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 #define DEBUG_TYPE "instcombine"
15 #include "InstCombine.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Target/TargetLibraryInfo.h"
21 using namespace PatternMatch;
23 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
24 /// expression. If so, decompose it, returning some value X, such that Val is
27 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
29 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
30 Offset = CI->getZExtValue();
32 return ConstantInt::get(Val->getType(), 0);
35 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
36 // Cannot look past anything that might overflow.
37 OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
38 if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
44 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
45 if (I->getOpcode() == Instruction::Shl) {
46 // This is a value scaled by '1 << the shift amt'.
47 Scale = UINT64_C(1) << RHS->getZExtValue();
49 return I->getOperand(0);
52 if (I->getOpcode() == Instruction::Mul) {
53 // This value is scaled by 'RHS'.
54 Scale = RHS->getZExtValue();
56 return I->getOperand(0);
59 if (I->getOpcode() == Instruction::Add) {
60 // We have X+C. Check to see if we really have (X*C2)+C1,
61 // where C1 is divisible by C2.
64 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
65 Offset += RHS->getZExtValue();
72 // Otherwise, we can't look past this.
78 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
79 /// try to eliminate the cast by moving the type information into the alloc.
80 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
82 // This requires DataLayout to get the alloca alignment and size information.
85 PointerType *PTy = cast<PointerType>(CI.getType());
87 BuilderTy AllocaBuilder(*Builder);
88 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
90 // Get the type really allocated and the type casted to.
91 Type *AllocElTy = AI.getAllocatedType();
92 Type *CastElTy = PTy->getElementType();
93 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
95 unsigned AllocElTyAlign = DL->getABITypeAlignment(AllocElTy);
96 unsigned CastElTyAlign = DL->getABITypeAlignment(CastElTy);
97 if (CastElTyAlign < AllocElTyAlign) return 0;
99 // If the allocation has multiple uses, only promote it if we are strictly
100 // increasing the alignment of the resultant allocation. If we keep it the
101 // same, we open the door to infinite loops of various kinds.
102 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
104 uint64_t AllocElTySize = DL->getTypeAllocSize(AllocElTy);
105 uint64_t CastElTySize = DL->getTypeAllocSize(CastElTy);
106 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
108 // If the allocation has multiple uses, only promote it if we're not
109 // shrinking the amount of memory being allocated.
110 uint64_t AllocElTyStoreSize = DL->getTypeStoreSize(AllocElTy);
111 uint64_t CastElTyStoreSize = DL->getTypeStoreSize(CastElTy);
112 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return 0;
114 // See if we can satisfy the modulus by pulling a scale out of the array
116 unsigned ArraySizeScale;
117 uint64_t ArrayOffset;
118 Value *NumElements = // See if the array size is a decomposable linear expr.
119 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
121 // If we can now satisfy the modulus, by using a non-1 scale, we really can
123 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
124 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
126 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
131 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
132 // Insert before the alloca, not before the cast.
133 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
136 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
137 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
139 Amt = AllocaBuilder.CreateAdd(Amt, Off);
142 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
143 New->setAlignment(AI.getAlignment());
146 // If the allocation has multiple real uses, insert a cast and change all
147 // things that used it to use the new cast. This will also hack on CI, but it
149 if (!AI.hasOneUse()) {
150 // New is the allocation instruction, pointer typed. AI is the original
151 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
152 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
153 ReplaceInstUsesWith(AI, NewCast);
155 return ReplaceInstUsesWith(CI, New);
158 /// EvaluateInDifferentType - Given an expression that
159 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
160 /// insert the code to evaluate the expression.
161 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
163 if (Constant *C = dyn_cast<Constant>(V)) {
164 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
165 // If we got a constantexpr back, try to simplify it with DL info.
166 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
167 C = ConstantFoldConstantExpression(CE, DL, TLI);
171 // Otherwise, it must be an instruction.
172 Instruction *I = cast<Instruction>(V);
173 Instruction *Res = 0;
174 unsigned Opc = I->getOpcode();
176 case Instruction::Add:
177 case Instruction::Sub:
178 case Instruction::Mul:
179 case Instruction::And:
180 case Instruction::Or:
181 case Instruction::Xor:
182 case Instruction::AShr:
183 case Instruction::LShr:
184 case Instruction::Shl:
185 case Instruction::UDiv:
186 case Instruction::URem: {
187 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
188 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
189 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
192 case Instruction::Trunc:
193 case Instruction::ZExt:
194 case Instruction::SExt:
195 // If the source type of the cast is the type we're trying for then we can
196 // just return the source. There's no need to insert it because it is not
198 if (I->getOperand(0)->getType() == Ty)
199 return I->getOperand(0);
201 // Otherwise, must be the same type of cast, so just reinsert a new one.
202 // This also handles the case of zext(trunc(x)) -> zext(x).
203 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
204 Opc == Instruction::SExt);
206 case Instruction::Select: {
207 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
208 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
209 Res = SelectInst::Create(I->getOperand(0), True, False);
212 case Instruction::PHI: {
213 PHINode *OPN = cast<PHINode>(I);
214 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
215 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
216 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
217 NPN->addIncoming(V, OPN->getIncomingBlock(i));
223 // TODO: Can handle more cases here.
224 llvm_unreachable("Unreachable!");
228 return InsertNewInstWith(Res, *I);
232 /// This function is a wrapper around CastInst::isEliminableCastPair. It
233 /// simply extracts arguments and returns what that function returns.
234 static Instruction::CastOps
235 isEliminableCastPair(
236 const CastInst *CI, ///< The first cast instruction
237 unsigned opcode, ///< The opcode of the second cast instruction
238 Type *DstTy, ///< The target type for the second cast instruction
239 const DataLayout *DL ///< The target data for pointer size
242 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
243 Type *MidTy = CI->getType(); // B from above
245 // Get the opcodes of the two Cast instructions
246 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
247 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
248 Type *SrcIntPtrTy = DL && SrcTy->isPtrOrPtrVectorTy() ?
249 DL->getIntPtrType(SrcTy) : 0;
250 Type *MidIntPtrTy = DL && MidTy->isPtrOrPtrVectorTy() ?
251 DL->getIntPtrType(MidTy) : 0;
252 Type *DstIntPtrTy = DL && DstTy->isPtrOrPtrVectorTy() ?
253 DL->getIntPtrType(DstTy) : 0;
254 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
255 DstTy, SrcIntPtrTy, MidIntPtrTy,
258 // We don't want to form an inttoptr or ptrtoint that converts to an integer
259 // type that differs from the pointer size.
