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, 170 ReturnValueSlot(), 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 SanitizerScope SanScope(this); 452 EmitCheck(Builder.getFalse(), "builtin_unreachable", 453 EmitCheckSourceLocation(E->getExprLoc()), 454 ArrayRef<llvm::Value *>(), CRK_Unrecoverable); 455 } else 456 Builder.CreateUnreachable(); 457 458 // We do need to preserve an insertion point. 459 EmitBlock(createBasicBlock("unreachable.cont")); 460 461 return RValue::get(nullptr); 462 } 463 464 case Builtin::BI__builtin_powi: 465 case Builtin::BI__builtin_powif: 466 case Builtin::BI__builtin_powil: { 467 Value *Base = EmitScalarExpr(E->getArg(0)); 468 Value *Exponent = EmitScalarExpr(E->getArg(1)); 469 llvm::Type *ArgType = Base->getType(); 470 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 471 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 472 } 473 474 case Builtin::BI__builtin_isgreater: 475 case Builtin::BI__builtin_isgreaterequal: 476 case Builtin::BI__builtin_isless: 477 case Builtin::BI__builtin_islessequal: 478 case Builtin::BI__builtin_islessgreater: 479 case Builtin::BI__builtin_isunordered: { 480 // Ordered comparisons: we know the arguments to these are matching scalar 481 // floating point values. 482 Value *LHS = EmitScalarExpr(E->getArg(0)); 483 Value *RHS = EmitScalarExpr(E->getArg(1)); 484 485 switch (BuiltinID) { 486 default: llvm_unreachable("Unknown ordered comparison"); 487 case Builtin::BI__builtin_isgreater: 488 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 489 break; 490 case Builtin::BI__builtin_isgreaterequal: 491 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 492 break; 493 case Builtin::BI__builtin_isless: 494 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 495 break; 496 case Builtin::BI__builtin_islessequal: 497 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 498 break; 499 case Builtin::BI__builtin_islessgreater: 500 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 501 break; 502 case Builtin::BI__builtin_isunordered: 503 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 504 break; 505 } 506 // ZExt bool to int type. 507 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 508 } 509 case Builtin::BI__builtin_isnan: { 510 Value *V = EmitScalarExpr(E->getArg(0)); 511 V = Builder.CreateFCmpUNO(V, V, "cmp"); 512 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 513 } 514 515 case Builtin::BI__builtin_isinf: { 516 // isinf(x) --> fabs(x) == infinity 517 Value *V = EmitScalarExpr(E->getArg(0)); 518 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 519 520 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 521 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 522 } 523 524 // TODO: BI__builtin_isinf_sign 525 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 526 527 case Builtin::BI__builtin_isnormal: { 528 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 529 Value *V = EmitScalarExpr(E->getArg(0)); 530 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 531 532 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 533 Value *IsLessThanInf = 534 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 535 APFloat Smallest = APFloat::getSmallestNormalized( 536 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 537 Value *IsNormal = 538 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 539 "isnormal"); 540 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 541 V = Builder.CreateAnd(V, IsNormal, "and"); 542 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 543 } 544 545 case Builtin::BI__builtin_isfinite: { 546 // isfinite(x) --> x == x && fabs(x) != infinity; 547 Value *V = EmitScalarExpr(E->getArg(0)); 548 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 549 550 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 551 Value *IsNotInf = 552 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 553 554 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 555 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 556 } 557 558 case Builtin::BI__builtin_fpclassify: { 559 Value *V = EmitScalarExpr(E->getArg(5)); 560 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 561 562 // Create Result 563 BasicBlock *Begin = Builder.GetInsertBlock(); 564 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 565 Builder.SetInsertPoint(End); 566 PHINode *Result = 567 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 568 "fpclassify_result"); 569 570 // if (V==0) return FP_ZERO 571 Builder.SetInsertPoint(Begin); 572 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 573 "iszero"); 574 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 575 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 576 Builder.CreateCondBr(IsZero, End, NotZero); 577 Result->addIncoming(ZeroLiteral, Begin); 578 579 // if (V != V) return FP_NAN 580 Builder.SetInsertPoint(NotZero); 581 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 582 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 583 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 584 Builder.CreateCondBr(IsNan, End, NotNan); 585 Result->addIncoming(NanLiteral, NotZero); 586 587 // if (fabs(V) == infinity) return FP_INFINITY 588 Builder.SetInsertPoint(NotNan); 589 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 590 Value *IsInf = 591 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 592 "isinf"); 593 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 594 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 595 Builder.CreateCondBr(IsInf, End, NotInf); 596 Result->addIncoming(InfLiteral, NotNan); 597 598 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 599 Builder.SetInsertPoint(NotInf); 600 APFloat Smallest = APFloat::getSmallestNormalized( 601 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 602 Value *IsNormal = 603 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 604 "isnormal"); 605 Value *NormalResult = 606 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 607 EmitScalarExpr(E->getArg(3))); 608 Builder.CreateBr(End); 609 Result->addIncoming(NormalResult, NotInf); 610 611 // return Result 612 Builder.SetInsertPoint(End); 613 return RValue::get(Result); 614 } 615 616 case Builtin::BIalloca: 617 case Builtin::BI_alloca: 618 case Builtin::BI__builtin_alloca: { 619 Value *Size = EmitScalarExpr(E->getArg(0)); 620 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 621 } 622 case Builtin::BIbzero: 623 case Builtin::BI__builtin_bzero: { 624 std::pair<llvm::Value*, unsigned> Dest = 625 EmitPointerWithAlignment(E->getArg(0)); 626 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 627 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 628 Dest.second, false); 629 return RValue::get(Dest.first); 630 } 631 case Builtin::BImemcpy: 632 case Builtin::BI__builtin_memcpy: { 633 std::pair<llvm::Value*, unsigned> Dest = 634 EmitPointerWithAlignment(E->getArg(0)); 635 std::pair<llvm::Value*, unsigned> Src = 636 EmitPointerWithAlignment(E->getArg(1)); 637 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 638 unsigned Align = std::min(Dest.second, Src.second); 639 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 640 return RValue::get(Dest.first); 641 } 642 643 case Builtin::BI__builtin___memcpy_chk: { 644 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 645 llvm::APSInt Size, DstSize; 646 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 647 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 648 break; 649 if (Size.ugt(DstSize)) 650 break; 651 std::pair<llvm::Value*, unsigned> Dest = 652 EmitPointerWithAlignment(E->getArg(0)); 653 std::pair<llvm::Value*, unsigned> Src = 654 EmitPointerWithAlignment(E->getArg(1)); 655 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 656 unsigned Align = std::min(Dest.second, Src.second); 657 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 658 return RValue::get(Dest.first); 659 } 660 661 case Builtin::BI__builtin_objc_memmove_collectable: { 662 Value *Address = EmitScalarExpr(E->getArg(0)); 663 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 664 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 665 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 666 Address, SrcAddr, SizeVal); 667 return RValue::get(Address); 668 } 669 670 case Builtin::BI__builtin___memmove_chk: { 671 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 672 llvm::APSInt Size, DstSize; 673 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 674 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 675 break; 676 if (Size.ugt(DstSize)) 677 break; 678 std::pair<llvm::Value*, unsigned> Dest = 679 EmitPointerWithAlignment(E->getArg(0)); 680 std::pair<llvm::Value*, unsigned> Src = 681 EmitPointerWithAlignment(E->getArg(1)); 682 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 683 unsigned Align = std::min(Dest.second, Src.second); 684 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 685 return RValue::get(Dest.first); 686 } 687 688 case Builtin::BImemmove: 689 case Builtin::BI__builtin_memmove: { 690 std::pair<llvm::Value*, unsigned> Dest = 691 EmitPointerWithAlignment(E->getArg(0)); 692 std::pair<llvm::Value*, unsigned> Src = 693 EmitPointerWithAlignment(E->getArg(1)); 694 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 695 unsigned Align = std::min(Dest.second, Src.second); 696 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 697 return RValue::get(Dest.first); 698 } 699 case Builtin::BImemset: 700 case Builtin::BI__builtin_memset: { 701 std::pair<llvm::Value*, unsigned> Dest = 702 EmitPointerWithAlignment(E->getArg(0)); 703 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 704 Builder.getInt8Ty()); 705 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 706 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 707 return RValue::get(Dest.first); 708 } 709 case Builtin::BI__builtin___memset_chk: { 710 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 711 llvm::APSInt Size, DstSize; 712 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 713 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 714 break; 715 if (Size.ugt(DstSize)) 716 break; 717 std::pair<llvm::Value*, unsigned> Dest = 718 EmitPointerWithAlignment(E->getArg(0)); 719 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 720 Builder.getInt8Ty()); 721 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 722 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 723 return RValue::get(Dest.first); 724 } 725 case Builtin::BI__builtin_dwarf_cfa: { 726 // The offset in bytes from the first argument to the CFA. 727 // 728 // Why on earth is this in the frontend? Is there any reason at 729 // all that the backend can't reasonably determine this while 730 // lowering llvm.eh.dwarf.cfa()? 731 // 732 // TODO: If there's a satisfactory reason, add a target hook for 733 // this instead of hard-coding 0, which is correct for most targets. 734 int32_t Offset = 0; 735 736 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 737 return RValue::get(Builder.CreateCall(F, 738 llvm::ConstantInt::get(Int32Ty, Offset))); 739 } 740 case Builtin::BI__builtin_return_address: { 741 Value *Depth = EmitScalarExpr(E->getArg(0)); 742 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 743 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 744 return RValue::get(Builder.CreateCall(F, Depth)); 745 } 746 case Builtin::BI__builtin_frame_address: { 747 Value *Depth = EmitScalarExpr(E->getArg(0)); 748 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 749 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 750 return RValue::get(Builder.CreateCall(F, Depth)); 751 } 752 case Builtin::BI__builtin_extract_return_addr: { 753 Value *Address = EmitScalarExpr(E->getArg(0)); 754 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 755 return RValue::get(Result); 756 } 757 case Builtin::BI__builtin_frob_return_addr: { 758 Value *Address = EmitScalarExpr(E->getArg(0)); 759 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 760 return RValue::get(Result); 761 } 762 case Builtin::BI__builtin_dwarf_sp_column: { 763 llvm::IntegerType *Ty 764 = cast<llvm::IntegerType>(ConvertType(E->getType())); 765 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 766 if (Column == -1) { 767 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 768 return RValue::get(llvm::UndefValue::get(Ty)); 769 } 770 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 771 } 772 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 773 Value *Address = EmitScalarExpr(E->getArg(0)); 774 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 775 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 776 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 777 } 778 case Builtin::BI__builtin_eh_return: { 779 Value *Int = EmitScalarExpr(E->getArg(0)); 780 Value *Ptr = EmitScalarExpr(E->getArg(1)); 781 782 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 783 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 784 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 785 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 786 ? Intrinsic::eh_return_i32 787 : Intrinsic::eh_return_i64); 788 Builder.CreateCall2(F, Int, Ptr); 789 Builder.CreateUnreachable(); 790 791 // We do need to preserve an insertion point. 792 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 793 794 return RValue::get(nullptr); 795 } 796 case Builtin::BI__builtin_unwind_init: { 797 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 798 return RValue::get(Builder.CreateCall(F)); 799 } 800 case Builtin::BI__builtin_extend_pointer: { 801 // Extends a pointer to the size of an _Unwind_Word, which is 802 // uint64_t on all platforms. Generally this gets poked into a 803 // register and eventually used as an address, so if the 804 // addressing registers are wider than pointers and the platform 805 // doesn't implicitly ignore high-order bits when doing 806 // addressing, we need to make sure we zext / sext based on 807 // the platform's expectations. 808 // 809 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 810 811 // Cast the pointer to intptr_t. 812 Value *Ptr = EmitScalarExpr(E->getArg(0)); 813 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 814 815 // If that's 64 bits, we're done. 816 if (IntPtrTy->getBitWidth() == 64) 817 return RValue::get(Result); 818 819 // Otherwise, ask the codegen data what to do. 820 if (getTargetHooks().extendPointerWithSExt()) 821 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 822 else 823 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 824 } 825 case Builtin::BI__builtin_setjmp: { 826 // Buffer is a void**. 827 Value *Buf = EmitScalarExpr(E->getArg(0)); 828 829 // Store the frame pointer to the setjmp buffer. 830 Value *FrameAddr = 831 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 832 ConstantInt::get(Int32Ty, 0)); 833 Builder.CreateStore(FrameAddr, Buf); 834 835 // Store the stack pointer to the setjmp buffer. 836 Value *StackAddr = 837 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 838 Value *StackSaveSlot = 839 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 840 Builder.CreateStore(StackAddr, StackSaveSlot); 841 842 // Call LLVM's EH setjmp, which is lightweight. 843 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 844 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 845 return RValue::get(Builder.CreateCall(F, Buf)); 846 } 847 case Builtin::BI__builtin_longjmp: { 848 Value *Buf = EmitScalarExpr(E->getArg(0)); 849 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 850 851 // Call LLVM's EH longjmp, which is lightweight. 852 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 853 854 // longjmp doesn't return; mark this as unreachable. 855 Builder.CreateUnreachable(); 856 857 // We do need to preserve an insertion point. 858 EmitBlock(createBasicBlock("longjmp.cont")); 859 860 return RValue::get(nullptr); 861 } 862 case Builtin::BI__sync_fetch_and_add: 863 case Builtin::BI__sync_fetch_and_sub: 864 case Builtin::BI__sync_fetch_and_or: 865 case Builtin::BI__sync_fetch_and_and: 866 case Builtin::BI__sync_fetch_and_xor: 867 case Builtin::BI__sync_add_and_fetch: 868 case Builtin::BI__sync_sub_and_fetch: 869 case Builtin::BI__sync_and_and_fetch: 870 case Builtin::BI__sync_or_and_fetch: 871 case Builtin::BI__sync_xor_and_fetch: 872 case Builtin::BI__sync_val_compare_and_swap: 873 case Builtin::BI__sync_bool_compare_and_swap: 874 case Builtin::BI__sync_lock_test_and_set: 875 case Builtin::BI__sync_lock_release: 876 case Builtin::BI__sync_swap: 877 llvm_unreachable("Shouldn't make it through sema"); 878 case Builtin::BI__sync_fetch_and_add_1: 879 case Builtin::BI__sync_fetch_and_add_2: 880 case Builtin::BI__sync_fetch_and_add_4: 881 case Builtin::BI__sync_fetch_and_add_8: 882 case Builtin::BI__sync_fetch_and_add_16: 883 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 884 case Builtin::BI__sync_fetch_and_sub_1: 885 case Builtin::BI__sync_fetch_and_sub_2: 886 case Builtin::BI__sync_fetch_and_sub_4: 887 case Builtin::BI__sync_fetch_and_sub_8: 888 case Builtin::BI__sync_fetch_and_sub_16: 889 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 890 case Builtin::BI__sync_fetch_and_or_1: 891 case Builtin::BI__sync_fetch_and_or_2: 892 case Builtin::BI__sync_fetch_and_or_4: 893 case Builtin::BI__sync_fetch_and_or_8: 894 case Builtin::BI__sync_fetch_and_or_16: 895 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 896 case Builtin::BI__sync_fetch_and_and_1: 897 case Builtin::BI__sync_fetch_and_and_2: 898 case Builtin::BI__sync_fetch_and_and_4: 899 case Builtin::BI__sync_fetch_and_and_8: 900 case Builtin::BI__sync_fetch_and_and_16: 901 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 902 case Builtin::BI__sync_fetch_and_xor_1: 903 case Builtin::BI__sync_fetch_and_xor_2: 904 case Builtin::BI__sync_fetch_and_xor_4: 905 case Builtin::BI__sync_fetch_and_xor_8: 906 case Builtin::BI__sync_fetch_and_xor_16: 907 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 908 909 // Clang extensions: not overloaded yet. 910 case Builtin::BI__sync_fetch_and_min: 911 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 912 case Builtin::BI__sync_fetch_and_max: 913 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 914 case Builtin::BI__sync_fetch_and_umin: 915 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 916 case Builtin::BI__sync_fetch_and_umax: 917 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 918 919 case Builtin::BI__sync_add_and_fetch_1: 920 case Builtin::BI__sync_add_and_fetch_2: 921 case Builtin::BI__sync_add_and_fetch_4: 922 case Builtin::BI__sync_add_and_fetch_8: 923 case Builtin::BI__sync_add_and_fetch_16: 924 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 925 llvm::Instruction::Add); 926 case Builtin::BI__sync_sub_and_fetch_1: 927 case Builtin::BI__sync_sub_and_fetch_2: 928 case Builtin::BI__sync_sub_and_fetch_4: 929 case Builtin::BI__sync_sub_and_fetch_8: 930 case Builtin::BI__sync_sub_and_fetch_16: 931 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 932 llvm::Instruction::Sub); 933 case Builtin::BI__sync_and_and_fetch_1: 934 case Builtin::BI__sync_and_and_fetch_2: 935 case Builtin::BI__sync_and_and_fetch_4: 936 case Builtin::BI__sync_and_and_fetch_8: 937 case Builtin::BI__sync_and_and_fetch_16: 938 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 939 llvm::Instruction::And); 940 case Builtin::BI__sync_or_and_fetch_1: 941 case Builtin::BI__sync_or_and_fetch_2: 942 case Builtin::BI__sync_or_and_fetch_4: 943 case Builtin::BI__sync_or_and_fetch_8: 944 case Builtin::BI__sync_or_and_fetch_16: 945 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 946 llvm::Instruction::Or); 947 case Builtin::BI__sync_xor_and_fetch_1: 948 case Builtin::BI__sync_xor_and_fetch_2: 949 case Builtin::BI__sync_xor_and_fetch_4: 950 case Builtin::BI__sync_xor_and_fetch_8: 951 case Builtin::BI__sync_xor_and_fetch_16: 952 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 953 llvm::Instruction::Xor); 954 955 case Builtin::BI__sync_val_compare_and_swap_1: 956 case Builtin::BI__sync_val_compare_and_swap_2: 957 case Builtin::BI__sync_val_compare_and_swap_4: 958 case Builtin::BI__sync_val_compare_and_swap_8: 959 case Builtin::BI__sync_val_compare_and_swap_16: { 960 QualType T = E->getType(); 961 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 962 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 963 964 llvm::IntegerType *IntType = 965 llvm::IntegerType::get(getLLVMContext(), 966 getContext().getTypeSize(T)); 967 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 968 969 Value *Args[3]; 970 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 971 Args[1] = EmitScalarExpr(E->getArg(1)); 972 llvm::Type *ValueType = Args[1]->getType(); 973 Args[1] = EmitToInt(*this, Args[1], T, IntType); 974 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 975 976 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 977 llvm::SequentiallyConsistent, 978 llvm::SequentiallyConsistent); 979 Result = Builder.CreateExtractValue(Result, 0); 980 Result = EmitFromInt(*this, Result, T, ValueType); 981 return RValue::get(Result); 982 } 983 984 case Builtin::BI__sync_bool_compare_and_swap_1: 985 case Builtin::BI__sync_bool_compare_and_swap_2: 986 case Builtin::BI__sync_bool_compare_and_swap_4: 987 case Builtin::BI__sync_bool_compare_and_swap_8: 988 case Builtin::BI__sync_bool_compare_and_swap_16: { 989 QualType T = E->getArg(1)->getType(); 990 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 991 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 992 993 llvm::IntegerType *IntType = 994 llvm::IntegerType::get(getLLVMContext(), 995 getContext().getTypeSize(T)); 996 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 997 998 Value *Args[3]; 999 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1000 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1001 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1002 1003 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1004 llvm::SequentiallyConsistent, 1005 llvm::SequentiallyConsistent); 1006 Value *Result = Builder.CreateExtractValue(Pair, 1); 1007 // zext bool to int. 1008 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1009 return RValue::get(Result); 1010 } 1011 1012 case Builtin::BI__sync_swap_1: 1013 case Builtin::BI__sync_swap_2: 1014 case Builtin::BI__sync_swap_4: 1015 case Builtin::BI__sync_swap_8: 1016 case Builtin::BI__sync_swap_16: 1017 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1018 1019 case Builtin::BI__sync_lock_test_and_set_1: 1020 case Builtin::BI__sync_lock_test_and_set_2: 1021 case Builtin::BI__sync_lock_test_and_set_4: 1022 case Builtin::BI__sync_lock_test_and_set_8: 1023 case Builtin::BI__sync_lock_test_and_set_16: 1024 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1025 1026 case Builtin::BI__sync_lock_release_1: 1027 case Builtin::BI__sync_lock_release_2: 1028 case Builtin::BI__sync_lock_release_4: 1029 case Builtin::BI__sync_lock_release_8: 1030 case Builtin::BI__sync_lock_release_16: { 1031 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1032 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1033 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1034 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1035 StoreSize.getQuantity() * 8); 1036 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1037 llvm::StoreInst *Store = 1038 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1039 Store->setAlignment(StoreSize.getQuantity()); 1040 Store->setAtomic(llvm::Release); 1041 return RValue::get(nullptr); 1042 } 1043 1044 case Builtin::BI__sync_synchronize: { 1045 // We assume this is supposed to correspond to a C++0x-style 1046 // sequentially-consistent fence (i.e. this is only usable for 1047 // synchonization, not device I/O or anything like that). This intrinsic 1048 // is really badly designed in the sense that in theory, there isn't 1049 // any way to safely use it... but in practice, it mostly works 1050 // to use it with non-atomic loads and stores to get acquire/release 1051 // semantics. 1052 Builder.CreateFence(llvm::SequentiallyConsistent); 1053 return RValue::get(nullptr); 1054 } 1055 1056 case Builtin::BI__c11_atomic_is_lock_free: 1057 case Builtin::BI__atomic_is_lock_free: { 1058 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1059 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1060 // _Atomic(T) is always properly-aligned. 1061 const char *LibCallName = "__atomic_is_lock_free"; 1062 CallArgList Args; 1063 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1064 getContext().getSizeType()); 1065 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1066 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1067 getContext().VoidPtrTy); 1068 else 1069 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1070 getContext().VoidPtrTy); 1071 const CGFunctionInfo &FuncInfo = 1072 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1073 FunctionType::ExtInfo(), 1074 RequiredArgs::All); 1075 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1076 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1077 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1078 } 1079 1080 case Builtin::BI__atomic_test_and_set: { 1081 // Look at the argument type to determine whether this is a volatile 1082 // operation. The parameter type is always volatile. 1083 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1084 bool Volatile = 1085 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1086 1087 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1088 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1089 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1090 Value *NewVal = Builder.getInt8(1); 1091 Value *Order = EmitScalarExpr(E->getArg(1)); 1092 if (isa<llvm::ConstantInt>(Order)) { 1093 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1094 AtomicRMWInst *Result = nullptr; 1095 switch (ord) { 1096 case 0: // memory_order_relaxed 1097 default: // invalid order 1098 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1099 Ptr, NewVal, 1100 llvm::Monotonic); 1101 break; 1102 case 1: // memory_order_consume 1103 case 2: // memory_order_acquire 1104 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1105 Ptr, NewVal, 1106 llvm::Acquire); 1107 break; 1108 case 3: // memory_order_release 1109 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1110 Ptr, NewVal, 1111 llvm::Release); 1112 break; 1113 case 4: // memory_order_acq_rel 1114 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1115 Ptr, NewVal, 1116 llvm::AcquireRelease); 1117 break; 1118 case 5: // memory_order_seq_cst 1119 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1120 Ptr, NewVal, 1121 llvm::SequentiallyConsistent); 1122 break; 1123 } 1124 Result->setVolatile(Volatile); 1125 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1126 } 1127 1128 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1129 1130 llvm::BasicBlock *BBs[5] = { 1131 createBasicBlock("monotonic", CurFn), 1132 createBasicBlock("acquire", CurFn), 1133 createBasicBlock("release", CurFn), 1134 createBasicBlock("acqrel", CurFn), 1135 createBasicBlock("seqcst", CurFn) 1136 }; 1137 llvm::AtomicOrdering Orders[5] = { 1138 llvm::Monotonic, llvm::Acquire, llvm::Release, 1139 llvm::AcquireRelease, llvm::SequentiallyConsistent 1140 }; 1141 1142 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1143 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1144 1145 Builder.SetInsertPoint(ContBB); 1146 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1147 1148 for (unsigned i = 0; i < 5; ++i) { 1149 Builder.SetInsertPoint(BBs[i]); 1150 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1151 Ptr, NewVal, Orders[i]); 1152 RMW->setVolatile(Volatile); 1153 Result->addIncoming(RMW, BBs[i]); 1154 Builder.CreateBr(ContBB); 1155 } 1156 1157 SI->addCase(Builder.getInt32(0), BBs[0]); 1158 SI->addCase(Builder.getInt32(1), BBs[1]); 1159 SI->addCase(Builder.getInt32(2), BBs[1]); 1160 SI->addCase(Builder.getInt32(3), BBs[2]); 1161 SI->addCase(Builder.getInt32(4), BBs[3]); 1162 SI->addCase(Builder.getInt32(5), BBs[4]); 1163 1164 Builder.SetInsertPoint(ContBB); 1165 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1166 } 1167 1168 case Builtin::BI__atomic_clear: { 1169 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1170 bool Volatile = 1171 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1172 1173 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1174 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1175 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1176 Value *NewVal = Builder.getInt8(0); 1177 Value *Order = EmitScalarExpr(E->getArg(1)); 1178 if (isa<llvm::ConstantInt>(Order)) { 1179 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1180 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1181 Store->setAlignment(1); 1182 switch (ord) { 1183 case 0: // memory_order_relaxed 1184 default: // invalid order 1185 Store->setOrdering(llvm::Monotonic); 1186 break; 1187 case 3: // memory_order_release 1188 Store->setOrdering(llvm::Release); 1189 break; 1190 case 5: // memory_order_seq_cst 1191 Store->setOrdering(llvm::SequentiallyConsistent); 1192 break; 1193 } 1194 return RValue::get(nullptr); 1195 } 1196 1197 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1198 1199 llvm::BasicBlock *BBs[3] = { 1200 createBasicBlock("monotonic", CurFn), 1201 createBasicBlock("release", CurFn), 1202 createBasicBlock("seqcst", CurFn) 1203 }; 1204 llvm::AtomicOrdering Orders[3] = { 1205 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1206 }; 1207 1208 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1209 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1210 1211 for (unsigned i = 0; i < 3; ++i) { 1212 Builder.SetInsertPoint(BBs[i]); 1213 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1214 Store->setAlignment(1); 1215 Store->setOrdering(Orders[i]); 1216 Builder.CreateBr(ContBB); 1217 } 1218 1219 SI->addCase(Builder.getInt32(0), BBs[0]); 1220 SI->addCase(Builder.getInt32(3), BBs[1]); 1221 SI->addCase(Builder.getInt32(5), BBs[2]); 1222 1223 Builder.SetInsertPoint(ContBB); 1224 return RValue::get(nullptr); 1225 } 1226 1227 case Builtin::BI__atomic_thread_fence: 1228 case Builtin::BI__atomic_signal_fence: 1229 case Builtin::BI__c11_atomic_thread_fence: 1230 case Builtin::BI__c11_atomic_signal_fence: { 1231 llvm::SynchronizationScope Scope; 1232 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1233 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1234 Scope = llvm::SingleThread; 1235 else 1236 Scope = llvm::CrossThread; 1237 Value *Order = EmitScalarExpr(E->getArg(0)); 1238 if (isa<llvm::ConstantInt>(Order)) { 1239 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1240 switch (ord) { 1241 case 0: // memory_order_relaxed 1242 default: // invalid order 1243 break; 1244 case 1: // memory_order_consume 1245 case 2: // memory_order_acquire 1246 Builder.CreateFence(llvm::Acquire, Scope); 1247 break; 1248 case 3: // memory_order_release 1249 Builder.CreateFence(llvm::Release, Scope); 1250 break; 1251 case 4: // memory_order_acq_rel 1252 Builder.CreateFence(llvm::AcquireRelease, Scope); 1253 break; 1254 case 5: // memory_order_seq_cst 1255 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1256 break; 1257 } 1258 return RValue::get(nullptr); 1259 } 1260 1261 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1262 AcquireBB = createBasicBlock("acquire", CurFn); 1263 ReleaseBB = createBasicBlock("release", CurFn); 1264 AcqRelBB = createBasicBlock("acqrel", CurFn); 1265 SeqCstBB = createBasicBlock("seqcst", CurFn); 1266 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1267 1268 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1269 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1270 1271 Builder.SetInsertPoint(AcquireBB); 1272 Builder.CreateFence(llvm::Acquire, Scope); 1273 Builder.CreateBr(ContBB); 1274 SI->addCase(Builder.getInt32(1), AcquireBB); 1275 SI->addCase(Builder.getInt32(2), AcquireBB); 1276 1277 Builder.SetInsertPoint(ReleaseBB); 1278 Builder.CreateFence(llvm::Release, Scope); 1279 Builder.CreateBr(ContBB); 1280 SI->addCase(Builder.getInt32(3), ReleaseBB); 1281 1282 Builder.SetInsertPoint(AcqRelBB); 1283 Builder.CreateFence(llvm::AcquireRelease, Scope); 1284 Builder.CreateBr(ContBB); 1285 SI->addCase(Builder.getInt32(4), AcqRelBB); 1286 1287 Builder.SetInsertPoint(SeqCstBB); 1288 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1289 Builder.CreateBr(ContBB); 1290 SI->addCase(Builder.getInt32(5), SeqCstBB); 1291 1292 Builder.SetInsertPoint(ContBB); 1293 return RValue::get(nullptr); 1294 } 1295 1296 // Library functions with special handling. 