CGBuiltin.cpp revision a6b8b2c09610b8bc4330e948ece8b940c2386406 
1//===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
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 contains code to emit Builtin calls as LLVM code.
11//
12//===----------------------------------------------------------------------===//
13
14#include "TargetInfo.h"
15#include "CodeGenFunction.h"
16#include "CodeGenModule.h"
17#include "CGObjCRuntime.h"
18#include "clang/Basic/TargetInfo.h"
19#include "clang/AST/APValue.h"
20#include "clang/AST/ASTContext.h"
21#include "clang/AST/Decl.h"
22#include "clang/Basic/TargetBuiltins.h"
23#include "llvm/Intrinsics.h"
24#include "llvm/Target/TargetData.h"
25
26using namespace clang;
27using namespace CodeGen;
28using namespace llvm;
29
30/// getBuiltinLibFunction - Given a builtin id for a function like
31/// "__builtin_fabsf", return a Function* for "fabsf".
32llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
33                                                  unsigned BuiltinID) {
34  assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
35
36  // Get the name, skip over the __builtin_ prefix (if necessary).
37  StringRef Name;
38  GlobalDecl D(FD);
39
40  // If the builtin has been declared explicitly with an assembler label,
41  // use the mangled name. This differs from the plain label on platforms
42  // that prefix labels.
43  if (FD->hasAttr<AsmLabelAttr>())
44    Name = getMangledName(D);
45  else
46    Name = Context.BuiltinInfo.GetName(BuiltinID) + 10;
47
48  llvm::FunctionType *Ty =
49    cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
50
51  return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
52}
53
54/// Emit the conversions required to turn the given value into an
55/// integer of the given size.
56static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
57                        QualType T, llvm::IntegerType *IntType) {
58  V = CGF.EmitToMemory(V, T);
59
60  if (V->getType()->isPointerTy())
61    return CGF.Builder.CreatePtrToInt(V, IntType);
62
63  assert(V->getType() == IntType);
64  return V;
65}
66
67static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
68                          QualType T, llvm::Type *ResultType) {
69  V = CGF.EmitFromMemory(V, T);
70
71  if (ResultType->isPointerTy())
72    return CGF.Builder.CreateIntToPtr(V, ResultType);
73
74  assert(V->getType() == ResultType);
75  return V;
76}
77
78/// Utility to insert an atomic instruction based on Instrinsic::ID
79/// and the expression node.
80static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
81                               llvm::AtomicRMWInst::BinOp Kind,
82                               const CallExpr *E) {
83  QualType T = E->getType();
84  assert(E->getArg(0)->getType()->isPointerType());
85  assert(CGF.getContext().hasSameUnqualifiedType(T,
86                                  E->getArg(0)->getType()->getPointeeType()));
87  assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
88
89  llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
90  unsigned AddrSpace =
91    cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
92
93  llvm::IntegerType *IntType =
94    llvm::IntegerType::get(CGF.getLLVMContext(),
95                           CGF.getContext().getTypeSize(T));
96  llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
97
98  llvm::Value *Args[2];
99  Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
100  Args[1] = CGF.EmitScalarExpr(E->getArg(1));
101  llvm::Type *ValueType = Args[1]->getType();
102  Args[1] = EmitToInt(CGF, Args[1], T, IntType);
103
104  llvm::Value *Result =
105      CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
106                                  llvm::SequentiallyConsistent);
107  Result = EmitFromInt(CGF, Result, T, ValueType);
108  return RValue::get(Result);
109}
110
111/// Utility to insert an atomic instruction based Instrinsic::ID and
112/// the expression node, where the return value is the result of the
113/// operation.
114static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
115                                   llvm::AtomicRMWInst::BinOp Kind,
116                                   const CallExpr *E,
117                                   Instruction::BinaryOps Op) {
118  QualType T = E->getType();
119  assert(E->getArg(0)->getType()->isPointerType());
120  assert(CGF.getContext().hasSameUnqualifiedType(T,
121                                  E->getArg(0)->getType()->getPointeeType()));
122  assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
123
124  llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
125  unsigned AddrSpace =
126    cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
127
128  llvm::IntegerType *IntType =
129    llvm::IntegerType::get(CGF.getLLVMContext(),
130                           CGF.getContext().getTypeSize(T));
131  llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
132
133  llvm::Value *Args[2];
134  Args[1] = CGF.EmitScalarExpr(E->getArg(1));
135  llvm::Type *ValueType = Args[1]->getType();
136  Args[1] = EmitToInt(CGF, Args[1], T, IntType);
137  Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
138
139  llvm::Value *Result =
140      CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
141                                  llvm::SequentiallyConsistent);
142  Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
143  Result = EmitFromInt(CGF, Result, T, ValueType);
144  return RValue::get(Result);
145}
146
147/// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy,
148/// which must be a scalar floating point type.
149static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) {
150  const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>();
151  assert(ValTyP && "isn't scalar fp type!");
152
153  StringRef FnName;
154  switch (ValTyP->getKind()) {
155  default: llvm_unreachable("Isn't a scalar fp type!");
156  case BuiltinType::Float:      FnName = "fabsf"; break;
157  case BuiltinType::Double:     FnName = "fabs"; break;
158  case BuiltinType::LongDouble: FnName = "fabsl"; break;
159  }
160
161  // The prototype is something that takes and returns whatever V's type is.
162  llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(),
163                                                   false);
164  llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName);
165
166  return CGF.Builder.CreateCall(Fn, V, "abs");
167}
168
169static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn,
170                              const CallExpr *E, llvm::Value *calleeValue) {
171  return CGF.EmitCall(E->getCallee()->getType(), calleeValue,
172                      ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn);
173}
174
175RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
176                                        unsigned BuiltinID, const CallExpr *E) {
177  // See if we can constant fold this builtin.  If so, don't emit it at all.
178  Expr::EvalResult Result;
179  if (E->Evaluate(Result, CGM.getContext()) &&
180      !Result.hasSideEffects()) {
181    if (Result.Val.isInt())
182      return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
183                                                Result.Val.getInt()));
184    if (Result.Val.isFloat())
185      return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
186                                               Result.Val.getFloat()));
187  }
188
189  switch (BuiltinID) {
190  default: break;  // Handle intrinsics and libm functions below.
191  case Builtin::BI__builtin___CFStringMakeConstantString:
192  case Builtin::BI__builtin___NSStringMakeConstantString:
193    return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0));
194  case Builtin::BI__builtin_stdarg_start:
195  case Builtin::BI__builtin_va_start:
196  case Builtin::BI__builtin_va_end: {
197    Value *ArgValue = EmitVAListRef(E->getArg(0));
198    llvm::Type *DestType = Int8PtrTy;
199    if (ArgValue->getType() != DestType)
200      ArgValue = Builder.CreateBitCast(ArgValue, DestType,
201                                       ArgValue->getName().data());
202
203    Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ?
204      Intrinsic::vaend : Intrinsic::vastart;
205    return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue));
206  }
207  case Builtin::BI__builtin_va_copy: {
208    Value *DstPtr = EmitVAListRef(E->getArg(0));
209    Value *SrcPtr = EmitVAListRef(E->getArg(1));
210
211    llvm::Type *Type = Int8PtrTy;
212
213    DstPtr = Builder.CreateBitCast(DstPtr, Type);
214    SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
215    return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy),
216                                           DstPtr, SrcPtr));
217  }
218  case Builtin::BI__builtin_abs: {
219    Value *ArgValue = EmitScalarExpr(E->getArg(0));
220
221    Value *NegOp = Builder.CreateNeg(ArgValue, "neg");
222    Value *CmpResult =
223    Builder.CreateICmpSGE(ArgValue,
224                          llvm::Constant::getNullValue(ArgValue->getType()),
225                                                            "abscond");
226    Value *Result =
227      Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs");
228
229    return RValue::get(Result);
230  }
231  case Builtin::BI__builtin_ctz:
232  case Builtin::BI__builtin_ctzl:
233  case Builtin::BI__builtin_ctzll: {
234    Value *ArgValue = EmitScalarExpr(E->getArg(0));
235
236    llvm::Type *ArgType = ArgValue->getType();
237    Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
238
239    llvm::Type *ResultType = ConvertType(E->getType());
240    Value *Result = Builder.CreateCall(F, ArgValue);
241    if (Result->getType() != ResultType)
242      Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
243                                     "cast");
244    return RValue::get(Result);
245  }
246  case Builtin::BI__builtin_clz:
247  case Builtin::BI__builtin_clzl:
248  case Builtin::BI__builtin_clzll: {
249    Value *ArgValue = EmitScalarExpr(E->getArg(0));
250
251    llvm::Type *ArgType = ArgValue->getType();
252    Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
253
254    llvm::Type *ResultType = ConvertType(E->getType());
255    Value *Result = Builder.CreateCall(F, ArgValue);
256    if (Result->getType() != ResultType)
257      Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
258                                     "cast");
259    return RValue::get(Result);
260  }
261  case Builtin::BI__builtin_ffs:
262  case Builtin::BI__builtin_ffsl:
263  case Builtin::BI__builtin_ffsll: {
264    // ffs(x) -> x ? cttz(x) + 1 : 0
265    Value *ArgValue = EmitScalarExpr(E->getArg(0));
266
267    llvm::Type *ArgType = ArgValue->getType();
268    Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
269
270    llvm::Type *ResultType = ConvertType(E->getType());
271    Value *Tmp = Builder.CreateAdd(Builder.CreateCall(F, ArgValue),
272                                   llvm::ConstantInt::get(ArgType, 1));
273    Value *Zero = llvm::Constant::getNullValue(ArgType);
274    Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
275    Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
276    if (Result->getType() != ResultType)
277      Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
278                                     "cast");
279    return RValue::get(Result);
280  }
281  case Builtin::BI__builtin_parity:
282  case Builtin::BI__builtin_parityl:
283  case Builtin::BI__builtin_parityll: {
284    // parity(x) -> ctpop(x) & 1
285    Value *ArgValue = EmitScalarExpr(E->getArg(0));
286
287    llvm::Type *ArgType = ArgValue->getType();
288    Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
289
290    llvm::Type *ResultType = ConvertType(E->getType());
291    Value *Tmp = Builder.CreateCall(F, ArgValue);
292    Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
293    if (Result->getType() != ResultType)
294      Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
295                                     "cast");
296    return RValue::get(Result);
297  }
298  case Builtin::BI__builtin_popcount:
299  case Builtin::BI__builtin_popcountl:
300  case Builtin::BI__builtin_popcountll: {
301    Value *ArgValue = EmitScalarExpr(E->getArg(0));
302
303    llvm::Type *ArgType = ArgValue->getType();
304    Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
305
306    llvm::Type *ResultType = ConvertType(E->getType());
307    Value *Result = Builder.CreateCall(F, ArgValue);
308    if (Result->getType() != ResultType)
309      Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
310                                     "cast");
311    return RValue::get(Result);
312  }
313  case Builtin::BI__builtin_expect: {
314    Value *ArgValue = EmitScalarExpr(E->getArg(0));
315    llvm::Type *ArgType = ArgValue->getType();
316
317    Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
318    Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
319
320    Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue,
321                                        "expval");
322    return RValue::get(Result);
323  }
324  case Builtin::BI__builtin_bswap32:
325  case Builtin::BI__builtin_bswap64: {
326    Value *ArgValue = EmitScalarExpr(E->getArg(0));
327    llvm::Type *ArgType = ArgValue->getType();
328    Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType);
329    return RValue::get(Builder.CreateCall(F, ArgValue));
330  }
331  case Builtin::BI__builtin_object_size: {
332    // We pass this builtin onto the optimizer so that it can
333    // figure out the object size in more complex cases.
334    llvm::Type *ResType = ConvertType(E->getType());
335
336    // LLVM only supports 0 and 2, make sure that we pass along that
337    // as a boolean.
