InstCombineCalls.cpp revision 26b482d7a76df3f67675ce852daed0eba709c63e 
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    // If the normal result of the add is dead, and the RHS is a constant, we
528    // can transform this into a range comparison.
529    // overflow = uadd a, -4  -->  overflow = icmp ugt a, 3
530    if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS))
531      if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(II->use_back()))
532        if (II->hasOneUse() && EVI->getNumIndices() == 1 && !EVI->use_empty() &&
533            *EVI->idx_begin() == 1) {  // Extract of overflow result.
534          Builder->SetInsertPoint(EVI);
535          Value *R = Builder->CreateICmpUGT(LHS, ConstantExpr::getNot(CI));
536          R->takeName(EVI);
537          ReplaceInstUsesWith(*EVI, R);
538          return II;
539        }
540
541  }
542  // FALL THROUGH uadd into sadd
543  case Intrinsic::sadd_with_overflow:
544    // Canonicalize constants into the RHS.
545    if (isa<Constant>(II->getArgOperand(0)) &&
546        !isa<Constant>(II->getArgOperand(1))) {
547      Value *LHS = II->getArgOperand(0);
548      II->setArgOperand(0, II->getArgOperand(1));
549      II->setArgOperand(1, LHS);
550      return II;
551    }
552
553    // X + undef -> undef
554    if (isa<UndefValue>(II->getArgOperand(1)))
555      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
556
557    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
558      // X + 0 -> {X, false}
559      if (RHS->isZero()) {
560        Constant *V[] = {
561          UndefValue::get(II->getArgOperand(0)->getType()),
562          ConstantInt::getFalse(II->getContext())
563        };
564        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
565        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
566      }
567    }
568
569    break;
570  case Intrinsic::usub_with_overflow:
571  case Intrinsic::ssub_with_overflow:
572    // undef - X -> undef
573    // X - undef -> undef
574    if (isa<UndefValue>(II->getArgOperand(0)) ||
575        isa<UndefValue>(II->getArgOperand(1)))
576      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
577
578    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
579      // X - 0 -> {X, false}
580      if (RHS->isZero()) {
581        Constant *V[] = {
582          UndefValue::get(II->getArgOperand(0)->getType()),
583          ConstantInt::getFalse(II->getContext())
584        };
585        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
586        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
587      }
588    }
589    break;
590  case Intrinsic::umul_with_overflow:
591  case Intrinsic::smul_with_overflow:
592    // Canonicalize constants into the RHS.
593    if (isa<Constant>(II->getArgOperand(0)) &&
594        !isa<Constant>(II->getArgOperand(1))) {
595      Value *LHS = II->getArgOperand(0);
596      II->setArgOperand(0, II->getArgOperand(1));
597      II->setArgOperand(1, LHS);
598      return II;
599    }
600
601    // X * undef -> undef
602    if (isa<UndefValue>(II->getArgOperand(1)))
603      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
604
605    if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
606      // X*0 -> {0, false}
607      if (RHSI->isZero())
608        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
609
610      // X * 1 -> {X, false}
611      if (RHSI->equalsInt(1)) {
612        Constant *V[] = {
613          UndefValue::get(II->getArgOperand(0)->getType()),
614          ConstantInt::getFalse(II->getContext())
615        };
616        Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false);
617        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
618      }
619    }
620    break;
621  case Intrinsic::ppc_altivec_lvx:
622  case Intrinsic::ppc_altivec_lvxl:
623  case Intrinsic::x86_sse_loadu_ps:
624  case Intrinsic::x86_sse2_loadu_pd:
625  case Intrinsic::x86_sse2_loadu_dq:
626    // Turn PPC lvx     -> load if the pointer is known aligned.
627    // Turn X86 loadups -> load if the pointer is known aligned.
628    if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
629      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
630                                         PointerType::getUnqual(II->getType()));
631      return new LoadInst(Ptr);
632    }
633    break;
634  case Intrinsic::ppc_altivec_stvx:
635  case Intrinsic::ppc_altivec_stvxl:
636    // Turn stvx -> store if the pointer is known aligned.