1297 case Builtin::BIsqrt: 1298 case Builtin::BIsqrtf: 1299 case Builtin::BIsqrtl: { 1300 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1301 // in finite- or unsafe-math mode (the intrinsic has different semantics 1302 // for handling negative numbers compared to the library function, so 1303 // -fmath-errno=0 is not enough). 1304 if (!FD->hasAttr<ConstAttr>()) 1305 break; 1306 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1307 CGM.getCodeGenOpts().NoNaNsFPMath)) 1308 break; 1309 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1310 llvm::Type *ArgType = Arg0->getType(); 1311 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1312 return RValue::get(Builder.CreateCall(F, Arg0)); 1313 } 1314 1315 case Builtin::BIpow: 1316 case Builtin::BIpowf: 1317 case Builtin::BIpowl: { 1318 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1319 if (!FD->hasAttr<ConstAttr>()) 1320 break; 1321 Value *Base = EmitScalarExpr(E->getArg(0)); 1322 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1323 llvm::Type *ArgType = Base->getType(); 1324 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1325 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1326 } 1327 1328 case Builtin::BIfma: 1329 case Builtin::BIfmaf: 1330 case Builtin::BIfmal: 1331 case Builtin::BI__builtin_fma: 1332 case Builtin::BI__builtin_fmaf: 1333 case Builtin::BI__builtin_fmal: { 1334 // Rewrite fma to intrinsic. 1335 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1336 llvm::Type *ArgType = FirstArg->getType(); 1337 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1338 return RValue::get(Builder.CreateCall3(F, FirstArg, 1339 EmitScalarExpr(E->getArg(1)), 1340 EmitScalarExpr(E->getArg(2)))); 1341 } 1342 1343 case Builtin::BI__builtin_signbit: 1344 case Builtin::BI__builtin_signbitf: 1345 case Builtin::BI__builtin_signbitl: { 1346 LLVMContext &C = CGM.getLLVMContext(); 1347 1348 Value *Arg = EmitScalarExpr(E->getArg(0)); 1349 llvm::Type *ArgTy = Arg->getType(); 1350 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1351 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1352 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1353 if (ArgTy->isPPC_FP128Ty()) { 1354 // The higher-order double comes first, and so we need to truncate the 1355 // pair to extract the overall sign. The order of the pair is the same 1356 // in both little- and big-Endian modes. 1357 ArgWidth >>= 1; 1358 ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1359 BCArg = Builder.CreateTrunc(BCArg, ArgIntTy); 1360 } 1361 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1362 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1363 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1364 } 1365 case Builtin::BI__builtin_annotation: { 1366 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1367 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1368 AnnVal->getType()); 1369 1370 // Get the annotation string, go through casts. Sema requires this to be a 1371 // non-wide string literal, potentially casted, so the cast<> is safe. 1372 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1373 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1374 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1375 } 1376 case Builtin::BI__builtin_addcb: 1377 case Builtin::BI__builtin_addcs: 1378 case Builtin::BI__builtin_addc: 1379 case Builtin::BI__builtin_addcl: 1380 case Builtin::BI__builtin_addcll: 1381 case Builtin::BI__builtin_subcb: 1382 case Builtin::BI__builtin_subcs: 1383 case Builtin::BI__builtin_subc: 1384 case Builtin::BI__builtin_subcl: 1385 case Builtin::BI__builtin_subcll: { 1386 1387 // We translate all of these builtins from expressions of the form: 1388 // int x = ..., y = ..., carryin = ..., carryout, result; 1389 // result = __builtin_addc(x, y, carryin, &carryout); 1390 // 1391 // to LLVM IR of the form: 1392 // 1393 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1394 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1395 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1396 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1397 // i32 %carryin) 1398 // %result = extractvalue {i32, i1} %tmp2, 0 1399 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1400 // %tmp3 = or i1 %carry1, %carry2 1401 // %tmp4 = zext i1 %tmp3 to i32 1402 // store i32 %tmp4, i32* %carryout 1403 1404 // Scalarize our inputs. 1405 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1406 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1407 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1408 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1409 EmitPointerWithAlignment(E->getArg(3)); 1410 1411 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1412 llvm::Intrinsic::ID IntrinsicId; 1413 switch (BuiltinID) { 1414 default: llvm_unreachable("Unknown multiprecision builtin id."); 1415 case Builtin::BI__builtin_addcb: 1416 case Builtin::BI__builtin_addcs: 1417 case Builtin::BI__builtin_addc: 1418 case Builtin::BI__builtin_addcl: 1419 case Builtin::BI__builtin_addcll: 1420 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1421 break; 1422 case Builtin::BI__builtin_subcb: 1423 case Builtin::BI__builtin_subcs: 1424 case Builtin::BI__builtin_subc: 1425 case Builtin::BI__builtin_subcl: 1426 case Builtin::BI__builtin_subcll: 1427 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1428 break; 1429 } 1430 1431 // Construct our resulting LLVM IR expression. 1432 llvm::Value *Carry1; 1433 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1434 X, Y, Carry1); 1435 llvm::Value *Carry2; 1436 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1437 Sum1, Carryin, Carry2); 1438 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1439 X->getType()); 1440 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1441 CarryOutPtr.first); 1442 CarryOutStore->setAlignment(CarryOutPtr.second); 1443 return RValue::get(Sum2); 1444 } 1445 case Builtin::BI__builtin_uadd_overflow: 1446 case Builtin::BI__builtin_uaddl_overflow: 1447 case Builtin::BI__builtin_uaddll_overflow: 1448 case Builtin::BI__builtin_usub_overflow: 1449 case Builtin::BI__builtin_usubl_overflow: 1450 case Builtin::BI__builtin_usubll_overflow: 1451 case Builtin::BI__builtin_umul_overflow: 1452 case Builtin::BI__builtin_umull_overflow: 1453 case Builtin::BI__builtin_umulll_overflow: 1454 case Builtin::BI__builtin_sadd_overflow: 1455 case Builtin::BI__builtin_saddl_overflow: 1456 case Builtin::BI__builtin_saddll_overflow: 1457 case Builtin::BI__builtin_ssub_overflow: 1458 case Builtin::BI__builtin_ssubl_overflow: 1459 case Builtin::BI__builtin_ssubll_overflow: 1460 case Builtin::BI__builtin_smul_overflow: 1461 case Builtin::BI__builtin_smull_overflow: 1462 case Builtin::BI__builtin_smulll_overflow: { 1463 1464 // We translate all of these builtins directly to the relevant llvm IR node. 1465 1466 // Scalarize our inputs. 1467 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1468 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1469 std::pair<llvm::Value *, unsigned> SumOutPtr = 1470 EmitPointerWithAlignment(E->getArg(2)); 1471 1472 // Decide which of the overflow intrinsics we are lowering to: 1473 llvm::Intrinsic::ID IntrinsicId; 1474 switch (BuiltinID) { 1475 default: llvm_unreachable("Unknown security overflow builtin id."); 1476 case Builtin::BI__builtin_uadd_overflow: 1477 case Builtin::BI__builtin_uaddl_overflow: 1478 case Builtin::BI__builtin_uaddll_overflow: 1479 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1480 break; 1481 case Builtin::BI__builtin_usub_overflow: 1482 case Builtin::BI__builtin_usubl_overflow: 1483 case Builtin::BI__builtin_usubll_overflow: 1484 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1485 break; 1486 case Builtin::BI__builtin_umul_overflow: 1487 case Builtin::BI__builtin_umull_overflow: 1488 case Builtin::BI__builtin_umulll_overflow: 1489 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1490 break; 1491 case Builtin::BI__builtin_sadd_overflow: 1492 case Builtin::BI__builtin_saddl_overflow: 1493 case Builtin::BI__builtin_saddll_overflow: 1494 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1495 break; 1496 case Builtin::BI__builtin_ssub_overflow: 1497 case Builtin::BI__builtin_ssubl_overflow: 1498 case Builtin::BI__builtin_ssubll_overflow: 1499 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1500 break; 1501 case Builtin::BI__builtin_smul_overflow: 1502 case Builtin::BI__builtin_smull_overflow: 1503 case Builtin::BI__builtin_smulll_overflow: 1504 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1505 break; 1506 } 1507 1508 1509 llvm::Value *Carry; 1510 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1511 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1512 SumOutStore->setAlignment(SumOutPtr.second); 1513 1514 return RValue::get(Carry); 1515 } 1516 case Builtin::BI__builtin_addressof: 1517 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1518 case Builtin::BI__builtin_operator_new: 1519 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1520 E->getArg(0), false); 1521 case Builtin::BI__builtin_operator_delete: 1522 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1523 E->getArg(0), true); 1524 case Builtin::BI__noop: 1525 // __noop always evaluates to an integer literal zero. 1526 return RValue::get(ConstantInt::get(IntTy, 0)); 1527 case Builtin::BI__assume: 1528 // Until LLVM supports assumptions at the IR level, this becomes nothing. 1529 return RValue::get(nullptr); 1530 case Builtin::BI_InterlockedExchange: 1531 case Builtin::BI_InterlockedExchangePointer: 1532 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1533 case Builtin::BI_InterlockedCompareExchangePointer: { 1534 llvm::Type *RTy; 1535 llvm::IntegerType *IntType = 1536 IntegerType::get(getLLVMContext(), 1537 getContext().getTypeSize(E->getType())); 1538 llvm::Type *IntPtrType = IntType->getPointerTo(); 1539 1540 llvm::Value *Destination = 1541 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1542 1543 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1544 RTy = Exchange->getType(); 1545 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1546 1547 llvm::Value *Comparand = 1548 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1549 1550 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1551 SequentiallyConsistent, 1552 SequentiallyConsistent); 1553 Result->setVolatile(true); 1554 1555 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1556 0), 1557 RTy)); 1558 } 1559 case Builtin::BI_InterlockedCompareExchange: { 1560 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1561 EmitScalarExpr(E->getArg(0)), 1562 EmitScalarExpr(E->getArg(2)), 1563 EmitScalarExpr(E->getArg(1)), 1564 SequentiallyConsistent, 1565 SequentiallyConsistent); 1566 CXI->setVolatile(true); 1567 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1568 } 1569 case Builtin::BI_InterlockedIncrement: { 1570 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1571 AtomicRMWInst::Add, 1572 EmitScalarExpr(E->getArg(0)), 1573 ConstantInt::get(Int32Ty, 1), 1574 llvm::SequentiallyConsistent); 1575 RMWI->setVolatile(true); 1576 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1577 } 1578 case Builtin::BI_InterlockedDecrement: { 1579 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1580 AtomicRMWInst::Sub, 1581 EmitScalarExpr(E->getArg(0)), 1582 ConstantInt::get(Int32Ty, 1), 1583 llvm::SequentiallyConsistent); 1584 RMWI->setVolatile(true); 1585 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1586 } 1587 case Builtin::BI_InterlockedExchangeAdd: { 1588 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1589 AtomicRMWInst::Add, 1590 EmitScalarExpr(E->getArg(0)), 1591 EmitScalarExpr(E->getArg(1)), 1592 llvm::SequentiallyConsistent); 1593 RMWI->setVolatile(true); 1594 return RValue::get(RMWI); 1595 } 1596 } 1597 1598 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1599 // the call using the normal call path, but using the unmangled 1600 // version of the function name. 1601 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1602 return emitLibraryCall(*this, FD, E, 1603 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1604 1605 // If this is a predefined lib function (e.g. malloc), emit the call 1606 // using exactly the normal call path. 1607 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1608 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1609 1610 // See if we have a target specific intrinsic. 1611 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1612 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1613 if (const char *Prefix = 1614 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1615 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1616 // NOTE we dont need to perform a compatibility flag check here since the 1617 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1618 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1619 if (IntrinsicID == Intrinsic::not_intrinsic) 1620 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1621 } 1622 1623 if (IntrinsicID != Intrinsic::not_intrinsic) { 1624 SmallVector<Value*, 16> Args; 1625 1626 // Find out if any arguments are required to be integer constant 1627 // expressions. 1628 unsigned ICEArguments = 0; 1629 ASTContext::GetBuiltinTypeError Error; 1630 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1631 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1632 1633 Function *F = CGM.getIntrinsic(IntrinsicID); 1634 llvm::FunctionType *FTy = F->getFunctionType(); 1635 1636 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1637 Value *ArgValue; 1638 // If this is a normal argument, just emit it as a scalar. 1639 if ((ICEArguments & (1 << i)) == 0) { 1640 ArgValue = EmitScalarExpr(E->getArg(i)); 1641 } else { 1642 // If this is required to be a constant, constant fold it so that we 1643 // know that the generated intrinsic gets a ConstantInt. 1644 llvm::APSInt Result; 1645 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1646 assert(IsConst && "Constant arg isn't actually constant?"); 1647 (void)IsConst; 1648 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1649 } 1650 1651 // If the intrinsic arg type is different from the builtin arg type 1652 // we need to do a bit cast. 1653 llvm::Type *PTy = FTy->getParamType(i); 1654 if (PTy != ArgValue->getType()) { 1655 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1656 "Must be able to losslessly bit cast to param"); 1657 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1658 } 1659 1660 Args.push_back(ArgValue); 1661 } 1662 1663 Value *V = Builder.CreateCall(F, Args); 1664 QualType BuiltinRetType = E->getType(); 1665 1666 llvm::Type *RetTy = VoidTy; 1667 if (!BuiltinRetType->isVoidType()) 1668 RetTy = ConvertType(BuiltinRetType); 1669 1670 if (RetTy != V->getType()) { 1671 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1672 "Must be able to losslessly bit cast result type"); 1673 V = Builder.CreateBitCast(V, RetTy); 1674 } 1675 1676 return RValue::get(V); 1677 } 1678 1679 // See if we have a target specific builtin that needs to be lowered. 1680 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1681 return RValue::get(V); 1682 1683 ErrorUnsupported(E, "builtin function"); 1684 1685 // Unknown builtin, for now just dump it out and return undef. 1686 return GetUndefRValue(E->getType()); 1687 } 1688 1689 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1690 const CallExpr *E) { 1691 switch (getTarget().getTriple().getArch()) { 1692 case llvm::Triple::arm: 1693 case llvm::Triple::armeb: 1694 case llvm::Triple::thumb: 1695 case llvm::Triple::thumbeb: 1696 return EmitARMBuiltinExpr(BuiltinID, E); 1697 case llvm::Triple::aarch64: 1698 case llvm::Triple::aarch64_be: 1699 return EmitAArch64BuiltinExpr(BuiltinID, E); 1700 case llvm::Triple::x86: 1701 case llvm::Triple::x86_64: 1702 return EmitX86BuiltinExpr(BuiltinID, E); 1703 case llvm::Triple::ppc: 1704 case llvm::Triple::ppc64: 1705 case llvm::Triple::ppc64le: 1706 return EmitPPCBuiltinExpr(BuiltinID, E); 1707 case llvm::Triple::r600: 1708 return EmitR600BuiltinExpr(BuiltinID, E); 1709 default: 1710 return nullptr; 1711 } 1712 } 1713 1714 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1715 NeonTypeFlags TypeFlags, 1716 bool V1Ty=false) { 1717 int IsQuad = TypeFlags.isQuad(); 1718 switch (TypeFlags.getEltType()) { 1719 case NeonTypeFlags::Int8: 1720 case NeonTypeFlags::Poly8: 1721 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1722 case NeonTypeFlags::Int16: 1723 case NeonTypeFlags::Poly16: 1724 case NeonTypeFlags::Float16: 1725 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1726 case NeonTypeFlags::Int32: 1727 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1728 case NeonTypeFlags::Int64: 1729 case NeonTypeFlags::Poly64: 1730 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1731 case NeonTypeFlags::Poly128: 1732 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1733 // There is a lot of i128 and f128 API missing. 1734 // so we use v16i8 to represent poly128 and get pattern matched. 1735 return llvm::VectorType::get(CGF->Int8Ty, 16); 1736 case NeonTypeFlags::Float32: 1737 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1738 case NeonTypeFlags::Float64: 1739 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1740 } 1741 llvm_unreachable("Unknown vector element type!"); 1742 } 1743 1744 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1745 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1746 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1747 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1748 } 1749 1750 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1751 const char *name, 1752 unsigned shift, bool rightshift) { 1753 unsigned j = 0; 1754 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1755 ai != ae; ++ai, ++j) 1756 if (shift > 0 && shift == j) 1757 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1758 else 1759 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1760 1761 return Builder.CreateCall(F, Ops, name); 1762 } 1763 1764 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1765 bool neg) { 1766 int SV = cast<ConstantInt>(V)->getSExtValue(); 1767 1768 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1769 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1770 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1771 } 1772 1773 // \brief Right-shift a vector by a constant. 1774 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1775 llvm::Type *Ty, bool usgn, 1776 const char *name) { 1777 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1778 1779 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1780 int EltSize = VTy->getScalarSizeInBits(); 1781 1782 Vec = Builder.CreateBitCast(Vec, Ty); 1783 1784 // lshr/ashr are undefined when the shift amount is equal to the vector 1785 // element size. 1786 if (ShiftAmt == EltSize) { 1787 if (usgn) { 1788 // Right-shifting an unsigned value by its size yields 0. 1789 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1790 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1791 } else { 1792 // Right-shifting a signed value by its size is equivalent 1793 // to a shift of size-1. 1794 --ShiftAmt; 1795 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1796 } 1797 } 1798 1799 Shift = EmitNeonShiftVector(Shift, Ty, false); 1800 if (usgn) 1801 return Builder.CreateLShr(Vec, Shift, name); 1802 else 1803 return Builder.CreateAShr(Vec, Shift, name); 1804 } 1805 1806 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1807 /// alignment of the type referenced by the pointer. Skip over implicit 1808 /// casts. 1809 std::pair<llvm::Value*, unsigned> 1810 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1811 assert(Addr->getType()->isPointerType()); 1812 Addr = Addr->IgnoreParens(); 1813 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1814 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1815 ICE->getSubExpr()->getType()->isPointerType()) { 1816 std::pair<llvm::Value*, unsigned> Ptr = 1817 EmitPointerWithAlignment(ICE->getSubExpr()); 1818 Ptr.first = Builder.CreateBitCast(Ptr.first, 1819 ConvertType(Addr->getType())); 1820 return Ptr; 1821 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1822 LValue LV = EmitLValue(ICE->getSubExpr()); 1823 unsigned Align = LV.getAlignment().getQuantity(); 1824 if (!Align) { 1825 // FIXME: Once LValues are fixed to always set alignment, 1826 // zap this code. 1827 QualType PtTy = ICE->getSubExpr()->getType(); 1828 if (!PtTy->isIncompleteType()) 1829 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1830 else 1831 Align = 1; 1832 } 1833 return std::make_pair(LV.getAddress(), Align); 1834 } 1835 } 1836 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1837 if (UO->getOpcode() == UO_AddrOf) { 1838 LValue LV = EmitLValue(UO->getSubExpr()); 1839 unsigned Align = LV.getAlignment().getQuantity(); 1840 if (!Align) { 1841 // FIXME: Once LValues are fixed to always set alignment, 1842 // zap this code. 1843 QualType PtTy = UO->getSubExpr()->getType(); 1844 if (!PtTy->isIncompleteType()) 1845 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1846 else 1847 Align = 1; 1848 } 1849 return std::make_pair(LV.getAddress(), Align); 1850 } 1851 } 1852 1853 unsigned Align = 1; 1854 QualType PtTy = Addr->getType()->getPointeeType(); 1855 if (!PtTy->isIncompleteType()) 1856 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1857 1858 return std::make_pair(EmitScalarExpr(Addr), Align); 1859 } 1860 1861 enum { 1862 AddRetType = (1 << 0), 1863 Add1ArgType = (1 << 1), 1864 Add2ArgTypes = (1 << 2), 1865 1866 VectorizeRetType = (1 << 3), 1867 VectorizeArgTypes = (1 << 4), 1868 1869 InventFloatType = (1 << 5), 1870 UnsignedAlts = (1 << 6), 1871 1872 Use64BitVectors = (1 << 7), 1873 Use128BitVectors = (1 << 8), 1874 1875 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 1876 VectorRet = AddRetType | VectorizeRetType, 1877 VectorRetGetArgs01 = 1878 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 1879 FpCmpzModifiers = 1880 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 1881 }; 1882 1883 struct NeonIntrinsicInfo { 1884 unsigned BuiltinID; 1885 unsigned LLVMIntrinsic; 1886 unsigned AltLLVMIntrinsic; 1887 const char *NameHint; 1888 unsigned TypeModifier; 1889 1890 bool operator<(unsigned RHSBuiltinID) const { 1891 return BuiltinID < RHSBuiltinID; 1892 } 1893 }; 1894 1895 #define NEONMAP0(NameBase) \ 1896 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 1897 1898 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 1899 { NEON:: BI__builtin_neon_ ## NameBase, \ 1900 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 1901 1902 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 1903 { NEON:: BI__builtin_neon_ ## NameBase, \ 1904 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 1905 #NameBase, TypeModifier } 1906 1907 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 1908 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1909 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1910 NEONMAP1(vabs_v, arm_neon_vabs, 0), 1911 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 1912 NEONMAP0(vaddhn_v), 1913 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 1914 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 1915 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 1916 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 1917 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 1918 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 1919 NEONMAP1(vcage_v, arm_neon_vacge, 0), 1920 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 1921 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 1922 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 1923 NEONMAP1(vcale_v, arm_neon_vacge, 0), 1924 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 1925 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 1926 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 1927 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 1928 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 1929 NEONMAP1(vclz_v, ctlz, Add1ArgType), 1930 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 1931 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 1932 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 1933 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 1934 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 1935 NEONMAP0(vcvt_f32_v), 1936 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1937 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1938 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1939 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1940 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1941 NEONMAP0(vcvt_s32_v), 1942 NEONMAP0(vcvt_s64_v), 1943 NEONMAP0(vcvt_u32_v), 1944 NEONMAP0(vcvt_u64_v), 1945 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 1946 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 1947 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 1948 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 1949 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 1950 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 1951 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 1952 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 1953 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 1954 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 1955 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 1956 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 1957 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 1958 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 1959 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 1960 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 1961 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 1962 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 1963 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 1964 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 1965 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 