InstCombineCalls.cpp revision aa5abe88d6aa445afa593476a665e3ab14b3524c 
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/Support/CallSite.h"
16#include "llvm/Target/TargetData.h"
17#include "llvm/Analysis/MemoryBuiltins.h"
18#include "llvm/Transforms/Utils/BuildLibCalls.h"
19#include "llvm/Transforms/Utils/Local.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 Type *getPromotedType(Type *Ty) {
25  if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
26    if (ITy->getBitWidth() < 32)
27      return Type::getInt32Ty(Ty->getContext());
28  }
29  return Ty;
30}
31
32
33Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
34  unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD);
35  unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD);
36  unsigned MinAlign = std::min(DstAlign, SrcAlign);
37  unsigned CopyAlign = MI->getAlignment();
38
39  if (CopyAlign < MinAlign) {
40    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
41                                             MinAlign, false));
42    return MI;
43  }
44
45  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
46  // load/store.
47  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
48  if (MemOpLength == 0) return 0;
49
50  // Source and destination pointer types are always "i8*" for intrinsic.  See
51  // if the size is something we can handle with a single primitive load/store.
52  // A single load+store correctly handles overlapping memory in the memmove
53  // case.
54  unsigned Size = MemOpLength->getZExtValue();
55  if (Size == 0) return MI;  // Delete this mem transfer.
56
57  if (Size > 8 || (Size&(Size-1)))
58    return 0;  // If not 1/2/4/8 bytes, exit.
59
60  // Use an integer load+store unless we can find something better.
61  unsigned SrcAddrSp =
62    cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
63  unsigned DstAddrSp =
64    cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
65
66  IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
67  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
68  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
69
70  // Memcpy forces the use of i8* for the source and destination.  That means
71  // that if you're using memcpy to move one double around, you'll get a cast
72  // from double* to i8*.  We'd much rather use a double load+store rather than
73  // an i64 load+store, here because this improves the odds that the source or
74  // dest address will be promotable.  See if we can find a better type than the
75  // integer datatype.
76  Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
77  if (StrippedDest != MI->getArgOperand(0)) {
78    Type *SrcETy = cast<PointerType>(StrippedDest->getType())
79                                    ->getElementType();
80    if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
81      // The SrcETy might be something like {{{double}}} or [1 x double].  Rip
82      // down through these levels if so.
83      while (!SrcETy->isSingleValueType()) {
84        if (StructType *STy = dyn_cast<StructType>(SrcETy)) {
85          if (STy->getNumElements() == 1)
86            SrcETy = STy->getElementType(0);
87          else
88            break;
89        } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
90          if (ATy->getNumElements() == 1)
91            SrcETy = ATy->getElementType();
92          else
93            break;
94        } else
95          break;
96      }
97
98      if (SrcETy->isSingleValueType()) {
99        NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
100        NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
101      }
102    }
103  }
104
105
106  // If the memcpy/memmove provides better alignment info than we can
107  // infer, use it.
108  SrcAlign = std::max(SrcAlign, CopyAlign);
109  DstAlign = std::max(DstAlign, CopyAlign);
110
111  Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
112  Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
113  LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
114  L->setAlignment(SrcAlign);
115  StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
116  S->setAlignment(DstAlign);
117
118  // Set the size of the copy to 0, it will be deleted on the next iteration.
119  MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
120  return MI;
121}
122
123Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
124  unsigned Alignment = getKnownAlignment(MI->getDest(), TD);
125  if (MI->getAlignment() < Alignment) {
126    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
127                                             Alignment, false));
128    return MI;
129  }
130
131  // Extract the length and alignment and fill if they are constant.
132  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
133  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
134  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
135    return 0;
136  uint64_t Len = LenC->getZExtValue();
137  Alignment = MI->getAlignment();
138
139  // If the length is zero, this is a no-op
140  if (Len == 0) return MI; // memset(d,c,0,a) -> noop
141
142  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
143  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
144    Type *ITy = IntegerType::get(MI->getContext(), Len*8);  // n=1 -> i8.
145
146    Value *Dest = MI->getDest();
147    unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
148    Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
149    Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
150
151    // Alignment 0 is identity for alignment 1 for memset, but not store.
152    if (Alignment == 0) Alignment = 1;
153
154    // Extract the fill value and store.
155    uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
156    StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
157                                        MI->isVolatile());
158    S->setAlignment(Alignment);
159
160    // Set the size of the copy to 0, it will be deleted on the next iteration.
161    MI->setLength(Constant::getNullValue(LenC->getType()));
162    return MI;
163  }
164
165  return 0;
166}
167
168/// visitCallInst - CallInst simplification.  This mostly only handles folding
169/// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
170/// the heavy lifting.
171///
172Instruction *InstCombiner::visitCallInst(CallInst &CI) {
173  if (isFreeCall(&CI))
174    return visitFree(CI);
175  if (isMalloc(&CI))
176    return visitMalloc(CI);
177
178  // If the caller function is nounwind, mark the call as nounwind, even if the
179  // callee isn't.
180  if (CI.getParent()->getParent()->doesNotThrow() &&
181      !CI.doesNotThrow()) {
182    CI.setDoesNotThrow();
183    return &CI;
184  }
185
186  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
187  if (!II) return visitCallSite(&CI);
188
189  // Intrinsics cannot occur in an invoke, so handle them here instead of in
190  // visitCallSite.
191  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
192    bool Changed = false;
193
194    // memmove/cpy/set of zero bytes is a noop.
195    if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
196      if (NumBytes->isNullValue())
197        return EraseInstFromFunction(CI);
198
199      if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
200        if (CI->getZExtValue() == 1) {
201          // Replace the instruction with just byte operations.  We would
202          // transform other cases to loads/stores, but we don't know if
203          // alignment is sufficient.
