InstCombineCalls.cpp revision 74b64611c4b74a633239f2a9bfd6be07cdedff88 
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    return Align;
63  }
64  case Instruction::Alloca: {
65    AllocaInst *AI = cast<AllocaInst>(V);
66    // If there is a requested alignment and if this is an alloca, round up.
67    if (AI->getAlignment() >= PrefAlign)
68      return AI->getAlignment();
69    AI->setAlignment(PrefAlign);
70    return PrefAlign;
71  }
72  }
73
74  if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
75    // If there is a large requested alignment and we can, bump up the alignment
76    // of the global.
77    if (GV->isDeclaration()) return Align;
78
79    if (GV->getAlignment() >= PrefAlign)
80      return GV->getAlignment();
81    // We can only increase the alignment of the global if it has no alignment
82    // specified or if it is not assigned a section.  If it is assigned a
83    // section, the global could be densely packed with other objects in the
84    // section, increasing the alignment could cause padding issues.
85    if (!GV->hasSection() || GV->getAlignment() == 0)
86      GV->setAlignment(PrefAlign);
87    return GV->getAlignment();
88  }
89
90  return Align;
91}
92
93/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
94/// we can determine, return it, otherwise return 0.  If PrefAlign is specified,
95/// and it is more than the alignment of the ultimate object, see if we can
96/// increase the alignment of the ultimate object, making this check succeed.
97unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
98                                                  unsigned PrefAlign) {
99  assert(V->getType()->isPointerTy() &&
100         "GetOrEnforceKnownAlignment expects a pointer!");
101  unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 64;
102  APInt Mask = APInt::getAllOnesValue(BitWidth);
103  APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
104  ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
105  unsigned TrailZ = KnownZero.countTrailingOnes();
106
107  // Avoid trouble with rediculously large TrailZ values, such as
108  // those computed from a null pointer.
109  TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
110
111  unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
112
113  // LLVM doesn't support alignments larger than this currently.
114  Align = std::min(Align, +Value::MaximumAlignment);
115
116  if (PrefAlign > Align)
117    Align = EnforceKnownAlignment(V, Align, PrefAlign);
118
119    // We don't need to make any adjustment.
120  return Align;
121}
122
123Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
124  unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(0));
125  unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(1));
126  unsigned MinAlign = std::min(DstAlign, SrcAlign);
127  unsigned CopyAlign = MI->getAlignment();
128
129  if (CopyAlign < MinAlign) {
130    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
131                                             MinAlign, false));
132    return MI;
133  }
134
135  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
136  // load/store.
137  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
138  if (MemOpLength == 0) return 0;
139
140  // Source and destination pointer types are always "i8*" for intrinsic.  See
141  // if the size is something we can handle with a single primitive load/store.
142  // A single load+store correctly handles overlapping memory in the memmove
143  // case.
144  unsigned Size = MemOpLength->getZExtValue();
145  if (Size == 0) return MI;  // Delete this mem transfer.
146
147  if (Size > 8 || (Size&(Size-1)))
148    return 0;  // If not 1/2/4/8 bytes, exit.
149
150  // Use an integer load+store unless we can find something better.
151  unsigned SrcAddrSp =
152    cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
153  unsigned DstAddrSp =
154    cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
155
156  const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
157  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
158  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
159
160  // Memcpy forces the use of i8* for the source and destination.  That means
161  // that if you're using memcpy to move one double around, you'll get a cast
162  // from double* to i8*.  We'd much rather use a double load+store rather than
163  // an i64 load+store, here because this improves the odds that the source or
164  // dest address will be promotable.  See if we can find a better type than the
165  // integer datatype.
166  Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
167  if (StrippedDest != MI->getArgOperand(0)) {
168    const Type *SrcETy = cast<PointerType>(StrippedDest->getType())
169                                    ->getElementType();
170    if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
171      // The SrcETy might be something like {{{double}}} or [1 x double].  Rip
172      // down through these levels if so.
173      while (!SrcETy->isSingleValueType()) {
174        if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
175          if (STy->getNumElements() == 1)
176            SrcETy = STy->getElementType(0);
177          else
178            break;
179        } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
180          if (ATy->getNumElements() == 1)
181            SrcETy = ATy->getElementType();
182          else
183            break;
184        } else
185          break;
186      }
187
188      if (SrcETy->isSingleValueType()) {
189        NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
190        NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
191      }
192    }
193  }
194
195
196  // If the memcpy/memmove provides better alignment info than we can
197  // infer, use it.
198  SrcAlign = std::max(SrcAlign, CopyAlign);
199  DstAlign = std::max(DstAlign, CopyAlign);
200
201  Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
202  Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
203  Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign);
204  InsertNewInstBefore(L, *MI);
205  InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign),
206                      *MI);
207
208  // Set the size of the copy to 0, it will be deleted on the next iteration.
209  MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
210  return MI;
211}
212
213Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
214  unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
215  if (MI->getAlignment() < Alignment) {
216    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
217                                             Alignment, false));
218    return MI;
219  }
220
221  // Extract the length and alignment and fill if they are constant.
