• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 "CodeGenFunction.h"
15 #include "CGObjCRuntime.h"
16 #include "CodeGenModule.h"
17 #include "TargetInfo.h"
18 #include "clang/AST/ASTContext.h"
19 #include "clang/AST/Decl.h"
20 #include "clang/Basic/TargetBuiltins.h"
21 #include "clang/Basic/TargetInfo.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Intrinsics.h"
24 
25 using namespace clang;
26 using namespace CodeGen;
27 using namespace llvm;
28 
29 /// getBuiltinLibFunction - Given a builtin id for a function like
30 /// "__builtin_fabsf", return a Function* for "fabsf".
getBuiltinLibFunction(const FunctionDecl * FD,unsigned BuiltinID)31 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
32                                                   unsigned BuiltinID) {
33   assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
34 
35   // Get the name, skip over the __builtin_ prefix (if necessary).
36   StringRef Name;
37   GlobalDecl D(FD);
38 
39   // If the builtin has been declared explicitly with an assembler label,
40   // use the mangled name. This differs from the plain label on platforms
41   // that prefix labels.
42   if (FD->hasAttr<AsmLabelAttr>())
43     Name = getMangledName(D);
44   else
45     Name = Context.BuiltinInfo.GetName(BuiltinID) + 10;
46 
47   llvm::FunctionType *Ty =
48     cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
49 
50   return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
51 }
52 
53 /// Emit the conversions required to turn the given value into an
54 /// integer of the given size.
EmitToInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::IntegerType * IntType)55 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
56                         QualType T, llvm::IntegerType *IntType) {
57   V = CGF.EmitToMemory(V, T);
58 
59   if (V->getType()->isPointerTy())
60     return CGF.Builder.CreatePtrToInt(V, IntType);
61 
62   assert(V->getType() == IntType);
63   return V;
64 }
65 
EmitFromInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::Type * ResultType)66 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
67                           QualType T, llvm::Type *ResultType) {
68   V = CGF.EmitFromMemory(V, T);
69 
70   if (ResultType->isPointerTy())
71     return CGF.Builder.CreateIntToPtr(V, ResultType);
72 
73   assert(V->getType() == ResultType);
74   return V;
75 }
76 
77 /// Utility to insert an atomic instruction based on Instrinsic::ID
78 /// and the expression node.
EmitBinaryAtomic(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E)79 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
80                                llvm::AtomicRMWInst::BinOp Kind,
81                                const CallExpr *E) {
82   QualType T = E->getType();
83   assert(E->getArg(0)->getType()->isPointerType());
84   assert(CGF.getContext().hasSameUnqualifiedType(T,
85                                   E->getArg(0)->getType()->getPointeeType()));
86   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
87 
88   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
89   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
90 
91   llvm::IntegerType *IntType =
92     llvm::IntegerType::get(CGF.getLLVMContext(),
93                            CGF.getContext().getTypeSize(T));
94   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
95 
96   llvm::Value *Args[2];
97   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
98   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
99   llvm::Type *ValueType = Args[1]->getType();
100   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
101 
102   llvm::Value *Result =
103       CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
104                                   llvm::SequentiallyConsistent);
105   Result = EmitFromInt(CGF, Result, T, ValueType);
106   return RValue::get(Result);
107 }
108 
109 /// Utility to insert an atomic instruction based Instrinsic::ID and
110 /// the expression node, where the return value is the result of the
111 /// operation.
EmitBinaryAtomicPost(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E,Instruction::BinaryOps Op)112 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
113                                    llvm::AtomicRMWInst::BinOp Kind,
114                                    const CallExpr *E,
115                                    Instruction::BinaryOps Op) {
116   QualType T = E->getType();
117   assert(E->getArg(0)->getType()->isPointerType());
118   assert(CGF.getContext().hasSameUnqualifiedType(T,
119                                   E->getArg(0)->getType()->getPointeeType()));
120   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
121 
122   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
123   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
124 
125   llvm::IntegerType *IntType =
126     llvm::IntegerType::get(CGF.getLLVMContext(),
127                            CGF.getContext().getTypeSize(T));
128   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
129 
130   llvm::Value *Args[2];
131   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
132   llvm::Type *ValueType = Args[1]->getType();
133   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
134   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
135 
136   llvm::Value *Result =
137       CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
138                                   llvm::SequentiallyConsistent);
139   Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
140   Result = EmitFromInt(CGF, Result, T, ValueType);
141   return RValue::get(Result);
142 }
143 
144 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy,
145 /// which must be a scalar floating point type.
EmitFAbs(CodeGenFunction & CGF,Value * V,QualType ValTy)146 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) {
147   const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>();
148   assert(ValTyP && "isn't scalar fp type!");
149 
150   StringRef FnName;
151   switch (ValTyP->getKind()) {
152   default: llvm_unreachable("Isn't a scalar fp type!");
153   case BuiltinType::Float:      FnName = "fabsf"; break;
154   case BuiltinType::Double:     FnName = "fabs"; break;
155   case BuiltinType::LongDouble: FnName = "fabsl"; break;
156   }
157 
158   // The prototype is something that takes and returns whatever V's type is.
159   llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(),
160                                                    false);
161   llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName);
162 
163   return CGF.EmitNounwindRuntimeCall(Fn, V, "abs");
164 }
165 
emitLibraryCall(CodeGenFunction & CGF,const FunctionDecl * Fn,const CallExpr * E,llvm::Value * calleeValue)166 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn,
167                               const CallExpr *E, llvm::Value *calleeValue) {
168   return CGF.EmitCall(E->getCallee()->getType(), calleeValue,
169                       ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn);
170 }
171 
172 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
173 /// depending on IntrinsicID.
174 ///
175 /// \arg CGF The current codegen function.
176 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
177 /// \arg X The first argument to the llvm.*.with.overflow.*.
178 /// \arg Y The second argument to the llvm.*.with.overflow.*.
179 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
180 /// \returns The result (i.e. sum/product) returned by the intrinsic.
EmitOverflowIntrinsic(CodeGenFunction & CGF,const llvm::Intrinsic::ID IntrinsicID,llvm::Value * X,llvm::Value * Y,llvm::Value * & Carry)181 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
182                                           const llvm::Intrinsic::ID IntrinsicID,
183                                           llvm::Value *X, llvm::Value *Y,
184                                           llvm::Value *&Carry) {
185   // Make sure we have integers of the same width.
186   assert(X->getType() == Y->getType() &&
187          "Arguments must be the same type. (Did you forget to make sure both "
188          "arguments have the same integer width?)");
189 
190   llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
191   llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y);
192   Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
193   return CGF.Builder.CreateExtractValue(Tmp, 0);
194 }
195 
EmitBuiltinExpr(const FunctionDecl * FD,unsigned BuiltinID,const CallExpr * E)196 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
197                                         unsigned BuiltinID, const CallExpr *E) {
198   // See if we can constant fold this builtin.  If so, don't emit it at all.
199   Expr::EvalResult Result;
200   if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
201       !Result.hasSideEffects()) {
202     if (Result.Val.isInt())
203       return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
204                                                 Result.Val.getInt()));
205     if (Result.Val.isFloat())
206       return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
207                                                Result.Val.getFloat()));
208   }
209 
210   switch (BuiltinID) {
211   default: break;  // Handle intrinsics and libm functions below.
212   case Builtin::BI__builtin___CFStringMakeConstantString:
213   case Builtin::BI__builtin___NSStringMakeConstantString:
214     return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0));
215   case Builtin::BI__builtin_stdarg_start:
216   case Builtin::BI__builtin_va_start:
217   case Builtin::BI__builtin_va_end: {
218     Value *ArgValue = EmitVAListRef(E->getArg(0));
219     llvm::Type *DestType = Int8PtrTy;
220     if (ArgValue->getType() != DestType)
221       ArgValue = Builder.CreateBitCast(ArgValue, DestType,
222                                        ArgValue->getName().data());
223 
224     Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ?
225       Intrinsic::vaend : Intrinsic::vastart;
226     return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue));
227   }
228   case Builtin::BI__builtin_va_copy: {
229     Value *DstPtr = EmitVAListRef(E->getArg(0));
230     Value *SrcPtr = EmitVAListRef(E->getArg(1));
231 
232     llvm::Type *Type = Int8PtrTy;
233 
234     DstPtr = Builder.CreateBitCast(DstPtr, Type);
235     SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
236     return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy),
237                                            DstPtr, SrcPtr));
238   }
239   case Builtin::BI__builtin_abs:
240   case Builtin::BI__builtin_labs:
241   case Builtin::BI__builtin_llabs: {
242     Value *ArgValue = EmitScalarExpr(E->getArg(0));
243 
244     Value *NegOp = Builder.CreateNeg(ArgValue, "neg");
245     Value *CmpResult =
246     Builder.CreateICmpSGE(ArgValue,
247                           llvm::Constant::getNullValue(ArgValue->getType()),
248                                                             "abscond");
249     Value *Result =
250       Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs");
251 
252     return RValue::get(Result);
253   }
254 
255   case Builtin::BI__builtin_conj:
256   case Builtin::BI__builtin_conjf:
257   case Builtin::BI__builtin_conjl: {
258     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
259     Value *Real = ComplexVal.first;
260     Value *Imag = ComplexVal.second;
261     Value *Zero =
262       Imag->getType()->isFPOrFPVectorTy()
263         ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType())
264         : llvm::Constant::getNullValue(Imag->getType());
265 
266     Imag = Builder.CreateFSub(Zero, Imag, "sub");
267     return RValue::getComplex(std::make_pair(Real, Imag));
268   }
269   case Builtin::BI__builtin_creal:
270   case Builtin::BI__builtin_crealf:
271   case Builtin::BI__builtin_creall:
272   case Builtin::BIcreal:
273   case Builtin::BIcrealf:
274   case Builtin::BIcreall: {
275     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
276     return RValue::get(ComplexVal.first);
277   }
278 
279   case Builtin::BI__builtin_cimag:
280   case Builtin::BI__builtin_cimagf:
281   case Builtin::BI__builtin_cimagl:
282   case Builtin::BIcimag:
283   case Builtin::BIcimagf:
284   case Builtin::BIcimagl: {
285     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
286     return RValue::get(ComplexVal.second);
287   }
288 
289   case Builtin::BI__builtin_ctzs:
290   case Builtin::BI__builtin_ctz:
291   case Builtin::BI__builtin_ctzl:
292   case Builtin::BI__builtin_ctzll: {
293     Value *ArgValue = EmitScalarExpr(E->getArg(0));
294 
295     llvm::Type *ArgType = ArgValue->getType();
296     Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
297 
298     llvm::Type *ResultType = ConvertType(E->getType());
299     Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef());
300     Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
301     if (Result->getType() != ResultType)
302       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
303                                      "cast");
304     return RValue::get(Result);
305   }
306   case Builtin::BI__builtin_clzs:
307   case Builtin::BI__builtin_clz:
308   case Builtin::BI__builtin_clzl:
309   case Builtin::BI__builtin_clzll: {
310     Value *ArgValue = EmitScalarExpr(E->getArg(0));
311 
312     llvm::Type *ArgType = ArgValue->getType();
313     Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
314 
315     llvm::Type *ResultType = ConvertType(E->getType());
316     Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef());
317     Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
318     if (Result->getType() != ResultType)
319       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
320                                      "cast");
321     return RValue::get(Result);
322   }
323   case Builtin::BI__builtin_ffs:
324   case Builtin::BI__builtin_ffsl:
325   case Builtin::BI__builtin_ffsll: {
326     // ffs(x) -> x ? cttz(x) + 1 : 0
327     Value *ArgValue = EmitScalarExpr(E->getArg(0));
328 
329     llvm::Type *ArgType = ArgValue->getType();
330     Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
331 
332     llvm::Type *ResultType = ConvertType(E->getType());
333     Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue,
334                                                        Builder.getTrue()),
335                                    llvm::ConstantInt::get(ArgType, 1));
336     Value *Zero = llvm::Constant::getNullValue(ArgType);
337     Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
338     Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
339     if (Result->getType() != ResultType)
340       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
341                                      "cast");
342     return RValue::get(Result);
343   }
344   case Builtin::BI__builtin_parity:
345   case Builtin::BI__builtin_parityl:
346   case Builtin::BI__builtin_parityll: {
347     // parity(x) -> ctpop(x) & 1
348     Value *ArgValue = EmitScalarExpr(E->getArg(0));
349 
350     llvm::Type *ArgType = ArgValue->getType();
351     Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
352 
353     llvm::Type *ResultType = ConvertType(E->getType());
354     Value *Tmp = Builder.CreateCall(F, ArgValue);
355     Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
356     if (Result->getType() != ResultType)
357       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
358                                      "cast");
359     return RValue::get(Result);
360   }
361   case Builtin::BI__builtin_popcount:
362   case Builtin::BI__builtin_popcountl:
363   case Builtin::BI__builtin_popcountll: {
364     Value *ArgValue = EmitScalarExpr(E->getArg(0));
365 
366     llvm::Type *ArgType = ArgValue->getType();
367     Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
368 
369     llvm::Type *ResultType = ConvertType(E->getType());
370     Value *Result = Builder.CreateCall(F, ArgValue);
371     if (Result->getType() != ResultType)
372       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
373                                      "cast");
374     return RValue::get(Result);
375   }
376   case Builtin::BI__builtin_expect: {
377     Value *ArgValue = EmitScalarExpr(E->getArg(0));
378     llvm::Type *ArgType = ArgValue->getType();
379 
380     Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
381     Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
382 
383     Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue,
384                                         "expval");
385     return RValue::get(Result);
386   }
387   case Builtin::BI__builtin_bswap16:
388   case Builtin::BI__builtin_bswap32:
389   case Builtin::BI__builtin_bswap64: {
390     Value *ArgValue = EmitScalarExpr(E->getArg(0));
391     llvm::Type *ArgType = ArgValue->getType();
392     Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType);
393     return RValue::get(Builder.CreateCall(F, ArgValue));
394   }
395   case Builtin::BI__builtin_object_size: {
396     // We rely on constant folding to deal with expressions with side effects.