1966 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 1967 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 1968 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 1969 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 1970 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 1971 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 1972 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 1973 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 1974 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 1975 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 1976 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 1977 NEONMAP0(vcvtq_f32_v), 1978 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1979 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1980 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1981 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1982 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1983 NEONMAP0(vcvtq_s32_v), 1984 NEONMAP0(vcvtq_s64_v), 1985 NEONMAP0(vcvtq_u32_v), 1986 NEONMAP0(vcvtq_u64_v), 1987 NEONMAP0(vext_v), 1988 NEONMAP0(vextq_v), 1989 NEONMAP0(vfma_v), 1990 NEONMAP0(vfmaq_v), 1991 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 1992 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 1993 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 1994 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 1995 NEONMAP0(vld1_dup_v), 1996 NEONMAP1(vld1_v, arm_neon_vld1, 0), 1997 NEONMAP0(vld1q_dup_v), 1998 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 1999 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2000 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2001 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2002 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2003 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2004 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2005 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2006 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2007 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2008 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2009 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2010 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2011 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2012 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2013 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2014 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2015 NEONMAP0(vmovl_v), 2016 NEONMAP0(vmovn_v), 2017 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2018 NEONMAP0(vmull_v), 2019 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2020 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2021 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2022 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2023 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2024 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2025 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2026 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2027 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2028 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2029 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2030 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2031 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2032 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2033 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2034 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2035 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2036 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2037 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2038 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2039 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2040 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2041 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2042 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2043 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2044 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2045 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2046 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2047 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2048 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2049 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2050 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2051 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2052 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2053 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2054 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2055 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2056 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2057 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2058 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2059 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2060 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2061 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2062 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2063 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2064 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2065 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2066 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2067 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2068 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2069 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2070 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2071 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2072 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2073 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2074 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2075 NEONMAP0(vshl_n_v), 2076 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2077 NEONMAP0(vshll_n_v), 2078 NEONMAP0(vshlq_n_v), 2079 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2080 NEONMAP0(vshr_n_v), 2081 NEONMAP0(vshrn_n_v), 2082 NEONMAP0(vshrq_n_v), 2083 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2084 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2085 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2086 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2087 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2088 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2089 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2090 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2091 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2092 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2093 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2094 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2095 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2096 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2097 NEONMAP0(vsubhn_v), 2098 NEONMAP0(vtrn_v), 2099 NEONMAP0(vtrnq_v), 2100 NEONMAP0(vtst_v), 2101 NEONMAP0(vtstq_v), 2102 NEONMAP0(vuzp_v), 2103 NEONMAP0(vuzpq_v), 2104 NEONMAP0(vzip_v), 2105 NEONMAP0(vzipq_v) 2106 }; 2107 2108 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2109 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2110 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2111 NEONMAP0(vaddhn_v), 2112 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2113 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2114 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2115 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2116 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2117 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2118 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2119 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2120 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2121 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2122 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2123 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2124 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2125 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2126 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2127 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2128 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2129 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2130 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2131 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2132 NEONMAP0(vcvt_f32_v), 2133 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2134 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2135 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2136 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2137 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2138 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2139 NEONMAP0(vcvtq_f32_v), 2140 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2141 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2142 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2143 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2144 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2145 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2146 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2147 NEONMAP0(vext_v), 2148 NEONMAP0(vextq_v), 2149 NEONMAP0(vfma_v), 2150 NEONMAP0(vfmaq_v), 2151 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2152 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2153 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2154 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2155 NEONMAP0(vmovl_v), 2156 NEONMAP0(vmovn_v), 2157 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2158 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2159 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2160 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2161 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2162 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2163 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2164 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2165 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2166 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2167 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2168 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2169 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2170 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2171 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2172 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2173 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2174 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2175 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2176 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2177 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2178 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2179 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2180 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2181 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2182 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2183 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2184 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2185 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2186 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2187 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2188 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2189 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2190 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2191 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2192 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2193 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2194 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2195 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2196 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2197 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2198 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2199 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2200 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2201 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2202 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2203 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2204 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2205 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2206 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2207 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2208 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2209 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2210 NEONMAP0(vshl_n_v), 2211 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2212 NEONMAP0(vshll_n_v), 2213 NEONMAP0(vshlq_n_v), 2214 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2215 NEONMAP0(vshr_n_v), 2216 NEONMAP0(vshrn_n_v), 2217 NEONMAP0(vshrq_n_v), 2218 NEONMAP0(vsubhn_v), 2219 NEONMAP0(vtst_v), 2220 NEONMAP0(vtstq_v), 2221 }; 2222 2223 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2224 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2225 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2226 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2227 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2228 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2229 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2230 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2231 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2232 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2233 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2234 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2235 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2236 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2237 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2238 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2239 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2240 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2241 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2242 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2243 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2244 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2245 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2246 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2247 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2248 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2249 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2250 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2251 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2252 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2253 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2254 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2255 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2256 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2257 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2258 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2259 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2260 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2261 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2262 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2263 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2264 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2265 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2266 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2267 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2268 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2269 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2270 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2271 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2272 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2273 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2274 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2275 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2276 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2277 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2278 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2279 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2280 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2281 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2282 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2283 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2284 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2285 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2286 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2287 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2288 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2289 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2290 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2291 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2292 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2293 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2294 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2295 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2296 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2297 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2298 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2299 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2300 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2301 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2302 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2303 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2304 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2305 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2306 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2307 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2308 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2309 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2310 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2311 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2312 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2313 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2314 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2315 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2316 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2317 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2318 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2319 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2320 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2321 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2322 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2323 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2324 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2325 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2326 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2327 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2328 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2329 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2330 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2331 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2332 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2333 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2334 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2335 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2336 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2337 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2338 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2339 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2340 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2341 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2342 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2343 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2344 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2345 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2346 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2347 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2348 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2349 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2350 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2351 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2352 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2353 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2354 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2355 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2356 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2357 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2358 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2359 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2360 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2361 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2362 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2363 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2364 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2365 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2366 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2367 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2368 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2369 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2370 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2371 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2372 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2373 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2374 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2375 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2376 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2377 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2378 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2379 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2380 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2381 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2382 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2383 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2384 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2385 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2386 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2387 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2388 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2389 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2390 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2391 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2392 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2393 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2394 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2395 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2396 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2397 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2398 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2399 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2400 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2401 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2402 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2403 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2404 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2405 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2406 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2407 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2408 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2409 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2410 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2411 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2412 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2413 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2414 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2415 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2416 }; 2417 2418 #undef NEONMAP0 2419 #undef NEONMAP1 2420 #undef NEONMAP2 2421 2422 static bool NEONSIMDIntrinsicsProvenSorted = false; 2423 2424 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2425 static bool AArch64SISDIntrinsicsProvenSorted = false; 2426 2427 2428 static const NeonIntrinsicInfo * 2429 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2430 unsigned BuiltinID, bool &MapProvenSorted) { 2431 2432 #ifndef NDEBUG 2433 if (!MapProvenSorted) { 2434 // FIXME: use std::is_sorted once C++11 is allowed 2435 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2436 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2437 MapProvenSorted = true; 2438 } 2439 #endif 2440 2441 const NeonIntrinsicInfo *Builtin = 2442 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2443 2444 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2445 return Builtin; 2446 2447 return nullptr; 2448 } 2449 2450 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2451 unsigned Modifier, 2452 llvm::Type *ArgType, 2453 const CallExpr *E) { 2454 int VectorSize = 0; 2455 if (Modifier & Use64BitVectors) 2456 VectorSize = 64; 2457 else if (Modifier & Use128BitVectors) 2458 VectorSize = 128; 2459 2460 // Return type. 2461 SmallVector<llvm::Type *, 3> Tys; 2462 if (Modifier & AddRetType) { 2463 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2464 if (Modifier & VectorizeRetType) 2465 Ty = llvm::VectorType::get( 2466 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2467 2468 Tys.push_back(Ty); 2469 } 2470 2471 // Arguments. 2472 if (Modifier & VectorizeArgTypes) { 2473 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2474 ArgType = llvm::VectorType::get(ArgType, Elts); 2475 } 2476 2477 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2478 Tys.push_back(ArgType); 2479 2480 if (Modifier & Add2ArgTypes) 2481 Tys.push_back(ArgType); 2482 2483 if (Modifier & InventFloatType) 2484 Tys.push_back(FloatTy); 2485 2486 return CGM.getIntrinsic(IntrinsicID, Tys); 2487 } 2488 2489 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2490 const NeonIntrinsicInfo &SISDInfo, 2491 SmallVectorImpl<Value *> &Ops, 2492 const CallExpr *E) { 2493 unsigned BuiltinID = SISDInfo.BuiltinID; 2494 unsigned int Int = SISDInfo.LLVMIntrinsic; 2495 unsigned Modifier = SISDInfo.TypeModifier; 2496 const char *s = SISDInfo.NameHint; 2497 2498 switch (BuiltinID) { 2499 case NEON::BI__builtin_neon_vcled_s64: 2500 case NEON::BI__builtin_neon_vcled_u64: 2501 case NEON::BI__builtin_neon_vcles_f32: 2502 case NEON::BI__builtin_neon_vcled_f64: 2503 case NEON::BI__builtin_neon_vcltd_s64: 2504 case NEON::BI__builtin_neon_vcltd_u64: 2505 case NEON::BI__builtin_neon_vclts_f32: 2506 case NEON::BI__builtin_neon_vcltd_f64: 2507 case NEON::BI__builtin_neon_vcales_f32: 2508 case NEON::BI__builtin_neon_vcaled_f64: 2509 case NEON::BI__builtin_neon_vcalts_f32: 2510 case NEON::BI__builtin_neon_vcaltd_f64: 2511 // Only one direction of comparisons actually exist, cmle is actually a cmge 2512 // with swapped operands. The table gives us the right intrinsic but we 2513 // still need to do the swap. 2514 std::swap(Ops[0], Ops[1]); 2515 break; 2516 } 2517 2518 assert(Int && "Generic code assumes a valid intrinsic"); 2519 2520 // Determine the type(s) of this overloaded AArch64 intrinsic. 2521 const Expr *Arg = E->getArg(0); 2522 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2523 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2524 2525 int j = 0; 2526 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2527 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2528 ai != ae; ++ai, ++j) { 2529 llvm::Type *ArgTy = ai->getType(); 2530 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2531 ArgTy->getPrimitiveSizeInBits()) 2532 continue; 2533 2534 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2535 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2536 // it before inserting. 2537 Ops[j] = 2538 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2539 Ops[j] = 2540 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2541 } 2542 2543 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2544 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2545 if (ResultType->getPrimitiveSizeInBits() < 2546 Result->getType()->getPrimitiveSizeInBits()) 2547 return CGF.Builder.CreateExtractElement(Result, C0); 2548 2549 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2550 } 2551 2552 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2553 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2554 const char *NameHint, unsigned Modifier, const CallExpr *E, 2555 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2556 // Get the last argument, which specifies the vector type. 2557 llvm::APSInt NeonTypeConst; 2558 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2559 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2560 return nullptr; 2561 2562 // Determine the type of this overloaded NEON intrinsic. 