1 //===- InstCombineCasts.cpp -----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for cast operations.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Target/TargetData.h"
16 #include "llvm/Support/PatternMatch.h"
18 using namespace PatternMatch;
20 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
21 /// expression. If so, decompose it, returning some value X, such that Val is
24 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
26 assert(Val->getType()->isInteger(32) && "Unexpected allocation size type!");
27 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
28 Offset = CI->getZExtValue();
30 return ConstantInt::get(Type::getInt32Ty(Val->getContext()), 0);
33 if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
34 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
35 if (I->getOpcode() == Instruction::Shl) {
36 // This is a value scaled by '1 << the shift amt'.
37 Scale = 1U << RHS->getZExtValue();
39 return I->getOperand(0);
42 if (I->getOpcode() == Instruction::Mul) {
43 // This value is scaled by 'RHS'.
44 Scale = RHS->getZExtValue();
46 return I->getOperand(0);
49 if (I->getOpcode() == Instruction::Add) {
50 // We have X+C. Check to see if we really have (X*C2)+C1,
51 // where C1 is divisible by C2.
54 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
55 Offset += RHS->getZExtValue();
62 // Otherwise, we can't look past this.
68 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
69 /// try to eliminate the cast by moving the type information into the alloc.
70 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
72 // This requires TargetData to get the alloca alignment and size information.
75 const PointerType *PTy = cast<PointerType>(CI.getType());
77 BuilderTy AllocaBuilder(*Builder);
78 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
80 // Get the type really allocated and the type casted to.
81 const Type *AllocElTy = AI.getAllocatedType();
82 const Type *CastElTy = PTy->getElementType();
83 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
85 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
86 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
87 if (CastElTyAlign < AllocElTyAlign) return 0;
89 // If the allocation has multiple uses, only promote it if we are strictly
90 // increasing the alignment of the resultant allocation. If we keep it the
91 // same, we open the door to infinite loops of various kinds. (A reference
92 // from a dbg.declare doesn't count as a use for this purpose.)
93 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
94 CastElTyAlign == AllocElTyAlign) return 0;
96 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
97 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
98 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
100 // See if we can satisfy the modulus by pulling a scale out of the array
102 unsigned ArraySizeScale;
104 Value *NumElements = // See if the array size is a decomposable linear expr.
105 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
107 // If we can now satisfy the modulus, by using a non-1 scale, we really can
109 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
110 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
112 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
117 Amt = ConstantInt::get(Type::getInt32Ty(CI.getContext()), Scale);
118 // Insert before the alloca, not before the cast.
119 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
122 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
123 Value *Off = ConstantInt::get(Type::getInt32Ty(CI.getContext()),
125 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
128 AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
129 New->setAlignment(AI.getAlignment());
132 // If the allocation has one real use plus a dbg.declare, just remove the
134 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
135 EraseInstFromFunction(*(Instruction*)DI);
137 // If the allocation has multiple real uses, insert a cast and change all
138 // things that used it to use the new cast. This will also hack on CI, but it
140 else if (!AI.hasOneUse()) {
141 // New is the allocation instruction, pointer typed. AI is the original
142 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
143 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
144 AI.replaceAllUsesWith(NewCast);
146 return ReplaceInstUsesWith(CI, New);
150 /// CanEvaluateInDifferentType - Return true if we can take the specified value
151 /// and return it as type Ty without inserting any new casts and without
152 /// changing the computed value. This is used by code that tries to decide
153 /// whether promoting or shrinking integer operations to wider or smaller types
154 /// will allow us to eliminate a truncate or extend.
156 /// This is a truncation operation if Ty is smaller than V->getType(), or a zero
157 /// extension operation if Ty is larger.
159 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
160 /// should return true if trunc(V) can be computed by computing V in the smaller
161 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
162 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
163 /// efficiently truncated.
165 /// If CastOpc is zext, we are asking if the low bits of the value can be
166 /// computed in a larger type, which is then and'd to get the final result.
167 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
169 unsigned &NumCastsRemoved) {
170 assert(CastOpc == Instruction::ZExt || CastOpc == Instruction::Trunc);
172 // We can always evaluate constants in another type.
173 if (isa<Constant>(V))
176 Instruction *I = dyn_cast<Instruction>(V);
177 if (!I) return false;
179 const Type *OrigTy = V->getType();
181 // If this is an extension or truncate, we can often eliminate it.
182 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
183 // If this is a cast from the destination type, we can trivially eliminate
184 // it, and this will remove a cast overall.
185 if (I->getOperand(0)->getType() == Ty) {
186 // If the first operand is itself a cast, and is eliminable, do not count
187 // this as an eliminable cast. We would prefer to eliminate those two
189 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
195 // We can't extend or shrink something that has multiple uses: doing so would
196 // require duplicating the instruction in general, which isn't profitable.
