InstCombineCalls.cpp revision e3305b17502c2a34152d4f50607b685eb2cadd21 
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/// computeAllocSize - compute the object size allocated by an allocation
169/// site. Returns 0 if the size is not constant (in SizeValue), 1 if the size
170/// is constant (in Size), and 2 if the size could not be determined within the
171/// given maximum Penalty that the computation would incurr at run-time.
172static int computeAllocSize(Value *Alloc, uint64_t &Size, Value* &SizeValue,
173                            uint64_t Penalty, TargetData *TD,
174                            InstCombiner::BuilderTy *Builder) {
175  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Alloc)) {
176    if (GV->hasUniqueInitializer()) {
177      Constant *C = GV->getInitializer();
178      Size = TD->getTypeAllocSize(C->getType());
179      return 1;
180    }
181    // Can't determine size of the GV.
182    return 2;
183
184  } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Alloc)) {
185    if (!AI->getAllocatedType()->isSized())
186      return 2;
187
188    Size = TD->getTypeAllocSize(AI->getAllocatedType());
189    if (!AI->isArrayAllocation())
190      return 1; // we are done
191
192    Value *ArraySize = AI->getArraySize();
193    if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
194      Size *= C->getZExtValue();
195      return 1;
196    }
197
198    if (Penalty < 2)
199      return 2;
200
201    SizeValue = Builder->CreateMul(Builder->getInt64(Size), ArraySize);
202    return 0;
203
204  } else if (CallInst *MI = extractMallocCall(Alloc)) {
205    SizeValue = MI->getArgOperand(0);
206    if (ConstantInt *CI = dyn_cast<ConstantInt>(SizeValue)) {
207      Size = CI->getZExtValue();
208      return 1;
209    }
210    return 0;
211
212  } else if (CallInst *MI = extractCallocCall(Alloc)) {
213    Value *Arg1 = MI->getArgOperand(0);
214    Value *Arg2 = MI->getArgOperand(1);
215    if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Arg1)) {
216      if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Arg2)) {
217        Size = (CI1->getValue() * CI2->getValue()).getZExtValue();
218        return 1;
219      }
220    }
221
222    if (Penalty < 2)
223      return 2;
224
225    SizeValue = Builder->CreateMul(Arg1, Arg2);
226    return 0;
227  }
228
229  DEBUG(errs() << "computeAllocSize failed:\n");
230  DEBUG(Alloc->dump());
231  return 2;
232}
233
234/// visitCallInst - CallInst simplification.  This mostly only handles folding
235/// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
236/// the heavy lifting.
237///
238Instruction *InstCombiner::visitCallInst(CallInst &CI) {
239  if (isFreeCall(&CI))
240    return visitFree(CI);
241  if (extractMallocCall(&CI) || extractCallocCall(&CI))
242    return visitMalloc(CI);
243
244  // If the caller function is nounwind, mark the call as nounwind, even if the
245  // callee isn't.
246  if (CI.getParent()->getParent()->doesNotThrow() &&
247      !CI.doesNotThrow()) {
248    CI.setDoesNotThrow();
249    return &CI;
250  }
251
252  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
253  if (!II) return visitCallSite(&CI);
254
255  // Intrinsics cannot occur in an invoke, so handle them here instead of in
256  // visitCallSite.
257  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
258    bool Changed = false;
259
260    // memmove/cpy/set of zero bytes is a noop.
261    if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
262      if (NumBytes->isNullValue())
263        return EraseInstFromFunction(CI);
264
265      if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
266        if (CI->getZExtValue() == 1) {
267          // Replace the instruction with just byte operations.  We would
268          // transform other cases to loads/stores, but we don't know if
269          // alignment is sufficient.
270        }
271    }
272
273    // No other transformations apply to volatile transfers.
274    if (MI->isVolatile())
275      return 0;
276
277    // If we have a memmove and the source operation is a constant global,
278    // then the source and dest pointers can't alias, so we can change this
279    // into a call to memcpy.
280    if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
281      if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
282        if (GVSrc->isConstant()) {
283          Module *M = CI.getParent()->getParent()->getParent();
284          Intrinsic::ID MemCpyID = Intrinsic::memcpy;
285          Type *Tys[3] = { CI.getArgOperand(0)->getType(),
286                           CI.getArgOperand(1)->getType(),
287                           CI.getArgOperand(2)->getType() };
288          CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
289          Changed = true;
290        }
291    }
292
293    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
294      // memmove(x,x,size) -> noop.
295      if (MTI->getSource() == MTI->getDest())
296        return EraseInstFromFunction(CI);
297    }
298
299    // If we can determine a pointer alignment that is bigger than currently
300    // set, update the alignment.
301    if (isa<MemTransferInst>(MI)) {
302      if (Instruction *I = SimplifyMemTransfer(MI))
303        return I;
304    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
305      if (Instruction *I = SimplifyMemSet(MSI))
306        return I;
307    }
308
309    if (Changed) return II;
310  }
311
312  switch (II->getIntrinsicID()) {
313  default: break;
314  case Intrinsic::objectsize: {
315    // We need target data for just about everything so depend on it.
316    if (!TD) return 0;
317
318    Type *ReturnTy = CI.getType();
319    uint64_t Penalty = cast<ConstantInt>(II->getArgOperand(2))->getZExtValue();
320
321    // Get to the real allocated thing and offset as fast as possible.
322    Value *Op1 = II->getArgOperand(0)->stripPointerCasts();
323
324    uint64_t Offset = 0;
325    Value *OffsetValue;
326    bool ConstOffset = true;
327
328    // Try to look through constant GEPs.
329    if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) {
330      if (!GEP->hasAllConstantIndices()) return 0;
331
332      // Get the current byte offset into the thing. Use the original
333      // operand in case we're looking through a bitcast.
