• 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 "clang/CodeGen/CGFunctionInfo.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Intrinsics.h"
25 
26 using namespace clang;
27 using namespace CodeGen;
28 using namespace llvm;
29 
30 /// getBuiltinLibFunction - Given a builtin id for a function like
31 /// "__builtin_fabsf", return a Function* for "fabsf".
getBuiltinLibFunction(const FunctionDecl * FD,unsigned BuiltinID)32 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
33                                                   unsigned BuiltinID) {
34   assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
35 
36   // Get the name, skip over the __builtin_ prefix (if necessary).
37   StringRef Name;
38   GlobalDecl D(FD);
39 
40   // If the builtin has been declared explicitly with an assembler label,
41   // use the mangled name. This differs from the plain label on platforms
42   // that prefix labels.
43   if (FD->hasAttr<AsmLabelAttr>())
44     Name = getMangledName(D);
45   else
46     Name = Context.BuiltinInfo.GetName(BuiltinID) + 10;
47 
48   llvm::FunctionType *Ty =
49     cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
50 
51   return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
52 }
53 
54 /// Emit the conversions required to turn the given value into an
55 /// integer of the given size.
EmitToInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::IntegerType * IntType)56 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
57                         QualType T, llvm::IntegerType *IntType) {
58   V = CGF.EmitToMemory(V, T);
59 
60   if (V->getType()->isPointerTy())
61     return CGF.Builder.CreatePtrToInt(V, IntType);
62 
63   assert(V->getType() == IntType);
64   return V;
65 }
66 
EmitFromInt(CodeGenFunction & CGF,llvm::Value * V,QualType T,llvm::Type * ResultType)67 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
68                           QualType T, llvm::Type *ResultType) {
69   V = CGF.EmitFromMemory(V, T);
70 
71   if (ResultType->isPointerTy())
72     return CGF.Builder.CreateIntToPtr(V, ResultType);
73 
74   assert(V->getType() == ResultType);
75   return V;
76 }
77 
78 /// Utility to insert an atomic instruction based on Instrinsic::ID
79 /// and the expression node.
EmitBinaryAtomic(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E)80 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
81                                llvm::AtomicRMWInst::BinOp Kind,
82                                const CallExpr *E) {
83   QualType T = E->getType();
84   assert(E->getArg(0)->getType()->isPointerType());
85   assert(CGF.getContext().hasSameUnqualifiedType(T,
86                                   E->getArg(0)->getType()->getPointeeType()));
87   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
88 
89   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
90   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
91 
92   llvm::IntegerType *IntType =
93     llvm::IntegerType::get(CGF.getLLVMContext(),
94                            CGF.getContext().getTypeSize(T));
95   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
96 
97   llvm::Value *Args[2];
98   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
99   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
100   llvm::Type *ValueType = Args[1]->getType();
101   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
102 
103   llvm::Value *Result =
104       CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
105                                   llvm::SequentiallyConsistent);
106   Result = EmitFromInt(CGF, Result, T, ValueType);
107   return RValue::get(Result);
108 }
109 
110 /// Utility to insert an atomic instruction based Instrinsic::ID and
111 /// the expression node, where the return value is the result of the
112 /// operation.
EmitBinaryAtomicPost(CodeGenFunction & CGF,llvm::AtomicRMWInst::BinOp Kind,const CallExpr * E,Instruction::BinaryOps Op)113 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
114                                    llvm::AtomicRMWInst::BinOp Kind,
115                                    const CallExpr *E,
116                                    Instruction::BinaryOps Op) {
117   QualType T = E->getType();
118   assert(E->getArg(0)->getType()->isPointerType());
119   assert(CGF.getContext().hasSameUnqualifiedType(T,
120                                   E->getArg(0)->getType()->getPointeeType()));
121   assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
122 
123   llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
124   unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
125 
126   llvm::IntegerType *IntType =
127     llvm::IntegerType::get(CGF.getLLVMContext(),
128                            CGF.getContext().getTypeSize(T));
129   llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
130 
131   llvm::Value *Args[2];
132   Args[1] = CGF.EmitScalarExpr(E->getArg(1));
133   llvm::Type *ValueType = Args[1]->getType();
134   Args[1] = EmitToInt(CGF, Args[1], T, IntType);
135   Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
136 
137   llvm::Value *Result =
138       CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1],
139                                   llvm::SequentiallyConsistent);
140   Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
141   Result = EmitFromInt(CGF, Result, T, ValueType);
142   return RValue::get(Result);
143 }
144 
145 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy,
146 /// which must be a scalar floating point type.
EmitFAbs(CodeGenFunction & CGF,Value * V,QualType ValTy)147 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) {
148   const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>();
149   assert(ValTyP && "isn't scalar fp type!");
150 
151   StringRef FnName;
152   switch (ValTyP->getKind()) {
153   default: llvm_unreachable("Isn't a scalar fp type!");
154   case BuiltinType::Float:      FnName = "fabsf"; break;
155   case BuiltinType::Double:     FnName = "fabs"; break;
156   case BuiltinType::LongDouble: FnName = "fabsl"; break;
157   }
158 
159   // The prototype is something that takes and returns whatever V's type is.
160   llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(),
161                                                    false);
162   llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName);
163 
164   return CGF.EmitNounwindRuntimeCall(Fn, V, "abs");
165 }
166 
emitLibraryCall(CodeGenFunction & CGF,const FunctionDecl * Fn,const CallExpr * E,llvm::Value * calleeValue)167 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn,
168                               const CallExpr *E, llvm::Value *calleeValue) {
169   return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E->getLocStart(),
170                       ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn);
171 }
172 
173 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
174 /// depending on IntrinsicID.
175 ///
176 /// \arg CGF The current codegen function.
177 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
178 /// \arg X The first argument to the llvm.*.with.overflow.*.
179 /// \arg Y The second argument to the llvm.*.with.overflow.*.
180 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
181 /// \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)182 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
183                                           const llvm::Intrinsic::ID IntrinsicID,
184                                           llvm::Value *X, llvm::Value *Y,
185                                           llvm::Value *&Carry) {
186   // Make sure we have integers of the same width.
187   assert(X->getType() == Y->getType() &&
188          "Arguments must be the same type. (Did you forget to make sure both "
189          "arguments have the same integer width?)");
190 
191   llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
192   llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y);
193   Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
194   return CGF.Builder.CreateExtractValue(Tmp, 0);
195 }
196 
EmitBuiltinExpr(const FunctionDecl * FD,unsigned BuiltinID,const CallExpr * E)197 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
198                                         unsigned BuiltinID, const CallExpr *E) {
199   // See if we can constant fold this builtin.  If so, don't emit it at all.
200   Expr::EvalResult Result;
201   if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
202       !Result.hasSideEffects()) {
203     if (Result.Val.isInt())
204       return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
205                                                 Result.Val.getInt()));
206     if (Result.Val.isFloat())
207       return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
208                                                Result.Val.getFloat()));
209   }
210 
211   switch (BuiltinID) {
212   default: break;  // Handle intrinsics and libm functions below.
213   case Builtin::BI__builtin___CFStringMakeConstantString:
214   case Builtin::BI__builtin___NSStringMakeConstantString:
215     return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr));
216   case Builtin::BI__builtin_stdarg_start:
217   case Builtin::BI__builtin_va_start:
218   case Builtin::BI__va_start:
219   case Builtin::BI__builtin_va_end: {
220     Value *ArgValue = (BuiltinID == Builtin::BI__va_start)
221                           ? EmitScalarExpr(E->getArg(0))
222                           : EmitVAListRef(E->getArg(0));
223     llvm::Type *DestType = Int8PtrTy;
224     if (ArgValue->getType() != DestType)
225       ArgValue = Builder.CreateBitCast(ArgValue, DestType,
226                                        ArgValue->getName().data());
227 
228     Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ?
229       Intrinsic::vaend : Intrinsic::vastart;
230     return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue));
231   }
232   case Builtin::BI__builtin_va_copy: {
233     Value *DstPtr = EmitVAListRef(E->getArg(0));
234     Value *SrcPtr = EmitVAListRef(E->getArg(1));
235 
236     llvm::Type *Type = Int8PtrTy;
237 
238     DstPtr = Builder.CreateBitCast(DstPtr, Type);
239     SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
240     return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy),
241                                            DstPtr, SrcPtr));
242   }
243   case Builtin::BI__builtin_abs:
244   case Builtin::BI__builtin_labs:
245   case Builtin::BI__builtin_llabs: {
246     Value *ArgValue = EmitScalarExpr(E->getArg(0));
247 
248     Value *NegOp = Builder.CreateNeg(ArgValue, "neg");
249     Value *CmpResult =
250     Builder.CreateICmpSGE(ArgValue,
251                           llvm::Constant::getNullValue(ArgValue->getType()),
252                                                             "abscond");
253     Value *Result =
254       Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs");
255 
256     return RValue::get(Result);
257   }
258 
259   case Builtin::BI__builtin_conj:
260   case Builtin::BI__builtin_conjf:
261   case Builtin::BI__builtin_conjl: {
262     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
263     Value *Real = ComplexVal.first;
264     Value *Imag = ComplexVal.second;
265     Value *Zero =
266       Imag->getType()->isFPOrFPVectorTy()
267         ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType())
268         : llvm::Constant::getNullValue(Imag->getType());
269 
270     Imag = Builder.CreateFSub(Zero, Imag, "sub");
271     return RValue::getComplex(std::make_pair(Real, Imag));
272   }
273   case Builtin::BI__builtin_creal:
274   case Builtin::BI__builtin_crealf:
275   case Builtin::BI__builtin_creall:
276   case Builtin::BIcreal:
277   case Builtin::BIcrealf:
278   case Builtin::BIcreall: {
279     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
280     return RValue::get(ComplexVal.first);
281   }
282 
283   case Builtin::BI__builtin_cimag:
284   case Builtin::BI__builtin_cimagf:
285   case Builtin::BI__builtin_cimagl:
286   case Builtin::BIcimag:
287   case Builtin::BIcimagf:
288   case Builtin::BIcimagl: {
289     ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
290     return RValue::get(ComplexVal.second);
291   }
292 
293   case Builtin::BI__builtin_ctzs:
294   case Builtin::BI__builtin_ctz:
295   case Builtin::BI__builtin_ctzl:
296   case Builtin::BI__builtin_ctzll: {
297     Value *ArgValue = EmitScalarExpr(E->getArg(0));
298 
299     llvm::Type *ArgType = ArgValue->getType();
300     Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
301 
302     llvm::Type *ResultType = ConvertType(E->getType());
303     Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
304     Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
305     if (Result->getType() != ResultType)
306       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
307                                      "cast");
308     return RValue::get(Result);
309   }
310   case Builtin::BI__builtin_clzs:
311   case Builtin::BI__builtin_clz:
312   case Builtin::BI__builtin_clzl:
313   case Builtin::BI__builtin_clzll: {
314     Value *ArgValue = EmitScalarExpr(E->getArg(0));
315 
316     llvm::Type *ArgType = ArgValue->getType();
317     Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
318 
319     llvm::Type *ResultType = ConvertType(E->getType());
320     Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
321     Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef);
322     if (Result->getType() != ResultType)
323       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
324                                      "cast");
325     return RValue::get(Result);
326   }
327   case Builtin::BI__builtin_ffs:
328   case Builtin::BI__builtin_ffsl:
329   case Builtin::BI__builtin_ffsll: {
330     // ffs(x) -> x ? cttz(x) + 1 : 0
331     Value *ArgValue = EmitScalarExpr(E->getArg(0));
332 
333     llvm::Type *ArgType = ArgValue->getType();
334     Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
335 
336     llvm::Type *ResultType = ConvertType(E->getType());
337     Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue,
338                                                        Builder.getTrue()),
339                                    llvm::ConstantInt::get(ArgType, 1));
340     Value *Zero = llvm::Constant::getNullValue(ArgType);
341     Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
342     Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
343     if (Result->getType() != ResultType)
344       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
345                                      "cast");
346     return RValue::get(Result);
347   }
348   case Builtin::BI__builtin_parity:
349   case Builtin::BI__builtin_parityl:
350   case Builtin::BI__builtin_parityll: {
351     // parity(x) -> ctpop(x) & 1
352     Value *ArgValue = EmitScalarExpr(E->getArg(0));
353 
354     llvm::Type *ArgType = ArgValue->getType();
355     Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
356 
357     llvm::Type *ResultType = ConvertType(E->getType());
358     Value *Tmp = Builder.CreateCall(F, ArgValue);
359     Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
360     if (Result->getType() != ResultType)
361       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
362                                      "cast");
363     return RValue::get(Result);
364   }
365   case Builtin::BI__builtin_popcount:
366   case Builtin::BI__builtin_popcountl:
367   case Builtin::BI__builtin_popcountll: {
368     Value *ArgValue = EmitScalarExpr(E->getArg(0));
369 
370     llvm::Type *ArgType = ArgValue->getType();
371     Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
372 
373     llvm::Type *ResultType = ConvertType(E->getType());
374     Value *Result = Builder.CreateCall(F, ArgValue);
375     if (Result->getType() != ResultType)
376       Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
377                                      "cast");
378     return RValue::get(Result);
379   }
380   case Builtin::BI__builtin_expect: {
381     Value *ArgValue = EmitScalarExpr(E->getArg(0));
382     llvm::Type *ArgType = ArgValue->getType();
383 
384     Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
385     Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
386 
387     Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue,
388                                         "expval");
389     return RValue::get(Result);
390   }
391   case Builtin::BI__builtin_bswap16:
392   case Builtin::BI__builtin_bswap32:
393   case Builtin::BI__builtin_bswap64: {
394     Value *ArgValue = EmitScalarExpr(E->getArg(0));
395     llvm::Type *ArgType = ArgValue->getType();
396     Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType);
397     return RValue::get(Builder.CreateCall(F, ArgValue));
398   }
399   case Builtin::BI__builtin_object_size: {
400     // We rely on constant folding to deal with expressions with side effects.
401     assert(!E->getArg(0)->HasSideEffects(getContext()) &&
402            "should have been constant folded");
403 
404     // We pass this builtin onto the optimizer so that it can
405     // figure out the object size in more complex cases.
406     llvm::Type *ResType = ConvertType(E->getType());
407 
408     // LLVM only supports 0 and 2, make sure that we pass along that
409     // as a boolean.
410     Value *Ty = EmitScalarExpr(E->getArg(1));
411     ConstantInt *CI = dyn_cast<ConstantInt>(Ty);
412     assert(CI);
413     uint64_t val = CI->getZExtValue();
414     CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1);
415     // FIXME: Get right address space.
416     llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) };
417     Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys);
418     return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI));
419   }
420   case Builtin::BI__builtin_prefetch: {
421     Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
422     // FIXME: Technically these constants should of type 'int', yes?
423     RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
424       llvm::ConstantInt::get(Int32Ty, 0);
425     Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
426       llvm::ConstantInt::get(Int32Ty, 3);
427     Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
428     Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
429     return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data));
430   }
431   case Builtin::BI__builtin_readcyclecounter: {
432     Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
433     return RValue::get(Builder.CreateCall(F));
434   }
435   case Builtin::BI__builtin___clear_cache: {
436     Value *Begin = EmitScalarExpr(E->getArg(0));
437     Value *End = EmitScalarExpr(E->getArg(1));
438     Value *F = CGM.getIntrinsic(Intrinsic::clear_cache);
439     return RValue::get(Builder.CreateCall2(F, Begin, End));
440   }
441   case Builtin::BI__builtin_trap: {
442     Value *F = CGM.getIntrinsic(Intrinsic::trap);
443     return RValue::get(Builder.CreateCall(F));
444   }
445   case Builtin::BI__debugbreak: {
446     Value *F = CGM.getIntrinsic(Intrinsic::debugtrap);
447     return RValue::get(Builder.CreateCall(F));
448   }
449   case Builtin::BI__builtin_unreachable: {
450     if (SanOpts->Unreachable)
451       EmitCheck(Builder.getFalse(), "builtin_unreachable",
452                 EmitCheckSourceLocation(E->getExprLoc()),
453                 ArrayRef<llvm::Value *>(), CRK_Unrecoverable);
454     else
455       Builder.CreateUnreachable();
456 
457     // We do need to preserve an insertion point.
458     EmitBlock(createBasicBlock("unreachable.cont"));
459 
460     return RValue::get(nullptr);
461   }
462 
463   case Builtin::BI__builtin_powi:
464   case Builtin::BI__builtin_powif:
465   case Builtin::BI__builtin_powil: {
466     Value *Base = EmitScalarExpr(E->getArg(0));
467     Value *Exponent = EmitScalarExpr(E->getArg(1));
468     llvm::Type *ArgType = Base->getType();
469     Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType);
470     return RValue::get(Builder.CreateCall2(F, Base, Exponent));
471   }
472 
473   case Builtin::BI__builtin_isgreater:
474   case Builtin::BI__builtin_isgreaterequal:
475   case Builtin::BI__builtin_isless:
476   case Builtin::BI__builtin_islessequal:
477   case Builtin::BI__builtin_islessgreater:
478   case Builtin::BI__builtin_isunordered: {
479     // Ordered comparisons: we know the arguments to these are matching scalar
480     // floating point values.
481     Value *LHS = EmitScalarExpr(E->getArg(0));
482     Value *RHS = EmitScalarExpr(E->getArg(1));
483 
484     switch (BuiltinID) {
485     default: llvm_unreachable("Unknown ordered comparison");
486     case Builtin::BI__builtin_isgreater:
487       LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
488       break;
489     case Builtin::BI__builtin_isgreaterequal:
490       LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
491       break;
492     case Builtin::BI__builtin_isless:
493       LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
494       break;
495     case Builtin::BI__builtin_islessequal:
496       LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
497       break;
498     case Builtin::BI__builtin_islessgreater:
499       LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
500       break;
501     case Builtin::BI__builtin_isunordered:
502       LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
503       break;
504     }
505     // ZExt bool to int type.
506     return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
507   }
508   case Builtin::BI__builtin_isnan: {
509     Value *V = EmitScalarExpr(E->getArg(0));
510     V = Builder.CreateFCmpUNO(V, V, "cmp");
511     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
512   }
513 
514   case Builtin::BI__builtin_isinf: {
515     // isinf(x) --> fabs(x) == infinity
516     Value *V = EmitScalarExpr(E->getArg(0));
517     V = EmitFAbs(*this, V, E->getArg(0)->getType());
518 
519     V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf");
520     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
521   }
522 
523   // TODO: BI__builtin_isinf_sign
524   //   isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0
525 
526   case Builtin::BI__builtin_isnormal: {
527     // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
528     Value *V = EmitScalarExpr(E->getArg(0));
529     Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
530 
531     Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
532     Value *IsLessThanInf =
533       Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
534     APFloat Smallest = APFloat::getSmallestNormalized(
535                    getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
536     Value *IsNormal =
537       Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
538                             "isnormal");
539     V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
540     V = Builder.CreateAnd(V, IsNormal, "and");
541     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
542   }
543 
544   case Builtin::BI__builtin_isfinite: {
545     // isfinite(x) --> x == x && fabs(x) != infinity;
546     Value *V = EmitScalarExpr(E->getArg(0));
547     Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
548 
549     Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType());
550     Value *IsNotInf =
551       Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
552 
553     V = Builder.CreateAnd(Eq, IsNotInf, "and");
554     return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
555   }
556 
557   case Builtin::BI__builtin_fpclassify: {
558     Value *V = EmitScalarExpr(E->getArg(5));
559     llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
560 
561     // Create Result
562     BasicBlock *Begin = Builder.GetInsertBlock();
563     BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
564     Builder.SetInsertPoint(End);
565     PHINode *Result =
566       Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
567                         "fpclassify_result");
568 
569     // if (V==0) return FP_ZERO
570     Builder.SetInsertPoint(Begin);
571     Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
572                                           "iszero");
573     Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
574     BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
575     Builder.CreateCondBr(IsZero, End, NotZero);
576     Result->addIncoming(ZeroLiteral, Begin);
577 
578     // if (V != V) return FP_NAN
579     Builder.SetInsertPoint(NotZero);
580     Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
581     Value *NanLiteral = EmitScalarExpr(E->getArg(0));
582     BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
583     Builder.CreateCondBr(IsNan, End, NotNan);
584     Result->addIncoming(NanLiteral, NotZero);
585 
586     // if (fabs(V) == infinity) return FP_INFINITY
587     Builder.SetInsertPoint(NotNan);
588     Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType());
589     Value *IsInf =
590       Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
591                             "isinf");
592     Value *InfLiteral = EmitScalarExpr(E->getArg(1));
593     BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
594     Builder.CreateCondBr(IsInf, End, NotInf);
595     Result->addIncoming(InfLiteral, NotNan);
596 
597     // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
598     Builder.SetInsertPoint(NotInf);
599     APFloat Smallest = APFloat::getSmallestNormalized(
600         getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
601     Value *IsNormal =
602       Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
603                             "isnormal");
604     Value *NormalResult =
605       Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
606                            EmitScalarExpr(E->getArg(3)));
607     Builder.CreateBr(End);
608     Result->addIncoming(NormalResult, NotInf);
609 
610     // return Result
611     Builder.SetInsertPoint(End);
612     return RValue::get(Result);
613   }
614 
615   case Builtin::BIalloca:
616   case Builtin::BI_alloca:
617   case Builtin::BI__builtin_alloca: {
618     Value *Size = EmitScalarExpr(E->getArg(0));
619     return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size));
620   }
621   case Builtin::BIbzero:
622   case Builtin::BI__builtin_bzero: {
623     std::pair<llvm::Value*, unsigned> Dest =
624         EmitPointerWithAlignment(E->getArg(0));
625     Value *SizeVal = EmitScalarExpr(E->getArg(1));
626     Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal,
627                          Dest.second, false);
628     return RValue::get(Dest.first);
629   }
630   case Builtin::BImemcpy:
631   case Builtin::BI__builtin_memcpy: {
632     std::pair<llvm::Value*, unsigned> Dest =
633         EmitPointerWithAlignment(E->getArg(0));
634     std::pair<llvm::Value*, unsigned> Src =
635         EmitPointerWithAlignment(E->getArg(1));
636     Value *SizeVal = EmitScalarExpr(E->getArg(2));
637     unsigned Align = std::min(Dest.second, Src.second);
638     Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
639     return RValue::get(Dest.first);
640   }
641 
642   case Builtin::BI__builtin___memcpy_chk: {
643     // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
644     llvm::APSInt Size, DstSize;
645     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
646         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
647       break;
648     if (Size.ugt(DstSize))
649       break;
650     std::pair<llvm::Value*, unsigned> Dest =
651         EmitPointerWithAlignment(E->getArg(0));
652     std::pair<llvm::Value*, unsigned> Src =
653         EmitPointerWithAlignment(E->getArg(1));
654     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
655     unsigned Align = std::min(Dest.second, Src.second);
656     Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false);
657     return RValue::get(Dest.first);
658   }
659 
660   case Builtin::BI__builtin_objc_memmove_collectable: {
661     Value *Address = EmitScalarExpr(E->getArg(0));
662     Value *SrcAddr = EmitScalarExpr(E->getArg(1));
663     Value *SizeVal = EmitScalarExpr(E->getArg(2));
664     CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
665                                                   Address, SrcAddr, SizeVal);
666     return RValue::get(Address);
667   }
668 
669   case Builtin::BI__builtin___memmove_chk: {
670     // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
671     llvm::APSInt Size, DstSize;
672     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
673         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
674       break;
675     if (Size.ugt(DstSize))
676       break;
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 = llvm::ConstantInt::get(Builder.getContext(), Size);
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 
687   case Builtin::BImemmove:
688   case Builtin::BI__builtin_memmove: {
689     std::pair<llvm::Value*, unsigned> Dest =
690         EmitPointerWithAlignment(E->getArg(0));
691     std::pair<llvm::Value*, unsigned> Src =
692         EmitPointerWithAlignment(E->getArg(1));
693     Value *SizeVal = EmitScalarExpr(E->getArg(2));
694     unsigned Align = std::min(Dest.second, Src.second);
695     Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false);
696     return RValue::get(Dest.first);
697   }
698   case Builtin::BImemset:
699   case Builtin::BI__builtin_memset: {
700     std::pair<llvm::Value*, unsigned> Dest =
701         EmitPointerWithAlignment(E->getArg(0));
702     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
703                                          Builder.getInt8Ty());
704     Value *SizeVal = EmitScalarExpr(E->getArg(2));
705     Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
706     return RValue::get(Dest.first);
707   }
708   case Builtin::BI__builtin___memset_chk: {
709     // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
710     llvm::APSInt Size, DstSize;
711     if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) ||
712         !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext()))
713       break;
714     if (Size.ugt(DstSize))
715       break;
716     std::pair<llvm::Value*, unsigned> Dest =
717         EmitPointerWithAlignment(E->getArg(0));
718     Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
719                                          Builder.getInt8Ty());
720     Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
721     Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false);
722     return RValue::get(Dest.first);
723   }
724   case Builtin::BI__builtin_dwarf_cfa: {
725     // The offset in bytes from the first argument to the CFA.
726     //
727     // Why on earth is this in the frontend?  Is there any reason at
728     // all that the backend can't reasonably determine this while
729     // lowering llvm.eh.dwarf.cfa()?
730     //
731     // TODO: If there's a satisfactory reason, add a target hook for
732     // this instead of hard-coding 0, which is correct for most targets.
733     int32_t Offset = 0;
734 
735     Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
736     return RValue::get(Builder.CreateCall(F,
737                                       llvm::ConstantInt::get(Int32Ty, Offset)));
738   }
739   case Builtin::BI__builtin_return_address: {
740     Value *Depth = EmitScalarExpr(E->getArg(0));
741     Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
742     Value *F = CGM.getIntrinsic(Intrinsic::returnaddress);
743     return RValue::get(Builder.CreateCall(F, Depth));
744   }
745   case Builtin::BI__builtin_frame_address: {
746     Value *Depth = EmitScalarExpr(E->getArg(0));
747     Depth = Builder.CreateIntCast(Depth, Int32Ty, false);
748     Value *F = CGM.getIntrinsic(Intrinsic::frameaddress);
749     return RValue::get(Builder.CreateCall(F, Depth));
750   }
751   case Builtin::BI__builtin_extract_return_addr: {
752     Value *Address = EmitScalarExpr(E->getArg(0));
753     Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
754     return RValue::get(Result);
755   }
756   case Builtin::BI__builtin_frob_return_addr: {
757     Value *Address = EmitScalarExpr(E->getArg(0));
758     Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
759     return RValue::get(Result);
760   }
761   case Builtin::BI__builtin_dwarf_sp_column: {
762     llvm::IntegerType *Ty
763       = cast<llvm::IntegerType>(ConvertType(E->getType()));
764     int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
765     if (Column == -1) {
766       CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
767       return RValue::get(llvm::UndefValue::get(Ty));
768     }
769     return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
770   }
771   case Builtin::BI__builtin_init_dwarf_reg_size_table: {
772     Value *Address = EmitScalarExpr(E->getArg(0));
773     if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
774       CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
775     return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
776   }
777   case Builtin::BI__builtin_eh_return: {
778     Value *Int = EmitScalarExpr(E->getArg(0));
779     Value *Ptr = EmitScalarExpr(E->getArg(1));
780 
781     llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
782     assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
783            "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
784     Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32
785                                   ? Intrinsic::eh_return_i32
786                                   : Intrinsic::eh_return_i64);
787     Builder.CreateCall2(F, Int, Ptr);
788     Builder.CreateUnreachable();
789 
790     // We do need to preserve an insertion point.
791     EmitBlock(createBasicBlock("builtin_eh_return.cont"));
792 
793     return RValue::get(nullptr);
794   }
795   case Builtin::BI__builtin_unwind_init: {
796     Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
797     return RValue::get(Builder.CreateCall(F));
798   }
799   case Builtin::BI__builtin_extend_pointer: {
800     // Extends a pointer to the size of an _Unwind_Word, which is
801     // uint64_t on all platforms.  Generally this gets poked into a
802     // register and eventually used as an address, so if the
803     // addressing registers are wider than pointers and the platform
804     // doesn't implicitly ignore high-order bits when doing
805     // addressing, we need to make sure we zext / sext based on
806     // the platform's expectations.
807     //
808     // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
809 
810     // Cast the pointer to intptr_t.
811     Value *Ptr = EmitScalarExpr(E->getArg(0));
812     Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
813 
814     // If that's 64 bits, we're done.
815     if (IntPtrTy->getBitWidth() == 64)
816       return RValue::get(Result);
817 
818     // Otherwise, ask the codegen data what to do.
819     if (getTargetHooks().extendPointerWithSExt())
820       return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
821     else
822       return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
823   }
824   case Builtin::BI__builtin_setjmp: {
825     // Buffer is a void**.
826     Value *Buf = EmitScalarExpr(E->getArg(0));
827 
828     // Store the frame pointer to the setjmp buffer.