397     assert(!E->getArg(0)->HasSideEffects(getContext()) &&
398            "should have been constant folded");
399 
400     // We pass this builtin onto the optimizer so that it can
401     // figure out the object size in more complex cases.
402     llvm::Type *ResType = ConvertType(E->getType());
403 
404     // LLVM only supports 0 and 2, make sure that we pass along that
405     // as a boolean.
406     Value *Ty = EmitScalarExpr(E->getArg(1));
407     ConstantInt *CI = dyn_cast<ConstantInt>(Ty);
408     assert(CI);
409     uint64_t val = CI->getZExtValue();
410     CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1);
411 
412     Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType);
413     return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI));
414   }
415   case Builtin::BI__builtin_prefetch: {
416     Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
417     // FIXME: Technically these constants should of type 'int', yes?
418     RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
419       llvm::ConstantInt::get(Int32Ty, 0);
420     Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
421       llvm::ConstantInt::get(Int32Ty, 3);
422     Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
423     Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
424     return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data));
425   }
426   case Builtin::BI__builtin_readcyclecounter: {
427     Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
428     return RValue::get(Builder.CreateCall(F));
429   }
430   case Builtin::BI__builtin_trap: {
431     Value *F = CGM.getIntrinsic(Intrinsic::trap);
432     return RValue::get(Builder.CreateCall(F));
433   }
434   case Builtin::BI__debugbreak: {
435     Value *F = CGM.getIntrinsic(Intrinsic::debugtrap);
436     return RValue::get(Builder.CreateCall(F));
437   }
438   case Builtin::BI__builtin_unreachable: {
439     if (SanOpts->Unreachable)
440       EmitCheck(Builder.getFalse(), "builtin_unreachable",
441                 EmitCheckSourceLocation(E->getExprLoc()),
442                 ArrayRef<llvm::Value *>(), CRK_Unrecoverable);
443     else
444       Builder.CreateUnreachable();
445 
446     // We do need to preserve an insertion point.
447     EmitBlock(createBasicBlock("unreachable.cont"));
448 
449     return RValue::get(0);
450   }
451 
452   case Builtin::BI__builtin_powi:
453   case Builtin::BI__builtin_powif:
454   case Builtin::BI__builtin_powil: {
455     Value *Base = EmitScalarExpr(E->getArg(0));
456     Value *Exponent = EmitScalarExpr(E->getArg(1));
457     llvm::Type *ArgType = Base->getType();
458     Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
459     return RValue::get(Builder.CreateCall2(F, Base, Exponent));
460   }
461 
462   case Builtin::BI__builtin_isgreater:
463   case Builtin::BI__builtin_isgreaterequal:
464   case Builtin::BI__builtin_isless:
465   case Builtin::BI__builtin_islessequal:
466   case Builtin::BI__builtin_islessgreater:
467   case Builtin::BI__builtin_isunordered: {
468     // Ordered comparisons: we know the arguments to these are matching scalar
469     // floating point values.
470     Value *LHS = EmitScalarExpr(E->getArg(0));
471     Value *RHS = EmitScalarExpr(E->getArg(1));
472 
473     switch (BuiltinID) {
474     default: llvm_unreachable("Unknown ordered comparison");
475     case Builtin::BI__builtin_isgreater:
476       LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
477       break;
478     case Builtin::BI__builtin_isgreaterequal:
479       LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
480       break;
481     case Builtin::BI__builtin_isless:
482       LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
483       break;
484     case Builtin::BI__builtin_islessequal:
485       LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
486       break;
487     case Builtin::BI__builtin_islessgreater:
488       LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
489       break;
490     case Builtin::BI__builtin_isunordered:
491       LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
492       break;
493     }
494     // ZExt bool to int type.
495     return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
496   }
497   case Builtin::BI__builtin_isnan: {
498     Value *V = EmitScalarExpr(E->getArg(0));
499     V = Builder.CreateFCmpUNO(V, V, "cmp");
500     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
501   }
502 
503   case Builtin::BI__builtin_isinf: {
504     // isinf(x) --> fabs(x) == infinity
505     Value *V = EmitScalarExpr(E->getArg(0));
506     V = EmitFAbs(*this, V, E->getArg(0)->getType());
507 
508     V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf");
509     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
510   }
511 
512   // TODO: BI__builtin_isinf_sign
513   //   isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0
514 
515   case Builtin::BI__builtin_isnormal: {
516     // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
517     Value *V = EmitScalarExpr(E->getArg(0));
518     Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
519 
520     Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
521     Value *IsLessThanInf =
522       Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
523     APFloat Smallest = APFloat::getSmallestNormalized(
524                    getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
525     Value *IsNormal =
526       Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
527                             "isnormal");
528     V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
529     V = Builder.CreateAnd(V, IsNormal, "and");
530     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
531   }
532 
533   case Builtin::BI__builtin_isfinite: {
534     // isfinite(x) --> x == x && fabs(x) != infinity;
535     Value *V = EmitScalarExpr(E->getArg(0));
536     Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
537 
538     Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
539     Value *IsNotInf =
540       Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
541 
542     V = Builder.CreateAnd(Eq, IsNotInf, "and");
543     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
544   }
545 
546   case Builtin::BI__builtin_fpclassify: {
547     Value *V = EmitScalarExpr(E->getArg(5));
548     llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
549 
550     // Create Result
551     BasicBlock *Begin = Builder.GetInsertBlock();
552     BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
553     Builder.SetInsertPoint(End);
554     PHINode *Result =
555       Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
556                         "fpclassify_result");
557 
558     // if (V==0) return FP_ZERO
559     Builder.SetInsertPoint(Begin);
560     Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
561                                           "iszero");
562     Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
563     BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
564     Builder.CreateCondBr(IsZero, End, NotZero);
565     Result->addIncoming(ZeroLiteral, Begin);
566 
567     // if (V != V) return FP_NAN
568     Builder.SetInsertPoint(NotZero);
569     Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
570     Value *NanLiteral = EmitScalarExpr(E->getArg(0));
571     BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
572     Builder.CreateCondBr(IsNan, End, NotNan);
573     Result->addIncoming(NanLiteral, NotZero);
574 
575     // if (fabs(V) == infinity) return FP_INFINITY
576     Builder.SetInsertPoint(NotNan);
577     Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType());
578     Value *IsInf =
579       Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
580                             "isinf");
581     Value *InfLiteral = EmitScalarExpr(E->getArg(1));
582     BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
583     Builder.CreateCondBr(IsInf, End, NotInf);
584     Result->addIncoming(InfLiteral, NotNan);
585 
586     // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
587     Builder.SetInsertPoint(NotInf);
588     APFloat Smallest = APFloat::getSmallestNormalized(
589         getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
590     Value *IsNormal =
591       Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
592                             "isnormal");
593     Value *NormalResult =
594       Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
595                            EmitScalarExpr(E->getArg(3)));
596     Builder.CreateBr(End);
597     Result->addIncoming(NormalResult, NotInf);
598 
599     // return Result
600     Builder.SetInsertPoint(End);
601     return RValue::get(Result);
602   }
603 
604   case Builtin::BIalloca:
605   case Builtin::BI__builtin_alloca: {
606     Value *Size = EmitScalarExpr(E->getArg(0));
607     return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size));
608   }
609   case Builtin::BIbzero:
610   case Builtin::BI__builtin_bzero: {
611     std::pair<llvm::Value*, unsigned> Dest =
612         EmitPointerWithAlignment(E->getArg(0));
613     Value *SizeVal = EmitScalarExpr(E->getArg(1));
614     Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal,
615                          Dest.second, false);
616     return RValue::get(Dest.first);
617   }
618   case Builtin::BImemcpy:
619   case Builtin::BI__builtin_memcpy: {
620     std::pair<llvm::Value*, unsigned> Dest =
621         EmitPointerWithAlignment(E->getArg(0));
622     std::pair<llvm::Value*, unsigned> Src =
623         EmitPointerWithAlignment(E->getArg(1));
624     Value *SizeVal = EmitScalarExpr(E->getArg(2));
625     unsigned Align = std::min(Dest.second, Src.second);
626     Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
627     return RValue::get(Dest.first);
628   }
629 
630   case Builtin::BI__builtin___memcpy_chk: {
631     // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
632     llvm::APSInt Size, DstSize;
633     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
634         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
635       break;
636     if (Size.ugt(DstSize))
637       break;
638     std::pair<llvm::Value*, unsigned> Dest =
639         EmitPointerWithAlignment(E->getArg(0));
640     std::pair<llvm::Value*, unsigned> Src =
641         EmitPointerWithAlignment(E->getArg(1));
642     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
643     unsigned Align = std::min(Dest.second, Src.second);
644     Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
645     return RValue::get(Dest.first);
646   }
647 
648   case Builtin::BI__builtin_objc_memmove_collectable: {
649     Value *Address = EmitScalarExpr(E->getArg(0));
650     Value *SrcAddr = EmitScalarExpr(E->getArg(1));
651     Value *SizeVal = EmitScalarExpr(E->getArg(2));
652     CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
653                                                   Address, SrcAddr, SizeVal);
654     return RValue::get(Address);
655   }
656 
657   case Builtin::BI__builtin___memmove_chk: {
658     // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
659     llvm::APSInt Size, DstSize;
660     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
661         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
662       break;
663     if (Size.ugt(DstSize))
664       break;
665     std::pair<llvm::Value*, unsigned> Dest =
666         EmitPointerWithAlignment(E->getArg(0));
667     std::pair<llvm::Value*, unsigned> Src =
668         EmitPointerWithAlignment(E->getArg(1));
669     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
670     unsigned Align = std::min(Dest.second, Src.second);
671     Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false);
672     return RValue::get(Dest.first);
673   }
674 
675   case Builtin::BImemmove:
676   case Builtin::BI__builtin_memmove: {
677     std::pair<llvm::Value*, unsigned> Dest =
678         EmitPointerWithAlignment(E->getArg(0));
679     std::pair<llvm::Value*, unsigned> Src =
680         EmitPointerWithAlignment(E->getArg(1));
681     Value *SizeVal = EmitScalarExpr(E->getArg(2));
682     unsigned Align = std::min(Dest.second, Src.second);
683     Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false);
684     return RValue::get(Dest.first);
685   }
686   case Builtin::BImemset:
687   case Builtin::BI__builtin_memset: {
688     std::pair<llvm::Value*, unsigned> Dest =
689         EmitPointerWithAlignment(E->getArg(0));
690     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
691                                          Builder.getInt8Ty());
692     Value *SizeVal = EmitScalarExpr(E->getArg(2));
693     Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
694     return RValue::get(Dest.first);
695   }
696   case Builtin::BI__builtin___memset_chk: {
697     // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
698     llvm::APSInt Size, DstSize;
699     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
700         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
701       break;
702     if (Size.ugt(DstSize))
703       break;
704     std::pair<llvm::Value*, unsigned> Dest =
705         EmitPointerWithAlignment(E->getArg(0));
706     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
707                                          Builder.getInt8Ty());
708     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
709     Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
710     return RValue::get(Dest.first);
711   }
712   case Builtin::BI__builtin_dwarf_cfa: {
713     // The offset in bytes from the first argument to the CFA.
714     //
715     // Why on earth is this in the frontend?  Is there any reason at
716     // all that the backend can't reasonably determine this while
717     // lowering llvm.eh.dwarf.cfa()?
718     //
719     // TODO: If there's a satisfactory reason, add a target hook for
720     // this instead of hard-coding 0, which is correct for most targets.
721     int32_t Offset = 0;
722 
723     Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
724     return RValue::get(Builder.CreateCall(F,
725                                       llvm::ConstantInt::get(Int32Ty, Offset)));
726   }
727   case Builtin::BI__builtin_return_address: {
728     Value *Depth = EmitScalarExpr(E->getArg(0));
729     Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
730     Value *F = CGM.getIntrinsic(Intrinsic::returnaddress);
731     return RValue::get(Builder.CreateCall(F, Depth));
732   }
733   case Builtin::BI__builtin_frame_address: {
734     Value *Depth = EmitScalarExpr(E->getArg(0));
735     Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
736     Value *F = CGM.getIntrinsic(Intrinsic::frameaddress);
737     return RValue::get(Builder.CreateCall(F, Depth));
738   }
739   case Builtin::BI__builtin_extract_return_addr: {
740     Value *Address = EmitScalarExpr(E->getArg(0));
741     Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
742     return RValue::get(Result);
743   }
744   case Builtin::BI__builtin_frob_return_addr: {
745     Value *Address = EmitScalarExpr(E->getArg(0));
746     Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
747     return RValue::get(Result);
748   }
749   case Builtin::BI__builtin_dwarf_sp_column: {
750     llvm::IntegerType *Ty
751       = cast<llvm::IntegerType>(ConvertType(E->getType()));
752     int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
753     if (Column == -1) {
754       CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
755       return RValue::get(llvm::UndefValue::get(Ty));
756     }
757     return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
758   }
759   case Builtin::BI__builtin_init_dwarf_reg_size_table: {
760     Value *Address = EmitScalarExpr(E->getArg(0));
761     if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
762       CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
763     return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
764   }
765   case Builtin::BI__builtin_eh_return: {
766     Value *Int = EmitScalarExpr(E->getArg(0));
767     Value *Ptr = EmitScalarExpr(E->getArg(1));
768 
769     llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
770     assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
771            "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
772     Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32
773                                   ? Intrinsic::eh_return_i32
774                                   : Intrinsic::eh_return_i64);
775     Builder.CreateCall2(F, Int, Ptr);
776     Builder.CreateUnreachable();
777 
778     // We do need to preserve an insertion point.
