1//===- InstCombineCasts.cpp -----------------------------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the visit functions for cast operations.
11//
12//===----------------------------------------------------------------------===//
13
14#include "InstCombine.h"
15#include "llvm/Analysis/ConstantFolding.h"
16#include "llvm/IR/DataLayout.h"
17#include "llvm/Support/PatternMatch.h"
18#include "llvm/Target/TargetLibraryInfo.h"
19using namespace llvm;
20using namespace PatternMatch;
21
22/// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
23/// expression.  If so, decompose it, returning some value X, such that Val is
24/// X*Scale+Offset.
25///
26static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
27                                        uint64_t &Offset) {
28  if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
29    Offset = CI->getZExtValue();
30    Scale  = 0;
31    return ConstantInt::get(Val->getType(), 0);
32  }
33
34  if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
35    // Cannot look past anything that might overflow.
36    OverflowingBinaryOperator *OBI = dyn_cast<OverflowingBinaryOperator>(Val);
37    if (OBI && !OBI->hasNoUnsignedWrap() && !OBI->hasNoSignedWrap()) {
38      Scale = 1;
39      Offset = 0;
40      return Val;
41    }
42
43    if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
44      if (I->getOpcode() == Instruction::Shl) {
45        // This is a value scaled by '1 << the shift amt'.
46        Scale = UINT64_C(1) << RHS->getZExtValue();
47        Offset = 0;
48        return I->getOperand(0);
49      }
50
51      if (I->getOpcode() == Instruction::Mul) {
52        // This value is scaled by 'RHS'.
53        Scale = RHS->getZExtValue();
54        Offset = 0;
55        return I->getOperand(0);
56      }
57
58      if (I->getOpcode() == Instruction::Add) {
59        // We have X+C.  Check to see if we really have (X*C2)+C1,
60        // where C1 is divisible by C2.
61        unsigned SubScale;
62        Value *SubVal =
63          DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
64        Offset += RHS->getZExtValue();
65        Scale = SubScale;
66        return SubVal;
67      }
68    }
69  }
70
71  // Otherwise, we can't look past this.
72  Scale = 1;
73  Offset = 0;
74  return Val;
75}
76
77/// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
78/// try to eliminate the cast by moving the type information into the alloc.
79Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
80                                                   AllocaInst &AI) {
81  // This requires DataLayout to get the alloca alignment and size information.
82  if (!TD) return 0;
83
84  PointerType *PTy = cast<PointerType>(CI.getType());
85
86  BuilderTy AllocaBuilder(*Builder);
87  AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
88
89  // Get the type really allocated and the type casted to.
90  Type *AllocElTy = AI.getAllocatedType();
91  Type *CastElTy = PTy->getElementType();
92  if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
93
94  unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
95  unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
96  if (CastElTyAlign < AllocElTyAlign) return 0;
97
98  // If the allocation has multiple uses, only promote it if we are strictly
99  // increasing the alignment of the resultant allocation.  If we keep it the
100  // same, we open the door to infinite loops of various kinds.
101  if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
102
103  uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
104  uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
105  if (CastElTySize == 0 || AllocElTySize == 0) return 0;
106
107  // If the allocation has multiple uses, only promote it if we're not
108  // shrinking the amount of memory being allocated.
109  uint64_t AllocElTyStoreSize = TD->getTypeStoreSize(AllocElTy);
110  uint64_t CastElTyStoreSize = TD->getTypeStoreSize(CastElTy);
111  if (!AI.hasOneUse() && CastElTyStoreSize < AllocElTyStoreSize) return 0;
112
113  // See if we can satisfy the modulus by pulling a scale out of the array
114  // size argument.
115  unsigned ArraySizeScale;
116  uint64_t ArrayOffset;
117  Value *NumElements = // See if the array size is a decomposable linear expr.
118    DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
119
120  // If we can now satisfy the modulus, by using a non-1 scale, we really can
121  // do the xform.
122  if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
123      (AllocElTySize*ArrayOffset   ) % CastElTySize != 0) return 0;
124
125  unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
126  Value *Amt = 0;
127  if (Scale == 1) {
128    Amt = NumElements;
129  } else {
130    Amt = ConstantInt::get(AI.getArraySize()->getType(), Scale);
131    // Insert before the alloca, not before the cast.
132    Amt = AllocaBuilder.CreateMul(Amt, NumElements);
133  }
134
135  if (uint64_t Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
136    Value *Off = ConstantInt::get(AI.getArraySize()->getType(),
137                                  Offset, true);
138    Amt = AllocaBuilder.CreateAdd(Amt, Off);
139  }
140
141  AllocaInst *New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
142  New->setAlignment(AI.getAlignment());
143  New->takeName(&AI);
144
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
147  // will die soon.
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);
153  }
154  return ReplaceInstUsesWith(CI, New);
155}
156
157/// EvaluateInDifferentType - Given an expression that
158/// CanEvaluateTruncated or CanEvaluateSExtd returns true for, actually
159/// insert the code to evaluate the expression.
160Value *InstCombiner::EvaluateInDifferentType(Value *V, Type *Ty,
161                                             bool isSigned) {
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 TD info.
165    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
166      C = ConstantFoldConstantExpression(CE, TD, TLI);
167    return C;
168  }
169
170  // Otherwise, it must be an instruction.
171  Instruction *I = cast<Instruction>(V);
172  Instruction *Res = 0;
173  unsigned Opc = I->getOpcode();
174  switch (Opc) {
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);
189    break;
190  }
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
196    // new.
197    if (I->getOperand(0)->getType() == Ty)
198      return I->getOperand(0);
199
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);
204    break;
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);
209    break;
210  }
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) {
215      Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
216      NPN->addIncoming(V, OPN->getIncomingBlock(i));
217    }
218    Res = NPN;
219    break;
220  }
221  default:
222    // TODO: Can handle more cases here.
223    llvm_unreachable("Unreachable!");
224  }
225
226  Res->takeName(I);
227  return InsertNewInstWith(Res, *I);
228}
229
230
231/// This function is a wrapper around CastInst::isEliminableCastPair. It
232/// simply extracts arguments and returns what that function returns.
233static Instruction::CastOps
234isEliminableCastPair(
235  const CastInst *CI, ///< The first cast instruction
236  unsigned opcode,       ///< The opcode of the second cast instruction
237  Type *DstTy,     ///< The target type for the second cast instruction
238  DataLayout *TD         ///< The target data for pointer size
239) {
240
241  Type *SrcTy = CI->getOperand(0)->getType();   // A from above
242  Type *MidTy = CI->getType();                  // B from above
243
244  // Get the opcodes of the two Cast instructions
245  Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
246  Instruction::CastOps secondOp = Instruction::CastOps(opcode);
247  Type *SrcIntPtrTy = TD && SrcTy->isPtrOrPtrVectorTy() ?
248    TD->getIntPtrType(SrcTy) : 0;
249  Type *MidIntPtrTy = TD && MidTy->isPtrOrPtrVectorTy() ?
250    TD->getIntPtrType(MidTy) : 0;
251  Type *DstIntPtrTy = TD && DstTy->isPtrOrPtrVectorTy() ?
252    TD->getIntPtrType(DstTy) : 0;
253  unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
254                                                DstTy, SrcIntPtrTy, MidIntPtrTy,
255                                                DstIntPtrTy);
256
257  // We don't want to form an inttoptr or ptrtoint that converts to an integer
258  // type that differs from the pointer size.
259  if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
260      (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
261    Res = 0;
262
263  return Instruction::CastOps(Res);
264}
265
266/// ShouldOptimizeCast - Return true if the cast from "V to Ty" actually
267/// results in any code being generated and is interesting to optimize out. If
268/// the cast can be eliminated by some other simple transformation, we prefer
269/// to do the simplification first.
270bool InstCombiner::ShouldOptimizeCast(Instruction::CastOps opc, const Value *V,
271                                      Type *Ty) {
272  // Noop casts and casts of constants should be eliminated trivially.
273  if (V->getType() == Ty || isa<Constant>(V)) return false;
274
275  // If this is another cast that can be eliminated, we prefer to have it
276  // eliminated.
277  if (const CastInst *CI = dyn_cast<CastInst>(V))
278    if (isEliminableCastPair(CI, opc, Ty, TD))
279      return false;
280
281  // If this is a vector sext from a compare, then we don't want to break the
282  // idiom where each element of the extended vector is either zero or all ones.
283  if (opc == Instruction::SExt && isa<CmpInst>(V) && Ty->isVectorTy())
284    return false;
285
286  return true;
287}
288
289
290/// @brief Implement the transforms common to all CastInst visitors.
291Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
292  Value *Src = CI.getOperand(0);
293
294  // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
295  // eliminate it now.
