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 "CGCXXABI.h" 16 #include "CGObjCRuntime.h" 17 #include "CodeGenModule.h" 18 #include "TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "clang/Basic/TargetInfo.h" 23 #include "clang/CodeGen/CGFunctionInfo.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include <sstream> 30 31 using namespace clang; 32 using namespace CodeGen; 33 using namespace llvm; 34 35 /// getBuiltinLibFunction - Given a builtin id for a function like 36 /// "__builtin_fabsf", return a Function* for "fabsf". 37 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 38 unsigned BuiltinID) { 39 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 40 41 // Get the name, skip over the __builtin_ prefix (if necessary). 42 StringRef Name; 43 GlobalDecl D(FD); 44 45 // If the builtin has been declared explicitly with an assembler label, 46 // use the mangled name. This differs from the plain label on platforms 47 // that prefix labels. 48 if (FD->hasAttr<AsmLabelAttr>()) 49 Name = getMangledName(D); 50 else 51 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 52 53 llvm::FunctionType *Ty = 54 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 55 56 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 57 } 58 59 /// Emit the conversions required to turn the given value into an 60 /// integer of the given size. 61 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 62 QualType T, llvm::IntegerType *IntType) { 63 V = CGF.EmitToMemory(V, T); 64 65 if (V->getType()->isPointerTy()) 66 return CGF.Builder.CreatePtrToInt(V, IntType); 67 68 assert(V->getType() == IntType); 69 return V; 70 } 71 72 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 73 QualType T, llvm::Type *ResultType) { 74 V = CGF.EmitFromMemory(V, T); 75 76 if (ResultType->isPointerTy()) 77 return CGF.Builder.CreateIntToPtr(V, ResultType); 78 79 assert(V->getType() == ResultType); 80 return V; 81 } 82 83 /// Utility to insert an atomic instruction based on Instrinsic::ID 84 /// and the expression node. 85 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 86 llvm::AtomicRMWInst::BinOp Kind, 87 const CallExpr *E) { 88 QualType T = E->getType(); 89 assert(E->getArg(0)->getType()->isPointerType()); 90 assert(CGF.getContext().hasSameUnqualifiedType(T, 91 E->getArg(0)->getType()->getPointeeType())); 92 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 93 94 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 95 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 96 97 llvm::IntegerType *IntType = 98 llvm::IntegerType::get(CGF.getLLVMContext(), 99 CGF.getContext().getTypeSize(T)); 100 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 101 102 llvm::Value *Args[2]; 103 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 104 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 105 llvm::Type *ValueType = Args[1]->getType(); 106 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 107 108 llvm::Value *Result = 109 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 110 llvm::SequentiallyConsistent); 111 Result = EmitFromInt(CGF, Result, T, ValueType); 112 return RValue::get(Result); 113 } 114 115 /// Utility to insert an atomic instruction based Instrinsic::ID and 116 /// the expression node, where the return value is the result of the 117 /// operation. 118 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 119 llvm::AtomicRMWInst::BinOp Kind, 120 const CallExpr *E, 121 Instruction::BinaryOps Op, 122 bool Invert = false) { 123 QualType T = E->getType(); 124 assert(E->getArg(0)->getType()->isPointerType()); 125 assert(CGF.getContext().hasSameUnqualifiedType(T, 126 E->getArg(0)->getType()->getPointeeType())); 127 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 128 129 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 130 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 131 132 llvm::IntegerType *IntType = 133 llvm::IntegerType::get(CGF.getLLVMContext(), 134 CGF.getContext().getTypeSize(T)); 135 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 136 137 llvm::Value *Args[2]; 138 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 139 llvm::Type *ValueType = Args[1]->getType(); 140 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 141 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 142 143 llvm::Value *Result = 144 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 145 llvm::SequentiallyConsistent); 146 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 147 if (Invert) 148 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 149 llvm::ConstantInt::get(IntType, -1)); 150 Result = EmitFromInt(CGF, Result, T, ValueType); 151 return RValue::get(Result); 152 } 153 154 /// EmitFAbs - Emit a call to @llvm.fabs(). 155 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 156 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 157 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 158 Call->setDoesNotAccessMemory(); 159 return Call; 160 } 161 162 /// Emit the computation of the sign bit for a floating point value. Returns 163 /// the i1 sign bit value. 164 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 165 LLVMContext &C = CGF.CGM.getLLVMContext(); 166 167 llvm::Type *Ty = V->getType(); 168 int Width = Ty->getPrimitiveSizeInBits(); 169 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 170 V = CGF.Builder.CreateBitCast(V, IntTy); 171 if (Ty->isPPC_FP128Ty()) { 172 // The higher-order double comes first, and so we need to truncate the 173 // pair to extract the overall sign. The order of the pair is the same 174 // in both little- and big-Endian modes. 175 Width >>= 1; 176 IntTy = llvm::IntegerType::get(C, Width); 177 V = CGF.Builder.CreateTrunc(V, IntTy); 178 } 179 Value *Zero = llvm::Constant::getNullValue(IntTy); 180 return CGF.Builder.CreateICmpSLT(V, Zero); 181 } 182 183 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 184 const CallExpr *E, llvm::Value *calleeValue) { 185 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 186 ReturnValueSlot(), Fn); 187 } 188 189 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 190 /// depending on IntrinsicID. 191 /// 192 /// \arg CGF The current codegen function. 193 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 194 /// \arg X The first argument to the llvm.*.with.overflow.*. 195 /// \arg Y The second argument to the llvm.*.with.overflow.*. 196 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 197 /// \returns The result (i.e. sum/product) returned by the intrinsic. 198 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 199 const llvm::Intrinsic::ID IntrinsicID, 200 llvm::Value *X, llvm::Value *Y, 201 llvm::Value *&Carry) { 202 // Make sure we have integers of the same width. 203 assert(X->getType() == Y->getType() && 204 "Arguments must be the same type. (Did you forget to make sure both " 205 "arguments have the same integer width?)"); 206 207 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 208 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 209 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 210 return CGF.Builder.CreateExtractValue(Tmp, 0); 211 } 212 213 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 214 unsigned BuiltinID, const CallExpr *E, 215 ReturnValueSlot ReturnValue) { 216 // See if we can constant fold this builtin. If so, don't emit it at all. 217 Expr::EvalResult Result; 218 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 219 !Result.hasSideEffects()) { 220 if (Result.Val.isInt()) 221 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 222 Result.Val.getInt())); 223 if (Result.Val.isFloat()) 224 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 225 Result.Val.getFloat())); 226 } 227 228 switch (BuiltinID) { 229 default: break; // Handle intrinsics and libm functions below. 230 case Builtin::BI__builtin___CFStringMakeConstantString: 231 case Builtin::BI__builtin___NSStringMakeConstantString: 232 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 233 case Builtin::BI__builtin_stdarg_start: 234 case Builtin::BI__builtin_va_start: 235 case Builtin::BI__va_start: 236 case Builtin::BI__builtin_va_end: { 237 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 238 ? EmitScalarExpr(E->getArg(0)) 239 : EmitVAListRef(E->getArg(0)); 240 llvm::Type *DestType = Int8PtrTy; 241 if (ArgValue->getType() != DestType) 242 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 243 ArgValue->getName().data()); 244 245 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 246 Intrinsic::vaend : Intrinsic::vastart; 247 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 248 } 249 case Builtin::BI__builtin_va_copy: { 250 Value *DstPtr = EmitVAListRef(E->getArg(0)); 251 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 252 253 llvm::Type *Type = Int8PtrTy; 254 255 DstPtr = Builder.CreateBitCast(DstPtr, Type); 256 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 257 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 258 {DstPtr, SrcPtr})); 259 } 260 case Builtin::BI__builtin_abs: 261 case Builtin::BI__builtin_labs: 262 case Builtin::BI__builtin_llabs: { 263 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 264 265 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 266 Value *CmpResult = 267 Builder.CreateICmpSGE(ArgValue, 268 llvm::Constant::getNullValue(ArgValue->getType()), 269 "abscond"); 270 Value *Result = 271 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 272 273 return RValue::get(Result); 274 } 275 case Builtin::BI__builtin_fabs: 276 case Builtin::BI__builtin_fabsf: 277 case Builtin::BI__builtin_fabsl: { 278 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 279 Value *Result = EmitFAbs(*this, Arg1); 280 return RValue::get(Result); 281 } 282 case Builtin::BI__builtin_fmod: 283 case Builtin::BI__builtin_fmodf: 284 case Builtin::BI__builtin_fmodl: { 285 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 286 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 287 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 288 return RValue::get(Result); 289 } 290 291 case Builtin::BI__builtin_conj: 292 case Builtin::BI__builtin_conjf: 293 case Builtin::BI__builtin_conjl: { 294 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 295 Value *Real = ComplexVal.first; 296 Value *Imag = ComplexVal.second; 297 Value *Zero = 298 Imag->getType()->isFPOrFPVectorTy() 299 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 300 : llvm::Constant::getNullValue(Imag->getType()); 301 302 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 303 return RValue::getComplex(std::make_pair(Real, Imag)); 304 } 305 case Builtin::BI__builtin_creal: 306 case Builtin::BI__builtin_crealf: 307 case Builtin::BI__builtin_creall: 308 case Builtin::BIcreal: 309 case Builtin::BIcrealf: 310 case Builtin::BIcreall: { 311 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 312 return RValue::get(ComplexVal.first); 313 } 314 315 case Builtin::BI__builtin_cimag: 316 case Builtin::BI__builtin_cimagf: 317 case Builtin::BI__builtin_cimagl: 318 case Builtin::BIcimag: 319 case Builtin::BIcimagf: 320 case Builtin::BIcimagl: { 321 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 322 return RValue::get(ComplexVal.second); 323 } 324 325 case Builtin::BI__builtin_ctzs: 326 case Builtin::BI__builtin_ctz: 327 case Builtin::BI__builtin_ctzl: 328 case Builtin::BI__builtin_ctzll: { 329 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 330 331 llvm::Type *ArgType = ArgValue->getType(); 332 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 333 334 llvm::Type *ResultType = ConvertType(E->getType()); 335 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 336 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 337 if (Result->getType() != ResultType) 338 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 339 "cast"); 340 return RValue::get(Result); 341 } 342 case Builtin::BI__builtin_clzs: 343 case Builtin::BI__builtin_clz: 344 case Builtin::BI__builtin_clzl: 345 case Builtin::BI__builtin_clzll: { 346 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 347 348 llvm::Type *ArgType = ArgValue->getType(); 349 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 350 351 llvm::Type *ResultType = ConvertType(E->getType()); 352 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 353 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 354 if (Result->getType() != ResultType) 355 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 356 "cast"); 357 return RValue::get(Result); 358 } 359 case Builtin::BI__builtin_ffs: 360 case Builtin::BI__builtin_ffsl: 361 case Builtin::BI__builtin_ffsll: { 362 // ffs(x) -> x ? cttz(x) + 1 : 0 363 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 364 365 llvm::Type *ArgType = ArgValue->getType(); 366 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 367 368 llvm::Type *ResultType = ConvertType(E->getType()); 369 Value *Tmp = 370 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 371 llvm::ConstantInt::get(ArgType, 1)); 372 Value *Zero = llvm::Constant::getNullValue(ArgType); 373 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 374 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 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_parity: 381 case Builtin::BI__builtin_parityl: 382 case Builtin::BI__builtin_parityll: { 383 // parity(x) -> ctpop(x) & 1 384 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 385 386 llvm::Type *ArgType = ArgValue->getType(); 387 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 388 389 llvm::Type *ResultType = ConvertType(E->getType()); 390 Value *Tmp = Builder.CreateCall(F, ArgValue); 391 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 392 if (Result->getType() != ResultType) 393 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 394 "cast"); 395 return RValue::get(Result); 396 } 397 case Builtin::BI__builtin_popcount: 398 case Builtin::BI__builtin_popcountl: 399 case Builtin::BI__builtin_popcountll: { 400 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 401 402 llvm::Type *ArgType = ArgValue->getType(); 403 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 404 405 llvm::Type *ResultType = ConvertType(E->getType()); 406 Value *Result = Builder.CreateCall(F, ArgValue); 407 if (Result->getType() != ResultType) 408 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 409 "cast"); 410 return RValue::get(Result); 411 } 412 case Builtin::BI__builtin_expect: { 413 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 414 llvm::Type *ArgType = ArgValue->getType(); 415 416 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 417 // Don't generate llvm.expect on -O0 as the backend won't use it for 418 // anything. 419 // Note, we still IRGen ExpectedValue because it could have side-effects. 420 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 421 return RValue::get(ArgValue); 422 423 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 424 Value *Result = 425 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 426 return RValue::get(Result); 427 } 428 case Builtin::BI__builtin_assume_aligned: { 429 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 430 Value *OffsetValue = 431 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 432 433 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 434 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 435 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 436 437 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 438 return RValue::get(PtrValue); 439 } 440 case Builtin::BI__assume: 441 case Builtin::BI__builtin_assume: { 442 if (E->getArg(0)->HasSideEffects(getContext())) 443 return RValue::get(nullptr); 444 445 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 446 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 447 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 448 } 449 case Builtin::BI__builtin_bswap16: 450 case Builtin::BI__builtin_bswap32: 451 case Builtin::BI__builtin_bswap64: { 452 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 453 llvm::Type *ArgType = ArgValue->getType(); 454 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 455 return RValue::get(Builder.CreateCall(F, ArgValue)); 456 } 457 case Builtin::BI__builtin_object_size: { 458 // We rely on constant folding to deal with expressions with side effects. 459 assert(!E->getArg(0)->HasSideEffects(getContext()) && 460 "should have been constant folded"); 461 462 // We pass this builtin onto the optimizer so that it can 463 // figure out the object size in more complex cases. 464 llvm::Type *ResType = ConvertType(E->getType()); 465 466 // LLVM only supports 0 and 2, make sure that we pass along that 467 // as a boolean. 468 Value *Ty = EmitScalarExpr(E->getArg(1)); 469 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 470 assert(CI); 471 uint64_t val = CI->getZExtValue(); 472 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 473 // FIXME: Get right address space. 474 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 475 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 476 return RValue::get( 477 Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0)), CI})); 478 } 479 case Builtin::BI__builtin_prefetch: { 480 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 481 // FIXME: Technically these constants should of type 'int', yes? 482 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 483 llvm::ConstantInt::get(Int32Ty, 0); 484 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 485 llvm::ConstantInt::get(Int32Ty, 3); 486 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 487 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 488 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 489 } 490 case Builtin::BI__builtin_readcyclecounter: { 491 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 492 return RValue::get(Builder.CreateCall(F, {})); 493 } 494 case Builtin::BI__builtin___clear_cache: { 495 Value *Begin = EmitScalarExpr(E->getArg(0)); 496 Value *End = EmitScalarExpr(E->getArg(1)); 497 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 498 return RValue::get(Builder.CreateCall(F, {Begin, End})); 499 } 500 case Builtin::BI__builtin_trap: { 501 Value *F = CGM.getIntrinsic(Intrinsic::trap); 502 return RValue::get(Builder.CreateCall(F, {})); 503 } 504 case Builtin::BI__debugbreak: { 505 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 506 return RValue::get(Builder.CreateCall(F, {})); 507 } 508 case Builtin::BI__builtin_unreachable: { 509 if (SanOpts.has(SanitizerKind::Unreachable)) { 510 SanitizerScope SanScope(this); 511 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 512 SanitizerKind::Unreachable), 513 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 514 None); 515 } else 516 Builder.CreateUnreachable(); 517 518 // We do need to preserve an insertion point. 519 EmitBlock(createBasicBlock("unreachable.cont")); 520 521 return RValue::get(nullptr); 522 } 523 524 case Builtin::BI__builtin_powi: 525 case Builtin::BI__builtin_powif: 526 case Builtin::BI__builtin_powil: { 527 Value *Base = EmitScalarExpr(E->getArg(0)); 528 Value *Exponent = EmitScalarExpr(E->getArg(1)); 529 llvm::Type *ArgType = Base->getType(); 530 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 531 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 532 } 533 534 case Builtin::BI__builtin_isgreater: 535 case Builtin::BI__builtin_isgreaterequal: 536 case Builtin::BI__builtin_isless: 537 case Builtin::BI__builtin_islessequal: 538 case Builtin::BI__builtin_islessgreater: 539 case Builtin::BI__builtin_isunordered: { 540 // Ordered comparisons: we know the arguments to these are matching scalar 541 // floating point values. 542 Value *LHS = EmitScalarExpr(E->getArg(0)); 543 Value *RHS = EmitScalarExpr(E->getArg(1)); 544 545 switch (BuiltinID) { 546 default: llvm_unreachable("Unknown ordered comparison"); 547 case Builtin::BI__builtin_isgreater: 548 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 549 break; 550 case Builtin::BI__builtin_isgreaterequal: 551 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 552 break; 553 case Builtin::BI__builtin_isless: 554 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 555 break; 556 case Builtin::BI__builtin_islessequal: 557 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 558 break; 559 case Builtin::BI__builtin_islessgreater: 560 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 561 break; 562 case Builtin::BI__builtin_isunordered: 563 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 564 break; 565 } 566 // ZExt bool to int type. 567 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 568 } 569 case Builtin::BI__builtin_isnan: { 570 Value *V = EmitScalarExpr(E->getArg(0)); 571 V = Builder.CreateFCmpUNO(V, V, "cmp"); 572 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 573 } 574 575 case Builtin::BI__builtin_isinf: { 576 // isinf(x) --> fabs(x) == infinity 577 Value *V = EmitScalarExpr(E->getArg(0)); 578 V = EmitFAbs(*this, V); 579 580 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 581 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 582 } 583 584 case Builtin::BI__builtin_isinf_sign: { 585 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 586 Value *Arg = EmitScalarExpr(E->getArg(0)); 587 Value *AbsArg = EmitFAbs(*this, Arg); 588 Value *IsInf = Builder.CreateFCmpOEQ( 589 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 590 Value *IsNeg = EmitSignBit(*this, Arg); 591 592 llvm::Type *IntTy = ConvertType(E->getType()); 593 Value *Zero = Constant::getNullValue(IntTy); 594 Value *One = ConstantInt::get(IntTy, 1); 595 Value *NegativeOne = ConstantInt::get(IntTy, -1); 596 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 597 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 598 return RValue::get(Result); 599 } 600 601 case Builtin::BI__builtin_isnormal: { 602 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 603 Value *V = EmitScalarExpr(E->getArg(0)); 604 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 605 606 Value *Abs = EmitFAbs(*this, V); 607 Value *IsLessThanInf = 608 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 609 APFloat Smallest = APFloat::getSmallestNormalized( 610 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 611 Value *IsNormal = 612 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 613 "isnormal"); 614 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 615 V = Builder.CreateAnd(V, IsNormal, "and"); 616 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 617 } 618 619 case Builtin::BI__builtin_isfinite: { 620 // isfinite(x) --> x == x && fabs(x) != infinity; 621 Value *V = EmitScalarExpr(E->getArg(0)); 622 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 623 624 Value *Abs = EmitFAbs(*this, V); 625 Value *IsNotInf = 626 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 627 628 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 629 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 630 } 631 632 case Builtin::BI__builtin_fpclassify: { 633 Value *V = EmitScalarExpr(E->getArg(5)); 634 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 635 636 // Create Result 637 BasicBlock *Begin = Builder.GetInsertBlock(); 638 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 639 Builder.SetInsertPoint(End); 640 PHINode *Result = 641 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 642 "fpclassify_result"); 643 644 // if (V==0) return FP_ZERO 645 Builder.SetInsertPoint(Begin); 646 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 647 "iszero"); 648 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 649 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 650 Builder.CreateCondBr(IsZero, End, NotZero); 651 Result->addIncoming(ZeroLiteral, Begin); 652 653 // if (V != V) return FP_NAN 654 Builder.SetInsertPoint(NotZero); 655 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 656 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 657 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 658 Builder.CreateCondBr(IsNan, End, NotNan); 659 Result->addIncoming(NanLiteral, NotZero); 660 661 // if (fabs(V) == infinity) return FP_INFINITY 662 Builder.SetInsertPoint(NotNan); 663 Value *VAbs = EmitFAbs(*this, V); 664 Value *IsInf = 665 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 666 "isinf"); 667 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 668 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 669 Builder.CreateCondBr(IsInf, End, NotInf); 670 Result->addIncoming(InfLiteral, NotNan); 671 672 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 673 Builder.SetInsertPoint(NotInf); 674 APFloat Smallest = APFloat::getSmallestNormalized( 675 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 676 Value *IsNormal = 677 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 678 "isnormal"); 679 Value *NormalResult = 680 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 681 EmitScalarExpr(E->getArg(3))); 682 Builder.CreateBr(End); 683 Result->addIncoming(NormalResult, NotInf); 684 685 // return Result 686 Builder.SetInsertPoint(End); 687 return RValue::get(Result); 688 } 689 690 case Builtin::BIalloca: 691 case Builtin::BI_alloca: 692 case Builtin::BI__builtin_alloca: { 693 Value *Size = EmitScalarExpr(E->getArg(0)); 694 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 695 } 696 case Builtin::BIbzero: 697 case Builtin::BI__builtin_bzero: { 698 std::pair<llvm::Value*, unsigned> Dest = 699 EmitPointerWithAlignment(E->getArg(0)); 700 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 701 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 702 E->getArg(0)->getExprLoc(), FD, 0); 703 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 704 Dest.second, false); 705 return RValue::get(Dest.first); 706 } 707 case Builtin::BImemcpy: 708 case Builtin::BI__builtin_memcpy: { 709 std::pair<llvm::Value*, unsigned> Dest = 710 EmitPointerWithAlignment(E->getArg(0)); 711 std::pair<llvm::Value*, unsigned> Src = 712 EmitPointerWithAlignment(E->getArg(1)); 713 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 714 unsigned Align = std::min(Dest.second, Src.second); 715 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 716 E->getArg(0)->getExprLoc(), FD, 0); 717 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 718 E->getArg(1)->getExprLoc(), FD, 1); 719 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 720 return RValue::get(Dest.first); 721 } 722 723 case Builtin::BI__builtin___memcpy_chk: { 724 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 725 llvm::APSInt Size, DstSize; 726 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 727 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 728 break; 729 if (Size.ugt(DstSize)) 730 break; 731 std::pair<llvm::Value*, unsigned> Dest = 732 EmitPointerWithAlignment(E->getArg(0)); 733 std::pair<llvm::Value*, unsigned> Src = 734 EmitPointerWithAlignment(E->getArg(1)); 735 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 736 unsigned Align = std::min(Dest.second, Src.second); 737 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 738 return RValue::get(Dest.first); 739 } 740 741 case Builtin::BI__builtin_objc_memmove_collectable: { 742 Value *Address = EmitScalarExpr(E->getArg(0)); 743 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 744 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 745 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 746 Address, SrcAddr, SizeVal); 747 return RValue::get(Address); 748 } 749 750 case Builtin::BI__builtin___memmove_chk: { 751 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 752 llvm::APSInt Size, DstSize; 753 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 754 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 755 break; 756 if (Size.ugt(DstSize)) 757 break; 758 std::pair<llvm::Value*, unsigned> Dest = 759 EmitPointerWithAlignment(E->getArg(0)); 760 std::pair<llvm::Value*, unsigned> Src = 761 EmitPointerWithAlignment(E->getArg(1)); 762 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 763 unsigned Align = std::min(Dest.second, Src.second); 764 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 765 return RValue::get(Dest.first); 766 } 767 768 case Builtin::BImemmove: 769 case Builtin::BI__builtin_memmove: { 770 std::pair<llvm::Value*, unsigned> Dest = 771 EmitPointerWithAlignment(E->getArg(0)); 772 std::pair<llvm::Value*, unsigned> Src = 773 EmitPointerWithAlignment(E->getArg(1)); 774 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 775 unsigned Align = std::min(Dest.second, Src.second); 776 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 777 E->getArg(0)->getExprLoc(), FD, 0); 778 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 779 E->getArg(1)->getExprLoc(), FD, 1); 780 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 781 return RValue::get(Dest.first); 782 } 783 case Builtin::BImemset: 784 case Builtin::BI__builtin_memset: { 785 std::pair<llvm::Value*, unsigned> Dest = 786 EmitPointerWithAlignment(E->getArg(0)); 787 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 788 Builder.getInt8Ty()); 789 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 790 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 791 E->getArg(0)->getExprLoc(), FD, 0); 792 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 793 return RValue::get(Dest.first); 794 } 795 case Builtin::BI__builtin___memset_chk: { 796 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 797 llvm::APSInt Size, DstSize; 798 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 799 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 800 break; 801 if (Size.ugt(DstSize)) 802 break; 803 std::pair<llvm::Value*, unsigned> Dest = 804 EmitPointerWithAlignment(E->getArg(0)); 805 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 806 Builder.getInt8Ty()); 807 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 808 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 809 return RValue::get(Dest.first); 810 } 811 case Builtin::BI__builtin_dwarf_cfa: { 812 // The offset in bytes from the first argument to the CFA. 813 // 814 // Why on earth is this in the frontend? Is there any reason at 815 // all that the backend can't reasonably determine this while 816 // lowering llvm.eh.dwarf.cfa()? 817 // 818 // TODO: If there's a satisfactory reason, add a target hook for 819 // this instead of hard-coding 0, which is correct for most targets. 820 int32_t Offset = 0; 821 822 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 823 return RValue::get(Builder.CreateCall(F, 824 llvm::ConstantInt::get(Int32Ty, Offset))); 825 } 826 case Builtin::BI__builtin_return_address: { 827 Value *Depth = EmitScalarExpr(E->getArg(0)); 828 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 829 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 830 return RValue::get(Builder.CreateCall(F, Depth)); 831 } 832 case Builtin::BI__builtin_frame_address: { 833 Value *Depth = EmitScalarExpr(E->getArg(0)); 834 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 835 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 836 return RValue::get(Builder.CreateCall(F, Depth)); 837 } 838 case Builtin::BI__builtin_extract_return_addr: { 839 Value *Address = EmitScalarExpr(E->getArg(0)); 840 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 841 return RValue::get(Result); 842 } 843 case Builtin::BI__builtin_frob_return_addr: { 844 Value *Address = EmitScalarExpr(E->getArg(0)); 845 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 846 return RValue::get(Result); 847 } 848 case Builtin::BI__builtin_dwarf_sp_column: { 849 llvm::IntegerType *Ty 850 = cast<llvm::IntegerType>(ConvertType(E->getType())); 851 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 852 if (Column == -1) { 853 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 854 return RValue::get(llvm::UndefValue::get(Ty)); 855 } 856 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 857 } 858 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 859 Value *Address = EmitScalarExpr(E->getArg(0)); 860 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 861 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 862 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 863 } 864 case Builtin::BI__builtin_eh_return: { 865 Value *Int = EmitScalarExpr(E->getArg(0)); 866 Value *Ptr = EmitScalarExpr(E->getArg(1)); 867 868 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 869 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 870 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 871 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 872 ? Intrinsic::eh_return_i32 873 : Intrinsic::eh_return_i64); 874 Builder.CreateCall(F, {Int, Ptr}); 875 Builder.CreateUnreachable(); 876 877 // We do need to preserve an insertion point. 878 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 879 880 return RValue::get(nullptr); 881 } 882 case Builtin::BI__builtin_unwind_init: { 883 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 884 return RValue::get(Builder.CreateCall(F, {})); 885 } 886 case Builtin::BI__builtin_extend_pointer: { 887 // Extends a pointer to the size of an _Unwind_Word, which is 888 // uint64_t on all platforms. Generally this gets poked into a 889 // register and eventually used as an address, so if the 890 // addressing registers are wider than pointers and the platform 891 // doesn't implicitly ignore high-order bits when doing 892 // addressing, we need to make sure we zext / sext based on 893 // the platform's expectations. 894 // 895 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 896 897 // Cast the pointer to intptr_t. 898 Value *Ptr = EmitScalarExpr(E->getArg(0)); 899 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 900 901 // If that's 64 bits, we're done. 902 if (IntPtrTy->getBitWidth() == 64) 903 return RValue::get(Result); 904 905 // Otherwise, ask the codegen data what to do. 906 if (getTargetHooks().extendPointerWithSExt()) 907 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 908 else 909 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 910 } 911 case Builtin::BI__builtin_setjmp: { 912 // Buffer is a void**. 913 Value *Buf = EmitScalarExpr(E->getArg(0)); 914 915 // Store the frame pointer to the setjmp buffer. 916 Value *FrameAddr = 917 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 918 ConstantInt::get(Int32Ty, 0)); 919 Builder.CreateStore(FrameAddr, Buf); 920 921 // Store the stack pointer to the setjmp buffer. 922 Value *StackAddr = 923 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave), {}); 924 Value *StackSaveSlot = 925 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 926 Builder.CreateStore(StackAddr, StackSaveSlot); 927 928 // Call LLVM's EH setjmp, which is lightweight. 929 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 930 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 931 return RValue::get(Builder.CreateCall(F, Buf)); 932 } 933 case Builtin::BI__builtin_longjmp: { 934 Value *Buf = EmitScalarExpr(E->getArg(0)); 935 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 936 937 // Call LLVM's EH longjmp, which is lightweight. 