1 //===- InstCombineCasts.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for cast operations.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/PatternMatch.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
20 using namespace PatternMatch;
22 #define DEBUG_TYPE "instcombine"
24 /// 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 PointerType *PTy = cast<PointerType>(CI.getType());
85 BuilderTy AllocaBuilder(*Builder);
86 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
88 // Get the type really allocated and the type casted to.
89 Type *AllocElTy = AI.getAllocatedType();
90 Type *CastElTy = PTy->getElementType();
91 if (!AllocElTy->isSized() || !CastElTy->isSized()) return nullptr;
93 unsigned AllocElTyAlign = DL.getABITypeAlignment(AllocElTy);
94 unsigned CastElTyAlign = DL.getABITypeAlignment(CastElTy);
95 if (CastElTyAlign < AllocElTyAlign) return nullptr;
97 // If the allocation has multiple uses, only promote it if we are strictly
98 // increasing the alignment of the resultant allocation. If we keep it the
99 // same, we open the door to infinite loops of various kinds.
100 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return nullptr;
102 uint64_t AllocElTySize = DL.getTypeAllocSize(AllocElTy);
103 uint64_t CastElTySize = DL.getTypeAllocSize(CastElTy);
104 if (CastElTySize == 0 || AllocElTySize == 0) return nullptr;
106 // If the allocation has multiple uses, only promote it if we're not
107 // shrinking the amount of memory being allocated.
108 uint64_t AllocElTyStoreSize = DL.getTypeStoreSize(AllocElTy);
109 uint64_t CastElTyStoreSize = DL.getTypeStoreSize(CastElTy);
110 if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return nullptr;
112 // See if we can satisfy the modulus by pulling a scale out of the array
114 unsigned ArraySizeScale;
115 uint64_t ArrayOffset;
116 Value *NumElements = // See if the array size is a decomposable linear expr.
117 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
119 // If we can now satisfy the modulus, by using a non-1 scale, we really can
121 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
122 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return nullptr;
124 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
125 Value *Amt = nullptr;
129 Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
130 // Insert before the alloca, not before the cast.
131 Amt = AllocaBuilder.CreateMul(Amt, NumElements);
134 if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
135 Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
137 Amt = AllocaBuilder.CreateAdd(Amt, Off);
140 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
141 New->setAlignment(AI.getAlignment());
143 New->setUsedWithInAlloca(AI.isUsedWithInAlloca());
145 // If the allocation has multiple real uses, insert a cast and change all
146 // things that used it to use the new cast. This will also hack on CI, but it
148 if (!AI.hasOneUse()) {
149 // New is the allocation instruction, pointer typed. AI is the original
150 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
151 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
152 ReplaceInstUsesWith(AI, NewCast);
154 return ReplaceInstUsesWith(CI, New);
157 /// EvaluateInDifferentType - Given an expression that
158 /// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
159 /// insert the code to evaluate the expression.
160 Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
162 if (Constant *C = dyn_cast<Constant>(V)) {
163 C = ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
164 // If we got a constantexpr back, try to simplify it with DL info.
165 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
166 C = ConstantFoldConstantExpression(CE, DL, TLI);
170 // Otherwise, it must be an instruction.
171 Instruction *I = cast<Instruction>(V);
172 Instruction *Res = nullptr;
173 unsigned Opc = I->getOpcode();
175 case Instruction::Add:
176 case Instruction::Sub:
177 case Instruction::Mul:
178 case Instruction::And:
179 case Instruction::Or:
180 case Instruction::Xor:
181 case Instruction::AShr:
182 case Instruction::LShr:
183 case Instruction::Shl:
184 case Instruction::UDiv:
185 case Instruction::URem: {
186 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
187 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
188 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
191 case Instruction::Trunc:
192 case Instruction::ZExt:
193 case Instruction::SExt:
194 // If the source type of the cast is the type we're trying for then we can
195 // just return the source. There's no need to insert it because it is not
197 if (I->getOperand(0)->getType() == Ty)
198 return I->getOperand(0);
200 // Otherwise, must be the same type of cast, so just reinsert a new one.
201 // This also handles the case of zext(trunc(x)) -> zext(x).
202 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
203 Opc == Instruction::SExt);
205 case Instruction::Select: {
206 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
207 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
208 Res = SelectInst::Create(I->getOperand(0), True, False);
211 case Instruction::PHI: {
212 PHINode *OPN = cast<PHINode>(I);
213 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
214 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
216 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
217 NPN->addIncoming(V, OPN->getIncomingBlock(i));
223 // TODO: Can handle more cases here.
224 llvm_unreachable("Unreachable!");
228 return InsertNewInstWith(Res, *I);
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(const CastInst *CI, ///< First cast instruction
236 unsigned opcode, ///< Opcode for the second cast
237 Type *DstTy, ///< Target type for the second cast
238 const DataLayout &DL) {
239 Type *SrcTy = CI->getOperand(0)->getType(); // A from above
240 Type *MidTy = CI->getType(); // B from above
242 // Get the opcodes of the two Cast instructions
243 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
244 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
246 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
248 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
250 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
251 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
252 DstTy, SrcIntPtrTy, MidIntPtrTy,
255 // We don't want to form an inttoptr or ptrtoint that converts to an integer
256 // type that differs from the pointer size.
257 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
258 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
261 return Instruction::CastOps(Res);
264 /// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
265 /// results in any code being generated and is interesting to optimize out. If
266 /// the cast can be eliminated by some other simple transformation, we prefer
267 /// to do the simplification first.
268 bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
270 // Noop casts and casts of constants should be eliminated trivially.
271 if (V->getType() == Ty || isa<Constant>(V)) return false;
273 // If this is another cast that can be eliminated, we prefer to have it
275 if (const CastInst *CI = dyn_cast<CastInst>(V))
276 if (isEliminableCastPair(CI, opc, Ty, DL))
279 // If this is a vector sext from a compare, then we don't want to break the
280 // idiom where each element of the extended vector is either zero or all ones.
281 if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
288 /// @brief Implement the transforms common to all CastInst visitors.
289 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
290 Value *Src = CI.getOperand(0);
292 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
294 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
295 if (Instruction::CastOps opc =
296 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), DL)) {
297 // The first cast (CSrc) is eliminable so we need to fix up or replace
298 // the second cast (CI). CSrc will then have a good chance of being dead.
299 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
303 // If we are casting a select then fold the cast into the select
304 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
305 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
308 // If we are casting a PHI then fold the cast into the PHI
309 if (isa<PHINode>(Src)) {
310 // We don't do this if this would create a PHI node with an illegal type if
311 // it is currently legal.
312 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
313 ShouldChangeType(CI.getType(), Src->getType()))
314 if (Instruction *NV = FoldOpIntoPhi(CI))
321 /// CanEvaluateTruncated - Return true if we can evaluate the specified
322 /// expression tree as type Ty instead of its larger type, and arrive with the
323 /// same value. This is used by code that tries to eliminate truncates.
