• 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(getTarget().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(getTarget().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     // Transform a call to pow* into a @llvm.pow.* intrinsic call.
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     break;
1305   }
1306 
1307   case Builtin::BIfma:
1308   case Builtin::BIfmaf:
1309   case Builtin::BIfmal:
1310   case Builtin::BI__builtin_fma:
1311   case Builtin::BI__builtin_fmaf:
1312   case Builtin::BI__builtin_fmal: {
1313     // Rewrite fma to intrinsic.
1314     Value *FirstArg = EmitScalarExpr(E->getArg(0));
1315     llvm::Type *ArgType = FirstArg->getType();
1316     Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType);
1317     return RValue::get(Builder.CreateCall3(F, FirstArg,
1318                                               EmitScalarExpr(E->getArg(1)),
1319                                               EmitScalarExpr(E->getArg(2))));
1320   }
1321 
1322   case Builtin::BI__builtin_signbit:
1323   case Builtin::BI__builtin_signbitf:
1324   case Builtin::BI__builtin_signbitl: {
1325     LLVMContext &C = CGM.getLLVMContext();
1326 
1327     Value *Arg = EmitScalarExpr(E->getArg(0));
1328     llvm::Type *ArgTy = Arg->getType();
1329     if (ArgTy->isPPC_FP128Ty())
1330       break; // FIXME: I'm not sure what the right implementation is here.
1331     int ArgWidth = ArgTy->getPrimitiveSizeInBits();
1332     llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth);
1333     Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy);
1334     Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy);
1335     Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp);
1336     return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType())));
1337   }
1338   case Builtin::BI__builtin_annotation: {
1339     llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
1340     llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
1341                                       AnnVal->getType());
1342 
1343     // Get the annotation string, go through casts. Sema requires this to be a
1344     // non-wide string literal, potentially casted, so the cast<> is safe.
1345     const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
1346     StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
1347     return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
1348   }
1349   case Builtin::BI__builtin_addcb:
1350   case Builtin::BI__builtin_addcs:
1351   case Builtin::BI__builtin_addc:
1352   case Builtin::BI__builtin_addcl:
1353   case Builtin::BI__builtin_addcll:
1354   case Builtin::BI__builtin_subcb:
1355   case Builtin::BI__builtin_subcs:
1356   case Builtin::BI__builtin_subc:
1357   case Builtin::BI__builtin_subcl:
1358   case Builtin::BI__builtin_subcll: {
1359 
1360     // We translate all of these builtins from expressions of the form:
1361     //   int x = ..., y = ..., carryin = ..., carryout, result;
1362     //   result = __builtin_addc(x, y, carryin, &carryout);
1363     //
1364     // to LLVM IR of the form:
1365     //
1366     //   %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
1367     //   %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
1368     //   %carry1 = extractvalue {i32, i1} %tmp1, 1
1369     //   %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
1370     //                                                       i32 %carryin)
1371     //   %result = extractvalue {i32, i1} %tmp2, 0
1372     //   %carry2 = extractvalue {i32, i1} %tmp2, 1
1373     //   %tmp3 = or i1 %carry1, %carry2
1374     //   %tmp4 = zext i1 %tmp3 to i32
1375     //   store i32 %tmp4, i32* %carryout
1376 
1377     // Scalarize our inputs.
1378     llvm::Value *X = EmitScalarExpr(E->getArg(0));
1379     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1380     llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
1381     std::pair<llvm::Value*, unsigned> CarryOutPtr =
1382       EmitPointerWithAlignment(E->getArg(3));
1383 
1384     // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
1385     llvm::Intrinsic::ID IntrinsicId;
1386     switch (BuiltinID) {
1387     default: llvm_unreachable("Unknown multiprecision builtin id.");
1388     case Builtin::BI__builtin_addcb:
1389     case Builtin::BI__builtin_addcs:
1390     case Builtin::BI__builtin_addc:
1391     case Builtin::BI__builtin_addcl:
1392     case Builtin::BI__builtin_addcll:
1393       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1394       break;
1395     case Builtin::BI__builtin_subcb:
1396     case Builtin::BI__builtin_subcs:
1397     case Builtin::BI__builtin_subc:
1398     case Builtin::BI__builtin_subcl:
1399     case Builtin::BI__builtin_subcll:
1400       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1401       break;
1402     }
1403 
1404     // Construct our resulting LLVM IR expression.
1405     llvm::Value *Carry1;
1406     llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
1407                                               X, Y, Carry1);
1408     llvm::Value *Carry2;
1409     llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
1410                                               Sum1, Carryin, Carry2);
1411     llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
1412                                                X->getType());
1413     llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut,
1414                                                          CarryOutPtr.first);
1415     CarryOutStore->setAlignment(CarryOutPtr.second);
1416     return RValue::get(Sum2);
1417   }
1418   case Builtin::BI__builtin_uadd_overflow:
1419   case Builtin::BI__builtin_uaddl_overflow:
1420   case Builtin::BI__builtin_uaddll_overflow:
1421   case Builtin::BI__builtin_usub_overflow:
1422   case Builtin::BI__builtin_usubl_overflow:
1423   case Builtin::BI__builtin_usubll_overflow:
1424   case Builtin::BI__builtin_umul_overflow:
1425   case Builtin::BI__builtin_umull_overflow:
1426   case Builtin::BI__builtin_umulll_overflow:
1427   case Builtin::BI__builtin_sadd_overflow:
1428   case Builtin::BI__builtin_saddl_overflow:
1429   case Builtin::BI__builtin_saddll_overflow:
1430   case Builtin::BI__builtin_ssub_overflow:
1431   case Builtin::BI__builtin_ssubl_overflow:
1432   case Builtin::BI__builtin_ssubll_overflow:
1433   case Builtin::BI__builtin_smul_overflow:
1434   case Builtin::BI__builtin_smull_overflow:
1435   case Builtin::BI__builtin_smulll_overflow: {
1436 
1437     // We translate all of these builtins directly to the relevant llvm IR node.
1438 
1439     // Scalarize our inputs.
1440     llvm::Value *X = EmitScalarExpr(E->getArg(0));
1441     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1442     std::pair<llvm::Value *, unsigned> SumOutPtr =
1443       EmitPointerWithAlignment(E->getArg(2));
1444 
1445     // Decide which of the overflow intrinsics we are lowering to:
1446     llvm::Intrinsic::ID IntrinsicId;
1447     switch (BuiltinID) {
1448     default: llvm_unreachable("Unknown security overflow builtin id.");
1449     case Builtin::BI__builtin_uadd_overflow:
1450     case Builtin::BI__builtin_uaddl_overflow:
1451     case Builtin::BI__builtin_uaddll_overflow:
1452       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1453       break;
1454     case Builtin::BI__builtin_usub_overflow:
1455     case Builtin::BI__builtin_usubl_overflow:
1456     case Builtin::BI__builtin_usubll_overflow:
1457       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1458       break;
1459     case Builtin::BI__builtin_umul_overflow:
1460     case Builtin::BI__builtin_umull_overflow:
1461     case Builtin::BI__builtin_umulll_overflow:
1462       IntrinsicId = llvm::Intrinsic::umul_with_overflow;
1463       break;
1464     case Builtin::BI__builtin_sadd_overflow:
1465     case Builtin::BI__builtin_saddl_overflow:
1466     case Builtin::BI__builtin_saddll_overflow:
1467       IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
1468       break;
1469     case Builtin::BI__builtin_ssub_overflow:
1470     case Builtin::BI__builtin_ssubl_overflow:
1471     case Builtin::BI__builtin_ssubll_overflow:
1472       IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
1473       break;
1474     case Builtin::BI__builtin_smul_overflow:
1475     case Builtin::BI__builtin_smull_overflow:
1476     case Builtin::BI__builtin_smulll_overflow:
1477       IntrinsicId = llvm::Intrinsic::smul_with_overflow;
1478       break;
1479     }
1480 
1481 
1482     llvm::Value *Carry;
1483     llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
1484     llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first);
1485     SumOutStore->setAlignment(SumOutPtr.second);
1486 
1487     return RValue::get(Carry);
1488   }
1489   case Builtin::BI__builtin_addressof:
1490     return RValue::get(EmitLValue(E->getArg(0)).getAddress());
1491   case Builtin::BI__noop:
1492     return RValue::get(0);
1493   }
1494 
1495   // If this is an alias for a lib function (e.g. __builtin_sin), emit
1496   // the call using the normal call path, but using the unmangled
1497   // version of the function name.
1498   if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
1499     return emitLibraryCall(*this, FD, E,
1500                            CGM.getBuiltinLibFunction(FD, BuiltinID));
1501 
1502   // If this is a predefined lib function (e.g. malloc), emit the call
1503   // using exactly the normal call path.
1504   if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
1505     return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee()));
1506 
1507   // See if we have a target specific intrinsic.
1508   const char *Name = getContext().BuiltinInfo.GetName(BuiltinID);
1509   Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
1510   if (const char *Prefix =
1511       llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch()))
1512     IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name);
1513 
1514   if (IntrinsicID != Intrinsic::not_intrinsic) {
1515     SmallVector<Value*, 16> Args;
1516 
1517     // Find out if any arguments are required to be integer constant
1518     // expressions.
1519     unsigned ICEArguments = 0;
1520     ASTContext::GetBuiltinTypeError Error;
1521     getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1522     assert(Error == ASTContext::GE_None && "Should not codegen an error");
1523 
1524     Function *F = CGM.getIntrinsic(IntrinsicID);
1525     llvm::FunctionType *FTy = F->getFunctionType();
1526 
1527     for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
1528       Value *ArgValue;
1529       // If this is a normal argument, just emit it as a scalar.
