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 EmitR600BuiltinExpr(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 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3283 const CallExpr *E) { 3284 if (auto Hint = GetValueForARMHint(BuiltinID)) 3285 return Hint; 3286 3287 if (BuiltinID == ARM::BI__emit) { 3288 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3289 llvm::FunctionType *FTy = 3290 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3291 3292 APSInt Value; 3293 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3294 llvm_unreachable("Sema will ensure that the parameter is constant"); 3295 3296 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3297 3298 llvm::InlineAsm *Emit = 3299 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3300 /*SideEffects=*/true) 3301 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3302 /*SideEffects=*/true); 3303 3304 return Builder.CreateCall(Emit, {}); 3305 } 3306 3307 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3308 Value *Option = EmitScalarExpr(E->getArg(0)); 3309 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3310 } 3311 3312 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3313 Value *Address = EmitScalarExpr(E->getArg(0)); 3314 Value *RW = EmitScalarExpr(E->getArg(1)); 3315 Value *IsData = EmitScalarExpr(E->getArg(2)); 3316 3317 // Locality is not supported on ARM target 3318 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3319 3320 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3321 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3322 } 3323 3324 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3325 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3326 EmitScalarExpr(E->getArg(0)), 3327 "rbit"); 3328 } 3329 3330 if (BuiltinID == ARM::BI__clear_cache) { 3331 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3332 const FunctionDecl *FD = E->getDirectCallee(); 3333 SmallVector<Value*, 2> Ops; 3334 for (unsigned i = 0; i < 2; i++) 3335 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3336 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3337 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3338 StringRef Name = FD->getName(); 3339 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3340 } 3341 3342 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3343 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3344 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3345 getContext().getTypeSize(E->getType()) == 64) || 3346 BuiltinID == ARM::BI__ldrexd) { 3347 Function *F; 3348 3349 switch (BuiltinID) { 3350 default: llvm_unreachable("unexpected builtin"); 3351 case ARM::BI__builtin_arm_ldaex: 3352 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3353 break; 3354 case ARM::BI__builtin_arm_ldrexd: 3355 case ARM::BI__builtin_arm_ldrex: 3356 case ARM::BI__ldrexd: 3357 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3358 break; 3359 } 3360 3361 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3362 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3363 "ldrexd"); 3364 3365 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3366 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3367 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3368 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3369 3370 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3371 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3372 Val = Builder.CreateOr(Val, Val1); 3373 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3374 } 3375 3376 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3377 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3378 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3379 3380 QualType Ty = E->getType(); 3381 llvm::Type *RealResTy = ConvertType(Ty); 3382 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3383 getContext().getTypeSize(Ty)); 3384 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3385 3386 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3387 ? Intrinsic::arm_ldaex 3388 : Intrinsic::arm_ldrex, 3389 LoadAddr->getType()); 3390 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3391 3392 if (RealResTy->isPointerTy()) 3393 return Builder.CreateIntToPtr(Val, RealResTy); 3394 else { 3395 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3396 return Builder.CreateBitCast(Val, RealResTy); 3397 } 3398 } 3399 3400 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3401 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3402 BuiltinID == ARM::BI__builtin_arm_strex) && 3403 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3404 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3405 ? Intrinsic::arm_stlexd 3406 : Intrinsic::arm_strexd); 3407 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3408 3409 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3410 Value *Val = EmitScalarExpr(E->getArg(0)); 3411 Builder.CreateStore(Val, Tmp); 3412 3413 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3414 Val = Builder.CreateLoad(LdPtr); 3415 3416 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3417 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3418 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3419 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3420 } 3421 3422 if (BuiltinID == ARM::BI__builtin_arm_strex || 3423 BuiltinID == ARM::BI__builtin_arm_stlex) { 3424 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3425 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3426 3427 QualType Ty = E->getArg(0)->getType(); 3428 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3429 getContext().getTypeSize(Ty)); 3430 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3431 3432 if (StoreVal->getType()->isPointerTy()) 3433 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3434 else { 3435 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3436 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3437 } 3438 3439 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3440 ? Intrinsic::arm_stlex 3441 : Intrinsic::arm_strex, 3442 StoreAddr->getType()); 3443 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3444 } 3445 3446 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3447 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3448 return Builder.CreateCall(F, {}); 3449 } 3450 3451 // CRC32 3452 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3453 switch (BuiltinID) { 3454 case ARM::BI__builtin_arm_crc32b: 3455 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3456 case ARM::BI__builtin_arm_crc32cb: 3457 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3458 case ARM::BI__builtin_arm_crc32h: 3459 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3460 case ARM::BI__builtin_arm_crc32ch: 3461 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3462 case ARM::BI__builtin_arm_crc32w: 3463 case ARM::BI__builtin_arm_crc32d: 3464 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3465 case ARM::BI__builtin_arm_crc32cw: 3466 case ARM::BI__builtin_arm_crc32cd: 3467 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3468 } 3469 3470 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3471 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3472 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3473 3474 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3475 // intrinsics, hence we need different codegen for these cases. 3476 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3477 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3478 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3479 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3480 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3481 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3482 3483 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3484 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3485 return Builder.CreateCall(F, {Res, Arg1b}); 3486 } else { 3487 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3488 3489 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3490 return Builder.CreateCall(F, {Arg0, Arg1}); 3491 } 3492 } 3493 3494 // Find out if any arguments are required to be integer constant 3495 // expressions. 3496 unsigned ICEArguments = 0; 3497 ASTContext::GetBuiltinTypeError Error; 3498 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3499 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3500 3501 SmallVector<Value*, 4> Ops; 3502 llvm::Value *Align = nullptr; 3503 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3504 if (i == 0) { 3505 switch (BuiltinID) { 3506 case NEON::BI__builtin_neon_vld1_v: 3507 case NEON::BI__builtin_neon_vld1q_v: 3508 case NEON::BI__builtin_neon_vld1q_lane_v: 3509 case NEON::BI__builtin_neon_vld1_lane_v: 3510 case NEON::BI__builtin_neon_vld1_dup_v: 3511 case NEON::BI__builtin_neon_vld1q_dup_v: 3512 case NEON::BI__builtin_neon_vst1_v: 3513 case NEON::BI__builtin_neon_vst1q_v: 3514 case NEON::BI__builtin_neon_vst1q_lane_v: 3515 case NEON::BI__builtin_neon_vst1_lane_v: 3516 case NEON::BI__builtin_neon_vst2_v: 3517 case NEON::BI__builtin_neon_vst2q_v: 3518 case NEON::BI__builtin_neon_vst2_lane_v: 3519 case NEON::BI__builtin_neon_vst2q_lane_v: 3520 case NEON::BI__builtin_neon_vst3_v: 3521 case NEON::BI__builtin_neon_vst3q_v: 3522 case NEON::BI__builtin_neon_vst3_lane_v: 3523 case NEON::BI__builtin_neon_vst3q_lane_v: 3524 case NEON::BI__builtin_neon_vst4_v: 3525 case NEON::BI__builtin_neon_vst4q_v: 3526 case NEON::BI__builtin_neon_vst4_lane_v: 3527 case NEON::BI__builtin_neon_vst4q_lane_v: 3528 // Get the alignment for the argument in addition to the value; 3529 // we'll use it later. 3530 std::pair<llvm::Value*, unsigned> Src = 3531 EmitPointerWithAlignment(E->getArg(0)); 3532 Ops.push_back(Src.first); 3533 Align = Builder.getInt32(Src.second); 3534 continue; 3535 } 3536 } 3537 if (i == 1) { 3538 switch (BuiltinID) { 3539 case NEON::BI__builtin_neon_vld2_v: 3540 case NEON::BI__builtin_neon_vld2q_v: 3541 case NEON::BI__builtin_neon_vld3_v: 3542 case NEON::BI__builtin_neon_vld3q_v: 3543 case NEON::BI__builtin_neon_vld4_v: 3544 case NEON::BI__builtin_neon_vld4q_v: 3545 case NEON::BI__builtin_neon_vld2_lane_v: 3546 case NEON::BI__builtin_neon_vld2q_lane_v: 3547 case NEON::BI__builtin_neon_vld3_lane_v: 3548 case NEON::BI__builtin_neon_vld3q_lane_v: 3549 case NEON::BI__builtin_neon_vld4_lane_v: 3550 case NEON::BI__builtin_neon_vld4q_lane_v: 3551 case NEON::BI__builtin_neon_vld2_dup_v: 3552 case NEON::BI__builtin_neon_vld3_dup_v: 3553 case NEON::BI__builtin_neon_vld4_dup_v: 3554 // Get the alignment for the argument in addition to the value; 3555 // we'll use it later. 3556 std::pair<llvm::Value*, unsigned> Src = 3557 EmitPointerWithAlignment(E->getArg(1)); 3558 Ops.push_back(Src.first); 3559 Align = Builder.getInt32(Src.second); 3560 continue; 3561 } 3562 } 3563 3564 if ((ICEArguments & (1 << i)) == 0) { 3565 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3566 } else { 3567 // If this is required to be a constant, constant fold it so that we know 3568 // that the generated intrinsic gets a ConstantInt. 3569 llvm::APSInt Result; 3570 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 3571 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 3572 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 3573 } 3574 } 3575 3576 switch (BuiltinID) { 3577 default: break; 3578 // vget_lane and vset_lane are not overloaded and do not have an extra 3579 // argument that specifies the vector type. 3580 case NEON::BI__builtin_neon_vget_lane_i8: 3581 case NEON::BI__builtin_neon_vget_lane_i16: 3582 case NEON::BI__builtin_neon_vget_lane_i32: 3583 case NEON::BI__builtin_neon_vget_lane_i64: 3584 case NEON::BI__builtin_neon_vget_lane_f32: 3585 case NEON::BI__builtin_neon_vgetq_lane_i8: 3586 case NEON::BI__builtin_neon_vgetq_lane_i16: 3587 case NEON::BI__builtin_neon_vgetq_lane_i32: 3588 case NEON::BI__builtin_neon_vgetq_lane_i64: 3589 case NEON::BI__builtin_neon_vgetq_lane_f32: 3590 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3591 "vget_lane"); 3592 case NEON::BI__builtin_neon_vset_lane_i8: 3593 case NEON::BI__builtin_neon_vset_lane_i16: 3594 case NEON::BI__builtin_neon_vset_lane_i32: 3595 case NEON::BI__builtin_neon_vset_lane_i64: 3596 case NEON::BI__builtin_neon_vset_lane_f32: 3597 case NEON::BI__builtin_neon_vsetq_lane_i8: 3598 case NEON::BI__builtin_neon_vsetq_lane_i16: 3599 case NEON::BI__builtin_neon_vsetq_lane_i32: 3600 case NEON::BI__builtin_neon_vsetq_lane_i64: 3601 case NEON::BI__builtin_neon_vsetq_lane_f32: 3602 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3603 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3604 3605 // Non-polymorphic crypto instructions also not overloaded 3606 case NEON::BI__builtin_neon_vsha1h_u32: 3607 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3608 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3609 "vsha1h"); 3610 case NEON::BI__builtin_neon_vsha1cq_u32: 3611 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3612 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3613 "vsha1h"); 3614 case NEON::BI__builtin_neon_vsha1pq_u32: 3615 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3616 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3617 "vsha1h"); 3618 case NEON::BI__builtin_neon_vsha1mq_u32: 3619 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3620 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3621 "vsha1h"); 3622 } 3623 3624 // Get the last argument, which specifies the vector type. 3625 llvm::APSInt Result; 3626 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3627 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3628 return nullptr; 3629 3630 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3631 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3632 // Determine the overloaded type of this builtin. 3633 llvm::Type *Ty; 3634 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3635 Ty = FloatTy; 3636 else 3637 Ty = DoubleTy; 3638 3639 // Determine whether this is an unsigned conversion or not. 3640 bool usgn = Result.getZExtValue() == 1; 3641 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3642 3643 // Call the appropriate intrinsic. 3644 Function *F = CGM.getIntrinsic(Int, Ty); 3645 return Builder.CreateCall(F, Ops, "vcvtr"); 3646 } 3647 3648 // Determine the type of this overloaded NEON intrinsic. 3649 NeonTypeFlags Type(Result.getZExtValue()); 3650 bool usgn = Type.isUnsigned(); 3651 bool rightShift = false; 3652 3653 llvm::VectorType *VTy = GetNeonType(this, Type); 3654 llvm::Type *Ty = VTy; 3655 if (!Ty) 3656 return nullptr; 3657 3658 // Many NEON builtins have identical semantics and uses in ARM and 3659 // AArch64. Emit these in a single function. 3660 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3661 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3662 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3663 if (Builtin) 3664 return EmitCommonNeonBuiltinExpr( 3665 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3666 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3667 3668 unsigned Int; 3669 switch (BuiltinID) { 3670 default: return nullptr; 3671 case NEON::BI__builtin_neon_vld1q_lane_v: 3672 // Handle 64-bit integer elements as a special case. Use shuffles of 3673 // one-element vectors to avoid poor code for i64 in the backend. 3674 if (VTy->getElementType()->isIntegerTy(64)) { 3675 // Extract the other lane. 3676 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3677 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3678 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3679 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3680 // Load the value as a one-element vector. 3681 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3682 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3683 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3684 // Combine them. 3685 SmallVector<Constant*, 2> Indices; 3686 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3687 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3688 SV = llvm::ConstantVector::get(Indices); 3689 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3690 } 3691 // fall through 3692 case NEON::BI__builtin_neon_vld1_lane_v: { 3693 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3694 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3695 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3696 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3697 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3698 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3699 } 3700 case NEON::BI__builtin_neon_vld2_dup_v: 3701 case NEON::BI__builtin_neon_vld3_dup_v: 3702 case NEON::BI__builtin_neon_vld4_dup_v: { 3703 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3704 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3705 switch (BuiltinID) { 3706 case NEON::BI__builtin_neon_vld2_dup_v: 3707 Int = Intrinsic::arm_neon_vld2; 3708 break; 3709 case NEON::BI__builtin_neon_vld3_dup_v: 3710 Int = Intrinsic::arm_neon_vld3; 3711 break; 3712 case NEON::BI__builtin_neon_vld4_dup_v: 3713 Int = Intrinsic::arm_neon_vld4; 3714 break; 3715 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3716 } 3717 Function *F = CGM.getIntrinsic(Int, Ty); 3718 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3719 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3720 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3721 return Builder.CreateStore(Ops[1], Ops[0]); 3722 } 3723 switch (BuiltinID) { 3724 case NEON::BI__builtin_neon_vld2_dup_v: 3725 Int = Intrinsic::arm_neon_vld2lane; 3726 break; 3727 case NEON::BI__builtin_neon_vld3_dup_v: 3728 Int = Intrinsic::arm_neon_vld3lane; 3729 break; 3730 case NEON::BI__builtin_neon_vld4_dup_v: 3731 Int = Intrinsic::arm_neon_vld4lane; 3732 break; 3733 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3734 } 3735 Function *F = CGM.getIntrinsic(Int, Ty); 3736 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3737 3738 SmallVector<Value*, 6> Args; 3739 Args.push_back(Ops[1]); 3740 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3741 3742 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3743 Args.push_back(CI); 3744 Args.push_back(Align); 3745 3746 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3747 // splat lane 0 to all elts in each vector of the result. 