325 /// Ty will always be a type smaller than V. We should return true if trunc(V)
326 /// can be computed by computing V in the smaller type. If V is an instruction,
327 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
328 /// makes sense if x and y can be efficiently truncated.
330 /// This function works on both vectors and scalars.
332 static bool CanEvaluateTruncated(Value *V, Type *Ty, InstCombiner &IC,
334 // We can always evaluate constants in another type.
335 if (isa<Constant>(V))
338 Instruction *I = dyn_cast<Instruction>(V);
339 if (!I) return false;
341 Type *OrigTy = V->getType();
343 // If this is an extension from the dest type, we can eliminate it, even if it
344 // has multiple uses.
345 if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
346 I->getOperand(0)->getType() == Ty)
349 // We can't extend or shrink something that has multiple uses: doing so would
350 // require duplicating the instruction in general, which isn't profitable.
351 if (!I->hasOneUse()) return false;
353 unsigned Opc = I->getOpcode();
355 case Instruction::Add:
356 case Instruction::Sub:
357 case Instruction::Mul:
358 case Instruction::And:
359 case Instruction::Or:
360 case Instruction::Xor:
361 // These operators can all arbitrarily be extended or truncated.
362 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
363 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
365 case Instruction::UDiv:
366 case Instruction::URem: {
367 // UDiv and URem can be truncated if all the truncated bits are zero.
368 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
369 uint32_t BitWidth = Ty->getScalarSizeInBits();
370 if (BitWidth < OrigBitWidth) {
371 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
372 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
373 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
374 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
375 CanEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
380 case Instruction::Shl:
381 // If we are truncating the result of this SHL, and if it's a shift of a
382 // constant amount, we can always perform a SHL in a smaller type.
383 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
384 uint32_t BitWidth = Ty->getScalarSizeInBits();
385 if (CI->getLimitedValue(BitWidth) < BitWidth)
386 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
389 case Instruction::LShr:
390 // If this is a truncate of a logical shr, we can truncate it to a smaller
391 // lshr iff we know that the bits we would otherwise be shifting in are
393 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
394 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
395 uint32_t BitWidth = Ty->getScalarSizeInBits();
396 if (IC.MaskedValueIsZero(I->getOperand(0),
397 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth), 0, CxtI) &&
398 CI->getLimitedValue(BitWidth) < BitWidth) {
399 return CanEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI);
403 case Instruction::Trunc:
404 // trunc(trunc(x)) -> trunc(x)
406 case Instruction::ZExt:
407 case Instruction::SExt:
408 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
409 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
411 case Instruction::Select: {
412 SelectInst *SI = cast<SelectInst>(I);
413 return CanEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
414 CanEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
416 case Instruction::PHI: {
417 // We can change a phi if we can change all operands. Note that we never
418 // get into trouble with cyclic PHIs here because we only consider
419 // instructions with a single use.
420 PHINode *PN = cast<PHINode>(I);
421 for (Value *IncValue : PN->incoming_values())
422 if (!CanEvaluateTruncated(IncValue, Ty, IC, CxtI))
427 // TODO: Can handle more cases here.
434 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
435 if (Instruction *Result = commonCastTransforms(CI))
438 // Test if the trunc is the user of a select which is part of a
439 // minimum or maximum operation. If so, don't do any more simplification.
440 // Even simplifying demanded bits can break the canonical form of a
443 if (SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0)))
444 if (matchSelectPattern(SI, LHS, RHS) != SPF_UNKNOWN)
447 // See if we can simplify any instructions used by the input whose sole
448 // purpose is to compute bits we don't care about.
449 if (SimplifyDemandedInstructionBits(CI))
452 Value *Src = CI.getOperand(0);
453 Type *DestTy = CI.getType(), *SrcTy = Src->getType();
455 // Attempt to truncate the entire input expression tree to the destination
456 // type. Only do this if the dest type is a simple type, don't convert the
457 // expression tree to something weird like i93 unless the source is also
459 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
460 CanEvaluateTruncated(Src, DestTy, *this, &CI)) {
462 // If this cast is a truncate, evaluting in a different type always
463 // eliminates the cast, so it is always a win.
464 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
465 " to avoid cast: " << CI << '\n');
466 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
467 assert(Res->getType() == DestTy);
468 return ReplaceInstUsesWith(CI, Res);
471 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
472 if (DestTy->getScalarSizeInBits() == 1) {
473 Constant *One = ConstantInt::get(Src->getType(), 1);
474 Src = Builder->CreateAnd(Src, One);
475 Value *Zero = Constant::getNullValue(Src->getType());
476 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
479 // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
480 Value *A = nullptr; ConstantInt *Cst = nullptr;
481 if (Src->hasOneUse() &&
482 match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
483 // We have three types to worry about here, the type of A, the source of
484 // the truncate (MidSize), and the destination of the truncate. We know that
485 // ASize < MidSize and MidSize > ResultSize, but don't know the relation
486 // between ASize and ResultSize.
487 unsigned ASize = A->getType()->getPrimitiveSizeInBits();
489 // If the shift amount is larger than the size of A, then the result is
490 // known to be zero because all the input bits got shifted out.
491 if (Cst->getZExtValue() >= ASize)
492 return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
494 // Since we're doing an lshr and a zero extend, and know that the shift
495 // amount is smaller than ASize, it is always safe to do the shift in A's
496 // type, then zero extend or truncate to the result.
497 Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
498 Shift->takeName(Src);
499 return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
502 // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
503 // type isn't non-native.
504 if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
505 ShouldChangeType(Src->getType(), CI.getType()) &&
506 match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
507 Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
508 return BinaryOperator::CreateAnd(NewTrunc,
509 ConstantExpr::getTrunc(Cst, CI.getType()));
515 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
516 /// in order to eliminate the icmp.
517 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
519 // If we are just checking for a icmp eq of a single bit and zext'ing it
520 // to an integer, then shift the bit to the appropriate place and then
521 // cast to integer to avoid the comparison.
522 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
523 const APInt &Op1CV = Op1C->getValue();
525 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
526 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
527 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
528 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
529 if (!DoXform) return ICI;
531 Value *In = ICI->getOperand(0);
532 Value *Sh = ConstantInt::get(In->getType(),
533 In->getType()->getScalarSizeInBits()-1);
534 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
535 if (In->getType() != CI.getType())
536 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
538 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
539 Constant *One = ConstantInt::get(In->getType(), 1);
540 In = Builder->CreateXor(In, One, In->getName()+".not");
543 return ReplaceInstUsesWith(CI, In);
546 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
547 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
548 // zext (X == 1) to i32 --> X iff X has only the low bit set.
549 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
550 // zext (X != 0) to i32 --> X iff X has only the low bit set.
