explained below in the various case statements.
1219 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1220 if (OpI && OpI->hasOneUse()) {
1221 Value *LHSOrig = lookThroughFPExtensions(OpI->getOperand(0));
1222 Value *RHSOrig = lookThroughFPExtensions(OpI->getOperand(1));
1223 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1224 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1225 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1226 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1227 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1228 switch (OpI->getOpcode()) {
1230 case Instruction::FAdd:
1231 case Instruction::FSub:
1232 // For addition and subtraction, the infinitely precise result can
1233 // essentially be arbitrarily wide; proving that double rounding
1234 // will not occur because the result of OpI is exact (as we will for
1235 // FMul, for example) is hopeless. However, we *can* nonetheless
1236 // frequently know that double rounding cannot occur (or that it is
1237 // innocuous) by taking advantage of the specific structure of
1238 // infinitely-precise results that admit double rounding.
1240 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1241 // to represent both sources, we can guarantee that the double
1242 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1243 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1244 // for proof of this fact).
1246 // Note: Figueroa does not consider the case where DstFormat !=
1247 // SrcFormat. It's possible (likely even!) that this analysis
1248 // could be tightened for those cases, but they are rare (the main
1249 // case of interest here is (float)((double)float + float)).
1250 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1251 if (LHSOrig->getType() != CI.getType())
1252 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1253 if (RHSOrig->getType() != CI.getType())
1254 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1256 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1257 RI->copyFastMathFlags(OpI);
1261 case Instruction::FMul:
1262 // For multiplication, the infinitely precise result has at most
1263 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1264 // that such a value can be exactly represented, then no double
1265 // rounding can possibly occur; we can safely perform the operation
1266 // in the destination format if it can represent both sources.
1267 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1268 if (LHSOrig->getType() != CI.getType())
1269 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1270 if (RHSOrig->getType() != CI.getType())
1271 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1273 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1274 RI->copyFastMathFlags(OpI);
1278 case Instruction::FDiv:
1279 // For division, we use again use the bound from Figueroa's
1280 // dissertation. I am entirely certain that this bound can be
1281 // tightened in the unbalanced operand case by an analysis based on
1282 // the diophantine rational approximation bound, but the well-known
1283 // condition used here is a good conservative first pass.
1284 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1285 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1286 if (LHSOrig->getType() != CI.getType())
1287 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1288 if (RHSOrig->getType() != CI.getType())
1289 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1291 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1292 RI->copyFastMathFlags(OpI);
1296 case Instruction::FRem:
1297 // Remainder is straightforward. Remainder is always exact, so the
1298 // type of OpI doesn't enter into things at all. We simply evaluate
1299 // in whichever source type is larger, then convert to the
1300 // destination type.
1301 if (SrcWidth == OpWidth)
1303 if (LHSWidth < SrcWidth)
1304 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1305 else if (RHSWidth <= SrcWidth)
1306 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1307 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1308 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1309 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1310 RI->copyFastMathFlags(OpI);
1311 return CastInst::CreateFPCast(ExactResult, CI.getType());
1315 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1316 if (BinaryOperator::isFNeg(OpI)) {
1317 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1319 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1320 RI->copyFastMathFlags(OpI);
1325 // (fptrunc (select cond, R1, Cst)) -->
1326 // (select cond, (fptrunc R1), (fptrunc Cst))
1328 // - but only if this isn't part of a min/max operation, else we'll
1329 // ruin min/max canonical form which is to have the select and
1330 // compare's operands be of the same type with no casts to look through.
