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/ADT/Statistic.h"
16#include "llvm/Analysis/MemoryBuiltins.h"
17#include "llvm/IR/CallSite.h"
18#include "llvm/IR/DataLayout.h"
19#include "llvm/IR/PatternMatch.h"
20#include "llvm/Transforms/Utils/BuildLibCalls.h"
21#include "llvm/Transforms/Utils/Local.h"
22using namespace llvm;
23using namespace PatternMatch;
24
25#define DEBUG_TYPE "instcombine"
26
27STATISTIC(NumSimplified, "Number of library calls simplified");
28
29/// getPromotedType - Return the specified type promoted as it would be to pass
30/// though a va_arg area.
31static Type *getPromotedType(Type *Ty) {
32  if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
33    if (ITy->getBitWidth() < 32)
34      return Type::getInt32Ty(Ty->getContext());
35  }
36  return Ty;
37}
38
39/// reduceToSingleValueType - Given an aggregate type which ultimately holds a
40/// single scalar element, like {{{type}}} or [1 x type], return type.
41static Type *reduceToSingleValueType(Type *T) {
42  while (!T->isSingleValueType()) {
43    if (StructType *STy = dyn_cast<StructType>(T)) {
44      if (STy->getNumElements() == 1)
45        T = STy->getElementType(0);
46      else
47        break;
48    } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) {
49      if (ATy->getNumElements() == 1)
50        T = ATy->getElementType();
51      else
52        break;
53    } else
54      break;
55  }
56
57  return T;
58}
59
60Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
61  unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL);
62  unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL);
63  unsigned MinAlign = std::min(DstAlign, SrcAlign);
64  unsigned CopyAlign = MI->getAlignment();
65
66  if (CopyAlign < MinAlign) {
67    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
68                                             MinAlign, false));
69    return MI;
70  }
71
72  // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
73  // load/store.
74  ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2));
75  if (!MemOpLength) return nullptr;
76
77  // Source and destination pointer types are always "i8*" for intrinsic.  See
78  // if the size is something we can handle with a single primitive load/store.
79  // A single load+store correctly handles overlapping memory in the memmove
80  // case.
81  uint64_t Size = MemOpLength->getLimitedValue();
82  assert(Size && "0-sized memory transferring should be removed already.");
83
84  if (Size > 8 || (Size&(Size-1)))
85    return nullptr;  // If not 1/2/4/8 bytes, exit.
86
87  // Use an integer load+store unless we can find something better.
88  unsigned SrcAddrSp =
89    cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
90  unsigned DstAddrSp =
91    cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace();
92
93  IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3);
94  Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp);
95  Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp);
96
97  // Memcpy forces the use of i8* for the source and destination.  That means
98  // that if you're using memcpy to move one double around, you'll get a cast
99  // from double* to i8*.  We'd much rather use a double load+store rather than
100  // an i64 load+store, here because this improves the odds that the source or
101  // dest address will be promotable.  See if we can find a better type than the
102  // integer datatype.
103  Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts();
104  MDNode *CopyMD = nullptr;
105  if (StrippedDest != MI->getArgOperand(0)) {
106    Type *SrcETy = cast<PointerType>(StrippedDest->getType())
107                                    ->getElementType();
108    if (DL && SrcETy->isSized() && DL->getTypeStoreSize(SrcETy) == Size) {
109      // The SrcETy might be something like {{{double}}} or [1 x double].  Rip
110      // down through these levels if so.
111      SrcETy = reduceToSingleValueType(SrcETy);
112
113      if (SrcETy->isSingleValueType()) {
114        NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp);
115        NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp);
116
117        // If the memcpy has metadata describing the members, see if we can
118        // get the TBAA tag describing our copy.
119        if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) {
120          if (M->getNumOperands() == 3 &&
121              M->getOperand(0) &&
122              isa<ConstantInt>(M->getOperand(0)) &&
123              cast<ConstantInt>(M->getOperand(0))->isNullValue() &&
124              M->getOperand(1) &&
125              isa<ConstantInt>(M->getOperand(1)) &&
126              cast<ConstantInt>(M->getOperand(1))->getValue() == Size &&
127              M->getOperand(2) &&
128              isa<MDNode>(M->getOperand(2)))
129            CopyMD = cast<MDNode>(M->getOperand(2));
130        }
131      }
132    }
133  }
134
135  // If the memcpy/memmove provides better alignment info than we can
136  // infer, use it.
137  SrcAlign = std::max(SrcAlign, CopyAlign);
138  DstAlign = std::max(DstAlign, CopyAlign);
139
140  Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy);
141  Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy);
142  LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile());
143  L->setAlignment(SrcAlign);
144  if (CopyMD)
145    L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
146  StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile());
147  S->setAlignment(DstAlign);
148  if (CopyMD)
149    S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
150
151  // Set the size of the copy to 0, it will be deleted on the next iteration.
152  MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType()));
153  return MI;
154}
155
156Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
157  unsigned Alignment = getKnownAlignment(MI->getDest(), DL);
158  if (MI->getAlignment() < Alignment) {
159    MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
160                                             Alignment, false));
161    return MI;
162  }
163
164  // Extract the length and alignment and fill if they are constant.
165  ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
166  ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
167  if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8))
168    return nullptr;
169  uint64_t Len = LenC->getLimitedValue();
170  Alignment = MI->getAlignment();
171  assert(Len && "0-sized memory setting should be removed already.");
172
173  // memset(s,c,n) -> store s, c (for n=1,2,4,8)
174  if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
175    Type *ITy = IntegerType::get(MI->getContext(), Len*8);  // n=1 -> i8.
176
177    Value *Dest = MI->getDest();
178    unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace();
179    Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
180    Dest = Builder->CreateBitCast(Dest, NewDstPtrTy);
181
182    // Alignment 0 is identity for alignment 1 for memset, but not store.
183    if (Alignment == 0) Alignment = 1;
184
185    // Extract the fill value and store.
186    uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
187    StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest,
188                                        MI->isVolatile());
189    S->setAlignment(Alignment);
190
191    // Set the size of the copy to 0, it will be deleted on the next iteration.
192    MI->setLength(Constant::getNullValue(LenC->getType()));
193    return MI;
194  }
195
196  return nullptr;
197}
198
199/// visitCallInst - CallInst simplification.  This mostly only handles folding
200/// of intrinsic instructions.  For normal calls, it allows visitCallSite to do
201/// the heavy lifting.
202///
203Instruction *InstCombiner::visitCallInst(CallInst &CI) {
204  if (isFreeCall(&CI, TLI))
205    return visitFree(CI);
206
207  // If the caller function is nounwind, mark the call as nounwind, even if the
208  // callee isn't.
209  if (CI.getParent()->getParent()->doesNotThrow() &&
210      !CI.doesNotThrow()) {
211    CI.setDoesNotThrow();
212    return &CI;
213  }
214
215  IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
216  if (!II) return visitCallSite(&CI);
217
218  // Intrinsics cannot occur in an invoke, so handle them here instead of in
219  // visitCallSite.
220  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
221    bool Changed = false;
222
223    // memmove/cpy/set of zero bytes is a noop.
224    if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
225      if (NumBytes->isNullValue())
226        return EraseInstFromFunction(CI);
227
228      if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
229        if (CI->getZExtValue() == 1) {
230          // Replace the instruction with just byte operations.  We would
231          // transform other cases to loads/stores, but we don't know if
232          // alignment is sufficient.
233        }
234    }
235
236    // No other transformations apply to volatile transfers.
237    if (MI->isVolatile())
238      return nullptr;
239
240    // If we have a memmove and the source operation is a constant global,
241    // then the source and dest pointers can't alias, so we can change this
242    // into a call to memcpy.