779     EmitBlock(createBasicBlock("builtin_eh_return.cont"));
780 
781     return RValue::get(0);
782   }
783   case Builtin::BI__builtin_unwind_init: {
784     Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
785     return RValue::get(Builder.CreateCall(F));
786   }
787   case Builtin::BI__builtin_extend_pointer: {
788     // Extends a pointer to the size of an _Unwind_Word, which is
789     // uint64_t on all platforms.  Generally this gets poked into a
790     // register and eventually used as an address, so if the
791     // addressing registers are wider than pointers and the platform
792     // doesn't implicitly ignore high-order bits when doing
793     // addressing, we need to make sure we zext / sext based on
794     // the platform's expectations.
795     //
796     // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
797 
798     // Cast the pointer to intptr_t.
799     Value *Ptr = EmitScalarExpr(E->getArg(0));
800     Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
801 
802     // If that's 64 bits, we're done.
803     if (IntPtrTy->getBitWidth() == 64)
804       return RValue::get(Result);
805 
806     // Otherwise, ask the codegen data what to do.
807     if (getTargetHooks().extendPointerWithSExt())
808       return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
809     else
810       return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
811   }
812   case Builtin::BI__builtin_setjmp: {
813     // Buffer is a void**.
814     Value *Buf = EmitScalarExpr(E->getArg(0));
815 
816     // Store the frame pointer to the setjmp buffer.
817     Value *FrameAddr =
818       Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
819                          ConstantInt::get(Int32Ty, 0));
820     Builder.CreateStore(FrameAddr, Buf);
821 
822     // Store the stack pointer to the setjmp buffer.
823     Value *StackAddr =
824       Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
825     Value *StackSaveSlot =
826       Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2));
827     Builder.CreateStore(StackAddr, StackSaveSlot);
828 
829     // Call LLVM's EH setjmp, which is lightweight.
830     Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
831     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
832     return RValue::get(Builder.CreateCall(F, Buf));
833   }
834   case Builtin::BI__builtin_longjmp: {
835     Value *Buf = EmitScalarExpr(E->getArg(0));
836     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
837 
838     // Call LLVM's EH longjmp, which is lightweight.
839     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
840 
841     // longjmp doesn't return; mark this as unreachable.
842     Builder.CreateUnreachable();
843 
844     // We do need to preserve an insertion point.
845     EmitBlock(createBasicBlock("longjmp.cont"));
846 
847     return RValue::get(0);
848   }
849   case Builtin::BI__sync_fetch_and_add:
850   case Builtin::BI__sync_fetch_and_sub:
851   case Builtin::BI__sync_fetch_and_or:
852   case Builtin::BI__sync_fetch_and_and:
853   case Builtin::BI__sync_fetch_and_xor:
854   case Builtin::BI__sync_add_and_fetch:
855   case Builtin::BI__sync_sub_and_fetch:
856   case Builtin::BI__sync_and_and_fetch:
857   case Builtin::BI__sync_or_and_fetch:
858   case Builtin::BI__sync_xor_and_fetch:
859   case Builtin::BI__sync_val_compare_and_swap:
860   case Builtin::BI__sync_bool_compare_and_swap:
861   case Builtin::BI__sync_lock_test_and_set:
862   case Builtin::BI__sync_lock_release:
863   case Builtin::BI__sync_swap:
864     llvm_unreachable("Shouldn't make it through sema");
865   case Builtin::BI__sync_fetch_and_add_1:
866   case Builtin::BI__sync_fetch_and_add_2:
867   case Builtin::BI__sync_fetch_and_add_4:
868   case Builtin::BI__sync_fetch_and_add_8:
869   case Builtin::BI__sync_fetch_and_add_16:
870     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
871   case Builtin::BI__sync_fetch_and_sub_1:
872   case Builtin::BI__sync_fetch_and_sub_2:
873   case Builtin::BI__sync_fetch_and_sub_4:
874   case Builtin::BI__sync_fetch_and_sub_8:
875   case Builtin::BI__sync_fetch_and_sub_16:
876     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
877   case Builtin::BI__sync_fetch_and_or_1:
878   case Builtin::BI__sync_fetch_and_or_2:
879   case Builtin::BI__sync_fetch_and_or_4:
880   case Builtin::BI__sync_fetch_and_or_8:
881   case Builtin::BI__sync_fetch_and_or_16:
882     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
883   case Builtin::BI__sync_fetch_and_and_1:
884   case Builtin::BI__sync_fetch_and_and_2:
885   case Builtin::BI__sync_fetch_and_and_4:
886   case Builtin::BI__sync_fetch_and_and_8:
887   case Builtin::BI__sync_fetch_and_and_16:
888     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
889   case Builtin::BI__sync_fetch_and_xor_1:
890   case Builtin::BI__sync_fetch_and_xor_2:
891   case Builtin::BI__sync_fetch_and_xor_4:
892   case Builtin::BI__sync_fetch_and_xor_8:
893   case Builtin::BI__sync_fetch_and_xor_16:
894     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
895 
896   // Clang extensions: not overloaded yet.
897   case Builtin::BI__sync_fetch_and_min:
898     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
899   case Builtin::BI__sync_fetch_and_max:
900     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
901   case Builtin::BI__sync_fetch_and_umin:
902     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
903   case Builtin::BI__sync_fetch_and_umax:
904     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
905 
906   case Builtin::BI__sync_add_and_fetch_1:
907   case Builtin::BI__sync_add_and_fetch_2:
908   case Builtin::BI__sync_add_and_fetch_4:
909   case Builtin::BI__sync_add_and_fetch_8:
910   case Builtin::BI__sync_add_and_fetch_16:
911     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
912                                 llvm::Instruction::Add);
913   case Builtin::BI__sync_sub_and_fetch_1:
914   case Builtin::BI__sync_sub_and_fetch_2:
915   case Builtin::BI__sync_sub_and_fetch_4:
916   case Builtin::BI__sync_sub_and_fetch_8:
917   case Builtin::BI__sync_sub_and_fetch_16:
918     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
919                                 llvm::Instruction::Sub);
920   case Builtin::BI__sync_and_and_fetch_1:
921   case Builtin::BI__sync_and_and_fetch_2:
922   case Builtin::BI__sync_and_and_fetch_4:
923   case Builtin::BI__sync_and_and_fetch_8:
924   case Builtin::BI__sync_and_and_fetch_16:
925     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
926                                 llvm::Instruction::And);
927   case Builtin::BI__sync_or_and_fetch_1:
928   case Builtin::BI__sync_or_and_fetch_2:
929   case Builtin::BI__sync_or_and_fetch_4:
930   case Builtin::BI__sync_or_and_fetch_8:
931   case Builtin::BI__sync_or_and_fetch_16:
932     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
933                                 llvm::Instruction::Or);
934   case Builtin::BI__sync_xor_and_fetch_1:
935   case Builtin::BI__sync_xor_and_fetch_2:
936   case Builtin::BI__sync_xor_and_fetch_4:
937   case Builtin::BI__sync_xor_and_fetch_8:
938   case Builtin::BI__sync_xor_and_fetch_16:
939     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
940                                 llvm::Instruction::Xor);
941 
942   case Builtin::BI__sync_val_compare_and_swap_1:
943   case Builtin::BI__sync_val_compare_and_swap_2:
944   case Builtin::BI__sync_val_compare_and_swap_4:
945   case Builtin::BI__sync_val_compare_and_swap_8:
946   case Builtin::BI__sync_val_compare_and_swap_16: {
947     QualType T = E->getType();
948     llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
949     unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
950 
951     llvm::IntegerType *IntType =
952       llvm::IntegerType::get(getLLVMContext(),
953                              getContext().getTypeSize(T));
954     llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
955 
956     Value *Args[3];
957     Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
958     Args[1] = EmitScalarExpr(E->getArg(1));
959     llvm::Type *ValueType = Args[1]->getType();
960     Args[1] = EmitToInt(*this, Args[1], T, IntType);
961     Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
962 
963     Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
964                                                 llvm::SequentiallyConsistent);
965     Result = EmitFromInt(*this, Result, T, ValueType);
966     return RValue::get(Result);
967   }
968 
969   case Builtin::BI__sync_bool_compare_and_swap_1:
970   case Builtin::BI__sync_bool_compare_and_swap_2:
971   case Builtin::BI__sync_bool_compare_and_swap_4:
972   case Builtin::BI__sync_bool_compare_and_swap_8:
973   case Builtin::BI__sync_bool_compare_and_swap_16: {
974     QualType T = E->getArg(1)->getType();
975     llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
976     unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
977 
978     llvm::IntegerType *IntType =
979       llvm::IntegerType::get(getLLVMContext(),
980                              getContext().getTypeSize(T));
981     llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
982 
983     Value *Args[3];
984     Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
985     Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType);
986     Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
987 
988     Value *OldVal = Args[1];
989     Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
990                                                  llvm::SequentiallyConsistent);
991     Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal);
992     // zext bool to int.
993     Result = Builder.CreateZExt(Result, ConvertType(E->getType()));
994     return RValue::get(Result);
995   }
996 
997   case Builtin::BI__sync_swap_1:
998   case Builtin::BI__sync_swap_2:
999   case Builtin::BI__sync_swap_4:
1000   case Builtin::BI__sync_swap_8:
1001   case Builtin::BI__sync_swap_16:
1002     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1003 
1004   case Builtin::BI__sync_lock_test_and_set_1:
1005   case Builtin::BI__sync_lock_test_and_set_2:
1006   case Builtin::BI__sync_lock_test_and_set_4:
1007   case Builtin::BI__sync_lock_test_and_set_8:
1008   case Builtin::BI__sync_lock_test_and_set_16:
1009     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1010 
1011   case Builtin::BI__sync_lock_release_1:
1012   case Builtin::BI__sync_lock_release_2:
1013   case Builtin::BI__sync_lock_release_4:
1014   case Builtin::BI__sync_lock_release_8:
1015   case Builtin::BI__sync_lock_release_16: {
1016     Value *Ptr = EmitScalarExpr(E->getArg(0));
1017     QualType ElTy = E->getArg(0)->getType()->getPointeeType();
1018     CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
1019     llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
1020                                              StoreSize.getQuantity() * 8);
1021     Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
1022     llvm::StoreInst *Store =
1023       Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr);
1024     Store->setAlignment(StoreSize.getQuantity());
1025     Store->setAtomic(llvm::Release);
1026     return RValue::get(0);
1027   }
1028 
1029   case Builtin::BI__sync_synchronize: {
1030     // We assume this is supposed to correspond to a C++0x-style
1031     // sequentially-consistent fence (i.e. this is only usable for
1032     // synchonization, not device I/O or anything like that). This intrinsic
1033     // is really badly designed in the sense that in theory, there isn't
1034     // any way to safely use it... but in practice, it mostly works
1035     // to use it with non-atomic loads and stores to get acquire/release
1036     // semantics.
1037     Builder.CreateFence(llvm::SequentiallyConsistent);
1038     return RValue::get(0);
1039   }
1040 
1041   case Builtin::BI__c11_atomic_is_lock_free:
1042   case Builtin::BI__atomic_is_lock_free: {
1043     // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
1044     // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
1045     // _Atomic(T) is always properly-aligned.
1046     const char *LibCallName = "__atomic_is_lock_free";
1047     CallArgList Args;
1048     Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
1049              getContext().getSizeType());
1050     if (BuiltinID == Builtin::BI__atomic_is_lock_free)
1051       Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
1052                getContext().VoidPtrTy);
1053     else
1054       Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
1055                getContext().VoidPtrTy);
1056     const CGFunctionInfo &FuncInfo =
1057         CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args,
1058                                                FunctionType::ExtInfo(),
1059                                                RequiredArgs::All);
1060     llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
1061     llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
1062     return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args);
1063   }
1064 
1065   case Builtin::BI__atomic_test_and_set: {
1066     // Look at the argument type to determine whether this is a volatile
1067     // operation. The parameter type is always volatile.