338    Value *Ty = EmitScalarExpr(E->getArg(1));
339    ConstantInt *CI = dyn_cast<ConstantInt>(Ty);
340    assert(CI);
341    uint64_t val = CI->getZExtValue();
342    CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1);
343
344    Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType);
345    return RValue::get(Builder.CreateCall2(F,
346                                           EmitScalarExpr(E->getArg(0)),
347                                           CI));
348  }
349  case Builtin::BI__builtin_prefetch: {
350    Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
351    // FIXME: Technically these constants should of type 'int', yes?
352    RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
353      llvm::ConstantInt::get(Int32Ty, 0);
354    Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
355      llvm::ConstantInt::get(Int32Ty, 3);
356    Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
357    Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
358    return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data));
359  }
360  case Builtin::BI__builtin_trap: {
361    Value *F = CGM.getIntrinsic(Intrinsic::trap);
362    return RValue::get(Builder.CreateCall(F));
363  }
364  case Builtin::BI__builtin_unreachable: {
365    if (CatchUndefined)
366      EmitBranch(getTrapBB());
367    else
368      Builder.CreateUnreachable();
369
370    // We do need to preserve an insertion point.
371    EmitBlock(createBasicBlock("unreachable.cont"));
372
373    return RValue::get(0);
374  }
375
376  case Builtin::BI__builtin_powi:
377  case Builtin::BI__builtin_powif:
378  case Builtin::BI__builtin_powil: {
379    Value *Base = EmitScalarExpr(E->getArg(0));
380    Value *Exponent = EmitScalarExpr(E->getArg(1));
381    llvm::Type *ArgType = Base->getType();
382    Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
383    return RValue::get(Builder.CreateCall2(F, Base, Exponent));
384  }
385
386  case Builtin::BI__builtin_isgreater:
387  case Builtin::BI__builtin_isgreaterequal:
388  case Builtin::BI__builtin_isless:
389  case Builtin::BI__builtin_islessequal:
390  case Builtin::BI__builtin_islessgreater:
391  case Builtin::BI__builtin_isunordered: {
392    // Ordered comparisons: we know the arguments to these are matching scalar
393    // floating point values.
394    Value *LHS = EmitScalarExpr(E->getArg(0));
395    Value *RHS = EmitScalarExpr(E->getArg(1));
396
397    switch (BuiltinID) {
398    default: llvm_unreachable("Unknown ordered comparison");
399    case Builtin::BI__builtin_isgreater:
400      LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
401      break;
402    case Builtin::BI__builtin_isgreaterequal:
403      LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
404      break;
405    case Builtin::BI__builtin_isless:
406      LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
407      break;
408    case Builtin::BI__builtin_islessequal:
409      LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
410      break;
411    case Builtin::BI__builtin_islessgreater:
412      LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
413      break;
414    case Builtin::BI__builtin_isunordered:
415      LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
416      break;
417    }
418    // ZExt bool to int type.
419    return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
420  }
421  case Builtin::BI__builtin_isnan: {
422    Value *V = EmitScalarExpr(E->getArg(0));
423    V = Builder.CreateFCmpUNO(V, V, "cmp");
424    return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
425  }
426
427  case Builtin::BI__builtin_isinf: {
428    // isinf(x) --> fabs(x) == infinity
429    Value *V = EmitScalarExpr(E->getArg(0));
430    V = EmitFAbs(*this, V, E->getArg(0)->getType());
431
432    V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf");
433    return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
434  }
435
436  // TODO: BI__builtin_isinf_sign
437  //   isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0
438
439  case Builtin::BI__builtin_isnormal: {
440    // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
441    Value *V = EmitScalarExpr(E->getArg(0));
442    Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
443
444    Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
445    Value *IsLessThanInf =
446      Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
447    APFloat Smallest = APFloat::getSmallestNormalized(
448                   getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
449    Value *IsNormal =
450      Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
451                            "isnormal");
452    V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
453    V = Builder.CreateAnd(V, IsNormal, "and");
454    return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
455  }
456
457  case Builtin::BI__builtin_isfinite: {
458    // isfinite(x) --> x == x && fabs(x) != infinity;
459    Value *V = EmitScalarExpr(E->getArg(0));
460    Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
461
462    Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
463    Value *IsNotInf =
464      Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
465
466    V = Builder.CreateAnd(Eq, IsNotInf, "and");
467    return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
468  }
469
470  case Builtin::BI__builtin_fpclassify: {
471    Value *V = EmitScalarExpr(E->getArg(5));
472    llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
473
474    // Create Result
475    BasicBlock *Begin = Builder.GetInsertBlock();
476    BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
477    Builder.SetInsertPoint(End);
478    PHINode *Result =
479      Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
480                        "fpclassify_result");
481
482    // if (V==0) return FP_ZERO
483    Builder.SetInsertPoint(Begin);
484    Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
485                                          "iszero");
486    Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
487    BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
488    Builder.CreateCondBr(IsZero, End, NotZero);
489    Result->addIncoming(ZeroLiteral, Begin);
490
491    // if (V != V) return FP_NAN
492    Builder.SetInsertPoint(NotZero);
493    Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
494    Value *NanLiteral = EmitScalarExpr(E->getArg(0));
495    BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
496    Builder.CreateCondBr(IsNan, End, NotNan);
497    Result->addIncoming(NanLiteral, NotZero);
498
499    // if (fabs(V) == infinity) return FP_INFINITY
500    Builder.SetInsertPoint(NotNan);
501    Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType());
502    Value *IsInf =
503      Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
504                            "isinf");
505    Value *InfLiteral = EmitScalarExpr(E->getArg(1));
506    BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
507    Builder.CreateCondBr(IsInf, End, NotInf);
508    Result->addIncoming(InfLiteral, NotNan);
509
510    // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
511    Builder.SetInsertPoint(NotInf);
512    APFloat Smallest = APFloat::getSmallestNormalized(
513        getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
514    Value *IsNormal =
515      Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
516                            "isnormal");
517    Value *NormalResult =
518      Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
519                           EmitScalarExpr(E->getArg(3)));
520    Builder.CreateBr(End);
521    Result->addIncoming(NormalResult, NotInf);
522
523    // return Result
524    Builder.SetInsertPoint(End);
525    return RValue::get(Result);
526  }
527
528  case Builtin::BIalloca:
529  case Builtin::BI__builtin_alloca: {
530    Value *Size = EmitScalarExpr(E->getArg(0));
531    return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size));
532  }
533  case Builtin::BIbzero:
534  case Builtin::BI__builtin_bzero: {
535    Value *Address = EmitScalarExpr(E->getArg(0));
536    Value *SizeVal = EmitScalarExpr(E->getArg(1));
537    Builder.CreateMemSet(Address, Builder.getInt8(0), SizeVal, 1, false);
538    return RValue::get(Address);
539  }
540  case Builtin::BImemcpy:
541  case Builtin::BI__builtin_memcpy: {
542    Value *Address = EmitScalarExpr(E->getArg(0));
543    Value *SrcAddr = EmitScalarExpr(E->getArg(1));
544    Value *SizeVal = EmitScalarExpr(E->getArg(2));
545    Builder.CreateMemCpy(Address, SrcAddr, SizeVal, 1, false);
546    return RValue::get(Address);
547  }
548
549  case Builtin::BI__builtin___memcpy_chk: {
550    // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
551    llvm::APSInt Size, DstSize;
552    if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
553        !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
554      break;
555    if (Size.ugt(DstSize))
556      break;
557    Value *Dest = EmitScalarExpr(E->getArg(0));
558    Value *Src = EmitScalarExpr(E->getArg(1));
559    Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
560    Builder.CreateMemCpy(Dest, Src, SizeVal, 1, false);
561    return RValue::get(Dest);
562  }
563
564  case Builtin::BI__builtin_objc_memmove_collectable: {
565    Value *Address = EmitScalarExpr(E->getArg(0));
566    Value *SrcAddr = EmitScalarExpr(E->getArg(1));
567    Value *SizeVal = EmitScalarExpr(E->getArg(2));
568    CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
569                                                  Address, SrcAddr, SizeVal);
570    return RValue::get(Address);
571  }
572
573  case Builtin::BI__builtin___memmove_chk: {
574    // fold __builtin_memmove_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
575    llvm::APSInt Size, DstSize;
576    if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
577        !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
578      break;
579    if (Size.ugt(DstSize))
580      break;
581    Value *Dest = EmitScalarExpr(E->getArg(0));
582    Value *Src = EmitScalarExpr(E->getArg(1));
583    Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
584    Builder.CreateMemMove(Dest, Src, SizeVal, 1, false);
585    return RValue::get(Dest);
586  }
587
588  case Builtin::BImemmove:
589  case Builtin::BI__builtin_memmove: {
590    Value *Address = EmitScalarExpr(E->getArg(0));
591    Value *SrcAddr = EmitScalarExpr(E->getArg(1));
592    Value *SizeVal = EmitScalarExpr(E->getArg(2));
593    Builder.CreateMemMove(Address, SrcAddr, SizeVal, 1, false);
594    return RValue::get(Address);
595  }
596  case Builtin::BImemset:
597  case Builtin::BI__builtin_memset: {
598    Value *Address = EmitScalarExpr(E->getArg(0));
599    Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
600                                         Builder.getInt8Ty());
601    Value *SizeVal = EmitScalarExpr(E->getArg(2));
602    Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false);
603    return RValue::get(Address);
604  }
605  case Builtin::BI__builtin___memset_chk: {
606    // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
607    llvm::APSInt Size, DstSize;
608    if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
609        !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
610      break;
611    if (Size.ugt(DstSize))
612      break;
613    Value *Address = EmitScalarExpr(E->getArg(0));
614    Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
615                                         Builder.getInt8Ty());
616    Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
617    Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false);
618
619    return RValue::get(Address);
620  }
621  case Builtin::BI__builtin_dwarf_cfa: {
622    // The offset in bytes from the first argument to the CFA.
623    //
624    // Why on earth is this in the frontend?  Is there any reason at
625    // all that the backend can't reasonably determine this while
626    // lowering llvm.eh.dwarf.cfa()?
627    //
628    // TODO: If there's a satisfactory reason, add a target hook for
629    // this instead of hard-coding 0, which is correct for most targets.
630    int32_t Offset = 0;
631
632    Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
633    return RValue::get(Builder.CreateCall(F,
634                                      llvm::ConstantInt::get(Int32Ty, Offset)));
635  }
636  case Builtin::BI__builtin_return_address: {
637    Value *Depth = EmitScalarExpr(E->getArg(0));
638    Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
639    Value *F = CGM.getIntrinsic(Intrinsic::returnaddress);
640    return RValue::get(Builder.CreateCall(F, Depth));
641  }
642  case Builtin::BI__builtin_frame_address: {
643    Value *Depth = EmitScalarExpr(E->getArg(0));
644    Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
645    Value *F = CGM.getIntrinsic(Intrinsic::frameaddress);
646    return RValue::get(Builder.CreateCall(F, Depth));
647  }
648  case Builtin::BI__builtin_extract_return_addr: {
649    Value *Address = EmitScalarExpr(E->getArg(0));
650    Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
651    return RValue::get(Result);
652  }
653  case Builtin::BI__builtin_frob_return_addr: {
654    Value *Address = EmitScalarExpr(E->getArg(0));
655    Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
656    return RValue::get(Result);
657  }
658  case Builtin::BI__builtin_dwarf_sp_column: {
659    llvm::IntegerType *Ty
660      = cast<llvm::IntegerType>(ConvertType(E->getType()));
661    int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
662    if (Column == -1) {
663      CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
664      return RValue::get(llvm::UndefValue::get(Ty));
665    }
666    return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
667  }
668  case Builtin::BI__builtin_init_dwarf_reg_size_table: {
669    Value *Address = EmitScalarExpr(E->getArg(0));
670    if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
671      CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
672    return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
673  }
674  case Builtin::BI__builtin_eh_return: {
675    Value *Int = EmitScalarExpr(E->getArg(0));
676    Value *Ptr = EmitScalarExpr(E->getArg(1));
677
678    llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
679    assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
680           "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
681    Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32
682                                  ? Intrinsic::eh_return_i32
683                                  : Intrinsic::eh_return_i64);
684    Builder.CreateCall2(F, Int, Ptr);
685    Builder.CreateUnreachable();
686
687    // We do need to preserve an insertion point.
