InstCombineCalls.cpp revision 165dac08d1bb8428b32a5f39cdd3dbee2888987f 
1//===- InstCombineCalls.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 visitCall and visitInvoke functions.
11//
12//===----------------------------------------------------------------------===//
13
14#include "InstCombine.h"
15#include "llvm/IntrinsicInst.h"
16#include "llvm/Support/CallSite.h"
17#include "llvm/Target/TargetData.h"
18#include "llvm/Analysis/MemoryBuiltins.h"
19#include "llvm/Transforms/Utils/BuildLibCalls.h"
20using namespace llvm;
21
22/// getPromotedType - Return the specified type promoted as it would be to pass
23/// though a va_arg area.
24static const Type *getPromotedType(const Type *Ty) {
25  if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
26    if (ITy->getBitWidth() < 32)
27      return Type::getInt32Ty(Ty->getContext());
28  }
29  return Ty;
30}
31
32/// EnforceKnownAlignment - If the specified pointer points to an object that
33/// we control, modify the object's alignment to PrefAlign. This isn't
34/// often possible though. If alignment is important, a more reliable approach
35/// is to simply align all global variables and allocation instructions to
36/// their preferred alignment from the beginning.
37///
38static unsigned EnforceKnownAlignment(Value *V,
39                                      unsigned Align, unsigned PrefAlign) {
40
41  User *U = dyn_cast<User>(V);
42  if (!U) return Align;
43
44  switch (Operator::getOpcode(U)) {
45  default: break;
46  case Instruction::BitCast:
47    return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
48  case Instruction::GetElementPtr: {
49    // If all indexes are zero, it is just the alignment of the base pointer.
50    bool AllZeroOperands = true;
51    for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
52      if (!isa<Constant>(*i) ||
53          !cast<Constant>(*i)->isNullValue()) {
54        AllZeroOperands = false;
55        break;
56      }
57
58    if (AllZeroOperands) {
59      // Treat this like a bitcast.
60      return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
61    }
62    break;
63  }
64  }
65
66  if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
67    // If there is a large requested alignment and we can, bump up the alignment
68    // of the global.
69    if (!GV->isDeclaration()) {
70      if (GV->getAlignment() >= PrefAlign)
71        Align = GV->getAlignment();
72      else {
73        GV->setAlignment(PrefAlign);
74        Align = PrefAlign;
75      }
76    }
77  } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
78    // If there is a requested alignment and if this is an alloca, round up.
79    if (AI->getAlignment() >= PrefAlign)
80      Align = AI->getAlignment();
81    else {
82      AI->setAlignment(PrefAlign);
83      Align = PrefAlign;
84    }
85  }
86
87  return Align;
88}
89
90/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
91/// we can determine, return it, otherwise return 0.  If PrefAlign is specified,
92/// and it is more than the alignment of the ultimate object, see if we can
93/// increase the alignment of the ultimate object, making this check succeed.
94unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
95                                                  unsigned PrefAlign) {
96  unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
97                      sizeof(PrefAlign) * CHAR_BIT;
98  APInt Mask = APInt::getAllOnesValue(BitWidth);
99  APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
100  ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
101  unsigned TrailZ = KnownZero.countTrailingOnes();
102  unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
103
104  if (PrefAlign > Align)
105    Align = EnforceKnownAlignment(V, Align, PrefAlign);
106
107    // We don't need to make any adjustment.
108  return Align;
109}
110
111Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
112  unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(0));
113  unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
114  unsigned MinAlign = std::min(DstAlign, SrcAlign);
115  unsigned CopyAlign = MI->getAlignment();
116
117  if (CopyAlign < MinAlign) {
118    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
119                                             MinAlign, false));
120    return MI;
121  }
122
123  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
124  // load/store.
125  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(2));
126  if (MemOpLength == 0) return 0;
127
128  // Source and destination pointer types are always "i8*" for intrinsic.  See
129  // if the size is something we can handle with a single primitive load/store.
130  // A single load+store correctly handles overlapping memory in the memmove
131  // case.
132  unsigned Size = MemOpLength->getZExtValue();
133  if (Size == 0) return MI;  // Delete this mem transfer.
134
135  if (Size > 8 || (Size&(Size-1)))
136    return 0;  // If not 1/2/4/8 bytes, exit.
137
138  // Use an integer load+store unless we can find something better.
139  unsigned SrcAddrSp =
140    cast<PointerType>(MI->getOperand(1)->getType())->getAddressSpace();
141  unsigned DstAddrSp =
142    cast<PointerType>(MI->getOperand(0)->getType())->getAddressSpace();
143
144  const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
145  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
146  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
147
148  // Memcpy forces the use of i8* for the source and destination.  That means
149  // that if you're using memcpy to move one double around, you'll get a cast
150  // from double* to i8*.  We'd much rather use a double load+store rather than
151  // an i64 load+store, here because this improves the odds that the source or
152  // dest address will be promotable.  See if we can find a better type than the
153  // integer datatype.
154  Value *StrippedDest = MI->getOperand(0)->stripPointerCasts();
155  if (StrippedDest != MI->getOperand(0)) {
156    const Type *SrcETy = cast<PointerType>(StrippedDest->getType())
157                                    ->getElementType();
158    if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
159      // The SrcETy might be something like {{{double}}} or [1 x double].  Rip
160      // down through these levels if so.