260 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
261 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
264 return Instruction::CastOps(Res);
267 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
268 /// results in any code being generated and is interesting to optimize out. If
269 /// the cast can be eliminated by some other simple transformation, we prefer
270 /// to do the simplification first.
271 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
273 // Noop casts and casts of constants should be eliminated trivially.
274 if (V->getType() == Ty || isa<Constant>(V)) return false;
276 // If this is another cast that can be eliminated, we prefer to have it
278 if (const CastInst *CI = dyn_cast<CastInst>(V))
279 if (isEliminableCastPair(CI, opc, Ty, DL))
282 // If this is a vector sext from a compare, then we don't want to break the
283 // idiom where each element of the extended vector is either zero or all ones.
284 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
291 /// @brief Implement the transforms common to all CastInst visitors.
292 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
293 Value *Src = CI.getOperand(0);
295 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
297 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
298 if (Instruction::CastOps opc =
299 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
300 // The first cast (CSrc) is eliminable so we need to fix up or replace
301 // the second cast (CI). CSrc will then have a good chance of being dead.
302 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
306 // If we are casting a select then fold the cast into the select
307 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
308 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
311 // If we are casting a PHI then fold the cast into the PHI
312 if (isa<PHINode>(Src)) {
313 // We don't do this if this would create a PHI node with an illegal type if
314 // it is currently legal.
315 if (!Src->getType()->isIntegerTy() ||
316 !CI.getType()->isIntegerTy() ||
317 ShouldChangeType(CI.getType(), Src->getType()))
318 if (Instruction *NV = FoldOpIntoPhi(CI))
325 /// CanEvaluateTruncated - Return true if we can evaluate the specified
326 /// expression tree as type Ty instead of its larger type, and arrive with the
327 /// same value. This is used by code that tries to eliminate truncates.
329 /// Ty will always be a type smaller than V. We should return true if trunc(V)
330 /// can be computed by computing V in the smaller type. If V is an instruction,
331 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
332 /// makes sense if x and y can be efficiently truncated.
334 /// This function works on both vectors and scalars.
336 static bool CanEvaluateTruncated(Value *V, Type *Ty) {
337 // We can always evaluate constants in another type.
338 if (isa<Constant>(V))
341 Instruction *I = dyn_cast<Instruction>(V);
342 if (!I) return false;
344 Type *OrigTy = V->getType();
346 // If this is an extension from the dest type, we can eliminate it, even if it
347 // has multiple uses.
348 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
349 I->getOperand(0)->getType() == Ty)
352 // We can't extend or shrink something that has multiple uses: doing so would
353 // require duplicating the instruction in general, which isn't profitable.
354 if (!I->hasOneUse()) return false;
356 unsigned Opc = I->getOpcode();
358 case Instruction::Add:
359 case Instruction::Sub:
360 case Instruction::Mul:
361 case Instruction::And:
362 case Instruction::Or:
363 case Instruction::Xor:
364 // These operators can all arbitrarily be extended or truncated.
365 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
366 CanEvaluateTruncated(I->getOperand(1), Ty);
368 case Instruction::UDiv:
369 case Instruction::URem: {
370 // UDiv and URem can be truncated if all the truncated bits are zero.
371 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
372 uint32_t BitWidth = Ty->getScalarSizeInBits();
373 if (BitWidth < OrigBitWidth) {
374 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
375 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
376 MaskedValueIsZero(I->getOperand(1), Mask)) {
377 return CanEvaluateTruncated(I->getOperand(0), Ty) &&
378 CanEvaluateTruncated(I->getOperand(1), Ty);
383 case Instruction::Shl:
384 // If we are truncating the result of this SHL, and if it's a shift of a
385 // constant amount, we can always perform a SHL in a smaller type.
386 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
387 uint32_t BitWidth = Ty->getScalarSizeInBits();
388 if (CI->getLimitedValue(BitWidth) < BitWidth)
389 return CanEvaluateTruncated(I->getOperand(0), Ty);
392 case Instruction::LShr:
393 // If this is a truncate of a logical shr, we can truncate it to a smaller
394 // lshr iff we know that the bits we would otherwise be shifting in are
396 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
397 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
398 uint32_t BitWidth = Ty->getScalarSizeInBits();
399 if (MaskedValueIsZero(I->getOperand(0),
400 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
401 CI->getLimitedValue(BitWidth) < BitWidth) {
402 return CanEvaluateTruncated(I->getOperand(0), Ty);
406 case Instruction::Trunc:
407 // trunc(trunc(x)) -> trunc(x)
409 case Instruction::ZExt:
410 case Instruction::SExt:
411 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
412 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
414 case Instruction::Select: {
415 SelectInst *SI = cast<SelectInst>(I);
416 return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
417 CanEvaluateTruncated(SI->getFalseValue(), Ty);
419 case Instruction::PHI: {
420 // We can change a phi if we can change all operands. Note that we never
421 // get into trouble with cyclic PHIs here because we only consider
422 // instructions with a single use.
423 PHINode *PN = cast<PHINode>(I);
424 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
425 if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
430 // TODO: Can handle more cases here.
437 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
438 if (Instruction *Result = commonCastTransforms(CI))
441 // See if we can simplify any instructions used by the input whose sole
442 // purpose is to compute bits we don't care about.
443 if (SimplifyDemandedInstructionBits(CI))
446 Value *Src = CI.getOperand(0);
447 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
449 // Attempt to truncate the entire input expression tree to the destination
450 // type. Only do this if the dest type is a simple type, don't convert the
451 // expression tree to something weird like i93 unless the source is also
453 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
454 CanEvaluateTruncated(Src, DestTy)) {
456 // If this cast is a truncate, evaluting in a different type always
457 // eliminates the cast, so it is always a win.
458 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
459 " to avoid cast: " << CI << '\n');
460 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
461 assert(Res->getType() == DestTy);
462 return ReplaceInstUsesWith(CI, Res);
465 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
466 if (DestTy->getScalarSizeInBits() == 1) {
467 Constant *One = ConstantInt::get(Src->getType(), 1);
468 Src = Builder->CreateAnd(Src, One);
469 Value *Zero = Constant::getNullValue(Src->getType());
470 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
473 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
474 Value *A = 0; ConstantInt *Cst = 0;
475 if (Src->hasOneUse() &&
476 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
477 // We have three types to worry about here, the type of A, the source of
478 // the truncate (MidSize), and the destination of the truncate. We know that
479 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
480 // between ASize and ResultSize.