551 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
552 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
553 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
554 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
555 // This only works for EQ and NE
557 // If Op1C some other power of two, convert:
558 uint32_t BitWidth = Op1C->getType()->getBitWidth();
559 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
560 computeKnownBits(ICI->getOperand(0), KnownZero, KnownOne, 0, &CI);
562 APInt KnownZeroMask(~KnownZero);
563 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
564 if (!DoXform) return ICI;
566 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
567 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
568 // (X&4) == 2 --> false
569 // (X&4) != 2 --> true
570 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
572 Res = ConstantExpr::getZExt(Res, CI.getType());
573 return ReplaceInstUsesWith(CI, Res);
576 uint32_t ShiftAmt = KnownZeroMask.logBase2();
577 Value *In = ICI->getOperand(0);
579 // Perform a logical shr by shiftamt.
580 // Insert the shift to put the result in the low bit.
581 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
582 In->getName()+".lobit");
585 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
586 Constant *One = ConstantInt::get(In->getType(), 1);
587 In = Builder->CreateXor(In, One);
590 if (CI.getType() == In->getType())
591 return ReplaceInstUsesWith(CI, In);
592 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
597 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
598 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
599 // may lead to additional simplifications.
600 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
601 if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
602 uint32_t BitWidth = ITy->getBitWidth();
603 Value *LHS = ICI->getOperand(0);
604 Value *RHS = ICI->getOperand(1);
606 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
607 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
608 computeKnownBits(LHS, KnownZeroLHS, KnownOneLHS, 0, &CI);
609 computeKnownBits(RHS, KnownZeroRHS, KnownOneRHS, 0, &CI);
611 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
612 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
613 APInt UnknownBit = ~KnownBits;
614 if (UnknownBit.countPopulation() == 1) {
615 if (!DoXform) return ICI;
617 Value *Result = Builder->CreateXor(LHS, RHS);
619 // Mask off any bits that are set and won't be shifted away.
620 if (KnownOneLHS.uge(UnknownBit))
621 Result = Builder->CreateAnd(Result,
622 ConstantInt::get(ITy, UnknownBit));
624 // Shift the bit we're testing down to the lsb.
625 Result = Builder->CreateLShr(
626 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
628 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
629 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
630 Result->takeName(ICI);
631 return ReplaceInstUsesWith(CI, Result);
640 /// CanEvaluateZExtd - Determine if the specified value can be computed in the
641 /// specified wider type and produce the same low bits. If not, return false.
643 /// If this function returns true, it can also return a non-zero number of bits
644 /// (in BitsToClear) which indicates that the value it computes is correct for
645 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
646 /// out. For example, to promote something like:
648 /// %B = trunc i64 %A to i32
649 /// %C = lshr i32 %B, 8
650 /// %E = zext i32 %C to i64
652 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
653 /// set to 8 to indicate that the promoted value needs to have bits 24-31
654 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
655 /// clear the top bits anyway, doing this has no extra cost.
657 /// This function works on both vectors and scalars.
658 static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
659 InstCombiner &IC, Instruction *CxtI) {
661 if (isa<Constant>(V))
664 Instruction *I = dyn_cast<Instruction>(V);
665 if (!I) return false;
667 // If the input is a truncate from the destination type, we can trivially
669 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
672 // We can't extend or shrink something that has multiple uses: doing so would
673 // require duplicating the instruction in general, which isn't profitable.
674 if (!I->hasOneUse()) return false;
676 unsigned Opc = I->getOpcode(), Tmp;
678 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
679 case Instruction::SExt: // zext(sext(x)) -> sext(x).
680 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
682 case Instruction::And:
683 case Instruction::Or:
684 case Instruction::Xor:
685 case Instruction::Add:
686 case Instruction::Sub:
687 case Instruction::Mul:
688 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
689 !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
691 // These can all be promoted if neither operand has 'bits to clear'.
692 if (BitsToClear == 0 && Tmp == 0)
695 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
696 // other side, BitsToClear is ok.
698 (Opc == Instruction::And || Opc == Instruction::Or ||
699 Opc == Instruction::Xor)) {
700 // We use MaskedValueIsZero here for generality, but the case we care
701 // about the most is constant RHS.
702 unsigned VSize = V->getType()->getScalarSizeInBits();
703 if (IC.MaskedValueIsZero(I->getOperand(1),
704 APInt::getHighBitsSet(VSize, BitsToClear),
709 // Otherwise, we don't know how to analyze this BitsToClear case yet.
712 case Instruction::Shl:
713 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
714 // upper bits we can reduce BitsToClear by the shift amount.
715 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
716 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
718 uint64_t ShiftAmt = Amt->getZExtValue();
719 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
723 case Instruction::LShr:
724 // We can promote lshr(x, cst) if we can promote x. This requires the
725 // ultimate 'and' to clear out the high zero bits we're clearing out though.
726 if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
727 if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
729 BitsToClear += Amt->getZExtValue();
730 if (BitsToClear > V->getType()->getScalarSizeInBits())
731 BitsToClear = V->getType()->getScalarSizeInBits();
734 // Cannot promote variable LSHR.
736 case Instruction::Select:
737 if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
738 !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
739 // TODO: If important, we could handle the case when the BitsToClear are
740 // known zero in the disagreeing side.
745 case Instruction::PHI: {
746 // We can change a phi if we can change all operands. Note that we never
747 // get into trouble with cyclic PHIs here because we only consider
748 // instructions with a single use.
749 PHINode *PN = cast<PHINode>(I);
750 if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
752 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
753 if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
754 // TODO: If important, we could handle the case when the BitsToClear
755 // are known zero in the disagreeing input.
761 // TODO: Can handle more cases here.
766 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
767 // If this zero extend is only used by a truncate, let the truncate be
768 // eliminated before we try to optimize this zext.
769 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
772 // If one of the common conversion will work, do it.
773 if (Instruction *Result = commonCastTransforms(CI))
776 // See if we can simplify any instructions used by the input whose sole
777 // purpose is to compute bits we don't care about.
778 if (SimplifyDemandedInstructionBits(CI))
781 Value *Src = CI.getOperand(0);
782 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
784 // Attempt to extend the entire input expression tree to the destination
785 // type. Only do this if the dest type is a simple type, don't convert the
786 // expression tree to something weird like i93 unless the source is also
788 unsigned BitsToClear;
789 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
790 CanEvaluateZExtd(Src, DestTy, BitsToClear, *this, &CI)) {
791 assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
792 "Unreasonable BitsToClear");
794 // Okay, we can transform this! Insert the new expression now.
795 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
796 " to avoid zero extend: " << CI);
797 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
798 assert(Res->getType() == DestTy);
800 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
801 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
803 // If the high bits are already filled with zeros, just replace this
804 // cast with the result.
805 if (MaskedValueIsZero(Res,
806 APInt::getHighBitsSet(DestBitSize,
807 DestBitSize-SrcBitsKept),
809 return ReplaceInstUsesWith(CI, Res);
811 // We need to emit an AND to clear the high bits.