688    EmitBlock(createBasicBlock("builtin_eh_return.cont"));
689
690    return RValue::get(0);
691  }
692  case Builtin::BI__builtin_unwind_init: {
693    Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
694    return RValue::get(Builder.CreateCall(F));
695  }
696  case Builtin::BI__builtin_extend_pointer: {
697    // Extends a pointer to the size of an _Unwind_Word, which is
698    // uint64_t on all platforms.  Generally this gets poked into a
699    // register and eventually used as an address, so if the
700    // addressing registers are wider than pointers and the platform
701    // doesn't implicitly ignore high-order bits when doing
702    // addressing, we need to make sure we zext / sext based on
703    // the platform's expectations.
704    //
705    // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
706
707    // Cast the pointer to intptr_t.
708    Value *Ptr = EmitScalarExpr(E->getArg(0));
709    Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
710
711    // If that's 64 bits, we're done.
712    if (IntPtrTy->getBitWidth() == 64)
713      return RValue::get(Result);
714
715    // Otherwise, ask the codegen data what to do.
716    if (getTargetHooks().extendPointerWithSExt())
717      return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
718    else
719      return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
720  }
721  case Builtin::BI__builtin_setjmp: {
722    // Buffer is a void**.
723    Value *Buf = EmitScalarExpr(E->getArg(0));
724
725    // Store the frame pointer to the setjmp buffer.
726    Value *FrameAddr =
727      Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
728                         ConstantInt::get(Int32Ty, 0));
729    Builder.CreateStore(FrameAddr, Buf);
730
731    // Store the stack pointer to the setjmp buffer.
732    Value *StackAddr =
733      Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
734    Value *StackSaveSlot =
735      Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2));
736    Builder.CreateStore(StackAddr, StackSaveSlot);
737
738    // Call LLVM's EH setjmp, which is lightweight.
739    Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
740    Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
741    return RValue::get(Builder.CreateCall(F, Buf));
742  }
743  case Builtin::BI__builtin_longjmp: {
744    Value *Buf = EmitScalarExpr(E->getArg(0));
745    Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
746
747    // Call LLVM's EH longjmp, which is lightweight.
748    Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
749
750    // longjmp doesn't return; mark this as unreachable.
751    Builder.CreateUnreachable();
752
753    // We do need to preserve an insertion point.
754    EmitBlock(createBasicBlock("longjmp.cont"));
755
756    return RValue::get(0);
757  }
758  case Builtin::BI__sync_fetch_and_add:
759  case Builtin::BI__sync_fetch_and_sub:
760  case Builtin::BI__sync_fetch_and_or:
761  case Builtin::BI__sync_fetch_and_and:
762  case Builtin::BI__sync_fetch_and_xor:
763  case Builtin::BI__sync_add_and_fetch:
764  case Builtin::BI__sync_sub_and_fetch:
765  case Builtin::BI__sync_and_and_fetch:
766  case Builtin::BI__sync_or_and_fetch:
767  case Builtin::BI__sync_xor_and_fetch:
768  case Builtin::BI__sync_val_compare_and_swap:
769  case Builtin::BI__sync_bool_compare_and_swap:
770  case Builtin::BI__sync_lock_test_and_set:
771  case Builtin::BI__sync_lock_release:
772  case Builtin::BI__sync_swap:
773    llvm_unreachable("Shouldn't make it through sema");
774  case Builtin::BI__sync_fetch_and_add_1:
775  case Builtin::BI__sync_fetch_and_add_2:
776  case Builtin::BI__sync_fetch_and_add_4:
777  case Builtin::BI__sync_fetch_and_add_8:
778  case Builtin::BI__sync_fetch_and_add_16:
779    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
780  case Builtin::BI__sync_fetch_and_sub_1:
781  case Builtin::BI__sync_fetch_and_sub_2:
782  case Builtin::BI__sync_fetch_and_sub_4:
783  case Builtin::BI__sync_fetch_and_sub_8:
784  case Builtin::BI__sync_fetch_and_sub_16:
785    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
786  case Builtin::BI__sync_fetch_and_or_1:
787  case Builtin::BI__sync_fetch_and_or_2:
788  case Builtin::BI__sync_fetch_and_or_4:
789  case Builtin::BI__sync_fetch_and_or_8:
790  case Builtin::BI__sync_fetch_and_or_16:
791    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
792  case Builtin::BI__sync_fetch_and_and_1:
793  case Builtin::BI__sync_fetch_and_and_2:
794  case Builtin::BI__sync_fetch_and_and_4:
795  case Builtin::BI__sync_fetch_and_and_8:
796  case Builtin::BI__sync_fetch_and_and_16:
797    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
798  case Builtin::BI__sync_fetch_and_xor_1:
799  case Builtin::BI__sync_fetch_and_xor_2:
800  case Builtin::BI__sync_fetch_and_xor_4:
801  case Builtin::BI__sync_fetch_and_xor_8:
802  case Builtin::BI__sync_fetch_and_xor_16:
803    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
804
805  // Clang extensions: not overloaded yet.
806  case Builtin::BI__sync_fetch_and_min:
807    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
808  case Builtin::BI__sync_fetch_and_max:
809    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
810  case Builtin::BI__sync_fetch_and_umin:
811    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
812  case Builtin::BI__sync_fetch_and_umax:
813    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
814
815  case Builtin::BI__sync_add_and_fetch_1:
816  case Builtin::BI__sync_add_and_fetch_2:
817  case Builtin::BI__sync_add_and_fetch_4:
818  case Builtin::BI__sync_add_and_fetch_8:
819  case Builtin::BI__sync_add_and_fetch_16:
820    return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
821                                llvm::Instruction::Add);
822  case Builtin::BI__sync_sub_and_fetch_1:
823  case Builtin::BI__sync_sub_and_fetch_2:
824  case Builtin::BI__sync_sub_and_fetch_4:
825  case Builtin::BI__sync_sub_and_fetch_8:
826  case Builtin::BI__sync_sub_and_fetch_16:
827    return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
828                                llvm::Instruction::Sub);
829  case Builtin::BI__sync_and_and_fetch_1:
830  case Builtin::BI__sync_and_and_fetch_2:
831  case Builtin::BI__sync_and_and_fetch_4:
832  case Builtin::BI__sync_and_and_fetch_8:
833  case Builtin::BI__sync_and_and_fetch_16:
834    return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
835                                llvm::Instruction::And);
836  case Builtin::BI__sync_or_and_fetch_1:
837  case Builtin::BI__sync_or_and_fetch_2:
838  case Builtin::BI__sync_or_and_fetch_4:
839  case Builtin::BI__sync_or_and_fetch_8:
840  case Builtin::BI__sync_or_and_fetch_16:
841    return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
842                                llvm::Instruction::Or);
843  case Builtin::BI__sync_xor_and_fetch_1:
844  case Builtin::BI__sync_xor_and_fetch_2:
845  case Builtin::BI__sync_xor_and_fetch_4:
846  case Builtin::BI__sync_xor_and_fetch_8:
847  case Builtin::BI__sync_xor_and_fetch_16:
848    return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
849                                llvm::Instruction::Xor);
850
851  case Builtin::BI__sync_val_compare_and_swap_1:
852  case Builtin::BI__sync_val_compare_and_swap_2:
853  case Builtin::BI__sync_val_compare_and_swap_4:
854  case Builtin::BI__sync_val_compare_and_swap_8:
855  case Builtin::BI__sync_val_compare_and_swap_16: {
856    QualType T = E->getType();
857    llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
858    unsigned AddrSpace =
859      cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
860
861    llvm::IntegerType *IntType =
862      llvm::IntegerType::get(getLLVMContext(),
863                             getContext().getTypeSize(T));
864    llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
865
866    Value *Args[3];
867    Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
868    Args[1] = EmitScalarExpr(E->getArg(1));
869    llvm::Type *ValueType = Args[1]->getType();
870    Args[1] = EmitToInt(*this, Args[1], T, IntType);
871    Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
872
873    Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
874                                                llvm::SequentiallyConsistent);
875    Result = EmitFromInt(*this, Result, T, ValueType);
876    return RValue::get(Result);
877  }
878
879  case Builtin::BI__sync_bool_compare_and_swap_1:
880  case Builtin::BI__sync_bool_compare_and_swap_2:
881  case Builtin::BI__sync_bool_compare_and_swap_4:
882  case Builtin::BI__sync_bool_compare_and_swap_8:
883  case Builtin::BI__sync_bool_compare_and_swap_16: {
884    QualType T = E->getArg(1)->getType();
885    llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
886    unsigned AddrSpace =
887      cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
888
889    llvm::IntegerType *IntType =
890      llvm::IntegerType::get(getLLVMContext(),
891                             getContext().getTypeSize(T));
892    llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
893
894    Value *Args[3];
895    Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
896    Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType);
897    Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
898
899    Value *OldVal = Args[1];
900    Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
901                                                 llvm::SequentiallyConsistent);
902    Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal);
903    // zext bool to int.
904    Result = Builder.CreateZExt(Result, ConvertType(E->getType()));
905    return RValue::get(Result);
906  }
907
908  case Builtin::BI__sync_swap_1:
909  case Builtin::BI__sync_swap_2:
910  case Builtin::BI__sync_swap_4:
911  case Builtin::BI__sync_swap_8:
912  case Builtin::BI__sync_swap_16:
913    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
914
915  case Builtin::BI__sync_lock_test_and_set_1:
916  case Builtin::BI__sync_lock_test_and_set_2:
917  case Builtin::BI__sync_lock_test_and_set_4:
918  case Builtin::BI__sync_lock_test_and_set_8:
919  case Builtin::BI__sync_lock_test_and_set_16:
920    return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
921
922  case Builtin::BI__sync_lock_release_1:
923  case Builtin::BI__sync_lock_release_2:
924  case Builtin::BI__sync_lock_release_4:
925  case Builtin::BI__sync_lock_release_8:
926  case Builtin::BI__sync_lock_release_16: {
927    Value *Ptr = EmitScalarExpr(E->getArg(0));
928    llvm::Type *ElLLVMTy =
929      cast<llvm::PointerType>(Ptr->getType())->getElementType();
930    llvm::StoreInst *Store =
931      Builder.CreateStore(llvm::Constant::getNullValue(ElLLVMTy), Ptr);
932    QualType ElTy = E->getArg(0)->getType()->getPointeeType();
933    CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
934    Store->setAlignment(StoreSize.getQuantity());
935    Store->setAtomic(llvm::Release);
936    return RValue::get(0);
937  }
938
939  case Builtin::BI__sync_synchronize: {
940    // We assume this is supposed to correspond to a C++0x-style
941    // sequentially-consistent fence (i.e. this is only usable for
942    // synchonization, not device I/O or anything like that). This intrinsic
943    // is really badly designed in the sense that in theory, there isn't
944    // any way to safely use it... but in practice, it mostly works
945    // to use it with non-atomic loads and stores to get acquire/release
946    // semantics.
947    Builder.CreateFence(llvm::SequentiallyConsistent);
948    return RValue::get(0);
949  }
950
951    // Library functions with special handling.
952  case Builtin::BIsqrt:
953  case Builtin::BIsqrtf:
954  case Builtin::BIsqrtl: {
955    // TODO: there is currently no set of optimizer flags
956    // sufficient for us to rewrite sqrt to @llvm.sqrt.
957    // -fmath-errno=0 is not good enough; we need finiteness.
958    // We could probably precondition the call with an ult
959    // against 0, but is that worth the complexity?
960    break;
961  }
962
963  case Builtin::BIpow:
964  case Builtin::BIpowf:
965  case Builtin::BIpowl: {
966    // Rewrite sqrt to intrinsic if allowed.
967    if (!FD->hasAttr<ConstAttr>())
968      break;
969    Value *Base = EmitScalarExpr(E->getArg(0));
970    Value *Exponent = EmitScalarExpr(E->getArg(1));
971    llvm::Type *ArgType = Base->getType();
972    Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType);
973    return RValue::get(Builder.CreateCall2(F, Base, Exponent));
974  }
975
976  case Builtin::BIfma:
977  case Builtin::BIfmaf:
978  case Builtin::BIfmal:
979  case Builtin::BI__builtin_fma:
980  case Builtin::BI__builtin_fmaf:
981  case Builtin::BI__builtin_fmal: {
982    // Rewrite fma to intrinsic.
