1 //===- InstCombineCasts.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for cast operations.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/PatternMatch.h"
18 #include "llvm/Target/TargetLibraryInfo.h"
20 using namespace PatternMatch;
22 #define DEBUG_TYPE "instcombine"
24 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
25 /// expression. 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 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
80 /// try to eliminate the cast by moving the type information into the alloc.
81 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
83 // This requires DataLayout to get the alloca alignment and size information.
84 if (!DL) return nullptr;
86 PointerType *PTy = cast<PointerType>(CI.getType());
88 BuilderTy AllocaBuilder(*Builder);
89 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
91 // Get the type really allocated and the type casted to.
92 Type *AllocElTy = AI.getAllocatedType();
93 Type *CastElTy = PTy->getElementType();
94 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
96 unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy);
97 unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy);
98 if (CastElTyAlign < AllocElTyAlign) return nullptr;
100 // If the allocation has multiple uses, only promote it if we are strictly
101 // increasing the alignment of the resultant allocation. If we keep it the
102 // same, we open the door to infinite loops of various kinds.
103 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
105 uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy);
106 uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy);
107 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
109 // If the allocation has multiple uses, only promote it if we're not
110 // shrinking the amount of memory being allocated.
111 uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy);
112 uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy);
113 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
115 // See if we can satisfy the modulus by pulling a scale out of the array
117 unsigned ArraySizeScale;
118 uint64_t ArrayOffset;
119 Value *NumElements = // See if the array size is a decomposable linear expr.
120 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
122 // If we can now satisfy the modulus, by using a non-1 scale, we really can
124 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
125 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
127 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
128 Value *Amt = nullptr;
132 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
133 // Insert before the alloca, not before the cast.
134 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
137 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
138 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
140 Amt = AllocaBuilder.CreateAdd(Amt, Off);
143 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
144 New->setAlignment(AI.getAlignment());
146 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
148 // If the allocation has multiple real uses, insert a cast and change all
149 // things that used it to use the new cast. This will also hack on CI, but it
151 if (!AI.hasOneUse()) {
152 // New is the allocation instruction, pointer typed. AI is the original
153 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
154 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
155 ReplaceInstUsesWith(AI, NewCast);
157 return ReplaceInstUsesWith(CI, New);
160 /// EvaluateInDifferentType - Given an expression that
161 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
162 /// insert the code to evaluate the expression.
163 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
165 if (Constant *C = dyn_cast<Constant>(V)) {
166 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
167 // If we got a constantexpr back, try to simplify it with DL info.
168 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
169 C = ConstantFoldConstantExpression(CE, DL, TLI);
173 // Otherwise, it must be an instruction.
174 Instruction *I = cast<Instruction>(V);
175 Instruction *Res = nullptr;
176 unsigned Opc = I->getOpcode();
178 case Instruction::Add:
179 case Instruction::Sub:
180 case Instruction::Mul:
181 case Instruction::And:
182 case Instruction::Or:
183 case Instruction::Xor:
184 case Instruction::AShr:
185 case Instruction::LShr:
186 case Instruction::Shl:
187 case Instruction::UDiv:
188 case Instruction::URem: {
189 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
190 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
191 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
194 case Instruction::Trunc:
195 case Instruction::ZExt:
196 case Instruction::SExt:
197 // If the source type of the cast is the type we're trying for then we can
198 // just return the source. There's no need to insert it because it is not
200 if (I->getOperand(0)->getType() == Ty)
201 return I->getOperand(0);
203 // Otherwise, must be the same type of cast, so just reinsert a new one.
204 // This also handles the case of zext(trunc(x)) -> zext(x).
205 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
206 Opc == Instruction::SExt);
208 case Instruction::Select: {
209 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
210 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
211 Res = SelectInst::Create(I->getOperand(0), True, False);
214 case Instruction::PHI: {
215 PHINode *OPN = cast<PHINode>(I);
216 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
217 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
218 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
219 NPN->addIncoming(V, OPN->getIncomingBlock(i));
225 // TODO: Can handle more cases here.
226 llvm_unreachable("Unreachable!");
230 return InsertNewInstWith(Res, *I);
234 /// This function is a wrapper around CastInst::isEliminableCastPair. It
235 /// simply extracts arguments and returns what that function returns.
236 static Instruction::CastOps
237 isEliminableCastPair(
238 const CastInst *CI, ///< The first cast instruction
239 unsigned opcode, ///< The opcode of the second cast instruction
240 Type *DstTy, ///< The target type for the second cast instruction
241 const DataLayout *DL ///< The target data for pointer size
244 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
245 Type *MidTy = CI->getType(); // B from above
247 // Get the opcodes of the two Cast instructions
248 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
249 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
250 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ?
251 DL->getIntPtrType(SrcTy) : nullptr;
252 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ?
253 DL->getIntPtrType(MidTy) : nullptr;
254 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ?
255 DL->getIntPtrType(DstTy) : nullptr;
256 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
257 DstTy, SrcIntPtrTy, MidIntPtrTy,
260 // We don't want to form an inttoptr or ptrtoint that converts to an integer
261 // type that differs from the pointer size.
262 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
263 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
266 return Instruction::CastOps(Res);
269 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
270 /// results in any code being generated and is interesting to optimize out. If
271 /// the cast can be eliminated by some other simple transformation, we prefer
272 /// to do the simplification first.
273 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
275 // Noop casts and casts of constants should be eliminated trivially.
276 if (V->getType() == Ty || isa<Constant>(V)) return false;
278 // If this is another cast that can be eliminated, we prefer to have it
280 if (const CastInst *CI = dyn_cast<CastInst>(V))
281 if (isEliminableCastPair(CI, opc, Ty, DL))
284 // If this is a vector sext from a compare, then we don't want to break the
285 // idiom where each element of the extended vector is either zero or all ones.
286 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
293 /// @brief Implement the transforms common to all CastInst visitors.
294 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
295 Value *Src = CI.getOperand(0);
297 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
299 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
300 if (Instruction::CastOps opc =
301 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
302 // The first cast (CSrc) is eliminable so we need to fix up or replace
303 // the second cast (CI). CSrc will then have a good chance of being dead.