938 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 939 940 // longjmp doesn't return; mark this as unreachable. 941 Builder.CreateUnreachable(); 942 943 // We do need to preserve an insertion point. 944 EmitBlock(createBasicBlock("longjmp.cont")); 945 946 return RValue::get(nullptr); 947 } 948 case Builtin::BI__sync_fetch_and_add: 949 case Builtin::BI__sync_fetch_and_sub: 950 case Builtin::BI__sync_fetch_and_or: 951 case Builtin::BI__sync_fetch_and_and: 952 case Builtin::BI__sync_fetch_and_xor: 953 case Builtin::BI__sync_fetch_and_nand: 954 case Builtin::BI__sync_add_and_fetch: 955 case Builtin::BI__sync_sub_and_fetch: 956 case Builtin::BI__sync_and_and_fetch: 957 case Builtin::BI__sync_or_and_fetch: 958 case Builtin::BI__sync_xor_and_fetch: 959 case Builtin::BI__sync_nand_and_fetch: 960 case Builtin::BI__sync_val_compare_and_swap: 961 case Builtin::BI__sync_bool_compare_and_swap: 962 case Builtin::BI__sync_lock_test_and_set: 963 case Builtin::BI__sync_lock_release: 964 case Builtin::BI__sync_swap: 965 llvm_unreachable("Shouldn't make it through sema"); 966 case Builtin::BI__sync_fetch_and_add_1: 967 case Builtin::BI__sync_fetch_and_add_2: 968 case Builtin::BI__sync_fetch_and_add_4: 969 case Builtin::BI__sync_fetch_and_add_8: 970 case Builtin::BI__sync_fetch_and_add_16: 971 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 972 case Builtin::BI__sync_fetch_and_sub_1: 973 case Builtin::BI__sync_fetch_and_sub_2: 974 case Builtin::BI__sync_fetch_and_sub_4: 975 case Builtin::BI__sync_fetch_and_sub_8: 976 case Builtin::BI__sync_fetch_and_sub_16: 977 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 978 case Builtin::BI__sync_fetch_and_or_1: 979 case Builtin::BI__sync_fetch_and_or_2: 980 case Builtin::BI__sync_fetch_and_or_4: 981 case Builtin::BI__sync_fetch_and_or_8: 982 case Builtin::BI__sync_fetch_and_or_16: 983 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 984 case Builtin::BI__sync_fetch_and_and_1: 985 case Builtin::BI__sync_fetch_and_and_2: 986 case Builtin::BI__sync_fetch_and_and_4: 987 case Builtin::BI__sync_fetch_and_and_8: 988 case Builtin::BI__sync_fetch_and_and_16: 989 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 990 case Builtin::BI__sync_fetch_and_xor_1: 991 case Builtin::BI__sync_fetch_and_xor_2: 992 case Builtin::BI__sync_fetch_and_xor_4: 993 case Builtin::BI__sync_fetch_and_xor_8: 994 case Builtin::BI__sync_fetch_and_xor_16: 995 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 996 case Builtin::BI__sync_fetch_and_nand_1: 997 case Builtin::BI__sync_fetch_and_nand_2: 998 case Builtin::BI__sync_fetch_and_nand_4: 999 case Builtin::BI__sync_fetch_and_nand_8: 1000 case Builtin::BI__sync_fetch_and_nand_16: 1001 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1002 1003 // Clang extensions: not overloaded yet. 1004 case Builtin::BI__sync_fetch_and_min: 1005 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1006 case Builtin::BI__sync_fetch_and_max: 1007 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1008 case Builtin::BI__sync_fetch_and_umin: 1009 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1010 case Builtin::BI__sync_fetch_and_umax: 1011 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1012 1013 case Builtin::BI__sync_add_and_fetch_1: 1014 case Builtin::BI__sync_add_and_fetch_2: 1015 case Builtin::BI__sync_add_and_fetch_4: 1016 case Builtin::BI__sync_add_and_fetch_8: 1017 case Builtin::BI__sync_add_and_fetch_16: 1018 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1019 llvm::Instruction::Add); 1020 case Builtin::BI__sync_sub_and_fetch_1: 1021 case Builtin::BI__sync_sub_and_fetch_2: 1022 case Builtin::BI__sync_sub_and_fetch_4: 1023 case Builtin::BI__sync_sub_and_fetch_8: 1024 case Builtin::BI__sync_sub_and_fetch_16: 1025 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1026 llvm::Instruction::Sub); 1027 case Builtin::BI__sync_and_and_fetch_1: 1028 case Builtin::BI__sync_and_and_fetch_2: 1029 case Builtin::BI__sync_and_and_fetch_4: 1030 case Builtin::BI__sync_and_and_fetch_8: 1031 case Builtin::BI__sync_and_and_fetch_16: 1032 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1033 llvm::Instruction::And); 1034 case Builtin::BI__sync_or_and_fetch_1: 1035 case Builtin::BI__sync_or_and_fetch_2: 1036 case Builtin::BI__sync_or_and_fetch_4: 1037 case Builtin::BI__sync_or_and_fetch_8: 1038 case Builtin::BI__sync_or_and_fetch_16: 1039 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1040 llvm::Instruction::Or); 1041 case Builtin::BI__sync_xor_and_fetch_1: 1042 case Builtin::BI__sync_xor_and_fetch_2: 1043 case Builtin::BI__sync_xor_and_fetch_4: 1044 case Builtin::BI__sync_xor_and_fetch_8: 1045 case Builtin::BI__sync_xor_and_fetch_16: 1046 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1047 llvm::Instruction::Xor); 1048 case Builtin::BI__sync_nand_and_fetch_1: 1049 case Builtin::BI__sync_nand_and_fetch_2: 1050 case Builtin::BI__sync_nand_and_fetch_4: 1051 case Builtin::BI__sync_nand_and_fetch_8: 1052 case Builtin::BI__sync_nand_and_fetch_16: 1053 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1054 llvm::Instruction::And, true); 1055 1056 case Builtin::BI__sync_val_compare_and_swap_1: 1057 case Builtin::BI__sync_val_compare_and_swap_2: 1058 case Builtin::BI__sync_val_compare_and_swap_4: 1059 case Builtin::BI__sync_val_compare_and_swap_8: 1060 case Builtin::BI__sync_val_compare_and_swap_16: { 1061 QualType T = E->getType(); 1062 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1063 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1064 1065 llvm::IntegerType *IntType = 1066 llvm::IntegerType::get(getLLVMContext(), 1067 getContext().getTypeSize(T)); 1068 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1069 1070 Value *Args[3]; 1071 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1072 Args[1] = EmitScalarExpr(E->getArg(1)); 1073 llvm::Type *ValueType = Args[1]->getType(); 1074 Args[1] = EmitToInt(*this, Args[1], T, IntType); 1075 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1076 1077 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1078 llvm::SequentiallyConsistent, 1079 llvm::SequentiallyConsistent); 1080 Result = Builder.CreateExtractValue(Result, 0); 1081 Result = EmitFromInt(*this, Result, T, ValueType); 1082 return RValue::get(Result); 1083 } 1084 1085 case Builtin::BI__sync_bool_compare_and_swap_1: 1086 case Builtin::BI__sync_bool_compare_and_swap_2: 1087 case Builtin::BI__sync_bool_compare_and_swap_4: 1088 case Builtin::BI__sync_bool_compare_and_swap_8: 1089 case Builtin::BI__sync_bool_compare_and_swap_16: { 1090 QualType T = E->getArg(1)->getType(); 1091 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1092 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1093 1094 llvm::IntegerType *IntType = 1095 llvm::IntegerType::get(getLLVMContext(), 1096 getContext().getTypeSize(T)); 1097 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1098 1099 Value *Args[3]; 1100 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1101 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1102 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1103 1104 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1105 llvm::SequentiallyConsistent, 1106 llvm::SequentiallyConsistent); 1107 Value *Result = Builder.CreateExtractValue(Pair, 1); 1108 // zext bool to int. 1109 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1110 return RValue::get(Result); 1111 } 1112 1113 case Builtin::BI__sync_swap_1: 1114 case Builtin::BI__sync_swap_2: 1115 case Builtin::BI__sync_swap_4: 1116 case Builtin::BI__sync_swap_8: 1117 case Builtin::BI__sync_swap_16: 1118 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1119 1120 case Builtin::BI__sync_lock_test_and_set_1: 1121 case Builtin::BI__sync_lock_test_and_set_2: 1122 case Builtin::BI__sync_lock_test_and_set_4: 1123 case Builtin::BI__sync_lock_test_and_set_8: 1124 case Builtin::BI__sync_lock_test_and_set_16: 1125 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1126 1127 case Builtin::BI__sync_lock_release_1: 1128 case Builtin::BI__sync_lock_release_2: 1129 case Builtin::BI__sync_lock_release_4: 1130 case Builtin::BI__sync_lock_release_8: 1131 case Builtin::BI__sync_lock_release_16: { 1132 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1133 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1134 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1135 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1136 StoreSize.getQuantity() * 8); 1137 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1138 llvm::StoreInst *Store = 1139 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1140 Store->setAlignment(StoreSize.getQuantity()); 1141 Store->setAtomic(llvm::Release); 1142 return RValue::get(nullptr); 1143 } 1144 1145 case Builtin::BI__sync_synchronize: { 1146 // We assume this is supposed to correspond to a C++0x-style 1147 // sequentially-consistent fence (i.e. this is only usable for 1148 // synchonization, not device I/O or anything like that). This intrinsic 1149 // is really badly designed in the sense that in theory, there isn't 1150 // any way to safely use it... but in practice, it mostly works 1151 // to use it with non-atomic loads and stores to get acquire/release 1152 // semantics. 1153 Builder.CreateFence(llvm::SequentiallyConsistent); 1154 return RValue::get(nullptr); 1155 } 1156 1157 case Builtin::BI__c11_atomic_is_lock_free: 1158 case Builtin::BI__atomic_is_lock_free: { 1159 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1160 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1161 // _Atomic(T) is always properly-aligned. 1162 const char *LibCallName = "__atomic_is_lock_free"; 1163 CallArgList Args; 1164 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1165 getContext().getSizeType()); 1166 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1167 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1168 getContext().VoidPtrTy); 1169 else 1170 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1171 getContext().VoidPtrTy); 1172 const CGFunctionInfo &FuncInfo = 1173 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1174 FunctionType::ExtInfo(), 1175 RequiredArgs::All); 1176 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1177 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1178 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1179 } 1180 1181 case Builtin::BI__atomic_test_and_set: { 1182 // Look at the argument type to determine whether this is a volatile 1183 // operation. The parameter type is always volatile. 1184 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1185 bool Volatile = 1186 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1187 1188 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1189 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1190 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1191 Value *NewVal = Builder.getInt8(1); 1192 Value *Order = EmitScalarExpr(E->getArg(1)); 1193 if (isa<llvm::ConstantInt>(Order)) { 1194 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1195 AtomicRMWInst *Result = nullptr; 1196 switch (ord) { 1197 case 0: // memory_order_relaxed 1198 default: // invalid order 1199 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1200 Ptr, NewVal, 1201 llvm::Monotonic); 1202 break; 1203 case 1: // memory_order_consume 1204 case 2: // memory_order_acquire 1205 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1206 Ptr, NewVal, 1207 llvm::Acquire); 1208 break; 1209 case 3: // memory_order_release 1210 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1211 Ptr, NewVal, 1212 llvm::Release); 1213 break; 1214 case 4: // memory_order_acq_rel 1215 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1216 Ptr, NewVal, 1217 llvm::AcquireRelease); 1218 break; 1219 case 5: // memory_order_seq_cst 1220 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1221 Ptr, NewVal, 1222 llvm::SequentiallyConsistent); 1223 break; 1224 } 1225 Result->setVolatile(Volatile); 1226 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1227 } 1228 1229 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1230 1231 llvm::BasicBlock *BBs[5] = { 1232 createBasicBlock("monotonic", CurFn), 1233 createBasicBlock("acquire", CurFn), 1234 createBasicBlock("release", CurFn), 1235 createBasicBlock("acqrel", CurFn), 1236 createBasicBlock("seqcst", CurFn) 1237 }; 1238 llvm::AtomicOrdering Orders[5] = { 1239 llvm::Monotonic, llvm::Acquire, llvm::Release, 1240 llvm::AcquireRelease, llvm::SequentiallyConsistent 1241 }; 1242 1243 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1244 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1245 1246 Builder.SetInsertPoint(ContBB); 1247 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1248 1249 for (unsigned i = 0; i < 5; ++i) { 1250 Builder.SetInsertPoint(BBs[i]); 1251 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1252 Ptr, NewVal, Orders[i]); 1253 RMW->setVolatile(Volatile); 1254 Result->addIncoming(RMW, BBs[i]); 1255 Builder.CreateBr(ContBB); 1256 } 1257 1258 SI->addCase(Builder.getInt32(0), BBs[0]); 1259 SI->addCase(Builder.getInt32(1), BBs[1]); 1260 SI->addCase(Builder.getInt32(2), BBs[1]); 1261 SI->addCase(Builder.getInt32(3), BBs[2]); 1262 SI->addCase(Builder.getInt32(4), BBs[3]); 1263 SI->addCase(Builder.getInt32(5), BBs[4]); 1264 1265 Builder.SetInsertPoint(ContBB); 1266 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1267 } 1268 1269 case Builtin::BI__atomic_clear: { 1270 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1271 bool Volatile = 1272 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1273 1274 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1275 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1276 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1277 Value *NewVal = Builder.getInt8(0); 1278 Value *Order = EmitScalarExpr(E->getArg(1)); 1279 if (isa<llvm::ConstantInt>(Order)) { 1280 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1281 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1282 Store->setAlignment(1); 1283 switch (ord) { 1284 case 0: // memory_order_relaxed 1285 default: // invalid order 1286 Store->setOrdering(llvm::Monotonic); 1287 break; 1288 case 3: // memory_order_release 1289 Store->setOrdering(llvm::Release); 1290 break; 1291 case 5: // memory_order_seq_cst 1292 Store->setOrdering(llvm::SequentiallyConsistent); 1293 break; 1294 } 1295 return RValue::get(nullptr); 1296 } 1297 1298 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1299 1300 llvm::BasicBlock *BBs[3] = { 1301 createBasicBlock("monotonic", CurFn), 1302 createBasicBlock("release", CurFn), 1303 createBasicBlock("seqcst", CurFn) 1304 }; 1305 llvm::AtomicOrdering Orders[3] = { 1306 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1307 }; 1308 1309 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1310 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1311 1312 for (unsigned i = 0; i < 3; ++i) { 1313 Builder.SetInsertPoint(BBs[i]); 1314 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1315 Store->setAlignment(1); 1316 Store->setOrdering(Orders[i]); 1317 Builder.CreateBr(ContBB); 1318 } 1319 1320 SI->addCase(Builder.getInt32(0), BBs[0]); 1321 SI->addCase(Builder.getInt32(3), BBs[1]); 1322 SI->addCase(Builder.getInt32(5), BBs[2]); 1323 1324 Builder.SetInsertPoint(ContBB); 1325 return RValue::get(nullptr); 1326 } 1327 1328 case Builtin::BI__atomic_thread_fence: 1329 case Builtin::BI__atomic_signal_fence: 1330 case Builtin::BI__c11_atomic_thread_fence: 1331 case Builtin::BI__c11_atomic_signal_fence: { 1332 llvm::SynchronizationScope Scope; 1333 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1334 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1335 Scope = llvm::SingleThread; 1336 else 1337 Scope = llvm::CrossThread; 1338 Value *Order = EmitScalarExpr(E->getArg(0)); 1339 if (isa<llvm::ConstantInt>(Order)) { 1340 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1341 switch (ord) { 1342 case 0: // memory_order_relaxed 1343 default: // invalid order 1344 break; 1345 case 1: // memory_order_consume 1346 case 2: // memory_order_acquire 1347 Builder.CreateFence(llvm::Acquire, Scope); 1348 break; 1349 case 3: // memory_order_release 1350 Builder.CreateFence(llvm::Release, Scope); 1351 break; 1352 case 4: // memory_order_acq_rel 1353 Builder.CreateFence(llvm::AcquireRelease, Scope); 1354 break; 1355 case 5: // memory_order_seq_cst 1356 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1357 break; 1358 } 1359 return RValue::get(nullptr); 1360 } 1361 1362 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1363 AcquireBB = createBasicBlock("acquire", CurFn); 1364 ReleaseBB = createBasicBlock("release", CurFn); 1365 AcqRelBB = createBasicBlock("acqrel", CurFn); 1366 SeqCstBB = createBasicBlock("seqcst", CurFn); 1367 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1368 1369 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1370 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1371 1372 Builder.SetInsertPoint(AcquireBB); 1373 Builder.CreateFence(llvm::Acquire, Scope); 1374 Builder.CreateBr(ContBB); 1375 SI->addCase(Builder.getInt32(1), AcquireBB); 1376 SI->addCase(Builder.getInt32(2), AcquireBB); 1377 1378 Builder.SetInsertPoint(ReleaseBB); 1379 Builder.CreateFence(llvm::Release, Scope); 1380 Builder.CreateBr(ContBB); 1381 SI->addCase(Builder.getInt32(3), ReleaseBB); 1382 1383 Builder.SetInsertPoint(AcqRelBB); 1384 Builder.CreateFence(llvm::AcquireRelease, Scope); 1385 Builder.CreateBr(ContBB); 1386 SI->addCase(Builder.getInt32(4), AcqRelBB); 1387 1388 Builder.SetInsertPoint(SeqCstBB); 1389 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1390 Builder.CreateBr(ContBB); 1391 SI->addCase(Builder.getInt32(5), SeqCstBB); 1392 1393 Builder.SetInsertPoint(ContBB); 1394 return RValue::get(nullptr); 1395 } 1396 1397 // Library functions with special handling. 1398 case Builtin::BIsqrt: 1399 case Builtin::BIsqrtf: 1400 case Builtin::BIsqrtl: { 1401 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1402 // in finite- or unsafe-math mode (the intrinsic has different semantics 1403 // for handling negative numbers compared to the library function, so 1404 // -fmath-errno=0 is not enough). 1405 if (!FD->hasAttr<ConstAttr>()) 1406 break; 1407 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1408 CGM.getCodeGenOpts().NoNaNsFPMath)) 1409 break; 1410 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1411 llvm::Type *ArgType = Arg0->getType(); 1412 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1413 return RValue::get(Builder.CreateCall(F, Arg0)); 1414 } 1415 1416 case Builtin::BI__builtin_pow: 1417 case Builtin::BI__builtin_powf: 1418 case Builtin::BI__builtin_powl: 1419 case Builtin::BIpow: 1420 case Builtin::BIpowf: 1421 case Builtin::BIpowl: { 1422 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1423 if (!FD->hasAttr<ConstAttr>()) 1424 break; 1425 Value *Base = EmitScalarExpr(E->getArg(0)); 1426 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1427 llvm::Type *ArgType = Base->getType(); 1428 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1429 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1430 } 1431 1432 case Builtin::BIfma: 1433 case Builtin::BIfmaf: 1434 case Builtin::BIfmal: 1435 case Builtin::BI__builtin_fma: 1436 case Builtin::BI__builtin_fmaf: 1437 case Builtin::BI__builtin_fmal: { 1438 // Rewrite fma to intrinsic. 1439 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1440 llvm::Type *ArgType = FirstArg->getType(); 1441 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1442 return RValue::get( 1443 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1444 EmitScalarExpr(E->getArg(2))})); 1445 } 1446 1447 case Builtin::BI__builtin_signbit: 1448 case Builtin::BI__builtin_signbitf: 1449 case Builtin::BI__builtin_signbitl: { 1450 return RValue::get( 1451 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1452 ConvertType(E->getType()))); 1453 } 1454 case Builtin::BI__builtin_annotation: { 1455 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1456 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1457 AnnVal->getType()); 1458 1459 // Get the annotation string, go through casts. Sema requires this to be a 1460 // non-wide string literal, potentially casted, so the cast<> is safe. 1461 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1462 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1463 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1464 } 1465 case Builtin::BI__builtin_addcb: 1466 case Builtin::BI__builtin_addcs: 1467 case Builtin::BI__builtin_addc: 1468 case Builtin::BI__builtin_addcl: 1469 case Builtin::BI__builtin_addcll: 1470 case Builtin::BI__builtin_subcb: 1471 case Builtin::BI__builtin_subcs: 1472 case Builtin::BI__builtin_subc: 1473 case Builtin::BI__builtin_subcl: 1474 case Builtin::BI__builtin_subcll: { 1475 1476 // We translate all of these builtins from expressions of the form: 1477 // int x = ..., y = ..., carryin = ..., carryout, result; 1478 // result = __builtin_addc(x, y, carryin, &carryout); 1479 // 1480 // to LLVM IR of the form: 1481 // 1482 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1483 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1484 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1485 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1486 // i32 %carryin) 1487 // %result = extractvalue {i32, i1} %tmp2, 0 1488 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1489 // %tmp3 = or i1 %carry1, %carry2 1490 // %tmp4 = zext i1 %tmp3 to i32 1491 // store i32 %tmp4, i32* %carryout 1492 1493 // Scalarize our inputs. 1494 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1495 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1496 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1497 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1498 EmitPointerWithAlignment(E->getArg(3)); 1499 1500 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1501 llvm::Intrinsic::ID IntrinsicId; 1502 switch (BuiltinID) { 1503 default: llvm_unreachable("Unknown multiprecision builtin id."); 1504 case Builtin::BI__builtin_addcb: 1505 case Builtin::BI__builtin_addcs: 1506 case Builtin::BI__builtin_addc: 1507 case Builtin::BI__builtin_addcl: 1508 case Builtin::BI__builtin_addcll: 1509 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1510 break; 1511 case Builtin::BI__builtin_subcb: 1512 case Builtin::BI__builtin_subcs: 1513 case Builtin::BI__builtin_subc: 1514 case Builtin::BI__builtin_subcl: 1515 case Builtin::BI__builtin_subcll: 1516 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1517 break; 1518 } 1519 1520 // Construct our resulting LLVM IR expression. 1521 llvm::Value *Carry1; 1522 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1523 X, Y, Carry1); 1524 llvm::Value *Carry2; 1525 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1526 Sum1, Carryin, Carry2); 1527 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1528 X->getType()); 1529 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1530 CarryOutPtr.first); 1531 CarryOutStore->setAlignment(CarryOutPtr.second); 1532 return RValue::get(Sum2); 1533 } 1534 case Builtin::BI__builtin_uadd_overflow: 1535 case Builtin::BI__builtin_uaddl_overflow: 1536 case Builtin::BI__builtin_uaddll_overflow: 1537 case Builtin::BI__builtin_usub_overflow: 1538 case Builtin::BI__builtin_usubl_overflow: 1539 case Builtin::BI__builtin_usubll_overflow: 1540 case Builtin::BI__builtin_umul_overflow: 1541 case Builtin::BI__builtin_umull_overflow: 1542 case Builtin::BI__builtin_umulll_overflow: 1543 case Builtin::BI__builtin_sadd_overflow: 1544 case Builtin::BI__builtin_saddl_overflow: 1545 case Builtin::BI__builtin_saddll_overflow: 1546 case Builtin::BI__builtin_ssub_overflow: 1547 case Builtin::BI__builtin_ssubl_overflow: 1548 case Builtin::BI__builtin_ssubll_overflow: 1549 case Builtin::BI__builtin_smul_overflow: 1550 case Builtin::BI__builtin_smull_overflow: 1551 case Builtin::BI__builtin_smulll_overflow: { 1552 1553 // We translate all of these builtins directly to the relevant llvm IR node. 1554 1555 // Scalarize our inputs. 1556 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1557 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1558 std::pair<llvm::Value *, unsigned> SumOutPtr = 1559 EmitPointerWithAlignment(E->getArg(2)); 1560 1561 // Decide which of the overflow intrinsics we are lowering to: 1562 llvm::Intrinsic::ID IntrinsicId; 1563 switch (BuiltinID) { 1564 default: llvm_unreachable("Unknown security overflow builtin id."); 1565 case Builtin::BI__builtin_uadd_overflow: 1566 case Builtin::BI__builtin_uaddl_overflow: 1567 case Builtin::BI__builtin_uaddll_overflow: 1568 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1569 break; 1570 case Builtin::BI__builtin_usub_overflow: 1571 case Builtin::BI__builtin_usubl_overflow: 1572 case Builtin::BI__builtin_usubll_overflow: 1573 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1574 break; 1575 case Builtin::BI__builtin_umul_overflow: 1576 case Builtin::BI__builtin_umull_overflow: 1577 case Builtin::BI__builtin_umulll_overflow: 1578 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1579 break; 1580 case Builtin::BI__builtin_sadd_overflow: 1581 case Builtin::BI__builtin_saddl_overflow: 1582 case Builtin::BI__builtin_saddll_overflow: 1583 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1584 break; 1585 case Builtin::BI__builtin_ssub_overflow: 1586 case Builtin::BI__builtin_ssubl_overflow: 1587 case Builtin::BI__builtin_ssubll_overflow: 1588 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1589 break; 1590 case Builtin::BI__builtin_smul_overflow: 1591 case Builtin::BI__builtin_smull_overflow: 1592 case Builtin::BI__builtin_smulll_overflow: 1593 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1594 break; 1595 } 1596 1597 1598 llvm::Value *Carry; 1599 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1600 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1601 SumOutStore->setAlignment(SumOutPtr.second); 1602 1603 return RValue::get(Carry); 1604 } 1605 case Builtin::BI__builtin_addressof: 1606 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1607 case Builtin::BI__builtin_operator_new: 1608 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1609 E->getArg(0), false); 1610 case Builtin::BI__builtin_operator_delete: 1611 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1612 E->getArg(0), true); 1613 case Builtin::BI__noop: 1614 // __noop always evaluates to an integer literal zero. 1615 return RValue::get(ConstantInt::get(IntTy, 0)); 1616 case Builtin::BI__builtin_call_with_static_chain: { 1617 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1618 const Expr *Chain = E->getArg(1); 1619 return EmitCall(Call->getCallee()->getType(), 1620 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1621 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1622 } 1623 case Builtin::BI_InterlockedExchange: 1624 case Builtin::BI_InterlockedExchangePointer: 1625 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1626 case Builtin::BI_InterlockedCompareExchangePointer: { 1627 llvm::Type *RTy; 1628 llvm::IntegerType *IntType = 1629 IntegerType::get(getLLVMContext(), 1630 getContext().getTypeSize(E->getType())); 1631 llvm::Type *IntPtrType = IntType->getPointerTo(); 1632 1633 llvm::Value *Destination = 1634 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1635 1636 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1637 RTy = Exchange->getType(); 1638 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1639 1640 llvm::Value *Comparand = 1641 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1642 1643 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1644 SequentiallyConsistent, 1645 SequentiallyConsistent); 1646 Result->setVolatile(true); 1647 1648 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1649 0), 1650 RTy)); 1651 } 1652 case Builtin::BI_InterlockedCompareExchange: { 1653 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1654 EmitScalarExpr(E->getArg(0)), 1655 EmitScalarExpr(E->getArg(2)), 1656 EmitScalarExpr(E->getArg(1)), 1657 SequentiallyConsistent, 1658 SequentiallyConsistent); 1659 CXI->setVolatile(true); 1660 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1661 } 1662 case Builtin::BI_InterlockedIncrement: { 1663 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1664 AtomicRMWInst::Add, 1665 EmitScalarExpr(E->getArg(0)), 1666 ConstantInt::get(Int32Ty, 1), 1667 llvm::SequentiallyConsistent); 1668 RMWI->setVolatile(true); 1669 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1670 } 1671 case Builtin::BI_InterlockedDecrement: { 1672 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1673 AtomicRMWInst::Sub, 1674 EmitScalarExpr(E->getArg(0)), 1675 ConstantInt::get(Int32Ty, 1), 1676 llvm::SequentiallyConsistent); 1677 RMWI->setVolatile(true); 1678 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1679 } 1680 case Builtin::BI_InterlockedExchangeAdd: { 1681 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1682 AtomicRMWInst::Add, 1683 EmitScalarExpr(E->getArg(0)), 1684 EmitScalarExpr(E->getArg(1)), 1685 llvm::SequentiallyConsistent); 1686 RMWI->setVolatile(true); 1687 return RValue::get(RMWI); 1688 } 1689 case Builtin::BI__readfsdword: { 1690 Value *IntToPtr = 1691 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1692 llvm::PointerType::get(CGM.Int32Ty, 257)); 1693 LoadInst *Load = 1694 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1695 return RValue::get(Load); 1696 } 1697 1698 case Builtin::BI__exception_code: 1699 case Builtin::BI_exception_code: 1700 return RValue::get(EmitSEHExceptionCode()); 1701 case Builtin::BI__exception_info: 1702 case Builtin::BI_exception_info: 1703 return RValue::get(EmitSEHExceptionInfo()); 1704 case Builtin::BI__abnormal_termination: 1705 case Builtin::BI_abnormal_termination: 1706 return RValue::get(EmitSEHAbnormalTermination()); 1707 case Builtin::BI_setjmpex: { 1708 if (getTarget().getTriple().isOSMSVCRT()) { 1709 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1710 llvm::AttributeSet ReturnsTwiceAttr = 1711 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1712 llvm::Attribute::ReturnsTwice); 1713 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1714 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1715 "_setjmpex", ReturnsTwiceAttr); 1716 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1717 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1718 llvm::Value *FrameAddr = 1719 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1720 ConstantInt::get(Int32Ty, 0)); 1721 llvm::Value *Args[] = {Buf, FrameAddr}; 1722 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1723 CS.setAttributes(ReturnsTwiceAttr); 1724 return RValue::get(CS.getInstruction()); 1725 } 1726 break; 1727 } 1728 case Builtin::BI_setjmp: { 1729 if (getTarget().getTriple().isOSMSVCRT()) { 1730 llvm::AttributeSet ReturnsTwiceAttr = 1731 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1732 llvm::Attribute::ReturnsTwice); 1733 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1734 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1735 llvm::CallSite CS; 1736 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1737 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1738 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1739 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1740 "_setjmp3", ReturnsTwiceAttr); 1741 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1742 llvm::Value *Args[] = {Buf, Count}; 1743 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1744 } else { 1745 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1746 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1747 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1748 "_setjmp", ReturnsTwiceAttr); 1749 llvm::Value *FrameAddr = 1750 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1751 ConstantInt::get(Int32Ty, 0)); 1752 llvm::Value *Args[] = {Buf, FrameAddr}; 1753 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1754 } 1755 CS.setAttributes(ReturnsTwiceAttr); 1756 return RValue::get(CS.getInstruction()); 1757 } 1758 break; 1759 } 1760 1761 case Builtin::BI__GetExceptionInfo: { 1762 if (llvm::GlobalVariable *GV = 1763 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1764 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1765 break; 1766 } 1767 } 1768 1769 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1770 // the call using the normal call path, but using the unmangled 1771 // version of the function name. 1772 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1773 return emitLibraryCall(*this, FD, E, 1774 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1775 1776 // If this is a predefined lib function (e.g. malloc), emit the call 1777 // using exactly the normal call path. 1778 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1779 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1780 1781 // See if we have a target specific intrinsic. 1782 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1783 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1784 if (const char *Prefix = 1785 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1786 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1787 // NOTE we dont need to perform a compatibility flag check here since the 1788 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1789 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1790 if (IntrinsicID == Intrinsic::not_intrinsic) 1791 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1792 } 1793 1794 if (IntrinsicID != Intrinsic::not_intrinsic) { 1795 SmallVector<Value*, 16> Args; 1796 1797 // Find out if any arguments are required to be integer constant 1798 // expressions. 1799 unsigned ICEArguments = 0; 1800 ASTContext::GetBuiltinTypeError Error; 1801 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1802 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1803 1804 Function *F = CGM.getIntrinsic(IntrinsicID); 1805 llvm::FunctionType *FTy = F->getFunctionType(); 1806 1807 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1808 Value *ArgValue; 1809 // If this is a normal argument, just emit it as a scalar. 1810 if ((ICEArguments & (1 << i)) == 0) { 1811 ArgValue = EmitScalarExpr(E->getArg(i)); 1812 } else { 1813 // If this is required to be a constant, constant fold it so that we 1814 // know that the generated intrinsic gets a ConstantInt. 