3748 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3749 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3750 Value *Elt = Builder.CreateBitCast(Val, Ty); 3751 Elt = EmitNeonSplat(Elt, CI); 3752 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3753 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3754 } 3755 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3756 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3757 return Builder.CreateStore(Ops[1], Ops[0]); 3758 } 3759 case NEON::BI__builtin_neon_vqrshrn_n_v: 3760 Int = 3761 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3762 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3763 1, true); 3764 case NEON::BI__builtin_neon_vqrshrun_n_v: 3765 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3766 Ops, "vqrshrun_n", 1, true); 3767 case NEON::BI__builtin_neon_vqshrn_n_v: 3768 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3769 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3770 1, true); 3771 case NEON::BI__builtin_neon_vqshrun_n_v: 3772 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3773 Ops, "vqshrun_n", 1, true); 3774 case NEON::BI__builtin_neon_vrecpe_v: 3775 case NEON::BI__builtin_neon_vrecpeq_v: 3776 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3777 Ops, "vrecpe"); 3778 case NEON::BI__builtin_neon_vrshrn_n_v: 3779 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3780 Ops, "vrshrn_n", 1, true); 3781 case NEON::BI__builtin_neon_vrsra_n_v: 3782 case NEON::BI__builtin_neon_vrsraq_n_v: 3783 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3784 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3785 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3786 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3787 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 3788 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3789 case NEON::BI__builtin_neon_vsri_n_v: 3790 case NEON::BI__builtin_neon_vsriq_n_v: 3791 rightShift = true; 3792 case NEON::BI__builtin_neon_vsli_n_v: 3793 case NEON::BI__builtin_neon_vsliq_n_v: 3794 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3795 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3796 Ops, "vsli_n"); 3797 case NEON::BI__builtin_neon_vsra_n_v: 3798 case NEON::BI__builtin_neon_vsraq_n_v: 3799 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3800 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3801 return Builder.CreateAdd(Ops[0], Ops[1]); 3802 case NEON::BI__builtin_neon_vst1q_lane_v: 3803 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3804 // a one-element vector and avoid poor code for i64 in the backend. 3805 if (VTy->getElementType()->isIntegerTy(64)) { 3806 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3807 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3808 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3809 Ops[2] = Align; 3810 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3811 Ops[1]->getType()), Ops); 3812 } 3813 // fall through 3814 case NEON::BI__builtin_neon_vst1_lane_v: { 3815 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3816 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3817 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3818 StoreInst *St = Builder.CreateStore(Ops[1], 3819 Builder.CreateBitCast(Ops[0], Ty)); 3820 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3821 return St; 3822 } 3823 case NEON::BI__builtin_neon_vtbl1_v: 3824 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3825 Ops, "vtbl1"); 3826 case NEON::BI__builtin_neon_vtbl2_v: 3827 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3828 Ops, "vtbl2"); 3829 case NEON::BI__builtin_neon_vtbl3_v: 3830 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3831 Ops, "vtbl3"); 3832 case NEON::BI__builtin_neon_vtbl4_v: 3833 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3834 Ops, "vtbl4"); 3835 case NEON::BI__builtin_neon_vtbx1_v: 3836 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3837 Ops, "vtbx1"); 3838 case NEON::BI__builtin_neon_vtbx2_v: 3839 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3840 Ops, "vtbx2"); 3841 case NEON::BI__builtin_neon_vtbx3_v: 3842 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3843 Ops, "vtbx3"); 3844 case NEON::BI__builtin_neon_vtbx4_v: 3845 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3846 Ops, "vtbx4"); 3847 } 3848 } 3849 3850 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3851 const CallExpr *E, 3852 SmallVectorImpl<Value *> &Ops) { 3853 unsigned int Int = 0; 3854 const char *s = nullptr; 3855 3856 switch (BuiltinID) { 3857 default: 3858 return nullptr; 3859 case NEON::BI__builtin_neon_vtbl1_v: 3860 case NEON::BI__builtin_neon_vqtbl1_v: 3861 case NEON::BI__builtin_neon_vqtbl1q_v: 3862 case NEON::BI__builtin_neon_vtbl2_v: 3863 case NEON::BI__builtin_neon_vqtbl2_v: 3864 case NEON::BI__builtin_neon_vqtbl2q_v: 3865 case NEON::BI__builtin_neon_vtbl3_v: 3866 case NEON::BI__builtin_neon_vqtbl3_v: 3867 case NEON::BI__builtin_neon_vqtbl3q_v: 3868 case NEON::BI__builtin_neon_vtbl4_v: 3869 case NEON::BI__builtin_neon_vqtbl4_v: 3870 case NEON::BI__builtin_neon_vqtbl4q_v: 3871 break; 3872 case NEON::BI__builtin_neon_vtbx1_v: 3873 case NEON::BI__builtin_neon_vqtbx1_v: 3874 case NEON::BI__builtin_neon_vqtbx1q_v: 3875 case NEON::BI__builtin_neon_vtbx2_v: 3876 case NEON::BI__builtin_neon_vqtbx2_v: 3877 case NEON::BI__builtin_neon_vqtbx2q_v: 3878 case NEON::BI__builtin_neon_vtbx3_v: 3879 case NEON::BI__builtin_neon_vqtbx3_v: 3880 case NEON::BI__builtin_neon_vqtbx3q_v: 3881 case NEON::BI__builtin_neon_vtbx4_v: 3882 case NEON::BI__builtin_neon_vqtbx4_v: 3883 case NEON::BI__builtin_neon_vqtbx4q_v: 3884 break; 3885 } 3886 3887 assert(E->getNumArgs() >= 3); 3888 3889 // Get the last argument, which specifies the vector type. 3890 llvm::APSInt Result; 3891 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3892 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3893 return nullptr; 3894 3895 // Determine the type of this overloaded NEON intrinsic. 3896 NeonTypeFlags Type(Result.getZExtValue()); 3897 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3898 llvm::Type *Ty = VTy; 3899 if (!Ty) 3900 return nullptr; 3901 3902 unsigned nElts = VTy->getNumElements(); 3903 3904 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3905 3906 // AArch64 scalar builtins are not overloaded, they do not have an extra 3907 // argument that specifies the vector type, need to handle each case. 3908 SmallVector<Value *, 2> TblOps; 3909 switch (BuiltinID) { 3910 case NEON::BI__builtin_neon_vtbl1_v: { 3911 TblOps.push_back(Ops[0]); 3912 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3913 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3914 } 3915 case NEON::BI__builtin_neon_vtbl2_v: { 3916 TblOps.push_back(Ops[0]); 3917 TblOps.push_back(Ops[1]); 3918 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3919 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3920 } 3921 case NEON::BI__builtin_neon_vtbl3_v: { 3922 TblOps.push_back(Ops[0]); 3923 TblOps.push_back(Ops[1]); 3924 TblOps.push_back(Ops[2]); 3925 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3926 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3927 } 3928 case NEON::BI__builtin_neon_vtbl4_v: { 3929 TblOps.push_back(Ops[0]); 3930 TblOps.push_back(Ops[1]); 3931 TblOps.push_back(Ops[2]); 3932 TblOps.push_back(Ops[3]); 3933 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3934 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3935 } 3936 case NEON::BI__builtin_neon_vtbx1_v: { 3937 TblOps.push_back(Ops[1]); 3938 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3939 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3940 3941 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3942 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3943 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3944 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3945 3946 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3947 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3948 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3949 } 3950 case NEON::BI__builtin_neon_vtbx2_v: { 3951 TblOps.push_back(Ops[1]); 3952 TblOps.push_back(Ops[2]); 3953 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3954 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3955 } 3956 case NEON::BI__builtin_neon_vtbx3_v: { 3957 TblOps.push_back(Ops[1]); 3958 TblOps.push_back(Ops[2]); 3959 TblOps.push_back(Ops[3]); 3960 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3961 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3962 3963 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3964 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3965 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3966 TwentyFourV); 3967 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3968 3969 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3970 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3971 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3972 } 3973 case NEON::BI__builtin_neon_vtbx4_v: { 3974 TblOps.push_back(Ops[1]); 3975 TblOps.push_back(Ops[2]); 3976 TblOps.push_back(Ops[3]); 3977 TblOps.push_back(Ops[4]); 3978 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3979 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3980 } 3981 case NEON::BI__builtin_neon_vqtbl1_v: 3982 case NEON::BI__builtin_neon_vqtbl1q_v: 3983 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3984 case NEON::BI__builtin_neon_vqtbl2_v: 3985 case NEON::BI__builtin_neon_vqtbl2q_v: { 3986 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3987 case NEON::BI__builtin_neon_vqtbl3_v: 3988 case NEON::BI__builtin_neon_vqtbl3q_v: 3989 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3990 case NEON::BI__builtin_neon_vqtbl4_v: 3991 case NEON::BI__builtin_neon_vqtbl4q_v: 3992 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3993 case NEON::BI__builtin_neon_vqtbx1_v: 3994 case NEON::BI__builtin_neon_vqtbx1q_v: 3995 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3996 case NEON::BI__builtin_neon_vqtbx2_v: 3997 case NEON::BI__builtin_neon_vqtbx2q_v: 3998 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3999 case NEON::BI__builtin_neon_vqtbx3_v: 4000 case NEON::BI__builtin_neon_vqtbx3q_v: 4001 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4002 case NEON::BI__builtin_neon_vqtbx4_v: 4003 case NEON::BI__builtin_neon_vqtbx4q_v: 4004 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4005 } 4006 } 4007 4008 if (!Int) 4009 return nullptr; 4010 4011 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4012 return CGF.EmitNeonCall(F, Ops, s); 4013 } 4014 4015 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4016 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4017 Op = Builder.CreateBitCast(Op, Int16Ty); 4018 Value *V = UndefValue::get(VTy); 4019 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4020 Op = Builder.CreateInsertElement(V, Op, CI); 4021 return Op; 4022 } 4023 4024 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 4025 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4026 Op = Builder.CreateBitCast(Op, Int8Ty); 4027 Value *V = UndefValue::get(VTy); 4028 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4029 Op = Builder.CreateInsertElement(V, Op, CI); 4030 return Op; 4031 } 4032 4033 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4034 const CallExpr *E) { 4035 unsigned HintID = static_cast<unsigned>(-1); 4036 switch (BuiltinID) { 4037 default: break; 4038 case AArch64::BI__builtin_arm_nop: 4039 HintID = 0; 4040 break; 4041 case AArch64::BI__builtin_arm_yield: 4042 HintID = 1; 4043 break; 4044 case AArch64::BI__builtin_arm_wfe: 4045 HintID = 2; 4046 break; 4047 case AArch64::BI__builtin_arm_wfi: 4048 HintID = 3; 4049 break; 4050 case AArch64::BI__builtin_arm_sev: 4051 HintID = 4; 4052 break; 4053 case AArch64::BI__builtin_arm_sevl: 4054 HintID = 5; 4055 break; 4056 } 4057 4058 if (HintID != static_cast<unsigned>(-1)) { 4059 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4060 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4061 } 4062 4063 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4064 Value *Address = EmitScalarExpr(E->getArg(0)); 4065 Value *RW = EmitScalarExpr(E->getArg(1)); 4066 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4067 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4068 Value *IsData = EmitScalarExpr(E->getArg(4)); 4069 4070 Value *Locality = nullptr; 4071 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4072 // Temporal fetch, needs to convert cache level to locality. 4073 Locality = llvm::ConstantInt::get(Int32Ty, 4074 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4075 } else { 4076 // Streaming fetch. 4077 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4078 } 4079 4080 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4081 // PLDL3STRM or PLDL2STRM. 4082 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4083 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4084 } 4085 4086 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4087 assert((getContext().getTypeSize(E->getType()) == 32) && 4088 "rbit of unusual size!"); 4089 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4090 return Builder.CreateCall( 4091 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4092 } 4093 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4094 assert((getContext().getTypeSize(E->getType()) == 64) && 4095 "rbit of unusual size!"); 4096 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4097 return Builder.CreateCall( 4098 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4099 } 4100 4101 if (BuiltinID == AArch64::BI__clear_cache) { 4102 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4103 const FunctionDecl *FD = E->getDirectCallee(); 4104 SmallVector<Value*, 2> Ops; 4105 for (unsigned i = 0; i < 2; i++) 4106 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4107 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4108 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4109 StringRef Name = FD->getName(); 4110 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4111 } 4112 4113 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4114 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4115 getContext().getTypeSize(E->getType()) == 128) { 4116 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4117 ? Intrinsic::aarch64_ldaxp 4118 : Intrinsic::aarch64_ldxp); 4119 4120 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4121 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4122 "ldxp"); 4123 4124 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4125 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4126 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4127 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4128 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4129 4130 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4131 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4132 Val = Builder.CreateOr(Val, Val1); 4133 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4134 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4135 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4136 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4137 4138 QualType Ty = E->getType(); 4139 llvm::Type *RealResTy = ConvertType(Ty); 4140 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4141 getContext().getTypeSize(Ty)); 4142 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4143 4144 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4145 ? Intrinsic::aarch64_ldaxr 4146 : Intrinsic::aarch64_ldxr, 4147 LoadAddr->getType()); 4148 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4149 4150 if (RealResTy->isPointerTy()) 4151 return Builder.CreateIntToPtr(Val, RealResTy); 4152 4153 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4154 return Builder.CreateBitCast(Val, RealResTy); 4155 } 4156 4157 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4158 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4159 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4160 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4161 ? Intrinsic::aarch64_stlxp 4162 : Intrinsic::aarch64_stxp); 4163 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4164 4165 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4166 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4167 One); 4168 Value *Val = EmitScalarExpr(E->getArg(0)); 4169 Builder.CreateStore(Val, Tmp); 4170 4171 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4172 Val = Builder.CreateLoad(LdPtr); 4173 4174 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4175 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4176 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4177 Int8PtrTy); 4178 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4179 } 4180 4181 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4182 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4183 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4184 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4185 4186 QualType Ty = E->getArg(0)->getType(); 4187 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4188 getContext().getTypeSize(Ty)); 4189 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4190 4191 if (StoreVal->getType()->isPointerTy()) 4192 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4193 else { 4194 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4195 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4196 } 4197 4198 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4199 ? Intrinsic::aarch64_stlxr 4200 : Intrinsic::aarch64_stxr, 4201 StoreAddr->getType()); 4202 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4203 } 4204 4205 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4206 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4207 return Builder.CreateCall(F, {}); 4208 } 4209 4210 // CRC32 4211 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4212 switch (BuiltinID) { 4213 case AArch64::BI__builtin_arm_crc32b: 4214 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4215 case AArch64::BI__builtin_arm_crc32cb: 4216 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4217 case AArch64::BI__builtin_arm_crc32h: 4218 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4219 case AArch64::BI__builtin_arm_crc32ch: 4220 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4221 case AArch64::BI__builtin_arm_crc32w: 4222 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4223 case AArch64::BI__builtin_arm_crc32cw: 4224 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4225 case AArch64::BI__builtin_arm_crc32d: 4226 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4227 case AArch64::BI__builtin_arm_crc32cd: 4228 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4229 } 4230 4231 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4232 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4233 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4234 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4235 4236 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4237 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4238 4239 return Builder.CreateCall(F, {Arg0, Arg1}); 4240 } 4241 4242 // Find out if any arguments are required to be integer constant 4243 // expressions. 