243    if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
244      if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
245        if (GVSrc->isConstant()) {
246          Module *M = CI.getParent()->getParent()->getParent();
247          Intrinsic::ID MemCpyID = Intrinsic::memcpy;
248          Type *Tys[3] = { CI.getArgOperand(0)->getType(),
249                           CI.getArgOperand(1)->getType(),
250                           CI.getArgOperand(2)->getType() };
251          CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys));
252          Changed = true;
253        }
254    }
255
256    if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
257      // memmove(x,x,size) -> noop.
258      if (MTI->getSource() == MTI->getDest())
259        return EraseInstFromFunction(CI);
260    }
261
262    // If we can determine a pointer alignment that is bigger than currently
263    // set, update the alignment.
264    if (isa<MemTransferInst>(MI)) {
265      if (Instruction *I = SimplifyMemTransfer(MI))
266        return I;
267    } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
268      if (Instruction *I = SimplifyMemSet(MSI))
269        return I;
270    }
271
272    if (Changed) return II;
273  }
274
275  switch (II->getIntrinsicID()) {
276  default: break;
277  case Intrinsic::objectsize: {
278    uint64_t Size;
279    if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
280      return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
281    return nullptr;
282  }
283  case Intrinsic::bswap: {
284    Value *IIOperand = II->getArgOperand(0);
285    Value *X = nullptr;
286
287    // bswap(bswap(x)) -> x
288    if (match(IIOperand, m_BSwap(m_Value(X))))
289        return ReplaceInstUsesWith(CI, X);
290
291    // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
292    if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
293      unsigned C = X->getType()->getPrimitiveSizeInBits() -
294        IIOperand->getType()->getPrimitiveSizeInBits();
295      Value *CV = ConstantInt::get(X->getType(), C);
296      Value *V = Builder->CreateLShr(X, CV);
297      return new TruncInst(V, IIOperand->getType());
298    }
299    break;
300  }
301
302  case Intrinsic::powi:
303    if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
304      // powi(x, 0) -> 1.0
305      if (Power->isZero())
306        return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0));
307      // powi(x, 1) -> x
308      if (Power->isOne())
309        return ReplaceInstUsesWith(CI, II->getArgOperand(0));
310      // powi(x, -1) -> 1/x
311      if (Power->isAllOnesValue())
312        return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0),
313                                          II->getArgOperand(0));
314    }
315    break;
316  case Intrinsic::cttz: {
317    // If all bits below the first known one are known zero,
318    // this value is constant.
319    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
320    // FIXME: Try to simplify vectors of integers.
321    if (!IT) break;
322    uint32_t BitWidth = IT->getBitWidth();
323    APInt KnownZero(BitWidth, 0);
324    APInt KnownOne(BitWidth, 0);
325    computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne);
326    unsigned TrailingZeros = KnownOne.countTrailingZeros();
327    APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros));
328    if ((Mask & KnownZero) == Mask)
329      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
330                                 APInt(BitWidth, TrailingZeros)));
331
332    }
333    break;
334  case Intrinsic::ctlz: {
335    // If all bits above the first known one are known zero,
336    // this value is constant.
337    IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType());
338    // FIXME: Try to simplify vectors of integers.
339    if (!IT) break;
340    uint32_t BitWidth = IT->getBitWidth();
341    APInt KnownZero(BitWidth, 0);
342    APInt KnownOne(BitWidth, 0);
343    computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne);
344    unsigned LeadingZeros = KnownOne.countLeadingZeros();
345    APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros));
346    if ((Mask & KnownZero) == Mask)
347      return ReplaceInstUsesWith(CI, ConstantInt::get(IT,
348                                 APInt(BitWidth, LeadingZeros)));
349
350    }
351    break;
352  case Intrinsic::uadd_with_overflow: {
353    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
354    IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType());
355    uint32_t BitWidth = IT->getBitWidth();
356    APInt LHSKnownZero(BitWidth, 0);
357    APInt LHSKnownOne(BitWidth, 0);
358    computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
359    bool LHSKnownNegative = LHSKnownOne[BitWidth - 1];
360    bool LHSKnownPositive = LHSKnownZero[BitWidth - 1];
361
362    if (LHSKnownNegative || LHSKnownPositive) {
363      APInt RHSKnownZero(BitWidth, 0);
364      APInt RHSKnownOne(BitWidth, 0);
365      computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
366      bool RHSKnownNegative = RHSKnownOne[BitWidth - 1];
367      bool RHSKnownPositive = RHSKnownZero[BitWidth - 1];
368      if (LHSKnownNegative && RHSKnownNegative) {
369        // The sign bit is set in both cases: this MUST overflow.
370        // Create a simple add instruction, and insert it into the struct.
371        Value *Add = Builder->CreateAdd(LHS, RHS);
372        Add->takeName(&CI);
373        Constant *V[] = {
374          UndefValue::get(LHS->getType()),
375          ConstantInt::getTrue(II->getContext())
376        };
377        StructType *ST = cast<StructType>(II->getType());
378        Constant *Struct = ConstantStruct::get(ST, V);
379        return InsertValueInst::Create(Struct, Add, 0);
380      }
381
382      if (LHSKnownPositive && RHSKnownPositive) {
383        // The sign bit is clear in both cases: this CANNOT overflow.
384        // Create a simple add instruction, and insert it into the struct.
385        Value *Add = Builder->CreateNUWAdd(LHS, RHS);
386        Add->takeName(&CI);
387        Constant *V[] = {
388          UndefValue::get(LHS->getType()),
389          ConstantInt::getFalse(II->getContext())
390        };
391        StructType *ST = cast<StructType>(II->getType());
392        Constant *Struct = ConstantStruct::get(ST, V);
393        return InsertValueInst::Create(Struct, Add, 0);
394      }
395    }
396  }
397  // FALL THROUGH uadd into sadd
398  case Intrinsic::sadd_with_overflow:
399    // Canonicalize constants into the RHS.
400    if (isa<Constant>(II->getArgOperand(0)) &&
401        !isa<Constant>(II->getArgOperand(1))) {
402      Value *LHS = II->getArgOperand(0);
403      II->setArgOperand(0, II->getArgOperand(1));
404      II->setArgOperand(1, LHS);
405      return II;
406    }
407
408    // X + undef -> undef
409    if (isa<UndefValue>(II->getArgOperand(1)))
410      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
411
412    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
413      // X + 0 -> {X, false}
414      if (RHS->isZero()) {
415        Constant *V[] = {
416          UndefValue::get(II->getArgOperand(0)->getType()),
417          ConstantInt::getFalse(II->getContext())
418        };
419        Constant *Struct =
420          ConstantStruct::get(cast<StructType>(II->getType()), V);
421        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
422      }
423    }
424
425    // We can strength reduce reduce this signed add into a regular add if we
426    // can prove that it will never overflow.