983    Value *FirstArg = EmitScalarExpr(E->getArg(0));
984    llvm::Type *ArgType = FirstArg->getType();
985    Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType);
986    return RValue::get(Builder.CreateCall3(F, FirstArg,
987                                              EmitScalarExpr(E->getArg(1)),
988                                              EmitScalarExpr(E->getArg(2))));
989  }
990
991  case Builtin::BI__builtin_signbit:
992  case Builtin::BI__builtin_signbitf:
993  case Builtin::BI__builtin_signbitl: {
994    LLVMContext &C = CGM.getLLVMContext();
995
996    Value *Arg = EmitScalarExpr(E->getArg(0));
997    llvm::Type *ArgTy = Arg->getType();
998    if (ArgTy->isPPC_FP128Ty())
999      break; // FIXME: I'm not sure what the right implementation is here.
1000    int ArgWidth = ArgTy->getPrimitiveSizeInBits();
1001    llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth);
1002    Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy);
1003    Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy);
1004    Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp);
1005    return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType())));
1006  }
1007  case Builtin::BI__builtin_annotation: {
1008    llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
1009    llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
1010                                      AnnVal->getType());
1011
1012    // Get the annotation string, go through casts. Sema requires this to be a
1013    // non-wide string literal, potentially casted, so the cast<> is safe.
1014    const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
1015    llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
1016    return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
1017  }
1018  }
1019
1020  // If this is an alias for a lib function (e.g. __builtin_sin), emit
1021  // the call using the normal call path, but using the unmangled
1022  // version of the function name.
1023  if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
1024    return emitLibraryCall(*this, FD, E,
1025                           CGM.getBuiltinLibFunction(FD, BuiltinID));
1026
1027  // If this is a predefined lib function (e.g. malloc), emit the call
1028  // using exactly the normal call path.
1029  if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
1030    return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee()));
1031
1032  // See if we have a target specific intrinsic.
1033  const char *Name = getContext().BuiltinInfo.GetName(BuiltinID);
1034  Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
1035  if (const char *Prefix =
1036      llvm::Triple::getArchTypePrefix(Target.getTriple().getArch()))
1037    IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name);
1038
1039  if (IntrinsicID != Intrinsic::not_intrinsic) {
1040    SmallVector<Value*, 16> Args;
1041
1042    // Find out if any arguments are required to be integer constant
1043    // expressions.
1044    unsigned ICEArguments = 0;
1045    ASTContext::GetBuiltinTypeError Error;
1046    getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1047    assert(Error == ASTContext::GE_None && "Should not codegen an error");
1048
1049    Function *F = CGM.getIntrinsic(IntrinsicID);
1050    llvm::FunctionType *FTy = F->getFunctionType();
1051
1052    for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
1053      Value *ArgValue;
1054      // If this is a normal argument, just emit it as a scalar.
1055      if ((ICEArguments & (1 << i)) == 0) {
1056        ArgValue = EmitScalarExpr(E->getArg(i));
1057      } else {
1058        // If this is required to be a constant, constant fold it so that we
1059        // know that the generated intrinsic gets a ConstantInt.
1060        llvm::APSInt Result;
1061        bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
1062        assert(IsConst && "Constant arg isn't actually constant?");
1063        (void)IsConst;
1064        ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
1065      }
1066
1067      // If the intrinsic arg type is different from the builtin arg type
1068      // we need to do a bit cast.
1069      llvm::Type *PTy = FTy->getParamType(i);
1070      if (PTy != ArgValue->getType()) {
1071        assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
1072               "Must be able to losslessly bit cast to param");
1073        ArgValue = Builder.CreateBitCast(ArgValue, PTy);
1074      }
1075
1076      Args.push_back(ArgValue);
1077    }
1078
1079    Value *V = Builder.CreateCall(F, Args);
1080    QualType BuiltinRetType = E->getType();
1081
1082    llvm::Type *RetTy = llvm::Type::getVoidTy(getLLVMContext());
1083    if (!BuiltinRetType->isVoidType()) RetTy = ConvertType(BuiltinRetType);
1084
1085    if (RetTy != V->getType()) {
1086      assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
1087             "Must be able to losslessly bit cast result type");
1088      V = Builder.CreateBitCast(V, RetTy);
1089    }
1090
1091    return RValue::get(V);
1092  }
1093
1094  // See if we have a target specific builtin that needs to be lowered.
1095  if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
1096    return RValue::get(V);
1097
1098  ErrorUnsupported(E, "builtin function");
1099
1100  // Unknown builtin, for now just dump it out and return undef.
1101  if (hasAggregateLLVMType(E->getType()))
1102    return RValue::getAggregate(CreateMemTemp(E->getType()));
1103  return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
1104}
1105
1106Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
1107                                              const CallExpr *E) {
1108  switch (Target.getTriple().getArch()) {
1109  case llvm::Triple::arm:
1110  case llvm::Triple::thumb:
1111    return EmitARMBuiltinExpr(BuiltinID, E);
1112  case llvm::Triple::x86:
1113  case llvm::Triple::x86_64:
1114    return EmitX86BuiltinExpr(BuiltinID, E);
1115  case llvm::Triple::ppc:
1116  case llvm::Triple::ppc64:
1117    return EmitPPCBuiltinExpr(BuiltinID, E);
1118  default:
1119    return 0;
1120  }
1121}
1122
1123static llvm::VectorType *GetNeonType(LLVMContext &C, unsigned type, bool q) {
1124  switch (type) {
1125    default: break;
1126    case 0:
1127    case 5: return llvm::VectorType::get(llvm::Type::getInt8Ty(C), 8 << (int)q);
1128    case 6:
1129    case 7:
1130    case 1: return llvm::VectorType::get(llvm::Type::getInt16Ty(C),4 << (int)q);
1131    case 2: return llvm::VectorType::get(llvm::Type::getInt32Ty(C),2 << (int)q);
1132    case 3: return llvm::VectorType::get(llvm::Type::getInt64Ty(C),1 << (int)q);
1133    case 4: return llvm::VectorType::get(llvm::Type::getFloatTy(C),2 << (int)q);
1134  };
1135  return 0;
1136}
1137
1138Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
1139  unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements();
1140  SmallVector<Constant*, 16> Indices(nElts, C);
1141  Value* SV = llvm::ConstantVector::get(Indices);
1142  return Builder.CreateShuffleVector(V, V, SV, "lane");
1143}
1144
1145Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
1146                                     const char *name,
1147                                     unsigned shift, bool rightshift) {
1148  unsigned j = 0;
1149  for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
1150       ai != ae; ++ai, ++j)
1151    if (shift > 0 && shift == j)
1152      Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
1153    else
1154      Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
1155
1156  return Builder.CreateCall(F, Ops, name);
1157}
1158
1159Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
1160                                            bool neg) {
1161  ConstantInt *CI = cast<ConstantInt>(V);
1162  int SV = CI->getSExtValue();
1163
1164  llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1165  llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV);
1166  SmallVector<llvm::Constant*, 16> CV(VTy->getNumElements(), C);
1167  return llvm::ConstantVector::get(CV);
1168}
1169
1170/// GetPointeeAlignment - Given an expression with a pointer type, find the
1171/// alignment of the type referenced by the pointer.  Skip over implicit
1172/// casts.
1173static Value *GetPointeeAlignment(CodeGenFunction &CGF, const Expr *Addr) {
1174  unsigned Align = 1;
1175  // Check if the type is a pointer.  The implicit cast operand might not be.
1176  while (Addr->getType()->isPointerType()) {
1177    QualType PtTy = Addr->getType()->getPointeeType();
1178    unsigned NewA = CGF.getContext().getTypeAlignInChars(PtTy).getQuantity();
1179    if (NewA > Align)
1180      Align = NewA;
1181
1182    // If the address is an implicit cast, repeat with the cast operand.
1183    if (const ImplicitCastExpr *CastAddr = dyn_cast<ImplicitCastExpr>(Addr)) {
1184      Addr = CastAddr->getSubExpr();
1185      continue;
1186    }
1187    break;
1188  }
1189  return llvm::ConstantInt::get(CGF.Int32Ty, Align);
1190}
1191
1192Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
1193                                           const CallExpr *E) {
1194  if (BuiltinID == ARM::BI__clear_cache) {
1195    const FunctionDecl *FD = E->getDirectCallee();
1196    // Oddly people write this call without args on occasion and gcc accepts
1197    // it - it's also marked as varargs in the description file.
1198    SmallVector<Value*, 2> Ops;
1199    for (unsigned i = 0; i < E->getNumArgs(); i++)
1200      Ops.push_back(EmitScalarExpr(E->getArg(i)));
1201    llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
1202    llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
1203    StringRef Name = FD->getName();
1204    return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
1205  }
1206
1207  if (BuiltinID == ARM::BI__builtin_arm_ldrexd) {
1208    Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
1209
1210    Value *LdPtr = EmitScalarExpr(E->getArg(0));
1211    Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd");
1212
1213    Value *Val0 = Builder.CreateExtractValue(Val, 1);
1214    Value *Val1 = Builder.CreateExtractValue(Val, 0);
1215    Val0 = Builder.CreateZExt(Val0, Int64Ty);
1216    Val1 = Builder.CreateZExt(Val1, Int64Ty);
1217
1218    Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
1219    Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
1220    return Builder.CreateOr(Val, Val1);
1221  }
1222
1223  if (BuiltinID == ARM::BI__builtin_arm_strexd) {
1224    Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd);
1225    llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL);
1226
1227    Value *One = llvm::ConstantInt::get(Int32Ty, 1);
1228    Value *Tmp = Builder.CreateAlloca(Int64Ty, One);
1229    Value *Val = EmitScalarExpr(E->getArg(0));
1230    Builder.CreateStore(Val, Tmp);
1231
1232    Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
1233    Val = Builder.CreateLoad(LdPtr);
1234
1235    Value *Arg0 = Builder.CreateExtractValue(Val, 0);
1236    Value *Arg1 = Builder.CreateExtractValue(Val, 1);
1237    Value *StPtr = EmitScalarExpr(E->getArg(1));
1238    return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd");
1239  }
1240
1241  SmallVector<Value*, 4> Ops;
1242  for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++)
1243    Ops.push_back(EmitScalarExpr(E->getArg(i)));
1244
1245  // vget_lane and vset_lane are not overloaded and do not have an extra
1246  // argument that specifies the vector type.
1247  switch (BuiltinID) {
1248  default: break;
1249  case ARM::BI__builtin_neon_vget_lane_i8:
1250  case ARM::BI__builtin_neon_vget_lane_i16:
1251  case ARM::BI__builtin_neon_vget_lane_i32:
1252  case ARM::BI__builtin_neon_vget_lane_i64:
1253  case ARM::BI__builtin_neon_vget_lane_f32:
1254  case ARM::BI__builtin_neon_vgetq_lane_i8:
1255  case ARM::BI__builtin_neon_vgetq_lane_i16:
1256  case ARM::BI__builtin_neon_vgetq_lane_i32:
1257  case ARM::BI__builtin_neon_vgetq_lane_i64:
1258  case ARM::BI__builtin_neon_vgetq_lane_f32:
1259    return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
1260                                        "vget_lane");
1261  case ARM::BI__builtin_neon_vset_lane_i8:
1262  case ARM::BI__builtin_neon_vset_lane_i16:
1263  case ARM::BI__builtin_neon_vset_lane_i32:
1264  case ARM::BI__builtin_neon_vset_lane_i64:
1265  case ARM::BI__builtin_neon_vset_lane_f32:
1266  case ARM::BI__builtin_neon_vsetq_lane_i8:
1267  case ARM::BI__builtin_neon_vsetq_lane_i16:
1268  case ARM::BI__builtin_neon_vsetq_lane_i32:
1269  case ARM::BI__builtin_neon_vsetq_lane_i64:
1270  case ARM::BI__builtin_neon_vsetq_lane_f32:
1271    Ops.push_back(EmitScalarExpr(E->getArg(2)));
1272    return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
1273  }
1274
1275  // Get the last argument, which specifies the vector type.