1815 llvm::APSInt Result; 1816 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1817 assert(IsConst && "Constant arg isn't actually constant?"); 1818 (void)IsConst; 1819 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1820 } 1821 1822 // If the intrinsic arg type is different from the builtin arg type 1823 // we need to do a bit cast. 1824 llvm::Type *PTy = FTy->getParamType(i); 1825 if (PTy != ArgValue->getType()) { 1826 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1827 "Must be able to losslessly bit cast to param"); 1828 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1829 } 1830 1831 Args.push_back(ArgValue); 1832 } 1833 1834 Value *V = Builder.CreateCall(F, Args); 1835 QualType BuiltinRetType = E->getType(); 1836 1837 llvm::Type *RetTy = VoidTy; 1838 if (!BuiltinRetType->isVoidType()) 1839 RetTy = ConvertType(BuiltinRetType); 1840 1841 if (RetTy != V->getType()) { 1842 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1843 "Must be able to losslessly bit cast result type"); 1844 V = Builder.CreateBitCast(V, RetTy); 1845 } 1846 1847 return RValue::get(V); 1848 } 1849 1850 // See if we have a target specific builtin that needs to be lowered. 1851 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1852 return RValue::get(V); 1853 1854 ErrorUnsupported(E, "builtin function"); 1855 1856 // Unknown builtin, for now just dump it out and return undef. 1857 return GetUndefRValue(E->getType()); 1858 } 1859 1860 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1861 const CallExpr *E) { 1862 switch (getTarget().getTriple().getArch()) { 1863 case llvm::Triple::arm: 1864 case llvm::Triple::armeb: 1865 case llvm::Triple::thumb: 1866 case llvm::Triple::thumbeb: 1867 return EmitARMBuiltinExpr(BuiltinID, E); 1868 case llvm::Triple::aarch64: 1869 case llvm::Triple::aarch64_be: 1870 return EmitAArch64BuiltinExpr(BuiltinID, E); 1871 case llvm::Triple::x86: 1872 case llvm::Triple::x86_64: 1873 return EmitX86BuiltinExpr(BuiltinID, E); 1874 case llvm::Triple::ppc: 1875 case llvm::Triple::ppc64: 1876 case llvm::Triple::ppc64le: 1877 return EmitPPCBuiltinExpr(BuiltinID, E); 1878 case llvm::Triple::r600: 1879 case llvm::Triple::amdgcn: 1880 return EmitAMDGPUBuiltinExpr(BuiltinID, E); 1881 case llvm::Triple::systemz: 1882 return EmitSystemZBuiltinExpr(BuiltinID, E); 1883 default: 1884 return nullptr; 1885 } 1886 } 1887 1888 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1889 NeonTypeFlags TypeFlags, 1890 bool V1Ty=false) { 1891 int IsQuad = TypeFlags.isQuad(); 1892 switch (TypeFlags.getEltType()) { 1893 case NeonTypeFlags::Int8: 1894 case NeonTypeFlags::Poly8: 1895 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1896 case NeonTypeFlags::Int16: 1897 case NeonTypeFlags::Poly16: 1898 case NeonTypeFlags::Float16: 1899 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1900 case NeonTypeFlags::Int32: 1901 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1902 case NeonTypeFlags::Int64: 1903 case NeonTypeFlags::Poly64: 1904 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1905 case NeonTypeFlags::Poly128: 1906 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1907 // There is a lot of i128 and f128 API missing. 1908 // so we use v16i8 to represent poly128 and get pattern matched. 1909 return llvm::VectorType::get(CGF->Int8Ty, 16); 1910 case NeonTypeFlags::Float32: 1911 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1912 case NeonTypeFlags::Float64: 1913 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1914 } 1915 llvm_unreachable("Unknown vector element type!"); 1916 } 1917 1918 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1919 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1920 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1921 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1922 } 1923 1924 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1925 const char *name, 1926 unsigned shift, bool rightshift) { 1927 unsigned j = 0; 1928 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1929 ai != ae; ++ai, ++j) 1930 if (shift > 0 && shift == j) 1931 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1932 else 1933 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1934 1935 return Builder.CreateCall(F, Ops, name); 1936 } 1937 1938 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1939 bool neg) { 1940 int SV = cast<ConstantInt>(V)->getSExtValue(); 1941 1942 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1943 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1944 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1945 } 1946 1947 // \brief Right-shift a vector by a constant. 1948 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1949 llvm::Type *Ty, bool usgn, 1950 const char *name) { 1951 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1952 1953 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1954 int EltSize = VTy->getScalarSizeInBits(); 1955 1956 Vec = Builder.CreateBitCast(Vec, Ty); 1957 1958 // lshr/ashr are undefined when the shift amount is equal to the vector 1959 // element size. 1960 if (ShiftAmt == EltSize) { 1961 if (usgn) { 1962 // Right-shifting an unsigned value by its size yields 0. 1963 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1964 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1965 } else { 1966 // Right-shifting a signed value by its size is equivalent 1967 // to a shift of size-1. 1968 --ShiftAmt; 1969 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1970 } 1971 } 1972 1973 Shift = EmitNeonShiftVector(Shift, Ty, false); 1974 if (usgn) 1975 return Builder.CreateLShr(Vec, Shift, name); 1976 else 1977 return Builder.CreateAShr(Vec, Shift, name); 1978 } 1979 1980 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1981 /// alignment of the type referenced by the pointer. Skip over implicit 1982 /// casts. 1983 std::pair<llvm::Value*, unsigned> 1984 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1985 assert(Addr->getType()->isPointerType()); 1986 Addr = Addr->IgnoreParens(); 1987 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1988 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1989 ICE->getSubExpr()->getType()->isPointerType()) { 1990 std::pair<llvm::Value*, unsigned> Ptr = 1991 EmitPointerWithAlignment(ICE->getSubExpr()); 1992 Ptr.first = Builder.CreateBitCast(Ptr.first, 1993 ConvertType(Addr->getType())); 1994 return Ptr; 1995 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1996 LValue LV = EmitLValue(ICE->getSubExpr()); 1997 unsigned Align = LV.getAlignment().getQuantity(); 1998 if (!Align) { 1999 // FIXME: Once LValues are fixed to always set alignment, 2000 // zap this code. 2001 QualType PtTy = ICE->getSubExpr()->getType(); 2002 if (!PtTy->isIncompleteType()) 2003 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2004 else 2005 Align = 1; 2006 } 2007 return std::make_pair(LV.getAddress(), Align); 2008 } 2009 } 2010 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 2011 if (UO->getOpcode() == UO_AddrOf) { 2012 LValue LV = EmitLValue(UO->getSubExpr()); 2013 unsigned Align = LV.getAlignment().getQuantity(); 2014 if (!Align) { 2015 // FIXME: Once LValues are fixed to always set alignment, 2016 // zap this code. 2017 QualType PtTy = UO->getSubExpr()->getType(); 2018 if (!PtTy->isIncompleteType()) 2019 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2020 else 2021 Align = 1; 2022 } 2023 return std::make_pair(LV.getAddress(), Align); 2024 } 2025 } 2026 2027 unsigned Align = 1; 2028 QualType PtTy = Addr->getType()->getPointeeType(); 2029 if (!PtTy->isIncompleteType()) 2030 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2031 2032 return std::make_pair(EmitScalarExpr(Addr), Align); 2033 } 2034 2035 enum { 2036 AddRetType = (1 << 0), 2037 Add1ArgType = (1 << 1), 2038 Add2ArgTypes = (1 << 2), 2039 2040 VectorizeRetType = (1 << 3), 2041 VectorizeArgTypes = (1 << 4), 2042 2043 InventFloatType = (1 << 5), 2044 UnsignedAlts = (1 << 6), 2045 2046 Use64BitVectors = (1 << 7), 2047 Use128BitVectors = (1 << 8), 2048 2049 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2050 VectorRet = AddRetType | VectorizeRetType, 2051 VectorRetGetArgs01 = 2052 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2053 FpCmpzModifiers = 2054 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2055 }; 2056 2057 struct NeonIntrinsicInfo { 2058 unsigned BuiltinID; 2059 unsigned LLVMIntrinsic; 2060 unsigned AltLLVMIntrinsic; 2061 const char *NameHint; 2062 unsigned TypeModifier; 2063 2064 bool operator<(unsigned RHSBuiltinID) const { 2065 return BuiltinID < RHSBuiltinID; 2066 } 2067 }; 2068 2069 #define NEONMAP0(NameBase) \ 2070 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2071 2072 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2073 { NEON:: BI__builtin_neon_ ## NameBase, \ 2074 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2075 2076 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2077 { NEON:: BI__builtin_neon_ ## NameBase, \ 2078 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2079 #NameBase, TypeModifier } 2080 2081 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2082 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2083 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2084 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2085 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2086 NEONMAP0(vaddhn_v), 2087 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2088 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2089 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2090 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2091 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2092 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2093 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2094 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2095 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2096 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2097 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2098 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2099 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2100 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2101 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2102 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2103 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2104 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2105 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2106 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2107 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 2108 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2109 NEONMAP0(vcvt_f32_v), 2110 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2111 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2112 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2113 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2114 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2115 NEONMAP0(vcvt_s32_v), 2116 NEONMAP0(vcvt_s64_v), 2117 NEONMAP0(vcvt_u32_v), 2118 NEONMAP0(vcvt_u64_v), 2119 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2120 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2121 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2122 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2123 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2124 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2125 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2126 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2127 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2128 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2129 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2130 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2131 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2132 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2133 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2134 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2135 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2136 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2137 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2138 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2139 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2140 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2141 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2142 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2143 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2144 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2145 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2146 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2147 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2148 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2149 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2150 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2151 NEONMAP0(vcvtq_f32_v), 2152 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2153 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2154 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2155 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2156 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2157 NEONMAP0(vcvtq_s32_v), 2158 NEONMAP0(vcvtq_s64_v), 2159 NEONMAP0(vcvtq_u32_v), 2160 NEONMAP0(vcvtq_u64_v), 2161 NEONMAP0(vext_v), 2162 NEONMAP0(vextq_v), 2163 NEONMAP0(vfma_v), 2164 NEONMAP0(vfmaq_v), 2165 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2166 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2167 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2168 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2169 NEONMAP0(vld1_dup_v), 2170 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2171 NEONMAP0(vld1q_dup_v), 2172 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2173 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2174 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2175 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2176 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2177 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2178 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2179 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2180 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2181 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2182 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2183 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2184 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2185 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2186 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2187 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2188 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2189 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2190 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2191 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2192 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2193 NEONMAP0(vmovl_v), 2194 NEONMAP0(vmovn_v), 2195 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2196 NEONMAP0(vmull_v), 2197 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2198 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2199 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2200 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2201 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2202 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2203 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2204 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2205 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2206 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2207 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2208 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2209 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2210 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2211 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2212 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2213 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2214 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2215 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2216 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2217 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2218 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2219 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2220 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2221 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2222 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2223 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2224 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2225 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2226 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2227 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2228 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2229 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2230 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2231 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2232 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2233 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2234 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2235 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2236 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2237 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2238 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2239 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2240 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2241 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2242 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2243 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2244 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2245 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2246 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2247 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2248 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2249 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2250 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2251 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2252 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2253 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2254 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2255 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2256 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2257 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2258 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2259 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2260 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2261 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2262 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2263 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2264 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2265 NEONMAP0(vshl_n_v), 2266 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2267 NEONMAP0(vshll_n_v), 2268 NEONMAP0(vshlq_n_v), 2269 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2270 NEONMAP0(vshr_n_v), 2271 NEONMAP0(vshrn_n_v), 2272 NEONMAP0(vshrq_n_v), 2273 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2274 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2275 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2276 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2277 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2278 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2279 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2280 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2281 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2282 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2283 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2284 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2285 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2286 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2287 NEONMAP0(vsubhn_v), 2288 NEONMAP0(vtrn_v), 2289 NEONMAP0(vtrnq_v), 2290 NEONMAP0(vtst_v), 2291 NEONMAP0(vtstq_v), 2292 NEONMAP0(vuzp_v), 2293 NEONMAP0(vuzpq_v), 2294 NEONMAP0(vzip_v), 2295 NEONMAP0(vzipq_v) 2296 }; 2297 2298 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2299 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2300 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2301 NEONMAP0(vaddhn_v), 2302 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2303 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2304 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2305 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2306 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2307 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2308 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2309 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2310 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2311 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2312 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2313 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2314 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2315 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2316 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2317 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2318 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2319 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2320 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2321 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2322 NEONMAP0(vcvt_f32_v), 2323 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2324 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2325 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2326 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2327 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2328 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2329 NEONMAP0(vcvtq_f32_v), 2330 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2331 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2332 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2333 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2334 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2335 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2336 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2337 NEONMAP0(vext_v), 2338 NEONMAP0(vextq_v), 2339 NEONMAP0(vfma_v), 2340 NEONMAP0(vfmaq_v), 2341 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2342 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2343 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2344 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2345 NEONMAP0(vmovl_v), 2346 NEONMAP0(vmovn_v), 2347 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2348 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2349 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2350 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2351 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2352 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2353 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2354 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2355 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2356 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2357 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2358 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2359 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2360 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2361 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2362 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2363 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2364 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2365 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2366 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2367 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2368 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2369 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2370 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2371 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2372 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2373 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2374 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2375 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2376 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2377 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2378 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2379 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2380 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2381 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2382 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2383 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2384 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2385 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2386 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2387 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2388 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2389 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2390 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2391 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2392 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2393 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2394 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2395 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2396 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2397 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2398 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2399 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2400 NEONMAP0(vshl_n_v), 2401 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2402 NEONMAP0(vshll_n_v), 2403 NEONMAP0(vshlq_n_v), 2404 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2405 NEONMAP0(vshr_n_v), 2406 NEONMAP0(vshrn_n_v), 2407 NEONMAP0(vshrq_n_v), 2408 NEONMAP0(vsubhn_v), 2409 NEONMAP0(vtst_v), 2410 NEONMAP0(vtstq_v), 2411 }; 2412 2413 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2414 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2415 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2416 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2417 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2418 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2419 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2420 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2421 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2422 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2423 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2424 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2425 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2426 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2427 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2428 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2429 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2430 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2431 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2432 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2433 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2434 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2435 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2436 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2437 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2438 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2439 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2440 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2441 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2442 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2443 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2444 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2445 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2446 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2447 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2448 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2449 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2450 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2451 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2452 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2453 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2454 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2455 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2456 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2457 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2458 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2459 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2460 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2461 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2462 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2463 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2464 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2465 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2466 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2467 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2468 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2469 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2470 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2471 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2472 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2473 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2474 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2475 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2476 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2477 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2478 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2479 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2480 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2481 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2482 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2483 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2484 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2485 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2486 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2487 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2488 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2489 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2490 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2491 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2492 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2493 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2494 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2495 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2496 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2497 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2498 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2499 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2500 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2501 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2502 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2503 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2504 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2505 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2506 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2507 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2508 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2509 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2510 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2511 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2512 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2513 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2514 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2515 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2516 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2517 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2518 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2519 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2520 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2521 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2522 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2523 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2524 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2525 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2526 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2527 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2528 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2529 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2530 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2531 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2532 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2533 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2534 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2535 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2536 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2537 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2538 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2539 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2540 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2541 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2542 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2543 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2544 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2545 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2546 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2547 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2548 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2549 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2550 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2551 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2552 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2553 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2554 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2555 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2556 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2557 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2558 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2559 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2560 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2561 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2562 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2563 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2564 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2565 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2566 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2567 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2568 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2569 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2570 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2571 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2572 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2573 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2574 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2575 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2576 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2577 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2578 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2579 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2580 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2581 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2582 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2583 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2584 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2585 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2586 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2587 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2588 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2589 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2590 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2591 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2592 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2593 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2594 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2595 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2596 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2597 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2598 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2599 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2600 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2601 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2602 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2603 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2604 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2605 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2606 }; 2607 2608 #undef NEONMAP0 2609 #undef NEONMAP1 2610 #undef NEONMAP2 2611 2612 static bool NEONSIMDIntrinsicsProvenSorted = false; 2613 2614 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2615 static bool AArch64SISDIntrinsicsProvenSorted = false; 2616 2617 2618 static const NeonIntrinsicInfo * 2619 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2620 unsigned BuiltinID, bool &MapProvenSorted) { 2621 2622 #ifndef NDEBUG 2623 if (!MapProvenSorted) { 2624 // FIXME: use std::is_sorted once C++11 is allowed 2625 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2626 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2627 MapProvenSorted = true; 2628 } 2629 #endif 2630 2631 const NeonIntrinsicInfo *Builtin = 2632 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2633 2634 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2635 return Builtin; 2636 2637 return nullptr; 2638 } 2639 2640 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2641 unsigned Modifier, 2642 llvm::Type *ArgType, 2643 const CallExpr *E) { 2644 int VectorSize = 0; 2645 if (Modifier & Use64BitVectors) 2646 VectorSize = 64; 2647 else if (Modifier & Use128BitVectors) 2648 VectorSize = 128; 2649 2650 // Return type. 2651 SmallVector<llvm::Type *, 3> Tys; 2652 if (Modifier & AddRetType) { 2653 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2654 if (Modifier & VectorizeRetType) 2655 Ty = llvm::VectorType::get( 2656 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2657 2658 Tys.push_back(Ty); 2659 } 2660 2661 // Arguments. 