812 Constant *C = ConstantInt::get(Res->getType(),
813 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
814 return BinaryOperator::CreateAnd(Res, C);
817 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
818 // types and if the sizes are just right we can convert this into a logical
819 // 'and' which will be much cheaper than the pair of casts.
820 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
821 // TODO: Subsume this into EvaluateInDifferentType.
823 // Get the sizes of the types involved. We know that the intermediate type
824 // will be smaller than A or C, but don't know the relation between A and C.
825 Value *A = CSrc->getOperand(0);
826 unsigned SrcSize = A->getType()->getScalarSizeInBits();
827 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
828 unsigned DstSize = CI.getType()->getScalarSizeInBits();
829 // If we're actually extending zero bits, then if
830 // SrcSize < DstSize: zext(a & mask)
831 // SrcSize == DstSize: a & mask
832 // SrcSize > DstSize: trunc(a) & mask
833 if (SrcSize < DstSize) {
834 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
835 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
836 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
837 return new ZExtInst(And, CI.getType());
840 if (SrcSize == DstSize) {
841 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
842 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
845 if (SrcSize > DstSize) {
846 Value *Trunc = Builder->CreateTrunc(A, CI.getType());
847 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
848 return BinaryOperator::CreateAnd(Trunc,
849 ConstantInt::get(Trunc->getType(),
854 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
855 return transformZExtICmp(ICI, CI);
857 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
858 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
859 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
860 // of the (zext icmp) will be transformed.
861 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
862 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
863 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
864 (transformZExtICmp(LHS, CI, false) ||
865 transformZExtICmp(RHS, CI, false))) {
866 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
867 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
868 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
872 // zext(trunc(X) & C) -> (X & zext(C)).
876 match(SrcI, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
877 X->getType() == CI.getType())
878 return BinaryOperator::CreateAnd(X, ConstantExpr::getZExt(C, CI.getType()));
880 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
882 if (SrcI && match(SrcI, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
883 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
884 X->getType() == CI.getType()) {
885 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
886 return BinaryOperator::CreateXor(Builder->CreateAnd(X, ZC), ZC);
889 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
890 if (SrcI && SrcI->hasOneUse() &&
891 SrcI->getType()->getScalarType()->isIntegerTy(1) &&
892 match(SrcI, m_Not(m_Value(X))) && (!X->hasOneUse() || !isa<CmpInst>(X))) {
893 Value *New = Builder->CreateZExt(X, CI.getType());
894 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
900 /// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
901 /// in order to eliminate the icmp.
902 Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
903 Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
904 ICmpInst::Predicate Pred = ICI->getPredicate();
906 // Don't bother if Op1 isn't of vector or integer type.
907 if (!Op1->getType()->isIntOrIntVectorTy())
910 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
911 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if negative
912 // (x >s -1) ? -1 : 0 -> not (ashr x, 31) -> all ones if positive
913 if ((Pred == ICmpInst::ICMP_SLT && Op1C->isNullValue()) ||
914 (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
916 Value *Sh = ConstantInt::get(Op0->getType(),
917 Op0->getType()->getScalarSizeInBits()-1);
918 Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
919 if (In->getType() != CI.getType())
920 In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
922 if (Pred == ICmpInst::ICMP_SGT)
923 In = Builder->CreateNot(In, In->getName()+".not");
924 return ReplaceInstUsesWith(CI, In);
928 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
929 // If we know that only one bit of the LHS of the icmp can be set and we
930 // have an equality comparison with zero or a power of 2, we can transform
931 // the icmp and sext into bitwise/integer operations.
932 if (ICI->hasOneUse() &&
933 ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
934 unsigned BitWidth = Op1C->getType()->getBitWidth();
935 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
936 computeKnownBits(Op0, KnownZero, KnownOne, 0, &CI);
938 APInt KnownZeroMask(~KnownZero);
939 if (KnownZeroMask.isPowerOf2()) {
940 Value *In = ICI->getOperand(0);
942 // If the icmp tests for a known zero bit we can constant fold it.
943 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
944 Value *V = Pred == ICmpInst::ICMP_NE ?
945 ConstantInt::getAllOnesValue(CI.getType()) :
946 ConstantInt::getNullValue(CI.getType());
947 return ReplaceInstUsesWith(CI, V);
950 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
951 // sext ((x & 2^n) == 0) -> (x >> n) - 1
952 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
953 unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
954 // Perform a right shift to place the desired bit in the LSB.
956 In = Builder->CreateLShr(In,
957 ConstantInt::get(In->getType(), ShiftAmt));
959 // At this point "In" is either 1 or 0. Subtract 1 to turn
960 // {1, 0} -> {0, -1}.
961 In = Builder->CreateAdd(In,
962 ConstantInt::getAllOnesValue(In->getType()),
965 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
966 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
967 unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
968 // Perform a left shift to place the desired bit in the MSB.
970 In = Builder->CreateShl(In,
971 ConstantInt::get(In->getType(), ShiftAmt));
973 // Distribute the bit over the whole bit width.
974 In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
975 BitWidth - 1), "sext");
978 if (CI.getType() == In->getType())
979 return ReplaceInstUsesWith(CI, In);
980 return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
988 /// CanEvaluateSExtd - Return true if we can take the specified value
989 /// and return it as type Ty without inserting any new casts and without
990 /// changing the value of the common low bits. This is used by code that tries
991 /// to promote integer operations to a wider types will allow us to eliminate
994 /// This function works on both vectors and scalars.
996 static bool CanEvaluateSExtd(Value *V, Type *Ty) {
997 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
998 "Can't sign extend type to a smaller type");
999 // If this is a constant, it can be trivially promoted.
1000 if (isa<Constant>(V))
1003 Instruction *I = dyn_cast<Instruction>(V);
1004 if (!I) return false;
1006 // If this is a truncate from the dest type, we can trivially eliminate it.
1007 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1010 // We can't extend or shrink something that has multiple uses: doing so would
1011 // require duplicating the instruction in general, which isn't profitable.
1012 if (!I->hasOneUse()) return false;
1014 switch (I->getOpcode()) {
1015 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1016 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1017 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1019 case Instruction::And:
1020 case Instruction::Or:
1021 case Instruction::Xor:
1022 case Instruction::Add:
1023 case Instruction::Sub:
1024 case Instruction::Mul:
1025 // These operators can all arbitrarily be extended if their inputs can.
1026 return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1027 CanEvaluateSExtd(I->getOperand(1), Ty);
1029 //case Instruction::Shl: TODO
1030 //case Instruction::LShr: TODO
1032 case Instruction::Select:
1033 return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1034 CanEvaluateSExtd(I->getOperand(2), Ty);
1036 case Instruction::PHI: {
1037 // We can change a phi if we can change all operands. Note that we never
1038 // get into trouble with cyclic PHIs here because we only consider
1039 // instructions with a single use.
