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". 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. 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 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. 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. 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. 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 } 166 167 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. 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 } 196 197 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 1606 if (IntrinsicID != Intrinsic::not_intrinsic) { 1607 SmallVector<Value*, 16> Args; 1608 1609 // Find out if any arguments are required to be integer constant 1610 // expressions. 1611 unsigned ICEArguments = 0; 1612 ASTContext::GetBuiltinTypeError Error; 1613 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1614 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1615 1616 Function *F = CGM.getIntrinsic(IntrinsicID); 1617 llvm::FunctionType *FTy = F->getFunctionType(); 1618 1619 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1620 Value *ArgValue; 1621 // If this is a normal argument, just emit it as a scalar. 1622 if ((ICEArguments & (1 << i)) == 0) { 1623 ArgValue = EmitScalarExpr(E->getArg(i)); 1624 } else { 1625 // If this is required to be a constant, constant fold it so that we 1626 // know that the generated intrinsic gets a ConstantInt. 1627 llvm::APSInt Result; 1628 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1629 assert(IsConst && "Constant arg isn't actually constant?"); 1630 (void)IsConst; 1631 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1632 } 1633 1634 // If the intrinsic arg type is different from the builtin arg type 1635 // we need to do a bit cast. 1636 llvm::Type *PTy = FTy->getParamType(i); 1637 if (PTy != ArgValue->getType()) { 1638 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1639 "Must be able to losslessly bit cast to param"); 1640 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1641 } 1642 1643 Args.push_back(ArgValue); 1644 } 1645 1646 Value *V = Builder.CreateCall(F, Args); 1647 QualType BuiltinRetType = E->getType(); 1648 1649 llvm::Type *RetTy = VoidTy; 1650 if (!BuiltinRetType->isVoidType()) 1651 RetTy = ConvertType(BuiltinRetType); 1652 1653 if (RetTy != V->getType()) { 1654 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1655 "Must be able to losslessly bit cast result type"); 1656 V = Builder.CreateBitCast(V, RetTy); 1657 } 1658 1659 return RValue::get(V); 1660 } 1661 1662 // See if we have a target specific builtin that needs to be lowered. 1663 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1664 return RValue::get(V); 1665 1666 ErrorUnsupported(E, "builtin function"); 1667 1668 // Unknown builtin, for now just dump it out and return undef. 1669 return GetUndefRValue(E->getType()); 1670 } 1671 1672 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1673 const CallExpr *E) { 1674 switch (getTarget().getTriple().getArch()) { 1675 case llvm::Triple::arm: 1676 case llvm::Triple::armeb: 1677 case llvm::Triple::thumb: 1678 case llvm::Triple::thumbeb: 1679 return EmitARMBuiltinExpr(BuiltinID, E); 1680 case llvm::Triple::aarch64: 1681 case llvm::Triple::aarch64_be: 1682 case llvm::Triple::arm64: 1683 case llvm::Triple::arm64_be: 1684 return EmitAArch64BuiltinExpr(BuiltinID, E); 1685 case llvm::Triple::x86: 1686 case llvm::Triple::x86_64: 1687 return EmitX86BuiltinExpr(BuiltinID, E); 1688 case llvm::Triple::ppc: 1689 case llvm::Triple::ppc64: 1690 case llvm::Triple::ppc64le: 1691 return EmitPPCBuiltinExpr(BuiltinID, E); 1692 default: 1693 return nullptr; 1694 } 1695 } 1696 1697 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1698 NeonTypeFlags TypeFlags, 1699 bool V1Ty=false) { 1700 int IsQuad = TypeFlags.isQuad(); 1701 switch (TypeFlags.getEltType()) { 1702 case NeonTypeFlags::Int8: 1703 case NeonTypeFlags::Poly8: 1704 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1705 case NeonTypeFlags::Int16: 1706 case NeonTypeFlags::Poly16: 1707 case NeonTypeFlags::Float16: 1708 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1709 case NeonTypeFlags::Int32: 1710 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1711 case NeonTypeFlags::Int64: 1712 case NeonTypeFlags::Poly64: 1713 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1714 case NeonTypeFlags::Poly128: 1715 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1716 // There is a lot of i128 and f128 API missing. 1717 // so we use v16i8 to represent poly128 and get pattern matched. 1718 return llvm::VectorType::get(CGF->Int8Ty, 16); 1719 case NeonTypeFlags::Float32: 1720 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1721 case NeonTypeFlags::Float64: 1722 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1723 } 1724 llvm_unreachable("Unknown vector element type!"); 1725 } 1726 1727 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1728 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1729 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1730 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1731 } 1732 1733 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1734 const char *name, 1735 unsigned shift, bool rightshift) { 1736 unsigned j = 0; 1737 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1738 ai != ae; ++ai, ++j) 1739 if (shift > 0 && shift == j) 1740 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1741 else 1742 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1743 1744 return Builder.CreateCall(F, Ops, name); 1745 } 1746 1747 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1748 bool neg) { 1749 int SV = cast<ConstantInt>(V)->getSExtValue(); 1750 1751 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1752 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1753 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1754 } 1755 1756 // \brief Right-shift a vector by a constant. 1757 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1758 llvm::Type *Ty, bool usgn, 1759 const char *name) { 1760 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1761 1762 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1763 int EltSize = VTy->getScalarSizeInBits(); 1764 1765 Vec = Builder.CreateBitCast(Vec, Ty); 1766 1767 // lshr/ashr are undefined when the shift amount is equal to the vector 1768 // element size. 1769 if (ShiftAmt == EltSize) { 1770 if (usgn) { 1771 // Right-shifting an unsigned value by its size yields 0. 1772 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1773 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1774 } else { 1775 // Right-shifting a signed value by its size is equivalent 1776 // to a shift of size-1. 1777 --ShiftAmt; 1778 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1779 } 1780 } 1781 1782 Shift = EmitNeonShiftVector(Shift, Ty, false); 1783 if (usgn) 1784 return Builder.CreateLShr(Vec, Shift, name); 1785 else 1786 return Builder.CreateAShr(Vec, Shift, name); 1787 } 1788 1789 Value *CodeGenFunction::EmitConcatVectors(Value *Lo, Value *Hi, 1790 llvm::Type *ArgTy) { 1791 unsigned NumElts = ArgTy->getVectorNumElements(); 1792 SmallVector<Constant *, 16> Indices; 1793 for (unsigned i = 0; i < 2 * NumElts; ++i) 1794 Indices.push_back(ConstantInt::get(Int32Ty, i)); 1795 1796 Constant *Mask = ConstantVector::get(Indices); 1797 Value *LoCast = Builder.CreateBitCast(Lo, ArgTy); 1798 Value *HiCast = Builder.CreateBitCast(Hi, ArgTy); 1799 return Builder.CreateShuffleVector(LoCast, HiCast, Mask, "concat"); 1800 } 1801 1802 Value *CodeGenFunction::EmitExtractHigh(Value *Vec, llvm::Type *ResTy) { 1803 unsigned NumElts = ResTy->getVectorNumElements(); 1804 SmallVector<Constant *, 8> Indices; 1805 1806 llvm::Type *InTy = llvm::VectorType::get(ResTy->getVectorElementType(), 1807 NumElts * 2); 1808 Value *VecCast = Builder.CreateBitCast(Vec, InTy); 1809 1810 // extract_high is a shuffle on the second half of the input indices: E.g. 4, 1811 // 5, 6, 7 if we're extracting <4 x i16> from <8 x i16>. 1812 for (unsigned i = 0; i < NumElts; ++i) 1813 Indices.push_back(ConstantInt::get(Int32Ty, NumElts + i)); 1814 1815 Constant *Mask = ConstantVector::get(Indices); 1816 return Builder.CreateShuffleVector(VecCast, VecCast, Mask, "concat"); 1817 } 1818 1819 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1820 /// alignment of the type referenced by the pointer. Skip over implicit 1821 /// casts. 1822 std::pair<llvm::Value*, unsigned> 1823 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1824 assert(Addr->getType()->isPointerType()); 1825 Addr = Addr->IgnoreParens(); 1826 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1827 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1828 ICE->getSubExpr()->getType()->isPointerType()) { 1829 std::pair<llvm::Value*, unsigned> Ptr = 1830 EmitPointerWithAlignment(ICE->getSubExpr()); 1831 Ptr.first = Builder.CreateBitCast(Ptr.first, 1832 ConvertType(Addr->getType())); 1833 return Ptr; 1834 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1835 LValue LV = EmitLValue(ICE->getSubExpr()); 1836 unsigned Align = LV.getAlignment().getQuantity(); 1837 if (!Align) { 1838 // FIXME: Once LValues are fixed to always set alignment, 1839 // zap this code. 1840 QualType PtTy = ICE->getSubExpr()->getType(); 1841 if (!PtTy->isIncompleteType()) 1842 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1843 else 1844 Align = 1; 1845 } 1846 return std::make_pair(LV.getAddress(), Align); 1847 } 1848 } 1849 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1850 if (UO->getOpcode() == UO_AddrOf) { 1851 LValue LV = EmitLValue(UO->getSubExpr()); 1852 unsigned Align = LV.getAlignment().getQuantity(); 1853 if (!Align) { 1854 // FIXME: Once LValues are fixed to always set alignment, 1855 // zap this code. 1856 QualType PtTy = UO->getSubExpr()->getType(); 1857 if (!PtTy->isIncompleteType()) 1858 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1859 else 1860 Align = 1; 1861 } 1862 return std::make_pair(LV.getAddress(), Align); 1863 } 1864 } 1865 1866 unsigned Align = 1; 1867 QualType PtTy = Addr->getType()->getPointeeType(); 1868 if (!PtTy->isIncompleteType()) 1869 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1870 1871 return std::make_pair(EmitScalarExpr(Addr), Align); 1872 } 1873 1874 enum { 1875 AddRetType = (1 << 0), 1876 Add1ArgType = (1 << 1), 1877 Add2ArgTypes = (1 << 2), 1878 1879 VectorizeRetType = (1 << 3), 1880 VectorizeArgTypes = (1 << 4), 1881 1882 InventFloatType = (1 << 5), 1883 UnsignedAlts = (1 << 6), 1884 1885 Use64BitVectors = (1 << 7), 1886 Use128BitVectors = (1 << 8), 1887 1888 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 1889 VectorRet = AddRetType | VectorizeRetType, 1890 VectorRetGetArgs01 = 1891 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 1892 FpCmpzModifiers = 1893 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 1894 }; 1895 1896 struct NeonIntrinsicInfo { 1897 unsigned BuiltinID; 1898 unsigned LLVMIntrinsic; 1899 unsigned AltLLVMIntrinsic; 1900 const char *NameHint; 1901 unsigned TypeModifier; 1902 1903 bool operator<(unsigned RHSBuiltinID) const { 1904 return BuiltinID < RHSBuiltinID; 1905 } 1906 }; 1907 1908 #define NEONMAP0(NameBase) \ 1909 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 1910 1911 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 1912 { NEON:: BI__builtin_neon_ ## NameBase, \ 1913 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 1914 1915 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 1916 { NEON:: BI__builtin_neon_ ## NameBase, \ 1917 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 1918 #NameBase, TypeModifier } 1919 1920 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 1921 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1922 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1923 NEONMAP1(vabs_v, arm_neon_vabs, 0), 1924 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 1925 NEONMAP0(vaddhn_v), 1926 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 1927 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 1928 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 1929 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 1930 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 1931 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 1932 NEONMAP1(vcage_v, arm_neon_vacge, 0), 1933 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 1934 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 1935 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 1936 NEONMAP1(vcale_v, arm_neon_vacge, 0), 1937 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 1938 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 1939 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 1940 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 1941 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 1942 NEONMAP1(vclz_v, ctlz, Add1ArgType), 1943 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 1944 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 1945 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 1946 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 1947 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 1948 NEONMAP0(vcvt_f32_v), 1949 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1950 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1951 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1952 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1953 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1954 NEONMAP0(vcvt_s32_v), 1955 NEONMAP0(vcvt_s64_v), 1956 NEONMAP0(vcvt_u32_v), 1957 NEONMAP0(vcvt_u64_v), 1958 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 1959 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 1960 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 1961 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 1962 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 1963 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 1964 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 1965 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 1966 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 1967 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 1968 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 1969 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 1970 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 1971 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 1972 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 1973 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 1974 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 1975 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 1976 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 1977 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 1978 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 1979 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 1980 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 1981 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 1982 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 1983 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 1984 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 1985 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 1986 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 1987 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 1988 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 1989 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 1990 NEONMAP0(vcvtq_f32_v), 1991 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1992 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1993 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1994 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1995 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1996 NEONMAP0(vcvtq_s32_v), 1997 NEONMAP0(vcvtq_s64_v), 1998 NEONMAP0(vcvtq_u32_v), 1999 NEONMAP0(vcvtq_u64_v), 2000 NEONMAP0(vext_v), 2001 NEONMAP0(vextq_v), 2002 NEONMAP0(vfma_v), 2003 NEONMAP0(vfmaq_v), 2004 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2005 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2006 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2007 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2008 NEONMAP0(vld1_dup_v), 2009 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2010 NEONMAP0(vld1q_dup_v), 2011 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2012 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2013 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2014 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2015 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2016 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2017 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2018 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2019 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2020 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2021 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2022 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2023 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2024 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2025 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2026 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2027 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2028 NEONMAP0(vmovl_v), 2029 NEONMAP0(vmovn_v), 2030 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2031 NEONMAP0(vmull_v), 2032 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2033 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2034 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2035 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2036 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2037 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2038 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2039 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2040 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2041 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2042 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2043 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2044 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2045 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2046 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2047 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2048 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2049 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2050 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2051 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2052 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2053 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2054 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2055 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2056 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2057 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2058 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2059 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2060 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2061 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2062 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2063 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2064 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2065 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2066 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2067 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2068 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2069 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2070 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2071 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2072 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2073 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2074 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2075 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2076 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2077 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2078 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2079 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2080 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2081 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2082 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2083 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2084 NEONMAP0(vshl_n_v), 2085 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2086 NEONMAP0(vshll_n_v), 2087 NEONMAP0(vshlq_n_v), 2088 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2089 NEONMAP0(vshr_n_v), 2090 NEONMAP0(vshrn_n_v), 2091 NEONMAP0(vshrq_n_v), 2092 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2093 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2094 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2095 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2096 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2097 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2098 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2099 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2100 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2101 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2102 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2103 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2104 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2105 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2106 NEONMAP0(vsubhn_v), 2107 NEONMAP0(vtrn_v), 2108 NEONMAP0(vtrnq_v), 2109 NEONMAP0(vtst_v), 2110 NEONMAP0(vtstq_v), 2111 NEONMAP0(vuzp_v), 2112 NEONMAP0(vuzpq_v), 2113 NEONMAP0(vzip_v), 2114 NEONMAP0(vzipq_v) 2115 }; 2116 2117 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2118 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2119 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2120 NEONMAP0(vaddhn_v), 2121 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2122 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2123 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2124 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2125 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2126 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2127 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2128 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2129 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2130 