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 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 702 Dest.second, false); 703 return RValue::get(Dest.first); 704 } 705 case Builtin::BImemcpy: 706 case Builtin::BI__builtin_memcpy: { 707 std::pair<llvm::Value*, unsigned> Dest = 708 EmitPointerWithAlignment(E->getArg(0)); 709 std::pair<llvm::Value*, unsigned> Src = 710 EmitPointerWithAlignment(E->getArg(1)); 711 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 712 unsigned Align = std::min(Dest.second, Src.second); 713 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 714 return RValue::get(Dest.first); 715 } 716 717 case Builtin::BI__builtin___memcpy_chk: { 718 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 719 llvm::APSInt Size, DstSize; 720 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 721 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 722 break; 723 if (Size.ugt(DstSize)) 724 break; 725 std::pair<llvm::Value*, unsigned> Dest = 726 EmitPointerWithAlignment(E->getArg(0)); 727 std::pair<llvm::Value*, unsigned> Src = 728 EmitPointerWithAlignment(E->getArg(1)); 729 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 730 unsigned Align = std::min(Dest.second, Src.second); 731 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 732 return RValue::get(Dest.first); 733 } 734 735 case Builtin::BI__builtin_objc_memmove_collectable: { 736 Value *Address = EmitScalarExpr(E->getArg(0)); 737 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 738 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 739 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 740 Address, SrcAddr, SizeVal); 741 return RValue::get(Address); 742 } 743 744 case Builtin::BI__builtin___memmove_chk: { 745 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 746 llvm::APSInt Size, DstSize; 747 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 748 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 749 break; 750 if (Size.ugt(DstSize)) 751 break; 752 std::pair<llvm::Value*, unsigned> Dest = 753 EmitPointerWithAlignment(E->getArg(0)); 754 std::pair<llvm::Value*, unsigned> Src = 755 EmitPointerWithAlignment(E->getArg(1)); 756 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 757 unsigned Align = std::min(Dest.second, Src.second); 758 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 759 return RValue::get(Dest.first); 760 } 761 762 case Builtin::BImemmove: 763 case Builtin::BI__builtin_memmove: { 764 std::pair<llvm::Value*, unsigned> Dest = 765 EmitPointerWithAlignment(E->getArg(0)); 766 std::pair<llvm::Value*, unsigned> Src = 767 EmitPointerWithAlignment(E->getArg(1)); 768 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 769 unsigned Align = std::min(Dest.second, Src.second); 770 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 771 return RValue::get(Dest.first); 772 } 773 case Builtin::BImemset: 774 case Builtin::BI__builtin_memset: { 775 std::pair<llvm::Value*, unsigned> Dest = 776 EmitPointerWithAlignment(E->getArg(0)); 777 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 778 Builder.getInt8Ty()); 779 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 780 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 781 return RValue::get(Dest.first); 782 } 783 case Builtin::BI__builtin___memset_chk: { 784 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 785 llvm::APSInt Size, DstSize; 786 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 787 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 788 break; 789 if (Size.ugt(DstSize)) 790 break; 791 std::pair<llvm::Value*, unsigned> Dest = 792 EmitPointerWithAlignment(E->getArg(0)); 793 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 794 Builder.getInt8Ty()); 795 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 796 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 797 return RValue::get(Dest.first); 798 } 799 case Builtin::BI__builtin_dwarf_cfa: { 800 // The offset in bytes from the first argument to the CFA. 801 // 802 // Why on earth is this in the frontend? Is there any reason at 803 // all that the backend can't reasonably determine this while 804 // lowering llvm.eh.dwarf.cfa()? 805 // 806 // TODO: If there's a satisfactory reason, add a target hook for 807 // this instead of hard-coding 0, which is correct for most targets. 808 int32_t Offset = 0; 809 810 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 811 return RValue::get(Builder.CreateCall(F, 812 llvm::ConstantInt::get(Int32Ty, Offset))); 813 } 814 case Builtin::BI__builtin_return_address: { 815 Value *Depth = EmitScalarExpr(E->getArg(0)); 816 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 817 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 818 return RValue::get(Builder.CreateCall(F, Depth)); 819 } 820 case Builtin::BI__builtin_frame_address: { 821 Value *Depth = EmitScalarExpr(E->getArg(0)); 822 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 823 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 824 return RValue::get(Builder.CreateCall(F, Depth)); 825 } 826 case Builtin::BI__builtin_extract_return_addr: { 827 Value *Address = EmitScalarExpr(E->getArg(0)); 828 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 829 return RValue::get(Result); 830 } 831 case Builtin::BI__builtin_frob_return_addr: { 832 Value *Address = EmitScalarExpr(E->getArg(0)); 833 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 834 return RValue::get(Result); 835 } 836 case Builtin::BI__builtin_dwarf_sp_column: { 837 llvm::IntegerType *Ty 838 = cast<llvm::IntegerType>(ConvertType(E->getType())); 839 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 840 if (Column == -1) { 841 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 842 return RValue::get(llvm::UndefValue::get(Ty)); 843 } 844 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 845 } 846 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 847 Value *Address = EmitScalarExpr(E->getArg(0)); 848 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 849 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 850 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 851 } 852 case Builtin::BI__builtin_eh_return: { 853 Value *Int = EmitScalarExpr(E->getArg(0)); 854 Value *Ptr = EmitScalarExpr(E->getArg(1)); 855 856 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 857 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 858 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 859 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 860 ? Intrinsic::eh_return_i32 861 : Intrinsic::eh_return_i64); 862 Builder.CreateCall(F, {Int, Ptr}); 863 Builder.CreateUnreachable(); 864 865 // We do need to preserve an insertion point. 866 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 867 868 return RValue::get(nullptr); 869 } 870 case Builtin::BI__builtin_unwind_init: { 871 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 872 return RValue::get(Builder.CreateCall(F, {})); 873 } 874 case Builtin::BI__builtin_extend_pointer: { 875 // Extends a pointer to the size of an _Unwind_Word, which is 876 // uint64_t on all platforms. Generally this gets poked into a 877 // register and eventually used as an address, so if the 878 // addressing registers are wider than pointers and the platform 879 // doesn't implicitly ignore high-order bits when doing 880 // addressing, we need to make sure we zext / sext based on 881 // the platform's expectations. 882 // 883 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 884 885 // Cast the pointer to intptr_t. 886 Value *Ptr = EmitScalarExpr(E->getArg(0)); 887 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 888 889 // If that's 64 bits, we're done. 890 if (IntPtrTy->getBitWidth() == 64) 891 return RValue::get(Result); 892 893 // Otherwise, ask the codegen data what to do. 894 if (getTargetHooks().extendPointerWithSExt()) 895 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 896 else 897 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 898 } 899 case Builtin::BI__builtin_setjmp: { 900 // Buffer is a void**. 901 Value *Buf = EmitScalarExpr(E->getArg(0)); 902 903 // Store the frame pointer to the setjmp buffer. 904 Value *FrameAddr = 905 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 906 ConstantInt::get(Int32Ty, 0)); 907 Builder.CreateStore(FrameAddr, Buf); 908 909 // Store the stack pointer to the setjmp buffer. 910 Value *StackAddr = 911 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave), {}); 912 Value *StackSaveSlot = 913 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 914 Builder.CreateStore(StackAddr, StackSaveSlot); 915 916 // Call LLVM's EH setjmp, which is lightweight. 917 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 918 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 919 return RValue::get(Builder.CreateCall(F, Buf)); 920 } 921 case Builtin::BI__builtin_longjmp: { 922 Value *Buf = EmitScalarExpr(E->getArg(0)); 923 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 924 925 // Call LLVM's EH longjmp, which is lightweight. 926 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 927 928 // longjmp doesn't return; mark this as unreachable. 929 Builder.CreateUnreachable(); 930 931 // We do need to preserve an insertion point. 932 EmitBlock(createBasicBlock("longjmp.cont")); 933 934 return RValue::get(nullptr); 935 } 936 case Builtin::BI__sync_fetch_and_add: 937 case Builtin::BI__sync_fetch_and_sub: 938 case Builtin::BI__sync_fetch_and_or: 939 case Builtin::BI__sync_fetch_and_and: 940 case Builtin::BI__sync_fetch_and_xor: 941 case Builtin::BI__sync_fetch_and_nand: 942 case Builtin::BI__sync_add_and_fetch: 943 case Builtin::BI__sync_sub_and_fetch: 944 case Builtin::BI__sync_and_and_fetch: 945 case Builtin::BI__sync_or_and_fetch: 946 case Builtin::BI__sync_xor_and_fetch: 947 case Builtin::BI__sync_nand_and_fetch: 948 case Builtin::BI__sync_val_compare_and_swap: 949 case Builtin::BI__sync_bool_compare_and_swap: 950 case Builtin::BI__sync_lock_test_and_set: 951 case Builtin::BI__sync_lock_release: 952 case Builtin::BI__sync_swap: 953 llvm_unreachable("Shouldn't make it through sema"); 954 case Builtin::BI__sync_fetch_and_add_1: 955 case Builtin::BI__sync_fetch_and_add_2: 956 case Builtin::BI__sync_fetch_and_add_4: 957 case Builtin::BI__sync_fetch_and_add_8: 958 case Builtin::BI__sync_fetch_and_add_16: 959 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 960 case Builtin::BI__sync_fetch_and_sub_1: 961 case Builtin::BI__sync_fetch_and_sub_2: 962 case Builtin::BI__sync_fetch_and_sub_4: 963 case Builtin::BI__sync_fetch_and_sub_8: 964 case Builtin::BI__sync_fetch_and_sub_16: 965 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 966 case Builtin::BI__sync_fetch_and_or_1: 967 case Builtin::BI__sync_fetch_and_or_2: 968 case Builtin::BI__sync_fetch_and_or_4: 969 case Builtin::BI__sync_fetch_and_or_8: 970 case Builtin::BI__sync_fetch_and_or_16: 971 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 972 case Builtin::BI__sync_fetch_and_and_1: 973 case Builtin::BI__sync_fetch_and_and_2: 974 case Builtin::BI__sync_fetch_and_and_4: 975 case Builtin::BI__sync_fetch_and_and_8: 976 case Builtin::BI__sync_fetch_and_and_16: 977 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 978 case Builtin::BI__sync_fetch_and_xor_1: 979 case Builtin::BI__sync_fetch_and_xor_2: 980 case Builtin::BI__sync_fetch_and_xor_4: 981 case Builtin::BI__sync_fetch_and_xor_8: 982 case Builtin::BI__sync_fetch_and_xor_16: 983 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 984 case Builtin::BI__sync_fetch_and_nand_1: 985 case Builtin::BI__sync_fetch_and_nand_2: 986 case Builtin::BI__sync_fetch_and_nand_4: 987 case Builtin::BI__sync_fetch_and_nand_8: 988 case Builtin::BI__sync_fetch_and_nand_16: 989 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 990 991 // Clang extensions: not overloaded yet. 992 case Builtin::BI__sync_fetch_and_min: 993 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 994 case Builtin::BI__sync_fetch_and_max: 995 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 996 case Builtin::BI__sync_fetch_and_umin: 997 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 998 case Builtin::BI__sync_fetch_and_umax: 999 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1000 1001 case Builtin::BI__sync_add_and_fetch_1: 1002 case Builtin::BI__sync_add_and_fetch_2: 1003 case Builtin::BI__sync_add_and_fetch_4: 1004 case Builtin::BI__sync_add_and_fetch_8: 1005 case Builtin::BI__sync_add_and_fetch_16: 1006 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1007 llvm::Instruction::Add); 1008 case Builtin::BI__sync_sub_and_fetch_1: 1009 case Builtin::BI__sync_sub_and_fetch_2: 1010 case Builtin::BI__sync_sub_and_fetch_4: 1011 case Builtin::BI__sync_sub_and_fetch_8: 1012 case Builtin::BI__sync_sub_and_fetch_16: 1013 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1014 llvm::Instruction::Sub); 1015 case Builtin::BI__sync_and_and_fetch_1: 1016 case Builtin::BI__sync_and_and_fetch_2: 1017 case Builtin::BI__sync_and_and_fetch_4: 1018 case Builtin::BI__sync_and_and_fetch_8: 1019 case Builtin::BI__sync_and_and_fetch_16: 1020 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1021 llvm::Instruction::And); 1022 case Builtin::BI__sync_or_and_fetch_1: 1023 case Builtin::BI__sync_or_and_fetch_2: 1024 case Builtin::BI__sync_or_and_fetch_4: 1025 case Builtin::BI__sync_or_and_fetch_8: 1026 case Builtin::BI__sync_or_and_fetch_16: 1027 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1028 llvm::Instruction::Or); 1029 case Builtin::BI__sync_xor_and_fetch_1: 1030 case Builtin::BI__sync_xor_and_fetch_2: 1031 case Builtin::BI__sync_xor_and_fetch_4: 1032 case Builtin::BI__sync_xor_and_fetch_8: 1033 case Builtin::BI__sync_xor_and_fetch_16: 1034 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1035 llvm::Instruction::Xor); 1036 case Builtin::BI__sync_nand_and_fetch_1: 1037 case Builtin::BI__sync_nand_and_fetch_2: 1038 case Builtin::BI__sync_nand_and_fetch_4: 1039 case Builtin::BI__sync_nand_and_fetch_8: 1040 case Builtin::BI__sync_nand_and_fetch_16: 1041 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1042 llvm::Instruction::And, true); 1043 1044 case Builtin::BI__sync_val_compare_and_swap_1: 1045 case Builtin::BI__sync_val_compare_and_swap_2: 1046 case Builtin::BI__sync_val_compare_and_swap_4: 1047 case Builtin::BI__sync_val_compare_and_swap_8: 1048 case Builtin::BI__sync_val_compare_and_swap_16: { 1049 QualType T = E->getType(); 1050 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1051 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1052 1053 llvm::IntegerType *IntType = 1054 llvm::IntegerType::get(getLLVMContext(), 1055 getContext().getTypeSize(T)); 1056 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1057 1058 Value *Args[3]; 1059 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1060 Args[1] = EmitScalarExpr(E->getArg(1)); 1061 llvm::Type *ValueType = Args[1]->getType(); 1062 Args[1] = EmitToInt(*this, Args[1], T, IntType); 1063 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1064 1065 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1066 llvm::SequentiallyConsistent, 1067 llvm::SequentiallyConsistent); 1068 Result = Builder.CreateExtractValue(Result, 0); 1069 Result = EmitFromInt(*this, Result, T, ValueType); 1070 return RValue::get(Result); 1071 } 1072 1073 case Builtin::BI__sync_bool_compare_and_swap_1: 1074 case Builtin::BI__sync_bool_compare_and_swap_2: 1075 case Builtin::BI__sync_bool_compare_and_swap_4: 1076 case Builtin::BI__sync_bool_compare_and_swap_8: 1077 case Builtin::BI__sync_bool_compare_and_swap_16: { 1078 QualType T = E->getArg(1)->getType(); 1079 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1080 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1081 1082 llvm::IntegerType *IntType = 1083 llvm::IntegerType::get(getLLVMContext(), 1084 getContext().getTypeSize(T)); 1085 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1086 1087 Value *Args[3]; 1088 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1089 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1090 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1091 1092 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1093 llvm::SequentiallyConsistent, 1094 llvm::SequentiallyConsistent); 1095 Value *Result = Builder.CreateExtractValue(Pair, 1); 1096 // zext bool to int. 1097 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1098 return RValue::get(Result); 1099 } 1100 1101 case Builtin::BI__sync_swap_1: 1102 case Builtin::BI__sync_swap_2: 1103 case Builtin::BI__sync_swap_4: 1104 case Builtin::BI__sync_swap_8: 1105 case Builtin::BI__sync_swap_16: 1106 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1107 1108 case Builtin::BI__sync_lock_test_and_set_1: 1109 case Builtin::BI__sync_lock_test_and_set_2: 1110 case Builtin::BI__sync_lock_test_and_set_4: 1111 case Builtin::BI__sync_lock_test_and_set_8: 1112 case Builtin::BI__sync_lock_test_and_set_16: 1113 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1114 1115 case Builtin::BI__sync_lock_release_1: 1116 case Builtin::BI__sync_lock_release_2: 1117 case Builtin::BI__sync_lock_release_4: 1118 case Builtin::BI__sync_lock_release_8: 1119 case Builtin::BI__sync_lock_release_16: { 1120 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1121 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1122 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1123 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1124 StoreSize.getQuantity() * 8); 1125 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1126 llvm::StoreInst *Store = 1127 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1128 Store->setAlignment(StoreSize.getQuantity()); 1129 Store->setAtomic(llvm::Release); 1130 return RValue::get(nullptr); 1131 } 1132 1133 case Builtin::BI__sync_synchronize: { 1134 // We assume this is supposed to correspond to a C++0x-style 1135 // sequentially-consistent fence (i.e. this is only usable for 1136 // synchonization, not device I/O or anything like that). This intrinsic 1137 // is really badly designed in the sense that in theory, there isn't 1138 // any way to safely use it... but in practice, it mostly works 1139 // to use it with non-atomic loads and stores to get acquire/release 1140 // semantics. 1141 Builder.CreateFence(llvm::SequentiallyConsistent); 1142 return RValue::get(nullptr); 1143 } 1144 1145 case Builtin::BI__c11_atomic_is_lock_free: 1146 case Builtin::BI__atomic_is_lock_free: { 1147 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1148 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1149 // _Atomic(T) is always properly-aligned. 1150 const char *LibCallName = "__atomic_is_lock_free"; 1151 CallArgList Args; 1152 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1153 getContext().getSizeType()); 1154 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1155 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1156 getContext().VoidPtrTy); 1157 else 1158 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1159 getContext().VoidPtrTy); 1160 const CGFunctionInfo &FuncInfo = 1161 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1162 FunctionType::ExtInfo(), 1163 RequiredArgs::All); 1164 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1165 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1166 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1167 } 1168 1169 case Builtin::BI__atomic_test_and_set: { 1170 // Look at the argument type to determine whether this is a volatile 1171 // operation. The parameter type is always volatile. 1172 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1173 bool Volatile = 1174 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1175 1176 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1177 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1178 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1179 Value *NewVal = Builder.getInt8(1); 1180 Value *Order = EmitScalarExpr(E->getArg(1)); 1181 if (isa<llvm::ConstantInt>(Order)) { 1182 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1183 AtomicRMWInst *Result = nullptr; 1184 switch (ord) { 1185 case 0: // memory_order_relaxed 1186 default: // invalid order 1187 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1188 Ptr, NewVal, 1189 llvm::Monotonic); 1190 break; 1191 case 1: // memory_order_consume 1192 case 2: // memory_order_acquire 1193 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1194 Ptr, NewVal, 1195 llvm::Acquire); 1196 break; 1197 case 3: // memory_order_release 1198 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1199 Ptr, NewVal, 1200 llvm::Release); 1201 break; 1202 case 4: // memory_order_acq_rel 1203 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1204 Ptr, NewVal, 1205 llvm::AcquireRelease); 1206 break; 1207 case 5: // memory_order_seq_cst 1208 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1209 Ptr, NewVal, 1210 llvm::SequentiallyConsistent); 1211 break; 1212 } 1213 Result->setVolatile(Volatile); 1214 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1215 } 1216 1217 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1218 1219 llvm::BasicBlock *BBs[5] = { 1220 createBasicBlock("monotonic", CurFn), 1221 createBasicBlock("acquire", CurFn), 1222 createBasicBlock("release", CurFn), 1223 createBasicBlock("acqrel", CurFn), 1224 createBasicBlock("seqcst", CurFn) 1225 }; 1226 llvm::AtomicOrdering Orders[5] = { 1227 llvm::Monotonic, llvm::Acquire, llvm::Release, 1228 llvm::AcquireRelease, llvm::SequentiallyConsistent 1229 }; 1230 1231 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1232 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1233 1234 Builder.SetInsertPoint(ContBB); 1235 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1236 1237 for (unsigned i = 0; i < 5; ++i) { 1238 Builder.SetInsertPoint(BBs[i]); 1239 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1240 Ptr, NewVal, Orders[i]); 1241 RMW->setVolatile(Volatile); 1242 Result->addIncoming(RMW, BBs[i]); 1243 Builder.CreateBr(ContBB); 1244 } 1245 1246 SI->addCase(Builder.getInt32(0), BBs[0]); 1247 SI->addCase(Builder.getInt32(1), BBs[1]); 1248 SI->addCase(Builder.getInt32(2), BBs[1]); 1249 SI->addCase(Builder.getInt32(3), BBs[2]); 1250 SI->addCase(Builder.getInt32(4), BBs[3]); 1251 SI->addCase(Builder.getInt32(5), BBs[4]); 1252 1253 Builder.SetInsertPoint(ContBB); 1254 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1255 } 1256 1257 case Builtin::BI__atomic_clear: { 1258 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1259 bool Volatile = 1260 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1261 1262 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1263 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1264 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1265 Value *NewVal = Builder.getInt8(0); 1266 Value *Order = EmitScalarExpr(E->getArg(1)); 1267 if (isa<llvm::ConstantInt>(Order)) { 1268 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1269 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1270 Store->setAlignment(1); 1271 switch (ord) { 1272 case 0: // memory_order_relaxed 1273 default: // invalid order 1274 Store->setOrdering(llvm::Monotonic); 1275 break; 1276 case 3: // memory_order_release 1277 Store->setOrdering(llvm::Release); 1278 break; 1279 case 5: // memory_order_seq_cst 1280 Store->setOrdering(llvm::SequentiallyConsistent); 1281 break; 1282 } 1283 return RValue::get(nullptr); 1284 } 1285 1286 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1287 1288 llvm::BasicBlock *BBs[3] = { 1289 createBasicBlock("monotonic", CurFn), 1290 createBasicBlock("release", CurFn), 1291 createBasicBlock("seqcst", CurFn) 1292 }; 1293 llvm::AtomicOrdering Orders[3] = { 1294 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1295 }; 1296 1297 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1298 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1299 1300 for (unsigned i = 0; i < 3; ++i) { 1301 Builder.SetInsertPoint(BBs[i]); 1302 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1303 Store->setAlignment(1); 1304 Store->setOrdering(Orders[i]); 1305 Builder.CreateBr(ContBB); 1306 } 1307 1308 SI->addCase(Builder.getInt32(0), BBs[0]); 1309 SI->addCase(Builder.getInt32(3), BBs[1]); 1310 SI->addCase(Builder.getInt32(5), BBs[2]); 1311 1312 Builder.SetInsertPoint(ContBB); 1313 return RValue::get(nullptr); 1314 } 1315 1316 case Builtin::BI__atomic_thread_fence: 1317 case Builtin::BI__atomic_signal_fence: 1318 case Builtin::BI__c11_atomic_thread_fence: 1319 case Builtin::BI__c11_atomic_signal_fence: { 1320 llvm::SynchronizationScope Scope; 1321 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1322 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1323 Scope = llvm::SingleThread; 1324 else 1325 Scope = llvm::CrossThread; 1326 Value *Order = EmitScalarExpr(E->getArg(0)); 1327 if (isa<llvm::ConstantInt>(Order)) { 1328 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1329 switch (ord) { 1330 case 0: // memory_order_relaxed 1331 default: // invalid order 1332 break; 1333 case 1: // memory_order_consume 1334 case 2: // memory_order_acquire 1335 Builder.CreateFence(llvm::Acquire, Scope); 1336 break; 1337 case 3: // memory_order_release 1338 Builder.CreateFence(llvm::Release, Scope); 1339 break; 1340 case 4: // memory_order_acq_rel 1341 Builder.CreateFence(llvm::AcquireRelease, Scope); 1342 break; 1343 case 5: // memory_order_seq_cst 1344 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1345 break; 1346 } 1347 return RValue::get(nullptr); 1348 } 1349 1350 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1351 AcquireBB = createBasicBlock("acquire", CurFn); 1352 ReleaseBB = createBasicBlock("release", CurFn); 1353 AcqRelBB = createBasicBlock("acqrel", CurFn); 1354 SeqCstBB = createBasicBlock("seqcst", CurFn); 1355 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1356 1357 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1358 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1359 1360 Builder.SetInsertPoint(AcquireBB); 1361 Builder.CreateFence(llvm::Acquire, Scope); 1362 Builder.CreateBr(ContBB); 1363 SI->addCase(Builder.getInt32(1), AcquireBB); 1364 SI->addCase(Builder.getInt32(2), AcquireBB); 1365 1366 Builder.SetInsertPoint(ReleaseBB); 1367 Builder.CreateFence(llvm::Release, Scope); 1368 Builder.CreateBr(ContBB); 1369 SI->addCase(Builder.getInt32(3), ReleaseBB); 1370 1371 Builder.SetInsertPoint(AcqRelBB); 1372 Builder.CreateFence(llvm::AcquireRelease, Scope); 1373 Builder.CreateBr(ContBB); 1374 SI->addCase(Builder.getInt32(4), AcqRelBB); 1375 1376 Builder.SetInsertPoint(SeqCstBB); 1377 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1378 Builder.CreateBr(ContBB); 1379 SI->addCase(Builder.getInt32(5), SeqCstBB); 1380 1381 Builder.SetInsertPoint(ContBB); 1382 return RValue::get(nullptr); 1383 } 1384 1385 // Library functions with special handling. 