481 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
483 // If the shift amount is larger than the size of A, then the result is
484 // known to be zero because all the input bits got shifted out.
485 if (Cst->getZExtValue() >= ASize)
486 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
488 // Since we're doing an lshr and a zero extend, and know that the shift
489 // amount is smaller than ASize, it is always safe to do the shift in A's
490 // type, then zero extend or truncate to the result.
491 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
492 Shift->takeName(Src);
493 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
496 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
497 // type isn't non-native.
498 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
499 ShouldChangeType(Src->getType(), CI.getType()) &&
500 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
501 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
502 return BinaryOperator::CreateAnd(NewTrunc,
503 ConstantExpr::getTrunc(Cst, CI.getType()));
509 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
510 /// in order to eliminate the icmp.
511 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
513 // If we are just checking for a icmp eq of a single bit and zext'ing it
514 // to an integer, then shift the bit to the appropriate place and then
515 // cast to integer to avoid the comparison.
516 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
517 const APInt &Op1CV = Op1C->getValue();
519 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
520 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
521 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
522 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
523 if (!DoXform) return ICI;
525 Value *In = ICI->getOperand(0);
526 Value *Sh = ConstantInt::get(In->getType(),
527 In->getType()->getScalarSizeInBits()-1);
528 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
529 if (In->getType() != CI.getType())
530 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
532 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
533 Constant *One = ConstantInt::get(In->getType(), 1);
534 In = Builder->CreateXor(In, One, In->getName()+".not");
537 return ReplaceInstUsesWith(CI, In);
540 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
541 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
542 // zext (X == 1) to i32 --> X iff X has only the low bit set.
543 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
544 // zext (X != 0) to i32 --> X iff X has only the low bit set.
545 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
546 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
547 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
548 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
549 // This only works for EQ and NE
551 // If Op1C some other power of two, convert:
552 uint32_t BitWidth = Op1C->getType()->getBitWidth();
553 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
554 ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne);
556 APInt KnownZeroMask(~KnownZero);
557 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
558 if (!DoXform) return ICI;
560 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
561 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
562 // (X&4) == 2 --> false
563 // (X&4) != 2 --> true
564 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
566 Res = ConstantExpr::getZExt(Res, CI.getType());
567 return ReplaceInstUsesWith(CI, Res);
570 uint32_t ShiftAmt = KnownZeroMask.logBase2();
571 Value *In = ICI->getOperand(0);
573 // Perform a logical shr by shiftamt.
574 // Insert the shift to put the result in the low bit.
575 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
576 In->getName()+".lobit");
579 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
580 Constant *One = ConstantInt::get(In->getType(), 1);
581 In = Builder->CreateXor(In, One);
584 if (CI.getType() == In->getType())
585 return ReplaceInstUsesWith(CI, In);
586 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
591 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
592 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
593 // may lead to additional simplifications.
594 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
595 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
596 uint32_t BitWidth = ITy->getBitWidth();
597 Value *LHS = ICI->getOperand(0);
598 Value *RHS = ICI->getOperand(1);
600 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
601 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
602 ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS);
603 ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS);
605 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
606 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
607 APInt UnknownBit = ~KnownBits;
608 if (UnknownBit.countPopulation() == 1) {
609 if (!DoXform) return ICI;
611 Value *Result = Builder->CreateXor(LHS, RHS);
613 // Mask off any bits that are set and won't be shifted away.
614 if (KnownOneLHS.uge(UnknownBit))
615 Result = Builder->CreateAnd(Result,
616 ConstantInt::get(ITy, UnknownBit));
618 // Shift the bit we're testing down to the lsb.
619 Result = Builder->CreateLShr(
620 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
622 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
623 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
624 Result->takeName(ICI);
625 return ReplaceInstUsesWith(CI, Result);
634 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
635 /// specified wider type and produce the same low bits. If not, return false.
637 /// If this function returns true, it can also return a non-zero number of bits
638 /// (in BitsToClear) which indicates that the value it computes is correct for
639 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
640 /// out. For example, to promote something like:
642 /// %B = trunc i64 %A to i32
643 /// %C = lshr i32 %B, 8
644 /// %E = zext i32 %C to i64
646 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
647 /// set to 8 to indicate that the promoted value needs to have bits 24-31
648 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
649 /// clear the top bits anyway, doing this has no extra cost.
651 /// This function works on both vectors and scalars.
652 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
654 if (isa<Constant>(V))
657 Instruction *I = dyn_cast<Instruction>(V);
658 if (!I) return false;
660 // If the input is a truncate from the destination type, we can trivially
662 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
665 // We can't extend or shrink something that has multiple uses: doing so would
666 // require duplicating the instruction in general, which isn't profitable.
667 if (!I->hasOneUse()) return false;
669 unsigned Opc = I->getOpcode(), Tmp;
671 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
672 case Instruction::SExt: // zext(sext(x)) -> sext(x).
673 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
675 case Instruction::And:
676 case Instruction::Or:
677 case Instruction::Xor:
678 case Instruction::Add:
679 case Instruction::Sub:
680 case Instruction::Mul:
681 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
682 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
684 // These can all be promoted if neither operand has 'bits to clear'.
685 if (BitsToClear == 0 && Tmp == 0)
688 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
689 // other side, BitsToClear is ok.
691 (Opc == Instruction::And || Opc == Instruction::Or ||
692 Opc == Instruction::Xor)) {
693 // We use MaskedValueIsZero here for generality, but the case we care
694 // about the most is constant RHS.
695 unsigned VSize = V->getType()->getScalarSizeInBits();
696 if (MaskedValueIsZero(I->getOperand(1),
697 APInt::getHighBitsSet(VSize, BitsToClear)))
701 // Otherwise, we don't know how to analyze this BitsToClear case yet.
704 case Instruction::Shl:
705 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
706 // upper bits we can reduce BitsToClear by the shift amount.
707 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
708 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
710 uint64_t ShiftAmt = Amt->getZExtValue();
711 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
715 case Instruction::LShr:
716 // We can promote lshr(x, cst) if we can promote x. This requires the
717 // ultimate 'and' to clear out the high zero bits we're clearing out though.
718 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
719 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
721 BitsToClear += Amt->getZExtValue();
722 if (BitsToClear > V->getType()->getScalarSizeInBits())
723 BitsToClear = V->getType()->getScalarSizeInBits();
726 // Cannot promote variable LSHR.