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2131 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2132 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2133 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2134 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2135 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2136 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2137 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2138 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2139 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2140 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2141 NEONMAP0(vcvt_f32_v), 2142 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2143 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2144 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2145 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2146 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2147 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2148 NEONMAP0(vcvtq_f32_v), 2149 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2150 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2151 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2152 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2153 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2154 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2155 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2156 NEONMAP0(vext_v), 2157 NEONMAP0(vextq_v), 2158 NEONMAP0(vfma_v), 2159 NEONMAP0(vfmaq_v), 2160 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2161 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2162 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2163 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2164 NEONMAP0(vmovl_v), 2165 NEONMAP0(vmovn_v), 2166 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2167 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2168 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2169 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2170 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2171 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2172 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2173 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2174 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2175 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2176 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2177 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2178 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2179 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2180 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2181 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2182 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2183 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2184 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2185 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2186 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2187 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2188 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2189 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2190 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2191 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2192 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2193 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2194 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2195 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2196 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2197 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2198 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2199 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2200 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2201 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2202 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2203 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2204 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2205 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2206 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2207 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2208 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2209 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2210 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2211 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2212 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2213 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2214 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2215 NEONMAP0(vshl_n_v), 2216 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2217 NEONMAP0(vshll_n_v), 2218 NEONMAP0(vshlq_n_v), 2219 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2220 NEONMAP0(vshr_n_v), 2221 NEONMAP0(vshrn_n_v), 2222 NEONMAP0(vshrq_n_v), 2223 NEONMAP0(vsubhn_v), 2224 NEONMAP0(vtst_v), 2225 NEONMAP0(vtstq_v), 2226 }; 2227 2228 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2229 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2230 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2231 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2232 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2233 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2234 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2235 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2236 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2237 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2238 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2239 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2240 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2241 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2242 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2243 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2244 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2245 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2246 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2247 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2248 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2249 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2250 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2251 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2252 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2253 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2254 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2255 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2256 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2257 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2258 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2259 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2260 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2261 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2262 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2263 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2264 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2265 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2266 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2267 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2268 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2269 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2270 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2271 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2272 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2273 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2274 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2275 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2276 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2277 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2278 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2279 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2280 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2281 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2282 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2283 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2284 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2285 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2286 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2287 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2288 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2289 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2290 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2291 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2292 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2293 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2294 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2295 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2296 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2297 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2298 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2299 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2300 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2301 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2302 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2303 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2304 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2305 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2306 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2307 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2308 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2309 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2310 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2311 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2312 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2313 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2314 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2315 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2316 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2317 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2318 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2319 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2320 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2321 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2322 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2323 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2324 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2325 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2326 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2327 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2328 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2329 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2330 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2331 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2332 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2333 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2334 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2335 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2336 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2337 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2338 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2339 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2340 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2341 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2342 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2343 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2344 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2345 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2346 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2347 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2348 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2349 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2350 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2351 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2352 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2353 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2354 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2355 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2356 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2357 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2358 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2359 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2360 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2361 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2362 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2363 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2364 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2365 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2366 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2367 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2368 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2369 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2370 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2371 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2372 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2373 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2374 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2375 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2376 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2377 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2378 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2379 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2380 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2381 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2382 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2383 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2384 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2385 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2386 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2387 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2388 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2389 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2390 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2391 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2392 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2393 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2394 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2395 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2396 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2397 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2398 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2399 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2400 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2401 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2402 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2403 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2404 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2405 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2406 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2407 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2408 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2409 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2410 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2411 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2412 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2413 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2414 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2415 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2416 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2417 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2418 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2419 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2420 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2421 }; 2422 2423 #undef NEONMAP0 2424 #undef NEONMAP1 2425 #undef NEONMAP2 2426 2427 static bool NEONSIMDIntrinsicsProvenSorted = false; 2428 2429 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2430 static bool AArch64SISDIntrinsicsProvenSorted = false; 2431 2432 2433 static const NeonIntrinsicInfo * 2434 findNeonIntrinsicInMap(llvm::ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2435 unsigned BuiltinID, bool &MapProvenSorted) { 2436 2437 #ifndef NDEBUG 2438 if (!MapProvenSorted) { 2439 // FIXME: use std::is_sorted once C++11 is allowed 2440 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2441 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2442 MapProvenSorted = true; 2443 } 2444 #endif 2445 2446 const NeonIntrinsicInfo *Builtin = 2447 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2448 2449 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2450 return Builtin; 2451 2452 return nullptr; 2453 } 2454 2455 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2456 unsigned Modifier, 2457 llvm::Type *ArgType, 2458 const CallExpr *E) { 2459 int VectorSize = 0; 2460 if (Modifier & Use64BitVectors) 2461 VectorSize = 64; 2462 else if (Modifier & Use128BitVectors) 2463 VectorSize = 128; 2464 2465 // Return type. 2466 SmallVector<llvm::Type *, 3> Tys; 2467 if (Modifier & AddRetType) { 2468 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2469 if (Modifier & VectorizeRetType) 2470 Ty = llvm::VectorType::get( 2471 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2472 2473 Tys.push_back(Ty); 2474 } 2475 2476 // Arguments. 2477 if (Modifier & VectorizeArgTypes) { 2478 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2479 ArgType = llvm::VectorType::get(ArgType, Elts); 2480 } 2481 2482 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2483 Tys.push_back(ArgType); 2484 2485 if (Modifier & Add2ArgTypes) 2486 Tys.push_back(ArgType); 2487 2488 if (Modifier & InventFloatType) 2489 Tys.push_back(FloatTy); 2490 2491 return CGM.getIntrinsic(IntrinsicID, Tys); 2492 } 2493 2494 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2495 const NeonIntrinsicInfo &SISDInfo, 2496 SmallVectorImpl<Value *> &Ops, 2497 const CallExpr *E) { 2498 unsigned BuiltinID = SISDInfo.BuiltinID; 2499 unsigned int Int = SISDInfo.LLVMIntrinsic; 2500 unsigned Modifier = SISDInfo.TypeModifier; 2501 const char *s = SISDInfo.NameHint; 2502 2503 switch (BuiltinID) { 2504 case NEON::BI__builtin_neon_vcled_s64: 2505 case NEON::BI__builtin_neon_vcled_u64: 2506 case NEON::BI__builtin_neon_vcles_f32: 2507 case NEON::BI__builtin_neon_vcled_f64: 2508 case NEON::BI__builtin_neon_vcltd_s64: 2509 case NEON::BI__builtin_neon_vcltd_u64: 2510 case NEON::BI__builtin_neon_vclts_f32: 2511 case NEON::BI__builtin_neon_vcltd_f64: 2512 case NEON::BI__builtin_neon_vcales_f32: 2513 case NEON::BI__builtin_neon_vcaled_f64: 2514 case NEON::BI__builtin_neon_vcalts_f32: 2515 case NEON::BI__builtin_neon_vcaltd_f64: 2516 // Only one direction of comparisons actually exist, cmle is actually a cmge 2517 // with swapped operands. The table gives us the right intrinsic but we 2518 // still need to do the swap. 2519 std::swap(Ops[0], Ops[1]); 2520 break; 2521 } 2522 2523 assert(Int && "Generic code assumes a valid intrinsic"); 2524 2525 // Determine the type(s) of this overloaded AArch64 intrinsic. 2526 const Expr *Arg = E->getArg(0); 2527 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2528 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2529 2530 int j = 0; 2531 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2532 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2533 ai != ae; ++ai, ++j) { 2534 llvm::Type *ArgTy = ai->getType(); 2535 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2536 ArgTy->getPrimitiveSizeInBits()) 2537 continue; 2538 2539 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2540 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2541 // it before inserting. 2542 Ops[j] = 2543 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2544 Ops[j] = 2545 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2546 } 2547 2548 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2549 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2550 if (ResultType->getPrimitiveSizeInBits() < 2551 Result->getType()->getPrimitiveSizeInBits()) 2552 return CGF.Builder.CreateExtractElement(Result, C0); 2553 2554 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2555 } 2556 2557 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2558 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2559 const char *NameHint, unsigned Modifier, const CallExpr *E, 2560 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2561 // Get the last argument, which specifies the vector type. 2562 llvm::APSInt NeonTypeConst; 2563 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2564 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2565 return nullptr; 2566 2567 // Determine the type of this overloaded NEON intrinsic. 