637    if (GetOrEnforceKnownAlignment(II->getArgOperand(1), 16) >= 16) {
638      const Type *OpPtrTy =
639        PointerType::getUnqual(II->getArgOperand(0)->getType());
640      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
641      return new StoreInst(II->getArgOperand(0), Ptr);
642    }
643    break;
644  case Intrinsic::x86_sse_storeu_ps:
645  case Intrinsic::x86_sse2_storeu_pd:
646  case Intrinsic::x86_sse2_storeu_dq:
647    // Turn X86 storeu -> store if the pointer is known aligned.
648    if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) {
649      const Type *OpPtrTy =
650        PointerType::getUnqual(II->getArgOperand(1)->getType());
651      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
652      return new StoreInst(II->getArgOperand(1), Ptr);
653    }
654    break;
655
656  case Intrinsic::x86_sse_cvttss2si: {
657    // These intrinsics only demands the 0th element of its input vector.  If
658    // we can simplify the input based on that, do so now.
659    unsigned VWidth =
660      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
661    APInt DemandedElts(VWidth, 1);
662    APInt UndefElts(VWidth, 0);
663    if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
664                                              DemandedElts, UndefElts)) {
665      II->setArgOperand(0, V);
666      return II;
667    }
668    break;
669  }
670
671  case Intrinsic::ppc_altivec_vperm:
672    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
673    if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) {
674      assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
675
676      // Check that all of the elements are integer constants or undefs.
677      bool AllEltsOk = true;
678      for (unsigned i = 0; i != 16; ++i) {
679        if (!isa<ConstantInt>(Mask->getOperand(i)) &&
680            !isa<UndefValue>(Mask->getOperand(i))) {
681          AllEltsOk = false;
682          break;
683        }
684      }
685
686      if (AllEltsOk) {
687        // Cast the input vectors to byte vectors.
688        Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
689                                            Mask->getType());
690        Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
691                                            Mask->getType());
692        Value *Result = UndefValue::get(Op0->getType());
693
694        // Only extract each element once.
695        Value *ExtractedElts[32];
696        memset(ExtractedElts, 0, sizeof(ExtractedElts));
697
698        for (unsigned i = 0; i != 16; ++i) {
699          if (isa<UndefValue>(Mask->getOperand(i)))
700            continue;
701          unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
702          Idx &= 31;  // Match the hardware behavior.
703
704          if (ExtractedElts[Idx] == 0) {
705            ExtractedElts[Idx] =
706              Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
707                  ConstantInt::get(Type::getInt32Ty(II->getContext()),
708                                   Idx&15, false), "tmp");
709          }
710
711          // Insert this value into the result vector.
712          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
713                         ConstantInt::get(Type::getInt32Ty(II->getContext()),
714                                          i, false), "tmp");
715        }
716        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
717      }
718    }
719    break;
720
721  case Intrinsic::arm_neon_vld1:
722  case Intrinsic::arm_neon_vld2:
723  case Intrinsic::arm_neon_vld3:
724  case Intrinsic::arm_neon_vld4:
725  case Intrinsic::arm_neon_vld2lane:
726  case Intrinsic::arm_neon_vld3lane:
727  case Intrinsic::arm_neon_vld4lane:
728  case Intrinsic::arm_neon_vst1:
729  case Intrinsic::arm_neon_vst2:
730  case Intrinsic::arm_neon_vst3:
731  case Intrinsic::arm_neon_vst4:
732  case Intrinsic::arm_neon_vst2lane:
733  case Intrinsic::arm_neon_vst3lane:
734  case Intrinsic::arm_neon_vst4lane: {
735    unsigned MemAlign = GetOrEnforceKnownAlignment(II->getArgOperand(0));
736    unsigned AlignArg = II->getNumArgOperands() - 1;
737    ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
738    if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
739      II->setArgOperand(AlignArg,
740                        ConstantInt::get(Type::getInt32Ty(II->getContext()),
741                                         MemAlign, false));
742      return II;
743    }
744    break;
745  }
746
747  case Intrinsic::stackrestore: {
748    // If the save is right next to the restore, remove the restore.  This can
749    // happen when variable allocas are DCE'd.