2563 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2564 bool Usgn = Type.isUnsigned(); 2565 bool Quad = Type.isQuad(); 2566 2567 llvm::VectorType *VTy = GetNeonType(this, Type); 2568 llvm::Type *Ty = VTy; 2569 if (!Ty) 2570 return nullptr; 2571 2572 unsigned Int = LLVMIntrinsic; 2573 if ((Modifier & UnsignedAlts) && !Usgn) 2574 Int = AltLLVMIntrinsic; 2575 2576 switch (BuiltinID) { 2577 default: break; 2578 case NEON::BI__builtin_neon_vabs_v: 2579 case NEON::BI__builtin_neon_vabsq_v: 2580 if (VTy->getElementType()->isFloatingPointTy()) 2581 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2582 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2583 case NEON::BI__builtin_neon_vaddhn_v: { 2584 llvm::VectorType *SrcTy = 2585 llvm::VectorType::getExtendedElementVectorType(VTy); 2586 2587 // %sum = add <4 x i32> %lhs, %rhs 2588 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2589 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2590 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2591 2592 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2593 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2594 SrcTy->getScalarSizeInBits() / 2); 2595 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2596 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2597 2598 // %res = trunc <4 x i32> %high to <4 x i16> 2599 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2600 } 2601 case NEON::BI__builtin_neon_vcale_v: 2602 case NEON::BI__builtin_neon_vcaleq_v: 2603 case NEON::BI__builtin_neon_vcalt_v: 2604 case NEON::BI__builtin_neon_vcaltq_v: 2605 std::swap(Ops[0], Ops[1]); 2606 case NEON::BI__builtin_neon_vcage_v: 2607 case NEON::BI__builtin_neon_vcageq_v: 2608 case NEON::BI__builtin_neon_vcagt_v: 2609 case NEON::BI__builtin_neon_vcagtq_v: { 2610 llvm::Type *VecFlt = llvm::VectorType::get( 2611 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2612 VTy->getNumElements()); 2613 llvm::Type *Tys[] = { VTy, VecFlt }; 2614 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2615 return EmitNeonCall(F, Ops, NameHint); 2616 } 2617 case NEON::BI__builtin_neon_vclz_v: 2618 case NEON::BI__builtin_neon_vclzq_v: 2619 // We generate target-independent intrinsic, which needs a second argument 2620 // for whether or not clz of zero is undefined; on ARM it isn't. 2621 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2622 break; 2623 case NEON::BI__builtin_neon_vcvt_f32_v: 2624 case NEON::BI__builtin_neon_vcvtq_f32_v: 2625 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2626 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2627 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2628 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2629 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2630 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2631 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2632 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2633 bool Double = 2634 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2635 llvm::Type *FloatTy = 2636 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2637 : NeonTypeFlags::Float32, 2638 false, Quad)); 2639 llvm::Type *Tys[2] = { FloatTy, Ty }; 2640 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2641 Function *F = CGM.getIntrinsic(Int, Tys); 2642 return EmitNeonCall(F, Ops, "vcvt_n"); 2643 } 2644 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2645 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2646 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2647 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2648 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2649 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2650 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2651 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2652 bool Double = 2653 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2654 llvm::Type *FloatTy = 2655 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2656 : NeonTypeFlags::Float32, 2657 false, Quad)); 2658 llvm::Type *Tys[2] = { Ty, FloatTy }; 2659 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2660 return EmitNeonCall(F, Ops, "vcvt_n"); 2661 } 2662 case NEON::BI__builtin_neon_vcvt_s32_v: 2663 case NEON::BI__builtin_neon_vcvt_u32_v: 2664 case NEON::BI__builtin_neon_vcvt_s64_v: 2665 case NEON::BI__builtin_neon_vcvt_u64_v: 2666 case NEON::BI__builtin_neon_vcvtq_s32_v: 2667 case NEON::BI__builtin_neon_vcvtq_u32_v: 2668 case NEON::BI__builtin_neon_vcvtq_s64_v: 2669 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2670 bool Double = 2671 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2672 llvm::Type *FloatTy = 2673 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2674 : NeonTypeFlags::Float32, 2675 false, Quad)); 2676 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2677 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2678 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2679 } 2680 case NEON::BI__builtin_neon_vcvta_s32_v: 2681 case NEON::BI__builtin_neon_vcvta_s64_v: 2682 case NEON::BI__builtin_neon_vcvta_u32_v: 2683 case NEON::BI__builtin_neon_vcvta_u64_v: 2684 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2685 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2686 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2687 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2688 case NEON::BI__builtin_neon_vcvtn_s32_v: 2689 case NEON::BI__builtin_neon_vcvtn_s64_v: 2690 case NEON::BI__builtin_neon_vcvtn_u32_v: 2691 case NEON::BI__builtin_neon_vcvtn_u64_v: 2692 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2693 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2694 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2695 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2696 case NEON::BI__builtin_neon_vcvtp_s32_v: 2697 case NEON::BI__builtin_neon_vcvtp_s64_v: 2698 case NEON::BI__builtin_neon_vcvtp_u32_v: 2699 case NEON::BI__builtin_neon_vcvtp_u64_v: 2700 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2701 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2702 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2703 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2704 case NEON::BI__builtin_neon_vcvtm_s32_v: 2705 case NEON::BI__builtin_neon_vcvtm_s64_v: 2706 case NEON::BI__builtin_neon_vcvtm_u32_v: 2707 case NEON::BI__builtin_neon_vcvtm_u64_v: 2708 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2709 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2710 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2711 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2712 bool Double = 2713 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2714 llvm::Type *InTy = 2715 GetNeonType(this, 2716 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2717 : NeonTypeFlags::Float32, false, Quad)); 2718 llvm::Type *Tys[2] = { Ty, InTy }; 2719 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2720 } 2721 case NEON::BI__builtin_neon_vext_v: 2722 case NEON::BI__builtin_neon_vextq_v: { 2723 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2724 SmallVector<Constant*, 16> Indices; 2725 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2726 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2727 2728 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2729 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2730 Value *SV = llvm::ConstantVector::get(Indices); 2731 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2732 } 2733 case NEON::BI__builtin_neon_vfma_v: 2734 case NEON::BI__builtin_neon_vfmaq_v: { 2735 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2736 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2737 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2738 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2739 2740 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2741 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2742 } 2743 case NEON::BI__builtin_neon_vld1_v: 2744 case NEON::BI__builtin_neon_vld1q_v: 2745 Ops.push_back(Align); 2746 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2747 case NEON::BI__builtin_neon_vld2_v: 2748 case NEON::BI__builtin_neon_vld2q_v: 2749 case NEON::BI__builtin_neon_vld3_v: 2750 case NEON::BI__builtin_neon_vld3q_v: 2751 case NEON::BI__builtin_neon_vld4_v: 2752 case NEON::BI__builtin_neon_vld4q_v: { 2753 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2754 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2755 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2756 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2757 return Builder.CreateStore(Ops[1], Ops[0]); 2758 } 2759 case NEON::BI__builtin_neon_vld1_dup_v: 2760 case NEON::BI__builtin_neon_vld1q_dup_v: { 2761 Value *V = UndefValue::get(Ty); 2762 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2763 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2764 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2765 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2766 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2767 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2768 return EmitNeonSplat(Ops[0], CI); 2769 } 2770 case NEON::BI__builtin_neon_vld2_lane_v: 2771 case NEON::BI__builtin_neon_vld2q_lane_v: 2772 case NEON::BI__builtin_neon_vld3_lane_v: 2773 case NEON::BI__builtin_neon_vld3q_lane_v: 2774 case NEON::BI__builtin_neon_vld4_lane_v: 2775 case NEON::BI__builtin_neon_vld4q_lane_v: { 2776 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2777 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2778 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2779 Ops.push_back(Align); 2780 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2781 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2782 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2783 return Builder.CreateStore(Ops[1], Ops[0]); 2784 } 2785 case NEON::BI__builtin_neon_vmovl_v: { 2786 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2787 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2788 if (Usgn) 2789 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2790 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2791 } 2792 case NEON::BI__builtin_neon_vmovn_v: { 2793 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2794 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2795 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2796 } 2797 case NEON::BI__builtin_neon_vmull_v: 2798 // FIXME: the integer vmull operations could be emitted in terms of pure 2799 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2800 // hoisting the exts outside loops. Until global ISel comes along that can 2801 // see through such movement this leads to bad CodeGen. So we need an 2802 // intrinsic for now. 2803 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2804 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2805 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2806 case NEON::BI__builtin_neon_vpadal_v: 2807 case NEON::BI__builtin_neon_vpadalq_v: { 2808 // The source operand type has twice as many elements of half the size. 2809 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2810 llvm::Type *EltTy = 2811 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2812 llvm::Type *NarrowTy = 2813 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2814 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2815 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2816 } 2817 case NEON::BI__builtin_neon_vpaddl_v: 2818 case NEON::BI__builtin_neon_vpaddlq_v: { 2819 // The source operand type has twice as many elements of half the size. 2820 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2821 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2822 llvm::Type *NarrowTy = 2823 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2824 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2825 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2826 } 2827 case NEON::BI__builtin_neon_vqdmlal_v: 2828 case NEON::BI__builtin_neon_vqdmlsl_v: { 2829 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2830 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2831 MulOps, "vqdmlal"); 2832 2833 SmallVector<Value *, 2> AccumOps; 2834 AccumOps.push_back(Ops[0]); 2835 AccumOps.push_back(Mul); 2836 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2837 AccumOps, NameHint); 2838 } 2839 case NEON::BI__builtin_neon_vqshl_n_v: 2840 case NEON::BI__builtin_neon_vqshlq_n_v: 2841 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2842 1, false); 2843 case NEON::BI__builtin_neon_vqshlu_n_v: 2844 case NEON::BI__builtin_neon_vqshluq_n_v: 2845 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 2846 1, false); 2847 case NEON::BI__builtin_neon_vrecpe_v: 2848 case NEON::BI__builtin_neon_vrecpeq_v: 2849 case NEON::BI__builtin_neon_vrsqrte_v: 2850 case NEON::BI__builtin_neon_vrsqrteq_v: 2851 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2852 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2853 2854 case NEON::BI__builtin_neon_vrshr_n_v: 2855 case NEON::BI__builtin_neon_vrshrq_n_v: 2856 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 2857 1, true); 2858 case NEON::BI__builtin_neon_vshl_n_v: 2859 case NEON::BI__builtin_neon_vshlq_n_v: 2860 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2861 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2862 "vshl_n"); 2863 case NEON::BI__builtin_neon_vshll_n_v: { 2864 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2865 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2866 if (Usgn) 2867 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2868 else 2869 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2870 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2871 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2872 } 2873 case NEON::BI__builtin_neon_vshrn_n_v: { 2874 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2875 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2876 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2877 if (Usgn) 2878 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2879 else 2880 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2881 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2882 } 2883 case NEON::BI__builtin_neon_vshr_n_v: 2884 case NEON::BI__builtin_neon_vshrq_n_v: 2885 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2886 case NEON::BI__builtin_neon_vst1_v: 2887 case NEON::BI__builtin_neon_vst1q_v: 2888 case NEON::BI__builtin_neon_vst2_v: 2889 case NEON::BI__builtin_neon_vst2q_v: 2890 case NEON::BI__builtin_neon_vst3_v: 2891 case NEON::BI__builtin_neon_vst3q_v: 2892 case NEON::BI__builtin_neon_vst4_v: 2893 case NEON::BI__builtin_neon_vst4q_v: 2894 case NEON::BI__builtin_neon_vst2_lane_v: 2895 case NEON::BI__builtin_neon_vst2q_lane_v: 2896 case NEON::BI__builtin_neon_vst3_lane_v: 2897 case NEON::BI__builtin_neon_vst3q_lane_v: 2898 case NEON::BI__builtin_neon_vst4_lane_v: 2899 case NEON::BI__builtin_neon_vst4q_lane_v: 2900 Ops.push_back(Align); 2901 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 2902 case NEON::BI__builtin_neon_vsubhn_v: { 2903 llvm::VectorType *SrcTy = 2904 llvm::VectorType::getExtendedElementVectorType(VTy); 2905 2906 // %sum = add <4 x i32> %lhs, %rhs 2907 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2908 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2909 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2910 2911 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2912 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2913 SrcTy->getScalarSizeInBits() / 2); 2914 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2915 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2916 2917 // %res = trunc <4 x i32> %high to <4 x i16> 2918 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2919 } 2920 case NEON::BI__builtin_neon_vtrn_v: 2921 case NEON::BI__builtin_neon_vtrnq_v: { 2922 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2923 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2924 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2925 Value *SV = nullptr; 2926 2927 for (unsigned vi = 0; vi != 2; ++vi) { 2928 SmallVector<Constant*, 16> Indices; 2929 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2930 Indices.push_back(Builder.getInt32(i+vi)); 2931 Indices.push_back(Builder.getInt32(i+e+vi)); 2932 } 2933 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2934 SV = llvm::ConstantVector::get(Indices); 2935 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2936 SV = Builder.CreateStore(SV, Addr); 2937 } 2938 return SV; 2939 } 2940 case NEON::BI__builtin_neon_vtst_v: 2941 case NEON::BI__builtin_neon_vtstq_v: { 2942 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2943 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2944 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2945 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2946 ConstantAggregateZero::get(Ty)); 2947 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2948 } 2949 case NEON::BI__builtin_neon_vuzp_v: 2950 case NEON::BI__builtin_neon_vuzpq_v: { 2951 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2952 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2953 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2954 Value *SV = nullptr; 2955 2956 for (unsigned vi = 0; vi != 2; ++vi) { 2957 SmallVector<Constant*, 16> Indices; 2958 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2959 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2960 2961 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2962 SV = llvm::ConstantVector::get(Indices); 2963 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2964 SV = Builder.CreateStore(SV, Addr); 2965 } 2966 return SV; 2967 } 2968 case NEON::BI__builtin_neon_vzip_v: 2969 case NEON::BI__builtin_neon_vzipq_v: { 2970 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2971 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2972 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2973 Value *SV = nullptr; 2974 2975 for (unsigned vi = 0; vi != 2; ++vi) { 2976 SmallVector<Constant*, 16> Indices; 2977 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2978 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2979 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2980 } 2981 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2982 SV = llvm::ConstantVector::get(Indices); 2983 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2984 SV = Builder.CreateStore(SV, Addr); 2985 } 2986 return SV; 2987 } 2988 } 2989 2990 assert(Int && "Expected valid intrinsic number"); 2991 2992 // Determine the type(s) of this overloaded AArch64 intrinsic. 2993 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 2994 2995 Value *Result = EmitNeonCall(F, Ops, NameHint); 2996 llvm::Type *ResultType = ConvertType(E->getType()); 2997 // AArch64 intrinsic one-element vector type cast to 2998 // scalar type expected by the builtin 2999 return Builder.CreateBitCast(Result, ResultType, NameHint); 3000 } 3001 3002 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3003 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3004 const CmpInst::Predicate Ip, const Twine &Name) { 3005 llvm::Type *OTy = Op->getType(); 3006 3007 // FIXME: this is utterly horrific. We should not be looking at previous 3008 // codegen context to find out what needs doing. Unfortunately TableGen 3009 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3010 // (etc). 3011 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3012 OTy = BI->getOperand(0)->getType(); 3013 3014 Op = Builder.CreateBitCast(Op, OTy); 3015 if (OTy->getScalarType()->isFloatingPointTy()) { 3016 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3017 } else { 3018 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3019 } 3020 return Builder.CreateSExt(Op, Ty, Name); 3021 } 3022 3023 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3024 Value *ExtOp, Value *IndexOp, 3025 llvm::Type *ResTy, unsigned IntID, 3026 const char *Name) { 3027 SmallVector<Value *, 2> TblOps; 3028 if (ExtOp) 3029 TblOps.push_back(ExtOp); 3030 3031 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3032 SmallVector<Constant*, 16> Indices; 3033 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3034 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3035 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3036 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3037 } 3038 Value *SV = llvm::ConstantVector::get(Indices); 3039 3040 int PairPos = 0, End = Ops.size() - 1; 3041 while (PairPos < End) { 3042 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3043 Ops[PairPos+1], SV, Name)); 3044 PairPos += 2; 3045 } 3046 3047 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3048 // of the 128-bit lookup table with zero. 3049 if (PairPos == End) { 3050 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3051 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3052 ZeroTbl, SV, Name)); 3053 } 3054 3055 Function *TblF; 3056 TblOps.push_back(IndexOp); 3057 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3058 3059 return CGF.EmitNeonCall(TblF, TblOps, Name); 3060 } 3061 3062 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3063 const CallExpr *E) { 3064 unsigned HintID = static_cast<unsigned>(-1); 3065 switch (BuiltinID) { 3066 default: break; 3067 case ARM::BI__builtin_arm_nop: 3068 HintID = 0; 3069 break; 3070 case ARM::BI__builtin_arm_yield: 3071 case ARM::BI__yield: 3072 HintID = 1; 3073 break; 3074 case ARM::BI__builtin_arm_wfe: 3075 case ARM::BI__wfe: 3076 HintID = 2; 3077 break; 3078 case ARM::BI__builtin_arm_wfi: 3079 case ARM::BI__wfi: 3080 HintID = 3; 3081 break; 3082 case ARM::BI__builtin_arm_sev: 3083 case ARM::BI__sev: 3084 HintID = 4; 3085 break; 3086 case ARM::BI__builtin_arm_sevl: 3087 case ARM::BI__sevl: 3088 HintID = 5; 3089 break; 3090 } 3091 3092 if (HintID != static_cast<unsigned>(-1)) { 3093 Function *F = CGM.getIntrinsic(Intrinsic::arm_hint); 3094 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3095 } 3096 3097 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3098 Value *Address = EmitScalarExpr(E->getArg(0)); 3099 Value *RW = EmitScalarExpr(E->getArg(1)); 3100 Value *IsData = EmitScalarExpr(E->getArg(2)); 3101 3102 // Locality is not supported on ARM target 3103 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3104 3105 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3106 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3107 } 3108 3109 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3110 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3111 EmitScalarExpr(E->getArg(0)), 3112 "rbit"); 3113 } 3114 3115 if (BuiltinID == ARM::BI__clear_cache) { 3116 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3117 const FunctionDecl *FD = E->getDirectCallee(); 3118 SmallVector<Value*, 2> Ops; 3119 for (unsigned i = 0; i < 2; i++) 3120 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3121 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3122 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3123 StringRef Name = FD->getName(); 3124 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3125 } 3126 3127 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3128 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3129 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3130 getContext().getTypeSize(E->getType()) == 64) || 3131 BuiltinID == ARM::BI__ldrexd) { 3132 Function *F; 3133 3134 switch (BuiltinID) { 3135 default: llvm_unreachable("unexpected builtin"); 3136 case ARM::BI__builtin_arm_ldaex: 3137 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3138 break; 3139 case ARM::BI__builtin_arm_ldrexd: 3140 case ARM::BI__builtin_arm_ldrex: 3141 case ARM::BI__ldrexd: 3142 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3143 break; 3144 } 3145 3146 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3147 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3148 "ldrexd"); 3149 3150 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3151 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3152 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3153 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3154 3155 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3156 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3157 Val = Builder.CreateOr(Val, Val1); 3158 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3159 } 3160 3161 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3162 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3163 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3164 3165 QualType Ty = E->getType(); 3166 llvm::Type *RealResTy = ConvertType(Ty); 3167 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3168 getContext().getTypeSize(Ty)); 3169 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3170 3171 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3172 ? Intrinsic::arm_ldaex 3173 : Intrinsic::arm_ldrex, 3174 LoadAddr->getType()); 3175 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3176 3177 if (RealResTy->isPointerTy()) 3178 return Builder.CreateIntToPtr(Val, RealResTy); 3179 else { 3180 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3181 return Builder.CreateBitCast(Val, RealResTy); 3182 } 3183 } 3184 3185 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3186 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3187 BuiltinID == ARM::BI__builtin_arm_strex) && 3188 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3189 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3190 ? Intrinsic::arm_stlexd 3191 : Intrinsic::arm_strexd); 3192 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 3193 3194 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3195 Value *Val = EmitScalarExpr(E->getArg(0)); 3196 Builder.CreateStore(Val, Tmp); 3197 3198 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3199 Val = Builder.CreateLoad(LdPtr); 3200 3201 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3202 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3203 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3204 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3205 } 3206 3207 if (BuiltinID == ARM::BI__builtin_arm_strex || 3208 BuiltinID == ARM::BI__builtin_arm_stlex) { 3209 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3210 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3211 3212 QualType Ty = E->getArg(0)->getType(); 3213 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3214 getContext().getTypeSize(Ty)); 3215 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3216 3217 if (StoreVal->getType()->isPointerTy()) 3218 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3219 else { 3220 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3221 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3222 } 3223 3224 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3225 ? Intrinsic::arm_stlex 3226 : Intrinsic::arm_strex, 3227 StoreAddr->getType()); 3228 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3229 } 3230 3231 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3232 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3233 return Builder.CreateCall(F); 3234 } 3235 3236 // CRC32 3237 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3238 switch (BuiltinID) { 3239 case ARM::BI__builtin_arm_crc32b: 3240 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3241 case ARM::BI__builtin_arm_crc32cb: 3242 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3243 case ARM::BI__builtin_arm_crc32h: 3244 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3245 case ARM::BI__builtin_arm_crc32ch: 3246 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3247 case ARM::BI__builtin_arm_crc32w: 3248 case ARM::BI__builtin_arm_crc32d: 3249 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3250 case ARM::BI__builtin_arm_crc32cw: 3251 case ARM::BI__builtin_arm_crc32cd: 3252 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3253 } 3254 3255 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3256 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3257 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3258 3259 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3260 // intrinsics, hence we need different codegen for these cases. 3261 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3262 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3263 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3264 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3265 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3266 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3267 3268 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3269 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3270 return Builder.CreateCall2(F, Res, Arg1b); 3271 } else { 3272 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3273 3274 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3275 return Builder.CreateCall2(F, Arg0, Arg1); 3276 } 3277 } 3278 3279 SmallVector<Value*, 4> Ops; 3280 llvm::Value *Align = nullptr; 3281 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3282 if (i == 0) { 3283 switch (BuiltinID) { 3284 case NEON::BI__builtin_neon_vld1_v: 3285 case NEON::BI__builtin_neon_vld1q_v: 3286 case NEON::BI__builtin_neon_vld1q_lane_v: 3287 case NEON::BI__builtin_neon_vld1_lane_v: 3288 case NEON::BI__builtin_neon_vld1_dup_v: 3289 case NEON::BI__builtin_neon_vld1q_dup_v: 3290 case NEON::BI__builtin_neon_vst1_v: 3291 case NEON::BI__builtin_neon_vst1q_v: 3292 case NEON::BI__builtin_neon_vst1q_lane_v: 3293 case NEON::BI__builtin_neon_vst1_lane_v: 3294 case NEON::BI__builtin_neon_vst2_v: 3295 case NEON::BI__builtin_neon_vst2q_v: 3296 case NEON::BI__builtin_neon_vst2_lane_v: 3297 case NEON::BI__builtin_neon_vst2q_lane_v: 3298 case NEON::BI__builtin_neon_vst3_v: 3299 case NEON::BI__builtin_neon_vst3q_v: 3300 case NEON::BI__builtin_neon_vst3_lane_v: 3301 case NEON::BI__builtin_neon_vst3q_lane_v: 3302 case NEON::BI__builtin_neon_vst4_v: 3303 case NEON::BI__builtin_neon_vst4q_v: 3304 case NEON::BI__builtin_neon_vst4_lane_v: 3305 case NEON::BI__builtin_neon_vst4q_lane_v: 3306 // Get the alignment for the argument in addition to the value; 3307 // we'll use it later. 