1068     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1069     bool Volatile =
1070         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1071 
1072     Value *Ptr = EmitScalarExpr(E->getArg(0));
1073     unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1074     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1075     Value *NewVal = Builder.getInt8(1);
1076     Value *Order = EmitScalarExpr(E->getArg(1));
1077     if (isa<llvm::ConstantInt>(Order)) {
1078       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1079       AtomicRMWInst *Result = 0;
1080       switch (ord) {
1081       case 0:  // memory_order_relaxed
1082       default: // invalid order
1083         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1084                                          Ptr, NewVal,
1085                                          llvm::Monotonic);
1086         break;
1087       case 1:  // memory_order_consume
1088       case 2:  // memory_order_acquire
1089         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1090                                          Ptr, NewVal,
1091                                          llvm::Acquire);
1092         break;
1093       case 3:  // memory_order_release
1094         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1095                                          Ptr, NewVal,
1096                                          llvm::Release);
1097         break;
1098       case 4:  // memory_order_acq_rel
1099         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1100                                          Ptr, NewVal,
1101                                          llvm::AcquireRelease);
1102         break;
1103       case 5:  // memory_order_seq_cst
1104         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1105                                          Ptr, NewVal,
1106                                          llvm::SequentiallyConsistent);
1107         break;
1108       }
1109       Result->setVolatile(Volatile);
1110       return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1111     }
1112 
1113     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1114 
1115     llvm::BasicBlock *BBs[5] = {
1116       createBasicBlock("monotonic", CurFn),
1117       createBasicBlock("acquire", CurFn),
1118       createBasicBlock("release", CurFn),
1119       createBasicBlock("acqrel", CurFn),
1120       createBasicBlock("seqcst", CurFn)
1121     };
1122     llvm::AtomicOrdering Orders[5] = {
1123       llvm::Monotonic, llvm::Acquire, llvm::Release,
1124       llvm::AcquireRelease, llvm::SequentiallyConsistent
1125     };
1126 
1127     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1128     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1129 
1130     Builder.SetInsertPoint(ContBB);
1131     PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
1132 
1133     for (unsigned i = 0; i < 5; ++i) {
1134       Builder.SetInsertPoint(BBs[i]);
1135       AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1136                                                    Ptr, NewVal, Orders[i]);
1137       RMW->setVolatile(Volatile);
1138       Result->addIncoming(RMW, BBs[i]);
1139       Builder.CreateBr(ContBB);
1140     }
1141 
1142     SI->addCase(Builder.getInt32(0), BBs[0]);
1143     SI->addCase(Builder.getInt32(1), BBs[1]);
1144     SI->addCase(Builder.getInt32(2), BBs[1]);
1145     SI->addCase(Builder.getInt32(3), BBs[2]);
1146     SI->addCase(Builder.getInt32(4), BBs[3]);
1147     SI->addCase(Builder.getInt32(5), BBs[4]);
1148 
1149     Builder.SetInsertPoint(ContBB);
1150     return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1151   }
1152 
1153   case Builtin::BI__atomic_clear: {
1154     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1155     bool Volatile =
1156         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1157 
1158     Value *Ptr = EmitScalarExpr(E->getArg(0));
1159     unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1160     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1161     Value *NewVal = Builder.getInt8(0);
1162     Value *Order = EmitScalarExpr(E->getArg(1));
1163     if (isa<llvm::ConstantInt>(Order)) {
1164       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1165       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1166       Store->setAlignment(1);
1167       switch (ord) {
1168       case 0:  // memory_order_relaxed
1169       default: // invalid order
1170         Store->setOrdering(llvm::Monotonic);
1171         break;
1172       case 3:  // memory_order_release
1173         Store->setOrdering(llvm::Release);
1174         break;
1175       case 5:  // memory_order_seq_cst
1176         Store->setOrdering(llvm::SequentiallyConsistent);
1177         break;
1178       }
1179       return RValue::get(0);
1180     }
1181 
1182     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1183 
1184     llvm::BasicBlock *BBs[3] = {
1185       createBasicBlock("monotonic", CurFn),
1186       createBasicBlock("release", CurFn),
1187       createBasicBlock("seqcst", CurFn)
1188     };
1189     llvm::AtomicOrdering Orders[3] = {
1190       llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent
1191     };
1192 
1193     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1194     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1195 
1196     for (unsigned i = 0; i < 3; ++i) {
1197       Builder.SetInsertPoint(BBs[i]);
1198       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1199       Store->setAlignment(1);
1200       Store->setOrdering(Orders[i]);
1201       Builder.CreateBr(ContBB);
1202     }
1203 
1204     SI->addCase(Builder.getInt32(0), BBs[0]);
1205     SI->addCase(Builder.getInt32(3), BBs[1]);
1206     SI->addCase(Builder.getInt32(5), BBs[2]);
1207 
1208     Builder.SetInsertPoint(ContBB);
1209     return RValue::get(0);
1210   }
1211 
1212   case Builtin::BI__atomic_thread_fence:
1213   case Builtin::BI__atomic_signal_fence:
1214   case Builtin::BI__c11_atomic_thread_fence:
1215   case Builtin::BI__c11_atomic_signal_fence: {
1216     llvm::SynchronizationScope Scope;
1217     if (BuiltinID == Builtin::BI__atomic_signal_fence ||
1218         BuiltinID == Builtin::BI__c11_atomic_signal_fence)
1219       Scope = llvm::SingleThread;
1220     else
1221       Scope = llvm::CrossThread;
1222     Value *Order = EmitScalarExpr(E->getArg(0));
1223     if (isa<llvm::ConstantInt>(Order)) {
1224       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1225       switch (ord) {
1226       case 0:  // memory_order_relaxed
1227       default: // invalid order
1228         break;
1229       case 1:  // memory_order_consume
1230       case 2:  // memory_order_acquire
1231         Builder.CreateFence(llvm::Acquire, Scope);
1232         break;
1233       case 3:  // memory_order_release
1234         Builder.CreateFence(llvm::Release, Scope);
1235         break;
1236       case 4:  // memory_order_acq_rel
1237         Builder.CreateFence(llvm::AcquireRelease, Scope);
1238         break;
1239       case 5:  // memory_order_seq_cst
1240         Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1241         break;
1242       }
1243       return RValue::get(0);
1244     }
1245 
1246     llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
1247     AcquireBB = createBasicBlock("acquire", CurFn);
1248     ReleaseBB = createBasicBlock("release", CurFn);
1249     AcqRelBB = createBasicBlock("acqrel", CurFn);
1250     SeqCstBB = createBasicBlock("seqcst", CurFn);
1251     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1252 
1253     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1254     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
1255 
1256     Builder.SetInsertPoint(AcquireBB);
1257     Builder.CreateFence(llvm::Acquire, Scope);
1258     Builder.CreateBr(ContBB);
1259     SI->addCase(Builder.getInt32(1), AcquireBB);
1260     SI->addCase(Builder.getInt32(2), AcquireBB);
1261 
1262     Builder.SetInsertPoint(ReleaseBB);
1263     Builder.CreateFence(llvm::Release, Scope);
1264     Builder.CreateBr(ContBB);
1265     SI->addCase(Builder.getInt32(3), ReleaseBB);
1266 
1267     Builder.SetInsertPoint(AcqRelBB);
1268     Builder.CreateFence(llvm::AcquireRelease, Scope);
1269     Builder.CreateBr(ContBB);
1270     SI->addCase(Builder.getInt32(4), AcqRelBB);
1271 
1272     Builder.SetInsertPoint(SeqCstBB);
1273     Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1274     Builder.CreateBr(ContBB);
1275     SI->addCase(Builder.getInt32(5), SeqCstBB);
1276 
1277     Builder.SetInsertPoint(ContBB);
1278     return RValue::get(0);
1279   }
1280 
1281     // Library functions with special handling.
1282   case Builtin::BIsqrt:
1283   case Builtin::BIsqrtf:
1284   case Builtin::BIsqrtl: {
1285     // TODO: there is currently no set of optimizer flags
1286     // sufficient for us to rewrite sqrt to @llvm.sqrt.
1287     // -fmath-errno=0 is not good enough; we need finiteness.
1288     // We could probably precondition the call with an ult
1289     // against 0, but is that worth the complexity?
1290     break;
1291   }
1292 
1293   case Builtin::BIpow:
1294   case Builtin::BIpowf:
1295   case Builtin::BIpowl: {
1296     // Rewrite sqrt to intrinsic if allowed.
1297     if (!FD->hasAttr<ConstAttr>())
1298       break;
1299     Value *Base = EmitScalarExpr(E->getArg(0));
1300     Value *Exponent = EmitScalarExpr(E->getArg(1));
1301     llvm::Type *ArgType = Base->getType();
1302     Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType);
1303     return RValue::get(Builder.CreateCall2(F, Base, Exponent));
1304   }
1305 
1306   case Builtin::BIfma:
1307   case Builtin::BIfmaf:
1308   case Builtin::BIfmal:
1309   case Builtin::BI__builtin_fma:
1310   case Builtin::BI__builtin_fmaf:
1311   case Builtin::BI__builtin_fmal: {
1312     // Rewrite fma to intrinsic.
1313     Value *FirstArg = EmitScalarExpr(E->getArg(0));
1314     llvm::Type *ArgType = FirstArg->getType();
1315     Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType);
1316     return RValue::get(Builder.CreateCall3(F, FirstArg,
1317                                               EmitScalarExpr(E->getArg(1)),
1318                                               EmitScalarExpr(E->getArg(2))));
1319   }
1320 
1321   case Builtin::BI__builtin_signbit:
1322   case Builtin::BI__builtin_signbitf:
1323   case Builtin::BI__builtin_signbitl: {
1324     LLVMContext &C = CGM.getLLVMContext();
1325 
1326     Value *Arg = EmitScalarExpr(E->getArg(0));
1327     llvm::Type *ArgTy = Arg->getType();
1328     if (ArgTy->isPPC_FP128Ty())
1329       break; // FIXME: I'm not sure what the right implementation is here.
1330     int ArgWidth = ArgTy->getPrimitiveSizeInBits();
1331     llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth);
1332     Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy);
1333     Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy);
1334     Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp);
1335     return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType())));
1336   }
1337   case Builtin::BI__builtin_annotation: {
1338     llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
1339     llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
1340                                       AnnVal->getType());
1341 
1342     // Get the annotation string, go through casts. Sema requires this to be a
1343     // non-wide string literal, potentially casted, so the cast<> is safe.
1344     const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
1345     StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
1346     return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
1347   }
1348   case Builtin::BI__builtin_addcs:
1349   case Builtin::BI__builtin_addc:
1350   case Builtin::BI__builtin_addcl:
1351   case Builtin::BI__builtin_addcll:
1352   case Builtin::BI__builtin_subcs:
1353   case Builtin::BI__builtin_subc:
1354   case Builtin::BI__builtin_subcl:
1355   case Builtin::BI__builtin_subcll: {
1356 
1357     // We translate all of these builtins from expressions of the form:
1358     //   int x = ..., y = ..., carryin = ..., carryout, result;
1359     //   result = __builtin_addc(x, y, carryin, &carryout);
1360     //
1361     // to LLVM IR of the form:
1362     //
1363     //   %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
1364     //   %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
1365     //   %carry1 = extractvalue {i32, i1} %tmp1, 1
1366     //   %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
1367     //                                                       i32 %carryin)
1368     //   %result = extractvalue {i32, i1} %tmp2, 0
1369     //   %carry2 = extractvalue {i32, i1} %tmp2, 1
1370     //   %tmp3 = or i1 %carry1, %carry2
1371     //   %tmp4 = zext i1 %tmp3 to i32
1372     //   store i32 %tmp4, i32* %carryout
1373 
1374     // Scalarize our inputs.
1375     llvm::Value *X = EmitScalarExpr(E->getArg(0));
1376     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1377     llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
1378     std::pair<llvm::Value*, unsigned> CarryOutPtr =
1379       EmitPointerWithAlignment(E->getArg(3));
1380 
1381     // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
1382     llvm::Intrinsic::ID IntrinsicId;
1383     switch (BuiltinID) {
1384     default: llvm_unreachable("Unknown multiprecision builtin id.");
1385     case Builtin::BI__builtin_addcs:
1386     case Builtin::BI__builtin_addc:
1387     case Builtin::BI__builtin_addcl:
1388     case Builtin::BI__builtin_addcll:
1389       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1390       break;
1391     case Builtin::BI__builtin_subcs:
1392     case Builtin::BI__builtin_subc:
1393     case Builtin::BI__builtin_subcl:
1394     case Builtin::BI__builtin_subcll:
1395       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1396       break;
1397     }
1398 
1399     // Construct our resulting LLVM IR expression.
1400     llvm::Value *Carry1;
1401     llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
1402                                               X, Y, Carry1);
1403     llvm::Value *Carry2;
1404     llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
1405                                               Sum1, Carryin, Carry2);
1406     llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
1407                                                X->getType());
1408     llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut,
1409                                                          CarryOutPtr.first);
1410     CarryOutStore->setAlignment(CarryOutPtr.second);
1411     return RValue::get(Sum2);
1412   }
1413   case Builtin::BI__noop:
1414     return RValue::get(0);
1415   }
1416 
1417   // If this is an alias for a lib function (e.g. __builtin_sin), emit
1418   // the call using the normal call path, but using the unmangled
1419   // version of the function name.
1420   if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
1421     return emitLibraryCall(*this, FD, E,
1422                            CGM.getBuiltinLibFunction(FD, BuiltinID));
1423 
1424   // If this is a predefined lib function (e.g. malloc), emit the call
1425   // using exactly the normal call path.
1426   if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
1427     return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee()));
1428 
1429   // See if we have a target specific intrinsic.
1430   const char *Name = getContext().BuiltinInfo.GetName(BuiltinID);
1431   Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
1432   if (const char *Prefix =
1433       llvm::Triple::getArchTypePrefix(Target.getTriple().getArch()))
1434     IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name);
1435 
1436   if (IntrinsicID != Intrinsic::not_intrinsic) {
1437     SmallVector<Value*, 16> Args;
1438 
1439     // Find out if any arguments are required to be integer constant
1440     // expressions.
1441     unsigned ICEArguments = 0;
1442     ASTContext::GetBuiltinTypeError Error;
1443     getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1444     assert(Error == ASTContext::GE_None && "Should not codegen an error");
1445 
1446     Function *F = CGM.getIntrinsic(IntrinsicID);
1447     llvm::FunctionType *FTy = F->getFunctionType();
1448 
1449     for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
1450       Value *ArgValue;
1451       // If this is a normal argument, just emit it as a scalar.
