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());
147 // If the allocation has multiple real uses, insert a cast and change all
148 // things that used it to use the new cast. This will also hack on CI, but it
150 if (!AI.hasOneUse()) {
151 // New is the allocation instruction, pointer typed. AI is the original
152 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
153 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
154 ReplaceInstUsesWith(AI, NewCast);
156 return ReplaceInstUsesWith(CI, New);
159 /// EvaluateInDifferentType - Given an expression that
160 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
161 /// insert the code to evaluate the expression.
162 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
164 if (Constant *C = dyn_cast<Constant>(V)) {
165 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
166 // If we got a constantexpr back, try to simplify it with DL info.
167 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
168 C = ConstantFoldConstantExpression(CE, DL, TLI);
172 // Otherwise, it must be an instruction.
173 Instruction *I = cast<Instruction>(V);
174 Instruction *Res = nullptr;
175 unsigned Opc = I->getOpcode();
177 case Instruction::Add:
178 case Instruction::Sub:
179 case Instruction::Mul:
180 case Instruction::And:
181 case Instruction::Or:
182 case Instruction::Xor:
183 case Instruction::AShr:
184 case Instruction::LShr:
185 case Instruction::Shl:
186 case Instruction::UDiv:
187 case Instruction::URem: {
188 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
189 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
190 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
193 case Instruction::Trunc:
194 case Instruction::ZExt:
195 case Instruction::SExt:
196 // If the source type of the cast is the type we're trying for then we can
197 // just return the source. There's no need to insert it because it is not
199 if (I->getOperand(0)->getType() == Ty)
200 return I->getOperand(0);
202 // Otherwise, must be the same type of cast, so just reinsert a new one.
203 // This also handles the case of zext(trunc(x)) -> zext(x).
204 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
205 Opc == Instruction::SExt);
207 case Instruction::Select: {
208 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
209 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
210 Res = SelectInst::Create(I->getOperand(0), True, False);
213 case Instruction::PHI: {
214 PHINode *OPN = cast<PHINode>(I);
215 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
216 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
217 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
218 NPN->addIncoming(V, OPN->getIncomingBlock(i));
224 // TODO: Can handle more cases here.
225 llvm_unreachable("Unreachable!");
229 return InsertNewInstWith(Res, *I);
233 /// This function is a wrapper around CastInst::isEliminableCastPair. It
234 /// simply extracts arguments and returns what that function returns.
235 static Instruction::CastOps
236 isEliminableCastPair(
237 const CastInst *CI, ///< The first cast instruction
238 unsigned opcode, ///< The opcode of the second cast instruction
239 Type *DstTy, ///< The target type for the second cast instruction
240 const DataLayout *DL ///< The target data for pointer size
243 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
244 Type *MidTy = CI->getType(); // B from above
246 // Get the opcodes of the two Cast instructions
247 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
248 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
249 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ?
250 DL->getIntPtrType(SrcTy) : nullptr;
251 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ?
252 DL->getIntPtrType(MidTy) : nullptr;
253 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ?
254 DL->getIntPtrType(DstTy) : nullptr;
255 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
256 DstTy, SrcIntPtrTy, MidIntPtrTy,
259 // We don't want to form an inttoptr or ptrtoint that converts to an integer
260 // type that differs from the pointer size.
261 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
262 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
265 return Instruction::CastOps(Res);
268 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
269 /// results in any code being generated and is interesting to optimize out. If
270 /// the cast can be eliminated by some other simple transformation, we prefer
271 /// to do the simplification first.
272 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
274 // Noop casts and casts of constants should be eliminated trivially.
275 if (V->getType() == Ty || isa<Constant>(V)) return false;
277 // If this is another cast that can be eliminated, we prefer to have it
279 if (const CastInst *CI = dyn_cast<CastInst>(V))
280 if (isEliminableCastPair(CI, opc, Ty, DL))
283 // If this is a vector sext from a compare, then we don't want to break the
284 // idiom where each element of the extended vector is either zero or all ones.
285 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
292 /// @brief Implement the transforms common to all CastInst visitors.
293 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
294 Value *Src = CI.getOperand(0);
296 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
298 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
299 if (Instruction::CastOps opc =
300 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
301 // The first cast (CSrc) is eliminable so we need to fix up or replace
302 // the second cast (CI). CSrc will then have a good chance of being dead.
303 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
307 // If we are casting a select then fold the cast into the select
308 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
309 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
312 // If we are casting a PHI then fold the cast into the PHI
313 if (isa<PHINode>(Src)) {
314 // We don't do this if this would create a PHI node with an illegal type if
315 // it is currently legal.
316 if (!Src->getType()->isIntegerTy() ||
317 !CI.getType()->isIntegerTy() ||
318 ShouldChangeType(CI.getType(), Src->getType()))
319 if (Instruction *NV = FoldOpIntoPhi(CI))
326 /// CanEvaluateTruncated - Return true if we can evaluate the specified
327 /// expression tree as type Ty instead of its larger type, and arrive with the
328 /// same value. This is used by code that tries to eliminate truncates.
330 /// Ty will always be a type smaller than V. We should return true if trunc(V)
331 /// can be computed by computing V in the smaller type. If V is an instruction,
332 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
333 /// makes sense if x and y can be efficiently truncated.
335 /// This function works on both vectors and scalars.
337 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
338 // We can always evaluate constants in another type.
339 if (isa<Constant>(V))
342 Instruction *I = dyn_cast<Instruction>(V);
343 if (!I) return false;
345 Type *OrigTy = V->getType();
347 // If this is an extension from the dest type, we can eliminate it, even if it
348 // has multiple uses.
349 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
350 I->getOperand(0)->getType() == Ty)
353 // We can't extend or shrink something that has multiple uses: doing so would
354 // require duplicating the instruction in general, which isn't profitable.