3308 std::pair<llvm::Value*, unsigned> Src = 3309 EmitPointerWithAlignment(E->getArg(0)); 3310 Ops.push_back(Src.first); 3311 Align = Builder.getInt32(Src.second); 3312 continue; 3313 } 3314 } 3315 if (i == 1) { 3316 switch (BuiltinID) { 3317 case NEON::BI__builtin_neon_vld2_v: 3318 case NEON::BI__builtin_neon_vld2q_v: 3319 case NEON::BI__builtin_neon_vld3_v: 3320 case NEON::BI__builtin_neon_vld3q_v: 3321 case NEON::BI__builtin_neon_vld4_v: 3322 case NEON::BI__builtin_neon_vld4q_v: 3323 case NEON::BI__builtin_neon_vld2_lane_v: 3324 case NEON::BI__builtin_neon_vld2q_lane_v: 3325 case NEON::BI__builtin_neon_vld3_lane_v: 3326 case NEON::BI__builtin_neon_vld3q_lane_v: 3327 case NEON::BI__builtin_neon_vld4_lane_v: 3328 case NEON::BI__builtin_neon_vld4q_lane_v: 3329 case NEON::BI__builtin_neon_vld2_dup_v: 3330 case NEON::BI__builtin_neon_vld3_dup_v: 3331 case NEON::BI__builtin_neon_vld4_dup_v: 3332 // Get the alignment for the argument in addition to the value; 3333 // we'll use it later. 3334 std::pair<llvm::Value*, unsigned> Src = 3335 EmitPointerWithAlignment(E->getArg(1)); 3336 Ops.push_back(Src.first); 3337 Align = Builder.getInt32(Src.second); 3338 continue; 3339 } 3340 } 3341 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3342 } 3343 3344 switch (BuiltinID) { 3345 default: break; 3346 // vget_lane and vset_lane are not overloaded and do not have an extra 3347 // argument that specifies the vector type. 3348 case NEON::BI__builtin_neon_vget_lane_i8: 3349 case NEON::BI__builtin_neon_vget_lane_i16: 3350 case NEON::BI__builtin_neon_vget_lane_i32: 3351 case NEON::BI__builtin_neon_vget_lane_i64: 3352 case NEON::BI__builtin_neon_vget_lane_f32: 3353 case NEON::BI__builtin_neon_vgetq_lane_i8: 3354 case NEON::BI__builtin_neon_vgetq_lane_i16: 3355 case NEON::BI__builtin_neon_vgetq_lane_i32: 3356 case NEON::BI__builtin_neon_vgetq_lane_i64: 3357 case NEON::BI__builtin_neon_vgetq_lane_f32: 3358 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3359 "vget_lane"); 3360 case NEON::BI__builtin_neon_vset_lane_i8: 3361 case NEON::BI__builtin_neon_vset_lane_i16: 3362 case NEON::BI__builtin_neon_vset_lane_i32: 3363 case NEON::BI__builtin_neon_vset_lane_i64: 3364 case NEON::BI__builtin_neon_vset_lane_f32: 3365 case NEON::BI__builtin_neon_vsetq_lane_i8: 3366 case NEON::BI__builtin_neon_vsetq_lane_i16: 3367 case NEON::BI__builtin_neon_vsetq_lane_i32: 3368 case NEON::BI__builtin_neon_vsetq_lane_i64: 3369 case NEON::BI__builtin_neon_vsetq_lane_f32: 3370 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3371 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3372 3373 // Non-polymorphic crypto instructions also not overloaded 3374 case NEON::BI__builtin_neon_vsha1h_u32: 3375 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3376 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3377 "vsha1h"); 3378 case NEON::BI__builtin_neon_vsha1cq_u32: 3379 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3380 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3381 "vsha1h"); 3382 case NEON::BI__builtin_neon_vsha1pq_u32: 3383 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3384 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3385 "vsha1h"); 3386 case NEON::BI__builtin_neon_vsha1mq_u32: 3387 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3388 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3389 "vsha1h"); 3390 } 3391 3392 // Get the last argument, which specifies the vector type. 3393 llvm::APSInt Result; 3394 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3395 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3396 return nullptr; 3397 3398 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3399 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3400 // Determine the overloaded type of this builtin. 3401 llvm::Type *Ty; 3402 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3403 Ty = FloatTy; 3404 else 3405 Ty = DoubleTy; 3406 3407 // Determine whether this is an unsigned conversion or not. 3408 bool usgn = Result.getZExtValue() == 1; 3409 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3410 3411 // Call the appropriate intrinsic. 3412 Function *F = CGM.getIntrinsic(Int, Ty); 3413 return Builder.CreateCall(F, Ops, "vcvtr"); 3414 } 3415 3416 // Determine the type of this overloaded NEON intrinsic. 3417 NeonTypeFlags Type(Result.getZExtValue()); 3418 bool usgn = Type.isUnsigned(); 3419 bool rightShift = false; 3420 3421 llvm::VectorType *VTy = GetNeonType(this, Type); 3422 llvm::Type *Ty = VTy; 3423 if (!Ty) 3424 return nullptr; 3425 3426 // Many NEON builtins have identical semantics and uses in ARM and 3427 // AArch64. Emit these in a single function. 3428 ArrayRef<NeonIntrinsicInfo> IntrinsicMap(ARMSIMDIntrinsicMap); 3429 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3430 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3431 if (Builtin) 3432 return EmitCommonNeonBuiltinExpr( 3433 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3434 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3435 3436 unsigned Int; 3437 switch (BuiltinID) { 3438 default: return nullptr; 3439 case NEON::BI__builtin_neon_vld1q_lane_v: 3440 // Handle 64-bit integer elements as a special case. Use shuffles of 3441 // one-element vectors to avoid poor code for i64 in the backend. 3442 if (VTy->getElementType()->isIntegerTy(64)) { 3443 // Extract the other lane. 3444 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3445 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3446 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3447 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3448 // Load the value as a one-element vector. 3449 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3450 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3451 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3452 // Combine them. 3453 SmallVector<Constant*, 2> Indices; 3454 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3455 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3456 SV = llvm::ConstantVector::get(Indices); 3457 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3458 } 3459 // fall through 3460 case NEON::BI__builtin_neon_vld1_lane_v: { 3461 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3462 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3463 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3464 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3465 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3466 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3467 } 3468 case NEON::BI__builtin_neon_vld2_dup_v: 3469 case NEON::BI__builtin_neon_vld3_dup_v: 3470 case NEON::BI__builtin_neon_vld4_dup_v: { 3471 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3472 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3473 switch (BuiltinID) { 3474 case NEON::BI__builtin_neon_vld2_dup_v: 3475 Int = Intrinsic::arm_neon_vld2; 3476 break; 3477 case NEON::BI__builtin_neon_vld3_dup_v: 3478 Int = Intrinsic::arm_neon_vld3; 3479 break; 3480 case NEON::BI__builtin_neon_vld4_dup_v: 3481 Int = Intrinsic::arm_neon_vld4; 3482 break; 3483 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3484 } 3485 Function *F = CGM.getIntrinsic(Int, Ty); 3486 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3487 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3488 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3489 return Builder.CreateStore(Ops[1], Ops[0]); 3490 } 3491 switch (BuiltinID) { 3492 case NEON::BI__builtin_neon_vld2_dup_v: 3493 Int = Intrinsic::arm_neon_vld2lane; 3494 break; 3495 case NEON::BI__builtin_neon_vld3_dup_v: 3496 Int = Intrinsic::arm_neon_vld3lane; 3497 break; 3498 case NEON::BI__builtin_neon_vld4_dup_v: 3499 Int = Intrinsic::arm_neon_vld4lane; 3500 break; 3501 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3502 } 3503 Function *F = CGM.getIntrinsic(Int, Ty); 3504 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3505 3506 SmallVector<Value*, 6> Args; 3507 Args.push_back(Ops[1]); 3508 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3509 3510 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3511 Args.push_back(CI); 3512 Args.push_back(Align); 3513 3514 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3515 // splat lane 0 to all elts in each vector of the result. 3516 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3517 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3518 Value *Elt = Builder.CreateBitCast(Val, Ty); 3519 Elt = EmitNeonSplat(Elt, CI); 3520 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3521 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3522 } 3523 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3524 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3525 return Builder.CreateStore(Ops[1], Ops[0]); 3526 } 3527 case NEON::BI__builtin_neon_vqrshrn_n_v: 3528 Int = 3529 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3530 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3531 1, true); 3532 case NEON::BI__builtin_neon_vqrshrun_n_v: 3533 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3534 Ops, "vqrshrun_n", 1, true); 3535 case NEON::BI__builtin_neon_vqshrn_n_v: 3536 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3537 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3538 1, true); 3539 case NEON::BI__builtin_neon_vqshrun_n_v: 3540 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3541 Ops, "vqshrun_n", 1, true); 3542 case NEON::BI__builtin_neon_vrecpe_v: 3543 case NEON::BI__builtin_neon_vrecpeq_v: 3544 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3545 Ops, "vrecpe"); 3546 case NEON::BI__builtin_neon_vrshrn_n_v: 3547 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3548 Ops, "vrshrn_n", 1, true); 3549 case NEON::BI__builtin_neon_vrsra_n_v: 3550 case NEON::BI__builtin_neon_vrsraq_n_v: 3551 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3552 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3553 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3554 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3555 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3556 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3557 case NEON::BI__builtin_neon_vsri_n_v: 3558 case NEON::BI__builtin_neon_vsriq_n_v: 3559 rightShift = true; 3560 case NEON::BI__builtin_neon_vsli_n_v: 3561 case NEON::BI__builtin_neon_vsliq_n_v: 3562 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3563 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3564 Ops, "vsli_n"); 3565 case NEON::BI__builtin_neon_vsra_n_v: 3566 case NEON::BI__builtin_neon_vsraq_n_v: 3567 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3568 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3569 return Builder.CreateAdd(Ops[0], Ops[1]); 3570 case NEON::BI__builtin_neon_vst1q_lane_v: 3571 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3572 // a one-element vector and avoid poor code for i64 in the backend. 3573 if (VTy->getElementType()->isIntegerTy(64)) { 3574 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3575 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3576 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3577 Ops[2] = Align; 3578 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3579 Ops[1]->getType()), Ops); 3580 } 3581 // fall through 3582 case NEON::BI__builtin_neon_vst1_lane_v: { 3583 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3584 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3585 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3586 StoreInst *St = Builder.CreateStore(Ops[1], 3587 Builder.CreateBitCast(Ops[0], Ty)); 3588 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3589 return St; 3590 } 3591 case NEON::BI__builtin_neon_vtbl1_v: 3592 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3593 Ops, "vtbl1"); 3594 case NEON::BI__builtin_neon_vtbl2_v: 3595 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3596 Ops, "vtbl2"); 3597 case NEON::BI__builtin_neon_vtbl3_v: 3598 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3599 Ops, "vtbl3"); 3600 case NEON::BI__builtin_neon_vtbl4_v: 3601 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3602 Ops, "vtbl4"); 3603 case NEON::BI__builtin_neon_vtbx1_v: 3604 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3605 Ops, "vtbx1"); 3606 case NEON::BI__builtin_neon_vtbx2_v: 3607 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3608 Ops, "vtbx2"); 3609 case NEON::BI__builtin_neon_vtbx3_v: 3610 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3611 Ops, "vtbx3"); 3612 case NEON::BI__builtin_neon_vtbx4_v: 3613 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3614 Ops, "vtbx4"); 3615 } 3616 } 3617 3618 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3619 const CallExpr *E, 3620 SmallVectorImpl<Value *> &Ops) { 3621 unsigned int Int = 0; 3622 const char *s = nullptr; 3623 3624 switch (BuiltinID) { 3625 default: 3626 return nullptr; 3627 case NEON::BI__builtin_neon_vtbl1_v: 3628 case NEON::BI__builtin_neon_vqtbl1_v: 3629 case NEON::BI__builtin_neon_vqtbl1q_v: 3630 case NEON::BI__builtin_neon_vtbl2_v: 3631 case NEON::BI__builtin_neon_vqtbl2_v: 3632 case NEON::BI__builtin_neon_vqtbl2q_v: 3633 case NEON::BI__builtin_neon_vtbl3_v: 3634 case NEON::BI__builtin_neon_vqtbl3_v: 3635 case NEON::BI__builtin_neon_vqtbl3q_v: 3636 case NEON::BI__builtin_neon_vtbl4_v: 3637 case NEON::BI__builtin_neon_vqtbl4_v: 3638 case NEON::BI__builtin_neon_vqtbl4q_v: 3639 break; 3640 case NEON::BI__builtin_neon_vtbx1_v: 3641 case NEON::BI__builtin_neon_vqtbx1_v: 3642 case NEON::BI__builtin_neon_vqtbx1q_v: 3643 case NEON::BI__builtin_neon_vtbx2_v: 3644 case NEON::BI__builtin_neon_vqtbx2_v: 3645 case NEON::BI__builtin_neon_vqtbx2q_v: 3646 case NEON::BI__builtin_neon_vtbx3_v: 3647 case NEON::BI__builtin_neon_vqtbx3_v: 3648 case NEON::BI__builtin_neon_vqtbx3q_v: 3649 case NEON::BI__builtin_neon_vtbx4_v: 3650 case NEON::BI__builtin_neon_vqtbx4_v: 3651 case NEON::BI__builtin_neon_vqtbx4q_v: 3652 break; 3653 } 3654 3655 assert(E->getNumArgs() >= 3); 3656 3657 // Get the last argument, which specifies the vector type. 3658 llvm::APSInt Result; 3659 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3660 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3661 return nullptr; 3662 3663 // Determine the type of this overloaded NEON intrinsic. 3664 NeonTypeFlags Type(Result.getZExtValue()); 3665 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3666 llvm::Type *Ty = VTy; 3667 if (!Ty) 3668 return nullptr; 3669 3670 unsigned nElts = VTy->getNumElements(); 3671 3672 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3673 3674 // AArch64 scalar builtins are not overloaded, they do not have an extra 3675 // argument that specifies the vector type, need to handle each case. 3676 SmallVector<Value *, 2> TblOps; 3677 switch (BuiltinID) { 3678 case NEON::BI__builtin_neon_vtbl1_v: { 3679 TblOps.push_back(Ops[0]); 3680 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3681 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3682 } 3683 case NEON::BI__builtin_neon_vtbl2_v: { 3684 TblOps.push_back(Ops[0]); 3685 TblOps.push_back(Ops[1]); 3686 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3687 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3688 } 3689 case NEON::BI__builtin_neon_vtbl3_v: { 3690 TblOps.push_back(Ops[0]); 3691 TblOps.push_back(Ops[1]); 3692 TblOps.push_back(Ops[2]); 3693 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3694 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3695 } 3696 case NEON::BI__builtin_neon_vtbl4_v: { 3697 TblOps.push_back(Ops[0]); 3698 TblOps.push_back(Ops[1]); 3699 TblOps.push_back(Ops[2]); 3700 TblOps.push_back(Ops[3]); 3701 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3702 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3703 } 3704 case NEON::BI__builtin_neon_vtbx1_v: { 3705 TblOps.push_back(Ops[1]); 3706 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3707 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3708 3709 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3710 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3711 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3712 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3713 3714 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3715 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3716 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3717 } 3718 case NEON::BI__builtin_neon_vtbx2_v: { 3719 TblOps.push_back(Ops[1]); 3720 TblOps.push_back(Ops[2]); 3721 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3722 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3723 } 3724 case NEON::BI__builtin_neon_vtbx3_v: { 3725 TblOps.push_back(Ops[1]); 3726 TblOps.push_back(Ops[2]); 3727 TblOps.push_back(Ops[3]); 3728 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3729 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3730 3731 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3732 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3733 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3734 TwentyFourV); 3735 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3736 3737 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3738 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3739 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3740 } 3741 case NEON::BI__builtin_neon_vtbx4_v: { 3742 TblOps.push_back(Ops[1]); 3743 TblOps.push_back(Ops[2]); 3744 TblOps.push_back(Ops[3]); 3745 TblOps.push_back(Ops[4]); 3746 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3747 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3748 } 3749 case NEON::BI__builtin_neon_vqtbl1_v: 3750 case NEON::BI__builtin_neon_vqtbl1q_v: 3751 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3752 case NEON::BI__builtin_neon_vqtbl2_v: 3753 case NEON::BI__builtin_neon_vqtbl2q_v: { 3754 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3755 case NEON::BI__builtin_neon_vqtbl3_v: 3756 case NEON::BI__builtin_neon_vqtbl3q_v: 3757 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3758 case NEON::BI__builtin_neon_vqtbl4_v: 3759 case NEON::BI__builtin_neon_vqtbl4q_v: 3760 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3761 case NEON::BI__builtin_neon_vqtbx1_v: 3762 case NEON::BI__builtin_neon_vqtbx1q_v: 3763 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3764 case NEON::BI__builtin_neon_vqtbx2_v: 3765 case NEON::BI__builtin_neon_vqtbx2q_v: 3766 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3767 case NEON::BI__builtin_neon_vqtbx3_v: 3768 case NEON::BI__builtin_neon_vqtbx3q_v: 3769 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3770 case NEON::BI__builtin_neon_vqtbx4_v: 3771 case NEON::BI__builtin_neon_vqtbx4q_v: 3772 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3773 } 3774 } 3775 3776 if (!Int) 3777 return nullptr; 3778 3779 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3780 return CGF.EmitNeonCall(F, Ops, s); 3781 } 3782 3783 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3784 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3785 Op = Builder.CreateBitCast(Op, Int16Ty); 3786 Value *V = UndefValue::get(VTy); 3787 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3788 Op = Builder.CreateInsertElement(V, Op, CI); 3789 return Op; 3790 } 3791 3792 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3793 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3794 Op = Builder.CreateBitCast(Op, Int8Ty); 3795 Value *V = UndefValue::get(VTy); 3796 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3797 Op = Builder.CreateInsertElement(V, Op, CI); 3798 return Op; 3799 } 3800 3801 Value *CodeGenFunction:: 3802 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3803 const char *Name) { 3804 // i8 is not a legal types for AArch64, so we can't just use 3805 // a normal overloaded intrinsic call for these scalar types. Instead 3806 // we'll build 64-bit vectors w/ lane zero being our input values and 3807 // perform the operation on that. The back end can pattern match directly 3808 // to the scalar instruction. 3809 Ops[0] = vectorWrapScalar8(Ops[0]); 3810 Ops[1] = vectorWrapScalar8(Ops[1]); 3811 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3812 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3813 Constant *CI = ConstantInt::get(SizeTy, 0); 3814 return Builder.CreateExtractElement(V, CI, "lane0"); 3815 } 3816 3817 Value *CodeGenFunction:: 3818 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3819 const char *Name) { 3820 // i16 is not a legal types for AArch64, so we can't just use 3821 // a normal overloaded intrinsic call for these scalar types. Instead 3822 // we'll build 64-bit vectors w/ lane zero being our input values and 3823 // perform the operation on that. The back end can pattern match directly 3824 // to the scalar instruction. 3825 Ops[0] = vectorWrapScalar16(Ops[0]); 3826 Ops[1] = vectorWrapScalar16(Ops[1]); 3827 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3828 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3829 Constant *CI = ConstantInt::get(SizeTy, 0); 3830 return Builder.CreateExtractElement(V, CI, "lane0"); 3831 } 3832 3833 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3834 const CallExpr *E) { 3835 unsigned HintID = static_cast<unsigned>(-1); 3836 switch (BuiltinID) { 3837 default: break; 3838 case AArch64::BI__builtin_arm_nop: 3839 HintID = 0; 3840 break; 3841 case AArch64::BI__builtin_arm_yield: 3842 HintID = 1; 3843 break; 3844 case AArch64::BI__builtin_arm_wfe: 3845 HintID = 2; 3846 break; 3847 case AArch64::BI__builtin_arm_wfi: 3848 HintID = 3; 3849 break; 3850 case AArch64::BI__builtin_arm_sev: 3851 HintID = 4; 3852 break; 3853 case AArch64::BI__builtin_arm_sevl: 3854 HintID = 5; 3855 break; 3856 } 3857 3858 if (HintID != static_cast<unsigned>(-1)) { 3859 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 3860 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3861 } 3862 3863 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 3864 Value *Address = EmitScalarExpr(E->getArg(0)); 3865 Value *RW = EmitScalarExpr(E->getArg(1)); 3866 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 3867 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 3868 Value *IsData = EmitScalarExpr(E->getArg(4)); 3869 3870 Value *Locality = nullptr; 3871 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 3872 // Temporal fetch, needs to convert cache level to locality. 3873 Locality = llvm::ConstantInt::get(Int32Ty, 3874 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 3875 } else { 3876 // Streaming fetch. 3877 Locality = llvm::ConstantInt::get(Int32Ty, 0); 3878 } 3879 3880 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 3881 // PLDL3STRM or PLDL2STRM. 3882 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3883 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3884 } 3885 3886 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 3887 assert((getContext().getTypeSize(E->getType()) == 32) && 3888 "rbit of unusual size!"); 3889 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3890 return Builder.CreateCall( 3891 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3892 } 3893 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 3894 assert((getContext().getTypeSize(E->getType()) == 64) && 3895 "rbit of unusual size!"); 3896 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3897 return Builder.CreateCall( 3898 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3899 } 3900 3901 if (BuiltinID == AArch64::BI__clear_cache) { 3902 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3903 const FunctionDecl *FD = E->getDirectCallee(); 3904 SmallVector<Value*, 2> Ops; 3905 for (unsigned i = 0; i < 2; i++) 3906 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3907 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3908 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3909 StringRef Name = FD->getName(); 3910 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3911 } 3912 3913 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 3914 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 3915 getContext().getTypeSize(E->getType()) == 128) { 3916 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 3917 ? Intrinsic::aarch64_ldaxp 3918 : Intrinsic::aarch64_ldxp); 3919 3920 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3921 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3922 "ldxp"); 3923 3924 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3925 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3926 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 3927 Val0 = Builder.CreateZExt(Val0, Int128Ty); 3928 Val1 = Builder.CreateZExt(Val1, Int128Ty); 3929 3930 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 3931 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3932 Val = Builder.CreateOr(Val, Val1); 3933 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3934 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 3935 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 3936 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3937 3938 QualType Ty = E->getType(); 3939 llvm::Type *RealResTy = ConvertType(Ty); 3940 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3941 getContext().getTypeSize(Ty)); 3942 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3943 3944 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 3945 ? Intrinsic::aarch64_ldaxr 3946 : Intrinsic::aarch64_ldxr, 3947 LoadAddr->getType()); 3948 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 3949 3950 if (RealResTy->isPointerTy()) 3951 return Builder.CreateIntToPtr(Val, RealResTy); 3952 3953 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3954 return Builder.CreateBitCast(Val, RealResTy); 3955 } 3956 3957 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 3958 BuiltinID == AArch64::BI__builtin_arm_stlex) && 3959 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 3960 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 3961 ? Intrinsic::aarch64_stlxp 3962 : Intrinsic::aarch64_stxp); 3963 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, NULL); 3964 3965 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 3966 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 3967 One); 3968 Value *Val = EmitScalarExpr(E->getArg(0)); 3969 Builder.CreateStore(Val, Tmp); 3970 3971 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3972 Val = Builder.CreateLoad(LdPtr); 3973 3974 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3975 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3976 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 3977 Int8PtrTy); 3978 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 3979 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 3980 BuiltinID == AArch64::BI__builtin_arm_stlex) { 3981 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3982 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3983 3984 QualType Ty = E->getArg(0)->getType(); 3985 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3986 getContext().getTypeSize(Ty)); 3987 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3988 3989 if (StoreVal->getType()->isPointerTy()) 3990 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 3991 else { 3992 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3993 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 3994 } 3995 3996 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 3997 ? Intrinsic::aarch64_stlxr 3998 : Intrinsic::aarch64_stxr, 3999 StoreAddr->getType()); 4000 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4001 } 4002 4003 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4004 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4005 return Builder.CreateCall(F); 4006 } 4007 4008 // CRC32 4009 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4010 switch (BuiltinID) { 4011 case AArch64::BI__builtin_arm_crc32b: 4012 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4013 case AArch64::BI__builtin_arm_crc32cb: 4014 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4015 case AArch64::BI__builtin_arm_crc32h: 4016 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4017 case AArch64::BI__builtin_arm_crc32ch: 4018 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4019 case AArch64::BI__builtin_arm_crc32w: 4020 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4021 case AArch64::BI__builtin_arm_crc32cw: 4022 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4023 case AArch64::BI__builtin_arm_crc32d: 4024 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4025 case AArch64::BI__builtin_arm_crc32cd: 4026 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4027 } 4028 4029 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4030 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4031 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4032 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4033 4034 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4035 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4036 4037 return Builder.CreateCall2(F, Arg0, Arg1); 4038 } 4039 4040 llvm::SmallVector<Value*, 4> Ops; 4041 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4042 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4043 4044 ArrayRef<NeonIntrinsicInfo> SISDMap(AArch64SISDIntrinsicMap); 4045 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4046 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4047 4048 if (Builtin) { 4049 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4050 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4051 assert(Result && "SISD intrinsic should have been handled"); 4052 return Result; 4053 } 4054 4055 llvm::APSInt Result; 4056 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4057 NeonTypeFlags Type(0); 4058 if (Arg->isIntegerConstantExpr(Result, getContext())) 4059 // Determine the type of this overloaded NEON intrinsic. 