1452       if ((ICEArguments & (1 << i)) == 0) {
1453         ArgValue = EmitScalarExpr(E->getArg(i));
1454       } else {
1455         // If this is required to be a constant, constant fold it so that we
1456         // know that the generated intrinsic gets a ConstantInt.
1457         llvm::APSInt Result;
1458         bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
1459         assert(IsConst && "Constant arg isn't actually constant?");
1460         (void)IsConst;
1461         ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
1462       }
1463 
1464       // If the intrinsic arg type is different from the builtin arg type
1465       // we need to do a bit cast.
1466       llvm::Type *PTy = FTy->getParamType(i);
1467       if (PTy != ArgValue->getType()) {
1468         assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
1469                "Must be able to losslessly bit cast to param");
1470         ArgValue = Builder.CreateBitCast(ArgValue, PTy);
1471       }
1472 
1473       Args.push_back(ArgValue);
1474     }
1475 
1476     Value *V = Builder.CreateCall(F, Args);
1477     QualType BuiltinRetType = E->getType();
1478 
1479     llvm::Type *RetTy = VoidTy;
1480     if (!BuiltinRetType->isVoidType())
1481       RetTy = ConvertType(BuiltinRetType);
1482 
1483     if (RetTy != V->getType()) {
1484       assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
1485              "Must be able to losslessly bit cast result type");
1486       V = Builder.CreateBitCast(V, RetTy);
1487     }
1488 
1489     return RValue::get(V);
1490   }
1491 
1492   // See if we have a target specific builtin that needs to be lowered.
1493   if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
1494     return RValue::get(V);
1495 
1496   ErrorUnsupported(E, "builtin function");
1497 
1498   // Unknown builtin, for now just dump it out and return undef.
1499   return GetUndefRValue(E->getType());
1500 }
1501 
EmitTargetBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1502 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
1503                                               const CallExpr *E) {
1504   switch (Target.getTriple().getArch()) {
1505   case llvm::Triple::arm:
1506   case llvm::Triple::thumb:
1507     return EmitARMBuiltinExpr(BuiltinID, E);
1508   case llvm::Triple::x86:
1509   case llvm::Triple::x86_64:
1510     return EmitX86BuiltinExpr(BuiltinID, E);
1511   case llvm::Triple::ppc:
1512   case llvm::Triple::ppc64:
1513     return EmitPPCBuiltinExpr(BuiltinID, E);
1514   default:
1515     return 0;
1516   }
1517 }
1518 
GetNeonType(CodeGenFunction * CGF,NeonTypeFlags TypeFlags)1519 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
1520                                      NeonTypeFlags TypeFlags) {
1521   int IsQuad = TypeFlags.isQuad();
1522   switch (TypeFlags.getEltType()) {
1523   case NeonTypeFlags::Int8:
1524   case NeonTypeFlags::Poly8:
1525     return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad);
1526   case NeonTypeFlags::Int16:
1527   case NeonTypeFlags::Poly16:
1528   case NeonTypeFlags::Float16:
1529     return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad);
1530   case NeonTypeFlags::Int32:
1531     return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad);
1532   case NeonTypeFlags::Int64:
1533     return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad);
1534   case NeonTypeFlags::Float32:
1535     return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad);
1536   }
1537   llvm_unreachable("Invalid NeonTypeFlags element type!");
1538 }
1539 
EmitNeonSplat(Value * V,Constant * C)1540 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
1541   unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements();
1542   Value* SV = llvm::ConstantVector::getSplat(nElts, C);
1543   return Builder.CreateShuffleVector(V, V, SV, "lane");
1544 }
1545 
EmitNeonCall(Function * F,SmallVectorImpl<Value * > & Ops,const char * name,unsigned shift,bool rightshift)1546 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
1547                                      const char *name,
1548                                      unsigned shift, bool rightshift) {
1549   unsigned j = 0;
1550   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
1551        ai != ae; ++ai, ++j)
1552     if (shift > 0 && shift == j)
1553       Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
1554     else
1555       Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
1556 
1557   return Builder.CreateCall(F, Ops, name);
1558 }
1559 
EmitNeonShiftVector(Value * V,llvm::Type * Ty,bool neg)1560 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
1561                                             bool neg) {
1562   int SV = cast<ConstantInt>(V)->getSExtValue();
1563 
1564   llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1565   llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV);
1566   return llvm::ConstantVector::getSplat(VTy->getNumElements(), C);
1567 }
1568 
1569 /// GetPointeeAlignment - Given an expression with a pointer type, find the
1570 /// alignment of the type referenced by the pointer.  Skip over implicit
1571 /// casts.
1572 std::pair<llvm::Value*, unsigned>
EmitPointerWithAlignment(const Expr * Addr)1573 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) {
1574   assert(Addr->getType()->isPointerType());
1575   Addr = Addr->IgnoreParens();
1576   if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) {
1577     if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) &&
1578         ICE->getSubExpr()->getType()->isPointerType()) {
1579       std::pair<llvm::Value*, unsigned> Ptr =
1580           EmitPointerWithAlignment(ICE->getSubExpr());
1581       Ptr.first = Builder.CreateBitCast(Ptr.first,
1582                                         ConvertType(Addr->getType()));
1583       return Ptr;
1584     } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) {
1585       LValue LV = EmitLValue(ICE->getSubExpr());
1586       unsigned Align = LV.getAlignment().getQuantity();
1587       if (!Align) {
1588         // FIXME: Once LValues are fixed to always set alignment,
1589         // zap this code.
1590         QualType PtTy = ICE->getSubExpr()->getType();
1591         if (!PtTy->isIncompleteType())
1592           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1593         else
1594           Align = 1;
1595       }
1596       return std::make_pair(LV.getAddress(), Align);
1597     }
1598   }
1599   if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) {
1600     if (UO->getOpcode() == UO_AddrOf) {
1601       LValue LV = EmitLValue(UO->getSubExpr());
1602       unsigned Align = LV.getAlignment().getQuantity();
1603       if (!Align) {
1604         // FIXME: Once LValues are fixed to always set alignment,
1605         // zap this code.
1606         QualType PtTy = UO->getSubExpr()->getType();
1607         if (!PtTy->isIncompleteType())
1608           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1609         else
1610           Align = 1;
1611       }
1612       return std::make_pair(LV.getAddress(), Align);
1613     }
1614   }
1615 
1616   unsigned Align = 1;
1617   QualType PtTy = Addr->getType()->getPointeeType();
1618   if (!PtTy->isIncompleteType())
1619     Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1620 
1621   return std::make_pair(EmitScalarExpr(Addr), Align);
1622 }
1623 
EmitARMBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1624 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
1625                                            const CallExpr *E) {
1626   if (BuiltinID == ARM::BI__clear_cache) {
1627     const FunctionDecl *FD = E->getDirectCallee();
1628     // Oddly people write this call without args on occasion and gcc accepts
1629     // it - it's also marked as varargs in the description file.
1630     SmallVector<Value*, 2> Ops;
1631     for (unsigned i = 0; i < E->getNumArgs(); i++)
1632       Ops.push_back(EmitScalarExpr(E->getArg(i)));
1633     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
1634     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
1635     StringRef Name = FD->getName();
1636     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
1637   }
1638 
1639   if (BuiltinID == ARM::BI__builtin_arm_ldrexd) {
1640     Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
1641 
1642     Value *LdPtr = EmitScalarExpr(E->getArg(0));
1643     Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd");
1644 
1645     Value *Val0 = Builder.CreateExtractValue(Val, 1);
1646     Value *Val1 = Builder.CreateExtractValue(Val, 0);
1647     Val0 = Builder.CreateZExt(Val0, Int64Ty);
1648     Val1 = Builder.CreateZExt(Val1, Int64Ty);
1649 
1650     Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
1651     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
1652     return Builder.CreateOr(Val, Val1);
1653   }
1654 
1655   if (BuiltinID == ARM::BI__builtin_arm_strexd) {
1656     Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd);
1657     llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL);
1658 
1659     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
1660     Value *Tmp = Builder.CreateAlloca(Int64Ty, One);
1661     Value *Val = EmitScalarExpr(E->getArg(0));
1662     Builder.CreateStore(Val, Tmp);
1663 
1664     Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
1665     Val = Builder.CreateLoad(LdPtr);
1666 
1667     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
1668     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
1669     Value *StPtr = EmitScalarExpr(E->getArg(1));
1670     return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd");
1671   }
1672 
1673   SmallVector<Value*, 4> Ops;
1674   llvm::Value *Align = 0;
1675   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
1676     if (i == 0) {
1677       switch (BuiltinID) {
1678       case ARM::BI__builtin_neon_vld1_v:
1679       case ARM::BI__builtin_neon_vld1q_v:
1680       case ARM::BI__builtin_neon_vld1q_lane_v:
1681       case ARM::BI__builtin_neon_vld1_lane_v:
1682       case ARM::BI__builtin_neon_vld1_dup_v:
1683       case ARM::BI__builtin_neon_vld1q_dup_v:
1684       case ARM::BI__builtin_neon_vst1_v:
1685       case ARM::BI__builtin_neon_vst1q_v:
1686       case ARM::BI__builtin_neon_vst1q_lane_v:
1687       case ARM::BI__builtin_neon_vst1_lane_v:
1688       case ARM::BI__builtin_neon_vst2_v:
1689       case ARM::BI__builtin_neon_vst2q_v:
1690       case ARM::BI__builtin_neon_vst2_lane_v:
1691       case ARM::BI__builtin_neon_vst2q_lane_v:
1692       case ARM::BI__builtin_neon_vst3_v:
1693       case ARM::BI__builtin_neon_vst3q_v:
1694       case ARM::BI__builtin_neon_vst3_lane_v:
1695       case ARM::BI__builtin_neon_vst3q_lane_v:
1696       case ARM::BI__builtin_neon_vst4_v:
1697       case ARM::BI__builtin_neon_vst4q_v:
1698       case ARM::BI__builtin_neon_vst4_lane_v:
1699       case ARM::BI__builtin_neon_vst4q_lane_v:
1700         // Get the alignment for the argument in addition to the value;
1701         // we'll use it later.
1702         std::pair<llvm::Value*, unsigned> Src =
1703             EmitPointerWithAlignment(E->getArg(0));
1704         Ops.push_back(Src.first);
1705         Align = Builder.getInt32(Src.second);
1706         continue;
1707       }
1708     }
1709     if (i == 1) {
1710       switch (BuiltinID) {
1711       case ARM::BI__builtin_neon_vld2_v:
1712       case ARM::BI__builtin_neon_vld2q_v:
1713       case ARM::BI__builtin_neon_vld3_v:
1714       case ARM::BI__builtin_neon_vld3q_v:
1715       case ARM::BI__builtin_neon_vld4_v:
1716       case ARM::BI__builtin_neon_vld4q_v:
1717       case ARM::BI__builtin_neon_vld2_lane_v:
1718       case ARM::BI__builtin_neon_vld2q_lane_v:
1719       case ARM::BI__builtin_neon_vld3_lane_v:
1720       case ARM::BI__builtin_neon_vld3q_lane_v:
1721       case ARM::BI__builtin_neon_vld4_lane_v:
1722       case ARM::BI__builtin_neon_vld4q_lane_v:
1723       case ARM::BI__builtin_neon_vld2_dup_v:
1724       case ARM::BI__builtin_neon_vld3_dup_v:
1725       case ARM::BI__builtin_neon_vld4_dup_v:
1726         // Get the alignment for the argument in addition to the value;
1727         // we'll use it later.
1728         std::pair<llvm::Value*, unsigned> Src =
1729             EmitPointerWithAlignment(E->getArg(1));
1730         Ops.push_back(Src.first);
1731         Align = Builder.getInt32(Src.second);
1732         continue;
1733       }
1734     }
1735     Ops.push_back(EmitScalarExpr(E->getArg(i)));
1736   }
1737 
1738   // vget_lane and vset_lane are not overloaded and do not have an extra
1739   // argument that specifies the vector type.
1740   switch (BuiltinID) {
1741   default: break;
1742   case ARM::BI__builtin_neon_vget_lane_i8:
1743   case ARM::BI__builtin_neon_vget_lane_i16:
1744   case ARM::BI__builtin_neon_vget_lane_i32:
1745   case ARM::BI__builtin_neon_vget_lane_i64:
1746   case ARM::BI__builtin_neon_vget_lane_f32:
1747   case ARM::BI__builtin_neon_vgetq_lane_i8:
1748   case ARM::BI__builtin_neon_vgetq_lane_i16:
1749   case ARM::BI__builtin_neon_vgetq_lane_i32:
1750   case ARM::BI__builtin_neon_vgetq_lane_i64:
1751   case ARM::BI__builtin_neon_vgetq_lane_f32:
1752     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
1753                                         "vget_lane");
1754   case ARM::BI__builtin_neon_vset_lane_i8:
1755   case ARM::BI__builtin_neon_vset_lane_i16:
1756   case ARM::BI__builtin_neon_vset_lane_i32:
1757   case ARM::BI__builtin_neon_vset_lane_i64:
1758   case ARM::BI__builtin_neon_vset_lane_f32:
1759   case ARM::BI__builtin_neon_vsetq_lane_i8:
1760   case ARM::BI__builtin_neon_vsetq_lane_i16:
1761   case ARM::BI__builtin_neon_vsetq_lane_i32:
1762   case ARM::BI__builtin_neon_vsetq_lane_i64:
1763   case ARM::BI__builtin_neon_vsetq_lane_f32:
1764     Ops.push_back(EmitScalarExpr(E->getArg(2)));
1765     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
1766   }
1767 
1768   // Get the last argument, which specifies the vector type.