1530       if ((ICEArguments & (1 << i)) == 0) {
1531         ArgValue = EmitScalarExpr(E->getArg(i));
1532       } else {
1533         // If this is required to be a constant, constant fold it so that we
1534         // know that the generated intrinsic gets a ConstantInt.
1535         llvm::APSInt Result;
1536         bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
1537         assert(IsConst && "Constant arg isn't actually constant?");
1538         (void)IsConst;
1539         ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
1540       }
1541 
1542       // If the intrinsic arg type is different from the builtin arg type
1543       // we need to do a bit cast.
1544       llvm::Type *PTy = FTy->getParamType(i);
1545       if (PTy != ArgValue->getType()) {
1546         assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
1547                "Must be able to losslessly bit cast to param");
1548         ArgValue = Builder.CreateBitCast(ArgValue, PTy);
1549       }
1550 
1551       Args.push_back(ArgValue);
1552     }
1553 
1554     Value *V = Builder.CreateCall(F, Args);
1555     QualType BuiltinRetType = E->getType();
1556 
1557     llvm::Type *RetTy = VoidTy;
1558     if (!BuiltinRetType->isVoidType())
1559       RetTy = ConvertType(BuiltinRetType);
1560 
1561     if (RetTy != V->getType()) {
1562       assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
1563              "Must be able to losslessly bit cast result type");
1564       V = Builder.CreateBitCast(V, RetTy);
1565     }
1566 
1567     return RValue::get(V);
1568   }
1569 
1570   // See if we have a target specific builtin that needs to be lowered.
1571   if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
1572     return RValue::get(V);
1573 
1574   ErrorUnsupported(E, "builtin function");
1575 
1576   // Unknown builtin, for now just dump it out and return undef.
1577   return GetUndefRValue(E->getType());
1578 }
1579 
EmitTargetBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1580 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
1581                                               const CallExpr *E) {
1582   switch (getTarget().getTriple().getArch()) {
1583   case llvm::Triple::aarch64:
1584     return EmitAArch64BuiltinExpr(BuiltinID, E);
1585   case llvm::Triple::arm:
1586   case llvm::Triple::thumb:
1587     return EmitARMBuiltinExpr(BuiltinID, E);
1588   case llvm::Triple::x86:
1589   case llvm::Triple::x86_64:
1590     return EmitX86BuiltinExpr(BuiltinID, E);
1591   case llvm::Triple::ppc:
1592   case llvm::Triple::ppc64:
1593   case llvm::Triple::ppc64le:
1594     return EmitPPCBuiltinExpr(BuiltinID, E);
1595   default:
1596     return 0;
1597   }
1598 }
1599 
GetNeonType(CodeGenFunction * CGF,NeonTypeFlags TypeFlags)1600 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
1601                                      NeonTypeFlags TypeFlags) {
1602   int IsQuad = TypeFlags.isQuad();
1603   switch (TypeFlags.getEltType()) {
1604   case NeonTypeFlags::Int8:
1605   case NeonTypeFlags::Poly8:
1606     return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad);
1607   case NeonTypeFlags::Int16:
1608   case NeonTypeFlags::Poly16:
1609   case NeonTypeFlags::Float16:
1610     return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad);
1611   case NeonTypeFlags::Int32:
1612     return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad);
1613   case NeonTypeFlags::Int64:
1614     return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad);
1615   case NeonTypeFlags::Float32:
1616     return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad);
1617   case NeonTypeFlags::Float64:
1618     return llvm::VectorType::get(CGF->DoubleTy, 1 << IsQuad);
1619   }
1620   llvm_unreachable("Invalid NeonTypeFlags element type!");
1621 }
1622 
EmitNeonSplat(Value * V,Constant * C)1623 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
1624   unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements();
1625   Value* SV = llvm::ConstantVector::getSplat(nElts, C);
1626   return Builder.CreateShuffleVector(V, V, SV, "lane");
1627 }
1628 
EmitNeonCall(Function * F,SmallVectorImpl<Value * > & Ops,const char * name,unsigned shift,bool rightshift)1629 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
1630                                      const char *name,
1631                                      unsigned shift, bool rightshift) {
1632   unsigned j = 0;
1633   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
1634        ai != ae; ++ai, ++j)
1635     if (shift > 0 && shift == j)
1636       Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
1637     else
1638       Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
1639 
1640   return Builder.CreateCall(F, Ops, name);
1641 }
1642 
EmitNeonShiftVector(Value * V,llvm::Type * Ty,bool neg)1643 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
1644                                             bool neg) {
1645   int SV = cast<ConstantInt>(V)->getSExtValue();
1646 
1647   llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1648   llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV);
1649   return llvm::ConstantVector::getSplat(VTy->getNumElements(), C);
1650 }
1651 
1652 /// GetPointeeAlignment - Given an expression with a pointer type, find the
1653 /// alignment of the type referenced by the pointer.  Skip over implicit
1654 /// casts.
1655 std::pair<llvm::Value*, unsigned>
EmitPointerWithAlignment(const Expr * Addr)1656 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) {
1657   assert(Addr->getType()->isPointerType());
1658   Addr = Addr->IgnoreParens();
1659   if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) {
1660     if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) &&
1661         ICE->getSubExpr()->getType()->isPointerType()) {
1662       std::pair<llvm::Value*, unsigned> Ptr =
1663           EmitPointerWithAlignment(ICE->getSubExpr());
1664       Ptr.first = Builder.CreateBitCast(Ptr.first,
1665                                         ConvertType(Addr->getType()));
1666       return Ptr;
1667     } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) {
1668       LValue LV = EmitLValue(ICE->getSubExpr());
1669       unsigned Align = LV.getAlignment().getQuantity();
1670       if (!Align) {
1671         // FIXME: Once LValues are fixed to always set alignment,
1672         // zap this code.
1673         QualType PtTy = ICE->getSubExpr()->getType();
1674         if (!PtTy->isIncompleteType())
1675           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1676         else
1677           Align = 1;
1678       }
1679       return std::make_pair(LV.getAddress(), Align);
1680     }
1681   }
1682   if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) {
1683     if (UO->getOpcode() == UO_AddrOf) {
1684       LValue LV = EmitLValue(UO->getSubExpr());
1685       unsigned Align = LV.getAlignment().getQuantity();
1686       if (!Align) {
1687         // FIXME: Once LValues are fixed to always set alignment,
1688         // zap this code.
1689         QualType PtTy = UO->getSubExpr()->getType();
1690         if (!PtTy->isIncompleteType())
1691           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1692         else
1693           Align = 1;
1694       }
1695       return std::make_pair(LV.getAddress(), Align);
1696     }
1697   }
1698 
1699   unsigned Align = 1;
1700   QualType PtTy = Addr->getType()->getPointeeType();
1701   if (!PtTy->isIncompleteType())
1702     Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1703 
1704   return std::make_pair(EmitScalarExpr(Addr), Align);
1705 }
1706 
EmitAArch64BuiltinExpr(unsigned BuiltinID,const CallExpr * E)1707 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
1708                                                const CallExpr *E) {
1709   if (BuiltinID == AArch64::BI__clear_cache) {
1710     assert(E->getNumArgs() == 2 &&
1711            "Variadic __clear_cache slipped through on AArch64");
1712 
1713     const FunctionDecl *FD = E->getDirectCallee();
1714     SmallVector<Value *, 2> Ops;
1715     for (unsigned i = 0; i < E->getNumArgs(); i++)
1716       Ops.push_back(EmitScalarExpr(E->getArg(i)));
1717     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
1718     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
1719     StringRef Name = FD->getName();
1720     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
1721   }
1722 
1723   SmallVector<Value *, 4> Ops;
1724   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
1725     Ops.push_back(EmitScalarExpr(E->getArg(i)));
1726   }
1727 
1728   // Get the last argument, which specifies the vector type.
1729   llvm::APSInt Result;
1730   const Expr *Arg = E->getArg(E->getNumArgs() - 1);
1731   if (!Arg->isIntegerConstantExpr(Result, getContext()))
1732     return 0;
1733 
1734   // Determine the type of this overloaded NEON intrinsic.
1735   NeonTypeFlags Type(Result.getZExtValue());
1736   bool usgn = Type.isUnsigned();
1737 
1738   llvm::VectorType *VTy = GetNeonType(this, Type);
1739   llvm::Type *Ty = VTy;
1740   if (!Ty)
1741     return 0;
1742 
1743   unsigned Int;
1744   switch (BuiltinID) {
1745   default:
1746     return 0;
1747 
1748   // AArch64 builtins mapping to legacy ARM v7 builtins.
1749   // FIXME: the mapped builtins listed correspond to what has been tested
1750   // in aarch64-neon-intrinsics.c so far.