1332 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1334 (isa<ConstantFP>(SI->getOperand(1)) ||
1335 isa<ConstantFP>(SI->getOperand(2))) &&
1336 matchSelectPattern(SI, LHS, RHS).Flavor == SPF_UNKNOWN) {
1337 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1339 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1341 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1344 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1346 switch (II->getIntrinsicID()) {
1348 case Intrinsic::fabs: {
1349 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1350 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1352 Type *IntrinsicType[] = { CI.getType() };
1353 Function *Overload =
1354 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1355 II->getIntrinsicID(), IntrinsicType);
1357 Value *Args[] = { InnerTrunc };
1358 return CallInst::Create(Overload, Args, II->getName());
1366 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1367 return commonCastTransforms(CI);
1370 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1371 // This is safe if the intermediate type has enough bits in its mantissa to
1372 // accurately represent all values of X. For example, this won't work with
1373 // i64 -> float -> i64.
1374 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1375 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1377 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1379 Value *SrcI = OpI->getOperand(0);
1380 Type *FITy = FI.getType();
1381 Type *OpITy = OpI->getType();
1382 Type *SrcTy = SrcI->getType();
1383 bool IsInputSigned = isa<SIToFPInst>(OpI);
1384 bool IsOutputSigned = isa<FPToSIInst>(FI);
1386 // We can safely assume the conversion won't overflow the output range,
1387 // because (for example) (uint8_t)18293.f is undefined behavior.
1389 // Since we can assume the conversion won't overflow, our decision as to
1390 // whether the input will fit in the float should depend on the minimum
1391 // of the input range and output range.
1393 // This means this is also safe for a signed input and unsigned output, since
1394 // a negative input would lead to undefined behavior.
1395 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1396 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1397 int ActualSize = std::min(InputSize, OutputSize);
1399 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1400 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1401 if (IsInputSigned && IsOutputSigned)
1402 return new SExtInst(SrcI, FITy);
1403 return new ZExtInst(SrcI, FITy);
1405 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1406 return new TruncInst(SrcI, FITy);
1408 return ReplaceInstUsesWith(FI, SrcI);
1409 return new BitCastInst(SrcI, FITy);
1414 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1415 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1417 return commonCastTransforms(FI);
1419 if (Instruction *I = FoldItoFPtoI(FI))
1422 return commonCastTransforms(FI);
1425 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1426 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1428 return commonCastTransforms(FI);
1430 if (Instruction *I = FoldItoFPtoI(FI))
1433 return commonCastTransforms(FI);
1436 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1437 return commonCastTransforms(CI);
1440 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1441 return commonCastTransforms(CI);
1444 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1445 // If the source integer type is not the intptr_t type for this target, do a
1446 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1447 // cast to be exposed to other transforms.
1448 unsigned AS = CI.getAddressSpace();
1449 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1450 DL.getPointerSizeInBits(AS)) {
1451 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1452 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1453 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1455 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1456 return new IntToPtrInst(P, CI.getType());
1459 if (Instruction *I = commonCastTransforms(CI))
1465 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1466 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1467 Value *Src = CI.getOperand(0);
1469 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1470 // If casting the result of a getelementptr instruction with no offset, turn
1471 // this into a cast of the original pointer!
1472 if (GEP->hasAllZeroIndices() &&
1473 // If CI is an addrspacecast and GEP changes the poiner type, merging
1474 // GEP into CI would undo canonicalizing addrspacecast with different
1475 // pointer types, causing infinite loops.
1476 (!isa<AddrSpaceCastInst>(CI) ||
1477 GEP->getType() == GEP->getPointerOperand()->getType())) {
1478 // Changing the cast operand is usually not a good idea but it is safe
1479 // here because the pointer operand is being replaced with another
1480 // pointer operand so the opcode doesn't need to change.
1482 CI.setOperand(0, GEP->getOperand(0));
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.
1495 Type *Ty = CI.getType();
1496 unsigned AS = CI.getPointerAddressSpace();
1498 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1499 return commonPointerCastTransforms(CI);
1501 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1502 if (Ty->isVectorTy()) // Handle vectors of pointers.
1503 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1505 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1506 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1509 /// This input value (which is known to have vector type) is being zero extended
1510 /// or truncated to the specified vector type.