1386 case Builtin::BIsqrt: 1387 case Builtin::BIsqrtf: 1388 case Builtin::BIsqrtl: { 1389 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1390 // in finite- or unsafe-math mode (the intrinsic has different semantics 1391 // for handling negative numbers compared to the library function, so 1392 // -fmath-errno=0 is not enough). 1393 if (!FD->hasAttr<ConstAttr>()) 1394 break; 1395 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1396 CGM.getCodeGenOpts().NoNaNsFPMath)) 1397 break; 1398 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1399 llvm::Type *ArgType = Arg0->getType(); 1400 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1401 return RValue::get(Builder.CreateCall(F, Arg0)); 1402 } 1403 1404 case Builtin::BI__builtin_pow: 1405 case Builtin::BI__builtin_powf: 1406 case Builtin::BI__builtin_powl: 1407 case Builtin::BIpow: 1408 case Builtin::BIpowf: 1409 case Builtin::BIpowl: { 1410 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1411 if (!FD->hasAttr<ConstAttr>()) 1412 break; 1413 Value *Base = EmitScalarExpr(E->getArg(0)); 1414 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1415 llvm::Type *ArgType = Base->getType(); 1416 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1417 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1418 } 1419 1420 case Builtin::BIfma: 1421 case Builtin::BIfmaf: 1422 case Builtin::BIfmal: 1423 case Builtin::BI__builtin_fma: 1424 case Builtin::BI__builtin_fmaf: 1425 case Builtin::BI__builtin_fmal: { 1426 // Rewrite fma to intrinsic. 1427 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1428 llvm::Type *ArgType = FirstArg->getType(); 1429 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1430 return RValue::get( 1431 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1432 EmitScalarExpr(E->getArg(2))})); 1433 } 1434 1435 case Builtin::BI__builtin_signbit: 1436 case Builtin::BI__builtin_signbitf: 1437 case Builtin::BI__builtin_signbitl: { 1438 return RValue::get( 1439 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1440 ConvertType(E->getType()))); 1441 } 1442 case Builtin::BI__builtin_annotation: { 1443 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1444 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1445 AnnVal->getType()); 1446 1447 // Get the annotation string, go through casts. Sema requires this to be a 1448 // non-wide string literal, potentially casted, so the cast<> is safe. 1449 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1450 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1451 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1452 } 1453 case Builtin::BI__builtin_addcb: 1454 case Builtin::BI__builtin_addcs: 1455 case Builtin::BI__builtin_addc: 1456 case Builtin::BI__builtin_addcl: 1457 case Builtin::BI__builtin_addcll: 1458 case Builtin::BI__builtin_subcb: 1459 case Builtin::BI__builtin_subcs: 1460 case Builtin::BI__builtin_subc: 1461 case Builtin::BI__builtin_subcl: 1462 case Builtin::BI__builtin_subcll: { 1463 1464 // We translate all of these builtins from expressions of the form: 1465 // int x = ..., y = ..., carryin = ..., carryout, result; 1466 // result = __builtin_addc(x, y, carryin, &carryout); 1467 // 1468 // to LLVM IR of the form: 1469 // 1470 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1471 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1472 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1473 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1474 // i32 %carryin) 1475 // %result = extractvalue {i32, i1} %tmp2, 0 1476 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1477 // %tmp3 = or i1 %carry1, %carry2 1478 // %tmp4 = zext i1 %tmp3 to i32 1479 // store i32 %tmp4, i32* %carryout 1480 1481 // Scalarize our inputs. 1482 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1483 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1484 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1485 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1486 EmitPointerWithAlignment(E->getArg(3)); 1487 1488 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1489 llvm::Intrinsic::ID IntrinsicId; 1490 switch (BuiltinID) { 1491 default: llvm_unreachable("Unknown multiprecision builtin id."); 1492 case Builtin::BI__builtin_addcb: 1493 case Builtin::BI__builtin_addcs: 1494 case Builtin::BI__builtin_addc: 1495 case Builtin::BI__builtin_addcl: 1496 case Builtin::BI__builtin_addcll: 1497 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1498 break; 1499 case Builtin::BI__builtin_subcb: 1500 case Builtin::BI__builtin_subcs: 1501 case Builtin::BI__builtin_subc: 1502 case Builtin::BI__builtin_subcl: 1503 case Builtin::BI__builtin_subcll: 1504 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1505 break; 1506 } 1507 1508 // Construct our resulting LLVM IR expression. 1509 llvm::Value *Carry1; 1510 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1511 X, Y, Carry1); 1512 llvm::Value *Carry2; 1513 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1514 Sum1, Carryin, Carry2); 1515 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1516 X->getType()); 1517 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1518 CarryOutPtr.first); 1519 CarryOutStore->setAlignment(CarryOutPtr.second); 1520 return RValue::get(Sum2); 1521 } 1522 case Builtin::BI__builtin_uadd_overflow: 1523 case Builtin::BI__builtin_uaddl_overflow: 1524 case Builtin::BI__builtin_uaddll_overflow: 1525 case Builtin::BI__builtin_usub_overflow: 1526 case Builtin::BI__builtin_usubl_overflow: 1527 case Builtin::BI__builtin_usubll_overflow: 1528 case Builtin::BI__builtin_umul_overflow: 1529 case Builtin::BI__builtin_umull_overflow: 1530 case Builtin::BI__builtin_umulll_overflow: 1531 case Builtin::BI__builtin_sadd_overflow: 1532 case Builtin::BI__builtin_saddl_overflow: 1533 case Builtin::BI__builtin_saddll_overflow: 1534 case Builtin::BI__builtin_ssub_overflow: 1535 case Builtin::BI__builtin_ssubl_overflow: 1536 case Builtin::BI__builtin_ssubll_overflow: 1537 case Builtin::BI__builtin_smul_overflow: 1538 case Builtin::BI__builtin_smull_overflow: 1539 case Builtin::BI__builtin_smulll_overflow: { 1540 1541 // We translate all of these builtins directly to the relevant llvm IR node. 1542 1543 // Scalarize our inputs. 1544 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1545 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1546 std::pair<llvm::Value *, unsigned> SumOutPtr = 1547 EmitPointerWithAlignment(E->getArg(2)); 1548 1549 // Decide which of the overflow intrinsics we are lowering to: 1550 llvm::Intrinsic::ID IntrinsicId; 1551 switch (BuiltinID) { 1552 default: llvm_unreachable("Unknown security overflow builtin id."); 1553 case Builtin::BI__builtin_uadd_overflow: 1554 case Builtin::BI__builtin_uaddl_overflow: 1555 case Builtin::BI__builtin_uaddll_overflow: 1556 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1557 break; 1558 case Builtin::BI__builtin_usub_overflow: 1559 case Builtin::BI__builtin_usubl_overflow: 1560 case Builtin::BI__builtin_usubll_overflow: 1561 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1562 break; 1563 case Builtin::BI__builtin_umul_overflow: 1564 case Builtin::BI__builtin_umull_overflow: 1565 case Builtin::BI__builtin_umulll_overflow: 1566 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1567 break; 1568 case Builtin::BI__builtin_sadd_overflow: 1569 case Builtin::BI__builtin_saddl_overflow: 1570 case Builtin::BI__builtin_saddll_overflow: 1571 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1572 break; 1573 case Builtin::BI__builtin_ssub_overflow: 1574 case Builtin::BI__builtin_ssubl_overflow: 1575 case Builtin::BI__builtin_ssubll_overflow: 1576 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1577 break; 1578 case Builtin::BI__builtin_smul_overflow: 1579 case Builtin::BI__builtin_smull_overflow: 1580 case Builtin::BI__builtin_smulll_overflow: 1581 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1582 break; 1583 } 1584 1585 1586 llvm::Value *Carry; 1587 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1588 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1589 SumOutStore->setAlignment(SumOutPtr.second); 1590 1591 return RValue::get(Carry); 1592 } 1593 case Builtin::BI__builtin_addressof: 1594 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1595 case Builtin::BI__builtin_operator_new: 1596 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1597 E->getArg(0), false); 1598 case Builtin::BI__builtin_operator_delete: 1599 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1600 E->getArg(0), true); 1601 case Builtin::BI__noop: 1602 // __noop always evaluates to an integer literal zero. 1603 return RValue::get(ConstantInt::get(IntTy, 0)); 1604 case Builtin::BI__builtin_call_with_static_chain: { 1605 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1606 const Expr *Chain = E->getArg(1); 1607 return EmitCall(Call->getCallee()->getType(), 1608 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1609 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1610 } 1611 case Builtin::BI_InterlockedExchange: 1612 case Builtin::BI_InterlockedExchangePointer: 1613 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1614 case Builtin::BI_InterlockedCompareExchangePointer: { 1615 llvm::Type *RTy; 1616 llvm::IntegerType *IntType = 1617 IntegerType::get(getLLVMContext(), 1618 getContext().getTypeSize(E->getType())); 1619 llvm::Type *IntPtrType = IntType->getPointerTo(); 1620 1621 llvm::Value *Destination = 1622 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1623 1624 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1625 RTy = Exchange->getType(); 1626 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1627 1628 llvm::Value *Comparand = 1629 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1630 1631 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1632 SequentiallyConsistent, 1633 SequentiallyConsistent); 1634 Result->setVolatile(true); 1635 1636 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1637 0), 1638 RTy)); 1639 } 1640 case Builtin::BI_InterlockedCompareExchange: { 1641 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1642 EmitScalarExpr(E->getArg(0)), 1643 EmitScalarExpr(E->getArg(2)), 1644 EmitScalarExpr(E->getArg(1)), 1645 SequentiallyConsistent, 1646 SequentiallyConsistent); 1647 CXI->setVolatile(true); 1648 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1649 } 1650 case Builtin::BI_InterlockedIncrement: { 1651 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1652 AtomicRMWInst::Add, 1653 EmitScalarExpr(E->getArg(0)), 1654 ConstantInt::get(Int32Ty, 1), 1655 llvm::SequentiallyConsistent); 1656 RMWI->setVolatile(true); 1657 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1658 } 1659 case Builtin::BI_InterlockedDecrement: { 1660 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1661 AtomicRMWInst::Sub, 1662 EmitScalarExpr(E->getArg(0)), 1663 ConstantInt::get(Int32Ty, 1), 1664 llvm::SequentiallyConsistent); 1665 RMWI->setVolatile(true); 1666 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1667 } 1668 case Builtin::BI_InterlockedExchangeAdd: { 1669 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1670 AtomicRMWInst::Add, 1671 EmitScalarExpr(E->getArg(0)), 1672 EmitScalarExpr(E->getArg(1)), 1673 llvm::SequentiallyConsistent); 1674 RMWI->setVolatile(true); 1675 return RValue::get(RMWI); 1676 } 1677 case Builtin::BI__readfsdword: { 1678 Value *IntToPtr = 1679 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1680 llvm::PointerType::get(CGM.Int32Ty, 257)); 1681 LoadInst *Load = 1682 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1683 return RValue::get(Load); 1684 } 1685 1686 case Builtin::BI__exception_code: 1687 case Builtin::BI_exception_code: 1688 return RValue::get(EmitSEHExceptionCode()); 1689 case Builtin::BI__exception_info: 1690 case Builtin::BI_exception_info: 1691 return RValue::get(EmitSEHExceptionInfo()); 1692 case Builtin::BI__abnormal_termination: 1693 case Builtin::BI_abnormal_termination: 1694 return RValue::get(EmitSEHAbnormalTermination()); 1695 case Builtin::BI_setjmpex: { 1696 if (getTarget().getTriple().isOSMSVCRT()) { 1697 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1698 llvm::AttributeSet ReturnsTwiceAttr = 1699 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1700 llvm::Attribute::ReturnsTwice); 1701 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1702 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1703 "_setjmpex", ReturnsTwiceAttr); 1704 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1705 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1706 llvm::Value *FrameAddr = 1707 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1708 ConstantInt::get(Int32Ty, 0)); 1709 llvm::Value *Args[] = {Buf, FrameAddr}; 1710 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1711 CS.setAttributes(ReturnsTwiceAttr); 1712 return RValue::get(CS.getInstruction()); 1713 } 1714 break; 1715 } 1716 case Builtin::BI_setjmp: { 1717 if (getTarget().getTriple().isOSMSVCRT()) { 1718 llvm::AttributeSet ReturnsTwiceAttr = 1719 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1720 llvm::Attribute::ReturnsTwice); 1721 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1722 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1723 llvm::CallSite CS; 1724 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1725 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1726 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1727 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1728 "_setjmp3", ReturnsTwiceAttr); 1729 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1730 llvm::Value *Args[] = {Buf, Count}; 1731 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1732 } else { 1733 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1734 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1735 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1736 "_setjmp", ReturnsTwiceAttr); 1737 llvm::Value *FrameAddr = 1738 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1739 ConstantInt::get(Int32Ty, 0)); 1740 llvm::Value *Args[] = {Buf, FrameAddr}; 1741 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1742 } 1743 CS.setAttributes(ReturnsTwiceAttr); 1744 return RValue::get(CS.getInstruction()); 1745 } 1746 break; 1747 } 1748 1749 case Builtin::BI__GetExceptionInfo: { 1750 if (llvm::GlobalVariable *GV = 1751 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1752 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1753 break; 1754 } 1755 } 1756 1757 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1758 // the call using the normal call path, but using the unmangled 1759 // version of the function name. 1760 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1761 return emitLibraryCall(*this, FD, E, 1762 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1763 1764 // If this is a predefined lib function (e.g. malloc), emit the call 1765 // using exactly the normal call path. 1766 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1767 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1768 1769 // See if we have a target specific intrinsic. 1770 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1771 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1772 if (const char *Prefix = 1773 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1774 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1775 // NOTE we dont need to perform a compatibility flag check here since the 1776 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1777 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1778 if (IntrinsicID == Intrinsic::not_intrinsic) 1779 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1780 } 1781 1782 if (IntrinsicID != Intrinsic::not_intrinsic) { 1783 SmallVector<Value*, 16> Args; 1784 1785 // Find out if any arguments are required to be integer constant 1786 // expressions. 1787 unsigned ICEArguments = 0; 1788 ASTContext::GetBuiltinTypeError Error; 1789 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1790 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1791 1792 Function *F = CGM.getIntrinsic(IntrinsicID); 1793 llvm::FunctionType *FTy = F->getFunctionType(); 1794 1795 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1796 Value *ArgValue; 1797 // If this is a normal argument, just emit it as a scalar. 1798 if ((ICEArguments & (1 << i)) == 0) { 1799 ArgValue = EmitScalarExpr(E->getArg(i)); 1800 } else { 1801 // If this is required to be a constant, constant fold it so that we 1802 // know that the generated intrinsic gets a ConstantInt. 1803 llvm::APSInt Result; 1804 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1805 assert(IsConst && "Constant arg isn't actually constant?"); 1806 (void)IsConst; 1807 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1808 } 1809 1810 // If the intrinsic arg type is different from the builtin arg type 1811 // we need to do a bit cast. 1812 llvm::Type *PTy = FTy->getParamType(i); 1813 if (PTy != ArgValue->getType()) { 1814 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1815 "Must be able to losslessly bit cast to param"); 1816 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1817 } 1818 1819 Args.push_back(ArgValue); 1820 } 1821 1822 Value *V = Builder.CreateCall(F, Args); 1823 QualType BuiltinRetType = E->getType(); 1824 1825 llvm::Type *RetTy = VoidTy; 1826 if (!BuiltinRetType->isVoidType()) 1827 RetTy = ConvertType(BuiltinRetType); 1828 1829 if (RetTy != V->getType()) { 1830 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1831 "Must be able to losslessly bit cast result type"); 1832 V = Builder.CreateBitCast(V, RetTy); 1833 } 1834 1835 return RValue::get(V); 1836 } 1837 1838 // See if we have a target specific builtin that needs to be lowered. 1839 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1840 return RValue::get(V); 1841 1842 ErrorUnsupported(E, "builtin function"); 1843 1844 // Unknown builtin, for now just dump it out and return undef. 1845 return GetUndefRValue(E->getType()); 1846 } 1847 1848 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1849 const CallExpr *E) { 1850 switch (getTarget().getTriple().getArch()) { 1851 case llvm::Triple::arm: 1852 case llvm::Triple::armeb: 1853 case llvm::Triple::thumb: 1854 case llvm::Triple::thumbeb: 1855 return EmitARMBuiltinExpr(BuiltinID, E); 1856 case llvm::Triple::aarch64: 1857 case llvm::Triple::aarch64_be: 1858 return EmitAArch64BuiltinExpr(BuiltinID, E); 1859 case llvm::Triple::x86: 1860 case llvm::Triple::x86_64: 1861 return EmitX86BuiltinExpr(BuiltinID, E); 1862 case llvm::Triple::ppc: 1863 case llvm::Triple::ppc64: 1864 case llvm::Triple::ppc64le: 1865 return EmitPPCBuiltinExpr(BuiltinID, E); 1866 case llvm::Triple::r600: 1867 case llvm::Triple::amdgcn: 1868 return EmitR600BuiltinExpr(BuiltinID, E); 1869 case llvm::Triple::systemz: 1870 return EmitSystemZBuiltinExpr(BuiltinID, E); 1871 default: 1872 return nullptr; 1873 } 1874 } 1875 1876 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1877 NeonTypeFlags TypeFlags, 1878 bool V1Ty=false) { 1879 int IsQuad = TypeFlags.isQuad(); 1880 switch (TypeFlags.getEltType()) { 1881 case NeonTypeFlags::Int8: 1882 case NeonTypeFlags::Poly8: 1883 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1884 case NeonTypeFlags::Int16: 1885 case NeonTypeFlags::Poly16: 1886 case NeonTypeFlags::Float16: 1887 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1888 case NeonTypeFlags::Int32: 1889 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1890 case NeonTypeFlags::Int64: 1891 case NeonTypeFlags::Poly64: 1892 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1893 case NeonTypeFlags::Poly128: 1894 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1895 // There is a lot of i128 and f128 API missing. 1896 // so we use v16i8 to represent poly128 and get pattern matched. 1897 return llvm::VectorType::get(CGF->Int8Ty, 16); 1898 case NeonTypeFlags::Float32: 1899 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1900 case NeonTypeFlags::Float64: 1901 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1902 } 1903 llvm_unreachable("Unknown vector element type!"); 1904 } 1905 1906 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1907 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1908 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1909 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1910 } 1911 1912 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1913 const char *name, 1914 unsigned shift, bool rightshift) { 1915 unsigned j = 0; 1916 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1917 ai != ae; ++ai, ++j) 1918 if (shift > 0 && shift == j) 1919 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1920 else 1921 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1922 1923 return Builder.CreateCall(F, Ops, name); 1924 } 1925 1926 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1927 bool neg) { 1928 int SV = cast<ConstantInt>(V)->getSExtValue(); 1929 1930 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1931 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1932 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1933 } 1934 1935 // \brief Right-shift a vector by a constant. 1936 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1937 llvm::Type *Ty, bool usgn, 1938 const char *name) { 1939 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1940 1941 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1942 int EltSize = VTy->getScalarSizeInBits(); 1943 1944 Vec = Builder.CreateBitCast(Vec, Ty); 1945 1946 // lshr/ashr are undefined when the shift amount is equal to the vector 1947 // element size. 1948 if (ShiftAmt == EltSize) { 1949 if (usgn) { 1950 // Right-shifting an unsigned value by its size yields 0. 1951 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1952 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1953 } else { 1954 // Right-shifting a signed value by its size is equivalent 1955 // to a shift of size-1. 1956 --ShiftAmt; 1957 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1958 } 1959 } 1960 1961 Shift = EmitNeonShiftVector(Shift, Ty, false); 1962 if (usgn) 1963 return Builder.CreateLShr(Vec, Shift, name); 1964 else 1965 return Builder.CreateAShr(Vec, Shift, name); 1966 } 1967 1968 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1969 /// alignment of the type referenced by the pointer. Skip over implicit 1970 /// casts. 1971 std::pair<llvm::Value*, unsigned> 1972 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1973 assert(Addr->getType()->isPointerType()); 1974 Addr = Addr->IgnoreParens(); 1975 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1976 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1977 ICE->getSubExpr()->getType()->isPointerType()) { 1978 std::pair<llvm::Value*, unsigned> Ptr = 1979 EmitPointerWithAlignment(ICE->getSubExpr()); 1980 Ptr.first = Builder.CreateBitCast(Ptr.first, 1981 ConvertType(Addr->getType())); 1982 return Ptr; 1983 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1984 LValue LV = EmitLValue(ICE->getSubExpr()); 1985 unsigned Align = LV.getAlignment().getQuantity(); 1986 if (!Align) { 1987 // FIXME: Once LValues are fixed to always set alignment, 1988 // zap this code. 1989 QualType PtTy = ICE->getSubExpr()->getType(); 1990 if (!PtTy->isIncompleteType()) 1991 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1992 else 1993 Align = 1; 1994 } 1995 return std::make_pair(LV.getAddress(), Align); 1996 } 1997 } 1998 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1999 if (UO->getOpcode() == UO_AddrOf) { 2000 LValue LV = EmitLValue(UO->getSubExpr()); 2001 unsigned Align = LV.getAlignment().getQuantity(); 2002 if (!Align) { 2003 // FIXME: Once LValues are fixed to always set alignment, 2004 // zap this code. 2005 QualType PtTy = UO->getSubExpr()->getType(); 2006 if (!PtTy->isIncompleteType()) 2007 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2008 else 2009 Align = 1; 2010 } 2011 return std::make_pair(LV.getAddress(), Align); 2012 } 2013 } 2014 2015 unsigned Align = 1; 2016 QualType PtTy = Addr->getType()->getPointeeType(); 2017 if (!PtTy->isIncompleteType()) 2018 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2019 2020 return std::make_pair(EmitScalarExpr(Addr), Align); 2021 } 2022 2023 enum { 2024 AddRetType = (1 << 0), 2025 Add1ArgType = (1 << 1), 2026 Add2ArgTypes = (1 << 2), 2027 2028 VectorizeRetType = (1 << 3), 2029 VectorizeArgTypes = (1 << 4), 2030 2031 InventFloatType = (1 << 5), 2032 UnsignedAlts = (1 << 6), 2033 2034 Use64BitVectors = (1 << 7), 2035 Use128BitVectors = (1 << 8), 2036 2037 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2038 VectorRet = AddRetType | VectorizeRetType, 2039 VectorRetGetArgs01 = 2040 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2041 FpCmpzModifiers = 2042 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2043 }; 2044 2045 struct NeonIntrinsicInfo { 2046 unsigned BuiltinID; 2047 unsigned LLVMIntrinsic; 2048 unsigned AltLLVMIntrinsic; 2049 const char *NameHint; 2050 unsigned TypeModifier; 2051 2052 bool operator<(unsigned RHSBuiltinID) const { 2053 return BuiltinID < RHSBuiltinID; 2054 } 2055 }; 2056 2057 #define NEONMAP0(NameBase) \ 2058 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2059 2060 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2061 { NEON:: BI__builtin_neon_ ## NameBase, \ 2062 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2063 2064 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2065 { NEON:: BI__builtin_neon_ ## NameBase, \ 2066 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2067 #NameBase, TypeModifier } 2068 2069 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2070 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2071 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2072 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2073 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2074 NEONMAP0(vaddhn_v), 2075 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2076 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2077 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2078 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2079 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2080 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2081 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2082 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2083 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2084 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2085 