334      SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end());
335      if (!GEP->getPointerOperandType()->isPointerTy())
336        return 0;
337      Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops);
338
339      Op1 = GEP->getPointerOperand()->stripPointerCasts();
340    }
341
342    uint64_t Size;
343    Value *SizeValue;
344    int ConstAlloc = computeAllocSize(Op1, Size, SizeValue, Penalty, TD,
345                                      Builder);
346
347    // Do not return "I don't know" here. Later optimization passes could
348    // make it possible to evaluate objectsize to a constant.
349    if (ConstAlloc == 2)
350      return 0;
351
352    if (ConstOffset && ConstAlloc) {
353      if (Size < Offset) {
354        // Out of bounds
355        return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, 0));
356      }
357      return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset));
358
359    } else if (Penalty >= 2) {
360      if (ConstOffset)
361        OffsetValue = Builder->getInt64(Offset);
362      if (ConstAlloc)
363        SizeValue = Builder->getInt64(Size);
364
365      Value *Val = Builder->CreateSub(SizeValue, OffsetValue);
366      Val = Builder->CreateTrunc(Val, ReturnTy);
367      // return 0 if there's an overflow
368      Value *Cmp = Builder->CreateICmpULT(SizeValue, OffsetValue);
369      Val = Builder->CreateSelect(Cmp, ConstantInt::get(ReturnTy, 0), Val);
370      return ReplaceInstUsesWith(CI, Val);
371
372    } else
373      return 0;
374  }
375  case Intrinsic::bswap:
376    // bswap(bswap(x)) -> x
377    if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0)))
378      if (Operand->getIntrinsicID() == Intrinsic::bswap)
379        return ReplaceInstUsesWith(CI, Operand->getArgOperand(0));
380
381    // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
382    if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) {
383      if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0)))
384        if (Operand->getIntrinsicID() == Intrinsic::bswap) {
385          unsigned C = Operand->getType()->getPrimitiveSizeInBits() -
386                       TI->getType()->getPrimitiveSizeInBits();
387          Value *CV = ConstantInt::get(Operand->getType(), C);
388          Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV);
389          return new TruncInst(V, TI->getType());
390        }
391    }
392
393    break;
394  case Intrinsic::powi:
395    if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
396      // powi(x, 0) -> 1.0
397      if (Power->isZero())
398        return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
399      // powi(x, 1) -> x
400      if (Power->isOne())
401        return ReplaceInstUsesWith(CI, II->getArgOperand(0));
402      // powi(x, -1) -> 1/x
403      if (Power->isAllOnesValue())
404        return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
405                                          II->getArgOperand(0));
406    }
407    break;
408  case Intrinsic::cttz: {
409    // If all bits below the first known one are known zero,
410    // this value is constant.
411    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
412    // FIXME: Try to simplify vectors of integers.
413    if (!IT) break;
414    uint32_t BitWidth = IT->getBitWidth();
415    APInt KnownZero(BitWidth, 0);
416    APInt KnownOne(BitWidth, 0);
417    ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
418    unsigned TrailingZeros = KnownOne.countTrailingZeros();
419    APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
420    if ((Mask & KnownZero) == Mask)
421      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
422                                 APInt(BitWidth, TrailingZeros)));
423
424    }
425    break;
426  case Intrinsic::ctlz: {
427    // If all bits above the first known one are known zero,
428    // this value is constant.
429    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
430    // FIXME: Try to simplify vectors of integers.
431    if (!IT) break;
432    uint32_t BitWidth = IT->getBitWidth();
433    APInt KnownZero(BitWidth, 0);
434    APInt KnownOne(BitWidth, 0);
435    ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne);
436    unsigned LeadingZeros = KnownOne.countLeadingZeros();
437    APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
438    if ((Mask & KnownZero) == Mask)
439      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
440                                 APInt(BitWidth, LeadingZeros)));
441
442    }
443    break;
444  case Intrinsic::uadd_with_overflow: {
445    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
446    IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
447    uint32_t BitWidth = IT->getBitWidth();
448    APInt LHSKnownZero(BitWidth, 0);
449    APInt LHSKnownOne(BitWidth, 0);
450    ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
451    bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
452    bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
453
454    if (LHSKnownNegative || LHSKnownPositive) {
455      APInt RHSKnownZero(BitWidth, 0);
456      APInt RHSKnownOne(BitWidth, 0);
457      ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
458      bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
459      bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
460      if (LHSKnownNegative && RHSKnownNegative) {
461        // The sign bit is set in both cases: this MUST overflow.
462        // Create a simple add instruction, and insert it into the struct.
463        Value *Add = Builder->CreateAdd(LHS, RHS);
464        Add->takeName(&CI);
465        Constant *V[] = {
466          UndefValue::get(LHS->getType()),
467          ConstantInt::getTrue(II->getContext())
468        };
469        StructType *ST = cast<StructType>(II->getType());
470        Constant *Struct = ConstantStruct::get(ST, V);
471        return InsertValueInst::Create(Struct, Add, 0);
472      }
473
474      if (LHSKnownPositive && RHSKnownPositive) {
475        // The sign bit is clear in both cases: this CANNOT overflow.
476        // Create a simple add instruction, and insert it into the struct.
477        Value *Add = Builder->CreateNUWAdd(LHS, RHS);
478        Add->takeName(&CI);
479        Constant *V[] = {
480          UndefValue::get(LHS->getType()),
481          ConstantInt::getFalse(II->getContext())
482        };
483        StructType *ST = cast<StructType>(II->getType());
484        Constant *Struct = ConstantStruct::get(ST, V);
485        return InsertValueInst::Create(Struct, Add, 0);
486      }
487    }
488  }
489  // FALL THROUGH uadd into sadd
490  case Intrinsic::sadd_with_overflow:
491    // Canonicalize constants into the RHS.