829     Value *FrameAddr =
830       Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress),
831                          ConstantInt::get(Int32Ty, 0));
832     Builder.CreateStore(FrameAddr, Buf);
833 
834     // Store the stack pointer to the setjmp buffer.
835     Value *StackAddr =
836       Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
837     Value *StackSaveSlot =
838       Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2));
839     Builder.CreateStore(StackAddr, StackSaveSlot);
840 
841     // Call LLVM's EH setjmp, which is lightweight.
842     Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
843     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
844     return RValue::get(Builder.CreateCall(F, Buf));
845   }
846   case Builtin::BI__builtin_longjmp: {
847     Value *Buf = EmitScalarExpr(E->getArg(0));
848     Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
849 
850     // Call LLVM's EH longjmp, which is lightweight.
851     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
852 
853     // longjmp doesn't return; mark this as unreachable.
854     Builder.CreateUnreachable();
855 
856     // We do need to preserve an insertion point.
857     EmitBlock(createBasicBlock("longjmp.cont"));
858 
859     return RValue::get(nullptr);
860   }
861   case Builtin::BI__sync_fetch_and_add:
862   case Builtin::BI__sync_fetch_and_sub:
863   case Builtin::BI__sync_fetch_and_or:
864   case Builtin::BI__sync_fetch_and_and:
865   case Builtin::BI__sync_fetch_and_xor:
866   case Builtin::BI__sync_add_and_fetch:
867   case Builtin::BI__sync_sub_and_fetch:
868   case Builtin::BI__sync_and_and_fetch:
869   case Builtin::BI__sync_or_and_fetch:
870   case Builtin::BI__sync_xor_and_fetch:
871   case Builtin::BI__sync_val_compare_and_swap:
872   case Builtin::BI__sync_bool_compare_and_swap:
873   case Builtin::BI__sync_lock_test_and_set:
874   case Builtin::BI__sync_lock_release:
875   case Builtin::BI__sync_swap:
876     llvm_unreachable("Shouldn't make it through sema");
877   case Builtin::BI__sync_fetch_and_add_1:
878   case Builtin::BI__sync_fetch_and_add_2:
879   case Builtin::BI__sync_fetch_and_add_4:
880   case Builtin::BI__sync_fetch_and_add_8:
881   case Builtin::BI__sync_fetch_and_add_16:
882     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
883   case Builtin::BI__sync_fetch_and_sub_1:
884   case Builtin::BI__sync_fetch_and_sub_2:
885   case Builtin::BI__sync_fetch_and_sub_4:
886   case Builtin::BI__sync_fetch_and_sub_8:
887   case Builtin::BI__sync_fetch_and_sub_16:
888     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
889   case Builtin::BI__sync_fetch_and_or_1:
890   case Builtin::BI__sync_fetch_and_or_2:
891   case Builtin::BI__sync_fetch_and_or_4:
892   case Builtin::BI__sync_fetch_and_or_8:
893   case Builtin::BI__sync_fetch_and_or_16:
894     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
895   case Builtin::BI__sync_fetch_and_and_1:
896   case Builtin::BI__sync_fetch_and_and_2:
897   case Builtin::BI__sync_fetch_and_and_4:
898   case Builtin::BI__sync_fetch_and_and_8:
899   case Builtin::BI__sync_fetch_and_and_16:
900     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
901   case Builtin::BI__sync_fetch_and_xor_1:
902   case Builtin::BI__sync_fetch_and_xor_2:
903   case Builtin::BI__sync_fetch_and_xor_4:
904   case Builtin::BI__sync_fetch_and_xor_8:
905   case Builtin::BI__sync_fetch_and_xor_16:
906     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
907 
908   // Clang extensions: not overloaded yet.
909   case Builtin::BI__sync_fetch_and_min:
910     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
911   case Builtin::BI__sync_fetch_and_max:
912     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
913   case Builtin::BI__sync_fetch_and_umin:
914     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
915   case Builtin::BI__sync_fetch_and_umax:
916     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
917 
918   case Builtin::BI__sync_add_and_fetch_1:
919   case Builtin::BI__sync_add_and_fetch_2:
920   case Builtin::BI__sync_add_and_fetch_4:
921   case Builtin::BI__sync_add_and_fetch_8:
922   case Builtin::BI__sync_add_and_fetch_16:
923     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
924                                 llvm::Instruction::Add);
925   case Builtin::BI__sync_sub_and_fetch_1:
926   case Builtin::BI__sync_sub_and_fetch_2:
927   case Builtin::BI__sync_sub_and_fetch_4:
928   case Builtin::BI__sync_sub_and_fetch_8:
929   case Builtin::BI__sync_sub_and_fetch_16:
930     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
931                                 llvm::Instruction::Sub);
932   case Builtin::BI__sync_and_and_fetch_1:
933   case Builtin::BI__sync_and_and_fetch_2:
934   case Builtin::BI__sync_and_and_fetch_4:
935   case Builtin::BI__sync_and_and_fetch_8:
936   case Builtin::BI__sync_and_and_fetch_16:
937     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
938                                 llvm::Instruction::And);
939   case Builtin::BI__sync_or_and_fetch_1:
940   case Builtin::BI__sync_or_and_fetch_2:
941   case Builtin::BI__sync_or_and_fetch_4:
942   case Builtin::BI__sync_or_and_fetch_8:
943   case Builtin::BI__sync_or_and_fetch_16:
944     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
945                                 llvm::Instruction::Or);
946   case Builtin::BI__sync_xor_and_fetch_1:
947   case Builtin::BI__sync_xor_and_fetch_2:
948   case Builtin::BI__sync_xor_and_fetch_4:
949   case Builtin::BI__sync_xor_and_fetch_8:
950   case Builtin::BI__sync_xor_and_fetch_16:
951     return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
952                                 llvm::Instruction::Xor);
953 
954   case Builtin::BI__sync_val_compare_and_swap_1:
955   case Builtin::BI__sync_val_compare_and_swap_2:
956   case Builtin::BI__sync_val_compare_and_swap_4:
957   case Builtin::BI__sync_val_compare_and_swap_8:
958   case Builtin::BI__sync_val_compare_and_swap_16: {
959     QualType T = E->getType();
960     llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
961     unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
962 
963     llvm::IntegerType *IntType =
964       llvm::IntegerType::get(getLLVMContext(),
965                              getContext().getTypeSize(T));
966     llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
967 
968     Value *Args[3];
969     Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
970     Args[1] = EmitScalarExpr(E->getArg(1));
971     llvm::Type *ValueType = Args[1]->getType();
972     Args[1] = EmitToInt(*this, Args[1], T, IntType);
973     Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
974 
975     Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
976                                                 llvm::SequentiallyConsistent,
977                                                 llvm::SequentiallyConsistent);
978     Result = Builder.CreateExtractValue(Result, 0);
979     Result = EmitFromInt(*this, Result, T, ValueType);
980     return RValue::get(Result);
981   }
982 
983   case Builtin::BI__sync_bool_compare_and_swap_1:
984   case Builtin::BI__sync_bool_compare_and_swap_2:
985   case Builtin::BI__sync_bool_compare_and_swap_4:
986   case Builtin::BI__sync_bool_compare_and_swap_8:
987   case Builtin::BI__sync_bool_compare_and_swap_16: {
988     QualType T = E->getArg(1)->getType();
989     llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0));
990     unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
991 
992     llvm::IntegerType *IntType =
993       llvm::IntegerType::get(getLLVMContext(),
994                              getContext().getTypeSize(T));
995     llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
996 
997     Value *Args[3];
998     Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType);
999     Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType);
1000     Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType);
1001 
1002     Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2],
1003                                               llvm::SequentiallyConsistent,
1004                                               llvm::SequentiallyConsistent);
1005     Value *Result = Builder.CreateExtractValue(Pair, 1);
1006     // zext bool to int.
1007     Result = Builder.CreateZExt(Result, ConvertType(E->getType()));
1008     return RValue::get(Result);
1009   }
1010 
1011   case Builtin::BI__sync_swap_1:
1012   case Builtin::BI__sync_swap_2:
1013   case Builtin::BI__sync_swap_4:
1014   case Builtin::BI__sync_swap_8:
1015   case Builtin::BI__sync_swap_16:
1016     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1017 
1018   case Builtin::BI__sync_lock_test_and_set_1:
1019   case Builtin::BI__sync_lock_test_and_set_2:
1020   case Builtin::BI__sync_lock_test_and_set_4:
1021   case Builtin::BI__sync_lock_test_and_set_8:
1022   case Builtin::BI__sync_lock_test_and_set_16:
1023     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1024 
1025   case Builtin::BI__sync_lock_release_1:
1026   case Builtin::BI__sync_lock_release_2:
1027   case Builtin::BI__sync_lock_release_4:
1028   case Builtin::BI__sync_lock_release_8:
1029   case Builtin::BI__sync_lock_release_16: {
1030     Value *Ptr = EmitScalarExpr(E->getArg(0));
1031     QualType ElTy = E->getArg(0)->getType()->getPointeeType();
1032     CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
1033     llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
1034                                              StoreSize.getQuantity() * 8);
1035     Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
1036     llvm::StoreInst *Store =
1037       Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr);
1038     Store->setAlignment(StoreSize.getQuantity());
1039     Store->setAtomic(llvm::Release);
1040     return RValue::get(nullptr);
1041   }
1042 
1043   case Builtin::BI__sync_synchronize: {
1044     // We assume this is supposed to correspond to a C++0x-style
1045     // sequentially-consistent fence (i.e. this is only usable for
1046     // synchonization, not device I/O or anything like that). This intrinsic
1047     // is really badly designed in the sense that in theory, there isn't
1048     // any way to safely use it... but in practice, it mostly works
1049     // to use it with non-atomic loads and stores to get acquire/release
1050     // semantics.
1051     Builder.CreateFence(llvm::SequentiallyConsistent);
1052     return RValue::get(nullptr);
1053   }
1054 
1055   case Builtin::BI__c11_atomic_is_lock_free:
1056   case Builtin::BI__atomic_is_lock_free: {
1057     // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
1058     // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
1059     // _Atomic(T) is always properly-aligned.
1060     const char *LibCallName = "__atomic_is_lock_free";
1061     CallArgList Args;
1062     Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
1063              getContext().getSizeType());
1064     if (BuiltinID == Builtin::BI__atomic_is_lock_free)
1065       Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
1066                getContext().VoidPtrTy);
1067     else
1068       Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
1069                getContext().VoidPtrTy);
1070     const CGFunctionInfo &FuncInfo =
1071         CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args,
1072                                                FunctionType::ExtInfo(),
1073                                                RequiredArgs::All);
1074     llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
1075     llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
1076     return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args);
1077   }
1078 
1079   case Builtin::BI__atomic_test_and_set: {
1080     // Look at the argument type to determine whether this is a volatile
1081     // operation. The parameter type is always volatile.
1082     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1083     bool Volatile =
1084         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1085 
1086     Value *Ptr = EmitScalarExpr(E->getArg(0));
1087     unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1088     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1089     Value *NewVal = Builder.getInt8(1);
1090     Value *Order = EmitScalarExpr(E->getArg(1));
1091     if (isa<llvm::ConstantInt>(Order)) {
1092       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1093       AtomicRMWInst *Result = nullptr;
1094       switch (ord) {
1095       case 0:  // memory_order_relaxed
1096       default: // invalid order
1097         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1098                                          Ptr, NewVal,
1099                                          llvm::Monotonic);
1100         break;
1101       case 1:  // memory_order_consume
1102       case 2:  // memory_order_acquire
1103         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1104                                          Ptr, NewVal,
1105                                          llvm::Acquire);
1106         break;
1107       case 3:  // memory_order_release
1108         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1109                                          Ptr, NewVal,
1110                                          llvm::Release);
1111         break;
1112       case 4:  // memory_order_acq_rel
1113         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1114                                          Ptr, NewVal,
1115                                          llvm::AcquireRelease);
1116         break;
1117       case 5:  // memory_order_seq_cst
1118         Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1119                                          Ptr, NewVal,
1120                                          llvm::SequentiallyConsistent);
1121         break;
1122       }
1123       Result->setVolatile(Volatile);
1124       return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1125     }
1126 
1127     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1128 
1129     llvm::BasicBlock *BBs[5] = {
1130       createBasicBlock("monotonic", CurFn),
1131       createBasicBlock("acquire", CurFn),
1132       createBasicBlock("release", CurFn),
1133       createBasicBlock("acqrel", CurFn),
1134       createBasicBlock("seqcst", CurFn)
1135     };
1136     llvm::AtomicOrdering Orders[5] = {
1137       llvm::Monotonic, llvm::Acquire, llvm::Release,
1138       llvm::AcquireRelease, llvm::SequentiallyConsistent
1139     };
1140 
1141     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1142     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1143 
1144     Builder.SetInsertPoint(ContBB);
1145     PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
1146 
1147     for (unsigned i = 0; i < 5; ++i) {
1148       Builder.SetInsertPoint(BBs[i]);
1149       AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
1150                                                    Ptr, NewVal, Orders[i]);
1151       RMW->setVolatile(Volatile);
1152       Result->addIncoming(RMW, BBs[i]);
1153       Builder.CreateBr(ContBB);
1154     }
1155 
1156     SI->addCase(Builder.getInt32(0), BBs[0]);
1157     SI->addCase(Builder.getInt32(1), BBs[1]);
1158     SI->addCase(Builder.getInt32(2), BBs[1]);
1159     SI->addCase(Builder.getInt32(3), BBs[2]);
1160     SI->addCase(Builder.getInt32(4), BBs[3]);
1161     SI->addCase(Builder.getInt32(5), BBs[4]);
1162 
1163     Builder.SetInsertPoint(ContBB);
1164     return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
1165   }
1166 
1167   case Builtin::BI__atomic_clear: {
1168     QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
1169     bool Volatile =
1170         PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
1171 
1172     Value *Ptr = EmitScalarExpr(E->getArg(0));
1173     unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
1174     Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
1175     Value *NewVal = Builder.getInt8(0);
1176     Value *Order = EmitScalarExpr(E->getArg(1));
1177     if (isa<llvm::ConstantInt>(Order)) {
1178       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1179       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1180       Store->setAlignment(1);
1181       switch (ord) {
1182       case 0:  // memory_order_relaxed
1183       default: // invalid order
1184         Store->setOrdering(llvm::Monotonic);
1185         break;
1186       case 3:  // memory_order_release
1187         Store->setOrdering(llvm::Release);
1188         break;
1189       case 5:  // memory_order_seq_cst
1190         Store->setOrdering(llvm::SequentiallyConsistent);
1191         break;
1192       }
1193       return RValue::get(nullptr);
1194     }
1195 
1196     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1197 
1198     llvm::BasicBlock *BBs[3] = {
1199       createBasicBlock("monotonic", CurFn),
1200       createBasicBlock("release", CurFn),
1201       createBasicBlock("seqcst", CurFn)
1202     };
1203     llvm::AtomicOrdering Orders[3] = {
1204       llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent
1205     };
1206 
1207     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1208     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
1209 
1210     for (unsigned i = 0; i < 3; ++i) {
1211       Builder.SetInsertPoint(BBs[i]);
1212       StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
1213       Store->setAlignment(1);
1214       Store->setOrdering(Orders[i]);
1215       Builder.CreateBr(ContBB);
1216     }
1217 
1218     SI->addCase(Builder.getInt32(0), BBs[0]);
1219     SI->addCase(Builder.getInt32(3), BBs[1]);
1220     SI->addCase(Builder.getInt32(5), BBs[2]);
1221 
1222     Builder.SetInsertPoint(ContBB);
1223     return RValue::get(nullptr);
1224   }
1225 
1226   case Builtin::BI__atomic_thread_fence:
1227   case Builtin::BI__atomic_signal_fence:
1228   case Builtin::BI__c11_atomic_thread_fence:
1229   case Builtin::BI__c11_atomic_signal_fence: {
1230     llvm::SynchronizationScope Scope;
1231     if (BuiltinID == Builtin::BI__atomic_signal_fence ||
1232         BuiltinID == Builtin::BI__c11_atomic_signal_fence)
1233       Scope = llvm::SingleThread;
1234     else
1235       Scope = llvm::CrossThread;
1236     Value *Order = EmitScalarExpr(E->getArg(0));
1237     if (isa<llvm::ConstantInt>(Order)) {
1238       int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1239       switch (ord) {
1240       case 0:  // memory_order_relaxed
1241       default: // invalid order
1242         break;
1243       case 1:  // memory_order_consume
1244       case 2:  // memory_order_acquire
1245         Builder.CreateFence(llvm::Acquire, Scope);
1246         break;
1247       case 3:  // memory_order_release
1248         Builder.CreateFence(llvm::Release, Scope);
1249         break;
1250       case 4:  // memory_order_acq_rel
1251         Builder.CreateFence(llvm::AcquireRelease, Scope);
1252         break;
1253       case 5:  // memory_order_seq_cst
1254         Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1255         break;
1256       }
1257       return RValue::get(nullptr);
1258     }
1259 
1260     llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
1261     AcquireBB = createBasicBlock("acquire", CurFn);
1262     ReleaseBB = createBasicBlock("release", CurFn);
1263     AcqRelBB = createBasicBlock("acqrel", CurFn);
1264     SeqCstBB = createBasicBlock("seqcst", CurFn);
1265     llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1266 
1267     Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1268     llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
1269 
1270     Builder.SetInsertPoint(AcquireBB);
1271     Builder.CreateFence(llvm::Acquire, Scope);
1272     Builder.CreateBr(ContBB);
1273     SI->addCase(Builder.getInt32(1), AcquireBB);
1274     SI->addCase(Builder.getInt32(2), AcquireBB);
1275 
1276     Builder.SetInsertPoint(ReleaseBB);
1277     Builder.CreateFence(llvm::Release, Scope);
1278     Builder.CreateBr(ContBB);
1279     SI->addCase(Builder.getInt32(3), ReleaseBB);
1280 
1281     Builder.SetInsertPoint(AcqRelBB);
1282     Builder.CreateFence(llvm::AcquireRelease, Scope);
1283     Builder.CreateBr(ContBB);
1284     SI->addCase(Builder.getInt32(4), AcqRelBB);
1285 
1286     Builder.SetInsertPoint(SeqCstBB);
1287     Builder.CreateFence(llvm::SequentiallyConsistent, Scope);
1288     Builder.CreateBr(ContBB);
1289     SI->addCase(Builder.getInt32(5), SeqCstBB);
1290 
1291     Builder.SetInsertPoint(ContBB);
1292     return RValue::get(nullptr);
1293   }
1294 
1295     // Library functions with special handling.
1296   case Builtin::BIsqrt:
1297   case Builtin::BIsqrtf:
1298   case Builtin::BIsqrtl: {
1299     // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only
1300     // in finite- or unsafe-math mode (the intrinsic has different semantics
1301     // for handling negative numbers compared to the library function, so
1302     // -fmath-errno=0 is not enough).
1303     if (!FD->hasAttr<ConstAttr>())
1304       break;
1305     if (!(CGM.getCodeGenOpts().UnsafeFPMath ||
1306           CGM.getCodeGenOpts().NoNaNsFPMath))
1307       break;
1308     Value *Arg0 = EmitScalarExpr(E->getArg(0));
1309     llvm::Type *ArgType = Arg0->getType();
1310     Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType);
1311     return RValue::get(Builder.CreateCall(F, Arg0));
1312   }
1313 
1314   case Builtin::BIpow:
1315   case Builtin::BIpowf:
1316   case Builtin::BIpowl: {
1317     // Transform a call to pow* into a @llvm.pow.* intrinsic call.
1318     if (!FD->hasAttr<ConstAttr>())
1319       break;
1320     Value *Base = EmitScalarExpr(E->getArg(0));
1321     Value *Exponent = EmitScalarExpr(E->getArg(1));
1322     llvm::Type *ArgType = Base->getType();
1323     Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType);
1324     return RValue::get(Builder.CreateCall2(F, Base, Exponent));
1325   }
1326 
1327   case Builtin::BIfma:
1328   case Builtin::BIfmaf:
1329   case Builtin::BIfmal:
1330   case Builtin::BI__builtin_fma:
1331   case Builtin::BI__builtin_fmaf:
1332   case Builtin::BI__builtin_fmal: {
1333     // Rewrite fma to intrinsic.
1334     Value *FirstArg = EmitScalarExpr(E->getArg(0));
1335     llvm::Type *ArgType = FirstArg->getType();
1336     Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType);
1337     return RValue::get(Builder.CreateCall3(F, FirstArg,
1338                                               EmitScalarExpr(E->getArg(1)),
1339                                               EmitScalarExpr(E->getArg(2))));
1340   }
1341 
1342   case Builtin::BI__builtin_signbit:
1343   case Builtin::BI__builtin_signbitf:
1344   case Builtin::BI__builtin_signbitl: {
1345     LLVMContext &C = CGM.getLLVMContext();
1346 
1347     Value *Arg = EmitScalarExpr(E->getArg(0));
1348     llvm::Type *ArgTy = Arg->getType();
1349     if (ArgTy->isPPC_FP128Ty())
1350       break; // FIXME: I'm not sure what the right implementation is here.
1351     int ArgWidth = ArgTy->getPrimitiveSizeInBits();
1352     llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth);
1353     Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy);
1354     Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy);
1355     Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp);
1356     return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType())));
1357   }
1358   case Builtin::BI__builtin_annotation: {
1359     llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
1360     llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
1361                                       AnnVal->getType());
1362 
1363     // Get the annotation string, go through casts. Sema requires this to be a
1364     // non-wide string literal, potentially casted, so the cast<> is safe.
1365     const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
1366     StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
1367     return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc()));
1368   }
1369   case Builtin::BI__builtin_addcb:
1370   case Builtin::BI__builtin_addcs:
1371   case Builtin::BI__builtin_addc:
1372   case Builtin::BI__builtin_addcl:
1373   case Builtin::BI__builtin_addcll:
1374   case Builtin::BI__builtin_subcb:
1375   case Builtin::BI__builtin_subcs:
1376   case Builtin::BI__builtin_subc:
1377   case Builtin::BI__builtin_subcl:
1378   case Builtin::BI__builtin_subcll: {
1379 
1380     // We translate all of these builtins from expressions of the form:
1381     //   int x = ..., y = ..., carryin = ..., carryout, result;
1382     //   result = __builtin_addc(x, y, carryin, &carryout);
1383     //
1384     // to LLVM IR of the form:
1385     //
1386     //   %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
1387     //   %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
1388     //   %carry1 = extractvalue {i32, i1} %tmp1, 1
1389     //   %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
1390     //                                                       i32 %carryin)
1391     //   %result = extractvalue {i32, i1} %tmp2, 0
1392     //   %carry2 = extractvalue {i32, i1} %tmp2, 1
1393     //   %tmp3 = or i1 %carry1, %carry2
1394     //   %tmp4 = zext i1 %tmp3 to i32
1395     //   store i32 %tmp4, i32* %carryout
1396 
1397     // Scalarize our inputs.
1398     llvm::Value *X = EmitScalarExpr(E->getArg(0));
1399     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1400     llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
1401     std::pair<llvm::Value*, unsigned> CarryOutPtr =
1402       EmitPointerWithAlignment(E->getArg(3));
1403 
1404     // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
1405     llvm::Intrinsic::ID IntrinsicId;
1406     switch (BuiltinID) {
1407     default: llvm_unreachable("Unknown multiprecision builtin id.");
1408     case Builtin::BI__builtin_addcb:
1409     case Builtin::BI__builtin_addcs:
1410     case Builtin::BI__builtin_addc:
1411     case Builtin::BI__builtin_addcl:
1412     case Builtin::BI__builtin_addcll:
1413       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1414       break;
1415     case Builtin::BI__builtin_subcb:
1416     case Builtin::BI__builtin_subcs:
1417     case Builtin::BI__builtin_subc:
1418     case Builtin::BI__builtin_subcl:
1419     case Builtin::BI__builtin_subcll:
1420       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1421       break;
1422     }
1423 
1424     // Construct our resulting LLVM IR expression.
1425     llvm::Value *Carry1;
1426     llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
1427                                               X, Y, Carry1);
1428     llvm::Value *Carry2;
1429     llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
1430                                               Sum1, Carryin, Carry2);
1431     llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
1432                                                X->getType());
1433     llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut,
1434                                                          CarryOutPtr.first);
1435     CarryOutStore->setAlignment(CarryOutPtr.second);
1436     return RValue::get(Sum2);
1437   }
1438   case Builtin::BI__builtin_uadd_overflow:
1439   case Builtin::BI__builtin_uaddl_overflow:
1440   case Builtin::BI__builtin_uaddll_overflow:
1441   case Builtin::BI__builtin_usub_overflow:
1442   case Builtin::BI__builtin_usubl_overflow:
1443   case Builtin::BI__builtin_usubll_overflow:
1444   case Builtin::BI__builtin_umul_overflow:
1445   case Builtin::BI__builtin_umull_overflow:
1446   case Builtin::BI__builtin_umulll_overflow:
1447   case Builtin::BI__builtin_sadd_overflow:
1448   case Builtin::BI__builtin_saddl_overflow:
1449   case Builtin::BI__builtin_saddll_overflow:
1450   case Builtin::BI__builtin_ssub_overflow:
1451   case Builtin::BI__builtin_ssubl_overflow:
1452   case Builtin::BI__builtin_ssubll_overflow:
1453   case Builtin::BI__builtin_smul_overflow:
1454   case Builtin::BI__builtin_smull_overflow:
1455   case Builtin::BI__builtin_smulll_overflow: {
1456 
1457     // We translate all of these builtins directly to the relevant llvm IR node.
1458 
1459     // Scalarize our inputs.
1460     llvm::Value *X = EmitScalarExpr(E->getArg(0));
1461     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
1462     std::pair<llvm::Value *, unsigned> SumOutPtr =
1463       EmitPointerWithAlignment(E->getArg(2));
1464 
1465     // Decide which of the overflow intrinsics we are lowering to:
1466     llvm::Intrinsic::ID IntrinsicId;
1467     switch (BuiltinID) {
1468     default: llvm_unreachable("Unknown security overflow builtin id.");
1469     case Builtin::BI__builtin_uadd_overflow:
1470     case Builtin::BI__builtin_uaddl_overflow:
1471     case Builtin::BI__builtin_uaddll_overflow:
1472       IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
1473       break;
1474     case Builtin::BI__builtin_usub_overflow:
1475     case Builtin::BI__builtin_usubl_overflow:
1476     case Builtin::BI__builtin_usubll_overflow:
1477       IntrinsicId = llvm::Intrinsic::usub_with_overflow;
1478       break;
1479     case Builtin::BI__builtin_umul_overflow:
1480     case Builtin::BI__builtin_umull_overflow:
1481     case Builtin::BI__builtin_umulll_overflow:
1482       IntrinsicId = llvm::Intrinsic::umul_with_overflow;
1483       break;
1484     case Builtin::BI__builtin_sadd_overflow:
1485     case Builtin::BI__builtin_saddl_overflow:
1486     case Builtin::BI__builtin_saddll_overflow:
1487       IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
1488       break;
1489     case Builtin::BI__builtin_ssub_overflow:
1490     case Builtin::BI__builtin_ssubl_overflow:
1491     case Builtin::BI__builtin_ssubll_overflow:
1492       IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
1493       break;
1494     case Builtin::BI__builtin_smul_overflow:
1495     case Builtin::BI__builtin_smull_overflow:
1496     case Builtin::BI__builtin_smulll_overflow:
1497       IntrinsicId = llvm::Intrinsic::smul_with_overflow;
1498       break;
1499     }
1500 
1501 
1502     llvm::Value *Carry;
1503     llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
1504     llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first);
1505     SumOutStore->setAlignment(SumOutPtr.second);
1506 
1507     return RValue::get(Carry);
1508   }
1509   case Builtin::BI__builtin_addressof:
1510     return RValue::get(EmitLValue(E->getArg(0)).getAddress());
1511   case Builtin::BI__builtin_operator_new:
1512     return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(),
1513                                     E->getArg(0), false);
1514   case Builtin::BI__builtin_operator_delete:
1515     return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(),
1516                                     E->getArg(0), true);
1517   case Builtin::BI__noop:
1518     return RValue::get(nullptr);
1519   case Builtin::BI_InterlockedExchange:
1520   case Builtin::BI_InterlockedExchangePointer:
1521     return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
1522   case Builtin::BI_InterlockedCompareExchangePointer: {
1523     llvm::Type *RTy;
1524     llvm::IntegerType *IntType =
1525       IntegerType::get(getLLVMContext(),
1526                        getContext().getTypeSize(E->getType()));
1527     llvm::Type *IntPtrType = IntType->getPointerTo();
1528 
1529     llvm::Value *Destination =
1530       Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
1531 
1532     llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
1533     RTy = Exchange->getType();
1534     Exchange = Builder.CreatePtrToInt(Exchange, IntType);
1535 
1536     llvm::Value *Comparand =
1537       Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
1538 
1539     auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
1540                                               SequentiallyConsistent,
1541                                               SequentiallyConsistent);
1542     Result->setVolatile(true);
1543 
1544     return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
1545                                                                          0),
1546                                               RTy));
1547   }
1548   case Builtin::BI_InterlockedCompareExchange: {
1549     AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg(
1550         EmitScalarExpr(E->getArg(0)),
1551         EmitScalarExpr(E->getArg(2)),
1552         EmitScalarExpr(E->getArg(1)),
1553         SequentiallyConsistent,
1554         SequentiallyConsistent);
1555       CXI->setVolatile(true);
1556       return RValue::get(Builder.CreateExtractValue(CXI, 0));
1557   }
1558   case Builtin::BI_InterlockedIncrement: {
1559     AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1560       AtomicRMWInst::Add,
1561       EmitScalarExpr(E->getArg(0)),
1562       ConstantInt::get(Int32Ty, 1),
1563       llvm::SequentiallyConsistent);
1564     RMWI->setVolatile(true);
1565     return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1)));
1566   }
1567   case Builtin::BI_InterlockedDecrement: {
1568     AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1569       AtomicRMWInst::Sub,
1570       EmitScalarExpr(E->getArg(0)),
1571       ConstantInt::get(Int32Ty, 1),
1572       llvm::SequentiallyConsistent);
1573     RMWI->setVolatile(true);
1574     return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1)));
1575   }
1576   case Builtin::BI_InterlockedExchangeAdd: {
1577     AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
1578       AtomicRMWInst::Add,
1579       EmitScalarExpr(E->getArg(0)),
1580       EmitScalarExpr(E->getArg(1)),
1581       llvm::SequentiallyConsistent);
1582     RMWI->setVolatile(true);
1583     return RValue::get(RMWI);
1584   }
1585   }
1586 
1587   // If this is an alias for a lib function (e.g. __builtin_sin), emit
1588   // the call using the normal call path, but using the unmangled
1589   // version of the function name.