4244 unsigned ICEArguments = 0; 4245 ASTContext::GetBuiltinTypeError Error; 4246 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4247 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4248 4249 llvm::SmallVector<Value*, 4> Ops; 4250 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4251 if ((ICEArguments & (1 << i)) == 0) { 4252 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4253 } else { 4254 // If this is required to be a constant, constant fold it so that we know 4255 // that the generated intrinsic gets a ConstantInt. 4256 llvm::APSInt Result; 4257 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4258 assert(IsConst && "Constant arg isn't actually constant?"); 4259 (void)IsConst; 4260 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4261 } 4262 } 4263 4264 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4265 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4266 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4267 4268 if (Builtin) { 4269 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4270 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4271 assert(Result && "SISD intrinsic should have been handled"); 4272 return Result; 4273 } 4274 4275 llvm::APSInt Result; 4276 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4277 NeonTypeFlags Type(0); 4278 if (Arg->isIntegerConstantExpr(Result, getContext())) 4279 // Determine the type of this overloaded NEON intrinsic. 4280 Type = NeonTypeFlags(Result.getZExtValue()); 4281 4282 bool usgn = Type.isUnsigned(); 4283 bool quad = Type.isQuad(); 4284 4285 // Handle non-overloaded intrinsics first. 4286 switch (BuiltinID) { 4287 default: break; 4288 case NEON::BI__builtin_neon_vldrq_p128: { 4289 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4290 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4291 return Builder.CreateLoad(Ptr); 4292 } 4293 case NEON::BI__builtin_neon_vstrq_p128: { 4294 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4295 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4296 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4297 } 4298 case NEON::BI__builtin_neon_vcvts_u32_f32: 4299 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4300 usgn = true; 4301 // FALL THROUGH 4302 case NEON::BI__builtin_neon_vcvts_s32_f32: 4303 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4304 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4305 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4306 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4307 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4308 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4309 if (usgn) 4310 return Builder.CreateFPToUI(Ops[0], InTy); 4311 return Builder.CreateFPToSI(Ops[0], InTy); 4312 } 4313 case NEON::BI__builtin_neon_vcvts_f32_u32: 4314 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4315 usgn = true; 4316 // FALL THROUGH 4317 case NEON::BI__builtin_neon_vcvts_f32_s32: 4318 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4319 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4320 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4321 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4322 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4323 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4324 if (usgn) 4325 return Builder.CreateUIToFP(Ops[0], FTy); 4326 return Builder.CreateSIToFP(Ops[0], FTy); 4327 } 4328 case NEON::BI__builtin_neon_vpaddd_s64: { 4329 llvm::Type *Ty = 4330 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4331 Value *Vec = EmitScalarExpr(E->getArg(0)); 4332 // The vector is v2f64, so make sure it's bitcast to that. 4333 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4334 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4335 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4336 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4337 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4338 // Pairwise addition of a v2f64 into a scalar f64. 4339 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4340 } 4341 case NEON::BI__builtin_neon_vpaddd_f64: { 4342 llvm::Type *Ty = 4343 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4344 Value *Vec = EmitScalarExpr(E->getArg(0)); 4345 // The vector is v2f64, so make sure it's bitcast to that. 4346 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4347 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4348 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4349 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4350 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4351 // Pairwise addition of a v2f64 into a scalar f64. 4352 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4353 } 4354 case NEON::BI__builtin_neon_vpadds_f32: { 4355 llvm::Type *Ty = 4356 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4357 Value *Vec = EmitScalarExpr(E->getArg(0)); 4358 // The vector is v2f32, so make sure it's bitcast to that. 4359 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4360 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4361 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4362 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4363 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4364 // Pairwise addition of a v2f32 into a scalar f32. 4365 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4366 } 4367 case NEON::BI__builtin_neon_vceqzd_s64: 4368 case NEON::BI__builtin_neon_vceqzd_f64: 4369 case NEON::BI__builtin_neon_vceqzs_f32: 4370 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4371 return EmitAArch64CompareBuiltinExpr( 4372 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4373 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4374 case NEON::BI__builtin_neon_vcgezd_s64: 4375 case NEON::BI__builtin_neon_vcgezd_f64: 4376 case NEON::BI__builtin_neon_vcgezs_f32: 4377 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4378 return EmitAArch64CompareBuiltinExpr( 4379 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4380 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4381 case NEON::BI__builtin_neon_vclezd_s64: 4382 case NEON::BI__builtin_neon_vclezd_f64: 4383 case NEON::BI__builtin_neon_vclezs_f32: 4384 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4385 return EmitAArch64CompareBuiltinExpr( 4386 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4387 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4388 case NEON::BI__builtin_neon_vcgtzd_s64: 4389 case NEON::BI__builtin_neon_vcgtzd_f64: 4390 case NEON::BI__builtin_neon_vcgtzs_f32: 4391 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4392 return EmitAArch64CompareBuiltinExpr( 4393 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4394 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4395 case NEON::BI__builtin_neon_vcltzd_s64: 4396 case NEON::BI__builtin_neon_vcltzd_f64: 4397 case NEON::BI__builtin_neon_vcltzs_f32: 4398 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4399 return EmitAArch64CompareBuiltinExpr( 4400 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4401 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4402 4403 case NEON::BI__builtin_neon_vceqzd_u64: { 4404 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4405 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4406 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4407 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4408 llvm::Constant::getNullValue(Ty)); 4409 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4410 } 4411 case NEON::BI__builtin_neon_vceqd_f64: 4412 case NEON::BI__builtin_neon_vcled_f64: 4413 case NEON::BI__builtin_neon_vcltd_f64: 4414 case NEON::BI__builtin_neon_vcged_f64: 4415 case NEON::BI__builtin_neon_vcgtd_f64: { 4416 llvm::CmpInst::Predicate P; 4417 switch (BuiltinID) { 4418 default: llvm_unreachable("missing builtin ID in switch!"); 4419 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4420 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4421 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4422 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4423 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4424 } 4425 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4426 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4427 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4428 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4429 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4430 } 4431 case NEON::BI__builtin_neon_vceqs_f32: 4432 case NEON::BI__builtin_neon_vcles_f32: 4433 case NEON::BI__builtin_neon_vclts_f32: 4434 case NEON::BI__builtin_neon_vcges_f32: 4435 case NEON::BI__builtin_neon_vcgts_f32: { 4436 llvm::CmpInst::Predicate P; 4437 switch (BuiltinID) { 4438 default: llvm_unreachable("missing builtin ID in switch!"); 4439 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4440 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4441 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4442 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4443 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4444 } 4445 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4446 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4447 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4448 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4449 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4450 } 4451 case NEON::BI__builtin_neon_vceqd_s64: 4452 case NEON::BI__builtin_neon_vceqd_u64: 4453 case NEON::BI__builtin_neon_vcgtd_s64: 4454 case NEON::BI__builtin_neon_vcgtd_u64: 4455 case NEON::BI__builtin_neon_vcltd_s64: 4456 case NEON::BI__builtin_neon_vcltd_u64: 4457 case NEON::BI__builtin_neon_vcged_u64: 4458 case NEON::BI__builtin_neon_vcged_s64: 4459 case NEON::BI__builtin_neon_vcled_u64: 4460 case NEON::BI__builtin_neon_vcled_s64: { 4461 llvm::CmpInst::Predicate P; 4462 switch (BuiltinID) { 4463 default: llvm_unreachable("missing builtin ID in switch!"); 4464 case NEON::BI__builtin_neon_vceqd_s64: 4465 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4466 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4467 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4468 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4469 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4470 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4471 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4472 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4473 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4474 } 4475 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4476 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4477 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4478 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4479 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4480 } 4481 case NEON::BI__builtin_neon_vtstd_s64: 4482 case NEON::BI__builtin_neon_vtstd_u64: { 4483 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4484 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4485 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4486 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4487 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4488 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4489 llvm::Constant::getNullValue(Ty)); 4490 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4491 } 4492 case NEON::BI__builtin_neon_vset_lane_i8: 4493 case NEON::BI__builtin_neon_vset_lane_i16: 4494 case NEON::BI__builtin_neon_vset_lane_i32: 4495 case NEON::BI__builtin_neon_vset_lane_i64: 4496 case NEON::BI__builtin_neon_vset_lane_f32: 4497 case NEON::BI__builtin_neon_vsetq_lane_i8: 4498 case NEON::BI__builtin_neon_vsetq_lane_i16: 4499 case NEON::BI__builtin_neon_vsetq_lane_i32: 4500 case NEON::BI__builtin_neon_vsetq_lane_i64: 4501 case NEON::BI__builtin_neon_vsetq_lane_f32: 4502 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4503 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4504 case NEON::BI__builtin_neon_vset_lane_f64: 4505 // The vector type needs a cast for the v1f64 variant. 4506 Ops[1] = Builder.CreateBitCast(Ops[1], 4507 llvm::VectorType::get(DoubleTy, 1)); 4508 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4509 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4510 case NEON::BI__builtin_neon_vsetq_lane_f64: 4511 // The vector type needs a cast for the v2f64 variant. 