304 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
308 // If we are casting a select then fold the cast into the select
309 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
310 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
313 // If we are casting a PHI then fold the cast into the PHI
314 if (isa<PHINode>(Src)) {
315 // We don't do this if this would create a PHI node with an illegal type if
316 // it is currently legal.
317 if (!Src->getType()->isIntegerTy() ||
318 !CI.getType()->isIntegerTy() ||
319 ShouldChangeType(CI.getType(), Src->getType()))
320 if (Instruction *NV = FoldOpIntoPhi(CI))
327 /// CanEvaluateTruncated - Return true if we can evaluate the specified
328 /// expression tree as type Ty instead of its larger type, and arrive with the
329 /// same value. This is used by code that tries to eliminate truncates.
331 /// Ty will always be a type smaller than V. We should return true if trunc(V)
332 /// can be computed by computing V in the smaller type. If V is an instruction,
333 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
334 /// makes sense if x and y can be efficiently truncated.
336 /// This function works on both vectors and scalars.
338 static bool CanEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
340 // We can always evaluate constants in another type.
341 if (isa<Constant>(V))
344 Instruction *I = dyn_cast<Instruction>(V);
345 if (!I) return false;
347 Type *OrigTy = V->getType();
349 // If this is an extension from the dest type, we can eliminate it, even if it
350 // has multiple uses.
351 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
352 I->getOperand(0)->getType() == Ty)
355 // We can't extend or shrink something that has multiple uses: doing so would
356 // require duplicating the instruction in general, which isn't profitable.
357 if (!I->hasOneUse()) return false;
359 unsigned Opc = I->getOpcode();
361 case Instruction::Add:
362 case Instruction::Sub:
363 case Instruction::Mul:
364 case Instruction::And:
365 case Instruction::Or:
366 case Instruction::Xor:
367 // These operators can all arbitrarily be extended or truncated.
368 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
369 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
371 case Instruction::UDiv:
372 case Instruction::URem: {
373 // UDiv and URem can be truncated if all the truncated bits are zero.
374 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
375 uint32_t BitWidth = Ty->getScalarSizeInBits();
376 if (BitWidth < OrigBitWidth) {
377 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
378 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
379 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
380 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
381 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
386 case Instruction::Shl:
387 // If we are truncating the result of this SHL, and if it's a shift of a
388 // constant amount, we can always perform a SHL in a smaller type.
389 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
390 uint32_t BitWidth = Ty->getScalarSizeInBits();
391 if (CI->getLimitedValue(BitWidth) < BitWidth)
392 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
395 case Instruction::LShr:
396 // If this is a truncate of a logical shr, we can truncate it to a smaller
397 // lshr iff we know that the bits we would otherwise be shifting in are
399 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
400 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
401 uint32_t BitWidth = Ty->getScalarSizeInBits();
402 if (IC.MaskedValueIsZero(I->getOperand(0),
403 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
404 CI->getLimitedValue(BitWidth) < BitWidth) {
405 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
409 case Instruction::Trunc:
410 // trunc(trunc(x)) -> trunc(x)
412 case Instruction::ZExt:
413 case Instruction::SExt:
414 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
415 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
417 case Instruction::Select: {
418 SelectInst *SI = cast<SelectInst>(I);
419 return CanEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
420 CanEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
422 case Instruction::PHI: {
423 // We can change a phi if we can change all operands. Note that we never
424 // get into trouble with cyclic PHIs here because we only consider
425 // instructions with a single use.
426 PHINode *PN = cast<PHINode>(I);
427 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
428 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty, IC, CxtI))
433 // TODO: Can handle more cases here.
440 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
441 if (Instruction *Result = commonCastTransforms(CI))
444 // See if we can simplify any instructions used by the input whose sole
445 // purpose is to compute bits we don't care about.
446 if (SimplifyDemandedInstructionBits(CI))
449 Value *Src = CI.getOperand(0);
450 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
452 // Attempt to truncate the entire input expression tree to the destination
453 // type. Only do this if the dest type is a simple type, don't convert the
454 // expression tree to something weird like i93 unless the source is also
456 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
457 CanEvaluateTruncated(Src, DestTy, *this, &CI)) {
459 // If this cast is a truncate, evaluting in a different type always
460 // eliminates the cast, so it is always a win.
461 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
462 " to avoid cast: " << CI << '\n');
463 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
464 assert(Res->getType() == DestTy);
465 return ReplaceInstUsesWith(CI, Res);
468 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
469 if (DestTy->getScalarSizeInBits() == 1) {
470 Constant *One = ConstantInt::get(Src->getType(), 1);
471 Src = Builder->CreateAnd(Src, One);
472 Value *Zero = Constant::getNullValue(Src->getType());
473 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
476 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
477 Value *A = nullptr; ConstantInt *Cst = nullptr;
478 if (Src->hasOneUse() &&
479 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
480 // We have three types to worry about here, the type of A, the source of
481 // the truncate (MidSize), and the destination of the truncate. We know that
482 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
483 // between ASize and ResultSize.
484 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
486 // If the shift amount is larger than the size of A, then the result is
487 // known to be zero because all the input bits got shifted out.
488 if (Cst->getZExtValue() >= ASize)
489 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
491 // Since we're doing an lshr and a zero extend, and know that the shift
492 // amount is smaller than ASize, it is always safe to do the shift in A's
493 // type, then zero extend or truncate to the result.
494 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
495 Shift->takeName(Src);
496 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
499 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
500 // type isn't non-native.
501 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
502 ShouldChangeType(Src->getType(), CI.getType()) &&
503 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
504 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
505 return BinaryOperator::CreateAnd(NewTrunc,
506 ConstantExpr::getTrunc(Cst, CI.getType()));
512 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
513 /// in order to eliminate the icmp.
514 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
516 // If we are just checking for a icmp eq of a single bit and zext'ing it
517 // to an integer, then shift the bit to the appropriate place and then
518 // cast to integer to avoid the comparison.
519 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
520 const APInt &Op1CV = Op1C->getValue();
522 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
523 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
524 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
525 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
526 if (!DoXform) return ICI;
528 Value *In = ICI->getOperand(0);
529 Value *Sh = ConstantInt::get(In->getType(),
530 In->getType()->getScalarSizeInBits()-1);
531 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
532 if (In->getType() != CI.getType())
533 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
535 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
536 Constant *One = ConstantInt::get(In->getType(), 1);
537 In = Builder->CreateXor(In, One, In->getName()+".not");
540 return ReplaceInstUsesWith(CI, In);
543 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
544 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
545 // zext (X == 1) to i32 --> X iff X has only the low bit set.