355 if (!I->hasOneUse()) return false;
357 unsigned Opc = I->getOpcode();
359 case Instruction::Add:
360 case Instruction::Sub:
361 case Instruction::Mul:
362 case Instruction::And:
363 case Instruction::Or:
364 case Instruction::Xor:
365 // These operators can all arbitrarily be extended or truncated.
366 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
367 CanEvaluateTruncated(I->getOperand(1), Ty);
369 case Instruction::UDiv:
370 case Instruction::URem: {
371 // UDiv and URem can be truncated if all the truncated bits are zero.
372 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
373 uint32_t BitWidth = Ty->getScalarSizeInBits();
374 if (BitWidth < OrigBitWidth) {
375 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
376 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
377 MaskedValueIsZero(I->getOperand(1), Mask)) {
378 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
379 CanEvaluateTruncated(I->getOperand(1), Ty);
384 case Instruction::Shl:
385 // If we are truncating the result of this SHL, and if it's a shift of a
386 // constant amount, we can always perform a SHL in a smaller type.
387 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
388 uint32_t BitWidth = Ty->getScalarSizeInBits();
389 if (CI->getLimitedValue(BitWidth) < BitWidth)
390 return CanEvaluateTruncated(I->getOperand(0), Ty);
393 case Instruction::LShr:
394 // If this is a truncate of a logical shr, we can truncate it to a smaller
395 // lshr iff we know that the bits we would otherwise be shifting in are
397 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
398 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
399 uint32_t BitWidth = Ty->getScalarSizeInBits();
400 if (MaskedValueIsZero(I->getOperand(0),
401 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
402 CI->getLimitedValue(BitWidth) < BitWidth) {
403 return CanEvaluateTruncated(I->getOperand(0), Ty);
407 case Instruction::Trunc:
408 // trunc(trunc(x)) -> trunc(x)
410 case Instruction::ZExt:
411 case Instruction::SExt:
412 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
413 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
415 case Instruction::Select: {
416 SelectInst *SI = cast<SelectInst>(I);
417 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
418 CanEvaluateTruncated(SI->getFalseValue(), Ty);
420 case Instruction::PHI: {
421 // We can change a phi if we can change all operands. Note that we never
422 // get into trouble with cyclic PHIs here because we only consider
423 // instructions with a single use.
424 PHINode *PN = cast<PHINode>(I);
425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
426 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
431 // TODO: Can handle more cases here.
438 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
439 if (Instruction *Result = commonCastTransforms(CI))
442 // See if we can simplify any instructions used by the input whose sole
443 // purpose is to compute bits we don't care about.
444 if (SimplifyDemandedInstructionBits(CI))
447 Value *Src = CI.getOperand(0);
448 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
450 // Attempt to truncate the entire input expression tree to the destination
451 // type. Only do this if the dest type is a simple type, don't convert the
452 // expression tree to something weird like i93 unless the source is also
454 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
455 CanEvaluateTruncated(Src, DestTy)) {
457 // If this cast is a truncate, evaluting in a different type always
458 // eliminates the cast, so it is always a win.
459 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
460 " to avoid cast: " << CI << '\n');
461 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
462 assert(Res->getType() == DestTy);
463 return ReplaceInstUsesWith(CI, Res);
466 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
467 if (DestTy->getScalarSizeInBits() == 1) {
468 Constant *One = ConstantInt::get(Src->getType(), 1);
469 Src = Builder->CreateAnd(Src, One);
470 Value *Zero = Constant::getNullValue(Src->getType());
471 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
474 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
475 Value *A = nullptr; ConstantInt *Cst = nullptr;
476 if (Src->hasOneUse() &&
477 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
478 // We have three types to worry about here, the type of A, the source of
479 // the truncate (MidSize), and the destination of the truncate. We know that
480 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
481 // between ASize and ResultSize.
482 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
484 // If the shift amount is larger than the size of A, then the result is
485 // known to be zero because all the input bits got shifted out.
486 if (Cst->getZExtValue() >= ASize)
487 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
489 // Since we're doing an lshr and a zero extend, and know that the shift
490 // amount is smaller than ASize, it is always safe to do the shift in A's
491 // type, then zero extend or truncate to the result.
492 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
493 Shift->takeName(Src);
494 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
497 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
498 // type isn't non-native.
499 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
500 ShouldChangeType(Src->getType(), CI.getType()) &&
501 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
502 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
503 return BinaryOperator::CreateAnd(NewTrunc,
504 ConstantExpr::getTrunc(Cst, CI.getType()));
510 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
511 /// in order to eliminate the icmp.
512 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
514 // If we are just checking for a icmp eq of a single bit and zext'ing it
515 // to an integer, then shift the bit to the appropriate place and then
516 // cast to integer to avoid the comparison.
517 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
518 const APInt &Op1CV = Op1C->getValue();
520 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
521 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
522 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
523 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
524 if (!DoXform) return ICI;
526 Value *In = ICI->getOperand(0);
527 Value *Sh = ConstantInt::get(In->getType(),
528 In->getType()->getScalarSizeInBits()-1);
529 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
530 if (In->getType() != CI.getType())
531 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
533 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
534 Constant *One = ConstantInt::get(In->getType(), 1);
535 In = Builder->CreateXor(In, One, In->getName()+".not");
538 return ReplaceInstUsesWith(CI, In);
541 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
542 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
543 // zext (X == 1) to i32 --> X iff X has only the low bit set.
544 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
545 // zext (X != 0) to i32 --> X iff X has only the low bit set.