4512 Ops[1] = Builder.CreateBitCast(Ops[1], 4513 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4514 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4515 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4516 4517 case NEON::BI__builtin_neon_vget_lane_i8: 4518 case NEON::BI__builtin_neon_vdupb_lane_i8: 4519 Ops[0] = Builder.CreateBitCast(Ops[0], 4520 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4521 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4522 "vget_lane"); 4523 case NEON::BI__builtin_neon_vgetq_lane_i8: 4524 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4525 Ops[0] = Builder.CreateBitCast(Ops[0], 4526 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4527 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4528 "vgetq_lane"); 4529 case NEON::BI__builtin_neon_vget_lane_i16: 4530 case NEON::BI__builtin_neon_vduph_lane_i16: 4531 Ops[0] = Builder.CreateBitCast(Ops[0], 4532 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4533 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4534 "vget_lane"); 4535 case NEON::BI__builtin_neon_vgetq_lane_i16: 4536 case NEON::BI__builtin_neon_vduph_laneq_i16: 4537 Ops[0] = Builder.CreateBitCast(Ops[0], 4538 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4539 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4540 "vgetq_lane"); 4541 case NEON::BI__builtin_neon_vget_lane_i32: 4542 case NEON::BI__builtin_neon_vdups_lane_i32: 4543 Ops[0] = Builder.CreateBitCast( 4544 Ops[0], 4545 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4546 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4547 "vget_lane"); 4548 case NEON::BI__builtin_neon_vdups_lane_f32: 4549 Ops[0] = Builder.CreateBitCast(Ops[0], 4550 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4551 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4552 "vdups_lane"); 4553 case NEON::BI__builtin_neon_vgetq_lane_i32: 4554 case NEON::BI__builtin_neon_vdups_laneq_i32: 4555 Ops[0] = Builder.CreateBitCast(Ops[0], 4556 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4557 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4558 "vgetq_lane"); 4559 case NEON::BI__builtin_neon_vget_lane_i64: 4560 case NEON::BI__builtin_neon_vdupd_lane_i64: 4561 Ops[0] = Builder.CreateBitCast(Ops[0], 4562 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4563 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4564 "vget_lane"); 4565 case NEON::BI__builtin_neon_vdupd_lane_f64: 4566 Ops[0] = Builder.CreateBitCast(Ops[0], 4567 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4568 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4569 "vdupd_lane"); 4570 case NEON::BI__builtin_neon_vgetq_lane_i64: 4571 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4572 Ops[0] = Builder.CreateBitCast(Ops[0], 4573 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4574 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4575 "vgetq_lane"); 4576 case NEON::BI__builtin_neon_vget_lane_f32: 4577 Ops[0] = Builder.CreateBitCast(Ops[0], 4578 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4579 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4580 "vget_lane"); 4581 case NEON::BI__builtin_neon_vget_lane_f64: 4582 Ops[0] = Builder.CreateBitCast(Ops[0], 4583 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4584 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4585 "vget_lane"); 4586 case NEON::BI__builtin_neon_vgetq_lane_f32: 4587 case NEON::BI__builtin_neon_vdups_laneq_f32: 4588 Ops[0] = Builder.CreateBitCast(Ops[0], 4589 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4590 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4591 "vgetq_lane"); 4592 case NEON::BI__builtin_neon_vgetq_lane_f64: 4593 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4594 Ops[0] = Builder.CreateBitCast(Ops[0], 4595 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4596 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4597 "vgetq_lane"); 4598 case NEON::BI__builtin_neon_vaddd_s64: 4599 case NEON::BI__builtin_neon_vaddd_u64: 4600 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4601 case NEON::BI__builtin_neon_vsubd_s64: 4602 case NEON::BI__builtin_neon_vsubd_u64: 4603 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4604 case NEON::BI__builtin_neon_vqdmlalh_s16: 4605 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4606 SmallVector<Value *, 2> ProductOps; 4607 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4608 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4609 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4610 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4611 ProductOps, "vqdmlXl"); 4612 Constant *CI = ConstantInt::get(SizeTy, 0); 4613 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4614 4615 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4616 ? Intrinsic::aarch64_neon_sqadd 4617 : Intrinsic::aarch64_neon_sqsub; 4618 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4619 } 4620 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4621 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4622 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4623 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4624 Ops, "vqshlu_n"); 4625 } 4626 case NEON::BI__builtin_neon_vqshld_n_u64: 4627 case NEON::BI__builtin_neon_vqshld_n_s64: { 4628 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4629 ? Intrinsic::aarch64_neon_uqshl 4630 : Intrinsic::aarch64_neon_sqshl; 4631 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4632 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4633 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4634 } 4635 case NEON::BI__builtin_neon_vrshrd_n_u64: 4636 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4637 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4638 ? Intrinsic::aarch64_neon_urshl 4639 : Intrinsic::aarch64_neon_srshl; 4640 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4641 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4642 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4643 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4644 } 4645 case NEON::BI__builtin_neon_vrsrad_n_u64: 4646 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4647 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4648 ? Intrinsic::aarch64_neon_urshl 4649 : Intrinsic::aarch64_neon_srshl; 4650 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4651 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4652 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4653 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4654 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4655 } 4656 case NEON::BI__builtin_neon_vshld_n_s64: 4657 case NEON::BI__builtin_neon_vshld_n_u64: { 4658 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4659 return Builder.CreateShl( 4660 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4661 } 4662 case NEON::BI__builtin_neon_vshrd_n_s64: { 4663 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4664 return Builder.CreateAShr( 4665 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4666 Amt->getZExtValue())), 4667 "shrd_n"); 4668 } 4669 case NEON::BI__builtin_neon_vshrd_n_u64: { 4670 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4671 uint64_t ShiftAmt = Amt->getZExtValue(); 4672 // Right-shifting an unsigned value by its size yields 0. 4673 if (ShiftAmt == 64) 4674 return ConstantInt::get(Int64Ty, 0); 4675 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4676 "shrd_n"); 4677 } 4678 case NEON::BI__builtin_neon_vsrad_n_s64: { 4679 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4680 Ops[1] = Builder.CreateAShr( 4681 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4682 Amt->getZExtValue())), 4683 "shrd_n"); 4684 return Builder.CreateAdd(Ops[0], Ops[1]); 4685 } 4686 case NEON::BI__builtin_neon_vsrad_n_u64: { 4687 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4688 uint64_t ShiftAmt = Amt->getZExtValue(); 4689 // Right-shifting an unsigned value by its size yields 0. 4690 // As Op + 0 = Op, return Ops[0] directly. 4691 if (ShiftAmt == 64) 4692 return Ops[0]; 4693 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4694 "shrd_n"); 4695 return Builder.CreateAdd(Ops[0], Ops[1]); 4696 } 4697 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4698 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4699 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4700 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4701 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4702 "lane"); 4703 SmallVector<Value *, 2> ProductOps; 4704 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4705 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4706 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4707 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4708 ProductOps, "vqdmlXl"); 4709 Constant *CI = ConstantInt::get(SizeTy, 0); 4710 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4711 Ops.pop_back(); 4712 4713 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4714 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4715 ? Intrinsic::aarch64_neon_sqadd 4716 : Intrinsic::aarch64_neon_sqsub; 4717 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4718 } 4719 case NEON::BI__builtin_neon_vqdmlals_s32: 4720 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4721 SmallVector<Value *, 2> ProductOps; 4722 ProductOps.push_back(Ops[1]); 4723 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4724 Ops[1] = 4725 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4726 ProductOps, "vqdmlXl"); 4727 4728 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4729 ? Intrinsic::aarch64_neon_sqadd 4730 : Intrinsic::aarch64_neon_sqsub; 4731 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4732 } 4733 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4734 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4735 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4736 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4737 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4738 "lane"); 4739 SmallVector<Value *, 2> ProductOps; 4740 ProductOps.push_back(Ops[1]); 4741 ProductOps.push_back(Ops[2]); 4742 Ops[1] = 4743 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4744 ProductOps, "vqdmlXl"); 4745 Ops.pop_back(); 4746 4747 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4748 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4749 ? Intrinsic::aarch64_neon_sqadd 4750 : Intrinsic::aarch64_neon_sqsub; 4751 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4752 } 4753 } 4754 4755 llvm::VectorType *VTy = GetNeonType(this, Type); 4756 llvm::Type *Ty = VTy; 4757 if (!Ty) 4758 return nullptr; 4759 4760 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4761 // defer to common code if it's been added to our special map. 4762 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4763 AArch64SIMDIntrinsicsProvenSorted); 4764 4765 if (Builtin) 4766 return EmitCommonNeonBuiltinExpr( 4767 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4768 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4769 4770 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4771 return V; 4772 4773 unsigned Int; 4774 switch (BuiltinID) { 4775 default: return nullptr; 4776 case NEON::BI__builtin_neon_vbsl_v: 4777 case NEON::BI__builtin_neon_vbslq_v: { 4778 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4779 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4780 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4781 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4782 4783 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4784 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4785 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4786 return Builder.CreateBitCast(Ops[0], Ty); 4787 } 4788 case NEON::BI__builtin_neon_vfma_lane_v: 4789 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4790 // The ARM builtins (and instructions) have the addend as the first 4791 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4792 Value *Addend = Ops[0]; 4793 Value *Multiplicand = Ops[1]; 4794 Value *LaneSource = Ops[2]; 4795 Ops[0] = Multiplicand; 4796 Ops[1] = LaneSource; 4797 Ops[2] = Addend; 4798 4799 // Now adjust things to handle the lane access. 4800 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4801 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4802 VTy; 4803 llvm::Constant *cst = cast<Constant>(Ops[3]); 4804 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4805 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4806 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4807 4808 Ops.pop_back(); 4809 Int = Intrinsic::fma; 4810 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4811 } 4812 case NEON::BI__builtin_neon_vfma_laneq_v: { 4813 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4814 // v1f64 fma should be mapped to Neon scalar f64 fma 4815 if (VTy && VTy->getElementType() == DoubleTy) { 4816 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4817 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4818 llvm::Type *VTy = GetNeonType(this, 4819 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4820 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4821 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4822 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4823 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4824 return Builder.CreateBitCast(Result, Ty); 4825 } 4826 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4827 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4828 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4829 4830 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4831 VTy->getNumElements() * 2); 4832 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4833 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4834 cast<ConstantInt>(Ops[3])); 4835 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4836 4837 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4838 } 4839 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4840 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4841 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4842 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4843 4844 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4845 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4846 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4847 } 4848 case NEON::BI__builtin_neon_vfmas_lane_f32: 4849 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4850 case NEON::BI__builtin_neon_vfmad_lane_f64: 4851 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4852 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4853 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4854 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4855 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4856 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4857 } 4858 case NEON::BI__builtin_neon_vfms_v: 4859 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4860 // FIXME: probably remove when we no longer support aarch64_simd.h 4861 // (arm_neon.h delegates to vfma). 4862 4863 // The ARM builtins (and instructions) have the addend as the first 4864 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4865 Value *Subtrahend = Ops[0]; 4866 Value *Multiplicand = Ops[2]; 4867 Ops[0] = Multiplicand; 4868 Ops[2] = Subtrahend; 4869 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4870 Ops[1] = Builder.CreateFNeg(Ops[1]); 4871 Int = Intrinsic::fma; 4872 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4873 } 4874 case NEON::BI__builtin_neon_vmull_v: 4875 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4876 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4877 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4878 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4879 case NEON::BI__builtin_neon_vmax_v: 4880 case NEON::BI__builtin_neon_vmaxq_v: 4881 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4882 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4883 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4884 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4885 case NEON::BI__builtin_neon_vmin_v: 4886 case NEON::BI__builtin_neon_vminq_v: 4887 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4888 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4889 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4890 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4891 case NEON::BI__builtin_neon_vabd_v: 4892 case NEON::BI__builtin_neon_vabdq_v: 4893 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4894 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4895 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4896 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4897 case NEON::BI__builtin_neon_vpadal_v: 4898 case NEON::BI__builtin_neon_vpadalq_v: { 4899 unsigned ArgElts = VTy->getNumElements(); 4900 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4901 unsigned BitWidth = EltTy->getBitWidth(); 4902 llvm::Type *ArgTy = llvm::VectorType::get( 4903 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4904 llvm::Type* Tys[2] = { VTy, ArgTy }; 4905 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4906 SmallVector<llvm::Value*, 1> TmpOps; 4907 TmpOps.push_back(Ops[1]); 4908 Function *F = CGM.getIntrinsic(Int, Tys); 4909 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4910 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4911 return Builder.CreateAdd(tmp, addend); 4912 } 4913 case NEON::BI__builtin_neon_vpmin_v: 4914 case NEON::BI__builtin_neon_vpminq_v: 4915 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4916 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4917 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4918 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4919 case NEON::BI__builtin_neon_vpmax_v: 4920 case NEON::BI__builtin_neon_vpmaxq_v: 4921 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4922 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4923 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4924 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4925 case NEON::BI__builtin_neon_vminnm_v: 4926 case NEON::BI__builtin_neon_vminnmq_v: 4927 Int = Intrinsic::aarch64_neon_fminnm; 4928 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4929 case NEON::BI__builtin_neon_vmaxnm_v: 4930 case NEON::BI__builtin_neon_vmaxnmq_v: 4931 Int = Intrinsic::aarch64_neon_fmaxnm; 4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4933 case NEON::BI__builtin_neon_vrecpss_f32: { 4934 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4935 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4936 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4937 Ops, "vrecps"); 4938 } 4939 case NEON::BI__builtin_neon_vrecpsd_f64: { 4940 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4941 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4942 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4943 Ops, "vrecps"); 4944 } 4945 case NEON::BI__builtin_neon_vqshrun_n_v: 4946 Int = Intrinsic::aarch64_neon_sqshrun; 4947 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4948 case NEON::BI__builtin_neon_vqrshrun_n_v: 4949 Int = Intrinsic::aarch64_neon_sqrshrun; 4950 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4951 case NEON::BI__builtin_neon_vqshrn_n_v: 4952 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4953 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4954 case NEON::BI__builtin_neon_vrshrn_n_v: 4955 Int = Intrinsic::aarch64_neon_rshrn; 4956 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4957 case NEON::BI__builtin_neon_vqrshrn_n_v: 4958 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4959 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4960 case NEON::BI__builtin_neon_vrnda_v: 4961 case NEON::BI__builtin_neon_vrndaq_v: { 4962 Int = Intrinsic::round; 4963 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4964 } 4965 case NEON::BI__builtin_neon_vrndi_v: 4966 case NEON::BI__builtin_neon_vrndiq_v: { 4967 Int = Intrinsic::nearbyint; 