204        }
205    }
206
207    // No other transformations apply to volatile transfers.
208    if (MI->isVolatile())
209      return 0;
210
211    // If we have a memmove and the source operation is a constant global,
212    // then the source and dest pointers can't alias, so we can change this
213    // into a call to memcpy.
214    if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
215      if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
216        if (GVSrc->isConstant()) {
217          Module *M = CI.getParent()->getParent()->getParent();
218          Intrinsic::ID MemCpyID = Intrinsic::memcpy;
219          Type *Tys[3] = { CI.getArgOperand(0)->getType(),
220                           CI.getArgOperand(1)->getType(),
221                           CI.getArgOperand(2)->getType() };
222          CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
223          Changed = true;
224        }
225    }
226
227    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
228      // memmove(x,x,size) -> noop.
229      if (MTI->getSource() == MTI->getDest())
230        return EraseInstFromFunction(CI);
231    }
232
233    // If we can determine a pointer alignment that is bigger than currently
234    // set, update the alignment.
235    if (isa<MemTransferInst>(MI)) {
236      if (Instruction *I = SimplifyMemTransfer(MI))
237        return I;
238    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
239      if (Instruction *I = SimplifyMemSet(MSI))
240        return I;
241    }
242
243    if (Changed) return II;
244  }
245
246  switch (II->getIntrinsicID()) {
247  default: break;
248  case Intrinsic::objectsize: {
249    // We need target data for just about everything so depend on it.
250    if (!TD) break;
251
252    Type *ReturnTy = CI.getType();
253    uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL;
254
255    // Get to the real allocated thing and offset as fast as possible.
256    Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
257
258    uint64_t Offset = 0;
259    uint64_t Size = -1ULL;
260
261    // Try to look through constant GEPs.
262    if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) {
263      if (!GEP->hasAllConstantIndices()) break;
264
265      // Get the current byte offset into the thing. Use the original
266      // operand in case we're looking through a bitcast.
267      SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
268      if (!GEP->getPointerOperandType()->isPointerTy())
269        return 0;
270      Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops);
271
272      Op1 = GEP->getPointerOperand()->stripPointerCasts();
273
274      // Make sure we're not a constant offset from an external
275      // global.
276      if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1))
277        if (!GV->hasDefinitiveInitializer()) break;
278    }
279
280    // If we've stripped down to a single global variable that we
281    // can know the size of then just return that.
282    if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) {
283      if (GV->hasDefinitiveInitializer()) {
284        Constant *C = GV->getInitializer();
285        Size = TD->getTypeAllocSize(C->getType());
286      } else {
287        // Can't determine size of the GV.
288        Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow);
289        return ReplaceInstUsesWith(CI, RetVal);
290      }
291    } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
292      // Get alloca size.
293      if (AI->getAllocatedType()->isSized()) {
294        Size = TD->getTypeAllocSize(AI->getAllocatedType());
295        if (AI->isArrayAllocation()) {
296          const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize());
297          if (!C) break;
298          Size *= C->getZExtValue();
299        }
300      }
301    } else if (CallInst *MI = extractMallocCall(Op1)) {
302      // Get allocation size.
303      Type* MallocType = getMallocAllocatedType(MI);
304      if (MallocType && MallocType->isSized())
305        if (Value *NElems = getMallocArraySize(MI, TD, true))
306          if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
307            Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType);
308    }
309
310    // Do not return "I don't know" here. Later optimization passes could
311    // make it possible to evaluate objectsize to a constant.
312    if (Size == -1ULL)
313      break;
314
315    if (Size < Offset) {
316      // Out of bound reference? Negative index normalized to large
317      // index? Just return "I don't know".
318      return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow));
319    }
320    return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
321  }
322  case Intrinsic::bswap:
323    // bswap(bswap(x)) -> x
324    if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
325      if (Operand->getIntrinsicID() == Intrinsic::bswap)
326        return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
327
328    // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
329    if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
330      if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
331        if (Operand->getIntrinsicID() == Intrinsic::bswap) {
332          unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
333                       TI->getType()->getPrimitiveSizeInBits();
334          Value *CV = ConstantInt::get(Operand->getType(), C);
335          Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
336          return new TruncInst(V, TI->getType());
337        }
338    }
339
340    break;
341  case Intrinsic::powi:
342    if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
343      // powi(x, 0) -> 1.0
344      if (Power->isZero())
345        return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
346      // powi(x, 1) -> x
347      if (Power->isOne())
348        return ReplaceInstUsesWith(CI, II->getArgOperand(0));
349      // powi(x, -1) -> 1/x
350      if (Power->isAllOnesValue())
351        return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
352                                          II->getArgOperand(0));
353    }
354    break;
355  case Intrinsic::cttz: {
356    // If all bits below the first known one are known zero,
357    // this value is constant.
358    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
359    // FIXME: Try to simplify vectors of integers.
360    if (!IT) break;
361    uint32_t BitWidth = IT->getBitWidth();
362    APInt KnownZero(BitWidth, 0);
363    APInt KnownOne(BitWidth, 0);
364    ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
365                      KnownZero, KnownOne);
366    unsigned TrailingZeros = KnownOne.countTrailingZeros();
367    APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
368    if ((Mask & KnownZero) == Mask)
369      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
370                                 APInt(BitWidth, TrailingZeros)));
371
372    }
373    break;
374  case Intrinsic::ctlz: {
375    // If all bits above the first known one are known zero,
376    // this value is constant.
377    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
378    // FIXME: Try to simplify vectors of integers.