2662 if (Modifier & VectorizeArgTypes) { 2663 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2664 ArgType = llvm::VectorType::get(ArgType, Elts); 2665 } 2666 2667 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2668 Tys.push_back(ArgType); 2669 2670 if (Modifier & Add2ArgTypes) 2671 Tys.push_back(ArgType); 2672 2673 if (Modifier & InventFloatType) 2674 Tys.push_back(FloatTy); 2675 2676 return CGM.getIntrinsic(IntrinsicID, Tys); 2677 } 2678 2679 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2680 const NeonIntrinsicInfo &SISDInfo, 2681 SmallVectorImpl<Value *> &Ops, 2682 const CallExpr *E) { 2683 unsigned BuiltinID = SISDInfo.BuiltinID; 2684 unsigned int Int = SISDInfo.LLVMIntrinsic; 2685 unsigned Modifier = SISDInfo.TypeModifier; 2686 const char *s = SISDInfo.NameHint; 2687 2688 switch (BuiltinID) { 2689 case NEON::BI__builtin_neon_vcled_s64: 2690 case NEON::BI__builtin_neon_vcled_u64: 2691 case NEON::BI__builtin_neon_vcles_f32: 2692 case NEON::BI__builtin_neon_vcled_f64: 2693 case NEON::BI__builtin_neon_vcltd_s64: 2694 case NEON::BI__builtin_neon_vcltd_u64: 2695 case NEON::BI__builtin_neon_vclts_f32: 2696 case NEON::BI__builtin_neon_vcltd_f64: 2697 case NEON::BI__builtin_neon_vcales_f32: 2698 case NEON::BI__builtin_neon_vcaled_f64: 2699 case NEON::BI__builtin_neon_vcalts_f32: 2700 case NEON::BI__builtin_neon_vcaltd_f64: 2701 // Only one direction of comparisons actually exist, cmle is actually a cmge 2702 // with swapped operands. The table gives us the right intrinsic but we 2703 // still need to do the swap. 2704 std::swap(Ops[0], Ops[1]); 2705 break; 2706 } 2707 2708 assert(Int && "Generic code assumes a valid intrinsic"); 2709 2710 // Determine the type(s) of this overloaded AArch64 intrinsic. 2711 const Expr *Arg = E->getArg(0); 2712 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2713 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2714 2715 int j = 0; 2716 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2717 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2718 ai != ae; ++ai, ++j) { 2719 llvm::Type *ArgTy = ai->getType(); 2720 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2721 ArgTy->getPrimitiveSizeInBits()) 2722 continue; 2723 2724 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2725 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2726 // it before inserting. 2727 Ops[j] = 2728 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2729 Ops[j] = 2730 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2731 } 2732 2733 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2734 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2735 if (ResultType->getPrimitiveSizeInBits() < 2736 Result->getType()->getPrimitiveSizeInBits()) 2737 return CGF.Builder.CreateExtractElement(Result, C0); 2738 2739 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2740 } 2741 2742 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2743 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2744 const char *NameHint, unsigned Modifier, const CallExpr *E, 2745 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2746 // Get the last argument, which specifies the vector type. 2747 llvm::APSInt NeonTypeConst; 2748 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2749 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2750 return nullptr; 2751 2752 // Determine the type of this overloaded NEON intrinsic. 2753 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2754 bool Usgn = Type.isUnsigned(); 2755 bool Quad = Type.isQuad(); 2756 2757 llvm::VectorType *VTy = GetNeonType(this, Type); 2758 llvm::Type *Ty = VTy; 2759 if (!Ty) 2760 return nullptr; 2761 2762 unsigned Int = LLVMIntrinsic; 2763 if ((Modifier & UnsignedAlts) && !Usgn) 2764 Int = AltLLVMIntrinsic; 2765 2766 switch (BuiltinID) { 2767 default: break; 2768 case NEON::BI__builtin_neon_vabs_v: 2769 case NEON::BI__builtin_neon_vabsq_v: 2770 if (VTy->getElementType()->isFloatingPointTy()) 2771 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2772 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2773 case NEON::BI__builtin_neon_vaddhn_v: { 2774 llvm::VectorType *SrcTy = 2775 llvm::VectorType::getExtendedElementVectorType(VTy); 2776 2777 // %sum = add <4 x i32> %lhs, %rhs 2778 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2779 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2780 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2781 2782 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2783 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2784 SrcTy->getScalarSizeInBits() / 2); 2785 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2786 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2787 2788 // %res = trunc <4 x i32> %high to <4 x i16> 2789 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2790 } 2791 case NEON::BI__builtin_neon_vcale_v: 2792 case NEON::BI__builtin_neon_vcaleq_v: 2793 case NEON::BI__builtin_neon_vcalt_v: 2794 case NEON::BI__builtin_neon_vcaltq_v: 2795 std::swap(Ops[0], Ops[1]); 2796 case NEON::BI__builtin_neon_vcage_v: 2797 case NEON::BI__builtin_neon_vcageq_v: 2798 case NEON::BI__builtin_neon_vcagt_v: 2799 case NEON::BI__builtin_neon_vcagtq_v: { 2800 llvm::Type *VecFlt = llvm::VectorType::get( 2801 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2802 VTy->getNumElements()); 2803 llvm::Type *Tys[] = { VTy, VecFlt }; 2804 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2805 return EmitNeonCall(F, Ops, NameHint); 2806 } 2807 case NEON::BI__builtin_neon_vclz_v: 2808 case NEON::BI__builtin_neon_vclzq_v: 2809 // We generate target-independent intrinsic, which needs a second argument 2810 // for whether or not clz of zero is undefined; on ARM it isn't. 2811 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2812 break; 2813 case NEON::BI__builtin_neon_vcvt_f32_v: 2814 case NEON::BI__builtin_neon_vcvtq_f32_v: 2815 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2816 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2817 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2818 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2819 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2820 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2821 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2822 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2823 bool Double = 2824 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2825 llvm::Type *FloatTy = 2826 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2827 : NeonTypeFlags::Float32, 2828 false, Quad)); 2829 llvm::Type *Tys[2] = { FloatTy, Ty }; 2830 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2831 Function *F = CGM.getIntrinsic(Int, Tys); 2832 return EmitNeonCall(F, Ops, "vcvt_n"); 2833 } 2834 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2835 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2836 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2837 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2838 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2839 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2840 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2841 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2842 bool Double = 2843 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2844 llvm::Type *FloatTy = 2845 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2846 : NeonTypeFlags::Float32, 2847 false, Quad)); 2848 llvm::Type *Tys[2] = { Ty, FloatTy }; 2849 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2850 return EmitNeonCall(F, Ops, "vcvt_n"); 2851 } 2852 case NEON::BI__builtin_neon_vcvt_s32_v: 2853 case NEON::BI__builtin_neon_vcvt_u32_v: 2854 case NEON::BI__builtin_neon_vcvt_s64_v: 2855 case NEON::BI__builtin_neon_vcvt_u64_v: 2856 case NEON::BI__builtin_neon_vcvtq_s32_v: 2857 case NEON::BI__builtin_neon_vcvtq_u32_v: 2858 case NEON::BI__builtin_neon_vcvtq_s64_v: 2859 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2860 bool Double = 2861 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2862 llvm::Type *FloatTy = 2863 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2864 : NeonTypeFlags::Float32, 2865 false, Quad)); 2866 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2867 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2868 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2869 } 2870 case NEON::BI__builtin_neon_vcvta_s32_v: 2871 case NEON::BI__builtin_neon_vcvta_s64_v: 2872 case NEON::BI__builtin_neon_vcvta_u32_v: 2873 case NEON::BI__builtin_neon_vcvta_u64_v: 2874 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2875 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2876 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2877 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2878 case NEON::BI__builtin_neon_vcvtn_s32_v: 2879 case NEON::BI__builtin_neon_vcvtn_s64_v: 2880 case NEON::BI__builtin_neon_vcvtn_u32_v: 2881 case NEON::BI__builtin_neon_vcvtn_u64_v: 2882 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2883 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2884 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2885 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2886 case NEON::BI__builtin_neon_vcvtp_s32_v: 2887 case NEON::BI__builtin_neon_vcvtp_s64_v: 2888 case NEON::BI__builtin_neon_vcvtp_u32_v: 2889 case NEON::BI__builtin_neon_vcvtp_u64_v: 2890 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2891 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2892 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2893 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2894 case NEON::BI__builtin_neon_vcvtm_s32_v: 2895 case NEON::BI__builtin_neon_vcvtm_s64_v: 2896 case NEON::BI__builtin_neon_vcvtm_u32_v: 2897 case NEON::BI__builtin_neon_vcvtm_u64_v: 2898 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2899 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2900 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2901 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2902 bool Double = 2903 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2904 llvm::Type *InTy = 2905 GetNeonType(this, 2906 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2907 : NeonTypeFlags::Float32, false, Quad)); 2908 llvm::Type *Tys[2] = { Ty, InTy }; 2909 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2910 } 2911 case NEON::BI__builtin_neon_vext_v: 2912 case NEON::BI__builtin_neon_vextq_v: { 2913 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2914 SmallVector<Constant*, 16> Indices; 2915 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2916 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2917 2918 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2919 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2920 Value *SV = llvm::ConstantVector::get(Indices); 2921 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2922 } 2923 case NEON::BI__builtin_neon_vfma_v: 2924 case NEON::BI__builtin_neon_vfmaq_v: { 2925 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2926 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2927 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2928 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2929 2930 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2931 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 2932 } 2933 case NEON::BI__builtin_neon_vld1_v: 2934 case NEON::BI__builtin_neon_vld1q_v: 2935 Ops.push_back(Align); 2936 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2937 case NEON::BI__builtin_neon_vld2_v: 2938 case NEON::BI__builtin_neon_vld2q_v: 2939 case NEON::BI__builtin_neon_vld3_v: 2940 case NEON::BI__builtin_neon_vld3q_v: 2941 case NEON::BI__builtin_neon_vld4_v: 2942 case NEON::BI__builtin_neon_vld4q_v: { 2943 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2944 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 2945 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2946 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2947 return Builder.CreateStore(Ops[1], Ops[0]); 2948 } 2949 case NEON::BI__builtin_neon_vld1_dup_v: 2950 case NEON::BI__builtin_neon_vld1q_dup_v: { 2951 Value *V = UndefValue::get(Ty); 2952 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2953 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2954 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2955 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2956 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2957 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2958 return EmitNeonSplat(Ops[0], CI); 2959 } 2960 case NEON::BI__builtin_neon_vld2_lane_v: 2961 case NEON::BI__builtin_neon_vld2q_lane_v: 2962 case NEON::BI__builtin_neon_vld3_lane_v: 2963 case NEON::BI__builtin_neon_vld3q_lane_v: 2964 case NEON::BI__builtin_neon_vld4_lane_v: 2965 case NEON::BI__builtin_neon_vld4q_lane_v: { 2966 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2967 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2968 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2969 Ops.push_back(Align); 2970 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2971 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2972 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2973 return Builder.CreateStore(Ops[1], Ops[0]); 2974 } 2975 case NEON::BI__builtin_neon_vmovl_v: { 2976 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2977 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2978 if (Usgn) 2979 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2980 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2981 } 2982 case NEON::BI__builtin_neon_vmovn_v: { 2983 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2984 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2985 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2986 } 2987 case NEON::BI__builtin_neon_vmull_v: 2988 // FIXME: the integer vmull operations could be emitted in terms of pure 2989 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2990 // hoisting the exts outside loops. Until global ISel comes along that can 2991 // see through such movement this leads to bad CodeGen. So we need an 2992 // intrinsic for now. 2993 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2994 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2995 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2996 case NEON::BI__builtin_neon_vpadal_v: 2997 case NEON::BI__builtin_neon_vpadalq_v: { 2998 // The source operand type has twice as many elements of half the size. 2999 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3000 llvm::Type *EltTy = 3001 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3002 llvm::Type *NarrowTy = 3003 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3004 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3005 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3006 } 3007 case NEON::BI__builtin_neon_vpaddl_v: 3008 case NEON::BI__builtin_neon_vpaddlq_v: { 3009 // The source operand type has twice as many elements of half the size. 3010 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3011 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3012 llvm::Type *NarrowTy = 3013 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3014 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3015 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3016 } 3017 case NEON::BI__builtin_neon_vqdmlal_v: 3018 case NEON::BI__builtin_neon_vqdmlsl_v: { 3019 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3020 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 3021 MulOps, "vqdmlal"); 3022 3023 SmallVector<Value *, 2> AccumOps; 3024 AccumOps.push_back(Ops[0]); 3025 AccumOps.push_back(Mul); 3026 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 3027 AccumOps, NameHint); 3028 } 3029 case NEON::BI__builtin_neon_vqshl_n_v: 3030 case NEON::BI__builtin_neon_vqshlq_n_v: 3031 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3032 1, false); 3033 case NEON::BI__builtin_neon_vqshlu_n_v: 3034 case NEON::BI__builtin_neon_vqshluq_n_v: 3035 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3036 1, false); 3037 case NEON::BI__builtin_neon_vrecpe_v: 3038 case NEON::BI__builtin_neon_vrecpeq_v: 3039 case NEON::BI__builtin_neon_vrsqrte_v: 3040 case NEON::BI__builtin_neon_vrsqrteq_v: 3041 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3042 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3043 3044 case NEON::BI__builtin_neon_vrshr_n_v: 3045 case NEON::BI__builtin_neon_vrshrq_n_v: 3046 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3047 1, true); 3048 case NEON::BI__builtin_neon_vshl_n_v: 3049 case NEON::BI__builtin_neon_vshlq_n_v: 3050 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3051 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3052 "vshl_n"); 3053 case NEON::BI__builtin_neon_vshll_n_v: { 3054 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3055 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3056 if (Usgn) 3057 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3058 else 3059 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3060 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3061 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3062 } 3063 case NEON::BI__builtin_neon_vshrn_n_v: { 3064 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3065 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3066 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3067 if (Usgn) 3068 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3069 else 3070 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3071 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3072 } 3073 case NEON::BI__builtin_neon_vshr_n_v: 3074 case NEON::BI__builtin_neon_vshrq_n_v: 3075 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3076 case NEON::BI__builtin_neon_vst1_v: 3077 case NEON::BI__builtin_neon_vst1q_v: 3078 case NEON::BI__builtin_neon_vst2_v: 3079 case NEON::BI__builtin_neon_vst2q_v: 3080 case NEON::BI__builtin_neon_vst3_v: 3081 case NEON::BI__builtin_neon_vst3q_v: 3082 case NEON::BI__builtin_neon_vst4_v: 3083 case NEON::BI__builtin_neon_vst4q_v: 3084 case NEON::BI__builtin_neon_vst2_lane_v: 3085 case NEON::BI__builtin_neon_vst2q_lane_v: 3086 case NEON::BI__builtin_neon_vst3_lane_v: 3087 case NEON::BI__builtin_neon_vst3q_lane_v: 3088 case NEON::BI__builtin_neon_vst4_lane_v: 3089 case NEON::BI__builtin_neon_vst4q_lane_v: 3090 Ops.push_back(Align); 3091 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3092 case NEON::BI__builtin_neon_vsubhn_v: { 3093 llvm::VectorType *SrcTy = 3094 llvm::VectorType::getExtendedElementVectorType(VTy); 3095 3096 // %sum = add <4 x i32> %lhs, %rhs 3097 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3098 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3099 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3100 3101 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3102 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 3103 SrcTy->getScalarSizeInBits() / 2); 3104 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 3105 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3106 3107 // %res = trunc <4 x i32> %high to <4 x i16> 3108 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3109 } 3110 case NEON::BI__builtin_neon_vtrn_v: 3111 case NEON::BI__builtin_neon_vtrnq_v: { 3112 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3113 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3114 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3115 Value *SV = nullptr; 3116 3117 for (unsigned vi = 0; vi != 2; ++vi) { 3118 SmallVector<Constant*, 16> Indices; 3119 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3120 Indices.push_back(Builder.getInt32(i+vi)); 3121 Indices.push_back(Builder.getInt32(i+e+vi)); 3122 } 3123 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3124 SV = llvm::ConstantVector::get(Indices); 3125 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3126 SV = Builder.CreateStore(SV, Addr); 3127 } 3128 return SV; 3129 } 3130 case NEON::BI__builtin_neon_vtst_v: 3131 case NEON::BI__builtin_neon_vtstq_v: { 3132 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3133 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3134 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3135 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3136 ConstantAggregateZero::get(Ty)); 3137 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3138 } 3139 case NEON::BI__builtin_neon_vuzp_v: 3140 case NEON::BI__builtin_neon_vuzpq_v: { 3141 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3142 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3143 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3144 Value *SV = nullptr; 3145 3146 for (unsigned vi = 0; vi != 2; ++vi) { 3147 SmallVector<Constant*, 16> Indices; 3148 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3149 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3150 3151 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3152 SV = llvm::ConstantVector::get(Indices); 3153 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3154 SV = Builder.CreateStore(SV, Addr); 3155 } 3156 return SV; 3157 } 3158 case NEON::BI__builtin_neon_vzip_v: 3159 case NEON::BI__builtin_neon_vzipq_v: { 3160 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3161 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3162 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3163 Value *SV = nullptr; 3164 3165 for (unsigned vi = 0; vi != 2; ++vi) { 3166 SmallVector<Constant*, 16> Indices; 3167 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3168 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3169 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3170 } 3171 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3172 SV = llvm::ConstantVector::get(Indices); 3173 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3174 SV = Builder.CreateStore(SV, Addr); 3175 } 3176 return SV; 3177 } 3178 } 3179 3180 assert(Int && "Expected valid intrinsic number"); 3181 3182 // Determine the type(s) of this overloaded AArch64 intrinsic. 3183 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3184 3185 Value *Result = EmitNeonCall(F, Ops, NameHint); 3186 llvm::Type *ResultType = ConvertType(E->getType()); 3187 // AArch64 intrinsic one-element vector type cast to 3188 // scalar type expected by the builtin 3189 return Builder.CreateBitCast(Result, ResultType, NameHint); 3190 } 3191 3192 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3193 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3194 const CmpInst::Predicate Ip, const Twine &Name) { 3195 llvm::Type *OTy = Op->getType(); 3196 3197 // FIXME: this is utterly horrific. We should not be looking at previous 3198 // codegen context to find out what needs doing. Unfortunately TableGen 3199 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3200 // (etc). 3201 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3202 OTy = BI->getOperand(0)->getType(); 3203 3204 Op = Builder.CreateBitCast(Op, OTy); 3205 if (OTy->getScalarType()->isFloatingPointTy()) { 3206 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3207 } else { 3208 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3209 } 3210 return Builder.CreateSExt(Op, Ty, Name); 3211 } 3212 3213 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3214 Value *ExtOp, Value *IndexOp, 3215 llvm::Type *ResTy, unsigned IntID, 3216 const char *Name) { 3217 SmallVector<Value *, 2> TblOps; 3218 if (ExtOp) 3219 TblOps.push_back(ExtOp); 3220 3221 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3222 SmallVector<Constant*, 16> Indices; 3223 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3224 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3225 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3226 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3227 } 3228 Value *SV = llvm::ConstantVector::get(Indices); 3229 3230 int PairPos = 0, End = Ops.size() - 1; 3231 while (PairPos < End) { 3232 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3233 Ops[PairPos+1], SV, Name)); 3234 PairPos += 2; 3235 } 3236 3237 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3238 // of the 128-bit lookup table with zero. 3239 if (PairPos == End) { 3240 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3241 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3242 ZeroTbl, SV, Name)); 3243 } 3244 3245 Function *TblF; 3246 TblOps.push_back(IndexOp); 3247 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3248 3249 return CGF.EmitNeonCall(TblF, TblOps, Name); 3250 } 3251 3252 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3253 switch (BuiltinID) { 3254 default: 3255 return nullptr; 3256 case ARM::BI__builtin_arm_nop: 3257 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3258 llvm::ConstantInt::get(Int32Ty, 0)); 3259 case ARM::BI__builtin_arm_yield: 3260 case ARM::BI__yield: 3261 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3262 llvm::ConstantInt::get(Int32Ty, 1)); 3263 case ARM::BI__builtin_arm_wfe: 3264 case ARM::BI__wfe: 3265 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3266 llvm::ConstantInt::get(Int32Ty, 2)); 3267 case ARM::BI__builtin_arm_wfi: 3268 case ARM::BI__wfi: 3269 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3270 llvm::ConstantInt::get(Int32Ty, 3)); 3271 case ARM::BI__builtin_arm_sev: 3272 case ARM::BI__sev: 3273 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3274 llvm::ConstantInt::get(Int32Ty, 4)); 3275 case ARM::BI__builtin_arm_sevl: 3276 case ARM::BI__sevl: 3277 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3278 llvm::ConstantInt::get(Int32Ty, 5)); 3279 } 3280 } 3281 3282 // Generates the IR for the read/write special register builtin, 3283 // ValueType is the type of the value that is to be written or read, 3284 // RegisterType is the type of the register being written to or read from. 3285 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3286 const CallExpr *E, 3287 llvm::Type *RegisterType, 3288 llvm::Type *ValueType, bool IsRead) { 3289 // write and register intrinsics only support 32 and 64 bit operations. 3290 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3291 && "Unsupported size for register."); 3292 3293 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3294 CodeGen::CodeGenModule &CGM = CGF.CGM; 3295 LLVMContext &Context = CGM.getLLVMContext(); 3296 3297 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3298 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3299 3300 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3301 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3302 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3303 3304 llvm::Type *Types[] = { RegisterType }; 3305 3306 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3307 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3308 && "Can't fit 64-bit value in 32-bit register"); 3309 3310 if (IsRead) { 3311 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3312 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3313 3314 if (MixedTypes) 3315 // Read into 64 bit register and then truncate result to 32 bit. 3316 return Builder.CreateTrunc(Call, ValueType); 3317 3318 if (ValueType->isPointerTy()) 3319 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3320 return Builder.CreateIntToPtr(Call, ValueType); 3321 3322 return Call; 3323 } 3324 3325 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3326 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3327 if (MixedTypes) { 3328 // Extend 32 bit write value to 64 bit to pass to write. 