427    if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow) {
428      Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
429      if (WillNotOverflowSignedAdd(LHS, RHS)) {
430        Value *Add = Builder->CreateNSWAdd(LHS, RHS);
431        Add->takeName(&CI);
432        Constant *V[] = {UndefValue::get(Add->getType()), Builder->getFalse()};
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    break;
440  case Intrinsic::usub_with_overflow:
441  case Intrinsic::ssub_with_overflow:
442    // undef - X -> undef
443    // X - undef -> undef
444    if (isa<UndefValue>(II->getArgOperand(0)) ||
445        isa<UndefValue>(II->getArgOperand(1)))
446      return ReplaceInstUsesWith(CI, UndefValue::get(II->getType()));
447
448    if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
449      // X - 0 -> {X, false}
450      if (RHS->isZero()) {
451        Constant *V[] = {
452          UndefValue::get(II->getArgOperand(0)->getType()),
453          ConstantInt::getFalse(II->getContext())
454        };
455        Constant *Struct =
456          ConstantStruct::get(cast<StructType>(II->getType()), V);
457        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
458      }
459    }
460    break;
461  case Intrinsic::umul_with_overflow: {
462    Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
463    unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth();
464
465    APInt LHSKnownZero(BitWidth, 0);
466    APInt LHSKnownOne(BitWidth, 0);
467    computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
468    APInt RHSKnownZero(BitWidth, 0);
469    APInt RHSKnownOne(BitWidth, 0);
470    computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
471
472    // Get the largest possible values for each operand.
473    APInt LHSMax = ~LHSKnownZero;
474    APInt RHSMax = ~RHSKnownZero;
475
476    // If multiplying the maximum values does not overflow then we can turn
477    // this into a plain NUW mul.
478    bool Overflow;
479    LHSMax.umul_ov(RHSMax, Overflow);
480    if (!Overflow) {
481      Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow");
482      Constant *V[] = {
483        UndefValue::get(LHS->getType()),
484        Builder->getFalse()
485      };
486      Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V);
487      return InsertValueInst::Create(Struct, Mul, 0);
488    }
489  } // FALL THROUGH
490  case Intrinsic::smul_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 *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
505      // X*0 -> {0, false}
506      if (RHSI->isZero())
507        return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType()));
508
509      // X * 1 -> {X, false}
510      if (RHSI->equalsInt(1)) {
511        Constant *V[] = {
512          UndefValue::get(II->getArgOperand(0)->getType()),
513          ConstantInt::getFalse(II->getContext())
514        };
515        Constant *Struct =
516          ConstantStruct::get(cast<StructType>(II->getType()), V);
517        return InsertValueInst::Create(Struct, II->getArgOperand(0), 0);
518      }
519    }
520    break;
521  case Intrinsic::ppc_altivec_lvx:
522  case Intrinsic::ppc_altivec_lvxl:
523    // Turn PPC lvx -> load if the pointer is known aligned.
524    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
525      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0),
526                                         PointerType::getUnqual(II->getType()));
527      return new LoadInst(Ptr);
528    }
529    break;
530  case Intrinsic::ppc_altivec_stvx:
531  case Intrinsic::ppc_altivec_stvxl:
532    // Turn stvx -> store if the pointer is known aligned.
533    if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL) >= 16) {
534      Type *OpPtrTy =
535        PointerType::getUnqual(II->getArgOperand(0)->getType());
536      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy);
537      return new StoreInst(II->getArgOperand(0), Ptr);
538    }
539    break;
540  case Intrinsic::x86_sse_storeu_ps:
541  case Intrinsic::x86_sse2_storeu_pd:
542  case Intrinsic::x86_sse2_storeu_dq:
543    // Turn X86 storeu -> store if the pointer is known aligned.
544    if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL) >= 16) {
545      Type *OpPtrTy =
546        PointerType::getUnqual(II->getArgOperand(1)->getType());
547      Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy);
548      return new StoreInst(II->getArgOperand(1), Ptr);
549    }
550    break;
551
552  case Intrinsic::x86_sse_cvtss2si:
553  case Intrinsic::x86_sse_cvtss2si64:
554  case Intrinsic::x86_sse_cvttss2si:
555  case Intrinsic::x86_sse_cvttss2si64:
556  case Intrinsic::x86_sse2_cvtsd2si:
557  case Intrinsic::x86_sse2_cvtsd2si64:
558  case Intrinsic::x86_sse2_cvttsd2si:
559  case Intrinsic::x86_sse2_cvttsd2si64: {
560    // These intrinsics only demand the 0th element of their input vectors. If
561    // we can simplify the input based on that, do so now.
562    unsigned VWidth =
563      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
564    APInt DemandedElts(VWidth, 1);
565    APInt UndefElts(VWidth, 0);
566    if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0),
567                                              DemandedElts, UndefElts)) {
568      II->setArgOperand(0, V);
569      return II;
570    }
571    break;
572  }
573
574  // Constant fold <A x Bi> << Ci.
575  // FIXME: We don't handle _dq because it's a shift of an i128, but is
576  // represented in the IR as <2 x i64>. A per element shift is wrong.
577  case Intrinsic::x86_sse2_psll_d:
578  case Intrinsic::x86_sse2_psll_q:
579  case Intrinsic::x86_sse2_psll_w:
580  case Intrinsic::x86_sse2_pslli_d:
581  case Intrinsic::x86_sse2_pslli_q:
582  case Intrinsic::x86_sse2_pslli_w:
583  case Intrinsic::x86_avx2_psll_d:
584  case Intrinsic::x86_avx2_psll_q:
585  case Intrinsic::x86_avx2_psll_w:
586  case Intrinsic::x86_avx2_pslli_d:
587  case Intrinsic::x86_avx2_pslli_q:
588  case Intrinsic::x86_avx2_pslli_w:
589  case Intrinsic::x86_sse2_psrl_d:
590  case Intrinsic::x86_sse2_psrl_q:
591  case Intrinsic::x86_sse2_psrl_w:
592  case Intrinsic::x86_sse2_psrli_d:
593  case Intrinsic::x86_sse2_psrli_q:
594  case Intrinsic::x86_sse2_psrli_w:
595  case Intrinsic::x86_avx2_psrl_d:
596  case Intrinsic::x86_avx2_psrl_q:
597  case Intrinsic::x86_avx2_psrl_w:
598  case Intrinsic::x86_avx2_psrli_d:
599  case Intrinsic::x86_avx2_psrli_q:
600  case Intrinsic::x86_avx2_psrli_w: {
601    // Simplify if count is constant. To 0 if >= BitWidth,
602    // otherwise to shl/lshr.
603    auto CDV = dyn_cast<ConstantDataVector>(II->getArgOperand(1));
604    auto CInt = dyn_cast<ConstantInt>(II->getArgOperand(1));
605    if (!CDV && !CInt)
606      break;
607    ConstantInt *Count;
608    if (CDV)
609      Count = cast<ConstantInt>(CDV->getElementAsConstant(0));
610    else
611      Count = CInt;
612
613    auto Vec = II->getArgOperand(0);
614    auto VT = cast<VectorType>(Vec->getType());
615    if (Count->getZExtValue() >
616        VT->getElementType()->getPrimitiveSizeInBits() - 1)
617      return ReplaceInstUsesWith(
618          CI, ConstantAggregateZero::get(Vec->getType()));
619
620    bool isPackedShiftLeft = true;
621    switch (II->getIntrinsicID()) {
622    default : break;
623    case Intrinsic::x86_sse2_psrl_d:
624    case Intrinsic::x86_sse2_psrl_q:
625    case Intrinsic::x86_sse2_psrl_w:
626    case Intrinsic::x86_sse2_psrli_d:
627    case Intrinsic::x86_sse2_psrli_q:
628    case Intrinsic::x86_sse2_psrli_w:
629    case Intrinsic::x86_avx2_psrl_d:
630    case Intrinsic::x86_avx2_psrl_q:
631    case Intrinsic::x86_avx2_psrl_w:
632    case Intrinsic::x86_avx2_psrli_d:
633    case Intrinsic::x86_avx2_psrli_q:
634    case Intrinsic::x86_avx2_psrli_w: isPackedShiftLeft = false; break;
635    }
636
637    unsigned VWidth = VT->getNumElements();
638    // Get a constant vector of the same type as the first operand.