4060 Type = NeonTypeFlags(Result.getZExtValue()); 4061 4062 bool usgn = Type.isUnsigned(); 4063 bool quad = Type.isQuad(); 4064 4065 // Handle non-overloaded intrinsics first. 4066 switch (BuiltinID) { 4067 default: break; 4068 case NEON::BI__builtin_neon_vldrq_p128: { 4069 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4070 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4071 return Builder.CreateLoad(Ptr); 4072 } 4073 case NEON::BI__builtin_neon_vstrq_p128: { 4074 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4075 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4076 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4077 } 4078 case NEON::BI__builtin_neon_vcvts_u32_f32: 4079 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4080 usgn = true; 4081 // FALL THROUGH 4082 case NEON::BI__builtin_neon_vcvts_s32_f32: 4083 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4084 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4085 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4086 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4087 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4088 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4089 if (usgn) 4090 return Builder.CreateFPToUI(Ops[0], InTy); 4091 return Builder.CreateFPToSI(Ops[0], InTy); 4092 } 4093 case NEON::BI__builtin_neon_vcvts_f32_u32: 4094 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4095 usgn = true; 4096 // FALL THROUGH 4097 case NEON::BI__builtin_neon_vcvts_f32_s32: 4098 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4099 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4100 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4101 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4102 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4103 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4104 if (usgn) 4105 return Builder.CreateUIToFP(Ops[0], FTy); 4106 return Builder.CreateSIToFP(Ops[0], FTy); 4107 } 4108 case NEON::BI__builtin_neon_vpaddd_s64: { 4109 llvm::Type *Ty = 4110 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4111 Value *Vec = EmitScalarExpr(E->getArg(0)); 4112 // The vector is v2f64, so make sure it's bitcast to that. 4113 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4114 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4115 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4116 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4117 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4118 // Pairwise addition of a v2f64 into a scalar f64. 4119 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4120 } 4121 case NEON::BI__builtin_neon_vpaddd_f64: { 4122 llvm::Type *Ty = 4123 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4124 Value *Vec = EmitScalarExpr(E->getArg(0)); 4125 // The vector is v2f64, so make sure it's bitcast to that. 4126 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4127 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4128 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4129 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4130 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4131 // Pairwise addition of a v2f64 into a scalar f64. 4132 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4133 } 4134 case NEON::BI__builtin_neon_vpadds_f32: { 4135 llvm::Type *Ty = 4136 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4137 Value *Vec = EmitScalarExpr(E->getArg(0)); 4138 // The vector is v2f32, so make sure it's bitcast to that. 4139 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4140 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4141 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4142 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4143 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4144 // Pairwise addition of a v2f32 into a scalar f32. 4145 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4146 } 4147 case NEON::BI__builtin_neon_vceqzd_s64: 4148 case NEON::BI__builtin_neon_vceqzd_f64: 4149 case NEON::BI__builtin_neon_vceqzs_f32: 4150 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4151 return EmitAArch64CompareBuiltinExpr( 4152 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4153 ICmpInst::ICMP_EQ, "vceqz"); 4154 case NEON::BI__builtin_neon_vcgezd_s64: 4155 case NEON::BI__builtin_neon_vcgezd_f64: 4156 case NEON::BI__builtin_neon_vcgezs_f32: 4157 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4158 return EmitAArch64CompareBuiltinExpr( 4159 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4160 ICmpInst::ICMP_SGE, "vcgez"); 4161 case NEON::BI__builtin_neon_vclezd_s64: 4162 case NEON::BI__builtin_neon_vclezd_f64: 4163 case NEON::BI__builtin_neon_vclezs_f32: 4164 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4165 return EmitAArch64CompareBuiltinExpr( 4166 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4167 ICmpInst::ICMP_SLE, "vclez"); 4168 case NEON::BI__builtin_neon_vcgtzd_s64: 4169 case NEON::BI__builtin_neon_vcgtzd_f64: 4170 case NEON::BI__builtin_neon_vcgtzs_f32: 4171 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4172 return EmitAArch64CompareBuiltinExpr( 4173 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4174 ICmpInst::ICMP_SGT, "vcgtz"); 4175 case NEON::BI__builtin_neon_vcltzd_s64: 4176 case NEON::BI__builtin_neon_vcltzd_f64: 4177 case NEON::BI__builtin_neon_vcltzs_f32: 4178 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4179 return EmitAArch64CompareBuiltinExpr( 4180 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4181 ICmpInst::ICMP_SLT, "vcltz"); 4182 4183 case NEON::BI__builtin_neon_vceqzd_u64: { 4184 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4185 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4186 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4187 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4188 llvm::Constant::getNullValue(Ty)); 4189 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4190 } 4191 case NEON::BI__builtin_neon_vceqd_f64: 4192 case NEON::BI__builtin_neon_vcled_f64: 4193 case NEON::BI__builtin_neon_vcltd_f64: 4194 case NEON::BI__builtin_neon_vcged_f64: 4195 case NEON::BI__builtin_neon_vcgtd_f64: { 4196 llvm::CmpInst::Predicate P; 4197 switch (BuiltinID) { 4198 default: llvm_unreachable("missing builtin ID in switch!"); 4199 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4200 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4201 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4202 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4203 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4204 } 4205 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4206 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4207 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4208 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4209 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4210 } 4211 case NEON::BI__builtin_neon_vceqs_f32: 4212 case NEON::BI__builtin_neon_vcles_f32: 4213 case NEON::BI__builtin_neon_vclts_f32: 4214 case NEON::BI__builtin_neon_vcges_f32: 4215 case NEON::BI__builtin_neon_vcgts_f32: { 4216 llvm::CmpInst::Predicate P; 4217 switch (BuiltinID) { 4218 default: llvm_unreachable("missing builtin ID in switch!"); 4219 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4220 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4221 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4222 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4223 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4224 } 4225 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4226 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4227 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4228 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4229 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4230 } 4231 case NEON::BI__builtin_neon_vceqd_s64: 4232 case NEON::BI__builtin_neon_vceqd_u64: 4233 case NEON::BI__builtin_neon_vcgtd_s64: 4234 case NEON::BI__builtin_neon_vcgtd_u64: 4235 case NEON::BI__builtin_neon_vcltd_s64: 4236 case NEON::BI__builtin_neon_vcltd_u64: 4237 case NEON::BI__builtin_neon_vcged_u64: 4238 case NEON::BI__builtin_neon_vcged_s64: 4239 case NEON::BI__builtin_neon_vcled_u64: 4240 case NEON::BI__builtin_neon_vcled_s64: { 4241 llvm::CmpInst::Predicate P; 4242 switch (BuiltinID) { 4243 default: llvm_unreachable("missing builtin ID in switch!"); 4244 case NEON::BI__builtin_neon_vceqd_s64: 4245 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4246 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4247 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4248 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4249 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4250 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4251 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4252 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4253 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4254 } 4255 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4256 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4257 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4258 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4259 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4260 } 4261 case NEON::BI__builtin_neon_vtstd_s64: 4262 case NEON::BI__builtin_neon_vtstd_u64: { 4263 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4264 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4265 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4266 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4267 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4268 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4269 llvm::Constant::getNullValue(Ty)); 4270 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4271 } 4272 case NEON::BI__builtin_neon_vset_lane_i8: 4273 case NEON::BI__builtin_neon_vset_lane_i16: 4274 case NEON::BI__builtin_neon_vset_lane_i32: 4275 case NEON::BI__builtin_neon_vset_lane_i64: 4276 case NEON::BI__builtin_neon_vset_lane_f32: 4277 case NEON::BI__builtin_neon_vsetq_lane_i8: 4278 case NEON::BI__builtin_neon_vsetq_lane_i16: 4279 case NEON::BI__builtin_neon_vsetq_lane_i32: 4280 case NEON::BI__builtin_neon_vsetq_lane_i64: 4281 case NEON::BI__builtin_neon_vsetq_lane_f32: 4282 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4283 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4284 case NEON::BI__builtin_neon_vset_lane_f64: 4285 // The vector type needs a cast for the v1f64 variant. 4286 Ops[1] = Builder.CreateBitCast(Ops[1], 4287 llvm::VectorType::get(DoubleTy, 1)); 4288 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4289 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4290 case NEON::BI__builtin_neon_vsetq_lane_f64: 4291 // The vector type needs a cast for the v2f64 variant. 4292 Ops[1] = Builder.CreateBitCast(Ops[1], 4293 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4294 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4295 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4296 4297 case NEON::BI__builtin_neon_vget_lane_i8: 4298 case NEON::BI__builtin_neon_vdupb_lane_i8: 4299 Ops[0] = Builder.CreateBitCast(Ops[0], 4300 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4301 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4302 "vget_lane"); 4303 case NEON::BI__builtin_neon_vgetq_lane_i8: 4304 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4305 Ops[0] = Builder.CreateBitCast(Ops[0], 4306 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4307 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4308 "vgetq_lane"); 4309 case NEON::BI__builtin_neon_vget_lane_i16: 4310 case NEON::BI__builtin_neon_vduph_lane_i16: 4311 Ops[0] = Builder.CreateBitCast(Ops[0], 4312 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4313 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4314 "vget_lane"); 4315 case NEON::BI__builtin_neon_vgetq_lane_i16: 4316 case NEON::BI__builtin_neon_vduph_laneq_i16: 4317 Ops[0] = Builder.CreateBitCast(Ops[0], 4318 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4319 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4320 "vgetq_lane"); 4321 case NEON::BI__builtin_neon_vget_lane_i32: 4322 case NEON::BI__builtin_neon_vdups_lane_i32: 4323 Ops[0] = Builder.CreateBitCast( 4324 Ops[0], 4325 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4326 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4327 "vget_lane"); 4328 case NEON::BI__builtin_neon_vdups_lane_f32: 4329 Ops[0] = Builder.CreateBitCast(Ops[0], 4330 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4331 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4332 "vdups_lane"); 4333 case NEON::BI__builtin_neon_vgetq_lane_i32: 4334 case NEON::BI__builtin_neon_vdups_laneq_i32: 4335 Ops[0] = Builder.CreateBitCast(Ops[0], 4336 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4337 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4338 "vgetq_lane"); 4339 case NEON::BI__builtin_neon_vget_lane_i64: 4340 case NEON::BI__builtin_neon_vdupd_lane_i64: 4341 Ops[0] = Builder.CreateBitCast(Ops[0], 4342 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4343 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4344 "vget_lane"); 4345 case NEON::BI__builtin_neon_vdupd_lane_f64: 4346 Ops[0] = Builder.CreateBitCast(Ops[0], 4347 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4348 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4349 "vdupd_lane"); 4350 case NEON::BI__builtin_neon_vgetq_lane_i64: 4351 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4352 Ops[0] = Builder.CreateBitCast(Ops[0], 4353 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4354 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4355 "vgetq_lane"); 4356 case NEON::BI__builtin_neon_vget_lane_f32: 4357 Ops[0] = Builder.CreateBitCast(Ops[0], 4358 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4359 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4360 "vget_lane"); 4361 case NEON::BI__builtin_neon_vget_lane_f64: 4362 Ops[0] = Builder.CreateBitCast(Ops[0], 4363 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4364 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4365 "vget_lane"); 4366 case NEON::BI__builtin_neon_vgetq_lane_f32: 4367 case NEON::BI__builtin_neon_vdups_laneq_f32: 4368 Ops[0] = Builder.CreateBitCast(Ops[0], 4369 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4370 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4371 "vgetq_lane"); 4372 case NEON::BI__builtin_neon_vgetq_lane_f64: 4373 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4374 Ops[0] = Builder.CreateBitCast(Ops[0], 4375 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4376 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4377 "vgetq_lane"); 4378 case NEON::BI__builtin_neon_vaddd_s64: 4379 case NEON::BI__builtin_neon_vaddd_u64: 4380 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4381 case NEON::BI__builtin_neon_vsubd_s64: 4382 case NEON::BI__builtin_neon_vsubd_u64: 4383 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4384 case NEON::BI__builtin_neon_vqdmlalh_s16: 4385 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4386 SmallVector<Value *, 2> ProductOps; 4387 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4388 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4389 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4390 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4391 ProductOps, "vqdmlXl"); 4392 Constant *CI = ConstantInt::get(SizeTy, 0); 4393 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4394 4395 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4396 ? Intrinsic::aarch64_neon_sqadd 4397 : Intrinsic::aarch64_neon_sqsub; 4398 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4399 } 4400 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4401 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4402 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4403 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4404 Ops, "vqshlu_n"); 4405 } 4406 case NEON::BI__builtin_neon_vqshld_n_u64: 4407 case NEON::BI__builtin_neon_vqshld_n_s64: { 4408 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4409 ? Intrinsic::aarch64_neon_uqshl 4410 : Intrinsic::aarch64_neon_sqshl; 4411 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4412 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4413 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4414 } 4415 case NEON::BI__builtin_neon_vrshrd_n_u64: 4416 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4417 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4418 ? Intrinsic::aarch64_neon_urshl 4419 : Intrinsic::aarch64_neon_srshl; 4420 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4421 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4422 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4423 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4424 } 4425 case NEON::BI__builtin_neon_vrsrad_n_u64: 4426 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4427 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4428 ? Intrinsic::aarch64_neon_urshl 4429 : Intrinsic::aarch64_neon_srshl; 4430 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4431 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4432 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4433 Builder.CreateSExt(Ops[2], Int64Ty)); 4434 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4435 } 4436 case NEON::BI__builtin_neon_vshld_n_s64: 4437 case NEON::BI__builtin_neon_vshld_n_u64: { 4438 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4439 return Builder.CreateShl( 4440 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4441 } 4442 case NEON::BI__builtin_neon_vshrd_n_s64: { 4443 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4444 return Builder.CreateAShr( 4445 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4446 Amt->getZExtValue())), 4447 "shrd_n"); 4448 } 4449 case NEON::BI__builtin_neon_vshrd_n_u64: { 4450 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4451 uint64_t ShiftAmt = Amt->getZExtValue(); 4452 // Right-shifting an unsigned value by its size yields 0. 4453 if (ShiftAmt == 64) 4454 return ConstantInt::get(Int64Ty, 0); 4455 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4456 "shrd_n"); 4457 } 4458 case NEON::BI__builtin_neon_vsrad_n_s64: { 4459 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4460 Ops[1] = Builder.CreateAShr( 4461 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4462 Amt->getZExtValue())), 4463 "shrd_n"); 4464 return Builder.CreateAdd(Ops[0], Ops[1]); 4465 } 4466 case NEON::BI__builtin_neon_vsrad_n_u64: { 4467 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4468 uint64_t ShiftAmt = Amt->getZExtValue(); 4469 // Right-shifting an unsigned value by its size yields 0. 4470 // As Op + 0 = Op, return Ops[0] directly. 4471 if (ShiftAmt == 64) 4472 return Ops[0]; 4473 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4474 "shrd_n"); 4475 return Builder.CreateAdd(Ops[0], Ops[1]); 4476 } 4477 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4478 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4479 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4480 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4481 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4482 "lane"); 4483 SmallVector<Value *, 2> ProductOps; 4484 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4485 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4486 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4487 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4488 ProductOps, "vqdmlXl"); 4489 Constant *CI = ConstantInt::get(SizeTy, 0); 4490 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4491 Ops.pop_back(); 4492 4493 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4494 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4495 ? Intrinsic::aarch64_neon_sqadd 4496 : Intrinsic::aarch64_neon_sqsub; 4497 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4498 } 4499 case NEON::BI__builtin_neon_vqdmlals_s32: 4500 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4501 SmallVector<Value *, 2> ProductOps; 4502 ProductOps.push_back(Ops[1]); 4503 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4504 Ops[1] = 4505 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4506 ProductOps, "vqdmlXl"); 4507 4508 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4509 ? Intrinsic::aarch64_neon_sqadd 4510 : Intrinsic::aarch64_neon_sqsub; 4511 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4512 } 4513 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4514 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4515 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4516 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4517 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4518 "lane"); 4519 SmallVector<Value *, 2> ProductOps; 4520 ProductOps.push_back(Ops[1]); 4521 ProductOps.push_back(Ops[2]); 4522 Ops[1] = 4523 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4524 ProductOps, "vqdmlXl"); 4525 Ops.pop_back(); 4526 4527 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4528 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4529 ? Intrinsic::aarch64_neon_sqadd 4530 : Intrinsic::aarch64_neon_sqsub; 4531 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4532 } 4533 } 4534 4535 llvm::VectorType *VTy = GetNeonType(this, Type); 4536 llvm::Type *Ty = VTy; 4537 if (!Ty) 4538 return nullptr; 4539 4540 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4541 // defer to common code if it's been added to our special map. 4542 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4543 AArch64SIMDIntrinsicsProvenSorted); 4544 4545 if (Builtin) 4546 return EmitCommonNeonBuiltinExpr( 4547 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4548 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4549 4550 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4551 return V; 4552 4553 unsigned Int; 4554 switch (BuiltinID) { 4555 default: return nullptr; 4556 case NEON::BI__builtin_neon_vbsl_v: 4557 case NEON::BI__builtin_neon_vbslq_v: { 4558 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4559 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4560 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4561 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4562 4563 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4564 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4565 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4566 return Builder.CreateBitCast(Ops[0], Ty); 4567 } 4568 case NEON::BI__builtin_neon_vfma_lane_v: 4569 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4570 // The ARM builtins (and instructions) have the addend as the first 4571 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4572 Value *Addend = Ops[0]; 4573 Value *Multiplicand = Ops[1]; 4574 Value *LaneSource = Ops[2]; 4575 Ops[0] = Multiplicand; 4576 Ops[1] = LaneSource; 4577 Ops[2] = Addend; 4578 4579 // Now adjust things to handle the lane access. 4580 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4581 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4582 VTy; 4583 llvm::Constant *cst = cast<Constant>(Ops[3]); 4584 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4585 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4586 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4587 4588 Ops.pop_back(); 4589 Int = Intrinsic::fma; 4590 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4591 } 4592 case NEON::BI__builtin_neon_vfma_laneq_v: { 4593 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4594 // v1f64 fma should be mapped to Neon scalar f64 fma 4595 if (VTy && VTy->getElementType() == DoubleTy) { 4596 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4597 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4598 llvm::Type *VTy = GetNeonType(this, 4599 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4600 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4601 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4602 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4603 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4604 return Builder.CreateBitCast(Result, Ty); 4605 } 4606 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4607 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4608 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4609 4610 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4611 VTy->getNumElements() * 2); 4612 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4613 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4614 cast<ConstantInt>(Ops[3])); 4615 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4616 4617 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4618 } 4619 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4620 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4621 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4622 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4623 4624 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4625 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4626 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4627 } 4628 case NEON::BI__builtin_neon_vfmas_lane_f32: 4629 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4630 case NEON::BI__builtin_neon_vfmad_lane_f64: 4631 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4632 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4633 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4634 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4635 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4636 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4637 } 4638 case NEON::BI__builtin_neon_vfms_v: 4639 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4640 // FIXME: probably remove when we no longer support aarch64_simd.h 4641 // (arm_neon.h delegates to vfma). 