492    if (isa<Constant>(II->getArgOperand(0)) &&
493        !isa<Constant>(II->getArgOperand(1))) {
494      Value *LHS = II->getArgOperand(0);
495      II->setArgOperand(0, II->getArgOperand(1));
496      II->setArgOperand(1, LHS);
497      return II;
498    }
499
500    // X + undef -> undef
501    if (isa<UndefValue>(II->getArgOperand(1)))
502      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
503
504    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
505      // X + 0 -> {X, false}
506      if (RHS->isZero()) {
507        Constant *V[] = {
508          UndefValue::get(II->getArgOperand(0)->getType()),
509          ConstantInt::getFalse(II->getContext())
510        };
511        Constant *Struct =
512          ConstantStruct::get(cast<StructType>(II->getType()), V);
513        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
514      }
515    }
516    break;
517  case Intrinsic::usub_with_overflow:
518  case Intrinsic::ssub_with_overflow:
519    // undef - X -> undef
520    // X - undef -> undef
521    if (isa<UndefValue>(II->getArgOperand(0)) ||
522        isa<UndefValue>(II->getArgOperand(1)))
523      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
524
525    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
526      // X - 0 -> {X, false}
527      if (RHS->isZero()) {
528        Constant *V[] = {
529          UndefValue::get(II->getArgOperand(0)->getType()),
530          ConstantInt::getFalse(II->getContext())
531        };
532        Constant *Struct =
533          ConstantStruct::get(cast<StructType>(II->getType()), V);
534        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
535      }
536    }
537    break;
538  case Intrinsic::umul_with_overflow: {
539    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
540    unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
541
542    APInt LHSKnownZero(BitWidth, 0);
543    APInt LHSKnownOne(BitWidth, 0);
544    ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne);
545    APInt RHSKnownZero(BitWidth, 0);
546    APInt RHSKnownOne(BitWidth, 0);
547    ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne);
548
549    // Get the largest possible values for each operand.
550    APInt LHSMax = ~LHSKnownZero;
551    APInt RHSMax = ~RHSKnownZero;
552
553    // If multiplying the maximum values does not overflow then we can turn
554    // this into a plain NUW mul.
555    bool Overflow;
556    LHSMax.umul_ov(RHSMax, Overflow);
557    if (!Overflow) {
558      Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
559      Constant *V[] = {
560        UndefValue::get(LHS->getType()),
561        Builder->getFalse()
562      };
563      Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
564      return InsertValueInst::Create(Struct, Mul, 0);
565    }
566  } // FALL THROUGH
567  case Intrinsic::smul_with_overflow:
568    // Canonicalize constants into the RHS.
569    if (isa<Constant>(II->getArgOperand(0)) &&
570        !isa<Constant>(II->getArgOperand(1))) {
571      Value *LHS = II->getArgOperand(0);
572      II->setArgOperand(0, II->getArgOperand(1));
573      II->setArgOperand(1, LHS);
574      return II;
575    }
576
577    // X * undef -> undef
578    if (isa<UndefValue>(II->getArgOperand(1)))
579      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
580
581    if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
582      // X*0 -> {0, false}
583      if (RHSI->isZero())
584        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
585
586      // X * 1 -> {X, false}
587      if (RHSI->equalsInt(1)) {
588        Constant *V[] = {
589          UndefValue::get(II->getArgOperand(0)->getType()),
590          ConstantInt::getFalse(II->getContext())
591        };
592        Constant *Struct =
593          ConstantStruct::get(cast<StructType>(II->getType()), V);
594        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
595      }
596    }
597    break;
598  case Intrinsic::ppc_altivec_lvx:
599  case Intrinsic::ppc_altivec_lvxl:
600    // Turn PPC lvx -> load if the pointer is known aligned.
601    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
602      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
603                                         PointerType::getUnqual(II->getType()));
604      return new LoadInst(Ptr);
605    }
606    break;
607  case Intrinsic::ppc_altivec_stvx:
608  case Intrinsic::ppc_altivec_stvxl:
609    // Turn stvx -> store if the pointer is known aligned.
610    if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) {
611      Type *OpPtrTy =
612        PointerType::getUnqual(II->getArgOperand(0)->getType());
613      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
614      return new StoreInst(II->getArgOperand(0), Ptr);
615    }
616    break;
617  case Intrinsic::x86_sse_storeu_ps:
618  case Intrinsic::x86_sse2_storeu_pd:
619  case Intrinsic::x86_sse2_storeu_dq:
620    // Turn X86 storeu -> store if the pointer is known aligned.
621    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) {
622      Type *OpPtrTy =
623        PointerType::getUnqual(II->getArgOperand(1)->getType());
624      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
625      return new StoreInst(II->getArgOperand(1), Ptr);
626    }
627    break;
628
629  case Intrinsic::x86_sse_cvtss2si:
630  case Intrinsic::x86_sse_cvtss2si64:
631  case Intrinsic::x86_sse_cvttss2si:
632  case Intrinsic::x86_sse_cvttss2si64:
633  case Intrinsic::x86_sse2_cvtsd2si:
634  case Intrinsic::x86_sse2_cvtsd2si64:
635  case Intrinsic::x86_sse2_cvttsd2si:
636  case Intrinsic::x86_sse2_cvttsd2si64: {
637    // These intrinsics only demand the 0th element of their input vectors. If
638    // we can simplify the input based on that, do so now.