1769   llvm::APSInt Result;
1770   const Expr *Arg = E->getArg(E->getNumArgs()-1);
1771   if (!Arg->isIntegerConstantExpr(Result, getContext()))
1772     return 0;
1773 
1774   if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
1775       BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
1776     // Determine the overloaded type of this builtin.
1777     llvm::Type *Ty;
1778     if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
1779       Ty = FloatTy;
1780     else
1781       Ty = DoubleTy;
1782 
1783     // Determine whether this is an unsigned conversion or not.
1784     bool usgn = Result.getZExtValue() == 1;
1785     unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
1786 
1787     // Call the appropriate intrinsic.
1788     Function *F = CGM.getIntrinsic(Int, Ty);
1789     return Builder.CreateCall(F, Ops, "vcvtr");
1790   }
1791 
1792   // Determine the type of this overloaded NEON intrinsic.
1793   NeonTypeFlags Type(Result.getZExtValue());
1794   bool usgn = Type.isUnsigned();
1795   bool quad = Type.isQuad();
1796   bool rightShift = false;
1797 
1798   llvm::VectorType *VTy = GetNeonType(this, Type);
1799   llvm::Type *Ty = VTy;
1800   if (!Ty)
1801     return 0;
1802 
1803   unsigned Int;
1804   switch (BuiltinID) {
1805   default: return 0;
1806   case ARM::BI__builtin_neon_vbsl_v:
1807   case ARM::BI__builtin_neon_vbslq_v:
1808     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty),
1809                         Ops, "vbsl");
1810   case ARM::BI__builtin_neon_vabd_v:
1811   case ARM::BI__builtin_neon_vabdq_v:
1812     Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds;
1813     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
1814   case ARM::BI__builtin_neon_vabs_v:
1815   case ARM::BI__builtin_neon_vabsq_v:
1816     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty),
1817                         Ops, "vabs");
1818   case ARM::BI__builtin_neon_vaddhn_v:
1819     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty),
1820                         Ops, "vaddhn");
1821   case ARM::BI__builtin_neon_vcale_v:
1822     std::swap(Ops[0], Ops[1]);
1823   case ARM::BI__builtin_neon_vcage_v: {
1824     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged);
1825     return EmitNeonCall(F, Ops, "vcage");
1826   }
1827   case ARM::BI__builtin_neon_vcaleq_v:
1828     std::swap(Ops[0], Ops[1]);
1829   case ARM::BI__builtin_neon_vcageq_v: {
1830     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq);
1831     return EmitNeonCall(F, Ops, "vcage");
1832   }
1833   case ARM::BI__builtin_neon_vcalt_v:
1834     std::swap(Ops[0], Ops[1]);
1835   case ARM::BI__builtin_neon_vcagt_v: {
1836     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd);
1837     return EmitNeonCall(F, Ops, "vcagt");
1838   }
1839   case ARM::BI__builtin_neon_vcaltq_v:
1840     std::swap(Ops[0], Ops[1]);
1841   case ARM::BI__builtin_neon_vcagtq_v: {
1842     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq);
1843     return EmitNeonCall(F, Ops, "vcagt");
1844   }
1845   case ARM::BI__builtin_neon_vcls_v:
1846   case ARM::BI__builtin_neon_vclsq_v: {
1847     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty);
1848     return EmitNeonCall(F, Ops, "vcls");
1849   }
1850   case ARM::BI__builtin_neon_vclz_v:
1851   case ARM::BI__builtin_neon_vclzq_v: {
1852     // Generate target-independent intrinsic; also need to add second argument
1853     // for whether or not clz of zero is undefined; on ARM it isn't.
1854     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty);
1855     Ops.push_back(Builder.getInt1(Target.isCLZForZeroUndef()));
1856     return EmitNeonCall(F, Ops, "vclz");
1857   }
1858   case ARM::BI__builtin_neon_vcnt_v:
1859   case ARM::BI__builtin_neon_vcntq_v: {
1860     // generate target-independent intrinsic
1861     Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty);
1862     return EmitNeonCall(F, Ops, "vctpop");
1863   }
1864   case ARM::BI__builtin_neon_vcvt_f16_v: {
1865     assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad &&
1866            "unexpected vcvt_f16_v builtin");
1867     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf);
1868     return EmitNeonCall(F, Ops, "vcvt");
1869   }
1870   case ARM::BI__builtin_neon_vcvt_f32_f16: {
1871     assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad &&
1872            "unexpected vcvt_f32_f16 builtin");
1873     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp);
1874     return EmitNeonCall(F, Ops, "vcvt");
1875   }
1876   case ARM::BI__builtin_neon_vcvt_f32_v:
1877   case ARM::BI__builtin_neon_vcvtq_f32_v:
1878     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1879     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
1880     return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
1881                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
1882   case ARM::BI__builtin_neon_vcvt_s32_v:
1883   case ARM::BI__builtin_neon_vcvt_u32_v:
1884   case ARM::BI__builtin_neon_vcvtq_s32_v:
1885   case ARM::BI__builtin_neon_vcvtq_u32_v: {
1886     llvm::Type *FloatTy =
1887       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
1888     Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
1889     return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
1890                 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
1891   }
1892   case ARM::BI__builtin_neon_vcvt_n_f32_v:
1893   case ARM::BI__builtin_neon_vcvtq_n_f32_v: {
1894     llvm::Type *FloatTy =
1895       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
1896     llvm::Type *Tys[2] = { FloatTy, Ty };
1897     Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp
1898                : Intrinsic::arm_neon_vcvtfxs2fp;
1899     Function *F = CGM.getIntrinsic(Int, Tys);
1900     return EmitNeonCall(F, Ops, "vcvt_n");
1901   }
1902   case ARM::BI__builtin_neon_vcvt_n_s32_v:
1903   case ARM::BI__builtin_neon_vcvt_n_u32_v:
1904   case ARM::BI__builtin_neon_vcvtq_n_s32_v:
1905   case ARM::BI__builtin_neon_vcvtq_n_u32_v: {
1906     llvm::Type *FloatTy =
1907       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
1908     llvm::Type *Tys[2] = { Ty, FloatTy };
1909     Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu
1910                : Intrinsic::arm_neon_vcvtfp2fxs;
1911     Function *F = CGM.getIntrinsic(Int, Tys);
1912     return EmitNeonCall(F, Ops, "vcvt_n");
1913   }
1914   case ARM::BI__builtin_neon_vext_v:
1915   case ARM::BI__builtin_neon_vextq_v: {
1916     int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
1917     SmallVector<Constant*, 16> Indices;
1918     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
1919       Indices.push_back(ConstantInt::get(Int32Ty, i+CV));
1920 
1921     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1922     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1923     Value *SV = llvm::ConstantVector::get(Indices);
1924     return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext");
1925   }
1926   case ARM::BI__builtin_neon_vhadd_v:
1927   case ARM::BI__builtin_neon_vhaddq_v:
1928     Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds;
1929     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd");
1930   case ARM::BI__builtin_neon_vhsub_v:
1931   case ARM::BI__builtin_neon_vhsubq_v:
1932     Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs;
1933     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub");
1934   case ARM::BI__builtin_neon_vld1_v:
1935   case ARM::BI__builtin_neon_vld1q_v:
1936     Ops.push_back(Align);
1937     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty),
1938                         Ops, "vld1");
1939   case ARM::BI__builtin_neon_vld1q_lane_v:
1940     // Handle 64-bit integer elements as a special case.  Use shuffles of
1941     // one-element vectors to avoid poor code for i64 in the backend.
1942     if (VTy->getElementType()->isIntegerTy(64)) {
1943       // Extract the other lane.
1944       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1945       int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
1946       Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
1947       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
1948       // Load the value as a one-element vector.
1949       Ty = llvm::VectorType::get(VTy->getElementType(), 1);
1950       Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty);
1951       Value *Ld = Builder.CreateCall2(F, Ops[0], Align);
1952       // Combine them.
1953       SmallVector<Constant*, 2> Indices;
1954       Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane));
1955       Indices.push_back(ConstantInt::get(Int32Ty, Lane));
1956       SV = llvm::ConstantVector::get(Indices);
1957       return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane");
1958     }
1959     // fall through
1960   case ARM::BI__builtin_neon_vld1_lane_v: {
1961     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1962     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
1963     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1964     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
1965     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
1966     return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
1967   }
1968   case ARM::BI__builtin_neon_vld1_dup_v:
1969   case ARM::BI__builtin_neon_vld1q_dup_v: {
1970     Value *V = UndefValue::get(Ty);
1971     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
1972     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1973     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
1974     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
1975     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
1976     Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
1977     return EmitNeonSplat(Ops[0], CI);
1978   }
1979   case ARM::BI__builtin_neon_vld2_v:
1980   case ARM::BI__builtin_neon_vld2q_v: {
1981     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty);
1982     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2");
1983     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1984     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1985     return Builder.CreateStore(Ops[1], Ops[0]);
1986   }
1987   case ARM::BI__builtin_neon_vld3_v:
1988   case ARM::BI__builtin_neon_vld3q_v: {
1989     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty);
1990     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3");
1991     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
1992     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1993     return Builder.CreateStore(Ops[1], Ops[0]);
1994   }
1995   case ARM::BI__builtin_neon_vld4_v:
1996   case ARM::BI__builtin_neon_vld4q_v: {
1997     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty);
1998     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4");
1999     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2000     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2001     return Builder.CreateStore(Ops[1], Ops[0]);
2002   }
2003   case ARM::BI__builtin_neon_vld2_lane_v:
2004   case ARM::BI__builtin_neon_vld2q_lane_v: {
2005     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty);
2006     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2007     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2008     Ops.push_back(Align);
2009     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
2010     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2011     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2012     return Builder.CreateStore(Ops[1], Ops[0]);
2013   }
2014   case ARM::BI__builtin_neon_vld3_lane_v:
2015   case ARM::BI__builtin_neon_vld3q_lane_v: {
2016     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty);
2017     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2018     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2019     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
2020     Ops.push_back(Align);
2021     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
2022     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2023     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2024     return Builder.CreateStore(Ops[1], Ops[0]);
2025   }
2026   case ARM::BI__builtin_neon_vld4_lane_v:
2027   case ARM::BI__builtin_neon_vld4q_lane_v: {
2028     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty);
2029     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2030     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2031     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
2032     Ops[5] = Builder.CreateBitCast(Ops[5], Ty);
2033     Ops.push_back(Align);
2034     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
2035     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2036     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2037     return Builder.CreateStore(Ops[1], Ops[0]);
2038   }
2039   case ARM::BI__builtin_neon_vld2_dup_v:
2040   case ARM::BI__builtin_neon_vld3_dup_v:
2041   case ARM::BI__builtin_neon_vld4_dup_v: {
2042     // Handle 64-bit elements as a special-case.  There is no "dup" needed.
2043     if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
2044       switch (BuiltinID) {
2045       case ARM::BI__builtin_neon_vld2_dup_v:
2046         Int = Intrinsic::arm_neon_vld2;
2047         break;
2048       case ARM::BI__builtin_neon_vld3_dup_v:
2049         Int = Intrinsic::arm_neon_vld3;
2050         break;
2051       case ARM::BI__builtin_neon_vld4_dup_v:
2052         Int = Intrinsic::arm_neon_vld4;
2053         break;
2054       default: llvm_unreachable("unknown vld_dup intrinsic?");
2055       }
2056       Function *F = CGM.getIntrinsic(Int, Ty);
2057       Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup");
2058       Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2059       Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2060       return Builder.CreateStore(Ops[1], Ops[0]);
2061     }
2062     switch (BuiltinID) {
2063     case ARM::BI__builtin_neon_vld2_dup_v:
2064       Int = Intrinsic::arm_neon_vld2lane;
2065       break;
2066     case ARM::BI__builtin_neon_vld3_dup_v:
2067       Int = Intrinsic::arm_neon_vld3lane;
2068       break;
2069     case ARM::BI__builtin_neon_vld4_dup_v:
2070       Int = Intrinsic::arm_neon_vld4lane;
2071       break;
2072     default: llvm_unreachable("unknown vld_dup intrinsic?");
2073     }
2074     Function *F = CGM.getIntrinsic(Int, Ty);
2075     llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
2076 
2077     SmallVector<Value*, 6> Args;
2078     Args.push_back(Ops[1]);
2079     Args.append(STy->getNumElements(), UndefValue::get(Ty));
2080 
2081     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
2082     Args.push_back(CI);
2083     Args.push_back(Align);
2084 
2085     Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
2086     // splat lane 0 to all elts in each vector of the result.