379    if (!IT) break;
380    uint32_t BitWidth = IT->getBitWidth();
381    APInt KnownZero(BitWidth, 0);
382    APInt KnownOne(BitWidth, 0);
383    ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth),
384                      KnownZero, KnownOne);
385    unsigned LeadingZeros = KnownOne.countLeadingZeros();
386    APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
387    if ((Mask & KnownZero) == Mask)
388      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
389                                 APInt(BitWidth, LeadingZeros)));
390
391    }
392    break;
393  case Intrinsic::uadd_with_overflow: {
394    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
395    IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
396    uint32_t BitWidth = IT->getBitWidth();
397    APInt Mask = APInt::getSignBit(BitWidth);
398    APInt LHSKnownZero(BitWidth, 0);
399    APInt LHSKnownOne(BitWidth, 0);
400    ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
401    bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
402    bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
403
404    if (LHSKnownNegative || LHSKnownPositive) {
405      APInt RHSKnownZero(BitWidth, 0);
406      APInt RHSKnownOne(BitWidth, 0);
407      ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
408      bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
409      bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
410      if (LHSKnownNegative && RHSKnownNegative) {
411        // The sign bit is set in both cases: this MUST overflow.
412        // Create a simple add instruction, and insert it into the struct.
413        Value *Add = Builder->CreateAdd(LHS, RHS);
414        Add->takeName(&CI);
415        Constant *V[] = {
416          UndefValue::get(LHS->getType()),
417          ConstantInt::getTrue(II->getContext())
418        };
419        StructType *ST = cast<StructType>(II->getType());
420        Constant *Struct = ConstantStruct::get(ST, V);
421        return InsertValueInst::Create(Struct, Add, 0);
422      }
423
424      if (LHSKnownPositive && RHSKnownPositive) {
425        // The sign bit is clear in both cases: this CANNOT overflow.
426        // Create a simple add instruction, and insert it into the struct.
427        Value *Add = Builder->CreateNUWAdd(LHS, RHS);
428        Add->takeName(&CI);
429        Constant *V[] = {
430          UndefValue::get(LHS->getType()),
431          ConstantInt::getFalse(II->getContext())
432        };
433        StructType *ST = cast<StructType>(II->getType());
434        Constant *Struct = ConstantStruct::get(ST, V);
435        return InsertValueInst::Create(Struct, Add, 0);
436      }
437    }
438  }
439  // FALL THROUGH uadd into sadd
440  case Intrinsic::sadd_with_overflow:
441    // Canonicalize constants into the RHS.
442    if (isa<Constant>(II->getArgOperand(0)) &&
443        !isa<Constant>(II->getArgOperand(1))) {
444      Value *LHS = II->getArgOperand(0);
445      II->setArgOperand(0, II->getArgOperand(1));
446      II->setArgOperand(1, LHS);
447      return II;
448    }
449
450    // X + undef -> undef
451    if (isa<UndefValue>(II->getArgOperand(1)))
452      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
453
454    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
455      // X + 0 -> {X, false}
456      if (RHS->isZero()) {
457        Constant *V[] = {
458          UndefValue::get(II->getArgOperand(0)->getType()),
459          ConstantInt::getFalse(II->getContext())
460        };
461        Constant *Struct =
462          ConstantStruct::get(cast<StructType>(II->getType()), V);
463        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
464      }
465    }
466    break;
467  case Intrinsic::usub_with_overflow:
468  case Intrinsic::ssub_with_overflow:
469    // undef - X -> undef
470    // X - undef -> undef
471    if (isa<UndefValue>(II->getArgOperand(0)) ||
472        isa<UndefValue>(II->getArgOperand(1)))
473      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
474
475    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
476      // X - 0 -> {X, false}
477      if (RHS->isZero()) {
478        Constant *V[] = {
479          UndefValue::get(II->getArgOperand(0)->getType()),
480          ConstantInt::getFalse(II->getContext())
481        };
482        Constant *Struct =
483          ConstantStruct::get(cast<StructType>(II->getType()), V);
484        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
485      }
486    }
487    break;
488  case Intrinsic::umul_with_overflow: {
489    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
490    unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
491    APInt Mask = APInt::getAllOnesValue(BitWidth);
492
493    APInt LHSKnownZero(BitWidth, 0);
494    APInt LHSKnownOne(BitWidth, 0);
495    ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
496    APInt RHSKnownZero(BitWidth, 0);
497    APInt RHSKnownOne(BitWidth, 0);
498    ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
499
500    // Get the largest possible values for each operand.
501    APInt LHSMax = ~LHSKnownZero;
502    APInt RHSMax = ~RHSKnownZero;
503
504    // If multiplying the maximum values does not overflow then we can turn
505    // this into a plain NUW mul.
506    bool Overflow;
507    LHSMax.umul_ov(RHSMax, Overflow);
508    if (!Overflow) {
509      Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
510      Constant *V[] = {
511        UndefValue::get(LHS->getType()),
512        Builder->getFalse()
513      };
514      Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
515      return InsertValueInst::Create(Struct, Mul, 0);
516    }
517  } // FALL THROUGH
518  case Intrinsic::smul_with_overflow:
519    // Canonicalize constants into the RHS.
520    if (isa<Constant>(II->getArgOperand(0)) &&
521        !isa<Constant>(II->getArgOperand(1))) {
522      Value *LHS = II->getArgOperand(0);
523      II->setArgOperand(0, II->getArgOperand(1));
524      II->setArgOperand(1, LHS);
525      return II;
526    }
527
528    // X * undef -> undef
529    if (isa<UndefValue>(II->getArgOperand(1)))
530      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
531
532    if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
533      // X*0 -> {0, false}
534      if (RHSI->isZero())
535        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
536
537      // X * 1 -> {X, false}
538      if (RHSI->equalsInt(1)) {
539        Constant *V[] = {
540          UndefValue::get(II->getArgOperand(0)->getType()),
541          ConstantInt::getFalse(II->getContext())
542        };
543        Constant *Struct =
544          ConstantStruct::get(cast<StructType>(II->getType()), V);
545        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
546      }
547    }
548    break;
549  case Intrinsic::ppc_altivec_lvx:
550  case Intrinsic::ppc_altivec_lvxl:
551    // Turn PPC lvx -> load if the pointer is known aligned.