1511 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
1513 /// The source and destination vector types may have different element types.
1514 static Instruction *optimizeVectorResize(Value *InVal, VectorType *DestTy,
1516 // We can only do this optimization if the output is a multiple of the input
1517 // element size, or the input is a multiple of the output element size.
1518 // Convert the input type to have the same element type as the output.
1519 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1521 if (SrcTy->getElementType() != DestTy->getElementType()) {
1522 // The input types don't need to be identical, but for now they must be the
1523 // same size. There is no specific reason we couldn't handle things like
1524 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1526 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1527 DestTy->getElementType()->getPrimitiveSizeInBits())
1530 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1531 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1534 // Now that the element types match, get the shuffle mask and RHS of the
1535 // shuffle to use, which depends on whether we're increasing or decreasing the
1536 // size of the input.
1537 SmallVector<uint32_t, 16> ShuffleMask;
1540 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1541 // If we're shrinking the number of elements, just shuffle in the low
1542 // elements from the input and use undef as the second shuffle input.
1543 V2 = UndefValue::get(SrcTy);
1544 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1545 ShuffleMask.push_back(i);
1548 // If we're increasing the number of elements, shuffle in all of the
1549 // elements from InVal and fill the rest of the result elements with zeros
1550 // from a constant zero.
1551 V2 = Constant::getNullValue(SrcTy);
1552 unsigned SrcElts = SrcTy->getNumElements();
1553 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1554 ShuffleMask.push_back(i);
1556 // The excess elements reference the first element of the zero input.
1557 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1558 ShuffleMask.push_back(SrcElts);
1561 return new ShuffleVectorInst(InVal, V2,
1562 ConstantDataVector::get(V2->getContext(),
1566 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1567 return Value % Ty->getPrimitiveSizeInBits() == 0;
1570 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1571 return Value / Ty->getPrimitiveSizeInBits();
1574 /// V is a value which is inserted into a vector of VecEltTy.
1575 /// Look through the value to see if we can decompose it into
1576 /// insertions into the vector. See the example in the comment for
1577 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1578 /// The type of V is always a non-zero multiple of VecEltTy's size.
1579 /// Shift is the number of bits between the lsb of V and the lsb of
1582 /// This returns false if the pattern can't be matched or true if it can,
1583 /// filling in Elements with the elements found here.
1584 static bool collectInsertionElements(Value *V, unsigned Shift,
1585 SmallVectorImpl<Value *> &Elements,
1586 Type *VecEltTy, bool isBigEndian) {
1587 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1588 "Shift should be a multiple of the element type size");
1590 // Undef values never contribute useful bits to the result.
1591 if (isa<UndefValue>(V)) return true;
1593 // If we got down to a value of the right type, we win, try inserting into the
1595 if (V->getType() == VecEltTy) {
1596 // Inserting null doesn't actually insert any elements.
1597 if (Constant *C = dyn_cast<Constant>(V))
1598 if (C->isNullValue())
1601 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1603 ElementIndex = Elements.size() - ElementIndex - 1;
1605 // Fail if multiple elements are inserted into this slot.
1606 if (Elements[ElementIndex])
1609 Elements[ElementIndex] = V;
1613 if (Constant *C = dyn_cast<Constant>(V)) {
1614 // Figure out the # elements this provides, and bitcast it or slice it up
1616 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1618 // If the constant is the size of a vector element, we just need to bitcast
1619 // it to the right type so it gets properly inserted.
1621 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1622 Shift, Elements, VecEltTy, isBigEndian);
1624 // Okay, this is a constant that covers multiple elements. Slice it up into
1625 // pieces and insert each element-sized piece into the vector.