728 case Instruction::Select:
729 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
730 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
731 // TODO: If important, we could handle the case when the BitsToClear are
732 // known zero in the disagreeing side.
737 case Instruction::PHI: {
738 // We can change a phi if we can change all operands. Note that we never
739 // get into trouble with cyclic PHIs here because we only consider
740 // instructions with a single use.
741 PHINode *PN = cast<PHINode>(I);
742 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
744 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
745 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
746 // TODO: If important, we could handle the case when the BitsToClear
747 // are known zero in the disagreeing input.
753 // TODO: Can handle more cases here.
758 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
759 // If this zero extend is only used by a truncate, let the truncate be
760 // eliminated before we try to optimize this zext.
761 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
764 // If one of the common conversion will work, do it.
765 if (Instruction *Result = commonCastTransforms(CI))
768 // See if we can simplify any instructions used by the input whose sole
769 // purpose is to compute bits we don't care about.
770 if (SimplifyDemandedInstructionBits(CI))
773 Value *Src = CI.getOperand(0);
774 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
776 // Attempt to extend the entire input expression tree to the destination
777 // type. Only do this if the dest type is a simple type, don't convert the
778 // expression tree to something weird like i93 unless the source is also
780 unsigned BitsToClear;
781 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
782 CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
783 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
784 "Unreasonable BitsToClear");
786 // Okay, we can transform this! Insert the new expression now.
787 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
788 " to avoid zero extend: " << CI);
789 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
790 assert(Res->getType() == DestTy);
792 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
793 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
795 // If the high bits are already filled with zeros, just replace this
796 // cast with the result.
797 if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
798 DestBitSize-SrcBitsKept)))
799 return ReplaceInstUsesWith(CI, Res);
801 // We need to emit an AND to clear the high bits.
802 Constant *C = ConstantInt::get(Res->getType(),
803 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
804 return BinaryOperator::CreateAnd(Res, C);
807 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
808 // types and if the sizes are just right we can convert this into a logical
809 // 'and' which will be much cheaper than the pair of casts.
810 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
811 // TODO: Subsume this into EvaluateInDifferentType.
813 // Get the sizes of the types involved. We know that the intermediate type
814 // will be smaller than A or C, but don't know the relation between A and C.
815 Value *A = CSrc->getOperand(0);
816 unsigned SrcSize = A->getType()->getScalarSizeInBits();
817 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
818 unsigned DstSize = CI.getType()->getScalarSizeInBits();
819 // If we're actually extending zero bits, then if
820 // SrcSize < DstSize: zext(a & mask)
821 // SrcSize == DstSize: a & mask
822 // SrcSize > DstSize: trunc(a) & mask
823 if (SrcSize < DstSize) {
824 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
825 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
826 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
827 return new ZExtInst(And, CI.getType());
830 if (SrcSize == DstSize) {
831 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
832 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
835 if (SrcSize > DstSize) {
836 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
837 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
838 return BinaryOperator::CreateAnd(Trunc,
839 ConstantInt::get(Trunc->getType(),
844 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
845 return transformZExtICmp(ICI, CI);
847 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
848 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
849 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
850 // of the (zext icmp) will be transformed.
851 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
852 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
853 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
854 (transformZExtICmp(LHS, CI, false) ||
855 transformZExtICmp(RHS, CI, false))) {
856 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
857 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
858 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
862 // zext(trunc(X) & C) -> (X & zext(C)).
866 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
867 X->getType() == CI.getType())
868 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
870 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
872 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
873 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
874 X->getType() == CI.getType()) {
875 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
876 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
879 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
880 if (SrcI && SrcI->hasOneUse() &&
881 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
882 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
883 Value *New = Builder->CreateZExt(X, CI.getType());
884 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
890 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
891 /// in order to eliminate the icmp.
892 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
893 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
894 ICmpInst::Predicate Pred = ICI->getPredicate();
896 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
897 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
898 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
899 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
900 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
902 Value *Sh = ConstantInt::get(Op0->getType(),
903 Op0->getType()->getScalarSizeInBits()-1);
904 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
905 if (In->getType() != CI.getType())
906 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
908 if (Pred == ICmpInst::ICMP_SGT)
909 In = Builder->CreateNot(In, In->getName()+".not");
910 return ReplaceInstUsesWith(CI, In);
914 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
915 // If we know that only one bit of the LHS of the icmp can be set and we
916 // have an equality comparison with zero or a power of 2, we can transform
917 // the icmp and sext into bitwise/integer operations.
918 if (ICI->hasOneUse() &&
919 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
920 unsigned BitWidth = Op1C->getType()->getBitWidth();
921 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
922 ComputeMaskedBits(Op0, KnownZero, KnownOne);
924 APInt KnownZeroMask(~KnownZero);
925 if (KnownZeroMask.isPowerOf2()) {
926 Value *In = ICI->getOperand(0);
928 // If the icmp tests for a known zero bit we can constant fold it.
929 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
930 Value *V = Pred == ICmpInst::ICMP_NE ?
931 ConstantInt::getAllOnesValue(CI.getType()) :
932 ConstantInt::getNullValue(CI.getType());
933 return ReplaceInstUsesWith(CI, V);
936 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
937 // sext ((x & 2^n) == 0) -> (x >> n) - 1
938 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
939 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
940 // Perform a right shift to place the desired bit in the LSB.
942 In = Builder->CreateLShr(In,
943 ConstantInt::get(In->getType(), ShiftAmt));
945 // At this point "In" is either 1 or 0. Subtract 1 to turn
946 // {1, 0} -> {0, -1}.
947 In = Builder->CreateAdd(In,
948 ConstantInt::getAllOnesValue(In->getType()),
951 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
952 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
953 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
954 // Perform a left shift to place the desired bit in the MSB.
956 In = Builder->CreateShl(In,
957 ConstantInt::get(In->getType(), ShiftAmt));
959 // Distribute the bit over the whole bit width.
960 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
961 BitWidth - 1), "sext");
964 if (CI.getType() == In->getType())
965 return ReplaceInstUsesWith(CI, In);
966 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
974 /// CanEvaluateSExtd - Return true if we can take the specified value
975 /// and return it as type Ty without inserting any new casts and without
976 /// changing the value of the common low bits. This is used by code that tries
977 /// to promote integer operations to a wider types will allow us to eliminate
980 /// This function works on both vectors and scalars.
982 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
983 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
984 "Can't sign extend type to a smaller type");
985 // If this is a constant, it can be trivially promoted.