552    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
553      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
554                                         PointerType::getUnqual(II->getType()));
555      return new LoadInst(Ptr);
556    }
557    break;
558  case Intrinsic::ppc_altivec_stvx:
559  case Intrinsic::ppc_altivec_stvxl:
560    // Turn stvx -> store if the pointer is known aligned.
561    if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
562      Type *OpPtrTy =
563        PointerType::getUnqual(II->getArgOperand(0)->getType());
564      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
565      return new StoreInst(II->getArgOperand(0), Ptr);
566    }
567    break;
568  case Intrinsic::x86_sse_storeu_ps:
569  case Intrinsic::x86_sse2_storeu_pd:
570  case Intrinsic::x86_sse2_storeu_dq:
571    // Turn X86 storeu -> store if the pointer is known aligned.
572    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
573      Type *OpPtrTy =
574        PointerType::getUnqual(II->getArgOperand(1)->getType());
575      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
576      return new StoreInst(II->getArgOperand(1), Ptr);
577    }
578    break;
579
580  case Intrinsic::x86_sse_cvtss2si:
581  case Intrinsic::x86_sse_cvtss2si64:
582  case Intrinsic::x86_sse_cvttss2si:
583  case Intrinsic::x86_sse_cvttss2si64:
584  case Intrinsic::x86_sse2_cvtsd2si:
585  case Intrinsic::x86_sse2_cvtsd2si64:
586  case Intrinsic::x86_sse2_cvttsd2si:
587  case Intrinsic::x86_sse2_cvttsd2si64: {
588    // These intrinsics only demand the 0th element of their input vectors. If
589    // we can simplify the input based on that, do so now.
590    unsigned VWidth =
591      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
592    APInt DemandedElts(VWidth, 1);
593    APInt UndefElts(VWidth, 0);
594    if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
595                                              DemandedElts, UndefElts)) {
596      II->setArgOperand(0, V);
597      return II;
598    }
599    break;
600  }
601
602
603  case Intrinsic::x86_sse41_pmovsxbw:
604  case Intrinsic::x86_sse41_pmovsxwd:
605  case Intrinsic::x86_sse41_pmovsxdq:
606  case Intrinsic::x86_sse41_pmovzxbw:
607  case Intrinsic::x86_sse41_pmovzxwd:
608  case Intrinsic::x86_sse41_pmovzxdq: {
609    // pmov{s|z}x ignores the upper half of their input vectors.
610    unsigned VWidth =
611      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
612    unsigned LowHalfElts = VWidth / 2;
613    APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
614    APInt UndefElts(VWidth, 0);
615    if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
616                                                 InputDemandedElts,
617                                                 UndefElts)) {
618      II->setArgOperand(0, TmpV);
619      return II;
620    }
621    break;
622  }
623
624  case Intrinsic::ppc_altivec_vperm:
625    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
626    if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
627      assert(Mask->getType()->getVectorNumElements() == 16 &&
628             "Bad type for intrinsic!");
629
630      // Check that all of the elements are integer constants or undefs.
631      bool AllEltsOk = true;
632      for (unsigned i = 0; i != 16; ++i) {
633        Constant *Elt = Mask->getAggregateElement(i);
634        if (Elt == 0 ||
635            !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
636          AllEltsOk = false;
637          break;
638        }
639      }
640
641      if (AllEltsOk) {
642        // Cast the input vectors to byte vectors.
643        Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
644                                            Mask->getType());
645        Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
646                                            Mask->getType());
647        Value *Result = UndefValue::get(Op0->getType());
648
649        // Only extract each element once.
650        Value *ExtractedElts[32];
651        memset(ExtractedElts, 0, sizeof(ExtractedElts));
652
653        for (unsigned i = 0; i != 16; ++i) {
654          if (isa<UndefValue>(Mask->getAggregateElement(i)))
655            continue;
656          unsigned Idx =
657            cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
658          Idx &= 31;  // Match the hardware behavior.
659
660          if (ExtractedElts[Idx] == 0) {
661            ExtractedElts[Idx] =
662              Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
663                                            Builder->getInt32(Idx&15));
664          }
665
666          // Insert this value into the result vector.
667          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
668                                                Builder->getInt32(i));
669        }
670        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
671      }
672    }
673    break;
674
675  case Intrinsic::arm_neon_vld1:
676  case Intrinsic::arm_neon_vld2:
677  case Intrinsic::arm_neon_vld3:
678  case Intrinsic::arm_neon_vld4:
679  case Intrinsic::arm_neon_vld2lane:
680  case Intrinsic::arm_neon_vld3lane:
681  case Intrinsic::arm_neon_vld4lane:
682  case Intrinsic::arm_neon_vst1:
683  case Intrinsic::arm_neon_vst2:
684  case Intrinsic::arm_neon_vst3:
685  case Intrinsic::arm_neon_vst4:
686  case Intrinsic::arm_neon_vst2lane:
687  case Intrinsic::arm_neon_vst3lane:
688  case Intrinsic::arm_neon_vst4lane: {
689    unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
690    unsigned AlignArg = II->getNumArgOperands() - 1;
691    ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
692    if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
693      II->setArgOperand(AlignArg,
694                        ConstantInt::get(Type::getInt32Ty(II->getContext()),
695                                         MemAlign, false));
696      return II;
697    }
698    break;
699  }
700
701  case Intrinsic::stackrestore: {
702    // If the save is right next to the restore, remove the restore.  This can
703    // happen when variable allocas are DCE'd.