1751   case AArch64::BI__builtin_neon_vmul_v:
1752     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmul_v, E);
1753   case AArch64::BI__builtin_neon_vmulq_v:
1754     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmulq_v, E);
1755   case AArch64::BI__builtin_neon_vabd_v:
1756     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vabd_v, E);
1757   case AArch64::BI__builtin_neon_vabdq_v:
1758     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vabdq_v, E);
1759   case AArch64::BI__builtin_neon_vfma_v:
1760     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vfma_v, E);
1761   case AArch64::BI__builtin_neon_vfmaq_v:
1762     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vfmaq_v, E);
1763   case AArch64::BI__builtin_neon_vbsl_v:
1764     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vbsl_v, E);
1765   case AArch64::BI__builtin_neon_vbslq_v:
1766     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vbslq_v, E);
1767   case AArch64::BI__builtin_neon_vrsqrts_v:
1768     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrsqrts_v, E);
1769   case AArch64::BI__builtin_neon_vrsqrtsq_v:
1770     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrsqrtsq_v, E);
1771   case AArch64::BI__builtin_neon_vrecps_v:
1772     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrecps_v, E);
1773   case AArch64::BI__builtin_neon_vrecpsq_v:
1774     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrecpsq_v, E);
1775   case AArch64::BI__builtin_neon_vcage_v:
1776     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcage_v, E);
1777   case AArch64::BI__builtin_neon_vcale_v:
1778     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcale_v, E);
1779   case AArch64::BI__builtin_neon_vcaleq_v:
1780     std::swap(Ops[0], Ops[1]);
1781   case AArch64::BI__builtin_neon_vcageq_v: {
1782     Function *F;
1783     if (VTy->getElementType()->isIntegerTy(64))
1784       F = CGM.getIntrinsic(Intrinsic::aarch64_neon_vacgeq);
1785     else
1786       F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq);
1787     return EmitNeonCall(F, Ops, "vcage");
1788   }
1789   case AArch64::BI__builtin_neon_vcalt_v:
1790     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcalt_v, E);
1791   case AArch64::BI__builtin_neon_vcagt_v:
1792     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcagt_v, E);
1793   case AArch64::BI__builtin_neon_vcaltq_v:
1794     std::swap(Ops[0], Ops[1]);
1795   case AArch64::BI__builtin_neon_vcagtq_v: {
1796     Function *F;
1797     if (VTy->getElementType()->isIntegerTy(64))
1798       F = CGM.getIntrinsic(Intrinsic::aarch64_neon_vacgtq);
1799     else
1800       F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq);
1801     return EmitNeonCall(F, Ops, "vcagt");
1802   }
1803   case AArch64::BI__builtin_neon_vtst_v:
1804     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vtst_v, E);
1805   case AArch64::BI__builtin_neon_vtstq_v:
1806     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vtstq_v, E);
1807   case AArch64::BI__builtin_neon_vhadd_v:
1808     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhadd_v, E);
1809   case AArch64::BI__builtin_neon_vhaddq_v:
1810     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhaddq_v, E);
1811   case AArch64::BI__builtin_neon_vhsub_v:
1812     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhsub_v, E);
1813   case AArch64::BI__builtin_neon_vhsubq_v:
1814     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhsubq_v, E);
1815   case AArch64::BI__builtin_neon_vrhadd_v:
1816     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrhadd_v, E);
1817   case AArch64::BI__builtin_neon_vrhaddq_v:
1818     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrhaddq_v, E);
1819   case AArch64::BI__builtin_neon_vqadd_v:
1820     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqadd_v, E);
1821   case AArch64::BI__builtin_neon_vqaddq_v:
1822     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqaddq_v, E);
1823   case AArch64::BI__builtin_neon_vqsub_v:
1824     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqsub_v, E);
1825   case AArch64::BI__builtin_neon_vqsubq_v:
1826     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqsubq_v, E);
1827   case AArch64::BI__builtin_neon_vshl_v:
1828     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshl_v, E);
1829   case AArch64::BI__builtin_neon_vshlq_v:
1830     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshlq_v, E);
1831   case AArch64::BI__builtin_neon_vqshl_v:
1832     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqshl_v, E);
1833   case AArch64::BI__builtin_neon_vqshlq_v:
1834     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqshlq_v, E);
1835   case AArch64::BI__builtin_neon_vrshl_v:
1836     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrshl_v, E);
1837   case AArch64::BI__builtin_neon_vrshlq_v:
1838     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrshlq_v, E);
1839   case AArch64::BI__builtin_neon_vqrshl_v:
1840     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrshl_v, E);
1841   case AArch64::BI__builtin_neon_vqrshlq_v:
1842     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrshlq_v, E);
1843   case AArch64::BI__builtin_neon_vmax_v:
1844     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmax_v, E);
1845   case AArch64::BI__builtin_neon_vmaxq_v:
1846     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmaxq_v, E);
1847   case AArch64::BI__builtin_neon_vmin_v:
1848     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmin_v, E);
1849   case AArch64::BI__builtin_neon_vminq_v:
1850     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vminq_v, E);
1851   case AArch64::BI__builtin_neon_vpmax_v:
1852     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpmax_v, E);
1853   case AArch64::BI__builtin_neon_vpmin_v:
1854     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpmin_v, E);
1855   case AArch64::BI__builtin_neon_vpadd_v:
1856     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpadd_v, E);
1857   case AArch64::BI__builtin_neon_vqdmulh_v:
1858     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqdmulh_v, E);
1859   case AArch64::BI__builtin_neon_vqdmulhq_v:
1860     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqdmulhq_v, E);
1861   case AArch64::BI__builtin_neon_vqrdmulh_v:
1862     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrdmulh_v, E);
1863   case AArch64::BI__builtin_neon_vqrdmulhq_v:
1864     return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrdmulhq_v, E);
1865 
1866   // AArch64-only builtins
1867   case AArch64::BI__builtin_neon_vfms_v:
1868   case AArch64::BI__builtin_neon_vfmsq_v: {
1869     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
1870     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
1871     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
1872     Ops[1] = Builder.CreateFNeg(Ops[1]);
1873     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
1874 
1875     // LLVM's fma intrinsic puts the accumulator in the last position, but the
1876     // AArch64 intrinsic has it first.
1877     return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
1878   }
1879   case AArch64::BI__builtin_neon_vmaxnm_v:
1880   case AArch64::BI__builtin_neon_vmaxnmq_v: {
1881     Int = Intrinsic::aarch64_neon_vmaxnm;
1882     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
1883   }
1884   case AArch64::BI__builtin_neon_vminnm_v:
1885   case AArch64::BI__builtin_neon_vminnmq_v: {
1886     Int = Intrinsic::aarch64_neon_vminnm;
1887     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
1888   }
1889   case AArch64::BI__builtin_neon_vpmaxnm_v:
1890   case AArch64::BI__builtin_neon_vpmaxnmq_v: {
1891     Int = Intrinsic::aarch64_neon_vpmaxnm;
1892     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
1893   }
1894   case AArch64::BI__builtin_neon_vpminnm_v:
1895   case AArch64::BI__builtin_neon_vpminnmq_v: {
1896     Int = Intrinsic::aarch64_neon_vpminnm;
1897     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
1898   }
1899   case AArch64::BI__builtin_neon_vpmaxq_v: {
1900     Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs;
1901     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
1902   }
1903   case AArch64::BI__builtin_neon_vpminq_v: {
1904     Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins;
1905     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
1906   }
1907   case AArch64::BI__builtin_neon_vpaddq_v: {
1908     Int = Intrinsic::arm_neon_vpadd;
1909     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpadd");
1910   }
1911   case AArch64::BI__builtin_neon_vmulx_v:
1912   case AArch64::BI__builtin_neon_vmulxq_v: {
1913     Int = Intrinsic::aarch64_neon_vmulx;
1914     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
1915   }
1916   }
1917 }
1918 
EmitARMBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1919 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
1920                                            const CallExpr *E) {
1921   if (BuiltinID == ARM::BI__clear_cache) {
1922     assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
1923     const FunctionDecl *FD = E->getDirectCallee();
1924     SmallVector<Value*, 2> Ops;
1925     for (unsigned i = 0; i < 2; i++)
1926       Ops.push_back(EmitScalarExpr(E->getArg(i)));
1927     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
1928     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
1929     StringRef Name = FD->getName();
1930     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
1931   }
1932 
1933   if (BuiltinID == ARM::BI__builtin_arm_ldrexd ||
1934       (BuiltinID == ARM::BI__builtin_arm_ldrex &&
1935        getContext().getTypeSize(E->getType()) == 64)) {
1936     Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
1937 
1938     Value *LdPtr = EmitScalarExpr(E->getArg(0));
1939     Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
1940                                     "ldrexd");
1941 
1942     Value *Val0 = Builder.CreateExtractValue(Val, 1);
1943     Value *Val1 = Builder.CreateExtractValue(Val, 0);
1944     Val0 = Builder.CreateZExt(Val0, Int64Ty);
1945     Val1 = Builder.CreateZExt(Val1, Int64Ty);
1946 
1947     Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
1948     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
1949     Val = Builder.CreateOr(Val, Val1);
1950     return Builder.CreateBitCast(Val, ConvertType(E->getType()));
1951   }
1952 
1953   if (BuiltinID == ARM::BI__builtin_arm_ldrex) {
1954     Value *LoadAddr = EmitScalarExpr(E->getArg(0));