1276  llvm::APSInt Result;
1277  const Expr *Arg = E->getArg(E->getNumArgs()-1);
1278  if (!Arg->isIntegerConstantExpr(Result, getContext()))
1279    return 0;
1280
1281  if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
1282      BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
1283    // Determine the overloaded type of this builtin.
1284    llvm::Type *Ty;
1285    if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
1286      Ty = llvm::Type::getFloatTy(getLLVMContext());
1287    else
1288      Ty = llvm::Type::getDoubleTy(getLLVMContext());
1289
1290    // Determine whether this is an unsigned conversion or not.
1291    bool usgn = Result.getZExtValue() == 1;
1292    unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
1293
1294    // Call the appropriate intrinsic.
1295    Function *F = CGM.getIntrinsic(Int, Ty);
1296    return Builder.CreateCall(F, Ops, "vcvtr");
1297  }
1298
1299  // Determine the type of this overloaded NEON intrinsic.
1300  unsigned type = Result.getZExtValue();
1301  bool usgn = type & 0x08;
1302  bool quad = type & 0x10;
1303  bool poly = (type & 0x7) == 5 || (type & 0x7) == 6;
1304  (void)poly;  // Only used in assert()s.
1305  bool rightShift = false;
1306
1307  llvm::VectorType *VTy = GetNeonType(getLLVMContext(), type & 0x7, quad);
1308  llvm::Type *Ty = VTy;
1309  if (!Ty)
1310    return 0;
1311
1312  unsigned Int;
1313  switch (BuiltinID) {
1314  default: return 0;
1315  case ARM::BI__builtin_neon_vabd_v:
1316  case ARM::BI__builtin_neon_vabdq_v:
1317    Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds;
1318    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
1319  case ARM::BI__builtin_neon_vabs_v:
1320  case ARM::BI__builtin_neon_vabsq_v:
1321    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty),
1322                        Ops, "vabs");
1323  case ARM::BI__builtin_neon_vaddhn_v:
1324    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty),
1325                        Ops, "vaddhn");
1326  case ARM::BI__builtin_neon_vcale_v:
1327    std::swap(Ops[0], Ops[1]);
1328  case ARM::BI__builtin_neon_vcage_v: {
1329    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged);
1330    return EmitNeonCall(F, Ops, "vcage");
1331  }
1332  case ARM::BI__builtin_neon_vcaleq_v:
1333    std::swap(Ops[0], Ops[1]);
1334  case ARM::BI__builtin_neon_vcageq_v: {
1335    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq);
1336    return EmitNeonCall(F, Ops, "vcage");
1337  }
1338  case ARM::BI__builtin_neon_vcalt_v:
1339    std::swap(Ops[0], Ops[1]);
1340  case ARM::BI__builtin_neon_vcagt_v: {
1341    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd);
1342    return EmitNeonCall(F, Ops, "vcagt");
1343  }
1344  case ARM::BI__builtin_neon_vcaltq_v:
1345    std::swap(Ops[0], Ops[1]);
1346  case ARM::BI__builtin_neon_vcagtq_v: {
1347    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq);
1348    return EmitNeonCall(F, Ops, "vcagt");
1349  }
1350  case ARM::BI__builtin_neon_vcls_v:
1351  case ARM::BI__builtin_neon_vclsq_v: {
1352    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty);
1353    return EmitNeonCall(F, Ops, "vcls");
1354  }
1355  case ARM::BI__builtin_neon_vclz_v:
1356  case ARM::BI__builtin_neon_vclzq_v: {
1357    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vclz, Ty);
1358    return EmitNeonCall(F, Ops, "vclz");
1359  }
1360  case ARM::BI__builtin_neon_vcnt_v:
1361  case ARM::BI__builtin_neon_vcntq_v: {
1362    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcnt, Ty);
1363    return EmitNeonCall(F, Ops, "vcnt");
1364  }
1365  case ARM::BI__builtin_neon_vcvt_f16_v: {
1366    assert((type & 0x7) == 7 && !quad && "unexpected vcvt_f16_v builtin");
1367    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf);
1368    return EmitNeonCall(F, Ops, "vcvt");
1369  }
1370  case ARM::BI__builtin_neon_vcvt_f32_f16: {
1371    assert((type & 0x7) == 7 && !quad && "unexpected vcvt_f32_f16 builtin");
1372    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp);
1373    return EmitNeonCall(F, Ops, "vcvt");
1374  }
1375  case ARM::BI__builtin_neon_vcvt_f32_v:
1376  case ARM::BI__builtin_neon_vcvtq_f32_v: {
1377    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1378    Ty = GetNeonType(getLLVMContext(), 4, quad);
1379    return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
1380                : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
1381  }
1382  case ARM::BI__builtin_neon_vcvt_s32_v:
1383  case ARM::BI__builtin_neon_vcvt_u32_v:
1384  case ARM::BI__builtin_neon_vcvtq_s32_v:
1385  case ARM::BI__builtin_neon_vcvtq_u32_v: {
1386    Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(getLLVMContext(), 4, quad));
1387    return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
1388                : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
1389  }
1390  case ARM::BI__builtin_neon_vcvt_n_f32_v:
1391  case ARM::BI__builtin_neon_vcvtq_n_f32_v: {
1392    llvm::Type *Tys[2] = { GetNeonType(getLLVMContext(), 4, quad), Ty };
1393    Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp : Intrinsic::arm_neon_vcvtfxs2fp;
1394    Function *F = CGM.getIntrinsic(Int, Tys);
1395    return EmitNeonCall(F, Ops, "vcvt_n");
1396  }
1397  case ARM::BI__builtin_neon_vcvt_n_s32_v:
1398  case ARM::BI__builtin_neon_vcvt_n_u32_v:
1399  case ARM::BI__builtin_neon_vcvtq_n_s32_v:
1400  case ARM::BI__builtin_neon_vcvtq_n_u32_v: {
1401    llvm::Type *Tys[2] = { Ty, GetNeonType(getLLVMContext(), 4, quad) };
1402    Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu : Intrinsic::arm_neon_vcvtfp2fxs;
1403    Function *F = CGM.getIntrinsic(Int, Tys);
1404    return EmitNeonCall(F, Ops, "vcvt_n");
1405  }
1406  case ARM::BI__builtin_neon_vext_v:
1407  case ARM::BI__builtin_neon_vextq_v: {
1408    int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
1409    SmallVector<Constant*, 16> Indices;
1410    for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
1411      Indices.push_back(ConstantInt::get(Int32Ty, i+CV));
1412
1413    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1414    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1415    Value *SV = llvm::ConstantVector::get(Indices);
1416    return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext");
1417  }
1418  case ARM::BI__builtin_neon_vhadd_v:
1419  case ARM::BI__builtin_neon_vhaddq_v:
1420    Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds;
1421    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd");
1422  case ARM::BI__builtin_neon_vhsub_v:
1423  case ARM::BI__builtin_neon_vhsubq_v:
1424    Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs;
1425    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub");
1426  case ARM::BI__builtin_neon_vld1_v:
1427  case ARM::BI__builtin_neon_vld1q_v:
1428    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1429    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty),
1430                        Ops, "vld1");
1431  case ARM::BI__builtin_neon_vld1_lane_v:
1432  case ARM::BI__builtin_neon_vld1q_lane_v:
1433    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1434    Ty = llvm::PointerType::getUnqual(VTy->getElementType());
1435    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1436    Ops[0] = Builder.CreateLoad(Ops[0]);
1437    return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
1438  case ARM::BI__builtin_neon_vld1_dup_v:
1439  case ARM::BI__builtin_neon_vld1q_dup_v: {
1440    Value *V = UndefValue::get(Ty);
1441    Ty = llvm::PointerType::getUnqual(VTy->getElementType());
1442    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1443    Ops[0] = Builder.CreateLoad(Ops[0]);
1444    llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
1445    Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
1446    return EmitNeonSplat(Ops[0], CI);
1447  }
1448  case ARM::BI__builtin_neon_vld2_v:
1449  case ARM::BI__builtin_neon_vld2q_v: {
1450    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty);
1451    Value *Align = GetPointeeAlignment(*this, E->getArg(1));
1452    Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2");
1453    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1454    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1455    return Builder.CreateStore(Ops[1], Ops[0]);
1456  }
1457  case ARM::BI__builtin_neon_vld3_v:
1458  case ARM::BI__builtin_neon_vld3q_v: {
1459    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty);
1460    Value *Align = GetPointeeAlignment(*this, E->getArg(1));
1461    Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3");
1462    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1463    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1464    return Builder.CreateStore(Ops[1], Ops[0]);
1465  }
1466  case ARM::BI__builtin_neon_vld4_v:
1467  case ARM::BI__builtin_neon_vld4q_v: {
1468    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty);
1469    Value *Align = GetPointeeAlignment(*this, E->getArg(1));
1470    Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4");
1471    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1472    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1473    return Builder.CreateStore(Ops[1], Ops[0]);
1474  }
1475  case ARM::BI__builtin_neon_vld2_lane_v:
1476  case ARM::BI__builtin_neon_vld2q_lane_v: {
1477    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty);
1478    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1479    Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
1480    Ops.push_back(GetPointeeAlignment(*this, E->getArg(1)));
1481    Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
1482    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1483    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1484    return Builder.CreateStore(Ops[1], Ops[0]);
1485  }
1486  case ARM::BI__builtin_neon_vld3_lane_v:
1487  case ARM::BI__builtin_neon_vld3q_lane_v: {
1488    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty);
1489    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1490    Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
1491    Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
1492    Ops.push_back(GetPointeeAlignment(*this, E->getArg(1)));
1493    Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
1494    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1495    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1496    return Builder.CreateStore(Ops[1], Ops[0]);
1497  }
1498  case ARM::BI__builtin_neon_vld4_lane_v:
1499  case ARM::BI__builtin_neon_vld4q_lane_v: {
1500    Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty);
1501    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1502    Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
1503    Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
1504    Ops[5] = Builder.CreateBitCast(Ops[5], Ty);
1505    Ops.push_back(GetPointeeAlignment(*this, E->getArg(1)));
1506    Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
1507    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1508    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1509    return Builder.CreateStore(Ops[1], Ops[0]);
1510  }
1511  case ARM::BI__builtin_neon_vld2_dup_v:
1512  case ARM::BI__builtin_neon_vld3_dup_v:
1513  case ARM::BI__builtin_neon_vld4_dup_v: {
1514    // Handle 64-bit elements as a special-case.  There is no "dup" needed.
1515    if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
1516      switch (BuiltinID) {
1517      case ARM::BI__builtin_neon_vld2_dup_v:
1518        Int = Intrinsic::arm_neon_vld2;
1519        break;
1520      case ARM::BI__builtin_neon_vld3_dup_v:
1521        Int = Intrinsic::arm_neon_vld2;
1522        break;
1523      case ARM::BI__builtin_neon_vld4_dup_v:
1524        Int = Intrinsic::arm_neon_vld2;
1525        break;
1526      default: llvm_unreachable("unknown vld_dup intrinsic?");
1527      }
1528      Function *F = CGM.getIntrinsic(Int, Ty);
1529      Value *Align = GetPointeeAlignment(*this, E->getArg(1));
1530      Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup");
1531      Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1532      Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1533      return Builder.CreateStore(Ops[1], Ops[0]);
1534    }
1535    switch (BuiltinID) {
1536    case ARM::BI__builtin_neon_vld2_dup_v:
1537      Int = Intrinsic::arm_neon_vld2lane;
1538      break;
1539    case ARM::BI__builtin_neon_vld3_dup_v:
1540      Int = Intrinsic::arm_neon_vld2lane;
1541      break;
1542    case ARM::BI__builtin_neon_vld4_dup_v:
1543      Int = Intrinsic::arm_neon_vld2lane;
1544      break;
1545    default: llvm_unreachable("unknown vld_dup intrinsic?");
1546    }
1547    Function *F = CGM.getIntrinsic(Int, Ty);
1548    llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
1549
1550    SmallVector<Value*, 6> Args;
1551    Args.push_back(Ops[1]);
1552    Args.append(STy->getNumElements(), UndefValue::get(Ty));
1553
1554    llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
1555    Args.push_back(CI);
1556    Args.push_back(GetPointeeAlignment(*this, E->getArg(1)));
1557
1558    Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
1559    // splat lane 0 to all elts in each vector of the result.