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2086 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2087 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2088 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2089 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2090 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2091 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2092 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2093 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2094 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2095 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 2096 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2097 NEONMAP0(vcvt_f32_v), 2098 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2099 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2100 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2101 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2102 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2103 NEONMAP0(vcvt_s32_v), 2104 NEONMAP0(vcvt_s64_v), 2105 NEONMAP0(vcvt_u32_v), 2106 NEONMAP0(vcvt_u64_v), 2107 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2108 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2109 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2110 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2111 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2112 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2113 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2114 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2115 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2116 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2117 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2118 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2119 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2120 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2121 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2122 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2123 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2124 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2125 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2126 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2127 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2128 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2129 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2130 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2131 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2132 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2133 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2134 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2135 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2136 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2137 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2138 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2139 NEONMAP0(vcvtq_f32_v), 2140 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2141 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2142 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2143 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2144 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2145 NEONMAP0(vcvtq_s32_v), 2146 NEONMAP0(vcvtq_s64_v), 2147 NEONMAP0(vcvtq_u32_v), 2148 NEONMAP0(vcvtq_u64_v), 2149 NEONMAP0(vext_v), 2150 NEONMAP0(vextq_v), 2151 NEONMAP0(vfma_v), 2152 NEONMAP0(vfmaq_v), 2153 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2154 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2155 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2156 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2157 NEONMAP0(vld1_dup_v), 2158 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2159 NEONMAP0(vld1q_dup_v), 2160 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2161 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2162 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2163 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2164 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2165 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2166 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2167 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2168 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2169 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2170 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2171 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2172 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2173 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2174 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2175 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2176 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2177 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2178 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2179 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2180 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2181 NEONMAP0(vmovl_v), 2182 NEONMAP0(vmovn_v), 2183 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2184 NEONMAP0(vmull_v), 2185 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2186 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2187 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2188 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2189 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2190 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2191 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2192 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2193 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2194 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2195 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2196 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2197 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2198 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2199 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2200 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2201 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2202 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2203 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2204 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2205 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2206 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2207 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2208 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2209 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2210 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2211 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2212 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2213 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2214 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2215 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2216 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2217 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2218 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2219 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2220 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2221 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2222 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2223 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2224 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2225 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2226 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2227 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2228 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2229 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2230 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2231 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2232 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2233 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2234 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2235 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2236 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2237 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2238 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2239 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2240 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2241 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2242 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2243 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2244 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2245 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2246 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2247 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2248 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2249 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2250 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2251 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2252 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2253 NEONMAP0(vshl_n_v), 2254 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2255 NEONMAP0(vshll_n_v), 2256 NEONMAP0(vshlq_n_v), 2257 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2258 NEONMAP0(vshr_n_v), 2259 NEONMAP0(vshrn_n_v), 2260 NEONMAP0(vshrq_n_v), 2261 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2262 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2263 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2264 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2265 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2266 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2267 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2268 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2269 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2270 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2271 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2272 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2273 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2274 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2275 NEONMAP0(vsubhn_v), 2276 NEONMAP0(vtrn_v), 2277 NEONMAP0(vtrnq_v), 2278 NEONMAP0(vtst_v), 2279 NEONMAP0(vtstq_v), 2280 NEONMAP0(vuzp_v), 2281 NEONMAP0(vuzpq_v), 2282 NEONMAP0(vzip_v), 2283 NEONMAP0(vzipq_v) 2284 }; 2285 2286 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2287 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2288 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2289 NEONMAP0(vaddhn_v), 2290 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2291 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2292 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2293 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2294 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2295 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2296 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2297 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2298 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2299 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2300 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2301 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2302 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2303 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2304 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2305 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2306 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2307 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2308 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2309 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2310 NEONMAP0(vcvt_f32_v), 2311 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2312 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2313 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2314 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2315 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2316 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2317 NEONMAP0(vcvtq_f32_v), 2318 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2319 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2320 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2321 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2322 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2323 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2324 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2325 NEONMAP0(vext_v), 2326 NEONMAP0(vextq_v), 2327 NEONMAP0(vfma_v), 2328 NEONMAP0(vfmaq_v), 2329 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2330 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2331 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2332 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2333 NEONMAP0(vmovl_v), 2334 NEONMAP0(vmovn_v), 2335 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2336 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2337 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2338 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2339 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2340 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2341 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2342 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2343 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2344 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2345 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2346 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2347 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2348 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2349 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2350 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2351 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2352 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2353 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2354 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2355 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2356 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2357 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2358 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2359 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2360 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2361 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2362 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2363 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2364 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2365 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2366 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2367 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2368 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2369 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2370 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2371 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2372 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2373 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2374 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2375 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2376 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2377 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2378 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2379 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2380 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2381 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2382 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2383 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2384 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2385 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2386 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2387 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2388 NEONMAP0(vshl_n_v), 2389 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2390 NEONMAP0(vshll_n_v), 2391 NEONMAP0(vshlq_n_v), 2392 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2393 NEONMAP0(vshr_n_v), 2394 NEONMAP0(vshrn_n_v), 2395 NEONMAP0(vshrq_n_v), 2396 NEONMAP0(vsubhn_v), 2397 NEONMAP0(vtst_v), 2398 NEONMAP0(vtstq_v), 2399 }; 2400 2401 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2402 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2403 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2404 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2405 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2406 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2407 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2408 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2409 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2410 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2411 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2412 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2413 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2414 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2415 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2416 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2417 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2418 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2419 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2420 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2421 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2422 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2423 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2424 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2425 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2426 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2427 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2428 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2429 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2430 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2431 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2432 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2433 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2434 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2435 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2436 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2437 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2438 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2439 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2440 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2441 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2442 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2443 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2444 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2445 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2446 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2447 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2448 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2449 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2450 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2451 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2452 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2453 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2454 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2455 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2456 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2457 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2458 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2459 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2460 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2461 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2462 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2463 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2464 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2465 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2466 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2467 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2468 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2469 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2470 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2471 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2472 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2473 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2474 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2475 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2476 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2477 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2478 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2479 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2480 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2481 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2482 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2483 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2484 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2485 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2486 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2487 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2488 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2489 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2490 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2491 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2492 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2493 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2494 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2495 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2496 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2497 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2498 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2499 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2500 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2501 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2502 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2503 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2504 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2505 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2506 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2507 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2508 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2509 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2510 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2511 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2512 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2513 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2514 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2515 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2516 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2517 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2518 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2519 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2520 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2521 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2522 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2523 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2524 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2525 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2526 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2527 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2528 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2529 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2530 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2531 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2532 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2533 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2534 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2535 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2536 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2537 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2538 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2539 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2540 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2541 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2542 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2543 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2544 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2545 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2546 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2547 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2548 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2549 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2550 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2551 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2552 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2553 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2554 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2555 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2556 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2557 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2558 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2559 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2560 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2561 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2562 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2563 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2564 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2565 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2566 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2567 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2568 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2569 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2570 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2571 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2572 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2573 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2574 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2575 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2576 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2577 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2578 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2579 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2580 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2581 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2582 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2583 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2584 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2585 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2586 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2587 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2588 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2589 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2590 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2591 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2592 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2593 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2594 }; 2595 2596 #undef NEONMAP0 2597 #undef NEONMAP1 2598 #undef NEONMAP2 2599 2600 static bool NEONSIMDIntrinsicsProvenSorted = false; 2601 2602 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2603 static bool AArch64SISDIntrinsicsProvenSorted = false; 2604 2605 2606 static const NeonIntrinsicInfo * 2607 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2608 unsigned BuiltinID, bool &MapProvenSorted) { 2609 2610 #ifndef NDEBUG 2611 if (!MapProvenSorted) { 2612 // FIXME: use std::is_sorted once C++11 is allowed 2613 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2614 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2615 MapProvenSorted = true; 2616 } 2617 #endif 2618 2619 const NeonIntrinsicInfo *Builtin = 2620 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2621 2622 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2623 return Builtin; 2624 2625 return nullptr; 2626 } 2627 2628 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2629 unsigned Modifier, 2630 llvm::Type *ArgType, 2631 const CallExpr *E) { 2632 int VectorSize = 0; 2633 if (Modifier & Use64BitVectors) 2634 VectorSize = 64; 2635 else if (Modifier & Use128BitVectors) 2636 VectorSize = 128; 2637 2638 // Return type. 2639 SmallVector<llvm::Type *, 3> Tys; 2640 if (Modifier & AddRetType) { 2641 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2642 if (Modifier & VectorizeRetType) 2643 Ty = llvm::VectorType::get( 2644 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2645 2646 Tys.push_back(Ty); 2647 } 2648 2649 // Arguments. 