222  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
223  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
224  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
225    return 0;
226  uint64_t Len = LenC->getZExtValue();
227  Alignment = MI->getAlignment();
228
229  // If the length is zero, this is a no-op
230  if (Len == 0) return MI; // memset(d,c,0,a) -> noop
231
232  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
233  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
234    const Type *ITy = IntegerType::get(MI->getContext(), Len*8);  // n=1 -> i8.
235
236    Value *Dest = MI->getDest();
237    Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
238
239    // Alignment 0 is identity for alignment 1 for memset, but not store.
240    if (Alignment == 0) Alignment = 1;
241
242    // Extract the fill value and store.
243    uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
244    InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
245                                      Dest, false, Alignment), *MI);
246
247    // Set the size of the copy to 0, it will be deleted on the next iteration.
248    MI->setLength(Constant::getNullValue(LenC->getType()));
249    return MI;
250  }
251
252  return 0;
253}
254
255/// visitCallInst - CallInst simplification.  This mostly only handles folding
256/// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
257/// the heavy lifting.
258///
259Instruction *InstCombiner::visitCallInst(CallInst &CI) {
260  if (isFreeCall(&CI))
261    return visitFree(CI);
262  if (isMalloc(&CI))
263    return visitMalloc(CI);
264
265  // If the caller function is nounwind, mark the call as nounwind, even if the
266  // callee isn't.
267  if (CI.getParent()->getParent()->doesNotThrow() &&
268      !CI.doesNotThrow()) {
269    CI.setDoesNotThrow();
270    return &CI;
271  }
272
273  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
274  if (!II) return visitCallSite(&CI);
275
276  // Intrinsics cannot occur in an invoke, so handle them here instead of in
277  // visitCallSite.
278  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
279    bool Changed = false;
280
281    // memmove/cpy/set of zero bytes is a noop.
282    if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
283      if (NumBytes->isNullValue())
284        return EraseInstFromFunction(CI);
285
286      if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
287        if (CI->getZExtValue() == 1) {
288          // Replace the instruction with just byte operations.  We would
289          // transform other cases to loads/stores, but we don't know if
290          // alignment is sufficient.
291        }
292    }
293
294    // No other transformations apply to volatile transfers.
295    if (MI->isVolatile())
296      return 0;
297
298    // If we have a memmove and the source operation is a constant global,
299    // then the source and dest pointers can't alias, so we can change this
300    // into a call to memcpy.
301    if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
302      if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
303        if (GVSrc->isConstant()) {
304          Module *M = CI.getParent()->getParent()->getParent();
305          Intrinsic::ID MemCpyID = Intrinsic::memcpy;
306          const Type *Tys[3] = { CI.getArgOperand(0)->getType(),
307                                 CI.getArgOperand(1)->getType(),
308                                 CI.getArgOperand(2)->getType() };
309          CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3));
310          Changed = true;
311        }
312    }
313
314    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
315      // memmove(x,x,size) -> noop.
316      if (MTI->getSource() == MTI->getDest())
317        return EraseInstFromFunction(CI);
318    }
319
320    // If we can determine a pointer alignment that is bigger than currently
321    // set, update the alignment.
322    if (isa<MemTransferInst>(MI)) {
323      if (Instruction *I = SimplifyMemTransfer(MI))
324        return I;
325    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
326      if (Instruction *I = SimplifyMemSet(MSI))
327        return I;
328    }
329
330    if (Changed) return II;
331  }
332
333  switch (II->getIntrinsicID()) {
334  default: break;
335  case Intrinsic::objectsize: {
336    // We need target data for just about everything so depend on it.
337    if (!TD) break;
338
339    const Type *ReturnTy = CI.getType();
340    bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
341
342    // Get to the real allocated thing and offset as fast as possible.
343    Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
344
345    // If we've stripped down to a single global variable that we
346    // can know the size of then just return that.
347    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
348      if (GV->hasDefinitiveInitializer()) {
349        Constant *C = GV->getInitializer();
350        uint64_t GlobalSize = TD->getTypeAllocSize(C->getType());
351        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize));
352      } else {
353        // Can't determine size of the GV.
354        Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
355        return ReplaceInstUsesWith(CI, RetVal);
356      }
357    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
358      // Get alloca size.
359      if (AI->getAllocatedType()->isSized()) {
360        uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
361        if (AI->isArrayAllocation()) {
362          const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
363          if (!C) break;
364          AllocaSize *= C->getZExtValue();
365        }
366        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize));
367      }
368    } else if (CallInst *MI = extractMallocCall(Op1)) {
369      const Type* MallocType = getMallocAllocatedType(MI);
370      // Get alloca size.
371      if (MallocType && MallocType->isSized()) {
372        if (Value *NElems = getMallocArraySize(MI, TD, true)) {
373          if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
374        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy,
375               (NElements->getZExtValue() * TD->getTypeAllocSize(MallocType))));
376        }
377      }
378    } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) {
379      // Only handle constant GEPs here.
380      if (CE->getOpcode() != Instruction::GetElementPtr) break;
381      GEPOperator *GEP = cast<GEPOperator>(CE);
382
383      // Make sure we're not a constant offset from an external
384      // global.
385      Value *Operand = GEP->getPointerOperand();
386      Operand = Operand->stripPointerCasts();
387      if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand))
388        if (!GV->hasDefinitiveInitializer()) break;
389
390      // Get what we're pointing to and its size.