161      while (!SrcETy->isSingleValueType()) {
162        if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
163          if (STy->getNumElements() == 1)
164            SrcETy = STy->getElementType(0);
165          else
166            break;
167        } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
168          if (ATy->getNumElements() == 1)
169            SrcETy = ATy->getElementType();
170          else
171            break;
172        } else
173          break;
174      }
175
176      if (SrcETy->isSingleValueType()) {
177        NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
178        NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
179      }
180    }
181  }
182
183
184  // If the memcpy/memmove provides better alignment info than we can
185  // infer, use it.
186  SrcAlign = std::max(SrcAlign, CopyAlign);
187  DstAlign = std::max(DstAlign, CopyAlign);
188
189  Value *Src = Builder->CreateBitCast(MI->getOperand(1), NewSrcPtrTy);
190  Value *Dest = Builder->CreateBitCast(MI->getOperand(0), NewDstPtrTy);
191  Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign);
192  InsertNewInstBefore(L, *MI);
193  InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign),
194                      *MI);
195
196  // Set the size of the copy to 0, it will be deleted on the next iteration.
197  MI->setOperand(2, Constant::getNullValue(MemOpLength->getType()));
198  return MI;
199}
200
201Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
202  unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
203  if (MI->getAlignment() < Alignment) {
204    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
205                                             Alignment, false));
206    return MI;
207  }
208
209  // Extract the length and alignment and fill if they are constant.
210  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
211  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
212  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
213    return 0;
214  uint64_t Len = LenC->getZExtValue();
215  Alignment = MI->getAlignment();
216
217  // If the length is zero, this is a no-op
218  if (Len == 0) return MI; // memset(d,c,0,a) -> noop
219
220  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
221  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
222    const Type *ITy = IntegerType::get(MI->getContext(), Len*8);  // n=1 -> i8.
223
224    Value *Dest = MI->getDest();
225    Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
226
227    // Alignment 0 is identity for alignment 1 for memset, but not store.
228    if (Alignment == 0) Alignment = 1;
229
230    // Extract the fill value and store.
231    uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
232    InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
233                                      Dest, false, Alignment), *MI);
234
235    // Set the size of the copy to 0, it will be deleted on the next iteration.
236    MI->setLength(Constant::getNullValue(LenC->getType()));
237    return MI;
238  }
239
240  return 0;
241}
242
243/// visitCallInst - CallInst simplification.  This mostly only handles folding
244/// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
245/// the heavy lifting.
246///
247Instruction *InstCombiner::visitCallInst(CallInst &CI) {
248  if (isFreeCall(&CI))
249    return visitFree(CI);
250
251  // If the caller function is nounwind, mark the call as nounwind, even if the
252  // callee isn't.
253  if (CI.getParent()->getParent()->doesNotThrow() &&
254      !CI.doesNotThrow()) {
255    CI.setDoesNotThrow();
256    return &CI;
257  }
258
259  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
260  if (!II) return visitCallSite(&CI);
261
262  // Intrinsics cannot occur in an invoke, so handle them here instead of in
263  // visitCallSite.
264  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
265    bool Changed = false;
266
267    // memmove/cpy/set of zero bytes is a noop.
268    if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
269      if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
270
271      if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
272        if (CI->getZExtValue() == 1) {
273          // Replace the instruction with just byte operations.  We would
274          // transform other cases to loads/stores, but we don't know if
275          // alignment is sufficient.
276        }
277    }
278
279    // If we have a memmove and the source operation is a constant global,
280    // then the source and dest pointers can't alias, so we can change this
281    // into a call to memcpy.
282    if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
283      if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
284        if (GVSrc->isConstant()) {
285          Module *M = MMI->getParent()->getParent()->getParent();
286          Intrinsic::ID MemCpyID = Intrinsic::memcpy;
287          const Type *Tys[3] = { CI.getOperand(0)->getType(),
288                                 CI.getOperand(1)->getType(),
289                                 CI.getOperand(2)->getType() };
290          MMI->setCalledFunction(
291                        Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
292          Changed = true;
293        }
294    }
295
296    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
297      // memmove(x,x,size) -> noop.
298      if (MTI->getSource() == MTI->getDest())
299        return EraseInstFromFunction(CI);
300
301      // If we can determine a pointer alignment that is bigger than currently
302      // set, update the alignment.
303      if (Instruction *I = SimplifyMemTransfer(MTI))
304        return I;
305    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
306      if (Instruction *I = SimplifyMemSet(MSI))
307        return I;
308    }
309
310    if (Changed) return II;
311  }
312
313  switch (II->getIntrinsicID()) {
314  default: break;
315  case Intrinsic::objectsize: {
316    // We need target data for just about everything so depend on it.
317    if (!TD) break;
318
319    const Type *ReturnTy = CI.getType();
320    bool Min = (cast<ConstantInt>(II->getOperand(1))->getZExtValue() == 1);
321
322    // Get to the real allocated thing and offset as fast as possible.
323    Value *Op1 = II->getOperand(0)->stripPointerCasts();
324
325    // If we've stripped down to a single global variable that we
326    // can know the size of then just return that.
327    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
328      if (GV->hasDefinitiveInitializer()) {
329        Constant *C = GV->getInitializer();
330        uint64_t GlobalSize = TD->getTypeAllocSize(C->getType());
331        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize));
332      } else {
333        // Can't determine size of the GV.
334        Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
335        return ReplaceInstUsesWith(CI, RetVal);
336      }
337    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
338      // Get alloca size.