1040 PHINode *PN = cast<PHINode>(I);
1041 for (Value *IncValue : PN->incoming_values())
1042 if (!CanEvaluateSExtd(IncValue, Ty)) return false;
1046 // TODO: Can handle more cases here.
1053 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1054 // If this sign extend is only used by a truncate, let the truncate be
1055 // eliminated before we try to optimize this sext.
1056 if (CI.hasOneUse() && isa<TruncInst>(CI.user_back()))
1059 if (Instruction *I = commonCastTransforms(CI))
1062 // See if we can simplify any instructions used by the input whose sole
1063 // purpose is to compute bits we don't care about.
1064 if (SimplifyDemandedInstructionBits(CI))
1067 Value *Src = CI.getOperand(0);
1068 Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1070 // If we know that the value being extended is positive, we can use a zext
1072 bool KnownZero, KnownOne;
1073 ComputeSignBit(Src, KnownZero, KnownOne, 0, &CI);
1075 Value *ZExt = Builder->CreateZExt(Src, DestTy);
1076 return ReplaceInstUsesWith(CI, ZExt);
1079 // Attempt to extend the entire input expression tree to the destination
1080 // type. Only do this if the dest type is a simple type, don't convert the
1081 // expression tree to something weird like i93 unless the source is also
1083 if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1084 CanEvaluateSExtd(Src, DestTy)) {
1085 // Okay, we can transform this! Insert the new expression now.
1086 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1087 " to avoid sign extend: " << CI);
1088 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1089 assert(Res->getType() == DestTy);
1091 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1092 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1094 // If the high bits are already filled with sign bit, just replace this
1095 // cast with the result.
1096 if (ComputeNumSignBits(Res, 0, &CI) > DestBitSize - SrcBitSize)
1097 return ReplaceInstUsesWith(CI, Res);
1099 // We need to emit a shl + ashr to do the sign extend.
1100 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1101 return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1105 // If this input is a trunc from our destination, then turn sext(trunc(x))
1107 if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1108 if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1109 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1110 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1112 // We need to emit a shl + ashr to do the sign extend.
1113 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1114 Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1115 return BinaryOperator::CreateAShr(Res, ShAmt);
1118 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1119 return transformSExtICmp(ICI, CI);
1121 // If the input is a shl/ashr pair of a same constant, then this is a sign
1122 // extension from a smaller value. If we could trust arbitrary bitwidth
1123 // integers, we could turn this into a truncate to the smaller bit and then
1124 // use a sext for the whole extension. Since we don't, look deeper and check
1125 // for a truncate. If the source and dest are the same type, eliminate the
1126 // trunc and extend and just do shifts. For example, turn:
1127 // %a = trunc i32 %i to i8
1128 // %b = shl i8 %a, 6
1129 // %c = ashr i8 %b, 6
1130 // %d = sext i8 %c to i32
1132 // %a = shl i32 %i, 30
1133 // %d = ashr i32 %a, 30
1135 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1136 ConstantInt *BA = nullptr, *CA = nullptr;
1137 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1138 m_ConstantInt(CA))) &&
1139 BA == CA && A->getType() == CI.getType()) {
1140 unsigned MidSize = Src->getType()->getScalarSizeInBits();
1141 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1142 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1143 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1144 A = Builder->CreateShl(A, ShAmtV, CI.getName());
1145 return BinaryOperator::CreateAShr(A, ShAmtV);
1152 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1153 /// in the specified FP type without changing its value.
1154 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1156 APFloat F = CFP->getValueAPF();
1157 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1159 return ConstantFP::get(CFP->getContext(), F);
1163 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1164 /// through it until we get the source value.
1165 static Value *LookThroughFPExtensions(Value *V) {
1166 if (Instruction *I = dyn_cast<Instruction>(V))
1167 if (I->getOpcode() == Instruction::FPExt)
1168 return LookThroughFPExtensions(I->getOperand(0));
1170 // If this value is a constant, return the constant in the smallest FP type
1171 // that can accurately represent it. This allows us to turn
1172 // (float)((double)X+2.0) into x+2.0f.
1173 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1174 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1175 return V; // No constant folding of this.
1176 // See if the value can be truncated to half and then reextended.
1177 if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1179 // See if the value can be truncated to float and then reextended.
1180 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1182 if (CFP->getType()->isDoubleTy())
1183 return V; // Won't shrink.
1184 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1186 // Don't try to shrink to various long double types.
1192 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1193 if (Instruction *I = commonCastTransforms(CI))
1195 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1196 // simpilify this expression to avoid one or more of the trunc/extend
1197 // operations if we can do so without changing the numerical results.
1199 // The exact manner in which the widths of the operands interact to limit
1200 // what we can and cannot do safely varies from operation to operation, and
1201 // is explained below in the various case statements.
1202 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1203 if (OpI && OpI->hasOneUse()) {
1204 Value *LHSOrig = LookThroughFPExtensions(OpI->getOperand(0));
1205 Value *RHSOrig = LookThroughFPExtensions(OpI->getOperand(1));
1206 unsigned OpWidth = OpI->getType()->getFPMantissaWidth();
1207 unsigned LHSWidth = LHSOrig->getType()->getFPMantissaWidth();
1208 unsigned RHSWidth = RHSOrig->getType()->getFPMantissaWidth();
1209 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1210 unsigned DstWidth = CI.getType()->getFPMantissaWidth();
1211 switch (OpI->getOpcode()) {
1213 case Instruction::FAdd:
1214 case Instruction::FSub:
1215 // For addition and subtraction, the infinitely precise result can
1216 // essentially be arbitrarily wide; proving that double rounding
1217 // will not occur because the result of OpI is exact (as we will for
1218 // FMul, for example) is hopeless. However, we *can* nonetheless
1219 // frequently know that double rounding cannot occur (or that it is
1220 // innocuous) by taking advantage of the specific structure of
1221 // infinitely-precise results that admit double rounding.
1223 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1224 // to represent both sources, we can guarantee that the double
1225 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1226 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1227 // for proof of this fact).
1229 // Note: Figueroa does not consider the case where DstFormat !=
1230 // SrcFormat. It's possible (likely even!) that this analysis
1231 // could be tightened for those cases, but they are rare (the main
1232 // case of interest here is (float)((double)float + float)).
1233 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1234 if (LHSOrig->getType() != CI.getType())
1235 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1236 if (RHSOrig->getType() != CI.getType())
1237 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1239 BinaryOperator::Create(OpI->getOpcode(), LHSOrig, RHSOrig);
1240 RI->copyFastMathFlags(OpI);
1244 case Instruction::FMul:
1245 // For multiplication, the infinitely precise result has at most
1246 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1247 // that such a value can be exactly represented, then no double
1248 // rounding can possibly occur; we can safely perform the operation
1249 // in the destination format if it can represent both sources.