704    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
705      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
706        BasicBlock::iterator BI = SS;
707        if (&*++BI == II)
708          return EraseInstFromFunction(CI);
709      }
710    }
711
712    // Scan down this block to see if there is another stack restore in the
713    // same block without an intervening call/alloca.
714    BasicBlock::iterator BI = II;
715    TerminatorInst *TI = II->getParent()->getTerminator();
716    bool CannotRemove = false;
717    for (++BI; &*BI != TI; ++BI) {
718      if (isa<AllocaInst>(BI) || isMalloc(BI)) {
719        CannotRemove = true;
720        break;
721      }
722      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
723        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
724          // If there is a stackrestore below this one, remove this one.
725          if (II->getIntrinsicID() == Intrinsic::stackrestore)
726            return EraseInstFromFunction(CI);
727          // Otherwise, ignore the intrinsic.
728        } else {
729          // If we found a non-intrinsic call, we can't remove the stack
730          // restore.
731          CannotRemove = true;
732          break;
733        }
734      }
735    }
736
737    // If the stack restore is in a return, resume, or unwind block and if there
738    // are no allocas or calls between the restore and the return, nuke the
739    // restore.
740    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
741      return EraseInstFromFunction(CI);
742    break;
743  }
744  }
745
746  return visitCallSite(II);
747}
748
749// InvokeInst simplification
750//
751Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
752  return visitCallSite(&II);
753}
754
755/// isSafeToEliminateVarargsCast - If this cast does not affect the value
756/// passed through the varargs area, we can eliminate the use of the cast.
757static bool isSafeToEliminateVarargsCast(const CallSite CS,
758                                         const CastInst * const CI,
759                                         const TargetData * const TD,
760                                         const int ix) {
761  if (!CI->isLosslessCast())
762    return false;
763
764  // The size of ByVal arguments is derived from the type, so we
765  // can't change to a type with a different size.  If the size were
766  // passed explicitly we could avoid this check.
767  if (!CS.isByValArgument(ix))
768    return true;
769
770  Type* SrcTy =
771            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
772  Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
773  if (!SrcTy->isSized() || !DstTy->isSized())
774    return false;
775  if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
776    return false;
777  return true;
778}
779
780namespace {
781class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
782  InstCombiner *IC;
783protected:
784  void replaceCall(Value *With) {
785    NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
786  }
787  bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
788    if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
789      return true;
790    if (ConstantInt *SizeCI =
791                           dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
792      if (SizeCI->isAllOnesValue())
793        return true;
794      if (isString) {
795        uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
796        // If the length is 0 we don't know how long it is and so we can't
797        // remove the check.
798        if (Len == 0) return false;
799        return SizeCI->getZExtValue() >= Len;
800      }
801      if (ConstantInt *Arg = dyn_cast<ConstantInt>(
802                                                  CI->getArgOperand(SizeArgOp)))
803        return SizeCI->getZExtValue() >= Arg->getZExtValue();
804    }
805    return false;
806  }
807public:
808  InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
809  Instruction *NewInstruction;
810};
811} // end anonymous namespace
812
813// Try to fold some different type of calls here.
814// Currently we're only working with the checking functions, memcpy_chk,
815// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
816// strcat_chk and strncat_chk.
817Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
818  if (CI->getCalledFunction() == 0) return 0;
819
820  InstCombineFortifiedLibCalls Simplifier(this);
821  Simplifier.fold(CI, TD);
822  return Simplifier.NewInstruction;
823}
824
825static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
826  // Strip off at most one level of pointer casts, looking for an alloca.  This
827  // is good enough in practice and simpler than handling any number of casts.
828  Value *Underlying = TrampMem->stripPointerCasts();
829  if (Underlying != TrampMem &&
830      (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
831    return 0;
832  if (!isa<AllocaInst>(Underlying))
833    return 0;
834
835  IntrinsicInst *InitTrampoline = 0;
836  for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
837       I != E; I++) {
838    IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
839    if (!II)
840      return 0;
841    if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
842      if (InitTrampoline)
843        // More than one init_trampoline writes to this value.  Give up.
844        return 0;
845      InitTrampoline = II;
846      continue;
847    }
848    if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
849      // Allow any number of calls to adjust.trampoline.
850      continue;
851    return 0;
852  }
853
854  // No call to init.trampoline found.
855  if (!InitTrampoline)
856    return 0;
857
858  // Check that the alloca is being used in the expected way.
859  if (InitTrampoline->getOperand(0) != TrampMem)
860    return 0;
861
862  return InitTrampoline;
863}
864
865static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
866                                               Value *TrampMem) {
867  // Visit all the previous instructions in the basic block, and try to find a
868  // init.trampoline which has a direct path to the adjust.trampoline.
869  for (BasicBlock::iterator I = AdjustTramp,
870       E = AdjustTramp->getParent()->begin(); I != E; ) {
871    Instruction *Inst = --I;
872    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
873      if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
874          II->getOperand(0) == TrampMem)
875        return II;
876    if (Inst->mayWriteToMemory())
877      return 0;
878  }
879  return 0;
880}
881
882// Given a call to llvm.adjust.trampoline, find and return the corresponding
883// call to llvm.init.trampoline if the call to the trampoline can be optimized
884// to a direct call to a function.  Otherwise return NULL.
885//
886static IntrinsicInst *FindInitTrampoline(Value *Callee) {
887  Callee = Callee->stripPointerCasts();
888  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
889  if (!AdjustTramp ||
890      AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
891    return 0;
892
893  Value *TrampMem = AdjustTramp->getOperand(0);
894
895  if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
896    return IT;
897  if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
898    return IT;
899  return 0;
900}
901
902// visitCallSite - Improvements for call and invoke instructions.