339      if (AI->getAllocatedType()->isSized()) {
340        uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
341        if (AI->isArrayAllocation()) {
342          const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
343          if (!C) break;
344          AllocaSize *= C->getZExtValue();
345        }
346        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize));
347      }
348    } else if (CallInst *MI = extractMallocCall(Op1)) {
349      const Type* MallocType = getMallocAllocatedType(MI);
350      // Get alloca size.
351      if (MallocType && MallocType->isSized()) {
352        if (Value *NElems = getMallocArraySize(MI, TD, true)) {
353          if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
354        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy,
355               (NElements->getZExtValue() * TD->getTypeAllocSize(MallocType))));
356        }
357      }
358    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) {
359      // Only handle constant GEPs here.
360      if (CE->getOpcode() != Instruction::GetElementPtr) break;
361      GEPOperator *GEP = cast<GEPOperator>(CE);
362
363      // Make sure we're not a constant offset from an external
364      // global.
365      Value *Operand = GEP->getPointerOperand();
366      Operand = Operand->stripPointerCasts();
367      if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand))
368        if (!GV->hasDefinitiveInitializer()) break;
369
370      // Get what we're pointing to and its size.
371      const PointerType *BaseType =
372        cast<PointerType>(Operand->getType());
373      uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType());
374
375      // Get the current byte offset into the thing. Use the original
376      // operand in case we're looking through a bitcast.
377      SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end());
378      const PointerType *OffsetType =
379        cast<PointerType>(GEP->getPointerOperand()->getType());
380      uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size());
381
382      if (Size < Offset) {
383        // Out of bound reference? Negative index normalized to large
384        // index? Just return "I don't know".
385        Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
386        return ReplaceInstUsesWith(CI, RetVal);
387      }
388
389      Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset);
390      return ReplaceInstUsesWith(CI, RetVal);
391    }
392
393    // Do not return "I don't know" here. Later optimization passes could
394    // make it possible to evaluate objectsize to a constant.
395    break;
396  }
397  case Intrinsic::bswap:
398    // bswap(bswap(x)) -> x
399    if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(0)))
400      if (Operand->getIntrinsicID() == Intrinsic::bswap)
401        return ReplaceInstUsesWith(CI, Operand->getOperand(0));
402
403    // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
404    if (TruncInst *TI = dyn_cast<TruncInst>(II->getOperand(0))) {
405      if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
406        if (Operand->getIntrinsicID() == Intrinsic::bswap) {
407          unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
408                       TI->getType()->getPrimitiveSizeInBits();
409          Value *CV = ConstantInt::get(Operand->getType(), C);
410          Value *V = Builder->CreateLShr(Operand->getOperand(0), CV);
411          return new TruncInst(V, TI->getType());
412        }
413    }
414
415    break;
416  case Intrinsic::powi:
417    if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getOperand(1))) {
418      // powi(x, 0) -> 1.0
419      if (Power->isZero())
420        return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
421      // powi(x, 1) -> x
422      if (Power->isOne())
423        return ReplaceInstUsesWith(CI, II->getOperand(0));
424      // powi(x, -1) -> 1/x
425      if (Power->isAllOnesValue())
426        return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
427                                          II->getOperand(0));
428    }
429    break;
430  case Intrinsic::cttz: {
431    // If all bits below the first known one are known zero,
432    // this value is constant.
433    const IntegerType *IT = cast<IntegerType>(II->getOperand(0)->getType());
434    uint32_t BitWidth = IT->getBitWidth();
435    APInt KnownZero(BitWidth, 0);
436    APInt KnownOne(BitWidth, 0);
437    ComputeMaskedBits(II->getOperand(0), APInt::getAllOnesValue(BitWidth),
438                      KnownZero, KnownOne);
439    unsigned TrailingZeros = KnownOne.countTrailingZeros();
440    APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
441    if ((Mask & KnownZero) == Mask)
442      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
443                                 APInt(BitWidth, TrailingZeros)));
444
445    }
446    break;
447  case Intrinsic::ctlz: {
448    // If all bits above the first known one are known zero,
449    // this value is constant.
450    const IntegerType *IT = cast<IntegerType>(II->getOperand(0)->getType());
451    uint32_t BitWidth = IT->getBitWidth();
452    APInt KnownZero(BitWidth, 0);
453    APInt KnownOne(BitWidth, 0);
454    ComputeMaskedBits(II->getOperand(0), APInt::getAllOnesValue(BitWidth),
455                      KnownZero, KnownOne);
456    unsigned LeadingZeros = KnownOne.countLeadingZeros();
457    APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
458    if ((Mask & KnownZero) == Mask)
459      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
460                                 APInt(BitWidth, LeadingZeros)));
461
462    }
463    break;
464  case Intrinsic::uadd_with_overflow: {
465    Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
466    const IntegerType *IT = cast<IntegerType>(II->getOperand(0)->getType());
467    uint32_t BitWidth = IT->getBitWidth();
468    APInt Mask = APInt::getSignBit(BitWidth);
469    APInt LHSKnownZero(BitWidth, 0);
470    APInt LHSKnownOne(BitWidth, 0);
471    ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
472    bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
473    bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
474
475    if (LHSKnownNegative || LHSKnownPositive) {
476      APInt RHSKnownZero(BitWidth, 0);
477      APInt RHSKnownOne(BitWidth, 0);
478      ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
479      bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
480      bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
481      if (LHSKnownNegative && RHSKnownNegative) {
482        // The sign bit is set in both cases: this MUST overflow.