546 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
547 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
548 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
549 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
550 // This only works for EQ and NE
552 // If Op1C some other power of two, convert:
553 uint32_t BitWidth = Op1C->getType()->getBitWidth();
554 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
555 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne);
557 APInt KnownZeroMask(~KnownZero);
558 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
559 if (!DoXform) return ICI;
561 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
562 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
563 // (X&4) == 2 --> false
564 // (X&4) != 2 --> true
565 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
567 Res = ConstantExpr::getZExt(Res, CI.getType());
568 return ReplaceInstUsesWith(CI, Res);
571 uint32_t ShiftAmt = KnownZeroMask.logBase2();
572 Value *In = ICI->getOperand(0);
574 // Perform a logical shr by shiftamt.
575 // Insert the shift to put the result in the low bit.
576 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
577 In->getName()+".lobit");
580 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
581 Constant *One = ConstantInt::get(In->getType(), 1);
582 In = Builder->CreateXor(In, One);
585 if (CI.getType() == In->getType())
586 return ReplaceInstUsesWith(CI, In);
587 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
592 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
593 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
594 // may lead to additional simplifications.
595 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
596 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
597 uint32_t BitWidth = ITy->getBitWidth();
598 Value *LHS = ICI->getOperand(0);
599 Value *RHS = ICI->getOperand(1);
601 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
602 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
603 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS);
604 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS);
606 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
607 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
608 APInt UnknownBit = ~KnownBits;
609 if (UnknownBit.countPopulation() == 1) {
610 if (!DoXform) return ICI;
612 Value *Result = Builder->CreateXor(LHS, RHS);
614 // Mask off any bits that are set and won't be shifted away.
615 if (KnownOneLHS.uge(UnknownBit))
616 Result = Builder->CreateAnd(Result,
617 ConstantInt::get(ITy, UnknownBit));
619 // Shift the bit we're testing down to the lsb.
620 Result = Builder->CreateLShr(
621 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
623 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
624 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
625 Result->takeName(ICI);
626 return ReplaceInstUsesWith(CI, Result);
635 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
636 /// specified wider type and produce the same low bits. If not, return false.
638 /// If this function returns true, it can also return a non-zero number of bits
639 /// (in BitsToClear) which indicates that the value it computes is correct for
640 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
641 /// out. For example, to promote something like:
643 /// %B = trunc i64 %A to i32
644 /// %C = lshr i32 %B, 8
645 /// %E = zext i32 %C to i64
647 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
648 /// set to 8 to indicate that the promoted value needs to have bits 24-31
649 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
650 /// clear the top bits anyway, doing this has no extra cost.
652 /// This function works on both vectors and scalars.
653 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
655 if (isa<Constant>(V))
658 Instruction *I = dyn_cast<Instruction>(V);
659 if (!I) return false;
661 // If the input is a truncate from the destination type, we can trivially
663 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
666 // We can't extend or shrink something that has multiple uses: doing so would
667 // require duplicating the instruction in general, which isn't profitable.
668 if (!I->hasOneUse()) return false;
670 unsigned Opc = I->getOpcode(), Tmp;
672 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
673 case Instruction::SExt: // zext(sext(x)) -> sext(x).
674 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
676 case Instruction::And:
677 case Instruction::Or:
678 case Instruction::Xor:
679 case Instruction::Add:
680 case Instruction::Sub:
681 case Instruction::Mul:
682 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
683 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
685 // These can all be promoted if neither operand has 'bits to clear'.
686 if (BitsToClear == 0 && Tmp == 0)
689 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
690 // other side, BitsToClear is ok.
692 (Opc == Instruction::And || Opc == Instruction::Or ||
693 Opc == Instruction::Xor)) {
694 // We use MaskedValueIsZero here for generality, but the case we care
695 // about the most is constant RHS.
696 unsigned VSize = V->getType()->getScalarSizeInBits();
697 if (MaskedValueIsZero(I->getOperand(1),
698 APInt::getHighBitsSet(VSize, BitsToClear)))
702 // Otherwise, we don't know how to analyze this BitsToClear case yet.
705 case Instruction::Shl:
706 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
707 // upper bits we can reduce BitsToClear by the shift amount.
708 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
709 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
711 uint64_t ShiftAmt = Amt->getZExtValue();
712 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
716 case Instruction::LShr:
717 // We can promote lshr(x, cst) if we can promote x. This requires the
718 // ultimate 'and' to clear out the high zero bits we're clearing out though.
719 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
720 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
722 BitsToClear += Amt->getZExtValue();
723 if (BitsToClear > V->getType()->getScalarSizeInBits())
724 BitsToClear = V->getType()->getScalarSizeInBits();
727 // Cannot promote variable LSHR.
729 case Instruction::Select:
730 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
731 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
732 // TODO: If important, we could handle the case when the BitsToClear are
733 // known zero in the disagreeing side.
738 case Instruction::PHI: {
739 // We can change a phi if we can change all operands. Note that we never
740 // get into trouble with cyclic PHIs here because we only consider
741 // instructions with a single use.
742 PHINode *PN = cast<PHINode>(I);
743 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
745 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
746 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
747 // TODO: If important, we could handle the case when the BitsToClear
748 // are known zero in the disagreeing input.
754 // TODO: Can handle more cases here.
759 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
760 // If this zero extend is only used by a truncate, let the truncate be
761 // eliminated before we try to optimize this zext.
762 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
765 // If one of the common conversion will work, do it.
766 if (Instruction *Result = commonCastTransforms(CI))
769 // See if we can simplify any instructions used by the input whose sole
770 // purpose is to compute bits we don't care about.
771 if (SimplifyDemandedInstructionBits(CI))
774 Value *Src = CI.getOperand(0);
775 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
777 // Attempt to extend the entire input expression tree to the destination
778 // type. Only do this if the dest type is a simple type, don't convert the
779 // expression tree to something weird like i93 unless the source is also
781 unsigned BitsToClear;
782 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
783 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
784 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
785 "Unreasonable BitsToClear");
787 // Okay, we can transform this! Insert the new expression now.