639    unsigned VWidth =
640      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
641    APInt DemandedElts(VWidth, 1);
642    APInt UndefElts(VWidth, 0);
643    if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
644                                              DemandedElts, UndefElts)) {
645      II->setArgOperand(0, V);
646      return II;
647    }
648    break;
649  }
650
651
652  case Intrinsic::x86_sse41_pmovsxbw:
653  case Intrinsic::x86_sse41_pmovsxwd:
654  case Intrinsic::x86_sse41_pmovsxdq:
655  case Intrinsic::x86_sse41_pmovzxbw:
656  case Intrinsic::x86_sse41_pmovzxwd:
657  case Intrinsic::x86_sse41_pmovzxdq: {
658    // pmov{s|z}x ignores the upper half of their input vectors.
659    unsigned VWidth =
660      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
661    unsigned LowHalfElts = VWidth / 2;
662    APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
663    APInt UndefElts(VWidth, 0);
664    if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
665                                                 InputDemandedElts,
666                                                 UndefElts)) {
667      II->setArgOperand(0, TmpV);
668      return II;
669    }
670    break;
671  }
672
673  case Intrinsic::ppc_altivec_vperm:
674    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
675    if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
676      assert(Mask->getType()->getVectorNumElements() == 16 &&
677             "Bad type for intrinsic!");
678
679      // Check that all of the elements are integer constants or undefs.
680      bool AllEltsOk = true;
681      for (unsigned i = 0; i != 16; ++i) {
682        Constant *Elt = Mask->getAggregateElement(i);
683        if (Elt == 0 ||
684            !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
685          AllEltsOk = false;
686          break;
687        }
688      }
689
690      if (AllEltsOk) {
691        // Cast the input vectors to byte vectors.
692        Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
693                                            Mask->getType());
694        Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
695                                            Mask->getType());
696        Value *Result = UndefValue::get(Op0->getType());
697
698        // Only extract each element once.
699        Value *ExtractedElts[32];
700        memset(ExtractedElts, 0, sizeof(ExtractedElts));
701
702        for (unsigned i = 0; i != 16; ++i) {
703          if (isa<UndefValue>(Mask->getAggregateElement(i)))
704            continue;
705          unsigned Idx =
706            cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
707          Idx &= 31;  // Match the hardware behavior.
708
709          if (ExtractedElts[Idx] == 0) {
710            ExtractedElts[Idx] =
711              Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
712                                            Builder->getInt32(Idx&15));
713          }
714
715          // Insert this value into the result vector.
716          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
717                                                Builder->getInt32(i));
718        }
719        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
720      }
721    }
722    break;
723
724  case Intrinsic::arm_neon_vld1:
725  case Intrinsic::arm_neon_vld2:
726  case Intrinsic::arm_neon_vld3:
727  case Intrinsic::arm_neon_vld4:
728  case Intrinsic::arm_neon_vld2lane:
729  case Intrinsic::arm_neon_vld3lane:
730  case Intrinsic::arm_neon_vld4lane:
731  case Intrinsic::arm_neon_vst1:
732  case Intrinsic::arm_neon_vst2:
733  case Intrinsic::arm_neon_vst3:
734  case Intrinsic::arm_neon_vst4:
735  case Intrinsic::arm_neon_vst2lane:
736  case Intrinsic::arm_neon_vst3lane:
737  case Intrinsic::arm_neon_vst4lane: {
738    unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD);
739    unsigned AlignArg = II->getNumArgOperands() - 1;
740    ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
741    if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
742      II->setArgOperand(AlignArg,
743                        ConstantInt::get(Type::getInt32Ty(II->getContext()),
744                                         MemAlign, false));
745      return II;
746    }
747    break;
748  }
749
750  case Intrinsic::arm_neon_vmulls:
751  case Intrinsic::arm_neon_vmullu: {
752    Value *Arg0 = II->getArgOperand(0);
753    Value *Arg1 = II->getArgOperand(1);
754
755    // Handle mul by zero first:
756    if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
757      return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
758    }
759
760    // Check for constant LHS & RHS - in this case we just simplify.
761    bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu);
762    VectorType *NewVT = cast<VectorType>(II->getType());
763    unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth();
764    if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) {
765      if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
766        VectorType* VT = cast<VectorType>(CV0->getType());
767        SmallVector<Constant*, 4> NewElems;
768        for (unsigned i = 0; i < VT->getNumElements(); ++i) {
769          APInt CV0E =
770            (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue();
771          CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth);
772          APInt CV1E =
773            (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue();
774          CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth);
775          NewElems.push_back(
776            ConstantInt::get(NewVT->getElementType(), CV0E * CV1E));
777        }
778        return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems));
779      }
780
781      // Couldn't simplify - cannonicalize constant to the RHS.
782      std::swap(Arg0, Arg1);
783    }
784
785    // Handle mul by one:
786    if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
787      if (ConstantInt *Splat =
788            dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) {
789        if (Splat->isOne()) {
790          if (Zext)
791            return CastInst::CreateZExtOrBitCast(Arg0, II->getType());
792          // else
793          return CastInst::CreateSExtOrBitCast(Arg0, II->getType());
794        }
795      }
796    }
797
798    break;
799  }
800
801  case Intrinsic::stackrestore: {
802    // If the save is right next to the restore, remove the restore.  This can
803    // happen when variable allocas are DCE'd.
804    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
805      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
806        BasicBlock::iterator BI = SS;
807        if (&*++BI == II)
808          return EraseInstFromFunction(CI);
809      }
810    }
811
812    // Scan down this block to see if there is another stack restore in the
813    // same block without an intervening call/alloca.
814    BasicBlock::iterator BI = II;
815    TerminatorInst *TI = II->getParent()->getTerminator();
816    bool CannotRemove = false;
817    for (++BI; &*BI != TI; ++BI) {
818      if (isa<AllocaInst>(BI) || isMalloc(BI)) {
819        CannotRemove = true;
820        break;
821      }
822      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
823        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
824          // If there is a stackrestore below this one, remove this one.