483        // Create a simple add instruction, and insert it into the struct.
484        Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
485        Worklist.Add(Add);
486        Constant *V[] = {
487          UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
488        };
489        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
490        return InsertValueInst::Create(Struct, Add, 0);
491      }
492
493      if (LHSKnownPositive && RHSKnownPositive) {
494        // The sign bit is clear in both cases: this CANNOT overflow.
495        // Create a simple add instruction, and insert it into the struct.
496        Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
497        Worklist.Add(Add);
498        Constant *V[] = {
499          UndefValue::get(LHS->getType()),
500          ConstantInt::getFalse(II->getContext())
501        };
502        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
503        return InsertValueInst::Create(Struct, Add, 0);
504      }
505    }
506  }
507  // FALL THROUGH uadd into sadd
508  case Intrinsic::sadd_with_overflow:
509    // Canonicalize constants into the RHS.
510    if (isa<Constant>(II->getOperand(0)) &&
511        !isa<Constant>(II->getOperand(1))) {
512      Value *LHS = II->getOperand(0);
513      II->setOperand(0, II->getOperand(1));
514      II->setOperand(1, LHS);
515      return II;
516    }
517
518    // X + undef -> undef
519    if (isa<UndefValue>(II->getOperand(1)))
520      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
521
522    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(1))) {
523      // X + 0 -> {X, false}
524      if (RHS->isZero()) {
525        Constant *V[] = {
526          UndefValue::get(II->getOperand(0)->getType()),
527          ConstantInt::getFalse(II->getContext())
528        };
529        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
530        return InsertValueInst::Create(Struct, II->getOperand(0), 0);
531      }
532    }
533    break;
534  case Intrinsic::usub_with_overflow:
535  case Intrinsic::ssub_with_overflow:
536    // undef - X -> undef
537    // X - undef -> undef
538    if (isa<UndefValue>(II->getOperand(0)) ||
539        isa<UndefValue>(II->getOperand(1)))
540      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
541
542    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(1))) {
543      // X - 0 -> {X, false}
544      if (RHS->isZero()) {
545        Constant *V[] = {
546          UndefValue::get(II->getOperand(0)->getType()),
547          ConstantInt::getFalse(II->getContext())
548        };
549        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
550        return InsertValueInst::Create(Struct, II->getOperand(0), 0);
551      }
552    }
553    break;
554  case Intrinsic::umul_with_overflow:
555  case Intrinsic::smul_with_overflow:
556    // Canonicalize constants into the RHS.
557    if (isa<Constant>(II->getOperand(0)) &&
558        !isa<Constant>(II->getOperand(1))) {
559      Value *LHS = II->getOperand(0);
560      II->setOperand(0, II->getOperand(1));
561      II->setOperand(1, LHS);
562      return II;
563    }
564
565    // X * undef -> undef
566    if (isa<UndefValue>(II->getOperand(1)))
567      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
568
569    if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getOperand(1))) {
570      // X*0 -> {0, false}
571      if (RHSI->isZero())
572        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
573
574      // X * 1 -> {X, false}
575      if (RHSI->equalsInt(1)) {
576        Constant *V[] = {
577          UndefValue::get(II->getOperand(0)->getType()),
578          ConstantInt::getFalse(II->getContext())
579        };
580        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
581        return InsertValueInst::Create(Struct, II->getOperand(0), 0);
582      }
583    }
584    break;
585  case Intrinsic::ppc_altivec_lvx:
586  case Intrinsic::ppc_altivec_lvxl:
587  case Intrinsic::x86_sse_loadu_ps:
588  case Intrinsic::x86_sse2_loadu_pd:
589  case Intrinsic::x86_sse2_loadu_dq:
590    // Turn PPC lvx     -> load if the pointer is known aligned.
591    // Turn X86 loadups -> load if the pointer is known aligned.
592    if (GetOrEnforceKnownAlignment(II->getOperand(0), 16) >= 16) {
593      Value *Ptr = Builder->CreateBitCast(II->getOperand(0),
594                                         PointerType::getUnqual(II->getType()));
595      return new LoadInst(Ptr);
596    }
597    break;
598  case Intrinsic::ppc_altivec_stvx:
599  case Intrinsic::ppc_altivec_stvxl:
600    // Turn stvx -> store if the pointer is known aligned.
601    if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
602      const Type *OpPtrTy =
603        PointerType::getUnqual(II->getOperand(0)->getType());
604      Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
605      return new StoreInst(II->getOperand(0), Ptr);
606    }
607    break;
608  case Intrinsic::x86_sse_storeu_ps:
609  case Intrinsic::x86_sse2_storeu_pd:
610  case Intrinsic::x86_sse2_storeu_dq:
611    // Turn X86 storeu -> store if the pointer is known aligned.
612    if (GetOrEnforceKnownAlignment(II->getOperand(0), 16) >= 16) {
613      const Type *OpPtrTy =
614        PointerType::getUnqual(II->getOperand(1)->getType());
615      Value *Ptr = Builder->CreateBitCast(II->getOperand(0), OpPtrTy);
616      return new StoreInst(II->getOperand(1), Ptr);
617    }
618    break;
619
620  case Intrinsic::x86_sse_cvttss2si: {
621    // These intrinsics only demands the 0th element of its input vector.  If
622    // we can simplify the input based on that, do so now.