1626 if (!isa<IntegerType>(C->getType()))
1627 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1628 C->getType()->getPrimitiveSizeInBits()));
1629 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1630 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1632 for (unsigned i = 0; i != NumElts; ++i) {
1633 unsigned ShiftI = Shift+i*ElementSize;
1634 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1636 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1637 if (!collectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1644 if (!V->hasOneUse()) return false;
1646 Instruction *I = dyn_cast<Instruction>(V);
1647 if (!I) return false;
1648 switch (I->getOpcode()) {
1649 default: return false; // Unhandled case.
1650 case Instruction::BitCast:
1651 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1653 case Instruction::ZExt:
1654 if (!isMultipleOfTypeSize(
1655 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1658 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1660 case Instruction::Or:
1661 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1663 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1665 case Instruction::Shl: {
1666 // Must be shifting by a constant that is a multiple of the element size.
1667 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1668 if (!CI) return false;
1669 Shift += CI->getZExtValue();
1670 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1671 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1679 /// If the input is an 'or' instruction, we may be doing shifts and ors to
1680 /// assemble the elements of the vector manually.
1681 /// Try to rip the code out and replace it with insertelements. This is to
1682 /// optimize code like this:
1684 /// %tmp37 = bitcast float %inc to i32
1685 /// %tmp38 = zext i32 %tmp37 to i64
1686 /// %tmp31 = bitcast float %inc5 to i32
1687 /// %tmp32 = zext i32 %tmp31 to i64
1688 /// %tmp33 = shl i64 %tmp32, 32
1689 /// %ins35 = or i64 %tmp33, %tmp38
1690 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1692 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1693 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
1695 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1696 Value *IntInput = CI.getOperand(0);
1698 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1699 if (!collectInsertionElements(IntInput, 0, Elements,
1700 DestVecTy->getElementType(),
1701 IC.getDataLayout().isBigEndian()))
1704 // If we succeeded, we know that all of the element are specified by Elements
1705 // or are zero if Elements has a null entry. Recast this as a set of
1707 Value *Result = Constant::getNullValue(CI.getType());
1708 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1709 if (!Elements[i]) continue; // Unset element.
1711 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1712 IC.Builder->getInt32(i));
1719 /// See if we can optimize an integer->float/double bitcast.
1720 /// The various long double bitcasts can't get in here.
1721 static Instruction *optimizeIntToFloatBitCast(BitCastInst &CI, InstCombiner &IC,
1722 const DataLayout &DL) {
1723 Value *Src = CI.getOperand(0);
1724 Type *DestTy = CI.getType();
1726 // If this is a bitcast from int to float, check to see if the int is an
1727 // extraction from a vector.
1728 Value *VecInput = nullptr;
1729 // bitcast(trunc(bitcast(somevector)))
1730 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1731 isa<VectorType>(VecInput->getType())) {
1732 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1733 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1735 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1736 // If the element type of the vector doesn't match the result type,
1737 // bitcast it to be a vector type we can extract from.
1738 if (VecTy->getElementType() != DestTy) {
1739 VecTy = VectorType::get(DestTy,
1740 VecTy->getPrimitiveSizeInBits() / DestWidth);
1741 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1745 if (DL.isBigEndian())
1746 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1747 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1751 // bitcast(trunc(lshr(bitcast(somevector), cst))
1752 ConstantInt *ShAmt = nullptr;
1753 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1754 m_ConstantInt(ShAmt)))) &&
1755 isa<VectorType>(VecInput->getType())) {
1756 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1757 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1758 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1759 ShAmt->getZExtValue() % DestWidth == 0) {
1760 // If the element type of the vector doesn't match the result type,
1761 // bitcast it to be a vector type we can extract from.
1762 if (VecTy->getElementType() != DestTy) {
1763 VecTy = VectorType::get(DestTy,
1764 VecTy->getPrimitiveSizeInBits() / DestWidth);
1765 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1768 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1769 if (DL.isBigEndian())
1770 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1771 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1777 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1778 // If the operands are integer typed then apply the integer transforms,
1779 // otherwise just apply the common ones.