546 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
547 // zext (X != 0) to i32 --> X iff X has only the low bit set.
548 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
549 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
550 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
551 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
552 // This only works for EQ and NE
554 // If Op1C some other power of two, convert:
555 uint32_t BitWidth = Op1C->getType()->getBitWidth();
556 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
557 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
559 APInt KnownZeroMask(~KnownZero);
560 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
561 if (!DoXform) return ICI;
563 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
564 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
565 // (X&4) == 2 --> false
566 // (X&4) != 2 --> true
567 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
569 Res = ConstantExpr::getZExt(Res, CI.getType());
570 return ReplaceInstUsesWith(CI, Res);
573 uint32_t ShiftAmt = KnownZeroMask.logBase2();
574 Value *In = ICI->getOperand(0);
576 // Perform a logical shr by shiftamt.
577 // Insert the shift to put the result in the low bit.
578 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
579 In->getName()+".lobit");
582 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
583 Constant *One = ConstantInt::get(In->getType(), 1);
584 In = Builder->CreateXor(In, One);
587 if (CI.getType() == In->getType())
588 return ReplaceInstUsesWith(CI, In);
589 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
594 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
595 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
596 // may lead to additional simplifications.
597 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
598 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
599 uint32_t BitWidth = ITy->getBitWidth();
600 Value *LHS = ICI->getOperand(0);
601 Value *RHS = ICI->getOperand(1);
603 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
604 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
605 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
606 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
608 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
609 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
610 APInt UnknownBit = ~KnownBits;
611 if (UnknownBit.countPopulation() == 1) {
612 if (!DoXform) return ICI;
614 Value *Result = Builder->CreateXor(LHS, RHS);
616 // Mask off any bits that are set and won't be shifted away.
617 if (KnownOneLHS.uge(UnknownBit))
618 Result = Builder->CreateAnd(Result,
619 ConstantInt::get(ITy, UnknownBit));
621 // Shift the bit we're testing down to the lsb.
622 Result = Builder->CreateLShr(
623 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
625 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
626 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
627 Result->takeName(ICI);
628 return ReplaceInstUsesWith(CI, Result);
637 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
638 /// specified wider type and produce the same low bits. If not, return false.
640 /// If this function returns true, it can also return a non-zero number of bits
641 /// (in BitsToClear) which indicates that the value it computes is correct for
642 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
643 /// out. For example, to promote something like:
645 /// %B = trunc i64 %A to i32
646 /// %C = lshr i32 %B, 8
647 /// %E = zext i32 %C to i64
649 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
650 /// set to 8 to indicate that the promoted value needs to have bits 24-31
651 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
652 /// clear the top bits anyway, doing this has no extra cost.
654 /// This function works on both vectors and scalars.
655 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
656 InstCombiner &IC, Instruction *CxtI) {
658 if (isa<Constant>(V))
661 Instruction *I = dyn_cast<Instruction>(V);
662 if (!I) return false;
664 // If the input is a truncate from the destination type, we can trivially
666 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
669 // We can't extend or shrink something that has multiple uses: doing so would
670 // require duplicating the instruction in general, which isn't profitable.
671 if (!I->hasOneUse()) return false;
673 unsigned Opc = I->getOpcode(), Tmp;
675 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
676 case Instruction::SExt: // zext(sext(x)) -> sext(x).
677 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
679 case Instruction::And:
680 case Instruction::Or:
681 case Instruction::Xor:
682 case Instruction::Add:
683 case Instruction::Sub:
684 case Instruction::Mul:
685 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
686 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
688 // These can all be promoted if neither operand has 'bits to clear'.
689 if (BitsToClear == 0 && Tmp == 0)
692 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
693 // other side, BitsToClear is ok.
695 (Opc == Instruction::And || Opc == Instruction::Or ||
696 Opc == Instruction::Xor)) {
697 // We use MaskedValueIsZero here for generality, but the case we care
698 // about the most is constant RHS.
699 unsigned VSize = V->getType()->getScalarSizeInBits();
700 if (IC.MaskedValueIsZero(I->getOperand(1),
701 APInt::getHighBitsSet(VSize, BitsToClear),
706 // Otherwise, we don't know how to analyze this BitsToClear case yet.
709 case Instruction::Shl:
710 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
711 // upper bits we can reduce BitsToClear by the shift amount.
712 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
713 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
715 uint64_t ShiftAmt = Amt->getZExtValue();
716 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
720 case Instruction::LShr:
721 // We can promote lshr(x, cst) if we can promote x. This requires the
722 // ultimate 'and' to clear out the high zero bits we're clearing out though.
723 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
724 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
726 BitsToClear += Amt->getZExtValue();
727 if (BitsToClear > V->getType()->getScalarSizeInBits())
728 BitsToClear = V->getType()->getScalarSizeInBits();
731 // Cannot promote variable LSHR.
733 case Instruction::Select:
734 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
735 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
736 // TODO: If important, we could handle the case when the BitsToClear are
737 // known zero in the disagreeing side.
742 case Instruction::PHI: {
743 // We can change a phi if we can change all operands. Note that we never
744 // get into trouble with cyclic PHIs here because we only consider
745 // instructions with a single use.
746 PHINode *PN = cast<PHINode>(I);
747 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
749 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
750 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
751 // TODO: If important, we could handle the case when the BitsToClear
752 // are known zero in the disagreeing input.
758 // TODO: Can handle more cases here.
763 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
764 // If this zero extend is only used by a truncate, let the truncate be
765 // eliminated before we try to optimize this zext.
766 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
769 // If one of the common conversion will work, do it.
770 if (Instruction *Result = commonCastTransforms(CI))
773 // See if we can simplify any instructions used by the input whose sole
774 // purpose is to compute bits we don't care about.
775 if (SimplifyDemandedInstructionBits(CI))
778 Value *Src = CI.getOperand(0);
779 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
781 // Attempt to extend the entire input expression tree to the destination
782 // type. Only do this if the dest type is a simple type, don't convert the
783 // expression tree to something weird like i93 unless the source is also
785 unsigned BitsToClear;
786 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
787 CanEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
788 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
789 "Unreasonable BitsToClear");
791 // Okay, we can transform this! Insert the new expression now.