4968 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4969 } 4970 case NEON::BI__builtin_neon_vrndm_v: 4971 case NEON::BI__builtin_neon_vrndmq_v: { 4972 Int = Intrinsic::floor; 4973 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4974 } 4975 case NEON::BI__builtin_neon_vrndn_v: 4976 case NEON::BI__builtin_neon_vrndnq_v: { 4977 Int = Intrinsic::aarch64_neon_frintn; 4978 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4979 } 4980 case NEON::BI__builtin_neon_vrndp_v: 4981 case NEON::BI__builtin_neon_vrndpq_v: { 4982 Int = Intrinsic::ceil; 4983 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4984 } 4985 case NEON::BI__builtin_neon_vrndx_v: 4986 case NEON::BI__builtin_neon_vrndxq_v: { 4987 Int = Intrinsic::rint; 4988 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4989 } 4990 case NEON::BI__builtin_neon_vrnd_v: 4991 case NEON::BI__builtin_neon_vrndq_v: { 4992 Int = Intrinsic::trunc; 4993 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4994 } 4995 case NEON::BI__builtin_neon_vceqz_v: 4996 case NEON::BI__builtin_neon_vceqzq_v: 4997 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4998 ICmpInst::ICMP_EQ, "vceqz"); 4999 case NEON::BI__builtin_neon_vcgez_v: 5000 case NEON::BI__builtin_neon_vcgezq_v: 5001 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5002 ICmpInst::ICMP_SGE, "vcgez"); 5003 case NEON::BI__builtin_neon_vclez_v: 5004 case NEON::BI__builtin_neon_vclezq_v: 5005 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5006 ICmpInst::ICMP_SLE, "vclez"); 5007 case NEON::BI__builtin_neon_vcgtz_v: 5008 case NEON::BI__builtin_neon_vcgtzq_v: 5009 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5010 ICmpInst::ICMP_SGT, "vcgtz"); 5011 case NEON::BI__builtin_neon_vcltz_v: 5012 case NEON::BI__builtin_neon_vcltzq_v: 5013 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5014 ICmpInst::ICMP_SLT, "vcltz"); 5015 case NEON::BI__builtin_neon_vcvt_f64_v: 5016 case NEON::BI__builtin_neon_vcvtq_f64_v: 5017 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5018 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5019 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5020 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5021 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5022 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5023 "unexpected vcvt_f64_f32 builtin"); 5024 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5025 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5026 5027 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5028 } 5029 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5030 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5031 "unexpected vcvt_f32_f64 builtin"); 5032 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5033 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5034 5035 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5036 } 5037 case NEON::BI__builtin_neon_vcvt_s32_v: 5038 case NEON::BI__builtin_neon_vcvt_u32_v: 5039 case NEON::BI__builtin_neon_vcvt_s64_v: 5040 case NEON::BI__builtin_neon_vcvt_u64_v: 5041 case NEON::BI__builtin_neon_vcvtq_s32_v: 5042 case NEON::BI__builtin_neon_vcvtq_u32_v: 5043 case NEON::BI__builtin_neon_vcvtq_s64_v: 5044 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5045 bool Double = 5046 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5047 llvm::Type *InTy = 5048 GetNeonType(this, 5049 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5050 : NeonTypeFlags::Float32, false, quad)); 5051 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5052 if (usgn) 5053 return Builder.CreateFPToUI(Ops[0], Ty); 5054 return Builder.CreateFPToSI(Ops[0], Ty); 5055 } 5056 case NEON::BI__builtin_neon_vcvta_s32_v: 5057 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5058 case NEON::BI__builtin_neon_vcvta_u32_v: 5059 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5060 case NEON::BI__builtin_neon_vcvta_s64_v: 5061 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5062 case NEON::BI__builtin_neon_vcvta_u64_v: 5063 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5064 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5065 bool Double = 5066 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5067 llvm::Type *InTy = 5068 GetNeonType(this, 5069 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5070 : NeonTypeFlags::Float32, false, quad)); 5071 llvm::Type *Tys[2] = { Ty, InTy }; 5072 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5073 } 5074 case NEON::BI__builtin_neon_vcvtm_s32_v: 5075 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5076 case NEON::BI__builtin_neon_vcvtm_u32_v: 5077 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5078 case NEON::BI__builtin_neon_vcvtm_s64_v: 5079 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5080 case NEON::BI__builtin_neon_vcvtm_u64_v: 5081 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5082 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5083 bool Double = 5084 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5085 llvm::Type *InTy = 5086 GetNeonType(this, 5087 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5088 : NeonTypeFlags::Float32, false, quad)); 5089 llvm::Type *Tys[2] = { Ty, InTy }; 5090 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5091 } 5092 case NEON::BI__builtin_neon_vcvtn_s32_v: 5093 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5094 case NEON::BI__builtin_neon_vcvtn_u32_v: 5095 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5096 case NEON::BI__builtin_neon_vcvtn_s64_v: 5097 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5098 case NEON::BI__builtin_neon_vcvtn_u64_v: 5099 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5100 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5101 bool Double = 5102 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5103 llvm::Type *InTy = 5104 GetNeonType(this, 5105 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5106 : NeonTypeFlags::Float32, false, quad)); 5107 llvm::Type *Tys[2] = { Ty, InTy }; 5108 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5109 } 5110 case NEON::BI__builtin_neon_vcvtp_s32_v: 5111 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5112 case NEON::BI__builtin_neon_vcvtp_u32_v: 5113 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5114 case NEON::BI__builtin_neon_vcvtp_s64_v: 5115 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5116 case NEON::BI__builtin_neon_vcvtp_u64_v: 5117 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5118 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5119 bool Double = 5120 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5121 llvm::Type *InTy = 5122 GetNeonType(this, 5123 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5124 : NeonTypeFlags::Float32, false, quad)); 5125 llvm::Type *Tys[2] = { Ty, InTy }; 5126 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5127 } 5128 case NEON::BI__builtin_neon_vmulx_v: 5129 case NEON::BI__builtin_neon_vmulxq_v: { 5130 Int = Intrinsic::aarch64_neon_fmulx; 5131 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5132 } 5133 case NEON::BI__builtin_neon_vmul_lane_v: 5134 case NEON::BI__builtin_neon_vmul_laneq_v: { 5135 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5136 bool Quad = false; 5137 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5138 Quad = true; 5139 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5140 llvm::Type *VTy = GetNeonType(this, 5141 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5142 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5143 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5144 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5145 return Builder.CreateBitCast(Result, Ty); 5146 } 5147 case NEON::BI__builtin_neon_vnegd_s64: 5148 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5149 case NEON::BI__builtin_neon_vpmaxnm_v: 5150 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5151 Int = Intrinsic::aarch64_neon_fmaxnmp; 5152 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5153 } 5154 case NEON::BI__builtin_neon_vpminnm_v: 5155 case NEON::BI__builtin_neon_vpminnmq_v: { 5156 Int = Intrinsic::aarch64_neon_fminnmp; 5157 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5158 } 5159 case NEON::BI__builtin_neon_vsqrt_v: 5160 case NEON::BI__builtin_neon_vsqrtq_v: { 5161 Int = Intrinsic::sqrt; 5162 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5163 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5164 } 5165 case NEON::BI__builtin_neon_vrbit_v: 5166 case NEON::BI__builtin_neon_vrbitq_v: { 5167 Int = Intrinsic::aarch64_neon_rbit; 5168 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5169 } 5170 case NEON::BI__builtin_neon_vaddv_u8: 5171 // FIXME: These are handled by the AArch64 scalar code. 5172 usgn = true; 5173 // FALLTHROUGH 5174 case NEON::BI__builtin_neon_vaddv_s8: { 5175 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5176 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5177 VTy = 5178 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5179 llvm::Type *Tys[2] = { Ty, VTy }; 5180 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5181 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5182 return Builder.CreateTrunc(Ops[0], 5183 llvm::IntegerType::get(getLLVMContext(), 8)); 5184 } 5185 case NEON::BI__builtin_neon_vaddv_u16: 5186 usgn = true; 5187 // FALLTHROUGH 5188 case NEON::BI__builtin_neon_vaddv_s16: { 5189 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5190 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5191 VTy = 5192 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5193 llvm::Type *Tys[2] = { Ty, VTy }; 5194 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5195 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5196 return Builder.CreateTrunc(Ops[0], 5197 llvm::IntegerType::get(getLLVMContext(), 16)); 5198 } 5199 case NEON::BI__builtin_neon_vaddvq_u8: 5200 usgn = true; 5201 // FALLTHROUGH 5202 case NEON::BI__builtin_neon_vaddvq_s8: { 5203 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5204 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5205 VTy = 5206 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5207 llvm::Type *Tys[2] = { Ty, VTy }; 5208 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5209 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5210 return Builder.CreateTrunc(Ops[0], 5211 llvm::IntegerType::get(getLLVMContext(), 8)); 5212 } 5213 case NEON::BI__builtin_neon_vaddvq_u16: 5214 usgn = true; 5215 // FALLTHROUGH 5216 case NEON::BI__builtin_neon_vaddvq_s16: { 5217 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5218 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5219 VTy = 5220 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5221 llvm::Type *Tys[2] = { Ty, VTy }; 5222 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5223 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5224 return Builder.CreateTrunc(Ops[0], 5225 llvm::IntegerType::get(getLLVMContext(), 16)); 5226 } 5227 case NEON::BI__builtin_neon_vmaxv_u8: { 5228 Int = Intrinsic::aarch64_neon_umaxv; 5229 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5230 VTy = 5231 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5232 llvm::Type *Tys[2] = { Ty, VTy }; 5233 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5234 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5235 return Builder.CreateTrunc(Ops[0], 5236 llvm::IntegerType::get(getLLVMContext(), 8)); 5237 } 5238 case NEON::BI__builtin_neon_vmaxv_u16: { 5239 Int = Intrinsic::aarch64_neon_umaxv; 5240 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5241 VTy = 5242 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5243 llvm::Type *Tys[2] = { Ty, VTy }; 5244 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5245 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5246 return Builder.CreateTrunc(Ops[0], 5247 llvm::IntegerType::get(getLLVMContext(), 16)); 5248 } 5249 case NEON::BI__builtin_neon_vmaxvq_u8: { 5250 Int = Intrinsic::aarch64_neon_umaxv; 5251 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5252 VTy = 5253 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5254 llvm::Type *Tys[2] = { Ty, VTy }; 5255 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5256 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5257 return Builder.CreateTrunc(Ops[0], 5258 llvm::IntegerType::get(getLLVMContext(), 8)); 5259 } 5260 case NEON::BI__builtin_neon_vmaxvq_u16: { 5261 Int = Intrinsic::aarch64_neon_umaxv; 5262 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5263 VTy = 5264 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5265 llvm::Type *Tys[2] = { Ty, VTy }; 5266 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5267 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5268 return Builder.CreateTrunc(Ops[0], 5269 llvm::IntegerType::get(getLLVMContext(), 16)); 5270 } 5271 case NEON::BI__builtin_neon_vmaxv_s8: { 5272 Int = Intrinsic::aarch64_neon_smaxv; 5273 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5274 VTy = 5275 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5276 llvm::Type *Tys[2] = { Ty, VTy }; 5277 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5278 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5279 return Builder.CreateTrunc(Ops[0], 5280 llvm::IntegerType::get(getLLVMContext(), 8)); 5281 } 5282 case NEON::BI__builtin_neon_vmaxv_s16: { 5283 Int = Intrinsic::aarch64_neon_smaxv; 5284 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5285 VTy = 5286 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5287 llvm::Type *Tys[2] = { Ty, VTy }; 5288 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5289 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5290 return Builder.CreateTrunc(Ops[0], 5291 llvm::IntegerType::get(getLLVMContext(), 16)); 5292 } 5293 case NEON::BI__builtin_neon_vmaxvq_s8: { 5294 Int = Intrinsic::aarch64_neon_smaxv; 5295 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5296 VTy = 5297 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5298 llvm::Type *Tys[2] = { Ty, VTy }; 5299 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5300 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5301 return Builder.CreateTrunc(Ops[0], 5302 llvm::IntegerType::get(getLLVMContext(), 8)); 5303 } 5304 case NEON::BI__builtin_neon_vmaxvq_s16: { 5305 Int = Intrinsic::aarch64_neon_smaxv; 5306 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5307 VTy = 5308 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5309 llvm::Type *Tys[2] = { Ty, VTy }; 5310 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5311 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5312 return Builder.CreateTrunc(Ops[0], 5313 llvm::IntegerType::get(getLLVMContext(), 16)); 5314 } 5315 case NEON::BI__builtin_neon_vminv_u8: { 5316 Int = Intrinsic::aarch64_neon_uminv; 5317 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5318 VTy = 5319 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5320 llvm::Type *Tys[2] = { Ty, VTy }; 5321 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5322 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5323 return Builder.CreateTrunc(Ops[0], 5324 llvm::IntegerType::get(getLLVMContext(), 8)); 5325 } 5326 case NEON::BI__builtin_neon_vminv_u16: { 5327 Int = Intrinsic::aarch64_neon_uminv; 5328 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5329 VTy = 5330 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5331 llvm::Type *Tys[2] = { Ty, VTy }; 5332 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5333 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5334 return Builder.CreateTrunc(Ops[0], 5335 llvm::IntegerType::get(getLLVMContext(), 16)); 5336 } 5337 case NEON::BI__builtin_neon_vminvq_u8: { 5338 Int = Intrinsic::aarch64_neon_uminv; 5339 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5340 VTy = 5341 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5342 llvm::Type *Tys[2] = { Ty, VTy }; 5343 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5344 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5345 return Builder.CreateTrunc(Ops[0], 5346 llvm::IntegerType::get(getLLVMContext(), 8)); 5347 } 5348 case NEON::BI__builtin_neon_vminvq_u16: { 5349 Int = Intrinsic::aarch64_neon_uminv; 5350 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5351 VTy = 5352 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 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, "vminv"); 5356 return Builder.CreateTrunc(Ops[0], 5357 llvm::IntegerType::get(getLLVMContext(), 16)); 5358 } 5359 case NEON::BI__builtin_neon_vminv_s8: { 5360 Int = Intrinsic::aarch64_neon_sminv; 5361 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5362 VTy = 5363 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5364 llvm::Type *Tys[2] = { Ty, VTy }; 5365 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5366 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5367 return Builder.CreateTrunc(Ops[0], 5368 llvm::IntegerType::get(getLLVMContext(), 8)); 5369 } 5370 case NEON::BI__builtin_neon_vminv_s16: { 5371 Int = Intrinsic::aarch64_neon_sminv; 5372 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5373 VTy = 5374 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5375 llvm::Type *Tys[2] = { Ty, VTy }; 5376 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5377 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5378 return Builder.CreateTrunc(Ops[0], 5379 llvm::IntegerType::get(getLLVMContext(), 16)); 5380 } 5381 case NEON::BI__builtin_neon_vminvq_s8: { 5382 Int = Intrinsic::aarch64_neon_sminv; 5383 