1955 
1956     QualType Ty = E->getType();
1957     llvm::Type *RealResTy = ConvertType(Ty);
1958     llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(),
1959                                                   getContext().getTypeSize(Ty));
1960     LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo());
1961 
1962     Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrex, LoadAddr->getType());
1963     Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
1964 
1965     if (RealResTy->isPointerTy())
1966       return Builder.CreateIntToPtr(Val, RealResTy);
1967     else {
1968       Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
1969       return Builder.CreateBitCast(Val, RealResTy);
1970     }
1971   }
1972 
1973   if (BuiltinID == ARM::BI__builtin_arm_strexd ||
1974       (BuiltinID == ARM::BI__builtin_arm_strex &&
1975        getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
1976     Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd);
1977     llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL);
1978 
1979     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
1980     Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()),
1981                                       One);
1982     Value *Val = EmitScalarExpr(E->getArg(0));
1983     Builder.CreateStore(Val, Tmp);
1984 
1985     Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
1986     Val = Builder.CreateLoad(LdPtr);
1987 
1988     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
1989     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
1990     Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
1991     return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd");
1992   }
1993 
1994   if (BuiltinID == ARM::BI__builtin_arm_strex) {
1995     Value *StoreVal = EmitScalarExpr(E->getArg(0));
1996     Value *StoreAddr = EmitScalarExpr(E->getArg(1));
1997 
1998     QualType Ty = E->getArg(0)->getType();
1999     llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
2000                                                  getContext().getTypeSize(Ty));
2001     StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
2002 
2003     if (StoreVal->getType()->isPointerTy())
2004       StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
2005     else {
2006       StoreVal = Builder.CreateBitCast(StoreVal, StoreTy);
2007       StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
2008     }
2009 
2010     Function *F = CGM.getIntrinsic(Intrinsic::arm_strex, StoreAddr->getType());
2011     return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex");
2012   }
2013 
2014   if (BuiltinID == ARM::BI__builtin_arm_clrex) {
2015     Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
2016     return Builder.CreateCall(F);
2017   }
2018 
2019   SmallVector<Value*, 4> Ops;
2020   llvm::Value *Align = 0;
2021   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
2022     if (i == 0) {
2023       switch (BuiltinID) {
2024       case ARM::BI__builtin_neon_vld1_v:
2025       case ARM::BI__builtin_neon_vld1q_v:
2026       case ARM::BI__builtin_neon_vld1q_lane_v:
2027       case ARM::BI__builtin_neon_vld1_lane_v:
2028       case ARM::BI__builtin_neon_vld1_dup_v:
2029       case ARM::BI__builtin_neon_vld1q_dup_v:
2030       case ARM::BI__builtin_neon_vst1_v:
2031       case ARM::BI__builtin_neon_vst1q_v:
2032       case ARM::BI__builtin_neon_vst1q_lane_v:
2033       case ARM::BI__builtin_neon_vst1_lane_v:
2034       case ARM::BI__builtin_neon_vst2_v:
2035       case ARM::BI__builtin_neon_vst2q_v:
2036       case ARM::BI__builtin_neon_vst2_lane_v:
2037       case ARM::BI__builtin_neon_vst2q_lane_v:
2038       case ARM::BI__builtin_neon_vst3_v:
2039       case ARM::BI__builtin_neon_vst3q_v:
2040       case ARM::BI__builtin_neon_vst3_lane_v:
2041       case ARM::BI__builtin_neon_vst3q_lane_v:
2042       case ARM::BI__builtin_neon_vst4_v:
2043       case ARM::BI__builtin_neon_vst4q_v:
2044       case ARM::BI__builtin_neon_vst4_lane_v:
2045       case ARM::BI__builtin_neon_vst4q_lane_v:
2046         // Get the alignment for the argument in addition to the value;
2047         // we'll use it later.
2048         std::pair<llvm::Value*, unsigned> Src =
2049             EmitPointerWithAlignment(E->getArg(0));
2050         Ops.push_back(Src.first);
2051         Align = Builder.getInt32(Src.second);
2052         continue;
2053       }
2054     }
2055     if (i == 1) {
2056       switch (BuiltinID) {
2057       case ARM::BI__builtin_neon_vld2_v:
2058       case ARM::BI__builtin_neon_vld2q_v:
2059       case ARM::BI__builtin_neon_vld3_v:
2060       case ARM::BI__builtin_neon_vld3q_v:
2061       case ARM::BI__builtin_neon_vld4_v:
2062       case ARM::BI__builtin_neon_vld4q_v:
2063       case ARM::BI__builtin_neon_vld2_lane_v:
2064       case ARM::BI__builtin_neon_vld2q_lane_v:
2065       case ARM::BI__builtin_neon_vld3_lane_v:
2066       case ARM::BI__builtin_neon_vld3q_lane_v:
2067       case ARM::BI__builtin_neon_vld4_lane_v:
2068       case ARM::BI__builtin_neon_vld4q_lane_v:
2069       case ARM::BI__builtin_neon_vld2_dup_v:
2070       case ARM::BI__builtin_neon_vld3_dup_v:
2071       case ARM::BI__builtin_neon_vld4_dup_v:
2072         // Get the alignment for the argument in addition to the value;
2073         // we'll use it later.
2074         std::pair<llvm::Value*, unsigned> Src =
2075             EmitPointerWithAlignment(E->getArg(1));
2076         Ops.push_back(Src.first);
2077         Align = Builder.getInt32(Src.second);
2078         continue;
2079       }
2080     }
2081     Ops.push_back(EmitScalarExpr(E->getArg(i)));
2082   }
2083 
2084   // vget_lane and vset_lane are not overloaded and do not have an extra
2085   // argument that specifies the vector type.
2086   switch (BuiltinID) {
2087   default: break;
2088   case ARM::BI__builtin_neon_vget_lane_i8:
2089   case ARM::BI__builtin_neon_vget_lane_i16:
2090   case ARM::BI__builtin_neon_vget_lane_i32:
2091   case ARM::BI__builtin_neon_vget_lane_i64:
2092   case ARM::BI__builtin_neon_vget_lane_f32:
2093   case ARM::BI__builtin_neon_vgetq_lane_i8:
2094   case ARM::BI__builtin_neon_vgetq_lane_i16:
2095   case ARM::BI__builtin_neon_vgetq_lane_i32:
2096   case ARM::BI__builtin_neon_vgetq_lane_i64:
2097   case ARM::BI__builtin_neon_vgetq_lane_f32:
2098     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
2099                                         "vget_lane");
2100   case ARM::BI__builtin_neon_vset_lane_i8:
2101   case ARM::BI__builtin_neon_vset_lane_i16:
2102   case ARM::BI__builtin_neon_vset_lane_i32:
2103   case ARM::BI__builtin_neon_vset_lane_i64:
2104   case ARM::BI__builtin_neon_vset_lane_f32:
2105   case ARM::BI__builtin_neon_vsetq_lane_i8:
2106   case ARM::BI__builtin_neon_vsetq_lane_i16:
2107   case ARM::BI__builtin_neon_vsetq_lane_i32:
2108   case ARM::BI__builtin_neon_vsetq_lane_i64:
2109   case ARM::BI__builtin_neon_vsetq_lane_f32:
2110     Ops.push_back(EmitScalarExpr(E->getArg(2)));
2111     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
2112   }
2113 
2114   // Get the last argument, which specifies the vector type.
2115   llvm::APSInt Result;
2116   const Expr *Arg = E->getArg(E->getNumArgs()-1);
2117   if (!Arg->isIntegerConstantExpr(Result, getContext()))
2118     return 0;
2119 
2120   if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
2121       BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
2122     // Determine the overloaded type of this builtin.
2123     llvm::Type *Ty;
2124     if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
2125       Ty = FloatTy;
2126     else
2127       Ty = DoubleTy;
2128 
2129     // Determine whether this is an unsigned conversion or not.
2130     bool usgn = Result.getZExtValue() == 1;
2131     unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
2132 
2133     // Call the appropriate intrinsic.
2134     Function *F = CGM.getIntrinsic(Int, Ty);
2135     return Builder.CreateCall(F, Ops, "vcvtr");
2136   }
2137 
2138   // Determine the type of this overloaded NEON intrinsic.
2139   NeonTypeFlags Type(Result.getZExtValue());
2140   bool usgn = Type.isUnsigned();
2141   bool quad = Type.isQuad();
2142   bool rightShift = false;
2143 
2144   llvm::VectorType *VTy = GetNeonType(this, Type);
2145   llvm::Type *Ty = VTy;
2146   if (!Ty)
2147     return 0;
2148 
2149   unsigned Int;
2150   switch (BuiltinID) {
2151   default: return 0;
2152   case ARM::BI__builtin_neon_vbsl_v:
2153   case ARM::BI__builtin_neon_vbslq_v:
2154     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty),
2155                         Ops, "vbsl");
2156   case ARM::BI__builtin_neon_vabd_v:
2157   case ARM::BI__builtin_neon_vabdq_v:
2158     Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds;
2159     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
2160   case ARM::BI__builtin_neon_vabs_v:
2161   case ARM::BI__builtin_neon_vabsq_v:
2162     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty),
2163                         Ops, "vabs");
2164   case ARM::BI__builtin_neon_vaddhn_v:
2165     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty),
2166                         Ops, "vaddhn");
2167   case ARM::BI__builtin_neon_vcale_v:
2168     std::swap(Ops[0], Ops[1]);
2169   case ARM::BI__builtin_neon_vcage_v: {
2170     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged);
2171     return EmitNeonCall(F, Ops, "vcage");
2172   }
2173   case ARM::BI__builtin_neon_vcaleq_v:
2174     std::swap(Ops[0], Ops[1]);
2175   case ARM::BI__builtin_neon_vcageq_v: {
2176     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq);
2177     return EmitNeonCall(F, Ops, "vcage");
2178   }
2179   case ARM::BI__builtin_neon_vcalt_v:
2180     std::swap(Ops[0], Ops[1]);
2181   case ARM::BI__builtin_neon_vcagt_v: {
2182     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd);
2183     return EmitNeonCall(F, Ops, "vcagt");
2184   }
2185   case ARM::BI__builtin_neon_vcaltq_v:
2186     std::swap(Ops[0], Ops[1]);
2187   case ARM::BI__builtin_neon_vcagtq_v: {
2188     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq);
2189     return EmitNeonCall(F, Ops, "vcagt");
2190   }
2191   case ARM::BI__builtin_neon_vcls_v:
2192   case ARM::BI__builtin_neon_vclsq_v: {
2193     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty);
2194     return EmitNeonCall(F, Ops, "vcls");
2195   }
2196   case ARM::BI__builtin_neon_vclz_v:
2197   case ARM::BI__builtin_neon_vclzq_v: {
2198     // Generate target-independent intrinsic; also need to add second argument
2199     // for whether or not clz of zero is undefined; on ARM it isn't.