750    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
751      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
752        BasicBlock::iterator BI = SS;
753        if (&*++BI == II)
754          return EraseInstFromFunction(CI);
755      }
756    }
757
758    // Scan down this block to see if there is another stack restore in the
759    // same block without an intervening call/alloca.
760    BasicBlock::iterator BI = II;
761    TerminatorInst *TI = II->getParent()->getTerminator();
762    bool CannotRemove = false;
763    for (++BI; &*BI != TI; ++BI) {
764      if (isa<AllocaInst>(BI) || isMalloc(BI)) {
765        CannotRemove = true;
766        break;
767      }
768      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
769        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
770          // If there is a stackrestore below this one, remove this one.
771          if (II->getIntrinsicID() == Intrinsic::stackrestore)
772            return EraseInstFromFunction(CI);
773          // Otherwise, ignore the intrinsic.
774        } else {
775          // If we found a non-intrinsic call, we can't remove the stack
776          // restore.
777          CannotRemove = true;
778          break;
779        }
780      }
781    }
782
783    // If the stack restore is in a return/unwind block and if there are no
784    // allocas or calls between the restore and the return, nuke the restore.
785    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
786      return EraseInstFromFunction(CI);
787    break;
788  }
789  }
790
791  return visitCallSite(II);
792}
793
794// InvokeInst simplification
795//
796Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
797  return visitCallSite(&II);
798}
799
800/// isSafeToEliminateVarargsCast - If this cast does not affect the value
801/// passed through the varargs area, we can eliminate the use of the cast.
802static bool isSafeToEliminateVarargsCast(const CallSite CS,
803                                         const CastInst * const CI,
804                                         const TargetData * const TD,
805                                         const int ix) {
806  if (!CI->isLosslessCast())
807    return false;
808
809  // The size of ByVal arguments is derived from the type, so we
810  // can't change to a type with a different size.  If the size were
811  // passed explicitly we could avoid this check.
812  if (!CS.paramHasAttr(ix, Attribute::ByVal))
813    return true;
814
815  const Type* SrcTy =
816            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
817  const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
818  if (!SrcTy->isSized() || !DstTy->isSized())
819    return false;
820  if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
821    return false;
822  return true;
823}
824
825namespace {
826class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
827  InstCombiner *IC;
828protected:
829  void replaceCall(Value *With) {
830    NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
831  }
832  bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
833    if (ConstantInt *SizeCI =
834                           dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
835      if (SizeCI->isAllOnesValue())
836        return true;
837      if (isString)
838        return SizeCI->getZExtValue() >=
839               GetStringLength(CI->getArgOperand(SizeArgOp));
840      if (ConstantInt *Arg = dyn_cast<ConstantInt>(
841                                                  CI->getArgOperand(SizeArgOp)))
842        return SizeCI->getZExtValue() >= Arg->getZExtValue();
843    }
844    return false;
845  }
846public:
847  InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
848  Instruction *NewInstruction;
849};
850} // end anonymous namespace
851
852// Try to fold some different type of calls here.
853// Currently we're only working with the checking functions, memcpy_chk,
854// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
855// strcat_chk and strncat_chk.
856Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
857  if (CI->getCalledFunction() == 0) return 0;
858
859  InstCombineFortifiedLibCalls Simplifier(this);
860  Simplifier.fold(CI, TD);
861  return Simplifier.NewInstruction;
862}
863
864// visitCallSite - Improvements for call and invoke instructions.
865//
866Instruction *InstCombiner::visitCallSite(CallSite CS) {
867  bool Changed = false;
868
869  // If the callee is a constexpr cast of a function, attempt to move the cast
870  // to the arguments of the call/invoke.