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5384 VTy = 5385 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5386 llvm::Type *Tys[2] = { Ty, VTy }; 5387 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5388 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5389 return Builder.CreateTrunc(Ops[0], 5390 llvm::IntegerType::get(getLLVMContext(), 8)); 5391 } 5392 case NEON::BI__builtin_neon_vminvq_s16: { 5393 Int = Intrinsic::aarch64_neon_sminv; 5394 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5395 VTy = 5396 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 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, "vminv"); 5400 return Builder.CreateTrunc(Ops[0], 5401 llvm::IntegerType::get(getLLVMContext(), 16)); 5402 } 5403 case NEON::BI__builtin_neon_vmul_n_f64: { 5404 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5405 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5406 return Builder.CreateFMul(Ops[0], RHS); 5407 } 5408 case NEON::BI__builtin_neon_vaddlv_u8: { 5409 Int = Intrinsic::aarch64_neon_uaddlv; 5410 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5411 VTy = 5412 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5413 llvm::Type *Tys[2] = { Ty, VTy }; 5414 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5415 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5416 return Builder.CreateTrunc(Ops[0], 5417 llvm::IntegerType::get(getLLVMContext(), 16)); 5418 } 5419 case NEON::BI__builtin_neon_vaddlv_u16: { 5420 Int = Intrinsic::aarch64_neon_uaddlv; 5421 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5422 VTy = 5423 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5424 llvm::Type *Tys[2] = { Ty, VTy }; 5425 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5426 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5427 } 5428 case NEON::BI__builtin_neon_vaddlvq_u8: { 5429 Int = Intrinsic::aarch64_neon_uaddlv; 5430 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5431 VTy = 5432 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5433 llvm::Type *Tys[2] = { Ty, VTy }; 5434 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5435 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5436 return Builder.CreateTrunc(Ops[0], 5437 llvm::IntegerType::get(getLLVMContext(), 16)); 5438 } 5439 case NEON::BI__builtin_neon_vaddlvq_u16: { 5440 Int = Intrinsic::aarch64_neon_uaddlv; 5441 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5442 VTy = 5443 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5444 llvm::Type *Tys[2] = { Ty, VTy }; 5445 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5446 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5447 } 5448 case NEON::BI__builtin_neon_vaddlv_s8: { 5449 Int = Intrinsic::aarch64_neon_saddlv; 5450 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5451 VTy = 5452 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5453 llvm::Type *Tys[2] = { Ty, VTy }; 5454 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5455 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5456 return Builder.CreateTrunc(Ops[0], 5457 llvm::IntegerType::get(getLLVMContext(), 16)); 5458 } 5459 case NEON::BI__builtin_neon_vaddlv_s16: { 5460 Int = Intrinsic::aarch64_neon_saddlv; 5461 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5462 VTy = 5463 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5464 llvm::Type *Tys[2] = { Ty, VTy }; 5465 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5466 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5467 } 5468 case NEON::BI__builtin_neon_vaddlvq_s8: { 5469 Int = Intrinsic::aarch64_neon_saddlv; 5470 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5471 VTy = 5472 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5473 llvm::Type *Tys[2] = { Ty, VTy }; 5474 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5475 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5476 return Builder.CreateTrunc(Ops[0], 5477 llvm::IntegerType::get(getLLVMContext(), 16)); 5478 } 5479 case NEON::BI__builtin_neon_vaddlvq_s16: { 5480 Int = Intrinsic::aarch64_neon_saddlv; 5481 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5482 VTy = 5483 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5484 llvm::Type *Tys[2] = { Ty, VTy }; 5485 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5486 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5487 } 5488 case NEON::BI__builtin_neon_vsri_n_v: 5489 case NEON::BI__builtin_neon_vsriq_n_v: { 5490 Int = Intrinsic::aarch64_neon_vsri; 5491 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5492 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5493 } 5494 case NEON::BI__builtin_neon_vsli_n_v: 5495 case NEON::BI__builtin_neon_vsliq_n_v: { 5496 Int = Intrinsic::aarch64_neon_vsli; 5497 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5498 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5499 } 5500 case NEON::BI__builtin_neon_vsra_n_v: 5501 case NEON::BI__builtin_neon_vsraq_n_v: 5502 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5503 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5504 return Builder.CreateAdd(Ops[0], Ops[1]); 5505 case NEON::BI__builtin_neon_vrsra_n_v: 5506 case NEON::BI__builtin_neon_vrsraq_n_v: { 5507 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5508 SmallVector<llvm::Value*,2> TmpOps; 5509 TmpOps.push_back(Ops[1]); 5510 TmpOps.push_back(Ops[2]); 5511 Function* F = CGM.getIntrinsic(Int, Ty); 5512 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5513 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5514 return Builder.CreateAdd(Ops[0], tmp); 5515 } 5516 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5517 // of an Align parameter here. 5518 case NEON::BI__builtin_neon_vld1_x2_v: 5519 case NEON::BI__builtin_neon_vld1q_x2_v: 5520 case NEON::BI__builtin_neon_vld1_x3_v: 5521 case NEON::BI__builtin_neon_vld1q_x3_v: 5522 case NEON::BI__builtin_neon_vld1_x4_v: 5523 case NEON::BI__builtin_neon_vld1q_x4_v: { 5524 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5525 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5526 llvm::Type *Tys[2] = { VTy, PTy }; 5527 unsigned Int; 5528 switch (BuiltinID) { 5529 case NEON::BI__builtin_neon_vld1_x2_v: 5530 case NEON::BI__builtin_neon_vld1q_x2_v: 5531 Int = Intrinsic::aarch64_neon_ld1x2; 5532 break; 5533 case NEON::BI__builtin_neon_vld1_x3_v: 5534 case NEON::BI__builtin_neon_vld1q_x3_v: 5535 Int = Intrinsic::aarch64_neon_ld1x3; 5536 break; 5537 case NEON::BI__builtin_neon_vld1_x4_v: 5538 case NEON::BI__builtin_neon_vld1q_x4_v: 5539 Int = Intrinsic::aarch64_neon_ld1x4; 5540 break; 5541 } 5542 Function *F = CGM.getIntrinsic(Int, Tys); 5543 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5544 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5545 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5546 return Builder.CreateStore(Ops[1], Ops[0]); 5547 } 5548 case NEON::BI__builtin_neon_vst1_x2_v: 5549 case NEON::BI__builtin_neon_vst1q_x2_v: 5550 case NEON::BI__builtin_neon_vst1_x3_v: 5551 case NEON::BI__builtin_neon_vst1q_x3_v: 5552 case NEON::BI__builtin_neon_vst1_x4_v: 5553 case NEON::BI__builtin_neon_vst1q_x4_v: { 5554 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5555 llvm::Type *Tys[2] = { VTy, PTy }; 5556 unsigned Int; 5557 switch (BuiltinID) { 5558 case NEON::BI__builtin_neon_vst1_x2_v: 5559 case NEON::BI__builtin_neon_vst1q_x2_v: 5560 Int = Intrinsic::aarch64_neon_st1x2; 5561 break; 5562 case NEON::BI__builtin_neon_vst1_x3_v: 5563 case NEON::BI__builtin_neon_vst1q_x3_v: 5564 Int = Intrinsic::aarch64_neon_st1x3; 5565 break; 5566 case NEON::BI__builtin_neon_vst1_x4_v: 5567 case NEON::BI__builtin_neon_vst1q_x4_v: 5568 Int = Intrinsic::aarch64_neon_st1x4; 5569 break; 5570 } 5571 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5572 IntOps.push_back(Ops[0]); 5573 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5574 } 5575 case NEON::BI__builtin_neon_vld1_v: 5576 case NEON::BI__builtin_neon_vld1q_v: 5577 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5578 return Builder.CreateLoad(Ops[0]); 5579 case NEON::BI__builtin_neon_vst1_v: 5580 case NEON::BI__builtin_neon_vst1q_v: 5581 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5582 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5583 return Builder.CreateStore(Ops[1], Ops[0]); 5584 case NEON::BI__builtin_neon_vld1_lane_v: 5585 case NEON::BI__builtin_neon_vld1q_lane_v: 5586 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5587 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5588 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5589 Ops[0] = Builder.CreateLoad(Ops[0]); 5590 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5591 case NEON::BI__builtin_neon_vld1_dup_v: 5592 case NEON::BI__builtin_neon_vld1q_dup_v: { 5593 Value *V = UndefValue::get(Ty); 5594 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5595 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5596 Ops[0] = Builder.CreateLoad(Ops[0]); 5597 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5598 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5599 return EmitNeonSplat(Ops[0], CI); 5600 } 5601 case NEON::BI__builtin_neon_vst1_lane_v: 5602 case NEON::BI__builtin_neon_vst1q_lane_v: 5603 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5604 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5605 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5606 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5607 case NEON::BI__builtin_neon_vld2_v: 5608 case NEON::BI__builtin_neon_vld2q_v: { 5609 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5610 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5611 llvm::Type *Tys[2] = { VTy, PTy }; 5612 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5613 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5614 Ops[0] = Builder.CreateBitCast(Ops[0], 5615 llvm::PointerType::getUnqual(Ops[1]->getType())); 5616 return Builder.CreateStore(Ops[1], Ops[0]); 5617 } 5618 case NEON::BI__builtin_neon_vld3_v: 5619 case NEON::BI__builtin_neon_vld3q_v: { 5620 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5621 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5622 llvm::Type *Tys[2] = { VTy, PTy }; 5623 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5624 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5625 Ops[0] = Builder.CreateBitCast(Ops[0], 5626 llvm::PointerType::getUnqual(Ops[1]->getType())); 5627 return Builder.CreateStore(Ops[1], Ops[0]); 5628 } 5629 case NEON::BI__builtin_neon_vld4_v: 5630 case NEON::BI__builtin_neon_vld4q_v: { 5631 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5632 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5633 llvm::Type *Tys[2] = { VTy, PTy }; 5634 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5635 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5636 Ops[0] = Builder.CreateBitCast(Ops[0], 5637 llvm::PointerType::getUnqual(Ops[1]->getType())); 5638 return Builder.CreateStore(Ops[1], Ops[0]); 5639 } 5640 case NEON::BI__builtin_neon_vld2_dup_v: 5641 case NEON::BI__builtin_neon_vld2q_dup_v: { 5642 llvm::Type *PTy = 5643 llvm::PointerType::getUnqual(VTy->getElementType()); 5644 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5645 llvm::Type *Tys[2] = { VTy, PTy }; 5646 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5647 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5648 Ops[0] = Builder.CreateBitCast(Ops[0], 5649 llvm::PointerType::getUnqual(Ops[1]->getType())); 5650 return Builder.CreateStore(Ops[1], Ops[0]); 5651 } 5652 case NEON::BI__builtin_neon_vld3_dup_v: 5653 case NEON::BI__builtin_neon_vld3q_dup_v: { 5654 llvm::Type *PTy = 5655 llvm::PointerType::getUnqual(VTy->getElementType()); 5656 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5657 llvm::Type *Tys[2] = { VTy, PTy }; 5658 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5659 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5660 Ops[0] = Builder.CreateBitCast(Ops[0], 5661 llvm::PointerType::getUnqual(Ops[1]->getType())); 5662 return Builder.CreateStore(Ops[1], Ops[0]); 5663 } 5664 case NEON::BI__builtin_neon_vld4_dup_v: 5665 case NEON::BI__builtin_neon_vld4q_dup_v: { 5666 llvm::Type *PTy = 5667 llvm::PointerType::getUnqual(VTy->getElementType()); 5668 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5669 llvm::Type *Tys[2] = { VTy, PTy }; 5670 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5671 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5672 Ops[0] = Builder.CreateBitCast(Ops[0], 5673 llvm::PointerType::getUnqual(Ops[1]->getType())); 5674 return Builder.CreateStore(Ops[1], Ops[0]); 5675 } 5676 case NEON::BI__builtin_neon_vld2_lane_v: 5677 case NEON::BI__builtin_neon_vld2q_lane_v: { 5678 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5679 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5680 Ops.push_back(Ops[1]); 5681 Ops.erase(Ops.begin()+1); 5682 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5683 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5684 Ops[3] = Builder.CreateZExt(Ops[3], 5685 llvm::IntegerType::get(getLLVMContext(), 64)); 5686 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5687 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5688 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5689 return Builder.CreateStore(Ops[1], Ops[0]); 5690 } 5691 case NEON::BI__builtin_neon_vld3_lane_v: 5692 case NEON::BI__builtin_neon_vld3q_lane_v: { 5693 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5694 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5695 Ops.push_back(Ops[1]); 5696 Ops.erase(Ops.begin()+1); 5697 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5698 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5699 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5700 Ops[4] = Builder.CreateZExt(Ops[4], 5701 llvm::IntegerType::get(getLLVMContext(), 64)); 5702 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5703 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5704 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5705 return Builder.CreateStore(Ops[1], Ops[0]); 5706 } 5707 case NEON::BI__builtin_neon_vld4_lane_v: 5708 case NEON::BI__builtin_neon_vld4q_lane_v: { 5709 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5710 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5711 Ops.push_back(Ops[1]); 5712 Ops.erase(Ops.begin()+1); 5713 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5714 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5715 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5716 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5717 Ops[5] = Builder.CreateZExt(Ops[5], 5718 llvm::IntegerType::get(getLLVMContext(), 64)); 5719 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5720 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5721 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5722 return Builder.CreateStore(Ops[1], Ops[0]); 5723 } 5724 case NEON::BI__builtin_neon_vst2_v: 5725 case NEON::BI__builtin_neon_vst2q_v: { 5726 Ops.push_back(Ops[0]); 5727 Ops.erase(Ops.begin()); 5728 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5729 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5730 Ops, ""); 5731 } 5732 case NEON::BI__builtin_neon_vst2_lane_v: 5733 case NEON::BI__builtin_neon_vst2q_lane_v: { 5734 Ops.push_back(Ops[0]); 5735 Ops.erase(Ops.begin()); 5736 Ops[2] = Builder.CreateZExt(Ops[2], 5737 llvm::IntegerType::get(getLLVMContext(), 64)); 5738 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5739 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5740 Ops, ""); 5741 } 5742 case NEON::BI__builtin_neon_vst3_v: 5743 case NEON::BI__builtin_neon_vst3q_v: { 5744 Ops.push_back(Ops[0]); 5745 Ops.erase(Ops.begin()); 5746 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5747 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5748 Ops, ""); 5749 } 5750 case NEON::BI__builtin_neon_vst3_lane_v: 5751 case NEON::BI__builtin_neon_vst3q_lane_v: { 5752 Ops.push_back(Ops[0]); 5753 Ops.erase(Ops.begin()); 5754 Ops[3] = Builder.CreateZExt(Ops[3], 5755 llvm::IntegerType::get(getLLVMContext(), 64)); 5756 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5757 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5758 Ops, ""); 5759 } 5760 case NEON::BI__builtin_neon_vst4_v: 5761 case NEON::BI__builtin_neon_vst4q_v: { 5762 Ops.push_back(Ops[0]); 5763 Ops.erase(Ops.begin()); 5764 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5765 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5766 Ops, ""); 5767 } 5768 case NEON::BI__builtin_neon_vst4_lane_v: 5769 case NEON::BI__builtin_neon_vst4q_lane_v: { 5770 Ops.push_back(Ops[0]); 5771 Ops.erase(Ops.begin()); 5772 Ops[4] = Builder.CreateZExt(Ops[4], 5773 llvm::IntegerType::get(getLLVMContext(), 64)); 5774 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5775 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5776 Ops, ""); 5777 } 5778 case NEON::BI__builtin_neon_vtrn_v: 5779 case NEON::BI__builtin_neon_vtrnq_v: { 5780 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5781 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5782 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5783 Value *SV = nullptr; 5784 5785 for (unsigned vi = 0; vi != 2; ++vi) { 5786 SmallVector<Constant*, 16> Indices; 5787 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5788 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5789 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5790 } 5791 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5792 SV = llvm::ConstantVector::get(Indices); 5793 