986 if (isa<Constant>(V))
989 Instruction *I = dyn_cast<Instruction>(V);
990 if (!I) return false;
992 // If this is a truncate from the dest type, we can trivially eliminate it.
993 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
996 // We can't extend or shrink something that has multiple uses: doing so would
997 // require duplicating the instruction in general, which isn't profitable.
998 if (!I->hasOneUse()) return false;
1000 switch (I->getOpcode()) {
1001 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1002 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1003 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1005 case Instruction::And:
1006 case Instruction::Or:
1007 case Instruction::Xor:
1008 case Instruction::Add:
1009 case Instruction::Sub:
1010 case Instruction::Mul:
1011 // These operators can all arbitrarily be extended if their inputs can.
1012 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1013 CanEvaluateSExtd(I->getOperand(1), Ty);
1015 //case Instruction::Shl: TODO
1016 //case Instruction::LShr: TODO
1018 case Instruction::Select:
1019 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1020 CanEvaluateSExtd(I->getOperand(2), Ty);
1022 case Instruction::PHI: {
1023 // We can change a phi if we can change all operands. Note that we never
1024 // get into trouble with cyclic PHIs here because we only consider
1025 // instructions with a single use.
1026 PHINode *PN = cast<PHINode>(I);
1027 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1028 if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1032 // TODO: Can handle more cases here.
1039 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1040 // If this sign extend is only used by a truncate, let the truncate be
1041 // eliminated before we try to optimize this sext.
1042 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1045 if (Instruction *I = commonCastTransforms(CI))
1048 // See if we can simplify any instructions used by the input whose sole
1049 // purpose is to compute bits we don't care about.
1050 if (SimplifyDemandedInstructionBits(CI))
1053 Value *Src = CI.getOperand(0);
1054 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1056 // Attempt to extend the entire input expression tree to the destination
1057 // type. Only do this if the dest type is a simple type, don't convert the
1058 // expression tree to something weird like i93 unless the source is also
1060 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1061 CanEvaluateSExtd(Src, DestTy)) {
1062 // Okay, we can transform this! Insert the new expression now.
1063 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1064 " to avoid sign extend: " << CI);
1065 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1066 assert(Res->getType() == DestTy);
1068 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1069 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1071 // If the high bits are already filled with sign bit, just replace this
1072 // cast with the result.
1073 if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1074 return ReplaceInstUsesWith(CI, Res);
1076 // We need to emit a shl + ashr to do the sign extend.
1077 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1078 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1082 // If this input is a trunc from our destination, then turn sext(trunc(x))
1084 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1085 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1086 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1087 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1089 // We need to emit a shl + ashr to do the sign extend.
1090 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1091 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1092 return BinaryOperator::CreateAShr(Res, ShAmt);
1095 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1096 return transformSExtICmp(ICI, CI);
1098 // If the input is a shl/ashr pair of a same constant, then this is a sign
1099 // extension from a smaller value. If we could trust arbitrary bitwidth
1100 // integers, we could turn this into a truncate to the smaller bit and then
1101 // use a sext for the whole extension. Since we don't, look deeper and check
1102 // for a truncate. If the source and dest are the same type, eliminate the
1103 // trunc and extend and just do shifts. For example, turn:
1104 // %a = trunc i32 %i to i8
1105 // %b = shl i8 %a, 6
1106 // %c = ashr i8 %b, 6
1107 // %d = sext i8 %c to i32
1109 // %a = shl i32 %i, 30
1110 // %d = ashr i32 %a, 30
1112 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1113 ConstantInt *BA = 0, *CA = 0;
1114 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1115 m_ConstantInt(CA))) &&
1116 BA == CA && A->getType() == CI.getType()) {
1117 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1118 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1119 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1120 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1121 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1122 return BinaryOperator::CreateAShr(A, ShAmtV);
1129 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1130 /// in the specified FP type without changing its value.
1131 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1133 APFloat F = CFP->getValueAPF();
1134 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1136 return ConstantFP::get(CFP->getContext(), F);
1140 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1141 /// through it until we get the source value.
1142 static Value *LookThroughFPExtensions(Value *V) {
1143 if (Instruction *I = dyn_cast<Instruction>(V))
1144 if (I->getOpcode() == Instruction::FPExt)
1145 return LookThroughFPExtensions(I->getOperand(0));
1147 // If this value is a constant, return the constant in the smallest FP type
1148 // that can accurately represent it. This allows us to turn
1149 // (float)((double)X+2.0) into x+2.0f.
1150 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1151 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1152 return V; // No constant folding of this.
1153 // See if the value can be truncated to half and then reextended.
1154 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1156 // See if the value can be truncated to float and then reextended.
1157 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1159 if (CFP->getType()->isDoubleTy())
1160 return V; // Won't shrink.
1161 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1163 // Don't try to shrink to various long double types.
1169 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1170 if (Instruction *I = commonCastTransforms(CI))
1172 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1173 // simpilify this expression to avoid one or more of the trunc/extend
1174 // operations if we can do so without changing the numerical results.
1176 // The exact manner in which the widths of the operands interact to limit
1177 // what we can and cannot do safely varies from operation to operation, and
1178 // is explained below in the various case statements.
1179 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1180 if (OpI && OpI->hasOneUse()) {
1181 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1182 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1183 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1184 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1185 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1186 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1187 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1188 switch (OpI->getOpcode()) {
1190 case Instruction::FAdd:
1191 case Instruction::FSub:
1192 // For addition and subtraction, the infinitely precise result can
1193 // essentially be arbitrarily wide; proving that double rounding
1194 // will not occur because the result of OpI is exact (as we will for
1195 // FMul, for example) is hopeless. However, we *can* nonetheless
1196 // frequently know that double rounding cannot occur (or that it is
1197 // innocuous) by taking advantage of the specific structure of
1198 // infinitely-precise results that admit double rounding.
1200 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1201 // to represent both sources, we can guarantee that the double
1202 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1203 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1204 // for proof of this fact).
1206 // Note: Figueroa does not consider the case where DstFormat !=
1207 // SrcFormat. It's possible (likely even!) that this analysis
1208 // could be tightened for those cases, but they are rare (the main
1209 // case of interest here is (float)((double)float + float)).
1210 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1211 if (LHSOrig->getType() != CI.getType())
1212 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1213 if (RHSOrig->getType() != CI.getType())
1214 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1216 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1217 RI->copyFastMathFlags(OpI);
1221 case Instruction::FMul:
1222 // For multiplication, the infinitely precise result has at most
1223 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1224 // that such a value can be exactly represented, then no double
1225 // rounding can possibly occur; we can safely perform the operation
1226 // in the destination format if it can represent both sources.