825          if (II->getIntrinsicID() == Intrinsic::stackrestore)
826            return EraseInstFromFunction(CI);
827          // Otherwise, ignore the intrinsic.
828        } else {
829          // If we found a non-intrinsic call, we can't remove the stack
830          // restore.
831          CannotRemove = true;
832          break;
833        }
834      }
835    }
836
837    // If the stack restore is in a return, resume, or unwind block and if there
838    // are no allocas or calls between the restore and the return, nuke the
839    // restore.
840    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
841      return EraseInstFromFunction(CI);
842    break;
843  }
844  }
845
846  return visitCallSite(II);
847}
848
849// InvokeInst simplification
850//
851Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
852  return visitCallSite(&II);
853}
854
855/// isSafeToEliminateVarargsCast - If this cast does not affect the value
856/// passed through the varargs area, we can eliminate the use of the cast.
857static bool isSafeToEliminateVarargsCast(const CallSite CS,
858                                         const CastInst * const CI,
859                                         const TargetData * const TD,
860                                         const int ix) {
861  if (!CI->isLosslessCast())
862    return false;
863
864  // The size of ByVal arguments is derived from the type, so we
865  // can't change to a type with a different size.  If the size were
866  // passed explicitly we could avoid this check.
867  if (!CS.isByValArgument(ix))
868    return true;
869
870  Type* SrcTy =
871            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
872  Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
873  if (!SrcTy->isSized() || !DstTy->isSized())
874    return false;
875  if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
876    return false;
877  return true;
878}
879
880namespace {
881class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls {
882  InstCombiner *IC;
883protected:
884  void replaceCall(Value *With) {
885    NewInstruction = IC->ReplaceInstUsesWith(*CI, With);
886  }
887  bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const {
888    if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp))
889      return true;
890    if (ConstantInt *SizeCI =
891                           dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) {
892      if (SizeCI->isAllOnesValue())
893        return true;
894      if (isString) {
895        uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp));
896        // If the length is 0 we don't know how long it is and so we can't
897        // remove the check.
898        if (Len == 0) return false;
899        return SizeCI->getZExtValue() >= Len;
900      }
901      if (ConstantInt *Arg = dyn_cast<ConstantInt>(
902                                                  CI->getArgOperand(SizeArgOp)))
903        return SizeCI->getZExtValue() >= Arg->getZExtValue();
904    }
905    return false;
906  }
907public:
908  InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { }
909  Instruction *NewInstruction;
910};
911} // end anonymous namespace
912
913// Try to fold some different type of calls here.
914// Currently we're only working with the checking functions, memcpy_chk,
915// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
916// strcat_chk and strncat_chk.
917Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) {
918  if (CI->getCalledFunction() == 0) return 0;
919
920  InstCombineFortifiedLibCalls Simplifier(this);
921  Simplifier.fold(CI, TD);
922  return Simplifier.NewInstruction;
923}
924
925static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
926  // Strip off at most one level of pointer casts, looking for an alloca.  This
927  // is good enough in practice and simpler than handling any number of casts.
928  Value *Underlying = TrampMem->stripPointerCasts();
929  if (Underlying != TrampMem &&
930      (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem))
931    return 0;
932  if (!isa<AllocaInst>(Underlying))
933    return 0;
934
935  IntrinsicInst *InitTrampoline = 0;
936  for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end();
937       I != E; I++) {
938    IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I);
939    if (!II)
940      return 0;
941    if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
942      if (InitTrampoline)
943        // More than one init_trampoline writes to this value.  Give up.
944        return 0;
945      InitTrampoline = II;
946      continue;
947    }
948    if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
949      // Allow any number of calls to adjust.trampoline.
950      continue;
951    return 0;
952  }
953
954  // No call to init.trampoline found.
955  if (!InitTrampoline)
956    return 0;
957
958  // Check that the alloca is being used in the expected way.
959  if (InitTrampoline->getOperand(0) != TrampMem)
960    return 0;
961
962  return InitTrampoline;
963}
964
965static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
966                                               Value *TrampMem) {
967  // Visit all the previous instructions in the basic block, and try to find a
968  // init.trampoline which has a direct path to the adjust.trampoline.
969  for (BasicBlock::iterator I = AdjustTramp,
970       E = AdjustTramp->getParent()->begin(); I != E; ) {
971    Instruction *Inst = --I;
972    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
973      if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
974          II->getOperand(0) == TrampMem)
975        return II;
976    if (Inst->mayWriteToMemory())
977      return 0;
978  }
979  return 0;
980}
981
982// Given a call to llvm.adjust.trampoline, find and return the corresponding
983// call to llvm.init.trampoline if the call to the trampoline can be optimized
984// to a direct call to a function.  Otherwise return NULL.
985//
986static IntrinsicInst *FindInitTrampoline(Value *Callee) {
987  Callee = Callee->stripPointerCasts();
988  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
989  if (!AdjustTramp ||
990      AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
991    return 0;
992
993  Value *TrampMem = AdjustTramp->getOperand(0);
994
995  if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
996    return IT;
997  if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
998    return IT;
999  return 0;
1000}
1001
1002// visitCallSite - Improvements for call and invoke instructions.
1003//
1004Instruction *InstCombiner::visitCallSite(CallSite CS) {
1005  bool Changed = false;
1006
1007  // If the callee is a pointer to a function, attempt to move any casts to the
1008  // arguments of the call/invoke.
1009  Value *Callee = CS.getCalledValue();
1010  if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1011    return 0;
1012
1013  if (Function *CalleeF = dyn_cast<Function>(Callee))
1014    // If the call and callee calling conventions don't match, this call must
1015    // be unreachable, as the call is undefined.