792 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
793 " to avoid zero extend: " << CI);
794 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
795 assert(Res->getType() == DestTy);
797 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
798 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
800 // If the high bits are already filled with zeros, just replace this
801 // cast with the result.
802 if (MaskedValueIsZero(Res,
803 APInt::getHighBitsSet(DestBitSize,
804 DestBitSize-SrcBitsKept),
806 return ReplaceInstUsesWith(CI, Res);
808 // We need to emit an AND to clear the high bits.
809 Constant *C = ConstantInt::get(Res->getType(),
810 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
811 return BinaryOperator::CreateAnd(Res, C);
814 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
815 // types and if the sizes are just right we can convert this into a logical
816 // 'and' which will be much cheaper than the pair of casts.
817 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
818 // TODO: Subsume this into EvaluateInDifferentType.
820 // Get the sizes of the types involved. We know that the intermediate type
821 // will be smaller than A or C, but don't know the relation between A and C.
822 Value *A = CSrc->getOperand(0);
823 unsigned SrcSize = A->getType()->getScalarSizeInBits();
824 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
825 unsigned DstSize = CI.getType()->getScalarSizeInBits();
826 // If we're actually extending zero bits, then if
827 // SrcSize < DstSize: zext(a & mask)
828 // SrcSize == DstSize: a & mask
829 // SrcSize > DstSize: trunc(a) & mask
830 if (SrcSize < DstSize) {
831 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
832 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
833 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
834 return new ZExtInst(And, CI.getType());
837 if (SrcSize == DstSize) {
838 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
839 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
842 if (SrcSize > DstSize) {
843 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
844 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
845 return BinaryOperator::CreateAnd(Trunc,
846 ConstantInt::get(Trunc->getType(),
851 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
852 return transformZExtICmp(ICI, CI);
854 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
855 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
856 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
857 // of the (zext icmp) will be transformed.
858 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
859 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
860 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
861 (transformZExtICmp(LHS, CI, false) ||
862 transformZExtICmp(RHS, CI, false))) {
863 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
864 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
865 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
869 // zext(trunc(X) & C) -> (X & zext(C)).
873 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
874 X->getType() == CI.getType())
875 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
877 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
879 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
880 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
881 X->getType() == CI.getType()) {
882 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
883 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
886 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
887 if (SrcI && SrcI->hasOneUse() &&
888 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
889 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
890 Value *New = Builder->CreateZExt(X, CI.getType());
891 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
897 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
898 /// in order to eliminate the icmp.
899 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
900 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
901 ICmpInst::Predicate Pred = ICI->getPredicate();
903 // Don't bother if Op1 isn't of vector or integer type.
904 if (!Op1->getType()->isIntOrIntVectorTy())
907 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
908 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
909 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
910 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
911 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
913 Value *Sh = ConstantInt::get(Op0->getType(),
914 Op0->getType()->getScalarSizeInBits()-1);
915 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
916 if (In->getType() != CI.getType())
917 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
919 if (Pred == ICmpInst::ICMP_SGT)
920 In = Builder->CreateNot(In, In->getName()+".not");
921 return ReplaceInstUsesWith(CI, In);
925 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
926 // If we know that only one bit of the LHS of the icmp can be set and we
927 // have an equality comparison with zero or a power of 2, we can transform
928 // the icmp and sext into bitwise/integer operations.
929 if (ICI->hasOneUse() &&
930 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
931 unsigned BitWidth = Op1C->getType()->getBitWidth();
932 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
933 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
935 APInt KnownZeroMask(~KnownZero);
936 if (KnownZeroMask.isPowerOf2()) {
937 Value *In = ICI->getOperand(0);
939 // If the icmp tests for a known zero bit we can constant fold it.
940 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
941 Value *V = Pred == ICmpInst::ICMP_NE ?
942 ConstantInt::getAllOnesValue(CI.getType()) :
943 ConstantInt::getNullValue(CI.getType());
944 return ReplaceInstUsesWith(CI, V);
947 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
948 // sext ((x & 2^n) == 0) -> (x >> n) - 1
949 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
950 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
951 // Perform a right shift to place the desired bit in the LSB.
953 In = Builder->CreateLShr(In,
954 ConstantInt::get(In->getType(), ShiftAmt));
956 // At this point "In" is either 1 or 0. Subtract 1 to turn
957 // {1, 0} -> {0, -1}.
958 In = Builder->CreateAdd(In,
959 ConstantInt::getAllOnesValue(In->getType()),
962 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
963 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
964 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
965 // Perform a left shift to place the desired bit in the MSB.
967 In = Builder->CreateShl(In,
968 ConstantInt::get(In->getType(), ShiftAmt));
970 // Distribute the bit over the whole bit width.
971 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
972 BitWidth - 1), "sext");
975 if (CI.getType() == In->getType())
976 return ReplaceInstUsesWith(CI, In);
977 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
985 /// CanEvaluateSExtd - Return true if we can take the specified value
986 /// and return it as type Ty without inserting any new casts and without
987 /// changing the value of the common low bits. This is used by code that tries
988 /// to promote integer operations to a wider types will allow us to eliminate
991 /// This function works on both vectors and scalars.
993 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
994 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
995 "Can't sign extend type to a smaller type");
996 // If this is a constant, it can be trivially promoted.
997 if (isa<Constant>(V))
1000 Instruction *I = dyn_cast<Instruction>(V);
1001 if (!I) return false;
1003 // If this is a truncate from the dest type, we can trivially eliminate it.
1004 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1007 // We can't extend or shrink something that has multiple uses: doing so would
1008 // require duplicating the instruction in general, which isn't profitable.
1009 if (!I->hasOneUse()) return false;
1011 switch (I->getOpcode()) {
1012 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1013 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1014 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1016 case Instruction::And:
1017 case Instruction::Or:
1018 case Instruction::Xor:
1019 case Instruction::Add:
1020 case Instruction::Sub:
1021 case Instruction::Mul:
1022 // These operators can all arbitrarily be extended if their inputs can.