296  if (CastInst *CSrc = dyn_cast<CastInst>(Src)) {   // A->B->C cast
297    if (Instruction::CastOps opc =
298        isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
299      // The first cast (CSrc) is eliminable so we need to fix up or replace
300      // the second cast (CI). CSrc will then have a good chance of being dead.
301      return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
302    }
303  }
304
305  // If we are casting a select then fold the cast into the select
306  if (SelectInst *SI = dyn_cast<SelectInst>(Src))
307    if (Instruction *NV = FoldOpIntoSelect(CI, SI))
308      return NV;
309
310  // If we are casting a PHI then fold the cast into the PHI
311  if (isa<PHINode>(Src)) {
312    // We don't do this if this would create a PHI node with an illegal type if
313    // it is currently legal.
314    if (!Src->getType()->isIntegerTy() ||
315        !CI.getType()->isIntegerTy() ||
316        ShouldChangeType(CI.getType(), Src->getType()))
317      if (Instruction *NV = FoldOpIntoPhi(CI))
318        return NV;
319  }
320
321  return 0;
322}
323
324/// CanEvaluateTruncated - Return true if we can evaluate the specified
325/// expression tree as type Ty instead of its larger type, and arrive with the
326/// same value.  This is used by code that tries to eliminate truncates.
327///
328/// Ty will always be a type smaller than V.  We should return true if trunc(V)
329/// can be computed by computing V in the smaller type.  If V is an instruction,
330/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
331/// makes sense if x and y can be efficiently truncated.
332///
333/// This function works on both vectors and scalars.
334///
335static bool CanEvaluateTruncated(Value *V, Type *Ty) {
336  // We can always evaluate constants in another type.
337  if (isa<Constant>(V))
338    return true;
339
340  Instruction *I = dyn_cast<Instruction>(V);
341  if (!I) return false;
342
343  Type *OrigTy = V->getType();
344
345  // If this is an extension from the dest type, we can eliminate it, even if it
346  // has multiple uses.
347  if ((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
348      I->getOperand(0)->getType() == Ty)
349    return true;
350
351  // We can't extend or shrink something that has multiple uses: doing so would
352  // require duplicating the instruction in general, which isn't profitable.
353  if (!I->hasOneUse()) return false;
354
355  unsigned Opc = I->getOpcode();
356  switch (Opc) {
357  case Instruction::Add:
358  case Instruction::Sub:
359  case Instruction::Mul:
360  case Instruction::And:
361  case Instruction::Or:
362  case Instruction::Xor:
363    // These operators can all arbitrarily be extended or truncated.
364    return CanEvaluateTruncated(I->getOperand(0), Ty) &&
365           CanEvaluateTruncated(I->getOperand(1), Ty);
366
367  case Instruction::UDiv:
368  case Instruction::URem: {
369    // UDiv and URem can be truncated if all the truncated bits are zero.
370    uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
371    uint32_t BitWidth = Ty->getScalarSizeInBits();
372    if (BitWidth < OrigBitWidth) {
373      APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
374      if (MaskedValueIsZero(I->getOperand(0), Mask) &&
375          MaskedValueIsZero(I->getOperand(1), Mask)) {
376        return CanEvaluateTruncated(I->getOperand(0), Ty) &&
377               CanEvaluateTruncated(I->getOperand(1), Ty);
378      }
379    }
380    break;
381  }
382  case Instruction::Shl:
383    // If we are truncating the result of this SHL, and if it's a shift of a
384    // constant amount, we can always perform a SHL in a smaller type.
385    if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
386      uint32_t BitWidth = Ty->getScalarSizeInBits();
387      if (CI->getLimitedValue(BitWidth) < BitWidth)
388        return CanEvaluateTruncated(I->getOperand(0), Ty);
389    }
390    break;
391  case Instruction::LShr:
392    // If this is a truncate of a logical shr, we can truncate it to a smaller
393    // lshr iff we know that the bits we would otherwise be shifting in are
394    // already zeros.
395    if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
396      uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
397      uint32_t BitWidth = Ty->getScalarSizeInBits();
398      if (MaskedValueIsZero(I->getOperand(0),
399            APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
400          CI->getLimitedValue(BitWidth) < BitWidth) {
401        return CanEvaluateTruncated(I->getOperand(0), Ty);
402      }
403    }
404    break;
405  case Instruction::Trunc:
406    // trunc(trunc(x)) -> trunc(x)
407    return true;
408  case Instruction::ZExt:
409  case Instruction::SExt:
410    // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
411    // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
412    return true;
413  case Instruction::Select: {
414    SelectInst *SI = cast<SelectInst>(I);
415    return CanEvaluateTruncated(SI->getTrueValue(), Ty) &&
416           CanEvaluateTruncated(SI->getFalseValue(), Ty);
417  }
418  case Instruction::PHI: {
419    // We can change a phi if we can change all operands.  Note that we never
420    // get into trouble with cyclic PHIs here because we only consider
421    // instructions with a single use.
422    PHINode *PN = cast<PHINode>(I);
423    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
424      if (!CanEvaluateTruncated(PN->getIncomingValue(i), Ty))
425        return false;
426    return true;
427  }
428  default:
429    // TODO: Can handle more cases here.
430    break;
431  }
432
433  return false;
434}
435
436Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
437  if (Instruction *Result = commonCastTransforms(CI))
438    return Result;
439
440  // See if we can simplify any instructions used by the input whose sole
441  // purpose is to compute bits we don't care about.
442  if (SimplifyDemandedInstructionBits(CI))
443    return &CI;
444
445  Value *Src = CI.getOperand(0);
446  Type *DestTy = CI.getType(), *SrcTy = Src->getType();
447
448  // Attempt to truncate the entire input expression tree to the destination
449  // type.   Only do this if the dest type is a simple type, don't convert the
450  // expression tree to something weird like i93 unless the source is also
451  // strange.
452  if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
453      CanEvaluateTruncated(Src, DestTy)) {
454
455    // If this cast is a truncate, evaluting in a different type always
456    // eliminates the cast, so it is always a win.
457    DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
458          " to avoid cast: " << CI << '\n');
459    Value *Res = EvaluateInDifferentType(Src, DestTy, false);
460    assert(Res->getType() == DestTy);
461    return ReplaceInstUsesWith(CI, Res);
462  }
463
464  // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0), likewise for vector.
465  if (DestTy->getScalarSizeInBits() == 1) {
466    Constant *One = ConstantInt::get(Src->getType(), 1);
467    Src = Builder->CreateAnd(Src, One);
468    Value *Zero = Constant::getNullValue(Src->getType());
469    return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
470  }
471
472  // Transform trunc(lshr (zext A), Cst) to eliminate one type conversion.
473  Value *A = 0; ConstantInt *Cst = 0;
474  if (Src->hasOneUse() &&
475      match(Src, m_LShr(m_ZExt(m_Value(A)), m_ConstantInt(Cst)))) {
476    // We have three types to worry about here, the type of A, the source of
477    // the truncate (MidSize), and the destination of the truncate. We know that
478    // ASize < MidSize   and MidSize > ResultSize, but don't know the relation
479    // between ASize and ResultSize.
480    unsigned ASize = A->getType()->getPrimitiveSizeInBits();
481
482    // If the shift amount is larger than the size of A, then the result is
483    // known to be zero because all the input bits got shifted out.
484    if (Cst->getZExtValue() >= ASize)
485      return ReplaceInstUsesWith(CI, Constant::getNullValue(CI.getType()));
486
487    // Since we're doing an lshr and a zero extend, and know that the shift
488    // amount is smaller than ASize, it is always safe to do the shift in A's
489    // type, then zero extend or truncate to the result.
490    Value *Shift = Builder->CreateLShr(A, Cst->getZExtValue());
491    Shift->takeName(Src);
492    return CastInst::CreateIntegerCast(Shift, CI.getType(), false);
493  }
494
495  // Transform "trunc (and X, cst)" -> "and (trunc X), cst" so long as the dest
496  // type isn't non-native.
497  if (Src->hasOneUse() && isa<IntegerType>(Src->getType()) &&
498      ShouldChangeType(Src->getType(), CI.getType()) &&
499      match(Src, m_And(m_Value(A), m_ConstantInt(Cst)))) {
500    Value *NewTrunc = Builder->CreateTrunc(A, CI.getType(), A->getName()+".tr");
501    return BinaryOperator::CreateAnd(NewTrunc,
502                                     ConstantExpr::getTrunc(Cst, CI.getType()));
503  }
504
505  return 0;
506}
507
508/// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
509/// in order to eliminate the icmp.
510Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
511                                             bool DoXform) {
512  // If we are just checking for a icmp eq of a single bit and zext'ing it
513  // to an integer, then shift the bit to the appropriate place and then
514  // cast to integer to avoid the comparison.