197 if (!I->hasOneUse()) return false;
199 unsigned Opc = I->getOpcode();
201 case Instruction::Add:
202 case Instruction::Sub:
203 case Instruction::Mul:
204 case Instruction::And:
205 case Instruction::Or:
206 case Instruction::Xor:
207 // These operators can all arbitrarily be extended or truncated.
208 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
210 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
213 case Instruction::UDiv:
214 case Instruction::URem: {
215 // UDiv and URem can be truncated if all the truncated bits are zero.
216 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
217 uint32_t BitWidth = Ty->getScalarSizeInBits();
218 if (BitWidth < OrigBitWidth) {
219 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
220 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
221 MaskedValueIsZero(I->getOperand(1), Mask)) {
222 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
224 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
230 case Instruction::Shl:
231 // If we are truncating the result of this SHL, and if it's a shift of a
232 // constant amount, we can always perform a SHL in a smaller type.
233 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
234 uint32_t BitWidth = Ty->getScalarSizeInBits();
235 if (BitWidth < OrigTy->getScalarSizeInBits() &&
236 CI->getLimitedValue(BitWidth) < BitWidth)
237 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
241 case Instruction::LShr:
242 // If this is a truncate of a logical shr, we can truncate it to a smaller
243 // lshr iff we know that the bits we would otherwise be shifting in are
245 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
246 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
247 uint32_t BitWidth = Ty->getScalarSizeInBits();
248 if (BitWidth < OrigBitWidth &&
249 MaskedValueIsZero(I->getOperand(0),
250 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
251 CI->getLimitedValue(BitWidth) < BitWidth) {
252 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
257 case Instruction::ZExt:
258 case Instruction::SExt:
259 case Instruction::Trunc:
260 // If this is the same kind of case as our original (e.g. zext+zext), we
261 // can safely replace it. Note that replacing it does not reduce the number
262 // of casts in the input.
266 // sext (zext ty1), ty2 -> zext ty2
267 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
270 case Instruction::Select: {
271 SelectInst *SI = cast<SelectInst>(I);
272 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
274 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
277 case Instruction::PHI: {
278 // We can change a phi if we can change all operands. Note that we never
279 // get into trouble with cyclic PHIs here because we only consider
280 // instructions with a single use.
281 PHINode *PN = cast<PHINode>(I);
282 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
283 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
289 // TODO: Can handle more cases here.
296 /// CanEvaluateSExtd - Return true if we can take the specified value
297 /// and return it as type Ty without inserting any new casts and without
298 /// changing the value of the common low bits. This is used by code that tries
299 /// to promote integer operations to a wider types will allow us to eliminate
302 /// This returns 0 if we can't do this or the number of sign bits that would be
303 /// set if we can. For example, CanEvaluateSExtd(i16 1, i64) would return 63,
304 /// because the computation can be extended (to "i64 1") and the resulting
305 /// computation has 63 equal sign bits.
307 /// This function works on both vectors and scalars. For vectors, the result is
308 /// the number of bits known sign extended in each element.
310 static unsigned CanEvaluateSExtd(Value *V, const Type *Ty,
311 unsigned &NumCastsRemoved, TargetData *TD) {
312 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
313 "Can't sign extend type to a smaller type");
314 // If this is a constant, return the number of sign bits the extended version
316 if (Constant *C = dyn_cast<Constant>(V))
317 return ComputeNumSignBits(ConstantExpr::getSExt(C, Ty), TD);
319 Instruction *I = dyn_cast<Instruction>(V);
322 // If this is a truncate from the destination type, we can trivially eliminate
323 // it, and this will remove a cast overall.
324 if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty) {
325 // If the operand of the truncate is itself a cast, and is eliminable, do
326 // not count this as an eliminable cast. We would prefer to eliminate those
328 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
330 return ComputeNumSignBits(I->getOperand(0), TD);
333 // We can't extend or shrink something that has multiple uses: doing so would
334 // require duplicating the instruction in general, which isn't profitable.
335 if (!I->hasOneUse()) return 0;
337 const Type *OrigTy = V->getType();
339 unsigned Opc = I->getOpcode();
342 case Instruction::And:
343 case Instruction::Or:
344 case Instruction::Xor:
345 // These operators can all arbitrarily be extended or truncated.
346 Tmp1 = CanEvaluateSExtd(I->getOperand(0), Ty, NumCastsRemoved, TD);
347 if (Tmp1 == 0) return 0;
348 Tmp2 = CanEvaluateSExtd(I->getOperand(1), Ty, NumCastsRemoved, TD);
349 return std::min(Tmp1, Tmp2);
350 case Instruction::Add:
351 case Instruction::Sub:
352 // Add/Sub can have at most one carry/borrow bit.