1590   if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
1591     return emitLibraryCall(*this, FD, E,
1592                            CGM.getBuiltinLibFunction(FD, BuiltinID));
1593 
1594   // If this is a predefined lib function (e.g. malloc), emit the call
1595   // using exactly the normal call path.
1596   if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
1597     return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee()));
1598 
1599   // See if we have a target specific intrinsic.
1600   const char *Name = getContext().BuiltinInfo.GetName(BuiltinID);
1601   Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
1602   if (const char *Prefix =
1603           llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) {
1604     IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name);
1605     // NOTE we dont need to perform a compatibility flag check here since the
1606     // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
1607     // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
1608     if (IntrinsicID == Intrinsic::not_intrinsic)
1609       IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name);
1610   }
1611 
1612   if (IntrinsicID != Intrinsic::not_intrinsic) {
1613     SmallVector<Value*, 16> Args;
1614 
1615     // Find out if any arguments are required to be integer constant
1616     // expressions.
1617     unsigned ICEArguments = 0;
1618     ASTContext::GetBuiltinTypeError Error;
1619     getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
1620     assert(Error == ASTContext::GE_None && "Should not codegen an error");
1621 
1622     Function *F = CGM.getIntrinsic(IntrinsicID);
1623     llvm::FunctionType *FTy = F->getFunctionType();
1624 
1625     for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
1626       Value *ArgValue;
1627       // If this is a normal argument, just emit it as a scalar.
1628       if ((ICEArguments & (1 << i)) == 0) {
1629         ArgValue = EmitScalarExpr(E->getArg(i));
1630       } else {
1631         // If this is required to be a constant, constant fold it so that we
1632         // know that the generated intrinsic gets a ConstantInt.
1633         llvm::APSInt Result;
1634         bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext());
1635         assert(IsConst && "Constant arg isn't actually constant?");
1636         (void)IsConst;
1637         ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result);
1638       }
1639 
1640       // If the intrinsic arg type is different from the builtin arg type
1641       // we need to do a bit cast.
1642       llvm::Type *PTy = FTy->getParamType(i);
1643       if (PTy != ArgValue->getType()) {
1644         assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
1645                "Must be able to losslessly bit cast to param");
1646         ArgValue = Builder.CreateBitCast(ArgValue, PTy);
1647       }
1648 
1649       Args.push_back(ArgValue);
1650     }
1651 
1652     Value *V = Builder.CreateCall(F, Args);
1653     QualType BuiltinRetType = E->getType();
1654 
1655     llvm::Type *RetTy = VoidTy;
1656     if (!BuiltinRetType->isVoidType())
1657       RetTy = ConvertType(BuiltinRetType);
1658 
1659     if (RetTy != V->getType()) {
1660       assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
1661              "Must be able to losslessly bit cast result type");
1662       V = Builder.CreateBitCast(V, RetTy);
1663     }
1664 
1665     return RValue::get(V);
1666   }
1667 
1668   // See if we have a target specific builtin that needs to be lowered.
1669   if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E))
1670     return RValue::get(V);
1671 
1672   ErrorUnsupported(E, "builtin function");
1673 
1674   // Unknown builtin, for now just dump it out and return undef.
1675   return GetUndefRValue(E->getType());
1676 }
1677 
EmitTargetBuiltinExpr(unsigned BuiltinID,const CallExpr * E)1678 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
1679                                               const CallExpr *E) {
1680   switch (getTarget().getTriple().getArch()) {
1681   case llvm::Triple::arm:
1682   case llvm::Triple::armeb:
1683   case llvm::Triple::thumb:
1684   case llvm::Triple::thumbeb:
1685     return EmitARMBuiltinExpr(BuiltinID, E);
1686   case llvm::Triple::aarch64:
1687   case llvm::Triple::aarch64_be:
1688   case llvm::Triple::arm64:
1689   case llvm::Triple::arm64_be:
1690     return EmitAArch64BuiltinExpr(BuiltinID, E);
1691   case llvm::Triple::x86:
1692   case llvm::Triple::x86_64:
1693     return EmitX86BuiltinExpr(BuiltinID, E);
1694   case llvm::Triple::ppc:
1695   case llvm::Triple::ppc64:
1696   case llvm::Triple::ppc64le:
1697     return EmitPPCBuiltinExpr(BuiltinID, E);
1698   case llvm::Triple::r600:
1699     return EmitR600BuiltinExpr(BuiltinID, E);
1700   default:
1701     return nullptr;
1702   }
1703 }
1704 
GetNeonType(CodeGenFunction * CGF,NeonTypeFlags TypeFlags,bool V1Ty=false)1705 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF,
1706                                      NeonTypeFlags TypeFlags,
1707                                      bool V1Ty=false) {
1708   int IsQuad = TypeFlags.isQuad();
1709   switch (TypeFlags.getEltType()) {
1710   case NeonTypeFlags::Int8:
1711   case NeonTypeFlags::Poly8:
1712     return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
1713   case NeonTypeFlags::Int16:
1714   case NeonTypeFlags::Poly16:
1715   case NeonTypeFlags::Float16:
1716     return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
1717   case NeonTypeFlags::Int32:
1718     return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
1719   case NeonTypeFlags::Int64:
1720   case NeonTypeFlags::Poly64:
1721     return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
1722   case NeonTypeFlags::Poly128:
1723     // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
1724     // There is a lot of i128 and f128 API missing.
1725     // so we use v16i8 to represent poly128 and get pattern matched.
1726     return llvm::VectorType::get(CGF->Int8Ty, 16);
1727   case NeonTypeFlags::Float32:
1728     return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
1729   case NeonTypeFlags::Float64:
1730     return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
1731   }
1732   llvm_unreachable("Unknown vector element type!");
1733 }
1734 
EmitNeonSplat(Value * V,Constant * C)1735 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
1736   unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements();
1737   Value* SV = llvm::ConstantVector::getSplat(nElts, C);
1738   return Builder.CreateShuffleVector(V, V, SV, "lane");
1739 }
1740 
EmitNeonCall(Function * F,SmallVectorImpl<Value * > & Ops,const char * name,unsigned shift,bool rightshift)1741 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
1742                                      const char *name,
1743                                      unsigned shift, bool rightshift) {
1744   unsigned j = 0;
1745   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
1746        ai != ae; ++ai, ++j)
1747     if (shift > 0 && shift == j)
1748       Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
1749     else
1750       Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
1751 
1752   return Builder.CreateCall(F, Ops, name);
1753 }
1754 
EmitNeonShiftVector(Value * V,llvm::Type * Ty,bool neg)1755 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
1756                                             bool neg) {
1757   int SV = cast<ConstantInt>(V)->getSExtValue();
1758 
1759   llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1760   llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV);
1761   return llvm::ConstantVector::getSplat(VTy->getNumElements(), C);
1762 }
1763 
1764 // \brief Right-shift a vector by a constant.
EmitNeonRShiftImm(Value * Vec,Value * Shift,llvm::Type * Ty,bool usgn,const char * name)1765 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift,
1766                                           llvm::Type *Ty, bool usgn,
1767                                           const char *name) {
1768   llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
1769 
1770   int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
1771   int EltSize = VTy->getScalarSizeInBits();
1772 
1773   Vec = Builder.CreateBitCast(Vec, Ty);
1774 
1775   // lshr/ashr are undefined when the shift amount is equal to the vector
1776   // element size.
1777   if (ShiftAmt == EltSize) {
1778     if (usgn) {
1779       // Right-shifting an unsigned value by its size yields 0.
1780       llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0);
1781       return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero);
1782     } else {
1783       // Right-shifting a signed value by its size is equivalent
1784       // to a shift of size-1.
1785       --ShiftAmt;
1786       Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
1787     }
1788   }
1789 
1790   Shift = EmitNeonShiftVector(Shift, Ty, false);
1791   if (usgn)
1792     return Builder.CreateLShr(Vec, Shift, name);
1793   else
1794     return Builder.CreateAShr(Vec, Shift, name);
1795 }
1796 
1797 /// GetPointeeAlignment - Given an expression with a pointer type, find the
1798 /// alignment of the type referenced by the pointer.  Skip over implicit
1799 /// casts.
1800 std::pair<llvm::Value*, unsigned>
EmitPointerWithAlignment(const Expr * Addr)1801 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) {
1802   assert(Addr->getType()->isPointerType());
1803   Addr = Addr->IgnoreParens();
1804   if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) {
1805     if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) &&
1806         ICE->getSubExpr()->getType()->isPointerType()) {
1807       std::pair<llvm::Value*, unsigned> Ptr =
1808           EmitPointerWithAlignment(ICE->getSubExpr());
1809       Ptr.first = Builder.CreateBitCast(Ptr.first,
1810                                         ConvertType(Addr->getType()));
1811       return Ptr;
1812     } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) {
1813       LValue LV = EmitLValue(ICE->getSubExpr());
1814       unsigned Align = LV.getAlignment().getQuantity();
1815       if (!Align) {
1816         // FIXME: Once LValues are fixed to always set alignment,
1817         // zap this code.
1818         QualType PtTy = ICE->getSubExpr()->getType();
1819         if (!PtTy->isIncompleteType())
1820           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1821         else
1822           Align = 1;
1823       }
1824       return std::make_pair(LV.getAddress(), Align);
1825     }
1826   }
1827   if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) {
1828     if (UO->getOpcode() == UO_AddrOf) {
1829       LValue LV = EmitLValue(UO->getSubExpr());
1830       unsigned Align = LV.getAlignment().getQuantity();
1831       if (!Align) {
1832         // FIXME: Once LValues are fixed to always set alignment,
1833         // zap this code.
1834         QualType PtTy = UO->getSubExpr()->getType();
1835         if (!PtTy->isIncompleteType())
1836           Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1837         else
1838           Align = 1;
1839       }
1840       return std::make_pair(LV.getAddress(), Align);
1841     }
1842   }
1843 
1844   unsigned Align = 1;
1845   QualType PtTy = Addr->getType()->getPointeeType();
1846   if (!PtTy->isIncompleteType())
1847     Align = getContext().getTypeAlignInChars(PtTy).getQuantity();
1848 
1849   return std::make_pair(EmitScalarExpr(Addr), Align);
1850 }
1851 
1852 enum {
1853   AddRetType = (1 << 0),
1854   Add1ArgType = (1 << 1),
1855   Add2ArgTypes = (1 << 2),
1856 
1857   VectorizeRetType = (1 << 3),
1858   VectorizeArgTypes = (1 << 4),
1859 
1860   InventFloatType = (1 << 5),
1861   UnsignedAlts = (1 << 6),
1862 
1863   Use64BitVectors = (1 << 7),
1864   Use128BitVectors = (1 << 8),
1865 
1866   Vectorize1ArgType = Add1ArgType | VectorizeArgTypes,
1867   VectorRet = AddRetType | VectorizeRetType,
1868   VectorRetGetArgs01 =
1869       AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes,
1870   FpCmpzModifiers =
1871       AddRetType | VectorizeRetType | Add1ArgType | InventFloatType
1872 };
1873 
1874  struct NeonIntrinsicInfo {
1875   unsigned BuiltinID;
1876   unsigned LLVMIntrinsic;
1877   unsigned AltLLVMIntrinsic;
1878   const char *NameHint;
1879   unsigned TypeModifier;
1880 
operator <NeonIntrinsicInfo1881   bool operator<(unsigned RHSBuiltinID) const {
1882     return BuiltinID < RHSBuiltinID;
1883   }
1884 };
1885 
1886 #define NEONMAP0(NameBase) \
1887   { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 }
1888 
1889 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
1890   { NEON:: BI__builtin_neon_ ## NameBase, \
1891       Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier }
1892 
1893 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
1894   { NEON:: BI__builtin_neon_ ## NameBase, \
1895       Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
1896       #NameBase, TypeModifier }
1897 
1898 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
1899   NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
1900   NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
1901   NEONMAP1(vabs_v, arm_neon_vabs, 0),
1902   NEONMAP1(vabsq_v, arm_neon_vabs, 0),
1903   NEONMAP0(vaddhn_v),
1904   NEONMAP1(vaesdq_v, arm_neon_aesd, 0),
1905   NEONMAP1(vaeseq_v, arm_neon_aese, 0),
1906   NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0),
1907   NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0),
1908   NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
1909   NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
1910   NEONMAP1(vcage_v, arm_neon_vacge, 0),
1911   NEONMAP1(vcageq_v, arm_neon_vacge, 0),
1912   NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
1913   NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
1914   NEONMAP1(vcale_v, arm_neon_vacge, 0),
1915   NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
1916   NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
1917   NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
1918   NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
1919   NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
1920   NEONMAP1(vclz_v, ctlz, Add1ArgType),
1921   NEONMAP1(vclzq_v, ctlz, Add1ArgType),
1922   NEONMAP1(vcnt_v, ctpop, Add1ArgType),
1923   NEONMAP1(vcntq_v, ctpop, Add1ArgType),
1924   NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0),
1925   NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
1926   NEONMAP0(vcvt_f32_v),
1927   NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
1928   NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
1929   NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
1930   NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
1931   NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
1932   NEONMAP0(vcvt_s32_v),
1933   NEONMAP0(vcvt_s64_v),
1934   NEONMAP0(vcvt_u32_v),
1935   NEONMAP0(vcvt_u64_v),
1936   NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
1937   NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
1938   NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
1939   NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
1940   NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
1941   NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
1942   NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
1943   NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
1944   NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
1945   NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
1946   NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
1947   NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
1948   NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
1949   NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
1950   NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
1951   NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
1952   NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
1953   NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
1954   NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
1955   NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
1956   NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
1957   NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
1958   NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
1959   NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
1960   NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
1961   NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
1962   NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
1963   NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
1964   NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
1965   NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
1966   NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
1967   NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
1968   NEONMAP0(vcvtq_f32_v),
1969   NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
1970   NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
1971   NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
1972   NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
1973   NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
1974   NEONMAP0(vcvtq_s32_v),
1975   NEONMAP0(vcvtq_s64_v),
1976   NEONMAP0(vcvtq_u32_v),
1977   NEONMAP0(vcvtq_u64_v),
1978   NEONMAP0(vext_v),
1979   NEONMAP0(vextq_v),
1980   NEONMAP0(vfma_v),
1981   NEONMAP0(vfmaq_v),
1982   NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
1983   NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
1984   NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
1985   NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
1986   NEONMAP0(vld1_dup_v),
1987   NEONMAP1(vld1_v, arm_neon_vld1, 0),
1988   NEONMAP0(vld1q_dup_v),
1989   NEONMAP1(vld1q_v, arm_neon_vld1, 0),
1990   NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
1991   NEONMAP1(vld2_v, arm_neon_vld2, 0),
1992   NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
1993   NEONMAP1(vld2q_v, arm_neon_vld2, 0),
1994   NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
1995   NEONMAP1(vld3_v, arm_neon_vld3, 0),
1996   NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
1997   NEONMAP1(vld3q_v, arm_neon_vld3, 0),
1998   NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
1999   NEONMAP1(vld4_v, arm_neon_vld4, 0),
2000   NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
2001   NEONMAP1(vld4q_v, arm_neon_vld4, 0),
2002   NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
2003   NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
2004   NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
2005   NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
2006   NEONMAP0(vmovl_v),
2007   NEONMAP0(vmovn_v),
2008   NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
2009   NEONMAP0(vmull_v),
2010   NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
2011   NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
2012   NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
2013   NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
2014   NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
2015   NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
2016   NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
2017   NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
2018   NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
2019   NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
2020   NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
2021   NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
2022   NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts),
2023   NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0),
2024   NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0),
2025   NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
2026   NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
2027   NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
2028   NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
2029   NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
2030   NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
2031   NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
2032   NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
2033   NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
2034   NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
2035   NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
2036   NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
2037   NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
2038   NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
2039   NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
2040   NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
2041   NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts),
2042   NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
2043   NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
2044   NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
2045   NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
2046   NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
2047   NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
2048   NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
2049   NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
2050   NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
2051   NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
2052   NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
2053   NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
2054   NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
2055   NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
2056   NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0),
2057   NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0),
2058   NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0),
2059   NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0),
2060   NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0),
2061   NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0),
2062   NEONMAP0(vshl_n_v),
2063   NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
2064   NEONMAP0(vshll_n_v),
2065   NEONMAP0(vshlq_n_v),
2066   NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
2067   NEONMAP0(vshr_n_v),
2068   NEONMAP0(vshrn_n_v),
2069   NEONMAP0(vshrq_n_v),
2070   NEONMAP1(vst1_v, arm_neon_vst1, 0),
2071   NEONMAP1(vst1q_v, arm_neon_vst1, 0),
2072   NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
2073   NEONMAP1(vst2_v, arm_neon_vst2, 0),
2074   NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
2075   NEONMAP1(vst2q_v, arm_neon_vst2, 0),
2076   NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
2077   NEONMAP1(vst3_v, arm_neon_vst3, 0),
2078   NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
2079   NEONMAP1(vst3q_v, arm_neon_vst3, 0),
2080   NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
2081   NEONMAP1(vst4_v, arm_neon_vst4, 0),
2082   NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
2083   NEONMAP1(vst4q_v, arm_neon_vst4, 0),
2084   NEONMAP0(vsubhn_v),
2085   NEONMAP0(vtrn_v),
2086   NEONMAP0(vtrnq_v),
2087   NEONMAP0(vtst_v),
2088   NEONMAP0(vtstq_v),
2089   NEONMAP0(vuzp_v),
2090   NEONMAP0(vuzpq_v),
2091   NEONMAP0(vzip_v),
2092   NEONMAP0(vzipq_v)
2093 };
2094 
2095 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
2096   NEONMAP1(vabs_v, aarch64_neon_abs, 0),
2097   NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
2098   NEONMAP0(vaddhn_v),
2099   NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0),
2100   NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0),
2101   NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0),
2102   NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0),
2103   NEONMAP1(vcage_v, aarch64_neon_facge, 0),
2104   NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
2105   NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
2106   NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
2107   NEONMAP1(vcale_v, aarch64_neon_facge, 0),
2108   NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
2109   NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
2110   NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
2111   NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
2112   NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
2113   NEONMAP1(vclz_v, ctlz, Add1ArgType),
2114   NEONMAP1(vclzq_v, ctlz, Add1ArgType),
2115   NEONMAP1(vcnt_v, ctpop, Add1ArgType),
2116   NEONMAP1(vcntq_v, ctpop, Add1ArgType),
2117   NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0),
2118   NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
2119   NEONMAP0(vcvt_f32_v),
2120   NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2121   NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2122   NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
2123   NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
2124   NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
2125   NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
2126   NEONMAP0(vcvtq_f32_v),
2127   NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2128   NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
2129   NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
2130   NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
2131   NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
2132   NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
2133   NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
2134   NEONMAP0(vext_v),
2135   NEONMAP0(vextq_v),
2136   NEONMAP0(vfma_v),
2137   NEONMAP0(vfmaq_v),
2138   NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
2139   NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
2140   NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
2141   NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
2142   NEONMAP0(vmovl_v),
2143   NEONMAP0(vmovn_v),
2144   NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
2145   NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
2146   NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
2147   NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
2148   NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
2149   NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
2150   NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
2151   NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
2152   NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
2153   NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
2154   NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
2155   NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
2156   NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
2157   NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
2158   NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
2159   NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
2160   NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
2161   NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
2162   NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
2163   NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
2164   NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
2165   NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
2166   NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
2167   NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
2168   NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
2169   NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
2170   NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
2171   NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
2172   NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
2173   NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
2174   NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
2175   NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
2176   NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
2177   NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
2178   NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
2179   NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
2180   NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
2181   NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
2182   NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
2183   NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
2184   NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
2185   NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
2186   NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
2187   NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0),
2188   NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0),
2189   NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0),
2190   NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0),
2191   NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0),
2192   NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0),
2193   NEONMAP0(vshl_n_v),
2194   NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
2195   NEONMAP0(vshll_n_v),
2196   NEONMAP0(vshlq_n_v),
2197   NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
2198   NEONMAP0(vshr_n_v),
2199   NEONMAP0(vshrn_n_v),
2200   NEONMAP0(vshrq_n_v),
2201   NEONMAP0(vsubhn_v),
2202   NEONMAP0(vtst_v),
2203   NEONMAP0(vtstq_v),
2204 };
2205 
2206 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = {
2207   NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
2208   NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
2209   NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
2210   NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
2211   NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
2212   NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
2213   NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
2214   NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
2215   NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
2216   NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2217   NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
2218   NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
2219   NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
2220   NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
2221   NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2222   NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2223   NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
2224   NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
2225   NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
2226   NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
2227   NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
2228   NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
2229   NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
2230   NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
2231   NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
2232   NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
2233   NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
2234   NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
2235   NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
2236   NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
2237   NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
2238   NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
2239   NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
2240   NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
2241   NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
2242   NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
2243   NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
2244   NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
2245   NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
2246   NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
2247   NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
2248   NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
2249   NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
2250   NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
2251   NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
2252   NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
2253   NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
2254   NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
2255   NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
2256   NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2257   NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2258   NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2259   NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2260   NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
2261   NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
2262   NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2263   NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2264   NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
2265   NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
2266   NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2267   NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2268   NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2269   NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2270   NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
2271   NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
2272   NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2273   NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
2274   NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
2275   NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
2276   NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
2277   NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
2278   NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
2279   NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2280   NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
2281   NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2282   NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
2283   NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2284   NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
2285   NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2286   NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
2287   NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
2288   NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
2289   NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
2290   NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
2291   NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
2292   NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
2293   NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
2294   NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
2295   NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
2296   NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
2297   NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
2298   NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
2299   NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
2300   NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
2301   NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
2302   NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
2303   NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
2304   NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
2305   NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
2306   NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
2307   NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
2308   NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
2309   NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
2310   NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
2311   NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
2312   NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
2313   NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
2314   NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
2315   NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
2316   NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
2317   NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
2318   NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
2319   NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
2320   NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
2321   NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
2322   NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
2323   NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
2324   NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
2325   NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
2326   NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
2327   NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
2328   NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
2329   NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
2330   NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
2331   NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
2332   NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
2333   NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
2334   NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
2335   NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
2336   NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
2337   NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2338   NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2339   NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2340   NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2341   NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
2342   NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
2343   NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2344   NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2345   NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
2346   NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
2347   NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
2348   NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
2349   NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
2350   NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
2351   NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
2352   NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
2353   NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
2354   NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
2355   NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
2356   NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
2357   NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
2358   NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
2359   NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
2360   NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
2361   NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
2362   NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
2363   NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
2364   NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
2365   NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
2366   NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
2367   NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
2368   NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
2369   NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
2370   NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
2371   NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
2372   NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
2373   NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
2374   NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
2375   NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
2376   NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
2377   NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
2378   NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
2379   NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
2380   NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
2381   NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
2382   NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
2383   NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
2384   NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
2385   NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
2386   NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
2387   NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
2388   NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
2389   NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
2390   NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
2391   NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
2392   NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
2393   NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
2394   NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
2395   NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
2396   NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
2397   NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
2398   NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
2399 };
2400 
2401 #undef NEONMAP0
2402 #undef NEONMAP1
2403 #undef NEONMAP2
2404 
2405 static bool NEONSIMDIntrinsicsProvenSorted = false;
2406 
2407 static bool AArch64SIMDIntrinsicsProvenSorted = false;
2408 static bool AArch64SISDIntrinsicsProvenSorted = false;
2409 
2410 
2411 static const NeonIntrinsicInfo *
findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap,unsigned BuiltinID,bool & MapProvenSorted)2412 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap,
2413                        unsigned BuiltinID, bool &MapProvenSorted) {
2414 
2415 #ifndef NDEBUG
2416   if (!MapProvenSorted) {
2417     // FIXME: use std::is_sorted once C++11 is allowed
2418     for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i)
2419       assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID);
2420     MapProvenSorted = true;
2421   }
2422 #endif
2423 
2424   const NeonIntrinsicInfo *Builtin =
2425       std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID);
2426 
2427   if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
2428     return Builtin;
2429 
2430   return nullptr;
2431 }
2432 
LookupNeonLLVMIntrinsic(unsigned IntrinsicID,unsigned Modifier,llvm::Type * ArgType,const CallExpr * E)2433 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
2434                                                    unsigned Modifier,
2435                                                    llvm::Type *ArgType,
2436                                                    const CallExpr *E) {
2437   int VectorSize = 0;
2438   if (Modifier & Use64BitVectors)
2439     VectorSize = 64;
2440   else if (Modifier & Use128BitVectors)
2441     VectorSize = 128;
2442 
2443   // Return type.
2444   SmallVector<llvm::Type *, 3> Tys;
2445   if (Modifier & AddRetType) {
2446     llvm::Type *Ty = ConvertType(E->getCallReturnType());
2447     if (Modifier & VectorizeRetType)
2448       Ty = llvm::VectorType::get(
2449           Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1);
2450 
2451     Tys.push_back(Ty);
2452   }
2453 
2454   // Arguments.
2455   if (Modifier & VectorizeArgTypes) {
2456     int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1;
2457     ArgType = llvm::VectorType::get(ArgType, Elts);
2458   }
2459 
2460   if (Modifier & (Add1ArgType | Add2ArgTypes))
2461     Tys.push_back(ArgType);
2462 
2463   if (Modifier & Add2ArgTypes)
2464     Tys.push_back(ArgType);
2465 
2466   if (Modifier & InventFloatType)
2467     Tys.push_back(FloatTy);
2468 
2469   return CGM.getIntrinsic(IntrinsicID, Tys);
2470 }
2471 
EmitCommonNeonSISDBuiltinExpr(CodeGenFunction & CGF,const NeonIntrinsicInfo & SISDInfo,SmallVectorImpl<Value * > & Ops,const CallExpr * E)2472 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF,
2473                                             const NeonIntrinsicInfo &SISDInfo,
2474                                             SmallVectorImpl<Value *> &Ops,
2475                                             const CallExpr *E) {
2476   unsigned BuiltinID = SISDInfo.BuiltinID;
2477   unsigned int Int = SISDInfo.LLVMIntrinsic;
2478   unsigned Modifier = SISDInfo.TypeModifier;
2479   const char *s = SISDInfo.NameHint;
2480 
2481   switch (BuiltinID) {
2482   case NEON::BI__builtin_neon_vcled_s64:
2483   case NEON::BI__builtin_neon_vcled_u64:
2484   case NEON::BI__builtin_neon_vcles_f32:
2485   case NEON::BI__builtin_neon_vcled_f64:
2486   case NEON::BI__builtin_neon_vcltd_s64:
2487   case NEON::BI__builtin_neon_vcltd_u64:
2488   case NEON::BI__builtin_neon_vclts_f32:
2489   case NEON::BI__builtin_neon_vcltd_f64:
2490   case NEON::BI__builtin_neon_vcales_f32:
2491   case NEON::BI__builtin_neon_vcaled_f64:
2492   case NEON::BI__builtin_neon_vcalts_f32:
2493   case NEON::BI__builtin_neon_vcaltd_f64:
2494     // Only one direction of comparisons actually exist, cmle is actually a cmge
2495     // with swapped operands. The table gives us the right intrinsic but we
2496     // still need to do the swap.