639    auto VTCI = ConstantInt::get(VT->getElementType(), Count->getZExtValue());
640    if (isPackedShiftLeft)
641      return BinaryOperator::CreateShl(Vec,
642          Builder->CreateVectorSplat(VWidth, VTCI));
643
644    return BinaryOperator::CreateLShr(Vec,
645        Builder->CreateVectorSplat(VWidth, VTCI));
646  }
647
648  case Intrinsic::x86_sse41_pmovsxbw:
649  case Intrinsic::x86_sse41_pmovsxwd:
650  case Intrinsic::x86_sse41_pmovsxdq:
651  case Intrinsic::x86_sse41_pmovzxbw:
652  case Intrinsic::x86_sse41_pmovzxwd:
653  case Intrinsic::x86_sse41_pmovzxdq: {
654    // pmov{s|z}x ignores the upper half of their input vectors.
655    unsigned VWidth =
656      cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements();
657    unsigned LowHalfElts = VWidth / 2;
658    APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts));
659    APInt UndefElts(VWidth, 0);
660    if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0),
661                                                 InputDemandedElts,
662                                                 UndefElts)) {
663      II->setArgOperand(0, TmpV);
664      return II;
665    }
666    break;
667  }
668
669  case Intrinsic::x86_sse4a_insertqi: {
670    // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
671    // ones undef
672    // TODO: eventually we should lower this intrinsic to IR
673    if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
674      if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
675        if (CIWidth->equalsInt(64) && CIStart->isZero()) {
676          Value *Vec = II->getArgOperand(1);
677          Value *Undef = UndefValue::get(Vec->getType());
678          const uint32_t Mask[] = { 0, 2 };
679          return ReplaceInstUsesWith(
680              CI,
681              Builder->CreateShuffleVector(
682                  Vec, Undef, ConstantDataVector::get(
683                                  II->getContext(), ArrayRef<uint32_t>(Mask))));
684
685        } else if (auto Source =
686                       dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
687          if (Source->hasOneUse() &&
688              Source->getArgOperand(1) == II->getArgOperand(1)) {
689            // If the source of the insert has only one use and it's another
690            // insert (and they're both inserting from the same vector), try to
691            // bundle both together.
692            auto CISourceWidth =
693                dyn_cast<ConstantInt>(Source->getArgOperand(2));
694            auto CISourceStart =
695                dyn_cast<ConstantInt>(Source->getArgOperand(3));
696            if (CISourceStart && CISourceWidth) {
697              unsigned Start = CIStart->getZExtValue();
698              unsigned Width = CIWidth->getZExtValue();
699              unsigned End = Start + Width;
700              unsigned SourceStart = CISourceStart->getZExtValue();
701              unsigned SourceWidth = CISourceWidth->getZExtValue();
702              unsigned SourceEnd = SourceStart + SourceWidth;
703              unsigned NewStart, NewWidth;
704              bool ShouldReplace = false;
705              if (Start <= SourceStart && SourceStart <= End) {
706                NewStart = Start;
707                NewWidth = std::max(End, SourceEnd) - NewStart;
708                ShouldReplace = true;
709              } else if (SourceStart <= Start && Start <= SourceEnd) {
710                NewStart = SourceStart;
711                NewWidth = std::max(SourceEnd, End) - NewStart;
712                ShouldReplace = true;
713              }
714
715              if (ShouldReplace) {
716                Constant *ConstantWidth = ConstantInt::get(
717                    II->getArgOperand(2)->getType(), NewWidth, false);
718                Constant *ConstantStart = ConstantInt::get(
719                    II->getArgOperand(3)->getType(), NewStart, false);
720                Value *Args[4] = { Source->getArgOperand(0),
721                                   II->getArgOperand(1), ConstantWidth,
722                                   ConstantStart };
723                Module *M = CI.getParent()->getParent()->getParent();
724                Value *F =
725                    Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
726                return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
727              }
728            }
729          }
730        }
731      }
732    }
733    break;
734  }
735
736  case Intrinsic::x86_sse41_pblendvb:
737  case Intrinsic::x86_sse41_blendvps:
738  case Intrinsic::x86_sse41_blendvpd:
739  case Intrinsic::x86_avx_blendv_ps_256:
740  case Intrinsic::x86_avx_blendv_pd_256:
741  case Intrinsic::x86_avx2_pblendvb: {
742    // Convert blendv* to vector selects if the mask is constant.
743    // This optimization is convoluted because the intrinsic is defined as
744    // getting a vector of floats or doubles for the ps and pd versions.
745    // FIXME: That should be changed.
746    Value *Mask = II->getArgOperand(2);
747    if (auto C = dyn_cast<ConstantDataVector>(Mask)) {
748      auto Tyi1 = Builder->getInt1Ty();
749      auto SelectorType = cast<VectorType>(Mask->getType());
750      auto EltTy = SelectorType->getElementType();
751      unsigned Size = SelectorType->getNumElements();
752      unsigned BitWidth =
753          EltTy->isFloatTy()
754              ? 32
755              : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth());
756      assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) &&
757             "Wrong arguments for variable blend intrinsic");
758      SmallVector<Constant *, 32> Selectors;
759      for (unsigned I = 0; I < Size; ++I) {
760        // The intrinsics only read the top bit
761        uint64_t Selector;
762        if (BitWidth == 8)
763          Selector = C->getElementAsInteger(I);
764        else
765          Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue();
766        Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1)));
767      }
768      auto NewSelector = ConstantVector::get(Selectors);
769      return SelectInst::Create(NewSelector, II->getArgOperand(1),
770                                II->getArgOperand(0), "blendv");
771    } else {
772      break;
773    }
774  }
775
776  case Intrinsic::x86_avx_vpermilvar_ps:
777  case Intrinsic::x86_avx_vpermilvar_ps_256:
778  case Intrinsic::x86_avx_vpermilvar_pd:
779  case Intrinsic::x86_avx_vpermilvar_pd_256: {
780    // Convert vpermil* to shufflevector if the mask is constant.
781    Value *V = II->getArgOperand(1);
782    unsigned Size = cast<VectorType>(V->getType())->getNumElements();
783    assert(Size == 8 || Size == 4 || Size == 2);
784    uint32_t Indexes[8];
785    if (auto C = dyn_cast<ConstantDataVector>(V)) {
786      // The intrinsics only read one or two bits, clear the rest.
787      for (unsigned I = 0; I < Size; ++I) {
788        uint32_t Index = C->getElementAsInteger(I) & 0x3;
789        if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
790            II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
791          Index >>= 1;
792        Indexes[I] = Index;
793      }
794    } else if (isa<ConstantAggregateZero>(V)) {
795      for (unsigned I = 0; I < Size; ++I)
796        Indexes[I] = 0;
797    } else {
798      break;
799    }
800    // The _256 variants are a bit trickier since the mask bits always index
801    // into the corresponding 128 half. In order to convert to a generic
802    // shuffle, we have to make that explicit.
803    if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
804        II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
805      for (unsigned I = Size / 2; I < Size; ++I)
806        Indexes[I] += Size / 2;
807    }
808    auto NewC =
809        ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
810    auto V1 = II->getArgOperand(0);
811    auto V2 = UndefValue::get(V1->getType());
812    auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
813    return ReplaceInstUsesWith(CI, Shuffle);
814  }
815
816  case Intrinsic::ppc_altivec_vperm:
817    // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
818    // Note that ppc_altivec_vperm has a big-endian bias, so when creating
819    // a vectorshuffle for little endian, we must undo the transformation
820    // performed on vec_perm in altivec.h.  That is, we must complement
821    // the permutation mask with respect to 31 and reverse the order of
822    // V1 and V2.