353 Tmp1 = CanEvaluateSExtd(I->getOperand(0), Ty, NumCastsRemoved, TD);
354 if (Tmp1 == 0) return 0;
355 Tmp2 = CanEvaluateSExtd(I->getOperand(1), Ty, NumCastsRemoved, TD);
356 if (Tmp2 == 0) return 0;
357 return std::min(Tmp1, Tmp2)-1;
358 case Instruction::Mul:
359 // These operators can all arbitrarily be extended or truncated.
360 if (!CanEvaluateSExtd(I->getOperand(0), Ty, NumCastsRemoved, TD))
362 if (!CanEvaluateSExtd(I->getOperand(1), Ty, NumCastsRemoved, TD))
364 return 1; // IMPROVE?
366 //case Instruction::Shl: TODO
367 //case Instruction::LShr: TODO
368 //case Instruction::Trunc: TODO
370 case Instruction::SExt:
371 case Instruction::ZExt: {
372 // sext(sext(x)) -> sext(x)
373 // sext(zext(x)) -> zext(x)
374 // Note that replacing a cast does not reduce the number of casts in the
376 unsigned InSignBits = ComputeNumSignBits(I, TD);
377 unsigned ExtBits = Ty->getScalarSizeInBits()-OrigTy->getScalarSizeInBits();
378 // We'll end up extending it all the way out.
379 return InSignBits+ExtBits;
381 case Instruction::Select: {
382 SelectInst *SI = cast<SelectInst>(I);
383 Tmp1 = CanEvaluateSExtd(SI->getTrueValue(), Ty, NumCastsRemoved, TD);
384 if (Tmp1 == 0) return 0;
385 Tmp2 = CanEvaluateSExtd(SI->getFalseValue(), Ty, NumCastsRemoved,TD);
386 return std::min(Tmp1, Tmp2);
388 case Instruction::PHI: {
389 // We can change a phi if we can change all operands. Note that we never
390 // get into trouble with cyclic PHIs here because we only consider
391 // instructions with a single use.
392 PHINode *PN = cast<PHINode>(I);
393 unsigned Result = ~0U;
394 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
395 Result = std::min(Result,
396 CanEvaluateSExtd(PN->getIncomingValue(i), Ty,
397 NumCastsRemoved, TD));
398 if (Result == 0) return 0;
403 // TODO: Can handle more cases here.
411 /// EvaluateInDifferentType - Given an expression that
412 /// CanEvaluateInDifferentType or CanEvaluateSExtd returns true for, actually
413 /// insert the code to evaluate the expression.
414 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
416 if (Constant *C = dyn_cast<Constant>(V))
417 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
419 // Otherwise, it must be an instruction.
420 Instruction *I = cast<Instruction>(V);
421 Instruction *Res = 0;
422 unsigned Opc = I->getOpcode();
424 case Instruction::Add:
425 case Instruction::Sub:
426 case Instruction::Mul:
427 case Instruction::And:
428 case Instruction::Or:
429 case Instruction::Xor:
430 case Instruction::AShr:
431 case Instruction::LShr:
432 case Instruction::Shl:
433 case Instruction::UDiv:
434 case Instruction::URem: {
435 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
436 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
437 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
440 case Instruction::Trunc:
441 case Instruction::ZExt:
442 case Instruction::SExt:
443 // If the source type of the cast is the type we're trying for then we can
444 // just return the source. There's no need to insert it because it is not
446 if (I->getOperand(0)->getType() == Ty)
447 return I->getOperand(0);
449 // Otherwise, must be the same type of cast, so just reinsert a new one.
450 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),Ty);
452 case Instruction::Select: {
453 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
454 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
455 Res = SelectInst::Create(I->getOperand(0), True, False);
458 case Instruction::PHI: {
459 PHINode *OPN = cast<PHINode>(I);
460 PHINode *NPN = PHINode::Create(Ty);
461 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
462 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
463 NPN->addIncoming(V, OPN->getIncomingBlock(i));
469 // TODO: Can handle more cases here.
470 llvm_unreachable("Unreachable!");
475 return InsertNewInstBefore(Res, *I);
479 /// This function is a wrapper around CastInst::isEliminableCastPair. It
480 /// simply extracts arguments and returns what that function returns.
481 static Instruction::CastOps
482 isEliminableCastPair(
483 const CastInst *CI, ///< The first cast instruction
484 unsigned opcode, ///< The opcode of the second cast instruction
485 const Type *DstTy, ///< The target type for the second cast instruction
486 TargetData *TD ///< The target data for pointer size
489 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
490 const Type *MidTy = CI->getType(); // B from above
492 // Get the opcodes of the two Cast instructions
493 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
494 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
496 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
498 TD ? TD->getIntPtrType(CI->getContext()) : 0);
500 // We don't want to form an inttoptr or ptrtoint that converts to an integer
501 // type that differs from the pointer size.