1227 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1228 if (LHSOrig->getType() != CI.getType())
1229 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1230 if (RHSOrig->getType() != CI.getType())
1231 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1233 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1234 RI->copyFastMathFlags(OpI);
1238 case Instruction::FDiv:
1239 // For division, we use again use the bound from Figueroa's
1240 // dissertation. I am entirely certain that this bound can be
1241 // tightened in the unbalanced operand case by an analysis based on
1242 // the diophantine rational approximation bound, but the well-known
1243 // condition used here is a good conservative first pass.
1244 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1245 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1246 if (LHSOrig->getType() != CI.getType())
1247 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1248 if (RHSOrig->getType() != CI.getType())
1249 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1251 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1252 RI->copyFastMathFlags(OpI);
1256 case Instruction::FRem:
1257 // Remainder is straightforward. Remainder is always exact, so the
1258 // type of OpI doesn't enter into things at all. We simply evaluate
1259 // in whichever source type is larger, then convert to the
1260 // destination type.
1261 if (LHSWidth < SrcWidth)
1262 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1263 else if (RHSWidth <= SrcWidth)
1264 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1265 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1266 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1267 RI->copyFastMathFlags(OpI);
1268 return CastInst::CreateFPCast(ExactResult, CI.getType());
1271 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1272 if (BinaryOperator::isFNeg(OpI)) {
1273 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1275 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1276 RI->copyFastMathFlags(OpI);
1281 // (fptrunc (select cond, R1, Cst)) -->
1282 // (select cond, (fptrunc R1), (fptrunc Cst))
1283 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1285 (isa<ConstantFP>(SI->getOperand(1)) ||
1286 isa<ConstantFP>(SI->getOperand(2)))) {
1287 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1289 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1291 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1294 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1296 switch (II->getIntrinsicID()) {
1298 case Intrinsic::fabs: {
1299 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1300 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1302 Type *IntrinsicType[] = { CI.getType() };
1303 Function *Overload =
1304 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1305 II->getIntrinsicID(), IntrinsicType);
1307 Value *Args[] = { InnerTrunc };
1308 return CallInst::Create(Overload, Args, II->getName());
1313 // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1314 // Note that we restrict this transformation based on
1315 // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because
1316 // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the
1317 // single-precision intrinsic can be expanded in the backend.
1318 CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1319 if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1320 (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) ||
1321 Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) &&
1322 Call->getNumArgOperands() == 1 &&
1323 Call->hasOneUse()) {
1324 CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1325 if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1326 CI.getType()->isFloatTy() &&
1327 Call->getType()->isDoubleTy() &&
1328 Arg->getType()->isDoubleTy() &&
1329 Arg->getOperand(0)->getType()->isFloatTy()) {
1330 Function *Callee = Call->getCalledFunction();
1331 Module *M = CI.getParent()->getParent()->getParent();
1332 Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ?
1333 Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) :
1334 M->getOrInsertFunction("sqrtf", Callee->getAttributes(),
1335 Builder->getFloatTy(), Builder->getFloatTy(),
1337 CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1339 ret->setAttributes(Callee->getAttributes());
1342 // Remove the old Call. With -fmath-errno, it won't get marked readnone.
1343 ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1344 EraseInstFromFunction(*Call);
1352 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1353 return commonCastTransforms(CI);
1356 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1357 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1359 return commonCastTransforms(FI);
1361 // fptoui(uitofp(X)) --> X
1362 // fptoui(sitofp(X)) --> X
1363 // This is safe if the intermediate type has enough bits in its mantissa to
1364 // accurately represent all values of X. For example, do not do this with
1365 // i64->float->i64. This is also safe for sitofp case, because any negative
1366 // 'X' value would cause an undefined result for the fptoui.
1367 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1368 OpI->getOperand(0)->getType() == FI.getType() &&
1369 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1370 OpI->getType()->getFPMantissaWidth())
1371 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1373 return commonCastTransforms(FI);
1376 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1377 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1379 return commonCastTransforms(FI);
1381 // fptosi(sitofp(X)) --> X
1382 // fptosi(uitofp(X)) --> X
1383 // This is safe if the intermediate type has enough bits in its mantissa to
1384 // accurately represent all values of X. For example, do not do this with
1385 // i64->float->i64. This is also safe for sitofp case, because any negative
1386 // 'X' value would cause an undefined result for the fptoui.
1387 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1388 OpI->getOperand(0)->getType() == FI.getType() &&
1389 (int)FI.getType()->getScalarSizeInBits() <=
1390 OpI->getType()->getFPMantissaWidth())
1391 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1393 return commonCastTransforms(FI);
1396 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1397 return commonCastTransforms(CI);
1400 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1401 return commonCastTransforms(CI);
1404 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1405 // If the source integer type is not the intptr_t type for this target, do a
1406 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1407 // cast to be exposed to other transforms.
1410 unsigned AS = CI.getAddressSpace();
1411 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1412 DL->getPointerSizeInBits(AS)) {
1413 Type *Ty = DL->getIntPtrType(CI.getContext(), AS);
1414 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1415 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1417 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1418 return new IntToPtrInst(P, CI.getType());
1422 if (Instruction *I = commonCastTransforms(CI))
1428 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1429 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1430 Value *Src = CI.getOperand(0);
1432 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1433 // If casting the result of a getelementptr instruction with no offset, turn
1434 // this into a cast of the original pointer!
1435 if (GEP->hasAllZeroIndices()) {
1436 // Changing the cast operand is usually not a good idea but it is safe
1437 // here because the pointer operand is being replaced with another
1438 // pointer operand so the opcode doesn't need to change.
1440 CI.setOperand(0, GEP->getOperand(0));
1445 return commonCastTransforms(CI);
1447 // If the GEP has a single use, and the base pointer is a bitcast, and the
1448 // GEP computes a constant offset, see if we can convert these three
1449 // instructions into fewer. This typically happens with unions and other
1450 // non-type-safe code.
1451 unsigned AS = GEP->getPointerAddressSpace();
1452 unsigned OffsetBits = DL->getPointerSizeInBits(AS);
1453 APInt Offset(OffsetBits, 0);
1454 BitCastInst *BCI = dyn_cast<BitCastInst>(GEP->getOperand(0));
1455 if (GEP->hasOneUse() &&
1457 GEP->accumulateConstantOffset(*DL, Offset)) {
1458 // Get the base pointer input of the bitcast, and the type it points to.