823    if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) {
824      assert(Mask->getType()->getVectorNumElements() == 16 &&
825             "Bad type for intrinsic!");
826
827      // Check that all of the elements are integer constants or undefs.
828      bool AllEltsOk = true;
829      for (unsigned i = 0; i != 16; ++i) {
830        Constant *Elt = Mask->getAggregateElement(i);
831        if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) {
832          AllEltsOk = false;
833          break;
834        }
835      }
836
837      if (AllEltsOk) {
838        // Cast the input vectors to byte vectors.
839        Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0),
840                                            Mask->getType());
841        Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1),
842                                            Mask->getType());
843        Value *Result = UndefValue::get(Op0->getType());
844
845        // Only extract each element once.
846        Value *ExtractedElts[32];
847        memset(ExtractedElts, 0, sizeof(ExtractedElts));
848
849        for (unsigned i = 0; i != 16; ++i) {
850          if (isa<UndefValue>(Mask->getAggregateElement(i)))
851            continue;
852          unsigned Idx =
853            cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue();
854          Idx &= 31;  // Match the hardware behavior.
855          if (DL && DL->isLittleEndian())
856            Idx = 31 - Idx;
857
858          if (!ExtractedElts[Idx]) {
859            Value *Op0ToUse = (DL && DL->isLittleEndian()) ? Op1 : Op0;
860            Value *Op1ToUse = (DL && DL->isLittleEndian()) ? Op0 : Op1;
861            ExtractedElts[Idx] =
862              Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse,
863                                            Builder->getInt32(Idx&15));
864          }
865
866          // Insert this value into the result vector.
867          Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
868                                                Builder->getInt32(i));
869        }
870        return CastInst::Create(Instruction::BitCast, Result, CI.getType());
871      }
872    }
873    break;
874
875  case Intrinsic::arm_neon_vld1:
876  case Intrinsic::arm_neon_vld2:
877  case Intrinsic::arm_neon_vld3:
878  case Intrinsic::arm_neon_vld4:
879  case Intrinsic::arm_neon_vld2lane:
880  case Intrinsic::arm_neon_vld3lane:
881  case Intrinsic::arm_neon_vld4lane:
882  case Intrinsic::arm_neon_vst1:
883  case Intrinsic::arm_neon_vst2:
884  case Intrinsic::arm_neon_vst3:
885  case Intrinsic::arm_neon_vst4:
886  case Intrinsic::arm_neon_vst2lane:
887  case Intrinsic::arm_neon_vst3lane:
888  case Intrinsic::arm_neon_vst4lane: {
889    unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL);
890    unsigned AlignArg = II->getNumArgOperands() - 1;
891    ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg));
892    if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) {
893      II->setArgOperand(AlignArg,
894                        ConstantInt::get(Type::getInt32Ty(II->getContext()),
895                                         MemAlign, false));
896      return II;
897    }
898    break;
899  }
900
901  case Intrinsic::arm_neon_vmulls:
902  case Intrinsic::arm_neon_vmullu:
903  case Intrinsic::aarch64_neon_smull:
904  case Intrinsic::aarch64_neon_umull: {
905    Value *Arg0 = II->getArgOperand(0);
906    Value *Arg1 = II->getArgOperand(1);
907
908    // Handle mul by zero first:
909    if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
910      return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
911    }
912
913    // Check for constant LHS & RHS - in this case we just simplify.
914    bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu ||
915                 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull);
916    VectorType *NewVT = cast<VectorType>(II->getType());
917    if (Constant *CV0 = dyn_cast<Constant>(Arg0)) {
918      if (Constant *CV1 = dyn_cast<Constant>(Arg1)) {
919        CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext);
920        CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext);
921
922        return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1));
923      }
924
925      // Couldn't simplify - canonicalize constant to the RHS.
926      std::swap(Arg0, Arg1);
927    }
928
929    // Handle mul by one:
930    if (Constant *CV1 = dyn_cast<Constant>(Arg1))
931      if (ConstantInt *Splat =
932              dyn_cast_or_null<ConstantInt>(CV1->getSplatValue()))
933        if (Splat->isOne())
934          return CastInst::CreateIntegerCast(Arg0, II->getType(),
935                                             /*isSigned=*/!Zext);
936
937    break;
938  }
939
940  case Intrinsic::AMDGPU_rcp: {
941    if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
942      const APFloat &ArgVal = C->getValueAPF();
943      APFloat Val(ArgVal.getSemantics(), 1.0);
944      APFloat::opStatus Status = Val.divide(ArgVal,
945                                            APFloat::rmNearestTiesToEven);
946      // Only do this if it was exact and therefore not dependent on the
947      // rounding mode.
948      if (Status == APFloat::opOK)
949        return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
950    }
951
952    break;
953  }
954  case Intrinsic::stackrestore: {
955    // If the save is right next to the restore, remove the restore.  This can
956    // happen when variable allocas are DCE'd.
957    if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
958      if (SS->getIntrinsicID() == Intrinsic::stacksave) {
959        BasicBlock::iterator BI = SS;
960        if (&*++BI == II)
961          return EraseInstFromFunction(CI);
962      }
963    }
964
965    // Scan down this block to see if there is another stack restore in the
966    // same block without an intervening call/alloca.
967    BasicBlock::iterator BI = II;
968    TerminatorInst *TI = II->getParent()->getTerminator();
969    bool CannotRemove = false;
970    for (++BI; &*BI != TI; ++BI) {
971      if (isa<AllocaInst>(BI)) {
972        CannotRemove = true;
973        break;
974      }
975      if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
976        if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
977          // If there is a stackrestore below this one, remove this one.
978          if (II->getIntrinsicID() == Intrinsic::stackrestore)
979            return EraseInstFromFunction(CI);
980          // Otherwise, ignore the intrinsic.
981        } else {
982          // If we found a non-intrinsic call, we can't remove the stack
983          // restore.
984          CannotRemove = true;
985          break;
986        }
987      }
988    }
989
990    // If the stack restore is in a return, resume, or unwind block and if there
991    // are no allocas or calls between the restore and the return, nuke the
992    // restore.
993    if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI)))
994      return EraseInstFromFunction(CI);
995    break;
996  }
997  }
998
999  return visitCallSite(II);
1000}
1001
1002// InvokeInst simplification
1003//
1004Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
1005  return visitCallSite(&II);
1006}
1007
1008/// isSafeToEliminateVarargsCast - If this cast does not affect the value
1009/// passed through the varargs area, we can eliminate the use of the cast.
1010static bool isSafeToEliminateVarargsCast(const CallSite CS,
1011                                         const CastInst * const CI,
1012                                         const DataLayout * const DL,
1013                                         const int ix) {
1014  if (!CI->isLosslessCast())
1015    return false;
1016
1017  // The size of ByVal or InAlloca arguments is derived from the type, so we
1018  // can't change to a type with a different size.  If the size were
1019  // passed explicitly we could avoid this check.
1020  if (!CS.isByValOrInAllocaArgument(ix))
1021    return true;
1022
1023  Type* SrcTy =
1024            cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
1025  Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
1026  if (!SrcTy->isSized() || !DstTy->isSized())
1027    return false;
1028  if (!DL || DL->getTypeAllocSize(SrcTy) != DL->getTypeAllocSize(DstTy))
1029    return false;
1030  return true;
1031}
1032
1033// Try to fold some different type of calls here.
1034// Currently we're only working with the checking functions, memcpy_chk,
1035// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk,
1036// strcat_chk and strncat_chk.