391      const PointerType *BaseType =
392        cast<PointerType>(Operand->getType());
393      uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType());
394
395      // Get the current byte offset into the thing. Use the original
396      // operand in case we're looking through a bitcast.
397      SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end());
398      const PointerType *OffsetType =
399        cast<PointerType>(GEP->getPointerOperand()->getType());
400      uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size());
401
402      if (Size < Offset) {
403        // Out of bound reference? Negative index normalized to large
404        // index? Just return "I don't know".
405        Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
406        return ReplaceInstUsesWith(CI, RetVal);
407      }
408
409      Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset);
410      return ReplaceInstUsesWith(CI, RetVal);
411    }
412
413    // Do not return "I don't know" here. Later optimization passes could
414    // make it possible to evaluate objectsize to a constant.
415    break;
416  }
417  case Intrinsic::bswap:
418    // bswap(bswap(x)) -> x
419    if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
420      if (Operand->getIntrinsicID() == Intrinsic::bswap)
421        return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
422
423    // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
424    if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
425      if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
426        if (Operand->getIntrinsicID() == Intrinsic::bswap) {
427          unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
428                       TI->getType()->getPrimitiveSizeInBits();
429          Value *CV = ConstantInt::get(Operand->getType(), C);
430          Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
431          return new TruncInst(V, TI->getType());
432        }
433    }
434
435    break;
436  case Intrinsic::powi:
437    if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
438      // powi(x, 0) -> 1.0
439      if (Power->isZero())
440        return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
441      // powi(x, 1) -> x
442      if (Power->isOne())
443        return ReplaceInstUsesWith(CI, II->getArgOperand(0));
444      // powi(x, -1) -> 1/x
445      if (Power->isAllOnesValue())
446        return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
447                                          II->getArgOperand(0));
448    }
449    break;
450  case Intrinsic::cttz: {
451    // If all bits below the first known one are known zero,
452    // this value is constant.
453    const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
454    uint32_t BitWidth = IT->getBitWidth();
455    APInt KnownZero(BitWidth, 0);
456    APInt KnownOne(BitWidth, 0);
457    ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
458                      KnownZero, KnownOne);
459    unsigned TrailingZeros = KnownOne.countTrailingZeros();
460    APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
461    if ((Mask & KnownZero) == Mask)
462      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
463                                 APInt(BitWidth, TrailingZeros)));
464
465    }
466    break;
467  case Intrinsic::ctlz: {
468    // If all bits above the first known one are known zero,
469    // this value is constant.
470    const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
471    uint32_t BitWidth = IT->getBitWidth();
472    APInt KnownZero(BitWidth, 0);
473    APInt KnownOne(BitWidth, 0);
474    ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
475                      KnownZero, KnownOne);
476    unsigned LeadingZeros = KnownOne.countLeadingZeros();
477    APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
478    if ((Mask & KnownZero) == Mask)
479      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
480                                 APInt(BitWidth, LeadingZeros)));
481
482    }
483    break;
484  case Intrinsic::uadd_with_overflow: {
485    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
486    const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
487    uint32_t BitWidth = IT->getBitWidth();
488    APInt Mask = APInt::getSignBit(BitWidth);
489    APInt LHSKnownZero(BitWidth, 0);
490    APInt LHSKnownOne(BitWidth, 0);
491    ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
492    bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
493    bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
494
495    if (LHSKnownNegative || LHSKnownPositive) {
496      APInt RHSKnownZero(BitWidth, 0);
497      APInt RHSKnownOne(BitWidth, 0);
498      ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
499      bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
500      bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
501      if (LHSKnownNegative && RHSKnownNegative) {
502        // The sign bit is set in both cases: this MUST overflow.
503        // Create a simple add instruction, and insert it into the struct.
504        Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI);
505        Worklist.Add(Add);
506        Constant *V[] = {
507          UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext())
508        };
509        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
510        return InsertValueInst::Create(Struct, Add, 0);
511      }
512
513      if (LHSKnownPositive && RHSKnownPositive) {
514        // The sign bit is clear in both cases: this CANNOT overflow.
515        // Create a simple add instruction, and insert it into the struct.
516        Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI);
517        Worklist.Add(Add);
518        Constant *V[] = {
519          UndefValue::get(LHS->getType()),
520          ConstantInt::getFalse(II->getContext())
521        };
522        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
523        return InsertValueInst::Create(Struct, Add, 0);
524      }
525    }
526  }
527  // FALL THROUGH uadd into sadd
528  case Intrinsic::sadd_with_overflow:
529    // Canonicalize constants into the RHS.