2200     Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty);
2201     Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
2202     return EmitNeonCall(F, Ops, "vclz");
2203   }
2204   case ARM::BI__builtin_neon_vcnt_v:
2205   case ARM::BI__builtin_neon_vcntq_v: {
2206     // generate target-independent intrinsic
2207     Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty);
2208     return EmitNeonCall(F, Ops, "vctpop");
2209   }
2210   case ARM::BI__builtin_neon_vcvt_f16_v: {
2211     assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad &&
2212            "unexpected vcvt_f16_v builtin");
2213     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf);
2214     return EmitNeonCall(F, Ops, "vcvt");
2215   }
2216   case ARM::BI__builtin_neon_vcvt_f32_f16: {
2217     assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad &&
2218            "unexpected vcvt_f32_f16 builtin");
2219     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp);
2220     return EmitNeonCall(F, Ops, "vcvt");
2221   }
2222   case ARM::BI__builtin_neon_vcvt_f32_v:
2223   case ARM::BI__builtin_neon_vcvtq_f32_v:
2224     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2225     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
2226     return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
2227                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
2228   case ARM::BI__builtin_neon_vcvt_s32_v:
2229   case ARM::BI__builtin_neon_vcvt_u32_v:
2230   case ARM::BI__builtin_neon_vcvtq_s32_v:
2231   case ARM::BI__builtin_neon_vcvtq_u32_v: {
2232     llvm::Type *FloatTy =
2233       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
2234     Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
2235     return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
2236                 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
2237   }
2238   case ARM::BI__builtin_neon_vcvt_n_f32_v:
2239   case ARM::BI__builtin_neon_vcvtq_n_f32_v: {
2240     llvm::Type *FloatTy =
2241       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
2242     llvm::Type *Tys[2] = { FloatTy, Ty };
2243     Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp
2244                : Intrinsic::arm_neon_vcvtfxs2fp;
2245     Function *F = CGM.getIntrinsic(Int, Tys);
2246     return EmitNeonCall(F, Ops, "vcvt_n");
2247   }
2248   case ARM::BI__builtin_neon_vcvt_n_s32_v:
2249   case ARM::BI__builtin_neon_vcvt_n_u32_v:
2250   case ARM::BI__builtin_neon_vcvtq_n_s32_v:
2251   case ARM::BI__builtin_neon_vcvtq_n_u32_v: {
2252     llvm::Type *FloatTy =
2253       GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad));
2254     llvm::Type *Tys[2] = { Ty, FloatTy };
2255     Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu
2256                : Intrinsic::arm_neon_vcvtfp2fxs;
2257     Function *F = CGM.getIntrinsic(Int, Tys);
2258     return EmitNeonCall(F, Ops, "vcvt_n");
2259   }
2260   case ARM::BI__builtin_neon_vext_v:
2261   case ARM::BI__builtin_neon_vextq_v: {
2262     int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
2263     SmallVector<Constant*, 16> Indices;
2264     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2265       Indices.push_back(ConstantInt::get(Int32Ty, i+CV));
2266 
2267     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2268     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2269     Value *SV = llvm::ConstantVector::get(Indices);
2270     return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext");
2271   }
2272   case ARM::BI__builtin_neon_vhadd_v:
2273   case ARM::BI__builtin_neon_vhaddq_v:
2274     Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds;
2275     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd");
2276   case ARM::BI__builtin_neon_vhsub_v:
2277   case ARM::BI__builtin_neon_vhsubq_v:
2278     Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs;
2279     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub");
2280   case ARM::BI__builtin_neon_vld1_v:
2281   case ARM::BI__builtin_neon_vld1q_v:
2282     Ops.push_back(Align);
2283     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty),
2284                         Ops, "vld1");
2285   case ARM::BI__builtin_neon_vld1q_lane_v:
2286     // Handle 64-bit integer elements as a special case.  Use shuffles of
2287     // one-element vectors to avoid poor code for i64 in the backend.
2288     if (VTy->getElementType()->isIntegerTy(64)) {
2289       // Extract the other lane.
2290       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2291       int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
2292       Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
2293       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
2294       // Load the value as a one-element vector.
2295       Ty = llvm::VectorType::get(VTy->getElementType(), 1);
2296       Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty);
2297       Value *Ld = Builder.CreateCall2(F, Ops[0], Align);
2298       // Combine them.
2299       SmallVector<Constant*, 2> Indices;
2300       Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane));
2301       Indices.push_back(ConstantInt::get(Int32Ty, Lane));
2302       SV = llvm::ConstantVector::get(Indices);
2303       return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane");
2304     }
2305     // fall through
2306   case ARM::BI__builtin_neon_vld1_lane_v: {
2307     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2308     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
2309     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2310     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
2311     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2312     return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
2313   }
2314   case ARM::BI__builtin_neon_vld1_dup_v:
2315   case ARM::BI__builtin_neon_vld1q_dup_v: {
2316     Value *V = UndefValue::get(Ty);
2317     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
2318     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2319     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
2320     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2321     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
2322     Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
2323     return EmitNeonSplat(Ops[0], CI);
2324   }
2325   case ARM::BI__builtin_neon_vld2_v:
2326   case ARM::BI__builtin_neon_vld2q_v: {
2327     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty);
2328     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2");
2329     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2330     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2331     return Builder.CreateStore(Ops[1], Ops[0]);
2332   }
2333   case ARM::BI__builtin_neon_vld3_v:
2334   case ARM::BI__builtin_neon_vld3q_v: {
2335     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty);
2336     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3");
2337     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2338     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2339     return Builder.CreateStore(Ops[1], Ops[0]);
2340   }
2341   case ARM::BI__builtin_neon_vld4_v:
2342   case ARM::BI__builtin_neon_vld4q_v: {
2343     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty);
2344     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4");
2345     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2346     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2347     return Builder.CreateStore(Ops[1], Ops[0]);
2348   }
2349   case ARM::BI__builtin_neon_vld2_lane_v:
2350   case ARM::BI__builtin_neon_vld2q_lane_v: {
2351     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty);
2352     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2353     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2354     Ops.push_back(Align);
2355     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
2356     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2357     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2358     return Builder.CreateStore(Ops[1], Ops[0]);
2359   }
2360   case ARM::BI__builtin_neon_vld3_lane_v:
2361   case ARM::BI__builtin_neon_vld3q_lane_v: {
2362     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty);
2363     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2364     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2365     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
2366     Ops.push_back(Align);
2367     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
2368     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2369     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2370     return Builder.CreateStore(Ops[1], Ops[0]);
2371   }
2372   case ARM::BI__builtin_neon_vld4_lane_v:
2373   case ARM::BI__builtin_neon_vld4q_lane_v: {
2374     Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty);
2375     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2376     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
2377     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
2378     Ops[5] = Builder.CreateBitCast(Ops[5], Ty);
2379     Ops.push_back(Align);
2380     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
2381     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2382     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2383     return Builder.CreateStore(Ops[1], Ops[0]);
2384   }
2385   case ARM::BI__builtin_neon_vld2_dup_v:
2386   case ARM::BI__builtin_neon_vld3_dup_v:
2387   case ARM::BI__builtin_neon_vld4_dup_v: {
2388     // Handle 64-bit elements as a special-case.  There is no "dup" needed.
2389     if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
2390       switch (BuiltinID) {
2391       case ARM::BI__builtin_neon_vld2_dup_v:
2392         Int = Intrinsic::arm_neon_vld2;
2393         break;
2394       case ARM::BI__builtin_neon_vld3_dup_v:
2395         Int = Intrinsic::arm_neon_vld3;
2396         break;
2397       case ARM::BI__builtin_neon_vld4_dup_v:
2398         Int = Intrinsic::arm_neon_vld4;
2399         break;
2400       default: llvm_unreachable("unknown vld_dup intrinsic?");
2401       }
2402       Function *F = CGM.getIntrinsic(Int, Ty);
2403       Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup");
2404       Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2405       Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2406       return Builder.CreateStore(Ops[1], Ops[0]);
2407     }
2408     switch (BuiltinID) {
2409     case ARM::BI__builtin_neon_vld2_dup_v:
2410       Int = Intrinsic::arm_neon_vld2lane;
2411       break;
2412     case ARM::BI__builtin_neon_vld3_dup_v:
2413       Int = Intrinsic::arm_neon_vld3lane;
2414       break;
2415     case ARM::BI__builtin_neon_vld4_dup_v:
2416       Int = Intrinsic::arm_neon_vld4lane;
2417       break;
2418     default: llvm_unreachable("unknown vld_dup intrinsic?");
2419     }
2420     Function *F = CGM.getIntrinsic(Int, Ty);
2421     llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
2422 
2423     SmallVector<Value*, 6> Args;
2424     Args.push_back(Ops[1]);
2425     Args.append(STy->getNumElements(), UndefValue::get(Ty));
2426 
2427     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
2428     Args.push_back(CI);
2429     Args.push_back(Align);
2430 
2431     Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
2432     // splat lane 0 to all elts in each vector of the result.