2650 if (Modifier & VectorizeArgTypes) { 2651 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2652 ArgType = llvm::VectorType::get(ArgType, Elts); 2653 } 2654 2655 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2656 Tys.push_back(ArgType); 2657 2658 if (Modifier & Add2ArgTypes) 2659 Tys.push_back(ArgType); 2660 2661 if (Modifier & InventFloatType) 2662 Tys.push_back(FloatTy); 2663 2664 return CGM.getIntrinsic(IntrinsicID, Tys); 2665 } 2666 2667 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2668 const NeonIntrinsicInfo &SISDInfo, 2669 SmallVectorImpl<Value *> &Ops, 2670 const CallExpr *E) { 2671 unsigned BuiltinID = SISDInfo.BuiltinID; 2672 unsigned int Int = SISDInfo.LLVMIntrinsic; 2673 unsigned Modifier = SISDInfo.TypeModifier; 2674 const char *s = SISDInfo.NameHint; 2675 2676 switch (BuiltinID) { 2677 case NEON::BI__builtin_neon_vcled_s64: 2678 case NEON::BI__builtin_neon_vcled_u64: 2679 case NEON::BI__builtin_neon_vcles_f32: 2680 case NEON::BI__builtin_neon_vcled_f64: 2681 case NEON::BI__builtin_neon_vcltd_s64: 2682 case NEON::BI__builtin_neon_vcltd_u64: 2683 case NEON::BI__builtin_neon_vclts_f32: 2684 case NEON::BI__builtin_neon_vcltd_f64: 2685 case NEON::BI__builtin_neon_vcales_f32: 2686 case NEON::BI__builtin_neon_vcaled_f64: 2687 case NEON::BI__builtin_neon_vcalts_f32: 2688 case NEON::BI__builtin_neon_vcaltd_f64: 2689 // Only one direction of comparisons actually exist, cmle is actually a cmge 2690 // with swapped operands. The table gives us the right intrinsic but we 2691 // still need to do the swap. 2692 std::swap(Ops[0], Ops[1]); 2693 break; 2694 } 2695 2696 assert(Int && "Generic code assumes a valid intrinsic"); 2697 2698 // Determine the type(s) of this overloaded AArch64 intrinsic. 2699 const Expr *Arg = E->getArg(0); 2700 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2701 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2702 2703 int j = 0; 2704 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2705 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2706 ai != ae; ++ai, ++j) { 2707 llvm::Type *ArgTy = ai->getType(); 2708 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2709 ArgTy->getPrimitiveSizeInBits()) 2710 continue; 2711 2712 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2713 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2714 // it before inserting. 2715 Ops[j] = 2716 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2717 Ops[j] = 2718 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2719 } 2720 2721 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2722 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2723 if (ResultType->getPrimitiveSizeInBits() < 2724 Result->getType()->getPrimitiveSizeInBits()) 2725 return CGF.Builder.CreateExtractElement(Result, C0); 2726 2727 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2728 } 2729 2730 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2731 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2732 const char *NameHint, unsigned Modifier, const CallExpr *E, 2733 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2734 // Get the last argument, which specifies the vector type. 2735 llvm::APSInt NeonTypeConst; 2736 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2737 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2738 return nullptr; 2739 2740 // Determine the type of this overloaded NEON intrinsic. 2741 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2742 bool Usgn = Type.isUnsigned(); 2743 bool Quad = Type.isQuad(); 2744 2745 llvm::VectorType *VTy = GetNeonType(this, Type); 2746 llvm::Type *Ty = VTy; 2747 if (!Ty) 2748 return nullptr; 2749 2750 unsigned Int = LLVMIntrinsic; 2751 if ((Modifier & UnsignedAlts) && !Usgn) 2752 Int = AltLLVMIntrinsic; 2753 2754 switch (BuiltinID) { 2755 default: break; 2756 case NEON::BI__builtin_neon_vabs_v: 2757 case NEON::BI__builtin_neon_vabsq_v: 2758 if (VTy->getElementType()->isFloatingPointTy()) 2759 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2760 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2761 case NEON::BI__builtin_neon_vaddhn_v: { 2762 llvm::VectorType *SrcTy = 2763 llvm::VectorType::getExtendedElementVectorType(VTy); 2764 2765 // %sum = add <4 x i32> %lhs, %rhs 2766 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2767 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2768 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2769 2770 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2771 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2772 SrcTy->getScalarSizeInBits() / 2); 2773 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2774 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2775 2776 // %res = trunc <4 x i32> %high to <4 x i16> 2777 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2778 } 2779 case NEON::BI__builtin_neon_vcale_v: 2780 case NEON::BI__builtin_neon_vcaleq_v: 2781 case NEON::BI__builtin_neon_vcalt_v: 2782 case NEON::BI__builtin_neon_vcaltq_v: 2783 std::swap(Ops[0], Ops[1]); 2784 case NEON::BI__builtin_neon_vcage_v: 2785 case NEON::BI__builtin_neon_vcageq_v: 2786 case NEON::BI__builtin_neon_vcagt_v: 2787 case NEON::BI__builtin_neon_vcagtq_v: { 2788 llvm::Type *VecFlt = llvm::VectorType::get( 2789 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2790 VTy->getNumElements()); 2791 llvm::Type *Tys[] = { VTy, VecFlt }; 2792 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2793 return EmitNeonCall(F, Ops, NameHint); 2794 } 2795 case NEON::BI__builtin_neon_vclz_v: 2796 case NEON::BI__builtin_neon_vclzq_v: 2797 // We generate target-independent intrinsic, which needs a second argument 2798 // for whether or not clz of zero is undefined; on ARM it isn't. 2799 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2800 break; 2801 case NEON::BI__builtin_neon_vcvt_f32_v: 2802 case NEON::BI__builtin_neon_vcvtq_f32_v: 2803 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2804 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2805 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2806 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2807 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2808 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2809 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2810 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2811 bool Double = 2812 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2813 llvm::Type *FloatTy = 2814 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2815 : NeonTypeFlags::Float32, 2816 false, Quad)); 2817 llvm::Type *Tys[2] = { FloatTy, Ty }; 2818 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2819 Function *F = CGM.getIntrinsic(Int, Tys); 2820 return EmitNeonCall(F, Ops, "vcvt_n"); 2821 } 2822 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2823 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2824 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2825 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2826 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2827 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2828 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2829 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2830 bool Double = 2831 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2832 llvm::Type *FloatTy = 2833 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2834 : NeonTypeFlags::Float32, 2835 false, Quad)); 2836 llvm::Type *Tys[2] = { Ty, FloatTy }; 2837 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2838 return EmitNeonCall(F, Ops, "vcvt_n"); 2839 } 2840 case NEON::BI__builtin_neon_vcvt_s32_v: 2841 case NEON::BI__builtin_neon_vcvt_u32_v: 2842 case NEON::BI__builtin_neon_vcvt_s64_v: 2843 case NEON::BI__builtin_neon_vcvt_u64_v: 2844 case NEON::BI__builtin_neon_vcvtq_s32_v: 2845 case NEON::BI__builtin_neon_vcvtq_u32_v: 2846 case NEON::BI__builtin_neon_vcvtq_s64_v: 2847 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2848 bool Double = 2849 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2850 llvm::Type *FloatTy = 2851 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2852 : NeonTypeFlags::Float32, 2853 false, Quad)); 2854 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2855 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2856 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2857 } 2858 case NEON::BI__builtin_neon_vcvta_s32_v: 2859 case NEON::BI__builtin_neon_vcvta_s64_v: 2860 case NEON::BI__builtin_neon_vcvta_u32_v: 2861 case NEON::BI__builtin_neon_vcvta_u64_v: 2862 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2863 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2864 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2865 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2866 case NEON::BI__builtin_neon_vcvtn_s32_v: 2867 case NEON::BI__builtin_neon_vcvtn_s64_v: 2868 case NEON::BI__builtin_neon_vcvtn_u32_v: 2869 case NEON::BI__builtin_neon_vcvtn_u64_v: 2870 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2871 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2872 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2873 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2874 case NEON::BI__builtin_neon_vcvtp_s32_v: 2875 case NEON::BI__builtin_neon_vcvtp_s64_v: 2876 case NEON::BI__builtin_neon_vcvtp_u32_v: 2877 case NEON::BI__builtin_neon_vcvtp_u64_v: 2878 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2879 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2880 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2881 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2882 case NEON::BI__builtin_neon_vcvtm_s32_v: 2883 case NEON::BI__builtin_neon_vcvtm_s64_v: 2884 case NEON::BI__builtin_neon_vcvtm_u32_v: 2885 case NEON::BI__builtin_neon_vcvtm_u64_v: 2886 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2887 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2888 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2889 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2890 bool Double = 2891 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2892 llvm::Type *InTy = 2893 GetNeonType(this, 2894 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2895 : NeonTypeFlags::Float32, false, Quad)); 2896 llvm::Type *Tys[2] = { Ty, InTy }; 2897 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2898 } 2899 case NEON::BI__builtin_neon_vext_v: 2900 case NEON::BI__builtin_neon_vextq_v: { 2901 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2902 SmallVector<Constant*, 16> Indices; 2903 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2904 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2905 2906 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2907 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2908 Value *SV = llvm::ConstantVector::get(Indices); 2909 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2910 } 2911 case NEON::BI__builtin_neon_vfma_v: 2912 case NEON::BI__builtin_neon_vfmaq_v: { 2913 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2914 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2915 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2916 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2917 2918 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2919 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 2920 } 2921 case NEON::BI__builtin_neon_vld1_v: 2922 case NEON::BI__builtin_neon_vld1q_v: 2923 Ops.push_back(Align); 2924 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2925 case NEON::BI__builtin_neon_vld2_v: 2926 case NEON::BI__builtin_neon_vld2q_v: 2927 case NEON::BI__builtin_neon_vld3_v: 2928 case NEON::BI__builtin_neon_vld3q_v: 2929 case NEON::BI__builtin_neon_vld4_v: 2930 case NEON::BI__builtin_neon_vld4q_v: { 2931 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2932 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 2933 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2934 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2935 return Builder.CreateStore(Ops[1], Ops[0]); 2936 } 2937 case NEON::BI__builtin_neon_vld1_dup_v: 2938 case NEON::BI__builtin_neon_vld1q_dup_v: { 2939 Value *V = UndefValue::get(Ty); 2940 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2941 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2942 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2943 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2944 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2945 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2946 return EmitNeonSplat(Ops[0], CI); 2947 } 2948 case NEON::BI__builtin_neon_vld2_lane_v: 2949 case NEON::BI__builtin_neon_vld2q_lane_v: 2950 case NEON::BI__builtin_neon_vld3_lane_v: 2951 case NEON::BI__builtin_neon_vld3q_lane_v: 2952 case NEON::BI__builtin_neon_vld4_lane_v: 2953 case NEON::BI__builtin_neon_vld4q_lane_v: { 2954 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2955 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2956 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2957 Ops.push_back(Align); 2958 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2959 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2960 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2961 return Builder.CreateStore(Ops[1], Ops[0]); 2962 } 2963 case NEON::BI__builtin_neon_vmovl_v: { 2964 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2965 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2966 if (Usgn) 2967 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2968 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2969 } 2970 case NEON::BI__builtin_neon_vmovn_v: { 2971 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2972 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2973 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2974 } 2975 case NEON::BI__builtin_neon_vmull_v: 2976 // FIXME: the integer vmull operations could be emitted in terms of pure 2977 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2978 // hoisting the exts outside loops. Until global ISel comes along that can 2979 // see through such movement this leads to bad CodeGen. So we need an 2980 // intrinsic for now. 2981 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2982 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2983 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2984 case NEON::BI__builtin_neon_vpadal_v: 2985 case NEON::BI__builtin_neon_vpadalq_v: { 2986 // The source operand type has twice as many elements of half the size. 2987 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2988 llvm::Type *EltTy = 2989 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2990 llvm::Type *NarrowTy = 2991 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2992 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2993 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2994 } 2995 case NEON::BI__builtin_neon_vpaddl_v: 2996 case NEON::BI__builtin_neon_vpaddlq_v: { 2997 // The source operand type has twice as many elements of half the size. 2998 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2999 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3000 llvm::Type *NarrowTy = 3001 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3002 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3003 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3004 } 3005 case NEON::BI__builtin_neon_vqdmlal_v: 3006 case NEON::BI__builtin_neon_vqdmlsl_v: { 3007 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3008 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 3009 MulOps, "vqdmlal"); 3010 3011 SmallVector<Value *, 2> AccumOps; 3012 AccumOps.push_back(Ops[0]); 3013 AccumOps.push_back(Mul); 3014 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 3015 AccumOps, NameHint); 3016 } 3017 case NEON::BI__builtin_neon_vqshl_n_v: 3018 case NEON::BI__builtin_neon_vqshlq_n_v: 3019 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3020 1, false); 3021 case NEON::BI__builtin_neon_vqshlu_n_v: 3022 case NEON::BI__builtin_neon_vqshluq_n_v: 3023 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3024 1, false); 3025 case NEON::BI__builtin_neon_vrecpe_v: 3026 case NEON::BI__builtin_neon_vrecpeq_v: 3027 case NEON::BI__builtin_neon_vrsqrte_v: 3028 case NEON::BI__builtin_neon_vrsqrteq_v: 3029 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3030 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3031 3032 case NEON::BI__builtin_neon_vrshr_n_v: 3033 case NEON::BI__builtin_neon_vrshrq_n_v: 3034 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3035 1, true); 3036 case NEON::BI__builtin_neon_vshl_n_v: 3037 case NEON::BI__builtin_neon_vshlq_n_v: 3038 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3039 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3040 "vshl_n"); 3041 case NEON::BI__builtin_neon_vshll_n_v: { 3042 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3043 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3044 if (Usgn) 3045 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3046 else 3047 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3048 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3049 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3050 } 3051 case NEON::BI__builtin_neon_vshrn_n_v: { 3052 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3053 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3054 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3055 if (Usgn) 3056 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3057 else 3058 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3059 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3060 } 3061 case NEON::BI__builtin_neon_vshr_n_v: 3062 case NEON::BI__builtin_neon_vshrq_n_v: 3063 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3064 case NEON::BI__builtin_neon_vst1_v: 3065 case NEON::BI__builtin_neon_vst1q_v: 3066 case NEON::BI__builtin_neon_vst2_v: 3067 case NEON::BI__builtin_neon_vst2q_v: 3068 case NEON::BI__builtin_neon_vst3_v: 3069 case NEON::BI__builtin_neon_vst3q_v: 3070 case NEON::BI__builtin_neon_vst4_v: 3071 case NEON::BI__builtin_neon_vst4q_v: 3072 case NEON::BI__builtin_neon_vst2_lane_v: 3073 case NEON::BI__builtin_neon_vst2q_lane_v: 3074 case NEON::BI__builtin_neon_vst3_lane_v: 3075 case NEON::BI__builtin_neon_vst3q_lane_v: 3076 case NEON::BI__builtin_neon_vst4_lane_v: 3077 case NEON::BI__builtin_neon_vst4q_lane_v: 3078 Ops.push_back(Align); 3079 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3080 case NEON::BI__builtin_neon_vsubhn_v: { 3081 llvm::VectorType *SrcTy = 3082 llvm::VectorType::getExtendedElementVectorType(VTy); 3083 3084 // %sum = add <4 x i32> %lhs, %rhs 3085 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3086 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3087 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3088 3089 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3090 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 3091 SrcTy->getScalarSizeInBits() / 2); 3092 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 3093 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3094 3095 // %res = trunc <4 x i32> %high to <4 x i16> 3096 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3097 } 3098 case NEON::BI__builtin_neon_vtrn_v: 3099 case NEON::BI__builtin_neon_vtrnq_v: { 3100 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3101 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3102 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3103 Value *SV = nullptr; 3104 3105 for (unsigned vi = 0; vi != 2; ++vi) { 3106 SmallVector<Constant*, 16> Indices; 3107 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3108 Indices.push_back(Builder.getInt32(i+vi)); 3109 Indices.push_back(Builder.getInt32(i+e+vi)); 3110 } 3111 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3112 SV = llvm::ConstantVector::get(Indices); 3113 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3114 SV = Builder.CreateStore(SV, Addr); 3115 } 3116 return SV; 3117 } 3118 case NEON::BI__builtin_neon_vtst_v: 3119 case NEON::BI__builtin_neon_vtstq_v: { 3120 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3121 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3122 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3123 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3124 ConstantAggregateZero::get(Ty)); 3125 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3126 } 3127 case NEON::BI__builtin_neon_vuzp_v: 3128 case NEON::BI__builtin_neon_vuzpq_v: { 3129 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3130 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3131 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3132 Value *SV = nullptr; 3133 3134 for (unsigned vi = 0; vi != 2; ++vi) { 3135 SmallVector<Constant*, 16> Indices; 3136 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3137 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3138 3139 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3140 SV = llvm::ConstantVector::get(Indices); 3141 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3142 SV = Builder.CreateStore(SV, Addr); 3143 } 3144 return SV; 3145 } 3146 case NEON::BI__builtin_neon_vzip_v: 3147 case NEON::BI__builtin_neon_vzipq_v: { 3148 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3149 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3150 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3151 Value *SV = nullptr; 3152 3153 for (unsigned vi = 0; vi != 2; ++vi) { 3154 SmallVector<Constant*, 16> Indices; 3155 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3156 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3157 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3158 } 3159 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3160 SV = llvm::ConstantVector::get(Indices); 3161 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3162 SV = Builder.CreateStore(SV, Addr); 3163 } 3164 return SV; 3165 } 3166 } 3167 3168 assert(Int && "Expected valid intrinsic number"); 3169 3170 // Determine the type(s) of this overloaded AArch64 intrinsic. 3171 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3172 3173 Value *Result = EmitNeonCall(F, Ops, NameHint); 3174 llvm::Type *ResultType = ConvertType(E->getType()); 3175 // AArch64 intrinsic one-element vector type cast to 3176 // scalar type expected by the builtin 3177 return Builder.CreateBitCast(Result, ResultType, NameHint); 3178 } 3179 3180 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3181 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3182 const CmpInst::Predicate Ip, const Twine &Name) { 3183 llvm::Type *OTy = Op->getType(); 3184 3185 // FIXME: this is utterly horrific. We should not be looking at previous 3186 // codegen context to find out what needs doing. Unfortunately TableGen 3187 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3188 // (etc). 