515  if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
516    const APInt &Op1CV = Op1C->getValue();
517
518    // zext (x <s  0) to i32 --> x>>u31      true if signbit set.
519    // zext (x >s -1) to i32 --> (x>>u31)^1  true if signbit clear.
520    if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
521        (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
522      if (!DoXform) return ICI;
523
524      Value *In = ICI->getOperand(0);
525      Value *Sh = ConstantInt::get(In->getType(),
526                                   In->getType()->getScalarSizeInBits()-1);
527      In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
528      if (In->getType() != CI.getType())
529        In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/);
530
531      if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
532        Constant *One = ConstantInt::get(In->getType(), 1);
533        In = Builder->CreateXor(In, One, In->getName()+".not");
534      }
535
536      return ReplaceInstUsesWith(CI, In);
537    }
538
539    // zext (X == 0) to i32 --> X^1      iff X has only the low bit set.
540    // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
541    // zext (X == 1) to i32 --> X        iff X has only the low bit set.
542    // zext (X == 2) to i32 --> X>>1     iff X has only the 2nd bit set.
543    // zext (X != 0) to i32 --> X        iff X has only the low bit set.
544    // zext (X != 0) to i32 --> X>>1     iff X has only the 2nd bit set.
545    // zext (X != 1) to i32 --> X^1      iff X has only the low bit set.
546    // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
547    if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
548        // This only works for EQ and NE
549        ICI->isEquality()) {
550      // If Op1C some other power of two, convert:
551      uint32_t BitWidth = Op1C->getType()->getBitWidth();
552      APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
553      ComputeMaskedBits(ICI->getOperand(0), KnownZero, KnownOne);
554
555      APInt KnownZeroMask(~KnownZero);
556      if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
557        if (!DoXform) return ICI;
558
559        bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
560        if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
561          // (X&4) == 2 --> false
562          // (X&4) != 2 --> true
563          Constant *Res = ConstantInt::get(Type::getInt1Ty(CI.getContext()),
564                                           isNE);
565          Res = ConstantExpr::getZExt(Res, CI.getType());
566          return ReplaceInstUsesWith(CI, Res);
567        }
568
569        uint32_t ShiftAmt = KnownZeroMask.logBase2();
570        Value *In = ICI->getOperand(0);
571        if (ShiftAmt) {
572          // Perform a logical shr by shiftamt.
573          // Insert the shift to put the result in the low bit.
574          In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
575                                   In->getName()+".lobit");
576        }
577
578        if ((Op1CV != 0) == isNE) { // Toggle the low bit.
579          Constant *One = ConstantInt::get(In->getType(), 1);
580          In = Builder->CreateXor(In, One);
581        }
582
583        if (CI.getType() == In->getType())
584          return ReplaceInstUsesWith(CI, In);
585        return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
586      }
587    }
588  }
589
590  // icmp ne A, B is equal to xor A, B when A and B only really have one bit.
591  // It is also profitable to transform icmp eq into not(xor(A, B)) because that
592  // may lead to additional simplifications.
593  if (ICI->isEquality() && CI.getType() == ICI->getOperand(0)->getType()) {
594    if (IntegerType *ITy = dyn_cast<IntegerType>(CI.getType())) {
595      uint32_t BitWidth = ITy->getBitWidth();
596      Value *LHS = ICI->getOperand(0);
597      Value *RHS = ICI->getOperand(1);
598
599      APInt KnownZeroLHS(BitWidth, 0), KnownOneLHS(BitWidth, 0);
600      APInt KnownZeroRHS(BitWidth, 0), KnownOneRHS(BitWidth, 0);
601      ComputeMaskedBits(LHS, KnownZeroLHS, KnownOneLHS);
602      ComputeMaskedBits(RHS, KnownZeroRHS, KnownOneRHS);
603
604      if (KnownZeroLHS == KnownZeroRHS && KnownOneLHS == KnownOneRHS) {
605        APInt KnownBits = KnownZeroLHS | KnownOneLHS;
606        APInt UnknownBit = ~KnownBits;
607        if (UnknownBit.countPopulation() == 1) {
608          if (!DoXform) return ICI;
609
610          Value *Result = Builder->CreateXor(LHS, RHS);
611
612          // Mask off any bits that are set and won't be shifted away.
613          if (KnownOneLHS.uge(UnknownBit))
614            Result = Builder->CreateAnd(Result,
615                                        ConstantInt::get(ITy, UnknownBit));
616
617          // Shift the bit we're testing down to the lsb.
618          Result = Builder->CreateLShr(
619               Result, ConstantInt::get(ITy, UnknownBit.countTrailingZeros()));
620
621          if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
622            Result = Builder->CreateXor(Result, ConstantInt::get(ITy, 1));
623          Result->takeName(ICI);
624          return ReplaceInstUsesWith(CI, Result);
625        }
626      }
627    }
628  }
629
630  return 0;
631}
632
633/// CanEvaluateZExtd - Determine if the specified value can be computed in the
634/// specified wider type and produce the same low bits.  If not, return false.
635///
636/// If this function returns true, it can also return a non-zero number of bits
637/// (in BitsToClear) which indicates that the value it computes is correct for
638/// the zero extend, but that the additional BitsToClear bits need to be zero'd
639/// out.  For example, to promote something like:
640///
641///   %B = trunc i64 %A to i32
642///   %C = lshr i32 %B, 8
643///   %E = zext i32 %C to i64
644///
645/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
646/// set to 8 to indicate that the promoted value needs to have bits 24-31
647/// cleared in addition to bits 32-63.  Since an 'and' will be generated to
648/// clear the top bits anyway, doing this has no extra cost.
649///
650/// This function works on both vectors and scalars.
651static bool CanEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear) {
652  BitsToClear = 0;
653  if (isa<Constant>(V))
654    return true;
655
656  Instruction *I = dyn_cast<Instruction>(V);
657  if (!I) return false;
658
659  // If the input is a truncate from the destination type, we can trivially
660  // eliminate it.
661  if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
662    return true;
663
664  // We can't extend or shrink something that has multiple uses: doing so would
665  // require duplicating the instruction in general, which isn't profitable.
666  if (!I->hasOneUse()) return false;
667
668  unsigned Opc = I->getOpcode(), Tmp;
669  switch (Opc) {
670  case Instruction::ZExt:  // zext(zext(x)) -> zext(x).
671  case Instruction::SExt:  // zext(sext(x)) -> sext(x).
672  case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
673    return true;
674  case Instruction::And:
675  case Instruction::Or:
676  case Instruction::Xor:
677  case Instruction::Add:
678  case Instruction::Sub:
679  case Instruction::Mul:
680  case Instruction::Shl:
681    if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear) ||
682        !CanEvaluateZExtd(I->getOperand(1), Ty, Tmp))
683      return false;
684    // These can all be promoted if neither operand has 'bits to clear'.
685    if (BitsToClear == 0 && Tmp == 0)
686      return true;
687
688    // If the operation is an AND/OR/XOR and the bits to clear are zero in the
689    // other side, BitsToClear is ok.
690    if (Tmp == 0 &&
691        (Opc == Instruction::And || Opc == Instruction::Or ||
692         Opc == Instruction::Xor)) {
693      // We use MaskedValueIsZero here for generality, but the case we care
694      // about the most is constant RHS.
695      unsigned VSize = V->getType()->getScalarSizeInBits();
696      if (MaskedValueIsZero(I->getOperand(1),
697                            APInt::getHighBitsSet(VSize, BitsToClear)))
698        return true;
699    }
700
701    // Otherwise, we don't know how to analyze this BitsToClear case yet.
702    return false;
703
704  case Instruction::LShr:
705    // We can promote lshr(x, cst) if we can promote x.  This requires the
706    // ultimate 'and' to clear out the high zero bits we're clearing out though.
707    if (ConstantInt *Amt = dyn_cast<ConstantInt>(I->getOperand(1))) {
708      if (!CanEvaluateZExtd(I->getOperand(0), Ty, BitsToClear))
709        return false;
710      BitsToClear += Amt->getZExtValue();
711      if (BitsToClear > V->getType()->getScalarSizeInBits())
712        BitsToClear = V->getType()->getScalarSizeInBits();
713      return true;
714    }
715    // Cannot promote variable LSHR.
716    return false;
717  case Instruction::Select:
718    if (!CanEvaluateZExtd(I->getOperand(1), Ty, Tmp) ||
719        !CanEvaluateZExtd(I->getOperand(2), Ty, BitsToClear) ||
720        // TODO: If important, we could handle the case when the BitsToClear are
721        // known zero in the disagreeing side.
722        Tmp != BitsToClear)
723      return false;
724    return true;
725
726  case Instruction::PHI: {
727    // We can change a phi if we can change all operands.  Note that we never
728    // get into trouble with cyclic PHIs here because we only consider
729    // instructions with a single use.