788 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
789 " to avoid zero extend: " << CI);
790 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
791 assert(Res->getType() == DestTy);
793 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
794 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
796 // If the high bits are already filled with zeros, just replace this
797 // cast with the result.
798 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
799 DestBitSize-SrcBitsKept)))
800 return ReplaceInstUsesWith(CI, Res);
802 // We need to emit an AND to clear the high bits.
803 Constant *C = ConstantInt::get(Res->getType(),
804 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
805 return BinaryOperator::CreateAnd(Res, C);
808 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
809 // types and if the sizes are just right we can convert this into a logical
810 // 'and' which will be much cheaper than the pair of casts.
811 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
812 // TODO: Subsume this into EvaluateInDifferentType.
814 // Get the sizes of the types involved. We know that the intermediate type
815 // will be smaller than A or C, but don't know the relation between A and C.
816 Value *A = CSrc->getOperand(0);
817 unsigned SrcSize = A->getType()->getScalarSizeInBits();
818 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
819 unsigned DstSize = CI.getType()->getScalarSizeInBits();
820 // If we're actually extending zero bits, then if
821 // SrcSize < DstSize: zext(a & mask)
822 // SrcSize == DstSize: a & mask
823 // SrcSize > DstSize: trunc(a) & mask
824 if (SrcSize < DstSize) {
825 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
826 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
827 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
828 return new ZExtInst(And, CI.getType());
831 if (SrcSize == DstSize) {
832 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
833 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
836 if (SrcSize > DstSize) {
837 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
838 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
839 return BinaryOperator::CreateAnd(Trunc,
840 ConstantInt::get(Trunc->getType(),
845 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
846 return transformZExtICmp(ICI, CI);
848 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
849 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
850 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
851 // of the (zext icmp) will be transformed.
852 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
853 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
854 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
855 (transformZExtICmp(LHS, CI, false) ||
856 transformZExtICmp(RHS, CI, false))) {
857 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
858 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
859 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
863 // zext(trunc(X) & C) -> (X & zext(C)).
867 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
868 X->getType() == CI.getType())
869 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
871 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
873 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
874 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
875 X->getType() == CI.getType()) {
876 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
877 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
880 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
881 if (SrcI && SrcI->hasOneUse() &&
882 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
883 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
884 Value *New = Builder->CreateZExt(X, CI.getType());
885 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
891 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
892 /// in order to eliminate the icmp.
893 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
894 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
895 ICmpInst::Predicate Pred = ICI->getPredicate();
897 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
898 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
899 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
900 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
901 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
903 Value *Sh = ConstantInt::get(Op0->getType(),
904 Op0->getType()->getScalarSizeInBits()-1);
905 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
906 if (In->getType() != CI.getType())
907 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
909 if (Pred == ICmpInst::ICMP_SGT)
910 In = Builder->CreateNot(In, In->getName()+".not");
911 return ReplaceInstUsesWith(CI, In);
915 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
916 // If we know that only one bit of the LHS of the icmp can be set and we
917 // have an equality comparison with zero or a power of 2, we can transform
918 // the icmp and sext into bitwise/integer operations.
919 if (ICI->hasOneUse() &&
920 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
921 unsigned BitWidth = Op1C->getType()->getBitWidth();
922 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
923 ComputeMaskedBits(Op0, KnownZero, KnownOne);
925 APInt KnownZeroMask(~KnownZero);
926 if (KnownZeroMask.isPowerOf2()) {
927 Value *In = ICI->getOperand(0);
929 // If the icmp tests for a known zero bit we can constant fold it.
930 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
931 Value *V = Pred == ICmpInst::ICMP_NE ?
932 ConstantInt::getAllOnesValue(CI.getType()) :
933 ConstantInt::getNullValue(CI.getType());
934 return ReplaceInstUsesWith(CI, V);
937 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
938 // sext ((x & 2^n) == 0) -> (x >> n) - 1
939 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
940 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
941 // Perform a right shift to place the desired bit in the LSB.
943 In = Builder->CreateLShr(In,
944 ConstantInt::get(In->getType(), ShiftAmt));
946 // At this point "In" is either 1 or 0. Subtract 1 to turn
947 // {1, 0} -> {0, -1}.
948 In = Builder->CreateAdd(In,
949 ConstantInt::getAllOnesValue(In->getType()),
952 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
953 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
954 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
955 // Perform a left shift to place the desired bit in the MSB.
957 In = Builder->CreateShl(In,
958 ConstantInt::get(In->getType(), ShiftAmt));
960 // Distribute the bit over the whole bit width.
961 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
962 BitWidth - 1), "sext");
965 if (CI.getType() == In->getType())
966 return ReplaceInstUsesWith(CI, In);
967 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
975 /// CanEvaluateSExtd - Return true if we can take the specified value
976 /// and return it as type Ty without inserting any new casts and without
977 /// changing the value of the common low bits. This is used by code that tries
978 /// to promote integer operations to a wider types will allow us to eliminate
981 /// This function works on both vectors and scalars.
983 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
984 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
985 "Can't sign extend type to a smaller type");
986 // If this is a constant, it can be trivially promoted.
987 if (isa<Constant>(V))
990 Instruction *I = dyn_cast<Instruction>(V);
991 if (!I) return false;
993 // If this is a truncate from the dest type, we can trivially eliminate it.
994 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
997 // We can't extend or shrink something that has multiple uses: doing so would
998 // require duplicating the instruction in general, which isn't profitable.