1016    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1017        // Only do this for calls to a function with a body.  A prototype may
1018        // not actually end up matching the implementation's calling conv for a
1019        // variety of reasons (e.g. it may be written in assembly).
1020        !CalleeF->isDeclaration()) {
1021      Instruction *OldCall = CS.getInstruction();
1022      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1023                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1024                                  OldCall);
1025      // If OldCall dues not return void then replaceAllUsesWith undef.
1026      // This allows ValueHandlers and custom metadata to adjust itself.
1027      if (!OldCall->getType()->isVoidTy())
1028        ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1029      if (isa<CallInst>(OldCall))
1030        return EraseInstFromFunction(*OldCall);
1031
1032      // We cannot remove an invoke, because it would change the CFG, just
1033      // change the callee to a null pointer.
1034      cast<InvokeInst>(OldCall)->setCalledFunction(
1035                                    Constant::getNullValue(CalleeF->getType()));
1036      return 0;
1037    }
1038
1039  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1040    // This instruction is not reachable, just remove it.  We insert a store to
1041    // undef so that we know that this code is not reachable, despite the fact
1042    // that we can't modify the CFG here.
1043    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1044               UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1045                  CS.getInstruction());
1046
1047    // If CS does not return void then replaceAllUsesWith undef.
1048    // This allows ValueHandlers and custom metadata to adjust itself.
1049    if (!CS.getInstruction()->getType()->isVoidTy())
1050      ReplaceInstUsesWith(*CS.getInstruction(),
1051                          UndefValue::get(CS.getInstruction()->getType()));
1052
1053    if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1054      // Don't break the CFG, insert a dummy cond branch.
1055      BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
1056                         ConstantInt::getTrue(Callee->getContext()), II);
1057    }
1058    return EraseInstFromFunction(*CS.getInstruction());
1059  }
1060
1061  if (IntrinsicInst *II = FindInitTrampoline(Callee))
1062    return transformCallThroughTrampoline(CS, II);
1063
1064  PointerType *PTy = cast<PointerType>(Callee->getType());
1065  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1066  if (FTy->isVarArg()) {
1067    int ix = FTy->getNumParams();
1068    // See if we can optimize any arguments passed through the varargs area of
1069    // the call.
1070    for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
1071           E = CS.arg_end(); I != E; ++I, ++ix) {
1072      CastInst *CI = dyn_cast<CastInst>(*I);
1073      if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
1074        *I = CI->getOperand(0);
1075        Changed = true;
1076      }
1077    }
1078  }
1079
1080  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1081    // Inline asm calls cannot throw - mark them 'nounwind'.
1082    CS.setDoesNotThrow();
1083    Changed = true;
1084  }
1085
1086  // Try to optimize the call if possible, we require TargetData for most of
1087  // this.  None of these calls are seen as possibly dead so go ahead and
1088  // delete the instruction now.
1089  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1090    Instruction *I = tryOptimizeCall(CI, TD);
1091    // If we changed something return the result, etc. Otherwise let
1092    // the fallthrough check.
1093    if (I) return EraseInstFromFunction(*I);
1094  }
1095
1096  return Changed ? CS.getInstruction() : 0;
1097}
1098
1099// transformConstExprCastCall - If the callee is a constexpr cast of a function,
1100// attempt to move the cast to the arguments of the call/invoke.
1101//
1102bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1103  Function *Callee =
1104    dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1105  if (Callee == 0)
1106    return false;
1107  Instruction *Caller = CS.getInstruction();
1108  const AttrListPtr &CallerPAL = CS.getAttributes();
1109
1110  // Okay, this is a cast from a function to a different type.  Unless doing so
1111  // would cause a type conversion of one of our arguments, change this call to
1112  // be a direct call with arguments casted to the appropriate types.
1113  //
1114  FunctionType *FT = Callee->getFunctionType();
1115  Type *OldRetTy = Caller->getType();
1116  Type *NewRetTy = FT->getReturnType();
1117
1118  if (NewRetTy->isStructTy())
1119    return false; // TODO: Handle multiple return values.
1120
1121  // Check to see if we are changing the return type...
1122  if (OldRetTy != NewRetTy) {
1123    if (Callee->isDeclaration() &&
1124        // Conversion is ok if changing from one pointer type to another or from
1125        // a pointer to an integer of the same size.
1126        !((OldRetTy->isPointerTy() || !TD ||
1127           OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
1128          (NewRetTy->isPointerTy() || !TD ||
1129           NewRetTy == TD->getIntPtrType(Caller->getContext()))))
1130      return false;   // Cannot transform this return value.
1131
1132    if (!Caller->use_empty() &&
1133        // void -> non-void is handled specially
1134        !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy))
1135      return false;   // Cannot transform this return value.
1136
1137    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1138      Attributes RAttrs = CallerPAL.getRetAttributes();
1139      if (RAttrs & Attribute::typeIncompatible(NewRetTy))
1140        return false;   // Attribute not compatible with transformed value.
1141    }
1142
1143    // If the callsite is an invoke instruction, and the return value is used by
1144    // a PHI node in a successor, we cannot change the return type of the call
1145    // because there is no place to put the cast instruction (without breaking
1146    // the critical edge).  Bail out in this case.
1147    if (!Caller->use_empty())
1148      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1149        for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
1150             UI != E; ++UI)
1151          if (PHINode *PN = dyn_cast<PHINode>(*UI))
1152            if (PN->getParent() == II->getNormalDest() ||
1153                PN->getParent() == II->getUnwindDest())
1154              return false;
1155  }
1156
1157  unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
1158  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1159
1160  CallSite::arg_iterator AI = CS.arg_begin();
1161  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1162    Type *ParamTy = FT->getParamType(i);
1163    Type *ActTy = (*AI)->getType();
1164
1165    if (!CastInst::isCastable(ActTy, ParamTy))
1166      return false;   // Cannot transform this parameter value.