730    PHINode *PN = cast<PHINode>(I);
731    if (!CanEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear))
732      return false;
733    for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
734      if (!CanEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp) ||
735          // TODO: If important, we could handle the case when the BitsToClear
736          // are known zero in the disagreeing input.
737          Tmp != BitsToClear)
738        return false;
739    return true;
740  }
741  default:
742    // TODO: Can handle more cases here.
743    return false;
744  }
745}
746
747Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
748  // If this zero extend is only used by a truncate, let the truncate be
749  // eliminated before we try to optimize this zext.
750  if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
751    return 0;
752
753  // If one of the common conversion will work, do it.
754  if (Instruction *Result = commonCastTransforms(CI))
755    return Result;
756
757  // See if we can simplify any instructions used by the input whose sole
758  // purpose is to compute bits we don't care about.
759  if (SimplifyDemandedInstructionBits(CI))
760    return &CI;
761
762  Value *Src = CI.getOperand(0);
763  Type *SrcTy = Src->getType(), *DestTy = CI.getType();
764
765  // Attempt to extend the entire input expression tree to the destination
766  // type.   Only do this if the dest type is a simple type, don't convert the
767  // expression tree to something weird like i93 unless the source is also
768  // strange.
769  unsigned BitsToClear;
770  if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
771      CanEvaluateZExtd(Src, DestTy, BitsToClear)) {
772    assert(BitsToClear < SrcTy->getScalarSizeInBits() &&
773           "Unreasonable BitsToClear");
774
775    // Okay, we can transform this!  Insert the new expression now.
776    DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
777          " to avoid zero extend: " << CI);
778    Value *Res = EvaluateInDifferentType(Src, DestTy, false);
779    assert(Res->getType() == DestTy);
780
781    uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits()-BitsToClear;
782    uint32_t DestBitSize = DestTy->getScalarSizeInBits();
783
784    // If the high bits are already filled with zeros, just replace this
785    // cast with the result.
786    if (MaskedValueIsZero(Res, APInt::getHighBitsSet(DestBitSize,
787                                                     DestBitSize-SrcBitsKept)))
788      return ReplaceInstUsesWith(CI, Res);
789
790    // We need to emit an AND to clear the high bits.
791    Constant *C = ConstantInt::get(Res->getType(),
792                               APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
793    return BinaryOperator::CreateAnd(Res, C);
794  }
795
796  // If this is a TRUNC followed by a ZEXT then we are dealing with integral
797  // types and if the sizes are just right we can convert this into a logical
798  // 'and' which will be much cheaper than the pair of casts.
799  if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) {   // A->B->C cast
800    // TODO: Subsume this into EvaluateInDifferentType.
801
802    // Get the sizes of the types involved.  We know that the intermediate type
803    // will be smaller than A or C, but don't know the relation between A and C.
804    Value *A = CSrc->getOperand(0);
805    unsigned SrcSize = A->getType()->getScalarSizeInBits();
806    unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
807    unsigned DstSize = CI.getType()->getScalarSizeInBits();
808    // If we're actually extending zero bits, then if
809    // SrcSize <  DstSize: zext(a & mask)
810    // SrcSize == DstSize: a & mask
811    // SrcSize  > DstSize: trunc(a) & mask
812    if (SrcSize < DstSize) {
813      APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
814      Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
815      Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
816      return new ZExtInst(And, CI.getType());
817    }
818
819    if (SrcSize == DstSize) {
820      APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
821      return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
822                                                           AndValue));
823    }
824    if (SrcSize > DstSize) {
825      Value *Trunc = Builder->CreateTrunc(A, CI.getType());
826      APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
827      return BinaryOperator::CreateAnd(Trunc,
828                                       ConstantInt::get(Trunc->getType(),
829                                                        AndValue));
830    }
831  }
832
833  if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
834    return transformZExtICmp(ICI, CI);
835
836  BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
837  if (SrcI && SrcI->getOpcode() == Instruction::Or) {
838    // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
839    // of the (zext icmp) will be transformed.
840    ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
841    ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
842    if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
843        (transformZExtICmp(LHS, CI, false) ||
844         transformZExtICmp(RHS, CI, false))) {
845      Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());
846      Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());
847      return BinaryOperator::Create(Instruction::Or, LCast, RCast);
848    }
849  }
850
851  // zext(trunc(t) & C) -> (t & zext(C)).
852  if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
853    if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
854      if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
855        Value *TI0 = TI->getOperand(0);
856        if (TI0->getType() == CI.getType())
857          return
858            BinaryOperator::CreateAnd(TI0,
859                                ConstantExpr::getZExt(C, CI.getType()));
860      }
861
862  // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
863  if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
864    if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
865      if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
866        if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
867            And->getOperand(1) == C)
868          if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
869            Value *TI0 = TI->getOperand(0);
870            if (TI0->getType() == CI.getType()) {
871              Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
872              Value *NewAnd = Builder->CreateAnd(TI0, ZC);
873              return BinaryOperator::CreateXor(NewAnd, ZC);
874            }
875          }
876
877  // zext (xor i1 X, true) to i32  --> xor (zext i1 X to i32), 1
878  Value *X;
879  if (SrcI && SrcI->hasOneUse() && SrcI->getType()->isIntegerTy(1) &&
880      match(SrcI, m_Not(m_Value(X))) &&
881      (!X->hasOneUse() || !isa<CmpInst>(X))) {
882    Value *New = Builder->CreateZExt(X, CI.getType());
883    return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
884  }
885
886  return 0;
887}
888
889/// transformSExtICmp - Transform (sext icmp) to bitwise / integer operations
890/// in order to eliminate the icmp.
891Instruction *InstCombiner::transformSExtICmp(ICmpInst *ICI, Instruction &CI) {
892  Value *Op0 = ICI->getOperand(0), *Op1 = ICI->getOperand(1);
893  ICmpInst::Predicate Pred = ICI->getPredicate();
894
895  if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
896    // (x <s  0) ? -1 : 0 -> ashr x, 31        -> all ones if negative
897    // (x >s -1) ? -1 : 0 -> not (ashr x, 31)  -> all ones if positive
898    if ((Pred == ICmpInst::ICMP_SLT && Op1C->isZero()) ||
899        (Pred == ICmpInst::ICMP_SGT && Op1C->isAllOnesValue())) {
900
901      Value *Sh = ConstantInt::get(Op0->getType(),
902                                   Op0->getType()->getScalarSizeInBits()-1);
903      Value *In = Builder->CreateAShr(Op0, Sh, Op0->getName()+".lobit");
904      if (In->getType() != CI.getType())
905        In = Builder->CreateIntCast(In, CI.getType(), true/*SExt*/);
906
907      if (Pred == ICmpInst::ICMP_SGT)
908        In = Builder->CreateNot(In, In->getName()+".not");
909      return ReplaceInstUsesWith(CI, In);
910    }
911
912    // If we know that only one bit of the LHS of the icmp can be set and we
913    // have an equality comparison with zero or a power of 2, we can transform
914    // the icmp and sext into bitwise/integer operations.
915    if (ICI->hasOneUse() &&
916        ICI->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
917      unsigned BitWidth = Op1C->getType()->getBitWidth();
918      APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
919      ComputeMaskedBits(Op0, KnownZero, KnownOne);
920
921      APInt KnownZeroMask(~KnownZero);
922      if (KnownZeroMask.isPowerOf2()) {
923        Value *In = ICI->getOperand(0);
924
925        // If the icmp tests for a known zero bit we can constant fold it.
926        if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
927          Value *V = Pred == ICmpInst::ICMP_NE ?
928                       ConstantInt::getAllOnesValue(CI.getType()) :
929                       ConstantInt::getNullValue(CI.getType());
930          return ReplaceInstUsesWith(CI, V);
931        }
932
933        if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
934          // sext ((x & 2^n) == 0)   -> (x >> n) - 1
935          // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
936          unsigned ShiftAmt = KnownZeroMask.countTrailingZeros();
937          // Perform a right shift to place the desired bit in the LSB.
938          if (ShiftAmt)
939            In = Builder->CreateLShr(In,
940                                     ConstantInt::get(In->getType(), ShiftAmt));
941
942          // At this point "In" is either 1 or 0. Subtract 1 to turn
943          // {1, 0} -> {0, -1}.
944          In = Builder->CreateAdd(In,
945                                  ConstantInt::getAllOnesValue(In->getType()),
946                                  "sext");
947        } else {
948          // sext ((x & 2^n) != 0)   -> (x << bitwidth-n) a>> bitwidth-1
949          // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
950          unsigned ShiftAmt = KnownZeroMask.countLeadingZeros();
951          // Perform a left shift to place the desired bit in the MSB.