903//
904Instruction *InstCombiner::visitCallSite(CallSite CS) {
905  bool Changed = false;
906
907  // If the callee is a pointer to a function, attempt to move any casts to the
908  // arguments of the call/invoke.
909  Value *Callee = CS.getCalledValue();
910  if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
911    return 0;
912
913  if (Function *CalleeF = dyn_cast<Function>(Callee))
914    // If the call and callee calling conventions don't match, this call must
915    // be unreachable, as the call is undefined.
916    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
917        // Only do this for calls to a function with a body.  A prototype may
918        // not actually end up matching the implementation's calling conv for a
919        // variety of reasons (e.g. it may be written in assembly).
920        !CalleeF->isDeclaration()) {
921      Instruction *OldCall = CS.getInstruction();
922      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
923                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
924                                  OldCall);
925      // If OldCall dues not return void then replaceAllUsesWith undef.
926      // This allows ValueHandlers and custom metadata to adjust itself.
927      if (!OldCall->getType()->isVoidTy())
928        ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
929      if (isa<CallInst>(OldCall))
930        return EraseInstFromFunction(*OldCall);
931
932      // We cannot remove an invoke, because it would change the CFG, just
933      // change the callee to a null pointer.
934      cast<InvokeInst>(OldCall)->setCalledFunction(
935                                    Constant::getNullValue(CalleeF->getType()));
936      return 0;
937    }
938
939  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
940    // This instruction is not reachable, just remove it.  We insert a store to
941    // undef so that we know that this code is not reachable, despite the fact
942    // that we can't modify the CFG here.
943    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
944               UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
945                  CS.getInstruction());
946
947    // If CS does not return void then replaceAllUsesWith undef.
948    // This allows ValueHandlers and custom metadata to adjust itself.
949    if (!CS.getInstruction()->getType()->isVoidTy())
950      ReplaceInstUsesWith(*CS.getInstruction(),
951                          UndefValue::get(CS.getInstruction()->getType()));
952
953    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
954      // Don't break the CFG, insert a dummy cond branch.
955      BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
956                         ConstantInt::getTrue(Callee->getContext()), II);
957    }
958    return EraseInstFromFunction(*CS.getInstruction());
959  }
960
961  if (IntrinsicInst *II = FindInitTrampoline(Callee))
962    return transformCallThroughTrampoline(CS, II);
963
964  PointerType *PTy = cast<PointerType>(Callee->getType());
965  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
966  if (FTy->isVarArg()) {
967    int ix = FTy->getNumParams();
968    // See if we can optimize any arguments passed through the varargs area of
969    // the call.
970    for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
971           E = CS.arg_end(); I != E; ++I, ++ix) {
972      CastInst *CI = dyn_cast<CastInst>(*I);
973      if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
974        *I = CI->getOperand(0);
975        Changed = true;
976      }
977    }
978  }
979
980  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
981    // Inline asm calls cannot throw - mark them 'nounwind'.
982    CS.setDoesNotThrow();
983    Changed = true;
984  }
985
986  // Try to optimize the call if possible, we require TargetData for most of
987  // this.  None of these calls are seen as possibly dead so go ahead and
988  // delete the instruction now.
989  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
990    Instruction *I = tryOptimizeCall(CI, TD);
991    // If we changed something return the result, etc. Otherwise let
992    // the fallthrough check.
993    if (I) return EraseInstFromFunction(*I);
994  }
995
996  return Changed ? CS.getInstruction() : 0;
997}
998
999// transformConstExprCastCall - If the callee is a constexpr cast of a function,
1000// attempt to move the cast to the arguments of the call/invoke.
1001//
1002bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1003  Function *Callee =
1004    dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1005  if (Callee == 0)
1006    return false;
1007  Instruction *Caller = CS.getInstruction();
1008  const AttrListPtr &CallerPAL = CS.getAttributes();
1009
1010  // Okay, this is a cast from a function to a different type.  Unless doing so
1011  // would cause a type conversion of one of our arguments, change this call to
1012  // be a direct call with arguments casted to the appropriate types.
1013  //
1014  FunctionType *FT = Callee->getFunctionType();
1015  Type *OldRetTy = Caller->getType();
1016  Type *NewRetTy = FT->getReturnType();
1017
1018  if (NewRetTy->isStructTy())
1019    return false; // TODO: Handle multiple return values.
1020
1021  // Check to see if we are changing the return type...
1022  if (OldRetTy != NewRetTy) {
1023    if (Callee->isDeclaration() &&
1024        // Conversion is ok if changing from one pointer type to another or from
1025        // a pointer to an integer of the same size.
1026        !((OldRetTy->isPointerTy() || !TD ||
1027           OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
1028          (NewRetTy->isPointerTy() || !TD ||
1029           NewRetTy == TD->getIntPtrType(Caller->getContext()))))
1030      return false;   // Cannot transform this return value.
1031
1032    if (!Caller->use_empty() &&
1033        // void -> non-void is handled specially
1034        !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
1035      return false;   // Cannot transform this return value.
1036
1037    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1038      Attributes RAttrs = CallerPAL.getRetAttributes();
1039      if (RAttrs & Attribute::typeIncompatible(NewRetTy))
1040        return false;   // Attribute not compatible with transformed value.
1041    }
1042
1043    // If the callsite is an invoke instruction, and the return value is used by
1044    // a PHI node in a successor, we cannot change the return type of the call
1045    // because there is no place to put the cast instruction (without breaking
1046    // the critical edge).  Bail out in this case.