3189 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3190 OTy = BI->getOperand(0)->getType(); 3191 3192 Op = Builder.CreateBitCast(Op, OTy); 3193 if (OTy->getScalarType()->isFloatingPointTy()) { 3194 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3195 } else { 3196 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3197 } 3198 return Builder.CreateSExt(Op, Ty, Name); 3199 } 3200 3201 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3202 Value *ExtOp, Value *IndexOp, 3203 llvm::Type *ResTy, unsigned IntID, 3204 const char *Name) { 3205 SmallVector<Value *, 2> TblOps; 3206 if (ExtOp) 3207 TblOps.push_back(ExtOp); 3208 3209 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3210 SmallVector<Constant*, 16> Indices; 3211 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3212 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3213 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3214 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3215 } 3216 Value *SV = llvm::ConstantVector::get(Indices); 3217 3218 int PairPos = 0, End = Ops.size() - 1; 3219 while (PairPos < End) { 3220 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3221 Ops[PairPos+1], SV, Name)); 3222 PairPos += 2; 3223 } 3224 3225 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3226 // of the 128-bit lookup table with zero. 3227 if (PairPos == End) { 3228 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3229 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3230 ZeroTbl, SV, Name)); 3231 } 3232 3233 Function *TblF; 3234 TblOps.push_back(IndexOp); 3235 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3236 3237 return CGF.EmitNeonCall(TblF, TblOps, Name); 3238 } 3239 3240 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3241 switch (BuiltinID) { 3242 default: 3243 return nullptr; 3244 case ARM::BI__builtin_arm_nop: 3245 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3246 llvm::ConstantInt::get(Int32Ty, 0)); 3247 case ARM::BI__builtin_arm_yield: 3248 case ARM::BI__yield: 3249 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3250 llvm::ConstantInt::get(Int32Ty, 1)); 3251 case ARM::BI__builtin_arm_wfe: 3252 case ARM::BI__wfe: 3253 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3254 llvm::ConstantInt::get(Int32Ty, 2)); 3255 case ARM::BI__builtin_arm_wfi: 3256 case ARM::BI__wfi: 3257 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3258 llvm::ConstantInt::get(Int32Ty, 3)); 3259 case ARM::BI__builtin_arm_sev: 3260 case ARM::BI__sev: 3261 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3262 llvm::ConstantInt::get(Int32Ty, 4)); 3263 case ARM::BI__builtin_arm_sevl: 3264 case ARM::BI__sevl: 3265 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3266 llvm::ConstantInt::get(Int32Ty, 5)); 3267 } 3268 } 3269 3270 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3271 const CallExpr *E) { 3272 if (auto Hint = GetValueForARMHint(BuiltinID)) 3273 return Hint; 3274 3275 if (BuiltinID == ARM::BI__emit) { 3276 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3277 llvm::FunctionType *FTy = 3278 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3279 3280 APSInt Value; 3281 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3282 llvm_unreachable("Sema will ensure that the parameter is constant"); 3283 3284 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3285 3286 llvm::InlineAsm *Emit = 3287 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3288 /*SideEffects=*/true) 3289 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3290 /*SideEffects=*/true); 3291 3292 return Builder.CreateCall(Emit, {}); 3293 } 3294 3295 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3296 Value *Option = EmitScalarExpr(E->getArg(0)); 3297 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3298 } 3299 3300 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3301 Value *Address = EmitScalarExpr(E->getArg(0)); 3302 Value *RW = EmitScalarExpr(E->getArg(1)); 3303 Value *IsData = EmitScalarExpr(E->getArg(2)); 3304 3305 // Locality is not supported on ARM target 3306 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3307 3308 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3309 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3310 } 3311 3312 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3313 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3314 EmitScalarExpr(E->getArg(0)), 3315 "rbit"); 3316 } 3317 3318 if (BuiltinID == ARM::BI__clear_cache) { 3319 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3320 const FunctionDecl *FD = E->getDirectCallee(); 3321 SmallVector<Value*, 2> Ops; 3322 for (unsigned i = 0; i < 2; i++) 3323 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3324 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3325 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3326 StringRef Name = FD->getName(); 3327 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3328 } 3329 3330 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3331 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3332 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3333 getContext().getTypeSize(E->getType()) == 64) || 3334 BuiltinID == ARM::BI__ldrexd) { 3335 Function *F; 3336 3337 switch (BuiltinID) { 3338 default: llvm_unreachable("unexpected builtin"); 3339 case ARM::BI__builtin_arm_ldaex: 3340 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3341 break; 3342 case ARM::BI__builtin_arm_ldrexd: 3343 case ARM::BI__builtin_arm_ldrex: 3344 case ARM::BI__ldrexd: 3345 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3346 break; 3347 } 3348 3349 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3350 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3351 "ldrexd"); 3352 3353 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3354 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3355 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3356 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3357 3358 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3359 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3360 Val = Builder.CreateOr(Val, Val1); 3361 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3362 } 3363 3364 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3365 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3366 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3367 3368 QualType Ty = E->getType(); 3369 llvm::Type *RealResTy = ConvertType(Ty); 3370 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3371 getContext().getTypeSize(Ty)); 3372 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3373 3374 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3375 ? Intrinsic::arm_ldaex 3376 : Intrinsic::arm_ldrex, 3377 LoadAddr->getType()); 3378 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3379 3380 if (RealResTy->isPointerTy()) 3381 return Builder.CreateIntToPtr(Val, RealResTy); 3382 else { 3383 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3384 return Builder.CreateBitCast(Val, RealResTy); 3385 } 3386 } 3387 3388 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3389 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3390 BuiltinID == ARM::BI__builtin_arm_strex) && 3391 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3392 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3393 ? Intrinsic::arm_stlexd 3394 : Intrinsic::arm_strexd); 3395 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3396 3397 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3398 Value *Val = EmitScalarExpr(E->getArg(0)); 3399 Builder.CreateStore(Val, Tmp); 3400 3401 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3402 Val = Builder.CreateLoad(LdPtr); 3403 3404 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3405 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3406 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3407 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3408 } 3409 3410 if (BuiltinID == ARM::BI__builtin_arm_strex || 3411 BuiltinID == ARM::BI__builtin_arm_stlex) { 3412 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3413 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3414 3415 QualType Ty = E->getArg(0)->getType(); 3416 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3417 getContext().getTypeSize(Ty)); 3418 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3419 3420 if (StoreVal->getType()->isPointerTy()) 3421 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3422 else { 3423 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3424 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3425 } 3426 3427 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3428 ? Intrinsic::arm_stlex 3429 : Intrinsic::arm_strex, 3430 StoreAddr->getType()); 3431 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3432 } 3433 3434 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3435 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3436 return Builder.CreateCall(F, {}); 3437 } 3438 3439 // CRC32 3440 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3441 switch (BuiltinID) { 3442 case ARM::BI__builtin_arm_crc32b: 3443 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3444 case ARM::BI__builtin_arm_crc32cb: 3445 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3446 case ARM::BI__builtin_arm_crc32h: 3447 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3448 case ARM::BI__builtin_arm_crc32ch: 3449 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3450 case ARM::BI__builtin_arm_crc32w: 3451 case ARM::BI__builtin_arm_crc32d: 3452 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3453 case ARM::BI__builtin_arm_crc32cw: 3454 case ARM::BI__builtin_arm_crc32cd: 3455 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3456 } 3457 3458 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3459 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3460 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3461 3462 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3463 // intrinsics, hence we need different codegen for these cases. 3464 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3465 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3466 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3467 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3468 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3469 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3470 3471 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3472 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3473 return Builder.CreateCall(F, {Res, Arg1b}); 3474 } else { 3475 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3476 3477 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3478 return Builder.CreateCall(F, {Arg0, Arg1}); 3479 } 3480 } 3481 3482 SmallVector<Value*, 4> Ops; 3483 llvm::Value *Align = nullptr; 3484 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3485 if (i == 0) { 3486 switch (BuiltinID) { 3487 case NEON::BI__builtin_neon_vld1_v: 3488 case NEON::BI__builtin_neon_vld1q_v: 3489 case NEON::BI__builtin_neon_vld1q_lane_v: 3490 case NEON::BI__builtin_neon_vld1_lane_v: 3491 case NEON::BI__builtin_neon_vld1_dup_v: 3492 case NEON::BI__builtin_neon_vld1q_dup_v: 3493 case NEON::BI__builtin_neon_vst1_v: 3494 case NEON::BI__builtin_neon_vst1q_v: 3495 case NEON::BI__builtin_neon_vst1q_lane_v: 3496 case NEON::BI__builtin_neon_vst1_lane_v: 3497 case NEON::BI__builtin_neon_vst2_v: 3498 case NEON::BI__builtin_neon_vst2q_v: 3499 case NEON::BI__builtin_neon_vst2_lane_v: 3500 case NEON::BI__builtin_neon_vst2q_lane_v: 3501 case NEON::BI__builtin_neon_vst3_v: 3502 case NEON::BI__builtin_neon_vst3q_v: 3503 case NEON::BI__builtin_neon_vst3_lane_v: 3504 case NEON::BI__builtin_neon_vst3q_lane_v: 3505 case NEON::BI__builtin_neon_vst4_v: 3506 case NEON::BI__builtin_neon_vst4q_v: 3507 case NEON::BI__builtin_neon_vst4_lane_v: 3508 case NEON::BI__builtin_neon_vst4q_lane_v: 3509 // Get the alignment for the argument in addition to the value; 3510 // we'll use it later. 3511 std::pair<llvm::Value*, unsigned> Src = 3512 EmitPointerWithAlignment(E->getArg(0)); 3513 Ops.push_back(Src.first); 3514 Align = Builder.getInt32(Src.second); 3515 continue; 3516 } 3517 } 3518 if (i == 1) { 3519 switch (BuiltinID) { 3520 case NEON::BI__builtin_neon_vld2_v: 3521 case NEON::BI__builtin_neon_vld2q_v: 3522 case NEON::BI__builtin_neon_vld3_v: 3523 case NEON::BI__builtin_neon_vld3q_v: 3524 case NEON::BI__builtin_neon_vld4_v: 3525 case NEON::BI__builtin_neon_vld4q_v: 3526 case NEON::BI__builtin_neon_vld2_lane_v: 3527 case NEON::BI__builtin_neon_vld2q_lane_v: 3528 case NEON::BI__builtin_neon_vld3_lane_v: 3529 case NEON::BI__builtin_neon_vld3q_lane_v: 3530 case NEON::BI__builtin_neon_vld4_lane_v: 3531 case NEON::BI__builtin_neon_vld4q_lane_v: 3532 case NEON::BI__builtin_neon_vld2_dup_v: 3533 case NEON::BI__builtin_neon_vld3_dup_v: 3534 case NEON::BI__builtin_neon_vld4_dup_v: 3535 // Get the alignment for the argument in addition to the value; 3536 // we'll use it later. 3537 std::pair<llvm::Value*, unsigned> Src = 3538 EmitPointerWithAlignment(E->getArg(1)); 3539 Ops.push_back(Src.first); 3540 Align = Builder.getInt32(Src.second); 3541 continue; 3542 } 3543 } 3544 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3545 } 3546 3547 switch (BuiltinID) { 3548 default: break; 3549 // vget_lane and vset_lane are not overloaded and do not have an extra 3550 // argument that specifies the vector type. 3551 case NEON::BI__builtin_neon_vget_lane_i8: 3552 case NEON::BI__builtin_neon_vget_lane_i16: 3553 case NEON::BI__builtin_neon_vget_lane_i32: 3554 case NEON::BI__builtin_neon_vget_lane_i64: 3555 case NEON::BI__builtin_neon_vget_lane_f32: 3556 case NEON::BI__builtin_neon_vgetq_lane_i8: 3557 case NEON::BI__builtin_neon_vgetq_lane_i16: 3558 case NEON::BI__builtin_neon_vgetq_lane_i32: 3559 case NEON::BI__builtin_neon_vgetq_lane_i64: 3560 case NEON::BI__builtin_neon_vgetq_lane_f32: 3561 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3562 "vget_lane"); 3563 case NEON::BI__builtin_neon_vset_lane_i8: 3564 case NEON::BI__builtin_neon_vset_lane_i16: 3565 case NEON::BI__builtin_neon_vset_lane_i32: 3566 case NEON::BI__builtin_neon_vset_lane_i64: 3567 case NEON::BI__builtin_neon_vset_lane_f32: 3568 case NEON::BI__builtin_neon_vsetq_lane_i8: 3569 case NEON::BI__builtin_neon_vsetq_lane_i16: 3570 case NEON::BI__builtin_neon_vsetq_lane_i32: 3571 case NEON::BI__builtin_neon_vsetq_lane_i64: 3572 case NEON::BI__builtin_neon_vsetq_lane_f32: 3573 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3574 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3575 3576 // Non-polymorphic crypto instructions also not overloaded 3577 case NEON::BI__builtin_neon_vsha1h_u32: 3578 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3579 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3580 "vsha1h"); 3581 case NEON::BI__builtin_neon_vsha1cq_u32: 3582 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3583 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3584 "vsha1h"); 3585 case NEON::BI__builtin_neon_vsha1pq_u32: 3586 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3587 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3588 "vsha1h"); 3589 case NEON::BI__builtin_neon_vsha1mq_u32: 3590 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3591 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3592 "vsha1h"); 3593 } 3594 3595 // Get the last argument, which specifies the vector type. 3596 llvm::APSInt Result; 3597 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3598 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3599 return nullptr; 3600 3601 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3602 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3603 // Determine the overloaded type of this builtin. 3604 llvm::Type *Ty; 3605 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3606 Ty = FloatTy; 3607 else 3608 Ty = DoubleTy; 3609 3610 // Determine whether this is an unsigned conversion or not. 3611 bool usgn = Result.getZExtValue() == 1; 3612 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3613 3614 // Call the appropriate intrinsic. 3615 Function *F = CGM.getIntrinsic(Int, Ty); 3616 return Builder.CreateCall(F, Ops, "vcvtr"); 3617 } 3618 3619 // Determine the type of this overloaded NEON intrinsic. 3620 NeonTypeFlags Type(Result.getZExtValue()); 3621 bool usgn = Type.isUnsigned(); 3622 bool rightShift = false; 3623 3624 llvm::VectorType *VTy = GetNeonType(this, Type); 3625 llvm::Type *Ty = VTy; 3626 if (!Ty) 3627 return nullptr; 3628 3629 // Many NEON builtins have identical semantics and uses in ARM and 3630 // AArch64. Emit these in a single function. 3631 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3632 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3633 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3634 if (Builtin) 3635 return EmitCommonNeonBuiltinExpr( 3636 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3637 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3638 3639 unsigned Int; 3640 switch (BuiltinID) { 3641 default: return nullptr; 3642 case NEON::BI__builtin_neon_vld1q_lane_v: 3643 // Handle 64-bit integer elements as a special case. Use shuffles of 3644 // one-element vectors to avoid poor code for i64 in the backend. 3645 if (VTy->getElementType()->isIntegerTy(64)) { 3646 // Extract the other lane. 3647 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3648 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3649 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3650 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3651 // Load the value as a one-element vector. 3652 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3653 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3654 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3655 // Combine them. 3656 SmallVector<Constant*, 2> Indices; 3657 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3658 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3659 SV = llvm::ConstantVector::get(Indices); 3660 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3661 } 3662 // fall through 3663 case NEON::BI__builtin_neon_vld1_lane_v: { 3664 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3665 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3666 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3667 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3668 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3669 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3670 } 3671 case NEON::BI__builtin_neon_vld2_dup_v: 3672 case NEON::BI__builtin_neon_vld3_dup_v: 3673 case NEON::BI__builtin_neon_vld4_dup_v: { 3674 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3675 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3676 switch (BuiltinID) { 3677 case NEON::BI__builtin_neon_vld2_dup_v: 3678 Int = Intrinsic::arm_neon_vld2; 3679 break; 3680 case NEON::BI__builtin_neon_vld3_dup_v: 3681 Int = Intrinsic::arm_neon_vld3; 3682 break; 3683 case NEON::BI__builtin_neon_vld4_dup_v: 3684 Int = Intrinsic::arm_neon_vld4; 3685 break; 3686 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3687 } 3688 Function *F = CGM.getIntrinsic(Int, Ty); 3689 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3690 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3691 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3692 return Builder.CreateStore(Ops[1], Ops[0]); 3693 } 3694 switch (BuiltinID) { 3695 case NEON::BI__builtin_neon_vld2_dup_v: 3696 Int = Intrinsic::arm_neon_vld2lane; 3697 break; 3698 case NEON::BI__builtin_neon_vld3_dup_v: 3699 Int = Intrinsic::arm_neon_vld3lane; 3700 break; 3701 case NEON::BI__builtin_neon_vld4_dup_v: 3702 Int = Intrinsic::arm_neon_vld4lane; 3703 break; 3704 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3705 } 3706 Function *F = CGM.getIntrinsic(Int, Ty); 3707 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3708 3709 SmallVector<Value*, 6> Args; 3710 Args.push_back(Ops[1]); 3711 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3712 3713 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3714 Args.push_back(CI); 3715 Args.push_back(Align); 3716 3717 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3718 // splat lane 0 to all elts in each vector of the result. 3719 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3720 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3721 Value *Elt = Builder.CreateBitCast(Val, Ty); 3722 Elt = EmitNeonSplat(Elt, CI); 3723 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3724 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3725 } 3726 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3727 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3728 return Builder.CreateStore(Ops[1], Ops[0]); 3729 } 3730 case NEON::BI__builtin_neon_vqrshrn_n_v: 3731 Int = 3732 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3733 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3734 1, true); 3735 case NEON::BI__builtin_neon_vqrshrun_n_v: 3736 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3737 Ops, "vqrshrun_n", 1, true); 3738 case NEON::BI__builtin_neon_vqshrn_n_v: 3739 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3740 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3741 1, true); 3742 case NEON::BI__builtin_neon_vqshrun_n_v: 3743 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3744 Ops, "vqshrun_n", 1, true); 3745 case NEON::BI__builtin_neon_vrecpe_v: 3746 case NEON::BI__builtin_neon_vrecpeq_v: 3747 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3748 Ops, "vrecpe"); 3749 case NEON::BI__builtin_neon_vrshrn_n_v: 3750 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3751 Ops, "vrshrn_n", 1, true); 3752 case NEON::BI__builtin_neon_vrsra_n_v: 3753 case NEON::BI__builtin_neon_vrsraq_n_v: 3754 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3755 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3756 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3757 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3758 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 3759 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3760 case NEON::BI__builtin_neon_vsri_n_v: 3761 case NEON::BI__builtin_neon_vsriq_n_v: 3762 rightShift = true; 3763 case NEON::BI__builtin_neon_vsli_n_v: 3764 case NEON::BI__builtin_neon_vsliq_n_v: 3765 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3766 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3767 Ops, "vsli_n"); 3768 case NEON::BI__builtin_neon_vsra_n_v: 3769 case NEON::BI__builtin_neon_vsraq_n_v: 3770 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3771 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3772 return Builder.CreateAdd(Ops[0], Ops[1]); 3773 case NEON::BI__builtin_neon_vst1q_lane_v: 3774 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3775 // a one-element vector and avoid poor code for i64 in the backend. 3776 if (VTy->getElementType()->isIntegerTy(64)) { 3777 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3778 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3779 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3780 Ops[2] = Align; 3781 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3782 Ops[1]->getType()), Ops); 3783 } 3784 // fall through 3785 case NEON::BI__builtin_neon_vst1_lane_v: { 3786 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3787 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3788 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3789 StoreInst *St = Builder.CreateStore(Ops[1], 3790 Builder.CreateBitCast(Ops[0], Ty)); 3791 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3792 return St; 3793 } 3794 case NEON::BI__builtin_neon_vtbl1_v: 3795 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3796 Ops, "vtbl1"); 3797 case NEON::BI__builtin_neon_vtbl2_v: 3798 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3799 Ops, "vtbl2"); 3800 case NEON::BI__builtin_neon_vtbl3_v: 3801 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3802 Ops, "vtbl3"); 3803 case NEON::BI__builtin_neon_vtbl4_v: 3804 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3805 Ops, "vtbl4"); 3806 case NEON::BI__builtin_neon_vtbx1_v: 3807 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3808 Ops, "vtbx1"); 3809 case NEON::BI__builtin_neon_vtbx2_v: 3810 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3811 Ops, "vtbx2"); 3812 case NEON::BI__builtin_neon_vtbx3_v: 3813 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3814 Ops, "vtbx3"); 3815 case NEON::BI__builtin_neon_vtbx4_v: 3816 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3817 Ops, "vtbx4"); 3818 } 3819 } 3820 3821 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3822 const CallExpr *E, 3823 SmallVectorImpl<Value *> &Ops) { 3824 unsigned int Int = 0; 3825 const char *s = nullptr; 3826 3827 switch (BuiltinID) { 3828 default: 3829 return nullptr; 3830 case NEON::BI__builtin_neon_vtbl1_v: 3831 case NEON::BI__builtin_neon_vqtbl1_v: 3832 case NEON::BI__builtin_neon_vqtbl1q_v: 3833 case NEON::BI__builtin_neon_vtbl2_v: 3834 case NEON::BI__builtin_neon_vqtbl2_v: 3835 case NEON::BI__builtin_neon_vqtbl2q_v: 3836 case NEON::BI__builtin_neon_vtbl3_v: 3837 case NEON::BI__builtin_neon_vqtbl3_v: 3838 case NEON::BI__builtin_neon_vqtbl3q_v: 3839 case NEON::BI__builtin_neon_vtbl4_v: 3840 case NEON::BI__builtin_neon_vqtbl4_v: 3841 case NEON::BI__builtin_neon_vqtbl4q_v: 3842 break; 3843 case NEON::BI__builtin_neon_vtbx1_v: 3844 case NEON::BI__builtin_neon_vqtbx1_v: 3845 case NEON::BI__builtin_neon_vqtbx1q_v: 3846 case NEON::BI__builtin_neon_vtbx2_v: 3847 case NEON::BI__builtin_neon_vqtbx2_v: 3848 case NEON::BI__builtin_neon_vqtbx2q_v: 3849 case NEON::BI__builtin_neon_vtbx3_v: 3850 case NEON::BI__builtin_neon_vqtbx3_v: 3851 case NEON::BI__builtin_neon_vqtbx3q_v: 3852 case NEON::BI__builtin_neon_vtbx4_v: 3853 case NEON::BI__builtin_neon_vqtbx4_v: 3854 case NEON::BI__builtin_neon_vqtbx4q_v: 3855 break; 3856 } 3857 3858 assert(E->getNumArgs() >= 3); 3859 3860 // Get the last argument, which specifies the vector type. 3861 llvm::APSInt Result; 3862 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3863 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3864 return nullptr; 3865 3866 // Determine the type of this overloaded NEON intrinsic. 