871  if (transformConstExprCastCall(CS)) return 0;
872
873  Value *Callee = CS.getCalledValue();
874
875  if (Function *CalleeF = dyn_cast<Function>(Callee))
876    // If the call and callee calling conventions don't match, this call must
877    // be unreachable, as the call is undefined.
878    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
879        // Only do this for calls to a function with a body.  A prototype may
880        // not actually end up matching the implementation's calling conv for a
881        // variety of reasons (e.g. it may be written in assembly).
882        !CalleeF->isDeclaration()) {
883      Instruction *OldCall = CS.getInstruction();
884      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
885                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
886                                  OldCall);
887      // If OldCall dues not return void then replaceAllUsesWith undef.
888      // This allows ValueHandlers and custom metadata to adjust itself.
889      if (!OldCall->getType()->isVoidTy())
890        OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
891      if (isa<CallInst>(OldCall))
892        return EraseInstFromFunction(*OldCall);
893
894      // We cannot remove an invoke, because it would change the CFG, just
895      // change the callee to a null pointer.
896      cast<InvokeInst>(OldCall)->setCalledFunction(
897                                    Constant::getNullValue(CalleeF->getType()));
898      return 0;
899    }
900
901  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
902    // This instruction is not reachable, just remove it.  We insert a store to
903    // undef so that we know that this code is not reachable, despite the fact
904    // that we can't modify the CFG here.
905    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
906               UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
907                  CS.getInstruction());
908
909    // If CS does not return void then replaceAllUsesWith undef.
910    // This allows ValueHandlers and custom metadata to adjust itself.
911    if (!CS.getInstruction()->getType()->isVoidTy())
912      CS.getInstruction()->
913        replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
914
915    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
916      // Don't break the CFG, insert a dummy cond branch.
917      BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
918                         ConstantInt::getTrue(Callee->getContext()), II);
919    }
920    return EraseInstFromFunction(*CS.getInstruction());
921  }
922
923  if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
924    if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
925      if (In->getIntrinsicID() == Intrinsic::init_trampoline)
926        return transformCallThroughTrampoline(CS);
927
928  const PointerType *PTy = cast<PointerType>(Callee->getType());
929  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
930  if (FTy->isVarArg()) {
931    int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
932    // See if we can optimize any arguments passed through the varargs area of
933    // the call.
934    for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
935           E = CS.arg_end(); I != E; ++I, ++ix) {
936      CastInst *CI = dyn_cast<CastInst>(*I);
937      if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
938        *I = CI->getOperand(0);
939        Changed = true;
940      }
941    }
942  }
943
944  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
945    // Inline asm calls cannot throw - mark them 'nounwind'.
946    CS.setDoesNotThrow();
947    Changed = true;
948  }
949
950  // Try to optimize the call if possible, we require TargetData for most of
951  // this.  None of these calls are seen as possibly dead so go ahead and
952  // delete the instruction now.
953  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
954    Instruction *I = tryOptimizeCall(CI, TD);
955    // If we changed something return the result, etc. Otherwise let
956    // the fallthrough check.
957    if (I) return EraseInstFromFunction(*I);
958  }
959
960  return Changed ? CS.getInstruction() : 0;
961}
962
963// transformConstExprCastCall - If the callee is a constexpr cast of a function,
964// attempt to move the cast to the arguments of the call/invoke.
965//
966bool InstCombiner::transformConstExprCastCall(CallSite CS) {
967  if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
968  ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
969  if (CE->getOpcode() != Instruction::BitCast ||
970      !isa<Function>(CE->getOperand(0)))
971    return false;
972  Function *Callee = cast<Function>(CE->getOperand(0));
973  Instruction *Caller = CS.getInstruction();
974  const AttrListPtr &CallerPAL = CS.getAttributes();
975
976  // Okay, this is a cast from a function to a different type.  Unless doing so
977  // would cause a type conversion of one of our arguments, change this call to
978  // be a direct call with arguments casted to the appropriate types.