1459 Value *OrigBase = BCI->getOperand(0);
1460 SmallVector<Value*, 8> NewIndices;
1461 if (FindElementAtOffset(OrigBase->getType(),
1462 Offset.getSExtValue(),
1464 // If we were able to index down into an element, create the GEP
1465 // and bitcast the result. This eliminates one bitcast, potentially
1467 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1468 Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1469 Builder->CreateGEP(OrigBase, NewIndices);
1470 NGEP->takeName(GEP);
1472 if (isa<BitCastInst>(CI))
1473 return new BitCastInst(NGEP, CI.getType());
1474 assert(isa<PtrToIntInst>(CI));
1475 return new PtrToIntInst(NGEP, CI.getType());
1480 return commonCastTransforms(CI);
1483 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1484 // If the destination integer type is not the intptr_t type for this target,
1485 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1486 // to be exposed to other transforms.
1489 return commonPointerCastTransforms(CI);
1491 Type *Ty = CI.getType();
1492 unsigned AS = CI.getPointerAddressSpace();
1494 if (Ty->getScalarSizeInBits() == DL->getPointerSizeInBits(AS))
1495 return commonPointerCastTransforms(CI);
1497 Type *PtrTy = DL->getIntPtrType(CI.getContext(), AS);
1498 if (Ty->isVectorTy()) // Handle vectors of pointers.
1499 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1501 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1502 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1505 /// OptimizeVectorResize - This input value (which is known to have vector type)
1506 /// is being zero extended or truncated to the specified vector type. Try to
1507 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1509 /// The source and destination vector types may have different element types.
1510 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1512 // We can only do this optimization if the output is a multiple of the input
1513 // element size, or the input is a multiple of the output element size.
1514 // Convert the input type to have the same element type as the output.
1515 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1517 if (SrcTy->getElementType() != DestTy->getElementType()) {
1518 // The input types don't need to be identical, but for now they must be the
1519 // same size. There is no specific reason we couldn't handle things like
1520 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1522 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1523 DestTy->getElementType()->getPrimitiveSizeInBits())
1526 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1527 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1530 // Now that the element types match, get the shuffle mask and RHS of the
1531 // shuffle to use, which depends on whether we're increasing or decreasing the
1532 // size of the input.
1533 SmallVector<uint32_t, 16> ShuffleMask;
1536 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1537 // If we're shrinking the number of elements, just shuffle in the low
1538 // elements from the input and use undef as the second shuffle input.
1539 V2 = UndefValue::get(SrcTy);
1540 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1541 ShuffleMask.push_back(i);
1544 // If we're increasing the number of elements, shuffle in all of the
1545 // elements from InVal and fill the rest of the result elements with zeros
1546 // from a constant zero.
1547 V2 = Constant::getNullValue(SrcTy);
1548 unsigned SrcElts = SrcTy->getNumElements();
1549 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1550 ShuffleMask.push_back(i);
1552 // The excess elements reference the first element of the zero input.
1553 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1554 ShuffleMask.push_back(SrcElts);
1557 return new ShuffleVectorInst(InVal, V2,
1558 ConstantDataVector::get(V2->getContext(),
1562 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1563 return Value % Ty->getPrimitiveSizeInBits() == 0;
1566 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1567 return Value / Ty->getPrimitiveSizeInBits();
1570 /// CollectInsertionElements - V is a value which is inserted into a vector of
1571 /// VecEltTy. Look through the value to see if we can decompose it into
1572 /// insertions into the vector. See the example in the comment for
1573 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1574 /// The type of V is always a non-zero multiple of VecEltTy's size.
1575 /// Shift is the number of bits between the lsb of V and the lsb of
1578 /// This returns false if the pattern can't be matched or true if it can,
1579 /// filling in Elements with the elements found here.
1580 static bool CollectInsertionElements(Value *V, unsigned Shift,
1581 SmallVectorImpl<Value*> &Elements,
1582 Type *VecEltTy, InstCombiner &IC) {
1583 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1584 "Shift should be a multiple of the element type size");
1586 // Undef values never contribute useful bits to the result.
1587 if (isa<UndefValue>(V)) return true;
1589 // If we got down to a value of the right type, we win, try inserting into the
1591 if (V->getType() == VecEltTy) {
1592 // Inserting null doesn't actually insert any elements.
1593 if (Constant *C = dyn_cast<Constant>(V))
1594 if (C->isNullValue())
1597 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1598 if (IC.getDataLayout()->isBigEndian())
1599 ElementIndex = Elements.size() - ElementIndex - 1;
1601 // Fail if multiple elements are inserted into this slot.
1602 if (Elements[ElementIndex] != 0)
1605 Elements[ElementIndex] = V;
1609 if (Constant *C = dyn_cast<Constant>(V)) {
1610 // Figure out the # elements this provides, and bitcast it or slice it up
1612 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1614 // If the constant is the size of a vector element, we just need to bitcast
1615 // it to the right type so it gets properly inserted.
1617 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1618 Shift, Elements, VecEltTy, IC);
1620 // Okay, this is a constant that covers multiple elements. Slice it up into
1621 // pieces and insert each element-sized piece into the vector.
1622 if (!isa<IntegerType>(C->getType()))
1623 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1624 C->getType()->getPrimitiveSizeInBits()));
1625 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1626 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1628 for (unsigned i = 0; i != NumElts; ++i) {
1629 unsigned ShiftI = Shift+i*ElementSize;
1630 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1632 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1633 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy, IC))
1639 if (!V->hasOneUse()) return false;
1641 Instruction *I = dyn_cast<Instruction>(V);
1642 if (I == 0) return false;
1643 switch (I->getOpcode()) {
1644 default: return false; // Unhandled case.
1645 case Instruction::BitCast:
1646 return CollectInsertionElements(I->getOperand(0), Shift,
1647 Elements, VecEltTy, IC);
1648 case Instruction::ZExt:
1649 if (!isMultipleOfTypeSize(
1650 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1653 return CollectInsertionElements(I->getOperand(0), Shift,
1654 Elements, VecEltTy, IC);
1655 case Instruction::Or:
1656 return CollectInsertionElements(I->getOperand(0), Shift,
1657 Elements, VecEltTy, IC) &&
1658 CollectInsertionElements(I->getOperand(1), Shift,
1659 Elements, VecEltTy, IC);
1660 case Instruction::Shl: {
1661 // Must be shifting by a constant that is a multiple of the element size.