3867 NeonTypeFlags Type(Result.getZExtValue()); 3868 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3869 llvm::Type *Ty = VTy; 3870 if (!Ty) 3871 return nullptr; 3872 3873 unsigned nElts = VTy->getNumElements(); 3874 3875 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3876 3877 // AArch64 scalar builtins are not overloaded, they do not have an extra 3878 // argument that specifies the vector type, need to handle each case. 3879 SmallVector<Value *, 2> TblOps; 3880 switch (BuiltinID) { 3881 case NEON::BI__builtin_neon_vtbl1_v: { 3882 TblOps.push_back(Ops[0]); 3883 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3884 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3885 } 3886 case NEON::BI__builtin_neon_vtbl2_v: { 3887 TblOps.push_back(Ops[0]); 3888 TblOps.push_back(Ops[1]); 3889 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3890 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3891 } 3892 case NEON::BI__builtin_neon_vtbl3_v: { 3893 TblOps.push_back(Ops[0]); 3894 TblOps.push_back(Ops[1]); 3895 TblOps.push_back(Ops[2]); 3896 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3897 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3898 } 3899 case NEON::BI__builtin_neon_vtbl4_v: { 3900 TblOps.push_back(Ops[0]); 3901 TblOps.push_back(Ops[1]); 3902 TblOps.push_back(Ops[2]); 3903 TblOps.push_back(Ops[3]); 3904 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3905 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3906 } 3907 case NEON::BI__builtin_neon_vtbx1_v: { 3908 TblOps.push_back(Ops[1]); 3909 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3910 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3911 3912 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3913 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3914 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3915 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3916 3917 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3918 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3919 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3920 } 3921 case NEON::BI__builtin_neon_vtbx2_v: { 3922 TblOps.push_back(Ops[1]); 3923 TblOps.push_back(Ops[2]); 3924 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3925 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3926 } 3927 case NEON::BI__builtin_neon_vtbx3_v: { 3928 TblOps.push_back(Ops[1]); 3929 TblOps.push_back(Ops[2]); 3930 TblOps.push_back(Ops[3]); 3931 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3932 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3933 3934 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3935 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3936 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3937 TwentyFourV); 3938 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3939 3940 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3941 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3942 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3943 } 3944 case NEON::BI__builtin_neon_vtbx4_v: { 3945 TblOps.push_back(Ops[1]); 3946 TblOps.push_back(Ops[2]); 3947 TblOps.push_back(Ops[3]); 3948 TblOps.push_back(Ops[4]); 3949 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3950 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3951 } 3952 case NEON::BI__builtin_neon_vqtbl1_v: 3953 case NEON::BI__builtin_neon_vqtbl1q_v: 3954 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3955 case NEON::BI__builtin_neon_vqtbl2_v: 3956 case NEON::BI__builtin_neon_vqtbl2q_v: { 3957 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3958 case NEON::BI__builtin_neon_vqtbl3_v: 3959 case NEON::BI__builtin_neon_vqtbl3q_v: 3960 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3961 case NEON::BI__builtin_neon_vqtbl4_v: 3962 case NEON::BI__builtin_neon_vqtbl4q_v: 3963 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3964 case NEON::BI__builtin_neon_vqtbx1_v: 3965 case NEON::BI__builtin_neon_vqtbx1q_v: 3966 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3967 case NEON::BI__builtin_neon_vqtbx2_v: 3968 case NEON::BI__builtin_neon_vqtbx2q_v: 3969 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3970 case NEON::BI__builtin_neon_vqtbx3_v: 3971 case NEON::BI__builtin_neon_vqtbx3q_v: 3972 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3973 case NEON::BI__builtin_neon_vqtbx4_v: 3974 case NEON::BI__builtin_neon_vqtbx4q_v: 3975 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3976 } 3977 } 3978 3979 if (!Int) 3980 return nullptr; 3981 3982 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3983 return CGF.EmitNeonCall(F, Ops, s); 3984 } 3985 3986 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3987 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3988 Op = Builder.CreateBitCast(Op, Int16Ty); 3989 Value *V = UndefValue::get(VTy); 3990 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3991 Op = Builder.CreateInsertElement(V, Op, CI); 3992 return Op; 3993 } 3994 3995 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3996 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3997 Op = Builder.CreateBitCast(Op, Int8Ty); 3998 Value *V = UndefValue::get(VTy); 3999 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4000 Op = Builder.CreateInsertElement(V, Op, CI); 4001 return Op; 4002 } 4003 4004 Value *CodeGenFunction:: 4005 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 4006 const char *Name) { 4007 // i8 is not a legal types for AArch64, so we can't just use 4008 // a normal overloaded intrinsic call for these scalar types. Instead 4009 // we'll build 64-bit vectors w/ lane zero being our input values and 4010 // perform the operation on that. The back end can pattern match directly 4011 // to the scalar instruction. 4012 Ops[0] = vectorWrapScalar8(Ops[0]); 4013 Ops[1] = vectorWrapScalar8(Ops[1]); 4014 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4015 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 4016 Constant *CI = ConstantInt::get(SizeTy, 0); 4017 return Builder.CreateExtractElement(V, CI, "lane0"); 4018 } 4019 4020 Value *CodeGenFunction:: 4021 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 4022 const char *Name) { 4023 // i16 is not a legal types for AArch64, so we can't just use 4024 // a normal overloaded intrinsic call for these scalar types. Instead 4025 // we'll build 64-bit vectors w/ lane zero being our input values and 4026 // perform the operation on that. The back end can pattern match directly 4027 // to the scalar instruction. 4028 Ops[0] = vectorWrapScalar16(Ops[0]); 4029 Ops[1] = vectorWrapScalar16(Ops[1]); 4030 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4031 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 4032 Constant *CI = ConstantInt::get(SizeTy, 0); 4033 return Builder.CreateExtractElement(V, CI, "lane0"); 4034 } 4035 4036 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4037 const CallExpr *E) { 4038 unsigned HintID = static_cast<unsigned>(-1); 4039 switch (BuiltinID) { 4040 default: break; 4041 case AArch64::BI__builtin_arm_nop: 4042 HintID = 0; 4043 break; 4044 case AArch64::BI__builtin_arm_yield: 4045 HintID = 1; 4046 break; 4047 case AArch64::BI__builtin_arm_wfe: 4048 HintID = 2; 4049 break; 4050 case AArch64::BI__builtin_arm_wfi: 4051 HintID = 3; 4052 break; 4053 case AArch64::BI__builtin_arm_sev: 4054 HintID = 4; 4055 break; 4056 case AArch64::BI__builtin_arm_sevl: 4057 HintID = 5; 4058 break; 4059 } 4060 4061 if (HintID != static_cast<unsigned>(-1)) { 4062 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4063 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4064 } 4065 4066 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4067 Value *Address = EmitScalarExpr(E->getArg(0)); 4068 Value *RW = EmitScalarExpr(E->getArg(1)); 4069 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4070 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4071 Value *IsData = EmitScalarExpr(E->getArg(4)); 4072 4073 Value *Locality = nullptr; 4074 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4075 // Temporal fetch, needs to convert cache level to locality. 4076 Locality = llvm::ConstantInt::get(Int32Ty, 4077 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4078 } else { 4079 // Streaming fetch. 4080 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4081 } 4082 4083 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4084 // PLDL3STRM or PLDL2STRM. 4085 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4086 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4087 } 4088 4089 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4090 assert((getContext().getTypeSize(E->getType()) == 32) && 4091 "rbit of unusual size!"); 4092 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4093 return Builder.CreateCall( 4094 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4095 } 4096 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4097 assert((getContext().getTypeSize(E->getType()) == 64) && 4098 "rbit of unusual size!"); 4099 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4100 return Builder.CreateCall( 4101 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4102 } 4103 4104 if (BuiltinID == AArch64::BI__clear_cache) { 4105 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4106 const FunctionDecl *FD = E->getDirectCallee(); 4107 SmallVector<Value*, 2> Ops; 4108 for (unsigned i = 0; i < 2; i++) 4109 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4110 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4111 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4112 StringRef Name = FD->getName(); 4113 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4114 } 4115 4116 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4117 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4118 getContext().getTypeSize(E->getType()) == 128) { 4119 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4120 ? Intrinsic::aarch64_ldaxp 4121 : Intrinsic::aarch64_ldxp); 4122 4123 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4124 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4125 "ldxp"); 4126 4127 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4128 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4129 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4130 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4131 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4132 4133 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4134 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4135 Val = Builder.CreateOr(Val, Val1); 4136 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4137 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4138 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4139 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4140 4141 QualType Ty = E->getType(); 4142 llvm::Type *RealResTy = ConvertType(Ty); 4143 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4144 getContext().getTypeSize(Ty)); 4145 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4146 4147 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4148 ? Intrinsic::aarch64_ldaxr 4149 : Intrinsic::aarch64_ldxr, 4150 LoadAddr->getType()); 4151 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4152 4153 if (RealResTy->isPointerTy()) 4154 return Builder.CreateIntToPtr(Val, RealResTy); 4155 4156 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4157 return Builder.CreateBitCast(Val, RealResTy); 4158 } 4159 4160 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4161 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4162 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4163 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4164 ? Intrinsic::aarch64_stlxp 4165 : Intrinsic::aarch64_stxp); 4166 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4167 4168 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4169 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4170 One); 4171 Value *Val = EmitScalarExpr(E->getArg(0)); 4172 Builder.CreateStore(Val, Tmp); 4173 4174 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4175 Val = Builder.CreateLoad(LdPtr); 4176 4177 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4178 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4179 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4180 Int8PtrTy); 4181 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4182 } 4183 4184 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4185 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4186 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4187 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4188 4189 QualType Ty = E->getArg(0)->getType(); 4190 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4191 getContext().getTypeSize(Ty)); 4192 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4193 4194 if (StoreVal->getType()->isPointerTy()) 4195 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4196 else { 4197 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4198 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4199 } 4200 4201 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4202 ? Intrinsic::aarch64_stlxr 4203 : Intrinsic::aarch64_stxr, 4204 StoreAddr->getType()); 4205 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4206 } 4207 4208 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4209 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4210 return Builder.CreateCall(F, {}); 4211 } 4212 4213 // CRC32 4214 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4215 switch (BuiltinID) { 4216 case AArch64::BI__builtin_arm_crc32b: 4217 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4218 case AArch64::BI__builtin_arm_crc32cb: 4219 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4220 case AArch64::BI__builtin_arm_crc32h: 4221 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4222 case AArch64::BI__builtin_arm_crc32ch: 4223 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4224 case AArch64::BI__builtin_arm_crc32w: 4225 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4226 case AArch64::BI__builtin_arm_crc32cw: 4227 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4228 case AArch64::BI__builtin_arm_crc32d: 4229 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4230 case AArch64::BI__builtin_arm_crc32cd: 4231 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4232 } 4233 4234 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4235 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4236 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4237 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4238 4239 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4240 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4241 4242 return Builder.CreateCall(F, {Arg0, Arg1}); 4243 } 4244 4245 llvm::SmallVector<Value*, 4> Ops; 4246 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4247 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4248 4249 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4250 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4251 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4252 4253 if (Builtin) { 4254 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4255 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4256 assert(Result && "SISD intrinsic should have been handled"); 4257 return Result; 4258 } 4259 4260 llvm::APSInt Result; 4261 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4262 NeonTypeFlags Type(0); 4263 if (Arg->isIntegerConstantExpr(Result, getContext())) 4264 // Determine the type of this overloaded NEON intrinsic. 4265 Type = NeonTypeFlags(Result.getZExtValue()); 4266 4267 bool usgn = Type.isUnsigned(); 4268 bool quad = Type.isQuad(); 4269 4270 // Handle non-overloaded intrinsics first. 4271 switch (BuiltinID) { 4272 default: break; 4273 case NEON::BI__builtin_neon_vldrq_p128: { 4274 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4275 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4276 return Builder.CreateLoad(Ptr); 4277 } 4278 case NEON::BI__builtin_neon_vstrq_p128: { 4279 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4280 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4281 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4282 } 4283 case NEON::BI__builtin_neon_vcvts_u32_f32: 4284 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4285 usgn = true; 4286 // FALL THROUGH 4287 case NEON::BI__builtin_neon_vcvts_s32_f32: 4288 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4289 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4290 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4291 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4292 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4293 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4294 if (usgn) 4295 return Builder.CreateFPToUI(Ops[0], InTy); 4296 return Builder.CreateFPToSI(Ops[0], InTy); 4297 } 4298 case NEON::BI__builtin_neon_vcvts_f32_u32: 4299 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4300 usgn = true; 4301 // FALL THROUGH 4302 case NEON::BI__builtin_neon_vcvts_f32_s32: 4303 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 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], InTy); 4309 if (usgn) 4310 return Builder.CreateUIToFP(Ops[0], FTy); 4311 return Builder.CreateSIToFP(Ops[0], FTy); 4312 } 4313 case NEON::BI__builtin_neon_vpaddd_s64: { 4314 llvm::Type *Ty = 4315 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4316 Value *Vec = EmitScalarExpr(E->getArg(0)); 4317 // The vector is v2f64, so make sure it's bitcast to that. 4318 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4319 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4320 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4321 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4322 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4323 // Pairwise addition of a v2f64 into a scalar f64. 4324 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4325 } 4326 case NEON::BI__builtin_neon_vpaddd_f64: { 4327 llvm::Type *Ty = 4328 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4329 Value *Vec = EmitScalarExpr(E->getArg(0)); 4330 // The vector is v2f64, so make sure it's bitcast to that. 4331 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4332 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4333 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4334 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4335 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4336 // Pairwise addition of a v2f64 into a scalar f64. 4337 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4338 } 4339 case NEON::BI__builtin_neon_vpadds_f32: { 4340 llvm::Type *Ty = 4341 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4342 Value *Vec = EmitScalarExpr(E->getArg(0)); 4343 // The vector is v2f32, so make sure it's bitcast to that. 4344 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4345 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4346 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4347 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4348 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4349 // Pairwise addition of a v2f32 into a scalar f32. 4350 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4351 } 4352 case NEON::BI__builtin_neon_vceqzd_s64: 4353 case NEON::BI__builtin_neon_vceqzd_f64: 4354 case NEON::BI__builtin_neon_vceqzs_f32: 4355 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4356 return EmitAArch64CompareBuiltinExpr( 4357 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4358 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4359 case NEON::BI__builtin_neon_vcgezd_s64: 4360 case NEON::BI__builtin_neon_vcgezd_f64: 4361 case NEON::BI__builtin_neon_vcgezs_f32: 4362 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4363 return EmitAArch64CompareBuiltinExpr( 4364 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4365 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4366 case NEON::BI__builtin_neon_vclezd_s64: 4367 case NEON::BI__builtin_neon_vclezd_f64: 4368 case NEON::BI__builtin_neon_vclezs_f32: 4369 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4370 return EmitAArch64CompareBuiltinExpr( 4371 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4372 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4373 case NEON::BI__builtin_neon_vcgtzd_s64: 4374 case NEON::BI__builtin_neon_vcgtzd_f64: 4375 case NEON::BI__builtin_neon_vcgtzs_f32: 4376 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4377 return EmitAArch64CompareBuiltinExpr( 4378 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4379 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4380 case NEON::BI__builtin_neon_vcltzd_s64: 4381 case NEON::BI__builtin_neon_vcltzd_f64: 4382 case NEON::BI__builtin_neon_vcltzs_f32: 4383 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4384 return EmitAArch64CompareBuiltinExpr( 4385 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4386 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4387 4388 case NEON::BI__builtin_neon_vceqzd_u64: { 4389 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4390 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4391 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4392 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4393 llvm::Constant::getNullValue(Ty)); 4394 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4395 } 4396 case NEON::BI__builtin_neon_vceqd_f64: 4397 case NEON::BI__builtin_neon_vcled_f64: 4398 case NEON::BI__builtin_neon_vcltd_f64: 4399 case NEON::BI__builtin_neon_vcged_f64: 4400 case NEON::BI__builtin_neon_vcgtd_f64: { 4401 llvm::CmpInst::Predicate P; 4402 switch (BuiltinID) { 4403 default: llvm_unreachable("missing builtin ID in switch!"); 4404 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4405 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4406 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4407 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4408 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4409 } 4410 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4411 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4412 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4413 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4414 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4415 } 4416 case NEON::BI__builtin_neon_vceqs_f32: 4417 case NEON::BI__builtin_neon_vcles_f32: 4418 case NEON::BI__builtin_neon_vclts_f32: 4419 case NEON::BI__builtin_neon_vcges_f32: 4420 case NEON::BI__builtin_neon_vcgts_f32: { 4421 llvm::CmpInst::Predicate P; 4422 switch (BuiltinID) { 4423 default: llvm_unreachable("missing builtin ID in switch!"); 4424 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4425 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4426 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4427 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4428 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4429 } 4430 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4431 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4432 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4433 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4434 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4435 } 4436 case NEON::BI__builtin_neon_vceqd_s64: 4437 case NEON::BI__builtin_neon_vceqd_u64: 4438 case NEON::BI__builtin_neon_vcgtd_s64: 4439 case NEON::BI__builtin_neon_vcgtd_u64: 4440 case NEON::BI__builtin_neon_vcltd_s64: 4441 case NEON::BI__builtin_neon_vcltd_u64: 4442 case NEON::BI__builtin_neon_vcged_u64: 4443 case NEON::BI__builtin_neon_vcged_s64: 4444 case NEON::BI__builtin_neon_vcled_u64: 4445 case NEON::BI__builtin_neon_vcled_s64: { 4446 llvm::CmpInst::Predicate P; 4447 switch (BuiltinID) { 4448 default: llvm_unreachable("missing builtin ID in switch!"); 4449 case NEON::BI__builtin_neon_vceqd_s64: 4450 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4451 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4452 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4453 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4454 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4455 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4456 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4457 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4458 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4459 } 4460 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4461 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4462 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4463 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4464 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4465 } 4466 case NEON::BI__builtin_neon_vtstd_s64: 4467 case NEON::BI__builtin_neon_vtstd_u64: { 4468 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4469 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4470 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4471 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4472 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4473 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4474 llvm::Constant::getNullValue(Ty)); 4475 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4476 } 4477 case NEON::BI__builtin_neon_vset_lane_i8: 4478 case NEON::BI__builtin_neon_vset_lane_i16: 4479 case NEON::BI__builtin_neon_vset_lane_i32: 4480 case NEON::BI__builtin_neon_vset_lane_i64: 4481 case NEON::BI__builtin_neon_vset_lane_f32: 4482 case NEON::BI__builtin_neon_vsetq_lane_i8: 4483 case NEON::BI__builtin_neon_vsetq_lane_i16: 4484 case NEON::BI__builtin_neon_vsetq_lane_i32: 4485 case NEON::BI__builtin_neon_vsetq_lane_i64: 4486 case NEON::BI__builtin_neon_vsetq_lane_f32: 4487 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4488 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4489 case NEON::BI__builtin_neon_vset_lane_f64: 4490 // The vector type needs a cast for the v1f64 variant. 4491 Ops[1] = Builder.CreateBitCast(Ops[1], 4492 llvm::VectorType::get(DoubleTy, 1)); 4493 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4494 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4495 case NEON::BI__builtin_neon_vsetq_lane_f64: 4496 // The vector type needs a cast for the v2f64 variant. 