530    if (isa<Constant>(II->getArgOperand(0)) &&
531        !isa<Constant>(II->getArgOperand(1))) {
532      Value *LHS = II->getArgOperand(0);
533      II->setArgOperand(0, II->getArgOperand(1));
534      II->setArgOperand(1, LHS);
535      return II;
536    }
537
538    // X + undef -> undef
539    if (isa<UndefValue>(II->getArgOperand(1)))
540      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
541
542    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
543      // X + 0 -> {X, false}
544      if (RHS->isZero()) {
545        Constant *V[] = {
546          UndefValue::get(II->getArgOperand(0)->getType()),
547          ConstantInt::getFalse(II->getContext())
548        };
549        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
550        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
551      }
552    }
553    break;
554  case Intrinsic::usub_with_overflow:
555  case Intrinsic::ssub_with_overflow:
556    // undef - X -> undef
557    // X - undef -> undef
558    if (isa<UndefValue>(II->getArgOperand(0)) ||
559        isa<UndefValue>(II->getArgOperand(1)))
560      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
561
562    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
563      // X - 0 -> {X, false}
564      if (RHS->isZero()) {
565        Constant *V[] = {
566          UndefValue::get(II->getArgOperand(0)->getType()),
567          ConstantInt::getFalse(II->getContext())
568        };
569        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
570        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
571      }
572    }
573    break;
574  case Intrinsic::umul_with_overflow:
575  case Intrinsic::smul_with_overflow:
576    // Canonicalize constants into the RHS.
577    if (isa<Constant>(II->getArgOperand(0)) &&
578        !isa<Constant>(II->getArgOperand(1))) {
579      Value *LHS = II->getArgOperand(0);
580      II->setArgOperand(0, II->getArgOperand(1));
581      II->setArgOperand(1, LHS);
582      return II;
583    }
584
585    // X * undef -> undef
586    if (isa<UndefValue>(II->getArgOperand(1)))
587      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
588
589    if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
590      // X*0 -> {0, false}
591      if (RHSI->isZero())
592        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
593
594      // X * 1 -> {X, false}
595      if (RHSI->equalsInt(1)) {
596        Constant *V[] = {
597          UndefValue::get(II->getArgOperand(0)->getType()),
598          ConstantInt::getFalse(II->getContext())
599        };
600        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
601        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
602      }
603    }
604    break;
605  case Intrinsic::ppc_altivec_lvx:
606  case Intrinsic::ppc_altivec_lvxl:
607  case Intrinsic::x86_sse_loadu_ps:
608  case Intrinsic::x86_sse2_loadu_pd:
609  case Intrinsic::x86_sse2_loadu_dq:
610    // Turn PPC lvx     -> load if the pointer is known aligned.
611    // Turn X86 loadups -> load if the pointer is known aligned.
612    if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
613      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
614                                         PointerType::getUnqual(II->getType()));
615      return new LoadInst(Ptr);
616    }
617    break;
618  case Intrinsic::ppc_altivec_stvx:
619  case Intrinsic::ppc_altivec_stvxl:
620    // Turn stvx -> store if the pointer is known aligned.
621    if (GetOrEnforceKnownAlignment(II->getArgOperand(1), 16) >= 16) {
622      const Type *OpPtrTy =
623        PointerType::getUnqual(II->getArgOperand(0)->getType());
624      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
625      return new StoreInst(II->getArgOperand(0), Ptr);
626    }
627    break;
628  case Intrinsic::x86_sse_storeu_ps:
629  case Intrinsic::x86_sse2_storeu_pd:
630  case Intrinsic::x86_sse2_storeu_dq:
631    // Turn X86 storeu -> store if the pointer is known aligned.
632    if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
633      const Type *OpPtrTy =
634        PointerType::getUnqual(II->getArgOperand(1)->getType());
635      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
636      return new StoreInst(II->getArgOperand(1), Ptr);
637    }
638    break;
639
640  case Intrinsic::x86_sse_cvttss2si: {
641    // These intrinsics only demands the 0th element of its input vector.  If
642    // we can simplify the input based on that, do so now.
643    unsigned VWidth =
644      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
645    APInt DemandedElts(VWidth, 1);
646    APInt UndefElts(VWidth, 0);
647    if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
648                                              DemandedElts, UndefElts)) {
649      II->setArgOperand(0, V);
650      return II;
651    }
652    break;
653  }
654
655  case Intrinsic::ppc_altivec_vperm:
656    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
657    if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
658      assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
659
660      // Check that all of the elements are integer constants or undefs.
661      bool AllEltsOk = true;
662      for (unsigned i = 0; i != 16; ++i) {
663        if (!isa<ConstantInt>(Mask->getOperand(i)) &&
664            !isa<UndefValue>(Mask->getOperand(i))) {
665          AllEltsOk = false;
666          break;
667        }
668      }
669
670      if (AllEltsOk) {
671        // Cast the input vectors to byte vectors.
672        Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
673                                            Mask->getType());
674        Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
675                                            Mask->getType());
676        Value *Result = UndefValue::get(Op0->getType());
677
678        // Only extract each element once.
679        Value *ExtractedElts[32];
680        memset(ExtractedElts, 0, sizeof(ExtractedElts));
681
682        for (unsigned i = 0; i != 16; ++i) {
683          if (isa<UndefValue>(Mask->getOperand(i)))
684            continue;
685          unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
686          Idx &= 31;  // Match the hardware behavior.
687
688          if (ExtractedElts[Idx] == 0) {
689            ExtractedElts[Idx] =
690              Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
691                  ConstantInt::get(Type::getInt32Ty(II->getContext()),
692                                   Idx&15, false), "tmp");
693          }
694
695          // Insert this value into the result vector.