3329 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3330 return Builder.CreateCall(F, { Metadata, ArgValue }); 3331 } 3332 3333 if (ValueType->isPointerTy()) { 3334 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3335 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3336 return Builder.CreateCall(F, { Metadata, ArgValue }); 3337 } 3338 3339 return Builder.CreateCall(F, { Metadata, ArgValue }); 3340 } 3341 3342 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3343 const CallExpr *E) { 3344 if (auto Hint = GetValueForARMHint(BuiltinID)) 3345 return Hint; 3346 3347 if (BuiltinID == ARM::BI__emit) { 3348 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3349 llvm::FunctionType *FTy = 3350 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3351 3352 APSInt Value; 3353 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3354 llvm_unreachable("Sema will ensure that the parameter is constant"); 3355 3356 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3357 3358 llvm::InlineAsm *Emit = 3359 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3360 /*SideEffects=*/true) 3361 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3362 /*SideEffects=*/true); 3363 3364 return Builder.CreateCall(Emit, {}); 3365 } 3366 3367 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3368 Value *Option = EmitScalarExpr(E->getArg(0)); 3369 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3370 } 3371 3372 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3373 Value *Address = EmitScalarExpr(E->getArg(0)); 3374 Value *RW = EmitScalarExpr(E->getArg(1)); 3375 Value *IsData = EmitScalarExpr(E->getArg(2)); 3376 3377 // Locality is not supported on ARM target 3378 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3379 3380 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3381 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3382 } 3383 3384 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3385 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3386 EmitScalarExpr(E->getArg(0)), 3387 "rbit"); 3388 } 3389 3390 if (BuiltinID == ARM::BI__clear_cache) { 3391 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3392 const FunctionDecl *FD = E->getDirectCallee(); 3393 SmallVector<Value*, 2> Ops; 3394 for (unsigned i = 0; i < 2; i++) 3395 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3396 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3397 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3398 StringRef Name = FD->getName(); 3399 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3400 } 3401 3402 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3403 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3404 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3405 getContext().getTypeSize(E->getType()) == 64) || 3406 BuiltinID == ARM::BI__ldrexd) { 3407 Function *F; 3408 3409 switch (BuiltinID) { 3410 default: llvm_unreachable("unexpected builtin"); 3411 case ARM::BI__builtin_arm_ldaex: 3412 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3413 break; 3414 case ARM::BI__builtin_arm_ldrexd: 3415 case ARM::BI__builtin_arm_ldrex: 3416 case ARM::BI__ldrexd: 3417 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3418 break; 3419 } 3420 3421 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3422 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3423 "ldrexd"); 3424 3425 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3426 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3427 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3428 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3429 3430 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3431 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3432 Val = Builder.CreateOr(Val, Val1); 3433 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3434 } 3435 3436 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3437 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3438 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3439 3440 QualType Ty = E->getType(); 3441 llvm::Type *RealResTy = ConvertType(Ty); 3442 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3443 getContext().getTypeSize(Ty)); 3444 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3445 3446 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3447 ? Intrinsic::arm_ldaex 3448 : Intrinsic::arm_ldrex, 3449 LoadAddr->getType()); 3450 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3451 3452 if (RealResTy->isPointerTy()) 3453 return Builder.CreateIntToPtr(Val, RealResTy); 3454 else { 3455 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3456 return Builder.CreateBitCast(Val, RealResTy); 3457 } 3458 } 3459 3460 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3461 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3462 BuiltinID == ARM::BI__builtin_arm_strex) && 3463 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3464 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3465 ? Intrinsic::arm_stlexd 3466 : Intrinsic::arm_strexd); 3467 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3468 3469 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3470 Value *Val = EmitScalarExpr(E->getArg(0)); 3471 Builder.CreateStore(Val, Tmp); 3472 3473 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3474 Val = Builder.CreateLoad(LdPtr); 3475 3476 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3477 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3478 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3479 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3480 } 3481 3482 if (BuiltinID == ARM::BI__builtin_arm_strex || 3483 BuiltinID == ARM::BI__builtin_arm_stlex) { 3484 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3485 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3486 3487 QualType Ty = E->getArg(0)->getType(); 3488 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3489 getContext().getTypeSize(Ty)); 3490 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3491 3492 if (StoreVal->getType()->isPointerTy()) 3493 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3494 else { 3495 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3496 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3497 } 3498 3499 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3500 ? Intrinsic::arm_stlex 3501 : Intrinsic::arm_strex, 3502 StoreAddr->getType()); 3503 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3504 } 3505 3506 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3507 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3508 return Builder.CreateCall(F, {}); 3509 } 3510 3511 // CRC32 3512 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3513 switch (BuiltinID) { 3514 case ARM::BI__builtin_arm_crc32b: 3515 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3516 case ARM::BI__builtin_arm_crc32cb: 3517 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3518 case ARM::BI__builtin_arm_crc32h: 3519 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3520 case ARM::BI__builtin_arm_crc32ch: 3521 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3522 case ARM::BI__builtin_arm_crc32w: 3523 case ARM::BI__builtin_arm_crc32d: 3524 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3525 case ARM::BI__builtin_arm_crc32cw: 3526 case ARM::BI__builtin_arm_crc32cd: 3527 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3528 } 3529 3530 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3531 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3532 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3533 3534 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3535 // intrinsics, hence we need different codegen for these cases. 3536 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3537 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3538 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3539 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3540 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3541 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3542 3543 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3544 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3545 return Builder.CreateCall(F, {Res, Arg1b}); 3546 } else { 3547 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3548 3549 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3550 return Builder.CreateCall(F, {Arg0, Arg1}); 3551 } 3552 } 3553 3554 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3555 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3556 BuiltinID == ARM::BI__builtin_arm_rsrp || 3557 BuiltinID == ARM::BI__builtin_arm_wsr || 3558 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3559 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3560 3561 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3562 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3563 BuiltinID == ARM::BI__builtin_arm_rsrp; 3564 3565 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3566 BuiltinID == ARM::BI__builtin_arm_wsrp; 3567 3568 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3569 BuiltinID == ARM::BI__builtin_arm_wsr64; 3570 3571 llvm::Type *ValueType; 3572 llvm::Type *RegisterType; 3573 if (IsPointerBuiltin) { 3574 ValueType = VoidPtrTy; 3575 RegisterType = Int32Ty; 3576 } else if (Is64Bit) { 3577 ValueType = RegisterType = Int64Ty; 3578 } else { 3579 ValueType = RegisterType = Int32Ty; 3580 } 3581 3582 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3583 } 3584 3585 // Find out if any arguments are required to be integer constant 3586 // expressions. 3587 unsigned ICEArguments = 0; 3588 ASTContext::GetBuiltinTypeError Error; 3589 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3590 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3591 3592 SmallVector<Value*, 4> Ops; 3593 llvm::Value *Align = nullptr; 3594 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3595 if (i == 0) { 3596 switch (BuiltinID) { 3597 case NEON::BI__builtin_neon_vld1_v: 3598 case NEON::BI__builtin_neon_vld1q_v: 3599 case NEON::BI__builtin_neon_vld1q_lane_v: 3600 case NEON::BI__builtin_neon_vld1_lane_v: 3601 case NEON::BI__builtin_neon_vld1_dup_v: 3602 case NEON::BI__builtin_neon_vld1q_dup_v: 3603 case NEON::BI__builtin_neon_vst1_v: 3604 case NEON::BI__builtin_neon_vst1q_v: 3605 case NEON::BI__builtin_neon_vst1q_lane_v: 3606 case NEON::BI__builtin_neon_vst1_lane_v: 3607 case NEON::BI__builtin_neon_vst2_v: 3608 case NEON::BI__builtin_neon_vst2q_v: 3609 case NEON::BI__builtin_neon_vst2_lane_v: 3610 case NEON::BI__builtin_neon_vst2q_lane_v: 3611 case NEON::BI__builtin_neon_vst3_v: 3612 case NEON::BI__builtin_neon_vst3q_v: 3613 case NEON::BI__builtin_neon_vst3_lane_v: 3614 case NEON::BI__builtin_neon_vst3q_lane_v: 3615 case NEON::BI__builtin_neon_vst4_v: 3616 case NEON::BI__builtin_neon_vst4q_v: 3617 case NEON::BI__builtin_neon_vst4_lane_v: 3618 case NEON::BI__builtin_neon_vst4q_lane_v: 3619 // Get the alignment for the argument in addition to the value; 3620 // we'll use it later. 3621 std::pair<llvm::Value*, unsigned> Src = 3622 EmitPointerWithAlignment(E->getArg(0)); 3623 Ops.push_back(Src.first); 3624 Align = Builder.getInt32(Src.second); 3625 continue; 3626 } 3627 } 3628 if (i == 1) { 3629 switch (BuiltinID) { 3630 case NEON::BI__builtin_neon_vld2_v: 3631 case NEON::BI__builtin_neon_vld2q_v: 3632 case NEON::BI__builtin_neon_vld3_v: 3633 case NEON::BI__builtin_neon_vld3q_v: 3634 case NEON::BI__builtin_neon_vld4_v: 3635 case NEON::BI__builtin_neon_vld4q_v: 3636 case NEON::BI__builtin_neon_vld2_lane_v: 3637 case NEON::BI__builtin_neon_vld2q_lane_v: 3638 case NEON::BI__builtin_neon_vld3_lane_v: 3639 case NEON::BI__builtin_neon_vld3q_lane_v: 3640 case NEON::BI__builtin_neon_vld4_lane_v: 3641 case NEON::BI__builtin_neon_vld4q_lane_v: 3642 case NEON::BI__builtin_neon_vld2_dup_v: 3643 case NEON::BI__builtin_neon_vld3_dup_v: 3644 case NEON::BI__builtin_neon_vld4_dup_v: 3645 // Get the alignment for the argument in addition to the value; 3646 // we'll use it later. 3647 std::pair<llvm::Value*, unsigned> Src = 3648 EmitPointerWithAlignment(E->getArg(1)); 3649 Ops.push_back(Src.first); 3650 Align = Builder.getInt32(Src.second); 3651 continue; 3652 } 3653 } 3654 3655 if ((ICEArguments & (1 << i)) == 0) { 3656 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3657 } else { 3658 // If this is required to be a constant, constant fold it so that we know 3659 // that the generated intrinsic gets a ConstantInt. 3660 llvm::APSInt Result; 3661 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 3662 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 3663 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 3664 } 3665 } 3666 3667 switch (BuiltinID) { 3668 default: break; 3669 // vget_lane and vset_lane are not overloaded and do not have an extra 3670 // argument that specifies the vector type. 3671 case NEON::BI__builtin_neon_vget_lane_i8: 3672 case NEON::BI__builtin_neon_vget_lane_i16: 3673 case NEON::BI__builtin_neon_vget_lane_i32: 3674 case NEON::BI__builtin_neon_vget_lane_i64: 3675 case NEON::BI__builtin_neon_vget_lane_f32: 3676 case NEON::BI__builtin_neon_vgetq_lane_i8: 3677 case NEON::BI__builtin_neon_vgetq_lane_i16: 3678 case NEON::BI__builtin_neon_vgetq_lane_i32: 3679 case NEON::BI__builtin_neon_vgetq_lane_i64: 3680 case NEON::BI__builtin_neon_vgetq_lane_f32: 3681 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3682 "vget_lane"); 3683 case NEON::BI__builtin_neon_vset_lane_i8: 3684 case NEON::BI__builtin_neon_vset_lane_i16: 3685 case NEON::BI__builtin_neon_vset_lane_i32: 3686 case NEON::BI__builtin_neon_vset_lane_i64: 3687 case NEON::BI__builtin_neon_vset_lane_f32: 3688 case NEON::BI__builtin_neon_vsetq_lane_i8: 3689 case NEON::BI__builtin_neon_vsetq_lane_i16: 3690 case NEON::BI__builtin_neon_vsetq_lane_i32: 3691 case NEON::BI__builtin_neon_vsetq_lane_i64: 3692 case NEON::BI__builtin_neon_vsetq_lane_f32: 3693 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3694 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3695 3696 // Non-polymorphic crypto instructions also not overloaded 3697 case NEON::BI__builtin_neon_vsha1h_u32: 3698 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3699 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3700 "vsha1h"); 3701 case NEON::BI__builtin_neon_vsha1cq_u32: 3702 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3703 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3704 "vsha1h"); 3705 case NEON::BI__builtin_neon_vsha1pq_u32: 3706 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3707 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3708 "vsha1h"); 3709 case NEON::BI__builtin_neon_vsha1mq_u32: 3710 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3711 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3712 "vsha1h"); 3713 } 3714 3715 // Get the last argument, which specifies the vector type. 3716 llvm::APSInt Result; 3717 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3718 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3719 return nullptr; 3720 3721 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3722 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3723 // Determine the overloaded type of this builtin. 3724 llvm::Type *Ty; 3725 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3726 Ty = FloatTy; 3727 else 3728 Ty = DoubleTy; 3729 3730 // Determine whether this is an unsigned conversion or not. 3731 bool usgn = Result.getZExtValue() == 1; 3732 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3733 3734 // Call the appropriate intrinsic. 3735 Function *F = CGM.getIntrinsic(Int, Ty); 3736 return Builder.CreateCall(F, Ops, "vcvtr"); 3737 } 3738 3739 // Determine the type of this overloaded NEON intrinsic. 3740 NeonTypeFlags Type(Result.getZExtValue()); 3741 bool usgn = Type.isUnsigned(); 3742 bool rightShift = false; 3743 3744 llvm::VectorType *VTy = GetNeonType(this, Type); 3745 llvm::Type *Ty = VTy; 3746 if (!Ty) 3747 return nullptr; 3748 3749 // Many NEON builtins have identical semantics and uses in ARM and 3750 // AArch64. Emit these in a single function. 3751 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3752 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3753 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3754 if (Builtin) 3755 return EmitCommonNeonBuiltinExpr( 3756 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3757 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3758 3759 unsigned Int; 3760 switch (BuiltinID) { 3761 default: return nullptr; 3762 case NEON::BI__builtin_neon_vld1q_lane_v: 3763 // Handle 64-bit integer elements as a special case. Use shuffles of 3764 // one-element vectors to avoid poor code for i64 in the backend. 3765 if (VTy->getElementType()->isIntegerTy(64)) { 3766 // Extract the other lane. 3767 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3768 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3769 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3770 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3771 // Load the value as a one-element vector. 3772 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3773 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3774 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3775 // Combine them. 3776 SmallVector<Constant*, 2> Indices; 3777 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3778 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3779 SV = llvm::ConstantVector::get(Indices); 3780 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3781 } 3782 // fall through 3783 case NEON::BI__builtin_neon_vld1_lane_v: { 3784 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3785 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3786 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3787 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3788 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3789 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3790 } 3791 case NEON::BI__builtin_neon_vld2_dup_v: 3792 case NEON::BI__builtin_neon_vld3_dup_v: 3793 case NEON::BI__builtin_neon_vld4_dup_v: { 3794 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3795 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3796 switch (BuiltinID) { 3797 case NEON::BI__builtin_neon_vld2_dup_v: 3798 Int = Intrinsic::arm_neon_vld2; 3799 break; 3800 case NEON::BI__builtin_neon_vld3_dup_v: 3801 Int = Intrinsic::arm_neon_vld3; 3802 break; 3803 case NEON::BI__builtin_neon_vld4_dup_v: 3804 Int = Intrinsic::arm_neon_vld4; 3805 break; 3806 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3807 } 3808 Function *F = CGM.getIntrinsic(Int, Ty); 3809 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3810 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3811 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3812 return Builder.CreateStore(Ops[1], Ops[0]); 3813 } 3814 switch (BuiltinID) { 3815 case NEON::BI__builtin_neon_vld2_dup_v: 3816 Int = Intrinsic::arm_neon_vld2lane; 3817 break; 3818 case NEON::BI__builtin_neon_vld3_dup_v: 3819 Int = Intrinsic::arm_neon_vld3lane; 3820 break; 3821 case NEON::BI__builtin_neon_vld4_dup_v: 3822 Int = Intrinsic::arm_neon_vld4lane; 3823 break; 3824 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3825 } 3826 Function *F = CGM.getIntrinsic(Int, Ty); 3827 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3828 3829 SmallVector<Value*, 6> Args; 3830 Args.push_back(Ops[1]); 3831 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3832 3833 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3834 Args.push_back(CI); 3835 Args.push_back(Align); 3836 3837 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3838 // splat lane 0 to all elts in each vector of the result. 3839 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3840 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3841 Value *Elt = Builder.CreateBitCast(Val, Ty); 3842 Elt = EmitNeonSplat(Elt, CI); 3843 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3844 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3845 } 3846 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3847 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3848 return Builder.CreateStore(Ops[1], Ops[0]); 3849 } 3850 case NEON::BI__builtin_neon_vqrshrn_n_v: 3851 Int = 3852 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3853 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3854 1, true); 3855 case NEON::BI__builtin_neon_vqrshrun_n_v: 3856 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3857 Ops, "vqrshrun_n", 1, true); 3858 case NEON::BI__builtin_neon_vqshrn_n_v: 3859 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3860 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3861 1, true); 3862 case NEON::BI__builtin_neon_vqshrun_n_v: 3863 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3864 Ops, "vqshrun_n", 1, true); 3865 case NEON::BI__builtin_neon_vrecpe_v: 3866 case NEON::BI__builtin_neon_vrecpeq_v: 3867 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3868 Ops, "vrecpe"); 3869 case NEON::BI__builtin_neon_vrshrn_n_v: 3870 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3871 Ops, "vrshrn_n", 1, true); 3872 case NEON::BI__builtin_neon_vrsra_n_v: 3873 case NEON::BI__builtin_neon_vrsraq_n_v: 3874 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3875 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3876 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3877 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3878 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 3879 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3880 case NEON::BI__builtin_neon_vsri_n_v: 3881 case NEON::BI__builtin_neon_vsriq_n_v: 3882 rightShift = true; 3883 case NEON::BI__builtin_neon_vsli_n_v: 3884 case NEON::BI__builtin_neon_vsliq_n_v: 3885 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3886 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3887 Ops, "vsli_n"); 3888 case NEON::BI__builtin_neon_vsra_n_v: 3889 case NEON::BI__builtin_neon_vsraq_n_v: 3890 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3891 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3892 return Builder.CreateAdd(Ops[0], Ops[1]); 3893 case NEON::BI__builtin_neon_vst1q_lane_v: 3894 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3895 // a one-element vector and avoid poor code for i64 in the backend. 3896 if (VTy->getElementType()->isIntegerTy(64)) { 3897 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3898 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3899 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3900 Ops[2] = Align; 3901 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3902 Ops[1]->getType()), Ops); 3903 } 3904 // fall through 3905 case NEON::BI__builtin_neon_vst1_lane_v: { 3906 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3907 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3908 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3909 StoreInst *St = Builder.CreateStore(Ops[1], 3910 Builder.CreateBitCast(Ops[0], Ty)); 3911 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3912 return St; 3913 } 3914 case NEON::BI__builtin_neon_vtbl1_v: 3915 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3916 Ops, "vtbl1"); 3917 case NEON::BI__builtin_neon_vtbl2_v: 3918 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3919 Ops, "vtbl2"); 3920 case NEON::BI__builtin_neon_vtbl3_v: 3921 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3922 Ops, "vtbl3"); 3923 case NEON::BI__builtin_neon_vtbl4_v: 3924 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3925 Ops, "vtbl4"); 3926 case NEON::BI__builtin_neon_vtbx1_v: 3927 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3928 Ops, "vtbx1"); 3929 case NEON::BI__builtin_neon_vtbx2_v: 3930 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3931 Ops, "vtbx2"); 3932 case NEON::BI__builtin_neon_vtbx3_v: 3933 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3934 Ops, "vtbx3"); 3935 case NEON::BI__builtin_neon_vtbx4_v: 3936 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3937 Ops, "vtbx4"); 3938 } 3939 } 3940 3941 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3942 const CallExpr *E, 3943 SmallVectorImpl<Value *> &Ops) { 3944 unsigned int Int = 0; 3945 const char *s = nullptr; 3946 3947 switch (BuiltinID) { 3948 default: 3949 return nullptr; 3950 case NEON::BI__builtin_neon_vtbl1_v: 3951 case NEON::BI__builtin_neon_vqtbl1_v: 3952 case NEON::BI__builtin_neon_vqtbl1q_v: 3953 case NEON::BI__builtin_neon_vtbl2_v: 3954 case NEON::BI__builtin_neon_vqtbl2_v: 3955 case NEON::BI__builtin_neon_vqtbl2q_v: 3956 case NEON::BI__builtin_neon_vtbl3_v: 3957 case NEON::BI__builtin_neon_vqtbl3_v: 3958 case NEON::BI__builtin_neon_vqtbl3q_v: 3959 case NEON::BI__builtin_neon_vtbl4_v: 3960 case NEON::BI__builtin_neon_vqtbl4_v: 3961 case NEON::BI__builtin_neon_vqtbl4q_v: 3962 break; 3963 case NEON::BI__builtin_neon_vtbx1_v: 3964 case NEON::BI__builtin_neon_vqtbx1_v: 3965 case NEON::BI__builtin_neon_vqtbx1q_v: 3966 case NEON::BI__builtin_neon_vtbx2_v: 3967 case NEON::BI__builtin_neon_vqtbx2_v: 3968 case NEON::BI__builtin_neon_vqtbx2q_v: 3969 case NEON::BI__builtin_neon_vtbx3_v: 3970 case NEON::BI__builtin_neon_vqtbx3_v: 3971 case NEON::BI__builtin_neon_vqtbx3q_v: 3972 case NEON::BI__builtin_neon_vtbx4_v: 3973 case NEON::BI__builtin_neon_vqtbx4_v: 3974 case NEON::BI__builtin_neon_vqtbx4q_v: 3975 break; 3976 } 3977 3978 assert(E->getNumArgs() >= 3); 3979 3980 // Get the last argument, which specifies the vector type. 3981 llvm::APSInt Result; 3982 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3983 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3984 return nullptr; 3985 3986 // Determine the type of this overloaded NEON intrinsic. 3987 NeonTypeFlags Type(Result.getZExtValue()); 3988 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3989 llvm::Type *Ty = VTy; 3990 if (!Ty) 3991 return nullptr; 3992 3993 unsigned nElts = VTy->getNumElements(); 3994 3995 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3996 3997 // AArch64 scalar builtins are not overloaded, they do not have an extra 3998 // argument that specifies the vector type, need to handle each case. 3999 SmallVector<Value *, 2> TblOps; 4000 switch (BuiltinID) { 4001 case NEON::BI__builtin_neon_vtbl1_v: { 4002 TblOps.push_back(Ops[0]); 4003 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 4004 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4005 } 4006 case NEON::BI__builtin_neon_vtbl2_v: { 4007 TblOps.push_back(Ops[0]); 4008 TblOps.push_back(Ops[1]); 4009 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 4010 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4011 } 4012 case NEON::BI__builtin_neon_vtbl3_v: { 4013 TblOps.push_back(Ops[0]); 4014 TblOps.push_back(Ops[1]); 4015 TblOps.push_back(Ops[2]); 4016 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 4017 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4018 } 4019 case NEON::BI__builtin_neon_vtbl4_v: { 4020 TblOps.push_back(Ops[0]); 4021 TblOps.push_back(Ops[1]); 4022 TblOps.push_back(Ops[2]); 4023 TblOps.push_back(Ops[3]); 4024 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 4025 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4026 } 4027 case NEON::BI__builtin_neon_vtbx1_v: { 4028 TblOps.push_back(Ops[1]); 4029 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 4030 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4031 4032 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 4033 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 4034 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4035 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4036 4037 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4038 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4039 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4040 } 4041 case NEON::BI__builtin_neon_vtbx2_v: { 4042 TblOps.push_back(Ops[1]); 4043 TblOps.push_back(Ops[2]); 4044 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 4045 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 4046 } 4047 case NEON::BI__builtin_neon_vtbx3_v: { 4048 TblOps.push_back(Ops[1]); 4049 TblOps.push_back(Ops[2]); 4050 TblOps.push_back(Ops[3]); 4051 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 4052 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4053 4054 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 4055 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 4056 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4057 TwentyFourV); 4058 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4059 4060 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4061 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4062 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4063 } 4064 case NEON::BI__builtin_neon_vtbx4_v: { 4065 TblOps.push_back(Ops[1]); 4066 TblOps.push_back(Ops[2]); 4067 TblOps.push_back(Ops[3]); 4068 TblOps.push_back(Ops[4]); 4069 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 4070 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 4071 } 4072 case NEON::BI__builtin_neon_vqtbl1_v: 4073 case NEON::BI__builtin_neon_vqtbl1q_v: 4074 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4075 case NEON::BI__builtin_neon_vqtbl2_v: 4076 case NEON::BI__builtin_neon_vqtbl2q_v: { 4077 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4078 case NEON::BI__builtin_neon_vqtbl3_v: 4079 case NEON::BI__builtin_neon_vqtbl3q_v: 4080 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4081 case NEON::BI__builtin_neon_vqtbl4_v: 4082 case NEON::BI__builtin_neon_vqtbl4q_v: 4083 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4084 case NEON::BI__builtin_neon_vqtbx1_v: 4085 case NEON::BI__builtin_neon_vqtbx1q_v: 4086 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4087 case NEON::BI__builtin_neon_vqtbx2_v: 4088 case NEON::BI__builtin_neon_vqtbx2q_v: 4089 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4090 case NEON::BI__builtin_neon_vqtbx3_v: 4091 case NEON::BI__builtin_neon_vqtbx3q_v: 4092 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4093 case NEON::BI__builtin_neon_vqtbx4_v: 4094 case NEON::BI__builtin_neon_vqtbx4q_v: 4095 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4096 } 4097 } 4098 4099 if (!Int) 4100 return nullptr; 4101 4102 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4103 return CGF.EmitNeonCall(F, Ops, s); 4104 } 4105 4106 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4107 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4108 Op = Builder.CreateBitCast(Op, Int16Ty); 4109 Value *V = UndefValue::get(VTy); 4110 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4111 Op = Builder.CreateInsertElement(V, Op, CI); 4112 return Op; 4113 } 4114 4115 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 4116 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4117 Op = Builder.CreateBitCast(Op, Int8Ty); 4118 Value *V = UndefValue::get(VTy); 4119 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4120 Op = Builder.CreateInsertElement(V, Op, CI); 4121 return Op; 4122 } 4123 4124 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4125 const CallExpr *E) { 4126 unsigned HintID = static_cast<unsigned>(-1); 4127 switch (BuiltinID) { 4128 default: break; 4129 case AArch64::BI__builtin_arm_nop: 4130 HintID = 0; 4131 break; 4132 case AArch64::BI__builtin_arm_yield: 4133 HintID = 1; 4134 break; 4135 case AArch64::BI__builtin_arm_wfe: 4136 HintID = 2; 4137 break; 4138 case AArch64::BI__builtin_arm_wfi: 4139 HintID = 3; 4140 break; 4141 case AArch64::BI__builtin_arm_sev: 4142 HintID = 4; 4143 break; 4144 case AArch64::BI__builtin_arm_sevl: 4145 HintID = 5; 4146 break; 4147 } 4148 4149 if (HintID != static_cast<unsigned>(-1)) { 4150 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4151 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4152 } 4153 4154 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4155 Value *Address = EmitScalarExpr(E->getArg(0)); 4156 Value *RW = EmitScalarExpr(E->getArg(1)); 4157 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4158 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4159 Value *IsData = EmitScalarExpr(E->getArg(4)); 4160 4161 Value *Locality = nullptr; 4162 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4163 // Temporal fetch, needs to convert cache level to locality. 4164 Locality = llvm::ConstantInt::get(Int32Ty, 4165 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4166 } else { 4167 // Streaming fetch. 4168 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4169 } 4170 4171 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4172 // PLDL3STRM or PLDL2STRM. 4173 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4174 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4175 } 4176 4177 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4178 assert((getContext().getTypeSize(E->getType()) == 32) && 4179 "rbit of unusual size!"); 4180 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4181 return Builder.CreateCall( 4182 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4183 } 4184 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4185 assert((getContext().getTypeSize(E->getType()) == 64) && 4186 "rbit of unusual size!"); 4187 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4188 return Builder.CreateCall( 4189 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4190 } 4191 4192 if (BuiltinID == AArch64::BI__clear_cache) { 4193 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4194 const FunctionDecl *FD = E->getDirectCallee(); 4195 SmallVector<Value*, 2> Ops; 4196 for (unsigned i = 0; i < 2; i++) 4197 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4198 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4199 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4200 StringRef Name = FD->getName(); 4201 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4202 } 4203 4204 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4205 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4206 getContext().getTypeSize(E->getType()) == 128) { 4207 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4208 ? Intrinsic::aarch64_ldaxp 4209 : Intrinsic::aarch64_ldxp); 4210 4211 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4212 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4213 "ldxp"); 4214 4215 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4216 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4217 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4218 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4219 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4220 4221 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4222 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4223 Val = Builder.CreateOr(Val, Val1); 4224 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4225 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4226 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4227 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4228 4229 QualType Ty = E->getType(); 4230 llvm::Type *RealResTy = ConvertType(Ty); 4231 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4232 getContext().getTypeSize(Ty)); 4233 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4234 4235 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4236 ? Intrinsic::aarch64_ldaxr 4237 : Intrinsic::aarch64_ldxr, 4238 LoadAddr->getType()); 4239 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4240 4241 if (RealResTy->isPointerTy()) 4242 return Builder.CreateIntToPtr(Val, RealResTy); 4243 4244 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4245 return Builder.CreateBitCast(Val, RealResTy); 4246 } 4247 4248 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4249 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4250 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4251 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4252 ? Intrinsic::aarch64_stlxp 4253 : Intrinsic::aarch64_stxp); 4254 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4255 4256 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4257 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4258 One); 4259 Value *Val = EmitScalarExpr(E->getArg(0)); 4260 Builder.CreateStore(Val, Tmp); 4261 4262 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4263 Val = Builder.CreateLoad(LdPtr); 4264 4265 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4266 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4267 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4268 Int8PtrTy); 4269 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4270 } 4271 4272 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4273 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4274 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4275 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4276 4277 QualType Ty = E->getArg(0)->getType(); 4278 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4279 getContext().getTypeSize(Ty)); 4280 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4281 4282 if (StoreVal->getType()->isPointerTy()) 4283 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4284 else { 4285 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4286 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4287 } 4288 4289 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4290 ? Intrinsic::aarch64_stlxr 4291 : Intrinsic::aarch64_stxr, 4292 StoreAddr->getType()); 4293 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4294 } 4295 4296 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4297 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4298 return Builder.CreateCall(F, {}); 4299 } 4300 4301 // CRC32 4302 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4303 switch (BuiltinID) { 4304 case AArch64::BI__builtin_arm_crc32b: 4305 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4306 case AArch64::BI__builtin_arm_crc32cb: 4307 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4308 case AArch64::BI__builtin_arm_crc32h: 4309 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4310 case AArch64::BI__builtin_arm_crc32ch: 4311 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4312 case AArch64::BI__builtin_arm_crc32w: 4313 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4314 case AArch64::BI__builtin_arm_crc32cw: 4315 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4316 case AArch64::BI__builtin_arm_crc32d: 4317 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4318 case AArch64::BI__builtin_arm_crc32cd: 4319 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4320 } 4321 4322 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4323 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4324 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4325 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4326 4327 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4328 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4329 4330 return Builder.CreateCall(F, {Arg0, Arg1}); 4331 } 4332 4333 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4334 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4335 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4336 BuiltinID == AArch64::BI__builtin_arm_wsr || 4337 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4338 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4339 4340 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4341 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4342 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4343 4344 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4345 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4346 4347 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4348 BuiltinID != AArch64::BI__builtin_arm_wsr; 4349 4350 llvm::Type *ValueType; 4351 llvm::Type *RegisterType = Int64Ty; 4352 if (IsPointerBuiltin) { 4353 ValueType = VoidPtrTy; 4354 } else if (Is64Bit) { 4355 ValueType = Int64Ty; 4356 } else { 4357 ValueType = Int32Ty; 4358 } 4359 4360 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4361 } 4362 4363 // Find out if any arguments are required to be integer constant 4364 // expressions. 