2497     std::swap(Ops[0], Ops[1]);
2498     break;
2499   }
2500 
2501   assert(Int && "Generic code assumes a valid intrinsic");
2502 
2503   // Determine the type(s) of this overloaded AArch64 intrinsic.
2504   const Expr *Arg = E->getArg(0);
2505   llvm::Type *ArgTy = CGF.ConvertType(Arg->getType());
2506   Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
2507 
2508   int j = 0;
2509   ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
2510   for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
2511        ai != ae; ++ai, ++j) {
2512     llvm::Type *ArgTy = ai->getType();
2513     if (Ops[j]->getType()->getPrimitiveSizeInBits() ==
2514              ArgTy->getPrimitiveSizeInBits())
2515       continue;
2516 
2517     assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy());
2518     // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate
2519     // it before inserting.
2520     Ops[j] =
2521         CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType());
2522     Ops[j] =
2523         CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0);
2524   }
2525 
2526   Value *Result = CGF.EmitNeonCall(F, Ops, s);
2527   llvm::Type *ResultType = CGF.ConvertType(E->getType());
2528   if (ResultType->getPrimitiveSizeInBits() <
2529       Result->getType()->getPrimitiveSizeInBits())
2530     return CGF.Builder.CreateExtractElement(Result, C0);
2531 
2532   return CGF.Builder.CreateBitCast(Result, ResultType, s);
2533 }
2534 
EmitCommonNeonBuiltinExpr(unsigned BuiltinID,unsigned LLVMIntrinsic,unsigned AltLLVMIntrinsic,const char * NameHint,unsigned Modifier,const CallExpr * E,SmallVectorImpl<llvm::Value * > & Ops,llvm::Value * Align)2535 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
2536     unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic,
2537     const char *NameHint, unsigned Modifier, const CallExpr *E,
2538     SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) {
2539   // Get the last argument, which specifies the vector type.
2540   llvm::APSInt NeonTypeConst;
2541   const Expr *Arg = E->getArg(E->getNumArgs() - 1);
2542   if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext()))
2543     return nullptr;
2544 
2545   // Determine the type of this overloaded NEON intrinsic.
2546   NeonTypeFlags Type(NeonTypeConst.getZExtValue());
2547   bool Usgn = Type.isUnsigned();
2548   bool Quad = Type.isQuad();
2549 
2550   llvm::VectorType *VTy = GetNeonType(this, Type);
2551   llvm::Type *Ty = VTy;
2552   if (!Ty)
2553     return nullptr;
2554 
2555   unsigned Int = LLVMIntrinsic;
2556   if ((Modifier & UnsignedAlts) && !Usgn)
2557     Int = AltLLVMIntrinsic;
2558 
2559   switch (BuiltinID) {
2560   default: break;
2561   case NEON::BI__builtin_neon_vabs_v:
2562   case NEON::BI__builtin_neon_vabsq_v:
2563     if (VTy->getElementType()->isFloatingPointTy())
2564       return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs");
2565     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs");
2566   case NEON::BI__builtin_neon_vaddhn_v: {
2567     llvm::VectorType *SrcTy =
2568         llvm::VectorType::getExtendedElementVectorType(VTy);
2569 
2570     // %sum = add <4 x i32> %lhs, %rhs
2571     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
2572     Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
2573     Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
2574 
2575     // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
2576     Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
2577                                        SrcTy->getScalarSizeInBits() / 2);
2578     ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
2579     Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
2580 
2581     // %res = trunc <4 x i32> %high to <4 x i16>
2582     return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
2583   }
2584   case NEON::BI__builtin_neon_vcale_v:
2585   case NEON::BI__builtin_neon_vcaleq_v:
2586   case NEON::BI__builtin_neon_vcalt_v:
2587   case NEON::BI__builtin_neon_vcaltq_v:
2588     std::swap(Ops[0], Ops[1]);
2589   case NEON::BI__builtin_neon_vcage_v:
2590   case NEON::BI__builtin_neon_vcageq_v:
2591   case NEON::BI__builtin_neon_vcagt_v:
2592   case NEON::BI__builtin_neon_vcagtq_v: {
2593     llvm::Type *VecFlt = llvm::VectorType::get(
2594         VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy,
2595         VTy->getNumElements());
2596     llvm::Type *Tys[] = { VTy, VecFlt };
2597     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
2598     return EmitNeonCall(F, Ops, NameHint);
2599   }
2600   case NEON::BI__builtin_neon_vclz_v:
2601   case NEON::BI__builtin_neon_vclzq_v:
2602     // We generate target-independent intrinsic, which needs a second argument
2603     // for whether or not clz of zero is undefined; on ARM it isn't.
2604     Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
2605     break;
2606   case NEON::BI__builtin_neon_vcvt_f32_v:
2607   case NEON::BI__builtin_neon_vcvtq_f32_v:
2608     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2609     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad));
2610     return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
2611                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
2612   case NEON::BI__builtin_neon_vcvt_n_f32_v:
2613   case NEON::BI__builtin_neon_vcvt_n_f64_v:
2614   case NEON::BI__builtin_neon_vcvtq_n_f32_v:
2615   case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
2616     bool Double =
2617       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2618     llvm::Type *FloatTy =
2619         GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2620                                                : NeonTypeFlags::Float32,
2621                                         false, Quad));
2622     llvm::Type *Tys[2] = { FloatTy, Ty };
2623     Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
2624     Function *F = CGM.getIntrinsic(Int, Tys);
2625     return EmitNeonCall(F, Ops, "vcvt_n");
2626   }
2627   case NEON::BI__builtin_neon_vcvt_n_s32_v:
2628   case NEON::BI__builtin_neon_vcvt_n_u32_v:
2629   case NEON::BI__builtin_neon_vcvt_n_s64_v:
2630   case NEON::BI__builtin_neon_vcvt_n_u64_v:
2631   case NEON::BI__builtin_neon_vcvtq_n_s32_v:
2632   case NEON::BI__builtin_neon_vcvtq_n_u32_v:
2633   case NEON::BI__builtin_neon_vcvtq_n_s64_v:
2634   case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
2635     bool Double =
2636       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2637     llvm::Type *FloatTy =
2638         GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2639                                                : NeonTypeFlags::Float32,
2640                                         false, Quad));
2641     llvm::Type *Tys[2] = { Ty, FloatTy };
2642     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
2643     return EmitNeonCall(F, Ops, "vcvt_n");
2644   }
2645   case NEON::BI__builtin_neon_vcvt_s32_v:
2646   case NEON::BI__builtin_neon_vcvt_u32_v:
2647   case NEON::BI__builtin_neon_vcvt_s64_v:
2648   case NEON::BI__builtin_neon_vcvt_u64_v:
2649   case NEON::BI__builtin_neon_vcvtq_s32_v:
2650   case NEON::BI__builtin_neon_vcvtq_u32_v:
2651   case NEON::BI__builtin_neon_vcvtq_s64_v:
2652   case NEON::BI__builtin_neon_vcvtq_u64_v: {
2653     bool Double =
2654       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2655     llvm::Type *FloatTy =
2656         GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64
2657                                                : NeonTypeFlags::Float32,
2658                                         false, Quad));
2659     Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
2660     return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
2661                 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
2662   }
2663   case NEON::BI__builtin_neon_vcvta_s32_v:
2664   case NEON::BI__builtin_neon_vcvta_s64_v:
2665   case NEON::BI__builtin_neon_vcvta_u32_v:
2666   case NEON::BI__builtin_neon_vcvta_u64_v:
2667   case NEON::BI__builtin_neon_vcvtaq_s32_v:
2668   case NEON::BI__builtin_neon_vcvtaq_s64_v:
2669   case NEON::BI__builtin_neon_vcvtaq_u32_v:
2670   case NEON::BI__builtin_neon_vcvtaq_u64_v:
2671   case NEON::BI__builtin_neon_vcvtn_s32_v:
2672   case NEON::BI__builtin_neon_vcvtn_s64_v:
2673   case NEON::BI__builtin_neon_vcvtn_u32_v:
2674   case NEON::BI__builtin_neon_vcvtn_u64_v:
2675   case NEON::BI__builtin_neon_vcvtnq_s32_v:
2676   case NEON::BI__builtin_neon_vcvtnq_s64_v:
2677   case NEON::BI__builtin_neon_vcvtnq_u32_v:
2678   case NEON::BI__builtin_neon_vcvtnq_u64_v:
2679   case NEON::BI__builtin_neon_vcvtp_s32_v:
2680   case NEON::BI__builtin_neon_vcvtp_s64_v:
2681   case NEON::BI__builtin_neon_vcvtp_u32_v:
2682   case NEON::BI__builtin_neon_vcvtp_u64_v:
2683   case NEON::BI__builtin_neon_vcvtpq_s32_v:
2684   case NEON::BI__builtin_neon_vcvtpq_s64_v:
2685   case NEON::BI__builtin_neon_vcvtpq_u32_v:
2686   case NEON::BI__builtin_neon_vcvtpq_u64_v:
2687   case NEON::BI__builtin_neon_vcvtm_s32_v:
2688   case NEON::BI__builtin_neon_vcvtm_s64_v:
2689   case NEON::BI__builtin_neon_vcvtm_u32_v:
2690   case NEON::BI__builtin_neon_vcvtm_u64_v:
2691   case NEON::BI__builtin_neon_vcvtmq_s32_v:
2692   case NEON::BI__builtin_neon_vcvtmq_s64_v:
2693   case NEON::BI__builtin_neon_vcvtmq_u32_v:
2694   case NEON::BI__builtin_neon_vcvtmq_u64_v: {
2695     bool Double =
2696       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
2697     llvm::Type *InTy =
2698       GetNeonType(this,
2699                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
2700                                 : NeonTypeFlags::Float32, false, Quad));
2701     llvm::Type *Tys[2] = { Ty, InTy };
2702     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
2703   }
2704   case NEON::BI__builtin_neon_vext_v:
2705   case NEON::BI__builtin_neon_vextq_v: {
2706     int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
2707     SmallVector<Constant*, 16> Indices;
2708     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2709       Indices.push_back(ConstantInt::get(Int32Ty, i+CV));
2710 
2711     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2712     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2713     Value *SV = llvm::ConstantVector::get(Indices);
2714     return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext");
2715   }
2716   case NEON::BI__builtin_neon_vfma_v:
2717   case NEON::BI__builtin_neon_vfmaq_v: {
2718     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
2719     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2720     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2721     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2722 
2723     // NEON intrinsic puts accumulator first, unlike the LLVM fma.
2724     return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
2725   }
2726   case NEON::BI__builtin_neon_vld1_v:
2727   case NEON::BI__builtin_neon_vld1q_v:
2728     Ops.push_back(Align);
2729     return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1");
2730   case NEON::BI__builtin_neon_vld2_v:
2731   case NEON::BI__builtin_neon_vld2q_v:
2732   case NEON::BI__builtin_neon_vld3_v:
2733   case NEON::BI__builtin_neon_vld3q_v:
2734   case NEON::BI__builtin_neon_vld4_v:
2735   case NEON::BI__builtin_neon_vld4q_v: {
2736     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty);
2737     Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint);
2738     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2739     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2740     return Builder.CreateStore(Ops[1], Ops[0]);
2741   }
2742   case NEON::BI__builtin_neon_vld1_dup_v:
2743   case NEON::BI__builtin_neon_vld1q_dup_v: {
2744     Value *V = UndefValue::get(Ty);
2745     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
2746     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2747     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
2748     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
2749     llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
2750     Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
2751     return EmitNeonSplat(Ops[0], CI);
2752   }
2753   case NEON::BI__builtin_neon_vld2_lane_v:
2754   case NEON::BI__builtin_neon_vld2q_lane_v:
2755   case NEON::BI__builtin_neon_vld3_lane_v:
2756   case NEON::BI__builtin_neon_vld3q_lane_v:
2757   case NEON::BI__builtin_neon_vld4_lane_v:
2758   case NEON::BI__builtin_neon_vld4q_lane_v: {
2759     Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty);
2760     for (unsigned I = 2; I < Ops.size() - 1; ++I)
2761       Ops[I] = Builder.CreateBitCast(Ops[I], Ty);
2762     Ops.push_back(Align);
2763     Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint);
2764     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
2765     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2766     return Builder.CreateStore(Ops[1], Ops[0]);
2767   }
2768   case NEON::BI__builtin_neon_vmovl_v: {
2769     llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy);
2770     Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
2771     if (Usgn)
2772       return Builder.CreateZExt(Ops[0], Ty, "vmovl");
2773     return Builder.CreateSExt(Ops[0], Ty, "vmovl");
2774   }
2775   case NEON::BI__builtin_neon_vmovn_v: {
2776     llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy);
2777     Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
2778     return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
2779   }
2780   case NEON::BI__builtin_neon_vmull_v:
2781     // FIXME: the integer vmull operations could be emitted in terms of pure
2782     // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of
2783     // hoisting the exts outside loops. Until global ISel comes along that can
2784     // see through such movement this leads to bad CodeGen. So we need an
2785     // intrinsic for now.
2786     Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
2787     Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
2788     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
2789   case NEON::BI__builtin_neon_vpadal_v:
2790   case NEON::BI__builtin_neon_vpadalq_v: {
2791     // The source operand type has twice as many elements of half the size.
2792     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2793     llvm::Type *EltTy =
2794       llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2795     llvm::Type *NarrowTy =
2796       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2797     llvm::Type *Tys[2] = { Ty, NarrowTy };
2798     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
2799   }
2800   case NEON::BI__builtin_neon_vpaddl_v:
2801   case NEON::BI__builtin_neon_vpaddlq_v: {
2802     // The source operand type has twice as many elements of half the size.
2803     unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
2804     llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
2805     llvm::Type *NarrowTy =
2806       llvm::VectorType::get(EltTy, VTy->getNumElements() * 2);
2807     llvm::Type *Tys[2] = { Ty, NarrowTy };
2808     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
2809   }
2810   case NEON::BI__builtin_neon_vqdmlal_v:
2811   case NEON::BI__builtin_neon_vqdmlsl_v: {
2812     SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end());
2813     Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty),
2814                               MulOps, "vqdmlal");
2815 
2816     SmallVector<Value *, 2> AccumOps;
2817     AccumOps.push_back(Ops[0]);
2818     AccumOps.push_back(Mul);
2819     return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty),
2820                         AccumOps, NameHint);
2821   }
2822   case NEON::BI__builtin_neon_vqshl_n_v:
2823   case NEON::BI__builtin_neon_vqshlq_n_v:
2824     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
2825                         1, false);
2826   case NEON::BI__builtin_neon_vrecpe_v:
2827   case NEON::BI__builtin_neon_vrecpeq_v:
2828   case NEON::BI__builtin_neon_vrsqrte_v:
2829   case NEON::BI__builtin_neon_vrsqrteq_v:
2830     Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic;
2831     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
2832 
2833   case NEON::BI__builtin_neon_vshl_n_v:
2834   case NEON::BI__builtin_neon_vshlq_n_v:
2835     Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
2836     return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
2837                              "vshl_n");
2838   case NEON::BI__builtin_neon_vshll_n_v: {
2839     llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy);
2840     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
2841     if (Usgn)
2842       Ops[0] = Builder.CreateZExt(Ops[0], VTy);
2843     else
2844       Ops[0] = Builder.CreateSExt(Ops[0], VTy);
2845     Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false);
2846     return Builder.CreateShl(Ops[0], Ops[1], "vshll_n");
2847   }
2848   case NEON::BI__builtin_neon_vshrn_n_v: {
2849     llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy);
2850     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
2851     Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false);
2852     if (Usgn)
2853       Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]);
2854     else
2855       Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]);
2856     return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n");
2857   }
2858   case NEON::BI__builtin_neon_vshr_n_v:
2859   case NEON::BI__builtin_neon_vshrq_n_v:
2860     return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n");
2861   case NEON::BI__builtin_neon_vst1_v:
2862   case NEON::BI__builtin_neon_vst1q_v:
2863   case NEON::BI__builtin_neon_vst2_v:
2864   case NEON::BI__builtin_neon_vst2q_v:
2865   case NEON::BI__builtin_neon_vst3_v:
2866   case NEON::BI__builtin_neon_vst3q_v:
2867   case NEON::BI__builtin_neon_vst4_v:
2868   case NEON::BI__builtin_neon_vst4q_v:
2869   case NEON::BI__builtin_neon_vst2_lane_v:
2870   case NEON::BI__builtin_neon_vst2q_lane_v:
2871   case NEON::BI__builtin_neon_vst3_lane_v:
2872   case NEON::BI__builtin_neon_vst3q_lane_v:
2873   case NEON::BI__builtin_neon_vst4_lane_v:
2874   case NEON::BI__builtin_neon_vst4q_lane_v:
2875     Ops.push_back(Align);
2876     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "");
2877   case NEON::BI__builtin_neon_vsubhn_v: {
2878     llvm::VectorType *SrcTy =
2879         llvm::VectorType::getExtendedElementVectorType(VTy);
2880 
2881     // %sum = add <4 x i32> %lhs, %rhs
2882     Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
2883     Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
2884     Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
2885 
2886     // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
2887     Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(),
2888                                        SrcTy->getScalarSizeInBits() / 2);
2889     ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt);
2890     Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
2891 
2892     // %res = trunc <4 x i32> %high to <4 x i16>
2893     return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
2894   }
2895   case NEON::BI__builtin_neon_vtrn_v:
2896   case NEON::BI__builtin_neon_vtrnq_v: {
2897     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2898     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2899     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2900     Value *SV = nullptr;
2901 
2902     for (unsigned vi = 0; vi != 2; ++vi) {
2903       SmallVector<Constant*, 16> Indices;
2904       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2905         Indices.push_back(Builder.getInt32(i+vi));
2906         Indices.push_back(Builder.getInt32(i+e+vi));
2907       }
2908       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2909       SV = llvm::ConstantVector::get(Indices);
2910       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
2911       SV = Builder.CreateStore(SV, Addr);
2912     }
2913     return SV;
2914   }
2915   case NEON::BI__builtin_neon_vtst_v:
2916   case NEON::BI__builtin_neon_vtstq_v: {
2917     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
2918     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2919     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
2920     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
2921                                 ConstantAggregateZero::get(Ty));
2922     return Builder.CreateSExt(Ops[0], Ty, "vtst");
2923   }
2924   case NEON::BI__builtin_neon_vuzp_v:
2925   case NEON::BI__builtin_neon_vuzpq_v: {
2926     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2927     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2928     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2929     Value *SV = nullptr;
2930 
2931     for (unsigned vi = 0; vi != 2; ++vi) {
2932       SmallVector<Constant*, 16> Indices;
2933       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
2934         Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
2935 
2936       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2937       SV = llvm::ConstantVector::get(Indices);
2938       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
2939       SV = Builder.CreateStore(SV, Addr);
2940     }
2941     return SV;
2942   }
2943   case NEON::BI__builtin_neon_vzip_v:
2944   case NEON::BI__builtin_neon_vzipq_v: {
2945     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
2946     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
2947     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
2948     Value *SV = nullptr;
2949 
2950     for (unsigned vi = 0; vi != 2; ++vi) {
2951       SmallVector<Constant*, 16> Indices;
2952       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
2953         Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
2954         Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
2955       }
2956       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
2957       SV = llvm::ConstantVector::get(Indices);
2958       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
2959       SV = Builder.CreateStore(SV, Addr);
2960     }
2961     return SV;
2962   }
2963   }
2964 
2965   assert(Int && "Expected valid intrinsic number");
2966 
2967   // Determine the type(s) of this overloaded AArch64 intrinsic.
2968   Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E);
2969 
2970   Value *Result = EmitNeonCall(F, Ops, NameHint);
2971   llvm::Type *ResultType = ConvertType(E->getType());
2972   // AArch64 intrinsic one-element vector type cast to
2973   // scalar type expected by the builtin
2974   return Builder.CreateBitCast(Result, ResultType, NameHint);
2975 }
2976 
EmitAArch64CompareBuiltinExpr(Value * Op,llvm::Type * Ty,const CmpInst::Predicate Fp,const CmpInst::Predicate Ip,const Twine & Name)2977 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr(
2978     Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp,
2979     const CmpInst::Predicate Ip, const Twine &Name) {
2980   llvm::Type *OTy = Op->getType();
2981 
2982   // FIXME: this is utterly horrific. We should not be looking at previous
2983   // codegen context to find out what needs doing. Unfortunately TableGen
2984   // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32
2985   // (etc).
2986   if (BitCastInst *BI = dyn_cast<BitCastInst>(Op))
2987     OTy = BI->getOperand(0)->getType();
2988 
2989   Op = Builder.CreateBitCast(Op, OTy);
2990   if (OTy->getScalarType()->isFloatingPointTy()) {
2991     Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy));
2992   } else {
2993     Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy));
2994   }
2995   return Builder.CreateSExt(Op, Ty, Name);
2996 }
2997 
packTBLDVectorList(CodeGenFunction & CGF,ArrayRef<Value * > Ops,Value * ExtOp,Value * IndexOp,llvm::Type * ResTy,unsigned IntID,const char * Name)2998 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
2999                                  Value *ExtOp, Value *IndexOp,
3000                                  llvm::Type *ResTy, unsigned IntID,
3001                                  const char *Name) {
3002   SmallVector<Value *, 2> TblOps;
3003   if (ExtOp)
3004     TblOps.push_back(ExtOp);
3005 
3006   // Build a vector containing sequential number like (0, 1, 2, ..., 15)
3007   SmallVector<Constant*, 16> Indices;
3008   llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType());
3009   for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) {
3010     Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i));
3011     Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1));
3012   }
3013   Value *SV = llvm::ConstantVector::get(Indices);
3014 
3015   int PairPos = 0, End = Ops.size() - 1;
3016   while (PairPos < End) {
3017     TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
3018                                                      Ops[PairPos+1], SV, Name));
3019     PairPos += 2;
3020   }
3021 
3022   // If there's an odd number of 64-bit lookup table, fill the high 64-bit
3023   // of the 128-bit lookup table with zero.
3024   if (PairPos == End) {
3025     Value *ZeroTbl = ConstantAggregateZero::get(TblTy);
3026     TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
3027                                                      ZeroTbl, SV, Name));
3028   }
3029 
3030   Function *TblF;
3031   TblOps.push_back(IndexOp);
3032   TblF = CGF.CGM.getIntrinsic(IntID, ResTy);
3033 
3034   return CGF.EmitNeonCall(TblF, TblOps, Name);
3035 }
3036 
EmitARMBuiltinExpr(unsigned BuiltinID,const CallExpr * E)3037 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
3038                                            const CallExpr *E) {
3039   unsigned HintID = static_cast<unsigned>(-1);
3040   switch (BuiltinID) {
3041   default: break;
3042   case ARM::BI__builtin_arm_yield:
3043   case ARM::BI__yield:
3044     HintID = 1;
3045     break;
3046   case ARM::BI__builtin_arm_wfe:
3047   case ARM::BI__wfe:
3048     HintID = 2;
3049     break;
3050   case ARM::BI__builtin_arm_wfi:
3051   case ARM::BI__wfi:
3052     HintID = 3;
3053     break;
3054   case ARM::BI__builtin_arm_sev:
3055   case ARM::BI__sev:
3056     HintID = 4;
3057     break;
3058   case ARM::BI__builtin_arm_sevl:
3059   case ARM::BI__sevl:
3060     HintID = 5;
3061     break;
3062   }
3063 
3064   if (HintID != static_cast<unsigned>(-1)) {
3065     Function *F = CGM.getIntrinsic(Intrinsic::arm_hint);
3066     return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID));
3067   }
3068 
3069   if (BuiltinID == ARM::BI__builtin_arm_rbit) {
3070     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit),
3071                                                EmitScalarExpr(E->getArg(0)),
3072                               "rbit");
3073   }
3074 
3075   if (BuiltinID == ARM::BI__clear_cache) {
3076     assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
3077     const FunctionDecl *FD = E->getDirectCallee();
3078     SmallVector<Value*, 2> Ops;
3079     for (unsigned i = 0; i < 2; i++)
3080       Ops.push_back(EmitScalarExpr(E->getArg(i)));
3081     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
3082     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
3083     StringRef Name = FD->getName();
3084     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
3085   }
3086 
3087   if (BuiltinID == ARM::BI__builtin_arm_ldrexd ||
3088       ((BuiltinID == ARM::BI__builtin_arm_ldrex ||
3089         BuiltinID == ARM::BI__builtin_arm_ldaex) &&
3090        getContext().getTypeSize(E->getType()) == 64) ||
3091       BuiltinID == ARM::BI__ldrexd) {
3092     Function *F;
3093 
3094     switch (BuiltinID) {
3095     default: llvm_unreachable("unexpected builtin");
3096     case ARM::BI__builtin_arm_ldaex:
3097       F = CGM.getIntrinsic(Intrinsic::arm_ldaexd);
3098       break;
3099     case ARM::BI__builtin_arm_ldrexd:
3100     case ARM::BI__builtin_arm_ldrex:
3101     case ARM::BI__ldrexd:
3102       F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
3103       break;
3104     }
3105 
3106     Value *LdPtr = EmitScalarExpr(E->getArg(0));
3107     Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
3108                                     "ldrexd");
3109 
3110     Value *Val0 = Builder.CreateExtractValue(Val, 1);
3111     Value *Val1 = Builder.CreateExtractValue(Val, 0);
3112     Val0 = Builder.CreateZExt(Val0, Int64Ty);
3113     Val1 = Builder.CreateZExt(Val1, Int64Ty);
3114 
3115     Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
3116     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
3117     Val = Builder.CreateOr(Val, Val1);
3118     return Builder.CreateBitCast(Val, ConvertType(E->getType()));
3119   }
3120 
3121   if (BuiltinID == ARM::BI__builtin_arm_ldrex ||
3122       BuiltinID == ARM::BI__builtin_arm_ldaex) {
3123     Value *LoadAddr = EmitScalarExpr(E->getArg(0));
3124 
3125     QualType Ty = E->getType();
3126     llvm::Type *RealResTy = ConvertType(Ty);
3127     llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(),
3128                                                   getContext().getTypeSize(Ty));
3129     LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo());
3130 
3131     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex
3132                                        ? Intrinsic::arm_ldaex
3133                                        : Intrinsic::arm_ldrex,
3134                                    LoadAddr->getType());
3135     Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
3136 
3137     if (RealResTy->isPointerTy())
3138       return Builder.CreateIntToPtr(Val, RealResTy);
3139     else {
3140       Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
3141       return Builder.CreateBitCast(Val, RealResTy);
3142     }
3143   }
3144 
3145   if (BuiltinID == ARM::BI__builtin_arm_strexd ||
3146       ((BuiltinID == ARM::BI__builtin_arm_stlex ||
3147         BuiltinID == ARM::BI__builtin_arm_strex) &&
3148        getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
3149     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
3150                                        ? Intrinsic::arm_stlexd
3151                                        : Intrinsic::arm_strexd);
3152     llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL);
3153 
3154     Value *Tmp = CreateMemTemp(E->getArg(0)->getType());
3155     Value *Val = EmitScalarExpr(E->getArg(0));
3156     Builder.CreateStore(Val, Tmp);
3157 
3158     Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
3159     Val = Builder.CreateLoad(LdPtr);
3160 
3161     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
3162     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
3163     Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
3164     return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd");
3165   }
3166 
3167   if (BuiltinID == ARM::BI__builtin_arm_strex ||
3168       BuiltinID == ARM::BI__builtin_arm_stlex) {
3169     Value *StoreVal = EmitScalarExpr(E->getArg(0));
3170     Value *StoreAddr = EmitScalarExpr(E->getArg(1));
3171 
3172     QualType Ty = E->getArg(0)->getType();
3173     llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
3174                                                  getContext().getTypeSize(Ty));
3175     StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
3176 
3177     if (StoreVal->getType()->isPointerTy())
3178       StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
3179     else {
3180       StoreVal = Builder.CreateBitCast(StoreVal, StoreTy);
3181       StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
3182     }
3183 
3184     Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
3185                                        ? Intrinsic::arm_stlex
3186                                        : Intrinsic::arm_strex,
3187                                    StoreAddr->getType());
3188     return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex");
3189   }
3190 
3191   if (BuiltinID == ARM::BI__builtin_arm_clrex) {
3192     Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
3193     return Builder.CreateCall(F);
3194   }
3195 
3196   // CRC32
3197   Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
3198   switch (BuiltinID) {
3199   case ARM::BI__builtin_arm_crc32b:
3200     CRCIntrinsicID = Intrinsic::arm_crc32b; break;
3201   case ARM::BI__builtin_arm_crc32cb:
3202     CRCIntrinsicID = Intrinsic::arm_crc32cb; break;
3203   case ARM::BI__builtin_arm_crc32h:
3204     CRCIntrinsicID = Intrinsic::arm_crc32h; break;
3205   case ARM::BI__builtin_arm_crc32ch:
3206     CRCIntrinsicID = Intrinsic::arm_crc32ch; break;
3207   case ARM::BI__builtin_arm_crc32w:
3208   case ARM::BI__builtin_arm_crc32d:
3209     CRCIntrinsicID = Intrinsic::arm_crc32w; break;
3210   case ARM::BI__builtin_arm_crc32cw:
3211   case ARM::BI__builtin_arm_crc32cd:
3212     CRCIntrinsicID = Intrinsic::arm_crc32cw; break;
3213   }
3214 
3215   if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
3216     Value *Arg0 = EmitScalarExpr(E->getArg(0));
3217     Value *Arg1 = EmitScalarExpr(E->getArg(1));
3218 
3219     // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w
3220     // intrinsics, hence we need different codegen for these cases.