1560    for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1561      Value *Val = Builder.CreateExtractValue(Ops[1], i);
1562      Value *Elt = Builder.CreateBitCast(Val, Ty);
1563      Elt = EmitNeonSplat(Elt, CI);
1564      Elt = Builder.CreateBitCast(Elt, Val->getType());
1565      Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
1566    }
1567    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1568    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1569    return Builder.CreateStore(Ops[1], Ops[0]);
1570  }
1571  case ARM::BI__builtin_neon_vmax_v:
1572  case ARM::BI__builtin_neon_vmaxq_v:
1573    Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs;
1574    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
1575  case ARM::BI__builtin_neon_vmin_v:
1576  case ARM::BI__builtin_neon_vminq_v:
1577    Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins;
1578    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
1579  case ARM::BI__builtin_neon_vmovl_v: {
1580    llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
1581    Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
1582    if (usgn)
1583      return Builder.CreateZExt(Ops[0], Ty, "vmovl");
1584    return Builder.CreateSExt(Ops[0], Ty, "vmovl");
1585  }
1586  case ARM::BI__builtin_neon_vmovn_v: {
1587    llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
1588    Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
1589    return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
1590  }
1591  case ARM::BI__builtin_neon_vmul_v:
1592  case ARM::BI__builtin_neon_vmulq_v:
1593    assert(poly && "vmul builtin only supported for polynomial types");
1594    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty),
1595                        Ops, "vmul");
1596  case ARM::BI__builtin_neon_vmull_v:
1597    Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
1598    Int = poly ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
1599    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
1600  case ARM::BI__builtin_neon_vpadal_v:
1601  case ARM::BI__builtin_neon_vpadalq_v: {
1602    Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals;
1603    // The source operand type has twice as many elements of half the size.
1604    unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
1605    llvm::Type *EltTy =
1606      llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
1607    llvm::Type *NarrowTy =
1608      llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
1609    llvm::Type *Tys[2] = { Ty, NarrowTy };
1610    return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal");
1611  }
1612  case ARM::BI__builtin_neon_vpadd_v:
1613    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty),
1614                        Ops, "vpadd");
1615  case ARM::BI__builtin_neon_vpaddl_v:
1616  case ARM::BI__builtin_neon_vpaddlq_v: {
1617    Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls;
1618    // The source operand type has twice as many elements of half the size.
1619    unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
1620    llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
1621    llvm::Type *NarrowTy =
1622      llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
1623    llvm::Type *Tys[2] = { Ty, NarrowTy };
1624    return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
1625  }
1626  case ARM::BI__builtin_neon_vpmax_v:
1627    Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs;
1628    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
1629  case ARM::BI__builtin_neon_vpmin_v:
1630    Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins;
1631    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
1632  case ARM::BI__builtin_neon_vqabs_v:
1633  case ARM::BI__builtin_neon_vqabsq_v:
1634    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty),
1635                        Ops, "vqabs");
1636  case ARM::BI__builtin_neon_vqadd_v:
1637  case ARM::BI__builtin_neon_vqaddq_v:
1638    Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds;
1639    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd");
1640  case ARM::BI__builtin_neon_vqdmlal_v:
1641    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty),
1642                        Ops, "vqdmlal");
1643  case ARM::BI__builtin_neon_vqdmlsl_v:
1644    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty),
1645                        Ops, "vqdmlsl");
1646  case ARM::BI__builtin_neon_vqdmulh_v:
1647  case ARM::BI__builtin_neon_vqdmulhq_v:
1648    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty),
1649                        Ops, "vqdmulh");
1650  case ARM::BI__builtin_neon_vqdmull_v:
1651    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty),
1652                        Ops, "vqdmull");
1653  case ARM::BI__builtin_neon_vqmovn_v:
1654    Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns;
1655    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn");
1656  case ARM::BI__builtin_neon_vqmovun_v:
1657    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty),
1658                        Ops, "vqdmull");
1659  case ARM::BI__builtin_neon_vqneg_v:
1660  case ARM::BI__builtin_neon_vqnegq_v:
1661    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty),
1662                        Ops, "vqneg");
1663  case ARM::BI__builtin_neon_vqrdmulh_v:
1664  case ARM::BI__builtin_neon_vqrdmulhq_v:
1665    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty),
1666                        Ops, "vqrdmulh");
1667  case ARM::BI__builtin_neon_vqrshl_v:
1668  case ARM::BI__builtin_neon_vqrshlq_v:
1669    Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts;
1670    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl");
1671  case ARM::BI__builtin_neon_vqrshrn_n_v:
1672    Int = usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
1673    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
1674                        1, true);
1675  case ARM::BI__builtin_neon_vqrshrun_n_v:
1676    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
1677                        Ops, "vqrshrun_n", 1, true);
1678  case ARM::BI__builtin_neon_vqshl_v:
1679  case ARM::BI__builtin_neon_vqshlq_v:
1680    Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
1681    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl");
1682  case ARM::BI__builtin_neon_vqshl_n_v:
1683  case ARM::BI__builtin_neon_vqshlq_n_v:
1684    Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
1685    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
1686                        1, false);
1687  case ARM::BI__builtin_neon_vqshlu_n_v:
1688  case ARM::BI__builtin_neon_vqshluq_n_v:
1689    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty),
1690                        Ops, "vqshlu", 1, false);
1691  case ARM::BI__builtin_neon_vqshrn_n_v:
1692    Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
1693    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
1694                        1, true);
1695  case ARM::BI__builtin_neon_vqshrun_n_v:
1696    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
1697                        Ops, "vqshrun_n", 1, true);
1698  case ARM::BI__builtin_neon_vqsub_v:
1699  case ARM::BI__builtin_neon_vqsubq_v:
1700    Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs;
1701    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub");
1702  case ARM::BI__builtin_neon_vraddhn_v:
1703    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty),
1704                        Ops, "vraddhn");
1705  case ARM::BI__builtin_neon_vrecpe_v:
1706  case ARM::BI__builtin_neon_vrecpeq_v:
1707    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
1708                        Ops, "vrecpe");
1709  case ARM::BI__builtin_neon_vrecps_v:
1710  case ARM::BI__builtin_neon_vrecpsq_v:
1711    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty),
1712                        Ops, "vrecps");
1713  case ARM::BI__builtin_neon_vrhadd_v:
1714  case ARM::BI__builtin_neon_vrhaddq_v:
1715    Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds;
1716    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd");
1717  case ARM::BI__builtin_neon_vrshl_v:
1718  case ARM::BI__builtin_neon_vrshlq_v:
1719    Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
1720    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl");
1721  case ARM::BI__builtin_neon_vrshrn_n_v:
1722    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
1723                        Ops, "vrshrn_n", 1, true);
1724  case ARM::BI__builtin_neon_vrshr_n_v:
1725  case ARM::BI__builtin_neon_vrshrq_n_v:
1726    Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
1727    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true);
1728  case ARM::BI__builtin_neon_vrsqrte_v:
1729  case ARM::BI__builtin_neon_vrsqrteq_v:
1730    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty),
1731                        Ops, "vrsqrte");
1732  case ARM::BI__builtin_neon_vrsqrts_v:
1733  case ARM::BI__builtin_neon_vrsqrtsq_v:
1734    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty),
1735                        Ops, "vrsqrts");
1736  case ARM::BI__builtin_neon_vrsra_n_v:
1737  case ARM::BI__builtin_neon_vrsraq_n_v:
1738    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1739    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1740    Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
1741    Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
1742    Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]);
1743    return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
1744  case ARM::BI__builtin_neon_vrsubhn_v:
1745    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty),
1746                        Ops, "vrsubhn");
1747  case ARM::BI__builtin_neon_vshl_v:
1748  case ARM::BI__builtin_neon_vshlq_v:
1749    Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts;
1750    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl");
1751  case ARM::BI__builtin_neon_vshll_n_v:
1752    Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls;
1753    return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1);
1754  case ARM::BI__builtin_neon_vshl_n_v:
1755  case ARM::BI__builtin_neon_vshlq_n_v:
1756    Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
1757    return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], "vshl_n");
1758  case ARM::BI__builtin_neon_vshrn_n_v:
1759    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty),
1760                        Ops, "vshrn_n", 1, true);
1761  case ARM::BI__builtin_neon_vshr_n_v:
1762  case ARM::BI__builtin_neon_vshrq_n_v:
1763    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1764    Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
1765    if (usgn)
1766      return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n");
1767    else
1768      return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n");
1769  case ARM::BI__builtin_neon_vsri_n_v:
1770  case ARM::BI__builtin_neon_vsriq_n_v:
1771    rightShift = true;
1772  case ARM::BI__builtin_neon_vsli_n_v:
1773  case ARM::BI__builtin_neon_vsliq_n_v:
1774    Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
1775    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
1776                        Ops, "vsli_n");
1777  case ARM::BI__builtin_neon_vsra_n_v:
1778  case ARM::BI__builtin_neon_vsraq_n_v:
1779    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1780    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1781    Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false);
1782    if (usgn)
1783      Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n");
1784    else
1785      Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n");
1786    return Builder.CreateAdd(Ops[0], Ops[1]);
1787  case ARM::BI__builtin_neon_vst1_v:
1788  case ARM::BI__builtin_neon_vst1q_v:
1789    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1790    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty),
1791                        Ops, "");
1792  case ARM::BI__builtin_neon_vst1_lane_v:
1793  case ARM::BI__builtin_neon_vst1q_lane_v:
1794    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1795    Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
1796    Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1797    return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty));
1798  case ARM::BI__builtin_neon_vst2_v:
1799  case ARM::BI__builtin_neon_vst2q_v:
1800    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1801    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty),
1802                        Ops, "");
1803  case ARM::BI__builtin_neon_vst2_lane_v:
1804  case ARM::BI__builtin_neon_vst2q_lane_v:
1805    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1806    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty),
1807                        Ops, "");
1808  case ARM::BI__builtin_neon_vst3_v:
1809  case ARM::BI__builtin_neon_vst3q_v:
1810    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1811    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty),
1812                        Ops, "");
1813  case ARM::BI__builtin_neon_vst3_lane_v:
1814  case ARM::BI__builtin_neon_vst3q_lane_v:
1815    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1816    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty),
1817                        Ops, "");
1818  case ARM::BI__builtin_neon_vst4_v:
1819  case ARM::BI__builtin_neon_vst4q_v:
1820    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1821    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty),
1822                        Ops, "");
1823  case ARM::BI__builtin_neon_vst4_lane_v:
1824  case ARM::BI__builtin_neon_vst4q_lane_v:
1825    Ops.push_back(GetPointeeAlignment(*this, E->getArg(0)));
1826    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty),
1827                        Ops, "");
1828  case ARM::BI__builtin_neon_vsubhn_v:
1829    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty),
1830                        Ops, "vsubhn");
1831  case ARM::BI__builtin_neon_vtbl1_v:
1832    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
1833                        Ops, "vtbl1");
1834  case ARM::BI__builtin_neon_vtbl2_v:
1835    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
1836                        Ops, "vtbl2");
1837  case ARM::BI__builtin_neon_vtbl3_v:
1838    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
1839                        Ops, "vtbl3");
1840  case ARM::BI__builtin_neon_vtbl4_v:
1841    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
1842                        Ops, "vtbl4");
1843  case ARM::BI__builtin_neon_vtbx1_v:
1844    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
1845                        Ops, "vtbx1");
1846  case ARM::BI__builtin_neon_vtbx2_v:
1847    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
1848                        Ops, "vtbx2");
1849  case ARM::BI__builtin_neon_vtbx3_v:
1850    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
1851                        Ops, "vtbx3");
1852  case ARM::BI__builtin_neon_vtbx4_v:
1853    return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
1854                        Ops, "vtbx4");
1855  case ARM::BI__builtin_neon_vtst_v:
1856  case ARM::BI__builtin_neon_vtstq_v: {
1857    Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1858    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1859    Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
1860    Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
1861                                ConstantAggregateZero::get(Ty));
1862    return Builder.CreateSExt(Ops[0], Ty, "vtst");
1863  }
1864  case ARM::BI__builtin_neon_vtrn_v:
1865  case ARM::BI__builtin_neon_vtrnq_v: {
1866    Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
1867    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1868    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1869    Value *SV = 0;
1870
1871    for (unsigned vi = 0; vi != 2; ++vi) {
1872      SmallVector<Constant*, 16> Indices;
1873      for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
1874        Indices.push_back(ConstantInt::get(Int32Ty, i+vi));
1875        Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi));
1876      }
1877      Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
1878      SV = llvm::ConstantVector::get(Indices);
1879      SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
1880      SV = Builder.CreateStore(SV, Addr);
1881    }
1882    return SV;
1883  }
1884  case ARM::BI__builtin_neon_vuzp_v:
1885  case ARM::BI__builtin_neon_vuzpq_v: {
1886    Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
1887    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1888    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1889    Value *SV = 0;
1890
1891    for (unsigned vi = 0; vi != 2; ++vi) {
1892      SmallVector<Constant*, 16> Indices;
1893      for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
1894        Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
1895
1896      Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
1897      SV = llvm::ConstantVector::get(Indices);
1898      SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
1899      SV = Builder.CreateStore(SV, Addr);
1900    }
1901    return SV;
1902  }
1903  case ARM::BI__builtin_neon_vzip_v:
1904  case ARM::BI__builtin_neon_vzipq_v: {
1905    Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
1906    Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1907    Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1908    Value *SV = 0;
1909
1910    for (unsigned vi = 0; vi != 2; ++vi) {
1911      SmallVector<Constant*, 16> Indices;
1912      for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
1913        Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
1914        Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
1915      }
1916      Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
1917      SV = llvm::ConstantVector::get(Indices);
1918      SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
1919      SV = Builder.CreateStore(SV, Addr);
1920    }
1921    return SV;
1922  }
1923  }
1924}
1925
1926llvm::Value *CodeGenFunction::
1927BuildVector(const SmallVectorImpl<llvm::Value*> &Ops) {
1928  assert((Ops.size() & (Ops.size() - 1)) == 0 &&
1929         "Not a power-of-two sized vector!");
1930  bool AllConstants = true;
1931  for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
1932    AllConstants &= isa<Constant>(Ops[i]);
1933
1934  // If this is a constant vector, create a ConstantVector.