1023 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1024 CanEvaluateSExtd(I->getOperand(1), Ty);
1026 //case Instruction::Shl: TODO
1027 //case Instruction::LShr: TODO
1029 case Instruction::Select:
1030 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1031 CanEvaluateSExtd(I->getOperand(2), Ty);
1033 case Instruction::PHI: {
1034 // We can change a phi if we can change all operands. Note that we never
1035 // get into trouble with cyclic PHIs here because we only consider
1036 // instructions with a single use.
1037 PHINode *PN = cast<PHINode>(I);
1038 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1039 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1043 // TODO: Can handle more cases here.
1050 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1051 // If this sign extend is only used by a truncate, let the truncate be
1052 // eliminated before we try to optimize this sext.
1053 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1056 if (Instruction *I = commonCastTransforms(CI))
1059 // See if we can simplify any instructions used by the input whose sole
1060 // purpose is to compute bits we don't care about.
1061 if (SimplifyDemandedInstructionBits(CI))
1064 Value *Src = CI.getOperand(0);
1065 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1067 // Attempt to extend the entire input expression tree to the destination
1068 // type. Only do this if the dest type is a simple type, don't convert the
1069 // expression tree to something weird like i93 unless the source is also
1071 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1072 CanEvaluateSExtd(Src, DestTy)) {
1073 // Okay, we can transform this! Insert the new expression now.
1074 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1075 " to avoid sign extend: " << CI);
1076 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1077 assert(Res->getType() == DestTy);
1079 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1080 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1082 // If the high bits are already filled with sign bit, just replace this
1083 // cast with the result.
1084 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1085 return ReplaceInstUsesWith(CI, Res);
1087 // We need to emit a shl + ashr to do the sign extend.
1088 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1089 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1093 // If this input is a trunc from our destination, then turn sext(trunc(x))
1095 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1096 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1097 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1098 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1100 // We need to emit a shl + ashr to do the sign extend.
1101 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1102 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1103 return BinaryOperator::CreateAShr(Res, ShAmt);
1106 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1107 return transformSExtICmp(ICI, CI);
1109 // If the input is a shl/ashr pair of a same constant, then this is a sign
1110 // extension from a smaller value. If we could trust arbitrary bitwidth
1111 // integers, we could turn this into a truncate to the smaller bit and then
1112 // use a sext for the whole extension. Since we don't, look deeper and check
1113 // for a truncate. If the source and dest are the same type, eliminate the
1114 // trunc and extend and just do shifts. For example, turn:
1115 // %a = trunc i32 %i to i8
1116 // %b = shl i8 %a, 6
1117 // %c = ashr i8 %b, 6
1118 // %d = sext i8 %c to i32
1120 // %a = shl i32 %i, 30
1121 // %d = ashr i32 %a, 30
1123 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1124 ConstantInt *BA = nullptr, *CA = nullptr;
1125 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1126 m_ConstantInt(CA))) &&
1127 BA == CA && A->getType() == CI.getType()) {
1128 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1129 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1130 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1131 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1132 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1133 return BinaryOperator::CreateAShr(A, ShAmtV);
1140 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1141 /// in the specified FP type without changing its value.
1142 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1144 APFloat F = CFP->getValueAPF();
1145 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1147 return ConstantFP::get(CFP->getContext(), F);
1151 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1152 /// through it until we get the source value.
1153 static Value *LookThroughFPExtensions(Value *V) {
1154 if (Instruction *I = dyn_cast<Instruction>(V))
1155 if (I->getOpcode() == Instruction::FPExt)
1156 return LookThroughFPExtensions(I->getOperand(0));
1158 // If this value is a constant, return the constant in the smallest FP type
1159 // that can accurately represent it. This allows us to turn
1160 // (float)((double)X+2.0) into x+2.0f.
1161 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1162 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1163 return V; // No constant folding of this.
1164 // See if the value can be truncated to half and then reextended.
1165 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1167 // See if the value can be truncated to float and then reextended.
1168 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1170 if (CFP->getType()->isDoubleTy())
1171 return V; // Won't shrink.
1172 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1174 // Don't try to shrink to various long double types.
1180 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1181 if (Instruction *I = commonCastTransforms(CI))
1183 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1184 // simpilify this expression to avoid one or more of the trunc/extend
1185 // operations if we can do so without changing the numerical results.
1187 // The exact manner in which the widths of the operands interact to limit
1188 // what we can and cannot do safely varies from operation to operation, and
1189 // is explained below in the various case statements.
1190 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1191 if (OpI && OpI->hasOneUse()) {
1192 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1193 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1194 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1195 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1196 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1197 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1198 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1199 switch (OpI->getOpcode()) {
1201 case Instruction::FAdd:
1202 case Instruction::FSub:
1203 // For addition and subtraction, the infinitely precise result can
1204 // essentially be arbitrarily wide; proving that double rounding
1205 // will not occur because the result of OpI is exact (as we will for
1206 // FMul, for example) is hopeless. However, we *can* nonetheless
1207 // frequently know that double rounding cannot occur (or that it is
1208 // innocuous) by taking advantage of the specific structure of
1209 // infinitely-precise results that admit double rounding.
1211 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1212 // to represent both sources, we can guarantee that the double
1213 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1214 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1215 // for proof of this fact).
1217 // Note: Figueroa does not consider the case where DstFormat !=
1218 // SrcFormat. It's possible (likely even!) that this analysis
1219 // could be tightened for those cases, but they are rare (the main
1220 // case of interest here is (float)((double)float + float)).
1221 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1222 if (LHSOrig->getType() != CI.getType())
1223 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1224 if (RHSOrig->getType() != CI.getType())
1225 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1227 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1228 RI->copyFastMathFlags(OpI);
1232 case Instruction::FMul:
1233 // For multiplication, the infinitely precise result has at most
1234 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1235 // that such a value can be exactly represented, then no double
1236 // rounding can possibly occur; we can safely perform the operation
1237 // in the destination format if it can represent both sources.
1238 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1239 if (LHSOrig->getType() != CI.getType())
1240 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1241 if (RHSOrig->getType() != CI.getType())
1242 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1244 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1245 RI->copyFastMathFlags(OpI);
1249 case Instruction::FDiv:
1250 // For division, we use again use the bound from Figueroa's
1251 // dissertation. I am entirely certain that this bound can be
1252 // tightened in the unbalanced operand case by an analysis based on
1253 // the diophantine rational approximation bound, but the well-known
1254 // condition used here is a good conservative first pass.