952          if (ShiftAmt)
953            In = Builder->CreateShl(In,
954                                    ConstantInt::get(In->getType(), ShiftAmt));
955
956          // Distribute the bit over the whole bit width.
957          In = Builder->CreateAShr(In, ConstantInt::get(In->getType(),
958                                                        BitWidth - 1), "sext");
959        }
960
961        if (CI.getType() == In->getType())
962          return ReplaceInstUsesWith(CI, In);
963        return CastInst::CreateIntegerCast(In, CI.getType(), true/*SExt*/);
964      }
965    }
966  }
967
968  // vector (x <s 0) ? -1 : 0 -> ashr x, 31   -> all ones if signed.
969  if (VectorType *VTy = dyn_cast<VectorType>(CI.getType())) {
970    if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_Zero()) &&
971        Op0->getType() == CI.getType()) {
972      Type *EltTy = VTy->getElementType();
973
974      // splat the shift constant to a constant vector.
975      Constant *VSh = ConstantInt::get(VTy, EltTy->getScalarSizeInBits()-1);
976      Value *In = Builder->CreateAShr(Op0, VSh, Op0->getName()+".lobit");
977      return ReplaceInstUsesWith(CI, In);
978    }
979  }
980
981  return 0;
982}
983
984/// CanEvaluateSExtd - Return true if we can take the specified value
985/// and return it as type Ty without inserting any new casts and without
986/// changing the value of the common low bits.  This is used by code that tries
987/// to promote integer operations to a wider types will allow us to eliminate
988/// the extension.
989///
990/// This function works on both vectors and scalars.
991///
992static bool CanEvaluateSExtd(Value *V, Type *Ty) {
993  assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
994         "Can't sign extend type to a smaller type");
995  // If this is a constant, it can be trivially promoted.
996  if (isa<Constant>(V))
997    return true;
998
999  Instruction *I = dyn_cast<Instruction>(V);
1000  if (!I) return false;
1001
1002  // If this is a truncate from the dest type, we can trivially eliminate it.
1003  if (isa<TruncInst>(I) && I->getOperand(0)->getType() == Ty)
1004    return true;
1005
1006  // We can't extend or shrink something that has multiple uses: doing so would
1007  // require duplicating the instruction in general, which isn't profitable.
1008  if (!I->hasOneUse()) return false;
1009
1010  switch (I->getOpcode()) {
1011  case Instruction::SExt:  // sext(sext(x)) -> sext(x)
1012  case Instruction::ZExt:  // sext(zext(x)) -> zext(x)
1013  case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1014    return true;
1015  case Instruction::And:
1016  case Instruction::Or:
1017  case Instruction::Xor:
1018  case Instruction::Add:
1019  case Instruction::Sub:
1020  case Instruction::Mul:
1021    // These operators can all arbitrarily be extended if their inputs can.
1022    return CanEvaluateSExtd(I->getOperand(0), Ty) &&
1023           CanEvaluateSExtd(I->getOperand(1), Ty);
1024
1025  //case Instruction::Shl:   TODO
1026  //case Instruction::LShr:  TODO
1027
1028  case Instruction::Select:
1029    return CanEvaluateSExtd(I->getOperand(1), Ty) &&
1030           CanEvaluateSExtd(I->getOperand(2), Ty);
1031
1032  case Instruction::PHI: {
1033    // We can change a phi if we can change all operands.  Note that we never
1034    // get into trouble with cyclic PHIs here because we only consider
1035    // instructions with a single use.
1036    PHINode *PN = cast<PHINode>(I);
1037    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1038      if (!CanEvaluateSExtd(PN->getIncomingValue(i), Ty)) return false;
1039    return true;
1040  }
1041  default:
1042    // TODO: Can handle more cases here.
1043    break;
1044  }
1045
1046  return false;
1047}
1048
1049Instruction *InstCombiner::visitSExt(SExtInst &CI) {
1050  // If this sign extend is only used by a truncate, let the truncate be
1051  // eliminated before we try to optimize this sext.
1052  if (CI.hasOneUse() && isa<TruncInst>(CI.use_back()))
1053    return 0;
1054
1055  if (Instruction *I = commonCastTransforms(CI))
1056    return I;
1057
1058  // See if we can simplify any instructions used by the input whose sole
1059  // purpose is to compute bits we don't care about.
1060  if (SimplifyDemandedInstructionBits(CI))
1061    return &CI;
1062
1063  Value *Src = CI.getOperand(0);
1064  Type *SrcTy = Src->getType(), *DestTy = CI.getType();
1065
1066  // Attempt to extend the entire input expression tree to the destination
1067  // type.   Only do this if the dest type is a simple type, don't convert the
1068  // expression tree to something weird like i93 unless the source is also
1069  // strange.
1070  if ((DestTy->isVectorTy() || ShouldChangeType(SrcTy, DestTy)) &&
1071      CanEvaluateSExtd(Src, DestTy)) {
1072    // Okay, we can transform this!  Insert the new expression now.
1073    DEBUG(dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1074          " to avoid sign extend: " << CI);
1075    Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1076    assert(Res->getType() == DestTy);
1077
1078    uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1079    uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1080
1081    // If the high bits are already filled with sign bit, just replace this
1082    // cast with the result.
1083    if (ComputeNumSignBits(Res) > DestBitSize - SrcBitSize)
1084      return ReplaceInstUsesWith(CI, Res);
1085
1086    // We need to emit a shl + ashr to do the sign extend.
1087    Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1088    return BinaryOperator::CreateAShr(Builder->CreateShl(Res, ShAmt, "sext"),
1089                                      ShAmt);
1090  }
1091
1092  // If this input is a trunc from our destination, then turn sext(trunc(x))
1093  // into shifts.
1094  if (TruncInst *TI = dyn_cast<TruncInst>(Src))
1095    if (TI->hasOneUse() && TI->getOperand(0)->getType() == DestTy) {
1096      uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
1097      uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1098
1099      // We need to emit a shl + ashr to do the sign extend.
1100      Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1101      Value *Res = Builder->CreateShl(TI->getOperand(0), ShAmt, "sext");
1102      return BinaryOperator::CreateAShr(Res, ShAmt);
1103    }
1104
1105  if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
1106    return transformSExtICmp(ICI, CI);
1107
1108  // If the input is a shl/ashr pair of a same constant, then this is a sign
1109  // extension from a smaller value.  If we could trust arbitrary bitwidth
1110  // integers, we could turn this into a truncate to the smaller bit and then
1111  // use a sext for the whole extension.  Since we don't, look deeper and check
1112  // for a truncate.  If the source and dest are the same type, eliminate the
1113  // trunc and extend and just do shifts.  For example, turn:
1114  //   %a = trunc i32 %i to i8
1115  //   %b = shl i8 %a, 6
1116  //   %c = ashr i8 %b, 6
1117  //   %d = sext i8 %c to i32
1118  // into:
1119  //   %a = shl i32 %i, 30
1120  //   %d = ashr i32 %a, 30
1121  Value *A = 0;
1122  // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1123  ConstantInt *BA = 0, *CA = 0;
1124  if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_ConstantInt(BA)),
1125                        m_ConstantInt(CA))) &&
1126      BA == CA && A->getType() == CI.getType()) {
1127    unsigned MidSize = Src->getType()->getScalarSizeInBits();
1128    unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
1129    unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
1130    Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
1131    A = Builder->CreateShl(A, ShAmtV, CI.getName());
1132    return BinaryOperator::CreateAShr(A, ShAmtV);
1133  }
1134
1135  return 0;
1136}
1137
1138
1139/// FitsInFPType - Return a Constant* for the specified FP constant if it fits
1140/// in the specified FP type without changing its value.
1141static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1142  bool losesInfo;
1143  APFloat F = CFP->getValueAPF();
1144  (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1145  if (!losesInfo)
1146    return ConstantFP::get(CFP->getContext(), F);
1147  return 0;
1148}
1149
1150/// LookThroughFPExtensions - If this is an fp extension instruction, look
1151/// through it until we get the source value.
1152static Value *LookThroughFPExtensions(Value *V) {
1153  if (Instruction *I = dyn_cast<Instruction>(V))
1154    if (I->getOpcode() == Instruction::FPExt)
1155      return LookThroughFPExtensions(I->getOperand(0));
1156
1157  // If this value is a constant, return the constant in the smallest FP type
1158  // that can accurately represent it.  This allows us to turn
1159  // (float)((double)X+2.0) into x+2.0f.
1160  if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
1161    if (CFP->getType() == Type::getPPC_FP128Ty(V->getContext()))
1162      return V;  // No constant folding of this.
1163    // See if the value can be truncated to half and then reextended.
1164    if (Value *V = FitsInFPType(CFP, APFloat::IEEEhalf))
1165      return V;
1166    // See if the value can be truncated to float and then reextended.