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5794 SV = Builder.CreateStore(SV, Addr); 5795 } 5796 return SV; 5797 } 5798 case NEON::BI__builtin_neon_vuzp_v: 5799 case NEON::BI__builtin_neon_vuzpq_v: { 5800 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5801 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5802 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5803 Value *SV = nullptr; 5804 5805 for (unsigned vi = 0; vi != 2; ++vi) { 5806 SmallVector<Constant*, 16> Indices; 5807 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5808 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5809 5810 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5811 SV = llvm::ConstantVector::get(Indices); 5812 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5813 SV = Builder.CreateStore(SV, Addr); 5814 } 5815 return SV; 5816 } 5817 case NEON::BI__builtin_neon_vzip_v: 5818 case NEON::BI__builtin_neon_vzipq_v: { 5819 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5820 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5821 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5822 Value *SV = nullptr; 5823 5824 for (unsigned vi = 0; vi != 2; ++vi) { 5825 SmallVector<Constant*, 16> Indices; 5826 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5827 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5828 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5829 } 5830 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5831 SV = llvm::ConstantVector::get(Indices); 5832 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5833 SV = Builder.CreateStore(SV, Addr); 5834 } 5835 return SV; 5836 } 5837 case NEON::BI__builtin_neon_vqtbl1q_v: { 5838 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5839 Ops, "vtbl1"); 5840 } 5841 case NEON::BI__builtin_neon_vqtbl2q_v: { 5842 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5843 Ops, "vtbl2"); 5844 } 5845 case NEON::BI__builtin_neon_vqtbl3q_v: { 5846 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5847 Ops, "vtbl3"); 5848 } 5849 case NEON::BI__builtin_neon_vqtbl4q_v: { 5850 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5851 Ops, "vtbl4"); 5852 } 5853 case NEON::BI__builtin_neon_vqtbx1q_v: { 5854 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5855 Ops, "vtbx1"); 5856 } 5857 case NEON::BI__builtin_neon_vqtbx2q_v: { 5858 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5859 Ops, "vtbx2"); 5860 } 5861 case NEON::BI__builtin_neon_vqtbx3q_v: { 5862 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5863 Ops, "vtbx3"); 5864 } 5865 case NEON::BI__builtin_neon_vqtbx4q_v: { 5866 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5867 Ops, "vtbx4"); 5868 } 5869 case NEON::BI__builtin_neon_vsqadd_v: 5870 case NEON::BI__builtin_neon_vsqaddq_v: { 5871 Int = Intrinsic::aarch64_neon_usqadd; 5872 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5873 } 5874 case NEON::BI__builtin_neon_vuqadd_v: 5875 case NEON::BI__builtin_neon_vuqaddq_v: { 5876 Int = Intrinsic::aarch64_neon_suqadd; 5877 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5878 } 5879 } 5880 } 5881 5882 llvm::Value *CodeGenFunction:: 5883 BuildVector(ArrayRef<llvm::Value*> Ops) { 5884 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5885 "Not a power-of-two sized vector!"); 5886 bool AllConstants = true; 5887 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5888 AllConstants &= isa<Constant>(Ops[i]); 5889 5890 // If this is a constant vector, create a ConstantVector. 5891 if (AllConstants) { 5892 SmallVector<llvm::Constant*, 16> CstOps; 5893 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5894 CstOps.push_back(cast<Constant>(Ops[i])); 5895 return llvm::ConstantVector::get(CstOps); 5896 } 5897 5898 // Otherwise, insertelement the values to build the vector. 5899 Value *Result = 5900 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5901 5902 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5903 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5904 5905 return Result; 5906 } 5907 5908 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5909 const CallExpr *E) { 5910 SmallVector<Value*, 4> Ops; 5911 5912 // Find out if any arguments are required to be integer constant expressions. 5913 unsigned ICEArguments = 0; 5914 ASTContext::GetBuiltinTypeError Error; 5915 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5916 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5917 5918 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5919 // If this is a normal argument, just emit it as a scalar. 5920 if ((ICEArguments & (1 << i)) == 0) { 5921 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5922 continue; 5923 } 5924 5925 // If this is required to be a constant, constant fold it so that we know 5926 // that the generated intrinsic gets a ConstantInt. 5927 llvm::APSInt Result; 5928 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5929 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5930 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5931 } 5932 5933 switch (BuiltinID) { 5934 default: return nullptr; 5935 case X86::BI_mm_prefetch: { 5936 Value *Address = EmitScalarExpr(E->getArg(0)); 5937 Value *RW = ConstantInt::get(Int32Ty, 0); 5938 Value *Locality = EmitScalarExpr(E->getArg(1)); 5939 Value *Data = ConstantInt::get(Int32Ty, 1); 5940 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5941 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 5942 } 5943 case X86::BI__builtin_ia32_vec_init_v8qi: 5944 case X86::BI__builtin_ia32_vec_init_v4hi: 5945 case X86::BI__builtin_ia32_vec_init_v2si: 5946 return Builder.CreateBitCast(BuildVector(Ops), 5947 llvm::Type::getX86_MMXTy(getLLVMContext())); 5948 case X86::BI__builtin_ia32_vec_ext_v2si: 5949 return Builder.CreateExtractElement(Ops[0], 5950 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5951 case X86::BI__builtin_ia32_ldmxcsr: { 5952 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5953 Builder.CreateStore(Ops[0], Tmp); 5954 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5955 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5956 } 5957 case X86::BI__builtin_ia32_stmxcsr: { 5958 Value *Tmp = CreateMemTemp(E->getType()); 5959 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5960 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5961 return Builder.CreateLoad(Tmp, "stmxcsr"); 5962 } 5963 case X86::BI__builtin_ia32_storehps: 5964 case X86::BI__builtin_ia32_storelps: { 5965 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5966 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5967 5968 // cast val v2i64 5969 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5970 5971 // extract (0, 1) 5972 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5973 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5974 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5975 5976 // cast pointer to i64 & store 5977 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5978 return Builder.CreateStore(Ops[1], Ops[0]); 5979 } 5980 case X86::BI__builtin_ia32_palignr128: 5981 case X86::BI__builtin_ia32_palignr256: { 5982 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5983 5984 unsigned NumElts = 5985 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 5986 assert(NumElts % 16 == 0); 5987 unsigned NumLanes = NumElts / 16; 5988 unsigned NumLaneElts = NumElts / NumLanes; 5989 5990 // If palignr is shifting the pair of vectors more than the size of two 5991 // lanes, emit zero. 5992 if (ShiftVal >= (2 * NumLaneElts)) 5993 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5994 5995 // If palignr is shifting the pair of input vectors more than one lane, 5996 // but less than two lanes, convert to shifting in zeroes. 5997 if (ShiftVal > NumLaneElts) { 5998 ShiftVal -= NumLaneElts; 5999 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6000 } 6001 6002 SmallVector<llvm::Constant*, 32> Indices; 6003 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6004 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6005 for (unsigned i = 0; i != NumLaneElts; ++i) { 6006 unsigned Idx = ShiftVal + i; 6007 if (Idx >= NumLaneElts) 6008 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6009 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6010 } 6011 } 6012 6013 Value* SV = llvm::ConstantVector::get(Indices); 6014 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6015 } 6016 case X86::BI__builtin_ia32_pslldqi256: { 6017 // Shift value is in bits so divide by 8. 6018 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6019 6020 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6021 if (shiftVal >= 16) 6022 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6023 6024 SmallVector<llvm::Constant*, 32> Indices; 6025 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6026 for (unsigned l = 0; l != 32; l += 16) { 6027 for (unsigned i = 0; i != 16; ++i) { 6028 unsigned Idx = 32 + i - shiftVal; 6029 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6030 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6031 } 6032 } 6033 6034 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6035 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6036 Value *Zero = llvm::Constant::getNullValue(VecTy); 6037 6038 Value *SV = llvm::ConstantVector::get(Indices); 6039 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6040 llvm::Type *ResultType = ConvertType(E->getType()); 6041 return Builder.CreateBitCast(SV, ResultType, "cast"); 6042 } 6043 case X86::BI__builtin_ia32_psrldqi256: { 6044 // Shift value is in bits so divide by 8. 6045 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6046 6047 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6048 if (shiftVal >= 16) 6049 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6050 6051 SmallVector<llvm::Constant*, 32> Indices; 6052 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6053 for (unsigned l = 0; l != 32; l += 16) { 6054 for (unsigned i = 0; i != 16; ++i) { 6055 unsigned Idx = i + shiftVal; 6056 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6057 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6058 } 6059 } 6060 6061 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6062 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6063 Value *Zero = llvm::Constant::getNullValue(VecTy); 6064 6065 Value *SV = llvm::ConstantVector::get(Indices); 6066 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6067 llvm::Type *ResultType = ConvertType(E->getType()); 6068 return Builder.CreateBitCast(SV, ResultType, "cast"); 6069 } 6070 case X86::BI__builtin_ia32_movntps: 6071 case X86::BI__builtin_ia32_movntps256: 6072 case X86::BI__builtin_ia32_movntpd: 6073 case X86::BI__builtin_ia32_movntpd256: 6074 case X86::BI__builtin_ia32_movntdq: 6075 case X86::BI__builtin_ia32_movntdq256: 6076 case X86::BI__builtin_ia32_movnti: 6077 case X86::BI__builtin_ia32_movnti64: { 6078 llvm::MDNode *Node = llvm::MDNode::get( 6079 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6080 6081 // Convert the type of the pointer to a pointer to the stored type. 6082 Value *BC = Builder.CreateBitCast(Ops[0], 6083 llvm::PointerType::getUnqual(Ops[1]->getType()), 6084 "cast"); 6085 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6086 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6087 6088 // If the operand is an integer, we can't assume alignment. Otherwise, 6089 // assume natural alignment. 6090 QualType ArgTy = E->getArg(1)->getType(); 6091 unsigned Align; 6092 if (ArgTy->isIntegerType()) 6093 Align = 1; 6094 else 6095 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6096 SI->setAlignment(Align); 6097 return SI; 6098 } 6099 // 3DNow! 6100 case X86::BI__builtin_ia32_pswapdsf: 6101 case X86::BI__builtin_ia32_pswapdsi: { 6102 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6103 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6104 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6105 return Builder.CreateCall(F, Ops, "pswapd"); 6106 } 6107 case X86::BI__builtin_ia32_rdrand16_step: 6108 case X86::BI__builtin_ia32_rdrand32_step: 6109 case X86::BI__builtin_ia32_rdrand64_step: 6110 case X86::BI__builtin_ia32_rdseed16_step: 6111 case X86::BI__builtin_ia32_rdseed32_step: 6112 case X86::BI__builtin_ia32_rdseed64_step: { 6113 Intrinsic::ID ID; 6114 switch (BuiltinID) { 6115 default: llvm_unreachable("Unsupported intrinsic!"); 6116 case X86::BI__builtin_ia32_rdrand16_step: 6117 ID = Intrinsic::x86_rdrand_16; 6118 break; 6119 case X86::BI__builtin_ia32_rdrand32_step: 6120 ID = Intrinsic::x86_rdrand_32; 6121 break; 6122 case X86::BI__builtin_ia32_rdrand64_step: 6123 ID = Intrinsic::x86_rdrand_64; 6124 break; 6125 case X86::BI__builtin_ia32_rdseed16_step: 6126 ID = Intrinsic::x86_rdseed_16; 6127 break; 6128 case X86::BI__builtin_ia32_rdseed32_step: 6129 ID = Intrinsic::x86_rdseed_32; 6130 break; 6131 case X86::BI__builtin_ia32_rdseed64_step: 6132 ID = Intrinsic::x86_rdseed_64; 6133 break; 6134 } 6135 6136 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {}); 6137 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6138 return Builder.CreateExtractValue(Call, 1); 6139 } 6140 // SSE comparison intrisics 6141 case X86::BI__builtin_ia32_cmpeqps: 6142 case X86::BI__builtin_ia32_cmpltps: 6143 case X86::BI__builtin_ia32_cmpleps: 6144 case X86::BI__builtin_ia32_cmpunordps: 6145 case X86::BI__builtin_ia32_cmpneqps: 6146 case X86::BI__builtin_ia32_cmpnltps: 6147 case X86::BI__builtin_ia32_cmpnleps: 6148 case X86::BI__builtin_ia32_cmpordps: 6149 case X86::BI__builtin_ia32_cmpeqss: 6150 case X86::BI__builtin_ia32_cmpltss: 6151 case X86::BI__builtin_ia32_cmpless: 6152 case X86::BI__builtin_ia32_cmpunordss: 6153 case X86::BI__builtin_ia32_cmpneqss: 6154 case X86::BI__builtin_ia32_cmpnltss: 6155 case X86::BI__builtin_ia32_cmpnless: 6156 case X86::BI__builtin_ia32_cmpordss: 6157 case X86::BI__builtin_ia32_cmpeqpd: 6158 case X86::BI__builtin_ia32_cmpltpd: 6159 case X86::BI__builtin_ia32_cmplepd: 6160 case X86::BI__builtin_ia32_cmpunordpd: 6161 case X86::BI__builtin_ia32_cmpneqpd: 6162 case X86::BI__builtin_ia32_cmpnltpd: 6163 case X86::BI__builtin_ia32_cmpnlepd: 6164 case X86::BI__builtin_ia32_cmpordpd: 6165 case X86::BI__builtin_ia32_cmpeqsd: 6166 case X86::BI__builtin_ia32_cmpltsd: 6167 case X86::BI__builtin_ia32_cmplesd: 6168 case X86::BI__builtin_ia32_cmpunordsd: 6169 case X86::BI__builtin_ia32_cmpneqsd: 6170 case X86::BI__builtin_ia32_cmpnltsd: 6171 case X86::BI__builtin_ia32_cmpnlesd: 6172 case X86::BI__builtin_ia32_cmpordsd: 6173 // These exist so that the builtin that takes an immediate can be bounds 6174 // checked by clang to avoid passing bad immediates to the backend. Since 6175 // AVX has a larger immediate than SSE we would need separate builtins to 6176 // do the different bounds checking. Rather than create a clang specific 6177 // SSE only builtin, this implements eight separate builtins to match gcc 6178 // implementation. 6179 6180 // Choose the immediate. 