999 if (!I->hasOneUse()) return false;
1001 switch (I->getOpcode()) {
1002 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1003 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1004 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1006 case Instruction::And:
1007 case Instruction::Or:
1008 case Instruction::Xor:
1009 case Instruction::Add:
1010 case Instruction::Sub:
1011 case Instruction::Mul:
1012 // These operators can all arbitrarily be extended if their inputs can.
1013 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1014 CanEvaluateSExtd(I->getOperand(1), Ty);
1016 //case Instruction::Shl: TODO
1017 //case Instruction::LShr: TODO
1019 case Instruction::Select:
1020 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1021 CanEvaluateSExtd(I->getOperand(2), Ty);
1023 case Instruction::PHI: {
1024 // We can change a phi if we can change all operands. Note that we never
1025 // get into trouble with cyclic PHIs here because we only consider
1026 // instructions with a single use.
1027 PHINode *PN = cast<PHINode>(I);
1028 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1029 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1033 // TODO: Can handle more cases here.
1040 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1041 // If this sign extend is only used by a truncate, let the truncate be
1042 // eliminated before we try to optimize this sext.
1043 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1046 if (Instruction *I = commonCastTransforms(CI))
1049 // See if we can simplify any instructions used by the input whose sole
1050 // purpose is to compute bits we don't care about.
1051 if (SimplifyDemandedInstructionBits(CI))
1054 Value *Src = CI.getOperand(0);
1055 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1057 // Attempt to extend the entire input expression tree to the destination
1058 // type. Only do this if the dest type is a simple type, don't convert the
1059 // expression tree to something weird like i93 unless the source is also
1061 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1062 CanEvaluateSExtd(Src, DestTy)) {
1063 // Okay, we can transform this! Insert the new expression now.
1064 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1065 " to avoid sign extend: " << CI);
1066 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1067 assert(Res->getType() == DestTy);
1069 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1070 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1072 // If the high bits are already filled with sign bit, just replace this
1073 // cast with the result.
1074 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1075 return ReplaceInstUsesWith(CI, Res);
1077 // We need to emit a shl + ashr to do the sign extend.
1078 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1079 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1083 // If this input is a trunc from our destination, then turn sext(trunc(x))
1085 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1086 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1087 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1088 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1090 // We need to emit a shl + ashr to do the sign extend.
1091 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1092 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1093 return BinaryOperator::CreateAShr(Res, ShAmt);
1096 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1097 return transformSExtICmp(ICI, CI);
1099 // If the input is a shl/ashr pair of a same constant, then this is a sign
1100 // extension from a smaller value. If we could trust arbitrary bitwidth
1101 // integers, we could turn this into a truncate to the smaller bit and then
1102 // use a sext for the whole extension. Since we don't, look deeper and check
1103 // for a truncate. If the source and dest are the same type, eliminate the
1104 // trunc and extend and just do shifts. For example, turn:
1105 // %a = trunc i32 %i to i8
1106 // %b = shl i8 %a, 6
1107 // %c = ashr i8 %b, 6
1108 // %d = sext i8 %c to i32
1110 // %a = shl i32 %i, 30
1111 // %d = ashr i32 %a, 30
1113 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1114 ConstantInt *BA = nullptr, *CA = nullptr;
1115 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1116 m_ConstantInt(CA))) &&
1117 BA == CA && A->getType() == CI.getType()) {
1118 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1119 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1120 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1121 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1122 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1123 return BinaryOperator::CreateAShr(A, ShAmtV);
1130 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1131 /// in the specified FP type without changing its value.
1132 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1134 APFloat F = CFP->getValueAPF();
1135 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1137 return ConstantFP::get(CFP->getContext(), F);
1141 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1142 /// through it until we get the source value.
1143 static Value *LookThroughFPExtensions(Value *V) {
1144 if (Instruction *I = dyn_cast<Instruction>(V))
1145 if (I->getOpcode() == Instruction::FPExt)
1146 return LookThroughFPExtensions(I->getOperand(0));
1148 // If this value is a constant, return the constant in the smallest FP type
1149 // that can accurately represent it. This allows us to turn
1150 // (float)((double)X+2.0) into x+2.0f.
1151 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1152 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1153 return V; // No constant folding of this.
1154 // See if the value can be truncated to half and then reextended.
1155 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1157 // See if the value can be truncated to float and then reextended.
1158 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1160 if (CFP->getType()->isDoubleTy())
1161 return V; // Won't shrink.
1162 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1164 // Don't try to shrink to various long double types.
1170 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1171 if (Instruction *I = commonCastTransforms(CI))
1173 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1174 // simpilify this expression to avoid one or more of the trunc/extend
1175 // operations if we can do so without changing the numerical results.
1177 // The exact manner in which the widths of the operands interact to limit
1178 // what we can and cannot do safely varies from operation to operation, and
1179 // is explained below in the various case statements.
1180 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1181 if (OpI && OpI->hasOneUse()) {
1182 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1183 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1184 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1185 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1186 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1187 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1188 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1189 switch (OpI->getOpcode()) {
1191 case Instruction::FAdd:
1192 case Instruction::FSub:
1193 // For addition and subtraction, the infinitely precise result can
1194 // essentially be arbitrarily wide; proving that double rounding
1195 // will not occur because the result of OpI is exact (as we will for
1196 // FMul, for example) is hopeless. However, we *can* nonetheless
1197 // frequently know that double rounding cannot occur (or that it is
1198 // innocuous) by taking advantage of the specific structure of
1199 // infinitely-precise results that admit double rounding.
1201 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1202 // to represent both sources, we can guarantee that the double
1203 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1204 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1205 // for proof of this fact).