979  //
980  const FunctionType *FT = Callee->getFunctionType();
981  const Type *OldRetTy = Caller->getType();
982  const Type *NewRetTy = FT->getReturnType();
983
984  if (NewRetTy->isStructTy())
985    return false; // TODO: Handle multiple return values.
986
987  // Check to see if we are changing the return type...
988  if (OldRetTy != NewRetTy) {
989    if (Callee->isDeclaration() &&
990        // Conversion is ok if changing from one pointer type to another or from
991        // a pointer to an integer of the same size.
992        !((OldRetTy->isPointerTy() || !TD ||
993           OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
994          (NewRetTy->isPointerTy() || !TD ||
995           NewRetTy == TD->getIntPtrType(Caller->getContext()))))
996      return false;   // Cannot transform this return value.
997
998    if (!Caller->use_empty() &&
999        // void -> non-void is handled specially
1000        !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
1001      return false;   // Cannot transform this return value.
1002
1003    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1004      Attributes RAttrs = CallerPAL.getRetAttributes();
1005      if (RAttrs & Attribute::typeIncompatible(NewRetTy))
1006        return false;   // Attribute not compatible with transformed value.
1007    }
1008
1009    // If the callsite is an invoke instruction, and the return value is used by
1010    // a PHI node in a successor, we cannot change the return type of the call
1011    // because there is no place to put the cast instruction (without breaking
1012    // the critical edge).  Bail out in this case.
1013    if (!Caller->use_empty())
1014      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1015        for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1016             UI != E; ++UI)
1017          if (PHINode *PN = dyn_cast<PHINode>(*UI))
1018            if (PN->getParent() == II->getNormalDest() ||
1019                PN->getParent() == II->getUnwindDest())
1020              return false;
1021  }
1022
1023  unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1024  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1025
1026  CallSite::arg_iterator AI = CS.arg_begin();
1027  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1028    const Type *ParamTy = FT->getParamType(i);
1029    const Type *ActTy = (*AI)->getType();
1030
1031    if (!CastInst::isCastable(ActTy, ParamTy))
1032      return false;   // Cannot transform this parameter value.
1033
1034    if (CallerPAL.getParamAttributes(i + 1)
1035        & Attribute::typeIncompatible(ParamTy))
1036      return false;   // Attribute not compatible with transformed value.
1037
1038    // Converting from one pointer type to another or between a pointer and an
1039    // integer of the same size is safe even if we do not have a body.
1040    bool isConvertible = ActTy == ParamTy ||
1041      (TD && ((ParamTy->isPointerTy() ||
1042      ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1043              (ActTy->isPointerTy() ||
1044              ActTy == TD->getIntPtrType(Caller->getContext()))));
1045    if (Callee->isDeclaration() && !isConvertible) return false;
1046  }
1047
1048  if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
1049      Callee->isDeclaration())
1050    return false;   // Do not delete arguments unless we have a function body.
1051
1052  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1053      !CallerPAL.isEmpty())
1054    // In this case we have more arguments than the new function type, but we
1055    // won't be dropping them.  Check that these extra arguments have attributes
1056    // that are compatible with being a vararg call argument.
1057    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1058      if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1059        break;
1060      Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1061      if (PAttrs & Attribute::VarArgsIncompatible)
1062        return false;
1063    }
1064
1065  // Okay, we decided that this is a safe thing to do: go ahead and start
1066  // inserting cast instructions as necessary...
1067  std::vector<Value*> Args;
1068  Args.reserve(NumActualArgs);
1069  SmallVector<AttributeWithIndex, 8> attrVec;
1070  attrVec.reserve(NumCommonArgs);
1071
1072  // Get any return attributes.
1073  Attributes RAttrs = CallerPAL.getRetAttributes();
1074
1075  // If the return value is not being used, the type may not be compatible
1076  // with the existing attributes.  Wipe out any problematic attributes.
1077  RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1078
1079  // Add the new return attributes.