2568 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2569 bool Usgn = Type.isUnsigned(); 2570 bool Quad = Type.isQuad(); 2571 2572 llvm::VectorType *VTy = GetNeonType(this, Type); 2573 llvm::Type *Ty = VTy; 2574 if (!Ty) 2575 return nullptr; 2576 2577 unsigned Int = LLVMIntrinsic; 2578 if ((Modifier & UnsignedAlts) && !Usgn) 2579 Int = AltLLVMIntrinsic; 2580 2581 switch (BuiltinID) { 2582 default: break; 2583 case NEON::BI__builtin_neon_vabs_v: 2584 case NEON::BI__builtin_neon_vabsq_v: 2585 if (VTy->getElementType()->isFloatingPointTy()) 2586 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2587 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2588 case NEON::BI__builtin_neon_vaddhn_v: { 2589 llvm::VectorType *SrcTy = 2590 llvm::VectorType::getExtendedElementVectorType(VTy); 2591 2592 // %sum = add <4 x i32> %lhs, %rhs 2593 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2594 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2595 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2596 2597 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2598 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2599 SrcTy->getScalarSizeInBits() / 2); 2600 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2601 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2602 2603 // %res = trunc <4 x i32> %high to <4 x i16> 2604 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2605 } 2606 case NEON::BI__builtin_neon_vcale_v: 2607 case NEON::BI__builtin_neon_vcaleq_v: 2608 case NEON::BI__builtin_neon_vcalt_v: 2609 case NEON::BI__builtin_neon_vcaltq_v: 2610 std::swap(Ops[0], Ops[1]); 2611 case NEON::BI__builtin_neon_vcage_v: 2612 case NEON::BI__builtin_neon_vcageq_v: 2613 case NEON::BI__builtin_neon_vcagt_v: 2614 case NEON::BI__builtin_neon_vcagtq_v: { 2615 llvm::Type *VecFlt = llvm::VectorType::get( 2616 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2617 VTy->getNumElements()); 2618 llvm::Type *Tys[] = { VTy, VecFlt }; 2619 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2620 return EmitNeonCall(F, Ops, NameHint); 2621 } 2622 case NEON::BI__builtin_neon_vclz_v: 2623 case NEON::BI__builtin_neon_vclzq_v: 2624 // We generate target-independent intrinsic, which needs a second argument 2625 // for whether or not clz of zero is undefined; on ARM it isn't. 2626 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2627 break; 2628 case NEON::BI__builtin_neon_vcvt_f32_v: 2629 case NEON::BI__builtin_neon_vcvtq_f32_v: 2630 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2631 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2632 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2633 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2634 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2635 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2636 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2637 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2638 bool Double = 2639 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2640 llvm::Type *FloatTy = 2641 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2642 : NeonTypeFlags::Float32, 2643 false, Quad)); 2644 llvm::Type *Tys[2] = { FloatTy, Ty }; 2645 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2646 Function *F = CGM.getIntrinsic(Int, Tys); 2647 return EmitNeonCall(F, Ops, "vcvt_n"); 2648 } 2649 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2650 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2651 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2652 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2653 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2654 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2655 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2656 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2657 bool Double = 2658 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2659 llvm::Type *FloatTy = 2660 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2661 : NeonTypeFlags::Float32, 2662 false, Quad)); 2663 llvm::Type *Tys[2] = { Ty, FloatTy }; 2664 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2665 return EmitNeonCall(F, Ops, "vcvt_n"); 2666 } 2667 case NEON::BI__builtin_neon_vcvt_s32_v: 2668 case NEON::BI__builtin_neon_vcvt_u32_v: 2669 case NEON::BI__builtin_neon_vcvt_s64_v: 2670 case NEON::BI__builtin_neon_vcvt_u64_v: 2671 case NEON::BI__builtin_neon_vcvtq_s32_v: 2672 case NEON::BI__builtin_neon_vcvtq_u32_v: 2673 case NEON::BI__builtin_neon_vcvtq_s64_v: 2674 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2675 bool Double = 2676 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2677 llvm::Type *FloatTy = 2678 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2679 : NeonTypeFlags::Float32, 2680 false, Quad)); 2681 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2682 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2683 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2684 } 2685 case NEON::BI__builtin_neon_vcvta_s32_v: 2686 case NEON::BI__builtin_neon_vcvta_s64_v: 2687 case NEON::BI__builtin_neon_vcvta_u32_v: 2688 case NEON::BI__builtin_neon_vcvta_u64_v: 2689 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2690 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2691 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2692 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2693 case NEON::BI__builtin_neon_vcvtn_s32_v: 2694 case NEON::BI__builtin_neon_vcvtn_s64_v: 2695 case NEON::BI__builtin_neon_vcvtn_u32_v: 2696 case NEON::BI__builtin_neon_vcvtn_u64_v: 2697 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2698 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2699 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2700 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2701 case NEON::BI__builtin_neon_vcvtp_s32_v: 2702 case NEON::BI__builtin_neon_vcvtp_s64_v: 2703 case NEON::BI__builtin_neon_vcvtp_u32_v: 2704 case NEON::BI__builtin_neon_vcvtp_u64_v: 2705 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2706 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2707 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2708 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2709 case NEON::BI__builtin_neon_vcvtm_s32_v: 2710 case NEON::BI__builtin_neon_vcvtm_s64_v: 2711 case NEON::BI__builtin_neon_vcvtm_u32_v: 2712 case NEON::BI__builtin_neon_vcvtm_u64_v: 2713 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2714 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2715 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2716 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2717 bool Double = 2718 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2719 llvm::Type *InTy = 2720 GetNeonType(this, 2721 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2722 : NeonTypeFlags::Float32, false, Quad)); 2723 llvm::Type *Tys[2] = { Ty, InTy }; 2724 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2725 } 2726 case NEON::BI__builtin_neon_vext_v: 2727 case NEON::BI__builtin_neon_vextq_v: { 2728 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2729 SmallVector<Constant*, 16> Indices; 2730 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2731 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2732 2733 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2734 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2735 Value *SV = llvm::ConstantVector::get(Indices); 2736 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2737 } 2738 case NEON::BI__builtin_neon_vfma_v: 2739 case NEON::BI__builtin_neon_vfmaq_v: { 2740 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2741 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2742 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2743 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2744 2745 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2746 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2747 } 2748 case NEON::BI__builtin_neon_vld1_v: 2749 case NEON::BI__builtin_neon_vld1q_v: 2750 Ops.push_back(Align); 2751 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2752 case NEON::BI__builtin_neon_vld2_v: 2753 case NEON::BI__builtin_neon_vld2q_v: 2754 case NEON::BI__builtin_neon_vld3_v: 2755 case NEON::BI__builtin_neon_vld3q_v: 2756 case NEON::BI__builtin_neon_vld4_v: 2757 case NEON::BI__builtin_neon_vld4q_v: { 2758 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2759 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2760 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2761 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2762 return Builder.CreateStore(Ops[1], Ops[0]); 2763 } 2764 case NEON::BI__builtin_neon_vld1_dup_v: 2765 case NEON::BI__builtin_neon_vld1q_dup_v: { 2766 Value *V = UndefValue::get(Ty); 2767 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2768 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2769 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2770 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2771 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2772 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2773 return EmitNeonSplat(Ops[0], CI); 2774 } 2775 case NEON::BI__builtin_neon_vld2_lane_v: 2776 case NEON::BI__builtin_neon_vld2q_lane_v: 2777 case NEON::BI__builtin_neon_vld3_lane_v: 2778 case NEON::BI__builtin_neon_vld3q_lane_v: 2779 case NEON::BI__builtin_neon_vld4_lane_v: 2780 case NEON::BI__builtin_neon_vld4q_lane_v: { 2781 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2782 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2783 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2784 Ops.push_back(Align); 2785 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2786 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2787 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2788 return Builder.CreateStore(Ops[1], Ops[0]); 2789 } 2790 case NEON::BI__builtin_neon_vmovl_v: { 2791 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2792 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2793 if (Usgn) 2794 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2795 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2796 } 2797 case NEON::BI__builtin_neon_vmovn_v: { 2798 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2799 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2800 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2801 } 2802 case NEON::BI__builtin_neon_vmull_v: 2803 // FIXME: the integer vmull operations could be emitted in terms of pure 2804 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2805 // hoisting the exts outside loops. Until global ISel comes along that can 2806 // see through such movement this leads to bad CodeGen. So we need an 2807 // intrinsic for now. 2808 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2809 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2810 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2811 case NEON::BI__builtin_neon_vpadal_v: 2812 case NEON::BI__builtin_neon_vpadalq_v: { 2813 // The source operand type has twice as many elements of half the size. 2814 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2815 llvm::Type *EltTy = 2816 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2817 llvm::Type *NarrowTy = 2818 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2819 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2820 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2821 } 2822 case NEON::BI__builtin_neon_vpaddl_v: 2823 case NEON::BI__builtin_neon_vpaddlq_v: { 2824 // The source operand type has twice as many elements of half the size. 2825 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2826 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2827 llvm::Type *NarrowTy = 2828 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2829 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2830 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2831 } 2832 case NEON::BI__builtin_neon_vqdmlal_v: 2833 case NEON::BI__builtin_neon_vqdmlsl_v: { 2834 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2835 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2836 MulOps, "vqdmlal"); 2837 2838 SmallVector<Value *, 2> AccumOps; 2839 AccumOps.push_back(Ops[0]); 2840 AccumOps.push_back(Mul); 2841 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2842 AccumOps, NameHint); 2843 } 2844 case NEON::BI__builtin_neon_vqshl_n_v: 2845 case NEON::BI__builtin_neon_vqshlq_n_v: 2846 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2847 1, false); 2848 case NEON::BI__builtin_neon_vrecpe_v: 2849 case NEON::BI__builtin_neon_vrecpeq_v: 2850 case NEON::BI__builtin_neon_vrsqrte_v: 2851 case NEON::BI__builtin_neon_vrsqrteq_v: 2852 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2853 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2854 2855 case NEON::BI__builtin_neon_vshl_n_v: 2856 case NEON::BI__builtin_neon_vshlq_n_v: 2857 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2858 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2859 "vshl_n"); 2860 case NEON::BI__builtin_neon_vshll_n_v: { 2861 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2862 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2863 if (Usgn) 2864 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2865 else 2866 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2867 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2868 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2869 } 2870 case NEON::BI__builtin_neon_vshrn_n_v: { 2871 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2872 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2873 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2874 if (Usgn) 2875 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2876 else 2877 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2878 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2879 } 2880 case NEON::BI__builtin_neon_vshr_n_v: 2881 case NEON::BI__builtin_neon_vshrq_n_v: 2882 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2883 case NEON::BI__builtin_neon_vst1_v: 2884 case NEON::BI__builtin_neon_vst1q_v: 2885 case NEON::BI__builtin_neon_vst2_v: 2886 case NEON::BI__builtin_neon_vst2q_v: 2887 case NEON::BI__builtin_neon_vst3_v: 2888 case NEON::BI__builtin_neon_vst3q_v: 2889 case NEON::BI__builtin_neon_vst4_v: 2890 case NEON::BI__builtin_neon_vst4q_v: 2891 case NEON::BI__builtin_neon_vst2_lane_v: 2892 case NEON::BI__builtin_neon_vst2q_lane_v: 2893 case NEON::BI__builtin_neon_vst3_lane_v: 2894 case NEON::BI__builtin_neon_vst3q_lane_v: 2895 case NEON::BI__builtin_neon_vst4_lane_v: 2896 case NEON::BI__builtin_neon_vst4q_lane_v: 2897 Ops.push_back(Align); 2898 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 2899 case NEON::BI__builtin_neon_vsubhn_v: { 2900 llvm::VectorType *SrcTy = 2901 llvm::VectorType::getExtendedElementVectorType(VTy); 2902 2903 // %sum = add <4 x i32> %lhs, %rhs 2904 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2905 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2906 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2907 2908 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2909 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2910 SrcTy->getScalarSizeInBits() / 2); 2911 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2912 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2913 2914 // %res = trunc <4 x i32> %high to <4 x i16> 2915 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2916 } 2917 case NEON::BI__builtin_neon_vtrn_v: 2918 case NEON::BI__builtin_neon_vtrnq_v: { 2919 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2920 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2921 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2922 Value *SV = nullptr; 2923 2924 for (unsigned vi = 0; vi != 2; ++vi) { 2925 SmallVector<Constant*, 16> Indices; 2926 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2927 Indices.push_back(Builder.getInt32(i+vi)); 2928 Indices.push_back(Builder.getInt32(i+e+vi)); 2929 } 2930 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2931 SV = llvm::ConstantVector::get(Indices); 2932 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2933 SV = Builder.CreateStore(SV, Addr); 2934 } 2935 return SV; 2936 } 2937 case NEON::BI__builtin_neon_vtst_v: 2938 case NEON::BI__builtin_neon_vtstq_v: { 2939 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2940 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2941 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2942 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2943 ConstantAggregateZero::get(Ty)); 2944 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2945 } 2946 case NEON::BI__builtin_neon_vuzp_v: 2947 case NEON::BI__builtin_neon_vuzpq_v: { 2948 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2949 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2950 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2951 Value *SV = nullptr; 2952 2953 for (unsigned vi = 0; vi != 2; ++vi) { 2954 SmallVector<Constant*, 16> Indices; 2955 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2956 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2957 2958 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2959 SV = llvm::ConstantVector::get(Indices); 2960 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2961 SV = Builder.CreateStore(SV, Addr); 2962 } 2963 return SV; 2964 } 2965 case NEON::BI__builtin_neon_vzip_v: 2966 case NEON::BI__builtin_neon_vzipq_v: { 2967 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2968 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2969 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2970 Value *SV = nullptr; 2971 2972 for (unsigned vi = 0; vi != 2; ++vi) { 2973 SmallVector<Constant*, 16> Indices; 2974 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2975 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2976 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2977 } 2978 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2979 SV = llvm::ConstantVector::get(Indices); 2980 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2981 SV = Builder.CreateStore(SV, Addr); 2982 } 2983 return SV; 2984 } 2985 } 2986 2987 assert(Int && "Expected valid intrinsic number"); 2988 2989 // Determine the type(s) of this overloaded AArch64 intrinsic. 2990 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 2991 2992 Value *Result = EmitNeonCall(F, Ops, NameHint); 2993 llvm::Type *ResultType = ConvertType(E->getType()); 2994 // AArch64 intrinsic one-element vector type cast to 2995 // scalar type expected by the builtin 2996 return Builder.CreateBitCast(Result, ResultType, NameHint); 2997 } 2998 2999 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3000 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3001 const CmpInst::Predicate Ip, const Twine &Name) { 3002 llvm::Type *OTy = Op->getType(); 3003 3004 // FIXME: this is utterly horrific. We should not be looking at previous 3005 // codegen context to find out what needs doing. Unfortunately TableGen 3006 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3007 // (etc). 3008 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3009 OTy = BI->getOperand(0)->getType(); 3010 3011 Op = Builder.CreateBitCast(Op, OTy); 3012 if (OTy->getScalarType()->isFloatingPointTy()) { 3013 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3014 } else { 3015 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3016 } 3017 return Builder.CreateSExt(Op, Ty, Name); 3018 } 3019 3020 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3021 Value *ExtOp, Value *IndexOp, 3022 llvm::Type *ResTy, unsigned IntID, 3023 const char *Name) { 3024 SmallVector<Value *, 2> TblOps; 3025 if (ExtOp) 3026 TblOps.push_back(ExtOp); 3027 3028 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3029 SmallVector<Constant*, 16> Indices; 3030 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3031 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3032 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3033 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3034 } 3035 Value *SV = llvm::ConstantVector::get(Indices); 3036 3037 int PairPos = 0, End = Ops.size() - 1; 3038 while (PairPos < End) { 3039 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3040 Ops[PairPos+1], SV, Name)); 3041 PairPos += 2; 3042 } 3043 3044 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3045 // of the 128-bit lookup table with zero. 3046 if (PairPos == End) { 3047 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3048 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3049 ZeroTbl, SV, Name)); 3050 } 3051 3052 Function *TblF; 3053 TblOps.push_back(IndexOp); 3054 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3055 3056 return CGF.EmitNeonCall(TblF, TblOps, Name); 3057 } 3058 3059 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3060 const CallExpr *E) { 3061 unsigned HintID = static_cast<unsigned>(-1); 3062 switch (BuiltinID) { 3063 default: break; 3064 case ARM::BI__yield: 3065 HintID = 1; 3066 break; 3067 case ARM::BI__wfe: 3068 HintID = 2; 3069 break; 3070 case ARM::BI__wfi: 3071 HintID = 3; 3072 break; 3073 case ARM::BI__sev: 3074 HintID = 4; 3075 break; 3076 case ARM::BI__sevl: 3077 HintID = 5; 3078 break; 3079 } 3080 3081 if (HintID != static_cast<unsigned>(-1)) { 3082 Function *F = CGM.getIntrinsic(Intrinsic::arm_hint); 3083 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3084 } 3085 3086 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3087 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3088 EmitScalarExpr(E->getArg(0)), 3089 "rbit"); 3090 } 3091 3092 if (BuiltinID == ARM::BI__clear_cache) { 3093 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3094 const FunctionDecl *FD = E->getDirectCallee(); 3095 SmallVector<Value*, 2> Ops; 3096 for (unsigned i = 0; i < 2; i++) 3097 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3098 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3099 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3100 StringRef Name = FD->getName(); 3101 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3102 } 3103 3104 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3105 (BuiltinID == ARM::BI__builtin_arm_ldrex && 3106 getContext().getTypeSize(E->getType()) == 64)) { 3107 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3108 3109 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3110 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3111 "ldrexd"); 3112 3113 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3114 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3115 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3116 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3117 3118 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3119 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3120 Val = Builder.CreateOr(Val, Val1); 3121 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3122 } 3123 3124 if (BuiltinID == ARM::BI__builtin_arm_ldrex) { 3125 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3126 3127 QualType Ty = E->getType(); 3128 llvm::Type *RealResTy = ConvertType(Ty); 3129 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3130 getContext().getTypeSize(Ty)); 3131 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3132 3133 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrex, LoadAddr->getType()); 3134 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3135 3136 if (RealResTy->isPointerTy()) 3137 return Builder.CreateIntToPtr(Val, RealResTy); 3138 else { 3139 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3140 return Builder.CreateBitCast(Val, RealResTy); 3141 } 3142 } 3143 3144 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3145 (BuiltinID == ARM::BI__builtin_arm_strex && 3146 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3147 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 3148 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 3149 3150 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3151 Value *Val = EmitScalarExpr(E->getArg(0)); 3152 Builder.CreateStore(Val, Tmp); 3153 3154 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3155 Val = Builder.CreateLoad(LdPtr); 3156 3157 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3158 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3159 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3160 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3161 } 3162 3163 if (BuiltinID == ARM::BI__builtin_arm_strex) { 3164 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3165 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3166 3167 QualType Ty = E->getArg(0)->getType(); 3168 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3169 getContext().getTypeSize(Ty)); 3170 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3171 3172 if (StoreVal->getType()->isPointerTy()) 3173 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3174 else { 3175 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3176 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3177 } 3178 3179 Function *F = CGM.getIntrinsic(Intrinsic::arm_strex, StoreAddr->getType()); 3180 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3181 } 3182 3183 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3184 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3185 return Builder.CreateCall(F); 3186 } 3187 3188 // CRC32 3189 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3190 switch (BuiltinID) { 3191 case ARM::BI__builtin_arm_crc32b: 3192 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3193 case ARM::BI__builtin_arm_crc32cb: 3194 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3195 case ARM::BI__builtin_arm_crc32h: 3196 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3197 case ARM::BI__builtin_arm_crc32ch: 3198 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3199 case ARM::BI__builtin_arm_crc32w: 3200 case ARM::BI__builtin_arm_crc32d: 3201 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3202 case ARM::BI__builtin_arm_crc32cw: 3203 case ARM::BI__builtin_arm_crc32cd: 3204 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3205 } 3206 3207 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3208 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3209 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3210 3211 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3212 // intrinsics, hence we need different codegen for these cases. 