1167    if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
1168      return V;
1169    if (CFP->getType()->isDoubleTy())
1170      return V;  // Won't shrink.
1171    if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
1172      return V;
1173    // Don't try to shrink to various long double types.
1174  }
1175
1176  return V;
1177}
1178
1179Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
1180  if (Instruction *I = commonCastTransforms(CI))
1181    return I;
1182
1183  // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
1184  // smaller than the destination type, we can eliminate the truncate by doing
1185  // the add as the smaller type.  This applies to fadd/fsub/fmul/fdiv as well
1186  // as many builtins (sqrt, etc).
1187  BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
1188  if (OpI && OpI->hasOneUse()) {
1189    switch (OpI->getOpcode()) {
1190    default: break;
1191    case Instruction::FAdd:
1192    case Instruction::FSub:
1193    case Instruction::FMul:
1194    case Instruction::FDiv:
1195    case Instruction::FRem:
1196      Type *SrcTy = OpI->getType();
1197      Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
1198      Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
1199      if (LHSTrunc->getType() != SrcTy &&
1200          RHSTrunc->getType() != SrcTy) {
1201        unsigned DstSize = CI.getType()->getScalarSizeInBits();
1202        // If the source types were both smaller than the destination type of
1203        // the cast, do this xform.
1204        if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
1205            RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
1206          LHSTrunc = Builder->CreateFPExt(LHSTrunc, CI.getType());
1207          RHSTrunc = Builder->CreateFPExt(RHSTrunc, CI.getType());
1208          return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
1209        }
1210      }
1211      break;
1212    }
1213
1214    // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1215    if (BinaryOperator::isFNeg(OpI)) {
1216      Value *InnerTrunc = Builder->CreateFPTrunc(OpI->getOperand(1),
1217                                                 CI.getType());
1218      return BinaryOperator::CreateFNeg(InnerTrunc);
1219    }
1220  }
1221
1222  IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
1223  if (II) {
1224    switch (II->getIntrinsicID()) {
1225      default: break;
1226      case Intrinsic::fabs: {
1227        // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1228        Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
1229                                                   CI.getType());
1230        Type *IntrinsicType[] = { CI.getType() };
1231        Function *Overload =
1232          Intrinsic::getDeclaration(CI.getParent()->getParent()->getParent(),
1233                                    II->getIntrinsicID(), IntrinsicType);
1234
1235        Value *Args[] = { InnerTrunc };
1236        return CallInst::Create(Overload, Args, II->getName());
1237      }
1238    }
1239  }
1240
1241  // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
1242  CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
1243  if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
1244      Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) &&
1245      Call->getNumArgOperands() == 1 &&
1246      Call->hasOneUse()) {
1247    CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
1248    if (Arg && Arg->getOpcode() == Instruction::FPExt &&
1249        CI.getType()->isFloatTy() &&
1250        Call->getType()->isDoubleTy() &&
1251        Arg->getType()->isDoubleTy() &&
1252        Arg->getOperand(0)->getType()->isFloatTy()) {
1253      Function *Callee = Call->getCalledFunction();
1254      Module *M = CI.getParent()->getParent()->getParent();
1255      Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
1256                                                   Callee->getAttributes(),
1257                                                   Builder->getFloatTy(),
1258                                                   Builder->getFloatTy(),
1259                                                   NULL);
1260      CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
1261                                       "sqrtfcall");
1262      ret->setAttributes(Callee->getAttributes());
1263
1264
1265      // Remove the old Call.  With -fmath-errno, it won't get marked readnone.
1266      ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
1267      EraseInstFromFunction(*Call);
1268      return ret;
1269    }
1270  }
1271
1272  return 0;
1273}
1274
1275Instruction *InstCombiner::visitFPExt(CastInst &CI) {
1276  return commonCastTransforms(CI);
1277}
1278
1279Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
1280  Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1281  if (OpI == 0)
1282    return commonCastTransforms(FI);
1283
1284  // fptoui(uitofp(X)) --> X
1285  // fptoui(sitofp(X)) --> X
1286  // This is safe if the intermediate type has enough bits in its mantissa to
1287  // accurately represent all values of X.  For example, do not do this with
1288  // i64->float->i64.  This is also safe for sitofp case, because any negative
1289  // 'X' value would cause an undefined result for the fptoui.
1290  if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1291      OpI->getOperand(0)->getType() == FI.getType() &&
1292      (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
1293                    OpI->getType()->getFPMantissaWidth())
1294    return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1295
1296  return commonCastTransforms(FI);
1297}
1298
1299Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
1300  Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
1301  if (OpI == 0)
1302    return commonCastTransforms(FI);
1303
1304  // fptosi(sitofp(X)) --> X
1305  // fptosi(uitofp(X)) --> X
1306  // This is safe if the intermediate type has enough bits in its mantissa to
1307  // accurately represent all values of X.  For example, do not do this with
1308  // i64->float->i64.  This is also safe for sitofp case, because any negative
1309  // 'X' value would cause an undefined result for the fptoui.
1310  if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
1311      OpI->getOperand(0)->getType() == FI.getType() &&
1312      (int)FI.getType()->getScalarSizeInBits() <=
1313                    OpI->getType()->getFPMantissaWidth())
1314    return ReplaceInstUsesWith(FI, OpI->getOperand(0));
1315
1316  return commonCastTransforms(FI);
1317}
1318
1319Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
1320  return commonCastTransforms(CI);
1321}
1322
1323Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
1324  return commonCastTransforms(CI);
1325}
1326
1327Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
1328  // If the source integer type is not the intptr_t type for this target, do a
1329  // trunc or zext to the intptr_t type, then inttoptr of it.  This allows the
1330  // cast to be exposed to other transforms.
1331  if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1332      TD->getPointerSizeInBits()) {
1333    Type *Ty = TD->getIntPtrType(CI.getContext());
1334    if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1335      Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1336
1337    Value *P = Builder->CreateZExtOrTrunc(CI.getOperand(0), Ty);
1338    return new IntToPtrInst(P, CI.getType());
1339  }
1340
1341  if (Instruction *I = commonCastTransforms(CI))
1342    return I;
1343
1344  return 0;
1345}
1346
1347/// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
1348Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
1349  Value *Src = CI.getOperand(0);
1350
1351  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
1352    // If casting the result of a getelementptr instruction with no offset, turn
1353    // this into a cast of the original pointer!
1354    if (GEP->hasAllZeroIndices()) {
1355      // Changing the cast operand is usually not a good idea but it is safe
1356      // here because the pointer operand is being replaced with another
1357      // pointer operand so the opcode doesn't need to change.
1358      Worklist.Add(GEP);
1359      CI.setOperand(0, GEP->getOperand(0));
1360      return &CI;
1361    }
1362
1363    // If the GEP has a single use, and the base pointer is a bitcast, and the
1364    // GEP computes a constant offset, see if we can convert these three
1365    // instructions into fewer.  This typically happens with unions and other
1366    // non-type-safe code.
1367    APInt Offset(TD ? TD->getPointerSizeInBits() : 1, 0);
1368    if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0)) &&
1369        GEP->accumulateConstantOffset(*TD, Offset)) {
1370      // Get the base pointer input of the bitcast, and the type it points to.
1371      Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
1372      Type *GEPIdxTy =
1373      cast<PointerType>(OrigBase->getType())->getElementType();
1374      SmallVector<Value*, 8> NewIndices;
1375      if (FindElementAtOffset(GEPIdxTy, Offset.getSExtValue(), NewIndices)) {
1376        // If we were able to index down into an element, create the GEP
1377        // and bitcast the result.  This eliminates one bitcast, potentially
1378        // two.
1379        Value *NGEP = cast<GEPOperator>(GEP)->isInBounds() ?
1380        Builder->CreateInBoundsGEP(OrigBase, NewIndices) :
1381        Builder->CreateGEP(OrigBase, NewIndices);
1382        NGEP->takeName(GEP);
1383
1384        if (isa<BitCastInst>(CI))
1385          return new BitCastInst(NGEP, CI.getType());
1386        assert(isa<PtrToIntInst>(CI));
1387        return new PtrToIntInst(NGEP, CI.getType());
1388      }
1389    }
1390  }
1391
1392  return commonCastTransforms(CI);
1393}
1394
1395Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
1396  // If the destination integer type is not the intptr_t type for this target,
1397  // do a ptrtoint to intptr_t then do a trunc or zext.  This allows the cast
1398  // to be exposed to other transforms.
1399  if (TD && CI.getType()->getScalarSizeInBits() != TD->getPointerSizeInBits()) {
1400    Type *Ty = TD->getIntPtrType(CI.getContext());
1401    if (CI.getType()->isVectorTy()) // Handle vectors of pointers.