696          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
697                         ConstantInt::get(Type::getInt32Ty(II->getContext()),
698                                          i, false), "tmp");
699        }
700        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
701      }
702    }
703    break;
704
705  case Intrinsic::arm_neon_vld1:
706  case Intrinsic::arm_neon_vld2:
707  case Intrinsic::arm_neon_vld3:
708  case Intrinsic::arm_neon_vld4:
709  case Intrinsic::arm_neon_vld2lane:
710  case Intrinsic::arm_neon_vld3lane:
711  case Intrinsic::arm_neon_vld4lane:
712  case Intrinsic::arm_neon_vst1:
713  case Intrinsic::arm_neon_vst2:
714  case Intrinsic::arm_neon_vst3:
715  case Intrinsic::arm_neon_vst4:
716  case Intrinsic::arm_neon_vst2lane:
717  case Intrinsic::arm_neon_vst3lane:
718  case Intrinsic::arm_neon_vst4lane: {
719    unsigned MemAlign = GetOrEnforceKnownAlignment(II->getArgOperand(0));
720    unsigned AlignArg = II->getNumArgOperands() - 1;
721    ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
722    if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
723      II->setArgOperand(AlignArg,
724                        ConstantInt::get(Type::getInt32Ty(II->getContext()),
725                                         MemAlign, false));
726      return II;
727    }
728    break;
729  }
730
731  case Intrinsic::stackrestore: {
732    // If the save is right next to the restore, remove the restore.  This can
733    // happen when variable allocas are DCE'd.
734    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
735      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
736        BasicBlock::iterator BI = SS;
737        if (&*++BI == II)
738          return EraseInstFromFunction(CI);
739      }
740    }
741
742    // Scan down this block to see if there is another stack restore in the
743    // same block without an intervening call/alloca.
744    BasicBlock::iterator BI = II;
745    TerminatorInst *TI = II->getParent()->getTerminator();
746    bool CannotRemove = false;
747    for (++BI; &*BI != TI; ++BI) {
748      if (isa<AllocaInst>(BI) || isMalloc(BI)) {
749        CannotRemove = true;
750        break;
751      }
752      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
753        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
754          // If there is a stackrestore below this one, remove this one.
755          if (II->getIntrinsicID() == Intrinsic::stackrestore)
756            return EraseInstFromFunction(CI);
757          // Otherwise, ignore the intrinsic.
758        } else {
759          // If we found a non-intrinsic call, we can't remove the stack
760          // restore.
761          CannotRemove = true;
762          break;
763        }
764      }
765    }
766
767    // If the stack restore is in a return/unwind block and if there are no
768    // allocas or calls between the restore and the return, nuke the restore.
769    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
770      return EraseInstFromFunction(CI);
771    break;
772  }
773  }
774
775  return visitCallSite(II);
776}
777
778// InvokeInst simplification
779//
780Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
781  return visitCallSite(&II);
782}
783
784/// isSafeToEliminateVarargsCast - If this cast does not affect the value
785/// passed through the varargs area, we can eliminate the use of the cast.
786static bool isSafeToEliminateVarargsCast(const CallSite CS,
787                                         const CastInst * const CI,
788                                         const TargetData * const TD,
789                                         const int ix) {
790  if (!CI->isLosslessCast())
791    return false;
792
793  // The size of ByVal arguments is derived from the type, so we
794  // can't change to a type with a different size.  If the size were
795  // passed explicitly we could avoid this check.
796  if (!CS.paramHasAttr(ix, Attribute::ByVal))
797    return true;
798
799  const Type* SrcTy =
800            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
801  const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
802  if (!SrcTy->isSized() || !DstTy->isSized())
803    return false;
804  if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
805    return false;
806  return true;
807}
808
809namespace {
810class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
811  InstCombiner *IC;
812protected:
813  void replaceCall(Value *With) {
814    NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
815  }
816  bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
817    if (ConstantInt *SizeCI =
818                           dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
819      if (SizeCI->isAllOnesValue())
820        return true;
821      if (isString)
822        return SizeCI->getZExtValue() >=
823               GetStringLength(CI->getArgOperand(SizeArgOp));
824      if (ConstantInt *Arg = dyn_cast<ConstantInt>(
825                                                  CI->getArgOperand(SizeArgOp)))
826        return SizeCI->getZExtValue() >= Arg->getZExtValue();
827    }
828    return false;
829  }
830public:
831  InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
832  Instruction *NewInstruction;
833};
834} // end anonymous namespace
835
836// Try to fold some different type of calls here.
837// Currently we're only working with the checking functions, memcpy_chk,
838// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
839// strcat_chk and strncat_chk.
840Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
841  if (CI->getCalledFunction() == 0) return 0;
842
843  InstCombineFortifiedLibCalls Simplifier(this);
844  Simplifier.fold(CI, TD);
845  return Simplifier.NewInstruction;
846}
847
848// visitCallSite - Improvements for call and invoke instructions.
849//
850Instruction *InstCombiner::visitCallSite(CallSite CS) {
851  bool Changed = false;
852
853  // If the callee is a constexpr cast of a function, attempt to move the cast
854  // to the arguments of the call/invoke.
855  if (transformConstExprCastCall(CS)) return 0;
856
857  Value *Callee = CS.getCalledValue();
858
859  if (Function *CalleeF = dyn_cast<Function>(Callee))
860    // If the call and callee calling conventions don't match, this call must
861    // be unreachable, as the call is undefined.