1037Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const DataLayout *DL) {
1038  if (!CI->getCalledFunction()) return nullptr;
1039
1040  if (Value *With = Simplifier->optimizeCall(CI)) {
1041    ++NumSimplified;
1042    return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With);
1043  }
1044
1045  return nullptr;
1046}
1047
1048static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) {
1049  // Strip off at most one level of pointer casts, looking for an alloca.  This
1050  // is good enough in practice and simpler than handling any number of casts.
1051  Value *Underlying = TrampMem->stripPointerCasts();
1052  if (Underlying != TrampMem &&
1053      (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem))
1054    return nullptr;
1055  if (!isa<AllocaInst>(Underlying))
1056    return nullptr;
1057
1058  IntrinsicInst *InitTrampoline = nullptr;
1059  for (User *U : TrampMem->users()) {
1060    IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1061    if (!II)
1062      return nullptr;
1063    if (II->getIntrinsicID() == Intrinsic::init_trampoline) {
1064      if (InitTrampoline)
1065        // More than one init_trampoline writes to this value.  Give up.
1066        return nullptr;
1067      InitTrampoline = II;
1068      continue;
1069    }
1070    if (II->getIntrinsicID() == Intrinsic::adjust_trampoline)
1071      // Allow any number of calls to adjust.trampoline.
1072      continue;
1073    return nullptr;
1074  }
1075
1076  // No call to init.trampoline found.
1077  if (!InitTrampoline)
1078    return nullptr;
1079
1080  // Check that the alloca is being used in the expected way.
1081  if (InitTrampoline->getOperand(0) != TrampMem)
1082    return nullptr;
1083
1084  return InitTrampoline;
1085}
1086
1087static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp,
1088                                               Value *TrampMem) {
1089  // Visit all the previous instructions in the basic block, and try to find a
1090  // init.trampoline which has a direct path to the adjust.trampoline.
1091  for (BasicBlock::iterator I = AdjustTramp,
1092       E = AdjustTramp->getParent()->begin(); I != E; ) {
1093    Instruction *Inst = --I;
1094    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1095      if (II->getIntrinsicID() == Intrinsic::init_trampoline &&
1096          II->getOperand(0) == TrampMem)
1097        return II;
1098    if (Inst->mayWriteToMemory())
1099      return nullptr;
1100  }
1101  return nullptr;
1102}
1103
1104// Given a call to llvm.adjust.trampoline, find and return the corresponding
1105// call to llvm.init.trampoline if the call to the trampoline can be optimized
1106// to a direct call to a function.  Otherwise return NULL.
1107//
1108static IntrinsicInst *FindInitTrampoline(Value *Callee) {
1109  Callee = Callee->stripPointerCasts();
1110  IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee);
1111  if (!AdjustTramp ||
1112      AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline)
1113    return nullptr;
1114
1115  Value *TrampMem = AdjustTramp->getOperand(0);
1116
1117  if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem))
1118    return IT;
1119  if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem))
1120    return IT;
1121  return nullptr;
1122}
1123
1124// visitCallSite - Improvements for call and invoke instructions.
1125//
1126Instruction *InstCombiner::visitCallSite(CallSite CS) {
1127  if (isAllocLikeFn(CS.getInstruction(), TLI))
1128    return visitAllocSite(*CS.getInstruction());
1129
1130  bool Changed = false;
1131
1132  // If the callee is a pointer to a function, attempt to move any casts to the
1133  // arguments of the call/invoke.
1134  Value *Callee = CS.getCalledValue();
1135  if (!isa<Function>(Callee) && transformConstExprCastCall(CS))
1136    return nullptr;
1137
1138  if (Function *CalleeF = dyn_cast<Function>(Callee))
1139    // If the call and callee calling conventions don't match, this call must
1140    // be unreachable, as the call is undefined.
1141    if (CalleeF->getCallingConv() != CS.getCallingConv() &&
1142        // Only do this for calls to a function with a body.  A prototype may
1143        // not actually end up matching the implementation's calling conv for a
1144        // variety of reasons (e.g. it may be written in assembly).
1145        !CalleeF->isDeclaration()) {
1146      Instruction *OldCall = CS.getInstruction();
1147      new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1148                UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1149                                  OldCall);
1150      // If OldCall does not return void then replaceAllUsesWith undef.
1151      // This allows ValueHandlers and custom metadata to adjust itself.
1152      if (!OldCall->getType()->isVoidTy())
1153        ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType()));
1154      if (isa<CallInst>(OldCall))
1155        return EraseInstFromFunction(*OldCall);
1156
1157      // We cannot remove an invoke, because it would change the CFG, just
1158      // change the callee to a null pointer.
1159      cast<InvokeInst>(OldCall)->setCalledFunction(
1160                                    Constant::getNullValue(CalleeF->getType()));
1161      return nullptr;
1162    }
1163
1164  if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1165    // If CS does not return void then replaceAllUsesWith undef.
1166    // This allows ValueHandlers and custom metadata to adjust itself.
1167    if (!CS.getInstruction()->getType()->isVoidTy())
1168      ReplaceInstUsesWith(*CS.getInstruction(),
1169                          UndefValue::get(CS.getInstruction()->getType()));
1170
1171    if (isa<InvokeInst>(CS.getInstruction())) {
1172      // Can't remove an invoke because we cannot change the CFG.
1173      return nullptr;
1174    }
1175
1176    // This instruction is not reachable, just remove it.  We insert a store to
1177    // undef so that we know that this code is not reachable, despite the fact
1178    // that we can't modify the CFG here.
1179    new StoreInst(ConstantInt::getTrue(Callee->getContext()),
1180                  UndefValue::get(Type::getInt1PtrTy(Callee->getContext())),
1181                  CS.getInstruction());
1182
1183    return EraseInstFromFunction(*CS.getInstruction());
1184  }
1185
1186  if (IntrinsicInst *II = FindInitTrampoline(Callee))
1187    return transformCallThroughTrampoline(CS, II);
1188
1189  PointerType *PTy = cast<PointerType>(Callee->getType());
1190  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1191  if (FTy->isVarArg()) {
1192    int ix = FTy->getNumParams();
1193    // See if we can optimize any arguments passed through the varargs area of
1194    // the call.
1195    for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(),
1196           E = CS.arg_end(); I != E; ++I, ++ix) {
1197      CastInst *CI = dyn_cast<CastInst>(*I);
1198      if (CI && isSafeToEliminateVarargsCast(CS, CI, DL, ix)) {
1199        *I = CI->getOperand(0);
1200        Changed = true;
1201      }
1202    }
1203  }
1204
1205  if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
1206    // Inline asm calls cannot throw - mark them 'nounwind'.
1207    CS.setDoesNotThrow();
1208    Changed = true;
1209  }
1210
1211  // Try to optimize the call if possible, we require DataLayout for most of
1212  // this.  None of these calls are seen as possibly dead so go ahead and
1213  // delete the instruction now.
1214  if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1215    Instruction *I = tryOptimizeCall(CI, DL);
1216    // If we changed something return the result, etc. Otherwise let
1217    // the fallthrough check.
1218    if (I) return EraseInstFromFunction(*I);
1219  }
1220
1221  return Changed ? CS.getInstruction() : nullptr;
1222}
1223
1224// transformConstExprCastCall - If the callee is a constexpr cast of a function,
1225// attempt to move the cast to the arguments of the call/invoke.
1226//
1227bool InstCombiner::transformConstExprCastCall(CallSite CS) {
1228  Function *Callee =
1229    dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
1230  if (!Callee)
1231    return false;
1232  Instruction *Caller = CS.getInstruction();
1233  const AttributeSet &CallerPAL = CS.getAttributes();
1234
1235  // Okay, this is a cast from a function to a different type.  Unless doing so
1236  // would cause a type conversion of one of our arguments, change this call to
1237  // be a direct call with arguments casted to the appropriate types.