3213 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3214 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3215 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3216 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3217 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3218 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3219 3220 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3221 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3222 return Builder.CreateCall2(F, Res, Arg1b); 3223 } else { 3224 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3225 3226 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3227 return Builder.CreateCall2(F, Arg0, Arg1); 3228 } 3229 } 3230 3231 SmallVector<Value*, 4> Ops; 3232 llvm::Value *Align = nullptr; 3233 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3234 if (i == 0) { 3235 switch (BuiltinID) { 3236 case NEON::BI__builtin_neon_vld1_v: 3237 case NEON::BI__builtin_neon_vld1q_v: 3238 case NEON::BI__builtin_neon_vld1q_lane_v: 3239 case NEON::BI__builtin_neon_vld1_lane_v: 3240 case NEON::BI__builtin_neon_vld1_dup_v: 3241 case NEON::BI__builtin_neon_vld1q_dup_v: 3242 case NEON::BI__builtin_neon_vst1_v: 3243 case NEON::BI__builtin_neon_vst1q_v: 3244 case NEON::BI__builtin_neon_vst1q_lane_v: 3245 case NEON::BI__builtin_neon_vst1_lane_v: 3246 case NEON::BI__builtin_neon_vst2_v: 3247 case NEON::BI__builtin_neon_vst2q_v: 3248 case NEON::BI__builtin_neon_vst2_lane_v: 3249 case NEON::BI__builtin_neon_vst2q_lane_v: 3250 case NEON::BI__builtin_neon_vst3_v: 3251 case NEON::BI__builtin_neon_vst3q_v: 3252 case NEON::BI__builtin_neon_vst3_lane_v: 3253 case NEON::BI__builtin_neon_vst3q_lane_v: 3254 case NEON::BI__builtin_neon_vst4_v: 3255 case NEON::BI__builtin_neon_vst4q_v: 3256 case NEON::BI__builtin_neon_vst4_lane_v: 3257 case NEON::BI__builtin_neon_vst4q_lane_v: 3258 // Get the alignment for the argument in addition to the value; 3259 // we'll use it later. 3260 std::pair<llvm::Value*, unsigned> Src = 3261 EmitPointerWithAlignment(E->getArg(0)); 3262 Ops.push_back(Src.first); 3263 Align = Builder.getInt32(Src.second); 3264 continue; 3265 } 3266 } 3267 if (i == 1) { 3268 switch (BuiltinID) { 3269 case NEON::BI__builtin_neon_vld2_v: 3270 case NEON::BI__builtin_neon_vld2q_v: 3271 case NEON::BI__builtin_neon_vld3_v: 3272 case NEON::BI__builtin_neon_vld3q_v: 3273 case NEON::BI__builtin_neon_vld4_v: 3274 case NEON::BI__builtin_neon_vld4q_v: 3275 case NEON::BI__builtin_neon_vld2_lane_v: 3276 case NEON::BI__builtin_neon_vld2q_lane_v: 3277 case NEON::BI__builtin_neon_vld3_lane_v: 3278 case NEON::BI__builtin_neon_vld3q_lane_v: 3279 case NEON::BI__builtin_neon_vld4_lane_v: 3280 case NEON::BI__builtin_neon_vld4q_lane_v: 3281 case NEON::BI__builtin_neon_vld2_dup_v: 3282 case NEON::BI__builtin_neon_vld3_dup_v: 3283 case NEON::BI__builtin_neon_vld4_dup_v: 3284 // Get the alignment for the argument in addition to the value; 3285 // we'll use it later. 3286 std::pair<llvm::Value*, unsigned> Src = 3287 EmitPointerWithAlignment(E->getArg(1)); 3288 Ops.push_back(Src.first); 3289 Align = Builder.getInt32(Src.second); 3290 continue; 3291 } 3292 } 3293 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3294 } 3295 3296 switch (BuiltinID) { 3297 default: break; 3298 // vget_lane and vset_lane are not overloaded and do not have an extra 3299 // argument that specifies the vector type. 3300 case NEON::BI__builtin_neon_vget_lane_i8: 3301 case NEON::BI__builtin_neon_vget_lane_i16: 3302 case NEON::BI__builtin_neon_vget_lane_i32: 3303 case NEON::BI__builtin_neon_vget_lane_i64: 3304 case NEON::BI__builtin_neon_vget_lane_f32: 3305 case NEON::BI__builtin_neon_vgetq_lane_i8: 3306 case NEON::BI__builtin_neon_vgetq_lane_i16: 3307 case NEON::BI__builtin_neon_vgetq_lane_i32: 3308 case NEON::BI__builtin_neon_vgetq_lane_i64: 3309 case NEON::BI__builtin_neon_vgetq_lane_f32: 3310 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3311 "vget_lane"); 3312 case NEON::BI__builtin_neon_vset_lane_i8: 3313 case NEON::BI__builtin_neon_vset_lane_i16: 3314 case NEON::BI__builtin_neon_vset_lane_i32: 3315 case NEON::BI__builtin_neon_vset_lane_i64: 3316 case NEON::BI__builtin_neon_vset_lane_f32: 3317 case NEON::BI__builtin_neon_vsetq_lane_i8: 3318 case NEON::BI__builtin_neon_vsetq_lane_i16: 3319 case NEON::BI__builtin_neon_vsetq_lane_i32: 3320 case NEON::BI__builtin_neon_vsetq_lane_i64: 3321 case NEON::BI__builtin_neon_vsetq_lane_f32: 3322 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3323 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3324 3325 // Non-polymorphic crypto instructions also not overloaded 3326 case NEON::BI__builtin_neon_vsha1h_u32: 3327 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3328 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3329 "vsha1h"); 3330 case NEON::BI__builtin_neon_vsha1cq_u32: 3331 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3332 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3333 "vsha1h"); 3334 case NEON::BI__builtin_neon_vsha1pq_u32: 3335 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3336 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3337 "vsha1h"); 3338 case NEON::BI__builtin_neon_vsha1mq_u32: 3339 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3340 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3341 "vsha1h"); 3342 } 3343 3344 // Get the last argument, which specifies the vector type. 3345 llvm::APSInt Result; 3346 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3347 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3348 return nullptr; 3349 3350 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3351 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3352 // Determine the overloaded type of this builtin. 3353 llvm::Type *Ty; 3354 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3355 Ty = FloatTy; 3356 else 3357 Ty = DoubleTy; 3358 3359 // Determine whether this is an unsigned conversion or not. 3360 bool usgn = Result.getZExtValue() == 1; 3361 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3362 3363 // Call the appropriate intrinsic. 3364 Function *F = CGM.getIntrinsic(Int, Ty); 3365 return Builder.CreateCall(F, Ops, "vcvtr"); 3366 } 3367 3368 // Determine the type of this overloaded NEON intrinsic. 3369 NeonTypeFlags Type(Result.getZExtValue()); 3370 bool usgn = Type.isUnsigned(); 3371 bool rightShift = false; 3372 3373 llvm::VectorType *VTy = GetNeonType(this, Type); 3374 llvm::Type *Ty = VTy; 3375 if (!Ty) 3376 return nullptr; 3377 3378 // Many NEON builtins have identical semantics and uses in ARM and 3379 // AArch64. Emit these in a single function. 3380 llvm::ArrayRef<NeonIntrinsicInfo> IntrinsicMap(ARMSIMDIntrinsicMap); 3381 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3382 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3383 if (Builtin) 3384 return EmitCommonNeonBuiltinExpr( 3385 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3386 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3387 3388 unsigned Int; 3389 switch (BuiltinID) { 3390 default: return nullptr; 3391 case NEON::BI__builtin_neon_vld1q_lane_v: 3392 // Handle 64-bit integer elements as a special case. Use shuffles of 3393 // one-element vectors to avoid poor code for i64 in the backend. 3394 if (VTy->getElementType()->isIntegerTy(64)) { 3395 // Extract the other lane. 3396 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3397 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3398 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3399 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3400 // Load the value as a one-element vector. 3401 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3402 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3403 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3404 // Combine them. 3405 SmallVector<Constant*, 2> Indices; 3406 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3407 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3408 SV = llvm::ConstantVector::get(Indices); 3409 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3410 } 3411 // fall through 3412 case NEON::BI__builtin_neon_vld1_lane_v: { 3413 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3414 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3415 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3416 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3417 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3418 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3419 } 3420 case NEON::BI__builtin_neon_vld2_dup_v: 3421 case NEON::BI__builtin_neon_vld3_dup_v: 3422 case NEON::BI__builtin_neon_vld4_dup_v: { 3423 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3424 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3425 switch (BuiltinID) { 3426 case NEON::BI__builtin_neon_vld2_dup_v: 3427 Int = Intrinsic::arm_neon_vld2; 3428 break; 3429 case NEON::BI__builtin_neon_vld3_dup_v: 3430 Int = Intrinsic::arm_neon_vld3; 3431 break; 3432 case NEON::BI__builtin_neon_vld4_dup_v: 3433 Int = Intrinsic::arm_neon_vld4; 3434 break; 3435 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3436 } 3437 Function *F = CGM.getIntrinsic(Int, Ty); 3438 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3439 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3440 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3441 return Builder.CreateStore(Ops[1], Ops[0]); 3442 } 3443 switch (BuiltinID) { 3444 case NEON::BI__builtin_neon_vld2_dup_v: 3445 Int = Intrinsic::arm_neon_vld2lane; 3446 break; 3447 case NEON::BI__builtin_neon_vld3_dup_v: 3448 Int = Intrinsic::arm_neon_vld3lane; 3449 break; 3450 case NEON::BI__builtin_neon_vld4_dup_v: 3451 Int = Intrinsic::arm_neon_vld4lane; 3452 break; 3453 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3454 } 3455 Function *F = CGM.getIntrinsic(Int, Ty); 3456 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3457 3458 SmallVector<Value*, 6> Args; 3459 Args.push_back(Ops[1]); 3460 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3461 3462 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3463 Args.push_back(CI); 3464 Args.push_back(Align); 3465 3466 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3467 // splat lane 0 to all elts in each vector of the result. 3468 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3469 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3470 Value *Elt = Builder.CreateBitCast(Val, Ty); 3471 Elt = EmitNeonSplat(Elt, CI); 3472 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3473 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3474 } 3475 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3476 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3477 return Builder.CreateStore(Ops[1], Ops[0]); 3478 } 3479 case NEON::BI__builtin_neon_vqrshrn_n_v: 3480 Int = 3481 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3482 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3483 1, true); 3484 case NEON::BI__builtin_neon_vqrshrun_n_v: 3485 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3486 Ops, "vqrshrun_n", 1, true); 3487 case NEON::BI__builtin_neon_vqshlu_n_v: 3488 case NEON::BI__builtin_neon_vqshluq_n_v: 3489 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 3490 Ops, "vqshlu", 1, false); 3491 case NEON::BI__builtin_neon_vqshrn_n_v: 3492 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3493 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3494 1, true); 3495 case NEON::BI__builtin_neon_vqshrun_n_v: 3496 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3497 Ops, "vqshrun_n", 1, true); 3498 case NEON::BI__builtin_neon_vrecpe_v: 3499 case NEON::BI__builtin_neon_vrecpeq_v: 3500 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3501 Ops, "vrecpe"); 3502 case NEON::BI__builtin_neon_vrshrn_n_v: 3503 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3504 Ops, "vrshrn_n", 1, true); 3505 case NEON::BI__builtin_neon_vrshr_n_v: 3506 case NEON::BI__builtin_neon_vrshrq_n_v: 3507 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3508 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 3509 case NEON::BI__builtin_neon_vrsra_n_v: 3510 case NEON::BI__builtin_neon_vrsraq_n_v: 3511 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3512 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3513 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3514 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3515 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3516 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3517 case NEON::BI__builtin_neon_vsri_n_v: 3518 case NEON::BI__builtin_neon_vsriq_n_v: 3519 rightShift = true; 3520 case NEON::BI__builtin_neon_vsli_n_v: 3521 case NEON::BI__builtin_neon_vsliq_n_v: 3522 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3523 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3524 Ops, "vsli_n"); 3525 case NEON::BI__builtin_neon_vsra_n_v: 3526 case NEON::BI__builtin_neon_vsraq_n_v: 3527 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3528 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3529 return Builder.CreateAdd(Ops[0], Ops[1]); 3530 case NEON::BI__builtin_neon_vst1q_lane_v: 3531 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3532 // a one-element vector and avoid poor code for i64 in the backend. 3533 if (VTy->getElementType()->isIntegerTy(64)) { 3534 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3535 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3536 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3537 Ops[2] = Align; 3538 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3539 Ops[1]->getType()), Ops); 3540 } 3541 // fall through 3542 case NEON::BI__builtin_neon_vst1_lane_v: { 3543 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3544 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3545 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3546 StoreInst *St = Builder.CreateStore(Ops[1], 3547 Builder.CreateBitCast(Ops[0], Ty)); 3548 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3549 return St; 3550 } 3551 case NEON::BI__builtin_neon_vtbl1_v: 3552 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3553 Ops, "vtbl1"); 3554 case NEON::BI__builtin_neon_vtbl2_v: 3555 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3556 Ops, "vtbl2"); 3557 case NEON::BI__builtin_neon_vtbl3_v: 3558 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3559 Ops, "vtbl3"); 3560 case NEON::BI__builtin_neon_vtbl4_v: 3561 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3562 Ops, "vtbl4"); 3563 case NEON::BI__builtin_neon_vtbx1_v: 3564 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3565 Ops, "vtbx1"); 3566 case NEON::BI__builtin_neon_vtbx2_v: 3567 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3568 Ops, "vtbx2"); 3569 case NEON::BI__builtin_neon_vtbx3_v: 3570 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3571 Ops, "vtbx3"); 3572 case NEON::BI__builtin_neon_vtbx4_v: 3573 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3574 Ops, "vtbx4"); 3575 } 3576 } 3577 3578 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3579 const CallExpr *E, 3580 SmallVectorImpl<Value *> &Ops) { 3581 unsigned int Int = 0; 3582 const char *s = nullptr; 3583 3584 switch (BuiltinID) { 3585 default: 3586 return nullptr; 3587 case NEON::BI__builtin_neon_vtbl1_v: 3588 case NEON::BI__builtin_neon_vqtbl1_v: 3589 case NEON::BI__builtin_neon_vqtbl1q_v: 3590 case NEON::BI__builtin_neon_vtbl2_v: 3591 case NEON::BI__builtin_neon_vqtbl2_v: 3592 case NEON::BI__builtin_neon_vqtbl2q_v: 3593 case NEON::BI__builtin_neon_vtbl3_v: 3594 case NEON::BI__builtin_neon_vqtbl3_v: 3595 case NEON::BI__builtin_neon_vqtbl3q_v: 3596 case NEON::BI__builtin_neon_vtbl4_v: 3597 case NEON::BI__builtin_neon_vqtbl4_v: 3598 case NEON::BI__builtin_neon_vqtbl4q_v: 3599 break; 3600 case NEON::BI__builtin_neon_vtbx1_v: 3601 case NEON::BI__builtin_neon_vqtbx1_v: 3602 case NEON::BI__builtin_neon_vqtbx1q_v: 3603 case NEON::BI__builtin_neon_vtbx2_v: 3604 case NEON::BI__builtin_neon_vqtbx2_v: 3605 case NEON::BI__builtin_neon_vqtbx2q_v: 3606 case NEON::BI__builtin_neon_vtbx3_v: 3607 case NEON::BI__builtin_neon_vqtbx3_v: 3608 case NEON::BI__builtin_neon_vqtbx3q_v: 3609 case NEON::BI__builtin_neon_vtbx4_v: 3610 case NEON::BI__builtin_neon_vqtbx4_v: 3611 case NEON::BI__builtin_neon_vqtbx4q_v: 3612 break; 3613 } 3614 3615 assert(E->getNumArgs() >= 3); 3616 3617 // Get the last argument, which specifies the vector type. 3618 llvm::APSInt Result; 3619 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3620 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3621 return nullptr; 3622 3623 // Determine the type of this overloaded NEON intrinsic. 3624 NeonTypeFlags Type(Result.getZExtValue()); 3625 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3626 llvm::Type *Ty = VTy; 3627 if (!Ty) 3628 return nullptr; 3629 3630 unsigned nElts = VTy->getNumElements(); 3631 3632 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3633 3634 // AArch64 scalar builtins are not overloaded, they do not have an extra 3635 // argument that specifies the vector type, need to handle each case. 3636 SmallVector<Value *, 2> TblOps; 3637 switch (BuiltinID) { 3638 case NEON::BI__builtin_neon_vtbl1_v: { 3639 TblOps.push_back(Ops[0]); 3640 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3641 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3642 } 3643 case NEON::BI__builtin_neon_vtbl2_v: { 3644 TblOps.push_back(Ops[0]); 3645 TblOps.push_back(Ops[1]); 3646 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3647 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3648 } 3649 case NEON::BI__builtin_neon_vtbl3_v: { 3650 TblOps.push_back(Ops[0]); 3651 TblOps.push_back(Ops[1]); 3652 TblOps.push_back(Ops[2]); 3653 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3654 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3655 } 3656 case NEON::BI__builtin_neon_vtbl4_v: { 3657 TblOps.push_back(Ops[0]); 3658 TblOps.push_back(Ops[1]); 3659 TblOps.push_back(Ops[2]); 3660 TblOps.push_back(Ops[3]); 3661 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3662 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3663 } 3664 case NEON::BI__builtin_neon_vtbx1_v: { 3665 TblOps.push_back(Ops[1]); 3666 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3667 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3668 3669 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3670 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3671 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3672 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3673 3674 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3675 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3676 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3677 } 3678 case NEON::BI__builtin_neon_vtbx2_v: { 3679 TblOps.push_back(Ops[1]); 3680 TblOps.push_back(Ops[2]); 3681 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3682 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3683 } 3684 case NEON::BI__builtin_neon_vtbx3_v: { 3685 TblOps.push_back(Ops[1]); 3686 TblOps.push_back(Ops[2]); 3687 TblOps.push_back(Ops[3]); 3688 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3689 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3690 3691 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3692 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3693 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3694 TwentyFourV); 3695 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3696 3697 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3698 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3699 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3700 } 3701 case NEON::BI__builtin_neon_vtbx4_v: { 3702 TblOps.push_back(Ops[1]); 3703 TblOps.push_back(Ops[2]); 3704 TblOps.push_back(Ops[3]); 3705 TblOps.push_back(Ops[4]); 3706 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3707 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3708 } 3709 case NEON::BI__builtin_neon_vqtbl1_v: 3710 case NEON::BI__builtin_neon_vqtbl1q_v: 3711 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3712 case NEON::BI__builtin_neon_vqtbl2_v: 3713 case NEON::BI__builtin_neon_vqtbl2q_v: { 3714 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3715 case NEON::BI__builtin_neon_vqtbl3_v: 3716 case NEON::BI__builtin_neon_vqtbl3q_v: 3717 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3718 case NEON::BI__builtin_neon_vqtbl4_v: 3719 case NEON::BI__builtin_neon_vqtbl4q_v: 3720 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3721 case NEON::BI__builtin_neon_vqtbx1_v: 3722 case NEON::BI__builtin_neon_vqtbx1q_v: 3723 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3724 case NEON::BI__builtin_neon_vqtbx2_v: 3725 case NEON::BI__builtin_neon_vqtbx2q_v: 3726 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3727 case NEON::BI__builtin_neon_vqtbx3_v: 3728 case NEON::BI__builtin_neon_vqtbx3q_v: 3729 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3730 case NEON::BI__builtin_neon_vqtbx4_v: 3731 case NEON::BI__builtin_neon_vqtbx4q_v: 3732 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3733 } 3734 } 3735 3736 if (!Int) 3737 return nullptr; 3738 3739 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3740 return CGF.EmitNeonCall(F, Ops, s); 3741 } 3742 3743 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3744 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3745 Op = Builder.CreateBitCast(Op, Int16Ty); 3746 Value *V = UndefValue::get(VTy); 3747 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3748 Op = Builder.CreateInsertElement(V, Op, CI); 3749 return Op; 3750 } 3751 3752 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3753 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3754 Op = Builder.CreateBitCast(Op, Int8Ty); 3755 Value *V = UndefValue::get(VTy); 3756 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3757 Op = Builder.CreateInsertElement(V, Op, CI); 3758 return Op; 3759 } 3760 3761 Value *CodeGenFunction:: 3762 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3763 const char *Name) { 3764 // i8 is not a legal types for AArch64, so we can't just use 3765 // a normal overloaed intrinsic call for these scalar types. Instead 3766 // we'll build 64-bit vectors w/ lane zero being our input values and 3767 // perform the operation on that. The back end can pattern match directly 3768 // to the scalar instruction. 