862    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
863        // Only do this for calls to a function with a body.  A prototype may
864        // not actually end up matching the implementation's calling conv for a
865        // variety of reasons (e.g. it may be written in assembly).
866        !CalleeF->isDeclaration()) {
867      Instruction *OldCall = CS.getInstruction();
868      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
869                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
870                                  OldCall);
871      // If OldCall dues not return void then replaceAllUsesWith undef.
872      // This allows ValueHandlers and custom metadata to adjust itself.
873      if (!OldCall->getType()->isVoidTy())
874        OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
875      if (isa<CallInst>(OldCall))
876        return EraseInstFromFunction(*OldCall);
877
878      // We cannot remove an invoke, because it would change the CFG, just
879      // change the callee to a null pointer.
880      cast<InvokeInst>(OldCall)->setCalledFunction(
881                                    Constant::getNullValue(CalleeF->getType()));
882      return 0;
883    }
884
885  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
886    // This instruction is not reachable, just remove it.  We insert a store to
887    // undef so that we know that this code is not reachable, despite the fact
888    // that we can't modify the CFG here.
889    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
890               UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
891                  CS.getInstruction());
892
893    // If CS does not return void then replaceAllUsesWith undef.
894    // This allows ValueHandlers and custom metadata to adjust itself.
895    if (!CS.getInstruction()->getType()->isVoidTy())
896      CS.getInstruction()->
897        replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
898
899    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
900      // Don't break the CFG, insert a dummy cond branch.
901      BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
902                         ConstantInt::getTrue(Callee->getContext()), II);
903    }
904    return EraseInstFromFunction(*CS.getInstruction());
905  }
906
907  if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
908    if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
909      if (In->getIntrinsicID() == Intrinsic::init_trampoline)
910        return transformCallThroughTrampoline(CS);
911
912  const PointerType *PTy = cast<PointerType>(Callee->getType());
913  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
914  if (FTy->isVarArg()) {
915    int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
916    // See if we can optimize any arguments passed through the varargs area of
917    // the call.
918    for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
919           E = CS.arg_end(); I != E; ++I, ++ix) {
920      CastInst *CI = dyn_cast<CastInst>(*I);
921      if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
922        *I = CI->getOperand(0);
923        Changed = true;
924      }
925    }
926  }
927
928  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
929    // Inline asm calls cannot throw - mark them 'nounwind'.
930    CS.setDoesNotThrow();
931    Changed = true;
932  }
933
934  // Try to optimize the call if possible, we require TargetData for most of
935  // this.  None of these calls are seen as possibly dead so go ahead and
936  // delete the instruction now.
937  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
938    Instruction *I = tryOptimizeCall(CI, TD);
939    // If we changed something return the result, etc. Otherwise let
940    // the fallthrough check.
941    if (I) return EraseInstFromFunction(*I);
942  }
943
944  return Changed ? CS.getInstruction() : 0;
945}
946
947// transformConstExprCastCall - If the callee is a constexpr cast of a function,
948// attempt to move the cast to the arguments of the call/invoke.
949//
950bool InstCombiner::transformConstExprCastCall(CallSite CS) {
951  if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
952  ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
953  if (CE->getOpcode() != Instruction::BitCast ||
954      !isa<Function>(CE->getOperand(0)))
955    return false;
956  Function *Callee = cast<Function>(CE->getOperand(0));
957  Instruction *Caller = CS.getInstruction();
958  const AttrListPtr &CallerPAL = CS.getAttributes();
959
960  // Okay, this is a cast from a function to a different type.  Unless doing so
961  // would cause a type conversion of one of our arguments, change this call to
962  // be a direct call with arguments casted to the appropriate types.
963  //
964  const FunctionType *FT = Callee->getFunctionType();
965  const Type *OldRetTy = Caller->getType();
966  const Type *NewRetTy = FT->getReturnType();
967
968  if (NewRetTy->isStructTy())
969    return false; // TODO: Handle multiple return values.
970
971  // Check to see if we are changing the return type...
972  if (OldRetTy != NewRetTy) {
973    if (Callee->isDeclaration() &&
974        // Conversion is ok if changing from one pointer type to another or from
975        // a pointer to an integer of the same size.
976        !((OldRetTy->isPointerTy() || !TD ||
977           OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
978          (NewRetTy->isPointerTy() || !TD ||
979           NewRetTy == TD->getIntPtrType(Caller->getContext()))))
980      return false;   // Cannot transform this return value.
981
982    if (!Caller->use_empty() &&
983        // void -> non-void is handled specially
984        !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
985      return false;   // Cannot transform this return value.
986
987    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
988      Attributes RAttrs = CallerPAL.getRetAttributes();
989      if (RAttrs & Attribute::typeIncompatible(NewRetTy))
990        return false;   // Attribute not compatible with transformed value.
991    }
992
993    // If the callsite is an invoke instruction, and the return value is used by
994    // a PHI node in a successor, we cannot change the return type of the call
995    // because there is no place to put the cast instruction (without breaking
996    // the critical edge).  Bail out in this case.