3221     if (BuiltinID == ARM::BI__builtin_arm_crc32d ||
3222         BuiltinID == ARM::BI__builtin_arm_crc32cd) {
3223       Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
3224       Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty);
3225       Value *Arg1b = Builder.CreateLShr(Arg1, C1);
3226       Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty);
3227 
3228       Function *F = CGM.getIntrinsic(CRCIntrinsicID);
3229       Value *Res = Builder.CreateCall2(F, Arg0, Arg1a);
3230       return Builder.CreateCall2(F, Res, Arg1b);
3231     } else {
3232       Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty);
3233 
3234       Function *F = CGM.getIntrinsic(CRCIntrinsicID);
3235       return Builder.CreateCall2(F, Arg0, Arg1);
3236     }
3237   }
3238 
3239   SmallVector<Value*, 4> Ops;
3240   llvm::Value *Align = nullptr;
3241   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
3242     if (i == 0) {
3243       switch (BuiltinID) {
3244       case NEON::BI__builtin_neon_vld1_v:
3245       case NEON::BI__builtin_neon_vld1q_v:
3246       case NEON::BI__builtin_neon_vld1q_lane_v:
3247       case NEON::BI__builtin_neon_vld1_lane_v:
3248       case NEON::BI__builtin_neon_vld1_dup_v:
3249       case NEON::BI__builtin_neon_vld1q_dup_v:
3250       case NEON::BI__builtin_neon_vst1_v:
3251       case NEON::BI__builtin_neon_vst1q_v:
3252       case NEON::BI__builtin_neon_vst1q_lane_v:
3253       case NEON::BI__builtin_neon_vst1_lane_v:
3254       case NEON::BI__builtin_neon_vst2_v:
3255       case NEON::BI__builtin_neon_vst2q_v:
3256       case NEON::BI__builtin_neon_vst2_lane_v:
3257       case NEON::BI__builtin_neon_vst2q_lane_v:
3258       case NEON::BI__builtin_neon_vst3_v:
3259       case NEON::BI__builtin_neon_vst3q_v:
3260       case NEON::BI__builtin_neon_vst3_lane_v:
3261       case NEON::BI__builtin_neon_vst3q_lane_v:
3262       case NEON::BI__builtin_neon_vst4_v:
3263       case NEON::BI__builtin_neon_vst4q_v:
3264       case NEON::BI__builtin_neon_vst4_lane_v:
3265       case NEON::BI__builtin_neon_vst4q_lane_v:
3266         // Get the alignment for the argument in addition to the value;
3267         // we'll use it later.
3268         std::pair<llvm::Value*, unsigned> Src =
3269             EmitPointerWithAlignment(E->getArg(0));
3270         Ops.push_back(Src.first);
3271         Align = Builder.getInt32(Src.second);
3272         continue;
3273       }
3274     }
3275     if (i == 1) {
3276       switch (BuiltinID) {
3277       case NEON::BI__builtin_neon_vld2_v:
3278       case NEON::BI__builtin_neon_vld2q_v:
3279       case NEON::BI__builtin_neon_vld3_v:
3280       case NEON::BI__builtin_neon_vld3q_v:
3281       case NEON::BI__builtin_neon_vld4_v:
3282       case NEON::BI__builtin_neon_vld4q_v:
3283       case NEON::BI__builtin_neon_vld2_lane_v:
3284       case NEON::BI__builtin_neon_vld2q_lane_v:
3285       case NEON::BI__builtin_neon_vld3_lane_v:
3286       case NEON::BI__builtin_neon_vld3q_lane_v:
3287       case NEON::BI__builtin_neon_vld4_lane_v:
3288       case NEON::BI__builtin_neon_vld4q_lane_v:
3289       case NEON::BI__builtin_neon_vld2_dup_v:
3290       case NEON::BI__builtin_neon_vld3_dup_v:
3291       case NEON::BI__builtin_neon_vld4_dup_v:
3292         // Get the alignment for the argument in addition to the value;
3293         // we'll use it later.
3294         std::pair<llvm::Value*, unsigned> Src =
3295             EmitPointerWithAlignment(E->getArg(1));
3296         Ops.push_back(Src.first);
3297         Align = Builder.getInt32(Src.second);
3298         continue;
3299       }
3300     }
3301     Ops.push_back(EmitScalarExpr(E->getArg(i)));
3302   }
3303 
3304   switch (BuiltinID) {
3305   default: break;
3306   // vget_lane and vset_lane are not overloaded and do not have an extra
3307   // argument that specifies the vector type.
3308   case NEON::BI__builtin_neon_vget_lane_i8:
3309   case NEON::BI__builtin_neon_vget_lane_i16:
3310   case NEON::BI__builtin_neon_vget_lane_i32:
3311   case NEON::BI__builtin_neon_vget_lane_i64:
3312   case NEON::BI__builtin_neon_vget_lane_f32:
3313   case NEON::BI__builtin_neon_vgetq_lane_i8:
3314   case NEON::BI__builtin_neon_vgetq_lane_i16:
3315   case NEON::BI__builtin_neon_vgetq_lane_i32:
3316   case NEON::BI__builtin_neon_vgetq_lane_i64:
3317   case NEON::BI__builtin_neon_vgetq_lane_f32:
3318     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
3319                                         "vget_lane");
3320   case NEON::BI__builtin_neon_vset_lane_i8:
3321   case NEON::BI__builtin_neon_vset_lane_i16:
3322   case NEON::BI__builtin_neon_vset_lane_i32:
3323   case NEON::BI__builtin_neon_vset_lane_i64:
3324   case NEON::BI__builtin_neon_vset_lane_f32:
3325   case NEON::BI__builtin_neon_vsetq_lane_i8:
3326   case NEON::BI__builtin_neon_vsetq_lane_i16:
3327   case NEON::BI__builtin_neon_vsetq_lane_i32:
3328   case NEON::BI__builtin_neon_vsetq_lane_i64:
3329   case NEON::BI__builtin_neon_vsetq_lane_f32:
3330     Ops.push_back(EmitScalarExpr(E->getArg(2)));
3331     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
3332 
3333   // Non-polymorphic crypto instructions also not overloaded
3334   case NEON::BI__builtin_neon_vsha1h_u32:
3335     Ops.push_back(EmitScalarExpr(E->getArg(0)));
3336     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops,
3337                         "vsha1h");
3338   case NEON::BI__builtin_neon_vsha1cq_u32:
3339     Ops.push_back(EmitScalarExpr(E->getArg(2)));
3340     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops,
3341                         "vsha1h");
3342   case NEON::BI__builtin_neon_vsha1pq_u32:
3343     Ops.push_back(EmitScalarExpr(E->getArg(2)));
3344     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops,
3345                         "vsha1h");
3346   case NEON::BI__builtin_neon_vsha1mq_u32:
3347     Ops.push_back(EmitScalarExpr(E->getArg(2)));
3348     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops,
3349                         "vsha1h");
3350   }
3351 
3352   // Get the last argument, which specifies the vector type.
3353   llvm::APSInt Result;
3354   const Expr *Arg = E->getArg(E->getNumArgs()-1);
3355   if (!Arg->isIntegerConstantExpr(Result, getContext()))
3356     return nullptr;
3357 
3358   if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
3359       BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
3360     // Determine the overloaded type of this builtin.
3361     llvm::Type *Ty;
3362     if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
3363       Ty = FloatTy;
3364     else
3365       Ty = DoubleTy;
3366 
3367     // Determine whether this is an unsigned conversion or not.
3368     bool usgn = Result.getZExtValue() == 1;
3369     unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
3370 
3371     // Call the appropriate intrinsic.
3372     Function *F = CGM.getIntrinsic(Int, Ty);
3373     return Builder.CreateCall(F, Ops, "vcvtr");
3374   }
3375 
3376   // Determine the type of this overloaded NEON intrinsic.
3377   NeonTypeFlags Type(Result.getZExtValue());
3378   bool usgn = Type.isUnsigned();
3379   bool rightShift = false;
3380 
3381   llvm::VectorType *VTy = GetNeonType(this, Type);
3382   llvm::Type *Ty = VTy;
3383   if (!Ty)
3384     return nullptr;
3385 
3386   // Many NEON builtins have identical semantics and uses in ARM and
3387   // AArch64. Emit these in a single function.
3388   ArrayRef<NeonIntrinsicInfo> IntrinsicMap(ARMSIMDIntrinsicMap);
3389   const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
3390       IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted);
3391   if (Builtin)
3392     return EmitCommonNeonBuiltinExpr(
3393         Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
3394         Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align);
3395 
3396   unsigned Int;
3397   switch (BuiltinID) {
3398   default: return nullptr;
3399   case NEON::BI__builtin_neon_vld1q_lane_v:
3400     // Handle 64-bit integer elements as a special case.  Use shuffles of
3401     // one-element vectors to avoid poor code for i64 in the backend.
3402     if (VTy->getElementType()->isIntegerTy(64)) {
3403       // Extract the other lane.
3404       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3405       int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
3406       Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
3407       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
3408       // Load the value as a one-element vector.
3409       Ty = llvm::VectorType::get(VTy->getElementType(), 1);
3410       Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty);
3411       Value *Ld = Builder.CreateCall2(F, Ops[0], Align);
3412       // Combine them.
3413       SmallVector<Constant*, 2> Indices;
3414       Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane));
3415       Indices.push_back(ConstantInt::get(Int32Ty, Lane));
3416       SV = llvm::ConstantVector::get(Indices);
3417       return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane");
3418     }
3419     // fall through
3420   case NEON::BI__builtin_neon_vld1_lane_v: {
3421     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3422     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
3423     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3424     LoadInst *Ld = Builder.CreateLoad(Ops[0]);
3425     Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
3426     return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
3427   }
3428   case NEON::BI__builtin_neon_vld2_dup_v:
3429   case NEON::BI__builtin_neon_vld3_dup_v:
3430   case NEON::BI__builtin_neon_vld4_dup_v: {
3431     // Handle 64-bit elements as a special-case.  There is no "dup" needed.
3432     if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
3433       switch (BuiltinID) {
3434       case NEON::BI__builtin_neon_vld2_dup_v:
3435         Int = Intrinsic::arm_neon_vld2;
3436         break;
3437       case NEON::BI__builtin_neon_vld3_dup_v:
3438         Int = Intrinsic::arm_neon_vld3;
3439         break;
3440       case NEON::BI__builtin_neon_vld4_dup_v:
3441         Int = Intrinsic::arm_neon_vld4;
3442         break;
3443       default: llvm_unreachable("unknown vld_dup intrinsic?");
3444       }
3445       Function *F = CGM.getIntrinsic(Int, Ty);
3446       Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup");
3447       Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3448       Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3449       return Builder.CreateStore(Ops[1], Ops[0]);
3450     }
3451     switch (BuiltinID) {
3452     case NEON::BI__builtin_neon_vld2_dup_v:
3453       Int = Intrinsic::arm_neon_vld2lane;
3454       break;
3455     case NEON::BI__builtin_neon_vld3_dup_v:
3456       Int = Intrinsic::arm_neon_vld3lane;
3457       break;
3458     case NEON::BI__builtin_neon_vld4_dup_v:
3459       Int = Intrinsic::arm_neon_vld4lane;
3460       break;
3461     default: llvm_unreachable("unknown vld_dup intrinsic?");
3462     }
3463     Function *F = CGM.getIntrinsic(Int, Ty);
3464     llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
3465 
3466     SmallVector<Value*, 6> Args;
3467     Args.push_back(Ops[1]);
3468     Args.append(STy->getNumElements(), UndefValue::get(Ty));
3469 
3470     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
3471     Args.push_back(CI);
3472     Args.push_back(Align);
3473 
3474     Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
3475     // splat lane 0 to all elts in each vector of the result.
3476     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3477       Value *Val = Builder.CreateExtractValue(Ops[1], i);
3478       Value *Elt = Builder.CreateBitCast(Val, Ty);
3479       Elt = EmitNeonSplat(Elt, CI);
3480       Elt = Builder.CreateBitCast(Elt, Val->getType());
3481       Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
3482     }
3483     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3484     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3485     return Builder.CreateStore(Ops[1], Ops[0]);
3486   }
3487   case NEON::BI__builtin_neon_vqrshrn_n_v:
3488     Int =
3489       usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
3490     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
3491                         1, true);
3492   case NEON::BI__builtin_neon_vqrshrun_n_v:
3493     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
3494                         Ops, "vqrshrun_n", 1, true);
3495   case NEON::BI__builtin_neon_vqshlu_n_v:
3496   case NEON::BI__builtin_neon_vqshluq_n_v:
3497     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty),
3498                         Ops, "vqshlu", 1, false);
3499   case NEON::BI__builtin_neon_vqshrn_n_v:
3500     Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
3501     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
3502                         1, true);
3503   case NEON::BI__builtin_neon_vqshrun_n_v:
3504     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
3505                         Ops, "vqshrun_n", 1, true);
3506   case NEON::BI__builtin_neon_vrecpe_v:
3507   case NEON::BI__builtin_neon_vrecpeq_v:
3508     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
3509                         Ops, "vrecpe");
3510   case NEON::BI__builtin_neon_vrshrn_n_v:
3511     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
3512                         Ops, "vrshrn_n", 1, true);
3513   case NEON::BI__builtin_neon_vrshr_n_v:
3514   case NEON::BI__builtin_neon_vrshrq_n_v:
3515     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
3516     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true);
3517   case NEON::BI__builtin_neon_vrsra_n_v:
3518   case NEON::BI__builtin_neon_vrsraq_n_v:
3519     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3520     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3521     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
3522     Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
3523     Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]);
3524     return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
3525   case NEON::BI__builtin_neon_vsri_n_v:
3526   case NEON::BI__builtin_neon_vsriq_n_v:
3527     rightShift = true;
3528   case NEON::BI__builtin_neon_vsli_n_v:
3529   case NEON::BI__builtin_neon_vsliq_n_v:
3530     Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
3531     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
3532                         Ops, "vsli_n");
3533   case NEON::BI__builtin_neon_vsra_n_v:
3534   case NEON::BI__builtin_neon_vsraq_n_v:
3535     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
3536     Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
3537     return Builder.CreateAdd(Ops[0], Ops[1]);
3538   case NEON::BI__builtin_neon_vst1q_lane_v:
3539     // Handle 64-bit integer elements as a special case.  Use a shuffle to get
3540     // a one-element vector and avoid poor code for i64 in the backend.
3541     if (VTy->getElementType()->isIntegerTy(64)) {
3542       Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3543       Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
3544       Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
3545       Ops[2] = Align;
3546       return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
3547                                                  Ops[1]->getType()), Ops);
3548     }
3549     // fall through
3550   case NEON::BI__builtin_neon_vst1_lane_v: {
3551     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
3552     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
3553     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
3554     StoreInst *St = Builder.CreateStore(Ops[1],
3555                                         Builder.CreateBitCast(Ops[0], Ty));
3556     St->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
3557     return St;
3558   }
3559   case NEON::BI__builtin_neon_vtbl1_v:
3560     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
3561                         Ops, "vtbl1");
3562   case NEON::BI__builtin_neon_vtbl2_v:
3563     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
3564                         Ops, "vtbl2");
3565   case NEON::BI__builtin_neon_vtbl3_v:
3566     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
3567                         Ops, "vtbl3");
3568   case NEON::BI__builtin_neon_vtbl4_v:
3569     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
3570                         Ops, "vtbl4");
3571   case NEON::BI__builtin_neon_vtbx1_v:
3572     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
3573                         Ops, "vtbx1");
3574   case NEON::BI__builtin_neon_vtbx2_v:
3575     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
3576                         Ops, "vtbx2");
3577   case NEON::BI__builtin_neon_vtbx3_v:
3578     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
3579                         Ops, "vtbx3");
3580   case NEON::BI__builtin_neon_vtbx4_v:
3581     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
3582                         Ops, "vtbx4");
3583   }
3584 }
3585 
EmitAArch64TblBuiltinExpr(CodeGenFunction & CGF,unsigned BuiltinID,const CallExpr * E,SmallVectorImpl<Value * > & Ops)3586 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID,
3587                                       const CallExpr *E,
3588                                       SmallVectorImpl<Value *> &Ops) {
3589   unsigned int Int = 0;
3590   const char *s = nullptr;
3591 
3592   switch (BuiltinID) {
3593   default:
3594     return nullptr;
3595   case NEON::BI__builtin_neon_vtbl1_v:
3596   case NEON::BI__builtin_neon_vqtbl1_v:
3597   case NEON::BI__builtin_neon_vqtbl1q_v:
3598   case NEON::BI__builtin_neon_vtbl2_v:
3599   case NEON::BI__builtin_neon_vqtbl2_v:
3600   case NEON::BI__builtin_neon_vqtbl2q_v:
3601   case NEON::BI__builtin_neon_vtbl3_v:
3602   case NEON::BI__builtin_neon_vqtbl3_v:
3603   case NEON::BI__builtin_neon_vqtbl3q_v:
3604   case NEON::BI__builtin_neon_vtbl4_v:
3605   case NEON::BI__builtin_neon_vqtbl4_v:
3606   case NEON::BI__builtin_neon_vqtbl4q_v:
3607     break;
3608   case NEON::BI__builtin_neon_vtbx1_v:
3609   case NEON::BI__builtin_neon_vqtbx1_v:
3610   case NEON::BI__builtin_neon_vqtbx1q_v:
3611   case NEON::BI__builtin_neon_vtbx2_v:
3612   case NEON::BI__builtin_neon_vqtbx2_v:
3613   case NEON::BI__builtin_neon_vqtbx2q_v:
3614   case NEON::BI__builtin_neon_vtbx3_v:
3615   case NEON::BI__builtin_neon_vqtbx3_v:
3616   case NEON::BI__builtin_neon_vqtbx3q_v:
3617   case NEON::BI__builtin_neon_vtbx4_v:
3618   case NEON::BI__builtin_neon_vqtbx4_v:
3619   case NEON::BI__builtin_neon_vqtbx4q_v:
3620     break;
3621   }
3622 
3623   assert(E->getNumArgs() >= 3);
3624 
3625   // Get the last argument, which specifies the vector type.
3626   llvm::APSInt Result;
3627   const Expr *Arg = E->getArg(E->getNumArgs() - 1);
3628   if (!Arg->isIntegerConstantExpr(Result, CGF.getContext()))
3629     return nullptr;
3630 
3631   // Determine the type of this overloaded NEON intrinsic.
3632   NeonTypeFlags Type(Result.getZExtValue());
3633   llvm::VectorType *VTy = GetNeonType(&CGF, Type);
3634   llvm::Type *Ty = VTy;
3635   if (!Ty)
3636     return nullptr;
3637 
3638   unsigned nElts = VTy->getNumElements();
3639 
3640   CodeGen::CGBuilderTy &Builder = CGF.Builder;
3641 
3642   // AArch64 scalar builtins are not overloaded, they do not have an extra
3643   // argument that specifies the vector type, need to handle each case.
3644   SmallVector<Value *, 2> TblOps;
3645   switch (BuiltinID) {
3646   case NEON::BI__builtin_neon_vtbl1_v: {
3647     TblOps.push_back(Ops[0]);
3648     return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty,
3649                               Intrinsic::aarch64_neon_tbl1, "vtbl1");
3650   }
3651   case NEON::BI__builtin_neon_vtbl2_v: {
3652     TblOps.push_back(Ops[0]);
3653     TblOps.push_back(Ops[1]);
3654     return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty,
3655                               Intrinsic::aarch64_neon_tbl1, "vtbl1");
3656   }
3657   case NEON::BI__builtin_neon_vtbl3_v: {
3658     TblOps.push_back(Ops[0]);
3659     TblOps.push_back(Ops[1]);
3660     TblOps.push_back(Ops[2]);
3661     return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty,
3662                               Intrinsic::aarch64_neon_tbl2, "vtbl2");
3663   }
3664   case NEON::BI__builtin_neon_vtbl4_v: {
3665     TblOps.push_back(Ops[0]);
3666     TblOps.push_back(Ops[1]);
3667     TblOps.push_back(Ops[2]);
3668     TblOps.push_back(Ops[3]);
3669     return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty,
3670                               Intrinsic::aarch64_neon_tbl2, "vtbl2");
3671   }
3672   case NEON::BI__builtin_neon_vtbx1_v: {
3673     TblOps.push_back(Ops[1]);
3674     Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty,
3675                                        Intrinsic::aarch64_neon_tbl1, "vtbl1");
3676 
3677     llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8);
3678     Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight);
3679     Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV);
3680     CmpRes = Builder.CreateSExt(CmpRes, Ty);
3681 
3682     Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
3683     Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
3684     return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
3685   }
3686   case NEON::BI__builtin_neon_vtbx2_v: {
3687     TblOps.push_back(Ops[1]);
3688     TblOps.push_back(Ops[2]);
3689     return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty,
3690                               Intrinsic::aarch64_neon_tbx1, "vtbx1");
3691   }
3692   case NEON::BI__builtin_neon_vtbx3_v: {
3693     TblOps.push_back(Ops[1]);
3694     TblOps.push_back(Ops[2]);
3695     TblOps.push_back(Ops[3]);
3696     Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty,
3697                                        Intrinsic::aarch64_neon_tbl2, "vtbl2");
3698 
3699     llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24);
3700     Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour);
3701     Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4],
3702                                            TwentyFourV);
3703     CmpRes = Builder.CreateSExt(CmpRes, Ty);
3704 
3705     Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
3706     Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
3707     return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
3708   }
3709   case NEON::BI__builtin_neon_vtbx4_v: {
3710     TblOps.push_back(Ops[1]);
3711     TblOps.push_back(Ops[2]);
3712     TblOps.push_back(Ops[3]);
3713     TblOps.push_back(Ops[4]);
3714     return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty,
3715                               Intrinsic::aarch64_neon_tbx2, "vtbx2");
3716   }
3717   case NEON::BI__builtin_neon_vqtbl1_v:
3718   case NEON::BI__builtin_neon_vqtbl1q_v:
3719     Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break;
3720   case NEON::BI__builtin_neon_vqtbl2_v:
3721   case NEON::BI__builtin_neon_vqtbl2q_v: {
3722     Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break;
3723   case NEON::BI__builtin_neon_vqtbl3_v:
3724   case NEON::BI__builtin_neon_vqtbl3q_v:
3725     Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break;
3726   case NEON::BI__builtin_neon_vqtbl4_v:
3727   case NEON::BI__builtin_neon_vqtbl4q_v:
3728     Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break;
3729   case NEON::BI__builtin_neon_vqtbx1_v:
3730   case NEON::BI__builtin_neon_vqtbx1q_v:
3731     Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break;
3732   case NEON::BI__builtin_neon_vqtbx2_v:
3733   case NEON::BI__builtin_neon_vqtbx2q_v:
3734     Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break;
3735   case NEON::BI__builtin_neon_vqtbx3_v:
3736   case NEON::BI__builtin_neon_vqtbx3q_v:
3737     Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break;
3738   case NEON::BI__builtin_neon_vqtbx4_v:
3739   case NEON::BI__builtin_neon_vqtbx4q_v:
3740     Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break;
3741   }
3742   }
3743 
3744   if (!Int)
3745     return nullptr;
3746 
3747   Function *F = CGF.CGM.getIntrinsic(Int, Ty);
3748   return CGF.EmitNeonCall(F, Ops, s);
3749 }
3750 
vectorWrapScalar16(Value * Op)3751 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) {
3752   llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
3753   Op = Builder.CreateBitCast(Op, Int16Ty);
3754   Value *V = UndefValue::get(VTy);
3755   llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
3756   Op = Builder.CreateInsertElement(V, Op, CI);
3757   return Op;
3758 }
3759 
vectorWrapScalar8(Value * Op)3760 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) {
3761   llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
3762   Op = Builder.CreateBitCast(Op, Int8Ty);
3763   Value *V = UndefValue::get(VTy);
3764   llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
3765   Op = Builder.CreateInsertElement(V, Op, CI);
3766   return Op;
3767 }
3768 
3769 Value *CodeGenFunction::
emitVectorWrappedScalar8Intrinsic(unsigned Int,SmallVectorImpl<Value * > & Ops,const char * Name)3770 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops,
3771                                   const char *Name) {
3772   // i8 is not a legal types for AArch64, so we can't just use
3773   // a normal overloaded intrinsic call for these scalar types. Instead
3774   // we'll build 64-bit vectors w/ lane zero being our input values and
3775   // perform the operation on that. The back end can pattern match directly
3776   // to the scalar instruction.
3777   Ops[0] = vectorWrapScalar8(Ops[0]);
3778   Ops[1] = vectorWrapScalar8(Ops[1]);
3779   llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
3780   Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
3781   Constant *CI = ConstantInt::get(SizeTy, 0);
3782   return Builder.CreateExtractElement(V, CI, "lane0");
3783 }
3784 
3785 Value *CodeGenFunction::
emitVectorWrappedScalar16Intrinsic(unsigned Int,SmallVectorImpl<Value * > & Ops,const char * Name)3786 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops,
3787                                    const char *Name) {
3788   // i16 is not a legal types for AArch64, so we can't just use
3789   // a normal overloaded intrinsic call for these scalar types. Instead
3790   // we'll build 64-bit vectors w/ lane zero being our input values and
3791   // perform the operation on that. The back end can pattern match directly
3792   // to the scalar instruction.