1255 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1256 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1257 if (LHSOrig->getType() != CI.getType())
1258 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1259 if (RHSOrig->getType() != CI.getType())
1260 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1262 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1263 RI->copyFastMathFlags(OpI);
1267 case Instruction::FRem:
1268 // Remainder is straightforward. Remainder is always exact, so the
1269 // type of OpI doesn't enter into things at all. We simply evaluate
1270 // in whichever source type is larger, then convert to the
1271 // destination type.
1272 Value *NewLHS = LHSOrig, *NewRHS = RHSOrig;
1273 if (LHSWidth < SrcWidth)
1274 NewLHS = Builder->CreateFPExt(NewLHS, RHSOrig->getType());
1275 else if (RHSWidth <= SrcWidth)
1276 NewRHS = Builder->CreateFPExt(NewRHS, LHSOrig->getType());
1277 if (NewLHS != LHSOrig || NewRHS != RHSOrig) {
1278 Value *ExactResult = Builder->CreateFRem(NewLHS, NewRHS);
1279 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1280 RI->copyFastMathFlags(OpI);
1281 return CastInst::CreateFPCast(ExactResult, CI.getType());
1285 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1286 if (BinaryOperator::isFNeg(OpI)) {
1287 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1289 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1290 RI->copyFastMathFlags(OpI);
1295 // (fptrunc (select cond, R1, Cst)) -->
1296 // (select cond, (fptrunc R1), (fptrunc Cst))
1297 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1299 (isa<ConstantFP>(SI->getOperand(1)) ||
1300 isa<ConstantFP>(SI->getOperand(2)))) {
1301 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1303 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1305 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1308 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1310 switch (II->getIntrinsicID()) {
1312 case Intrinsic::fabs: {
1313 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1314 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1316 Type *IntrinsicType[] = { CI.getType() };
1317 Function *Overload =
1318 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1319 II->getIntrinsicID(), IntrinsicType);
1321 Value *Args[] = { InnerTrunc };
1322 return CallInst::Create(Overload, Args, II->getName());
1330 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1331 return commonCastTransforms(CI);
1334 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1335 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1337 return commonCastTransforms(FI);
1339 // fptoui(uitofp(X)) --> X
1340 // fptoui(sitofp(X)) --> X
1341 // This is safe if the intermediate type has enough bits in its mantissa to
1342 // accurately represent all values of X. For example, do not do this with
1343 // i64->float->i64. This is also safe for sitofp case, because any negative
1344 // 'X' value would cause an undefined result for the fptoui.
1345 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1346 OpI->getOperand(0)->getType() == FI.getType() &&
1347 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1348 OpI->getType()->getFPMantissaWidth())
1349 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1351 return commonCastTransforms(FI);
1354 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1355 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1357 return commonCastTransforms(FI);
1359 // fptosi(sitofp(X)) --> X
1360 // fptosi(uitofp(X)) --> X
1361 // This is safe if the intermediate type has enough bits in its mantissa to
1362 // accurately represent all values of X. For example, do not do this with
1363 // i64->float->i64. This is also safe for sitofp case, because any negative
1364 // 'X' value would cause an undefined result for the fptoui.
1365 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1366 OpI->getOperand(0)->getType() == FI.getType() &&
1367 (int)FI.getType()->getScalarSizeInBits() <=
1368 OpI->getType()->getFPMantissaWidth())
1369 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1371 return commonCastTransforms(FI);
1374 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1375 return commonCastTransforms(CI);
1378 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1379 return commonCastTransforms(CI);
1382 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1383 // If the source integer type is not the intptr_t type for this target, do a
1384 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1385 // cast to be exposed to other transforms.
1388 unsigned AS = CI.getAddressSpace();
1389 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1390 DL->getPointerSizeInBits(AS)) {
1391 Type *Ty = DL->getIntPtrType(CI.getContext(), AS);
1392 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1393 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1395 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1396 return new IntToPtrInst(P, CI.getType());
1400 if (Instruction *I = commonCastTransforms(CI))
1406 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1407 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1408 Value *Src = CI.getOperand(0);
1410 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1411 // If casting the result of a getelementptr instruction with no offset, turn
1412 // this into a cast of the original pointer!
1413 if (GEP->hasAllZeroIndices() &&
1414 // If CI is an addrspacecast and GEP changes the poiner type, merging
1415 // GEP into CI would undo canonicalizing addrspacecast with different
1416 // pointer types, causing infinite loops.
1417 (!isa<AddrSpaceCastInst>(CI) ||
1418 GEP->getType() == GEP->getPointerOperand()->getType())) {
1419 // Changing the cast operand is usually not a good idea but it is safe
1420 // here because the pointer operand is being replaced with another
1421 // pointer operand so the opcode doesn't need to change.
1423 CI.setOperand(0, GEP->getOperand(0));
1428 return commonCastTransforms(CI);
1430 // If the GEP has a single use, and the base pointer is a bitcast, and the
1431 // GEP computes a constant offset, see if we can convert these three
1432 // instructions into fewer. This typically happens with unions and other
1433 // non-type-safe code.
1434 unsigned AS = GEP->getPointerAddressSpace();
1435 unsigned OffsetBits = DL->getPointerSizeInBits(AS);
1436 APInt Offset(OffsetBits, 0);
1437 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0));
1438 if (GEP->hasOneUse() &&
1440 GEP->accumulateConstantOffset(*DL, Offset)) {
1441 // Get the base pointer input of the bitcast, and the type it points to.
1442 Value *OrigBase = BCI->getOperand(0);
1443 SmallVector<Value*, 8> NewIndices;
1444 if (FindElementAtOffset(OrigBase->getType(),
1445 Offset.getSExtValue(),
1447 // If we were able to index down into an element, create the GEP
1448 // and bitcast the result. This eliminates one bitcast, potentially
1450 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1451 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1452 Builder->CreateGEP(OrigBase, NewIndices);
1453 NGEP->takeName(GEP);
1455 if (isa<BitCastInst>(CI))
1456 return new BitCastInst(NGEP, CI.getType());
1457 assert(isa<PtrToIntInst>(CI));
1458 return new PtrToIntInst(NGEP, CI.getType());
1463 return commonCastTransforms(CI);
1466 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1467 // If the destination integer type is not the intptr_t type for this target,
1468 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1469 // to be exposed to other transforms.