1167
1168    Attributes Attrs = CallerPAL.getParamAttributes(i + 1);
1169    if (Attrs & Attribute::typeIncompatible(ParamTy))
1170      return false;   // Attribute not compatible with transformed value.
1171
1172    // If the parameter is passed as a byval argument, then we have to have a
1173    // sized type and the sized type has to have the same size as the old type.
1174    if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) {
1175      PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1176      if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0)
1177        return false;
1178
1179      Type *CurElTy = cast<PointerType>(ActTy)->getElementType();
1180      if (TD->getTypeAllocSize(CurElTy) !=
1181          TD->getTypeAllocSize(ParamPTy->getElementType()))
1182        return false;
1183    }
1184
1185    // Converting from one pointer type to another or between a pointer and an
1186    // integer of the same size is safe even if we do not have a body.
1187    bool isConvertible = ActTy == ParamTy ||
1188      (TD && ((ParamTy->isPointerTy() ||
1189      ParamTy == TD->getIntPtrType(Caller->getContext())) &&
1190              (ActTy->isPointerTy() ||
1191              ActTy == TD->getIntPtrType(Caller->getContext()))));
1192    if (Callee->isDeclaration() && !isConvertible) return false;
1193  }
1194
1195  if (Callee->isDeclaration()) {
1196    // Do not delete arguments unless we have a function body.
1197    if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1198      return false;
1199
1200    // If the callee is just a declaration, don't change the varargsness of the
1201    // call.  We don't want to introduce a varargs call where one doesn't
1202    // already exist.
1203    PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1204    if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1205      return false;
1206
1207    // If both the callee and the cast type are varargs, we still have to make
1208    // sure the number of fixed parameters are the same or we have the same
1209    // ABI issues as if we introduce a varargs call.
1210    if (FT->isVarArg() &&
1211        cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1212        FT->getNumParams() !=
1213        cast<FunctionType>(APTy->getElementType())->getNumParams())
1214      return false;
1215  }
1216
1217  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1218      !CallerPAL.isEmpty())
1219    // In this case we have more arguments than the new function type, but we
1220    // won't be dropping them.  Check that these extra arguments have attributes
1221    // that are compatible with being a vararg call argument.
1222    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1223      if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
1224        break;
1225      Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
1226      if (PAttrs & Attribute::VarArgsIncompatible)
1227        return false;
1228    }
1229
1230
1231  // Okay, we decided that this is a safe thing to do: go ahead and start
1232  // inserting cast instructions as necessary.
1233  std::vector<Value*> Args;
1234  Args.reserve(NumActualArgs);
1235  SmallVector<AttributeWithIndex, 8> attrVec;
1236  attrVec.reserve(NumCommonArgs);
1237
1238  // Get any return attributes.
1239  Attributes RAttrs = CallerPAL.getRetAttributes();
1240
1241  // If the return value is not being used, the type may not be compatible
1242  // with the existing attributes.  Wipe out any problematic attributes.
1243  RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
1244
1245  // Add the new return attributes.
1246  if (RAttrs)
1247    attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
1248
1249  AI = CS.arg_begin();
1250  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1251    Type *ParamTy = FT->getParamType(i);
1252    if ((*AI)->getType() == ParamTy) {
1253      Args.push_back(*AI);
1254    } else {
1255      Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
1256          false, ParamTy, false);
1257      Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy));
1258    }
1259
1260    // Add any parameter attributes.
1261    if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1262      attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1263  }
1264
1265  // If the function takes more arguments than the call was taking, add them
1266  // now.
1267  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1268    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1269
1270  // If we are removing arguments to the function, emit an obnoxious warning.
1271  if (FT->getNumParams() < NumActualArgs) {
1272    if (!FT->isVarArg()) {
1273      errs() << "WARNING: While resolving call to function '"
1274             << Callee->getName() << "' arguments were dropped!\n";
1275    } else {
1276      // Add all of the arguments in their promoted form to the arg list.
1277      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1278        Type *PTy = getPromotedType((*AI)->getType());
1279        if (PTy != (*AI)->getType()) {
1280          // Must promote to pass through va_arg area!
1281          Instruction::CastOps opcode =
1282            CastInst::getCastOpcode(*AI, false, PTy, false);
1283          Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1284        } else {
1285          Args.push_back(*AI);
1286        }
1287
1288        // Add any parameter attributes.
1289        if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
1290          attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
1291      }
1292    }
1293  }
1294
1295  if (Attributes FnAttrs =  CallerPAL.getFnAttributes())
1296    attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
1297
1298  if (NewRetTy->isVoidTy())
1299    Caller->setName("");   // Void type should not have a name.
1300
1301  const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
1302                                                     attrVec.end());
1303
1304  Instruction *NC;
1305  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1306    NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1307                               II->getUnwindDest(), Args);
1308    NC->takeName(II);
1309    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1310    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1311  } else {
1312    CallInst *CI = cast<CallInst>(Caller);
1313    NC = Builder->CreateCall(Callee, Args);
1314    NC->takeName(CI);
1315    if (CI->isTailCall())
1316      cast<CallInst>(NC)->setTailCall();
1317    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1318    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1319  }
1320
1321  // Insert a cast of the return type as necessary.
1322  Value *NV = NC;
1323  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1324    if (!NV->getType()->isVoidTy()) {
1325      Instruction::CastOps opcode =
1326        CastInst::getCastOpcode(NC, false, OldRetTy, false);
1327      NV = NC = CastInst::Create(opcode, NC, OldRetTy);
1328      NC->setDebugLoc(Caller->getDebugLoc());
1329
1330      // If this is an invoke instruction, we should insert it after the first
1331      // non-phi, instruction in the normal successor block.