1402      Ty = VectorType::get(Ty, CI.getType()->getVectorNumElements());
1403
1404    Value *P = Builder->CreatePtrToInt(CI.getOperand(0), Ty);
1405    return CastInst::CreateIntegerCast(P, CI.getType(), /*isSigned=*/false);
1406  }
1407
1408  return commonPointerCastTransforms(CI);
1409}
1410
1411/// OptimizeVectorResize - This input value (which is known to have vector type)
1412/// is being zero extended or truncated to the specified vector type.  Try to
1413/// replace it with a shuffle (and vector/vector bitcast) if possible.
1414///
1415/// The source and destination vector types may have different element types.
1416static Instruction *OptimizeVectorResize(Value *InVal, VectorType *DestTy,
1417                                         InstCombiner &IC) {
1418  // We can only do this optimization if the output is a multiple of the input
1419  // element size, or the input is a multiple of the output element size.
1420  // Convert the input type to have the same element type as the output.
1421  VectorType *SrcTy = cast<VectorType>(InVal->getType());
1422
1423  if (SrcTy->getElementType() != DestTy->getElementType()) {
1424    // The input types don't need to be identical, but for now they must be the
1425    // same size.  There is no specific reason we couldn't handle things like
1426    // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
1427    // there yet.
1428    if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
1429        DestTy->getElementType()->getPrimitiveSizeInBits())
1430      return 0;
1431
1432    SrcTy = VectorType::get(DestTy->getElementType(), SrcTy->getNumElements());
1433    InVal = IC.Builder->CreateBitCast(InVal, SrcTy);
1434  }
1435
1436  // Now that the element types match, get the shuffle mask and RHS of the
1437  // shuffle to use, which depends on whether we're increasing or decreasing the
1438  // size of the input.
1439  SmallVector<uint32_t, 16> ShuffleMask;
1440  Value *V2;
1441
1442  if (SrcTy->getNumElements() > DestTy->getNumElements()) {
1443    // If we're shrinking the number of elements, just shuffle in the low
1444    // elements from the input and use undef as the second shuffle input.
1445    V2 = UndefValue::get(SrcTy);
1446    for (unsigned i = 0, e = DestTy->getNumElements(); i != e; ++i)
1447      ShuffleMask.push_back(i);
1448
1449  } else {
1450    // If we're increasing the number of elements, shuffle in all of the
1451    // elements from InVal and fill the rest of the result elements with zeros
1452    // from a constant zero.
1453    V2 = Constant::getNullValue(SrcTy);
1454    unsigned SrcElts = SrcTy->getNumElements();
1455    for (unsigned i = 0, e = SrcElts; i != e; ++i)
1456      ShuffleMask.push_back(i);
1457
1458    // The excess elements reference the first element of the zero input.
1459    for (unsigned i = 0, e = DestTy->getNumElements()-SrcElts; i != e; ++i)
1460      ShuffleMask.push_back(SrcElts);
1461  }
1462
1463  return new ShuffleVectorInst(InVal, V2,
1464                               ConstantDataVector::get(V2->getContext(),
1465                                                       ShuffleMask));
1466}
1467
1468static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
1469  return Value % Ty->getPrimitiveSizeInBits() == 0;
1470}
1471
1472static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
1473  return Value / Ty->getPrimitiveSizeInBits();
1474}
1475
1476/// CollectInsertionElements - V is a value which is inserted into a vector of
1477/// VecEltTy.  Look through the value to see if we can decompose it into
1478/// insertions into the vector.  See the example in the comment for
1479/// OptimizeIntegerToVectorInsertions for the pattern this handles.
1480/// The type of V is always a non-zero multiple of VecEltTy's size.
1481///
1482/// This returns false if the pattern can't be matched or true if it can,
1483/// filling in Elements with the elements found here.
1484static bool CollectInsertionElements(Value *V, unsigned ElementIndex,
1485                                     SmallVectorImpl<Value*> &Elements,
1486                                     Type *VecEltTy) {
1487  // Undef values never contribute useful bits to the result.
1488  if (isa<UndefValue>(V)) return true;
1489
1490  // If we got down to a value of the right type, we win, try inserting into the
1491  // right element.
1492  if (V->getType() == VecEltTy) {
1493    // Inserting null doesn't actually insert any elements.
1494    if (Constant *C = dyn_cast<Constant>(V))
1495      if (C->isNullValue())
1496        return true;
1497
1498    // Fail if multiple elements are inserted into this slot.
1499    if (ElementIndex >= Elements.size() || Elements[ElementIndex] != 0)
1500      return false;
1501
1502    Elements[ElementIndex] = V;
1503    return true;
1504  }
1505
1506  if (Constant *C = dyn_cast<Constant>(V)) {
1507    // Figure out the # elements this provides, and bitcast it or slice it up
1508    // as required.
1509    unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
1510                                        VecEltTy);
1511    // If the constant is the size of a vector element, we just need to bitcast
1512    // it to the right type so it gets properly inserted.
1513    if (NumElts == 1)
1514      return CollectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
1515                                      ElementIndex, Elements, VecEltTy);
1516
1517    // Okay, this is a constant that covers multiple elements.  Slice it up into
1518    // pieces and insert each element-sized piece into the vector.
1519    if (!isa<IntegerType>(C->getType()))
1520      C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
1521                                       C->getType()->getPrimitiveSizeInBits()));
1522    unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
1523    Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
1524
1525    for (unsigned i = 0; i != NumElts; ++i) {
1526      Constant *Piece = ConstantExpr::getLShr(C, ConstantInt::get(C->getType(),
1527                                                               i*ElementSize));
1528      Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
1529      if (!CollectInsertionElements(Piece, ElementIndex+i, Elements, VecEltTy))
1530        return false;
1531    }
1532    return true;
1533  }
1534
1535  if (!V->hasOneUse()) return false;
1536
1537  Instruction *I = dyn_cast<Instruction>(V);
1538  if (I == 0) return false;
1539  switch (I->getOpcode()) {
1540  default: return false; // Unhandled case.
1541  case Instruction::BitCast:
1542    return CollectInsertionElements(I->getOperand(0), ElementIndex,
1543                                    Elements, VecEltTy);
1544  case Instruction::ZExt:
1545    if (!isMultipleOfTypeSize(
1546                          I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
1547                              VecEltTy))
1548      return false;
1549    return CollectInsertionElements(I->getOperand(0), ElementIndex,
1550                                    Elements, VecEltTy);
1551  case Instruction::Or:
1552    return CollectInsertionElements(I->getOperand(0), ElementIndex,
1553                                    Elements, VecEltTy) &&
1554           CollectInsertionElements(I->getOperand(1), ElementIndex,
1555                                    Elements, VecEltTy);
1556  case Instruction::Shl: {
1557    // Must be shifting by a constant that is a multiple of the element size.
1558    ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
1559    if (CI == 0) return false;
1560    if (!isMultipleOfTypeSize(CI->getZExtValue(), VecEltTy)) return false;
1561    unsigned IndexShift = getTypeSizeIndex(CI->getZExtValue(), VecEltTy);
1562
1563    return CollectInsertionElements(I->getOperand(0), ElementIndex+IndexShift,
1564                                    Elements, VecEltTy);
1565  }
1566
1567  }
1568}
1569
1570
1571/// OptimizeIntegerToVectorInsertions - If the input is an 'or' instruction, we
1572/// may be doing shifts and ors to assemble the elements of the vector manually.
1573/// Try to rip the code out and replace it with insertelements.  This is to
1574/// optimize code like this:
1575///
1576///    %tmp37 = bitcast float %inc to i32
1577///    %tmp38 = zext i32 %tmp37 to i64
1578///    %tmp31 = bitcast float %inc5 to i32
1579///    %tmp32 = zext i32 %tmp31 to i64
1580///    %tmp33 = shl i64 %tmp32, 32
1581///    %ins35 = or i64 %tmp33, %tmp38
1582///    %tmp43 = bitcast i64 %ins35 to <2 x float>
1583///
1584/// Into two insertelements that do "buildvector{%inc, %inc5}".
1585static Value *OptimizeIntegerToVectorInsertions(BitCastInst &CI,
1586                                                InstCombiner &IC) {
1587  VectorType *DestVecTy = cast<VectorType>(CI.getType());
1588  Value *IntInput = CI.getOperand(0);
1589
1590  SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
1591  if (!CollectInsertionElements(IntInput, 0, Elements,
1592                                DestVecTy->getElementType()))
1593    return 0;
1594
1595  // If we succeeded, we know that all of the element are specified by Elements
1596  // or are zero if Elements has a null entry.  Recast this as a set of
1597  // insertions.
1598  Value *Result = Constant::getNullValue(CI.getType());
1599  for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
1600    if (Elements[i] == 0) continue;  // Unset element.