1935  if (AllConstants) {
1936    std::vector<llvm::Constant*> CstOps;
1937    for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1938      CstOps.push_back(cast<Constant>(Ops[i]));
1939    return llvm::ConstantVector::get(CstOps);
1940  }
1941
1942  // Otherwise, insertelement the values to build the vector.
1943  Value *Result =
1944    llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
1945
1946  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1947    Result = Builder.CreateInsertElement(Result, Ops[i],
1948               llvm::ConstantInt::get(llvm::Type::getInt32Ty(getLLVMContext()), i));
1949
1950  return Result;
1951}
1952
1953Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
1954                                           const CallExpr *E) {
1955  SmallVector<Value*, 4> Ops;
1956
1957  // Find out if any arguments are required to be integer constant expressions.
1958  unsigned ICEArguments = 0;
1959  ASTContext::GetBuiltinTypeError Error;
1960  getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1961  assert(Error == ASTContext::GE_None && "Should not codegen an error");
1962
1963  for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
1964    // If this is a normal argument, just emit it as a scalar.
1965    if ((ICEArguments & (1 << i)) == 0) {
1966      Ops.push_back(EmitScalarExpr(E->getArg(i)));
1967      continue;
1968    }
1969
1970    // If this is required to be a constant, constant fold it so that we know
1971    // that the generated intrinsic gets a ConstantInt.
1972    llvm::APSInt Result;
1973    bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
1974    assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
1975    Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
1976  }
1977
1978  switch (BuiltinID) {
1979  default: return 0;
1980  case X86::BI__builtin_ia32_pslldi128:
1981  case X86::BI__builtin_ia32_psllqi128:
1982  case X86::BI__builtin_ia32_psllwi128:
1983  case X86::BI__builtin_ia32_psradi128:
1984  case X86::BI__builtin_ia32_psrawi128:
1985  case X86::BI__builtin_ia32_psrldi128:
1986  case X86::BI__builtin_ia32_psrlqi128:
1987  case X86::BI__builtin_ia32_psrlwi128: {
1988    Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext");
1989    llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2);
1990    llvm::Value *Zero = llvm::ConstantInt::get(Int32Ty, 0);
1991    Ops[1] = Builder.CreateInsertElement(llvm::UndefValue::get(Ty),
1992                                         Ops[1], Zero, "insert");
1993    Ops[1] = Builder.CreateBitCast(Ops[1], Ops[0]->getType(), "bitcast");
1994    const char *name = 0;
1995    Intrinsic::ID ID = Intrinsic::not_intrinsic;
1996
1997    switch (BuiltinID) {
1998    default: llvm_unreachable("Unsupported shift intrinsic!");
1999    case X86::BI__builtin_ia32_pslldi128:
2000      name = "pslldi";
2001      ID = Intrinsic::x86_sse2_psll_d;
2002      break;
2003    case X86::BI__builtin_ia32_psllqi128:
2004      name = "psllqi";
2005      ID = Intrinsic::x86_sse2_psll_q;
2006      break;
2007    case X86::BI__builtin_ia32_psllwi128:
2008      name = "psllwi";
2009      ID = Intrinsic::x86_sse2_psll_w;
2010      break;
2011    case X86::BI__builtin_ia32_psradi128:
2012      name = "psradi";
2013      ID = Intrinsic::x86_sse2_psra_d;
2014      break;
2015    case X86::BI__builtin_ia32_psrawi128:
2016      name = "psrawi";
2017      ID = Intrinsic::x86_sse2_psra_w;
2018      break;
2019    case X86::BI__builtin_ia32_psrldi128:
2020      name = "psrldi";
2021      ID = Intrinsic::x86_sse2_psrl_d;
2022      break;
2023    case X86::BI__builtin_ia32_psrlqi128:
2024      name = "psrlqi";
2025      ID = Intrinsic::x86_sse2_psrl_q;
2026      break;
2027    case X86::BI__builtin_ia32_psrlwi128:
2028      name = "psrlwi";
2029      ID = Intrinsic::x86_sse2_psrl_w;
2030      break;
2031    }
2032    llvm::Function *F = CGM.getIntrinsic(ID);
2033    return Builder.CreateCall(F, Ops, name);
2034  }
2035  case X86::BI__builtin_ia32_vec_init_v8qi:
2036  case X86::BI__builtin_ia32_vec_init_v4hi:
2037  case X86::BI__builtin_ia32_vec_init_v2si:
2038    return Builder.CreateBitCast(BuildVector(Ops),
2039                                 llvm::Type::getX86_MMXTy(getLLVMContext()));
2040  case X86::BI__builtin_ia32_vec_ext_v2si:
2041    return Builder.CreateExtractElement(Ops[0],
2042                                  llvm::ConstantInt::get(Ops[1]->getType(), 0));
2043  case X86::BI__builtin_ia32_pslldi:
2044  case X86::BI__builtin_ia32_psllqi:
2045  case X86::BI__builtin_ia32_psllwi:
2046  case X86::BI__builtin_ia32_psradi:
2047  case X86::BI__builtin_ia32_psrawi:
2048  case X86::BI__builtin_ia32_psrldi:
2049  case X86::BI__builtin_ia32_psrlqi:
2050  case X86::BI__builtin_ia32_psrlwi: {
2051    Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext");
2052    llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 1);
2053    Ops[1] = Builder.CreateBitCast(Ops[1], Ty, "bitcast");
2054    const char *name = 0;
2055    Intrinsic::ID ID = Intrinsic::not_intrinsic;
2056
2057    switch (BuiltinID) {
2058    default: llvm_unreachable("Unsupported shift intrinsic!");
2059    case X86::BI__builtin_ia32_pslldi:
2060      name = "pslldi";
2061      ID = Intrinsic::x86_mmx_psll_d;
2062      break;
2063    case X86::BI__builtin_ia32_psllqi:
2064      name = "psllqi";
2065      ID = Intrinsic::x86_mmx_psll_q;
2066      break;
2067    case X86::BI__builtin_ia32_psllwi:
2068      name = "psllwi";
2069      ID = Intrinsic::x86_mmx_psll_w;
2070      break;
2071    case X86::BI__builtin_ia32_psradi:
2072      name = "psradi";
2073      ID = Intrinsic::x86_mmx_psra_d;
2074      break;
2075    case X86::BI__builtin_ia32_psrawi:
2076      name = "psrawi";
2077      ID = Intrinsic::x86_mmx_psra_w;
2078      break;
2079    case X86::BI__builtin_ia32_psrldi:
2080      name = "psrldi";
2081      ID = Intrinsic::x86_mmx_psrl_d;
2082      break;
2083    case X86::BI__builtin_ia32_psrlqi:
2084      name = "psrlqi";
2085      ID = Intrinsic::x86_mmx_psrl_q;
2086      break;
2087    case X86::BI__builtin_ia32_psrlwi:
2088      name = "psrlwi";
2089      ID = Intrinsic::x86_mmx_psrl_w;
2090      break;
2091    }
2092    llvm::Function *F = CGM.getIntrinsic(ID);
2093    return Builder.CreateCall(F, Ops, name);
2094  }
2095  case X86::BI__builtin_ia32_cmpps: {
2096    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps);
2097    return Builder.CreateCall(F, Ops, "cmpps");
2098  }
2099  case X86::BI__builtin_ia32_cmpss: {
2100    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss);
2101    return Builder.CreateCall(F, Ops, "cmpss");
2102  }
2103  case X86::BI__builtin_ia32_ldmxcsr: {
2104    llvm::Type *PtrTy = Int8PtrTy;
2105    Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2106    Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2107    Builder.CreateStore(Ops[0], Tmp);
2108    return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
2109                              Builder.CreateBitCast(Tmp, PtrTy));
2110  }
2111  case X86::BI__builtin_ia32_stmxcsr: {
2112    llvm::Type *PtrTy = Int8PtrTy;
2113    Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2114    Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2115    Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
2116                       Builder.CreateBitCast(Tmp, PtrTy));
2117    return Builder.CreateLoad(Tmp, "stmxcsr");
2118  }
2119  case X86::BI__builtin_ia32_cmppd: {
2120    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd);
2121    return Builder.CreateCall(F, Ops, "cmppd");
2122  }
2123  case X86::BI__builtin_ia32_cmpsd: {
2124    llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd);
2125    return Builder.CreateCall(F, Ops, "cmpsd");
2126  }
2127  case X86::BI__builtin_ia32_storehps:
2128  case X86::BI__builtin_ia32_storelps: {
2129    llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty);
2130    llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2131
2132    // cast val v2i64
2133    Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast");
2134
2135    // extract (0, 1)
2136    unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
2137    llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index);
2138    Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
2139
2140    // cast pointer to i64 & store
2141    Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
2142    return Builder.CreateStore(Ops[1], Ops[0]);
2143  }
2144  case X86::BI__builtin_ia32_palignr: {
2145    unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2146
2147    // If palignr is shifting the pair of input vectors less than 9 bytes,
2148    // emit a shuffle instruction.
2149    if (shiftVal <= 8) {
2150      SmallVector<llvm::Constant*, 8> Indices;
2151      for (unsigned i = 0; i != 8; ++i)
2152        Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2153
2154      Value* SV = llvm::ConstantVector::get(Indices);
2155      return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2156    }
2157
2158    // If palignr is shifting the pair of input vectors more than 8 but less
2159    // than 16 bytes, emit a logical right shift of the destination.