3769 Ops[0] = vectorWrapScalar8(Ops[0]); 3770 Ops[1] = vectorWrapScalar8(Ops[1]); 3771 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3772 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3773 Constant *CI = ConstantInt::get(SizeTy, 0); 3774 return Builder.CreateExtractElement(V, CI, "lane0"); 3775 } 3776 3777 Value *CodeGenFunction:: 3778 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3779 const char *Name) { 3780 // i16 is not a legal types for AArch64, so we can't just use 3781 // a normal overloaed intrinsic call for these scalar types. Instead 3782 // we'll build 64-bit vectors w/ lane zero being our input values and 3783 // perform the operation on that. The back end can pattern match directly 3784 // to the scalar instruction. 3785 Ops[0] = vectorWrapScalar16(Ops[0]); 3786 Ops[1] = vectorWrapScalar16(Ops[1]); 3787 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3788 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3789 Constant *CI = ConstantInt::get(SizeTy, 0); 3790 return Builder.CreateExtractElement(V, CI, "lane0"); 3791 } 3792 3793 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3794 const CallExpr *E) { 3795 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 3796 assert((getContext().getTypeSize(E->getType()) == 32) && 3797 "rbit of unusual size!"); 3798 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3799 return Builder.CreateCall( 3800 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3801 } 3802 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 3803 assert((getContext().getTypeSize(E->getType()) == 64) && 3804 "rbit of unusual size!"); 3805 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3806 return Builder.CreateCall( 3807 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3808 } 3809 3810 if (BuiltinID == AArch64::BI__clear_cache) { 3811 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3812 const FunctionDecl *FD = E->getDirectCallee(); 3813 SmallVector<Value*, 2> Ops; 3814 for (unsigned i = 0; i < 2; i++) 3815 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3816 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3817 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3818 StringRef Name = FD->getName(); 3819 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3820 } 3821 3822 if (BuiltinID == AArch64::BI__builtin_arm_ldrex && 3823 getContext().getTypeSize(E->getType()) == 128) { 3824 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_ldxp); 3825 3826 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3827 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3828 "ldxp"); 3829 3830 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3831 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3832 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 3833 Val0 = Builder.CreateZExt(Val0, Int128Ty); 3834 Val1 = Builder.CreateZExt(Val1, Int128Ty); 3835 3836 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 3837 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3838 Val = Builder.CreateOr(Val, Val1); 3839 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3840 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex) { 3841 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3842 3843 QualType Ty = E->getType(); 3844 llvm::Type *RealResTy = ConvertType(Ty); 3845 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3846 getContext().getTypeSize(Ty)); 3847 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3848 3849 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_ldxr, LoadAddr->getType()); 3850 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 3851 3852 if (RealResTy->isPointerTy()) 3853 return Builder.CreateIntToPtr(Val, RealResTy); 3854 3855 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3856 return Builder.CreateBitCast(Val, RealResTy); 3857 } 3858 3859 if (BuiltinID == AArch64::BI__builtin_arm_strex && 3860 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 3861 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_stxp); 3862 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, NULL); 3863 3864 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 3865 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 3866 One); 3867 Value *Val = EmitScalarExpr(E->getArg(0)); 3868 Builder.CreateStore(Val, Tmp); 3869 3870 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3871 Val = Builder.CreateLoad(LdPtr); 3872 3873 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3874 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3875 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 3876 Int8PtrTy); 3877 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 3878 } else if (BuiltinID == AArch64::BI__builtin_arm_strex) { 3879 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3880 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3881 3882 QualType Ty = E->getArg(0)->getType(); 3883 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3884 getContext().getTypeSize(Ty)); 3885 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3886 3887 if (StoreVal->getType()->isPointerTy()) 3888 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 3889 else { 3890 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3891 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 3892 } 3893 3894 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_stxr, StoreAddr->getType()); 3895 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 3896 } 3897 3898 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 3899 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 3900 return Builder.CreateCall(F); 3901 } 3902 3903 // CRC32 3904 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3905 switch (BuiltinID) { 3906 case AArch64::BI__builtin_arm_crc32b: 3907 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 3908 case AArch64::BI__builtin_arm_crc32cb: 3909 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 3910 case AArch64::BI__builtin_arm_crc32h: 3911 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 3912 case AArch64::BI__builtin_arm_crc32ch: 3913 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 3914 case AArch64::BI__builtin_arm_crc32w: 3915 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 3916 case AArch64::BI__builtin_arm_crc32cw: 3917 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 3918 case AArch64::BI__builtin_arm_crc32d: 3919 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 3920 case AArch64::BI__builtin_arm_crc32cd: 3921 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 3922 } 3923 3924 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3925 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3926 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3927 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3928 3929 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 3930 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 3931 3932 return Builder.CreateCall2(F, Arg0, Arg1); 3933 } 3934 3935 llvm::SmallVector<Value*, 4> Ops; 3936 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 3937 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3938 3939 llvm::ArrayRef<NeonIntrinsicInfo> SISDMap(AArch64SISDIntrinsicMap); 3940 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3941 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 3942 3943 if (Builtin) { 3944 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 3945 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 3946 assert(Result && "SISD intrinsic should have been handled"); 3947 return Result; 3948 } 3949 3950 llvm::APSInt Result; 3951 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3952 NeonTypeFlags Type(0); 3953 if (Arg->isIntegerConstantExpr(Result, getContext())) 3954 // Determine the type of this overloaded NEON intrinsic. 3955 Type = NeonTypeFlags(Result.getZExtValue()); 3956 3957 bool usgn = Type.isUnsigned(); 3958 bool quad = Type.isQuad(); 3959 3960 // Handle non-overloaded intrinsics first. 3961 switch (BuiltinID) { 3962 default: break; 3963 case NEON::BI__builtin_neon_vldrq_p128: { 3964 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 3965 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 3966 return Builder.CreateLoad(Ptr); 3967 } 3968 case NEON::BI__builtin_neon_vstrq_p128: { 3969 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 3970 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 3971 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 3972 } 3973 case NEON::BI__builtin_neon_vcvts_u32_f32: 3974 case NEON::BI__builtin_neon_vcvtd_u64_f64: 3975 usgn = true; 3976 // FALL THROUGH 3977 case NEON::BI__builtin_neon_vcvts_s32_f32: 3978 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 3979 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3980 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 3981 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 3982 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 3983 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 3984 if (usgn) 3985 return Builder.CreateFPToUI(Ops[0], InTy); 3986 return Builder.CreateFPToSI(Ops[0], InTy); 3987 } 3988 case NEON::BI__builtin_neon_vcvts_f32_u32: 3989 case NEON::BI__builtin_neon_vcvtd_f64_u64: 3990 usgn = true; 3991 // FALL THROUGH 3992 case NEON::BI__builtin_neon_vcvts_f32_s32: 3993 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 3994 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3995 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 3996 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 3997 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 3998 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 3999 if (usgn) 4000 return Builder.CreateUIToFP(Ops[0], FTy); 4001 return Builder.CreateSIToFP(Ops[0], FTy); 4002 } 4003 case NEON::BI__builtin_neon_vpaddd_s64: { 4004 llvm::Type *Ty = 4005 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4006 Value *Vec = EmitScalarExpr(E->getArg(0)); 4007 // The vector is v2f64, so make sure it's bitcast to that. 4008 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4009 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4010 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4011 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4012 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4013 // Pairwise addition of a v2f64 into a scalar f64. 4014 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4015 } 4016 case NEON::BI__builtin_neon_vpaddd_f64: { 4017 llvm::Type *Ty = 4018 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4019 Value *Vec = EmitScalarExpr(E->getArg(0)); 4020 // The vector is v2f64, so make sure it's bitcast to that. 4021 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4022 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4023 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4024 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4025 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4026 // Pairwise addition of a v2f64 into a scalar f64. 4027 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4028 } 4029 case NEON::BI__builtin_neon_vpadds_f32: { 4030 llvm::Type *Ty = 4031 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4032 Value *Vec = EmitScalarExpr(E->getArg(0)); 4033 // The vector is v2f32, so make sure it's bitcast to that. 4034 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4035 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4036 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4037 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4038 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4039 // Pairwise addition of a v2f32 into a scalar f32. 4040 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4041 } 4042 case NEON::BI__builtin_neon_vceqzd_s64: 4043 case NEON::BI__builtin_neon_vceqzd_f64: 4044 case NEON::BI__builtin_neon_vceqzs_f32: 4045 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4046 return EmitAArch64CompareBuiltinExpr( 4047 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4048 ICmpInst::ICMP_EQ, "vceqz"); 4049 case NEON::BI__builtin_neon_vcgezd_s64: 4050 case NEON::BI__builtin_neon_vcgezd_f64: 4051 case NEON::BI__builtin_neon_vcgezs_f32: 4052 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4053 return EmitAArch64CompareBuiltinExpr( 4054 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4055 ICmpInst::ICMP_SGE, "vcgez"); 4056 case NEON::BI__builtin_neon_vclezd_s64: 4057 case NEON::BI__builtin_neon_vclezd_f64: 4058 case NEON::BI__builtin_neon_vclezs_f32: 4059 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4060 return EmitAArch64CompareBuiltinExpr( 4061 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4062 ICmpInst::ICMP_SLE, "vclez"); 4063 case NEON::BI__builtin_neon_vcgtzd_s64: 4064 case NEON::BI__builtin_neon_vcgtzd_f64: 4065 case NEON::BI__builtin_neon_vcgtzs_f32: 4066 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4067 return EmitAArch64CompareBuiltinExpr( 4068 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4069 ICmpInst::ICMP_SGT, "vcgtz"); 4070 case NEON::BI__builtin_neon_vcltzd_s64: 4071 case NEON::BI__builtin_neon_vcltzd_f64: 4072 case NEON::BI__builtin_neon_vcltzs_f32: 4073 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4074 return EmitAArch64CompareBuiltinExpr( 4075 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4076 ICmpInst::ICMP_SLT, "vcltz"); 4077 4078 case NEON::BI__builtin_neon_vceqzd_u64: { 4079 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4080 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4081 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4082 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4083 llvm::Constant::getNullValue(Ty)); 4084 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4085 } 4086 case NEON::BI__builtin_neon_vceqd_f64: 4087 case NEON::BI__builtin_neon_vcled_f64: 4088 case NEON::BI__builtin_neon_vcltd_f64: 4089 case NEON::BI__builtin_neon_vcged_f64: 4090 case NEON::BI__builtin_neon_vcgtd_f64: { 4091 llvm::CmpInst::Predicate P; 4092 switch (BuiltinID) { 4093 default: llvm_unreachable("missing builtin ID in switch!"); 4094 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4095 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4096 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4097 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4098 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4099 } 4100 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4101 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4102 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4103 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4104 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4105 } 4106 case NEON::BI__builtin_neon_vceqs_f32: 4107 case NEON::BI__builtin_neon_vcles_f32: 4108 case NEON::BI__builtin_neon_vclts_f32: 4109 case NEON::BI__builtin_neon_vcges_f32: 4110 case NEON::BI__builtin_neon_vcgts_f32: { 4111 llvm::CmpInst::Predicate P; 4112 switch (BuiltinID) { 4113 default: llvm_unreachable("missing builtin ID in switch!"); 4114 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4115 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4116 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4117 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4118 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4119 } 4120 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4121 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4122 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4123 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4124 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4125 } 4126 case NEON::BI__builtin_neon_vceqd_s64: 4127 case NEON::BI__builtin_neon_vceqd_u64: 4128 case NEON::BI__builtin_neon_vcgtd_s64: 4129 case NEON::BI__builtin_neon_vcgtd_u64: 4130 case NEON::BI__builtin_neon_vcltd_s64: 4131 case NEON::BI__builtin_neon_vcltd_u64: 4132 case NEON::BI__builtin_neon_vcged_u64: 4133 case NEON::BI__builtin_neon_vcged_s64: 4134 case NEON::BI__builtin_neon_vcled_u64: 4135 case NEON::BI__builtin_neon_vcled_s64: { 4136 llvm::CmpInst::Predicate P; 4137 switch (BuiltinID) { 4138 default: llvm_unreachable("missing builtin ID in switch!"); 4139 case NEON::BI__builtin_neon_vceqd_s64: 4140 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4141 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4142 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4143 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4144 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4145 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4146 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4147 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4148 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4149 } 4150 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4151 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4152 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4153 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4154 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4155 } 4156 case NEON::BI__builtin_neon_vtstd_s64: 4157 case NEON::BI__builtin_neon_vtstd_u64: { 4158 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4159 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4160 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4161 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4162 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4163 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4164 llvm::Constant::getNullValue(Ty)); 4165 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4166 } 4167 case NEON::BI__builtin_neon_vset_lane_i8: 4168 case NEON::BI__builtin_neon_vset_lane_i16: 4169 case NEON::BI__builtin_neon_vset_lane_i32: 4170 case NEON::BI__builtin_neon_vset_lane_i64: 4171 case NEON::BI__builtin_neon_vset_lane_f32: 4172 case NEON::BI__builtin_neon_vsetq_lane_i8: 4173 case NEON::BI__builtin_neon_vsetq_lane_i16: 4174 case NEON::BI__builtin_neon_vsetq_lane_i32: 4175 case NEON::BI__builtin_neon_vsetq_lane_i64: 4176 case NEON::BI__builtin_neon_vsetq_lane_f32: 4177 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4178 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4179 case NEON::BI__builtin_neon_vset_lane_f64: 4180 // The vector type needs a cast for the v1f64 variant. 4181 Ops[1] = Builder.CreateBitCast(Ops[1], 4182 llvm::VectorType::get(DoubleTy, 1)); 4183 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4184 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4185 case NEON::BI__builtin_neon_vsetq_lane_f64: 4186 // The vector type needs a cast for the v2f64 variant. 