4365 unsigned ICEArguments = 0; 4366 ASTContext::GetBuiltinTypeError Error; 4367 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4368 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4369 4370 llvm::SmallVector<Value*, 4> Ops; 4371 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4372 if ((ICEArguments & (1 << i)) == 0) { 4373 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4374 } else { 4375 // If this is required to be a constant, constant fold it so that we know 4376 // that the generated intrinsic gets a ConstantInt. 4377 llvm::APSInt Result; 4378 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4379 assert(IsConst && "Constant arg isn't actually constant?"); 4380 (void)IsConst; 4381 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4382 } 4383 } 4384 4385 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4386 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4387 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4388 4389 if (Builtin) { 4390 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4391 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4392 assert(Result && "SISD intrinsic should have been handled"); 4393 return Result; 4394 } 4395 4396 llvm::APSInt Result; 4397 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4398 NeonTypeFlags Type(0); 4399 if (Arg->isIntegerConstantExpr(Result, getContext())) 4400 // Determine the type of this overloaded NEON intrinsic. 4401 Type = NeonTypeFlags(Result.getZExtValue()); 4402 4403 bool usgn = Type.isUnsigned(); 4404 bool quad = Type.isQuad(); 4405 4406 // Handle non-overloaded intrinsics first. 4407 switch (BuiltinID) { 4408 default: break; 4409 case NEON::BI__builtin_neon_vldrq_p128: { 4410 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4411 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4412 return Builder.CreateLoad(Ptr); 4413 } 4414 case NEON::BI__builtin_neon_vstrq_p128: { 4415 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4416 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4417 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4418 } 4419 case NEON::BI__builtin_neon_vcvts_u32_f32: 4420 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4421 usgn = true; 4422 // FALL THROUGH 4423 case NEON::BI__builtin_neon_vcvts_s32_f32: 4424 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4425 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4426 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4427 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4428 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4429 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4430 if (usgn) 4431 return Builder.CreateFPToUI(Ops[0], InTy); 4432 return Builder.CreateFPToSI(Ops[0], InTy); 4433 } 4434 case NEON::BI__builtin_neon_vcvts_f32_u32: 4435 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4436 usgn = true; 4437 // FALL THROUGH 4438 case NEON::BI__builtin_neon_vcvts_f32_s32: 4439 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4440 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4441 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4442 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4443 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4444 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4445 if (usgn) 4446 return Builder.CreateUIToFP(Ops[0], FTy); 4447 return Builder.CreateSIToFP(Ops[0], FTy); 4448 } 4449 case NEON::BI__builtin_neon_vpaddd_s64: { 4450 llvm::Type *Ty = 4451 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4452 Value *Vec = EmitScalarExpr(E->getArg(0)); 4453 // The vector is v2f64, so make sure it's bitcast to that. 4454 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4455 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4456 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4457 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4458 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4459 // Pairwise addition of a v2f64 into a scalar f64. 4460 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4461 } 4462 case NEON::BI__builtin_neon_vpaddd_f64: { 4463 llvm::Type *Ty = 4464 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4465 Value *Vec = EmitScalarExpr(E->getArg(0)); 4466 // The vector is v2f64, so make sure it's bitcast to that. 4467 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4468 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4469 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4470 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4471 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4472 // Pairwise addition of a v2f64 into a scalar f64. 4473 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4474 } 4475 case NEON::BI__builtin_neon_vpadds_f32: { 4476 llvm::Type *Ty = 4477 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4478 Value *Vec = EmitScalarExpr(E->getArg(0)); 4479 // The vector is v2f32, so make sure it's bitcast to that. 4480 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4481 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4482 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4483 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4484 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4485 // Pairwise addition of a v2f32 into a scalar f32. 4486 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4487 } 4488 case NEON::BI__builtin_neon_vceqzd_s64: 4489 case NEON::BI__builtin_neon_vceqzd_f64: 4490 case NEON::BI__builtin_neon_vceqzs_f32: 4491 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4492 return EmitAArch64CompareBuiltinExpr( 4493 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4494 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4495 case NEON::BI__builtin_neon_vcgezd_s64: 4496 case NEON::BI__builtin_neon_vcgezd_f64: 4497 case NEON::BI__builtin_neon_vcgezs_f32: 4498 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4499 return EmitAArch64CompareBuiltinExpr( 4500 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4501 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4502 case NEON::BI__builtin_neon_vclezd_s64: 4503 case NEON::BI__builtin_neon_vclezd_f64: 4504 case NEON::BI__builtin_neon_vclezs_f32: 4505 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4506 return EmitAArch64CompareBuiltinExpr( 4507 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4508 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4509 case NEON::BI__builtin_neon_vcgtzd_s64: 4510 case NEON::BI__builtin_neon_vcgtzd_f64: 4511 case NEON::BI__builtin_neon_vcgtzs_f32: 4512 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4513 return EmitAArch64CompareBuiltinExpr( 4514 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4515 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4516 case NEON::BI__builtin_neon_vcltzd_s64: 4517 case NEON::BI__builtin_neon_vcltzd_f64: 4518 case NEON::BI__builtin_neon_vcltzs_f32: 4519 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4520 return EmitAArch64CompareBuiltinExpr( 4521 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4522 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4523 4524 case NEON::BI__builtin_neon_vceqzd_u64: { 4525 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4526 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4527 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4528 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4529 llvm::Constant::getNullValue(Ty)); 4530 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4531 } 4532 case NEON::BI__builtin_neon_vceqd_f64: 4533 case NEON::BI__builtin_neon_vcled_f64: 4534 case NEON::BI__builtin_neon_vcltd_f64: 4535 case NEON::BI__builtin_neon_vcged_f64: 4536 case NEON::BI__builtin_neon_vcgtd_f64: { 4537 llvm::CmpInst::Predicate P; 4538 switch (BuiltinID) { 4539 default: llvm_unreachable("missing builtin ID in switch!"); 4540 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4541 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4542 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4543 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4544 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4545 } 4546 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4547 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4548 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4549 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4550 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4551 } 4552 case NEON::BI__builtin_neon_vceqs_f32: 4553 case NEON::BI__builtin_neon_vcles_f32: 4554 case NEON::BI__builtin_neon_vclts_f32: 4555 case NEON::BI__builtin_neon_vcges_f32: 4556 case NEON::BI__builtin_neon_vcgts_f32: { 4557 llvm::CmpInst::Predicate P; 4558 switch (BuiltinID) { 4559 default: llvm_unreachable("missing builtin ID in switch!"); 4560 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4561 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4562 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4563 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4564 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4565 } 4566 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4567 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4568 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4569 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4570 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4571 } 4572 case NEON::BI__builtin_neon_vceqd_s64: 4573 case NEON::BI__builtin_neon_vceqd_u64: 4574 case NEON::BI__builtin_neon_vcgtd_s64: 4575 case NEON::BI__builtin_neon_vcgtd_u64: 4576 case NEON::BI__builtin_neon_vcltd_s64: 4577 case NEON::BI__builtin_neon_vcltd_u64: 4578 case NEON::BI__builtin_neon_vcged_u64: 4579 case NEON::BI__builtin_neon_vcged_s64: 4580 case NEON::BI__builtin_neon_vcled_u64: 4581 case NEON::BI__builtin_neon_vcled_s64: { 4582 llvm::CmpInst::Predicate P; 4583 switch (BuiltinID) { 4584 default: llvm_unreachable("missing builtin ID in switch!"); 4585 case NEON::BI__builtin_neon_vceqd_s64: 4586 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4587 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4588 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4589 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4590 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4591 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4592 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4593 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4594 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4595 } 4596 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4597 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4598 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4599 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4600 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4601 } 4602 case NEON::BI__builtin_neon_vtstd_s64: 4603 case NEON::BI__builtin_neon_vtstd_u64: { 4604 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4605 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4606 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4607 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4608 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4609 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4610 llvm::Constant::getNullValue(Ty)); 4611 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4612 } 4613 case NEON::BI__builtin_neon_vset_lane_i8: 4614 case NEON::BI__builtin_neon_vset_lane_i16: 4615 case NEON::BI__builtin_neon_vset_lane_i32: 4616 case NEON::BI__builtin_neon_vset_lane_i64: 4617 case NEON::BI__builtin_neon_vset_lane_f32: 4618 case NEON::BI__builtin_neon_vsetq_lane_i8: 4619 case NEON::BI__builtin_neon_vsetq_lane_i16: 4620 case NEON::BI__builtin_neon_vsetq_lane_i32: 4621 case NEON::BI__builtin_neon_vsetq_lane_i64: 4622 case NEON::BI__builtin_neon_vsetq_lane_f32: 4623 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4624 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4625 case NEON::BI__builtin_neon_vset_lane_f64: 4626 // The vector type needs a cast for the v1f64 variant. 4627 Ops[1] = Builder.CreateBitCast(Ops[1], 4628 llvm::VectorType::get(DoubleTy, 1)); 4629 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4630 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4631 case NEON::BI__builtin_neon_vsetq_lane_f64: 4632 // The vector type needs a cast for the v2f64 variant. 4633 Ops[1] = Builder.CreateBitCast(Ops[1], 4634 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4635 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4636 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4637 4638 case NEON::BI__builtin_neon_vget_lane_i8: 4639 case NEON::BI__builtin_neon_vdupb_lane_i8: 4640 Ops[0] = Builder.CreateBitCast(Ops[0], 4641 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4642 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4643 "vget_lane"); 4644 case NEON::BI__builtin_neon_vgetq_lane_i8: 4645 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4646 Ops[0] = Builder.CreateBitCast(Ops[0], 4647 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4648 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4649 "vgetq_lane"); 4650 case NEON::BI__builtin_neon_vget_lane_i16: 4651 case NEON::BI__builtin_neon_vduph_lane_i16: 4652 Ops[0] = Builder.CreateBitCast(Ops[0], 4653 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4654 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4655 "vget_lane"); 4656 case NEON::BI__builtin_neon_vgetq_lane_i16: 4657 case NEON::BI__builtin_neon_vduph_laneq_i16: 4658 Ops[0] = Builder.CreateBitCast(Ops[0], 4659 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4660 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4661 "vgetq_lane"); 4662 case NEON::BI__builtin_neon_vget_lane_i32: 4663 case NEON::BI__builtin_neon_vdups_lane_i32: 4664 Ops[0] = Builder.CreateBitCast( 4665 Ops[0], 4666 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4667 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4668 "vget_lane"); 4669 case NEON::BI__builtin_neon_vdups_lane_f32: 4670 Ops[0] = Builder.CreateBitCast(Ops[0], 4671 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4672 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4673 "vdups_lane"); 4674 case NEON::BI__builtin_neon_vgetq_lane_i32: 4675 case NEON::BI__builtin_neon_vdups_laneq_i32: 4676 Ops[0] = Builder.CreateBitCast(Ops[0], 4677 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4678 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4679 "vgetq_lane"); 4680 case NEON::BI__builtin_neon_vget_lane_i64: 4681 case NEON::BI__builtin_neon_vdupd_lane_i64: 4682 Ops[0] = Builder.CreateBitCast(Ops[0], 4683 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4684 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4685 "vget_lane"); 4686 case NEON::BI__builtin_neon_vdupd_lane_f64: 4687 Ops[0] = Builder.CreateBitCast(Ops[0], 4688 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4689 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4690 "vdupd_lane"); 4691 case NEON::BI__builtin_neon_vgetq_lane_i64: 4692 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4693 Ops[0] = Builder.CreateBitCast(Ops[0], 4694 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4695 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4696 "vgetq_lane"); 4697 case NEON::BI__builtin_neon_vget_lane_f32: 4698 Ops[0] = Builder.CreateBitCast(Ops[0], 4699 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4700 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4701 "vget_lane"); 4702 case NEON::BI__builtin_neon_vget_lane_f64: 4703 Ops[0] = Builder.CreateBitCast(Ops[0], 4704 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4705 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4706 "vget_lane"); 4707 case NEON::BI__builtin_neon_vgetq_lane_f32: 4708 case NEON::BI__builtin_neon_vdups_laneq_f32: 4709 Ops[0] = Builder.CreateBitCast(Ops[0], 4710 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4711 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4712 "vgetq_lane"); 4713 case NEON::BI__builtin_neon_vgetq_lane_f64: 4714 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4715 Ops[0] = Builder.CreateBitCast(Ops[0], 4716 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4717 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4718 "vgetq_lane"); 4719 case NEON::BI__builtin_neon_vaddd_s64: 4720 case NEON::BI__builtin_neon_vaddd_u64: 4721 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4722 case NEON::BI__builtin_neon_vsubd_s64: 4723 case NEON::BI__builtin_neon_vsubd_u64: 4724 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4725 case NEON::BI__builtin_neon_vqdmlalh_s16: 4726 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4727 SmallVector<Value *, 2> ProductOps; 4728 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4729 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4730 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4731 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4732 ProductOps, "vqdmlXl"); 4733 Constant *CI = ConstantInt::get(SizeTy, 0); 4734 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4735 4736 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4737 ? Intrinsic::aarch64_neon_sqadd 4738 : Intrinsic::aarch64_neon_sqsub; 4739 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4740 } 4741 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4742 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4743 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4744 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4745 Ops, "vqshlu_n"); 4746 } 4747 case NEON::BI__builtin_neon_vqshld_n_u64: 4748 case NEON::BI__builtin_neon_vqshld_n_s64: { 4749 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4750 ? Intrinsic::aarch64_neon_uqshl 4751 : Intrinsic::aarch64_neon_sqshl; 4752 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4753 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4754 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4755 } 4756 case NEON::BI__builtin_neon_vrshrd_n_u64: 4757 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4758 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4759 ? Intrinsic::aarch64_neon_urshl 4760 : Intrinsic::aarch64_neon_srshl; 4761 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4762 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4763 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4764 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4765 } 4766 case NEON::BI__builtin_neon_vrsrad_n_u64: 4767 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4768 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4769 ? Intrinsic::aarch64_neon_urshl 4770 : Intrinsic::aarch64_neon_srshl; 4771 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4772 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4773 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4774 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4775 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4776 } 4777 case NEON::BI__builtin_neon_vshld_n_s64: 4778 case NEON::BI__builtin_neon_vshld_n_u64: { 4779 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4780 return Builder.CreateShl( 4781 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4782 } 4783 case NEON::BI__builtin_neon_vshrd_n_s64: { 4784 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4785 return Builder.CreateAShr( 4786 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4787 Amt->getZExtValue())), 4788 "shrd_n"); 4789 } 4790 case NEON::BI__builtin_neon_vshrd_n_u64: { 4791 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4792 uint64_t ShiftAmt = Amt->getZExtValue(); 4793 // Right-shifting an unsigned value by its size yields 0. 4794 if (ShiftAmt == 64) 4795 return ConstantInt::get(Int64Ty, 0); 4796 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4797 "shrd_n"); 4798 } 4799 case NEON::BI__builtin_neon_vsrad_n_s64: { 4800 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4801 Ops[1] = Builder.CreateAShr( 4802 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4803 Amt->getZExtValue())), 4804 "shrd_n"); 4805 return Builder.CreateAdd(Ops[0], Ops[1]); 4806 } 4807 case NEON::BI__builtin_neon_vsrad_n_u64: { 4808 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4809 uint64_t ShiftAmt = Amt->getZExtValue(); 4810 // Right-shifting an unsigned value by its size yields 0. 4811 // As Op + 0 = Op, return Ops[0] directly. 4812 if (ShiftAmt == 64) 4813 return Ops[0]; 4814 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4815 "shrd_n"); 4816 return Builder.CreateAdd(Ops[0], Ops[1]); 4817 } 4818 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4819 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4820 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4821 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4822 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4823 "lane"); 4824 SmallVector<Value *, 2> ProductOps; 4825 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4826 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4827 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4828 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4829 ProductOps, "vqdmlXl"); 4830 Constant *CI = ConstantInt::get(SizeTy, 0); 4831 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4832 Ops.pop_back(); 4833 4834 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4835 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4836 ? Intrinsic::aarch64_neon_sqadd 4837 : Intrinsic::aarch64_neon_sqsub; 4838 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4839 } 4840 case NEON::BI__builtin_neon_vqdmlals_s32: 4841 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4842 SmallVector<Value *, 2> ProductOps; 4843 ProductOps.push_back(Ops[1]); 4844 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4845 Ops[1] = 4846 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4847 ProductOps, "vqdmlXl"); 4848 4849 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4850 ? Intrinsic::aarch64_neon_sqadd 4851 : Intrinsic::aarch64_neon_sqsub; 4852 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4853 } 4854 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4855 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4856 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4857 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4858 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4859 "lane"); 4860 SmallVector<Value *, 2> ProductOps; 4861 ProductOps.push_back(Ops[1]); 4862 ProductOps.push_back(Ops[2]); 4863 Ops[1] = 4864 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4865 ProductOps, "vqdmlXl"); 4866 Ops.pop_back(); 4867 4868 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4869 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4870 ? Intrinsic::aarch64_neon_sqadd 4871 : Intrinsic::aarch64_neon_sqsub; 4872 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4873 } 4874 } 4875 4876 llvm::VectorType *VTy = GetNeonType(this, Type); 4877 llvm::Type *Ty = VTy; 4878 if (!Ty) 4879 return nullptr; 4880 4881 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4882 // defer to common code if it's been added to our special map. 4883 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4884 AArch64SIMDIntrinsicsProvenSorted); 4885 4886 if (Builtin) 4887 return EmitCommonNeonBuiltinExpr( 4888 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4889 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4890 4891 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4892 return V; 4893 4894 unsigned Int; 4895 switch (BuiltinID) { 4896 default: return nullptr; 4897 case NEON::BI__builtin_neon_vbsl_v: 4898 case NEON::BI__builtin_neon_vbslq_v: { 4899 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4900 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4901 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4902 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4903 4904 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4905 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4906 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4907 return Builder.CreateBitCast(Ops[0], Ty); 4908 } 4909 case NEON::BI__builtin_neon_vfma_lane_v: 4910 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4911 // The ARM builtins (and instructions) have the addend as the first 4912 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4913 Value *Addend = Ops[0]; 4914 Value *Multiplicand = Ops[1]; 4915 Value *LaneSource = Ops[2]; 4916 Ops[0] = Multiplicand; 4917 Ops[1] = LaneSource; 4918 Ops[2] = Addend; 4919 4920 // Now adjust things to handle the lane access. 4921 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4922 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4923 VTy; 4924 llvm::Constant *cst = cast<Constant>(Ops[3]); 4925 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4926 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4927 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4928 4929 Ops.pop_back(); 4930 Int = Intrinsic::fma; 4931 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4932 } 4933 case NEON::BI__builtin_neon_vfma_laneq_v: { 4934 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4935 // v1f64 fma should be mapped to Neon scalar f64 fma 4936 if (VTy && VTy->getElementType() == DoubleTy) { 4937 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4938 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4939 llvm::Type *VTy = GetNeonType(this, 4940 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4941 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4942 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4943 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4944 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4945 return Builder.CreateBitCast(Result, Ty); 4946 } 4947 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4948 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4949 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4950 4951 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4952 VTy->getNumElements() * 2); 4953 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4954 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4955 cast<ConstantInt>(Ops[3])); 4956 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4957 4958 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4959 } 4960 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4961 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4962 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4963 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4964 4965 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4966 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4967 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4968 } 4969 case NEON::BI__builtin_neon_vfmas_lane_f32: 4970 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4971 case NEON::BI__builtin_neon_vfmad_lane_f64: 4972 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4973 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4974 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4975 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4976 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4977 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4978 } 4979 case NEON::BI__builtin_neon_vfms_v: 4980 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4981 // FIXME: probably remove when we no longer support aarch64_simd.h 4982 // (arm_neon.h delegates to vfma). 4983 4984 // The ARM builtins (and instructions) have the addend as the first 4985 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4986 Value *Subtrahend = Ops[0]; 4987 Value *Multiplicand = Ops[2]; 4988 Ops[0] = Multiplicand; 4989 Ops[2] = Subtrahend; 4990 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4991 Ops[1] = Builder.CreateFNeg(Ops[1]); 4992 Int = Intrinsic::fma; 4993 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4994 } 4995 case NEON::BI__builtin_neon_vmull_v: 4996 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4997 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4998 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4999 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5000 case NEON::BI__builtin_neon_vmax_v: 5001 case NEON::BI__builtin_neon_vmaxq_v: 5002 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5003 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5004 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5005 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5006 case NEON::BI__builtin_neon_vmin_v: 5007 case NEON::BI__builtin_neon_vminq_v: 5008 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5009 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5010 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5011 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5012 case NEON::BI__builtin_neon_vabd_v: 5013 case NEON::BI__builtin_neon_vabdq_v: 5014 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5015 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5016 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5017 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5018 case NEON::BI__builtin_neon_vpadal_v: 5019 case NEON::BI__builtin_neon_vpadalq_v: { 5020 unsigned ArgElts = VTy->getNumElements(); 5021 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5022 unsigned BitWidth = EltTy->getBitWidth(); 5023 llvm::Type *ArgTy = llvm::VectorType::get( 5024 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5025 llvm::Type* Tys[2] = { VTy, ArgTy }; 5026 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5027 SmallVector<llvm::Value*, 1> TmpOps; 5028 TmpOps.push_back(Ops[1]); 5029 Function *F = CGM.getIntrinsic(Int, Tys); 5030 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5031 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5032 return Builder.CreateAdd(tmp, addend); 5033 } 5034 case NEON::BI__builtin_neon_vpmin_v: 5035 case NEON::BI__builtin_neon_vpminq_v: 5036 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5037 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5038 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5039 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5040 case NEON::BI__builtin_neon_vpmax_v: 5041 case NEON::BI__builtin_neon_vpmaxq_v: 5042 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5043 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5044 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5045 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5046 case NEON::BI__builtin_neon_vminnm_v: 5047 case NEON::BI__builtin_neon_vminnmq_v: 5048 Int = Intrinsic::aarch64_neon_fminnm; 5049 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5050 case NEON::BI__builtin_neon_vmaxnm_v: 5051 case NEON::BI__builtin_neon_vmaxnmq_v: 5052 Int = Intrinsic::aarch64_neon_fmaxnm; 5053 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5054 case NEON::BI__builtin_neon_vrecpss_f32: { 5055 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 5056 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5057 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 5058 Ops, "vrecps"); 5059 } 5060 case NEON::BI__builtin_neon_vrecpsd_f64: { 5061 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 5062 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5063 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 5064 Ops, "vrecps"); 5065 } 5066 case NEON::BI__builtin_neon_vqshrun_n_v: 5067 Int = Intrinsic::aarch64_neon_sqshrun; 5068 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5069 case NEON::BI__builtin_neon_vqrshrun_n_v: 5070 Int = Intrinsic::aarch64_neon_sqrshrun; 5071 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5072 case NEON::BI__builtin_neon_vqshrn_n_v: 5073 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5074 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5075 case NEON::BI__builtin_neon_vrshrn_n_v: 5076 Int = Intrinsic::aarch64_neon_rshrn; 5077 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5078 case NEON::BI__builtin_neon_vqrshrn_n_v: 5079 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5080 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5081 case NEON::BI__builtin_neon_vrnda_v: 5082 case NEON::BI__builtin_neon_vrndaq_v: { 5083 Int = Intrinsic::round; 5084 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5085 } 5086 case NEON::BI__builtin_neon_vrndi_v: 5087 case NEON::BI__builtin_neon_vrndiq_v: { 5088 Int = Intrinsic::nearbyint; 5089 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5090 } 5091 case NEON::BI__builtin_neon_vrndm_v: 5092 case NEON::BI__builtin_neon_vrndmq_v: { 5093 Int = Intrinsic::floor; 5094 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5095 } 5096 case NEON::BI__builtin_neon_vrndn_v: 5097 case NEON::BI__builtin_neon_vrndnq_v: { 5098 Int = Intrinsic::aarch64_neon_frintn; 5099 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5100 } 5101 case NEON::BI__builtin_neon_vrndp_v: 5102 case NEON::BI__builtin_neon_vrndpq_v: { 5103 Int = Intrinsic::ceil; 5104 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5105 } 5106 case