1662 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1663 if (CI == 0) return false;
1664 Shift += CI->getZExtValue();
1665 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1666 return CollectInsertionElements(I->getOperand(0), Shift,
1667 Elements, VecEltTy, IC);
1674 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1675 /// may be doing shifts and ors to assemble the elements of the vector manually.
1676 /// Try to rip the code out and replace it with insertelements. This is to
1677 /// optimize code like this:
1679 /// %tmp37 = bitcast float %inc to i32
1680 /// %tmp38 = zext i32 %tmp37 to i64
1681 /// %tmp31 = bitcast float %inc5 to i32
1682 /// %tmp32 = zext i32 %tmp31 to i64
1683 /// %tmp33 = shl i64 %tmp32, 32
1684 /// %ins35 = or i64 %tmp33, %tmp38
1685 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1687 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1688 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1690 // We need to know the target byte order to perform this optimization.
1691 if (!IC.getDataLayout()) return 0;
1693 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1694 Value *IntInput = CI.getOperand(0);
1696 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1697 if (!CollectInsertionElements(IntInput, 0, Elements,
1698 DestVecTy->getElementType(), IC))
1701 // If we succeeded, we know that all of the element are specified by Elements
1702 // or are zero if Elements has a null entry. Recast this as a set of
1704 Value *Result = Constant::getNullValue(CI.getType());
1705 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1706 if (Elements[i] == 0) continue; // Unset element.
1708 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1709 IC.Builder->getInt32(i));
1716 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1717 /// bitcast. The various long double bitcasts can't get in here.
1718 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1719 // We need to know the target byte order to perform this optimization.
1720 if (!IC.getDataLayout()) return 0;
1722 Value *Src = CI.getOperand(0);
1723 Type *DestTy = CI.getType();
1725 // If this is a bitcast from int to float, check to see if the int is an
1726 // extraction from a vector.
1727 Value *VecInput = 0;
1728 // bitcast(trunc(bitcast(somevector)))
1729 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1730 isa<VectorType>(VecInput->getType())) {
1731 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1732 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1734 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1735 // If the element type of the vector doesn't match the result type,
1736 // bitcast it to be a vector type we can extract from.
1737 if (VecTy->getElementType() != DestTy) {
1738 VecTy = VectorType::get(DestTy,
1739 VecTy->getPrimitiveSizeInBits() / DestWidth);
1740 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1744 if (IC.getDataLayout()->isBigEndian())
1745 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1746 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1750 // bitcast(trunc(lshr(bitcast(somevector), cst))
1751 ConstantInt *ShAmt = 0;
1752 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1753 m_ConstantInt(ShAmt)))) &&
1754 isa<VectorType>(VecInput->getType())) {
1755 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1756 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1757 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1758 ShAmt->getZExtValue() % DestWidth == 0) {
1759 // If the element type of the vector doesn't match the result type,
1760 // bitcast it to be a vector type we can extract from.
1761 if (VecTy->getElementType() != DestTy) {
1762 VecTy = VectorType::get(DestTy,
1763 VecTy->getPrimitiveSizeInBits() / DestWidth);
1764 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1767 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1768 if (IC.getDataLayout()->isBigEndian())
1769 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1770 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1776 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1777 // If the operands are integer typed then apply the integer transforms,
1778 // otherwise just apply the common ones.
1779 Value *Src = CI.getOperand(0);
1780 Type *SrcTy = Src->getType();
1781 Type *DestTy = CI.getType();
1783 // Get rid of casts from one type to the same type. These are useless and can
1784 // be replaced by the operand.
1785 if (DestTy == Src->getType())
1786 return ReplaceInstUsesWith(CI, Src);
1788 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1789 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1790 Type *DstElTy = DstPTy->getElementType();
1791 Type *SrcElTy = SrcPTy->getElementType();
1793 // If we are casting a alloca to a pointer to a type of the same
1794 // size, rewrite the allocation instruction to allocate the "right" type.
1795 // There is no need to modify malloc calls because it is their bitcast that
1796 // needs to be cleaned up.
1797 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1798 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1801 // If the source and destination are pointers, and this cast is equivalent
1802 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1803 // This can enhance SROA and other transforms that want type-safe pointers.
1804 Constant *ZeroUInt =
1805 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1806 unsigned NumZeros = 0;
1807 while (SrcElTy != DstElTy &&
1808 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1809 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1810 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1814 // If we found a path from the src to dest, create the getelementptr now.
1815 if (SrcElTy == DstElTy) {
1816 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1817 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1821 // Try to optimize int -> float bitcasts.
1822 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1823 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1826 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1827 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1828 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1829 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1830 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1831 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1834 if (isa<IntegerType>(SrcTy)) {
1835 // If this is a cast from an integer to vector, check to see if the input
1836 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1837 // the casts with a shuffle and (potentially) a bitcast.
1838 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1839 CastInst *SrcCast = cast<CastInst>(Src);
1840 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1841 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1842 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1843 cast<VectorType>(DestTy), *this))
1847 // If the input is an 'or' instruction, we may be doing shifts and ors to
1848 // assemble the elements of the vector manually. Try to rip the code out
1849 // and replace it with insertelements.
1850 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1851 return ReplaceInstUsesWith(CI, V);
1855 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1856 if (SrcVTy->getNumElements() == 1) {
1857 // If our destination is not a vector, then make this a straight
1858 // scalar-scalar cast.
1859 if (!DestTy->isVectorTy()) {
1861 Builder->CreateExtractElement(Src,
1862 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1863 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1866 // Otherwise, see if our source is an insert. If so, then use the scalar
1867 // component directly.
1868 if (InsertElementInst *IEI =
1869 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1870 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1875 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1876 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1877 // a bitcast to a vector with the same # elts.
1878 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1879 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1880 SVI->getType()->getNumElements() ==
1881 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1883 // If either of the operands is a cast from CI.getType(), then
1884 // evaluating the shuffle in the casted destination's type will allow
1885 // us to eliminate at least one cast.
1886 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1887 Tmp->getOperand(0)->getType() == DestTy) ||
1888 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1889 Tmp->getOperand(0)->getType() == DestTy)) {
1890 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1891 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1892 // Return a new shuffle vector. Use the same element ID's, as we
1893 // know the vector types match #elts.
1894 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1899 if (SrcTy->isPointerTy())
1900 return commonPointerCastTransforms(CI);
1901 return commonCastTransforms(CI);
1904 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1905 return commonPointerCastTransforms(CI);