623    unsigned VWidth =
624      cast<VectorType>(II->getOperand(0)->getType())->getNumElements();
625    APInt DemandedElts(VWidth, 1);
626    APInt UndefElts(VWidth, 0);
627    if (Value *V = SimplifyDemandedVectorElts(II->getOperand(0), DemandedElts,
628                                              UndefElts)) {
629      II->setOperand(0, V);
630      return II;
631    }
632    break;
633  }
634
635  case Intrinsic::ppc_altivec_vperm:
636    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
637    if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(2))) {
638      assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
639
640      // Check that all of the elements are integer constants or undefs.
641      bool AllEltsOk = true;
642      for (unsigned i = 0; i != 16; ++i) {
643        if (!isa<ConstantInt>(Mask->getOperand(i)) &&
644            !isa<UndefValue>(Mask->getOperand(i))) {
645          AllEltsOk = false;
646          break;
647        }
648      }
649
650      if (AllEltsOk) {
651        // Cast the input vectors to byte vectors.
652        Value *Op0 = Builder->CreateBitCast(II->getOperand(0), Mask->getType());
653        Value *Op1 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
654        Value *Result = UndefValue::get(Op0->getType());
655
656        // Only extract each element once.
657        Value *ExtractedElts[32];
658        memset(ExtractedElts, 0, sizeof(ExtractedElts));
659
660        for (unsigned i = 0; i != 16; ++i) {
661          if (isa<UndefValue>(Mask->getOperand(i)))
662            continue;
663          unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
664          Idx &= 31;  // Match the hardware behavior.
665
666          if (ExtractedElts[Idx] == 0) {
667            ExtractedElts[Idx] =
668              Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
669                  ConstantInt::get(Type::getInt32Ty(II->getContext()),
670                                   Idx&15, false), "tmp");
671          }
672
673          // Insert this value into the result vector.
674          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
675                         ConstantInt::get(Type::getInt32Ty(II->getContext()),
676                                          i, false), "tmp");
677        }
678        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
679      }
680    }
681    break;
682
683  case Intrinsic::stackrestore: {
684    // If the save is right next to the restore, remove the restore.  This can
685    // happen when variable allocas are DCE'd.
686    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(0))) {
687      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
688        BasicBlock::iterator BI = SS;
689        if (&*++BI == II)
690          return EraseInstFromFunction(CI);
691      }
692    }
693
694    // Scan down this block to see if there is another stack restore in the
695    // same block without an intervening call/alloca.
696    BasicBlock::iterator BI = II;
697    TerminatorInst *TI = II->getParent()->getTerminator();
698    bool CannotRemove = false;
699    for (++BI; &*BI != TI; ++BI) {
700      if (isa<AllocaInst>(BI) || isMalloc(BI)) {
701        CannotRemove = true;
702        break;
703      }
704      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
705        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
706          // If there is a stackrestore below this one, remove this one.
707          if (II->getIntrinsicID() == Intrinsic::stackrestore)
708            return EraseInstFromFunction(CI);
709          // Otherwise, ignore the intrinsic.
710        } else {
711          // If we found a non-intrinsic call, we can't remove the stack
712          // restore.
713          CannotRemove = true;
714          break;
715        }
716      }
717    }
718
719    // If the stack restore is in a return/unwind block and if there are no
720    // allocas or calls between the restore and the return, nuke the restore.
721    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
722      return EraseInstFromFunction(CI);
723    break;
724  }
725  }
726
727  return visitCallSite(II);
728}
729
730// InvokeInst simplification
731//
732Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
733  return visitCallSite(&II);
734}
735
736/// isSafeToEliminateVarargsCast - If this cast does not affect the value
737/// passed through the varargs area, we can eliminate the use of the cast.
738static bool isSafeToEliminateVarargsCast(const CallSite CS,
739                                         const CastInst * const CI,
740                                         const TargetData * const TD,
741                                         const int ix) {
742  if (!CI->isLosslessCast())
743    return false;
744
745  // The size of ByVal arguments is derived from the type, so we
746  // can't change to a type with a different size.  If the size were
747  // passed explicitly we could avoid this check.
748  if (!CS.paramHasAttr(ix, Attribute::ByVal))
749    return true;
750
751  const Type* SrcTy =
752            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
753  const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
754  if (!SrcTy->isSized() || !DstTy->isSized())
755    return false;
756  if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
757    return false;
758  return true;
759}
760
761namespace {
762class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
763  InstCombiner *IC;
764protected:
765  void replaceCall(Value *With) {
766    NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
767  }
768  bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
769    if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getOperand(SizeCIOp))) {
770      if (SizeCI->isAllOnesValue())
771        return true;
772      if (isString)
773        return SizeCI->getZExtValue() >=
774               GetStringLength(CI->getOperand(SizeArgOp));
775      if (ConstantInt *Arg = dyn_cast<ConstantInt>(CI->getOperand(SizeArgOp)))
776        return SizeCI->getZExtValue() >= Arg->getZExtValue();
777    }
778    return false;
779  }
780public:
781  InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
782  Instruction *NewInstruction;
783};
784} // end anonymous namespace
785
786// Try to fold some different type of calls here.
787// Currently we're only working with the checking functions, memcpy_chk,
788// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
789// strcat_chk and strncat_chk.
790Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
791  if (CI->getCalledFunction() == 0) return 0;
792
793  InstCombineFortifiedLibCalls Simplifier(this);
794  Simplifier.fold(CI, TD);
795  return Simplifier.NewInstruction;
796}
797
798// visitCallSite - Improvements for call and invoke instructions.