502 if ((Res == Instruction::IntToPtr &&
503 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
504 (Res == Instruction::PtrToInt &&
505 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
508 return Instruction::CastOps(Res);
511 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
512 /// in any code being generated. It does not require codegen if V is simple
513 /// enough or if the cast can be folded into other casts.
514 bool InstCombiner::ValueRequiresCast(Instruction::CastOps opcode,const Value *V,
516 if (V->getType() == Ty || isa<Constant>(V)) return false;
518 // If this is another cast that can be eliminated, it isn't codegen either.
519 if (const CastInst *CI = dyn_cast<CastInst>(V))
520 if (isEliminableCastPair(CI, opcode, Ty, TD))
526 /// @brief Implement the transforms common to all CastInst visitors.
527 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
528 Value *Src = CI.getOperand(0);
530 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
532 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
533 if (Instruction::CastOps opc =
534 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
535 // The first cast (CSrc) is eliminable so we need to fix up or replace
536 // the second cast (CI). CSrc will then have a good chance of being dead.
537 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
541 // If we are casting a select then fold the cast into the select
542 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
543 if (Instruction *NV = FoldOpIntoSelect(CI, SI))
546 // If we are casting a PHI then fold the cast into the PHI
547 if (isa<PHINode>(Src)) {
548 // We don't do this if this would create a PHI node with an illegal type if
549 // it is currently legal.
550 if (!isa<IntegerType>(Src->getType()) ||
551 !isa<IntegerType>(CI.getType()) ||
552 ShouldChangeType(CI.getType(), Src->getType()))
553 if (Instruction *NV = FoldOpIntoPhi(CI))
560 /// commonIntCastTransforms - This function implements the common transforms
561 /// for trunc, zext, and sext.
562 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
563 if (Instruction *Result = commonCastTransforms(CI))
566 // See if we can simplify any instructions used by the LHS whose sole
567 // purpose is to compute bits we don't care about.
568 if (SimplifyDemandedInstructionBits(CI))
571 // If the source isn't an instruction or has more than one use then we
572 // can't do anything more.
573 Instruction *Src = dyn_cast<Instruction>(CI.getOperand(0));
574 if (!Src || !Src->hasOneUse())
577 // Check to see if we can eliminate the cast by changing the entire
578 // computation chain to do the computation in the result type.
579 const Type *SrcTy = Src->getType();
580 const Type *DestTy = CI.getType();
582 // Only do this if the dest type is a simple type, don't convert the
583 // expression tree to something weird like i93 unless the source is also
585 if (!isa<VectorType>(DestTy) && !ShouldChangeType(SrcTy, DestTy))
588 // Attempt to propagate the cast into the instruction for int->int casts.
589 unsigned NumCastsRemoved = 0;
590 switch (CI.getOpcode()) {
591 default: assert(0 && "not an integer cast");
592 case Instruction::Trunc:
593 if (!CanEvaluateInDifferentType(Src, DestTy,
594 Instruction::Trunc, NumCastsRemoved))
597 // If this cast is a truncate, evaluting in a different type always
598 // eliminates the cast, so it is always a win.
600 case Instruction::ZExt:
601 if (!CanEvaluateInDifferentType(Src, DestTy,
602 Instruction::ZExt, NumCastsRemoved))
605 // If this is a zero-extension, we need to do an AND to maintain the clear
606 // top-part of the computation, so we require that the input have eliminated
607 // at least one cast.
608 if (NumCastsRemoved < 1)
611 case Instruction::SExt: {
612 // Check to see if we can do this transformation, and if so, how many bits
613 // of the promoted expression will be known copies of the sign bit in the
615 unsigned NumBitsSExt = CanEvaluateSExtd(Src, DestTy, NumCastsRemoved, TD);
616 if (NumBitsSExt == 0)
619 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
620 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
622 // Because this is a sign extension, we can always transform it by inserting
623 // two new shifts (to do the extension). However, this is only profitable
624 // if we've eliminated two or more casts from the input. If we know the
625 // result will be sign-extended enough to not require these shifts, we can
626 // always do the transformation.
627 if (NumCastsRemoved < 2 &&
628 NumBitsSExt <= DestBitSize-SrcBitSize)
631 // Okay, we can transform this! Insert the new expression now.
632 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
633 " to avoid sign extend: " << CI);
634 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
635 assert(Res->getType() == DestTy);
637 // If the high bits are already filled with sign bit, just replace this
638 // cast with the result.
639 if (NumBitsSExt > DestBitSize - SrcBitSize ||
640 ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
641 return ReplaceInstUsesWith(CI, Res);
643 // We need to emit a cast to truncate, then a cast to sext.
644 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
648 DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
649 " to avoid cast: " << CI);
650 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
651 assert(Res->getType() == DestTy);
653 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
654 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
655 switch (CI.getOpcode()) {
656 default: assert(0 && "Unknown cast type!");
657 case Instruction::Trunc:
658 // Just replace this cast with the result.