NEON::BI__builtin_neon_vrndx_v: 5107 case NEON::BI__builtin_neon_vrndxq_v: { 5108 Int = Intrinsic::rint; 5109 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5110 } 5111 case NEON::BI__builtin_neon_vrnd_v: 5112 case NEON::BI__builtin_neon_vrndq_v: { 5113 Int = Intrinsic::trunc; 5114 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5115 } 5116 case NEON::BI__builtin_neon_vceqz_v: 5117 case NEON::BI__builtin_neon_vceqzq_v: 5118 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5119 ICmpInst::ICMP_EQ, "vceqz"); 5120 case NEON::BI__builtin_neon_vcgez_v: 5121 case NEON::BI__builtin_neon_vcgezq_v: 5122 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5123 ICmpInst::ICMP_SGE, "vcgez"); 5124 case NEON::BI__builtin_neon_vclez_v: 5125 case NEON::BI__builtin_neon_vclezq_v: 5126 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5127 ICmpInst::ICMP_SLE, "vclez"); 5128 case NEON::BI__builtin_neon_vcgtz_v: 5129 case NEON::BI__builtin_neon_vcgtzq_v: 5130 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5131 ICmpInst::ICMP_SGT, "vcgtz"); 5132 case NEON::BI__builtin_neon_vcltz_v: 5133 case NEON::BI__builtin_neon_vcltzq_v: 5134 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5135 ICmpInst::ICMP_SLT, "vcltz"); 5136 case NEON::BI__builtin_neon_vcvt_f64_v: 5137 case NEON::BI__builtin_neon_vcvtq_f64_v: 5138 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5139 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5140 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5141 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5142 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5143 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5144 "unexpected vcvt_f64_f32 builtin"); 5145 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5146 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5147 5148 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5149 } 5150 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5151 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5152 "unexpected vcvt_f32_f64 builtin"); 5153 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5154 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5155 5156 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5157 } 5158 case NEON::BI__builtin_neon_vcvt_s32_v: 5159 case NEON::BI__builtin_neon_vcvt_u32_v: 5160 case NEON::BI__builtin_neon_vcvt_s64_v: 5161 case NEON::BI__builtin_neon_vcvt_u64_v: 5162 case NEON::BI__builtin_neon_vcvtq_s32_v: 5163 case NEON::BI__builtin_neon_vcvtq_u32_v: 5164 case NEON::BI__builtin_neon_vcvtq_s64_v: 5165 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5166 bool Double = 5167 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5168 llvm::Type *InTy = 5169 GetNeonType(this, 5170 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5171 : NeonTypeFlags::Float32, false, quad)); 5172 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5173 if (usgn) 5174 return Builder.CreateFPToUI(Ops[0], Ty); 5175 return Builder.CreateFPToSI(Ops[0], Ty); 5176 } 5177 case NEON::BI__builtin_neon_vcvta_s32_v: 5178 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5179 case NEON::BI__builtin_neon_vcvta_u32_v: 5180 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5181 case NEON::BI__builtin_neon_vcvta_s64_v: 5182 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5183 case NEON::BI__builtin_neon_vcvta_u64_v: 5184 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5185 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5186 bool Double = 5187 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5188 llvm::Type *InTy = 5189 GetNeonType(this, 5190 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5191 : NeonTypeFlags::Float32, false, quad)); 5192 llvm::Type *Tys[2] = { Ty, InTy }; 5193 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5194 } 5195 case NEON::BI__builtin_neon_vcvtm_s32_v: 5196 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5197 case NEON::BI__builtin_neon_vcvtm_u32_v: 5198 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5199 case NEON::BI__builtin_neon_vcvtm_s64_v: 5200 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5201 case NEON::BI__builtin_neon_vcvtm_u64_v: 5202 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5203 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5204 bool Double = 5205 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5206 llvm::Type *InTy = 5207 GetNeonType(this, 5208 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5209 : NeonTypeFlags::Float32, false, quad)); 5210 llvm::Type *Tys[2] = { Ty, InTy }; 5211 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5212 } 5213 case NEON::BI__builtin_neon_vcvtn_s32_v: 5214 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5215 case NEON::BI__builtin_neon_vcvtn_u32_v: 5216 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5217 case NEON::BI__builtin_neon_vcvtn_s64_v: 5218 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5219 case NEON::BI__builtin_neon_vcvtn_u64_v: 5220 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5221 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5222 bool Double = 5223 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5224 llvm::Type *InTy = 5225 GetNeonType(this, 5226 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5227 : NeonTypeFlags::Float32, false, quad)); 5228 llvm::Type *Tys[2] = { Ty, InTy }; 5229 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5230 } 5231 case NEON::BI__builtin_neon_vcvtp_s32_v: 5232 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5233 case NEON::BI__builtin_neon_vcvtp_u32_v: 5234 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5235 case NEON::BI__builtin_neon_vcvtp_s64_v: 5236 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5237 case NEON::BI__builtin_neon_vcvtp_u64_v: 5238 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5239 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5240 bool Double = 5241 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5242 llvm::Type *InTy = 5243 GetNeonType(this, 5244 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5245 : NeonTypeFlags::Float32, false, quad)); 5246 llvm::Type *Tys[2] = { Ty, InTy }; 5247 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5248 } 5249 case NEON::BI__builtin_neon_vmulx_v: 5250 case NEON::BI__builtin_neon_vmulxq_v: { 5251 Int = Intrinsic::aarch64_neon_fmulx; 5252 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5253 } 5254 case NEON::BI__builtin_neon_vmul_lane_v: 5255 case NEON::BI__builtin_neon_vmul_laneq_v: { 5256 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5257 bool Quad = false; 5258 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5259 Quad = true; 5260 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5261 llvm::Type *VTy = GetNeonType(this, 5262 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5263 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5264 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5265 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5266 return Builder.CreateBitCast(Result, Ty); 5267 } 5268 case NEON::BI__builtin_neon_vnegd_s64: 5269 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5270 case NEON::BI__builtin_neon_vpmaxnm_v: 5271 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5272 Int = Intrinsic::aarch64_neon_fmaxnmp; 5273 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5274 } 5275 case NEON::BI__builtin_neon_vpminnm_v: 5276 case NEON::BI__builtin_neon_vpminnmq_v: { 5277 Int = Intrinsic::aarch64_neon_fminnmp; 5278 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5279 } 5280 case NEON::BI__builtin_neon_vsqrt_v: 5281 case NEON::BI__builtin_neon_vsqrtq_v: { 5282 Int = Intrinsic::sqrt; 5283 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5284 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5285 } 5286 case NEON::BI__builtin_neon_vrbit_v: 5287 case NEON::BI__builtin_neon_vrbitq_v: { 5288 Int = Intrinsic::aarch64_neon_rbit; 5289 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5290 } 5291 case NEON::BI__builtin_neon_vaddv_u8: 5292 // FIXME: These are handled by the AArch64 scalar code. 5293 usgn = true; 5294 // FALLTHROUGH 5295 case NEON::BI__builtin_neon_vaddv_s8: { 5296 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5297 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5298 VTy = 5299 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5300 llvm::Type *Tys[2] = { Ty, VTy }; 5301 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5302 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5303 return Builder.CreateTrunc(Ops[0], 5304 llvm::IntegerType::get(getLLVMContext(), 8)); 5305 } 5306 case NEON::BI__builtin_neon_vaddv_u16: 5307 usgn = true; 5308 // FALLTHROUGH 5309 case NEON::BI__builtin_neon_vaddv_s16: { 5310 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5311 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5312 VTy = 5313 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5314 llvm::Type *Tys[2] = { Ty, VTy }; 5315 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5316 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5317 return Builder.CreateTrunc(Ops[0], 5318 llvm::IntegerType::get(getLLVMContext(), 16)); 5319 } 5320 case NEON::BI__builtin_neon_vaddvq_u8: 5321 usgn = true; 5322 // FALLTHROUGH 5323 case NEON::BI__builtin_neon_vaddvq_s8: { 5324 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5325 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5326 VTy = 5327 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5328 llvm::Type *Tys[2] = { Ty, VTy }; 5329 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5330 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5331 return Builder.CreateTrunc(Ops[0], 5332 llvm::IntegerType::get(getLLVMContext(), 8)); 5333 } 5334 case NEON::BI__builtin_neon_vaddvq_u16: 5335 usgn = true; 5336 // FALLTHROUGH 5337 case NEON::BI__builtin_neon_vaddvq_s16: { 5338 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5339 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5340 VTy = 5341 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5342 llvm::Type *Tys[2] = { Ty, VTy }; 5343 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5344 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5345 return Builder.CreateTrunc(Ops[0], 5346 llvm::IntegerType::get(getLLVMContext(), 16)); 5347 } 5348 case NEON::BI__builtin_neon_vmaxv_u8: { 5349 Int = Intrinsic::aarch64_neon_umaxv; 5350 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5351 VTy = 5352 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5353 llvm::Type *Tys[2] = { Ty, VTy }; 5354 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5355 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5356 return Builder.CreateTrunc(Ops[0], 5357 llvm::IntegerType::get(getLLVMContext(), 8)); 5358 } 5359 case NEON::BI__builtin_neon_vmaxv_u16: { 5360 Int = Intrinsic::aarch64_neon_umaxv; 5361 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5362 VTy = 5363 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5364 llvm::Type *Tys[2] = { Ty, VTy }; 5365 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5366 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5367 return Builder.CreateTrunc(Ops[0], 5368 llvm::IntegerType::get(getLLVMContext(), 16)); 5369 } 5370 case NEON::BI__builtin_neon_vmaxvq_u8: { 5371 Int = Intrinsic::aarch64_neon_umaxv; 5372 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5373 VTy = 5374 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5375 llvm::Type *Tys[2] = { Ty, VTy }; 5376 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5377 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5378 return Builder.CreateTrunc(Ops[0], 5379 llvm::IntegerType::get(getLLVMContext(), 8)); 5380 } 5381 case NEON::BI__builtin_neon_vmaxvq_u16: { 5382 Int = Intrinsic::aarch64_neon_umaxv; 5383 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5384 VTy = 5385 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5386 llvm::Type *Tys[2] = { Ty, VTy }; 5387 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5388 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5389 return Builder.CreateTrunc(Ops[0], 5390 llvm::IntegerType::get(getLLVMContext(), 16)); 5391 } 5392 case NEON::BI__builtin_neon_vmaxv_s8: { 5393 Int = Intrinsic::aarch64_neon_smaxv; 5394 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5395 VTy = 5396 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5397 llvm::Type *Tys[2] = { Ty, VTy }; 5398 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5399 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5400 return Builder.CreateTrunc(Ops[0], 5401 llvm::IntegerType::get(getLLVMContext(), 8)); 5402 } 5403 case NEON::BI__builtin_neon_vmaxv_s16: { 5404 Int = Intrinsic::aarch64_neon_smaxv; 5405 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5406 VTy = 5407 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5408 llvm::Type *Tys[2] = { Ty, VTy }; 5409 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5410 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5411 return Builder.CreateTrunc(Ops[0], 5412 llvm::IntegerType::get(getLLVMContext(), 16)); 5413 } 5414 case NEON::BI__builtin_neon_vmaxvq_s8: { 5415 Int = Intrinsic::aarch64_neon_smaxv; 5416 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5417 VTy = 5418 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5419 llvm::Type *Tys[2] = { Ty, VTy }; 5420 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5421 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5422 return Builder.CreateTrunc(Ops[0], 5423 llvm::IntegerType::get(getLLVMContext(), 8)); 5424 } 5425 case NEON::BI__builtin_neon_vmaxvq_s16: { 5426 Int = Intrinsic::aarch64_neon_smaxv; 5427 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5428 VTy = 5429 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5430 llvm::Type *Tys[2] = { Ty, VTy }; 5431 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5432 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5433 return Builder.CreateTrunc(Ops[0], 5434 llvm::IntegerType::get(getLLVMContext(), 16)); 5435 } 5436 case NEON::BI__builtin_neon_vminv_u8: { 5437 Int = Intrinsic::aarch64_neon_uminv; 5438 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5439 VTy = 5440 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5441 llvm::Type *Tys[2] = { Ty, VTy }; 5442 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5443 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5444 return Builder.CreateTrunc(Ops[0], 5445 llvm::IntegerType::get(getLLVMContext(), 8)); 5446 } 5447 case NEON::BI__builtin_neon_vminv_u16: { 5448 Int = Intrinsic::aarch64_neon_uminv; 5449 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5450 VTy = 5451 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5452 llvm::Type *Tys[2] = { Ty, VTy }; 5453 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5454 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5455 return Builder.CreateTrunc(Ops[0], 5456 llvm::IntegerType::get(getLLVMContext(), 16)); 5457 } 5458 case NEON::BI__builtin_neon_vminvq_u8: { 5459 Int = Intrinsic::aarch64_neon_uminv; 5460 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5461 VTy = 5462 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5463 llvm::Type *Tys[2] = { Ty, VTy }; 5464 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5465 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5466 return Builder.CreateTrunc(Ops[0], 5467 llvm::IntegerType::get(getLLVMContext(), 8)); 5468 } 5469 case NEON::BI__builtin_neon_vminvq_u16: { 5470 Int = Intrinsic::aarch64_neon_uminv; 5471 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5472 VTy = 5473 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5474 llvm::Type *Tys[2] = { Ty, VTy }; 5475 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5476 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5477 return Builder.CreateTrunc(Ops[0], 5478 llvm::IntegerType::get(getLLVMContext(), 16)); 5479 } 5480 case NEON::BI__builtin_neon_vminv_s8: { 5481 Int = Intrinsic::aarch64_neon_sminv; 5482 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5483 VTy = 5484 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5485 llvm::Type *Tys[2] = { Ty, VTy }; 5486 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5487 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5488 return Builder.CreateTrunc(Ops[0], 5489 llvm::IntegerType::get(getLLVMContext(), 8)); 5490 } 5491 case NEON::BI__builtin_neon_vminv_s16: { 5492 Int = Intrinsic::aarch64_neon_sminv; 5493 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5494 VTy = 5495 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5496 llvm::Type *Tys[2] = { Ty, VTy }; 5497 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5498 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5499 return Builder.CreateTrunc(Ops[0], 5500 llvm::IntegerType::get(getLLVMContext(), 16)); 5501 } 5502 case NEON::BI__builtin_neon_vminvq_s8: { 5503 Int = Intrinsic::aarch64_neon_sminv; 5504 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5505 VTy = 5506 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5507 llvm::Type *Tys[2] = { Ty, VTy }; 5508 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5509 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5510 return Builder.CreateTrunc(Ops[0], 5511 llvm::IntegerType::get(getLLVMContext(), 8)); 5512 } 5513 case NEON::BI__builtin_neon_vminvq_s16: { 5514 Int = Intrinsic::aarch64_neon_sminv; 5515 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5516 VTy = 5517 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5518 llvm::Type *Tys[2] = { Ty, VTy }; 5519 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5520 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5521 return Builder.CreateTrunc(Ops[0], 5522 llvm::IntegerType::get(getLLVMContext(), 16)); 5523 } 5524 case NEON::BI__builtin_neon_vmul_n_f64: { 5525 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5526 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5527 return Builder.CreateFMul(Ops[0], RHS); 5528 } 5529 case NEON::BI__builtin_neon_vaddlv_u8: { 5530 Int = Intrinsic::aarch64_neon_uaddlv; 5531 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5532 VTy = 5533 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5534 llvm::Type *Tys[2] = { Ty, VTy }; 5535 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5536 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5537 return Builder.CreateTrunc(Ops[0], 5538 llvm::IntegerType::get(getLLVMContext(), 16)); 5539 } 5540 case NEON::BI__builtin_neon_vaddlv_u16: { 5541 Int = Intrinsic::aarch64_neon_uaddlv; 5542 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5543 VTy = 5544 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5545 llvm::Type *Tys[2] = { Ty, VTy }; 5546 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5547 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5548 } 5549 case NEON::BI__builtin_neon_vaddlvq_u8: { 5550 Int = Intrinsic::aarch64_neon_uaddlv; 5551 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5552 VTy = 5553 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5554 llvm::Type *Tys[2] = { Ty, VTy }; 5555 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5556 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5557 return Builder.CreateTrunc(Ops[0], 5558 llvm::IntegerType::get(getLLVMContext(), 16)); 5559 } 5560 case NEON::BI__builtin_neon_vaddlvq_u16: { 5561 Int = Intrinsic::aarch64_neon_uaddlv; 5562 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5563 VTy = 5564 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5565 llvm::Type *Tys[2] = { Ty, VTy }; 5566 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5567 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5568 } 5569 case NEON::BI__builtin_neon_vaddlv_s8: { 5570 Int = Intrinsic::aarch64_neon_saddlv; 5571 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5572 VTy = 5573 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5574 llvm::Type *Tys[2] = { Ty, VTy }; 5575 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5576 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5577 return Builder.CreateTrunc(Ops[0], 5578 llvm::IntegerType::get(getLLVMContext(), 16)); 5579 } 5580 case NEON::BI__builtin_neon_vaddlv_s16: { 5581 Int = Intrinsic::aarch64_neon_saddlv; 5582 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5583 VTy = 5584 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5585 llvm::Type *Tys[2] = { Ty, VTy }; 5586 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5587 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5588 } 5589 case NEON::BI__builtin_neon_vaddlvq_s8: { 5590 Int = Intrinsic::aarch64_neon_saddlv; 5591 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5592 VTy = 5593 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5594 llvm::Type *Tys[2] = { Ty, VTy }; 5595 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5596 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5597 return Builder.CreateTrunc(Ops[0], 5598 llvm::IntegerType::get(getLLVMContext(), 16)); 5599 } 5600 case NEON::BI__builtin_neon_vaddlvq_s16: { 5601 Int = Intrinsic::aarch64_neon_saddlv; 5602 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5603 VTy = 5604 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5605 llvm::Type *Tys[2] = { Ty, VTy }; 5606 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5607 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5608 } 5609 case NEON::BI__builtin_neon_vsri_n_v: 5610 case NEON::BI__builtin_neon_vsriq_n_v: { 5611 Int = Intrinsic::aarch64_neon_vsri; 5612 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5613 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5614 } 5615 case NEON::BI__builtin_neon_vsli_n_v: 5616 case NEON::BI__builtin_neon_vsliq_n_v: { 5617 Int = Intrinsic::aarch64_neon_vsli; 5618 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5619 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5620 } 5621 case NEON::BI__builtin_neon_vsra_n_v: 5622 case NEON::BI__builtin_neon_vsraq_n_v: 5623 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5624 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5625 return Builder.CreateAdd(Ops[0], Ops[1]); 5626 case NEON::BI__builtin_neon_vrsra_n_v: 5627 case NEON::BI__builtin_neon_vrsraq_n_v: { 5628 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5629 SmallVector<llvm::Value*,2> TmpOps; 5630 TmpOps.push_back(Ops[1]); 5631 TmpOps.push_back(Ops[2]); 5632 Function* F = CGM.getIntrinsic(Int, Ty); 5633 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5634 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5635 return Builder.CreateAdd(Ops[0], tmp); 5636 } 5637 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5638 // of an Align parameter here. 5639 case NEON::BI__builtin_neon_vld1_x2_v: 5640 case NEON::BI__builtin_neon_vld1q_x2_v: 5641 case NEON::BI__builtin_neon_vld1_x3_v: 5642 case NEON::BI__builtin_neon_vld1q_x3_v: 5643 case NEON::BI__builtin_neon_vld1_x4_v: 5644 case NEON::BI__builtin_neon_vld1q_x4_v: { 5645 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5646 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5647 llvm::Type *Tys[2] = { VTy, PTy }; 5648 unsigned Int; 5649 switch (BuiltinID) { 5650 case NEON::BI__builtin_neon_vld1_x2_v: 5651 case NEON::BI__builtin_neon_vld1q_x2_v: 5652 Int = Intrinsic::aarch64_neon_ld1x2; 5653 break; 5654 case NEON::BI__builtin_neon_vld1_x3_v: 5655 case NEON::BI__builtin_neon_vld1q_x3_v: 5656 Int = Intrinsic::aarch64_neon_ld1x3; 5657 break; 5658 case NEON::BI__builtin_neon_vld1_x4_v: 5659 case NEON::BI__builtin_neon_vld1q_x4_v: 5660 Int = Intrinsic::aarch64_neon_ld1x4; 5661 break; 5662 } 5663 Function *F = CGM.getIntrinsic(Int, Tys); 5664 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5665 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5666 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5667 return Builder.CreateStore(Ops[1], Ops[0]); 5668 } 5669 case NEON::BI__builtin_neon_vst1_x2_v: 5670 case NEON::BI__builtin_neon_vst1q_x2_v: 5671 case NEON::BI__builtin_neon_vst1_x3_v: 5672 case NEON::BI__builtin_neon_vst1q_x3_v: 5673 case NEON::BI__builtin_neon_vst1_x4_v: 5674 case NEON::BI__builtin_neon_vst1q_x4_v: { 5675 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5676 llvm::Type *Tys[2] = { VTy, PTy }; 5677 unsigned Int; 5678 switch (BuiltinID) { 5679 case NEON::BI__builtin_neon_vst1_x2_v: 5680 case NEON::BI__builtin_neon_vst1q_x2_v: 5681 Int = Intrinsic::aarch64_neon_st1x2; 5682 break; 5683 case NEON::BI__builtin_neon_vst1_x3_v: 5684 case NEON::BI__builtin_neon_vst1q_x3_v: 5685 Int = Intrinsic::aarch64_neon_st1x3; 5686 break; 5687 case NEON::BI__builtin_neon_vst1_x4_v: 5688 case NEON::BI__builtin_neon_vst1q_x4_v: 5689 Int = Intrinsic::aarch64_neon_st1x4; 5690 break; 5691 } 5692 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5693 IntOps.push_back(Ops[0]); 5694 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5695 } 5696 case NEON::BI__builtin_neon_vld1_v: 5697 case NEON::BI__builtin_neon_vld1q_v: 5698 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5699 return Builder.CreateLoad(Ops[0]); 5700 case NEON::BI__builtin_neon_vst1_v: 5701 case NEON::BI__builtin_neon_vst1q_v: 5702 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5703 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5704 return Builder.CreateStore(Ops[1], Ops[0]); 5705 case NEON::BI__builtin_neon_vld1_lane_v: 5706 case NEON::BI__builtin_neon_vld1q_lane_v: 5707 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5708 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5709 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5710 Ops[0] = Builder.CreateLoad(Ops[0]); 5711 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5712 case NEON::BI__builtin_neon_vld1_dup_v: 5713 case NEON::BI__builtin_neon_vld1q_dup_v: { 5714 Value *V = UndefValue::get(Ty); 5715 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5716 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5717 Ops[0] = Builder.CreateLoad(Ops[0]); 5718 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5719 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5720 return EmitNeonSplat(Ops[0], CI); 5721 } 5722 case NEON::BI__builtin_neon_vst1_lane_v: 5723 case NEON::BI__builtin_neon_vst1q_lane_v: 5724 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5725 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5726 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5727 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5728 case NEON::BI__builtin_neon_vld2_v: 5729 case NEON::BI__builtin_neon_vld2q_v: { 5730 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5731 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5732 llvm::Type *Tys[2] = { VTy, PTy }; 5733 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5734 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5735 Ops[0] = Builder.CreateBitCast(Ops[0], 5736 llvm::PointerType::getUnqual(Ops[1]->getType())); 5737 return Builder.CreateStore(Ops[1], Ops[0]); 5738 } 5739 case NEON::BI__builtin_neon_vld3_v: 5740 case NEON::BI__builtin_neon_vld3q_v: { 5741 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5742 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5743 llvm::Type *Tys[2] = { VTy, PTy }; 5744 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5745 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5746 Ops[0] = Builder.CreateBitCast(Ops[0], 5747 llvm::PointerType::getUnqual(Ops[1]->getType())); 5748 return Builder.CreateStore(Ops[1], Ops[0]); 5749 } 5750 case NEON::BI__builtin_neon_vld4_v: 5751 case NEON::BI__builtin_neon_vld4q_v: { 5752 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5753 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5754 llvm::Type *Tys[2] = { VTy, PTy }; 5755 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5756 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5757 Ops[0] = Builder.CreateBitCast(Ops[0], 5758 llvm::PointerType::getUnqual(Ops[1]->getType())); 5759 return Builder.CreateStore(Ops[1], Ops[0]); 5760 } 5761 case NEON::BI__builtin_neon_vld2_dup_v: 5762 case NEON::BI__builtin_neon_vld2q_dup_v: { 5763 llvm::Type *PTy = 5764 llvm::PointerType::getUnqual(VTy->getElementType()); 5765 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5766 llvm::Type *Tys[2] = { VTy, PTy }; 5767 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5768 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5769 Ops[0] = Builder.CreateBitCast(Ops[0], 5770 llvm::PointerType::getUnqual(Ops[1]->getType())); 5771 return Builder.CreateStore(Ops[1], Ops[0]); 5772 } 5773 case NEON::BI__builtin_neon_vld3_dup_v: 5774 case NEON::BI__builtin_neon_vld3q_dup_v: { 5775 llvm::Type *PTy = 5776 llvm::PointerType::getUnqual(VTy->getElementType()); 5777 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5778 llvm::Type *Tys[2] = { VTy, PTy }; 5779 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5780 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5781 Ops[0] = Builder.CreateBitCast(Ops[0], 5782 llvm::PointerType::getUnqual(Ops[1]->getType())); 5783 return Builder.CreateStore(Ops[1], Ops[0]); 5784 } 5785 case NEON::BI__builtin_neon_vld4_dup_v: 5786 case NEON::BI__builtin_neon_vld4q_dup_v: { 5787 llvm::Type *PTy = 5788 llvm::PointerType::getUnqual(VTy->getElementType()); 5789 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5790 llvm::Type *Tys[2] = { VTy, PTy }; 5791 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5792 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5793 Ops[0] = Builder.CreateBitCast(Ops[0], 5794 llvm::PointerType::getUnqual(Ops[1]->getType())); 5795 return Builder.CreateStore(Ops[1], Ops[0]); 5796 } 5797 case NEON::BI__builtin_neon_vld2_lane_v: 5798 case NEON::BI__builtin_neon_vld2q_lane_v: { 5799 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5800 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5801 Ops.push_back(Ops[1]); 5802 Ops.erase(Ops.begin()+1); 5803 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5804 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5805 Ops[3] = Builder.CreateZExt(Ops[3], 5806 llvm::IntegerType::get(getLLVMContext(), 64)); 5807 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5808 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5809 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5810 return Builder.CreateStore(Ops[1], Ops[0]); 5811 } 5812 case NEON::BI__builtin_neon_vld3_lane_v: 5813 case NEON::BI__builtin_neon_vld3q_lane_v: { 5814 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5815 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5816 Ops.push_back(Ops[1]); 5817 Ops.erase(Ops.begin()+1); 5818 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5819 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5820 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5821 Ops[4] = Builder.CreateZExt(Ops[4], 5822 llvm::IntegerType::get(getLLVMContext(), 64)); 5823 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5824 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5825 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5826 return Builder.CreateStore(Ops[1], Ops[0]); 5827 } 5828 case NEON::BI__builtin_neon_vld4_lane_v: 5829 case NEON::BI__builtin_neon_vld4q_lane_v: { 5830 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5831 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5832 Ops.push_back(Ops[1]); 5833 Ops.erase(Ops.begin()+1); 5834 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5835 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5836 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5837 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5838 Ops[5] = Builder.CreateZExt(Ops[5], 5839 llvm::IntegerType::get(getLLVMContext(), 64)); 5840 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5841 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5842 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5843 return Builder.CreateStore(Ops[1], Ops[0]); 5844 } 5845 case NEON::BI__builtin_neon_vst2_v: 5846 case NEON::BI__builtin_neon_vst2q_v: { 5847 Ops.push_back(Ops[0]); 5848 Ops.erase(Ops.begin()); 5849 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5850 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5851 Ops, ""); 5852 } 5853 case NEON::BI__builtin_neon_vst2_lane_v: 5854 case NEON::BI__builtin_neon_vst2q_lane_v: { 5855 Ops.push_back(Ops[0]); 5856 Ops.erase(Ops.begin()); 5857 Ops[2] = Builder.CreateZExt(Ops[2], 5858 llvm::IntegerType::get(getLLVMContext(), 64)); 5859 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5860 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5861 Ops, ""); 5862 } 5863 case NEON::BI__builtin_neon_vst3_v: 5864 case NEON::BI__builtin_neon_vst3q_v: { 5865 Ops.push_back(Ops[0]); 5866 Ops.erase(Ops.begin()); 5867 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5868 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5869 Ops, ""); 5870 } 5871 case NEON::BI__builtin_neon_vst3_lane_v: 5872 case NEON::BI__builtin_neon_vst3q_lane_v: { 5873 Ops.push_back(Ops[0]); 5874 Ops.erase(Ops.begin()); 5875 Ops[3] = Builder.CreateZExt(Ops[3], 5876 llvm::IntegerType::get(getLLVMContext(), 64)); 5877 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5878 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5879 Ops, ""); 5880 } 5881 case NEON::BI__builtin_neon_vst4_v: 5882 case NEON::BI__builtin_neon_vst4q_v: { 5883 Ops.push_back(Ops[0]); 5884 Ops.erase(Ops.begin()); 5885 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5886 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5887 Ops, ""); 5888 } 5889 case NEON::BI__builtin_neon_vst4_lane_v: 5890 case NEON::BI__builtin_neon_vst4q_lane_v: { 5891 Ops.push_back(Ops[0]); 5892 Ops.erase(Ops.begin()); 5893 Ops[4] = Builder.CreateZExt(Ops[4], 5894 llvm::IntegerType::get(getLLVMContext(), 64)); 5895 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5896 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5897 Ops, ""); 5898 } 5899 case NEON::BI__builtin_neon_vtrn_v: 5900 case NEON::BI__builtin_neon_vtrnq_v: { 5901 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5902 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5903 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5904 Value *SV = nullptr; 5905 5906 for (unsigned vi = 0; vi != 2; ++vi) { 5907 SmallVector<Constant*, 16> Indices; 5908 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5909 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5910 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5911 } 5912 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5913 SV = llvm::ConstantVector::get(Indices); 5914 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5915 SV = Builder.CreateStore(SV, Addr); 5916 } 5917 return SV; 5918 } 5919 case NEON::BI__builtin_neon_vuzp_v: 5920 case NEON::BI__builtin_neon_vuzpq_v: { 5921 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5922 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5923 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5924 Value *SV = nullptr; 5925 5926 for (unsigned vi = 0; vi != 2; ++vi) { 5927 SmallVector<Constant*, 16> Indices; 5928 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5929 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5930 5931 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5932 SV = llvm::ConstantVector::get(Indices); 5933 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5934 SV = Builder.CreateStore(SV, Addr); 5935 } 5936 return SV; 5937 } 5938 case NEON::BI__builtin_neon_vzip_v: 5939 case NEON::BI__builtin_neon_vzipq_v: { 5940 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5941 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5942 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5943 Value *SV = nullptr; 5944 5945 for (unsigned vi = 0; vi != 2; ++vi) { 5946 SmallVector<Constant*, 16> Indices; 5947 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5948 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5949 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5950 } 5951 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5952 SV = llvm::ConstantVector::get(Indices); 5953 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5954 SV = Builder.CreateStore(SV, Addr); 5955 } 5956 return SV; 5957 } 5958 case NEON::BI__builtin_neon_vqtbl1q_v: { 5959 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5960 Ops, "vtbl1"); 5961 } 5962 case NEON::BI__builtin_neon_vqtbl2q_v: { 5963 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5964 Ops, "vtbl2"); 5965 } 5966 case NEON::BI__builtin_neon_vqtbl3q_v: { 5967 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5968 Ops, "vtbl3"); 5969 } 5970 case NEON::BI__builtin_neon_vqtbl4q_v: { 5971 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5972 Ops, "vtbl4"); 5973 } 5974 case NEON::BI__builtin_neon_vqtbx1q_v: { 5975 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5976 Ops, "vtbx1"); 5977 } 5978 case NEON::BI__builtin_neon_vqtbx2q_v: { 5979 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5980 Ops, "vtbx2"); 5981 } 5982 case NEON::BI__builtin_neon_vqtbx3q_v: { 5983 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5984 Ops, "vtbx3"); 5985 } 5986 case NEON::BI__builtin_neon_vqtbx4q_v: { 5987 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5988 Ops, "vtbx4"); 5989 } 5990 case NEON::BI__builtin_neon_vsqadd_v: 5991 case NEON::BI__builtin_neon_vsqaddq_v: { 5992 Int = Intrinsic::aarch64_neon_usqadd; 5993 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5994 } 5995 case NEON::BI__builtin_neon_vuqadd_v: 5996 case NEON::BI__builtin_neon_vuqaddq_v: { 5997 Int = Intrinsic::aarch64_neon_suqadd; 5998 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5999 } 6000 } 6001 } 6002 6003 llvm::Value *CodeGenFunction:: 6004 BuildVector(ArrayRef<llvm::Value*> Ops) { 6005 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6006 "Not a power-of-two sized vector!"); 6007 bool AllConstants = true; 6008 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6009 AllConstants &= isa<Constant>(Ops[i]); 6010 6011 // If this is a constant vector, create a ConstantVector. 