3793   Ops[0] = vectorWrapScalar16(Ops[0]);
3794   Ops[1] = vectorWrapScalar16(Ops[1]);
3795   llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
3796   Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
3797   Constant *CI = ConstantInt::get(SizeTy, 0);
3798   return Builder.CreateExtractElement(V, CI, "lane0");
3799 }
3800 
EmitAArch64BuiltinExpr(unsigned BuiltinID,const CallExpr * E)3801 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
3802                                                const CallExpr *E) {
3803   if (BuiltinID == AArch64::BI__builtin_arm_rbit) {
3804     assert((getContext().getTypeSize(E->getType()) == 32) &&
3805            "rbit of unusual size!");
3806     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
3807     return Builder.CreateCall(
3808         CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit");
3809   }
3810   if (BuiltinID == AArch64::BI__builtin_arm_rbit64) {
3811     assert((getContext().getTypeSize(E->getType()) == 64) &&
3812            "rbit of unusual size!");
3813     llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
3814     return Builder.CreateCall(
3815         CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit");
3816   }
3817 
3818   if (BuiltinID == AArch64::BI__clear_cache) {
3819     assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
3820     const FunctionDecl *FD = E->getDirectCallee();
3821     SmallVector<Value*, 2> Ops;
3822     for (unsigned i = 0; i < 2; i++)
3823       Ops.push_back(EmitScalarExpr(E->getArg(i)));
3824     llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
3825     llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
3826     StringRef Name = FD->getName();
3827     return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
3828   }
3829 
3830   if ((BuiltinID == AArch64::BI__builtin_arm_ldrex ||
3831       BuiltinID == AArch64::BI__builtin_arm_ldaex) &&
3832       getContext().getTypeSize(E->getType()) == 128) {
3833     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
3834                                        ? Intrinsic::aarch64_ldaxp
3835                                        : Intrinsic::aarch64_ldxp);
3836 
3837     Value *LdPtr = EmitScalarExpr(E->getArg(0));
3838     Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
3839                                     "ldxp");
3840 
3841     Value *Val0 = Builder.CreateExtractValue(Val, 1);
3842     Value *Val1 = Builder.CreateExtractValue(Val, 0);
3843     llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
3844     Val0 = Builder.CreateZExt(Val0, Int128Ty);
3845     Val1 = Builder.CreateZExt(Val1, Int128Ty);
3846 
3847     Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64);
3848     Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
3849     Val = Builder.CreateOr(Val, Val1);
3850     return Builder.CreateBitCast(Val, ConvertType(E->getType()));
3851   } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex ||
3852              BuiltinID == AArch64::BI__builtin_arm_ldaex) {
3853     Value *LoadAddr = EmitScalarExpr(E->getArg(0));
3854 
3855     QualType Ty = E->getType();
3856     llvm::Type *RealResTy = ConvertType(Ty);
3857     llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(),
3858                                                   getContext().getTypeSize(Ty));
3859     LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo());
3860 
3861     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
3862                                        ? Intrinsic::aarch64_ldaxr
3863                                        : Intrinsic::aarch64_ldxr,
3864                                    LoadAddr->getType());
3865     Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr");
3866 
3867     if (RealResTy->isPointerTy())
3868       return Builder.CreateIntToPtr(Val, RealResTy);
3869 
3870     Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
3871     return Builder.CreateBitCast(Val, RealResTy);
3872   }
3873 
3874   if ((BuiltinID == AArch64::BI__builtin_arm_strex ||
3875        BuiltinID == AArch64::BI__builtin_arm_stlex) &&
3876       getContext().getTypeSize(E->getArg(0)->getType()) == 128) {
3877     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
3878                                        ? Intrinsic::aarch64_stlxp
3879                                        : Intrinsic::aarch64_stxp);
3880     llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, NULL);
3881 
3882     Value *One = llvm::ConstantInt::get(Int32Ty, 1);
3883     Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()),
3884                                       One);
3885     Value *Val = EmitScalarExpr(E->getArg(0));
3886     Builder.CreateStore(Val, Tmp);
3887 
3888     Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
3889     Val = Builder.CreateLoad(LdPtr);
3890 
3891     Value *Arg0 = Builder.CreateExtractValue(Val, 0);
3892     Value *Arg1 = Builder.CreateExtractValue(Val, 1);
3893     Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)),
3894                                          Int8PtrTy);
3895     return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp");
3896   } else if (BuiltinID == AArch64::BI__builtin_arm_strex ||
3897              BuiltinID == AArch64::BI__builtin_arm_stlex) {
3898     Value *StoreVal = EmitScalarExpr(E->getArg(0));
3899     Value *StoreAddr = EmitScalarExpr(E->getArg(1));
3900 
3901     QualType Ty = E->getArg(0)->getType();
3902     llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
3903                                                  getContext().getTypeSize(Ty));
3904     StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
3905 
3906     if (StoreVal->getType()->isPointerTy())
3907       StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty);
3908     else {
3909       StoreVal = Builder.CreateBitCast(StoreVal, StoreTy);
3910       StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty);
3911     }
3912 
3913     Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
3914                                        ? Intrinsic::aarch64_stlxr
3915                                        : Intrinsic::aarch64_stxr,
3916                                    StoreAddr->getType());
3917     return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr");
3918   }
3919 
3920   if (BuiltinID == AArch64::BI__builtin_arm_clrex) {
3921     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex);
3922     return Builder.CreateCall(F);
3923   }
3924 
3925   // CRC32
3926   Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
3927   switch (BuiltinID) {
3928   case AArch64::BI__builtin_arm_crc32b:
3929     CRCIntrinsicID = Intrinsic::aarch64_crc32b; break;
3930   case AArch64::BI__builtin_arm_crc32cb:
3931     CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break;
3932   case AArch64::BI__builtin_arm_crc32h:
3933     CRCIntrinsicID = Intrinsic::aarch64_crc32h; break;
3934   case AArch64::BI__builtin_arm_crc32ch:
3935     CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break;
3936   case AArch64::BI__builtin_arm_crc32w:
3937     CRCIntrinsicID = Intrinsic::aarch64_crc32w; break;
3938   case AArch64::BI__builtin_arm_crc32cw:
3939     CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break;
3940   case AArch64::BI__builtin_arm_crc32d:
3941     CRCIntrinsicID = Intrinsic::aarch64_crc32x; break;
3942   case AArch64::BI__builtin_arm_crc32cd:
3943     CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break;
3944   }
3945 
3946   if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
3947     Value *Arg0 = EmitScalarExpr(E->getArg(0));
3948     Value *Arg1 = EmitScalarExpr(E->getArg(1));
3949     Function *F = CGM.getIntrinsic(CRCIntrinsicID);
3950 
3951     llvm::Type *DataTy = F->getFunctionType()->getParamType(1);
3952     Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy);
3953 
3954     return Builder.CreateCall2(F, Arg0, Arg1);
3955   }
3956 
3957   llvm::SmallVector<Value*, 4> Ops;
3958   for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++)
3959     Ops.push_back(EmitScalarExpr(E->getArg(i)));
3960 
3961   ArrayRef<NeonIntrinsicInfo> SISDMap(AArch64SISDIntrinsicMap);
3962   const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap(
3963       SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted);
3964 
3965   if (Builtin) {
3966     Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1)));
3967     Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E);
3968     assert(Result && "SISD intrinsic should have been handled");
3969     return Result;
3970   }
3971 
3972   llvm::APSInt Result;
3973   const Expr *Arg = E->getArg(E->getNumArgs()-1);
3974   NeonTypeFlags Type(0);
3975   if (Arg->isIntegerConstantExpr(Result, getContext()))
3976     // Determine the type of this overloaded NEON intrinsic.
3977     Type = NeonTypeFlags(Result.getZExtValue());
3978 
3979   bool usgn = Type.isUnsigned();
3980   bool quad = Type.isQuad();
3981 
3982   // Handle non-overloaded intrinsics first.
3983   switch (BuiltinID) {
3984   default: break;
3985   case NEON::BI__builtin_neon_vldrq_p128: {
3986     llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
3987     Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy);
3988     return Builder.CreateLoad(Ptr);
3989   }
3990   case NEON::BI__builtin_neon_vstrq_p128: {
3991     llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
3992     Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy);
3993     return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr);
3994   }
3995   case NEON::BI__builtin_neon_vcvts_u32_f32:
3996   case NEON::BI__builtin_neon_vcvtd_u64_f64:
3997     usgn = true;
3998     // FALL THROUGH
3999   case NEON::BI__builtin_neon_vcvts_s32_f32:
4000   case NEON::BI__builtin_neon_vcvtd_s64_f64: {
4001     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4002     bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
4003     llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
4004     llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
4005     Ops[0] = Builder.CreateBitCast(Ops[0], FTy);
4006     if (usgn)
4007       return Builder.CreateFPToUI(Ops[0], InTy);
4008     return Builder.CreateFPToSI(Ops[0], InTy);
4009   }
4010   case NEON::BI__builtin_neon_vcvts_f32_u32:
4011   case NEON::BI__builtin_neon_vcvtd_f64_u64:
4012     usgn = true;
4013     // FALL THROUGH
4014   case NEON::BI__builtin_neon_vcvts_f32_s32:
4015   case NEON::BI__builtin_neon_vcvtd_f64_s64: {
4016     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4017     bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
4018     llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
4019     llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
4020     Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
4021     if (usgn)
4022       return Builder.CreateUIToFP(Ops[0], FTy);
4023     return Builder.CreateSIToFP(Ops[0], FTy);
4024   }
4025   case NEON::BI__builtin_neon_vpaddd_s64: {
4026     llvm::Type *Ty =
4027       llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2);
4028     Value *Vec = EmitScalarExpr(E->getArg(0));
4029     // The vector is v2f64, so make sure it's bitcast to that.
4030     Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
4031     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4032     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4033     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4034     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4035     // Pairwise addition of a v2f64 into a scalar f64.
4036     return Builder.CreateAdd(Op0, Op1, "vpaddd");
4037   }
4038   case NEON::BI__builtin_neon_vpaddd_f64: {
4039     llvm::Type *Ty =
4040       llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2);
4041     Value *Vec = EmitScalarExpr(E->getArg(0));
4042     // The vector is v2f64, so make sure it's bitcast to that.
4043     Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
4044     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4045     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4046     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4047     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4048     // Pairwise addition of a v2f64 into a scalar f64.
4049     return Builder.CreateFAdd(Op0, Op1, "vpaddd");
4050   }
4051   case NEON::BI__builtin_neon_vpadds_f32: {
4052     llvm::Type *Ty =
4053       llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2);
4054     Value *Vec = EmitScalarExpr(E->getArg(0));
4055     // The vector is v2f32, so make sure it's bitcast to that.
4056     Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
4057     llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
4058     llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
4059     Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
4060     Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
4061     // Pairwise addition of a v2f32 into a scalar f32.
4062     return Builder.CreateFAdd(Op0, Op1, "vpaddd");
4063   }
4064   case NEON::BI__builtin_neon_vceqzd_s64:
4065   case NEON::BI__builtin_neon_vceqzd_f64:
4066   case NEON::BI__builtin_neon_vceqzs_f32:
4067     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4068     return EmitAArch64CompareBuiltinExpr(
4069         Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ,
4070         ICmpInst::ICMP_EQ, "vceqz");
4071   case NEON::BI__builtin_neon_vcgezd_s64:
4072   case NEON::BI__builtin_neon_vcgezd_f64:
4073   case NEON::BI__builtin_neon_vcgezs_f32:
4074     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4075     return EmitAArch64CompareBuiltinExpr(
4076         Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE,
4077         ICmpInst::ICMP_SGE, "vcgez");
4078   case NEON::BI__builtin_neon_vclezd_s64:
4079   case NEON::BI__builtin_neon_vclezd_f64:
4080   case NEON::BI__builtin_neon_vclezs_f32:
4081     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4082     return EmitAArch64CompareBuiltinExpr(
4083         Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE,
4084         ICmpInst::ICMP_SLE, "vclez");
4085   case NEON::BI__builtin_neon_vcgtzd_s64:
4086   case NEON::BI__builtin_neon_vcgtzd_f64:
4087   case NEON::BI__builtin_neon_vcgtzs_f32:
4088     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4089     return EmitAArch64CompareBuiltinExpr(
4090         Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT,
4091         ICmpInst::ICMP_SGT, "vcgtz");
4092   case NEON::BI__builtin_neon_vcltzd_s64:
4093   case NEON::BI__builtin_neon_vcltzd_f64:
4094   case NEON::BI__builtin_neon_vcltzs_f32:
4095     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4096     return EmitAArch64CompareBuiltinExpr(
4097         Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT,
4098         ICmpInst::ICMP_SLT, "vcltz");
4099 
4100   case NEON::BI__builtin_neon_vceqzd_u64: {
4101     llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext());
4102     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4103     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4104     Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0],
4105                                 llvm::Constant::getNullValue(Ty));
4106     return Builder.CreateSExt(Ops[0], Ty, "vceqzd");
4107   }
4108   case NEON::BI__builtin_neon_vceqd_f64:
4109   case NEON::BI__builtin_neon_vcled_f64:
4110   case NEON::BI__builtin_neon_vcltd_f64:
4111   case NEON::BI__builtin_neon_vcged_f64:
4112   case NEON::BI__builtin_neon_vcgtd_f64: {
4113     llvm::CmpInst::Predicate P;
4114     switch (BuiltinID) {
4115     default: llvm_unreachable("missing builtin ID in switch!");
4116     case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break;
4117     case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break;
4118     case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break;
4119     case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break;
4120     case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break;
4121     }
4122     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4123     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
4124     Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
4125     Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
4126     return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd");
4127   }
4128   case NEON::BI__builtin_neon_vceqs_f32:
4129   case NEON::BI__builtin_neon_vcles_f32:
4130   case NEON::BI__builtin_neon_vclts_f32:
4131   case NEON::BI__builtin_neon_vcges_f32:
4132   case NEON::BI__builtin_neon_vcgts_f32: {
4133     llvm::CmpInst::Predicate P;
4134     switch (BuiltinID) {
4135     default: llvm_unreachable("missing builtin ID in switch!");
4136     case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break;
4137     case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break;
4138     case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break;
4139     case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break;
4140     case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break;
4141     }
4142     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4143     Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
4144     Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy);
4145     Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
4146     return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd");
4147   }
4148   case NEON::BI__builtin_neon_vceqd_s64:
4149   case NEON::BI__builtin_neon_vceqd_u64:
4150   case NEON::BI__builtin_neon_vcgtd_s64:
4151   case NEON::BI__builtin_neon_vcgtd_u64:
4152   case NEON::BI__builtin_neon_vcltd_s64:
4153   case NEON::BI__builtin_neon_vcltd_u64:
4154   case NEON::BI__builtin_neon_vcged_u64:
4155   case NEON::BI__builtin_neon_vcged_s64:
4156   case NEON::BI__builtin_neon_vcled_u64:
4157   case NEON::BI__builtin_neon_vcled_s64: {
4158     llvm::CmpInst::Predicate P;
4159     switch (BuiltinID) {
4160     default: llvm_unreachable("missing builtin ID in switch!");
4161     case NEON::BI__builtin_neon_vceqd_s64:
4162     case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break;
4163     case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break;
4164     case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break;
4165     case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break;
4166     case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break;
4167     case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break;
4168     case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break;
4169     case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break;
4170     case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break;
4171     }
4172     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4173     Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
4174     Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
4175     Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]);
4176     return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd");
4177   }
4178   case NEON::BI__builtin_neon_vtstd_s64:
4179   case NEON::BI__builtin_neon_vtstd_u64: {
4180     llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext());
4181     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4182     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4183     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4184     Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
4185     Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
4186                                 llvm::Constant::getNullValue(Ty));
4187     return Builder.CreateSExt(Ops[0], Ty, "vtstd");
4188   }
4189   case NEON::BI__builtin_neon_vset_lane_i8:
4190   case NEON::BI__builtin_neon_vset_lane_i16:
4191   case NEON::BI__builtin_neon_vset_lane_i32:
4192   case NEON::BI__builtin_neon_vset_lane_i64:
4193   case NEON::BI__builtin_neon_vset_lane_f32:
4194   case NEON::BI__builtin_neon_vsetq_lane_i8:
4195   case NEON::BI__builtin_neon_vsetq_lane_i16:
4196   case NEON::BI__builtin_neon_vsetq_lane_i32:
4197   case NEON::BI__builtin_neon_vsetq_lane_i64:
4198   case NEON::BI__builtin_neon_vsetq_lane_f32:
4199     Ops.push_back(EmitScalarExpr(E->getArg(2)));
4200     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4201   case NEON::BI__builtin_neon_vset_lane_f64:
4202     // The vector type needs a cast for the v1f64 variant.
4203     Ops[1] = Builder.CreateBitCast(Ops[1],
4204                                    llvm::VectorType::get(DoubleTy, 1));
4205     Ops.push_back(EmitScalarExpr(E->getArg(2)));
4206     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4207   case NEON::BI__builtin_neon_vsetq_lane_f64:
4208     // The vector type needs a cast for the v2f64 variant.
4209     Ops[1] = Builder.CreateBitCast(Ops[1],
4210         llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2));
4211     Ops.push_back(EmitScalarExpr(E->getArg(2)));
4212     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
4213 
4214   case NEON::BI__builtin_neon_vget_lane_i8:
4215   case NEON::BI__builtin_neon_vdupb_lane_i8:
4216     Ops[0] = Builder.CreateBitCast(Ops[0],
4217         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8));
4218     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4219                                         "vget_lane");
4220   case NEON::BI__builtin_neon_vgetq_lane_i8:
4221   case NEON::BI__builtin_neon_vdupb_laneq_i8:
4222     Ops[0] = Builder.CreateBitCast(Ops[0],
4223         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16));
4224     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4225                                         "vgetq_lane");
4226   case NEON::BI__builtin_neon_vget_lane_i16:
4227   case NEON::BI__builtin_neon_vduph_lane_i16:
4228     Ops[0] = Builder.CreateBitCast(Ops[0],
4229         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4));
4230     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4231                                         "vget_lane");
4232   case NEON::BI__builtin_neon_vgetq_lane_i16:
4233   case NEON::BI__builtin_neon_vduph_laneq_i16:
4234     Ops[0] = Builder.CreateBitCast(Ops[0],
4235         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8));
4236     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4237                                         "vgetq_lane");
4238   case NEON::BI__builtin_neon_vget_lane_i32:
4239   case NEON::BI__builtin_neon_vdups_lane_i32:
4240     Ops[0] = Builder.CreateBitCast(
4241         Ops[0],
4242         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2));
4243     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4244                                         "vget_lane");
4245   case NEON::BI__builtin_neon_vdups_lane_f32:
4246     Ops[0] = Builder.CreateBitCast(Ops[0],
4247         llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2));
4248     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4249                                         "vdups_lane");
4250   case NEON::BI__builtin_neon_vgetq_lane_i32:
4251   case NEON::BI__builtin_neon_vdups_laneq_i32:
4252     Ops[0] = Builder.CreateBitCast(Ops[0],
4253         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4));
4254     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4255                                         "vgetq_lane");
4256   case NEON::BI__builtin_neon_vget_lane_i64:
4257   case NEON::BI__builtin_neon_vdupd_lane_i64:
4258     Ops[0] = Builder.CreateBitCast(Ops[0],
4259         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1));
4260     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4261                                         "vget_lane");
4262   case NEON::BI__builtin_neon_vdupd_lane_f64:
4263     Ops[0] = Builder.CreateBitCast(Ops[0],
4264         llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1));
4265     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4266                                         "vdupd_lane");
4267   case NEON::BI__builtin_neon_vgetq_lane_i64:
4268   case NEON::BI__builtin_neon_vdupd_laneq_i64:
4269     Ops[0] = Builder.CreateBitCast(Ops[0],
4270         llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2));
4271     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4272                                         "vgetq_lane");
4273   case NEON::BI__builtin_neon_vget_lane_f32:
4274     Ops[0] = Builder.CreateBitCast(Ops[0],
4275         llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2));
4276     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4277                                         "vget_lane");
4278   case NEON::BI__builtin_neon_vget_lane_f64:
4279     Ops[0] = Builder.CreateBitCast(Ops[0],
4280         llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1));
4281     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4282                                         "vget_lane");
4283   case NEON::BI__builtin_neon_vgetq_lane_f32:
4284   case NEON::BI__builtin_neon_vdups_laneq_f32:
4285     Ops[0] = Builder.CreateBitCast(Ops[0],
4286         llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4));
4287     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4288                                         "vgetq_lane");
4289   case NEON::BI__builtin_neon_vgetq_lane_f64:
4290   case NEON::BI__builtin_neon_vdupd_laneq_f64:
4291     Ops[0] = Builder.CreateBitCast(Ops[0],
4292         llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2));
4293     return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
4294                                         "vgetq_lane");
4295   case NEON::BI__builtin_neon_vaddd_s64:
4296   case NEON::BI__builtin_neon_vaddd_u64:
4297     return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd");
4298   case NEON::BI__builtin_neon_vsubd_s64:
4299   case NEON::BI__builtin_neon_vsubd_u64:
4300     return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd");
4301   case NEON::BI__builtin_neon_vqdmlalh_s16:
4302   case NEON::BI__builtin_neon_vqdmlslh_s16: {
4303     SmallVector<Value *, 2> ProductOps;
4304     ProductOps.push_back(vectorWrapScalar16(Ops[1]));
4305     ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2))));
4306     llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
4307     Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
4308                           ProductOps, "vqdmlXl");
4309     Constant *CI = ConstantInt::get(SizeTy, 0);
4310     Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
4311 
4312     unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
4313                                         ? Intrinsic::aarch64_neon_sqadd
4314                                         : Intrinsic::aarch64_neon_sqsub;
4315     return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl");
4316   }
4317   case NEON::BI__builtin_neon_vqshlud_n_s64: {
4318     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4319     Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
4320     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty),
4321                         Ops, "vqshlu_n");
4322   }
4323   case NEON::BI__builtin_neon_vqshld_n_u64:
4324   case NEON::BI__builtin_neon_vqshld_n_s64: {
4325     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64
4326                                    ? Intrinsic::aarch64_neon_uqshl
4327                                    : Intrinsic::aarch64_neon_sqshl;
4328     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4329     Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
4330     return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n");
4331   }
4332   case NEON::BI__builtin_neon_vrshrd_n_u64:
4333   case NEON::BI__builtin_neon_vrshrd_n_s64: {
4334     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64
4335                                    ? Intrinsic::aarch64_neon_urshl
4336                                    : Intrinsic::aarch64_neon_srshl;
4337     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4338     int SV = cast<ConstantInt>(Ops[1])->getSExtValue();
4339     Ops[1] = ConstantInt::get(Int64Ty, -SV);
4340     return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n");
4341   }
4342   case NEON::BI__builtin_neon_vrsrad_n_u64:
4343   case NEON::BI__builtin_neon_vrsrad_n_s64: {
4344     unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64
4345                                    ? Intrinsic::aarch64_neon_urshl
4346                                    : Intrinsic::aarch64_neon_srshl;
4347     Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
4348     Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2))));
4349     Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1],
4350                                  Builder.CreateSExt(Ops[2], Int64Ty));
4351     return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty));
4352   }
4353   case NEON::BI__builtin_neon_vshld_n_s64:
4354   case NEON::BI__builtin_neon_vshld_n_u64: {
4355     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4356     return Builder.CreateShl(
4357         Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n");
4358   }
4359   case NEON::BI__builtin_neon_vshrd_n_s64: {
4360     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4361     return Builder.CreateAShr(
4362         Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
4363                                                    Amt->getZExtValue())),
4364         "shrd_n");
4365   }
4366   case NEON::BI__builtin_neon_vshrd_n_u64: {
4367     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
4368     uint64_t ShiftAmt = Amt->getZExtValue();
4369     // Right-shifting an unsigned value by its size yields 0.
4370     if (ShiftAmt == 64)
4371       return ConstantInt::get(Int64Ty, 0);
4372     return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt),
4373                               "shrd_n");
4374   }
4375   case NEON::BI__builtin_neon_vsrad_n_s64: {
4376     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
4377     Ops[1] = Builder.CreateAShr(
4378         Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
4379                                                    Amt->getZExtValue())),
4380         "shrd_n");
4381     return Builder.CreateAdd(Ops[0], Ops[1]);
4382   }
4383   case NEON::BI__builtin_neon_vsrad_n_u64: {
4384     llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
4385     uint64_t ShiftAmt = Amt->getZExtValue();
4386     // Right-shifting an unsigned value by its size yields 0.
4387     // As Op + 0 = Op, return Ops[0] directly.
4388     if (ShiftAmt == 64)
4389       return Ops[0];
4390     Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt),
4391                                 "shrd_n");
4392     return Builder.CreateAdd(Ops[0], Ops[1]);
4393   }
4394   case NEON::BI__builtin_neon_vqdmlalh_lane_s16:
4395   case NEON::BI__builtin_neon_vqdmlalh_laneq_s16:
4396   case NEON::BI__builtin_neon_vqdmlslh_lane_s16:
4397   case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: {
4398     Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
4399                                           "lane");
4400     SmallVector<Value *, 2> ProductOps;
4401     ProductOps.push_back(vectorWrapScalar16(Ops[1]));
4402     ProductOps.push_back(vectorWrapScalar16(Ops[2]));
4403     llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
4404     Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
4405                           ProductOps, "vqdmlXl");
4406     Constant *CI = ConstantInt::get(SizeTy, 0);
4407     Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
4408     Ops.pop_back();
4409 
4410     unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 ||
4411                        BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16)
4412                           ? Intrinsic::aarch64_neon_sqadd
4413                           : Intrinsic::aarch64_neon_sqsub;
4414     return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl");
4415   }
4416   case NEON::BI__builtin_neon_vqdmlals_s32:
4417   case NEON::BI__builtin_neon_vqdmlsls_s32: {
4418     SmallVector<Value *, 2> ProductOps;
4419     ProductOps.push_back(Ops[1]);
4420     ProductOps.push_back(EmitScalarExpr(E->getArg(2)));
4421     Ops[1] =
4422         EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
4423                      ProductOps, "vqdmlXl");
4424 
4425     unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32
4426                                         ? Intrinsic::aarch64_neon_sqadd
4427                                         : Intrinsic::aarch64_neon_sqsub;
4428     return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl");
4429   }
4430   case NEON::BI__builtin_neon_vqdmlals_lane_s32:
4431   case NEON::BI__builtin_neon_vqdmlals_laneq_s32:
4432   case NEON::BI__builtin_neon_vqdmlsls_lane_s32:
4433   case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: {
4434     Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
4435                                           "lane");
4436     SmallVector<Value *, 2> ProductOps;
4437     ProductOps.push_back(Ops[1]);
4438     ProductOps.push_back(Ops[2]);
4439     Ops[1] =
4440         EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
4441                      ProductOps, "vqdmlXl");
4442     Ops.pop_back();
4443 
4444     unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 ||
4445                        BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32)
4446                           ? Intrinsic::aarch64_neon_sqadd
4447                           : Intrinsic::aarch64_neon_sqsub;
4448     return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl");
4449   }
4450   }
4451 
4452   llvm::VectorType *VTy = GetNeonType(this, Type);
4453   llvm::Type *Ty = VTy;
4454   if (!Ty)
4455     return nullptr;
4456 
4457   // Not all intrinsics handled by the common case work for AArch64 yet, so only
4458   // defer to common code if it's been added to our special map.
4459   Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
4460                                    AArch64SIMDIntrinsicsProvenSorted);
4461 
4462   if (Builtin)
4463     return EmitCommonNeonBuiltinExpr(
4464         Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
4465         Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr);
4466 
4467   if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops))
4468     return V;
4469 
4470   unsigned Int;
4471   switch (BuiltinID) {
4472   default: return nullptr;
4473   case NEON::BI__builtin_neon_vbsl_v:
4474   case NEON::BI__builtin_neon_vbslq_v: {
4475     llvm::Type *BitTy = llvm::VectorType::getInteger(VTy);
4476     Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl");
4477     Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl");
4478     Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl");
4479 
4480     Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl");
4481     Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl");
4482     Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl");
4483     return Builder.CreateBitCast(Ops[0], Ty);
4484   }
4485   case NEON::BI__builtin_neon_vfma_lane_v:
4486   case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types
4487     // The ARM builtins (and instructions) have the addend as the first
4488     // operand, but the 'fma' intrinsics have it last. Swap it around here.
4489     Value *Addend = Ops[0];
4490     Value *Multiplicand = Ops[1];
4491     Value *LaneSource = Ops[2];
4492     Ops[0] = Multiplicand;
4493     Ops[1] = LaneSource;
4494     Ops[2] = Addend;
4495 
4496     // Now adjust things to handle the lane access.
4497     llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ?
4498       llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) :
4499       VTy;
4500     llvm::Constant *cst = cast<Constant>(Ops[3]);
4501     Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst);
4502     Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy);
4503     Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane");
4504 
4505     Ops.pop_back();
4506     Int = Intrinsic::fma;
4507     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla");
4508   }
4509   case NEON::BI__builtin_neon_vfma_laneq_v: {
4510     llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
4511     // v1f64 fma should be mapped to Neon scalar f64 fma
4512     if (VTy && VTy->getElementType() == DoubleTy) {
4513       Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
4514       Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
4515       llvm::Type *VTy = GetNeonType(this,
4516         NeonTypeFlags(NeonTypeFlags::Float64, false, true));
4517       Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
4518       Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
4519       Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy);
4520       Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
4521       return Builder.CreateBitCast(Result, Ty);
4522     }
4523     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4524     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4525     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4526 
4527     llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(),
4528                                             VTy->getNumElements() * 2);
4529     Ops[2] = Builder.CreateBitCast(Ops[2], STy);
4530     Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(),
4531                                                cast<ConstantInt>(Ops[3]));
4532     Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane");
4533 
4534     return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]);
4535   }
4536   case NEON::BI__builtin_neon_vfmaq_laneq_v: {
4537     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4538     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4539     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
4540 
4541     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
4542     Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
4543     return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]);
4544   }
4545   case NEON::BI__builtin_neon_vfmas_lane_f32:
4546   case NEON::BI__builtin_neon_vfmas_laneq_f32:
4547   case NEON::BI__builtin_neon_vfmad_lane_f64:
4548   case NEON::BI__builtin_neon_vfmad_laneq_f64: {
4549     Ops.push_back(EmitScalarExpr(E->getArg(3)));
4550     llvm::Type *Ty = ConvertType(E->getCallReturnType());
4551     Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty);
4552     Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
4553     return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
4554   }
4555   case NEON::BI__builtin_neon_vfms_v:
4556   case NEON::BI__builtin_neon_vfmsq_v: {  // Only used for FP types
4557     // FIXME: probably remove when we no longer support aarch64_simd.h
4558     // (arm_neon.h delegates to vfma).
4559 
4560     // The ARM builtins (and instructions) have the addend as the first
4561     // operand, but the 'fma' intrinsics have it last. Swap it around here.