4187 Ops[1] = Builder.CreateBitCast(Ops[1], 4188 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4189 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4190 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4191 4192 case NEON::BI__builtin_neon_vget_lane_i8: 4193 case NEON::BI__builtin_neon_vdupb_lane_i8: 4194 Ops[0] = Builder.CreateBitCast(Ops[0], 4195 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4196 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4197 "vget_lane"); 4198 case NEON::BI__builtin_neon_vgetq_lane_i8: 4199 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4200 Ops[0] = Builder.CreateBitCast(Ops[0], 4201 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4202 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4203 "vgetq_lane"); 4204 case NEON::BI__builtin_neon_vget_lane_i16: 4205 case NEON::BI__builtin_neon_vduph_lane_i16: 4206 Ops[0] = Builder.CreateBitCast(Ops[0], 4207 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4208 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4209 "vget_lane"); 4210 case NEON::BI__builtin_neon_vgetq_lane_i16: 4211 case NEON::BI__builtin_neon_vduph_laneq_i16: 4212 Ops[0] = Builder.CreateBitCast(Ops[0], 4213 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4214 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4215 "vgetq_lane"); 4216 case NEON::BI__builtin_neon_vget_lane_i32: 4217 case NEON::BI__builtin_neon_vdups_lane_i32: 4218 Ops[0] = Builder.CreateBitCast( 4219 Ops[0], 4220 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4221 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4222 "vget_lane"); 4223 case NEON::BI__builtin_neon_vdups_lane_f32: 4224 Ops[0] = Builder.CreateBitCast(Ops[0], 4225 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4226 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4227 "vdups_lane"); 4228 case NEON::BI__builtin_neon_vgetq_lane_i32: 4229 case NEON::BI__builtin_neon_vdups_laneq_i32: 4230 Ops[0] = Builder.CreateBitCast(Ops[0], 4231 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4232 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4233 "vgetq_lane"); 4234 case NEON::BI__builtin_neon_vget_lane_i64: 4235 case NEON::BI__builtin_neon_vdupd_lane_i64: 4236 Ops[0] = Builder.CreateBitCast(Ops[0], 4237 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4238 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4239 "vget_lane"); 4240 case NEON::BI__builtin_neon_vdupd_lane_f64: 4241 Ops[0] = Builder.CreateBitCast(Ops[0], 4242 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4243 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4244 "vdupd_lane"); 4245 case NEON::BI__builtin_neon_vgetq_lane_i64: 4246 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4247 Ops[0] = Builder.CreateBitCast(Ops[0], 4248 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4249 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4250 "vgetq_lane"); 4251 case NEON::BI__builtin_neon_vget_lane_f32: 4252 Ops[0] = Builder.CreateBitCast(Ops[0], 4253 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4254 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4255 "vget_lane"); 4256 case NEON::BI__builtin_neon_vget_lane_f64: 4257 Ops[0] = Builder.CreateBitCast(Ops[0], 4258 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4259 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4260 "vget_lane"); 4261 case NEON::BI__builtin_neon_vgetq_lane_f32: 4262 case NEON::BI__builtin_neon_vdups_laneq_f32: 4263 Ops[0] = Builder.CreateBitCast(Ops[0], 4264 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4265 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4266 "vgetq_lane"); 4267 case NEON::BI__builtin_neon_vgetq_lane_f64: 4268 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4269 Ops[0] = Builder.CreateBitCast(Ops[0], 4270 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4271 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4272 "vgetq_lane"); 4273 case NEON::BI__builtin_neon_vaddd_s64: 4274 case NEON::BI__builtin_neon_vaddd_u64: 4275 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4276 case NEON::BI__builtin_neon_vsubd_s64: 4277 case NEON::BI__builtin_neon_vsubd_u64: 4278 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4279 case NEON::BI__builtin_neon_vqdmlalh_s16: 4280 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4281 SmallVector<Value *, 2> ProductOps; 4282 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4283 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4284 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4285 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4286 ProductOps, "vqdmlXl"); 4287 Constant *CI = ConstantInt::get(SizeTy, 0); 4288 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4289 4290 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4291 ? Intrinsic::aarch64_neon_sqadd 4292 : Intrinsic::aarch64_neon_sqsub; 4293 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4294 } 4295 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4296 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4297 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4298 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4299 Ops, "vqshlu_n"); 4300 } 4301 case NEON::BI__builtin_neon_vqshld_n_u64: 4302 case NEON::BI__builtin_neon_vqshld_n_s64: { 4303 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4304 ? Intrinsic::aarch64_neon_uqshl 4305 : Intrinsic::aarch64_neon_sqshl; 4306 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4307 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4308 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4309 } 4310 case NEON::BI__builtin_neon_vrshrd_n_u64: 4311 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4312 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4313 ? Intrinsic::aarch64_neon_urshl 4314 : Intrinsic::aarch64_neon_srshl; 4315 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4316 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4317 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4318 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4319 } 4320 case NEON::BI__builtin_neon_vrsrad_n_u64: 4321 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4322 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4323 ? Intrinsic::aarch64_neon_urshl 4324 : Intrinsic::aarch64_neon_srshl; 4325 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4326 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4327 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4328 Builder.CreateSExt(Ops[2], Int64Ty)); 4329 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4330 } 4331 case NEON::BI__builtin_neon_vshld_n_s64: 4332 case NEON::BI__builtin_neon_vshld_n_u64: { 4333 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4334 return Builder.CreateShl( 4335 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4336 } 4337 case NEON::BI__builtin_neon_vshrd_n_s64: { 4338 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4339 return Builder.CreateAShr( 4340 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4341 Amt->getZExtValue())), 4342 "shrd_n"); 4343 } 4344 case NEON::BI__builtin_neon_vshrd_n_u64: { 4345 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4346 uint64_t ShiftAmt = Amt->getZExtValue(); 4347 // Right-shifting an unsigned value by its size yields 0. 4348 if (ShiftAmt == 64) 4349 return ConstantInt::get(Int64Ty, 0); 4350 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4351 "shrd_n"); 4352 } 4353 case NEON::BI__builtin_neon_vsrad_n_s64: { 4354 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4355 Ops[1] = Builder.CreateAShr( 4356 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4357 Amt->getZExtValue())), 4358 "shrd_n"); 4359 return Builder.CreateAdd(Ops[0], Ops[1]); 4360 } 4361 case NEON::BI__builtin_neon_vsrad_n_u64: { 4362 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4363 uint64_t ShiftAmt = Amt->getZExtValue(); 4364 // Right-shifting an unsigned value by its size yields 0. 4365 // As Op + 0 = Op, return Ops[0] directly. 4366 if (ShiftAmt == 64) 4367 return Ops[0]; 4368 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4369 "shrd_n"); 4370 return Builder.CreateAdd(Ops[0], Ops[1]); 4371 } 4372 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4373 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4374 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4375 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4376 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4377 "lane"); 4378 SmallVector<Value *, 2> ProductOps; 4379 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4380 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4381 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4382 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4383 ProductOps, "vqdmlXl"); 4384 Constant *CI = ConstantInt::get(SizeTy, 0); 4385 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4386 Ops.pop_back(); 4387 4388 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4389 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4390 ? Intrinsic::aarch64_neon_sqadd 4391 : Intrinsic::aarch64_neon_sqsub; 4392 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4393 } 4394 case NEON::BI__builtin_neon_vqdmlals_s32: 4395 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4396 SmallVector<Value *, 2> ProductOps; 4397 ProductOps.push_back(Ops[1]); 4398 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4399 Ops[1] = 4400 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4401 ProductOps, "vqdmlXl"); 4402 4403 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4404 ? Intrinsic::aarch64_neon_sqadd 4405 : Intrinsic::aarch64_neon_sqsub; 4406 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4407 } 4408 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4409 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4410 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4411 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4412 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4413 "lane"); 4414 SmallVector<Value *, 2> ProductOps; 4415 ProductOps.push_back(Ops[1]); 4416 ProductOps.push_back(Ops[2]); 4417 Ops[1] = 4418 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4419 ProductOps, "vqdmlXl"); 4420 Ops.pop_back(); 4421 4422 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4423 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4424 ? Intrinsic::aarch64_neon_sqadd 4425 : Intrinsic::aarch64_neon_sqsub; 4426 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4427 } 4428 } 4429 4430 llvm::VectorType *VTy = GetNeonType(this, Type); 4431 llvm::Type *Ty = VTy; 4432 if (!Ty) 4433 return nullptr; 4434 4435 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4436 // defer to common code if it's been added to our special map. 4437 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4438 AArch64SIMDIntrinsicsProvenSorted); 4439 4440 if (Builtin) 4441 return EmitCommonNeonBuiltinExpr( 4442 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4443 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4444 4445 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4446 return V; 4447 4448 unsigned Int; 4449 switch (BuiltinID) { 4450 default: return nullptr; 4451 case NEON::BI__builtin_neon_vbsl_v: 4452 case NEON::BI__builtin_neon_vbslq_v: { 4453 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4454 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4455 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4456 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4457 4458 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4459 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4460 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4461 return Builder.CreateBitCast(Ops[0], Ty); 4462 } 4463 case NEON::BI__builtin_neon_vfma_lane_v: 4464 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4465 // The ARM builtins (and instructions) have the addend as the first 4466 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4467 Value *Addend = Ops[0]; 4468 Value *Multiplicand = Ops[1]; 4469 Value *LaneSource = Ops[2]; 4470 Ops[0] = Multiplicand; 4471 Ops[1] = LaneSource; 4472 Ops[2] = Addend; 4473 4474 // Now adjust things to handle the lane access. 4475 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4476 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4477 VTy; 4478 llvm::Constant *cst = cast<Constant>(Ops[3]); 4479 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4480 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4481 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4482 4483 Ops.pop_back(); 4484 Int = Intrinsic::fma; 4485 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4486 } 4487 case NEON::BI__builtin_neon_vfma_laneq_v: { 4488 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4489 // v1f64 fma should be mapped to Neon scalar f64 fma 4490 if (VTy && VTy->getElementType() == DoubleTy) { 4491 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4492 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4493 llvm::Type *VTy = GetNeonType(this, 4494 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4495 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4496 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4497 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4498 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4499 return Builder.CreateBitCast(Result, Ty); 4500 } 4501 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4502 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4503 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4504 4505 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4506 VTy->getNumElements() * 2); 4507 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4508 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4509 cast<ConstantInt>(Ops[3])); 4510 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4511 4512 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4513 } 4514 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4515 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4516 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4517 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4518 4519 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4520 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4521 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4522 } 4523 case NEON::BI__builtin_neon_vfmas_lane_f32: 4524 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4525 case NEON::BI__builtin_neon_vfmad_lane_f64: 4526 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4527 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4528 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4529 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4530 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4531 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4532 } 4533 case NEON::BI__builtin_neon_vfms_v: 4534 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4535 // FIXME: probably remove when we no longer support aarch64_simd.h 4536 // (arm_neon.h delegates to vfma). 4537 4538 // The ARM builtins (and instructions) have the addend as the first 4539 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4540 Value *Subtrahend = Ops[0]; 4541 Value *Multiplicand = Ops[2]; 4542 Ops[0] = Multiplicand; 4543 Ops[2] = Subtrahend; 4544 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4545 Ops[1] = Builder.CreateFNeg(Ops[1]); 4546 Int = Intrinsic::fma; 4547 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4548 } 4549 case NEON::BI__builtin_neon_vmull_v: 4550 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4551 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4552 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4553 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4554 case NEON::BI__builtin_neon_vmax_v: 4555 case NEON::BI__builtin_neon_vmaxq_v: 4556 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4557 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4558 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4559 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4560 case NEON::BI__builtin_neon_vmin_v: 4561 case NEON::BI__builtin_neon_vminq_v: 4562 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4563 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4564 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4565 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4566 case NEON::BI__builtin_neon_vabd_v: 4567 case NEON::BI__builtin_neon_vabdq_v: 4568 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4569 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4570 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4571 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4572 case NEON::BI__builtin_neon_vpadal_v: 4573 case NEON::BI__builtin_neon_vpadalq_v: { 4574 unsigned ArgElts = VTy->getNumElements(); 4575 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4576 unsigned BitWidth = EltTy->getBitWidth(); 4577 llvm::Type *ArgTy = llvm::VectorType::get( 4578 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4579 llvm::Type* Tys[2] = { VTy, ArgTy }; 4580 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4581 SmallVector<llvm::Value*, 1> TmpOps; 4582 TmpOps.push_back(Ops[1]); 4583 Function *F = CGM.getIntrinsic(Int, Tys); 4584 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4585 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4586 return Builder.CreateAdd(tmp, addend); 4587 } 4588 case NEON::BI__builtin_neon_vpmin_v: 4589 case NEON::BI__builtin_neon_vpminq_v: 4590 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4591 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4592 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4593 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4594 case NEON::BI__builtin_neon_vpmax_v: 4595 case NEON::BI__builtin_neon_vpmaxq_v: 4596 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4597 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4598 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4599 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4600 case NEON::BI__builtin_neon_vminnm_v: 4601 case NEON::BI__builtin_neon_vminnmq_v: 4602 Int = Intrinsic::aarch64_neon_fminnm; 4603 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4604 case NEON::BI__builtin_neon_vmaxnm_v: 4605 case NEON::BI__builtin_neon_vmaxnmq_v: 4606 Int = Intrinsic::aarch64_neon_fmaxnm; 4607 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4608 case NEON::BI__builtin_neon_vrecpss_f32: { 4609 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4610 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4611 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4612 Ops, "vrecps"); 4613 } 4614 case NEON::BI__builtin_neon_vrecpsd_f64: { 4615 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4616 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4617 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4618 Ops, "vrecps"); 4619 } 4620 case NEON::BI__builtin_neon_vrshr_n_v: 4621 case NEON::BI__builtin_neon_vrshrq_n_v: 4622 // FIXME: this can be shared with 32-bit ARM, but not AArch64 at the 4623 // moment. After the final merge it should be added to 4624 // EmitCommonNeonBuiltinExpr. 4625 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 4626 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 4627 case NEON::BI__builtin_neon_vqshlu_n_v: 4628 case NEON::BI__builtin_neon_vqshluq_n_v: 4629 // FIXME: AArch64 and ARM use different intrinsics for this, but are 4630 // essentially compatible. It should be in EmitCommonNeonBuiltinExpr after 4631 // the final merge. 4632 Int = Intrinsic::aarch64_neon_sqshlu; 4633 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 1, false); 4634 case NEON::BI__builtin_neon_vqshrun_n_v: 4635 // FIXME: as above 4636 Int = Intrinsic::aarch64_neon_sqshrun; 4637 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4638 case NEON::BI__builtin_neon_vqrshrun_n_v: 4639 // FIXME: and again. 4640 Int = Intrinsic::aarch64_neon_sqrshrun; 4641 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4642 case NEON::BI__builtin_neon_vqshrn_n_v: 4643 // FIXME: guess 4644 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4645 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4646 case NEON::BI__builtin_neon_vrshrn_n_v: 4647 // FIXME: there might be a pattern here. 4648 Int = Intrinsic::aarch64_neon_rshrn; 4649 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4650 case NEON::BI__builtin_neon_vqrshrn_n_v: 4651 // FIXME: another one 4652 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4653 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4654 case NEON::BI__builtin_neon_vrnda_v: 4655 case NEON::BI__builtin_neon_vrndaq_v: { 4656 Int = Intrinsic::round; 4657 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4658 } 4659 case NEON::BI__builtin_neon_vrndi_v: 4660 case NEON::BI__builtin_neon_vrndiq_v: { 4661 Int = Intrinsic::nearbyint; 4662 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4663 } 4664 case NEON::BI__builtin_neon_vrndm_v: 4665 case NEON::BI__builtin_neon_vrndmq_v: { 4666 Int = Intrinsic::floor; 4667 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4668 } 4669 case NEON::BI__builtin_neon_vrndn_v: 4670 case NEON::BI__builtin_neon_vrndnq_v: { 4671 Int = Intrinsic::aarch64_neon_frintn; 4672 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4673 } 4674 case NEON::BI__builtin_neon_vrndp_v: 4675 case NEON::BI__builtin_neon_vrndpq_v: { 4676 Int = Intrinsic::ceil; 4677 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4678 } 4679 case NEON::BI__builtin_neon_vrndx_v: 4680 case NEON::BI__builtin_neon_vrndxq_v: { 4681 Int = Intrinsic::rint; 4682 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4683 } 4684 case NEON::BI__builtin_neon_vrnd_v: 4685 case NEON::BI__builtin_neon_vrndq_v: { 4686 Int = Intrinsic::trunc; 4687 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4688 } 4689 case NEON::BI__builtin_neon_vceqz_v: 4690 case NEON::BI__builtin_neon_vceqzq_v: 4691 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4692 ICmpInst::ICMP_EQ, "vceqz"); 4693 case NEON::BI__builtin_neon_vcgez_v: 4694 case NEON::BI__builtin_neon_vcgezq_v: 4695 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4696 ICmpInst::ICMP_SGE, "vcgez"); 4697 case NEON::BI__builtin_neon_vclez_v: 4698 case NEON::BI__builtin_neon_vclezq_v: 4699 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4700 ICmpInst::ICMP_SLE, "vclez"); 4701 case NEON::BI__builtin_neon_vcgtz_v: 4702 case NEON::BI__builtin_neon_vcgtzq_v: 4703 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4704 ICmpInst::ICMP_SGT, "vcgtz"); 4705 case NEON::BI__builtin_neon_vcltz_v: 4706 case NEON::BI__builtin_neon_vcltzq_v: 4707 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4708 ICmpInst::ICMP_SLT, "vcltz"); 4709 case NEON::BI__builtin_neon_vcvt_f64_v: 4710 case NEON::BI__builtin_neon_vcvtq_f64_v: 4711 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4712 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4713 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4714 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4715 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4716 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4717 "unexpected vcvt_f64_f32 builtin"); 4718 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4719 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4720 4721 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4722 } 4723 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4724 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4725 "unexpected vcvt_f32_f64 builtin"); 4726 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4727 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4728 4729 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4730 } 4731 case NEON::BI__builtin_neon_vcvt_s32_v: 4732 case NEON::BI__builtin_neon_vcvt_u32_v: 4733 case