1250 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1251 if (LHSOrig->getType() != CI.getType())
1252 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1253 if (RHSOrig->getType() != CI.getType())
1254 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1256 BinaryOperator::CreateFMul(LHSOrig, RHSOrig);
1257 RI->copyFastMathFlags(OpI);
1261 case Instruction::FDiv:
1262 // For division, we use again use the bound from Figueroa's
1263 // dissertation. I am entirely certain that this bound can be
1264 // tightened in the unbalanced operand case by an analysis based on
1265 // the diophantine rational approximation bound, but the well-known
1266 // condition used here is a good conservative first pass.
1267 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1268 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1269 if (LHSOrig->getType() != CI.getType())
1270 LHSOrig = Builder->CreateFPExt(LHSOrig, CI.getType());
1271 if (RHSOrig->getType() != CI.getType())
1272 RHSOrig = Builder->CreateFPExt(RHSOrig, CI.getType());
1274 BinaryOperator::CreateFDiv(LHSOrig, RHSOrig);
1275 RI->copyFastMathFlags(OpI);
1279 case Instruction::FRem:
1280 // Remainder is straightforward. Remainder is always exact, so the
1281 // type of OpI doesn't enter into things at all. We simply evaluate
1282 // in whichever source type is larger, then convert to the
1283 // destination type.
1284 if (SrcWidth == OpWidth)
1286 if (LHSWidth < SrcWidth)
1287 LHSOrig = Builder->CreateFPExt(LHSOrig, RHSOrig->getType());
1288 else if (RHSWidth <= SrcWidth)
1289 RHSOrig = Builder->CreateFPExt(RHSOrig, LHSOrig->getType());
1290 if (LHSOrig != OpI->getOperand(0) || RHSOrig != OpI->getOperand(1)) {
1291 Value *ExactResult = Builder->CreateFRem(LHSOrig, RHSOrig);
1292 if (Instruction *RI = dyn_cast<Instruction>(ExactResult))
1293 RI->copyFastMathFlags(OpI);
1294 return CastInst::CreateFPCast(ExactResult, CI.getType());
1298 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1299 if (BinaryOperator::isFNeg(OpI)) {
1300 Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1302 Instruction *RI = BinaryOperator::CreateFNeg(InnerTrunc);
1303 RI->copyFastMathFlags(OpI);
1308 // (fptrunc (select cond, R1, Cst)) -->
1309 // (select cond, (fptrunc R1), (fptrunc Cst))
1310 SelectInst *SI = dyn_cast<SelectInst>(CI.getOperand(0));
1312 (isa<ConstantFP>(SI->getOperand(1)) ||
1313 isa<ConstantFP>(SI->getOperand(2)))) {
1314 Value *LHSTrunc = Builder->CreateFPTrunc(SI->getOperand(1),
1316 Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
1318 return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
1321 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1323 switch (II->getIntrinsicID()) {
1325 case Intrinsic::fabs: {
1326 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1327 Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1329 Type *IntrinsicType[] = { CI.getType() };
1330 Function *Overload =
1331 Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1332 II->getIntrinsicID(), IntrinsicType);
1334 Value *Args[] = { InnerTrunc };
1335 return CallInst::Create(Overload, Args, II->getName());
1343 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1344 return commonCastTransforms(CI);
1347 // fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1348 // This is safe if the intermediate type has enough bits in its mantissa to
1349 // accurately represent all values of X. For example, this won't work with
1350 // i64 -> float -> i64.
1351 Instruction *InstCombiner::FoldItoFPtoI(Instruction &FI) {
1352 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1354 Instruction *OpI = cast<Instruction>(FI.getOperand(0));
1356 Value *SrcI = OpI->getOperand(0);
1357 Type *FITy = FI.getType();
1358 Type *OpITy = OpI->getType();
1359 Type *SrcTy = SrcI->getType();
1360 bool IsInputSigned = isa<SIToFPInst>(OpI);
1361 bool IsOutputSigned = isa<FPToSIInst>(FI);
1363 // We can safely assume the conversion won't overflow the output range,
1364 // because (for example) (uint8_t)18293.f is undefined behavior.
1366 // Since we can assume the conversion won't overflow, our decision as to
1367 // whether the input will fit in the float should depend on the minimum
1368 // of the input range and output range.
1370 // This means this is also safe for a signed input and unsigned output, since
1371 // a negative input would lead to undefined behavior.
1372 int InputSize = (int)SrcTy->getScalarSizeInBits() - IsInputSigned;
1373 int OutputSize = (int)FITy->getScalarSizeInBits() - IsOutputSigned;
1374 int ActualSize = std::min(InputSize, OutputSize);
1376 if (ActualSize <= OpITy->getFPMantissaWidth()) {
1377 if (FITy->getScalarSizeInBits() > SrcTy->getScalarSizeInBits()) {
1378 if (IsInputSigned && IsOutputSigned)
1379 return new SExtInst(SrcI, FITy);
1380 return new ZExtInst(SrcI, FITy);
1382 if (FITy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits())
1383 return new TruncInst(SrcI, FITy);
1385 return ReplaceInstUsesWith(FI, SrcI);
1386 return new BitCastInst(SrcI, FITy);
1391 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1392 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1394 return commonCastTransforms(FI);
1396 if (Instruction *I = FoldItoFPtoI(FI))
1399 return commonCastTransforms(FI);
1402 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1403 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1405 return commonCastTransforms(FI);
1407 if (Instruction *I = FoldItoFPtoI(FI))
1410 return commonCastTransforms(FI);
1413 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1414 return commonCastTransforms(CI);
1417 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1418 return commonCastTransforms(CI);
1421 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1422 // If the source integer type is not the intptr_t type for this target, do a
1423 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1424 // cast to be exposed to other transforms.
1425 unsigned AS = CI.getAddressSpace();
1426 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1427 DL.getPointerSizeInBits(AS)) {
1428 Type *Ty = DL.getIntPtrType(CI.getContext(), AS);
1429 if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1430 Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1432 Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1433 return new IntToPtrInst(P, CI.getType());
1436 if (Instruction *I = commonCastTransforms(CI))
1442 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1443 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1444 Value *Src = CI.getOperand(0);
1446 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1447 // If casting the result of a getelementptr instruction with no offset, turn
1448 // this into a cast of the original pointer!
1449 if (GEP->hasAllZeroIndices() &&
1450 // If CI is an addrspacecast and GEP changes the poiner type, merging
1451 // GEP into CI would undo canonicalizing addrspacecast with different
1452 // pointer types, causing infinite loops.
1453 (!isa<AddrSpaceCastInst>(CI) ||
1454 GEP->getType() == GEP->getPointerOperand()->getType())) {
1455 // Changing the cast operand is usually not a good idea but it is safe
1456 // here because the pointer operand is being replaced with another
1457 // pointer operand so the opcode doesn't need to change.