2433     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2434       Value *Val = Builder.CreateExtractValue(Ops[1], i);
2435       Value *Elt = Builder.CreateBitCast(Val, Ty);
2436       Elt = EmitNeonSplat(Elt, CI);
2437       Elt = Builder.CreateBitCast(Elt, Val->getType());
2438       Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
2439     }
2440     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2441     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2442     return Builder.CreateStore(Ops[1], Ops[0]);
2443   }
2444   case ARM::BI__builtin_neon_vmax_v:
2445   case ARM::BI__builtin_neon_vmaxq_v:
2446     Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs;
2447     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
2448   case ARM::BI__builtin_neon_vmin_v:
2449   case ARM::BI__builtin_neon_vminq_v:
2450     Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins;
2451     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
2452   case ARM::BI__builtin_neon_vmovl_v: {
2453     llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
2454     Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
2455     if (usgn)
2456       return Builder.CreateZExt(Ops[0], Ty, "vmovl");
2457     return Builder.CreateSExt(Ops[0], Ty, "vmovl");
2458   }
2459   case ARM::BI__builtin_neon_vmovn_v: {
2460     llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
2461     Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
2462     return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
2463   }
2464   case ARM::BI__builtin_neon_vmul_v:
2465   case ARM::BI__builtin_neon_vmulq_v:
2466     assert(Type.isPoly() && "vmul builtin only supported for polynomial types");
2467     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty),
2468                         Ops, "vmul");
2469   case ARM::BI__builtin_neon_vmull_v:
2470     Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
2471     Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
2472     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
2473   case ARM::BI__builtin_neon_vfma_v:
2474   case ARM::BI__builtin_neon_vfmaq_v: {
2475     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
2476     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2477     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2478     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2479 
2480     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
2481     return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
2482   }
2483   case ARM::BI__builtin_neon_vpadal_v:
2484   case ARM::BI__builtin_neon_vpadalq_v: {
2485     Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals;
2486     // The source operand type has twice as many elements of half the size.
2487     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2488     llvm::Type *EltTy =
2489       llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2490     llvm::Type *NarrowTy =
2491       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2492     llvm::Type *Tys[2] = { Ty, NarrowTy };
2493     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal");
2494   }
2495   case ARM::BI__builtin_neon_vpadd_v:
2496     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty),
2497                         Ops, "vpadd");
2498   case ARM::BI__builtin_neon_vpaddl_v:
2499   case ARM::BI__builtin_neon_vpaddlq_v: {
2500     Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls;
2501     // The source operand type has twice as many elements of half the size.
2502     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2503     llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2504     llvm::Type *NarrowTy =
2505       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2506     llvm::Type *Tys[2] = { Ty, NarrowTy };
2507     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
2508   }
2509   case ARM::BI__builtin_neon_vpmax_v:
2510     Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs;
2511     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
2512   case ARM::BI__builtin_neon_vpmin_v:
2513     Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins;
2514     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
2515   case ARM::BI__builtin_neon_vqabs_v:
2516   case ARM::BI__builtin_neon_vqabsq_v:
2517     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty),
2518                         Ops, "vqabs");
2519   case ARM::BI__builtin_neon_vqadd_v:
2520   case ARM::BI__builtin_neon_vqaddq_v:
2521     Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds;
2522     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd");
2523   case ARM::BI__builtin_neon_vqdmlal_v:
2524     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty),
2525                         Ops, "vqdmlal");
2526   case ARM::BI__builtin_neon_vqdmlsl_v:
2527     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty),
2528                         Ops, "vqdmlsl");
2529   case ARM::BI__builtin_neon_vqdmulh_v:
2530   case ARM::BI__builtin_neon_vqdmulhq_v:
2531     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty),
2532                         Ops, "vqdmulh");
2533   case ARM::BI__builtin_neon_vqdmull_v:
2534     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty),
2535                         Ops, "vqdmull");
2536   case ARM::BI__builtin_neon_vqmovn_v:
2537     Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns;
2538     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn");
2539   case ARM::BI__builtin_neon_vqmovun_v:
2540     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty),
2541                         Ops, "vqdmull");
2542   case ARM::BI__builtin_neon_vqneg_v:
2543   case ARM::BI__builtin_neon_vqnegq_v:
2544     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty),
2545                         Ops, "vqneg");
2546   case ARM::BI__builtin_neon_vqrdmulh_v:
2547   case ARM::BI__builtin_neon_vqrdmulhq_v:
2548     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty),
2549                         Ops, "vqrdmulh");
2550   case ARM::BI__builtin_neon_vqrshl_v:
2551   case ARM::BI__builtin_neon_vqrshlq_v:
2552     Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts;
2553     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl");
2554   case ARM::BI__builtin_neon_vqrshrn_n_v:
2555     Int =
2556       usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
2557     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
2558                         1, true);
2559   case ARM::BI__builtin_neon_vqrshrun_n_v:
2560     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
2561                         Ops, "vqrshrun_n", 1, true);
2562   case ARM::BI__builtin_neon_vqshl_v:
2563   case ARM::BI__builtin_neon_vqshlq_v:
2564     Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
2565     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl");
2566   case ARM::BI__builtin_neon_vqshl_n_v:
2567   case ARM::BI__builtin_neon_vqshlq_n_v:
2568     Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts;
2569     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
2570                         1, false);
2571   case ARM::BI__builtin_neon_vqshlu_n_v:
2572   case ARM::BI__builtin_neon_vqshluq_n_v:
2573     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty),
2574                         Ops, "vqshlu", 1, false);
2575   case ARM::BI__builtin_neon_vqshrn_n_v:
2576     Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
2577     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
2578                         1, true);
2579   case ARM::BI__builtin_neon_vqshrun_n_v:
2580     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
2581                         Ops, "vqshrun_n", 1, true);
2582   case ARM::BI__builtin_neon_vqsub_v:
2583   case ARM::BI__builtin_neon_vqsubq_v:
2584     Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs;
2585     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub");
2586   case ARM::BI__builtin_neon_vraddhn_v:
2587     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty),
2588                         Ops, "vraddhn");
2589   case ARM::BI__builtin_neon_vrecpe_v:
2590   case ARM::BI__builtin_neon_vrecpeq_v:
2591     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
2592                         Ops, "vrecpe");
2593   case ARM::BI__builtin_neon_vrecps_v:
2594   case ARM::BI__builtin_neon_vrecpsq_v:
2595     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty),
2596                         Ops, "vrecps");
2597   case ARM::BI__builtin_neon_vrhadd_v:
2598   case ARM::BI__builtin_neon_vrhaddq_v:
2599     Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds;
2600     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd");
2601   case ARM::BI__builtin_neon_vrshl_v:
2602   case ARM::BI__builtin_neon_vrshlq_v:
2603     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2604     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl");
2605   case ARM::BI__builtin_neon_vrshrn_n_v:
2606     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
2607                         Ops, "vrshrn_n", 1, true);
2608   case ARM::BI__builtin_neon_vrshr_n_v:
2609   case ARM::BI__builtin_neon_vrshrq_n_v:
2610     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2611     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true);
2612   case ARM::BI__builtin_neon_vrsqrte_v:
2613   case ARM::BI__builtin_neon_vrsqrteq_v:
2614     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty),
2615                         Ops, "vrsqrte");
2616   case ARM::BI__builtin_neon_vrsqrts_v:
2617   case ARM::BI__builtin_neon_vrsqrtsq_v:
2618     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty),
2619                         Ops, "vrsqrts");
2620   case ARM::BI__builtin_neon_vrsra_n_v:
2621   case ARM::BI__builtin_neon_vrsraq_n_v:
2622     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2623     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2624     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
2625     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
2626     Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]);
2627     return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
2628   case ARM::BI__builtin_neon_vrsubhn_v:
2629     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty),
2630                         Ops, "vrsubhn");
2631   case ARM::BI__builtin_neon_vshl_v:
2632   case ARM::BI__builtin_neon_vshlq_v:
2633     Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts;
2634     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl");
2635   case ARM::BI__builtin_neon_vshll_n_v:
2636     Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls;
2637     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1);
2638   case ARM::BI__builtin_neon_vshl_n_v:
2639   case ARM::BI__builtin_neon_vshlq_n_v:
2640     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
2641     return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
2642                              "vshl_n");
2643   case ARM::BI__builtin_neon_vshrn_n_v:
2644     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty),
2645                         Ops, "vshrn_n", 1, true);
2646   case ARM::BI__builtin_neon_vshr_n_v:
2647   case ARM::BI__builtin_neon_vshrq_n_v:
2648     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2649     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
2650     if (usgn)
2651       return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n");
2652     else
2653       return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n");
2654   case ARM::BI__builtin_neon_vsri_n_v:
2655   case ARM::BI__builtin_neon_vsriq_n_v:
2656     rightShift = true;
2657   case ARM::BI__builtin_neon_vsli_n_v:
2658   case ARM::BI__builtin_neon_vsliq_n_v:
2659     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
2660     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
2661                         Ops, "vsli_n");
2662   case ARM::BI__builtin_neon_vsra_n_v:
2663   case ARM::BI__builtin_neon_vsraq_n_v:
2664     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2665     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2666     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false);
2667     if (usgn)
2668       Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n");
2669     else
2670       Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n");
2671     return Builder.CreateAdd(Ops[0], Ops[1]);
2672   case ARM::BI__builtin_neon_vst1_v:
2673   case ARM::BI__builtin_neon_vst1q_v:
2674     Ops.push_back(Align);
2675     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty),
2676                         Ops, "");
2677   case ARM::BI__builtin_neon_vst1q_lane_v:
2678     // Handle 64-bit integer elements as a special case.  Use a shuffle to get
2679     // a one-element vector and avoid poor code for i64 in the backend.