6012 if (AllConstants) { 6013 SmallVector<llvm::Constant*, 16> CstOps; 6014 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6015 CstOps.push_back(cast<Constant>(Ops[i])); 6016 return llvm::ConstantVector::get(CstOps); 6017 } 6018 6019 // Otherwise, insertelement the values to build the vector. 6020 Value *Result = 6021 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6022 6023 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6024 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6025 6026 return Result; 6027 } 6028 6029 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6030 const CallExpr *E) { 6031 SmallVector<Value*, 4> Ops; 6032 6033 // Find out if any arguments are required to be integer constant expressions. 6034 unsigned ICEArguments = 0; 6035 ASTContext::GetBuiltinTypeError Error; 6036 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6037 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6038 6039 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6040 // If this is a normal argument, just emit it as a scalar. 6041 if ((ICEArguments & (1 << i)) == 0) { 6042 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6043 continue; 6044 } 6045 6046 // If this is required to be a constant, constant fold it so that we know 6047 // that the generated intrinsic gets a ConstantInt. 6048 llvm::APSInt Result; 6049 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6050 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6051 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6052 } 6053 6054 switch (BuiltinID) { 6055 default: return nullptr; 6056 case X86::BI_mm_prefetch: { 6057 Value *Address = EmitScalarExpr(E->getArg(0)); 6058 Value *RW = ConstantInt::get(Int32Ty, 0); 6059 Value *Locality = EmitScalarExpr(E->getArg(1)); 6060 Value *Data = ConstantInt::get(Int32Ty, 1); 6061 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6062 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6063 } 6064 case X86::BI__builtin_ia32_vec_init_v8qi: 6065 case X86::BI__builtin_ia32_vec_init_v4hi: 6066 case X86::BI__builtin_ia32_vec_init_v2si: 6067 return Builder.CreateBitCast(BuildVector(Ops), 6068 llvm::Type::getX86_MMXTy(getLLVMContext())); 6069 case X86::BI__builtin_ia32_vec_ext_v2si: 6070 return Builder.CreateExtractElement(Ops[0], 6071 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6072 case X86::BI__builtin_ia32_ldmxcsr: { 6073 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 6074 Builder.CreateStore(Ops[0], Tmp); 6075 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6076 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6077 } 6078 case X86::BI__builtin_ia32_stmxcsr: { 6079 Value *Tmp = CreateMemTemp(E->getType()); 6080 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6081 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6082 return Builder.CreateLoad(Tmp, "stmxcsr"); 6083 } 6084 case X86::BI__builtin_ia32_storehps: 6085 case X86::BI__builtin_ia32_storelps: { 6086 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6087 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6088 6089 // cast val v2i64 6090 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6091 6092 // extract (0, 1) 6093 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6094 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6095 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6096 6097 // cast pointer to i64 & store 6098 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6099 return Builder.CreateStore(Ops[1], Ops[0]); 6100 } 6101 case X86::BI__builtin_ia32_palignr128: 6102 case X86::BI__builtin_ia32_palignr256: { 6103 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6104 6105 unsigned NumElts = 6106 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6107 assert(NumElts % 16 == 0); 6108 unsigned NumLanes = NumElts / 16; 6109 unsigned NumLaneElts = NumElts / NumLanes; 6110 6111 // If palignr is shifting the pair of vectors more than the size of two 6112 // lanes, emit zero. 6113 if (ShiftVal >= (2 * NumLaneElts)) 6114 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6115 6116 // If palignr is shifting the pair of input vectors more than one lane, 6117 // but less than two lanes, convert to shifting in zeroes. 6118 if (ShiftVal > NumLaneElts) { 6119 ShiftVal -= NumLaneElts; 6120 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6121 } 6122 6123 SmallVector<llvm::Constant*, 32> Indices; 6124 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6125 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6126 for (unsigned i = 0; i != NumLaneElts; ++i) { 6127 unsigned Idx = ShiftVal + i; 6128 if (Idx >= NumLaneElts) 6129 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6130 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6131 } 6132 } 6133 6134 Value* SV = llvm::ConstantVector::get(Indices); 6135 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6136 } 6137 case X86::BI__builtin_ia32_pslldqi256: { 6138 // Shift value is in bits so divide by 8. 6139 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6140 6141 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6142 if (shiftVal >= 16) 6143 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6144 6145 SmallVector<llvm::Constant*, 32> Indices; 6146 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6147 for (unsigned l = 0; l != 32; l += 16) { 6148 for (unsigned i = 0; i != 16; ++i) { 6149 unsigned Idx = 32 + i - shiftVal; 6150 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6151 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6152 } 6153 } 6154 6155 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6156 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6157 Value *Zero = llvm::Constant::getNullValue(VecTy); 6158 6159 Value *SV = llvm::ConstantVector::get(Indices); 6160 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6161 llvm::Type *ResultType = ConvertType(E->getType()); 6162 return Builder.CreateBitCast(SV, ResultType, "cast"); 6163 } 6164 case X86::BI__builtin_ia32_psrldqi256: { 6165 // Shift value is in bits so divide by 8. 6166 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6167 6168 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6169 if (shiftVal >= 16) 6170 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6171 6172 SmallVector<llvm::Constant*, 32> Indices; 6173 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6174 for (unsigned l = 0; l != 32; l += 16) { 6175 for (unsigned i = 0; i != 16; ++i) { 6176 unsigned Idx = i + shiftVal; 6177 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6178 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6179 } 6180 } 6181 6182 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6183 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6184 Value *Zero = llvm::Constant::getNullValue(VecTy); 6185 6186 Value *SV = llvm::ConstantVector::get(Indices); 6187 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6188 llvm::Type *ResultType = ConvertType(E->getType()); 6189 return Builder.CreateBitCast(SV, ResultType, "cast"); 6190 } 6191 case X86::BI__builtin_ia32_movntps: 6192 case X86::BI__builtin_ia32_movntps256: 6193 case X86::BI__builtin_ia32_movntpd: 6194 case X86::BI__builtin_ia32_movntpd256: 6195 case X86::BI__builtin_ia32_movntdq: 6196 case X86::BI__builtin_ia32_movntdq256: 6197 case X86::BI__builtin_ia32_movnti: 6198 case X86::BI__builtin_ia32_movnti64: { 6199 llvm::MDNode *Node = llvm::MDNode::get( 6200 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6201 6202 // Convert the type of the pointer to a pointer to the stored type. 6203 Value *BC = Builder.CreateBitCast(Ops[0], 6204 llvm::PointerType::getUnqual(Ops[1]->getType()), 6205 "cast"); 6206 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6207 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6208 6209 // If the operand is an integer, we can't assume alignment. Otherwise, 6210 // assume natural alignment. 6211 QualType ArgTy = E->getArg(1)->getType(); 6212 unsigned Align; 6213 if (ArgTy->isIntegerType()) 6214 Align = 1; 6215 else 6216 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6217 SI->setAlignment(Align); 6218 return SI; 6219 } 6220 // 3DNow! 6221 case X86::BI__builtin_ia32_pswapdsf: 6222 case X86::BI__builtin_ia32_pswapdsi: { 6223 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6224 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6225 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6226 return Builder.CreateCall(F, Ops, "pswapd"); 6227 } 6228 case X86::BI__builtin_ia32_rdrand16_step: 6229 case X86::BI__builtin_ia32_rdrand32_step: 6230 case X86::BI__builtin_ia32_rdrand64_step: 6231 case X86::BI__builtin_ia32_rdseed16_step: 6232 case X86::BI__builtin_ia32_rdseed32_step: 6233 case X86::BI__builtin_ia32_rdseed64_step: { 6234 Intrinsic::ID ID; 6235 switch (BuiltinID) { 6236 default: llvm_unreachable("Unsupported intrinsic!"); 6237 case X86::BI__builtin_ia32_rdrand16_step: 6238 ID = Intrinsic::x86_rdrand_16; 6239 break; 6240 case X86::BI__builtin_ia32_rdrand32_step: 6241 ID = Intrinsic::x86_rdrand_32; 6242 break; 6243 case X86::BI__builtin_ia32_rdrand64_step: 6244 ID = Intrinsic::x86_rdrand_64; 6245 break; 6246 case X86::BI__builtin_ia32_rdseed16_step: 6247 ID = Intrinsic::x86_rdseed_16; 6248 break; 6249 case X86::BI__builtin_ia32_rdseed32_step: 6250 ID = Intrinsic::x86_rdseed_32; 6251 break; 6252 case X86::BI__builtin_ia32_rdseed64_step: 6253 ID = Intrinsic::x86_rdseed_64; 6254 break; 6255 } 6256 6257 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {}); 6258 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6259 return Builder.CreateExtractValue(Call, 1); 6260 } 6261 // SSE comparison intrisics 6262 case X86::BI__builtin_ia32_cmpeqps: 6263 case X86::BI__builtin_ia32_cmpltps: 6264 case X86::BI__builtin_ia32_cmpleps: 6265 case X86::BI__builtin_ia32_cmpunordps: 6266 case X86::BI__builtin_ia32_cmpneqps: 6267 case X86::BI__builtin_ia32_cmpnltps: 6268 case X86::BI__builtin_ia32_cmpnleps: 6269 case X86::BI__builtin_ia32_cmpordps: 6270 case X86::BI__builtin_ia32_cmpeqss: 6271 case X86::BI__builtin_ia32_cmpltss: 6272 case X86::BI__builtin_ia32_cmpless: 6273 case X86::BI__builtin_ia32_cmpunordss: 6274 case X86::BI__builtin_ia32_cmpneqss: 6275 case X86::BI__builtin_ia32_cmpnltss: 6276 case X86::BI__builtin_ia32_cmpnless: 6277 case X86::BI__builtin_ia32_cmpordss: 6278 case X86::BI__builtin_ia32_cmpeqpd: 6279 case X86::BI__builtin_ia32_cmpltpd: 6280 case X86::BI__builtin_ia32_cmplepd: 6281 case X86::BI__builtin_ia32_cmpunordpd: 6282 case X86::BI__builtin_ia32_cmpneqpd: 6283 case X86::BI__builtin_ia32_cmpnltpd: 6284 case X86::BI__builtin_ia32_cmpnlepd: 6285 case X86::BI__builtin_ia32_cmpordpd: 6286 case X86::BI__builtin_ia32_cmpeqsd: 6287 case X86::BI__builtin_ia32_cmpltsd: 6288 case X86::BI__builtin_ia32_cmplesd: 6289 case X86::BI__builtin_ia32_cmpunordsd: 6290 case X86::BI__builtin_ia32_cmpneqsd: 6291 case X86::BI__builtin_ia32_cmpnltsd: 6292 case X86::BI__builtin_ia32_cmpnlesd: 6293 case X86::BI__builtin_ia32_cmpordsd: 6294 // These exist so that the builtin that takes an immediate can be bounds 6295 // checked by clang to avoid passing bad immediates to the backend. Since 6296 // AVX has a larger immediate than SSE we would need separate builtins to 6297 // do the different bounds checking. Rather than create a clang specific 6298 // SSE only builtin, this implements eight separate builtins to match gcc 6299 // implementation. 6300 6301 // Choose the immediate. 6302 unsigned Imm; 6303 switch (BuiltinID) { 6304 default: llvm_unreachable("Unsupported intrinsic!"); 6305 case X86::BI__builtin_ia32_cmpeqps: 6306 case X86::BI__builtin_ia32_cmpeqss: 6307 case X86::BI__builtin_ia32_cmpeqpd: 6308 case X86::BI__builtin_ia32_cmpeqsd: 6309 Imm = 0; 6310 break; 6311 case X86::BI__builtin_ia32_cmpltps: 6312 case X86::BI__builtin_ia32_cmpltss: 6313 case X86::BI__builtin_ia32_cmpltpd: 6314 case X86::BI__builtin_ia32_cmpltsd: 6315 Imm = 1; 6316 break; 6317 case X86::BI__builtin_ia32_cmpleps: 6318 case X86::BI__builtin_ia32_cmpless: 6319 case X86::BI__builtin_ia32_cmplepd: 6320 case X86::BI__builtin_ia32_cmplesd: 6321 Imm = 2; 6322 break; 6323 case X86::BI__builtin_ia32_cmpunordps: 6324 case X86::BI__builtin_ia32_cmpunordss: 6325 case X86::BI__builtin_ia32_cmpunordpd: 6326 case X86::BI__builtin_ia32_cmpunordsd: 6327 Imm = 3; 6328 break; 6329 case X86::BI__builtin_ia32_cmpneqps: 6330 case X86::BI__builtin_ia32_cmpneqss: 6331 case X86::BI__builtin_ia32_cmpneqpd: 6332 case X86::BI__builtin_ia32_cmpneqsd: 6333 Imm = 4; 6334 break; 6335 case X86::BI__builtin_ia32_cmpnltps: 6336 case X86::BI__builtin_ia32_cmpnltss: 6337 case X86::BI__builtin_ia32_cmpnltpd: 6338 case X86::BI__builtin_ia32_cmpnltsd: 6339 Imm = 5; 6340 break; 6341 case X86::BI__builtin_ia32_cmpnleps: 6342 case X86::BI__builtin_ia32_cmpnless: 6343 case X86::BI__builtin_ia32_cmpnlepd: 6344 case X86::BI__builtin_ia32_cmpnlesd: 6345 Imm = 6; 6346 break; 6347 case X86::BI__builtin_ia32_cmpordps: 6348 case X86::BI__builtin_ia32_cmpordss: 6349 case X86::BI__builtin_ia32_cmpordpd: 6350 case X86::BI__builtin_ia32_cmpordsd: 6351 Imm = 7; 6352 break; 6353 } 6354 6355 // Choose the intrinsic ID. 6356 const char *name; 6357 Intrinsic::ID ID; 6358 switch (BuiltinID) { 6359 default: llvm_unreachable("Unsupported intrinsic!"); 6360 case X86::BI__builtin_ia32_cmpeqps: 6361 case X86::BI__builtin_ia32_cmpltps: 6362 case X86::BI__builtin_ia32_cmpleps: 6363 case X86::BI__builtin_ia32_cmpunordps: 6364 case X86::BI__builtin_ia32_cmpneqps: 6365 case X86::BI__builtin_ia32_cmpnltps: 6366 case X86::BI__builtin_ia32_cmpnleps: 6367 case X86::BI__builtin_ia32_cmpordps: 6368 name = "cmpps"; 6369 ID = Intrinsic::x86_sse_cmp_ps; 6370 break; 6371 case X86::BI__builtin_ia32_cmpeqss: 6372 case X86::BI__builtin_ia32_cmpltss: 6373 case X86::BI__builtin_ia32_cmpless: 6374 case X86::BI__builtin_ia32_cmpunordss: 6375 case X86::BI__builtin_ia32_cmpneqss: 6376 case X86::BI__builtin_ia32_cmpnltss: 6377 case X86::BI__builtin_ia32_cmpnless: 6378 case X86::BI__builtin_ia32_cmpordss: 6379 name = "cmpss"; 6380 ID = Intrinsic::x86_sse_cmp_ss; 6381 break; 6382 case X86::BI__builtin_ia32_cmpeqpd: 6383 case X86::BI__builtin_ia32_cmpltpd: 6384 case X86::BI__builtin_ia32_cmplepd: 6385 case X86::BI__builtin_ia32_cmpunordpd: 6386 case X86::BI__builtin_ia32_cmpneqpd: 6387 case X86::BI__builtin_ia32_cmpnltpd: 6388 case X86::BI__builtin_ia32_cmpnlepd: 6389 case X86::BI__builtin_ia32_cmpordpd: 6390 name = "cmppd"; 6391 ID = Intrinsic::x86_sse2_cmp_pd; 6392 break; 6393 case X86::BI__builtin_ia32_cmpeqsd: 6394 case X86::BI__builtin_ia32_cmpltsd: 6395 case X86::BI__builtin_ia32_cmplesd: 6396 case X86::BI__builtin_ia32_cmpunordsd: 6397 case X86::BI__builtin_ia32_cmpneqsd: 6398 case X86::BI__builtin_ia32_cmpnltsd: 6399 case X86::BI__builtin_ia32_cmpnlesd: 6400 case X86::BI__builtin_ia32_cmpordsd: 6401 name = "cmpsd"; 6402 ID = Intrinsic::x86_sse2_cmp_sd; 6403 break; 6404 } 6405 6406 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6407 llvm::Function *F = CGM.getIntrinsic(ID); 6408 return Builder.CreateCall(F, Ops, name); 6409 } 6410 } 6411 6412 6413 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6414 const CallExpr *E) { 6415 SmallVector<Value*, 4> Ops; 6416 6417 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6418 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6419 6420 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6421 6422 switch (BuiltinID) { 6423 default: return nullptr; 6424 6425 // vec_ld, vec_lvsl, vec_lvsr 6426 case PPC::BI__builtin_altivec_lvx: 6427 case PPC::BI__builtin_altivec_lvxl: 6428 case PPC::BI__builtin_altivec_lvebx: 6429 case PPC::BI__builtin_altivec_lvehx: 6430 case PPC::BI__builtin_altivec_lvewx: 6431 case PPC::BI__builtin_altivec_lvsl: 6432 case PPC::BI__builtin_altivec_lvsr: 6433 case PPC::BI__builtin_vsx_lxvd2x: 6434 case PPC::BI__builtin_vsx_lxvw4x: 6435 { 6436 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6437 6438 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6439 Ops.pop_back(); 6440 6441 switch (BuiltinID) { 6442 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6443 case PPC::BI__builtin_altivec_lvx: 6444 ID = Intrinsic::ppc_altivec_lvx; 6445 break; 6446 case PPC::BI__builtin_altivec_lvxl: 6447 ID = Intrinsic::ppc_altivec_lvxl; 6448 break; 6449 case PPC::BI__builtin_altivec_lvebx: 6450 ID = Intrinsic::ppc_altivec_lvebx; 6451 break; 6452 case PPC::BI__builtin_altivec_lvehx: 6453 ID = Intrinsic::ppc_altivec_lvehx; 6454 break; 6455 case PPC::BI__builtin_altivec_lvewx: 6456 ID = Intrinsic::ppc_altivec_lvewx; 6457 break; 6458 case PPC::BI__builtin_altivec_lvsl: 6459 ID = Intrinsic::ppc_altivec_lvsl; 6460 break; 6461 case PPC::BI__builtin_altivec_lvsr: 6462 ID = Intrinsic::ppc_altivec_lvsr; 6463 break; 6464 case PPC::BI__builtin_vsx_lxvd2x: 6465 ID = Intrinsic::ppc_vsx_lxvd2x; 6466 break; 6467 case PPC::BI__builtin_vsx_lxvw4x: 6468 ID = Intrinsic::ppc_vsx_lxvw4x; 6469 break; 6470 } 6471 llvm::Function *F = CGM.getIntrinsic(ID); 6472 return Builder.CreateCall(F, Ops, ""); 6473 } 6474 6475 // vec_st 6476 case PPC::BI__builtin_altivec_stvx: 6477 case PPC::BI__builtin_altivec_stvxl: 6478 case PPC::BI__builtin_altivec_stvebx: 6479 case PPC::BI__builtin_altivec_stvehx: 6480 case PPC::BI__builtin_altivec_stvewx: 6481 case PPC::BI__builtin_vsx_stxvd2x: 6482 case PPC::BI__builtin_vsx_stxvw4x: 6483 { 6484 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6485 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6486 Ops.pop_back(); 6487 6488 switch (BuiltinID) { 6489 default: llvm_unreachable("Unsupported st intrinsic!"); 6490 case PPC::BI__builtin_altivec_stvx: 6491 ID = Intrinsic::ppc_altivec_stvx; 6492 break; 6493 case PPC::BI__builtin_altivec_stvxl: 6494 ID = Intrinsic::ppc_altivec_stvxl; 6495 break; 6496 case PPC::BI__builtin_altivec_stvebx: 6497 ID = Intrinsic::ppc_altivec_stvebx; 6498 break; 6499 case PPC::BI__builtin_altivec_stvehx: 6500 ID = Intrinsic::ppc_altivec_stvehx; 6501 break; 6502 case PPC::BI__builtin_altivec_stvewx: 6503 ID = Intrinsic::ppc_altivec_stvewx; 6504 break; 6505 case PPC::BI__builtin_vsx_stxvd2x: 6506 ID = Intrinsic::ppc_vsx_stxvd2x; 6507 break; 6508 case PPC::BI__builtin_vsx_stxvw4x: 6509 ID = Intrinsic::ppc_vsx_stxvw4x; 6510 break; 6511 } 6512 llvm::Function *F = CGM.getIntrinsic(ID); 6513 return Builder.CreateCall(F, Ops, ""); 6514 } 6515 } 6516 } 6517 6518 // Emit an intrinsic that has 1 float or double. 6519 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6520 const CallExpr *E, 6521 unsigned IntrinsicID) { 6522 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6523 6524 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6525 return CGF.Builder.CreateCall(F, Src0); 6526 } 6527 6528 // Emit an intrinsic that has 3 float or double operands. 6529 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6530 const CallExpr *E, 6531 unsigned IntrinsicID) { 6532 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6533 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6534 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6535 6536 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6537 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6538 } 6539 6540 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6541 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6542 const CallExpr *E, 6543 unsigned IntrinsicID) { 6544 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6545 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6546 6547 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6548 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6549 } 6550 6551 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 6552 const CallExpr *E) { 6553 switch (BuiltinID) { 6554 case AMDGPU::BI__builtin_amdgpu_div_scale: 6555 case AMDGPU::BI__builtin_amdgpu_div_scalef: { 6556 // Translate from the intrinsics's struct return to the builtin's out 6557 // argument. 6558 6559 std::pair<llvm::Value *, unsigned> FlagOutPtr 6560 = EmitPointerWithAlignment(E->getArg(3)); 6561 6562 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6563 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6564 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6565 6566 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6567 X->getType()); 6568 6569 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6570 6571 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6572 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6573 6574 llvm::Type *RealFlagType 6575 = FlagOutPtr.first->getType()->getPointerElementType(); 6576 6577 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6578 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6579 FlagStore->setAlignment(FlagOutPtr.second); 6580 return Result; 6581 } 6582 case AMDGPU::BI__builtin_amdgpu_div_fmas: 6583 case AMDGPU::BI__builtin_amdgpu_div_fmasf: { 6584 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6585 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6586 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6587 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6588 6589 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6590 Src0->getType()); 6591 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6592 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6593 } 6594 case AMDGPU::BI__builtin_amdgpu_div_fixup: 6595 case AMDGPU::BI__builtin_amdgpu_div_fixupf: 6596 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6597 case AMDGPU::BI__builtin_amdgpu_trig_preop: 6598 case AMDGPU::BI__builtin_amdgpu_trig_preopf: 6599 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6600 case AMDGPU::BI__builtin_amdgpu_rcp: 6601 case AMDGPU::BI__builtin_amdgpu_rcpf: 6602 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6603 case AMDGPU::BI__builtin_amdgpu_rsq: 6604 case AMDGPU::BI__builtin_amdgpu_rsqf: 6605 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6606 case AMDGPU::BI__builtin_amdgpu_rsq_clamped: 6607 case AMDGPU::BI__builtin_amdgpu_rsq_clampedf: 6608 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6609 case AMDGPU::BI__builtin_amdgpu_ldexp: 6610 case AMDGPU::BI__builtin_amdgpu_ldexpf: 6611 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6612 case AMDGPU::BI__builtin_amdgpu_class: 6613 case AMDGPU::BI__builtin_amdgpu_classf: 6614 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6615 default: 6616 return nullptr; 6617 } 6618 } 6619 6620 /// Handle a SystemZ function in which the final argument is a pointer 6621 /// to an int that receives the post-instruction CC value. At the LLVM level 6622 /// this is represented as a function that returns a {result, cc} pair. 6623 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6624 unsigned IntrinsicID, 6625 const CallExpr *E) { 6626 unsigned NumArgs = E->getNumArgs() - 1; 6627 SmallVector<Value *, 8> Args(NumArgs); 6628 for (unsigned I = 0; I < NumArgs; ++I) 6629 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6630 Value *CCPtr = CGF.EmitScalarExpr(E->getArg(NumArgs)); 6631 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6632 Value *Call = CGF.Builder.CreateCall(F, Args); 6633 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6634 CGF.Builder.CreateStore(CC, CCPtr); 6635 return CGF.Builder.CreateExtractValue(Call, 0); 6636 } 6637 6638 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6639 const CallExpr *E) { 6640 switch (BuiltinID) { 6641 case SystemZ::BI__builtin_tbegin: { 6642 Value *TDB = EmitScalarExpr(E->getArg(0)); 6643 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6644 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6645 return Builder.CreateCall(F, {TDB, Control}); 6646 } 6647 case SystemZ::BI__builtin_tbegin_nofloat: { 6648 Value *TDB = EmitScalarExpr(E->getArg(0)); 6649 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6650 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6651 return Builder.CreateCall(F, {TDB, Control}); 6652 } 6653 case SystemZ::BI__builtin_tbeginc: { 6654 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6655 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6656 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6657 return Builder.CreateCall(F, {TDB, Control}); 6658 } 6659 case SystemZ::BI__builtin_tabort: { 6660 Value *Data = EmitScalarExpr(E->getArg(0)); 6661 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6662 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6663 } 6664 case SystemZ::BI__builtin_non_tx_store: { 6665 Value *Address = EmitScalarExpr(E->getArg(0)); 6666 Value *Data = EmitScalarExpr(E->getArg(1)); 6667 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6668 return Builder.CreateCall(F, {Data, Address}); 6669 } 6670 6671 // Vector builtins. Note that most vector builtins are mapped automatically 6672 // to target-specific LLVM intrinsics. The ones handled specially here can 6673 // be represented via standard LLVM IR, which is preferable to enable common 6674 // LLVM optimizations. 6675 6676 case SystemZ::BI__builtin_s390_vpopctb: 6677 case SystemZ::BI__builtin_s390_vpopcth: 6678 case SystemZ::BI__builtin_s390_vpopctf: 6679 case SystemZ::BI__builtin_s390_vpopctg: { 6680 llvm::Type *ResultType = ConvertType(E->getType()); 6681 Value *X = EmitScalarExpr(E->getArg(0)); 6682 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6683 return Builder.CreateCall(F, X); 6684 } 6685 6686 case SystemZ::BI__builtin_s390_vclzb: 6687 case SystemZ::BI__builtin_s390_vclzh: 6688 case SystemZ::BI__builtin_s390_vclzf: 6689 case SystemZ::BI__builtin_s390_vclzg: { 6690 llvm::Type *ResultType = ConvertType(E->getType()); 6691 Value *X = EmitScalarExpr(E->getArg(0)); 6692 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6693 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6694 return Builder.CreateCall(F, {X, Undef}); 6695 } 6696 6697 case SystemZ::BI__builtin_s390_vctzb: 6698 case SystemZ::BI__builtin_s390_vctzh: 6699 case SystemZ::BI__builtin_s390_vctzf: 6700 case SystemZ::BI__builtin_s390_vctzg: { 6701 llvm::Type *ResultType = ConvertType(E->getType()); 6702 Value *X = EmitScalarExpr(E->getArg(0)); 6703 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6704 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6705 return Builder.CreateCall(F, {X, Undef}); 6706 } 6707 6708 case SystemZ::BI__builtin_s390_vfsqdb: { 6709 llvm::Type *ResultType = ConvertType(E->getType()); 6710 Value *X = EmitScalarExpr(E->getArg(0)); 6711 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6712 return Builder.CreateCall(F, X); 6713 } 6714 case SystemZ::BI__builtin_s390_vfmadb: { 6715 llvm::Type *ResultType = ConvertType(E->getType()); 6716 Value *X = EmitScalarExpr(E->getArg(0)); 6717 Value *Y = EmitScalarExpr(E->getArg(1)); 6718 Value *Z = EmitScalarExpr(E->getArg(2)); 6719 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6720 return Builder.CreateCall(F, {X, Y, Z}); 6721 } 6722 case SystemZ::BI__builtin_s390_vfmsdb: { 6723 llvm::Type *ResultType = ConvertType(E->getType()); 6724 Value *X = EmitScalarExpr(E->getArg(0)); 6725 Value *Y = EmitScalarExpr(E->getArg(1)); 6726 Value *Z = EmitScalarExpr(E->getArg(2)); 6727 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6728 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6729 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6730 } 6731 case SystemZ::BI__builtin_s390_vflpdb: { 6732 llvm::Type *ResultType = ConvertType(E->getType()); 6733 Value *X = EmitScalarExpr(E->getArg(0)); 6734 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6735 return Builder.CreateCall(F, X); 6736 } 6737 case SystemZ::BI__builtin_s390_vflndb: { 6738 llvm::Type *ResultType = ConvertType(E->getType()); 6739 Value *X = EmitScalarExpr(E->getArg(0)); 6740 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6741 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6742 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 6743 } 6744 case SystemZ::BI__builtin_s390_vfidb: { 6745 llvm::Type *ResultType = ConvertType(E->getType()); 6746 Value *X = EmitScalarExpr(E->getArg(0)); 6747 // Constant-fold the M4 and M5 mask arguments. 6748 llvm::APSInt M4, M5; 6749 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 6750 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 6751 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 6752 (void)IsConstM4; (void)IsConstM5; 6753 // Check whether this instance of vfidb can be represented via a LLVM 6754 // standard intrinsic. We only support some combinations of M4 and M5. 6755 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6756 switch (M4.getZExtValue()) { 6757 default: break; 6758 case 0: // IEEE-inexact exception allowed 6759 switch (M5.getZExtValue()) { 6760 default: break; 6761 case 0: ID = Intrinsic::rint; break; 6762 } 6763 break; 6764 case 4: // IEEE-inexact exception suppressed 6765 switch (M5.getZExtValue()) { 6766 default: break; 6767 case 0: ID = Intrinsic::nearbyint; break; 6768 case 1: ID = Intrinsic::round; break; 6769 case 5: ID = Intrinsic::trunc; break; 6770 case 6: ID = Intrinsic::ceil; break; 6771 case 7: ID = Intrinsic::floor; break; 6772 } 6773 break; 6774 } 6775 if (ID != Intrinsic::not_intrinsic) { 6776 Function *F = CGM.getIntrinsic(ID, ResultType); 6777 return Builder.CreateCall(F, X); 6778 } 6779 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 6780 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 6781 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 6782 return Builder.CreateCall(F, {X, M4Value, M5Value}); 6783 } 6784 6785 // Vector intrisincs that output the post-instruction CC value. 6786 6787 #define INTRINSIC_WITH_CC(NAME) \ 6788 case SystemZ::BI__builtin_##NAME: \ 6789 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 6790 6791 INTRINSIC_WITH_CC(s390_vpkshs); 6792 INTRINSIC_WITH_CC(s390_vpksfs); 6793 INTRINSIC_WITH_CC(s390_vpksgs); 6794 6795 INTRINSIC_WITH_CC(s390_vpklshs); 6796 INTRINSIC_WITH_CC(s390_vpklsfs); 6797 INTRINSIC_WITH_CC(s390_vpklsgs); 6798 6799 INTRINSIC_WITH_CC(s390_vceqbs); 6800 INTRINSIC_WITH_CC(s390_vceqhs); 6801 INTRINSIC_WITH_CC(s390_vceqfs); 6802 INTRINSIC_WITH_CC(s390_vceqgs); 6803 6804 INTRINSIC_WITH_CC(s390_vchbs); 6805 INTRINSIC_WITH_CC(s390_vchhs); 6806 INTRINSIC_WITH_CC(s390_vchfs); 6807 INTRINSIC_WITH_CC(s390_vchgs); 6808 6809 INTRINSIC_WITH_CC(s390_vchlbs); 6810 INTRINSIC_WITH_CC(s390_vchlhs); 6811 INTRINSIC_WITH_CC(s390_vchlfs); 6812 INTRINSIC_WITH_CC(s390_vchlgs); 6813 6814 INTRINSIC_WITH_CC(s390_vfaebs); 6815 INTRINSIC_WITH_CC(s390_vfaehs); 6816 INTRINSIC_WITH_CC(s390_vfaefs); 6817 6818 INTRINSIC_WITH_CC(s390_vfaezbs); 6819 INTRINSIC_WITH_CC(s390_vfaezhs); 6820 INTRINSIC_WITH_CC(s390_vfaezfs); 6821 6822 INTRINSIC_WITH_CC(s390_vfeebs); 6823 INTRINSIC_WITH_CC(s390_vfeehs); 6824 INTRINSIC_WITH_CC(s390_vfeefs); 6825 6826 INTRINSIC_WITH_CC(s390_vfeezbs); 6827 INTRINSIC_WITH_CC(s390_vfeezhs); 6828 INTRINSIC_WITH_CC(s390_vfeezfs); 6829 6830 INTRINSIC_WITH_CC(s390_vfenebs); 6831 INTRINSIC_WITH_CC(s390_vfenehs); 6832 INTRINSIC_WITH_CC(s390_vfenefs); 6833 6834 INTRINSIC_WITH_CC(s390_vfenezbs); 6835 INTRINSIC_WITH_CC(s390_vfenezhs); 6836 INTRINSIC_WITH_CC(s390_vfenezfs); 6837 6838 INTRINSIC_WITH_CC(s390_vistrbs); 6839 INTRINSIC_WITH_CC(s390_vistrhs); 6840 INTRINSIC_WITH_CC(s390_vistrfs); 6841 6842 INTRINSIC_WITH_CC(s390_vstrcbs); 6843 INTRINSIC_WITH_CC(s390_vstrchs); 6844 INTRINSIC_WITH_CC(s390_vstrcfs); 6845 6846 INTRINSIC_WITH_CC(s390_vstrczbs); 6847 INTRINSIC_WITH_CC(s390_vstrczhs); 6848 INTRINSIC_WITH_CC(s390_vstrczfs); 6849 6850 INTRINSIC_WITH_CC(s390_vfcedbs); 6851 INTRINSIC_WITH_CC(s390_vfchdbs); 6852 INTRINSIC_WITH_CC(s390_vfchedbs); 6853 6854 INTRINSIC_WITH_CC(s390_vftcidb); 6855 6856 #undef INTRINSIC_WITH_CC 6857 6858 default: 6859 return nullptr; 6860 } 6861 } 6862