1047    if (!Caller->use_empty())
1048      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1049        for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1050             UI != E; ++UI)
1051          if (PHINode *PN = dyn_cast<PHINode>(*UI))
1052            if (PN->getParent() == II->getNormalDest() ||
1053                PN->getParent() == II->getUnwindDest())
1054              return false;
1055  }
1056
1057  unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1058  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1059
1060  CallSite::arg_iterator AI = CS.arg_begin();
1061  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1062    Type *ParamTy = FT->getParamType(i);
1063    Type *ActTy = (*AI)->getType();
1064
1065    if (!CastInst::isCastable(ActTy, ParamTy))
1066      return false;   // Cannot transform this parameter value.
1067
1068    Attributes Attrs = CallerPAL.getParamAttributes(i + 1);
1069    if (Attrs & Attribute::typeIncompatible(ParamTy))
1070      return false;   // Attribute not compatible with transformed value.
1071
1072    // If the parameter is passed as a byval argument, then we have to have a
1073    // sized type and the sized type has to have the same size as the old type.
1074    if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
1075      PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1076      if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1077        return false;
1078
1079      Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1080      if (TD->getTypeAllocSize(CurElTy) !=
1081          TD->getTypeAllocSize(ParamPTy->getElementType()))
1082        return false;
1083    }
1084
1085    // Converting from one pointer type to another or between a pointer and an
1086    // integer of the same size is safe even if we do not have a body.
1087    bool isConvertible = ActTy == ParamTy ||
1088      (TD && ((ParamTy->isPointerTy() ||
1089      ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1090              (ActTy->isPointerTy() ||
1091              ActTy == TD->getIntPtrType(Caller->getContext()))));
1092    if (Callee->isDeclaration() && !isConvertible) return false;
1093  }
1094
1095  if (Callee->isDeclaration()) {
1096    // Do not delete arguments unless we have a function body.
1097    if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1098      return false;
1099
1100    // If the callee is just a declaration, don't change the varargsness of the
1101    // call.  We don't want to introduce a varargs call where one doesn't
1102    // already exist.
1103    PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1104    if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1105      return false;
1106
1107    // If both the callee and the cast type are varargs, we still have to make
1108    // sure the number of fixed parameters are the same or we have the same
1109    // ABI issues as if we introduce a varargs call.
1110    if (FT->isVarArg() &&
1111        cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1112        FT->getNumParams() !=
1113        cast<FunctionType>(APTy->getElementType())->getNumParams())
1114      return false;
1115  }
1116
1117  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1118      !CallerPAL.isEmpty())
1119    // In this case we have more arguments than the new function type, but we
1120    // won't be dropping them.  Check that these extra arguments have attributes
1121    // that are compatible with being a vararg call argument.
1122    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1123      if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1124        break;
1125      Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1126      if (PAttrs & Attribute::VarArgsIncompatible)
1127        return false;
1128    }
1129
1130
1131  // Okay, we decided that this is a safe thing to do: go ahead and start
1132  // inserting cast instructions as necessary.
1133  std::vector<Value*> Args;
1134  Args.reserve(NumActualArgs);
1135  SmallVector<AttributeWithIndex, 8> attrVec;
1136  attrVec.reserve(NumCommonArgs);
1137
1138  // Get any return attributes.
1139  Attributes RAttrs = CallerPAL.getRetAttributes();
1140
1141  // If the return value is not being used, the type may not be compatible
1142  // with the existing attributes.  Wipe out any problematic attributes.
1143  RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1144
1145  // Add the new return attributes.
1146  if (RAttrs)
1147    attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1148
1149  AI = CS.arg_begin();
1150  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1151    Type *ParamTy = FT->getParamType(i);
1152    if ((*AI)->getType() == ParamTy) {
1153      Args.push_back(*AI);
1154    } else {
1155      Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1156          false, ParamTy, false);
1157      Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy));
1158    }
1159
1160    // Add any parameter attributes.
1161    if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1162      attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1163  }
1164
1165  // If the function takes more arguments than the call was taking, add them
1166  // now.
1167  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1168    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1169
1170  // If we are removing arguments to the function, emit an obnoxious warning.
1171  if (FT->getNumParams() < NumActualArgs) {
1172    if (!FT->isVarArg()) {
1173      errs() << "WARNING: While resolving call to function '"
1174             << Callee->getName() << "' arguments were dropped!\n";
1175    } else {
1176      // Add all of the arguments in their promoted form to the arg list.
1177      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1178        Type *PTy = getPromotedType((*AI)->getType());
1179        if (PTy != (*AI)->getType()) {
1180          // Must promote to pass through va_arg area!
1181          Instruction::CastOps opcode =
1182            CastInst::getCastOpcode(*AI, false, PTy, false);
1183          Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1184        } else {
1185          Args.push_back(*AI);
1186        }
1187
1188        // Add any parameter attributes.
1189        if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1190          attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1191      }
1192    }
1193  }
1194
1195  if (Attributes FnAttrs =  CallerPAL.getFnAttributes())
1196    attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1197
1198  if (NewRetTy->isVoidTy())
1199    Caller->setName("");   // Void type should not have a name.
1200
1201  const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1202                                                     attrVec.end());
1203
1204  Instruction *NC;
1205  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1206    NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1207                               II->getUnwindDest(), Args);
1208    NC->takeName(II);
1209    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1210    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1211  } else {
1212    CallInst *CI = cast<CallInst>(Caller);
1213    NC = Builder->CreateCall(Callee, Args);
1214    NC->takeName(CI);
1215    if (CI->isTailCall())
1216      cast<CallInst>(NC)->setTailCall();
1217    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1218    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1219  }
1220
1221  // Insert a cast of the return type as necessary.