2680     if (VTy->getElementType()->isIntegerTy(64)) {
2681       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2682       Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
2683       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
2684       Ops[2] = Align;
2685       return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
2686                                                  Ops[1]->getType()), Ops);
2687     }
2688     // fall through
2689   case ARM::BI__builtin_neon_vst1_lane_v: {
2690     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2691     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
2692     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2693     StoreInst *St = Builder.CreateStore(Ops[1],
2694                                         Builder.CreateBitCast(Ops[0], Ty));
2695     St->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2696     return St;
2697   }
2698   case ARM::BI__builtin_neon_vst2_v:
2699   case ARM::BI__builtin_neon_vst2q_v:
2700     Ops.push_back(Align);
2701     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty),
2702                         Ops, "");
2703   case ARM::BI__builtin_neon_vst2_lane_v:
2704   case ARM::BI__builtin_neon_vst2q_lane_v:
2705     Ops.push_back(Align);
2706     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty),
2707                         Ops, "");
2708   case ARM::BI__builtin_neon_vst3_v:
2709   case ARM::BI__builtin_neon_vst3q_v:
2710     Ops.push_back(Align);
2711     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty),
2712                         Ops, "");
2713   case ARM::BI__builtin_neon_vst3_lane_v:
2714   case ARM::BI__builtin_neon_vst3q_lane_v:
2715     Ops.push_back(Align);
2716     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty),
2717                         Ops, "");
2718   case ARM::BI__builtin_neon_vst4_v:
2719   case ARM::BI__builtin_neon_vst4q_v:
2720     Ops.push_back(Align);
2721     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty),
2722                         Ops, "");
2723   case ARM::BI__builtin_neon_vst4_lane_v:
2724   case ARM::BI__builtin_neon_vst4q_lane_v:
2725     Ops.push_back(Align);
2726     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty),
2727                         Ops, "");
2728   case ARM::BI__builtin_neon_vsubhn_v:
2729     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty),
2730                         Ops, "vsubhn");
2731   case ARM::BI__builtin_neon_vtbl1_v:
2732     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
2733                         Ops, "vtbl1");
2734   case ARM::BI__builtin_neon_vtbl2_v:
2735     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
2736                         Ops, "vtbl2");
2737   case ARM::BI__builtin_neon_vtbl3_v:
2738     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
2739                         Ops, "vtbl3");
2740   case ARM::BI__builtin_neon_vtbl4_v:
2741     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
2742                         Ops, "vtbl4");
2743   case ARM::BI__builtin_neon_vtbx1_v:
2744     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
2745                         Ops, "vtbx1");
2746   case ARM::BI__builtin_neon_vtbx2_v:
2747     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
2748                         Ops, "vtbx2");
2749   case ARM::BI__builtin_neon_vtbx3_v:
2750     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
2751                         Ops, "vtbx3");
2752   case ARM::BI__builtin_neon_vtbx4_v:
2753     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
2754                         Ops, "vtbx4");
2755   case ARM::BI__builtin_neon_vtst_v:
2756   case ARM::BI__builtin_neon_vtstq_v: {
2757     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2758     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2759     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
2760     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
2761                                 ConstantAggregateZero::get(Ty));
2762     return Builder.CreateSExt(Ops[0], Ty, "vtst");
2763   }
2764   case ARM::BI__builtin_neon_vtrn_v:
2765   case ARM::BI__builtin_neon_vtrnq_v: {
2766     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2767     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2768     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2769     Value *SV = 0;
2770 
2771     for (unsigned vi = 0; vi != 2; ++vi) {
2772       SmallVector<Constant*, 16> Indices;
2773       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2774         Indices.push_back(Builder.getInt32(i+vi));
2775         Indices.push_back(Builder.getInt32(i+e+vi));
2776       }
2777       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2778       SV = llvm::ConstantVector::get(Indices);
2779       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
2780       SV = Builder.CreateStore(SV, Addr);
2781     }
2782     return SV;
2783   }
2784   case ARM::BI__builtin_neon_vuzp_v:
2785   case ARM::BI__builtin_neon_vuzpq_v: {
2786     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2787     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2788     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2789     Value *SV = 0;
2790 
2791     for (unsigned vi = 0; vi != 2; ++vi) {
2792       SmallVector<Constant*, 16> Indices;
2793       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2794         Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
2795 
2796       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2797       SV = llvm::ConstantVector::get(Indices);
2798       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
2799       SV = Builder.CreateStore(SV, Addr);
2800     }
2801     return SV;
2802   }
2803   case ARM::BI__builtin_neon_vzip_v:
2804   case ARM::BI__builtin_neon_vzipq_v: {
2805     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2806     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2807     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2808     Value *SV = 0;
2809 
2810     for (unsigned vi = 0; vi != 2; ++vi) {
2811       SmallVector<Constant*, 16> Indices;
2812       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2813         Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
2814         Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
2815       }
2816       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2817       SV = llvm::ConstantVector::get(Indices);
2818       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
2819       SV = Builder.CreateStore(SV, Addr);
2820     }
2821     return SV;
2822   }
2823   }
2824 }
2825 
2826 llvm::Value *CodeGenFunction::
BuildVector(ArrayRef<llvm::Value * > Ops)2827 BuildVector(ArrayRef<llvm::Value*> Ops) {
2828   assert((Ops.size() & (Ops.size() - 1)) == 0 &&
2829          "Not a power-of-two sized vector!");
2830   bool AllConstants = true;
2831   for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
2832     AllConstants &= isa<Constant>(Ops[i]);
2833 
2834   // If this is a constant vector, create a ConstantVector.
2835   if (AllConstants) {
2836     SmallVector<llvm::Constant*, 16> CstOps;
2837     for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2838       CstOps.push_back(cast<Constant>(Ops[i]));
2839     return llvm::ConstantVector::get(CstOps);
2840   }
2841 
2842   // Otherwise, insertelement the values to build the vector.
2843   Value *Result =
2844     llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
2845 
2846   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2847     Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
2848 
2849   return Result;
2850 }
2851 
EmitX86BuiltinExpr(unsigned BuiltinID,const CallExpr * E)2852 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
2853                                            const CallExpr *E) {
2854   SmallVector<Value*, 4> Ops;
2855 
2856   // Find out if any arguments are required to be integer constant expressions.
2857   unsigned ICEArguments = 0;
2858   ASTContext::GetBuiltinTypeError Error;
2859   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
2860   assert(Error == ASTContext::GE_None && "Should not codegen an error");
2861 
2862   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
2863     // If this is a normal argument, just emit it as a scalar.
2864     if ((ICEArguments & (1 << i)) == 0) {
2865       Ops.push_back(EmitScalarExpr(E->getArg(i)));
2866       continue;
2867     }
2868 
2869     // If this is required to be a constant, constant fold it so that we know
2870     // that the generated intrinsic gets a ConstantInt.
2871     llvm::APSInt Result;
2872     bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
2873     assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
2874     Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
2875   }
2876 
2877   switch (BuiltinID) {
2878   default: return 0;
2879   case X86::BI__builtin_ia32_vec_init_v8qi:
2880   case X86::BI__builtin_ia32_vec_init_v4hi:
2881   case X86::BI__builtin_ia32_vec_init_v2si:
2882     return Builder.CreateBitCast(BuildVector(Ops),
2883                                  llvm::Type::getX86_MMXTy(getLLVMContext()));
2884   case X86::BI__builtin_ia32_vec_ext_v2si:
2885     return Builder.CreateExtractElement(Ops[0],
2886                                   llvm::ConstantInt::get(Ops[1]->getType(), 0));
2887   case X86::BI__builtin_ia32_ldmxcsr: {
2888     llvm::Type *PtrTy = Int8PtrTy;
2889     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2890     Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2891     Builder.CreateStore(Ops[0], Tmp);
2892     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
2893                               Builder.CreateBitCast(Tmp, PtrTy));
2894   }
2895   case X86::BI__builtin_ia32_stmxcsr: {
2896     llvm::Type *PtrTy = Int8PtrTy;
2897     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
2898     Value *Tmp = Builder.CreateAlloca(Int32Ty, One);
2899     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
2900                        Builder.CreateBitCast(Tmp, PtrTy));
2901     return Builder.CreateLoad(Tmp, "stmxcsr");
2902   }
2903   case X86::BI__builtin_ia32_storehps:
2904   case X86::BI__builtin_ia32_storelps: {
2905     llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty);
2906     llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2907 
2908     // cast val v2i64
2909     Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast");
2910 
2911     // extract (0, 1)
2912     unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
2913     llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index);
2914     Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
2915 
2916     // cast pointer to i64 & store
2917     Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
2918     return Builder.CreateStore(Ops[1], Ops[0]);
2919   }
2920   case X86::BI__builtin_ia32_palignr: {
2921     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2922 
2923     // If palignr is shifting the pair of input vectors less than 9 bytes,
2924     // emit a shuffle instruction.
2925     if (shiftVal <= 8) {
2926       SmallVector<llvm::Constant*, 8> Indices;
2927       for (unsigned i = 0; i != 8; ++i)
2928         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2929 
2930       Value* SV = llvm::ConstantVector::get(Indices);
2931       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2932     }
2933 
2934     // If palignr is shifting the pair of input vectors more than 8 but less
2935     // than 16 bytes, emit a logical right shift of the destination.