2160    if (shiftVal < 16) {
2161      // MMX has these as 1 x i64 vectors for some odd optimization reasons.
2162      llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1);
2163
2164      Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2165      Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8);
2166
2167      // create i32 constant
2168      llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q);
2169      return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2170    }
2171
2172    // If palignr is shifting the pair of vectors more than 16 bytes, emit zero.
2173    return llvm::Constant::getNullValue(ConvertType(E->getType()));
2174  }
2175  case X86::BI__builtin_ia32_palignr128: {
2176    unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2177
2178    // If palignr is shifting the pair of input vectors less than 17 bytes,
2179    // emit a shuffle instruction.
2180    if (shiftVal <= 16) {
2181      SmallVector<llvm::Constant*, 16> Indices;
2182      for (unsigned i = 0; i != 16; ++i)
2183        Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2184
2185      Value* SV = llvm::ConstantVector::get(Indices);
2186      return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2187    }
2188
2189    // If palignr is shifting the pair of input vectors more than 16 but less
2190    // than 32 bytes, emit a logical right shift of the destination.
2191    if (shiftVal < 32) {
2192      llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2193
2194      Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2195      Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
2196
2197      // create i32 constant
2198      llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq);
2199      return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2200    }
2201
2202    // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
2203    return llvm::Constant::getNullValue(ConvertType(E->getType()));
2204  }
2205  case X86::BI__builtin_ia32_movntps:
2206  case X86::BI__builtin_ia32_movntpd:
2207  case X86::BI__builtin_ia32_movntdq:
2208  case X86::BI__builtin_ia32_movnti: {
2209    llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(),
2210                                           Builder.getInt32(1));
2211
2212    // Convert the type of the pointer to a pointer to the stored type.
2213    Value *BC = Builder.CreateBitCast(Ops[0],
2214                                llvm::PointerType::getUnqual(Ops[1]->getType()),
2215                                      "cast");
2216    StoreInst *SI = Builder.CreateStore(Ops[1], BC);
2217    SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
2218    SI->setAlignment(16);
2219    return SI;
2220  }
2221  // 3DNow!
2222  case X86::BI__builtin_ia32_pavgusb:
2223  case X86::BI__builtin_ia32_pf2id:
2224  case X86::BI__builtin_ia32_pfacc:
2225  case X86::BI__builtin_ia32_pfadd:
2226  case X86::BI__builtin_ia32_pfcmpeq:
2227  case X86::BI__builtin_ia32_pfcmpge:
2228  case X86::BI__builtin_ia32_pfcmpgt:
2229  case X86::BI__builtin_ia32_pfmax:
2230  case X86::BI__builtin_ia32_pfmin:
2231  case X86::BI__builtin_ia32_pfmul:
2232  case X86::BI__builtin_ia32_pfrcp:
2233  case X86::BI__builtin_ia32_pfrcpit1:
2234  case X86::BI__builtin_ia32_pfrcpit2:
2235  case X86::BI__builtin_ia32_pfrsqrt:
2236  case X86::BI__builtin_ia32_pfrsqit1:
2237  case X86::BI__builtin_ia32_pfrsqrtit1:
2238  case X86::BI__builtin_ia32_pfsub:
2239  case X86::BI__builtin_ia32_pfsubr:
2240  case X86::BI__builtin_ia32_pi2fd:
2241  case X86::BI__builtin_ia32_pmulhrw:
2242  case X86::BI__builtin_ia32_pf2iw:
2243  case X86::BI__builtin_ia32_pfnacc:
2244  case X86::BI__builtin_ia32_pfpnacc:
2245  case X86::BI__builtin_ia32_pi2fw:
2246  case X86::BI__builtin_ia32_pswapdsf:
2247  case X86::BI__builtin_ia32_pswapdsi: {
2248    const char *name = 0;
2249    Intrinsic::ID ID = Intrinsic::not_intrinsic;
2250    switch(BuiltinID) {
2251    case X86::BI__builtin_ia32_pavgusb:
2252      name = "pavgusb";
2253      ID = Intrinsic::x86_3dnow_pavgusb;
2254      break;
2255    case X86::BI__builtin_ia32_pf2id:
2256      name = "pf2id";
2257      ID = Intrinsic::x86_3dnow_pf2id;
2258      break;
2259    case X86::BI__builtin_ia32_pfacc:
2260      name = "pfacc";
2261      ID = Intrinsic::x86_3dnow_pfacc;
2262      break;
2263    case X86::BI__builtin_ia32_pfadd:
2264      name = "pfadd";
2265      ID = Intrinsic::x86_3dnow_pfadd;
2266      break;
2267    case X86::BI__builtin_ia32_pfcmpeq:
2268      name = "pfcmpeq";
2269      ID = Intrinsic::x86_3dnow_pfcmpeq;
2270      break;
2271    case X86::BI__builtin_ia32_pfcmpge:
2272      name = "pfcmpge";
2273      ID = Intrinsic::x86_3dnow_pfcmpge;
2274      break;
2275    case X86::BI__builtin_ia32_pfcmpgt:
2276      name = "pfcmpgt";
2277      ID = Intrinsic::x86_3dnow_pfcmpgt;
2278      break;
2279    case X86::BI__builtin_ia32_pfmax:
2280      name = "pfmax";
2281      ID = Intrinsic::x86_3dnow_pfmax;
2282      break;
2283    case X86::BI__builtin_ia32_pfmin:
2284      name = "pfmin";
2285      ID = Intrinsic::x86_3dnow_pfmin;
2286      break;
2287    case X86::BI__builtin_ia32_pfmul:
2288      name = "pfmul";
2289      ID = Intrinsic::x86_3dnow_pfmul;
2290      break;
2291    case X86::BI__builtin_ia32_pfrcp:
2292      name = "pfrcp";
2293      ID = Intrinsic::x86_3dnow_pfrcp;
2294      break;
2295    case X86::BI__builtin_ia32_pfrcpit1:
2296      name = "pfrcpit1";
2297      ID = Intrinsic::x86_3dnow_pfrcpit1;
2298      break;
2299    case X86::BI__builtin_ia32_pfrcpit2:
2300      name = "pfrcpit2";
2301      ID = Intrinsic::x86_3dnow_pfrcpit2;
2302      break;
2303    case X86::BI__builtin_ia32_pfrsqrt:
2304      name = "pfrsqrt";
2305      ID = Intrinsic::x86_3dnow_pfrsqrt;
2306      break;
2307    case X86::BI__builtin_ia32_pfrsqit1:
2308    case X86::BI__builtin_ia32_pfrsqrtit1:
2309      name = "pfrsqit1";
2310      ID = Intrinsic::x86_3dnow_pfrsqit1;
2311      break;
2312    case X86::BI__builtin_ia32_pfsub:
2313      name = "pfsub";
2314      ID = Intrinsic::x86_3dnow_pfsub;
2315      break;
2316    case X86::BI__builtin_ia32_pfsubr:
2317      name = "pfsubr";
2318      ID = Intrinsic::x86_3dnow_pfsubr;
2319      break;
2320    case X86::BI__builtin_ia32_pi2fd:
2321      name = "pi2fd";
2322      ID = Intrinsic::x86_3dnow_pi2fd;
2323      break;
2324    case X86::BI__builtin_ia32_pmulhrw:
2325      name = "pmulhrw";
2326      ID = Intrinsic::x86_3dnow_pmulhrw;
2327      break;
2328    case X86::BI__builtin_ia32_pf2iw:
2329      name = "pf2iw";
2330      ID = Intrinsic::x86_3dnowa_pf2iw;
2331      break;
2332    case X86::BI__builtin_ia32_pfnacc:
2333      name = "pfnacc";
2334      ID = Intrinsic::x86_3dnowa_pfnacc;
2335      break;
2336    case X86::BI__builtin_ia32_pfpnacc:
2337      name = "pfpnacc";
2338      ID = Intrinsic::x86_3dnowa_pfpnacc;
2339      break;
2340    case X86::BI__builtin_ia32_pi2fw:
2341      name = "pi2fw";
2342      ID = Intrinsic::x86_3dnowa_pi2fw;
2343      break;
2344    case X86::BI__builtin_ia32_pswapdsf:
2345    case X86::BI__builtin_ia32_pswapdsi:
2346      name = "pswapd";
2347      ID = Intrinsic::x86_3dnowa_pswapd;
2348      break;
2349    }
2350    llvm::Function *F = CGM.getIntrinsic(ID);
2351    return Builder.CreateCall(F, Ops, name);
2352  }
2353  }
2354}
2355
2356Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
2357                                           const CallExpr *E) {
2358  SmallVector<Value*, 4> Ops;
2359
2360  for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
2361    Ops.push_back(EmitScalarExpr(E->getArg(i)));
2362
2363  Intrinsic::ID ID = Intrinsic::not_intrinsic;
2364
2365  switch (BuiltinID) {
2366  default: return 0;
2367
2368  // vec_ld, vec_lvsl, vec_lvsr
2369  case PPC::BI__builtin_altivec_lvx:
2370  case PPC::BI__builtin_altivec_lvxl:
2371  case PPC::BI__builtin_altivec_lvebx:
2372  case PPC::BI__builtin_altivec_lvehx:
2373  case PPC::BI__builtin_altivec_lvewx:
2374  case PPC::BI__builtin_altivec_lvsl:
2375  case PPC::BI__builtin_altivec_lvsr:
2376  {
2377    Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
2378
2379    Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
2380    Ops.pop_back();
2381
2382    switch (BuiltinID) {
2383    default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
2384    case PPC::BI__builtin_altivec_lvx:
2385      ID = Intrinsic::ppc_altivec_lvx;
2386      break;
2387    case PPC::BI__builtin_altivec_lvxl:
2388      ID = Intrinsic::ppc_altivec_lvxl;
2389      break;
2390    case PPC::BI__builtin_altivec_lvebx:
2391      ID = Intrinsic::ppc_altivec_lvebx;
2392      break;
2393    case PPC::BI__builtin_altivec_lvehx:
2394      ID = Intrinsic::ppc_altivec_lvehx;
2395      break;
2396    case PPC::BI__builtin_altivec_lvewx:
2397      ID = Intrinsic::ppc_altivec_lvewx;
2398      break;
2399    case PPC::BI__builtin_altivec_lvsl:
2400      ID = Intrinsic::ppc_altivec_lvsl;
2401      break;
2402    case PPC::BI__builtin_altivec_lvsr:
2403      ID = Intrinsic::ppc_altivec_lvsr;
2404      break;
2405    }
2406    llvm::Function *F = CGM.getIntrinsic(ID);
2407    return Builder.CreateCall(F, Ops, "");
2408  }
2409
2410  // vec_st
2411  case PPC::BI__builtin_altivec_stvx:
2412  case PPC::BI__builtin_altivec_stvxl:
2413  case PPC::BI__builtin_altivec_stvebx:
2414  case PPC::BI__builtin_altivec_stvehx:
2415  case PPC::BI__builtin_altivec_stvewx:
2416  {
2417    Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
2418    Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
2419    Ops.pop_back();
2420
2421    switch (BuiltinID) {
2422    default: llvm_unreachable("Unsupported st intrinsic!");
2423    case PPC::BI__builtin_altivec_stvx:
2424      ID = Intrinsic::ppc_altivec_stvx;
2425      break;
2426    case PPC::BI__builtin_altivec_stvxl:
2427      ID = Intrinsic::ppc_altivec_stvxl;
2428      break;
2429    case PPC::BI__builtin_altivec_stvebx:
2430      ID = Intrinsic::ppc_altivec_stvebx;
2431      break;
2432    case PPC::BI__builtin_altivec_stvehx:
2433      ID = Intrinsic::ppc_altivec_stvehx;
2434      break;
2435    case PPC::BI__builtin_altivec_stvewx:
2436      ID = Intrinsic::ppc_altivec_stvewx;
2437      break;
2438    }
2439    llvm::Function *F = CGM.getIntrinsic(ID);
2440    return Builder.CreateCall(F, Ops, "");
2441  }
2442  }
2443  return 0;
2444}
2445