4562     Value *Subtrahend = Ops[0];
4563     Value *Multiplicand = Ops[2];
4564     Ops[0] = Multiplicand;
4565     Ops[2] = Subtrahend;
4566     Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
4567     Ops[1] = Builder.CreateFNeg(Ops[1]);
4568     Int = Intrinsic::fma;
4569     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls");
4570   }
4571   case NEON::BI__builtin_neon_vmull_v:
4572     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4573     Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull;
4574     if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull;
4575     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
4576   case NEON::BI__builtin_neon_vmax_v:
4577   case NEON::BI__builtin_neon_vmaxq_v:
4578     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4579     Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax;
4580     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax;
4581     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
4582   case NEON::BI__builtin_neon_vmin_v:
4583   case NEON::BI__builtin_neon_vminq_v:
4584     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4585     Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin;
4586     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin;
4587     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
4588   case NEON::BI__builtin_neon_vabd_v:
4589   case NEON::BI__builtin_neon_vabdq_v:
4590     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4591     Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd;
4592     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd;
4593     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
4594   case NEON::BI__builtin_neon_vpadal_v:
4595   case NEON::BI__builtin_neon_vpadalq_v: {
4596     unsigned ArgElts = VTy->getNumElements();
4597     llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType());
4598     unsigned BitWidth = EltTy->getBitWidth();
4599     llvm::Type *ArgTy = llvm::VectorType::get(
4600         llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts);
4601     llvm::Type* Tys[2] = { VTy, ArgTy };
4602     Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp;
4603     SmallVector<llvm::Value*, 1> TmpOps;
4604     TmpOps.push_back(Ops[1]);
4605     Function *F = CGM.getIntrinsic(Int, Tys);
4606     llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal");
4607     llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType());
4608     return Builder.CreateAdd(tmp, addend);
4609   }
4610   case NEON::BI__builtin_neon_vpmin_v:
4611   case NEON::BI__builtin_neon_vpminq_v:
4612     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4613     Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp;
4614     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp;
4615     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
4616   case NEON::BI__builtin_neon_vpmax_v:
4617   case NEON::BI__builtin_neon_vpmaxq_v:
4618     // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
4619     Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp;
4620     if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp;
4621     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
4622   case NEON::BI__builtin_neon_vminnm_v:
4623   case NEON::BI__builtin_neon_vminnmq_v:
4624     Int = Intrinsic::aarch64_neon_fminnm;
4625     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
4626   case NEON::BI__builtin_neon_vmaxnm_v:
4627   case NEON::BI__builtin_neon_vmaxnmq_v:
4628     Int = Intrinsic::aarch64_neon_fmaxnm;
4629     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
4630   case NEON::BI__builtin_neon_vrecpss_f32: {
4631     llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext());
4632     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4633     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type),
4634                         Ops, "vrecps");
4635   }
4636   case NEON::BI__builtin_neon_vrecpsd_f64: {
4637     llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext());
4638     Ops.push_back(EmitScalarExpr(E->getArg(1)));
4639     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type),
4640                         Ops, "vrecps");
4641   }
4642   case NEON::BI__builtin_neon_vrshr_n_v:
4643   case NEON::BI__builtin_neon_vrshrq_n_v:
4644     // FIXME: this can be shared with 32-bit ARM, but not AArch64 at the
4645     // moment. After the final merge it should be added to
4646     // EmitCommonNeonBuiltinExpr.
4647     Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
4648     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true);
4649   case NEON::BI__builtin_neon_vqshlu_n_v:
4650   case NEON::BI__builtin_neon_vqshluq_n_v:
4651     // FIXME: AArch64 and ARM use different intrinsics for this, but are
4652     // essentially compatible. It should be in EmitCommonNeonBuiltinExpr after
4653     // the final merge.
4654     Int = Intrinsic::aarch64_neon_sqshlu;
4655     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 1, false);
4656   case NEON::BI__builtin_neon_vqshrun_n_v:
4657     // FIXME: as above
4658     Int = Intrinsic::aarch64_neon_sqshrun;
4659     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n");
4660   case NEON::BI__builtin_neon_vqrshrun_n_v:
4661     // FIXME: and again.
4662     Int = Intrinsic::aarch64_neon_sqrshrun;
4663     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n");
4664   case NEON::BI__builtin_neon_vqshrn_n_v:
4665     // FIXME: guess
4666     Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn;
4667     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n");
4668   case NEON::BI__builtin_neon_vrshrn_n_v:
4669     // FIXME: there might be a pattern here.
4670     Int = Intrinsic::aarch64_neon_rshrn;
4671     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n");
4672   case NEON::BI__builtin_neon_vqrshrn_n_v:
4673     // FIXME: another one
4674     Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn;
4675     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n");
4676   case NEON::BI__builtin_neon_vrnda_v:
4677   case NEON::BI__builtin_neon_vrndaq_v: {
4678     Int = Intrinsic::round;
4679     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda");
4680   }
4681   case NEON::BI__builtin_neon_vrndi_v:
4682   case NEON::BI__builtin_neon_vrndiq_v: {
4683     Int = Intrinsic::nearbyint;
4684     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi");
4685   }
4686   case NEON::BI__builtin_neon_vrndm_v:
4687   case NEON::BI__builtin_neon_vrndmq_v: {
4688     Int = Intrinsic::floor;
4689     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm");
4690   }
4691   case NEON::BI__builtin_neon_vrndn_v:
4692   case NEON::BI__builtin_neon_vrndnq_v: {
4693     Int = Intrinsic::aarch64_neon_frintn;
4694     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn");
4695   }
4696   case NEON::BI__builtin_neon_vrndp_v:
4697   case NEON::BI__builtin_neon_vrndpq_v: {
4698     Int = Intrinsic::ceil;
4699     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp");
4700   }
4701   case NEON::BI__builtin_neon_vrndx_v:
4702   case NEON::BI__builtin_neon_vrndxq_v: {
4703     Int = Intrinsic::rint;
4704     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx");
4705   }
4706   case NEON::BI__builtin_neon_vrnd_v:
4707   case NEON::BI__builtin_neon_vrndq_v: {
4708     Int = Intrinsic::trunc;
4709     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz");
4710   }
4711   case NEON::BI__builtin_neon_vceqz_v:
4712   case NEON::BI__builtin_neon_vceqzq_v:
4713     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ,
4714                                          ICmpInst::ICMP_EQ, "vceqz");
4715   case NEON::BI__builtin_neon_vcgez_v:
4716   case NEON::BI__builtin_neon_vcgezq_v:
4717     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE,
4718                                          ICmpInst::ICMP_SGE, "vcgez");
4719   case NEON::BI__builtin_neon_vclez_v:
4720   case NEON::BI__builtin_neon_vclezq_v:
4721     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE,
4722                                          ICmpInst::ICMP_SLE, "vclez");
4723   case NEON::BI__builtin_neon_vcgtz_v:
4724   case NEON::BI__builtin_neon_vcgtzq_v:
4725     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT,
4726                                          ICmpInst::ICMP_SGT, "vcgtz");
4727   case NEON::BI__builtin_neon_vcltz_v:
4728   case NEON::BI__builtin_neon_vcltzq_v:
4729     return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT,
4730                                          ICmpInst::ICMP_SLT, "vcltz");
4731   case NEON::BI__builtin_neon_vcvt_f64_v:
4732   case NEON::BI__builtin_neon_vcvtq_f64_v:
4733     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4734     Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad));
4735     return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
4736                 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
4737   case NEON::BI__builtin_neon_vcvt_f64_f32: {
4738     assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&
4739            "unexpected vcvt_f64_f32 builtin");
4740     NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false);
4741     Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
4742 
4743     return Builder.CreateFPExt(Ops[0], Ty, "vcvt");
4744   }
4745   case NEON::BI__builtin_neon_vcvt_f32_f64: {
4746     assert(Type.getEltType() == NeonTypeFlags::Float32 &&
4747            "unexpected vcvt_f32_f64 builtin");
4748     NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true);
4749     Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
4750 
4751     return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt");
4752   }
4753   case NEON::BI__builtin_neon_vcvt_s32_v:
4754   case NEON::BI__builtin_neon_vcvt_u32_v:
4755   case NEON::BI__builtin_neon_vcvt_s64_v:
4756   case NEON::BI__builtin_neon_vcvt_u64_v:
4757   case NEON::BI__builtin_neon_vcvtq_s32_v:
4758   case NEON::BI__builtin_neon_vcvtq_u32_v:
4759   case NEON::BI__builtin_neon_vcvtq_s64_v:
4760   case NEON::BI__builtin_neon_vcvtq_u64_v: {
4761     bool Double =
4762       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
4763     llvm::Type *InTy =
4764       GetNeonType(this,
4765                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
4766                                 : NeonTypeFlags::Float32, false, quad));
4767     Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
4768     if (usgn)
4769       return Builder.CreateFPToUI(Ops[0], Ty);
4770     return Builder.CreateFPToSI(Ops[0], Ty);
4771   }
4772   case NEON::BI__builtin_neon_vcvta_s32_v:
4773   case NEON::BI__builtin_neon_vcvtaq_s32_v:
4774   case NEON::BI__builtin_neon_vcvta_u32_v:
4775   case NEON::BI__builtin_neon_vcvtaq_u32_v:
4776   case NEON::BI__builtin_neon_vcvta_s64_v:
4777   case NEON::BI__builtin_neon_vcvtaq_s64_v:
4778   case NEON::BI__builtin_neon_vcvta_u64_v:
4779   case NEON::BI__builtin_neon_vcvtaq_u64_v: {
4780     Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas;
4781     bool Double =
4782       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
4783     llvm::Type *InTy =
4784       GetNeonType(this,
4785                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
4786                                 : NeonTypeFlags::Float32, false, quad));
4787     llvm::Type *Tys[2] = { Ty, InTy };
4788     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
4789   }
4790   case NEON::BI__builtin_neon_vcvtm_s32_v:
4791   case NEON::BI__builtin_neon_vcvtmq_s32_v:
4792   case NEON::BI__builtin_neon_vcvtm_u32_v:
4793   case NEON::BI__builtin_neon_vcvtmq_u32_v:
4794   case NEON::BI__builtin_neon_vcvtm_s64_v:
4795   case NEON::BI__builtin_neon_vcvtmq_s64_v:
4796   case NEON::BI__builtin_neon_vcvtm_u64_v:
4797   case NEON::BI__builtin_neon_vcvtmq_u64_v: {
4798     Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms;
4799     bool Double =
4800       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
4801     llvm::Type *InTy =
4802       GetNeonType(this,
4803                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
4804                                 : NeonTypeFlags::Float32, false, quad));
4805     llvm::Type *Tys[2] = { Ty, InTy };
4806     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
4807   }
4808   case NEON::BI__builtin_neon_vcvtn_s32_v:
4809   case NEON::BI__builtin_neon_vcvtnq_s32_v:
4810   case NEON::BI__builtin_neon_vcvtn_u32_v:
4811   case NEON::BI__builtin_neon_vcvtnq_u32_v:
4812   case NEON::BI__builtin_neon_vcvtn_s64_v:
4813   case NEON::BI__builtin_neon_vcvtnq_s64_v:
4814   case NEON::BI__builtin_neon_vcvtn_u64_v:
4815   case NEON::BI__builtin_neon_vcvtnq_u64_v: {
4816     Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns;
4817     bool Double =
4818       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
4819     llvm::Type *InTy =
4820       GetNeonType(this,
4821                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
4822                                 : NeonTypeFlags::Float32, false, quad));
4823     llvm::Type *Tys[2] = { Ty, InTy };
4824     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
4825   }
4826   case NEON::BI__builtin_neon_vcvtp_s32_v:
4827   case NEON::BI__builtin_neon_vcvtpq_s32_v:
4828   case NEON::BI__builtin_neon_vcvtp_u32_v:
4829   case NEON::BI__builtin_neon_vcvtpq_u32_v:
4830   case NEON::BI__builtin_neon_vcvtp_s64_v:
4831   case NEON::BI__builtin_neon_vcvtpq_s64_v:
4832   case NEON::BI__builtin_neon_vcvtp_u64_v:
4833   case NEON::BI__builtin_neon_vcvtpq_u64_v: {
4834     Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps;
4835     bool Double =
4836       (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64);
4837     llvm::Type *InTy =
4838       GetNeonType(this,
4839                   NeonTypeFlags(Double ? NeonTypeFlags::Float64
4840                                 : NeonTypeFlags::Float32, false, quad));
4841     llvm::Type *Tys[2] = { Ty, InTy };
4842     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp");
4843   }
4844   case NEON::BI__builtin_neon_vmulx_v:
4845   case NEON::BI__builtin_neon_vmulxq_v: {
4846     Int = Intrinsic::aarch64_neon_fmulx;
4847     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
4848   }
4849   case NEON::BI__builtin_neon_vmul_lane_v:
4850   case NEON::BI__builtin_neon_vmul_laneq_v: {
4851     // v1f64 vmul_lane should be mapped to Neon scalar mul lane
4852     bool Quad = false;
4853     if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v)
4854       Quad = true;
4855     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
4856     llvm::Type *VTy = GetNeonType(this,
4857       NeonTypeFlags(NeonTypeFlags::Float64, false, Quad));
4858     Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
4859     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
4860     Value *Result = Builder.CreateFMul(Ops[0], Ops[1]);
4861     return Builder.CreateBitCast(Result, Ty);
4862   }
4863   case NEON::BI__builtin_neon_vnegd_s64:
4864     return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd");
4865   case NEON::BI__builtin_neon_vpmaxnm_v:
4866   case NEON::BI__builtin_neon_vpmaxnmq_v: {
4867     Int = Intrinsic::aarch64_neon_fmaxnmp;
4868     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
4869   }
4870   case NEON::BI__builtin_neon_vpminnm_v:
4871   case NEON::BI__builtin_neon_vpminnmq_v: {
4872     Int = Intrinsic::aarch64_neon_fminnmp;
4873     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
4874   }
4875   case NEON::BI__builtin_neon_vsqrt_v:
4876   case NEON::BI__builtin_neon_vsqrtq_v: {
4877     Int = Intrinsic::sqrt;
4878     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
4879     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt");
4880   }
4881   case NEON::BI__builtin_neon_vrbit_v:
4882   case NEON::BI__builtin_neon_vrbitq_v: {
4883     Int = Intrinsic::aarch64_neon_rbit;
4884     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit");
4885   }
4886   case NEON::BI__builtin_neon_vaddv_u8:
4887     // FIXME: These are handled by the AArch64 scalar code.
4888     usgn = true;
4889     // FALLTHROUGH
4890   case NEON::BI__builtin_neon_vaddv_s8: {
4891     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
4892     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4893     VTy =
4894       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
4895     llvm::Type *Tys[2] = { Ty, VTy };
4896     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4897     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
4898     return Builder.CreateTrunc(Ops[0],
4899              llvm::IntegerType::get(getLLVMContext(), 8));
4900   }
4901   case NEON::BI__builtin_neon_vaddv_u16:
4902     usgn = true;
4903     // FALLTHROUGH
4904   case NEON::BI__builtin_neon_vaddv_s16: {
4905     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
4906     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4907     VTy =
4908       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
4909     llvm::Type *Tys[2] = { Ty, VTy };
4910     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4911     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
4912     return Builder.CreateTrunc(Ops[0],
4913              llvm::IntegerType::get(getLLVMContext(), 16));
4914   }
4915   case NEON::BI__builtin_neon_vaddvq_u8:
4916     usgn = true;
4917     // FALLTHROUGH
4918   case NEON::BI__builtin_neon_vaddvq_s8: {
4919     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
4920     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4921     VTy =
4922       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
4923     llvm::Type *Tys[2] = { Ty, VTy };
4924     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4925     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
4926     return Builder.CreateTrunc(Ops[0],
4927              llvm::IntegerType::get(getLLVMContext(), 8));
4928   }
4929   case NEON::BI__builtin_neon_vaddvq_u16:
4930     usgn = true;
4931     // FALLTHROUGH
4932   case NEON::BI__builtin_neon_vaddvq_s16: {
4933     Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
4934     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4935     VTy =
4936       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
4937     llvm::Type *Tys[2] = { Ty, VTy };
4938     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4939     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
4940     return Builder.CreateTrunc(Ops[0],
4941              llvm::IntegerType::get(getLLVMContext(), 16));
4942   }
4943   case NEON::BI__builtin_neon_vmaxv_u8: {
4944     Int = Intrinsic::aarch64_neon_umaxv;
4945     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4946     VTy =
4947       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
4948     llvm::Type *Tys[2] = { Ty, VTy };
4949     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4950     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
4951     return Builder.CreateTrunc(Ops[0],
4952              llvm::IntegerType::get(getLLVMContext(), 8));
4953   }
4954   case NEON::BI__builtin_neon_vmaxv_u16: {
4955     Int = Intrinsic::aarch64_neon_umaxv;
4956     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4957     VTy =
4958       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
4959     llvm::Type *Tys[2] = { Ty, VTy };
4960     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4961     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
4962     return Builder.CreateTrunc(Ops[0],
4963              llvm::IntegerType::get(getLLVMContext(), 16));
4964   }
4965   case NEON::BI__builtin_neon_vmaxvq_u8: {
4966     Int = Intrinsic::aarch64_neon_umaxv;
4967     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4968     VTy =
4969       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
4970     llvm::Type *Tys[2] = { Ty, VTy };
4971     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4972     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
4973     return Builder.CreateTrunc(Ops[0],
4974              llvm::IntegerType::get(getLLVMContext(), 8));
4975   }
4976   case NEON::BI__builtin_neon_vmaxvq_u16: {
4977     Int = Intrinsic::aarch64_neon_umaxv;
4978     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4979     VTy =
4980       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
4981     llvm::Type *Tys[2] = { Ty, VTy };
4982     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4983     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
4984     return Builder.CreateTrunc(Ops[0],
4985              llvm::IntegerType::get(getLLVMContext(), 16));
4986   }
4987   case NEON::BI__builtin_neon_vmaxv_s8: {
4988     Int = Intrinsic::aarch64_neon_smaxv;
4989     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
4990     VTy =
4991       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
4992     llvm::Type *Tys[2] = { Ty, VTy };
4993     Ops.push_back(EmitScalarExpr(E->getArg(0)));
4994     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
4995     return Builder.CreateTrunc(Ops[0],
4996              llvm::IntegerType::get(getLLVMContext(), 8));
4997   }
4998   case NEON::BI__builtin_neon_vmaxv_s16: {
4999     Int = Intrinsic::aarch64_neon_smaxv;
5000     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5001     VTy =
5002       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5003     llvm::Type *Tys[2] = { Ty, VTy };
5004     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5005     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5006     return Builder.CreateTrunc(Ops[0],
5007              llvm::IntegerType::get(getLLVMContext(), 16));
5008   }
5009   case NEON::BI__builtin_neon_vmaxvq_s8: {
5010     Int = Intrinsic::aarch64_neon_smaxv;
5011     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5012     VTy =
5013       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5014     llvm::Type *Tys[2] = { Ty, VTy };
5015     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5016     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5017     return Builder.CreateTrunc(Ops[0],
5018              llvm::IntegerType::get(getLLVMContext(), 8));
5019   }
5020   case NEON::BI__builtin_neon_vmaxvq_s16: {
5021     Int = Intrinsic::aarch64_neon_smaxv;
5022     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5023     VTy =
5024       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5025     llvm::Type *Tys[2] = { Ty, VTy };
5026     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5027     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
5028     return Builder.CreateTrunc(Ops[0],
5029              llvm::IntegerType::get(getLLVMContext(), 16));
5030   }
5031   case NEON::BI__builtin_neon_vminv_u8: {
5032     Int = Intrinsic::aarch64_neon_uminv;
5033     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5034     VTy =
5035       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5036     llvm::Type *Tys[2] = { Ty, VTy };
5037     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5038     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5039     return Builder.CreateTrunc(Ops[0],
5040              llvm::IntegerType::get(getLLVMContext(), 8));
5041   }
5042   case NEON::BI__builtin_neon_vminv_u16: {
5043     Int = Intrinsic::aarch64_neon_uminv;
5044     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5045     VTy =
5046       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5047     llvm::Type *Tys[2] = { Ty, VTy };
5048     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5049     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5050     return Builder.CreateTrunc(Ops[0],
5051              llvm::IntegerType::get(getLLVMContext(), 16));
5052   }
5053   case NEON::BI__builtin_neon_vminvq_u8: {
5054     Int = Intrinsic::aarch64_neon_uminv;
5055     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5056     VTy =
5057       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5058     llvm::Type *Tys[2] = { Ty, VTy };
5059     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5060     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5061     return Builder.CreateTrunc(Ops[0],
5062              llvm::IntegerType::get(getLLVMContext(), 8));
5063   }
5064   case NEON::BI__builtin_neon_vminvq_u16: {
5065     Int = Intrinsic::aarch64_neon_uminv;
5066     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5067     VTy =
5068       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5069     llvm::Type *Tys[2] = { Ty, VTy };
5070     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5071     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5072     return Builder.CreateTrunc(Ops[0],
5073              llvm::IntegerType::get(getLLVMContext(), 16));
5074   }
5075   case NEON::BI__builtin_neon_vminv_s8: {
5076     Int = Intrinsic::aarch64_neon_sminv;
5077     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5078     VTy =
5079       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5080     llvm::Type *Tys[2] = { Ty, VTy };
5081     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5082     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5083     return Builder.CreateTrunc(Ops[0],
5084              llvm::IntegerType::get(getLLVMContext(), 8));
5085   }
5086   case NEON::BI__builtin_neon_vminv_s16: {
5087     Int = Intrinsic::aarch64_neon_sminv;
5088     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5089     VTy =
5090       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5091     llvm::Type *Tys[2] = { Ty, VTy };
5092     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5093     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5094     return Builder.CreateTrunc(Ops[0],
5095              llvm::IntegerType::get(getLLVMContext(), 16));
5096   }
5097   case NEON::BI__builtin_neon_vminvq_s8: {
5098     Int = Intrinsic::aarch64_neon_sminv;
5099     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5100     VTy =
5101       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5102     llvm::Type *Tys[2] = { Ty, VTy };
5103     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5104     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5105     return Builder.CreateTrunc(Ops[0],
5106              llvm::IntegerType::get(getLLVMContext(), 8));
5107   }
5108   case NEON::BI__builtin_neon_vminvq_s16: {
5109     Int = Intrinsic::aarch64_neon_sminv;
5110     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5111     VTy =
5112       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5113     llvm::Type *Tys[2] = { Ty, VTy };
5114     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5115     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
5116     return Builder.CreateTrunc(Ops[0],
5117              llvm::IntegerType::get(getLLVMContext(), 16));
5118   }
5119   case NEON::BI__builtin_neon_vmul_n_f64: {
5120     Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
5121     Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
5122     return Builder.CreateFMul(Ops[0], RHS);
5123   }
5124   case NEON::BI__builtin_neon_vaddlv_u8: {
5125     Int = Intrinsic::aarch64_neon_uaddlv;
5126     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5127     VTy =
5128       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5129     llvm::Type *Tys[2] = { Ty, VTy };
5130     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5131     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5132     return Builder.CreateTrunc(Ops[0],
5133              llvm::IntegerType::get(getLLVMContext(), 16));
5134   }
5135   case NEON::BI__builtin_neon_vaddlv_u16: {
5136     Int = Intrinsic::aarch64_neon_uaddlv;
5137     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5138     VTy =
5139       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5140     llvm::Type *Tys[2] = { Ty, VTy };
5141     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5142     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5143   }
5144   case NEON::BI__builtin_neon_vaddlvq_u8: {
5145     Int = Intrinsic::aarch64_neon_uaddlv;
5146     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5147     VTy =
5148       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5149     llvm::Type *Tys[2] = { Ty, VTy };
5150     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5151     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5152     return Builder.CreateTrunc(Ops[0],
5153              llvm::IntegerType::get(getLLVMContext(), 16));
5154   }
5155   case NEON::BI__builtin_neon_vaddlvq_u16: {
5156     Int = Intrinsic::aarch64_neon_uaddlv;
5157     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5158     VTy =
5159       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5160     llvm::Type *Tys[2] = { Ty, VTy };
5161     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5162     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5163   }
5164   case NEON::BI__builtin_neon_vaddlv_s8: {
5165     Int = Intrinsic::aarch64_neon_saddlv;
5166     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5167     VTy =
5168       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8);
5169     llvm::Type *Tys[2] = { Ty, VTy };
5170     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5171     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5172     return Builder.CreateTrunc(Ops[0],
5173              llvm::IntegerType::get(getLLVMContext(), 16));
5174   }
5175   case NEON::BI__builtin_neon_vaddlv_s16: {
5176     Int = Intrinsic::aarch64_neon_saddlv;
5177     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5178     VTy =
5179       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4);
5180     llvm::Type *Tys[2] = { Ty, VTy };
5181     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5182     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5183   }
5184   case NEON::BI__builtin_neon_vaddlvq_s8: {
5185     Int = Intrinsic::aarch64_neon_saddlv;
5186     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5187     VTy =
5188       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16);
5189     llvm::Type *Tys[2] = { Ty, VTy };
5190     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5191     Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5192     return Builder.CreateTrunc(Ops[0],
5193              llvm::IntegerType::get(getLLVMContext(), 16));
5194   }
5195   case NEON::BI__builtin_neon_vaddlvq_s16: {
5196     Int = Intrinsic::aarch64_neon_saddlv;
5197     Ty = llvm::IntegerType::get(getLLVMContext(), 32);
5198     VTy =
5199       llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8);
5200     llvm::Type *Tys[2] = { Ty, VTy };
5201     Ops.push_back(EmitScalarExpr(E->getArg(0)));
5202     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
5203   }
5204   case NEON::BI__builtin_neon_vsri_n_v:
5205   case NEON::BI__builtin_neon_vsriq_n_v: {
5206     Int = Intrinsic::aarch64_neon_vsri;
5207     llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
5208     return EmitNeonCall(Intrin, Ops, "vsri_n");
5209   }
5210   case NEON::BI__builtin_neon_vsli_n_v:
5211   case NEON::BI__builtin_neon_vsliq_n_v: {
5212     Int = Intrinsic::aarch64_neon_vsli;
5213     llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
5214     return EmitNeonCall(Intrin, Ops, "vsli_n");
5215   }
5216   case NEON::BI__builtin_neon_vsra_n_v:
5217   case NEON::BI__builtin_neon_vsraq_n_v:
5218     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5219     Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
5220     return Builder.CreateAdd(Ops[0], Ops[1]);
5221   case NEON::BI__builtin_neon_vrsra_n_v:
5222   case NEON::BI__builtin_neon_vrsraq_n_v: {
5223     Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
5224     SmallVector<llvm::Value*,2> TmpOps;
5225     TmpOps.push_back(Ops[1]);
5226     TmpOps.push_back(Ops[2]);
5227     Function* F = CGM.getIntrinsic(Int, Ty);
5228     llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true);
5229     Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
5230     return Builder.CreateAdd(Ops[0], tmp);
5231   }
5232     // FIXME: Sharing loads & stores with 32-bit is complicated by the absence
5233     // of an Align parameter here.