NEON::BI__builtin_neon_vcvt_s64_v: 4734 case NEON::BI__builtin_neon_vcvt_u64_v: 4735 case NEON::BI__builtin_neon_vcvtq_s32_v: 4736 case NEON::BI__builtin_neon_vcvtq_u32_v: 4737 case NEON::BI__builtin_neon_vcvtq_s64_v: 4738 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4739 bool Double = 4740 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4741 llvm::Type *InTy = 4742 GetNeonType(this, 4743 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4744 : NeonTypeFlags::Float32, false, quad)); 4745 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4746 if (usgn) 4747 return Builder.CreateFPToUI(Ops[0], Ty); 4748 return Builder.CreateFPToSI(Ops[0], Ty); 4749 } 4750 case NEON::BI__builtin_neon_vcvta_s32_v: 4751 case NEON::BI__builtin_neon_vcvtaq_s32_v: 4752 case NEON::BI__builtin_neon_vcvta_u32_v: 4753 case NEON::BI__builtin_neon_vcvtaq_u32_v: 4754 case NEON::BI__builtin_neon_vcvta_s64_v: 4755 case NEON::BI__builtin_neon_vcvtaq_s64_v: 4756 case NEON::BI__builtin_neon_vcvta_u64_v: 4757 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4758 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 4759 bool Double = 4760 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4761 llvm::Type *InTy = 4762 GetNeonType(this, 4763 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4764 : NeonTypeFlags::Float32, false, quad)); 4765 llvm::Type *Tys[2] = { Ty, InTy }; 4766 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 4767 } 4768 case NEON::BI__builtin_neon_vcvtm_s32_v: 4769 case NEON::BI__builtin_neon_vcvtmq_s32_v: 4770 case NEON::BI__builtin_neon_vcvtm_u32_v: 4771 case NEON::BI__builtin_neon_vcvtmq_u32_v: 4772 case NEON::BI__builtin_neon_vcvtm_s64_v: 4773 case NEON::BI__builtin_neon_vcvtmq_s64_v: 4774 case NEON::BI__builtin_neon_vcvtm_u64_v: 4775 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4776 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 4777 bool Double = 4778 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4779 llvm::Type *InTy = 4780 GetNeonType(this, 4781 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4782 : NeonTypeFlags::Float32, false, quad)); 4783 llvm::Type *Tys[2] = { Ty, InTy }; 4784 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 4785 } 4786 case NEON::BI__builtin_neon_vcvtn_s32_v: 4787 case NEON::BI__builtin_neon_vcvtnq_s32_v: 4788 case NEON::BI__builtin_neon_vcvtn_u32_v: 4789 case NEON::BI__builtin_neon_vcvtnq_u32_v: 4790 case NEON::BI__builtin_neon_vcvtn_s64_v: 4791 case NEON::BI__builtin_neon_vcvtnq_s64_v: 4792 case NEON::BI__builtin_neon_vcvtn_u64_v: 4793 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4794 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 4795 bool Double = 4796 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4797 llvm::Type *InTy = 4798 GetNeonType(this, 4799 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4800 : NeonTypeFlags::Float32, false, quad)); 4801 llvm::Type *Tys[2] = { Ty, InTy }; 4802 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 4803 } 4804 case NEON::BI__builtin_neon_vcvtp_s32_v: 4805 case NEON::BI__builtin_neon_vcvtpq_s32_v: 4806 case NEON::BI__builtin_neon_vcvtp_u32_v: 4807 case NEON::BI__builtin_neon_vcvtpq_u32_v: 4808 case NEON::BI__builtin_neon_vcvtp_s64_v: 4809 case NEON::BI__builtin_neon_vcvtpq_s64_v: 4810 case NEON::BI__builtin_neon_vcvtp_u64_v: 4811 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4812 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 4813 bool Double = 4814 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4815 llvm::Type *InTy = 4816 GetNeonType(this, 4817 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4818 : NeonTypeFlags::Float32, false, quad)); 4819 llvm::Type *Tys[2] = { Ty, InTy }; 4820 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 4821 } 4822 case NEON::BI__builtin_neon_vmulx_v: 4823 case NEON::BI__builtin_neon_vmulxq_v: { 4824 Int = Intrinsic::aarch64_neon_fmulx; 4825 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 4826 } 4827 case NEON::BI__builtin_neon_vmul_lane_v: 4828 case NEON::BI__builtin_neon_vmul_laneq_v: { 4829 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 4830 bool Quad = false; 4831 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 4832 Quad = true; 4833 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4834 llvm::Type *VTy = GetNeonType(this, 4835 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 4836 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4837 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 4838 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 4839 return Builder.CreateBitCast(Result, Ty); 4840 } 4841 case NEON::BI__builtin_neon_vnegd_s64: 4842 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 4843 case NEON::BI__builtin_neon_vpmaxnm_v: 4844 case NEON::BI__builtin_neon_vpmaxnmq_v: { 4845 Int = Intrinsic::aarch64_neon_fmaxnmp; 4846 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 4847 } 4848 case NEON::BI__builtin_neon_vpminnm_v: 4849 case NEON::BI__builtin_neon_vpminnmq_v: { 4850 Int = Intrinsic::aarch64_neon_fminnmp; 4851 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 4852 } 4853 case NEON::BI__builtin_neon_vsqrt_v: 4854 case NEON::BI__builtin_neon_vsqrtq_v: { 4855 Int = Intrinsic::sqrt; 4856 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4857 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 4858 } 4859 case NEON::BI__builtin_neon_vrbit_v: 4860 case NEON::BI__builtin_neon_vrbitq_v: { 4861 Int = Intrinsic::aarch64_neon_rbit; 4862 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 4863 } 4864 case NEON::BI__builtin_neon_vaddv_u8: 4865 // FIXME: These are handled by the AArch64 scalar code. 4866 usgn = true; 4867 // FALLTHROUGH 4868 case NEON::BI__builtin_neon_vaddv_s8: { 4869 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4870 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4871 VTy = 4872 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 4873 llvm::Type *Tys[2] = { Ty, VTy }; 4874 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4875 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4876 return Builder.CreateTrunc(Ops[0], 4877 llvm::IntegerType::get(getLLVMContext(), 8)); 4878 } 4879 case NEON::BI__builtin_neon_vaddv_u16: 4880 usgn = true; 4881 // FALLTHROUGH 4882 case NEON::BI__builtin_neon_vaddv_s16: { 4883 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4884 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4885 VTy = 4886 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 4887 llvm::Type *Tys[2] = { Ty, VTy }; 4888 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4889 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4890 return Builder.CreateTrunc(Ops[0], 4891 llvm::IntegerType::get(getLLVMContext(), 16)); 4892 } 4893 case NEON::BI__builtin_neon_vaddvq_u8: 4894 usgn = true; 4895 // FALLTHROUGH 4896 case NEON::BI__builtin_neon_vaddvq_s8: { 4897 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4898 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4899 VTy = 4900 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 4901 llvm::Type *Tys[2] = { Ty, VTy }; 4902 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4903 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4904 return Builder.CreateTrunc(Ops[0], 4905 llvm::IntegerType::get(getLLVMContext(), 8)); 4906 } 4907 case NEON::BI__builtin_neon_vaddvq_u16: 4908 usgn = true; 4909 // FALLTHROUGH 4910 case NEON::BI__builtin_neon_vaddvq_s16: { 4911 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4912 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4913 VTy = 4914 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 4915 llvm::Type *Tys[2] = { Ty, VTy }; 4916 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4917 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4918 return Builder.CreateTrunc(Ops[0], 4919 llvm::IntegerType::get(getLLVMContext(), 16)); 4920 } 4921 case NEON::BI__builtin_neon_vmaxv_u8: { 4922 Int = Intrinsic::aarch64_neon_umaxv; 4923 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4924 VTy = 4925 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 4926 llvm::Type *Tys[2] = { Ty, VTy }; 4927 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4928 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 4929 return Builder.CreateTrunc(Ops[0], 4930 llvm::IntegerType::get(getLLVMContext(), 8)); 4931 } 4932 case NEON::BI__builtin_neon_vmaxv_u16: { 4933 Int = Intrinsic::aarch64_neon_umaxv; 4934 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4935 VTy = 4936 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 4937 llvm::Type *Tys[2] = { Ty, VTy }; 4938 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4939 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 4940 return Builder.CreateTrunc(Ops[0], 4941 llvm::IntegerType::get(getLLVMContext(), 16)); 4942 } 4943 case NEON::BI__builtin_neon_vmaxvq_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), 16); 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_vmaxvq_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), 8); 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_vmaxv_s8: { 4966 Int = Intrinsic::aarch64_neon_smaxv; 4967 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4968 VTy = 4969 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 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_vmaxv_s16: { 4977 Int = Intrinsic::aarch64_neon_smaxv; 4978 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4979 VTy = 4980 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 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_vmaxvq_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), 16); 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_vmaxvq_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), 8); 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_vminv_u8: { 5010 Int = Intrinsic::aarch64_neon_uminv; 5011 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5012 VTy = 5013 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5014 llvm::Type *Tys[2] = { Ty, VTy }; 5015 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5016 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5017 return Builder.CreateTrunc(Ops[0], 5018 llvm::IntegerType::get(getLLVMContext(), 8)); 5019 } 5020 case NEON::BI__builtin_neon_vminv_u16: { 5021 Int = Intrinsic::aarch64_neon_uminv; 5022 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5023 VTy = 5024 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5025 llvm::Type *Tys[2] = { Ty, VTy }; 5026 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5027 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5028 return Builder.CreateTrunc(Ops[0], 5029 llvm::IntegerType::get(getLLVMContext(), 16)); 5030 } 5031 case NEON::BI__builtin_neon_vminvq_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), 16); 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_vminvq_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), 8); 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_vminv_s8: { 5054 Int = Intrinsic::aarch64_neon_sminv; 5055 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5056 VTy = 5057 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 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_vminv_s16: { 5065 Int = Intrinsic::aarch64_neon_sminv; 5066 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5067 VTy = 5068 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 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_vminvq_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), 16); 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_vminvq_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), 8); 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_vmul_n_f64: { 5098 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5099 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5100 return Builder.CreateFMul(Ops[0], RHS); 5101 } 5102 case NEON::BI__builtin_neon_vaddlv_u8: { 5103 Int = Intrinsic::aarch64_neon_uaddlv; 5104 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5105 VTy = 5106 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5107 llvm::Type *Tys[2] = { Ty, VTy }; 5108 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5109 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5110 return Builder.CreateTrunc(Ops[0], 5111 llvm::IntegerType::get(getLLVMContext(), 16)); 5112 } 5113 case NEON::BI__builtin_neon_vaddlv_u16: { 5114 Int = Intrinsic::aarch64_neon_uaddlv; 5115 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5116 VTy = 5117 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5118 llvm::Type *Tys[2] = { Ty, VTy }; 5119 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5120 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5121 } 5122 case NEON::BI__builtin_neon_vaddlvq_u8: { 5123 Int = Intrinsic::aarch64_neon_uaddlv; 5124 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5125 VTy = 5126 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5127 llvm::Type *Tys[2] = { Ty, VTy }; 5128 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5129 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5130 return Builder.CreateTrunc(Ops[0], 5131 llvm::IntegerType::get(getLLVMContext(), 16)); 5132 } 5133 case NEON::BI__builtin_neon_vaddlvq_u16: { 5134 Int = Intrinsic::aarch64_neon_uaddlv; 5135 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5136 VTy = 5137 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5138 llvm::Type *Tys[2] = { Ty, VTy }; 5139 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5140 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5141 } 5142 case NEON::BI__builtin_neon_vaddlv_s8: { 5143 Int = Intrinsic::aarch64_neon_saddlv; 5144 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5145 VTy = 5146 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5147 llvm::Type *Tys[2] = { Ty, VTy }; 5148 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5149 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5150 return Builder.CreateTrunc(Ops[0], 5151 llvm::IntegerType::get(getLLVMContext(), 16)); 5152 } 5153 case NEON::BI__builtin_neon_vaddlv_s16: { 5154 Int = Intrinsic::aarch64_neon_saddlv; 5155 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5156 VTy = 5157 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5158 llvm::Type *Tys[2] = { Ty, VTy }; 5159 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5160 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5161 } 5162 case NEON::BI__builtin_neon_vaddlvq_s8: { 5163 Int = Intrinsic::aarch64_neon_saddlv; 5164 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5165 VTy = 5166 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5167 llvm::Type *Tys[2] = { Ty, VTy }; 5168 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5169 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5170 return Builder.CreateTrunc(Ops[0], 5171 llvm::IntegerType::get(getLLVMContext(), 16)); 5172 } 5173 case NEON::BI__builtin_neon_vaddlvq_s16: { 5174 Int = Intrinsic::aarch64_neon_saddlv; 5175 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5176 VTy = 5177 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5178 llvm::Type *Tys[2] = { Ty, VTy }; 5179 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5180 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5181 } 5182 case NEON::BI__builtin_neon_vsri_n_v: 5183 case NEON::BI__builtin_neon_vsriq_n_v: { 5184 Int = Intrinsic::aarch64_neon_vsri; 5185 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5186 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5187 } 5188 case NEON::BI__builtin_neon_vsli_n_v: 5189 case NEON::BI__builtin_neon_vsliq_n_v: { 5190 Int = Intrinsic::aarch64_neon_vsli; 5191 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5192 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5193 } 5194 case NEON::BI__builtin_neon_vsra_n_v: 5195 case NEON::BI__builtin_neon_vsraq_n_v: 5196 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5197 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5198 return Builder.CreateAdd(Ops[0], Ops[1]); 5199 case NEON::BI__builtin_neon_vrsra_n_v: 5200 case NEON::BI__builtin_neon_vrsraq_n_v: { 5201 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5202 SmallVector<llvm::Value*,2> TmpOps; 5203 TmpOps.push_back(Ops[1]); 5204 TmpOps.push_back(Ops[2]); 5205 Function* F = CGM.getIntrinsic(Int, Ty); 5206 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5207 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5208 return Builder.CreateAdd(Ops[0], tmp); 5209 } 5210 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5211 // of an Align parameter here. 5212 case NEON::BI__builtin_neon_vld1_x2_v: 5213 case NEON::BI__builtin_neon_vld1q_x2_v: 5214 case NEON::BI__builtin_neon_vld1_x3_v: 5215 case NEON::BI__builtin_neon_vld1q_x3_v: 5216 case NEON::BI__builtin_neon_vld1_x4_v: 5217 case NEON::BI__builtin_neon_vld1q_x4_v: { 5218 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5219 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5220 llvm::Type *Tys[2] = { VTy, PTy }; 5221 unsigned Int; 5222 switch (BuiltinID) { 5223 case NEON::BI__builtin_neon_vld1_x2_v: 5224 case NEON::BI__builtin_neon_vld1q_x2_v: 5225 Int = Intrinsic::aarch64_neon_ld1x2; 5226 break; 5227 case NEON::BI__builtin_neon_vld1_x3_v: 5228 case NEON::BI__builtin_neon_vld1q_x3_v: 5229 Int = Intrinsic::aarch64_neon_ld1x3; 5230 break; 5231 case NEON::BI__builtin_neon_vld1_x4_v: 5232 case NEON::BI__builtin_neon_vld1q_x4_v: 5233 Int = Intrinsic::aarch64_neon_ld1x4; 5234 break; 5235 } 5236 Function *F = CGM.getIntrinsic(Int, Tys); 5237 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5238 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5239 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5240 return Builder.CreateStore(Ops[1], Ops[0]); 5241 } 5242 case NEON::BI__builtin_neon_vst1_x2_v: 5243 case NEON::BI__builtin_neon_vst1q_x2_v: 5244 case NEON::BI__builtin_neon_vst1_x3_v: 5245 case NEON::BI__builtin_neon_vst1q_x3_v: 5246 case NEON::BI__builtin_neon_vst1_x4_v: 5247 case NEON::BI__builtin_neon_vst1q_x4_v: { 5248 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5249 llvm::Type *Tys[2] = { VTy, PTy }; 5250 unsigned Int; 5251 switch (BuiltinID) { 5252 case NEON::BI__builtin_neon_vst1_x2_v: 5253 case NEON::BI__builtin_neon_vst1q_x2_v: 5254 Int = Intrinsic::aarch64_neon_st1x2; 5255 break; 5256 case NEON::BI__builtin_neon_vst1_x3_v: 5257 case NEON::BI__builtin_neon_vst1q_x3_v: 5258 Int = Intrinsic::aarch64_neon_st1x3; 5259 break; 5260 case NEON::BI__builtin_neon_vst1_x4_v: 5261 case NEON::BI__builtin_neon_vst1q_x4_v: 5262 Int = Intrinsic::aarch64_neon_st1x4; 5263 break; 5264 } 5265 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5266 IntOps.push_back(Ops[0]); 5267 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5268 } 5269 case NEON::BI__builtin_neon_vld1_v: 5270 case NEON::BI__builtin_neon_vld1q_v: 5271 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5272 return Builder.CreateLoad(Ops[0]); 5273 case NEON::BI__builtin_neon_vst1_v: 5274 case NEON::BI__builtin_neon_vst1q_v: 5275 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5276 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5277 return Builder.CreateStore(Ops[1], Ops[0]); 5278 case NEON::BI__builtin_neon_vld1_lane_v: 5279 case NEON::BI__builtin_neon_vld1q_lane_v: 5280 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5281 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5282 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5283 Ops[0] = Builder.CreateLoad(Ops[0]); 5284 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5285 case NEON::BI__builtin_neon_vld1_dup_v: 5286 case NEON::BI__builtin_neon_vld1q_dup_v: { 5287 Value *V = UndefValue::get(Ty); 5288 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5289 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5290 Ops[0] = Builder.CreateLoad(Ops[0]); 5291 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5292 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5293 return EmitNeonSplat(Ops[0], CI); 5294 } 5295 case NEON::BI__builtin_neon_vst1_lane_v: 5296 case NEON::BI__builtin_neon_vst1q_lane_v: 5297 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5298 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5299 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5300 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5301 case NEON::BI__builtin_neon_vld2_v: 5302 case NEON::BI__builtin_neon_vld2q_v: { 5303 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5304 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5305 llvm::Type *Tys[2] = { VTy, PTy }; 5306 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5307 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5308 Ops[0] = Builder.CreateBitCast(Ops[0], 5309 llvm::PointerType::getUnqual(Ops[1]->getType())); 5310 return Builder.CreateStore(Ops[1], Ops[0]); 5311 } 5312 case NEON::BI__builtin_neon_vld3_v: 5313 case NEON::BI__builtin_neon_vld3q_v: { 5314 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5315 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5316 llvm::Type *Tys[2] = { VTy, PTy }; 5317 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5318 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5319 Ops[0] = Builder.CreateBitCast(Ops[0], 5320 llvm::PointerType::getUnqual(Ops[1]->getType())); 5321 return Builder.CreateStore(Ops[1], Ops[0]); 5322 } 5323 case NEON::BI__builtin_neon_vld4_v: 5324 case NEON::BI__builtin_neon_vld4q_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_ld4, Tys); 5329 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 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_vld2_dup_v: 5335 case NEON::BI__builtin_neon_vld2q_dup_v: { 5336 llvm::Type *PTy = 5337 llvm::PointerType::getUnqual(VTy->getElementType()); 5338 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5339 llvm::Type *Tys[2] = { VTy, PTy }; 5340 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5341 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5342 Ops[0] = Builder.CreateBitCast(Ops[0], 5343 llvm::PointerType::getUnqual(Ops[1]->getType())); 5344 return Builder.CreateStore(Ops[1], Ops[0]); 5345 } 5346 case NEON::BI__builtin_neon_vld3_dup_v: 5347 case NEON::BI__builtin_neon_vld3q_dup_v: { 5348 llvm::Type *PTy = 5349 llvm::PointerType::getUnqual(VTy->getElementType()); 5350 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5351 llvm::Type *Tys[2] = { VTy, PTy }; 5352 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5353 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5354 