1222  Value *NV = NC;
1223  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1224    if (!NV->getType()->isVoidTy()) {
1225      Instruction::CastOps opcode =
1226        CastInst::getCastOpcode(NC, false, OldRetTy, false);
1227      NV = NC = CastInst::Create(opcode, NC, OldRetTy);
1228      NC->setDebugLoc(Caller->getDebugLoc());
1229
1230      // If this is an invoke instruction, we should insert it after the first
1231      // non-phi, instruction in the normal successor block.
1232      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1233        BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1234        InsertNewInstBefore(NC, *I);
1235      } else {
1236        // Otherwise, it's a call, just insert cast right after the call.
1237        InsertNewInstBefore(NC, *Caller);
1238      }
1239      Worklist.AddUsersToWorkList(*Caller);
1240    } else {
1241      NV = UndefValue::get(Caller->getType());
1242    }
1243  }
1244
1245  if (!Caller->use_empty())
1246    ReplaceInstUsesWith(*Caller, NV);
1247
1248  EraseInstFromFunction(*Caller);
1249  return true;
1250}
1251
1252// transformCallThroughTrampoline - Turn a call to a function created by
1253// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1254// underlying function.
1255//
1256Instruction *
1257InstCombiner::transformCallThroughTrampoline(CallSite CS,
1258                                             IntrinsicInst *Tramp) {
1259  Value *Callee = CS.getCalledValue();
1260  PointerType *PTy = cast<PointerType>(Callee->getType());
1261  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1262  const AttrListPtr &Attrs = CS.getAttributes();
1263
1264  // If the call already has the 'nest' attribute somewhere then give up -
1265  // otherwise 'nest' would occur twice after splicing in the chain.
1266  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1267    return 0;
1268
1269  assert(Tramp &&
1270         "transformCallThroughTrampoline called with incorrect CallSite.");
1271
1272  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1273  PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1274  FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1275
1276  const AttrListPtr &NestAttrs = NestF->getAttributes();
1277  if (!NestAttrs.isEmpty()) {
1278    unsigned NestIdx = 1;
1279    Type *NestTy = 0;
1280    Attributes NestAttr = Attribute::None;
1281
1282    // Look for a parameter marked with the 'nest' attribute.
1283    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1284         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1285      if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1286        // Record the parameter type and any other attributes.
1287        NestTy = *I;
1288        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1289        break;
1290      }
1291
1292    if (NestTy) {
1293      Instruction *Caller = CS.getInstruction();
1294      std::vector<Value*> NewArgs;
1295      NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1296
1297      SmallVector<AttributeWithIndex, 8> NewAttrs;
1298      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1299
1300      // Insert the nest argument into the call argument list, which may
1301      // mean appending it.  Likewise for attributes.
1302
1303      // Add any result attributes.
1304      if (Attributes Attr = Attrs.getRetAttributes())
1305        NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1306
1307      {
1308        unsigned Idx = 1;
1309        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1310        do {
1311          if (Idx == NestIdx) {
1312            // Add the chain argument and attributes.
1313            Value *NestVal = Tramp->getArgOperand(2);
1314            if (NestVal->getType() != NestTy)
1315              NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1316            NewArgs.push_back(NestVal);
1317            NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1318          }
1319
1320          if (I == E)
1321            break;
1322
1323          // Add the original argument and attributes.
1324          NewArgs.push_back(*I);
1325          if (Attributes Attr = Attrs.getParamAttributes(Idx))
1326            NewAttrs.push_back
1327              (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1328
1329          ++Idx, ++I;
1330        } while (1);
1331      }
1332
1333      // Add any function attributes.
1334      if (Attributes Attr = Attrs.getFnAttributes())
1335        NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1336
1337      // The trampoline may have been bitcast to a bogus type (FTy).
1338      // Handle this by synthesizing a new function type, equal to FTy
1339      // with the chain parameter inserted.
1340
1341      std::vector<Type*> NewTypes;
1342      NewTypes.reserve(FTy->getNumParams()+1);
1343
1344      // Insert the chain's type into the list of parameter types, which may
1345      // mean appending it.
1346      {
1347        unsigned Idx = 1;
1348        FunctionType::param_iterator I = FTy->param_begin(),
1349          E = FTy->param_end();
1350
1351        do {
1352          if (Idx == NestIdx)
1353            // Add the chain's type.
1354            NewTypes.push_back(NestTy);
1355
1356          if (I == E)
1357            break;
1358
1359          // Add the original type.
1360          NewTypes.push_back(*I);
1361
1362          ++Idx, ++I;
1363        } while (1);
1364      }
1365
1366      // Replace the trampoline call with a direct call.  Let the generic
1367      // code sort out any function type mismatches.
1368      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1369                                                FTy->isVarArg());
1370      Constant *NewCallee =
1371        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1372        NestF : ConstantExpr::getBitCast(NestF,
1373                                         PointerType::getUnqual(NewFTy));
1374      const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1375                                                   NewAttrs.end());
1376
1377      Instruction *NewCaller;
1378      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1379        NewCaller = InvokeInst::Create(NewCallee,
1380                                       II->getNormalDest(), II->getUnwindDest(),
1381                                       NewArgs);
1382        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1383        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1384      } else {
1385        NewCaller = CallInst::Create(NewCallee, NewArgs);
1386        if (cast<CallInst>(Caller)->isTailCall())
1387          cast<CallInst>(NewCaller)->setTailCall();
1388        cast<CallInst>(NewCaller)->
1389          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1390        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1391      }
1392
1393      return NewCaller;
1394    }
1395  }
1396
1397  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1398  // parameter, there is no need to adjust the argument list.  Let the generic
1399  // code sort out any function type mismatches.
1400  Constant *NewCallee =
1401    NestF->getType() == PTy ? NestF :
1402                              ConstantExpr::getBitCast(NestF, PTy);
1403  CS.setCalledFunction(NewCallee);
1404  return CS.getInstruction();
1405}
1406