1080  if (RAttrs)
1081    attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1082
1083  AI = CS.arg_begin();
1084  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1085    const Type *ParamTy = FT->getParamType(i);
1086    if ((*AI)->getType() == ParamTy) {
1087      Args.push_back(*AI);
1088    } else {
1089      Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1090          false, ParamTy, false);
1091      Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
1092    }
1093
1094    // Add any parameter attributes.
1095    if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1096      attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1097  }
1098
1099  // If the function takes more arguments than the call was taking, add them
1100  // now.
1101  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1102    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1103
1104  // If we are removing arguments to the function, emit an obnoxious warning.
1105  if (FT->getNumParams() < NumActualArgs) {
1106    if (!FT->isVarArg()) {
1107      errs() << "WARNING: While resolving call to function '"
1108             << Callee->getName() << "' arguments were dropped!\n";
1109    } else {
1110      // Add all of the arguments in their promoted form to the arg list.
1111      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1112        const Type *PTy = getPromotedType((*AI)->getType());
1113        if (PTy != (*AI)->getType()) {
1114          // Must promote to pass through va_arg area!
1115          Instruction::CastOps opcode =
1116            CastInst::getCastOpcode(*AI, false, PTy, false);
1117          Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
1118        } else {
1119          Args.push_back(*AI);
1120        }
1121
1122        // Add any parameter attributes.
1123        if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1124          attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1125      }
1126    }
1127  }
1128
1129  if (Attributes FnAttrs =  CallerPAL.getFnAttributes())
1130    attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1131
1132  if (NewRetTy->isVoidTy())
1133    Caller->setName("");   // Void type should not have a name.
1134
1135  const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1136                                                     attrVec.end());
1137
1138  Instruction *NC;
1139  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1140    NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
1141                            Args.begin(), Args.end(),
1142                            Caller->getName(), Caller);
1143    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1144    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1145  } else {
1146    NC = CallInst::Create(Callee, Args.begin(), Args.end(),
1147                          Caller->getName(), Caller);
1148    CallInst *CI = cast<CallInst>(Caller);
1149    if (CI->isTailCall())
1150      cast<CallInst>(NC)->setTailCall();
1151    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1152    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1153  }
1154
1155  // Insert a cast of the return type as necessary.
1156  Value *NV = NC;
1157  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1158    if (!NV->getType()->isVoidTy()) {
1159      Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
1160                                                            OldRetTy, false);
1161      NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
1162
1163      // If this is an invoke instruction, we should insert it after the first
1164      // non-phi, instruction in the normal successor block.
1165      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1166        BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
1167        InsertNewInstBefore(NC, *I);
1168      } else {
1169        // Otherwise, it's a call, just insert cast right after the call instr
1170        InsertNewInstBefore(NC, *Caller);
1171      }
1172      Worklist.AddUsersToWorkList(*Caller);
1173    } else {
1174      NV = UndefValue::get(Caller->getType());
1175    }
1176  }
1177
1178
1179  if (!Caller->use_empty())
1180    Caller->replaceAllUsesWith(NV);
1181
1182  EraseInstFromFunction(*Caller);
1183  return true;
1184}
1185
1186// transformCallThroughTrampoline - Turn a call to a function created by the
1187// init_trampoline intrinsic into a direct call to the underlying function.
1188//
1189Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
1190  Value *Callee = CS.getCalledValue();
1191  const PointerType *PTy = cast<PointerType>(Callee->getType());
1192  const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1193  const AttrListPtr &Attrs = CS.getAttributes();
1194
1195  // If the call already has the 'nest' attribute somewhere then give up -
1196  // otherwise 'nest' would occur twice after splicing in the chain.
1197  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1198    return 0;
1199
1200  IntrinsicInst *Tramp =
1201    cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
1202
1203  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1204  const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1205  const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1206
1207  const AttrListPtr &NestAttrs = NestF->getAttributes();
1208  if (!NestAttrs.isEmpty()) {
1209    unsigned NestIdx = 1;
1210    const Type *NestTy = 0;
1211    Attributes NestAttr = Attribute::None;
1212
1213    // Look for a parameter marked with the 'nest' attribute.