659 return ReplaceInstUsesWith(CI, Res);
660 case Instruction::ZExt: {
661 // If the high bits are already zero, just replace this cast with the
663 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
664 if (MaskedValueIsZero(Res, Mask))
665 return ReplaceInstUsesWith(CI, Res);
667 // We need to emit an AND to clear the high bits.
668 Constant *C = ConstantInt::get(CI.getContext(),
669 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
670 return BinaryOperator::CreateAnd(Res, C);
672 case Instruction::SExt: {
673 // If the high bits are already filled with sign bit, just replace this
674 // cast with the result.
675 unsigned NumSignBits = ComputeNumSignBits(Res);
676 if (NumSignBits > (DestBitSize - SrcBitSize))
677 return ReplaceInstUsesWith(CI, Res);
679 // We need to emit a cast to truncate, then a cast to sext.
680 return new SExtInst(Builder->CreateTrunc(Res, Src->getType()), DestTy);
685 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
686 if (Instruction *Result = commonIntCastTransforms(CI))
689 Value *Src = CI.getOperand(0);
690 const Type *DestTy = CI.getType();
692 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
693 if (DestTy->getScalarSizeInBits() == 1) {
694 Constant *One = ConstantInt::get(Src->getType(), 1);
695 Src = Builder->CreateAnd(Src, One, "tmp");
696 Value *Zero = Constant::getNullValue(Src->getType());
697 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
703 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
704 /// in order to eliminate the icmp.
705 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
707 // If we are just checking for a icmp eq of a single bit and zext'ing it
708 // to an integer, then shift the bit to the appropriate place and then
709 // cast to integer to avoid the comparison.
710 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
711 const APInt &Op1CV = Op1C->getValue();
713 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
714 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
715 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
716 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
717 if (!DoXform) return ICI;
719 Value *In = ICI->getOperand(0);
720 Value *Sh = ConstantInt::get(In->getType(),
721 In->getType()->getScalarSizeInBits()-1);
722 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
723 if (In->getType() != CI.getType())
724 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
726 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
727 Constant *One = ConstantInt::get(In->getType(), 1);
728 In = Builder->CreateXor(In, One, In->getName()+".not");
731 return ReplaceInstUsesWith(CI, In);
736 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
737 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
738 // zext (X == 1) to i32 --> X iff X has only the low bit set.
739 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
740 // zext (X != 0) to i32 --> X iff X has only the low bit set.
741 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
742 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
743 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
744 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
745 // This only works for EQ and NE
747 // If Op1C some other power of two, convert:
748 uint32_t BitWidth = Op1C->getType()->getBitWidth();
749 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
750 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
751 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
753 APInt KnownZeroMask(~KnownZero);
754 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
755 if (!DoXform) return ICI;
757 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
758 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
759 // (X&4) == 2 --> false
760 // (X&4) != 2 --> true
761 Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
763 Res = ConstantExpr::getZExt(Res, CI.getType());
764 return ReplaceInstUsesWith(CI, Res);
767 uint32_t ShiftAmt = KnownZeroMask.logBase2();
768 Value *In = ICI->getOperand(0);
770 // Perform a logical shr by shiftamt.
771 // Insert the shift to put the result in the low bit.
772 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
773 In->getName()+".lobit");
776 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
777 Constant *One = ConstantInt::get(In->getType(), 1);
778 In = Builder->CreateXor(In, One, "tmp");
781 if (CI.getType() == In->getType())
782 return ReplaceInstUsesWith(CI, In);
784 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
789 // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
790 // It is also profitable to transform icmp eq into not(xor(A, B)) because that
791 // may lead to additional simplifications.
792 if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
793 if (const IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
794 uint32_t BitWidth = ITy->getBitWidth();
795 Value *LHS = ICI->getOperand(0);
796 Value *RHS = ICI->getOperand(1);
798 APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
799 APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
800 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
801 ComputeMaskedBits(LHS, TypeMask, KnownZeroLHS, KnownOneLHS);
802 ComputeMaskedBits(RHS, TypeMask, KnownZeroRHS, KnownOneRHS);
804 if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
805 APInt KnownBits = KnownZeroLHS | KnownOneLHS;
806 APInt UnknownBit = ~KnownBits;
807 if (UnknownBit.countPopulation() == 1) {
808 if (!DoXform) return ICI;
810 Value *Result = Builder->CreateXor(LHS, RHS);
812 // Mask off any bits that are set and won't be shifted away.
813 if (KnownOneLHS.uge(UnknownBit))
814 Result = Builder->CreateAnd(Result,
815 ConstantInt::get(ITy, UnknownBit));
817 // Shift the bit we're testing down to the lsb.