799//
800Instruction *InstCombiner::visitCallSite(CallSite CS) {
801  bool Changed = false;
802
803  // If the callee is a constexpr cast of a function, attempt to move the cast
804  // to the arguments of the call/invoke.
805  if (transformConstExprCastCall(CS)) return 0;
806
807  Value *Callee = CS.getCalledValue();
808
809  if (Function *CalleeF = dyn_cast<Function>(Callee))
810    // If the call and callee calling conventions don't match, this call must
811    // be unreachable, as the call is undefined.
812    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
813        // Only do this for calls to a function with a body.  A prototype may
814        // not actually end up matching the implementation's calling conv for a
815        // variety of reasons (e.g. it may be written in assembly).
816        !CalleeF->isDeclaration()) {
817      Instruction *OldCall = CS.getInstruction();
818      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
819                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
820                                  OldCall);
821      // If OldCall dues not return void then replaceAllUsesWith undef.
822      // This allows ValueHandlers and custom metadata to adjust itself.
823      if (!OldCall->getType()->isVoidTy())
824        OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
825      if (isa<CallInst>(OldCall))
826        return EraseInstFromFunction(*OldCall);
827
828      // We cannot remove an invoke, because it would change the CFG, just
829      // change the callee to a null pointer.
830      cast<InvokeInst>(OldCall)->setCalledFunction(
831                                    Constant::getNullValue(CalleeF->getType()));
832      return 0;
833    }
834
835  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
836    // This instruction is not reachable, just remove it.  We insert a store to
837    // undef so that we know that this code is not reachable, despite the fact
838    // that we can't modify the CFG here.
839    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
840               UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
841                  CS.getInstruction());
842
843    // If CS does not return void then replaceAllUsesWith undef.
844    // This allows ValueHandlers and custom metadata to adjust itself.
845    if (!CS.getInstruction()->getType()->isVoidTy())
846      CS.getInstruction()->
847        replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
848
849    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
850      // Don't break the CFG, insert a dummy cond branch.
851      BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
852                         ConstantInt::getTrue(Callee->getContext()), II);
853    }
854    return EraseInstFromFunction(*CS.getInstruction());
855  }
856
857  if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
858    if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
859      if (In->getIntrinsicID() == Intrinsic::init_trampoline)
860        return transformCallThroughTrampoline(CS);
861
862  const PointerType *PTy = cast<PointerType>(Callee->getType());
863  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
864  if (FTy->isVarArg()) {
865    int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
866    // See if we can optimize any arguments passed through the varargs area of
867    // the call.
868    for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
869           E = CS.arg_end(); I != E; ++I, ++ix) {
870      CastInst *CI = dyn_cast<CastInst>(*I);
871      if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
872        *I = CI->getOperand(0);
873        Changed = true;
874      }
875    }
876  }
877
878  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
879    // Inline asm calls cannot throw - mark them 'nounwind'.
880    CS.setDoesNotThrow();
881    Changed = true;
882  }
883
884  // Try to optimize the call if possible, we require TargetData for most of
885  // this.  None of these calls are seen as possibly dead so go ahead and
886  // delete the instruction now.
887  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
888    Instruction *I = tryOptimizeCall(CI, TD);
889    // If we changed something return the result, etc. Otherwise let
890    // the fallthrough check.
891    if (I) return EraseInstFromFunction(*I);
892  }
893
894  return Changed ? CS.getInstruction() : 0;
895}
896
897// transformConstExprCastCall - If the callee is a constexpr cast of a function,
898// attempt to move the cast to the arguments of the call/invoke.
899//
900bool InstCombiner::transformConstExprCastCall(CallSite CS) {
901  if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
902  ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
903  if (CE->getOpcode() != Instruction::BitCast ||
904      !isa<Function>(CE->getOperand(0)))
905    return false;
906  Function *Callee = cast<Function>(CE->getOperand(0));
907  Instruction *Caller = CS.getInstruction();
908  const AttrListPtr &CallerPAL = CS.getAttributes();
909
910  // Okay, this is a cast from a function to a different type.  Unless doing so
911  // would cause a type conversion of one of our arguments, change this call to
912  // be a direct call with arguments casted to the appropriate types.
913  //
914  const FunctionType *FT = Callee->getFunctionType();
915  const Type *OldRetTy = Caller->getType();
916  const Type *NewRetTy = FT->getReturnType();
917
918  if (NewRetTy->isStructTy())
919    return false; // TODO: Handle multiple return values.
920
921  // Check to see if we are changing the return type...
922  if (OldRetTy != NewRetTy) {
923    if (Callee->isDeclaration() &&
924        // Conversion is ok if changing from one pointer type to another or from
925        // a pointer to an integer of the same size.
926        !((OldRetTy->isPointerTy() || !TD ||
927           OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
928          (NewRetTy->isPointerTy() || !TD ||
929           NewRetTy == TD->getIntPtrType(Caller->getContext()))))
930      return false;   // Cannot transform this return value.
931
932    if (!Caller->use_empty() &&
933        // void -> non-void is handled specially
934        !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
935      return false;   // Cannot transform this return value.
936
937    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
938      Attributes RAttrs = CallerPAL.getRetAttributes();
939      if (RAttrs & Attribute::typeIncompatible(NewRetTy))
940        return false;   // Attribute not compatible with transformed value.