997    if (!Caller->use_empty())
998      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
999        for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1000             UI != E; ++UI)
1001          if (PHINode *PN = dyn_cast<PHINode>(*UI))
1002            if (PN->getParent() == II->getNormalDest() ||
1003                PN->getParent() == II->getUnwindDest())
1004              return false;
1005  }
1006
1007  unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1008  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1009
1010  CallSite::arg_iterator AI = CS.arg_begin();
1011  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1012    const Type *ParamTy = FT->getParamType(i);
1013    const Type *ActTy = (*AI)->getType();
1014
1015    if (!CastInst::isCastable(ActTy, ParamTy))
1016      return false;   // Cannot transform this parameter value.
1017
1018    if (CallerPAL.getParamAttributes(i + 1)
1019        & Attribute::typeIncompatible(ParamTy))
1020      return false;   // Attribute not compatible with transformed value.
1021
1022    // Converting from one pointer type to another or between a pointer and an
1023    // integer of the same size is safe even if we do not have a body.
1024    bool isConvertible = ActTy == ParamTy ||
1025      (TD && ((ParamTy->isPointerTy() ||
1026      ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1027              (ActTy->isPointerTy() ||
1028              ActTy == TD->getIntPtrType(Caller->getContext()))));
1029    if (Callee->isDeclaration() && !isConvertible) return false;
1030  }
1031
1032  if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
1033      Callee->isDeclaration())
1034    return false;   // Do not delete arguments unless we have a function body.
1035
1036  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1037      !CallerPAL.isEmpty())
1038    // In this case we have more arguments than the new function type, but we
1039    // won't be dropping them.  Check that these extra arguments have attributes
1040    // that are compatible with being a vararg call argument.
1041    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1042      if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1043        break;
1044      Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1045      if (PAttrs & Attribute::VarArgsIncompatible)
1046        return false;
1047    }
1048
1049  // Okay, we decided that this is a safe thing to do: go ahead and start
1050  // inserting cast instructions as necessary...
1051  std::vector<Value*> Args;
1052  Args.reserve(NumActualArgs);
1053  SmallVector<AttributeWithIndex, 8> attrVec;
1054  attrVec.reserve(NumCommonArgs);
1055
1056  // Get any return attributes.
1057  Attributes RAttrs = CallerPAL.getRetAttributes();
1058
1059  // If the return value is not being used, the type may not be compatible
1060  // with the existing attributes.  Wipe out any problematic attributes.
1061  RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1062
1063  // Add the new return attributes.
1064  if (RAttrs)
1065    attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1066
1067  AI = CS.arg_begin();
1068  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1069    const Type *ParamTy = FT->getParamType(i);
1070    if ((*AI)->getType() == ParamTy) {
1071      Args.push_back(*AI);
1072    } else {
1073      Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1074          false, ParamTy, false);
1075      Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1076    }
1077
1078    // Add any parameter attributes.
1079    if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1080      attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1081  }
1082
1083  // If the function takes more arguments than the call was taking, add them
1084  // now.
1085  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1086    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1087
1088  // If we are removing arguments to the function, emit an obnoxious warning.
1089  if (FT->getNumParams() < NumActualArgs) {
1090    if (!FT->isVarArg()) {
1091      errs() << "WARNING: While resolving call to function '"
1092             << Callee->getName() << "' arguments were dropped!\n";
1093    } else {
1094      // Add all of the arguments in their promoted form to the arg list.
1095      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1096        const Type *PTy = getPromotedType((*AI)->getType());
1097        if (PTy != (*AI)->getType()) {
1098          // Must promote to pass through va_arg area!
1099          Instruction::CastOps opcode =
1100            CastInst::getCastOpcode(*AI, false, PTy, false);
1101          Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1102        } else {
1103          Args.push_back(*AI);
1104        }
1105
1106        // Add any parameter attributes.
1107        if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1108          attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1109      }
1110    }
1111  }
1112
1113  if (Attributes FnAttrs =  CallerPAL.getFnAttributes())
1114    attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1115
1116  if (NewRetTy->isVoidTy())
1117    Caller->setName("");   // Void type should not have a name.
1118
1119  const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1120                                                     attrVec.end());
1121
1122  Instruction *NC;
1123  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1124    NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
1125                            Args.begin(), Args.end(),
1126                            Caller->getName(), Caller);
1127    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1128    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1129  } else {
1130    NC = CallInst::Create(Callee, Args.begin(), Args.end(),
1131                          Caller->getName(), Caller);
1132    CallInst *CI = cast<CallInst>(Caller);
1133    if (CI->isTailCall())
1134      cast<CallInst>(NC)->setTailCall();
1135    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1136    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1137  }
1138
1139  // Insert a cast of the return type as necessary.
1140  Value *NV = NC;
1141  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1142    if (!NV->getType()->isVoidTy()) {
1143      Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
1144                                                            OldRetTy, false);
1145      NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1146
1147      // If this is an invoke instruction, we should insert it after the first
1148      // non-phi, instruction in the normal successor block.