4642 4643 // The ARM builtins (and instructions) have the addend as the first 4644 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4645 Value *Subtrahend = Ops[0]; 4646 Value *Multiplicand = Ops[2]; 4647 Ops[0] = Multiplicand; 4648 Ops[2] = Subtrahend; 4649 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4650 Ops[1] = Builder.CreateFNeg(Ops[1]); 4651 Int = Intrinsic::fma; 4652 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4653 } 4654 case NEON::BI__builtin_neon_vmull_v: 4655 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4656 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4657 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4658 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4659 case NEON::BI__builtin_neon_vmax_v: 4660 case NEON::BI__builtin_neon_vmaxq_v: 4661 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4662 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4663 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4664 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4665 case NEON::BI__builtin_neon_vmin_v: 4666 case NEON::BI__builtin_neon_vminq_v: 4667 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4668 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4669 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4670 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4671 case NEON::BI__builtin_neon_vabd_v: 4672 case NEON::BI__builtin_neon_vabdq_v: 4673 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4674 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4675 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4676 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4677 case NEON::BI__builtin_neon_vpadal_v: 4678 case NEON::BI__builtin_neon_vpadalq_v: { 4679 unsigned ArgElts = VTy->getNumElements(); 4680 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4681 unsigned BitWidth = EltTy->getBitWidth(); 4682 llvm::Type *ArgTy = llvm::VectorType::get( 4683 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4684 llvm::Type* Tys[2] = { VTy, ArgTy }; 4685 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4686 SmallVector<llvm::Value*, 1> TmpOps; 4687 TmpOps.push_back(Ops[1]); 4688 Function *F = CGM.getIntrinsic(Int, Tys); 4689 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4690 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4691 return Builder.CreateAdd(tmp, addend); 4692 } 4693 case NEON::BI__builtin_neon_vpmin_v: 4694 case NEON::BI__builtin_neon_vpminq_v: 4695 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4696 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4697 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4698 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4699 case NEON::BI__builtin_neon_vpmax_v: 4700 case NEON::BI__builtin_neon_vpmaxq_v: 4701 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4702 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4703 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4704 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4705 case NEON::BI__builtin_neon_vminnm_v: 4706 case NEON::BI__builtin_neon_vminnmq_v: 4707 Int = Intrinsic::aarch64_neon_fminnm; 4708 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4709 case NEON::BI__builtin_neon_vmaxnm_v: 4710 case NEON::BI__builtin_neon_vmaxnmq_v: 4711 Int = Intrinsic::aarch64_neon_fmaxnm; 4712 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4713 case NEON::BI__builtin_neon_vrecpss_f32: { 4714 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4715 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4716 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4717 Ops, "vrecps"); 4718 } 4719 case NEON::BI__builtin_neon_vrecpsd_f64: { 4720 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4721 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4722 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4723 Ops, "vrecps"); 4724 } 4725 case NEON::BI__builtin_neon_vqshrun_n_v: 4726 Int = Intrinsic::aarch64_neon_sqshrun; 4727 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4728 case NEON::BI__builtin_neon_vqrshrun_n_v: 4729 Int = Intrinsic::aarch64_neon_sqrshrun; 4730 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4731 case NEON::BI__builtin_neon_vqshrn_n_v: 4732 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4733 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4734 case NEON::BI__builtin_neon_vrshrn_n_v: 4735 Int = Intrinsic::aarch64_neon_rshrn; 4736 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4737 case NEON::BI__builtin_neon_vqrshrn_n_v: 4738 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4739 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4740 case NEON::BI__builtin_neon_vrnda_v: 4741 case NEON::BI__builtin_neon_vrndaq_v: { 4742 Int = Intrinsic::round; 4743 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4744 } 4745 case NEON::BI__builtin_neon_vrndi_v: 4746 case NEON::BI__builtin_neon_vrndiq_v: { 4747 Int = Intrinsic::nearbyint; 4748 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4749 } 4750 case NEON::BI__builtin_neon_vrndm_v: 4751 case NEON::BI__builtin_neon_vrndmq_v: { 4752 Int = Intrinsic::floor; 4753 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4754 } 4755 case NEON::BI__builtin_neon_vrndn_v: 4756 case NEON::BI__builtin_neon_vrndnq_v: { 4757 Int = Intrinsic::aarch64_neon_frintn; 4758 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4759 } 4760 case NEON::BI__builtin_neon_vrndp_v: 4761 case NEON::BI__builtin_neon_vrndpq_v: { 4762 Int = Intrinsic::ceil; 4763 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4764 } 4765 case NEON::BI__builtin_neon_vrndx_v: 4766 case NEON::BI__builtin_neon_vrndxq_v: { 4767 Int = Intrinsic::rint; 4768 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4769 } 4770 case NEON::BI__builtin_neon_vrnd_v: 4771 case NEON::BI__builtin_neon_vrndq_v: { 4772 Int = Intrinsic::trunc; 4773 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4774 } 4775 case NEON::BI__builtin_neon_vceqz_v: 4776 case NEON::BI__builtin_neon_vceqzq_v: 4777 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4778 ICmpInst::ICMP_EQ, "vceqz"); 4779 case NEON::BI__builtin_neon_vcgez_v: 4780 case NEON::BI__builtin_neon_vcgezq_v: 4781 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4782 ICmpInst::ICMP_SGE, "vcgez"); 4783 case NEON::BI__builtin_neon_vclez_v: 4784 case NEON::BI__builtin_neon_vclezq_v: 4785 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4786 ICmpInst::ICMP_SLE, "vclez"); 4787 case NEON::BI__builtin_neon_vcgtz_v: 4788 case NEON::BI__builtin_neon_vcgtzq_v: 4789 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4790 ICmpInst::ICMP_SGT, "vcgtz"); 4791 case NEON::BI__builtin_neon_vcltz_v: 4792 case NEON::BI__builtin_neon_vcltzq_v: 4793 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4794 ICmpInst::ICMP_SLT, "vcltz"); 4795 case NEON::BI__builtin_neon_vcvt_f64_v: 4796 case NEON::BI__builtin_neon_vcvtq_f64_v: 4797 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4798 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4799 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4800 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4801 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4802 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4803 "unexpected vcvt_f64_f32 builtin"); 4804 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4805 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4806 4807 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4808 } 4809 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4810 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4811 "unexpected vcvt_f32_f64 builtin"); 4812 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4813 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4814 4815 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4816 } 4817 case NEON::BI__builtin_neon_vcvt_s32_v: 4818 case NEON::BI__builtin_neon_vcvt_u32_v: 4819 case NEON::BI__builtin_neon_vcvt_s64_v: 4820 case NEON::BI__builtin_neon_vcvt_u64_v: 4821 case NEON::BI__builtin_neon_vcvtq_s32_v: 4822 case NEON::BI__builtin_neon_vcvtq_u32_v: 4823 case NEON::BI__builtin_neon_vcvtq_s64_v: 4824 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4825 bool Double = 4826 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4827 llvm::Type *InTy = 4828 GetNeonType(this, 4829 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4830 : NeonTypeFlags::Float32, false, quad)); 4831 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4832 if (usgn) 4833 return Builder.CreateFPToUI(Ops[0], Ty); 4834 return Builder.CreateFPToSI(Ops[0], Ty); 4835 } 4836 case NEON::BI__builtin_neon_vcvta_s32_v: 4837 case NEON::BI__builtin_neon_vcvtaq_s32_v: 4838 case NEON::BI__builtin_neon_vcvta_u32_v: 4839 case NEON::BI__builtin_neon_vcvtaq_u32_v: 4840 case NEON::BI__builtin_neon_vcvta_s64_v: 4841 case NEON::BI__builtin_neon_vcvtaq_s64_v: 4842 case NEON::BI__builtin_neon_vcvta_u64_v: 4843 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4844 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 4845 bool Double = 4846 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4847 llvm::Type *InTy = 4848 GetNeonType(this, 4849 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4850 : NeonTypeFlags::Float32, false, quad)); 4851 llvm::Type *Tys[2] = { Ty, InTy }; 4852 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 4853 } 4854 case NEON::BI__builtin_neon_vcvtm_s32_v: 4855 case NEON::BI__builtin_neon_vcvtmq_s32_v: 4856 case NEON::BI__builtin_neon_vcvtm_u32_v: 4857 case NEON::BI__builtin_neon_vcvtmq_u32_v: 4858 case NEON::BI__builtin_neon_vcvtm_s64_v: 4859 case NEON::BI__builtin_neon_vcvtmq_s64_v: 4860 case NEON::BI__builtin_neon_vcvtm_u64_v: 4861 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4862 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 4863 bool Double = 4864 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4865 llvm::Type *InTy = 4866 GetNeonType(this, 4867 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4868 : NeonTypeFlags::Float32, false, quad)); 4869 llvm::Type *Tys[2] = { Ty, InTy }; 4870 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 4871 } 4872 case NEON::BI__builtin_neon_vcvtn_s32_v: 4873 case NEON::BI__builtin_neon_vcvtnq_s32_v: 4874 case NEON::BI__builtin_neon_vcvtn_u32_v: 4875 case NEON::BI__builtin_neon_vcvtnq_u32_v: 4876 case NEON::BI__builtin_neon_vcvtn_s64_v: 4877 case NEON::BI__builtin_neon_vcvtnq_s64_v: 4878 case NEON::BI__builtin_neon_vcvtn_u64_v: 4879 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4880 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 4881 bool Double = 4882 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4883 llvm::Type *InTy = 4884 GetNeonType(this, 4885 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4886 : NeonTypeFlags::Float32, false, quad)); 4887 llvm::Type *Tys[2] = { Ty, InTy }; 4888 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 4889 } 4890 case NEON::BI__builtin_neon_vcvtp_s32_v: 4891 case NEON::BI__builtin_neon_vcvtpq_s32_v: 4892 case NEON::BI__builtin_neon_vcvtp_u32_v: 4893 case NEON::BI__builtin_neon_vcvtpq_u32_v: 4894 case NEON::BI__builtin_neon_vcvtp_s64_v: 4895 case NEON::BI__builtin_neon_vcvtpq_s64_v: 4896 case NEON::BI__builtin_neon_vcvtp_u64_v: 4897 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4898 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 4899 bool Double = 4900 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4901 llvm::Type *InTy = 4902 GetNeonType(this, 4903 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4904 : NeonTypeFlags::Float32, false, quad)); 4905 llvm::Type *Tys[2] = { Ty, InTy }; 4906 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 4907 } 4908 case NEON::BI__builtin_neon_vmulx_v: 4909 case NEON::BI__builtin_neon_vmulxq_v: { 4910 Int = Intrinsic::aarch64_neon_fmulx; 4911 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 4912 } 4913 case NEON::BI__builtin_neon_vmul_lane_v: 4914 case NEON::BI__builtin_neon_vmul_laneq_v: { 4915 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 4916 bool Quad = false; 4917 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 4918 Quad = true; 4919 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4920 llvm::Type *VTy = GetNeonType(this, 4921 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 4922 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4923 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 4924 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 4925 return Builder.CreateBitCast(Result, Ty); 4926 } 4927 case NEON::BI__builtin_neon_vnegd_s64: 4928 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 4929 case NEON::BI__builtin_neon_vpmaxnm_v: 4930 case NEON::BI__builtin_neon_vpmaxnmq_v: { 4931 Int = Intrinsic::aarch64_neon_fmaxnmp; 4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 4933 } 4934 case NEON::BI__builtin_neon_vpminnm_v: 4935 case NEON::BI__builtin_neon_vpminnmq_v: { 4936 Int = Intrinsic::aarch64_neon_fminnmp; 4937 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 4938 } 4939 case NEON::BI__builtin_neon_vsqrt_v: 4940 case NEON::BI__builtin_neon_vsqrtq_v: { 4941 Int = Intrinsic::sqrt; 4942 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4943 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 4944 } 4945 case NEON::BI__builtin_neon_vrbit_v: 4946 case NEON::BI__builtin_neon_vrbitq_v: { 4947 Int = Intrinsic::aarch64_neon_rbit; 4948 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 4949 } 4950 case NEON::BI__builtin_neon_vaddv_u8: 4951 // FIXME: These are handled by the AArch64 scalar code. 4952 usgn = true; 4953 // FALLTHROUGH 4954 case NEON::BI__builtin_neon_vaddv_s8: { 4955 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4956 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4957 VTy = 4958 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 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, "vaddv"); 4962 return Builder.CreateTrunc(Ops[0], 4963 llvm::IntegerType::get(getLLVMContext(), 8)); 4964 } 4965 case NEON::BI__builtin_neon_vaddv_u16: 4966 usgn = true; 4967 // FALLTHROUGH 4968 case NEON::BI__builtin_neon_vaddv_s16: { 4969 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4970 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4971 VTy = 4972 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 4973 llvm::Type *Tys[2] = { Ty, VTy }; 4974 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4975 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4976 return Builder.CreateTrunc(Ops[0], 4977 llvm::IntegerType::get(getLLVMContext(), 16)); 4978 } 4979 case NEON::BI__builtin_neon_vaddvq_u8: 4980 usgn = true; 4981 // FALLTHROUGH 4982 case NEON::BI__builtin_neon_vaddvq_s8: { 4983 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4984 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4985 VTy = 4986 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 4987 llvm::Type *Tys[2] = { Ty, VTy }; 4988 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4989 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4990 return Builder.CreateTrunc(Ops[0], 4991 llvm::IntegerType::get(getLLVMContext(), 8)); 4992 } 4993 case NEON::BI__builtin_neon_vaddvq_u16: 4994 usgn = true; 4995 // FALLTHROUGH 4996 case NEON::BI__builtin_neon_vaddvq_s16: { 4997 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4998 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4999 VTy = 5000 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5001 llvm::Type *Tys[2] = { Ty, VTy }; 5002 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5003 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5004 return Builder.CreateTrunc(Ops[0], 5005 llvm::IntegerType::get(getLLVMContext(), 16)); 5006 } 5007 case NEON::BI__builtin_neon_vmaxv_u8: { 5008 Int = Intrinsic::aarch64_neon_umaxv; 5009 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5010 VTy = 5011 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5012 llvm::Type *Tys[2] = { Ty, VTy }; 5013 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5014 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5015 return Builder.CreateTrunc(Ops[0], 5016 llvm::IntegerType::get(getLLVMContext(), 8)); 5017 } 5018 case NEON::BI__builtin_neon_vmaxv_u16: { 5019 Int = Intrinsic::aarch64_neon_umaxv; 5020 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5021 VTy = 5022 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5023 llvm::Type *Tys[2] = { Ty, VTy }; 5024 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5025 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5026 return Builder.CreateTrunc(Ops[0], 5027 llvm::IntegerType::get(getLLVMContext(), 16)); 5028 } 5029 case NEON::BI__builtin_neon_vmaxvq_u8: { 5030 Int = Intrinsic::aarch64_neon_umaxv; 5031 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5032 VTy = 5033 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5034 llvm::Type *Tys[2] = { Ty, VTy }; 5035 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5036 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5037 return Builder.CreateTrunc(Ops[0], 5038 llvm::IntegerType::get(getLLVMContext(), 8)); 5039 } 5040 case NEON::BI__builtin_neon_vmaxvq_u16: { 5041 Int = Intrinsic::aarch64_neon_umaxv; 5042 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5043 VTy = 5044 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5045 llvm::Type *Tys[2] = { Ty, VTy }; 5046 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5047 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5048 return Builder.CreateTrunc(Ops[0], 5049 llvm::IntegerType::get(getLLVMContext(), 16)); 5050 } 5051 case NEON::BI__builtin_neon_vmaxv_s8: { 5052 Int = Intrinsic::aarch64_neon_smaxv; 5053 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5054 VTy = 5055 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5056 llvm::Type *Tys[2] = { Ty, VTy }; 5057 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5058 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5059 return Builder.CreateTrunc(Ops[0], 5060 llvm::IntegerType::get(getLLVMContext(), 8)); 5061 } 5062 case NEON::BI__builtin_neon_vmaxv_s16: { 5063 Int = Intrinsic::aarch64_neon_smaxv; 5064 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5065 VTy = 5066 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5067 llvm::Type *Tys[2] = { Ty, VTy }; 5068 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5069 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5070 return Builder.CreateTrunc(Ops[0], 5071 llvm::IntegerType::get(getLLVMContext(), 16)); 5072 } 5073 case NEON::BI__builtin_neon_vmaxvq_s8: { 5074 Int = Intrinsic::aarch64_neon_smaxv; 5075 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5076 VTy = 5077 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5078 llvm::Type *Tys[2] = { Ty, VTy }; 5079 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5080 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5081 return Builder.CreateTrunc(Ops[0], 5082 llvm::IntegerType::get(getLLVMContext(), 8)); 5083 } 5084 case NEON::BI__builtin_neon_vmaxvq_s16: { 5085 Int = Intrinsic::aarch64_neon_smaxv; 5086 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5087 VTy = 5088 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5089 llvm::Type *Tys[2] = { Ty, VTy }; 5090 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5091 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5092 return Builder.CreateTrunc(Ops[0], 5093 llvm::IntegerType::get(getLLVMContext(), 16)); 5094 } 5095 case NEON::BI__builtin_neon_vminv_u8: { 5096 Int = Intrinsic::aarch64_neon_uminv; 5097 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5098 VTy = 5099 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5100 llvm::Type *Tys[2] = { Ty, VTy }; 5101 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5102 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5103 return Builder.CreateTrunc(Ops[0], 5104 llvm::IntegerType::get(getLLVMContext(), 8)); 5105 } 5106 case NEON::BI__builtin_neon_vminv_u16: { 5107 Int = Intrinsic::aarch64_neon_uminv; 5108 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5109 VTy = 5110 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5111 llvm::Type *Tys[2] = { Ty, VTy }; 5112 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5113 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5114 return Builder.CreateTrunc(Ops[0], 5115 llvm::IntegerType::get(getLLVMContext(), 16)); 5116 } 5117 case NEON::BI__builtin_neon_vminvq_u8: { 5118 Int = Intrinsic::aarch64_neon_uminv; 5119 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5120 VTy = 5121 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5122 llvm::Type *Tys[2] = { Ty, VTy }; 5123 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5124 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5125 return Builder.CreateTrunc(Ops[0], 5126 llvm::IntegerType::get(getLLVMContext(), 8)); 5127 } 5128 case NEON::BI__builtin_neon_vminvq_u16: { 5129 Int = Intrinsic::aarch64_neon_uminv; 5130 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5131 VTy = 5132 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5133 llvm::Type *Tys[2] = { Ty, VTy }; 5134 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5135 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5136 return Builder.CreateTrunc(Ops[0], 5137 llvm::IntegerType::get(getLLVMContext(), 16)); 5138 } 5139 case NEON::BI__builtin_neon_vminv_s8: { 5140 Int = Intrinsic::aarch64_neon_sminv; 5141 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5142 VTy = 5143 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5144 llvm::Type *Tys[2] = { Ty, VTy }; 5145 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5146 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5147 return Builder.CreateTrunc(Ops[0], 5148 llvm::IntegerType::get(getLLVMContext(), 8)); 5149 } 5150 case NEON::BI__builtin_neon_vminv_s16: { 5151 Int = Intrinsic::aarch64_neon_sminv; 5152 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5153 VTy = 5154 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5155 llvm::Type *Tys[2] = { Ty, VTy }; 5156 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5157 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5158 return Builder.CreateTrunc(Ops[0], 5159 llvm::IntegerType::get(getLLVMContext(), 16)); 5160 } 5161 case NEON::BI__builtin_neon_vminvq_s8: { 5162 Int = Intrinsic::aarch64_neon_sminv; 5163 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5164 VTy = 5165 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5166 llvm::Type *Tys[2] = { Ty, VTy }; 5167 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5168 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5169 return Builder.CreateTrunc(Ops[0], 5170 llvm::IntegerType::get(getLLVMContext(), 8)); 5171 } 5172 case NEON::BI__builtin_neon_vminvq_s16: { 5173 Int = Intrinsic::aarch64_neon_sminv; 5174 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5175 VTy = 5176 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5177 llvm::Type *Tys[2] = { Ty, VTy }; 5178 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5179 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5180 return Builder.CreateTrunc(Ops[0], 5181 llvm::IntegerType::get(getLLVMContext(), 16)); 5182 } 5183 case NEON::BI__builtin_neon_vmul_n_f64: { 5184 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5185 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5186 return Builder.CreateFMul(Ops[0], RHS); 5187 } 5188 case NEON::BI__builtin_neon_vaddlv_u8: { 5189 Int = Intrinsic::aarch64_neon_uaddlv; 5190 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5191 VTy = 5192 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5193 llvm::Type *Tys[2] = { Ty, VTy }; 5194 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5195 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5196 return Builder.CreateTrunc(Ops[0], 5197 llvm::IntegerType::get(getLLVMContext(), 16)); 5198 } 5199 case NEON::BI__builtin_neon_vaddlv_u16: { 5200 Int = Intrinsic::aarch64_neon_uaddlv; 5201 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5202 VTy = 5203 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5204 llvm::Type *Tys[2] = { Ty, VTy }; 5205 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5206 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5207 } 5208 case NEON::BI__builtin_neon_vaddlvq_u8: { 5209 Int = Intrinsic::aarch64_neon_uaddlv; 5210 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5211 VTy = 5212 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5213 llvm::Type *Tys[2] = { Ty, VTy }; 5214 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5215 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5216 return Builder.CreateTrunc(Ops[0], 5217 llvm::IntegerType::get(getLLVMContext(), 16)); 5218 } 5219 case NEON::BI__builtin_neon_vaddlvq_u16: { 5220 Int = Intrinsic::aarch64_neon_uaddlv; 5221 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5222 VTy = 5223 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5224 llvm::Type *Tys[2] = { Ty, VTy }; 5225 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5226 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5227 } 5228 case NEON::BI__builtin_neon_vaddlv_s8: { 5229 Int = Intrinsic::aarch64_neon_saddlv; 5230 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5231 VTy = 5232 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5233 llvm::Type *Tys[2] = { Ty, VTy }; 5234 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5235 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5236 return Builder.CreateTrunc(Ops[0], 5237 llvm::IntegerType::get(getLLVMContext(), 16)); 5238 } 5239 case NEON::BI__builtin_neon_vaddlv_s16: { 5240 Int = Intrinsic::aarch64_neon_saddlv; 5241 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5242 VTy = 5243 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5244 llvm::Type *Tys[2] = { Ty, VTy }; 5245 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5246 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5247 } 5248 case NEON::BI__builtin_neon_vaddlvq_s8: { 5249 Int = Intrinsic::aarch64_neon_saddlv; 5250 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5251 VTy = 5252 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5253 llvm::Type *Tys[2] = { Ty, VTy }; 5254 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5255 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5256 return Builder.CreateTrunc(Ops[0], 5257 llvm::IntegerType::get(getLLVMContext(), 16)); 5258 } 5259 case NEON::BI__builtin_neon_vaddlvq_s16: { 5260 Int = Intrinsic::aarch64_neon_saddlv; 5261 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5262 VTy = 5263 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5264 llvm::Type *Tys[2] = { Ty, VTy }; 5265 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5266 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5267 } 5268 case NEON::BI__builtin_neon_vsri_n_v: 5269 case NEON::BI__builtin_neon_vsriq_n_v: { 5270 Int = Intrinsic::aarch64_neon_vsri; 5271 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5272 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5273 } 5274 case NEON::BI__builtin_neon_vsli_n_v: 5275 case NEON::BI__builtin_neon_vsliq_n_v: { 5276 Int = Intrinsic::aarch64_neon_vsli; 5277 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5278 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5279 } 5280 case NEON::BI__builtin_neon_vsra_n_v: 5281 case NEON::BI__builtin_neon_vsraq_n_v: 5282 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5283 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5284 return Builder.CreateAdd(Ops[0], Ops[1]); 5285 case NEON::BI__builtin_neon_vrsra_n_v: 5286 case NEON::BI__builtin_neon_vrsraq_n_v: { 5287 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5288 SmallVector<llvm::Value*,2> TmpOps; 5289 TmpOps.push_back(Ops[1]); 5290 TmpOps.push_back(Ops[2]); 5291 Function* F = CGM.getIntrinsic(Int, Ty); 5292 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5293 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5294 return Builder.CreateAdd(Ops[0], tmp); 5295 } 5296 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5297 // of an Align parameter here. 5298 case NEON::BI__builtin_neon_vld1_x2_v: 5299 case NEON::BI__builtin_neon_vld1q_x2_v: 5300 case NEON::BI__builtin_neon_vld1_x3_v: 5301 case NEON::BI__builtin_neon_vld1q_x3_v: 5302 case NEON::BI__builtin_neon_vld1_x4_v: 5303 case NEON::BI__builtin_neon_vld1q_x4_v: { 5304 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5305 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5306 llvm::Type *Tys[2] = { VTy, PTy }; 5307 unsigned Int; 5308 switch (BuiltinID) { 5309 case NEON::BI__builtin_neon_vld1_x2_v: 5310 case NEON::BI__builtin_neon_vld1q_x2_v: 5311 Int = Intrinsic::aarch64_neon_ld1x2; 5312 break; 5313 case NEON::BI__builtin_neon_vld1_x3_v: 5314 case NEON::BI__builtin_neon_vld1q_x3_v: 5315 Int = Intrinsic::aarch64_neon_ld1x3; 5316 break; 5317 case NEON::BI__builtin_neon_vld1_x4_v: 5318 case NEON::BI__builtin_neon_vld1q_x4_v: 5319 Int = Intrinsic::aarch64_neon_ld1x4; 5320 break; 5321 } 5322 Function *F = CGM.getIntrinsic(Int, Tys); 5323 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5324 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5325 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5326 return Builder.CreateStore(Ops[1], Ops[0]); 5327 } 5328 case NEON::BI__builtin_neon_vst1_x2_v: 5329 case NEON::BI__builtin_neon_vst1q_x2_v: 5330 case NEON::BI__builtin_neon_vst1_x3_v: 5331 case NEON::BI__builtin_neon_vst1q_x3_v: 5332 case NEON::BI__builtin_neon_vst1_x4_v: 5333 case NEON::BI__builtin_neon_vst1q_x4_v: { 5334 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5335 llvm::Type *Tys[2] = { VTy, PTy }; 5336 unsigned Int; 5337 switch (BuiltinID) { 5338 case NEON::BI__builtin_neon_vst1_x2_v: 5339 case NEON::BI__builtin_neon_vst1q_x2_v: 5340 Int = Intrinsic::aarch64_neon_st1x2; 5341 break; 5342 case NEON::BI__builtin_neon_vst1_x3_v: 5343 case NEON::BI__builtin_neon_vst1q_x3_v: 5344 Int = Intrinsic::aarch64_neon_st1x3; 5345 break; 5346 case NEON::BI__builtin_neon_vst1_x4_v: 5347 case NEON::BI__builtin_neon_vst1q_x4_v: 5348 Int = Intrinsic::aarch64_neon_st1x4; 5349 break; 5350 } 5351 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5352 IntOps.push_back(Ops[0]); 5353 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5354 } 5355 case NEON::BI__builtin_neon_vld1_v: 5356 case NEON::BI__builtin_neon_vld1q_v: 5357 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5358 return Builder.CreateLoad(Ops[0]); 5359 case NEON::BI__builtin_neon_vst1_v: 5360 case NEON::BI__builtin_neon_vst1q_v: 5361 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5362 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5363 return Builder.CreateStore(Ops[1], Ops[0]); 5364 case NEON::BI__builtin_neon_vld1_lane_v: 5365 case NEON::BI__builtin_neon_vld1q_lane_v: 5366 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5367 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5368 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5369 Ops[0] = Builder.CreateLoad(Ops[0]); 