Ops[0] = Builder.CreateBitCast(Ops[0], 5355 llvm::PointerType::getUnqual(Ops[1]->getType())); 5356 return Builder.CreateStore(Ops[1], Ops[0]); 5357 } 5358 case NEON::BI__builtin_neon_vld4_dup_v: 5359 case NEON::BI__builtin_neon_vld4q_dup_v: { 5360 llvm::Type *PTy = 5361 llvm::PointerType::getUnqual(VTy->getElementType()); 5362 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5363 llvm::Type *Tys[2] = { VTy, PTy }; 5364 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5365 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5366 Ops[0] = Builder.CreateBitCast(Ops[0], 5367 llvm::PointerType::getUnqual(Ops[1]->getType())); 5368 return Builder.CreateStore(Ops[1], Ops[0]); 5369 } 5370 case NEON::BI__builtin_neon_vld2_lane_v: 5371 case NEON::BI__builtin_neon_vld2q_lane_v: { 5372 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5373 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5374 Ops.push_back(Ops[1]); 5375 Ops.erase(Ops.begin()+1); 5376 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5377 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5378 Ops[3] = Builder.CreateZExt(Ops[3], 5379 llvm::IntegerType::get(getLLVMContext(), 64)); 5380 Ops[1] = Builder.CreateCall(F, 5381 ArrayRef<Value*>(Ops).slice(1), "vld2_lane"); 5382 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5383 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5384 return Builder.CreateStore(Ops[1], Ops[0]); 5385 } 5386 case NEON::BI__builtin_neon_vld3_lane_v: 5387 case NEON::BI__builtin_neon_vld3q_lane_v: { 5388 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5389 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5390 Ops.push_back(Ops[1]); 5391 Ops.erase(Ops.begin()+1); 5392 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5393 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5394 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5395 Ops[4] = Builder.CreateZExt(Ops[4], 5396 llvm::IntegerType::get(getLLVMContext(), 64)); 5397 Ops[1] = Builder.CreateCall(F, 5398 ArrayRef<Value*>(Ops).slice(1), "vld3_lane"); 5399 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5400 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5401 return Builder.CreateStore(Ops[1], Ops[0]); 5402 } 5403 case NEON::BI__builtin_neon_vld4_lane_v: 5404 case NEON::BI__builtin_neon_vld4q_lane_v: { 5405 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5406 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5407 Ops.push_back(Ops[1]); 5408 Ops.erase(Ops.begin()+1); 5409 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5410 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5411 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5412 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5413 Ops[5] = Builder.CreateZExt(Ops[5], 5414 llvm::IntegerType::get(getLLVMContext(), 64)); 5415 Ops[1] = Builder.CreateCall(F, 5416 ArrayRef<Value*>(Ops).slice(1), "vld4_lane"); 5417 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5418 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5419 return Builder.CreateStore(Ops[1], Ops[0]); 5420 } 5421 case NEON::BI__builtin_neon_vst2_v: 5422 case NEON::BI__builtin_neon_vst2q_v: { 5423 Ops.push_back(Ops[0]); 5424 Ops.erase(Ops.begin()); 5425 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5426 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5427 Ops, ""); 5428 } 5429 case NEON::BI__builtin_neon_vst2_lane_v: 5430 case NEON::BI__builtin_neon_vst2q_lane_v: { 5431 Ops.push_back(Ops[0]); 5432 Ops.erase(Ops.begin()); 5433 Ops[2] = Builder.CreateZExt(Ops[2], 5434 llvm::IntegerType::get(getLLVMContext(), 64)); 5435 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5436 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5437 Ops, ""); 5438 } 5439 case NEON::BI__builtin_neon_vst3_v: 5440 case NEON::BI__builtin_neon_vst3q_v: { 5441 Ops.push_back(Ops[0]); 5442 Ops.erase(Ops.begin()); 5443 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5444 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5445 Ops, ""); 5446 } 5447 case NEON::BI__builtin_neon_vst3_lane_v: 5448 case NEON::BI__builtin_neon_vst3q_lane_v: { 5449 Ops.push_back(Ops[0]); 5450 Ops.erase(Ops.begin()); 5451 Ops[3] = Builder.CreateZExt(Ops[3], 5452 llvm::IntegerType::get(getLLVMContext(), 64)); 5453 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5454 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5455 Ops, ""); 5456 } 5457 case NEON::BI__builtin_neon_vst4_v: 5458 case NEON::BI__builtin_neon_vst4q_v: { 5459 Ops.push_back(Ops[0]); 5460 Ops.erase(Ops.begin()); 5461 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5462 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5463 Ops, ""); 5464 } 5465 case NEON::BI__builtin_neon_vst4_lane_v: 5466 case NEON::BI__builtin_neon_vst4q_lane_v: { 5467 Ops.push_back(Ops[0]); 5468 Ops.erase(Ops.begin()); 5469 Ops[4] = Builder.CreateZExt(Ops[4], 5470 llvm::IntegerType::get(getLLVMContext(), 64)); 5471 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5472 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5473 Ops, ""); 5474 } 5475 case NEON::BI__builtin_neon_vtrn_v: 5476 case NEON::BI__builtin_neon_vtrnq_v: { 5477 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5478 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5479 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5480 Value *SV = nullptr; 5481 5482 for (unsigned vi = 0; vi != 2; ++vi) { 5483 SmallVector<Constant*, 16> Indices; 5484 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5485 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5486 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5487 } 5488 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5489 SV = llvm::ConstantVector::get(Indices); 5490 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5491 SV = Builder.CreateStore(SV, Addr); 5492 } 5493 return SV; 5494 } 5495 case NEON::BI__builtin_neon_vuzp_v: 5496 case NEON::BI__builtin_neon_vuzpq_v: { 5497 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5498 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5499 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5500 Value *SV = nullptr; 5501 5502 for (unsigned vi = 0; vi != 2; ++vi) { 5503 SmallVector<Constant*, 16> Indices; 5504 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5505 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5506 5507 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5508 SV = llvm::ConstantVector::get(Indices); 5509 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5510 SV = Builder.CreateStore(SV, Addr); 5511 } 5512 return SV; 5513 } 5514 case NEON::BI__builtin_neon_vzip_v: 5515 case NEON::BI__builtin_neon_vzipq_v: { 5516 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5517 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5518 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5519 Value *SV = nullptr; 5520 5521 for (unsigned vi = 0; vi != 2; ++vi) { 5522 SmallVector<Constant*, 16> Indices; 5523 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5524 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5525 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5526 } 5527 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5528 SV = llvm::ConstantVector::get(Indices); 5529 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5530 SV = Builder.CreateStore(SV, Addr); 5531 } 5532 return SV; 5533 } 5534 case NEON::BI__builtin_neon_vqtbl1q_v: { 5535 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5536 Ops, "vtbl1"); 5537 } 5538 case NEON::BI__builtin_neon_vqtbl2q_v: { 5539 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5540 Ops, "vtbl2"); 5541 } 5542 case NEON::BI__builtin_neon_vqtbl3q_v: { 5543 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5544 Ops, "vtbl3"); 5545 } 5546 case NEON::BI__builtin_neon_vqtbl4q_v: { 5547 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5548 Ops, "vtbl4"); 5549 } 5550 case NEON::BI__builtin_neon_vqtbx1q_v: { 5551 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5552 Ops, "vtbx1"); 5553 } 5554 case NEON::BI__builtin_neon_vqtbx2q_v: { 5555 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5556 Ops, "vtbx2"); 5557 } 5558 case NEON::BI__builtin_neon_vqtbx3q_v: { 5559 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5560 Ops, "vtbx3"); 5561 } 5562 case NEON::BI__builtin_neon_vqtbx4q_v: { 5563 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5564 Ops, "vtbx4"); 5565 } 5566 case NEON::BI__builtin_neon_vsqadd_v: 5567 case NEON::BI__builtin_neon_vsqaddq_v: { 5568 Int = Intrinsic::aarch64_neon_usqadd; 5569 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5570 } 5571 case NEON::BI__builtin_neon_vuqadd_v: 5572 case NEON::BI__builtin_neon_vuqaddq_v: { 5573 Int = Intrinsic::aarch64_neon_suqadd; 5574 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5575 } 5576 } 5577 } 5578 5579 llvm::Value *CodeGenFunction:: 5580 BuildVector(ArrayRef<llvm::Value*> Ops) { 5581 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5582 "Not a power-of-two sized vector!"); 5583 bool AllConstants = true; 5584 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5585 AllConstants &= isa<Constant>(Ops[i]); 5586 5587 // If this is a constant vector, create a ConstantVector. 5588 if (AllConstants) { 5589 SmallVector<llvm::Constant*, 16> CstOps; 5590 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5591 CstOps.push_back(cast<Constant>(Ops[i])); 5592 return llvm::ConstantVector::get(CstOps); 5593 } 5594 5595 // Otherwise, insertelement the values to build the vector. 5596 Value *Result = 5597 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5598 5599 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5600 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5601 5602 return Result; 5603 } 5604 5605 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5606 const CallExpr *E) { 5607 SmallVector<Value*, 4> Ops; 5608 5609 // Find out if any arguments are required to be integer constant expressions. 5610 unsigned ICEArguments = 0; 5611 ASTContext::GetBuiltinTypeError Error; 5612 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5613 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5614 5615 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5616 // If this is a normal argument, just emit it as a scalar. 5617 if ((ICEArguments & (1 << i)) == 0) { 5618 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5619 continue; 5620 } 5621 5622 // If this is required to be a constant, constant fold it so that we know 5623 // that the generated intrinsic gets a ConstantInt. 5624 llvm::APSInt Result; 5625 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5626 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5627 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5628 } 5629 5630 switch (BuiltinID) { 5631 default: return nullptr; 5632 case X86::BI_mm_prefetch: { 5633 Value *Address = EmitScalarExpr(E->getArg(0)); 5634 Value *RW = ConstantInt::get(Int32Ty, 0); 5635 Value *Locality = EmitScalarExpr(E->getArg(1)); 5636 Value *Data = ConstantInt::get(Int32Ty, 1); 5637 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5638 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5639 } 5640 case X86::BI__builtin_ia32_vec_init_v8qi: 5641 case X86::BI__builtin_ia32_vec_init_v4hi: 5642 case X86::BI__builtin_ia32_vec_init_v2si: 5643 return Builder.CreateBitCast(BuildVector(Ops), 5644 llvm::Type::getX86_MMXTy(getLLVMContext())); 5645 case X86::BI__builtin_ia32_vec_ext_v2si: 5646 return Builder.CreateExtractElement(Ops[0], 5647 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5648 case X86::BI__builtin_ia32_ldmxcsr: { 5649 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5650 Builder.CreateStore(Ops[0], Tmp); 5651 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5652 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5653 } 5654 case X86::BI__builtin_ia32_stmxcsr: { 5655 Value *Tmp = CreateMemTemp(E->getType()); 5656 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5657 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5658 return Builder.CreateLoad(Tmp, "stmxcsr"); 5659 } 5660 case X86::BI__builtin_ia32_storehps: 5661 case X86::BI__builtin_ia32_storelps: { 5662 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5663 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5664 5665 // cast val v2i64 5666 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5667 5668 // extract (0, 1) 5669 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5670 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5671 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5672 5673 // cast pointer to i64 & store 5674 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5675 return Builder.CreateStore(Ops[1], Ops[0]); 5676 } 5677 case X86::BI__builtin_ia32_palignr: { 5678 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5679 5680 // If palignr is shifting the pair of input vectors less than 9 bytes, 5681 // emit a shuffle instruction. 5682 if (shiftVal <= 8) { 5683 SmallVector<llvm::Constant*, 8> Indices; 5684 for (unsigned i = 0; i != 8; ++i) 5685 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5686 5687 Value* SV = llvm::ConstantVector::get(Indices); 5688 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5689 } 5690 5691 // If palignr is shifting the pair of input vectors more than 8 but less 5692 // than 16 bytes, emit a logical right shift of the destination. 5693 if (shiftVal < 16) { 5694 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 5695 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 5696 5697 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5698 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 5699 5700 // create i32 constant 5701 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 5702 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5703 } 5704 5705 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 5706 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5707 } 5708 case X86::BI__builtin_ia32_palignr128: { 5709 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5710 5711 // If palignr is shifting the pair of input vectors less than 17 bytes, 5712 // emit a shuffle instruction. 5713 if (shiftVal <= 16) { 5714 SmallVector<llvm::Constant*, 16> Indices; 5715 for (unsigned i = 0; i != 16; ++i) 5716 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5717 5718 Value* SV = llvm::ConstantVector::get(Indices); 5719 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5720 } 5721 5722 // If palignr is shifting the pair of input vectors more than 16 but less 5723 // than 32 bytes, emit a logical right shift of the destination. 5724 if (shiftVal < 32) { 5725 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5726 5727 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5728 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5729 5730 // create i32 constant 5731 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 5732 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5733 } 5734 5735 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5736 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5737 } 5738 case X86::BI__builtin_ia32_palignr256: { 5739 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5740 5741 // If palignr is shifting the pair of input vectors less than 17 bytes, 5742 // emit a shuffle instruction. 5743 if (shiftVal <= 16) { 5744 SmallVector<llvm::Constant*, 32> Indices; 5745 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5746 for (unsigned l = 0; l != 2; ++l) { 5747 unsigned LaneStart = l * 16; 5748 unsigned LaneEnd = (l+1) * 16; 5749 for (unsigned i = 0; i != 16; ++i) { 5750 unsigned Idx = shiftVal + i + LaneStart; 5751 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 5752 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 5753 } 5754 } 5755 5756 Value* SV = llvm::ConstantVector::get(Indices); 5757 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5758 } 5759 5760 // If palignr is shifting the pair of input vectors more than 16 but less 5761 // than 32 bytes, emit a logical right shift of the destination. 5762 if (shiftVal < 32) { 5763 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 5764 5765 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5766 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5767 5768 // create i32 constant 5769 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 5770 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5771 } 5772 5773 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5774 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5775 } 5776 case X86::BI__builtin_ia32_movntps: 5777 case X86::BI__builtin_ia32_movntps256: 5778 case X86::BI__builtin_ia32_movntpd: 5779 case X86::BI__builtin_ia32_movntpd256: 5780 case X86::BI__builtin_ia32_movntdq: 5781 case X86::BI__builtin_ia32_movntdq256: 5782 case X86::BI__builtin_ia32_movnti: 5783 case X86::BI__builtin_ia32_movnti64: { 5784 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 5785 Builder.getInt32(1)); 5786 5787 // Convert the type of the pointer to a pointer to the stored type. 5788 Value *BC = Builder.CreateBitCast(Ops[0], 5789 llvm::PointerType::getUnqual(Ops[1]->getType()), 5790 "cast"); 5791 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 5792 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 5793 5794 // If the operand is an integer, we can't assume alignment. Otherwise, 5795 // assume natural alignment. 5796 QualType ArgTy = E->getArg(1)->getType(); 5797 unsigned Align; 5798 if (ArgTy->isIntegerType()) 5799 Align = 1; 5800 else 5801 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 5802 SI->setAlignment(Align); 5803 return SI; 5804 } 5805 // 3DNow! 5806 case X86::BI__builtin_ia32_pswapdsf: 5807 case X86::BI__builtin_ia32_pswapdsi: { 5808 const char *name = nullptr; 5809 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5810 switch(BuiltinID) { 5811 default: llvm_unreachable("Unsupported intrinsic!"); 5812 case X86::BI__builtin_ia32_pswapdsf: 5813 case X86::BI__builtin_ia32_pswapdsi: 5814 name = "pswapd"; 5815 ID = Intrinsic::x86_3dnowa_pswapd; 5816 break; 5817 } 5818 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 5819 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 5820 llvm::Function *F = CGM.getIntrinsic(ID); 5821 return Builder.CreateCall(F, Ops, name); 5822 } 5823 case X86::BI__builtin_ia32_rdrand16_step: 5824 case X86::BI__builtin_ia32_rdrand32_step: 5825 case X86::BI__builtin_ia32_rdrand64_step: 5826 case X86::BI__builtin_ia32_rdseed16_step: 5827 case X86::BI__builtin_ia32_rdseed32_step: 5828 case X86::BI__builtin_ia32_rdseed64_step: { 5829 Intrinsic::ID ID; 5830 switch (BuiltinID) { 5831 default: llvm_unreachable("Unsupported intrinsic!"); 5832 case X86::BI__builtin_ia32_rdrand16_step: 5833 ID = Intrinsic::x86_rdrand_16; 5834 break; 5835 case X86::BI__builtin_ia32_rdrand32_step: 5836 ID = Intrinsic::x86_rdrand_32; 5837 break; 5838 case X86::BI__builtin_ia32_rdrand64_step: 5839 ID = Intrinsic::x86_rdrand_64; 5840 break; 5841 case X86::BI__builtin_ia32_rdseed16_step: 5842 ID = Intrinsic::x86_rdseed_16; 5843 break; 5844 case X86::BI__builtin_ia32_rdseed32_step: 5845 ID = Intrinsic::x86_rdseed_32; 5846 break; 5847 case X86::BI__builtin_ia32_rdseed64_step: 5848 ID = Intrinsic::x86_rdseed_64; 5849 break; 5850 } 5851 5852 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 5853 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 5854 return Builder.CreateExtractValue(Call, 1); 5855 } 5856 // AVX2 broadcast 5857 case X86::BI__builtin_ia32_vbroadcastsi256: { 5858 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 5859 Builder.CreateStore(Ops[0], VecTmp); 5860 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 5861 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 5862 } 5863 } 5864 } 5865 5866 5867 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 5868 const CallExpr *E) { 5869 SmallVector<Value*, 4> Ops; 5870 5871 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 5872 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5873 5874 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5875 5876 switch (BuiltinID) { 5877 default: return nullptr; 5878 5879 // vec_ld, vec_lvsl, vec_lvsr 5880 case PPC::BI__builtin_altivec_lvx: 5881 case PPC::BI__builtin_altivec_lvxl: 5882 case PPC::BI__builtin_altivec_lvebx: 5883 case PPC::BI__builtin_altivec_lvehx: 5884 case PPC::BI__builtin_altivec_lvewx: 5885 case PPC::BI__builtin_altivec_lvsl: 5886 case PPC::BI__builtin_altivec_lvsr: 5887 { 5888 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 5889 5890 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 5891 Ops.pop_back(); 5892 5893 switch (BuiltinID) { 5894 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 5895 case PPC::BI__builtin_altivec_lvx: 5896 ID = Intrinsic::ppc_altivec_lvx; 5897 break; 5898 case PPC::BI__builtin_altivec_lvxl: 5899 ID = Intrinsic::ppc_altivec_lvxl; 5900 break; 5901 case PPC::BI__builtin_altivec_lvebx: 5902 ID = Intrinsic::ppc_altivec_lvebx; 5903 break; 5904 case PPC::BI__builtin_altivec_lvehx: 5905 ID = Intrinsic::ppc_altivec_lvehx; 5906 break; 5907 case PPC::BI__builtin_altivec_lvewx: 5908 ID = Intrinsic::ppc_altivec_lvewx; 5909 break; 5910 case PPC::BI__builtin_altivec_lvsl: 5911 ID = Intrinsic::ppc_altivec_lvsl; 5912 break; 5913 case PPC::BI__builtin_altivec_lvsr: 5914 ID = Intrinsic::ppc_altivec_lvsr; 5915 break; 5916 } 5917 llvm::Function *F = CGM.getIntrinsic(ID); 5918 return Builder.CreateCall(F, Ops, ""); 5919 } 5920 5921 // vec_st 5922 case PPC::BI__builtin_altivec_stvx: 5923 case PPC::BI__builtin_altivec_stvxl: 5924 case PPC::BI__builtin_altivec_stvebx: 5925 case PPC::BI__builtin_altivec_stvehx: 5926 case PPC::BI__builtin_altivec_stvewx: 5927 { 5928 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 5929 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 5930 Ops.pop_back(); 5931 5932 switch (BuiltinID) { 5933 default: llvm_unreachable("Unsupported st intrinsic!"); 5934 case PPC::BI__builtin_altivec_stvx: 5935 ID = Intrinsic::ppc_altivec_stvx; 5936 break; 5937 case PPC::BI__builtin_altivec_stvxl: 5938 ID = Intrinsic::ppc_altivec_stvxl; 5939 break; 5940 case PPC::BI__builtin_altivec_stvebx: 5941 ID = Intrinsic::ppc_altivec_stvebx; 5942 break; 5943 case PPC::BI__builtin_altivec_stvehx: 5944 ID = Intrinsic::ppc_altivec_stvehx; 5945 break; 5946 case PPC::BI__builtin_altivec_stvewx: 5947 ID = Intrinsic::ppc_altivec_stvewx; 5948 break; 5949 } 5950 llvm::Function *F = CGM.getIntrinsic(ID); 5951 return Builder.CreateCall(F, Ops, ""); 5952 } 5953 } 5954 } 5955