2087     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2088       Value *Val = Builder.CreateExtractValue(Ops[1], i);
2089       Value *Elt = Builder.CreateBitCast(Val, Ty);
2090       Elt = EmitNeonSplat(Elt, CI);
2091       Elt = Builder.CreateBitCast(Elt, Val->getType());
2092       Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
2093     }
2094     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2095     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2096     return Builder.CreateStore(Ops[1], Ops[0]);
2097   }
2098   case ARM::BI__builtin_neon_vmax_v:
2099   case ARM::BI__builtin_neon_vmaxq_v:
2100     Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs;
2101     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
2102   case ARM::BI__builtin_neon_vmin_v:
2103   case ARM::BI__builtin_neon_vminq_v:
2104     Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins;
2105     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
2106   case ARM::BI__builtin_neon_vmovl_v: {
2107     llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
2108     Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
2109     if (usgn)
2110       return Builder.CreateZExt(Ops[0], Ty, "vmovl");
2111     return Builder.CreateSExt(Ops[0], Ty, "vmovl");
2112   }
2113   case ARM::BI__builtin_neon_vmovn_v: {
2114     llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
2115     Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
2116     return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
2117   }
2118   case ARM::BI__builtin_neon_vmul_v:
2119   case ARM::BI__builtin_neon_vmulq_v:
2120     assert(Type.isPoly() && "vmul builtin only supported for polynomial types");
2121     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty),
2122                         Ops, "vmul");
2123   case ARM::BI__builtin_neon_vmull_v:
2124     Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
2125     Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
2126     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
2127   case ARM::BI__builtin_neon_vfma_v:
2128   case ARM::BI__builtin_neon_vfmaq_v: {
2129     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
2130     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2131     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2132     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2133 
2134     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
2135     return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
2136   }
2137   case ARM::BI__builtin_neon_vpadal_v:
2138   case ARM::BI__builtin_neon_vpadalq_v: {
2139     Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals;
2140     // The source operand type has twice as many elements of half the size.
2141     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2142     llvm::Type *EltTy =
2143       llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2144     llvm::Type *NarrowTy =
2145       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2146     llvm::Type *Tys[2] = { Ty, NarrowTy };
2147     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal");
2148   }
2149   case ARM::BI__builtin_neon_vpadd_v:
2150     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty),
2151                         Ops, "vpadd");
2152   case ARM::BI__builtin_neon_vpaddl_v:
2153   case ARM::BI__builtin_neon_vpaddlq_v: {
2154     Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls;
2155     // The source operand type has twice as many elements of half the size.
2156     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2157     llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2158     llvm::Type *NarrowTy =
2159       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2160     llvm::Type *Tys[2] = { Ty, NarrowTy };
2161     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
2162   }
2163   case ARM::BI__builtin_neon_vpmax_v:
2164     Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs;
2165     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
2166   case ARM::BI__builtin_neon_vpmin_v:
2167     Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins;
2168     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
2169   case ARM::BI__builtin_neon_vqabs_v:
2170   case ARM::BI__builtin_neon_vqabsq_v:
2171     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty),
2172                         Ops, "vqabs");
2173   case ARM::BI__builtin_neon_vqadd_v:
2174   case ARM::BI__builtin_neon_vqaddq_v:
2175     Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds;
2176     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd");
2177   case ARM::BI__builtin_neon_vqdmlal_v:
2178     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty),
2179                         Ops, "vqdmlal");
2180   case ARM::BI__builtin_neon_vqdmlsl_v:
2181     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty),
2182                         Ops, "vqdmlsl");
2183   case ARM::BI__builtin_neon_vqdmulh_v:
2184   case ARM::BI__builtin_neon_vqdmulhq_v:
2185     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty),
2186                         Ops, "vqdmulh");
2187   case ARM::BI__builtin_neon_vqdmull_v:
2188     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty),
2189                         Ops, "vqdmull");
2190   case ARM::BI__builtin_neon_vqmovn_v:
2191     Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns;
2192     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn");
2193   case ARM::BI__builtin_neon_vqmovun_v:
2194     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty),
2195                         Ops, "vqdmull");
2196   case ARM::BI__builtin_neon_vqneg_v:
2197   case ARM::BI__builtin_neon_vqnegq_v:
2198     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty),
2199                         Ops, "vqneg");
2200   case ARM::BI__builtin_neon_vqrdmulh_v:
2201   case ARM::BI__builtin_neon_vqrdmulhq_v:
2202     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty),
2203                         Ops, "vqrdmulh");
2204   case ARM::BI__builtin_neon_vqrshl_v:
2205   case ARM::BI__builtin_neon_vqrshlq_v:
2206     Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts;
2207     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl");
2208   case ARM::BI__builtin_neon_vqrshrn_n_v:
2209     Int =
2210       usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
2211     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
2212                         1, true);
2213   case ARM::BI__builtin_neon_vqrshrun_n_v:
2214     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
2215                         Ops, "vqrshrun_n", 1, true);
2216   case ARM::BI__builtin_neon_vqshl_v:
2217   case ARM::BI__builtin_neon_vqshlq_v:
2218     Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
2219     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl");
2220   case ARM::BI__builtin_neon_vqshl_n_v:
2221   case ARM::BI__builtin_neon_vqshlq_n_v:
2222     Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
2223     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
2224                         1, false);
2225   case ARM::BI__builtin_neon_vqshlu_n_v:
2226   case ARM::BI__builtin_neon_vqshluq_n_v:
2227     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty),
2228                         Ops, "vqshlu", 1, false);
2229   case ARM::BI__builtin_neon_vqshrn_n_v:
2230     Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
2231     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
2232                         1, true);
2233   case ARM::BI__builtin_neon_vqshrun_n_v:
2234     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
2235                         Ops, "vqshrun_n", 1, true);
2236   case ARM::BI__builtin_neon_vqsub_v:
2237   case ARM::BI__builtin_neon_vqsubq_v:
2238     Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs;
2239     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub");
2240   case ARM::BI__builtin_neon_vraddhn_v:
2241     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty),
2242                         Ops, "vraddhn");
2243   case ARM::BI__builtin_neon_vrecpe_v:
2244   case ARM::BI__builtin_neon_vrecpeq_v:
2245     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
2246                         Ops, "vrecpe");
2247   case ARM::BI__builtin_neon_vrecps_v:
2248   case ARM::BI__builtin_neon_vrecpsq_v:
2249     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty),
2250                         Ops, "vrecps");
2251   case ARM::BI__builtin_neon_vrhadd_v:
2252   case ARM::BI__builtin_neon_vrhaddq_v:
2253     Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds;
2254     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd");
2255   case ARM::BI__builtin_neon_vrshl_v:
2256   case ARM::BI__builtin_neon_vrshlq_v:
2257     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2258     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl");
2259   case ARM::BI__builtin_neon_vrshrn_n_v:
2260     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
2261                         Ops, "vrshrn_n", 1, true);
2262   case ARM::BI__builtin_neon_vrshr_n_v:
2263   case ARM::BI__builtin_neon_vrshrq_n_v:
2264     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2265     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true);
2266   case ARM::BI__builtin_neon_vrsqrte_v:
2267   case ARM::BI__builtin_neon_vrsqrteq_v:
2268     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty),
2269                         Ops, "vrsqrte");
2270   case ARM::BI__builtin_neon_vrsqrts_v:
2271   case ARM::BI__builtin_neon_vrsqrtsq_v:
2272     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty),
2273                         Ops, "vrsqrts");
2274   case ARM::BI__builtin_neon_vrsra_n_v:
2275   case ARM::BI__builtin_neon_vrsraq_n_v:
2276     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2277     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2278     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
2279     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2280     Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]);
2281     return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
2282   case ARM::BI__builtin_neon_vrsubhn_v:
2283     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty),
2284                         Ops, "vrsubhn");
2285   case ARM::BI__builtin_neon_vshl_v:
2286   case ARM::BI__builtin_neon_vshlq_v:
2287     Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts;
2288     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl");
2289   case ARM::BI__builtin_neon_vshll_n_v:
2290     Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls;
2291     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1);
2292   case ARM::BI__builtin_neon_vshl_n_v:
2293   case ARM::BI__builtin_neon_vshlq_n_v:
2294     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
2295     return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
2296                              "vshl_n");
2297   case ARM::BI__builtin_neon_vshrn_n_v:
2298     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty),
2299                         Ops, "vshrn_n", 1, true);
2300   case ARM::BI__builtin_neon_vshr_n_v:
2301   case ARM::BI__builtin_neon_vshrq_n_v:
2302     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2303     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
2304     if (usgn)
2305       return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n");
2306     else
2307       return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n");
2308   case ARM::BI__builtin_neon_vsri_n_v:
2309   case ARM::BI__builtin_neon_vsriq_n_v:
2310     rightShift = true;
2311   case ARM::BI__builtin_neon_vsli_n_v:
2312   case ARM::BI__builtin_neon_vsliq_n_v:
2313     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
2314     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
2315                         Ops, "vsli_n");
2316   case ARM::BI__builtin_neon_vsra_n_v:
2317   case ARM::BI__builtin_neon_vsraq_n_v:
2318     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2319     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2320     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false);
2321     if (usgn)
2322       Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n");
2323     else
2324       Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n");
2325     return Builder.CreateAdd(Ops[0], Ops[1]);
2326   case ARM::BI__builtin_neon_vst1_v:
2327   case ARM::BI__builtin_neon_vst1q_v:
2328     Ops.push_back(Align);
2329     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty),
2330                         Ops, "");
2331   case ARM::BI__builtin_neon_vst1q_lane_v:
2332     // Handle 64-bit integer elements as a special case.  Use a shuffle to get
2333     // a one-element vector and avoid poor code for i64 in the backend.
2334     if (VTy->getElementType()->isIntegerTy(64)) {
2335       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2336       Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
2337       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
2338       Ops[2] = Align;
2339       return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
2340                                                  Ops[1]->getType()), Ops);
2341     }
2342     // fall through
2343   case ARM::BI__builtin_neon_vst1_lane_v: {
2344     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2345     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
2346     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2347     StoreInst *St = Builder.CreateStore(Ops[1],
2348                                         Builder.CreateBitCast(Ops[0], Ty));
2349     St->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2350     return St;
2351   }
2352   case ARM::BI__builtin_neon_vst2_v:
2353   case ARM::BI__builtin_neon_vst2q_v:
2354     Ops.push_back(Align);
2355     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty),
2356                         Ops, "");
2357   case ARM::BI__builtin_neon_vst2_lane_v:
2358   case ARM::BI__builtin_neon_vst2q_lane_v:
2359     Ops.push_back(Align);
2360     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty),
2361                         Ops, "");
2362   case ARM::BI__builtin_neon_vst3_v:
2363   case ARM::BI__builtin_neon_vst3q_v:
2364     Ops.push_back(Align);
2365     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty),
2366                         Ops, "");
2367   case ARM::BI__builtin_neon_vst3_lane_v:
2368   case ARM::BI__builtin_neon_vst3q_lane_v:
2369     Ops.push_back(Align);
2370     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty),
2371                         Ops, "");
2372   case ARM::BI__builtin_neon_vst4_v:
2373   case ARM::BI__builtin_neon_vst4q_v:
2374     Ops.push_back(Align);
2375     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty),
2376                         Ops, "");
2377   case ARM::BI__builtin_neon_vst4_lane_v:
2378   case ARM::BI__builtin_neon_vst4q_lane_v:
2379     Ops.push_back(Align);
2380     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty),
2381                         Ops, "");
2382   case ARM::BI__builtin_neon_vsubhn_v:
2383     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty),
2384                         Ops, "vsubhn");
2385   case ARM::BI__builtin_neon_vtbl1_v:
2386     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
2387                         Ops, "vtbl1");
2388   case ARM::BI__builtin_neon_vtbl2_v:
2389     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
2390                         Ops, "vtbl2");
2391   case ARM::BI__builtin_neon_vtbl3_v:
2392     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
2393                         Ops, "vtbl3");
2394   case ARM::BI__builtin_neon_vtbl4_v:
2395     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
2396                         Ops, "vtbl4");
2397   case ARM::BI__builtin_neon_vtbx1_v:
2398     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
2399                         Ops, "vtbx1");
2400   case ARM::BI__builtin_neon_vtbx2_v:
2401     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
2402                         Ops, "vtbx2");
2403   case ARM::BI__builtin_neon_vtbx3_v:
2404     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
2405                         Ops, "vtbx3");
2406   case ARM::BI__builtin_neon_vtbx4_v:
2407     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
2408                         Ops, "vtbx4");
2409   case ARM::BI__builtin_neon_vtst_v:
2410   case ARM::BI__builtin_neon_vtstq_v: {
2411     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2412     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2413     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
2414     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
2415                                 ConstantAggregateZero::get(Ty));
2416     return Builder.CreateSExt(Ops[0], Ty, "vtst");
2417   }
2418   case ARM::BI__builtin_neon_vtrn_v:
2419   case ARM::BI__builtin_neon_vtrnq_v: {
2420     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2421     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2422     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2423     Value *SV = 0;
2424 
2425     for (unsigned vi = 0; vi != 2; ++vi) {
2426       SmallVector<Constant*, 16> Indices;
2427       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2428         Indices.push_back(Builder.getInt32(i+vi));
2429         Indices.push_back(Builder.getInt32(i+e+vi));
2430       }
2431       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2432       SV = llvm::ConstantVector::get(Indices);
2433       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
2434       SV = Builder.CreateStore(SV, Addr);
2435     }
2436     return SV;
2437   }
2438   case ARM::BI__builtin_neon_vuzp_v:
2439   case ARM::BI__builtin_neon_vuzpq_v: {
2440     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2441     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2442     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2443     Value *SV = 0;
2444 
2445     for (unsigned vi = 0; vi != 2; ++vi) {
2446       SmallVector<Constant*, 16> Indices;
2447       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2448         Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
2449 
2450       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2451       SV = llvm::ConstantVector::get(Indices);
2452       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
2453       SV = Builder.CreateStore(SV, Addr);
2454     }
2455     return SV;
2456   }
2457   case ARM::BI__builtin_neon_vzip_v:
2458   case ARM::BI__builtin_neon_vzipq_v: {
2459     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2460     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2461     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2462     Value *SV = 0;
2463 
2464     for (unsigned vi = 0; vi != 2; ++vi) {
2465       SmallVector<Constant*, 16> Indices;
2466       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2467         Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
2468         Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
2469       }
2470       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2471       SV = llvm::ConstantVector::get(Indices);
2472       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
2473       SV = Builder.CreateStore(SV, Addr);
2474     }
2475     return SV;
2476   }
2477   }
2478 }
2479 
2480 llvm::Value *CodeGenFunction::
BuildVector(ArrayRef<llvm::Value * > Ops)2481 BuildVector(ArrayRef<llvm::Value*> Ops) {
2482   assert((Ops.size() & (Ops.size() - 1)) == 0 &&
2483          "Not a power-of-two sized vector!");
2484   bool AllConstants = true;
2485   for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
2486     AllConstants &= isa<Constant>(Ops[i]);
2487 
2488   // If this is a constant vector, create a ConstantVector.