1207 // Note: Figueroa does not consider the case where DstFormat !=
1208 // SrcFormat. It's possible (likely even!) that this analysis
1209 // could be tightened for those cases, but they are rare (the main
1210 // case of interest here is (float)((double)float + float)).
1211 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1212 if (LHSOrig->getType() != CI.getType())
1213 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1214 if (RHSOrig->getType() != CI.getType())
1215 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1217 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1218 RI->copyFastMathFlags(OpI);
1222 case Instruction::FMul:
1223 // For multiplication, the infinitely precise result has at most
1224 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1225 // that such a value can be exactly represented, then no double
1226 // rounding can possibly occur; we can safely perform the operation
1227 // in the destination format if it can represent both sources.
1228 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1229 if (LHSOrig->getType() != CI.getType())
1230 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1231 if (RHSOrig->getType() != CI.getType())
1232 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1234 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1235 RI->copyFastMathFlags(OpI);
1239 case Instruction::FDiv:
1240 // For division, we use again use the bound from Figueroa's
1241 // dissertation. I am entirely certain that this bound can be
1242 // tightened in the unbalanced operand case by an analysis based on
1243 // the diophantine rational approximation bound, but the well-known
1244 // condition used here is a good conservative first pass.
1245 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1246 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1247 if (LHSOrig->getType() != CI.getType())
1248 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1249 if (RHSOrig->getType() != CI.getType())
1250 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1252 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1253 RI->copyFastMathFlags(OpI);
1257 case Instruction::FRem:
1258 // Remainder is straightforward. Remainder is always exact, so the
1259 // type of OpI doesn't enter into things at all. We simply evaluate
1260 // in whichever source type is larger, then convert to the
1261 // destination type.
1262 if (LHSWidth < SrcWidth)
1263 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1264 else if (RHSWidth <= SrcWidth)
1265 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1266 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1267 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1268 RI->copyFastMathFlags(OpI);
1269 return CastInst::CreateFPCast(ExactResult, CI.getType());
1272 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1273 if (BinaryOperator::isFNeg(OpI)) {
1274 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1276 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1277 RI->copyFastMathFlags(OpI);
1282 // (fptrunc (select cond, R1, Cst)) -->
1283 // (select cond, (fptrunc R1), (fptrunc Cst))
1284 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1286 (isa<ConstantFP>(SI->getOperand(1)) ||
1287 isa<ConstantFP>(SI->getOperand(2)))) {
1288 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1290 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1292 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1295 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1297 switch (II->getIntrinsicID()) {
1299 case Intrinsic::fabs: {
1300 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1301 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1303 Type *IntrinsicType[] = { CI.getType() };
1304 Function *Overload =
1305 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1306 II->getIntrinsicID(), IntrinsicType);
1308 Value *Args[] = { InnerTrunc };
1309 return CallInst::Create(Overload, Args, II->getName());
1314 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1315 // Note that we restrict this transformation based on
1316 // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because
1317 // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the
1318 // single-precision intrinsic can be expanded in the backend.
1319 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1320 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1321 (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) ||
1322 Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) &&
1323 Call->getNumArgOperands() == 1 &&
1324 Call->hasOneUse()) {
1325 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1326 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1327 CI.getType()->isFloatTy() &&
1328 Call->getType()->isDoubleTy() &&
1329 Arg->getType()->isDoubleTy() &&
1330 Arg->getOperand(0)->getType()->isFloatTy()) {
1331 Function *Callee = Call->getCalledFunction();
1332 Module *M = CI.getParent()->getParent()->getParent();
1333 Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ?
1334 Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) :
1335 M->getOrInsertFunction("sqrtf", Callee->getAttributes(),
1336 Builder->getFloatTy(), Builder->getFloatTy(),
1338 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1340 ret->setAttributes(Callee->getAttributes());
1343 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1344 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1345 EraseInstFromFunction(*Call);
1353 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1354 return commonCastTransforms(CI);
1357 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1358 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1360 return commonCastTransforms(FI);
1362 // fptoui(uitofp(X)) --> X
1363 // fptoui(sitofp(X)) --> X
1364 // This is safe if the intermediate type has enough bits in its mantissa to
1365 // accurately represent all values of X. For example, do not do this with
1366 // i64->float->i64. This is also safe for sitofp case, because any negative
1367 // 'X' value would cause an undefined result for the fptoui.
1368 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1369 OpI->getOperand(0)->getType() == FI.getType() &&
1370 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1371 OpI->getType()->getFPMantissaWidth())
1372 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1374 return commonCastTransforms(FI);
1377 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1378 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1380 return commonCastTransforms(FI);
1382 // fptosi(sitofp(X)) --> X
1383 // fptosi(uitofp(X)) --> X
1384 // This is safe if the intermediate type has enough bits in its mantissa to
1385 // accurately represent all values of X. For example, do not do this with
1386 // i64->float->i64. This is also safe for sitofp case, because any negative
1387 // 'X' value would cause an undefined result for the fptoui.
1388 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1389 OpI->getOperand(0)->getType() == FI.getType() &&
1390 (int)FI.getType()->getScalarSizeInBits() <=
1391 OpI->getType()->getFPMantissaWidth())
1392 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1394 return commonCastTransforms(FI);
1397 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1398 return commonCastTransforms(CI);
1401 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1402 return commonCastTransforms(CI);
1405 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1406 // If the source integer type is not the intptr_t type for this target, do a
1407 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1408 // cast to be exposed to other transforms.
1411 unsigned AS = CI.getAddressSpace();
1412 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1413 DL->getPointerSizeInBits(AS)) {
1414 Type *Ty = DL->getIntPtrType(CI.getContext(), AS);
1415 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1416 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1418 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1419 return new IntToPtrInst(P, CI.getType());
1423 if (Instruction *I = commonCastTransforms(CI))
1429 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1430 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1431 Value *Src = CI.getOperand(0);
1433 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1434 // If casting the result of a getelementptr instruction with no offset, turn
1435 // this into a cast of the original pointer!