941    }
942
943    // If the callsite is an invoke instruction, and the return value is used by
944    // a PHI node in a successor, we cannot change the return type of the call
945    // because there is no place to put the cast instruction (without breaking
946    // the critical edge).  Bail out in this case.
947    if (!Caller->use_empty())
948      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
949        for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
950             UI != E; ++UI)
951          if (PHINode *PN = dyn_cast<PHINode>(*UI))
952            if (PN->getParent() == II->getNormalDest() ||
953                PN->getParent() == II->getUnwindDest())
954              return false;
955  }
956
957  unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
958  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
959
960  CallSite::arg_iterator AI = CS.arg_begin();
961  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
962    const Type *ParamTy = FT->getParamType(i);
963    const Type *ActTy = (*AI)->getType();
964
965    if (!CastInst::isCastable(ActTy, ParamTy))
966      return false;   // Cannot transform this parameter value.
967
968    if (CallerPAL.getParamAttributes(i + 1)
969        & Attribute::typeIncompatible(ParamTy))
970      return false;   // Attribute not compatible with transformed value.
971
972    // Converting from one pointer type to another or between a pointer and an
973    // integer of the same size is safe even if we do not have a body.
974    bool isConvertible = ActTy == ParamTy ||
975      (TD && ((ParamTy->isPointerTy() ||
976      ParamTy == TD->getIntPtrType(Caller->getContext())) &&
977              (ActTy->isPointerTy() ||
978              ActTy == TD->getIntPtrType(Caller->getContext()))));
979    if (Callee->isDeclaration() && !isConvertible) return false;
980  }
981
982  if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
983      Callee->isDeclaration())
984    return false;   // Do not delete arguments unless we have a function body.
985
986  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
987      !CallerPAL.isEmpty())
988    // In this case we have more arguments than the new function type, but we
989    // won't be dropping them.  Check that these extra arguments have attributes
990    // that are compatible with being a vararg call argument.
991    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
992      if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
993        break;
994      Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
995      if (PAttrs & Attribute::VarArgsIncompatible)
996        return false;
997    }
998
999  // Okay, we decided that this is a safe thing to do: go ahead and start
1000  // inserting cast instructions as necessary...
1001  std::vector<Value*> Args;
1002  Args.reserve(NumActualArgs);
1003  SmallVector<AttributeWithIndex, 8> attrVec;
1004  attrVec.reserve(NumCommonArgs);
1005
1006  // Get any return attributes.
1007  Attributes RAttrs = CallerPAL.getRetAttributes();
1008
1009  // If the return value is not being used, the type may not be compatible
1010  // with the existing attributes.  Wipe out any problematic attributes.
1011  RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1012
1013  // Add the new return attributes.
1014  if (RAttrs)
1015    attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1016
1017  AI = CS.arg_begin();
1018  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1019    const Type *ParamTy = FT->getParamType(i);
1020    if ((*AI)->getType() == ParamTy) {
1021      Args.push_back(*AI);
1022    } else {
1023      Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1024          false, ParamTy, false);
1025      Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1026    }
1027
1028    // Add any parameter attributes.
1029    if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1030      attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1031  }
1032
1033  // If the function takes more arguments than the call was taking, add them
1034  // now.
1035  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1036    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1037
1038  // If we are removing arguments to the function, emit an obnoxious warning.
1039  if (FT->getNumParams() < NumActualArgs) {
1040    if (!FT->isVarArg()) {
1041      errs() << "WARNING: While resolving call to function '"
1042             << Callee->getName() << "' arguments were dropped!\n";
1043    } else {
1044      // Add all of the arguments in their promoted form to the arg list.
1045      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1046        const Type *PTy = getPromotedType((*AI)->getType());
1047        if (PTy != (*AI)->getType()) {
1048          // Must promote to pass through va_arg area!
1049          Instruction::CastOps opcode =
1050            CastInst::getCastOpcode(*AI, false, PTy, false);
1051          Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1052        } else {
1053          Args.push_back(*AI);
1054        }
1055
1056        // Add any parameter attributes.
1057        if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1058          attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1059      }
1060    }
1061  }
1062
1063  if (Attributes FnAttrs =  CallerPAL.getFnAttributes())
1064    attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1065
1066  if (NewRetTy->isVoidTy())
1067    Caller->setName("");   // Void type should not have a name.
1068
1069  const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1070                                                     attrVec.end());
1071
1072  Instruction *NC;
1073  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1074    NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
1075                            Args.begin(), Args.end(),
1076                            Caller->getName(), Caller);
1077    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1078    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1079  } else {
1080    NC = CallInst::Create(Callee, Args.begin(), Args.end(),
1081                          Caller->getName(), Caller);
1082    CallInst *CI = cast<CallInst>(Caller);
1083    if (CI->isTailCall())
1084      cast<CallInst>(NC)->setTailCall();
1085    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1086    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1087  }
1088
1089  // Insert a cast of the return type as necessary.
1090  Value *NV = NC;
1091  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1092    if (!NV->getType()->isVoidTy()) {
1093      Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
1094                                                            OldRetTy, false);
1095      NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1096
1097      // If this is an invoke instruction, we should insert it after the first
1098      // non-phi, instruction in the normal successor block.