2936     if (shiftVal < 16) {
2937       // MMX has these as 1 x i64 vectors for some odd optimization reasons.
2938       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1);
2939 
2940       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2941       Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8);
2942 
2943       // create i32 constant
2944       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q);
2945       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2946     }
2947 
2948     // If palignr is shifting the pair of vectors more than 16 bytes, emit zero.
2949     return llvm::Constant::getNullValue(ConvertType(E->getType()));
2950   }
2951   case X86::BI__builtin_ia32_palignr128: {
2952     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2953 
2954     // If palignr is shifting the pair of input vectors less than 17 bytes,
2955     // emit a shuffle instruction.
2956     if (shiftVal <= 16) {
2957       SmallVector<llvm::Constant*, 16> Indices;
2958       for (unsigned i = 0; i != 16; ++i)
2959         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
2960 
2961       Value* SV = llvm::ConstantVector::get(Indices);
2962       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
2963     }
2964 
2965     // If palignr is shifting the pair of input vectors more than 16 but less
2966     // than 32 bytes, emit a logical right shift of the destination.
2967     if (shiftVal < 32) {
2968       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
2969 
2970       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
2971       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
2972 
2973       // create i32 constant
2974       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq);
2975       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
2976     }
2977 
2978     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
2979     return llvm::Constant::getNullValue(ConvertType(E->getType()));
2980   }
2981   case X86::BI__builtin_ia32_palignr256: {
2982     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
2983 
2984     // If palignr is shifting the pair of input vectors less than 17 bytes,
2985     // emit a shuffle instruction.
2986     if (shiftVal <= 16) {
2987       SmallVector<llvm::Constant*, 32> Indices;
2988       // 256-bit palignr operates on 128-bit lanes so we need to handle that
2989       for (unsigned l = 0; l != 2; ++l) {
2990         unsigned LaneStart = l * 16;
2991         unsigned LaneEnd = (l+1) * 16;
2992         for (unsigned i = 0; i != 16; ++i) {
2993           unsigned Idx = shiftVal + i + LaneStart;
2994           if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand
2995           Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx));
2996         }
2997       }
2998 
2999       Value* SV = llvm::ConstantVector::get(Indices);
3000       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
3001     }
3002 
3003     // If palignr is shifting the pair of input vectors more than 16 but less
3004     // than 32 bytes, emit a logical right shift of the destination.
3005     if (shiftVal < 32) {
3006       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4);
3007 
3008       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
3009       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
3010 
3011       // create i32 constant
3012       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq);
3013       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
3014     }
3015 
3016     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
3017     return llvm::Constant::getNullValue(ConvertType(E->getType()));
3018   }
3019   case X86::BI__builtin_ia32_movntps:
3020   case X86::BI__builtin_ia32_movntps256:
3021   case X86::BI__builtin_ia32_movntpd:
3022   case X86::BI__builtin_ia32_movntpd256:
3023   case X86::BI__builtin_ia32_movntdq:
3024   case X86::BI__builtin_ia32_movntdq256:
3025   case X86::BI__builtin_ia32_movnti: {
3026     llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(),
3027                                            Builder.getInt32(1));
3028 
3029     // Convert the type of the pointer to a pointer to the stored type.
3030     Value *BC = Builder.CreateBitCast(Ops[0],
3031                                 llvm::PointerType::getUnqual(Ops[1]->getType()),
3032                                       "cast");
3033     StoreInst *SI = Builder.CreateStore(Ops[1], BC);
3034     SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
3035     SI->setAlignment(16);
3036     return SI;
3037   }
3038   // 3DNow!
3039   case X86::BI__builtin_ia32_pswapdsf:
3040   case X86::BI__builtin_ia32_pswapdsi: {
3041     const char *name = 0;
3042     Intrinsic::ID ID = Intrinsic::not_intrinsic;
3043     switch(BuiltinID) {
3044     default: llvm_unreachable("Unsupported intrinsic!");
3045     case X86::BI__builtin_ia32_pswapdsf:
3046     case X86::BI__builtin_ia32_pswapdsi:
3047       name = "pswapd";
3048       ID = Intrinsic::x86_3dnowa_pswapd;
3049       break;
3050     }
3051     llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
3052     Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
3053     llvm::Function *F = CGM.getIntrinsic(ID);
3054     return Builder.CreateCall(F, Ops, name);
3055   }
3056   case X86::BI__builtin_ia32_rdrand16_step:
3057   case X86::BI__builtin_ia32_rdrand32_step:
3058   case X86::BI__builtin_ia32_rdrand64_step:
3059   case X86::BI__builtin_ia32_rdseed16_step:
3060   case X86::BI__builtin_ia32_rdseed32_step:
3061   case X86::BI__builtin_ia32_rdseed64_step: {
3062     Intrinsic::ID ID;
3063     switch (BuiltinID) {
3064     default: llvm_unreachable("Unsupported intrinsic!");
3065     case X86::BI__builtin_ia32_rdrand16_step:
3066       ID = Intrinsic::x86_rdrand_16;
3067       break;
3068     case X86::BI__builtin_ia32_rdrand32_step:
3069       ID = Intrinsic::x86_rdrand_32;
3070       break;
3071     case X86::BI__builtin_ia32_rdrand64_step:
3072       ID = Intrinsic::x86_rdrand_64;
3073       break;
3074     case X86::BI__builtin_ia32_rdseed16_step:
3075       ID = Intrinsic::x86_rdseed_16;
3076       break;
3077     case X86::BI__builtin_ia32_rdseed32_step:
3078       ID = Intrinsic::x86_rdseed_32;
3079       break;
3080     case X86::BI__builtin_ia32_rdseed64_step:
3081       ID = Intrinsic::x86_rdseed_64;
3082       break;
3083     }
3084 
3085     Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
3086     Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]);
3087     return Builder.CreateExtractValue(Call, 1);
3088   }
3089   }
3090 }
3091 
3092 
EmitPPCBuiltinExpr(unsigned BuiltinID,const CallExpr * E)3093 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
3094                                            const CallExpr *E) {
3095   SmallVector<Value*, 4> Ops;
3096 
3097   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
3098     Ops.push_back(EmitScalarExpr(E->getArg(i)));
3099 
3100   Intrinsic::ID ID = Intrinsic::not_intrinsic;
3101 
3102   switch (BuiltinID) {
3103   default: return 0;
3104 
3105   // vec_ld, vec_lvsl, vec_lvsr
3106   case PPC::BI__builtin_altivec_lvx:
3107   case PPC::BI__builtin_altivec_lvxl:
3108   case PPC::BI__builtin_altivec_lvebx:
3109   case PPC::BI__builtin_altivec_lvehx:
3110   case PPC::BI__builtin_altivec_lvewx:
3111   case PPC::BI__builtin_altivec_lvsl:
3112   case PPC::BI__builtin_altivec_lvsr:
3113   {
3114     Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
3115 
3116     Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
3117     Ops.pop_back();
3118 
3119     switch (BuiltinID) {
3120     default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
3121     case PPC::BI__builtin_altivec_lvx:
3122       ID = Intrinsic::ppc_altivec_lvx;
3123       break;
3124     case PPC::BI__builtin_altivec_lvxl:
3125       ID = Intrinsic::ppc_altivec_lvxl;
3126       break;
3127     case PPC::BI__builtin_altivec_lvebx:
3128       ID = Intrinsic::ppc_altivec_lvebx;
3129       break;
3130     case PPC::BI__builtin_altivec_lvehx:
3131       ID = Intrinsic::ppc_altivec_lvehx;
3132       break;
3133     case PPC::BI__builtin_altivec_lvewx:
3134       ID = Intrinsic::ppc_altivec_lvewx;
3135       break;
3136     case PPC::BI__builtin_altivec_lvsl:
3137       ID = Intrinsic::ppc_altivec_lvsl;
3138       break;
3139     case PPC::BI__builtin_altivec_lvsr:
3140       ID = Intrinsic::ppc_altivec_lvsr;
3141       break;
3142     }
3143     llvm::Function *F = CGM.getIntrinsic(ID);
3144     return Builder.CreateCall(F, Ops, "");
3145   }
3146 
3147   // vec_st
3148   case PPC::BI__builtin_altivec_stvx:
3149   case PPC::BI__builtin_altivec_stvxl:
3150   case PPC::BI__builtin_altivec_stvebx:
3151   case PPC::BI__builtin_altivec_stvehx:
3152   case PPC::BI__builtin_altivec_stvewx:
3153   {
3154     Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
3155     Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
3156     Ops.pop_back();
3157 
3158     switch (BuiltinID) {
3159     default: llvm_unreachable("Unsupported st intrinsic!");
3160     case PPC::BI__builtin_altivec_stvx:
3161       ID = Intrinsic::ppc_altivec_stvx;
3162       break;
3163     case PPC::BI__builtin_altivec_stvxl:
3164       ID = Intrinsic::ppc_altivec_stvxl;
3165       break;
3166     case PPC::BI__builtin_altivec_stvebx:
3167       ID = Intrinsic::ppc_altivec_stvebx;
3168       break;
3169     case PPC::BI__builtin_altivec_stvehx:
3170       ID = Intrinsic::ppc_altivec_stvehx;
3171       break;
3172     case PPC::BI__builtin_altivec_stvewx:
3173       ID = Intrinsic::ppc_altivec_stvewx;
3174       break;
3175     }
3176     llvm::Function *F = CGM.getIntrinsic(ID);
3177     return Builder.CreateCall(F, Ops, "");
3178   }
3179   }
3180 }
3181