1332      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1333        BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1334        InsertNewInstBefore(NC, *I);
1335      } else {
1336        // Otherwise, it's a call, just insert cast right after the call.
1337        InsertNewInstBefore(NC, *Caller);
1338      }
1339      Worklist.AddUsersToWorkList(*Caller);
1340    } else {
1341      NV = UndefValue::get(Caller->getType());
1342    }
1343  }
1344
1345  if (!Caller->use_empty())
1346    ReplaceInstUsesWith(*Caller, NV);
1347
1348  EraseInstFromFunction(*Caller);
1349  return true;
1350}
1351
1352// transformCallThroughTrampoline - Turn a call to a function created by
1353// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1354// underlying function.
1355//
1356Instruction *
1357InstCombiner::transformCallThroughTrampoline(CallSite CS,
1358                                             IntrinsicInst *Tramp) {
1359  Value *Callee = CS.getCalledValue();
1360  PointerType *PTy = cast<PointerType>(Callee->getType());
1361  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1362  const AttrListPtr &Attrs = CS.getAttributes();
1363
1364  // If the call already has the 'nest' attribute somewhere then give up -
1365  // otherwise 'nest' would occur twice after splicing in the chain.
1366  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1367    return 0;
1368
1369  assert(Tramp &&
1370         "transformCallThroughTrampoline called with incorrect CallSite.");
1371
1372  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1373  PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1374  FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1375
1376  const AttrListPtr &NestAttrs = NestF->getAttributes();
1377  if (!NestAttrs.isEmpty()) {
1378    unsigned NestIdx = 1;
1379    Type *NestTy = 0;
1380    Attributes NestAttr = Attribute::None;
1381
1382    // Look for a parameter marked with the 'nest' attribute.
1383    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1384         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1385      if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
1386        // Record the parameter type and any other attributes.
1387        NestTy = *I;
1388        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1389        break;
1390      }
1391
1392    if (NestTy) {
1393      Instruction *Caller = CS.getInstruction();
1394      std::vector<Value*> NewArgs;
1395      NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
1396
1397      SmallVector<AttributeWithIndex, 8> NewAttrs;
1398      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1399
1400      // Insert the nest argument into the call argument list, which may
1401      // mean appending it.  Likewise for attributes.
1402
1403      // Add any result attributes.
1404      if (Attributes Attr = Attrs.getRetAttributes())
1405        NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
1406
1407      {
1408        unsigned Idx = 1;
1409        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1410        do {
1411          if (Idx == NestIdx) {
1412            // Add the chain argument and attributes.
1413            Value *NestVal = Tramp->getArgOperand(2);
1414            if (NestVal->getType() != NestTy)
1415              NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1416            NewArgs.push_back(NestVal);
1417            NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
1418          }
1419
1420          if (I == E)
1421            break;
1422
1423          // Add the original argument and attributes.
1424          NewArgs.push_back(*I);
1425          if (Attributes Attr = Attrs.getParamAttributes(Idx))
1426            NewAttrs.push_back
1427              (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
1428
1429          ++Idx, ++I;
1430        } while (1);
1431      }
1432
1433      // Add any function attributes.
1434      if (Attributes Attr = Attrs.getFnAttributes())
1435        NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
1436
1437      // The trampoline may have been bitcast to a bogus type (FTy).
1438      // Handle this by synthesizing a new function type, equal to FTy
1439      // with the chain parameter inserted.
1440
1441      std::vector<Type*> NewTypes;
1442      NewTypes.reserve(FTy->getNumParams()+1);
1443
1444      // Insert the chain's type into the list of parameter types, which may
1445      // mean appending it.
1446      {
1447        unsigned Idx = 1;
1448        FunctionType::param_iterator I = FTy->param_begin(),
1449          E = FTy->param_end();
1450
1451        do {
1452          if (Idx == NestIdx)
1453            // Add the chain's type.
1454            NewTypes.push_back(NestTy);
1455
1456          if (I == E)
1457            break;
1458
1459          // Add the original type.
1460          NewTypes.push_back(*I);
1461
1462          ++Idx, ++I;
1463        } while (1);
1464      }
1465
1466      // Replace the trampoline call with a direct call.  Let the generic
1467      // code sort out any function type mismatches.
1468      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1469                                                FTy->isVarArg());
1470      Constant *NewCallee =
1471        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1472        NestF : ConstantExpr::getBitCast(NestF,
1473                                         PointerType::getUnqual(NewFTy));
1474      const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
1475                                                   NewAttrs.end());
1476
1477      Instruction *NewCaller;
1478      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1479        NewCaller = InvokeInst::Create(NewCallee,
1480                                       II->getNormalDest(), II->getUnwindDest(),
1481                                       NewArgs);
1482        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1483        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1484      } else {
1485        NewCaller = CallInst::Create(NewCallee, NewArgs);
1486        if (cast<CallInst>(Caller)->isTailCall())
1487          cast<CallInst>(NewCaller)->setTailCall();
1488        cast<CallInst>(NewCaller)->
1489          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1490        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1491      }
1492
1493      return NewCaller;
1494    }
1495  }
1496
1497  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1498  // parameter, there is no need to adjust the argument list.  Let the generic
1499  // code sort out any function type mismatches.
1500  Constant *NewCallee =
1501    NestF->getType() == PTy ? NestF :
1502                              ConstantExpr::getBitCast(NestF, PTy);
1503  CS.setCalledFunction(NewCallee);
1504  return CS.getInstruction();
1505}
1506