818 Result = Builder->CreateLShr(
819 Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
821 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
822 Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
823 Result->takeName(ICI);
824 return ReplaceInstUsesWith(CI, Result);
833 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
834 // If one of the common conversion will work, do it.
835 if (Instruction *Result = commonIntCastTransforms(CI))
838 Value *Src = CI.getOperand(0);
840 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
841 // types and if the sizes are just right we can convert this into a logical
842 // 'and' which will be much cheaper than the pair of casts.
843 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
844 // Get the sizes of the types involved. We know that the intermediate type
845 // will be smaller than A or C, but don't know the relation between A and C.
846 Value *A = CSrc->getOperand(0);
847 unsigned SrcSize = A->getType()->getScalarSizeInBits();
848 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
849 unsigned DstSize = CI.getType()->getScalarSizeInBits();
850 // If we're actually extending zero bits, then if
851 // SrcSize < DstSize: zext(a & mask)
852 // SrcSize == DstSize: a & mask
853 // SrcSize > DstSize: trunc(a) & mask
854 if (SrcSize < DstSize) {
855 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
856 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
857 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
858 return new ZExtInst(And, CI.getType());
861 if (SrcSize == DstSize) {
862 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
863 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
866 if (SrcSize > DstSize) {
867 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
868 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
869 return BinaryOperator::CreateAnd(Trunc,
870 ConstantInt::get(Trunc->getType(),
875 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
876 return transformZExtICmp(ICI, CI);
878 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
879 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
880 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
881 // of the (zext icmp) will be transformed.
882 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
883 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
884 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
885 (transformZExtICmp(LHS, CI, false) ||
886 transformZExtICmp(RHS, CI, false))) {
887 Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
888 Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
889 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
893 // zext(trunc(t) & C) -> (t & zext(C)).
894 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
895 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
896 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
897 Value *TI0 = TI->getOperand(0);
898 if (TI0->getType() == CI.getType())
900 BinaryOperator::CreateAnd(TI0,
901 ConstantExpr::getZExt(C, CI.getType()));
904 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
905 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
906 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
907 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
908 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
909 And->getOperand(1) == C)
910 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
911 Value *TI0 = TI->getOperand(0);
912 if (TI0->getType() == CI.getType()) {
913 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
914 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
915 return BinaryOperator::CreateXor(NewAnd, ZC);
919 // zext (xor i1 X, true) to i32 --> xor (zext i1 X to i32), 1
921 if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isInteger(1) &&
922 match(SrcI, m_Not(m_Value(X))) &&
923 (!X->hasOneUse() || !isa<CmpInst>(X))) {
924 Value *New = Builder->CreateZExt(X, CI.getType());
925 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
931 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
932 if (Instruction *I = commonIntCastTransforms(CI))
935 Value *Src = CI.getOperand(0);
937 // Canonicalize sign-extend from i1 to a select.
938 if (Src->getType()->isInteger(1))
939 return SelectInst::Create(Src,
940 Constant::getAllOnesValue(CI.getType()),
941 Constant::getNullValue(CI.getType()));
943 // See if the value being truncated is already sign extended. If so, just
944 // eliminate the trunc/sext pair.
945 if (Operator::getOpcode(Src) == Instruction::Trunc) {
946 Value *Op = cast<User>(Src)->getOperand(0);
947 unsigned OpBits = Op->getType()->getScalarSizeInBits();
948 unsigned MidBits = Src->getType()->getScalarSizeInBits();
949 unsigned DestBits = CI.getType()->getScalarSizeInBits();
950 unsigned NumSignBits = ComputeNumSignBits(Op);
952 if (OpBits == DestBits) {
953 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
954 // bits, it is already ready.
955 if (NumSignBits > DestBits-MidBits)
956 return ReplaceInstUsesWith(CI, Op);
957 } else if (OpBits < DestBits) {
958 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
959 // bits, just sext from i32.
960 if (NumSignBits > OpBits-MidBits)
961 return new SExtInst(Op, CI.getType(), "tmp");
963 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
964 // bits, just truncate to i32.
965 if (NumSignBits > OpBits-MidBits)
966 return new TruncInst(Op, CI.getType(), "tmp");
970 // If the input is a shl/ashr pair of a same constant, then this is a sign
971 // extension from a smaller value. If we could trust arbitrary bitwidth
972 // integers, we could turn this into a truncate to the smaller bit and then
973 // use a sext for the whole extension. Since we don't, look deeper and check
974 // for a truncate. If the source and dest are the same type, eliminate the
975 // trunc and extend and just do shifts. For example, turn:
976 // %a = trunc i32 %i to i8
978 // %c = ashr i8 %b, 6
979 // %d = sext i8 %c to i32
981 // %a = shl i32 %i, 30
982 // %d = ashr i32 %a, 30
984 ConstantInt *BA = 0, *CA = 0;
985 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
986 m_ConstantInt(CA))) &&
987 BA == CA && isa<TruncInst>(A)) {
988 Value *I = cast<TruncInst>(A)->getOperand(0);
989 if (I->getType() == CI.getType()) {
990 unsigned MidSize = Src->getType()->getScalarSizeInBits();
991 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
992 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
993 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
994 I = Builder->CreateShl(I, ShAmtV, CI.getName());
995 return BinaryOperator::CreateAShr(I, ShAmtV);
1003 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1004 /// in the specified FP type without changing its value.