1238  //
1239  FunctionType *FT = Callee->getFunctionType();
1240  Type *OldRetTy = Caller->getType();
1241  Type *NewRetTy = FT->getReturnType();
1242
1243  // Check to see if we are changing the return type...
1244  if (OldRetTy != NewRetTy) {
1245
1246    if (NewRetTy->isStructTy())
1247      return false; // TODO: Handle multiple return values.
1248
1249    if (!CastInst::isBitCastable(NewRetTy, OldRetTy)) {
1250      if (Callee->isDeclaration())
1251        return false;   // Cannot transform this return value.
1252
1253      if (!Caller->use_empty() &&
1254          // void -> non-void is handled specially
1255          !NewRetTy->isVoidTy())
1256      return false;   // Cannot transform this return value.
1257    }
1258
1259    if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
1260      AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1261      if (RAttrs.
1262          hasAttributes(AttributeFuncs::
1263                        typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1264                        AttributeSet::ReturnIndex))
1265        return false;   // Attribute not compatible with transformed value.
1266    }
1267
1268    // If the callsite is an invoke instruction, and the return value is used by
1269    // a PHI node in a successor, we cannot change the return type of the call
1270    // because there is no place to put the cast instruction (without breaking
1271    // the critical edge).  Bail out in this case.
1272    if (!Caller->use_empty())
1273      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
1274        for (User *U : II->users())
1275          if (PHINode *PN = dyn_cast<PHINode>(U))
1276            if (PN->getParent() == II->getNormalDest() ||
1277                PN->getParent() == II->getUnwindDest())
1278              return false;
1279  }
1280
1281  unsigned NumActualArgs = CS.arg_size();
1282  unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
1283
1284  CallSite::arg_iterator AI = CS.arg_begin();
1285  for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
1286    Type *ParamTy = FT->getParamType(i);
1287    Type *ActTy = (*AI)->getType();
1288
1289    if (!CastInst::isBitCastable(ActTy, ParamTy))
1290      return false;   // Cannot transform this parameter value.
1291
1292    if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1).
1293          hasAttributes(AttributeFuncs::
1294                        typeIncompatible(ParamTy, i + 1), i + 1))
1295      return false;   // Attribute not compatible with transformed value.
1296
1297    if (CS.isInAllocaArgument(i))
1298      return false;   // Cannot transform to and from inalloca.
1299
1300    // If the parameter is passed as a byval argument, then we have to have a
1301    // sized type and the sized type has to have the same size as the old type.
1302    if (ParamTy != ActTy &&
1303        CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1,
1304                                                         Attribute::ByVal)) {
1305      PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy);
1306      if (!ParamPTy || !ParamPTy->getElementType()->isSized() || !DL)
1307        return false;
1308
1309      Type *CurElTy = ActTy->getPointerElementType();
1310      if (DL->getTypeAllocSize(CurElTy) !=
1311          DL->getTypeAllocSize(ParamPTy->getElementType()))
1312        return false;
1313    }
1314  }
1315
1316  if (Callee->isDeclaration()) {
1317    // Do not delete arguments unless we have a function body.
1318    if (FT->getNumParams() < NumActualArgs && !FT->isVarArg())
1319      return false;
1320
1321    // If the callee is just a declaration, don't change the varargsness of the
1322    // call.  We don't want to introduce a varargs call where one doesn't
1323    // already exist.
1324    PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType());
1325    if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg())
1326      return false;
1327
1328    // If both the callee and the cast type are varargs, we still have to make
1329    // sure the number of fixed parameters are the same or we have the same
1330    // ABI issues as if we introduce a varargs call.
1331    if (FT->isVarArg() &&
1332        cast<FunctionType>(APTy->getElementType())->isVarArg() &&
1333        FT->getNumParams() !=
1334        cast<FunctionType>(APTy->getElementType())->getNumParams())
1335      return false;
1336  }
1337
1338  if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
1339      !CallerPAL.isEmpty())
1340    // In this case we have more arguments than the new function type, but we
1341    // won't be dropping them.  Check that these extra arguments have attributes
1342    // that are compatible with being a vararg call argument.
1343    for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
1344      unsigned Index = CallerPAL.getSlotIndex(i - 1);
1345      if (Index <= FT->getNumParams())
1346        break;
1347
1348      // Check if it has an attribute that's incompatible with varargs.
1349      AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1);
1350      if (PAttrs.hasAttribute(Index, Attribute::StructRet))
1351        return false;
1352    }
1353
1354
1355  // Okay, we decided that this is a safe thing to do: go ahead and start
1356  // inserting cast instructions as necessary.
1357  std::vector<Value*> Args;
1358  Args.reserve(NumActualArgs);
1359  SmallVector<AttributeSet, 8> attrVec;
1360  attrVec.reserve(NumCommonArgs);
1361
1362  // Get any return attributes.
1363  AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex);
1364
1365  // If the return value is not being used, the type may not be compatible
1366  // with the existing attributes.  Wipe out any problematic attributes.
1367  RAttrs.
1368    removeAttributes(AttributeFuncs::
1369                     typeIncompatible(NewRetTy, AttributeSet::ReturnIndex),
1370                     AttributeSet::ReturnIndex);
1371
1372  // Add the new return attributes.
1373  if (RAttrs.hasAttributes())
1374    attrVec.push_back(AttributeSet::get(Caller->getContext(),
1375                                        AttributeSet::ReturnIndex, RAttrs));
1376
1377  AI = CS.arg_begin();
1378  for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
1379    Type *ParamTy = FT->getParamType(i);
1380
1381    if ((*AI)->getType() == ParamTy) {
1382      Args.push_back(*AI);
1383    } else {
1384      Args.push_back(Builder->CreateBitCast(*AI, ParamTy));
1385    }
1386
1387    // Add any parameter attributes.
1388    AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1389    if (PAttrs.hasAttributes())
1390      attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1,
1391                                          PAttrs));
1392  }
1393
1394  // If the function takes more arguments than the call was taking, add them
1395  // now.
1396  for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
1397    Args.push_back(Constant::getNullValue(FT->getParamType(i)));
1398
1399  // If we are removing arguments to the function, emit an obnoxious warning.
1400  if (FT->getNumParams() < NumActualArgs) {
1401    // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722
1402    if (FT->isVarArg()) {
1403      // Add all of the arguments in their promoted form to the arg list.
1404      for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
1405        Type *PTy = getPromotedType((*AI)->getType());
1406        if (PTy != (*AI)->getType()) {
1407          // Must promote to pass through va_arg area!
1408          Instruction::CastOps opcode =
1409            CastInst::getCastOpcode(*AI, false, PTy, false);
1410          Args.push_back(Builder->CreateCast(opcode, *AI, PTy));
1411        } else {
1412          Args.push_back(*AI);
1413        }
1414
1415        // Add any parameter attributes.
1416        AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1);
1417        if (PAttrs.hasAttributes())
1418          attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1,
1419                                              PAttrs));
1420      }
1421    }
1422  }
1423
1424  AttributeSet FnAttrs = CallerPAL.getFnAttributes();
1425  if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex))
1426    attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs));
1427
1428  if (NewRetTy->isVoidTy())
1429    Caller->setName("");   // Void type should not have a name.
1430
1431  const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(),
1432                                                       attrVec);
1433
1434  Instruction *NC;
1435  if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1436    NC = Builder->CreateInvoke(Callee, II->getNormalDest(),
1437                               II->getUnwindDest(), Args);
1438    NC->takeName(II);
1439    cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
1440    cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
1441  } else {
1442    CallInst *CI = cast<CallInst>(Caller);
1443    NC = Builder->CreateCall(Callee, Args);
1444    NC->takeName(CI);
1445    if (CI->isTailCall())
1446      cast<CallInst>(NC)->setTailCall();
1447    cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
1448    cast<CallInst>(NC)->setAttributes(NewCallerPAL);
1449  }
1450
1451  // Insert a cast of the return type as necessary.