2489   if (AllConstants) {
2490     SmallVector<llvm::Constant*, 16> CstOps;
2491     for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2492       CstOps.push_back(cast<Constant>(Ops[i]));
2493     return llvm::ConstantVector::get(CstOps);
2494   }
2495 
2496   // Otherwise, insertelement the values to build the vector.
2497   Value *Result =
2498     llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
2499 
2500   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2501     Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
2502 
2503   return Result;
2504 }
2505 
EmitX86BuiltinExpr(unsigned BuiltinID,const CallExpr * E)2506 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
2507                                            const CallExpr *E) {
2508   SmallVector<Value*, 4> Ops;
2509 
2510   // Find out if any arguments are required to be integer constant expressions.
2511   unsigned ICEArguments = 0;
2512   ASTContext::GetBuiltinTypeError Error;
2513   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
2514   assert(Error == ASTContext::GE_None && "Should not codegen an error");
2515 
2516   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
2517     // If this is a normal argument, just emit it as a scalar.
2518     if ((ICEArguments & (1 << i)) == 0) {
2519       Ops.push_back(EmitScalarExpr(E->getArg(i)));
2520       continue;
2521     }
2522 
2523     // If this is required to be a constant, constant fold it so that we know
2524     // that the generated intrinsic gets a ConstantInt.
2525     llvm::APSInt Result;
2526     bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
2527     assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
2528     Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
2529   }
2530 
2531   switch (BuiltinID) {
2532   default: return 0;
2533   case X86::BI__builtin_ia32_vec_init_v8qi:
2534   case X86::BI__builtin_ia32_vec_init_v4hi:
2535   case X86::BI__builtin_ia32_vec_init_v2si:
2536     return Builder.CreateBitCast(BuildVector(Ops),
2537                                  llvm::Type::getX86_MMXTy(getLLVMContext()));
2538   case X86::BI__builtin_ia32_vec_ext_v2si:
2539     return Builder.CreateExtractElement(Ops[0],
2540                                   llvm::ConstantInt::get(Ops[1]->getType(), 0));
2541   case X86::BI__builtin_ia32_ldmxcsr: {
2542     llvm::Type *PtrTy = Int8PtrTy;
2543     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2544     Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2545     Builder.CreateStore(Ops[0], Tmp);
2546     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
2547                               Builder.CreateBitCast(Tmp, PtrTy));
2548   }
2549   case X86::BI__builtin_ia32_stmxcsr: {
2550     llvm::Type *PtrTy = Int8PtrTy;
2551     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2552     Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2553     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
2554                        Builder.CreateBitCast(Tmp, PtrTy));
2555     return Builder.CreateLoad(Tmp, "stmxcsr");
2556   }
2557   case X86::BI__builtin_ia32_storehps:
2558   case X86::BI__builtin_ia32_storelps: {
2559     llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty);
2560     llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2561 
2562     // cast val v2i64
2563     Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast");
2564 
2565     // extract (0, 1)
2566     unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
2567     llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index);
2568     Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
2569 
2570     // cast pointer to i64 & store
2571     Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
2572     return Builder.CreateStore(Ops[1], Ops[0]);
2573   }
2574   case X86::BI__builtin_ia32_palignr: {
2575     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2576 
2577     // If palignr is shifting the pair of input vectors less than 9 bytes,
2578     // emit a shuffle instruction.
2579     if (shiftVal <= 8) {
2580       SmallVector<llvm::Constant*, 8> Indices;
2581       for (unsigned i = 0; i != 8; ++i)
2582         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2583 
2584       Value* SV = llvm::ConstantVector::get(Indices);
2585       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2586     }
2587 
2588     // If palignr is shifting the pair of input vectors more than 8 but less
2589     // than 16 bytes, emit a logical right shift of the destination.
2590     if (shiftVal < 16) {
2591       // MMX has these as 1 x i64 vectors for some odd optimization reasons.
2592       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1);
2593 
2594       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2595       Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8);
2596 
2597       // create i32 constant
2598       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q);
2599       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2600     }
2601 
2602     // If palignr is shifting the pair of vectors more than 16 bytes, emit zero.
2603     return llvm::Constant::getNullValue(ConvertType(E->getType()));
2604   }
2605   case X86::BI__builtin_ia32_palignr128: {
2606     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2607 
2608     // If palignr is shifting the pair of input vectors less than 17 bytes,
2609     // emit a shuffle instruction.
2610     if (shiftVal <= 16) {
2611       SmallVector<llvm::Constant*, 16> Indices;
2612       for (unsigned i = 0; i != 16; ++i)
2613         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2614 
2615       Value* SV = llvm::ConstantVector::get(Indices);
2616       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2617     }
2618 
2619     // If palignr is shifting the pair of input vectors more than 16 but less
2620     // than 32 bytes, emit a logical right shift of the destination.
2621     if (shiftVal < 32) {
2622       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2623 
2624       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2625       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
2626 
2627       // create i32 constant
2628       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq);
2629       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2630     }
2631 
2632     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
2633     return llvm::Constant::getNullValue(ConvertType(E->getType()));
2634   }
2635   case X86::BI__builtin_ia32_palignr256: {
2636     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2637 
2638     // If palignr is shifting the pair of input vectors less than 17 bytes,
2639     // emit a shuffle instruction.
2640     if (shiftVal <= 16) {
2641       SmallVector<llvm::Constant*, 32> Indices;
2642       // 256-bit palignr operates on 128-bit lanes so we need to handle that
2643       for (unsigned l = 0; l != 2; ++l) {
2644         unsigned LaneStart = l * 16;
2645         unsigned LaneEnd = (l+1) * 16;
2646         for (unsigned i = 0; i != 16; ++i) {
2647           unsigned Idx = shiftVal + i + LaneStart;
2648           if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand
2649           Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx));
2650         }
2651       }
2652 
2653       Value* SV = llvm::ConstantVector::get(Indices);
2654       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2655     }
2656 
2657     // If palignr is shifting the pair of input vectors more than 16 but less
2658     // than 32 bytes, emit a logical right shift of the destination.
2659     if (shiftVal < 32) {
2660       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4);
2661 
2662       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2663       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
2664 
2665       // create i32 constant
2666       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq);
2667       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2668     }
2669 
2670     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
2671     return llvm::Constant::getNullValue(ConvertType(E->getType()));
2672   }
2673   case X86::BI__builtin_ia32_movntps:
2674   case X86::BI__builtin_ia32_movntps256:
2675   case X86::BI__builtin_ia32_movntpd:
2676   case X86::BI__builtin_ia32_movntpd256:
2677   case X86::BI__builtin_ia32_movntdq:
2678   case X86::BI__builtin_ia32_movntdq256:
2679   case X86::BI__builtin_ia32_movnti: {
2680     llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(),
2681                                            Builder.getInt32(1));
2682 
2683     // Convert the type of the pointer to a pointer to the stored type.
2684     Value *BC = Builder.CreateBitCast(Ops[0],
2685                                 llvm::PointerType::getUnqual(Ops[1]->getType()),
2686                                       "cast");
2687     StoreInst *SI = Builder.CreateStore(Ops[1], BC);
2688     SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
2689     SI->setAlignment(16);
2690     return SI;
2691   }
2692   // 3DNow!
2693   case X86::BI__builtin_ia32_pswapdsf:
2694   case X86::BI__builtin_ia32_pswapdsi: {
2695     const char *name = 0;
2696     Intrinsic::ID ID = Intrinsic::not_intrinsic;
2697     switch(BuiltinID) {
2698     default: llvm_unreachable("Unsupported intrinsic!");
2699     case X86::BI__builtin_ia32_pswapdsf:
2700     case X86::BI__builtin_ia32_pswapdsi:
2701       name = "pswapd";
2702       ID = Intrinsic::x86_3dnowa_pswapd;
2703       break;
2704     }
2705     llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
2706     Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
2707     llvm::Function *F = CGM.getIntrinsic(ID);
2708     return Builder.CreateCall(F, Ops, name);
2709   }
2710   case X86::BI__builtin_ia32_rdrand16_step:
2711   case X86::BI__builtin_ia32_rdrand32_step:
2712   case X86::BI__builtin_ia32_rdrand64_step: {
2713     Intrinsic::ID ID;
2714     switch (BuiltinID) {
2715     default: llvm_unreachable("Unsupported intrinsic!");
2716     case X86::BI__builtin_ia32_rdrand16_step:
2717       ID = Intrinsic::x86_rdrand_16;
2718       break;
2719     case X86::BI__builtin_ia32_rdrand32_step:
2720       ID = Intrinsic::x86_rdrand_32;
2721       break;
2722     case X86::BI__builtin_ia32_rdrand64_step:
2723       ID = Intrinsic::x86_rdrand_64;
2724       break;
2725     }
2726 
2727     Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
2728     Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]);
2729     return Builder.CreateExtractValue(Call, 1);
2730   }
2731   }
2732 }
2733 
2734 
EmitPPCBuiltinExpr(unsigned BuiltinID,const CallExpr * E)2735 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
2736                                            const CallExpr *E) {
2737   SmallVector<Value*, 4> Ops;
2738 
2739   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
2740     Ops.push_back(EmitScalarExpr(E->getArg(i)));
2741 
2742   Intrinsic::ID ID = Intrinsic::not_intrinsic;
2743 
2744   switch (BuiltinID) {
2745   default: return 0;
2746 
2747   // vec_ld, vec_lvsl, vec_lvsr
2748   case PPC::BI__builtin_altivec_lvx:
2749   case PPC::BI__builtin_altivec_lvxl:
2750   case PPC::BI__builtin_altivec_lvebx:
2751   case PPC::BI__builtin_altivec_lvehx:
2752   case PPC::BI__builtin_altivec_lvewx:
2753   case PPC::BI__builtin_altivec_lvsl:
2754   case PPC::BI__builtin_altivec_lvsr:
2755   {
2756     Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
2757 
2758     Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
2759     Ops.pop_back();
2760 
2761     switch (BuiltinID) {
2762     default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
2763     case PPC::BI__builtin_altivec_lvx:
2764       ID = Intrinsic::ppc_altivec_lvx;
2765       break;
2766     case PPC::BI__builtin_altivec_lvxl:
2767       ID = Intrinsic::ppc_altivec_lvxl;
2768       break;
2769     case PPC::BI__builtin_altivec_lvebx:
2770       ID = Intrinsic::ppc_altivec_lvebx;
2771       break;
2772     case PPC::BI__builtin_altivec_lvehx:
2773       ID = Intrinsic::ppc_altivec_lvehx;
2774       break;
2775     case PPC::BI__builtin_altivec_lvewx:
2776       ID = Intrinsic::ppc_altivec_lvewx;
2777       break;
2778     case PPC::BI__builtin_altivec_lvsl:
2779       ID = Intrinsic::ppc_altivec_lvsl;
2780       break;
2781     case PPC::BI__builtin_altivec_lvsr:
2782       ID = Intrinsic::ppc_altivec_lvsr;
2783       break;
2784     }
2785     llvm::Function *F = CGM.getIntrinsic(ID);
2786     return Builder.CreateCall(F, Ops, "");
2787   }
2788 
2789   // vec_st
2790   case PPC::BI__builtin_altivec_stvx:
2791   case PPC::BI__builtin_altivec_stvxl:
2792   case PPC::BI__builtin_altivec_stvebx:
2793   case PPC::BI__builtin_altivec_stvehx:
2794   case PPC::BI__builtin_altivec_stvewx:
2795   {
2796     Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
2797     Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
2798     Ops.pop_back();
2799 
2800     switch (BuiltinID) {
2801     default: llvm_unreachable("Unsupported st intrinsic!");
2802     case PPC::BI__builtin_altivec_stvx:
2803       ID = Intrinsic::ppc_altivec_stvx;
2804       break;
2805     case PPC::BI__builtin_altivec_stvxl:
2806       ID = Intrinsic::ppc_altivec_stvxl;
2807       break;
2808     case PPC::BI__builtin_altivec_stvebx:
2809       ID = Intrinsic::ppc_altivec_stvebx;
2810       break;
2811     case PPC::BI__builtin_altivec_stvehx:
2812       ID = Intrinsic::ppc_altivec_stvehx;
2813       break;
2814     case PPC::BI__builtin_altivec_stvewx:
2815       ID = Intrinsic::ppc_altivec_stvewx;
2816       break;
2817     }
2818     llvm::Function *F = CGM.getIntrinsic(ID);
2819     return Builder.CreateCall(F, Ops, "");
2820   }
2821   }
2822 }
2823