5234   case NEON::BI__builtin_neon_vld1_x2_v:
5235   case NEON::BI__builtin_neon_vld1q_x2_v:
5236   case NEON::BI__builtin_neon_vld1_x3_v:
5237   case NEON::BI__builtin_neon_vld1q_x3_v:
5238   case NEON::BI__builtin_neon_vld1_x4_v:
5239   case NEON::BI__builtin_neon_vld1q_x4_v: {
5240     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5241     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5242     llvm::Type *Tys[2] = { VTy, PTy };
5243     unsigned Int;
5244     switch (BuiltinID) {
5245     case NEON::BI__builtin_neon_vld1_x2_v:
5246     case NEON::BI__builtin_neon_vld1q_x2_v:
5247       Int = Intrinsic::aarch64_neon_ld1x2;
5248       break;
5249     case NEON::BI__builtin_neon_vld1_x3_v:
5250     case NEON::BI__builtin_neon_vld1q_x3_v:
5251       Int = Intrinsic::aarch64_neon_ld1x3;
5252       break;
5253     case NEON::BI__builtin_neon_vld1_x4_v:
5254     case NEON::BI__builtin_neon_vld1q_x4_v:
5255       Int = Intrinsic::aarch64_neon_ld1x4;
5256       break;
5257     }
5258     Function *F = CGM.getIntrinsic(Int, Tys);
5259     Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN");
5260     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5261     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5262     return Builder.CreateStore(Ops[1], Ops[0]);
5263   }
5264   case NEON::BI__builtin_neon_vst1_x2_v:
5265   case NEON::BI__builtin_neon_vst1q_x2_v:
5266   case NEON::BI__builtin_neon_vst1_x3_v:
5267   case NEON::BI__builtin_neon_vst1q_x3_v:
5268   case NEON::BI__builtin_neon_vst1_x4_v:
5269   case NEON::BI__builtin_neon_vst1q_x4_v: {
5270     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType());
5271     llvm::Type *Tys[2] = { VTy, PTy };
5272     unsigned Int;
5273     switch (BuiltinID) {
5274     case NEON::BI__builtin_neon_vst1_x2_v:
5275     case NEON::BI__builtin_neon_vst1q_x2_v:
5276       Int = Intrinsic::aarch64_neon_st1x2;
5277       break;
5278     case NEON::BI__builtin_neon_vst1_x3_v:
5279     case NEON::BI__builtin_neon_vst1q_x3_v:
5280       Int = Intrinsic::aarch64_neon_st1x3;
5281       break;
5282     case NEON::BI__builtin_neon_vst1_x4_v:
5283     case NEON::BI__builtin_neon_vst1q_x4_v:
5284       Int = Intrinsic::aarch64_neon_st1x4;
5285       break;
5286     }
5287     SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end());
5288     IntOps.push_back(Ops[0]);
5289     return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, "");
5290   }
5291   case NEON::BI__builtin_neon_vld1_v:
5292   case NEON::BI__builtin_neon_vld1q_v:
5293     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
5294     return Builder.CreateLoad(Ops[0]);
5295   case NEON::BI__builtin_neon_vst1_v:
5296   case NEON::BI__builtin_neon_vst1q_v:
5297     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
5298     Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
5299     return Builder.CreateStore(Ops[1], Ops[0]);
5300   case NEON::BI__builtin_neon_vld1_lane_v:
5301   case NEON::BI__builtin_neon_vld1q_lane_v:
5302     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5303     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
5304     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5305     Ops[0] = Builder.CreateLoad(Ops[0]);
5306     return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
5307   case NEON::BI__builtin_neon_vld1_dup_v:
5308   case NEON::BI__builtin_neon_vld1q_dup_v: {
5309     Value *V = UndefValue::get(Ty);
5310     Ty = llvm::PointerType::getUnqual(VTy->getElementType());
5311     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5312     Ops[0] = Builder.CreateLoad(Ops[0]);
5313     llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
5314     Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
5315     return EmitNeonSplat(Ops[0], CI);
5316   }
5317   case NEON::BI__builtin_neon_vst1_lane_v:
5318   case NEON::BI__builtin_neon_vst1q_lane_v:
5319     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5320     Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
5321     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5322     return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty));
5323   case NEON::BI__builtin_neon_vld2_v:
5324   case NEON::BI__builtin_neon_vld2q_v: {
5325     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5326     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5327     llvm::Type *Tys[2] = { VTy, PTy };
5328     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys);
5329     Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
5330     Ops[0] = Builder.CreateBitCast(Ops[0],
5331                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5332     return Builder.CreateStore(Ops[1], Ops[0]);
5333   }
5334   case NEON::BI__builtin_neon_vld3_v:
5335   case NEON::BI__builtin_neon_vld3q_v: {
5336     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5337     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5338     llvm::Type *Tys[2] = { VTy, PTy };
5339     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys);
5340     Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
5341     Ops[0] = Builder.CreateBitCast(Ops[0],
5342                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5343     return Builder.CreateStore(Ops[1], Ops[0]);
5344   }
5345   case NEON::BI__builtin_neon_vld4_v:
5346   case NEON::BI__builtin_neon_vld4q_v: {
5347     llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
5348     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5349     llvm::Type *Tys[2] = { VTy, PTy };
5350     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys);
5351     Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
5352     Ops[0] = Builder.CreateBitCast(Ops[0],
5353                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5354     return Builder.CreateStore(Ops[1], Ops[0]);
5355   }
5356   case NEON::BI__builtin_neon_vld2_dup_v:
5357   case NEON::BI__builtin_neon_vld2q_dup_v: {
5358     llvm::Type *PTy =
5359       llvm::PointerType::getUnqual(VTy->getElementType());
5360     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5361     llvm::Type *Tys[2] = { VTy, PTy };
5362     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys);
5363     Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
5364     Ops[0] = Builder.CreateBitCast(Ops[0],
5365                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5366     return Builder.CreateStore(Ops[1], Ops[0]);
5367   }
5368   case NEON::BI__builtin_neon_vld3_dup_v:
5369   case NEON::BI__builtin_neon_vld3q_dup_v: {
5370     llvm::Type *PTy =
5371       llvm::PointerType::getUnqual(VTy->getElementType());
5372     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5373     llvm::Type *Tys[2] = { VTy, PTy };
5374     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys);
5375     Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
5376     Ops[0] = Builder.CreateBitCast(Ops[0],
5377                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5378     return Builder.CreateStore(Ops[1], Ops[0]);
5379   }
5380   case NEON::BI__builtin_neon_vld4_dup_v:
5381   case NEON::BI__builtin_neon_vld4q_dup_v: {
5382     llvm::Type *PTy =
5383       llvm::PointerType::getUnqual(VTy->getElementType());
5384     Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
5385     llvm::Type *Tys[2] = { VTy, PTy };
5386     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys);
5387     Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
5388     Ops[0] = Builder.CreateBitCast(Ops[0],
5389                 llvm::PointerType::getUnqual(Ops[1]->getType()));
5390     return Builder.CreateStore(Ops[1], Ops[0]);
5391   }
5392   case NEON::BI__builtin_neon_vld2_lane_v:
5393   case NEON::BI__builtin_neon_vld2q_lane_v: {
5394     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5395     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys);
5396     Ops.push_back(Ops[1]);
5397     Ops.erase(Ops.begin()+1);
5398     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5399     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5400     Ops[3] = Builder.CreateZExt(Ops[3],
5401                 llvm::IntegerType::get(getLLVMContext(), 64));
5402     Ops[1] = Builder.CreateCall(F,
5403                 ArrayRef<Value*>(Ops).slice(1), "vld2_lane");
5404     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5405     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5406     return Builder.CreateStore(Ops[1], Ops[0]);
5407   }
5408   case NEON::BI__builtin_neon_vld3_lane_v:
5409   case NEON::BI__builtin_neon_vld3q_lane_v: {
5410     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5411     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys);
5412     Ops.push_back(Ops[1]);
5413     Ops.erase(Ops.begin()+1);
5414     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5415     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5416     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
5417     Ops[4] = Builder.CreateZExt(Ops[4],
5418                 llvm::IntegerType::get(getLLVMContext(), 64));
5419     Ops[1] = Builder.CreateCall(F,
5420                 ArrayRef<Value*>(Ops).slice(1), "vld3_lane");
5421     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5422     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5423     return Builder.CreateStore(Ops[1], Ops[0]);
5424   }
5425   case NEON::BI__builtin_neon_vld4_lane_v:
5426   case NEON::BI__builtin_neon_vld4q_lane_v: {
5427     llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
5428     Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys);
5429     Ops.push_back(Ops[1]);
5430     Ops.erase(Ops.begin()+1);
5431     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5432     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5433     Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
5434     Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
5435     Ops[5] = Builder.CreateZExt(Ops[5],
5436                 llvm::IntegerType::get(getLLVMContext(), 64));
5437     Ops[1] = Builder.CreateCall(F,
5438                 ArrayRef<Value*>(Ops).slice(1), "vld4_lane");
5439     Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
5440     Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
5441     return Builder.CreateStore(Ops[1], Ops[0]);
5442   }
5443   case NEON::BI__builtin_neon_vst2_v:
5444   case NEON::BI__builtin_neon_vst2q_v: {
5445     Ops.push_back(Ops[0]);
5446     Ops.erase(Ops.begin());
5447     llvm::Type *Tys[2] = { VTy, Ops[2]->getType() };
5448     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys),
5449                         Ops, "");
5450   }
5451   case NEON::BI__builtin_neon_vst2_lane_v:
5452   case NEON::BI__builtin_neon_vst2q_lane_v: {
5453     Ops.push_back(Ops[0]);
5454     Ops.erase(Ops.begin());
5455     Ops[2] = Builder.CreateZExt(Ops[2],
5456                 llvm::IntegerType::get(getLLVMContext(), 64));
5457     llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
5458     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys),
5459                         Ops, "");
5460   }
5461   case NEON::BI__builtin_neon_vst3_v:
5462   case NEON::BI__builtin_neon_vst3q_v: {
5463     Ops.push_back(Ops[0]);
5464     Ops.erase(Ops.begin());
5465     llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
5466     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys),
5467                         Ops, "");
5468   }
5469   case NEON::BI__builtin_neon_vst3_lane_v:
5470   case NEON::BI__builtin_neon_vst3q_lane_v: {
5471     Ops.push_back(Ops[0]);
5472     Ops.erase(Ops.begin());
5473     Ops[3] = Builder.CreateZExt(Ops[3],
5474                 llvm::IntegerType::get(getLLVMContext(), 64));
5475     llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
5476     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys),
5477                         Ops, "");
5478   }
5479   case NEON::BI__builtin_neon_vst4_v:
5480   case NEON::BI__builtin_neon_vst4q_v: {
5481     Ops.push_back(Ops[0]);
5482     Ops.erase(Ops.begin());
5483     llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
5484     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys),
5485                         Ops, "");
5486   }
5487   case NEON::BI__builtin_neon_vst4_lane_v:
5488   case NEON::BI__builtin_neon_vst4q_lane_v: {
5489     Ops.push_back(Ops[0]);
5490     Ops.erase(Ops.begin());
5491     Ops[4] = Builder.CreateZExt(Ops[4],
5492                 llvm::IntegerType::get(getLLVMContext(), 64));
5493     llvm::Type *Tys[2] = { VTy, Ops[5]->getType() };
5494     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys),
5495                         Ops, "");
5496   }
5497   case NEON::BI__builtin_neon_vtrn_v:
5498   case NEON::BI__builtin_neon_vtrnq_v: {
5499     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5500     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5501     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5502     Value *SV = nullptr;
5503 
5504     for (unsigned vi = 0; vi != 2; ++vi) {
5505       SmallVector<Constant*, 16> Indices;
5506       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5507         Indices.push_back(ConstantInt::get(Int32Ty, i+vi));
5508         Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi));
5509       }
5510       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
5511       SV = llvm::ConstantVector::get(Indices);
5512       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
5513       SV = Builder.CreateStore(SV, Addr);
5514     }
5515     return SV;
5516   }
5517   case NEON::BI__builtin_neon_vuzp_v:
5518   case NEON::BI__builtin_neon_vuzpq_v: {
5519     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5520     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5521     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5522     Value *SV = nullptr;
5523 
5524     for (unsigned vi = 0; vi != 2; ++vi) {
5525       SmallVector<Constant*, 16> Indices;
5526       for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
5527         Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
5528 
5529       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
5530       SV = llvm::ConstantVector::get(Indices);
5531       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
5532       SV = Builder.CreateStore(SV, Addr);
5533     }
5534     return SV;
5535   }
5536   case NEON::BI__builtin_neon_vzip_v:
5537   case NEON::BI__builtin_neon_vzipq_v: {
5538     Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
5539     Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
5540     Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
5541     Value *SV = nullptr;
5542 
5543     for (unsigned vi = 0; vi != 2; ++vi) {
5544       SmallVector<Constant*, 16> Indices;
5545       for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
5546         Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
5547         Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
5548       }
5549       Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
5550       SV = llvm::ConstantVector::get(Indices);
5551       SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
5552       SV = Builder.CreateStore(SV, Addr);
5553     }
5554     return SV;
5555   }
5556   case NEON::BI__builtin_neon_vqtbl1q_v: {
5557     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty),
5558                         Ops, "vtbl1");
5559   }
5560   case NEON::BI__builtin_neon_vqtbl2q_v: {
5561     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty),
5562                         Ops, "vtbl2");
5563   }
5564   case NEON::BI__builtin_neon_vqtbl3q_v: {
5565     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty),
5566                         Ops, "vtbl3");
5567   }
5568   case NEON::BI__builtin_neon_vqtbl4q_v: {
5569     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty),
5570                         Ops, "vtbl4");
5571   }
5572   case NEON::BI__builtin_neon_vqtbx1q_v: {
5573     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty),
5574                         Ops, "vtbx1");
5575   }
5576   case NEON::BI__builtin_neon_vqtbx2q_v: {
5577     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty),
5578                         Ops, "vtbx2");
5579   }
5580   case NEON::BI__builtin_neon_vqtbx3q_v: {
5581     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty),
5582                         Ops, "vtbx3");
5583   }
5584   case NEON::BI__builtin_neon_vqtbx4q_v: {
5585     return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty),
5586                         Ops, "vtbx4");
5587   }
5588   case NEON::BI__builtin_neon_vsqadd_v:
5589   case NEON::BI__builtin_neon_vsqaddq_v: {
5590     Int = Intrinsic::aarch64_neon_usqadd;
5591     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd");
5592   }
5593   case NEON::BI__builtin_neon_vuqadd_v:
5594   case NEON::BI__builtin_neon_vuqaddq_v: {
5595     Int = Intrinsic::aarch64_neon_suqadd;
5596     return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd");
5597   }
5598   }
5599 }
5600 
5601 llvm::Value *CodeGenFunction::
BuildVector(ArrayRef<llvm::Value * > Ops)5602 BuildVector(ArrayRef<llvm::Value*> Ops) {
5603   assert((Ops.size() & (Ops.size() - 1)) == 0 &&
5604          "Not a power-of-two sized vector!");
5605   bool AllConstants = true;
5606   for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
5607     AllConstants &= isa<Constant>(Ops[i]);
5608 
5609   // If this is a constant vector, create a ConstantVector.
5610   if (AllConstants) {
5611     SmallVector<llvm::Constant*, 16> CstOps;
5612     for (unsigned i = 0, e = Ops.size(); i != e; ++i)
5613       CstOps.push_back(cast<Constant>(Ops[i]));
5614     return llvm::ConstantVector::get(CstOps);
5615   }
5616 
5617   // Otherwise, insertelement the values to build the vector.
5618   Value *Result =
5619     llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size()));
5620 
5621   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
5622     Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
5623 
5624   return Result;
5625 }
5626 
EmitX86BuiltinExpr(unsigned BuiltinID,const CallExpr * E)5627 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
5628                                            const CallExpr *E) {
5629   SmallVector<Value*, 4> Ops;
5630 
5631   // Find out if any arguments are required to be integer constant expressions.
5632   unsigned ICEArguments = 0;
5633   ASTContext::GetBuiltinTypeError Error;
5634   getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
5635   assert(Error == ASTContext::GE_None && "Should not codegen an error");
5636 
5637   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
5638     // If this is a normal argument, just emit it as a scalar.
5639     if ((ICEArguments & (1 << i)) == 0) {
5640       Ops.push_back(EmitScalarExpr(E->getArg(i)));
5641       continue;
5642     }
5643 
5644     // If this is required to be a constant, constant fold it so that we know
5645     // that the generated intrinsic gets a ConstantInt.
5646     llvm::APSInt Result;
5647     bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext());
5648     assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst;
5649     Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result));
5650   }
5651 
5652   switch (BuiltinID) {
5653   default: return nullptr;
5654   case X86::BI_mm_prefetch: {
5655     Value *Address = EmitScalarExpr(E->getArg(0));
5656     Value *RW = ConstantInt::get(Int32Ty, 0);
5657     Value *Locality = EmitScalarExpr(E->getArg(1));
5658     Value *Data = ConstantInt::get(Int32Ty, 1);
5659     Value *F = CGM.getIntrinsic(Intrinsic::prefetch);
5660     return Builder.CreateCall4(F, Address, RW, Locality, Data);
5661   }
5662   case X86::BI__builtin_ia32_vec_init_v8qi:
5663   case X86::BI__builtin_ia32_vec_init_v4hi:
5664   case X86::BI__builtin_ia32_vec_init_v2si:
5665     return Builder.CreateBitCast(BuildVector(Ops),
5666                                  llvm::Type::getX86_MMXTy(getLLVMContext()));
5667   case X86::BI__builtin_ia32_vec_ext_v2si:
5668     return Builder.CreateExtractElement(Ops[0],
5669                                   llvm::ConstantInt::get(Ops[1]->getType(), 0));
5670   case X86::BI__builtin_ia32_ldmxcsr: {
5671     Value *Tmp = CreateMemTemp(E->getArg(0)->getType());
5672     Builder.CreateStore(Ops[0], Tmp);
5673     return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
5674                               Builder.CreateBitCast(Tmp, Int8PtrTy));
5675   }
5676   case X86::BI__builtin_ia32_stmxcsr: {
5677     Value *Tmp = CreateMemTemp(E->getType());
5678     Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
5679                        Builder.CreateBitCast(Tmp, Int8PtrTy));
5680     return Builder.CreateLoad(Tmp, "stmxcsr");
5681   }
5682   case X86::BI__builtin_ia32_storehps:
5683   case X86::BI__builtin_ia32_storelps: {
5684     llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty);
5685     llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
5686 
5687     // cast val v2i64
5688     Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast");
5689 
5690     // extract (0, 1)
5691     unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
5692     llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index);
5693     Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
5694 
5695     // cast pointer to i64 & store
5696     Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
5697     return Builder.CreateStore(Ops[1], Ops[0]);
5698   }
5699   case X86::BI__builtin_ia32_palignr: {
5700     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
5701 
5702     // If palignr is shifting the pair of input vectors less than 9 bytes,
5703     // emit a shuffle instruction.
5704     if (shiftVal <= 8) {
5705       SmallVector<llvm::Constant*, 8> Indices;
5706       for (unsigned i = 0; i != 8; ++i)
5707         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
5708 
5709       Value* SV = llvm::ConstantVector::get(Indices);
5710       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
5711     }
5712 
5713     // If palignr is shifting the pair of input vectors more than 8 but less
5714     // than 16 bytes, emit a logical right shift of the destination.
5715     if (shiftVal < 16) {
5716       // MMX has these as 1 x i64 vectors for some odd optimization reasons.
5717       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1);
5718 
5719       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
5720       Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8);
5721 
5722       // create i32 constant
5723       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q);
5724       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
5725     }
5726 
5727     // If palignr is shifting the pair of vectors more than 16 bytes, emit zero.
5728     return llvm::Constant::getNullValue(ConvertType(E->getType()));
5729   }
5730   case X86::BI__builtin_ia32_palignr128: {
5731     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
5732 
5733     // If palignr is shifting the pair of input vectors less than 17 bytes,
5734     // emit a shuffle instruction.
5735     if (shiftVal <= 16) {
5736       SmallVector<llvm::Constant*, 16> Indices;
5737       for (unsigned i = 0; i != 16; ++i)
5738         Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i));
5739 
5740       Value* SV = llvm::ConstantVector::get(Indices);
5741       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
5742     }
5743 
5744     // If palignr is shifting the pair of input vectors more than 16 but less
5745     // than 32 bytes, emit a logical right shift of the destination.
5746     if (shiftVal < 32) {
5747       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2);
5748 
5749       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
5750       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
5751 
5752       // create i32 constant
5753       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq);
5754       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
5755     }
5756 
5757     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
5758     return llvm::Constant::getNullValue(ConvertType(E->getType()));
5759   }
5760   case X86::BI__builtin_ia32_palignr256: {
5761     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
5762 
5763     // If palignr is shifting the pair of input vectors less than 17 bytes,
5764     // emit a shuffle instruction.
5765     if (shiftVal <= 16) {
5766       SmallVector<llvm::Constant*, 32> Indices;
5767       // 256-bit palignr operates on 128-bit lanes so we need to handle that
5768       for (unsigned l = 0; l != 2; ++l) {
5769         unsigned LaneStart = l * 16;
5770         unsigned LaneEnd = (l+1) * 16;
5771         for (unsigned i = 0; i != 16; ++i) {
5772           unsigned Idx = shiftVal + i + LaneStart;
5773           if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand
5774           Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx));
5775         }
5776       }
5777 
5778       Value* SV = llvm::ConstantVector::get(Indices);
5779       return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
5780     }
5781 
5782     // If palignr is shifting the pair of input vectors more than 16 but less
5783     // than 32 bytes, emit a logical right shift of the destination.
5784     if (shiftVal < 32) {
5785       llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4);
5786 
5787       Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
5788       Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8);
5789 
5790       // create i32 constant
5791       llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq);
5792       return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr");
5793     }
5794 
5795     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
5796     return llvm::Constant::getNullValue(ConvertType(E->getType()));
5797   }
5798   case X86::BI__builtin_ia32_movntps:
5799   case X86::BI__builtin_ia32_movntps256:
5800   case X86::BI__builtin_ia32_movntpd:
5801   case X86::BI__builtin_ia32_movntpd256:
5802   case X86::BI__builtin_ia32_movntdq:
5803   case X86::BI__builtin_ia32_movntdq256:
5804   case X86::BI__builtin_ia32_movnti:
5805   case X86::BI__builtin_ia32_movnti64: {
5806     llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(),
5807                                            Builder.getInt32(1));
5808 
5809     // Convert the type of the pointer to a pointer to the stored type.
5810     Value *BC = Builder.CreateBitCast(Ops[0],
5811                                 llvm::PointerType::getUnqual(Ops[1]->getType()),
5812                                       "cast");
5813     StoreInst *SI = Builder.CreateStore(Ops[1], BC);
5814     SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
5815 
5816     // If the operand is an integer, we can't assume alignment. Otherwise,
5817     // assume natural alignment.
5818     QualType ArgTy = E->getArg(1)->getType();
5819     unsigned Align;
5820     if (ArgTy->isIntegerType())
5821       Align = 1;
5822     else
5823       Align = getContext().getTypeSizeInChars(ArgTy).getQuantity();
5824     SI->setAlignment(Align);
5825     return SI;
5826   }
5827   // 3DNow!
5828   case X86::BI__builtin_ia32_pswapdsf:
5829   case X86::BI__builtin_ia32_pswapdsi: {
5830     const char *name = nullptr;
5831     Intrinsic::ID ID = Intrinsic::not_intrinsic;
5832     switch(BuiltinID) {
5833     default: llvm_unreachable("Unsupported intrinsic!");
5834     case X86::BI__builtin_ia32_pswapdsf:
5835     case X86::BI__builtin_ia32_pswapdsi:
5836       name = "pswapd";
5837       ID = Intrinsic::x86_3dnowa_pswapd;
5838       break;
5839     }
5840     llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
5841     Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
5842     llvm::Function *F = CGM.getIntrinsic(ID);
5843     return Builder.CreateCall(F, Ops, name);
5844   }
5845   case X86::BI__builtin_ia32_rdrand16_step:
5846   case X86::BI__builtin_ia32_rdrand32_step:
5847   case X86::BI__builtin_ia32_rdrand64_step:
5848   case X86::BI__builtin_ia32_rdseed16_step:
5849   case X86::BI__builtin_ia32_rdseed32_step:
5850   case X86::BI__builtin_ia32_rdseed64_step: {
5851     Intrinsic::ID ID;
5852     switch (BuiltinID) {
5853     default: llvm_unreachable("Unsupported intrinsic!");
5854     case X86::BI__builtin_ia32_rdrand16_step:
5855       ID = Intrinsic::x86_rdrand_16;
5856       break;
5857     case X86::BI__builtin_ia32_rdrand32_step:
5858       ID = Intrinsic::x86_rdrand_32;
5859       break;
5860     case X86::BI__builtin_ia32_rdrand64_step:
5861       ID = Intrinsic::x86_rdrand_64;
5862       break;
5863     case X86::BI__builtin_ia32_rdseed16_step:
5864       ID = Intrinsic::x86_rdseed_16;
5865       break;
5866     case X86::BI__builtin_ia32_rdseed32_step:
5867       ID = Intrinsic::x86_rdseed_32;
5868       break;
5869     case X86::BI__builtin_ia32_rdseed64_step:
5870       ID = Intrinsic::x86_rdseed_64;
5871       break;
5872     }
5873 
5874     Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
5875     Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]);
5876     return Builder.CreateExtractValue(Call, 1);
5877   }
5878   // AVX2 broadcast
5879   case X86::BI__builtin_ia32_vbroadcastsi256: {
5880     Value *VecTmp = CreateMemTemp(E->getArg(0)->getType());
5881     Builder.CreateStore(Ops[0], VecTmp);
5882     Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128);
5883     return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy));
5884   }
5885   }
5886 }
5887 
5888 
EmitPPCBuiltinExpr(unsigned BuiltinID,const CallExpr * E)5889 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
5890                                            const CallExpr *E) {
5891   SmallVector<Value*, 4> Ops;
5892 
5893   for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
5894     Ops.push_back(EmitScalarExpr(E->getArg(i)));
5895 
5896   Intrinsic::ID ID = Intrinsic::not_intrinsic;
5897 
5898   switch (BuiltinID) {
5899   default: return nullptr;
5900 
5901   // vec_ld, vec_lvsl, vec_lvsr
5902   case PPC::BI__builtin_altivec_lvx:
5903   case PPC::BI__builtin_altivec_lvxl:
5904   case PPC::BI__builtin_altivec_lvebx:
5905   case PPC::BI__builtin_altivec_lvehx:
5906   case PPC::BI__builtin_altivec_lvewx:
5907   case PPC::BI__builtin_altivec_lvsl:
5908   case PPC::BI__builtin_altivec_lvsr:
5909   {
5910     Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
5911 
5912     Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]);
5913     Ops.pop_back();
5914 
5915     switch (BuiltinID) {
5916     default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
5917     case PPC::BI__builtin_altivec_lvx:
5918       ID = Intrinsic::ppc_altivec_lvx;
5919       break;
5920     case PPC::BI__builtin_altivec_lvxl:
5921       ID = Intrinsic::ppc_altivec_lvxl;
5922       break;
5923     case PPC::BI__builtin_altivec_lvebx:
5924       ID = Intrinsic::ppc_altivec_lvebx;
5925       break;
5926     case PPC::BI__builtin_altivec_lvehx:
5927       ID = Intrinsic::ppc_altivec_lvehx;
5928       break;
5929     case PPC::BI__builtin_altivec_lvewx:
5930       ID = Intrinsic::ppc_altivec_lvewx;
5931       break;
5932     case PPC::BI__builtin_altivec_lvsl:
5933       ID = Intrinsic::ppc_altivec_lvsl;
5934       break;
5935     case PPC::BI__builtin_altivec_lvsr:
5936       ID = Intrinsic::ppc_altivec_lvsr;
5937       break;
5938     }
5939     llvm::Function *F = CGM.getIntrinsic(ID);
5940     return Builder.CreateCall(F, Ops, "");
5941   }
5942 
5943   // vec_st
5944   case PPC::BI__builtin_altivec_stvx:
5945   case PPC::BI__builtin_altivec_stvxl:
5946   case PPC::BI__builtin_altivec_stvebx:
5947   case PPC::BI__builtin_altivec_stvehx:
5948   case PPC::BI__builtin_altivec_stvewx:
5949   {
5950     Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
5951     Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]);
5952     Ops.pop_back();
5953 
5954     switch (BuiltinID) {
5955     default: llvm_unreachable("Unsupported st intrinsic!");
5956     case PPC::BI__builtin_altivec_stvx:
5957       ID = Intrinsic::ppc_altivec_stvx;
5958       break;
5959     case PPC::BI__builtin_altivec_stvxl:
5960       ID = Intrinsic::ppc_altivec_stvxl;
5961       break;
5962     case PPC::BI__builtin_altivec_stvebx:
5963       ID = Intrinsic::ppc_altivec_stvebx;
5964       break;
5965     case PPC::BI__builtin_altivec_stvehx:
5966       ID = Intrinsic::ppc_altivec_stvehx;
5967       break;
5968     case PPC::BI__builtin_altivec_stvewx:
5969       ID = Intrinsic::ppc_altivec_stvewx;
5970       break;
5971     }
5972     llvm::Function *F = CGM.getIntrinsic(ID);
5973     return Builder.CreateCall(F, Ops, "");
5974   }
5975   }
5976 }
5977 
EmitR600BuiltinExpr(unsigned BuiltinID,const CallExpr * E)5978 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID,
5979                                             const CallExpr *E) {
5980   switch (BuiltinID) {
5981   case R600::BI__builtin_amdgpu_div_scale:
5982   case R600::BI__builtin_amdgpu_div_scalef: {
5983     // Translate from the intrinsics's struct return to the builtin's out
5984     // argument.
5985 
5986     std::pair<llvm::Value *, unsigned> FlagOutPtr
5987       = EmitPointerWithAlignment(E->getArg(3));
5988 
5989     llvm::Value *X = EmitScalarExpr(E->getArg(0));
5990     llvm::Value *Y = EmitScalarExpr(E->getArg(1));
5991     llvm::Value *Z = EmitScalarExpr(E->getArg(2));
5992 
5993     llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale,
5994                                            X->getType());
5995 
5996     llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z);
5997 
5998     llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0);
5999     llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1);
6000 
6001     llvm::Type *RealFlagType
6002       = FlagOutPtr.first->getType()->getPointerElementType();
6003 
6004     llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType);
6005     llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first);
6006     FlagStore->setAlignment(FlagOutPtr.second);
6007     return Result;
6008   } default:
6009     return nullptr;
6010   }
6011 }
6012