1005 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1007 APFloat F = CFP->getValueAPF();
1008 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1010 return ConstantFP::get(CFP->getContext(), F);
1014 /// LookThroughFPExtensions - If this is an fp extension instruction, look
1015 /// through it until we get the source value.
1016 static Value *LookThroughFPExtensions(Value *V) {
1017 if (Instruction *I = dyn_cast<Instruction>(V))
1018 if (I->getOpcode() == Instruction::FPExt)
1019 return LookThroughFPExtensions(I->getOperand(0));
1021 // If this value is a constant, return the constant in the smallest FP type
1022 // that can accurately represent it. This allows us to turn
1023 // (float)((double)X+2.0) into x+2.0f.
1024 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1025 if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1026 return V; // No constant folding of this.
1027 // See if the value can be truncated to float and then reextended.
1028 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1030 if (CFP->getType()->isDoubleTy())
1031 return V; // Won't shrink.
1032 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1034 // Don't try to shrink to various long double types.
1040 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1041 if (Instruction *I = commonCastTransforms(CI))
1044 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1045 // smaller than the destination type, we can eliminate the truncate by doing
1046 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well
1047 // as many builtins (sqrt, etc).
1048 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1049 if (OpI && OpI->hasOneUse()) {
1050 switch (OpI->getOpcode()) {
1052 case Instruction::FAdd:
1053 case Instruction::FSub:
1054 case Instruction::FMul:
1055 case Instruction::FDiv:
1056 case Instruction::FRem:
1057 const Type *SrcTy = OpI->getType();
1058 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1059 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1060 if (LHSTrunc->getType() != SrcTy &&
1061 RHSTrunc->getType() != SrcTy) {
1062 unsigned DstSize = CI.getType()->getScalarSizeInBits();
1063 // If the source types were both smaller than the destination type of
1064 // the cast, do this xform.
1065 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1066 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1067 LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1068 RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1069 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1078 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1079 return commonCastTransforms(CI);
1082 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1083 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1085 return commonCastTransforms(FI);
1087 // fptoui(uitofp(X)) --> X
1088 // fptoui(sitofp(X)) --> X
1089 // This is safe if the intermediate type has enough bits in its mantissa to
1090 // accurately represent all values of X. For example, do not do this with
1091 // i64->float->i64. This is also safe for sitofp case, because any negative
1092 // 'X' value would cause an undefined result for the fptoui.
1093 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1094 OpI->getOperand(0)->getType() == FI.getType() &&
1095 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1096 OpI->getType()->getFPMantissaWidth())
1097 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1099 return commonCastTransforms(FI);
1102 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1103 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1105 return commonCastTransforms(FI);
1107 // fptosi(sitofp(X)) --> X
1108 // fptosi(uitofp(X)) --> X
1109 // This is safe if the intermediate type has enough bits in its mantissa to
1110 // accurately represent all values of X. For example, do not do this with
1111 // i64->float->i64. This is also safe for sitofp case, because any negative
1112 // 'X' value would cause an undefined result for the fptoui.
1113 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1114 OpI->getOperand(0)->getType() == FI.getType() &&
1115 (int)FI.getType()->getScalarSizeInBits() <=
1116 OpI->getType()->getFPMantissaWidth())
1117 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1119 return commonCastTransforms(FI);
1122 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1123 return commonCastTransforms(CI);
1126 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1127 return commonCastTransforms(CI);
1130 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1131 // If the source integer type is larger than the intptr_t type for
1132 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
1133 // allows the trunc to be exposed to other transforms. Don't do this for
1134 // extending inttoptr's, because we don't know if the target sign or zero
1135 // extends to pointers.
1136 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
1137 TD->getPointerSizeInBits()) {
1138 Value *P = Builder->CreateTrunc(CI.getOperand(0),
1139 TD->getIntPtrType(CI.getContext()), "tmp");
1140 return new IntToPtrInst(P, CI.getType());
1143 if (Instruction *I = commonCastTransforms(CI))
1149 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1150 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1151 Value *Src = CI.getOperand(0);
1153 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1154 // If casting the result of a getelementptr instruction with no offset, turn
1155 // this into a cast of the original pointer!
1156 if (GEP->hasAllZeroIndices()) {
1157 // Changing the cast operand is usually not a good idea but it is safe
1158 // here because the pointer operand is being replaced with another
1159 // pointer operand so the opcode doesn't need to change.