1436 if (GEP->hasAllZeroIndices()) {
1437 // Changing the cast operand is usually not a good idea but it is safe
1438 // here because the pointer operand is being replaced with another
1439 // pointer operand so the opcode doesn't need to change.
1441 CI.setOperand(0, GEP->getOperand(0));
1446 return commonCastTransforms(CI);
1448 // If the GEP has a single use, and the base pointer is a bitcast, and the
1449 // GEP computes a constant offset, see if we can convert these three
1450 // instructions into fewer. This typically happens with unions and other
1451 // non-type-safe code.
1452 unsigned AS = GEP->getPointerAddressSpace();
1453 unsigned OffsetBits = DL->getPointerSizeInBits(AS);
1454 APInt Offset(OffsetBits, 0);
1455 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0));
1456 if (GEP->hasOneUse() &&
1458 GEP->accumulateConstantOffset(*DL, Offset)) {
1459 // Get the base pointer input of the bitcast, and the type it points to.
1460 Value *OrigBase = BCI->getOperand(0);
1461 SmallVector<Value*, 8> NewIndices;
1462 if (FindElementAtOffset(OrigBase->getType(),
1463 Offset.getSExtValue(),
1465 // If we were able to index down into an element, create the GEP
1466 // and bitcast the result. This eliminates one bitcast, potentially
1468 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1469 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1470 Builder->CreateGEP(OrigBase, NewIndices);
1471 NGEP->takeName(GEP);
1473 if (isa<BitCastInst>(CI))
1474 return new BitCastInst(NGEP, CI.getType());
1475 assert(isa<PtrToIntInst>(CI));
1476 return new PtrToIntInst(NGEP, CI.getType());
1481 return commonCastTransforms(CI);
1484 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1485 // If the destination integer type is not the intptr_t type for this target,
1486 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1487 // to be exposed to other transforms.
1490 return commonPointerCastTransforms(CI);
1492 Type *Ty = CI.getType();
1493 unsigned AS = CI.getPointerAddressSpace();
1495 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS))
1496 return commonPointerCastTransforms(CI);
1498 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS);
1499 if (Ty->isVectorTy()) // Handle vectors of pointers.
1500 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1502 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1503 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1506 /// OptimizeVectorResize - This input value (which is known to have vector type)
1507 /// is being zero extended or truncated to the specified vector type. Try to
1508 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1510 /// The source and destination vector types may have different element types.
1511 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1513 // We can only do this optimization if the output is a multiple of the input
1514 // element size, or the input is a multiple of the output element size.
1515 // Convert the input type to have the same element type as the output.
1516 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1518 if (SrcTy->getElementType() != DestTy->getElementType()) {
1519 // The input types don't need to be identical, but for now they must be the
1520 // same size. There is no specific reason we couldn't handle things like
1521 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1523 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1524 DestTy->getElementType()->getPrimitiveSizeInBits())
1527 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1528 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1531 // Now that the element types match, get the shuffle mask and RHS of the
1532 // shuffle to use, which depends on whether we're increasing or decreasing the
1533 // size of the input.
1534 SmallVector<uint32_t, 16> ShuffleMask;
1537 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1538 // If we're shrinking the number of elements, just shuffle in the low
1539 // elements from the input and use undef as the second shuffle input.
1540 V2 = UndefValue::get(SrcTy);
1541 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1542 ShuffleMask.push_back(i);
1545 // If we're increasing the number of elements, shuffle in all of the
1546 // elements from InVal and fill the rest of the result elements with zeros
1547 // from a constant zero.
1548 V2 = Constant::getNullValue(SrcTy);
1549 unsigned SrcElts = SrcTy->getNumElements();
1550 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1551 ShuffleMask.push_back(i);
1553 // The excess elements reference the first element of the zero input.
1554 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1555 ShuffleMask.push_back(SrcElts);
1558 return new ShuffleVectorInst(InVal, V2,
1559 ConstantDataVector::get(V2->getContext(),
1563 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1564 return Value % Ty->getPrimitiveSizeInBits() == 0;
1567 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1568 return Value / Ty->getPrimitiveSizeInBits();
1571 /// CollectInsertionElements - V is a value which is inserted into a vector of
1572 /// VecEltTy. Look through the value to see if we can decompose it into
1573 /// insertions into the vector. See the example in the comment for
1574 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1575 /// The type of V is always a non-zero multiple of VecEltTy's size.
1576 /// Shift is the number of bits between the lsb of V and the lsb of
1579 /// This returns false if the pattern can't be matched or true if it can,
1580 /// filling in Elements with the elements found here.
1581 static bool CollectInsertionElements(Value *V, unsigned Shift,
1582 SmallVectorImpl<Value*> &Elements,
1583 Type *VecEltTy, InstCombiner &IC) {
1584 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1585 "Shift should be a multiple of the element type size");
1587 // Undef values never contribute useful bits to the result.
1588 if (isa<UndefValue>(V)) return true;
1590 // If we got down to a value of the right type, we win, try inserting into the
1592 if (V->getType() == VecEltTy) {
1593 // Inserting null doesn't actually insert any elements.
1594 if (Constant *C = dyn_cast<Constant>(V))
1595 if (C->isNullValue())
1598 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1599 if (IC.getDataLayout()->isBigEndian())
1600 ElementIndex = Elements.size() - ElementIndex - 1;
1602 // Fail if multiple elements are inserted into this slot.