1099      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1100        BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1101        InsertNewInstBefore(NC, *I);
1102      } else {
1103        // Otherwise, it's a call, just insert cast right after the call instr
1104        InsertNewInstBefore(NC, *Caller);
1105      }
1106      Worklist.AddUsersToWorkList(*Caller);
1107    } else {
1108      NV = UndefValue::get(Caller->getType());
1109    }
1110  }
1111
1112
1113  if (!Caller->use_empty())
1114    Caller->replaceAllUsesWith(NV);
1115
1116  EraseInstFromFunction(*Caller);
1117  return true;
1118}
1119
1120// transformCallThroughTrampoline - Turn a call to a function created by the
1121// init_trampoline intrinsic into a direct call to the underlying function.
1122//
1123Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1124  Value *Callee = CS.getCalledValue();
1125  const PointerType *PTy = cast<PointerType>(Callee->getType());
1126  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1127  const AttrListPtr &Attrs = CS.getAttributes();
1128
1129  // If the call already has the 'nest' attribute somewhere then give up -
1130  // otherwise 'nest' would occur twice after splicing in the chain.
1131  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1132    return 0;
1133
1134  IntrinsicInst *Tramp =
1135    cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1136
1137  Function *NestF = cast<Function>(Tramp->getOperand(1)->stripPointerCasts());
1138  const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1139  const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1140
1141  const AttrListPtr &NestAttrs = NestF->getAttributes();
1142  if (!NestAttrs.isEmpty()) {
1143    unsigned NestIdx = 1;
1144    const Type *NestTy = 0;
1145    Attributes NestAttr = Attribute::None;
1146
1147    // Look for a parameter marked with the 'nest' attribute.
1148    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1149         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1150      if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1151        // Record the parameter type and any other attributes.
1152        NestTy = *I;
1153        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1154        break;
1155      }
1156
1157    if (NestTy) {
1158      Instruction *Caller = CS.getInstruction();
1159      std::vector<Value*> NewArgs;
1160      NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1161
1162      SmallVector<AttributeWithIndex, 8> NewAttrs;
1163      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1164
1165      // Insert the nest argument into the call argument list, which may
1166      // mean appending it.  Likewise for attributes.
1167
1168      // Add any result attributes.
1169      if (Attributes Attr = Attrs.getRetAttributes())
1170        NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1171
1172      {
1173        unsigned Idx = 1;
1174        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1175        do {
1176          if (Idx == NestIdx) {
1177            // Add the chain argument and attributes.
1178            Value *NestVal = Tramp->getOperand(2);
1179            if (NestVal->getType() != NestTy)
1180              NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
1181            NewArgs.push_back(NestVal);
1182            NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1183          }
1184
1185          if (I == E)
1186            break;
1187
1188          // Add the original argument and attributes.
1189          NewArgs.push_back(*I);
1190          if (Attributes Attr = Attrs.getParamAttributes(Idx))
1191            NewAttrs.push_back
1192              (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1193
1194          ++Idx, ++I;
1195        } while (1);
1196      }
1197
1198      // Add any function attributes.
1199      if (Attributes Attr = Attrs.getFnAttributes())
1200        NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1201
1202      // The trampoline may have been bitcast to a bogus type (FTy).
1203      // Handle this by synthesizing a new function type, equal to FTy
1204      // with the chain parameter inserted.
1205
1206      std::vector<const Type*> NewTypes;
1207      NewTypes.reserve(FTy->getNumParams()+1);
1208
1209      // Insert the chain's type into the list of parameter types, which may
1210      // mean appending it.
1211      {
1212        unsigned Idx = 1;
1213        FunctionType::param_iterator I = FTy->param_begin(),
1214          E = FTy->param_end();
1215
1216        do {
1217          if (Idx == NestIdx)
1218            // Add the chain's type.
1219            NewTypes.push_back(NestTy);
1220
1221          if (I == E)
1222            break;
1223
1224          // Add the original type.
1225          NewTypes.push_back(*I);
1226
1227          ++Idx, ++I;
1228        } while (1);
1229      }
1230
1231      // Replace the trampoline call with a direct call.  Let the generic
1232      // code sort out any function type mismatches.
1233      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1234                                                FTy->isVarArg());
1235      Constant *NewCallee =
1236        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1237        NestF : ConstantExpr::getBitCast(NestF,
1238                                         PointerType::getUnqual(NewFTy));
1239      const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1240                                                   NewAttrs.end());
1241
1242      Instruction *NewCaller;
1243      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1244        NewCaller = InvokeInst::Create(NewCallee,
1245                                       II->getNormalDest(), II->getUnwindDest(),
1246                                       NewArgs.begin(), NewArgs.end(),
1247                                       Caller->getName(), Caller);
1248        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1249        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1250      } else {
1251        NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
1252                                     Caller->getName(), Caller);
1253        if (cast<CallInst>(Caller)->isTailCall())
1254          cast<CallInst>(NewCaller)->setTailCall();
1255        cast<CallInst>(NewCaller)->
1256          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1257        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1258      }
1259      if (!Caller->getType()->isVoidTy())
1260        Caller->replaceAllUsesWith(NewCaller);
1261      Caller->eraseFromParent();
1262      Worklist.Remove(Caller);
1263      return 0;
1264    }
1265  }
1266
1267  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1268  // parameter, there is no need to adjust the argument list.  Let the generic
1269  // code sort out any function type mismatches.
1270  Constant *NewCallee =
1271    NestF->getType() == PTy ? NestF :
1272                              ConstantExpr::getBitCast(NestF, PTy);
1273  CS.setCalledFunction(NewCallee);
1274  return CS.getInstruction();
1275}
1276
1277