5370 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5371 case NEON::BI__builtin_neon_vld1_dup_v: 5372 case NEON::BI__builtin_neon_vld1q_dup_v: { 5373 Value *V = UndefValue::get(Ty); 5374 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5375 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5376 Ops[0] = Builder.CreateLoad(Ops[0]); 5377 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5378 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5379 return EmitNeonSplat(Ops[0], CI); 5380 } 5381 case NEON::BI__builtin_neon_vst1_lane_v: 5382 case NEON::BI__builtin_neon_vst1q_lane_v: 5383 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5384 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5385 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5386 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5387 case NEON::BI__builtin_neon_vld2_v: 5388 case NEON::BI__builtin_neon_vld2q_v: { 5389 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5390 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5391 llvm::Type *Tys[2] = { VTy, PTy }; 5392 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5393 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5394 Ops[0] = Builder.CreateBitCast(Ops[0], 5395 llvm::PointerType::getUnqual(Ops[1]->getType())); 5396 return Builder.CreateStore(Ops[1], Ops[0]); 5397 } 5398 case NEON::BI__builtin_neon_vld3_v: 5399 case NEON::BI__builtin_neon_vld3q_v: { 5400 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5401 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5402 llvm::Type *Tys[2] = { VTy, PTy }; 5403 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5404 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5405 Ops[0] = Builder.CreateBitCast(Ops[0], 5406 llvm::PointerType::getUnqual(Ops[1]->getType())); 5407 return Builder.CreateStore(Ops[1], Ops[0]); 5408 } 5409 case NEON::BI__builtin_neon_vld4_v: 5410 case NEON::BI__builtin_neon_vld4q_v: { 5411 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5412 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5413 llvm::Type *Tys[2] = { VTy, PTy }; 5414 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5415 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5416 Ops[0] = Builder.CreateBitCast(Ops[0], 5417 llvm::PointerType::getUnqual(Ops[1]->getType())); 5418 return Builder.CreateStore(Ops[1], Ops[0]); 5419 } 5420 case NEON::BI__builtin_neon_vld2_dup_v: 5421 case NEON::BI__builtin_neon_vld2q_dup_v: { 5422 llvm::Type *PTy = 5423 llvm::PointerType::getUnqual(VTy->getElementType()); 5424 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5425 llvm::Type *Tys[2] = { VTy, PTy }; 5426 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5427 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5428 Ops[0] = Builder.CreateBitCast(Ops[0], 5429 llvm::PointerType::getUnqual(Ops[1]->getType())); 5430 return Builder.CreateStore(Ops[1], Ops[0]); 5431 } 5432 case NEON::BI__builtin_neon_vld3_dup_v: 5433 case NEON::BI__builtin_neon_vld3q_dup_v: { 5434 llvm::Type *PTy = 5435 llvm::PointerType::getUnqual(VTy->getElementType()); 5436 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5437 llvm::Type *Tys[2] = { VTy, PTy }; 5438 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5439 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5440 Ops[0] = Builder.CreateBitCast(Ops[0], 5441 llvm::PointerType::getUnqual(Ops[1]->getType())); 5442 return Builder.CreateStore(Ops[1], Ops[0]); 5443 } 5444 case NEON::BI__builtin_neon_vld4_dup_v: 5445 case NEON::BI__builtin_neon_vld4q_dup_v: { 5446 llvm::Type *PTy = 5447 llvm::PointerType::getUnqual(VTy->getElementType()); 5448 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5449 llvm::Type *Tys[2] = { VTy, PTy }; 5450 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5451 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5452 Ops[0] = Builder.CreateBitCast(Ops[0], 5453 llvm::PointerType::getUnqual(Ops[1]->getType())); 5454 return Builder.CreateStore(Ops[1], Ops[0]); 5455 } 5456 case NEON::BI__builtin_neon_vld2_lane_v: 5457 case NEON::BI__builtin_neon_vld2q_lane_v: { 5458 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5459 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5460 Ops.push_back(Ops[1]); 5461 Ops.erase(Ops.begin()+1); 5462 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5463 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5464 Ops[3] = Builder.CreateZExt(Ops[3], 5465 llvm::IntegerType::get(getLLVMContext(), 64)); 5466 Ops[1] = Builder.CreateCall(F, 5467 ArrayRef<Value*>(Ops).slice(1), "vld2_lane"); 5468 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5469 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5470 return Builder.CreateStore(Ops[1], Ops[0]); 5471 } 5472 case NEON::BI__builtin_neon_vld3_lane_v: 5473 case NEON::BI__builtin_neon_vld3q_lane_v: { 5474 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5475 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5476 Ops.push_back(Ops[1]); 5477 Ops.erase(Ops.begin()+1); 5478 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5479 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5480 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5481 Ops[4] = Builder.CreateZExt(Ops[4], 5482 llvm::IntegerType::get(getLLVMContext(), 64)); 5483 Ops[1] = Builder.CreateCall(F, 5484 ArrayRef<Value*>(Ops).slice(1), "vld3_lane"); 5485 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5486 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5487 return Builder.CreateStore(Ops[1], Ops[0]); 5488 } 5489 case NEON::BI__builtin_neon_vld4_lane_v: 5490 case NEON::BI__builtin_neon_vld4q_lane_v: { 5491 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5492 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5493 Ops.push_back(Ops[1]); 5494 Ops.erase(Ops.begin()+1); 5495 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5496 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5497 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5498 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5499 Ops[5] = Builder.CreateZExt(Ops[5], 5500 llvm::IntegerType::get(getLLVMContext(), 64)); 5501 Ops[1] = Builder.CreateCall(F, 5502 ArrayRef<Value*>(Ops).slice(1), "vld4_lane"); 5503 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5504 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5505 return Builder.CreateStore(Ops[1], Ops[0]); 5506 } 5507 case NEON::BI__builtin_neon_vst2_v: 5508 case NEON::BI__builtin_neon_vst2q_v: { 5509 Ops.push_back(Ops[0]); 5510 Ops.erase(Ops.begin()); 5511 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5512 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5513 Ops, ""); 5514 } 5515 case NEON::BI__builtin_neon_vst2_lane_v: 5516 case NEON::BI__builtin_neon_vst2q_lane_v: { 5517 Ops.push_back(Ops[0]); 5518 Ops.erase(Ops.begin()); 5519 Ops[2] = Builder.CreateZExt(Ops[2], 5520 llvm::IntegerType::get(getLLVMContext(), 64)); 5521 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5522 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5523 Ops, ""); 5524 } 5525 case NEON::BI__builtin_neon_vst3_v: 5526 case NEON::BI__builtin_neon_vst3q_v: { 5527 Ops.push_back(Ops[0]); 5528 Ops.erase(Ops.begin()); 5529 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5530 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5531 Ops, ""); 5532 } 5533 case NEON::BI__builtin_neon_vst3_lane_v: 5534 case NEON::BI__builtin_neon_vst3q_lane_v: { 5535 Ops.push_back(Ops[0]); 5536 Ops.erase(Ops.begin()); 5537 Ops[3] = Builder.CreateZExt(Ops[3], 5538 llvm::IntegerType::get(getLLVMContext(), 64)); 5539 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5540 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5541 Ops, ""); 5542 } 5543 case NEON::BI__builtin_neon_vst4_v: 5544 case NEON::BI__builtin_neon_vst4q_v: { 5545 Ops.push_back(Ops[0]); 5546 Ops.erase(Ops.begin()); 5547 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5548 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5549 Ops, ""); 5550 } 5551 case NEON::BI__builtin_neon_vst4_lane_v: 5552 case NEON::BI__builtin_neon_vst4q_lane_v: { 5553 Ops.push_back(Ops[0]); 5554 Ops.erase(Ops.begin()); 5555 Ops[4] = Builder.CreateZExt(Ops[4], 5556 llvm::IntegerType::get(getLLVMContext(), 64)); 5557 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5558 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5559 Ops, ""); 5560 } 5561 case NEON::BI__builtin_neon_vtrn_v: 5562 case NEON::BI__builtin_neon_vtrnq_v: { 5563 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5564 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5565 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5566 Value *SV = nullptr; 5567 5568 for (unsigned vi = 0; vi != 2; ++vi) { 5569 SmallVector<Constant*, 16> Indices; 5570 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5571 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5572 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5573 } 5574 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5575 SV = llvm::ConstantVector::get(Indices); 5576 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5577 SV = Builder.CreateStore(SV, Addr); 5578 } 5579 return SV; 5580 } 5581 case NEON::BI__builtin_neon_vuzp_v: 5582 case NEON::BI__builtin_neon_vuzpq_v: { 5583 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5584 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5585 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5586 Value *SV = nullptr; 5587 5588 for (unsigned vi = 0; vi != 2; ++vi) { 5589 SmallVector<Constant*, 16> Indices; 5590 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5591 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5592 5593 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5594 SV = llvm::ConstantVector::get(Indices); 5595 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5596 SV = Builder.CreateStore(SV, Addr); 5597 } 5598 return SV; 5599 } 5600 case NEON::BI__builtin_neon_vzip_v: 5601 case NEON::BI__builtin_neon_vzipq_v: { 5602 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5603 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5604 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5605 Value *SV = nullptr; 5606 5607 for (unsigned vi = 0; vi != 2; ++vi) { 5608 SmallVector<Constant*, 16> Indices; 5609 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5610 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5611 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5612 } 5613 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5614 SV = llvm::ConstantVector::get(Indices); 5615 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5616 SV = Builder.CreateStore(SV, Addr); 5617 } 5618 return SV; 5619 } 5620 case NEON::BI__builtin_neon_vqtbl1q_v: { 5621 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5622 Ops, "vtbl1"); 5623 } 5624 case NEON::BI__builtin_neon_vqtbl2q_v: { 5625 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5626 Ops, "vtbl2"); 5627 } 5628 case NEON::BI__builtin_neon_vqtbl3q_v: { 5629 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5630 Ops, "vtbl3"); 5631 } 5632 case NEON::BI__builtin_neon_vqtbl4q_v: { 5633 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5634 Ops, "vtbl4"); 5635 } 5636 case NEON::BI__builtin_neon_vqtbx1q_v: { 5637 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5638 Ops, "vtbx1"); 5639 } 5640 case NEON::BI__builtin_neon_vqtbx2q_v: { 5641 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5642 Ops, "vtbx2"); 5643 } 5644 case NEON::BI__builtin_neon_vqtbx3q_v: { 5645 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5646 Ops, "vtbx3"); 5647 } 5648 case NEON::BI__builtin_neon_vqtbx4q_v: { 5649 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5650 Ops, "vtbx4"); 5651 } 5652 case NEON::BI__builtin_neon_vsqadd_v: 5653 case NEON::BI__builtin_neon_vsqaddq_v: { 5654 Int = Intrinsic::aarch64_neon_usqadd; 5655 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5656 } 5657 case NEON::BI__builtin_neon_vuqadd_v: 5658 case NEON::BI__builtin_neon_vuqaddq_v: { 5659 Int = Intrinsic::aarch64_neon_suqadd; 5660 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5661 } 5662 } 5663 } 5664 5665 llvm::Value *CodeGenFunction:: 5666 BuildVector(ArrayRef<llvm::Value*> Ops) { 5667 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5668 "Not a power-of-two sized vector!"); 5669 bool AllConstants = true; 5670 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5671 AllConstants &= isa<Constant>(Ops[i]); 5672 5673 // If this is a constant vector, create a ConstantVector. 5674 if (AllConstants) { 5675 SmallVector<llvm::Constant*, 16> CstOps; 5676 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5677 CstOps.push_back(cast<Constant>(Ops[i])); 5678 return llvm::ConstantVector::get(CstOps); 5679 } 5680 5681 // Otherwise, insertelement the values to build the vector. 5682 Value *Result = 5683 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5684 5685 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5686 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5687 5688 return Result; 5689 } 5690 5691 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5692 const CallExpr *E) { 5693 SmallVector<Value*, 4> Ops; 5694 5695 // Find out if any arguments are required to be integer constant expressions. 5696 unsigned ICEArguments = 0; 5697 ASTContext::GetBuiltinTypeError Error; 5698 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5699 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5700 5701 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5702 // If this is a normal argument, just emit it as a scalar. 5703 if ((ICEArguments & (1 << i)) == 0) { 5704 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5705 continue; 5706 } 5707 5708 // If this is required to be a constant, constant fold it so that we know 5709 // that the generated intrinsic gets a ConstantInt. 5710 llvm::APSInt Result; 5711 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5712 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5713 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5714 } 5715 5716 switch (BuiltinID) { 5717 default: return nullptr; 5718 case X86::BI_mm_prefetch: { 5719 Value *Address = EmitScalarExpr(E->getArg(0)); 5720 Value *RW = ConstantInt::get(Int32Ty, 0); 5721 Value *Locality = EmitScalarExpr(E->getArg(1)); 5722 Value *Data = ConstantInt::get(Int32Ty, 1); 5723 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5724 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5725 } 5726 case X86::BI__builtin_ia32_vec_init_v8qi: 5727 case X86::BI__builtin_ia32_vec_init_v4hi: 5728 case X86::BI__builtin_ia32_vec_init_v2si: 5729 return Builder.CreateBitCast(BuildVector(Ops), 5730 llvm::Type::getX86_MMXTy(getLLVMContext())); 5731 case X86::BI__builtin_ia32_vec_ext_v2si: 5732 return Builder.CreateExtractElement(Ops[0], 5733 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5734 case X86::BI__builtin_ia32_ldmxcsr: { 5735 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5736 Builder.CreateStore(Ops[0], Tmp); 5737 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5738 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5739 } 5740 case X86::BI__builtin_ia32_stmxcsr: { 5741 Value *Tmp = CreateMemTemp(E->getType()); 5742 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5743 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5744 return Builder.CreateLoad(Tmp, "stmxcsr"); 5745 } 5746 case X86::BI__builtin_ia32_storehps: 5747 case X86::BI__builtin_ia32_storelps: { 5748 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5749 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5750 5751 // cast val v2i64 5752 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5753 5754 // extract (0, 1) 5755 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5756 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5757 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5758 5759 // cast pointer to i64 & store 5760 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5761 return Builder.CreateStore(Ops[1], Ops[0]); 5762 } 5763 case X86::BI__builtin_ia32_palignr: { 5764 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5765 5766 // If palignr is shifting the pair of input vectors less than 9 bytes, 5767 // emit a shuffle instruction. 5768 if (shiftVal <= 8) { 5769 SmallVector<llvm::Constant*, 8> Indices; 5770 for (unsigned i = 0; i != 8; ++i) 5771 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5772 5773 Value* SV = llvm::ConstantVector::get(Indices); 5774 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5775 } 5776 5777 // If palignr is shifting the pair of input vectors more than 8 but less 5778 // than 16 bytes, emit a logical right shift of the destination. 5779 if (shiftVal < 16) { 5780 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 5781 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 5782 5783 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5784 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 5785 5786 // create i32 constant 5787 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 5788 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5789 } 5790 5791 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 5792 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5793 } 5794 case X86::BI__builtin_ia32_palignr128: { 5795 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5796 5797 // If palignr is shifting the pair of input vectors less than 17 bytes, 5798 // emit a shuffle instruction. 5799 if (shiftVal <= 16) { 5800 SmallVector<llvm::Constant*, 16> Indices; 5801 for (unsigned i = 0; i != 16; ++i) 5802 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5803 5804 Value* SV = llvm::ConstantVector::get(Indices); 5805 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5806 } 5807 5808 // If palignr is shifting the pair of input vectors more than 16 but less 5809 // than 32 bytes, emit a logical right shift of the destination. 5810 if (shiftVal < 32) { 5811 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5812 5813 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5814 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5815 5816 // create i32 constant 5817 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 5818 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5819 } 5820 5821 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5822 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5823 } 5824 case X86::BI__builtin_ia32_palignr256: { 5825 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5826 5827 // If palignr is shifting the pair of input vectors less than 17 bytes, 5828 // emit a shuffle instruction. 5829 if (shiftVal <= 16) { 5830 SmallVector<llvm::Constant*, 32> Indices; 5831 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5832 for (unsigned l = 0; l != 2; ++l) { 5833 unsigned LaneStart = l * 16; 5834 unsigned LaneEnd = (l+1) * 16; 5835 for (unsigned i = 0; i != 16; ++i) { 5836 unsigned Idx = shiftVal + i + LaneStart; 5837 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 5838 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 5839 } 5840 } 5841 5842 Value* SV = llvm::ConstantVector::get(Indices); 5843 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5844 } 5845 5846 // If palignr is shifting the pair of input vectors more than 16 but less 5847 // than 32 bytes, emit a logical right shift of the destination. 5848 if (shiftVal < 32) { 5849 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 5850 5851 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5852 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5853 5854 // create i32 constant 5855 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 5856 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 5857 } 5858 5859 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5860 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5861 } 5862 case X86::BI__builtin_ia32_movntps: 5863 case X86::BI__builtin_ia32_movntps256: 5864 case X86::BI__builtin_ia32_movntpd: 5865 case X86::BI__builtin_ia32_movntpd256: 5866 case X86::BI__builtin_ia32_movntdq: 5867 case X86::BI__builtin_ia32_movntdq256: 5868 case X86::BI__builtin_ia32_movnti: 5869 case X86::BI__builtin_ia32_movnti64: { 5870 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 5871 Builder.getInt32(1)); 5872 5873 // Convert the type of the pointer to a pointer to the stored type. 5874 Value *BC = Builder.CreateBitCast(Ops[0], 5875 llvm::PointerType::getUnqual(Ops[1]->getType()), 5876 "cast"); 5877 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 5878 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 5879 5880 // If the operand is an integer, we can't assume alignment. Otherwise, 5881 // assume natural alignment. 5882 QualType ArgTy = E->getArg(1)->getType(); 5883 unsigned Align; 5884 if (ArgTy->isIntegerType()) 5885 Align = 1; 5886 else 5887 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 5888 SI->setAlignment(Align); 5889 return SI; 5890 } 5891 // 3DNow! 5892 case X86::BI__builtin_ia32_pswapdsf: 5893 case X86::BI__builtin_ia32_pswapdsi: { 5894 const char *name = nullptr; 5895 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5896 switch(BuiltinID) { 5897 default: llvm_unreachable("Unsupported intrinsic!"); 5898 case X86::BI__builtin_ia32_pswapdsf: 5899 case X86::BI__builtin_ia32_pswapdsi: 5900 name = "pswapd"; 5901 ID = Intrinsic::x86_3dnowa_pswapd; 5902 break; 5903 } 5904 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 5905 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 5906 llvm::Function *F = CGM.getIntrinsic(ID); 5907 return Builder.CreateCall(F, Ops, name); 5908 } 5909 case X86::BI__builtin_ia32_rdrand16_step: 5910 case X86::BI__builtin_ia32_rdrand32_step: 5911 case X86::BI__builtin_ia32_rdrand64_step: 5912 case X86::BI__builtin_ia32_rdseed16_step: 5913 case X86::BI__builtin_ia32_rdseed32_step: 5914 case X86::BI__builtin_ia32_rdseed64_step: { 5915 Intrinsic::ID ID; 5916 switch (BuiltinID) { 5917 default: llvm_unreachable("Unsupported intrinsic!"); 5918 case X86::BI__builtin_ia32_rdrand16_step: 5919 ID = Intrinsic::x86_rdrand_16; 5920 break; 5921 case X86::BI__builtin_ia32_rdrand32_step: 5922 ID = Intrinsic::x86_rdrand_32; 5923 break; 5924 case X86::BI__builtin_ia32_rdrand64_step: 5925 ID = Intrinsic::x86_rdrand_64; 5926 break; 5927 case X86::BI__builtin_ia32_rdseed16_step: 5928 ID = Intrinsic::x86_rdseed_16; 5929 break; 5930 case X86::BI__builtin_ia32_rdseed32_step: 5931 ID = Intrinsic::x86_rdseed_32; 5932 break; 5933 case X86::BI__builtin_ia32_rdseed64_step: 5934 ID = Intrinsic::x86_rdseed_64; 5935 break; 5936 } 5937 5938 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 5939 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 5940 return Builder.CreateExtractValue(Call, 1); 5941 } 5942 // AVX2 broadcast 5943 case X86::BI__builtin_ia32_vbroadcastsi256: { 5944 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 5945 Builder.CreateStore(Ops[0], VecTmp); 5946 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 5947 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 5948 } 5949 } 5950 } 5951 5952 5953 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 5954 const CallExpr *E) { 5955 SmallVector<Value*, 4> Ops; 5956 5957 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 5958 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5959 5960 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5961 5962 switch (BuiltinID) { 5963 default: return nullptr; 5964 5965 // vec_ld, vec_lvsl, vec_lvsr 5966 case PPC::BI__builtin_altivec_lvx: 5967 case PPC::BI__builtin_altivec_lvxl: 5968 case PPC::BI__builtin_altivec_lvebx: 5969 case PPC::BI__builtin_altivec_lvehx: 5970 case PPC::BI__builtin_altivec_lvewx: 5971 case PPC::BI__builtin_altivec_lvsl: 5972 case PPC::BI__builtin_altivec_lvsr: 5973 { 5974 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 5975 5976 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 5977 Ops.pop_back(); 5978 5979 switch (BuiltinID) { 5980 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 5981 case PPC::BI__builtin_altivec_lvx: 5982 ID = Intrinsic::ppc_altivec_lvx; 5983 break; 5984 case PPC::BI__builtin_altivec_lvxl: 5985 ID = Intrinsic::ppc_altivec_lvxl; 5986 break; 5987 case PPC::BI__builtin_altivec_lvebx: 5988 ID = Intrinsic::ppc_altivec_lvebx; 5989 break; 5990 case PPC::BI__builtin_altivec_lvehx: 5991 ID = Intrinsic::ppc_altivec_lvehx; 5992 break; 5993 case PPC::BI__builtin_altivec_lvewx: 5994 ID = Intrinsic::ppc_altivec_lvewx; 5995 break; 5996 case PPC::BI__builtin_altivec_lvsl: 5997 ID = Intrinsic::ppc_altivec_lvsl; 5998 break; 5999 case PPC::BI__builtin_altivec_lvsr: 6000 ID = Intrinsic::ppc_altivec_lvsr; 6001 break; 6002 } 6003 llvm::Function *F = CGM.getIntrinsic(ID); 6004 return Builder.CreateCall(F, Ops, ""); 6005 } 6006 6007 // vec_st 6008 case PPC::BI__builtin_altivec_stvx: 6009 case PPC::BI__builtin_altivec_stvxl: 6010 case PPC::BI__builtin_altivec_stvebx: 6011 case PPC::BI__builtin_altivec_stvehx: 6012 case PPC::BI__builtin_altivec_stvewx: 6013 { 6014 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6015 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6016 Ops.pop_back(); 6017 6018 switch (BuiltinID) { 6019 default: llvm_unreachable("Unsupported st intrinsic!"); 6020 case PPC::BI__builtin_altivec_stvx: 6021 ID = Intrinsic::ppc_altivec_stvx; 6022 break; 6023 case PPC::BI__builtin_altivec_stvxl: 6024 ID = Intrinsic::ppc_altivec_stvxl; 6025 break; 6026 case PPC::BI__builtin_altivec_stvebx: 6027 ID = Intrinsic::ppc_altivec_stvebx; 6028 break; 6029 case PPC::BI__builtin_altivec_stvehx: 6030 ID = Intrinsic::ppc_altivec_stvehx; 6031 break; 6032 case PPC::BI__builtin_altivec_stvewx: 6033 ID = Intrinsic::ppc_altivec_stvewx; 6034 break; 6035 } 6036 llvm::Function *F = CGM.getIntrinsic(ID); 6037 return Builder.CreateCall(F, Ops, ""); 6038 } 6039 } 6040 } 6041 6042 // Emit an intrinsic that has 1 float or double. 6043 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6044 const CallExpr *E, 6045 unsigned IntrinsicID) { 6046 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6047 6048 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6049 return CGF.Builder.CreateCall(F, Src0); 6050 } 6051 6052 // Emit an intrinsic that has 3 float or double operands. 6053 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6054 const CallExpr *E, 6055 unsigned IntrinsicID) { 6056 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6057 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6058 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6059 6060 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6061 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6062 } 6063 6064 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6065 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6066 const CallExpr *E, 6067 unsigned IntrinsicID) { 6068 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6069 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6070 6071 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6072 return CGF.Builder.CreateCall2(F, Src0, Src1); 6073 } 6074 6075 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6076 const CallExpr *E) { 6077 switch (BuiltinID) { 6078 case R600::BI__builtin_amdgpu_div_scale: 6079 case R600::BI__builtin_amdgpu_div_scalef: { 6080 // Translate from the intrinsics's struct return to the builtin's out 6081 // argument. 6082 6083 std::pair<llvm::Value *, unsigned> FlagOutPtr 6084 = EmitPointerWithAlignment(E->getArg(3)); 6085 6086 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6087 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6088 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6089 6090 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6091 X->getType()); 6092 6093 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6094 6095 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6096 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6097 6098 llvm::Type *RealFlagType 6099 = FlagOutPtr.first->getType()->getPointerElementType(); 6100 6101 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6102 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6103 FlagStore->setAlignment(FlagOutPtr.second); 6104 return Result; 6105 } 6106 case R600::BI__builtin_amdgpu_div_fmas: 6107 case R600::BI__builtin_amdgpu_div_fmasf: 6108 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fmas); 6109 case R600::BI__builtin_amdgpu_div_fixup: 6110 case R600::BI__builtin_amdgpu_div_fixupf: 6111 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6112 case R600::BI__builtin_amdgpu_trig_preop: 6113 case R600::BI__builtin_amdgpu_trig_preopf: 6114 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6115 case R600::BI__builtin_amdgpu_rcp: 6116 case R600::BI__builtin_amdgpu_rcpf: 6117 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6118 case R600::BI__builtin_amdgpu_rsq: 6119 case R600::BI__builtin_amdgpu_rsqf: 6120 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6121 case R600::BI__builtin_amdgpu_rsq_clamped: 6122 case R600::BI__builtin_amdgpu_rsq_clampedf: 6123 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6124 case R600::BI__builtin_amdgpu_ldexp: 6125 case R600::BI__builtin_amdgpu_ldexpf: 6126 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6127 default: 6128 return nullptr; 6129 } 6130 } 6131