1601
1602    Result = IC.Builder->CreateInsertElement(Result, Elements[i],
1603                                             IC.Builder->getInt32(i));
1604  }
1605
1606  return Result;
1607}
1608
1609
1610/// OptimizeIntToFloatBitCast - See if we can optimize an integer->float/double
1611/// bitcast.  The various long double bitcasts can't get in here.
1612static Instruction *OptimizeIntToFloatBitCast(BitCastInst &CI,InstCombiner &IC){
1613  // We need to know the target byte order to perform this optimization.
1614  if (!IC.getDataLayout()) return 0;
1615
1616  Value *Src = CI.getOperand(0);
1617  Type *DestTy = CI.getType();
1618
1619  // If this is a bitcast from int to float, check to see if the int is an
1620  // extraction from a vector.
1621  Value *VecInput = 0;
1622  // bitcast(trunc(bitcast(somevector)))
1623  if (match(Src, m_Trunc(m_BitCast(m_Value(VecInput)))) &&
1624      isa<VectorType>(VecInput->getType())) {
1625    VectorType *VecTy = cast<VectorType>(VecInput->getType());
1626    unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1627
1628    if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0) {
1629      // If the element type of the vector doesn't match the result type,
1630      // bitcast it to be a vector type we can extract from.
1631      if (VecTy->getElementType() != DestTy) {
1632        VecTy = VectorType::get(DestTy,
1633                                VecTy->getPrimitiveSizeInBits() / DestWidth);
1634        VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1635      }
1636
1637      unsigned Elt = 0;
1638      if (IC.getDataLayout()->isBigEndian())
1639        Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1;
1640      return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1641    }
1642  }
1643
1644  // bitcast(trunc(lshr(bitcast(somevector), cst))
1645  ConstantInt *ShAmt = 0;
1646  if (match(Src, m_Trunc(m_LShr(m_BitCast(m_Value(VecInput)),
1647                                m_ConstantInt(ShAmt)))) &&
1648      isa<VectorType>(VecInput->getType())) {
1649    VectorType *VecTy = cast<VectorType>(VecInput->getType());
1650    unsigned DestWidth = DestTy->getPrimitiveSizeInBits();
1651    if (VecTy->getPrimitiveSizeInBits() % DestWidth == 0 &&
1652        ShAmt->getZExtValue() % DestWidth == 0) {
1653      // If the element type of the vector doesn't match the result type,
1654      // bitcast it to be a vector type we can extract from.
1655      if (VecTy->getElementType() != DestTy) {
1656        VecTy = VectorType::get(DestTy,
1657                                VecTy->getPrimitiveSizeInBits() / DestWidth);
1658        VecInput = IC.Builder->CreateBitCast(VecInput, VecTy);
1659      }
1660
1661      unsigned Elt = ShAmt->getZExtValue() / DestWidth;
1662      if (IC.getDataLayout()->isBigEndian())
1663        Elt = VecTy->getPrimitiveSizeInBits() / DestWidth - 1 - Elt;
1664      return ExtractElementInst::Create(VecInput, IC.Builder->getInt32(Elt));
1665    }
1666  }
1667  return 0;
1668}
1669
1670Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
1671  // If the operands are integer typed then apply the integer transforms,
1672  // otherwise just apply the common ones.
1673  Value *Src = CI.getOperand(0);
1674  Type *SrcTy = Src->getType();
1675  Type *DestTy = CI.getType();
1676
1677  // Get rid of casts from one type to the same type. These are useless and can
1678  // be replaced by the operand.
1679  if (DestTy == Src->getType())
1680    return ReplaceInstUsesWith(CI, Src);
1681
1682  if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
1683    PointerType *SrcPTy = cast<PointerType>(SrcTy);
1684    Type *DstElTy = DstPTy->getElementType();
1685    Type *SrcElTy = SrcPTy->getElementType();
1686
1687    // If the address spaces don't match, don't eliminate the bitcast, which is
1688    // required for changing types.
1689    if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
1690      return 0;
1691
1692    // If we are casting a alloca to a pointer to a type of the same
1693    // size, rewrite the allocation instruction to allocate the "right" type.
1694    // There is no need to modify malloc calls because it is their bitcast that
1695    // needs to be cleaned up.
1696    if (AllocaInst *AI = dyn_cast<AllocaInst>(Src))
1697      if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
1698        return V;
1699
1700    // If the source and destination are pointers, and this cast is equivalent
1701    // to a getelementptr X, 0, 0, 0...  turn it into the appropriate gep.
1702    // This can enhance SROA and other transforms that want type-safe pointers.
1703    Constant *ZeroUInt =
1704      Constant::getNullValue(Type::getInt32Ty(CI.getContext()));
1705    unsigned NumZeros = 0;
1706    while (SrcElTy != DstElTy &&
1707           isa<CompositeType>(SrcElTy) && !SrcElTy->isPointerTy() &&
1708           SrcElTy->getNumContainedTypes() /* not "{}" */) {
1709      SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
1710      ++NumZeros;
1711    }
1712
1713    // If we found a path from the src to dest, create the getelementptr now.
1714    if (SrcElTy == DstElTy) {
1715      SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
1716      return GetElementPtrInst::CreateInBounds(Src, Idxs);
1717    }
1718  }
1719
1720  // Try to optimize int -> float bitcasts.
1721  if ((DestTy->isFloatTy() || DestTy->isDoubleTy()) && isa<IntegerType>(SrcTy))
1722    if (Instruction *I = OptimizeIntToFloatBitCast(CI, *this))
1723      return I;
1724
1725  if (VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
1726    if (DestVTy->getNumElements() == 1 && !SrcTy->isVectorTy()) {
1727      Value *Elem = Builder->CreateBitCast(Src, DestVTy->getElementType());
1728      return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
1729                     Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1730      // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
1731    }
1732
1733    if (isa<IntegerType>(SrcTy)) {
1734      // If this is a cast from an integer to vector, check to see if the input
1735      // is a trunc or zext of a bitcast from vector.  If so, we can replace all
1736      // the casts with a shuffle and (potentially) a bitcast.
1737      if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
1738        CastInst *SrcCast = cast<CastInst>(Src);
1739        if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
1740          if (isa<VectorType>(BCIn->getOperand(0)->getType()))
1741            if (Instruction *I = OptimizeVectorResize(BCIn->getOperand(0),
1742                                               cast<VectorType>(DestTy), *this))
1743              return I;
1744      }
1745
1746      // If the input is an 'or' instruction, we may be doing shifts and ors to
1747      // assemble the elements of the vector manually.  Try to rip the code out
1748      // and replace it with insertelements.
1749      if (Value *V = OptimizeIntegerToVectorInsertions(CI, *this))
1750        return ReplaceInstUsesWith(CI, V);
1751    }
1752  }
1753
1754  if (VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
1755    if (SrcVTy->getNumElements() == 1) {
1756      // If our destination is not a vector, then make this a straight
1757      // scalar-scalar cast.
1758      if (!DestTy->isVectorTy()) {
1759        Value *Elem =
1760          Builder->CreateExtractElement(Src,
1761                     Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
1762        return CastInst::Create(Instruction::BitCast, Elem, DestTy);
1763      }
1764
1765      // Otherwise, see if our source is an insert. If so, then use the scalar
1766      // component directly.
1767      if (InsertElementInst *IEI =
1768            dyn_cast<InsertElementInst>(CI.getOperand(0)))
1769        return CastInst::Create(Instruction::BitCast, IEI->getOperand(1),
1770                                DestTy);
1771    }
1772  }
1773
1774  if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
1775    // Okay, we have (bitcast (shuffle ..)).  Check to see if this is
1776    // a bitcast to a vector with the same # elts.
1777    if (SVI->hasOneUse() && DestTy->isVectorTy() &&
1778        cast<VectorType>(DestTy)->getNumElements() ==
1779              SVI->getType()->getNumElements() &&
1780        SVI->getType()->getNumElements() ==
1781          cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
1782      BitCastInst *Tmp;
1783      // If either of the operands is a cast from CI.getType(), then
1784      // evaluating the shuffle in the casted destination's type will allow
1785      // us to eliminate at least one cast.
1786      if (((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(0))) &&
1787           Tmp->getOperand(0)->getType() == DestTy) ||
1788          ((Tmp = dyn_cast<BitCastInst>(SVI->getOperand(1))) &&
1789           Tmp->getOperand(0)->getType() == DestTy)) {
1790        Value *LHS = Builder->CreateBitCast(SVI->getOperand(0), DestTy);
1791        Value *RHS = Builder->CreateBitCast(SVI->getOperand(1), DestTy);
1792        // Return a new shuffle vector.  Use the same element ID's, as we
1793        // know the vector types match #elts.
1794        return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
1795      }
1796    }
1797  }
1798
1799  if (SrcTy->isPointerTy())
1800    return commonPointerCastTransforms(CI);
1801  return commonCastTransforms(CI);
1802}
1803