1161 CI.setOperand(0, GEP->getOperand(0));
1165 // If the GEP has a single use, and the base pointer is a bitcast, and the
1166 // GEP computes a constant offset, see if we can convert these three
1167 // instructions into fewer. This typically happens with unions and other
1168 // non-type-safe code.
1169 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1170 GEP->hasAllConstantIndices()) {
1171 // We are guaranteed to get a constant from EmitGEPOffset.
1172 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP));
1173 int64_t Offset = OffsetV->getSExtValue();
1175 // Get the base pointer input of the bitcast, and the type it points to.
1176 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1177 const Type *GEPIdxTy =
1178 cast<PointerType>(OrigBase->getType())->getElementType();
1179 SmallVector<Value*, 8> NewIndices;
1180 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices)) {
1181 // If we were able to index down into an element, create the GEP
1182 // and bitcast the result. This eliminates one bitcast, potentially
1184 Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1185 Builder->CreateInBoundsGEP(OrigBase,
1186 NewIndices.begin(), NewIndices.end()) :
1187 Builder->CreateGEP(OrigBase, NewIndices.begin(), NewIndices.end());
1188 NGEP->takeName(GEP);
1190 if (isa<BitCastInst>(CI))
1191 return new BitCastInst(NGEP, CI.getType());
1192 assert(isa<PtrToIntInst>(CI));
1193 return new PtrToIntInst(NGEP, CI.getType());
1198 return commonCastTransforms(CI);
1201 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1202 // If the destination integer type is smaller than the intptr_t type for
1203 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
1204 // trunc to be exposed to other transforms. Don't do this for extending
1205 // ptrtoint's, because we don't know if the target sign or zero extends its
1208 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
1209 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
1210 TD->getIntPtrType(CI.getContext()),
1212 return new TruncInst(P, CI.getType());
1215 return commonPointerCastTransforms(CI);
1218 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1219 // If the operands are integer typed then apply the integer transforms,
1220 // otherwise just apply the common ones.
1221 Value *Src = CI.getOperand(0);
1222 const Type *SrcTy = Src->getType();
1223 const Type *DestTy = CI.getType();
1225 // Get rid of casts from one type to the same type. These are useless and can
1226 // be replaced by the operand.
1227 if (DestTy == Src->getType())
1228 return ReplaceInstUsesWith(CI, Src);
1230 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1231 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
1232 const Type *DstElTy = DstPTy->getElementType();
1233 const Type *SrcElTy = SrcPTy->getElementType();
1235 // If the address spaces don't match, don't eliminate the bitcast, which is
1236 // required for changing types.
1237 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1240 // If we are casting a alloca to a pointer to a type of the same
1241 // size, rewrite the allocation instruction to allocate the "right" type.
1242 // There is no need to modify malloc calls because it is their bitcast that
1243 // needs to be cleaned up.
1244 if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1245 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1248 // If the source and destination are pointers, and this cast is equivalent
1249 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
1250 // This can enhance SROA and other transforms that want type-safe pointers.
1251 Constant *ZeroUInt =
1252 Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1253 unsigned NumZeros = 0;
1254 while (SrcElTy != DstElTy &&
1255 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
1256 SrcElTy->getNumContainedTypes() /* not "{}" */) {
1257 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1261 // If we found a path from the src to dest, create the getelementptr now.
1262 if (SrcElTy == DstElTy) {
1263 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1264 return GetElementPtrInst::CreateInBounds(Src, Idxs.begin(), Idxs.end(),"",
1265 ((Instruction*)NULL));
1269 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1270 if (DestVTy->getNumElements() == 1 && !isa<VectorType>(SrcTy)) {
1271 Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1272 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1273 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1274 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1278 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1279 if (SrcVTy->getNumElements() == 1 && !isa<VectorType>(DestTy)) {
1281 Builder->CreateExtractElement(Src,
1282 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1283 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1287 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1288 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
1289 // a bitconvert to a vector with the same # elts.
1290 if (SVI->hasOneUse() && isa<VectorType>(DestTy) &&
1291 cast<VectorType>(DestTy)->getNumElements() ==
1292 SVI->getType()->getNumElements() &&
1293 SVI->getType()->getNumElements() ==
1294 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1296 // If either of the operands is a cast from CI.getType(), then
1297 // evaluating the shuffle in the casted destination's type will allow
1298 // us to eliminate at least one cast.
1299 if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1300 Tmp->getOperand(0)->getType() == DestTy) ||
1301 ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1302 Tmp->getOperand(0)->getType() == DestTy)) {
1303 Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1304 Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1305 // Return a new shuffle vector. Use the same element ID's, as we
1306 // know the vector types match #elts.
1307 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1312 if (isa<PointerType>(SrcTy))
1313 return commonPointerCastTransforms(CI);
1314 return commonCastTransforms(CI);