1214    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1215         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1216      if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1217        // Record the parameter type and any other attributes.
1218        NestTy = *I;
1219        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1220        break;
1221      }
1222
1223    if (NestTy) {
1224      Instruction *Caller = CS.getInstruction();
1225      std::vector<Value*> NewArgs;
1226      NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1227
1228      SmallVector<AttributeWithIndex, 8> NewAttrs;
1229      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1230
1231      // Insert the nest argument into the call argument list, which may
1232      // mean appending it.  Likewise for attributes.
1233
1234      // Add any result attributes.
1235      if (Attributes Attr = Attrs.getRetAttributes())
1236        NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1237
1238      {
1239        unsigned Idx = 1;
1240        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1241        do {
1242          if (Idx == NestIdx) {
1243            // Add the chain argument and attributes.
1244            Value *NestVal = Tramp->getArgOperand(2);
1245            if (NestVal->getType() != NestTy)
1246              NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
1247            NewArgs.push_back(NestVal);
1248            NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1249          }
1250
1251          if (I == E)
1252            break;
1253
1254          // Add the original argument and attributes.
1255          NewArgs.push_back(*I);
1256          if (Attributes Attr = Attrs.getParamAttributes(Idx))
1257            NewAttrs.push_back
1258              (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1259
1260          ++Idx, ++I;
1261        } while (1);
1262      }
1263
1264      // Add any function attributes.
1265      if (Attributes Attr = Attrs.getFnAttributes())
1266        NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1267
1268      // The trampoline may have been bitcast to a bogus type (FTy).
1269      // Handle this by synthesizing a new function type, equal to FTy
1270      // with the chain parameter inserted.
1271
1272      std::vector<const Type*> NewTypes;
1273      NewTypes.reserve(FTy->getNumParams()+1);
1274
1275      // Insert the chain's type into the list of parameter types, which may
1276      // mean appending it.
1277      {
1278        unsigned Idx = 1;
1279        FunctionType::param_iterator I = FTy->param_begin(),
1280          E = FTy->param_end();
1281
1282        do {
1283          if (Idx == NestIdx)
1284            // Add the chain's type.
1285            NewTypes.push_back(NestTy);
1286
1287          if (I == E)
1288            break;
1289
1290          // Add the original type.
1291          NewTypes.push_back(*I);
1292
1293          ++Idx, ++I;
1294        } while (1);
1295      }
1296
1297      // Replace the trampoline call with a direct call.  Let the generic
1298      // code sort out any function type mismatches.
1299      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1300                                                FTy->isVarArg());
1301      Constant *NewCallee =
1302        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1303        NestF : ConstantExpr::getBitCast(NestF,
1304                                         PointerType::getUnqual(NewFTy));
1305      const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1306                                                   NewAttrs.end());
1307
1308      Instruction *NewCaller;
1309      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1310        NewCaller = InvokeInst::Create(NewCallee,
1311                                       II->getNormalDest(), II->getUnwindDest(),
1312                                       NewArgs.begin(), NewArgs.end(),
1313                                       Caller->getName(), Caller);
1314        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1315        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1316      } else {
1317        NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
1318                                     Caller->getName(), Caller);
1319        if (cast<CallInst>(Caller)->isTailCall())
1320          cast<CallInst>(NewCaller)->setTailCall();
1321        cast<CallInst>(NewCaller)->
1322          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1323        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1324      }
1325      if (!Caller->getType()->isVoidTy())
1326        Caller->replaceAllUsesWith(NewCaller);
1327      Caller->eraseFromParent();
1328      Worklist.Remove(Caller);
1329      return 0;
1330    }
1331  }
1332
1333  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1334  // parameter, there is no need to adjust the argument list.  Let the generic
1335  // code sort out any function type mismatches.
1336  Constant *NewCallee =
1337    NestF->getType() == PTy ? NestF :
1338                              ConstantExpr::getBitCast(NestF, PTy);
1339  CS.setCalledFunction(NewCallee);
1340  return CS.getInstruction();
1341}
1342
1343