1459 CI.setOperand(0, GEP->getOperand(0));
1464 return commonCastTransforms(CI);
1467 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1468 // If the destination integer type is not the intptr_t type for this target,
1469 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1470 // to be exposed to other transforms.
1472 Type *Ty = CI.getType();
1473 unsigned AS = CI.getPointerAddressSpace();
1475 if (Ty->getScalarSizeInBits() == DL.getPointerSizeInBits(AS))
1476 return commonPointerCastTransforms(CI);
1478 Type *PtrTy = DL.getIntPtrType(CI.getContext(), AS);
1479 if (Ty->isVectorTy()) // Handle vectors of pointers.
1480 PtrTy = VectorType::get(PtrTy, Ty->getVectorNumElements());
1482 Value *P = Builder->CreatePtrToInt(CI.getOperand(0), PtrTy);
1483 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1486 /// OptimizeVectorResize - This input value (which is known to have vector type)
1487 /// is being zero extended or truncated to the specified vector type. Try to
1488 /// replace it with a shuffle (and vector/vector bitcast) if possible.
1490 /// The source and destination vector types may have different element types.
1491 static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1493 // We can only do this optimization if the output is a multiple of the input
1494 // element size, or the input is a multiple of the output element size.
1495 // Convert the input type to have the same element type as the output.
1496 VectorType *SrcTy = cast<VectorType>(InVal->getType());
1498 if (SrcTy->getElementType() != DestTy->getElementType()) {
1499 // The input types don't need to be identical, but for now they must be the
1500 // same size. There is no specific reason we couldn't handle things like
1501 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1503 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1504 DestTy->getElementType()->getPrimitiveSizeInBits())
1507 SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1508 InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1511 // Now that the element types match, get the shuffle mask and RHS of the
1512 // shuffle to use, which depends on whether we're increasing or decreasing the
1513 // size of the input.
1514 SmallVector<uint32_t, 16> ShuffleMask;
1517 if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1518 // If we're shrinking the number of elements, just shuffle in the low
1519 // elements from the input and use undef as the second shuffle input.
1520 V2 = UndefValue::get(SrcTy);
1521 for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1522 ShuffleMask.push_back(i);
1525 // If we're increasing the number of elements, shuffle in all of the
1526 // elements from InVal and fill the rest of the result elements with zeros
1527 // from a constant zero.
1528 V2 = Constant::getNullValue(SrcTy);
1529 unsigned SrcElts = SrcTy->getNumElements();
1530 for (unsigned i = 0, e = SrcElts; i != e; ++i)
1531 ShuffleMask.push_back(i);
1533 // The excess elements reference the first element of the zero input.
1534 for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1535 ShuffleMask.push_back(SrcElts);
1538 return new ShuffleVectorInst(InVal, V2,
1539 ConstantDataVector::get(V2->getContext(),
1543 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1544 return Value % Ty->getPrimitiveSizeInBits() == 0;
1547 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1548 return Value / Ty->getPrimitiveSizeInBits();
1551 /// CollectInsertionElements - V is a value which is inserted into a vector of
1552 /// VecEltTy. Look through the value to see if we can decompose it into
1553 /// insertions into the vector. See the example in the comment for
1554 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
1555 /// The type of V is always a non-zero multiple of VecEltTy's size.
1556 /// Shift is the number of bits between the lsb of V and the lsb of
1559 /// This returns false if the pattern can't be matched or true if it can,
1560 /// filling in Elements with the elements found here.
1561 static bool CollectInsertionElements(Value *V, unsigned Shift,
1562 SmallVectorImpl<Value *> &Elements,
1563 Type *VecEltTy, bool isBigEndian) {
1564 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
1565 "Shift should be a multiple of the element type size");
1567 // Undef values never contribute useful bits to the result.
1568 if (isa<UndefValue>(V)) return true;
1570 // If we got down to a value of the right type, we win, try inserting into the
1572 if (V->getType() == VecEltTy) {
1573 // Inserting null doesn't actually insert any elements.
1574 if (Constant *C = dyn_cast<Constant>(V))
1575 if (C->isNullValue())
1578 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
1580 ElementIndex = Elements.size() - ElementIndex - 1;
1582 // Fail if multiple elements are inserted into this slot.
1583 if (Elements[ElementIndex])
1586 Elements[ElementIndex] = V;
1590 if (Constant *C = dyn_cast<Constant>(V)) {
1591 // Figure out the # elements this provides, and bitcast it or slice it up
1593 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1595 // If the constant is the size of a vector element, we just need to bitcast
1596 // it to the right type so it gets properly inserted.
1598 return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1599 Shift, Elements, VecEltTy, isBigEndian);
1601 // Okay, this is a constant that covers multiple elements. Slice it up into
1602 // pieces and insert each element-sized piece into the vector.
1603 if (!isa<IntegerType>(C->getType()))
1604 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1605 C->getType()->getPrimitiveSizeInBits()));
1606 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1607 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1609 for (unsigned i = 0; i != NumElts; ++i) {
1610 unsigned ShiftI = Shift+i*ElementSize;
1611 Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1613 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1614 if (!CollectInsertionElements(Piece, ShiftI, Elements, VecEltTy,
1621 if (!V->hasOneUse()) return false;
1623 Instruction *I = dyn_cast<Instruction>(V);
1624 if (!I) return false;
1625 switch (I->getOpcode()) {
1626 default: return false; // Unhandled case.
1627 case Instruction::BitCast:
1628 return CollectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1630 case Instruction::ZExt:
1631 if (!isMultipleOfTypeSize(
1632 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1635 return CollectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1637 case Instruction::Or:
1638 return CollectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1640 CollectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
1642 case Instruction::Shl: {
1643 // Must be shifting by a constant that is a multiple of the element size.
1644 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1645 if (!CI) return false;
1646 Shift += CI->getZExtValue();
1647 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
1648 return CollectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
1656 /// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1657 /// may be doing shifts and ors to assemble the elements of the vector manually.
1658 /// Try to rip the code out and replace it with insertelements. This is to
1659 /// optimize code like this:
1661 /// %tmp37 = bitcast float %inc to i32
1662 /// %tmp38 = zext i32 %tmp37 to i64
1663 /// %tmp31 = bitcast float %inc5 to i32
1664 /// %tmp32 = zext i32 %tmp31 to i64
1665 /// %tmp33 = shl i64 %tmp32, 32
1666 /// %ins35 = or i64 %tmp33, %tmp38
1667 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
1669 /// Into two insertelements that do "buildvector{%inc, %inc5}".
1670 static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1672 VectorType *DestVecTy = cast<VectorType>(CI.getType());
1673 Value *IntInput = CI.getOperand(0);
1675 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1676 if (!CollectInsertionElements(IntInput, 0, Elements,
1677 DestVecTy->getElementType(),
1678 IC.getDataLayout().isBigEndian()))
1681 // If we succeeded, we know that all of the element are specified by Elements
1682 // or are zero if Elements has a null entry. Recast this as a set of
1684 Value *Result = Constant::getNullValue(CI.getType());
1685 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1686 if (!Elements[i]) continue; // Unset element.