6181 unsigned Imm; 6182 switch (BuiltinID) { 6183 default: llvm_unreachable("Unsupported intrinsic!"); 6184 case X86::BI__builtin_ia32_cmpeqps: 6185 case X86::BI__builtin_ia32_cmpeqss: 6186 case X86::BI__builtin_ia32_cmpeqpd: 6187 case X86::BI__builtin_ia32_cmpeqsd: 6188 Imm = 0; 6189 break; 6190 case X86::BI__builtin_ia32_cmpltps: 6191 case X86::BI__builtin_ia32_cmpltss: 6192 case X86::BI__builtin_ia32_cmpltpd: 6193 case X86::BI__builtin_ia32_cmpltsd: 6194 Imm = 1; 6195 break; 6196 case X86::BI__builtin_ia32_cmpleps: 6197 case X86::BI__builtin_ia32_cmpless: 6198 case X86::BI__builtin_ia32_cmplepd: 6199 case X86::BI__builtin_ia32_cmplesd: 6200 Imm = 2; 6201 break; 6202 case X86::BI__builtin_ia32_cmpunordps: 6203 case X86::BI__builtin_ia32_cmpunordss: 6204 case X86::BI__builtin_ia32_cmpunordpd: 6205 case X86::BI__builtin_ia32_cmpunordsd: 6206 Imm = 3; 6207 break; 6208 case X86::BI__builtin_ia32_cmpneqps: 6209 case X86::BI__builtin_ia32_cmpneqss: 6210 case X86::BI__builtin_ia32_cmpneqpd: 6211 case X86::BI__builtin_ia32_cmpneqsd: 6212 Imm = 4; 6213 break; 6214 case X86::BI__builtin_ia32_cmpnltps: 6215 case X86::BI__builtin_ia32_cmpnltss: 6216 case X86::BI__builtin_ia32_cmpnltpd: 6217 case X86::BI__builtin_ia32_cmpnltsd: 6218 Imm = 5; 6219 break; 6220 case X86::BI__builtin_ia32_cmpnleps: 6221 case X86::BI__builtin_ia32_cmpnless: 6222 case X86::BI__builtin_ia32_cmpnlepd: 6223 case X86::BI__builtin_ia32_cmpnlesd: 6224 Imm = 6; 6225 break; 6226 case X86::BI__builtin_ia32_cmpordps: 6227 case X86::BI__builtin_ia32_cmpordss: 6228 case X86::BI__builtin_ia32_cmpordpd: 6229 case X86::BI__builtin_ia32_cmpordsd: 6230 Imm = 7; 6231 break; 6232 } 6233 6234 // Choose the intrinsic ID. 6235 const char *name; 6236 Intrinsic::ID ID; 6237 switch (BuiltinID) { 6238 default: llvm_unreachable("Unsupported intrinsic!"); 6239 case X86::BI__builtin_ia32_cmpeqps: 6240 case X86::BI__builtin_ia32_cmpltps: 6241 case X86::BI__builtin_ia32_cmpleps: 6242 case X86::BI__builtin_ia32_cmpunordps: 6243 case X86::BI__builtin_ia32_cmpneqps: 6244 case X86::BI__builtin_ia32_cmpnltps: 6245 case X86::BI__builtin_ia32_cmpnleps: 6246 case X86::BI__builtin_ia32_cmpordps: 6247 name = "cmpps"; 6248 ID = Intrinsic::x86_sse_cmp_ps; 6249 break; 6250 case X86::BI__builtin_ia32_cmpeqss: 6251 case X86::BI__builtin_ia32_cmpltss: 6252 case X86::BI__builtin_ia32_cmpless: 6253 case X86::BI__builtin_ia32_cmpunordss: 6254 case X86::BI__builtin_ia32_cmpneqss: 6255 case X86::BI__builtin_ia32_cmpnltss: 6256 case X86::BI__builtin_ia32_cmpnless: 6257 case X86::BI__builtin_ia32_cmpordss: 6258 name = "cmpss"; 6259 ID = Intrinsic::x86_sse_cmp_ss; 6260 break; 6261 case X86::BI__builtin_ia32_cmpeqpd: 6262 case X86::BI__builtin_ia32_cmpltpd: 6263 case X86::BI__builtin_ia32_cmplepd: 6264 case X86::BI__builtin_ia32_cmpunordpd: 6265 case X86::BI__builtin_ia32_cmpneqpd: 6266 case X86::BI__builtin_ia32_cmpnltpd: 6267 case X86::BI__builtin_ia32_cmpnlepd: 6268 case X86::BI__builtin_ia32_cmpordpd: 6269 name = "cmppd"; 6270 ID = Intrinsic::x86_sse2_cmp_pd; 6271 break; 6272 case X86::BI__builtin_ia32_cmpeqsd: 6273 case X86::BI__builtin_ia32_cmpltsd: 6274 case X86::BI__builtin_ia32_cmplesd: 6275 case X86::BI__builtin_ia32_cmpunordsd: 6276 case X86::BI__builtin_ia32_cmpneqsd: 6277 case X86::BI__builtin_ia32_cmpnltsd: 6278 case X86::BI__builtin_ia32_cmpnlesd: 6279 case X86::BI__builtin_ia32_cmpordsd: 6280 name = "cmpsd"; 6281 ID = Intrinsic::x86_sse2_cmp_sd; 6282 break; 6283 } 6284 6285 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6286 llvm::Function *F = CGM.getIntrinsic(ID); 6287 return Builder.CreateCall(F, Ops, name); 6288 } 6289 } 6290 6291 6292 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6293 const CallExpr *E) { 6294 SmallVector<Value*, 4> Ops; 6295 6296 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6297 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6298 6299 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6300 6301 switch (BuiltinID) { 6302 default: return nullptr; 6303 6304 // vec_ld, vec_lvsl, vec_lvsr 6305 case PPC::BI__builtin_altivec_lvx: 6306 case PPC::BI__builtin_altivec_lvxl: 6307 case PPC::BI__builtin_altivec_lvebx: 6308 case PPC::BI__builtin_altivec_lvehx: 6309 case PPC::BI__builtin_altivec_lvewx: 6310 case PPC::BI__builtin_altivec_lvsl: 6311 case PPC::BI__builtin_altivec_lvsr: 6312 case PPC::BI__builtin_vsx_lxvd2x: 6313 case PPC::BI__builtin_vsx_lxvw4x: 6314 { 6315 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6316 6317 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6318 Ops.pop_back(); 6319 6320 switch (BuiltinID) { 6321 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6322 case PPC::BI__builtin_altivec_lvx: 6323 ID = Intrinsic::ppc_altivec_lvx; 6324 break; 6325 case PPC::BI__builtin_altivec_lvxl: 6326 ID = Intrinsic::ppc_altivec_lvxl; 6327 break; 6328 case PPC::BI__builtin_altivec_lvebx: 6329 ID = Intrinsic::ppc_altivec_lvebx; 6330 break; 6331 case PPC::BI__builtin_altivec_lvehx: 6332 ID = Intrinsic::ppc_altivec_lvehx; 6333 break; 6334 case PPC::BI__builtin_altivec_lvewx: 6335 ID = Intrinsic::ppc_altivec_lvewx; 6336 break; 6337 case PPC::BI__builtin_altivec_lvsl: 6338 ID = Intrinsic::ppc_altivec_lvsl; 6339 break; 6340 case PPC::BI__builtin_altivec_lvsr: 6341 ID = Intrinsic::ppc_altivec_lvsr; 6342 break; 6343 case PPC::BI__builtin_vsx_lxvd2x: 6344 ID = Intrinsic::ppc_vsx_lxvd2x; 6345 break; 6346 case PPC::BI__builtin_vsx_lxvw4x: 6347 ID = Intrinsic::ppc_vsx_lxvw4x; 6348 break; 6349 } 6350 llvm::Function *F = CGM.getIntrinsic(ID); 6351 return Builder.CreateCall(F, Ops, ""); 6352 } 6353 6354 // vec_st 6355 case PPC::BI__builtin_altivec_stvx: 6356 case PPC::BI__builtin_altivec_stvxl: 6357 case PPC::BI__builtin_altivec_stvebx: 6358 case PPC::BI__builtin_altivec_stvehx: 6359 case PPC::BI__builtin_altivec_stvewx: 6360 case PPC::BI__builtin_vsx_stxvd2x: 6361 case PPC::BI__builtin_vsx_stxvw4x: 6362 { 6363 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6364 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6365 Ops.pop_back(); 6366 6367 switch (BuiltinID) { 6368 default: llvm_unreachable("Unsupported st intrinsic!"); 6369 case PPC::BI__builtin_altivec_stvx: 6370 ID = Intrinsic::ppc_altivec_stvx; 6371 break; 6372 case PPC::BI__builtin_altivec_stvxl: 6373 ID = Intrinsic::ppc_altivec_stvxl; 6374 break; 6375 case PPC::BI__builtin_altivec_stvebx: 6376 ID = Intrinsic::ppc_altivec_stvebx; 6377 break; 6378 case PPC::BI__builtin_altivec_stvehx: 6379 ID = Intrinsic::ppc_altivec_stvehx; 6380 break; 6381 case PPC::BI__builtin_altivec_stvewx: 6382 ID = Intrinsic::ppc_altivec_stvewx; 6383 break; 6384 case PPC::BI__builtin_vsx_stxvd2x: 6385 ID = Intrinsic::ppc_vsx_stxvd2x; 6386 break; 6387 case PPC::BI__builtin_vsx_stxvw4x: 6388 ID = Intrinsic::ppc_vsx_stxvw4x; 6389 break; 6390 } 6391 llvm::Function *F = CGM.getIntrinsic(ID); 6392 return Builder.CreateCall(F, Ops, ""); 6393 } 6394 } 6395 } 6396 6397 // Emit an intrinsic that has 1 float or double. 6398 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6399 const CallExpr *E, 6400 unsigned IntrinsicID) { 6401 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6402 6403 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6404 return CGF.Builder.CreateCall(F, Src0); 6405 } 6406 6407 // Emit an intrinsic that has 3 float or double operands. 6408 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6409 const CallExpr *E, 6410 unsigned IntrinsicID) { 6411 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6412 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6413 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6414 6415 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6416 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6417 } 6418 6419 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6420 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6421 const CallExpr *E, 6422 unsigned IntrinsicID) { 6423 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6424 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6425 6426 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6427 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6428 } 6429 6430 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6431 const CallExpr *E) { 6432 switch (BuiltinID) { 6433 case R600::BI__builtin_amdgpu_div_scale: 6434 case R600::BI__builtin_amdgpu_div_scalef: { 6435 // Translate from the intrinsics's struct return to the builtin's out 6436 // argument. 6437 6438 std::pair<llvm::Value *, unsigned> FlagOutPtr 6439 = EmitPointerWithAlignment(E->getArg(3)); 6440 6441 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6442 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6443 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6444 6445 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6446 X->getType()); 6447 6448 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6449 6450 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6451 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6452 6453 llvm::Type *RealFlagType 6454 = FlagOutPtr.first->getType()->getPointerElementType(); 6455 6456 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6457 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6458 FlagStore->setAlignment(FlagOutPtr.second); 6459 return Result; 6460 } 6461 case R600::BI__builtin_amdgpu_div_fmas: 6462 case R600::BI__builtin_amdgpu_div_fmasf: { 6463 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6464 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6465 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6466 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6467 6468 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6469 Src0->getType()); 6470 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6471 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6472 } 6473 case R600::BI__builtin_amdgpu_div_fixup: 6474 case R600::BI__builtin_amdgpu_div_fixupf: 6475 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6476 case R600::BI__builtin_amdgpu_trig_preop: 6477 case R600::BI__builtin_amdgpu_trig_preopf: 6478 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6479 case R600::BI__builtin_amdgpu_rcp: 6480 case R600::BI__builtin_amdgpu_rcpf: 6481 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6482 case R600::BI__builtin_amdgpu_rsq: 6483 case R600::BI__builtin_amdgpu_rsqf: 6484 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6485 case R600::BI__builtin_amdgpu_rsq_clamped: 6486 case R600::BI__builtin_amdgpu_rsq_clampedf: 6487 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6488 case R600::BI__builtin_amdgpu_ldexp: 6489 case R600::BI__builtin_amdgpu_ldexpf: 6490 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6491 case R600::BI__builtin_amdgpu_class: 6492 case R600::BI__builtin_amdgpu_classf: 6493 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6494 default: 6495 return nullptr; 6496 } 6497 } 6498 6499 /// Handle a SystemZ function in which the final argument is a pointer 6500 /// to an int that receives the post-instruction CC value. At the LLVM level 6501 /// this is represented as a function that returns a {result, cc} pair. 6502 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6503 unsigned IntrinsicID, 6504 const CallExpr *E) { 6505 unsigned NumArgs = E->getNumArgs() - 1; 6506 SmallVector<Value *, 8> Args(NumArgs); 6507 for (unsigned I = 0; I < NumArgs; ++I) 6508 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6509 Value *CCPtr = CGF.EmitScalarExpr(E->getArg(NumArgs)); 6510 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6511 Value *Call = CGF.Builder.CreateCall(F, Args); 6512 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6513 CGF.Builder.CreateStore(CC, CCPtr); 6514 return CGF.Builder.CreateExtractValue(Call, 0); 6515 } 6516 6517 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6518 const CallExpr *E) { 6519 switch (BuiltinID) { 6520 case SystemZ::BI__builtin_tbegin: { 6521 Value *TDB = EmitScalarExpr(E->getArg(0)); 6522 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6523 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6524 return Builder.CreateCall(F, {TDB, Control}); 6525 } 6526 case SystemZ::BI__builtin_tbegin_nofloat: { 6527 Value *TDB = EmitScalarExpr(E->getArg(0)); 6528 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6529 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6530 return Builder.CreateCall(F, {TDB, Control}); 6531 } 6532 case SystemZ::BI__builtin_tbeginc: { 6533 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6534 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6535 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6536 return Builder.CreateCall(F, {TDB, Control}); 6537 } 6538 case SystemZ::BI__builtin_tabort: { 6539 Value *Data = EmitScalarExpr(E->getArg(0)); 6540 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6541 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6542 } 6543 case SystemZ::BI__builtin_non_tx_store: { 6544 Value *Address = EmitScalarExpr(E->getArg(0)); 6545 Value *Data = EmitScalarExpr(E->getArg(1)); 6546 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6547 return Builder.CreateCall(F, {Data, Address}); 6548 } 6549 6550 // Vector builtins. Note that most vector builtins are mapped automatically 6551 // to target-specific LLVM intrinsics. The ones handled specially here can 6552 // be represented via standard LLVM IR, which is preferable to enable common 6553 // LLVM optimizations. 6554 6555 case SystemZ::BI__builtin_s390_vpopctb: 6556 case SystemZ::BI__builtin_s390_vpopcth: 6557 case SystemZ::BI__builtin_s390_vpopctf: 6558 case SystemZ::BI__builtin_s390_vpopctg: { 6559 llvm::Type *ResultType = ConvertType(E->getType()); 6560 Value *X = EmitScalarExpr(E->getArg(0)); 6561 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6562 return Builder.CreateCall(F, X); 6563 } 6564 6565 case SystemZ::BI__builtin_s390_vclzb: 6566 case SystemZ::BI__builtin_s390_vclzh: 6567 case SystemZ::BI__builtin_s390_vclzf: 6568 case SystemZ::BI__builtin_s390_vclzg: { 6569 llvm::Type *ResultType = ConvertType(E->getType()); 6570 Value *X = EmitScalarExpr(E->getArg(0)); 6571 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6572 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6573 return Builder.CreateCall(F, {X, Undef}); 6574 } 6575 6576 case SystemZ::BI__builtin_s390_vctzb: 6577 case SystemZ::BI__builtin_s390_vctzh: 6578 case SystemZ::BI__builtin_s390_vctzf: 6579 case SystemZ::BI__builtin_s390_vctzg: { 6580 llvm::Type *ResultType = ConvertType(E->getType()); 6581 Value *X = EmitScalarExpr(E->getArg(0)); 6582 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6583 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6584 return Builder.CreateCall(F, {X, Undef}); 6585 } 6586 6587 case SystemZ::BI__builtin_s390_vfsqdb: { 6588 llvm::Type *ResultType = ConvertType(E->getType()); 6589 Value *X = EmitScalarExpr(E->getArg(0)); 6590 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6591 return Builder.CreateCall(F, X); 6592 } 6593 case SystemZ::BI__builtin_s390_vfmadb: { 6594 llvm::Type *ResultType = ConvertType(E->getType()); 6595 Value *X = EmitScalarExpr(E->getArg(0)); 6596 Value *Y = EmitScalarExpr(E->getArg(1)); 6597 Value *Z = EmitScalarExpr(E->getArg(2)); 6598 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6599 return Builder.CreateCall(F, {X, Y, Z}); 6600 } 6601 case SystemZ::BI__builtin_s390_vfmsdb: { 6602 llvm::Type *ResultType = ConvertType(E->getType()); 6603 Value *X = EmitScalarExpr(E->getArg(0)); 6604 Value *Y = EmitScalarExpr(E->getArg(1)); 6605 Value *Z = EmitScalarExpr(E->getArg(2)); 6606 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6607 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6608 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6609 } 6610 case SystemZ::BI__builtin_s390_vflpdb: { 6611 llvm::Type *ResultType = ConvertType(E->getType()); 6612 Value *X = EmitScalarExpr(E->getArg(0)); 6613 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6614 return Builder.CreateCall(F, X); 6615 } 6616 case SystemZ::BI__builtin_s390_vflndb: { 6617 llvm::Type *ResultType = ConvertType(E->getType()); 6618 Value *X = EmitScalarExpr(E->getArg(0)); 6619 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6620 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6621 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 6622 } 6623 case SystemZ::BI__builtin_s390_vfidb: { 6624 llvm::Type *ResultType = ConvertType(E->getType()); 6625 Value *X = EmitScalarExpr(E->getArg(0)); 6626 // Constant-fold the M4 and M5 mask arguments. 6627 llvm::APSInt M4, M5; 6628 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 6629 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 6630 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 6631 (void)IsConstM4; (void)IsConstM5; 6632 // Check whether this instance of vfidb can be represented via a LLVM 6633 // standard intrinsic. We only support some combinations of M4 and M5. 6634 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6635 switch (M4.getZExtValue()) { 6636 default: break; 6637 case 0: // IEEE-inexact exception allowed 6638 switch (M5.getZExtValue()) { 6639 default: break; 6640 case 0: ID = Intrinsic::rint; break; 6641 } 6642 break; 6643 case 4: // IEEE-inexact exception suppressed 6644 switch (M5.getZExtValue()) { 6645 default: break; 6646 case 0: ID = Intrinsic::nearbyint; break; 6647 case 1: ID = Intrinsic::round; break; 6648 case 5: ID = Intrinsic::trunc; break; 6649 case 6: ID = Intrinsic::ceil; break; 6650 case 7: ID = Intrinsic::floor; break; 6651 } 6652 break; 6653 } 6654 if (ID != Intrinsic::not_intrinsic) { 6655 Function *F = CGM.getIntrinsic(ID, ResultType); 6656 return Builder.CreateCall(F, X); 6657 } 6658 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 6659 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 6660 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 6661 return Builder.CreateCall(F, {X, M4Value, M5Value}); 6662 } 6663 6664 // Vector intrisincs that output the post-instruction CC value. 6665 6666 #define INTRINSIC_WITH_CC(NAME) \ 6667 case SystemZ::BI__builtin_##NAME: \ 6668 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 6669 6670 INTRINSIC_WITH_CC(s390_vpkshs); 6671 INTRINSIC_WITH_CC(s390_vpksfs); 6672 INTRINSIC_WITH_CC(s390_vpksgs); 6673 6674 INTRINSIC_WITH_CC(s390_vpklshs); 6675 INTRINSIC_WITH_CC(s390_vpklsfs); 6676 INTRINSIC_WITH_CC(s390_vpklsgs); 6677 6678 INTRINSIC_WITH_CC(s390_vceqbs); 6679 INTRINSIC_WITH_CC(s390_vceqhs); 6680 INTRINSIC_WITH_CC(s390_vceqfs); 6681 INTRINSIC_WITH_CC(s390_vceqgs); 6682 6683 INTRINSIC_WITH_CC(s390_vchbs); 6684 INTRINSIC_WITH_CC(s390_vchhs); 6685 INTRINSIC_WITH_CC(s390_vchfs); 6686 INTRINSIC_WITH_CC(s390_vchgs); 6687 6688 INTRINSIC_WITH_CC(s390_vchlbs); 6689 INTRINSIC_WITH_CC(s390_vchlhs); 6690 INTRINSIC_WITH_CC(s390_vchlfs); 6691 INTRINSIC_WITH_CC(s390_vchlgs); 6692 6693 INTRINSIC_WITH_CC(s390_vfaebs); 6694 INTRINSIC_WITH_CC(s390_vfaehs); 6695 INTRINSIC_WITH_CC(s390_vfaefs); 6696 6697 INTRINSIC_WITH_CC(s390_vfaezbs); 6698 INTRINSIC_WITH_CC(s390_vfaezhs); 6699 INTRINSIC_WITH_CC(s390_vfaezfs); 6700 6701 INTRINSIC_WITH_CC(s390_vfeebs); 6702 INTRINSIC_WITH_CC(s390_vfeehs); 6703 INTRINSIC_WITH_CC(s390_vfeefs); 6704 6705 INTRINSIC_WITH_CC(s390_vfeezbs); 6706 INTRINSIC_WITH_CC(s390_vfeezhs); 6707 INTRINSIC_WITH_CC(s390_vfeezfs); 6708 6709 INTRINSIC_WITH_CC(s390_vfenebs); 6710 INTRINSIC_WITH_CC(s390_vfenehs); 6711 INTRINSIC_WITH_CC(s390_vfenefs); 6712 6713 INTRINSIC_WITH_CC(s390_vfenezbs); 6714 INTRINSIC_WITH_CC(s390_vfenezhs); 6715 INTRINSIC_WITH_CC(s390_vfenezfs); 6716 6717 INTRINSIC_WITH_CC(s390_vistrbs); 6718 INTRINSIC_WITH_CC(s390_vistrhs); 6719 INTRINSIC_WITH_CC(s390_vistrfs); 6720 6721 INTRINSIC_WITH_CC(s390_vstrcbs); 6722 INTRINSIC_WITH_CC(s390_vstrchs); 6723 INTRINSIC_WITH_CC(s390_vstrcfs); 6724 6725 INTRINSIC_WITH_CC(s390_vstrczbs); 6726 INTRINSIC_WITH_CC(s390_vstrczhs); 6727 INTRINSIC_WITH_CC(s390_vstrczfs); 6728 6729 INTRINSIC_WITH_CC(s390_vfcedbs); 6730 INTRINSIC_WITH_CC(s390_vfchdbs); 6731 INTRINSIC_WITH_CC(s390_vfchedbs); 6732 6733 INTRINSIC_WITH_CC(s390_vftcidb); 6734 6735 #undef INTRINSIC_WITH_CC 6736 6737 default: 6738 return nullptr; 6739 } 6740 } 6741