1603 if (Elements[ElementIndex])
1606 Elements[ElementIndex] = V;
1610 if (Constant *C = dyn_cast<Constant>(V)) {
1611 // Figure out the # elements this provides, and bitcast it or slice it up
1613 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1615 // If the constant is the size of a vector element, we just need to bitcast
1616 // it to the right type so it gets properly inserted.
1618 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1619 Shift, Elements, VecEltTy, IC);
1621 // Okay, this is a constant that covers multiple elements. Slice it up into
1622 // pieces and insert each element-sized piece into the vector.
1623 if (!isa<IntegerType>(C->getType()))
1624 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1625 C->getType()->getPrimitiveSizeInBits()));
1626 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1627 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1629 for (unsigned i = 0; i != NumElts; ++i) {
1630 unsigned ShiftI = Shift+i*ElementSize;
1631 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1633 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1634 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC))
1640 if (!V->hasOneUse()) return false;
1642 Instruction *I = dyn_cast<Instruction>(V);
1643 if (!I) return false;
1644 switch (I->getOpcode()) {
1645 default: return false; // Unhandled case.
1646 case Instruction::BitCast:
1647 return CollectInsertionElements(I->getOperand(0), Shift,
1648 Elements, VecEltTy, IC);
1649 case Instruction::ZExt:
1650 if (!isMultipleOfTypeSize(
1651 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1654 return CollectInsertionElements(I->getOperand(0), Shift,
1655 Elements, VecEltTy, IC);
1656 case Instruction::Or:
1657 return CollectInsertionElements(I->getOperand(0), Shift,
1658 Elements, VecEltTy, IC) &&
1659 CollectInsertionElements(I->getOperand(1), Shift,
1660 Elements, VecEltTy, IC);
1661 case Instruction::Shl: {
1662 // Must be shifting by a constant that is a multiple of the element size.
1663 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1664 if (!CI) return false;
1665 Shift += CI->getZExtValue();
1666 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1667 return CollectInsertionElements(I->getOperand(0), Shift,
1668 Elements, VecEltTy, IC);
1675 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1676 /// may be doing shifts and ors to assemble the elements of the vector manually.
1677 /// Try to rip the code out and replace it with insertelements. This is to
1678 /// optimize code like this:
1680 /// %tmp37 = bitcast float %inc to i32
1681 /// %tmp38 = zext i32 %tmp37 to i64
1682 /// %tmp31 = bitcast float %inc5 to i32
1683 /// %tmp32 = zext i32 %tmp31 to i64
1684 /// %tmp33 = shl i64 %tmp32, 32
1685 /// %ins35 = or i64 %tmp33, %tmp38
1686 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1688 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1689 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1691 // We need to know the target byte order to perform this optimization.
1692 if (!IC.getDataLayout()) return nullptr;
1694 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1695 Value *IntInput = CI.getOperand(0);
1697 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1698 if (!CollectInsertionElements(IntInput, 0, Elements,
1699 DestVecTy->getElementType(), IC))
1702 // If we succeeded, we know that all of the element are specified by Elements
1703 // or are zero if Elements has a null entry. Recast this as a set of
1705 Value *Result = Constant::getNullValue(CI.getType());
1706 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1707 if (!Elements[i]) continue; // Unset element.
1709 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1710 IC.Builder->getInt32(i));
1717 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1718 /// bitcast. The various long double bitcasts can't get in here.
1719 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1720 // We need to know the target byte order to perform this optimization.
1721 if (!IC.getDataLayout()) return nullptr;
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 (IC.getDataLayout()->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 (IC.getDataLayout()->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 Constant *ZeroUInt =
1806 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1807 unsigned NumZeros = 0;
1808 while (SrcElTy != DstElTy &&
1809 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1810 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1811 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1815 // If we found a path from the src to dest, create the getelementptr now.
1816 if (SrcElTy == DstElTy) {
1817 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1818 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1822 // Try to optimize int -> float bitcasts.
1823 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1824 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1827 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1828 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1829 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1830 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1831 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1832 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1835 if (isa<IntegerType>(SrcTy)) {
1836 // If this is a cast from an integer to vector, check to see if the input
1837 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1838 // the casts with a shuffle and (potentially) a bitcast.
1839 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1840 CastInst *SrcCast = cast<CastInst>(Src);
1841 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1842 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1843 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1844 cast<VectorType>(DestTy), *this))
1848 // If the input is an 'or' instruction, we may be doing shifts and ors to
1849 // assemble the elements of the vector manually. Try to rip the code out
1850 // and replace it with insertelements.
1851 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1852 return ReplaceInstUsesWith(CI, V);
1856 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1857 if (SrcVTy->getNumElements() == 1) {
1858 // If our destination is not a vector, then make this a straight
1859 // scalar-scalar cast.
1860 if (!DestTy->isVectorTy()) {
1862 Builder->CreateExtractElement(Src,
1863 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1864 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1867 // Otherwise, see if our source is an insert. If so, then use the scalar
1868 // component directly.
1869 if (InsertElementInst *IEI =
1870 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1871 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1876 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1877 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1878 // a bitcast to a vector with the same # elts.
1879 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1880 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1881 SVI->getType()->getNumElements() ==
1882 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1884 // If either of the operands is a cast from CI.getType(), then
1885 // evaluating the shuffle in the casted destination's type will allow
1886 // us to eliminate at least one cast.
1887 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1888 Tmp->getOperand(0)->getType() == DestTy) ||
1889 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1890 Tmp->getOperand(0)->getType() == DestTy)) {
1891 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1892 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1893 // Return a new shuffle vector. Use the same element ID's, as we
1894 // know the vector types match #elts.
1895 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1900 if (SrcTy->isPointerTy())
1901 return commonPointerCastTransforms(CI);
1902 return commonCastTransforms(CI);
1905 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1906 return commonPointerCastTransforms(CI);