1149      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1150        BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1151        InsertNewInstBefore(NC, *I);
1152      } else {
1153        // Otherwise, it's a call, just insert cast right after the call instr
1154        InsertNewInstBefore(NC, *Caller);
1155      }
1156      Worklist.AddUsersToWorkList(*Caller);
1157    } else {
1158      NV = UndefValue::get(Caller->getType());
1159    }
1160  }
1161
1162
1163  if (!Caller->use_empty())
1164    Caller->replaceAllUsesWith(NV);
1165
1166  EraseInstFromFunction(*Caller);
1167  return true;
1168}
1169
1170// transformCallThroughTrampoline - Turn a call to a function created by the
1171// init_trampoline intrinsic into a direct call to the underlying function.
1172//
1173Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1174  Value *Callee = CS.getCalledValue();
1175  const PointerType *PTy = cast<PointerType>(Callee->getType());
1176  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1177  const AttrListPtr &Attrs = CS.getAttributes();
1178
1179  // If the call already has the 'nest' attribute somewhere then give up -
1180  // otherwise 'nest' would occur twice after splicing in the chain.
1181  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1182    return 0;
1183
1184  IntrinsicInst *Tramp =
1185    cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1186
1187  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1188  const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1189  const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1190
1191  const AttrListPtr &NestAttrs = NestF->getAttributes();
1192  if (!NestAttrs.isEmpty()) {
1193    unsigned NestIdx = 1;
1194    const Type *NestTy = 0;
1195    Attributes NestAttr = Attribute::None;
1196
1197    // Look for a parameter marked with the 'nest' attribute.
1198    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1199         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1200      if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1201        // Record the parameter type and any other attributes.
1202        NestTy = *I;
1203        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1204        break;
1205      }
1206
1207    if (NestTy) {
1208      Instruction *Caller = CS.getInstruction();
1209      std::vector<Value*> NewArgs;
1210      NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1211
1212      SmallVector<AttributeWithIndex, 8> NewAttrs;
1213      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1214
1215      // Insert the nest argument into the call argument list, which may
1216      // mean appending it.  Likewise for attributes.
1217
1218      // Add any result attributes.
1219      if (Attributes Attr = Attrs.getRetAttributes())
1220        NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1221
1222      {
1223        unsigned Idx = 1;
1224        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1225        do {
1226          if (Idx == NestIdx) {
1227            // Add the chain argument and attributes.
1228            Value *NestVal = Tramp->getArgOperand(2);
1229            if (NestVal->getType() != NestTy)
1230              NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
1231            NewArgs.push_back(NestVal);
1232            NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1233          }
1234
1235          if (I == E)
1236            break;
1237
1238          // Add the original argument and attributes.
1239          NewArgs.push_back(*I);
1240          if (Attributes Attr = Attrs.getParamAttributes(Idx))
1241            NewAttrs.push_back
1242              (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1243
1244          ++Idx, ++I;
1245        } while (1);
1246      }
1247
1248      // Add any function attributes.
1249      if (Attributes Attr = Attrs.getFnAttributes())
1250        NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1251
1252      // The trampoline may have been bitcast to a bogus type (FTy).
1253      // Handle this by synthesizing a new function type, equal to FTy
1254      // with the chain parameter inserted.
1255
1256      std::vector<const Type*> NewTypes;
1257      NewTypes.reserve(FTy->getNumParams()+1);
1258
1259      // Insert the chain's type into the list of parameter types, which may
1260      // mean appending it.
1261      {
1262        unsigned Idx = 1;
1263        FunctionType::param_iterator I = FTy->param_begin(),
1264          E = FTy->param_end();
1265
1266        do {
1267          if (Idx == NestIdx)
1268            // Add the chain's type.
1269            NewTypes.push_back(NestTy);
1270
1271          if (I == E)
1272            break;
1273
1274          // Add the original type.
1275          NewTypes.push_back(*I);
1276
1277          ++Idx, ++I;
1278        } while (1);
1279      }
1280
1281      // Replace the trampoline call with a direct call.  Let the generic
1282      // code sort out any function type mismatches.
1283      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1284                                                FTy->isVarArg());
1285      Constant *NewCallee =
1286        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1287        NestF : ConstantExpr::getBitCast(NestF,
1288                                         PointerType::getUnqual(NewFTy));
1289      const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1290                                                   NewAttrs.end());
1291
1292      Instruction *NewCaller;
1293      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1294        NewCaller = InvokeInst::Create(NewCallee,
1295                                       II->getNormalDest(), II->getUnwindDest(),
1296                                       NewArgs.begin(), NewArgs.end(),
1297                                       Caller->getName(), Caller);
1298        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1299        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1300      } else {
1301        NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
1302                                     Caller->getName(), Caller);
1303        if (cast<CallInst>(Caller)->isTailCall())
1304          cast<CallInst>(NewCaller)->setTailCall();
1305        cast<CallInst>(NewCaller)->
1306          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1307        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1308      }
1309      if (!Caller->getType()->isVoidTy())
1310        Caller->replaceAllUsesWith(NewCaller);
1311      Caller->eraseFromParent();
1312      Worklist.Remove(Caller);
1313      return 0;
1314    }
1315  }
1316
1317  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1318  // parameter, there is no need to adjust the argument list.  Let the generic
1319  // code sort out any function type mismatches.
1320  Constant *NewCallee =
1321    NestF->getType() == PTy ? NestF :
1322                              ConstantExpr::getBitCast(NestF, PTy);
1323  CS.setCalledFunction(NewCallee);
1324  return CS.getInstruction();
1325}
1326
1327