1472 return commonPointerCastTransforms(CI);
1474 Type *Ty = CI.getType();
1475 unsigned AS = CI.getPointerAddressSpace();
1477 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS))
1478 return commonPointerCastTransforms(CI);
1480 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS);
1481 if (Ty->isVectorTy()) // Handle vectors of pointers.
1482 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1484 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1485 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1488 /// OptimizeVectorResize - This input value (which is known to have vector type)
1489 /// is being zero extended or truncated to the specified vector type. Try to
1490 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1492 /// The source and destination vector types may have different element types.
1493 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1495 // We can only do this optimization if the output is a multiple of the input
1496 // element size, or the input is a multiple of the output element size.
1497 // Convert the input type to have the same element type as the output.
1498 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1500 if (SrcTy->getElementType() != DestTy->getElementType()) {
1501 // The input types don't need to be identical, but for now they must be the
1502 // same size. There is no specific reason we couldn't handle things like
1503 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1505 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1506 DestTy->getElementType()->getPrimitiveSizeInBits())
1509 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1510 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1513 // Now that the element types match, get the shuffle mask and RHS of the
1514 // shuffle to use, which depends on whether we're increasing or decreasing the
1515 // size of the input.
1516 SmallVector<uint32_t, 16> ShuffleMask;
1519 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1520 // If we're shrinking the number of elements, just shuffle in the low
1521 // elements from the input and use undef as the second shuffle input.
1522 V2 = UndefValue::get(SrcTy);
1523 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1524 ShuffleMask.push_back(i);
1527 // If we're increasing the number of elements, shuffle in all of the
1528 // elements from InVal and fill the rest of the result elements with zeros
1529 // from a constant zero.
1530 V2 = Constant::getNullValue(SrcTy);
1531 unsigned SrcElts = SrcTy->getNumElements();
1532 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1533 ShuffleMask.push_back(i);
1535 // The excess elements reference the first element of the zero input.
1536 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1537 ShuffleMask.push_back(SrcElts);
1540 return new ShuffleVectorInst(InVal, V2,
1541 ConstantDataVector::get(V2->getContext(),
1545 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1546 return Value % Ty->getPrimitiveSizeInBits() == 0;
1549 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1550 return Value / Ty->getPrimitiveSizeInBits();
1553 /// CollectInsertionElements - V is a value which is inserted into a vector of
1554 /// VecEltTy. Look through the value to see if we can decompose it into
1555 /// insertions into the vector. See the example in the comment for
1556 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1557 /// The type of V is always a non-zero multiple of VecEltTy's size.
1558 /// Shift is the number of bits between the lsb of V and the lsb of
1561 /// This returns false if the pattern can't be matched or true if it can,
1562 /// filling in Elements with the elements found here.
1563 static bool CollectInsertionElements(Value *V, unsigned Shift,
1564 SmallVectorImpl<Value*> &Elements,
1565 Type *VecEltTy, InstCombiner &IC) {
1566 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1567 "Shift should be a multiple of the element type size");
1569 // Undef values never contribute useful bits to the result.
1570 if (isa<UndefValue>(V)) return true;
1572 // If we got down to a value of the right type, we win, try inserting into the
1574 if (V->getType() == VecEltTy) {
1575 // Inserting null doesn't actually insert any elements.
1576 if (Constant *C = dyn_cast<Constant>(V))
1577 if (C->isNullValue())
1580 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1581 if (IC.getDataLayout()->isBigEndian())
1582 ElementIndex = Elements.size() - ElementIndex - 1;
1584 // Fail if multiple elements are inserted into this slot.
1585 if (Elements[ElementIndex])
1588 Elements[ElementIndex] = V;
1592 if (Constant *C = dyn_cast<Constant>(V)) {
1593 // Figure out the # elements this provides, and bitcast it or slice it up
1595 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1597 // If the constant is the size of a vector element, we just need to bitcast
1598 // it to the right type so it gets properly inserted.
1600 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1601 Shift, Elements, VecEltTy, IC);
1603 // Okay, this is a constant that covers multiple elements. Slice it up into
1604 // pieces and insert each element-sized piece into the vector.
1605 if (!isa<IntegerType>(C->getType()))
1606 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1607 C->getType()->getPrimitiveSizeInBits()));
1608 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1609 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1611 for (unsigned i = 0; i != NumElts; ++i) {
1612 unsigned ShiftI = Shift+i*ElementSize;
1613 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1615 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1616 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC))
1622 if (!V->hasOneUse()) return false;
1624 Instruction *I = dyn_cast<Instruction>(V);
1625 if (!I) return false;
1626 switch (I->getOpcode()) {
1627 default: return false; // Unhandled case.
1628 case Instruction::BitCast:
1629 return CollectInsertionElements(I->getOperand(0), Shift,
1630 Elements, VecEltTy, IC);
1631 case Instruction::ZExt:
1632 if (!isMultipleOfTypeSize(
1633 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1636 return CollectInsertionElements(I->getOperand(0), Shift,
1637 Elements, VecEltTy, IC);
1638 case Instruction::Or:
1639 return CollectInsertionElements(I->getOperand(0), Shift,
1640 Elements, VecEltTy, IC) &&
1641 CollectInsertionElements(I->getOperand(1), Shift,
1642 Elements, VecEltTy, IC);
1643 case Instruction::Shl: {
1644 // Must be shifting by a constant that is a multiple of the element size.
1645 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1646 if (!CI) return false;
1647 Shift += CI->getZExtValue();
1648 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1649 return CollectInsertionElements(I->getOperand(0), Shift,
1650 Elements, VecEltTy, IC);
1657 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1658 /// may be doing shifts and ors to assemble the elements of the vector manually.
1659 /// Try to rip the code out and replace it with insertelements. This is to
1660 /// optimize code like this:
1662 /// %tmp37 = bitcast float %inc to i32
1663 /// %tmp38 = zext i32 %tmp37 to i64
1664 /// %tmp31 = bitcast float %inc5 to i32
1665 /// %tmp32 = zext i32 %tmp31 to i64
1666 /// %tmp33 = shl i64 %tmp32, 32
1667 /// %ins35 = or i64 %tmp33, %tmp38
1668 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1670 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1671 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1673 // We need to know the target byte order to perform this optimization.