1688 Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1689 IC.Builder->getInt32(i));
1696 /// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1697 /// bitcast. The various long double bitcasts can't get in here.
1698 static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI, InstCombiner &IC,
1699 const DataLayout &DL) {
1700 Value *Src = CI.getOperand(0);
1701 Type *DestTy = CI.getType();
1703 // If this is a bitcast from int to float, check to see if the int is an
1704 // extraction from a vector.
1705 Value *VecInput = nullptr;
1706 // bitcast(trunc(bitcast(somevector)))
1707 if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1708 isa<VectorType>(VecInput->getType())) {
1709 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1710 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1712 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1713 // If the element type of the vector doesn't match the result type,
1714 // bitcast it to be a vector type we can extract from.
1715 if (VecTy->getElementType() != DestTy) {
1716 VecTy = VectorType::get(DestTy,
1717 VecTy->getPrimitiveSizeInBits() / DestWidth);
1718 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1722 if (DL.isBigEndian())
1723 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1724 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1728 // bitcast(trunc(lshr(bitcast(somevector), cst))
1729 ConstantInt *ShAmt = nullptr;
1730 if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1731 m_ConstantInt(ShAmt)))) &&
1732 isa<VectorType>(VecInput->getType())) {
1733 VectorType *VecTy = cast<VectorType>(VecInput->getType());
1734 unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1735 if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1736 ShAmt->getZExtValue() % DestWidth == 0) {
1737 // If the element type of the vector doesn't match the result type,
1738 // bitcast it to be a vector type we can extract from.
1739 if (VecTy->getElementType() != DestTy) {
1740 VecTy = VectorType::get(DestTy,
1741 VecTy->getPrimitiveSizeInBits() / DestWidth);
1742 VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1745 unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1746 if (DL.isBigEndian())
1747 Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1748 return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1754 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1755 // If the operands are integer typed then apply the integer transforms,
1756 // otherwise just apply the common ones.
1757 Value *Src = CI.getOperand(0);
1758 Type *SrcTy = Src->getType();
1759 Type *DestTy = CI.getType();
1761 // Get rid of casts from one type to the same type. These are useless and can
1762 // be replaced by the operand.
1763 if (DestTy == Src->getType())
1764 return ReplaceInstUsesWith(CI, Src);
1766 if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1767 PointerType *SrcPTy = cast<PointerType>(SrcTy);
1768 Type *DstElTy = DstPTy->getElementType();
1769 Type *SrcElTy = SrcPTy->getElementType();
1771 // If we are casting a alloca to a pointer to a type of the same
1772 // size, rewrite the allocation instruction to allocate the "right" type.
1773 // There is no need to modify malloc calls because it is their bitcast that
1774 // needs to be cleaned up.
1775 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1776 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1779 // If the source and destination are pointers, and this cast is equivalent
1780 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1781 // This can enhance SROA and other transforms that want type-safe pointers.
1782 unsigned NumZeros = 0;
1783 while (SrcElTy != DstElTy &&
1784 isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1785 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1786 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(0U);
1790 // If we found a path from the src to dest, create the getelementptr now.
1791 if (SrcElTy == DstElTy) {
1792 SmallVector<Value *, 8> Idxs(NumZeros + 1, Builder->getInt32(0));
1793 return GetElementPtrInst::CreateInBounds(Src, Idxs);
1797 // Try to optimize int -> float bitcasts.
1798 if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1799 if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this, DL))
1802 if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1803 if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1804 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1805 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1806 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1807 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1810 if (isa<IntegerType>(SrcTy)) {
1811 // If this is a cast from an integer to vector, check to see if the input
1812 // is a trunc or zext of a bitcast from vector. If so, we can replace all
1813 // the casts with a shuffle and (potentially) a bitcast.
1814 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1815 CastInst *SrcCast = cast<CastInst>(Src);
1816 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1817 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1818 if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1819 cast<VectorType>(DestTy), *this))
1823 // If the input is an 'or' instruction, we may be doing shifts and ors to
1824 // assemble the elements of the vector manually. Try to rip the code out
1825 // and replace it with insertelements.
1826 if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1827 return ReplaceInstUsesWith(CI, V);
1831 if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1832 if (SrcVTy->getNumElements() == 1) {
1833 // If our destination is not a vector, then make this a straight
1834 // scalar-scalar cast.
1835 if (!DestTy->isVectorTy()) {
1837 Builder->CreateExtractElement(Src,
1838 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1839 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1842 // Otherwise, see if our source is an insert. If so, then use the scalar
1843 // component directly.
1844 if (InsertElementInst *IEI =
1845 dyn_cast<InsertElementInst>(CI.getOperand(0)))
1846 return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1851 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1852 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1853 // a bitcast to a vector with the same # elts.
1854 if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1855 DestTy->getVectorNumElements() == SVI->getType()->getNumElements() &&
1856 SVI->getType()->getNumElements() ==
1857 SVI->getOperand(0)->getType()->getVectorNumElements()) {
1859 // If either of the operands is a cast from CI.getType(), then
1860 // evaluating the shuffle in the casted destination's type will allow
1861 // us to eliminate at least one cast.
1862 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1863 Tmp->getOperand(0)->getType() == DestTy) ||
1864 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1865 Tmp->getOperand(0)->getType() == DestTy)) {
1866 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1867 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1868 // Return a new shuffle vector. Use the same element ID's, as we
1869 // know the vector types match #elts.
1870 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1875 if (SrcTy->isPointerTy())
1876 return commonPointerCastTransforms(CI);
1877 return commonCastTransforms(CI);
1880 Instruction *InstCombiner::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
1881 // If the destination pointer element type is not the same as the source's
1882 // first do a bitcast to the destination type, and then the addrspacecast.
1883 // This allows the cast to be exposed to other transforms.
1884 Value *Src = CI.getOperand(0);
1885 PointerType *SrcTy = cast<PointerType>(Src->getType()->getScalarType());
1886 PointerType *DestTy = cast<PointerType>(CI.getType()->getScalarType());
1888 Type *DestElemTy = DestTy->getElementType();
1889 if (SrcTy->getElementType() != DestElemTy) {
1890 Type *MidTy = PointerType::get(DestElemTy, SrcTy->getAddressSpace());
1891 if (VectorType *VT = dyn_cast<VectorType>(CI.getType())) {
1892 // Handle vectors of pointers.
1893 MidTy = VectorType::get(MidTy, VT->getNumElements());
1896 Value *NewBitCast = Builder->CreateBitCast(Src, MidTy);
1897 return new AddrSpaceCastInst(NewBitCast, CI.getType());
1900 return commonPointerCastTransforms(CI);