1780 Value *Src = CI.getOperand(0);
1781 Type *SrcTy = Src->getType();
1782 Type *DestTy = CI.getType();
1784 // Get rid of casts from one type to the same type. These are useless and can
1785 // be replaced by the operand.
1786 if (DestTy == Src->getType())
1787 return ReplaceInstUsesWith(CI, Src);
1789 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1790 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1791 Type *DstElTy = DstPTy->getElementType();
1792 Type *SrcElTy = SrcPTy->getElementType();
1794 // If we are casting a alloca to a pointer to a type of the same
1795 // size, rewrite the allocation instruction to allocate the "right" type.
1796 // There is no need to modify malloc calls because it is their bitcast that
1797 // needs to be cleaned up.
1798 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1799 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1802 // If the source and destination are pointers, and this cast is equivalent
1803 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1804 // This can enhance SROA and other transforms that want type-safe pointers.
1805 unsigned NumZeros = 0;
1806 while (SrcElTy != DstElTy &&
1807 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1808 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1809 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
1813 // If we found a path from the src to dest, create the getelementptr now.
1814 if (SrcElTy == DstElTy) {
1815 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0));
1816 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1820 // Try to optimize int -> float bitcasts.
1821 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1822 if (Instruction *I = optimizeIntToFloatBitCast(CI, *this, DL))
1825 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1826 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1827 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1828 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1829 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1830 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1833 if (isa<IntegerType>(SrcTy)) {
1834 // If this is a cast from an integer to vector, check to see if the input
1835 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1836 // the casts with a shuffle and (potentially) a bitcast.
1837 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1838 CastInst *SrcCast = cast<CastInst>(Src);
1839 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1840 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1841 if (Instruction *I = optimizeVectorResize(BCIn->getOperand(0),
1842 cast<VectorType>(DestTy), *this))
1846 // If the input is an 'or' instruction, we may be doing shifts and ors to
1847 // assemble the elements of the vector manually. Try to rip the code out
1848 // and replace it with insertelements.
1849 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
1850 return ReplaceInstUsesWith(CI, V);
1854 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1855 if (SrcVTy->getNumElements() == 1) {
1856 // If our destination is not a vector, then make this a straight
1857 // scalar-scalar cast.
1858 if (!DestTy->isVectorTy()) {
1860 Builder->CreateExtractElement(Src,
1861 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1862 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1865 // Otherwise, see if our source is an insert. If so, then use the scalar
1866 // component directly.
1867 if (InsertElementInst *IEI =
1868 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1869 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1874 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1875 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1876 // a bitcast to a vector with the same # elts.
1877 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1878 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1879 SVI->getType()->getNumElements() ==
1880 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1882 // If either of the operands is a cast from CI.getType(), then
1883 // evaluating the shuffle in the casted destination's type will allow
1884 // us to eliminate at least one cast.
1885 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1886 Tmp->getOperand(0)->getType() == DestTy) ||
1887 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1888 Tmp->getOperand(0)->getType() == DestTy)) {
1889 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1890 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1891 // Return a new shuffle vector. Use the same element ID's, as we
1892 // know the vector types match #elts.
1893 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1898 if (SrcTy->isPointerTy())
1899 return commonPointerCastTransforms(CI);
1900 return commonCastTransforms(CI);
1903 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1904 // If the destination pointer element type is not the same as the source's
1905 // first do a bitcast to the destination type, and then the addrspacecast.
1906 // This allows the cast to be exposed to other transforms.
1907 Value *Src = CI.getOperand(0);
1908 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
1909 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
1911 Type *DestElemTy = DestTy->getElementType();
1912 if (SrcTy->getElementType() != DestElemTy) {
1913 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
1914 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
1915 // Handle vectors of pointers.
1916 MidTy = VectorType::get(MidTy, VT->getNumElements());
1919 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
1920 return new AddrSpaceCastInst(NewBitCast, CI.getType());
1923 return commonPointerCastTransforms(CI);