1674 if (!IC.getDataLayout()) return nullptr;
1676 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1677 Value *IntInput = CI.getOperand(0);
1679 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1680 if (!CollectInsertionElements(IntInput, 0, Elements,
1681 DestVecTy->getElementType(), IC))
1684 // If we succeeded, we know that all of the element are specified by Elements
1685 // or are zero if Elements has a null entry. Recast this as a set of
1687 Value *Result = Constant::getNullValue(CI.getType());
1688 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1689 if (!Elements[i]) continue; // Unset element.
1691 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1692 IC.Builder->getInt32(i));
1699 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1700 /// bitcast. The various long double bitcasts can't get in here.
1701 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1702 // We need to know the target byte order to perform this optimization.
1703 if (!IC.getDataLayout()) return nullptr;
1705 Value *Src = CI.getOperand(0);
1706 Type *DestTy = CI.getType();
1708 // If this is a bitcast from int to float, check to see if the int is an
1709 // extraction from a vector.
1710 Value *VecInput = nullptr;
1711 // bitcast(trunc(bitcast(somevector)))
1712 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1713 isa<VectorType>(VecInput->getType())) {
1714 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1715 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1717 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1718 // If the element type of the vector doesn't match the result type,
1719 // bitcast it to be a vector type we can extract from.
1720 if (VecTy->getElementType() != DestTy) {
1721 VecTy = VectorType::get(DestTy,
1722 VecTy->getPrimitiveSizeInBits() / DestWidth);
1723 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1727 if (IC.getDataLayout()->isBigEndian())
1728 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1729 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1733 // bitcast(trunc(lshr(bitcast(somevector), cst))
1734 ConstantInt *ShAmt = nullptr;
1735 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1736 m_ConstantInt(ShAmt)))) &&
1737 isa<VectorType>(VecInput->getType())) {
1738 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1739 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1740 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1741 ShAmt->getZExtValue() % DestWidth == 0) {
1742 // If the element type of the vector doesn't match the result type,
1743 // bitcast it to be a vector type we can extract from.
1744 if (VecTy->getElementType() != DestTy) {
1745 VecTy = VectorType::get(DestTy,
1746 VecTy->getPrimitiveSizeInBits() / DestWidth);
1747 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1750 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1751 if (IC.getDataLayout()->isBigEndian())
1752 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1753 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1759 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1760 // If the operands are integer typed then apply the integer transforms,
1761 // otherwise just apply the common ones.
1762 Value *Src = CI.getOperand(0);
1763 Type *SrcTy = Src->getType();
1764 Type *DestTy = CI.getType();
1766 // Get rid of casts from one type to the same type. These are useless and can
1767 // be replaced by the operand.
1768 if (DestTy == Src->getType())
1769 return ReplaceInstUsesWith(CI, Src);
1771 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1772 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1773 Type *DstElTy = DstPTy->getElementType();
1774 Type *SrcElTy = SrcPTy->getElementType();
1776 // If we are casting a alloca to a pointer to a type of the same
1777 // size, rewrite the allocation instruction to allocate the "right" type.
1778 // There is no need to modify malloc calls because it is their bitcast that
1779 // needs to be cleaned up.
1780 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1781 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1784 // If the source and destination are pointers, and this cast is equivalent
1785 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1786 // This can enhance SROA and other transforms that want type-safe pointers.
1787 Constant *ZeroUInt =
1788 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1789 unsigned NumZeros = 0;
1790 while (SrcElTy != DstElTy &&
1791 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1792 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1793 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1797 // If we found a path from the src to dest, create the getelementptr now.
1798 if (SrcElTy == DstElTy) {
1799 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1800 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1804 // Try to optimize int -> float bitcasts.
1805 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1806 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1809 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1810 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1811 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1812 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1813 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1814 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1817 if (isa<IntegerType>(SrcTy)) {
1818 // If this is a cast from an integer to vector, check to see if the input
1819 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1820 // the casts with a shuffle and (potentially) a bitcast.
1821 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1822 CastInst *SrcCast = cast<CastInst>(Src);
1823 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1824 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1825 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1826 cast<VectorType>(DestTy), *this))
1830 // If the input is an 'or' instruction, we may be doing shifts and ors to
1831 // assemble the elements of the vector manually. Try to rip the code out
1832 // and replace it with insertelements.
1833 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1834 return ReplaceInstUsesWith(CI, V);
1838 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1839 if (SrcVTy->getNumElements() == 1) {
1840 // If our destination is not a vector, then make this a straight
1841 // scalar-scalar cast.
1842 if (!DestTy->isVectorTy()) {
1844 Builder->CreateExtractElement(Src,
1845 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1846 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1849 // Otherwise, see if our source is an insert. If so, then use the scalar
1850 // component directly.
1851 if (InsertElementInst *IEI =
1852 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1853 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1858 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1859 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1860 // a bitcast to a vector with the same # elts.
1861 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1862 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1863 SVI->getType()->getNumElements() ==
1864 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1866 // If either of the operands is a cast from CI.getType(), then
1867 // evaluating the shuffle in the casted destination's type will allow
1868 // us to eliminate at least one cast.
1869 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1870 Tmp->getOperand(0)->getType() == DestTy) ||
1871 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1872 Tmp->getOperand(0)->getType() == DestTy)) {
1873 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1874 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1875 // Return a new shuffle vector. Use the same element ID's, as we
1876 // know the vector types match #elts.
1877 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1882 if (SrcTy->isPointerTy())
1883 return commonPointerCastTransforms(CI);
1884 return commonCastTransforms(CI);
1887 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1888 // If the destination pointer element type is not the same as the source's
1889 // first do a bitcast to the destination type, and then the addrspacecast.
1890 // This allows the cast to be exposed to other transforms.
1891 Value *Src = CI.getOperand(0);
1892 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
1893 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
1895 Type *DestElemTy = DestTy->getElementType();
1896 if (SrcTy->getElementType() != DestElemTy) {
1897 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
1898 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
1899 // Handle vectors of pointers.
1900 MidTy = VectorType::get(MidTy, VT->getNumElements());
1903 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
1904 return new AddrSpaceCastInst(NewBitCast, CI.getType());
1907 return commonPointerCastTransforms(CI);