1452  Value *NV = NC;
1453  if (OldRetTy != NV->getType() && !Caller->use_empty()) {
1454    if (!NV->getType()->isVoidTy()) {
1455      NV = NC = CastInst::Create(CastInst::BitCast, NC, OldRetTy);
1456      NC->setDebugLoc(Caller->getDebugLoc());
1457
1458      // If this is an invoke instruction, we should insert it after the first
1459      // non-phi, instruction in the normal successor block.
1460      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1461        BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt();
1462        InsertNewInstBefore(NC, *I);
1463      } else {
1464        // Otherwise, it's a call, just insert cast right after the call.
1465        InsertNewInstBefore(NC, *Caller);
1466      }
1467      Worklist.AddUsersToWorkList(*Caller);
1468    } else {
1469      NV = UndefValue::get(Caller->getType());
1470    }
1471  }
1472
1473  if (!Caller->use_empty())
1474    ReplaceInstUsesWith(*Caller, NV);
1475  else if (Caller->hasValueHandle())
1476    ValueHandleBase::ValueIsRAUWd(Caller, NV);
1477
1478  EraseInstFromFunction(*Caller);
1479  return true;
1480}
1481
1482// transformCallThroughTrampoline - Turn a call to a function created by
1483// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the
1484// underlying function.
1485//
1486Instruction *
1487InstCombiner::transformCallThroughTrampoline(CallSite CS,
1488                                             IntrinsicInst *Tramp) {
1489  Value *Callee = CS.getCalledValue();
1490  PointerType *PTy = cast<PointerType>(Callee->getType());
1491  FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
1492  const AttributeSet &Attrs = CS.getAttributes();
1493
1494  // If the call already has the 'nest' attribute somewhere then give up -
1495  // otherwise 'nest' would occur twice after splicing in the chain.
1496  if (Attrs.hasAttrSomewhere(Attribute::Nest))
1497    return nullptr;
1498
1499  assert(Tramp &&
1500         "transformCallThroughTrampoline called with incorrect CallSite.");
1501
1502  Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts());
1503  PointerType *NestFPTy = cast<PointerType>(NestF->getType());
1504  FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
1505
1506  const AttributeSet &NestAttrs = NestF->getAttributes();
1507  if (!NestAttrs.isEmpty()) {
1508    unsigned NestIdx = 1;
1509    Type *NestTy = nullptr;
1510    AttributeSet NestAttr;
1511
1512    // Look for a parameter marked with the 'nest' attribute.
1513    for (FunctionType::param_iterator I = NestFTy->param_begin(),
1514         E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
1515      if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) {
1516        // Record the parameter type and any other attributes.
1517        NestTy = *I;
1518        NestAttr = NestAttrs.getParamAttributes(NestIdx);
1519        break;
1520      }
1521
1522    if (NestTy) {
1523      Instruction *Caller = CS.getInstruction();
1524      std::vector<Value*> NewArgs;
1525      NewArgs.reserve(CS.arg_size() + 1);
1526
1527      SmallVector<AttributeSet, 8> NewAttrs;
1528      NewAttrs.reserve(Attrs.getNumSlots() + 1);
1529
1530      // Insert the nest argument into the call argument list, which may
1531      // mean appending it.  Likewise for attributes.
1532
1533      // Add any result attributes.
1534      if (Attrs.hasAttributes(AttributeSet::ReturnIndex))
1535        NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1536                                             Attrs.getRetAttributes()));
1537
1538      {
1539        unsigned Idx = 1;
1540        CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
1541        do {
1542          if (Idx == NestIdx) {
1543            // Add the chain argument and attributes.
1544            Value *NestVal = Tramp->getArgOperand(2);
1545            if (NestVal->getType() != NestTy)
1546              NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest");
1547            NewArgs.push_back(NestVal);
1548            NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1549                                                 NestAttr));
1550          }
1551
1552          if (I == E)
1553            break;
1554
1555          // Add the original argument and attributes.
1556          NewArgs.push_back(*I);
1557          AttributeSet Attr = Attrs.getParamAttributes(Idx);
1558          if (Attr.hasAttributes(Idx)) {
1559            AttrBuilder B(Attr, Idx);
1560            NewAttrs.push_back(AttributeSet::get(Caller->getContext(),
1561                                                 Idx + (Idx >= NestIdx), B));
1562          }
1563
1564          ++Idx, ++I;
1565        } while (1);
1566      }
1567
1568      // Add any function attributes.
1569      if (Attrs.hasAttributes(AttributeSet::FunctionIndex))
1570        NewAttrs.push_back(AttributeSet::get(FTy->getContext(),
1571                                             Attrs.getFnAttributes()));
1572
1573      // The trampoline may have been bitcast to a bogus type (FTy).
1574      // Handle this by synthesizing a new function type, equal to FTy
1575      // with the chain parameter inserted.
1576
1577      std::vector<Type*> NewTypes;
1578      NewTypes.reserve(FTy->getNumParams()+1);
1579
1580      // Insert the chain's type into the list of parameter types, which may
1581      // mean appending it.
1582      {
1583        unsigned Idx = 1;
1584        FunctionType::param_iterator I = FTy->param_begin(),
1585          E = FTy->param_end();
1586
1587        do {
1588          if (Idx == NestIdx)
1589            // Add the chain's type.
1590            NewTypes.push_back(NestTy);
1591
1592          if (I == E)
1593            break;
1594
1595          // Add the original type.
1596          NewTypes.push_back(*I);
1597
1598          ++Idx, ++I;
1599        } while (1);
1600      }
1601
1602      // Replace the trampoline call with a direct call.  Let the generic
1603      // code sort out any function type mismatches.
1604      FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
1605                                                FTy->isVarArg());
1606      Constant *NewCallee =
1607        NestF->getType() == PointerType::getUnqual(NewFTy) ?
1608        NestF : ConstantExpr::getBitCast(NestF,
1609                                         PointerType::getUnqual(NewFTy));
1610      const AttributeSet &NewPAL =
1611          AttributeSet::get(FTy->getContext(), NewAttrs);
1612
1613      Instruction *NewCaller;
1614      if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
1615        NewCaller = InvokeInst::Create(NewCallee,
1616                                       II->getNormalDest(), II->getUnwindDest(),
1617                                       NewArgs);
1618        cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
1619        cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
1620      } else {
1621        NewCaller = CallInst::Create(NewCallee, NewArgs);
1622        if (cast<CallInst>(Caller)->isTailCall())
1623          cast<CallInst>(NewCaller)->setTailCall();
1624        cast<CallInst>(NewCaller)->
1625          setCallingConv(cast<CallInst>(Caller)->getCallingConv());
1626        cast<CallInst>(NewCaller)->setAttributes(NewPAL);
1627      }
1628
1629      return NewCaller;
1630    }
1631  }
1632
1633  // Replace the trampoline call with a direct call.  Since there is no 'nest'
1634  // parameter, there is no need to adjust the argument list.  Let the generic
1635  // code sort out any function type mismatches.
1636  Constant *NewCallee =
1637    NestF->getType() == PTy ? NestF :
1638                              ConstantExpr::getBitCast(NestF, PTy);
1639  CS.setCalledFunction(NewCallee);
1640  return CS.getInstruction();
1641}
1642