4497 Ops[1] = Builder.CreateBitCast(Ops[1], 4498 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4499 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4500 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4501 4502 case NEON::BI__builtin_neon_vget_lane_i8: 4503 case NEON::BI__builtin_neon_vdupb_lane_i8: 4504 Ops[0] = Builder.CreateBitCast(Ops[0], 4505 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4506 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4507 "vget_lane"); 4508 case NEON::BI__builtin_neon_vgetq_lane_i8: 4509 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4510 Ops[0] = Builder.CreateBitCast(Ops[0], 4511 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4512 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4513 "vgetq_lane"); 4514 case NEON::BI__builtin_neon_vget_lane_i16: 4515 case NEON::BI__builtin_neon_vduph_lane_i16: 4516 Ops[0] = Builder.CreateBitCast(Ops[0], 4517 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4518 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4519 "vget_lane"); 4520 case NEON::BI__builtin_neon_vgetq_lane_i16: 4521 case NEON::BI__builtin_neon_vduph_laneq_i16: 4522 Ops[0] = Builder.CreateBitCast(Ops[0], 4523 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4524 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4525 "vgetq_lane"); 4526 case NEON::BI__builtin_neon_vget_lane_i32: 4527 case NEON::BI__builtin_neon_vdups_lane_i32: 4528 Ops[0] = Builder.CreateBitCast( 4529 Ops[0], 4530 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4531 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4532 "vget_lane"); 4533 case NEON::BI__builtin_neon_vdups_lane_f32: 4534 Ops[0] = Builder.CreateBitCast(Ops[0], 4535 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4536 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4537 "vdups_lane"); 4538 case NEON::BI__builtin_neon_vgetq_lane_i32: 4539 case NEON::BI__builtin_neon_vdups_laneq_i32: 4540 Ops[0] = Builder.CreateBitCast(Ops[0], 4541 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4542 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4543 "vgetq_lane"); 4544 case NEON::BI__builtin_neon_vget_lane_i64: 4545 case NEON::BI__builtin_neon_vdupd_lane_i64: 4546 Ops[0] = Builder.CreateBitCast(Ops[0], 4547 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4548 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4549 "vget_lane"); 4550 case NEON::BI__builtin_neon_vdupd_lane_f64: 4551 Ops[0] = Builder.CreateBitCast(Ops[0], 4552 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4553 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4554 "vdupd_lane"); 4555 case NEON::BI__builtin_neon_vgetq_lane_i64: 4556 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4557 Ops[0] = Builder.CreateBitCast(Ops[0], 4558 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4559 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4560 "vgetq_lane"); 4561 case NEON::BI__builtin_neon_vget_lane_f32: 4562 Ops[0] = Builder.CreateBitCast(Ops[0], 4563 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4564 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4565 "vget_lane"); 4566 case NEON::BI__builtin_neon_vget_lane_f64: 4567 Ops[0] = Builder.CreateBitCast(Ops[0], 4568 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4569 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4570 "vget_lane"); 4571 case NEON::BI__builtin_neon_vgetq_lane_f32: 4572 case NEON::BI__builtin_neon_vdups_laneq_f32: 4573 Ops[0] = Builder.CreateBitCast(Ops[0], 4574 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4575 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4576 "vgetq_lane"); 4577 case NEON::BI__builtin_neon_vgetq_lane_f64: 4578 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4579 Ops[0] = Builder.CreateBitCast(Ops[0], 4580 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4581 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4582 "vgetq_lane"); 4583 case NEON::BI__builtin_neon_vaddd_s64: 4584 case NEON::BI__builtin_neon_vaddd_u64: 4585 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4586 case NEON::BI__builtin_neon_vsubd_s64: 4587 case NEON::BI__builtin_neon_vsubd_u64: 4588 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4589 case NEON::BI__builtin_neon_vqdmlalh_s16: 4590 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4591 SmallVector<Value *, 2> ProductOps; 4592 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4593 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4594 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4595 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4596 ProductOps, "vqdmlXl"); 4597 Constant *CI = ConstantInt::get(SizeTy, 0); 4598 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4599 4600 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4601 ? Intrinsic::aarch64_neon_sqadd 4602 : Intrinsic::aarch64_neon_sqsub; 4603 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4604 } 4605 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4606 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4607 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4608 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4609 Ops, "vqshlu_n"); 4610 } 4611 case NEON::BI__builtin_neon_vqshld_n_u64: 4612 case NEON::BI__builtin_neon_vqshld_n_s64: { 4613 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4614 ? Intrinsic::aarch64_neon_uqshl 4615 : Intrinsic::aarch64_neon_sqshl; 4616 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4617 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4618 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4619 } 4620 case NEON::BI__builtin_neon_vrshrd_n_u64: 4621 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4622 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4623 ? Intrinsic::aarch64_neon_urshl 4624 : Intrinsic::aarch64_neon_srshl; 4625 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4626 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4627 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4628 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4629 } 4630 case NEON::BI__builtin_neon_vrsrad_n_u64: 4631 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4632 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4633 ? Intrinsic::aarch64_neon_urshl 4634 : Intrinsic::aarch64_neon_srshl; 4635 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4636 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4637 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4638 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4639 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4640 } 4641 case NEON::BI__builtin_neon_vshld_n_s64: 4642 case NEON::BI__builtin_neon_vshld_n_u64: { 4643 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4644 return Builder.CreateShl( 4645 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4646 } 4647 case NEON::BI__builtin_neon_vshrd_n_s64: { 4648 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4649 return Builder.CreateAShr( 4650 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4651 Amt->getZExtValue())), 4652 "shrd_n"); 4653 } 4654 case NEON::BI__builtin_neon_vshrd_n_u64: { 4655 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4656 uint64_t ShiftAmt = Amt->getZExtValue(); 4657 // Right-shifting an unsigned value by its size yields 0. 4658 if (ShiftAmt == 64) 4659 return ConstantInt::get(Int64Ty, 0); 4660 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4661 "shrd_n"); 4662 } 4663 case NEON::BI__builtin_neon_vsrad_n_s64: { 4664 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4665 Ops[1] = Builder.CreateAShr( 4666 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4667 Amt->getZExtValue())), 4668 "shrd_n"); 4669 return Builder.CreateAdd(Ops[0], Ops[1]); 4670 } 4671 case NEON::BI__builtin_neon_vsrad_n_u64: { 4672 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4673 uint64_t ShiftAmt = Amt->getZExtValue(); 4674 // Right-shifting an unsigned value by its size yields 0. 4675 // As Op + 0 = Op, return Ops[0] directly. 4676 if (ShiftAmt == 64) 4677 return Ops[0]; 4678 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4679 "shrd_n"); 4680 return Builder.CreateAdd(Ops[0], Ops[1]); 4681 } 4682 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4683 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4684 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4685 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4686 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4687 "lane"); 4688 SmallVector<Value *, 2> ProductOps; 4689 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4690 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4691 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4692 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4693 ProductOps, "vqdmlXl"); 4694 Constant *CI = ConstantInt::get(SizeTy, 0); 4695 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4696 Ops.pop_back(); 4697 4698 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4699 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4700 ? Intrinsic::aarch64_neon_sqadd 4701 : Intrinsic::aarch64_neon_sqsub; 4702 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4703 } 4704 case NEON::BI__builtin_neon_vqdmlals_s32: 4705 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4706 SmallVector<Value *, 2> ProductOps; 4707 ProductOps.push_back(Ops[1]); 4708 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4709 Ops[1] = 4710 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4711 ProductOps, "vqdmlXl"); 4712 4713 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4714 ? Intrinsic::aarch64_neon_sqadd 4715 : Intrinsic::aarch64_neon_sqsub; 4716 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4717 } 4718 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4719 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4720 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4721 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4722 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4723 "lane"); 4724 SmallVector<Value *, 2> ProductOps; 4725 ProductOps.push_back(Ops[1]); 4726 ProductOps.push_back(Ops[2]); 4727 Ops[1] = 4728 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4729 ProductOps, "vqdmlXl"); 4730 Ops.pop_back(); 4731 4732 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4733 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4734 ? Intrinsic::aarch64_neon_sqadd 4735 : Intrinsic::aarch64_neon_sqsub; 4736 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4737 } 4738 } 4739 4740 llvm::VectorType *VTy = GetNeonType(this, Type); 4741 llvm::Type *Ty = VTy; 4742 if (!Ty) 4743 return nullptr; 4744 4745 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4746 // defer to common code if it's been added to our special map. 4747 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4748 AArch64SIMDIntrinsicsProvenSorted); 4749 4750 if (Builtin) 4751 return EmitCommonNeonBuiltinExpr( 4752 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4753 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4754 4755 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4756 return V; 4757 4758 unsigned Int; 4759 switch (BuiltinID) { 4760 default: return nullptr; 4761 case NEON::BI__builtin_neon_vbsl_v: 4762 case NEON::BI__builtin_neon_vbslq_v: { 4763 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4764 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4765 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4766 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4767 4768 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4769 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4770 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4771 return Builder.CreateBitCast(Ops[0], Ty); 4772 } 4773 case NEON::BI__builtin_neon_vfma_lane_v: 4774 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4775 // The ARM builtins (and instructions) have the addend as the first 4776 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4777 Value *Addend = Ops[0]; 4778 Value *Multiplicand = Ops[1]; 4779 Value *LaneSource = Ops[2]; 4780 Ops[0] = Multiplicand; 4781 Ops[1] = LaneSource; 4782 Ops[2] = Addend; 4783 4784 // Now adjust things to handle the lane access. 4785 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4786 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4787 VTy; 4788 llvm::Constant *cst = cast<Constant>(Ops[3]); 4789 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4790 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4791 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4792 4793 Ops.pop_back(); 4794 Int = Intrinsic::fma; 4795 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4796 } 4797 case NEON::BI__builtin_neon_vfma_laneq_v: { 4798 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4799 // v1f64 fma should be mapped to Neon scalar f64 fma 4800 if (VTy && VTy->getElementType() == DoubleTy) { 4801 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4802 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4803 llvm::Type *VTy = GetNeonType(this, 4804 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4805 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4806 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4807 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4808 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4809 return Builder.CreateBitCast(Result, Ty); 4810 } 4811 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4812 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4813 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4814 4815 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4816 VTy->getNumElements() * 2); 4817 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4818 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4819 cast<ConstantInt>(Ops[3])); 4820 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4821 4822 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4823 } 4824 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4825 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4826 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4827 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4828 4829 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4830 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4831 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4832 } 4833 case NEON::BI__builtin_neon_vfmas_lane_f32: 4834 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4835 case NEON::BI__builtin_neon_vfmad_lane_f64: 4836 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4837 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4838 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4839 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4840 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4841 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4842 } 4843 case NEON::BI__builtin_neon_vfms_v: 4844 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4845 // FIXME: probably remove when we no longer support aarch64_simd.h 4846 // (arm_neon.h delegates to vfma). 4847 4848 // The ARM builtins (and instructions) have the addend as the first 4849 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4850 Value *Subtrahend = Ops[0]; 4851 Value *Multiplicand = Ops[2]; 4852 Ops[0] = Multiplicand; 4853 Ops[2] = Subtrahend; 4854 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4855 Ops[1] = Builder.CreateFNeg(Ops[1]); 4856 Int = Intrinsic::fma; 4857 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4858 } 4859 case NEON::BI__builtin_neon_vmull_v: 4860 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4861 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4862 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4863 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4864 case NEON::BI__builtin_neon_vmax_v: 4865 case NEON::BI__builtin_neon_vmaxq_v: 4866 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4867 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4868 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4869 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4870 case NEON::BI__builtin_neon_vmin_v: 4871 case NEON::BI__builtin_neon_vminq_v: 4872 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4873 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4874 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4875 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4876 case NEON::BI__builtin_neon_vabd_v: 4877 case NEON::BI__builtin_neon_vabdq_v: 4878 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4879 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4880 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4881 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4882 case NEON::BI__builtin_neon_vpadal_v: 4883 case NEON::BI__builtin_neon_vpadalq_v: { 4884 unsigned ArgElts = VTy->getNumElements(); 4885 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4886 unsigned BitWidth = EltTy->getBitWidth(); 4887 llvm::Type *ArgTy = llvm::VectorType::get( 4888 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4889 llvm::Type* Tys[2] = { VTy, ArgTy }; 4890 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4891 SmallVector<llvm::Value*, 1> TmpOps; 4892 TmpOps.push_back(Ops[1]); 4893 Function *F = CGM.getIntrinsic(Int, Tys); 4894 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4895 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4896 return Builder.CreateAdd(tmp, addend); 4897 } 4898 case NEON::BI__builtin_neon_vpmin_v: 4899 case NEON::BI__builtin_neon_vpminq_v: 4900 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4901 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4902 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4903 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4904 case NEON::BI__builtin_neon_vpmax_v: 4905 case NEON::BI__builtin_neon_vpmaxq_v: 4906 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4907 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4908 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4909 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4910 case NEON::BI__builtin_neon_vminnm_v: 4911 case NEON::BI__builtin_neon_vminnmq_v: 4912 Int = Intrinsic::aarch64_neon_fminnm; 4913 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4914 case NEON::BI__builtin_neon_vmaxnm_v: 4915 case NEON::BI__builtin_neon_vmaxnmq_v: 4916 Int = Intrinsic::aarch64_neon_fmaxnm; 4917 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4918 case NEON::BI__builtin_neon_vrecpss_f32: { 4919 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4920 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4921 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4922 Ops, "vrecps"); 4923 } 4924 case NEON::BI__builtin_neon_vrecpsd_f64: { 4925 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4926 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4927 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4928 Ops, "vrecps"); 4929 } 4930 case NEON::BI__builtin_neon_vqshrun_n_v: 4931 Int = Intrinsic::aarch64_neon_sqshrun; 4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4933 case NEON::BI__builtin_neon_vqrshrun_n_v: 4934 Int = Intrinsic::aarch64_neon_sqrshrun; 4935 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4936 case NEON::BI__builtin_neon_vqshrn_n_v: 4937 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4938 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4939 case NEON::BI__builtin_neon_vrshrn_n_v: 4940 Int = Intrinsic::aarch64_neon_rshrn; 4941 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4942 case NEON::BI__builtin_neon_vqrshrn_n_v: 4943 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4944 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4945 case NEON::BI__builtin_neon_vrnda_v: 4946 case NEON::BI__builtin_neon_vrndaq_v: { 4947 Int = Intrinsic::round; 4948 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4949 } 4950 case NEON::BI__builtin_neon_vrndi_v: 4951 case NEON::BI__builtin_neon_vrndiq_v: { 4952 Int = Intrinsic::nearbyint; 4953 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4954 } 4955 case NEON::BI__builtin_neon_vrndm_v: 4956 case NEON::BI__builtin_neon_vrndmq_v: { 4957 Int = Intrinsic::floor; 4958 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4959 } 4960 case NEON::BI__builtin_neon_vrndn_v: 4961 case NEON::BI__builtin_neon_vrndnq_v: { 4962 Int = Intrinsic::aarch64_neon_frintn; 4963 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4964 } 4965 case NEON::BI__builtin_neon_vrndp_v: 4966 case NEON::BI__builtin_neon_vrndpq_v: { 4967 Int = Intrinsic::ceil; 4968 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4969 } 4970 case NEON::BI__builtin_neon_vrndx_v: 4971 case NEON::BI__builtin_neon_vrndxq_v: { 4972 Int = Intrinsic::rint; 4973 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4974 } 4975 case NEON::BI__builtin_neon_vrnd_v: 4976 case NEON::BI__builtin_neon_vrndq_v: { 4977 Int = Intrinsic::trunc; 4978 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4979 } 4980 case NEON::BI__builtin_neon_vceqz_v: 4981 case NEON::BI__builtin_neon_vceqzq_v: 4982 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4983 ICmpInst::ICMP_EQ, "vceqz"); 4984 case NEON::BI__builtin_neon_vcgez_v: 4985 case NEON::BI__builtin_neon_vcgezq_v: 4986 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4987 ICmpInst::ICMP_SGE, "vcgez"); 4988 case NEON::BI__builtin_neon_vclez_v: 4989 case NEON::BI__builtin_neon_vclezq_v: 4990 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4991 ICmpInst::ICMP_SLE, "vclez"); 4992 case NEON::BI__builtin_neon_vcgtz_v: 4993 case NEON::BI__builtin_neon_vcgtzq_v: 4994 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4995 ICmpInst::ICMP_SGT, "vcgtz"); 4996 case NEON::BI__builtin_neon_vcltz_v: 4997 case NEON::BI__builtin_neon_vcltzq_v: 4998 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4999 ICmpInst::ICMP_SLT, "vcltz"); 5000 case NEON::BI__builtin_neon_vcvt_f64_v: 5001 case NEON::BI__builtin_neon_vcvtq_f64_v: 5002 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5003 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5004 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5005 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5006 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5007 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5008 "unexpected vcvt_f64_f32 builtin"); 5009 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5010 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5011 5012 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5013 } 5014 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5015 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5016 "unexpected vcvt_f32_f64 builtin"); 5017 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5018 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5019 5020 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5021 } 5022 case NEON::BI__builtin_neon_vcvt_s32_v: 5023 case NEON::BI__builtin_neon_vcvt_u32_v: 5024 case NEON::BI__builtin_neon_vcvt_s64_v: 5025 case NEON::BI__builtin_neon_vcvt_u64_v: 5026 case NEON::BI__builtin_neon_vcvtq_s32_v: 5027 case NEON::BI__builtin_neon_vcvtq_u32_v: 5028 case NEON::BI__builtin_neon_vcvtq_s64_v: 5029 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5030 bool Double = 5031 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5032 llvm::Type *InTy = 5033 GetNeonType(this, 5034 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5035 : NeonTypeFlags::Float32, false, quad)); 5036 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5037 if (usgn) 5038 return Builder.CreateFPToUI(Ops[0], Ty); 5039 return Builder.CreateFPToSI(Ops[0], Ty); 5040 } 5041 case NEON::BI__builtin_neon_vcvta_s32_v: 5042 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5043 case NEON::BI__builtin_neon_vcvta_u32_v: 5044 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5045 case NEON::BI__builtin_neon_vcvta_s64_v: 5046 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5047 case NEON::BI__builtin_neon_vcvta_u64_v: 5048 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5049 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5050 bool Double = 5051 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5052 llvm::Type *InTy = 5053 GetNeonType(this, 5054 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5055 : NeonTypeFlags::Float32, false, quad)); 5056 llvm::Type *Tys[2] = { Ty, InTy }; 5057 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5058 } 5059 case NEON::BI__builtin_neon_vcvtm_s32_v: 5060 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5061 case NEON::BI__builtin_neon_vcvtm_u32_v: 5062 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5063 case NEON::BI__builtin_neon_vcvtm_s64_v: 5064 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5065 case NEON::BI__builtin_neon_vcvtm_u64_v: 5066 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5067 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5068 bool Double = 5069 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5070 llvm::Type *InTy = 5071 GetNeonType(this, 5072 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5073 : NeonTypeFlags::Float32, false, quad)); 5074 llvm::Type *Tys[2] = { Ty, InTy }; 5075 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5076 } 5077 case NEON::BI__builtin_neon_vcvtn_s32_v: 5078 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5079 case NEON::BI__builtin_neon_vcvtn_u32_v: 5080 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5081 case NEON::BI__builtin_neon_vcvtn_s64_v: 5082 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5083 case NEON::BI__builtin_neon_vcvtn_u64_v: 5084 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5085 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5086 bool Double = 5087 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5088 llvm::Type *InTy = 5089 GetNeonType(this, 5090 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5091 : NeonTypeFlags::Float32, false, quad)); 5092 llvm::Type *Tys[2] = { Ty, InTy }; 5093 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5094 } 5095 case NEON::BI__builtin_neon_vcvtp_s32_v: 5096 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5097 case NEON::BI__builtin_neon_vcvtp_u32_v: 5098 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5099 case NEON::BI__builtin_neon_vcvtp_s64_v: 5100 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5101 case NEON::BI__builtin_neon_vcvtp_u64_v: 5102 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5103 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5104 bool Double = 5105 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5106 llvm::Type *InTy = 5107 GetNeonType(this, 5108 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5109 : NeonTypeFlags::Float32, false, quad)); 5110 llvm::Type *Tys[2] = { Ty, InTy }; 5111 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5112 } 5113 case NEON::BI__builtin_neon_vmulx_v: 5114 case NEON::BI__builtin_neon_vmulxq_v: { 5115 Int = Intrinsic::aarch64_neon_fmulx; 5116 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5117 } 5118 case NEON::BI__builtin_neon_vmul_lane_v: 5119 case NEON::BI__builtin_neon_vmul_laneq_v: { 5120 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5121 bool Quad = false; 5122 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5123 Quad = true; 5124 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5125 llvm::Type *VTy = GetNeonType(this, 5126 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5127 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5128 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5129 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5130 return Builder.CreateBitCast(Result, Ty); 5131 } 5132 case NEON::BI__builtin_neon_vnegd_s64: 5133 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5134 case NEON::BI__builtin_neon_vpmaxnm_v: 5135 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5136 Int = Intrinsic::aarch64_neon_fmaxnmp; 5137 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5138 } 5139 case NEON::BI__builtin_neon_vpminnm_v: 5140 case NEON::BI__builtin_neon_vpminnmq_v: { 5141 Int = Intrinsic::aarch64_neon_fminnmp; 5142 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5143 } 5144 case NEON::BI__builtin_neon_vsqrt_v: 5145 case NEON::BI__builtin_neon_vsqrtq_v: { 5146 Int = Intrinsic::sqrt; 5147 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5148 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5149 } 5150 case NEON::BI__builtin_neon_vrbit_v: 5151 case NEON::BI__builtin_neon_vrbitq_v: { 5152 Int = Intrinsic::aarch64_neon_rbit; 5153 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5154 } 5155 case NEON::BI__builtin_neon_vaddv_u8: 5156 // FIXME: These are handled by the AArch64 scalar code. 5157 usgn = true; 5158 // FALLTHROUGH 5159 case NEON::BI__builtin_neon_vaddv_s8: { 5160 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5161 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5162 VTy = 5163 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5164 llvm::Type *Tys[2] = { Ty, VTy }; 5165 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5166 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5167 return Builder.CreateTrunc(Ops[0], 5168 llvm::IntegerType::get(getLLVMContext(), 8)); 5169 } 5170 case NEON::BI__builtin_neon_vaddv_u16: 5171 usgn = true; 5172 // FALLTHROUGH 5173 case NEON::BI__builtin_neon_vaddv_s16: { 5174 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5175 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5176 VTy = 5177 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5178 llvm::Type *Tys[2] = { Ty, VTy }; 5179 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5180 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5181 return Builder.CreateTrunc(Ops[0], 5182 llvm::IntegerType::get(getLLVMContext(), 16)); 5183 } 5184 case NEON::BI__builtin_neon_vaddvq_u8: 5185 usgn = true; 5186 // FALLTHROUGH 5187 case NEON::BI__builtin_neon_vaddvq_s8: { 5188 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5189 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5190 VTy = 5191 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5192 llvm::Type *Tys[2] = { Ty, VTy }; 5193 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5194 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5195 return Builder.CreateTrunc(Ops[0], 5196 llvm::IntegerType::get(getLLVMContext(), 8)); 5197 } 5198 case NEON::BI__builtin_neon_vaddvq_u16: 5199 usgn = true; 5200 // FALLTHROUGH 5201 case NEON::BI__builtin_neon_vaddvq_s16: { 5202 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5203 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5204 VTy = 5205 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5206 llvm::Type *Tys[2] = { Ty, VTy }; 5207 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5208 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5209 return Builder.CreateTrunc(Ops[0], 5210 llvm::IntegerType::get(getLLVMContext(), 16)); 5211 } 5212 case NEON::BI__builtin_neon_vmaxv_u8: { 5213 Int = Intrinsic::aarch64_neon_umaxv; 5214 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5215 VTy = 5216 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5217 llvm::Type *Tys[2] = { Ty, VTy }; 5218 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5219 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5220 return Builder.CreateTrunc(Ops[0], 5221 llvm::IntegerType::get(getLLVMContext(), 8)); 5222 } 5223 case NEON::BI__builtin_neon_vmaxv_u16: { 5224 Int = Intrinsic::aarch64_neon_umaxv; 5225 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5226 VTy = 5227 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5228 llvm::Type *Tys[2] = { Ty, VTy }; 5229 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5230 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5231 return Builder.CreateTrunc(Ops[0], 5232 llvm::IntegerType::get(getLLVMContext(), 16)); 5233 } 5234 case NEON::BI__builtin_neon_vmaxvq_u8: { 5235 Int = Intrinsic::aarch64_neon_umaxv; 5236 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5237 VTy = 5238 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5239 llvm::Type *Tys[2] = { Ty, VTy }; 5240 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5241 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5242 return Builder.CreateTrunc(Ops[0], 5243 llvm::IntegerType::get(getLLVMContext(), 8)); 5244 } 5245 case NEON::BI__builtin_neon_vmaxvq_u16: { 5246 Int = Intrinsic::aarch64_neon_umaxv; 5247 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5248 VTy = 5249 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5250 llvm::Type *Tys[2] = { Ty, VTy }; 5251 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5252 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5253 return Builder.CreateTrunc(Ops[0], 5254 llvm::IntegerType::get(getLLVMContext(), 16)); 5255 } 5256 case NEON::BI__builtin_neon_vmaxv_s8: { 5257 Int = Intrinsic::aarch64_neon_smaxv; 5258 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5259 VTy = 5260 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5261 llvm::Type *Tys[2] = { Ty, VTy }; 5262 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5263 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5264 return Builder.CreateTrunc(Ops[0], 5265 llvm::IntegerType::get(getLLVMContext(), 8)); 5266 } 5267 case NEON::BI__builtin_neon_vmaxv_s16: { 5268 Int = Intrinsic::aarch64_neon_smaxv; 5269 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5270 VTy = 5271 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5272 llvm::Type *Tys[2] = { Ty, VTy }; 5273 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5274 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5275 return Builder.CreateTrunc(Ops[0], 5276 llvm::IntegerType::get(getLLVMContext(), 16)); 5277 } 5278 case NEON::BI__builtin_neon_vmaxvq_s8: { 5279 Int = Intrinsic::aarch64_neon_smaxv; 5280 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5281 VTy = 5282 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5283 llvm::Type *Tys[2] = { Ty, VTy }; 5284 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5285 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5286 return Builder.CreateTrunc(Ops[0], 5287 llvm::IntegerType::get(getLLVMContext(), 8)); 5288 } 5289 case NEON::BI__builtin_neon_vmaxvq_s16: { 5290 Int = Intrinsic::aarch64_neon_smaxv; 5291 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5292 VTy = 5293 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5294 llvm::Type *Tys[2] = { Ty, VTy }; 5295 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5296 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5297 return Builder.CreateTrunc(Ops[0], 5298 llvm::IntegerType::get(getLLVMContext(), 16)); 5299 } 5300 case NEON::BI__builtin_neon_vminv_u8: { 5301 Int = Intrinsic::aarch64_neon_uminv; 5302 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5303 VTy = 5304 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5305 llvm::Type *Tys[2] = { Ty, VTy }; 5306 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5307 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5308 return Builder.CreateTrunc(Ops[0], 5309 llvm::IntegerType::get(getLLVMContext(), 8)); 5310 } 5311 case NEON::BI__builtin_neon_vminv_u16: { 5312 Int = Intrinsic::aarch64_neon_uminv; 5313 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5314 VTy = 5315 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5316 llvm::Type *Tys[2] = { Ty, VTy }; 5317 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5318 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5319 return Builder.CreateTrunc(Ops[0], 5320 llvm::IntegerType::get(getLLVMContext(), 16)); 5321 } 5322 case NEON::BI__builtin_neon_vminvq_u8: { 5323 Int = Intrinsic::aarch64_neon_uminv; 5324 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5325 VTy = 5326 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5327 llvm::Type *Tys[2] = { Ty, VTy }; 5328 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5329 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5330 return Builder.CreateTrunc(Ops[0], 5331 llvm::IntegerType::get(getLLVMContext(), 8)); 5332 } 5333 case NEON::BI__builtin_neon_vminvq_u16: { 5334 Int = Intrinsic::aarch64_neon_uminv; 5335 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5336 VTy = 5337 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5338 llvm::Type *Tys[2] = { Ty, VTy }; 5339 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5340 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5341 return Builder.CreateTrunc(Ops[0], 5342 llvm::IntegerType::get(getLLVMContext(), 16)); 5343 } 5344 case NEON::BI__builtin_neon_vminv_s8: { 5345 Int = Intrinsic::aarch64_neon_sminv; 5346 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5347 VTy = 5348 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5349 llvm::Type *Tys[2] = { Ty, VTy }; 5350 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5351 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5352 return Builder.CreateTrunc(Ops[0], 5353 llvm::IntegerType::get(getLLVMContext(), 8)); 5354 } 5355 case NEON::BI__builtin_neon_vminv_s16: { 5356 Int = Intrinsic::aarch64_neon_sminv; 5357 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5358 VTy = 5359 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5360 llvm::Type *Tys[2] = { Ty, VTy }; 5361 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5362 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5363 return Builder.CreateTrunc(Ops[0], 5364 llvm::IntegerType::get(getLLVMContext(), 16)); 5365 } 5366 case NEON::BI__builtin_neon_vminvq_s8: { 5367 Int = Intrinsic::aarch64_neon_sminv; 5368 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5369 VTy = 5370 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5371 llvm::Type *Tys[2] = { Ty, VTy }; 5372 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5373 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5374 return Builder.CreateTrunc(Ops[0], 5375 llvm::IntegerType::get(getLLVMContext(), 8)); 5376 } 5377 case NEON::BI__builtin_neon_vminvq_s16: { 5378 Int = Intrinsic::aarch64_neon_sminv; 5379 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5380 VTy = 5381 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5382 llvm::Type *Tys[2] = { Ty, VTy }; 5383 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5384 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5385 return Builder.CreateTrunc(Ops[0], 5386 llvm::IntegerType::get(getLLVMContext(), 16)); 5387 } 5388 case NEON::BI__builtin_neon_vmul_n_f64: { 5389 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5390 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5391 return Builder.CreateFMul(Ops[0], RHS); 5392 } 5393 case NEON::BI__builtin_neon_vaddlv_u8: { 5394 Int = Intrinsic::aarch64_neon_uaddlv; 5395 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5396 VTy = 5397 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5398 llvm::Type *Tys[2] = { Ty, VTy }; 5399 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5400 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5401 return Builder.CreateTrunc(Ops[0], 5402 llvm::IntegerType::get(getLLVMContext(), 16)); 5403 } 5404 case NEON::BI__builtin_neon_vaddlv_u16: { 5405 Int = Intrinsic::aarch64_neon_uaddlv; 5406 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5407 VTy = 5408 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5409 llvm::Type *Tys[2] = { Ty, VTy }; 5410 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5411 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5412 } 5413 case NEON::BI__builtin_neon_vaddlvq_u8: { 5414 Int = Intrinsic::aarch64_neon_uaddlv; 5415 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5416 VTy = 5417 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5418 llvm::Type *Tys[2] = { Ty, VTy }; 5419 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5420 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5421 return Builder.CreateTrunc(Ops[0], 5422 llvm::IntegerType::get(getLLVMContext(), 16)); 5423 } 5424 case NEON::BI__builtin_neon_vaddlvq_u16: { 5425 Int = Intrinsic::aarch64_neon_uaddlv; 5426 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5427 VTy = 5428 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5429 llvm::Type *Tys[2] = { Ty, VTy }; 5430 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5431 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5432 } 5433 case NEON::BI__builtin_neon_vaddlv_s8: { 5434 Int = Intrinsic::aarch64_neon_saddlv; 5435 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5436 VTy = 5437 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5438 llvm::Type *Tys[2] = { Ty, VTy }; 5439 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5440 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5441 return Builder.CreateTrunc(Ops[0], 5442 llvm::IntegerType::get(getLLVMContext(), 16)); 5443 } 5444 case NEON::BI__builtin_neon_vaddlv_s16: { 5445 Int = Intrinsic::aarch64_neon_saddlv; 5446 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5447 VTy = 5448 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5449 llvm::Type *Tys[2] = { Ty, VTy }; 5450 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5451 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5452 } 5453 case NEON::BI__builtin_neon_vaddlvq_s8: { 5454 Int = Intrinsic::aarch64_neon_saddlv; 5455 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5456 VTy = 5457 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5458 llvm::Type *Tys[2] = { Ty, VTy }; 5459 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5460 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5461 return Builder.CreateTrunc(Ops[0], 5462 llvm::IntegerType::get(getLLVMContext(), 16)); 5463 } 5464 case NEON::BI__builtin_neon_vaddlvq_s16: { 5465 Int = Intrinsic::aarch64_neon_saddlv; 5466 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5467 VTy = 5468 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5469 llvm::Type *Tys[2] = { Ty, VTy }; 5470 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5471 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5472 } 5473 case NEON::BI__builtin_neon_vsri_n_v: 5474 case NEON::BI__builtin_neon_vsriq_n_v: { 5475 Int = Intrinsic::aarch64_neon_vsri; 5476 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5477 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5478 } 5479 case NEON::BI__builtin_neon_vsli_n_v: 5480 case NEON::BI__builtin_neon_vsliq_n_v: { 5481 Int = Intrinsic::aarch64_neon_vsli; 5482 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5483 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5484 } 5485 case NEON::BI__builtin_neon_vsra_n_v: 5486 case NEON::BI__builtin_neon_vsraq_n_v: 5487 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5488 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5489 return Builder.CreateAdd(Ops[0], Ops[1]); 5490 case NEON::BI__builtin_neon_vrsra_n_v: 5491 case NEON::BI__builtin_neon_vrsraq_n_v: { 5492 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5493 SmallVector<llvm::Value*,2> TmpOps; 5494 TmpOps.push_back(Ops[1]); 5495 TmpOps.push_back(Ops[2]); 5496 Function* F = CGM.getIntrinsic(Int, Ty); 5497 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5498 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5499 return Builder.CreateAdd(Ops[0], tmp); 5500 } 5501 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5502 // of an Align parameter here. 5503 case NEON::BI__builtin_neon_vld1_x2_v: 5504 case NEON::BI__builtin_neon_vld1q_x2_v: 5505 case NEON::BI__builtin_neon_vld1_x3_v: 5506 case NEON::BI__builtin_neon_vld1q_x3_v: 5507 case NEON::BI__builtin_neon_vld1_x4_v: 5508 case NEON::BI__builtin_neon_vld1q_x4_v: { 5509 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5510 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5511 llvm::Type *Tys[2] = { VTy, PTy }; 5512 unsigned Int; 5513 switch (BuiltinID) { 5514 case NEON::BI__builtin_neon_vld1_x2_v: 5515 case NEON::BI__builtin_neon_vld1q_x2_v: 5516 Int = Intrinsic::aarch64_neon_ld1x2; 5517 break; 5518 case NEON::BI__builtin_neon_vld1_x3_v: 5519 case NEON::BI__builtin_neon_vld1q_x3_v: 5520 Int = Intrinsic::aarch64_neon_ld1x3; 5521 break; 5522 case NEON::BI__builtin_neon_vld1_x4_v: 5523 case NEON::BI__builtin_neon_vld1q_x4_v: 5524 Int = Intrinsic::aarch64_neon_ld1x4; 5525 break; 5526 } 5527 Function *F = CGM.getIntrinsic(Int, Tys); 5528 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5529 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5530 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5531 return Builder.CreateStore(Ops[1], Ops[0]); 5532 } 5533 case NEON::BI__builtin_neon_vst1_x2_v: 5534 case NEON::BI__builtin_neon_vst1q_x2_v: 5535 case NEON::BI__builtin_neon_vst1_x3_v: 5536 case NEON::BI__builtin_neon_vst1q_x3_v: 5537 case NEON::BI__builtin_neon_vst1_x4_v: 5538 case NEON::BI__builtin_neon_vst1q_x4_v: { 5539 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5540 llvm::Type *Tys[2] = { VTy, PTy }; 5541 unsigned Int; 5542 switch (BuiltinID) { 5543 case NEON::BI__builtin_neon_vst1_x2_v: 5544 case NEON::BI__builtin_neon_vst1q_x2_v: 5545 Int = Intrinsic::aarch64_neon_st1x2; 5546 break; 5547 case NEON::BI__builtin_neon_vst1_x3_v: 5548 case NEON::BI__builtin_neon_vst1q_x3_v: 5549 Int = Intrinsic::aarch64_neon_st1x3; 5550 break; 5551 case NEON::BI__builtin_neon_vst1_x4_v: 5552 case NEON::BI__builtin_neon_vst1q_x4_v: 5553 Int = Intrinsic::aarch64_neon_st1x4; 5554 break; 5555 } 5556 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5557 IntOps.push_back(Ops[0]); 5558 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5559 } 5560 case NEON::BI__builtin_neon_vld1_v: 5561 case NEON::BI__builtin_neon_vld1q_v: 5562 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5563 return Builder.CreateLoad(Ops[0]); 5564 case NEON::BI__builtin_neon_vst1_v: 5565 case NEON::BI__builtin_neon_vst1q_v: 5566 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5567 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5568 return Builder.CreateStore(Ops[1], Ops[0]); 5569 case NEON::BI__builtin_neon_vld1_lane_v: 5570 case NEON::BI__builtin_neon_vld1q_lane_v: 5571 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5572 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5573 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5574 Ops[0] = Builder.CreateLoad(Ops[0]); 5575 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5576 case NEON::BI__builtin_neon_vld1_dup_v: 5577 case NEON::BI__builtin_neon_vld1q_dup_v: { 5578 Value *V = UndefValue::get(Ty); 5579 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5580 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5581 Ops[0] = Builder.CreateLoad(Ops[0]); 5582 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5583 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5584 return EmitNeonSplat(Ops[0], CI); 5585 } 5586 case NEON::BI__builtin_neon_vst1_lane_v: 5587 case NEON::BI__builtin_neon_vst1q_lane_v: 5588 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5589 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5590 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5591 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5592 case NEON::BI__builtin_neon_vld2_v: 5593 case NEON::BI__builtin_neon_vld2q_v: { 5594 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5595 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5596 llvm::Type *Tys[2] = { VTy, PTy }; 5597 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5598 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5599 Ops[0] = Builder.CreateBitCast(Ops[0], 5600 llvm::PointerType::getUnqual(Ops[1]->getType())); 5601 return Builder.CreateStore(Ops[1], Ops[0]); 5602 } 5603 case NEON::BI__builtin_neon_vld3_v: 5604 case NEON::BI__builtin_neon_vld3q_v: { 5605 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5606 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5607 llvm::Type *Tys[2] = { VTy, PTy }; 5608 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5609 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5610 Ops[0] = Builder.CreateBitCast(Ops[0], 5611 llvm::PointerType::getUnqual(Ops[1]->getType())); 5612 return Builder.CreateStore(Ops[1], Ops[0]); 5613 } 5614 case NEON::BI__builtin_neon_vld4_v: 5615 case NEON::BI__builtin_neon_vld4q_v: { 5616 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5617 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5618 llvm::Type *Tys[2] = { VTy, PTy }; 5619 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5620 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5621 Ops[0] = Builder.CreateBitCast(Ops[0], 5622 llvm::PointerType::getUnqual(Ops[1]->getType())); 5623 return Builder.CreateStore(Ops[1], Ops[0]); 5624 } 5625 case NEON::BI__builtin_neon_vld2_dup_v: 5626 case NEON::BI__builtin_neon_vld2q_dup_v: { 5627 llvm::Type *PTy = 5628 llvm::PointerType::getUnqual(VTy->getElementType()); 5629 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5630 llvm::Type *Tys[2] = { VTy, PTy }; 5631 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5632 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5633 Ops[0] = Builder.CreateBitCast(Ops[0], 5634 llvm::PointerType::getUnqual(Ops[1]->getType())); 5635 return Builder.CreateStore(Ops[1], Ops[0]); 5636 } 5637 case NEON::BI__builtin_neon_vld3_dup_v: 5638 case NEON::BI__builtin_neon_vld3q_dup_v: { 5639 llvm::Type *PTy = 5640 llvm::PointerType::getUnqual(VTy->getElementType()); 5641 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5642 llvm::Type *Tys[2] = { VTy, PTy }; 5643 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5644 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5645 Ops[0] = Builder.CreateBitCast(Ops[0], 5646 llvm::PointerType::getUnqual(Ops[1]->getType())); 5647 return Builder.CreateStore(Ops[1], Ops[0]); 5648 } 5649 case NEON::BI__builtin_neon_vld4_dup_v: 5650 case NEON::BI__builtin_neon_vld4q_dup_v: { 5651 llvm::Type *PTy = 5652 llvm::PointerType::getUnqual(VTy->getElementType()); 5653 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5654 llvm::Type *Tys[2] = { VTy, PTy }; 5655 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5656 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5657 Ops[0] = Builder.CreateBitCast(Ops[0], 5658 llvm::PointerType::getUnqual(Ops[1]->getType())); 5659 return Builder.CreateStore(Ops[1], Ops[0]); 5660 } 5661 case NEON::BI__builtin_neon_vld2_lane_v: 5662 case NEON::BI__builtin_neon_vld2q_lane_v: { 5663 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5664 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5665 Ops.push_back(Ops[1]); 5666 Ops.erase(Ops.begin()+1); 5667 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5668 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5669 Ops[3] = Builder.CreateZExt(Ops[3], 5670 llvm::IntegerType::get(getLLVMContext(), 64)); 5671 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5672 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5673 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5674 return Builder.CreateStore(Ops[1], Ops[0]); 5675 } 5676 case NEON::BI__builtin_neon_vld3_lane_v: 5677 case NEON::BI__builtin_neon_vld3q_lane_v: { 5678 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5679 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, 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.CreateBitCast(Ops[3], Ty); 5685 Ops[4] = Builder.CreateZExt(Ops[4], 5686 llvm::IntegerType::get(getLLVMContext(), 64)); 5687 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5688 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5689 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5690 return Builder.CreateStore(Ops[1], Ops[0]); 5691 } 5692 case NEON::BI__builtin_neon_vld4_lane_v: 5693 case NEON::BI__builtin_neon_vld4q_lane_v: { 5694 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5695 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5696 Ops.push_back(Ops[1]); 5697 Ops.erase(Ops.begin()+1); 5698 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5699 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5700 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5701 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5702 Ops[5] = Builder.CreateZExt(Ops[5], 5703 llvm::IntegerType::get(getLLVMContext(), 64)); 5704 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5705 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5706 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5707 return Builder.CreateStore(Ops[1], Ops[0]); 5708 } 5709 case NEON::BI__builtin_neon_vst2_v: 5710 case NEON::BI__builtin_neon_vst2q_v: { 5711 Ops.push_back(Ops[0]); 5712 Ops.erase(Ops.begin()); 5713 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5714 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5715 Ops, ""); 5716 } 5717 case NEON::BI__builtin_neon_vst2_lane_v: 5718 case NEON::BI__builtin_neon_vst2q_lane_v: { 5719 Ops.push_back(Ops[0]); 5720 Ops.erase(Ops.begin()); 5721 Ops[2] = Builder.CreateZExt(Ops[2], 5722 llvm::IntegerType::get(getLLVMContext(), 64)); 5723 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5724 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5725 Ops, ""); 5726 } 5727 case NEON::BI__builtin_neon_vst3_v: 5728 case NEON::BI__builtin_neon_vst3q_v: { 5729 Ops.push_back(Ops[0]); 5730 Ops.erase(Ops.begin()); 5731 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5733 Ops, ""); 5734 } 5735 case NEON::BI__builtin_neon_vst3_lane_v: 5736 case NEON::BI__builtin_neon_vst3q_lane_v: { 5737 Ops.push_back(Ops[0]); 5738 Ops.erase(Ops.begin()); 5739 Ops[3] = Builder.CreateZExt(Ops[3], 5740 llvm::IntegerType::get(getLLVMContext(), 64)); 5741 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5742 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5743 Ops, ""); 5744 } 5745 case NEON::BI__builtin_neon_vst4_v: 5746 case NEON::BI__builtin_neon_vst4q_v: { 5747 Ops.push_back(Ops[0]); 5748 Ops.erase(Ops.begin()); 5749 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5750 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5751 Ops, ""); 5752 } 5753 case NEON::BI__builtin_neon_vst4_lane_v: 5754 case NEON::BI__builtin_neon_vst4q_lane_v: { 5755 Ops.push_back(Ops[0]); 5756 Ops.erase(Ops.begin()); 5757 Ops[4] = Builder.CreateZExt(Ops[4], 5758 llvm::IntegerType::get(getLLVMContext(), 64)); 5759 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5760 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5761 Ops, ""); 5762 } 5763 case NEON::BI__builtin_neon_vtrn_v: 5764 case NEON::BI__builtin_neon_vtrnq_v: { 5765 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5766 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5767 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5768 Value *SV = nullptr; 5769 5770 for (unsigned vi = 0; vi != 2; ++vi) { 5771 SmallVector<Constant*, 16> Indices; 5772 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5773 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5774 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5775 } 5776 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5777 SV = llvm::ConstantVector::get(Indices); 5778 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5779 SV = Builder.CreateStore(SV, Addr); 5780 } 5781 return SV; 5782 } 5783 case NEON::BI__builtin_neon_vuzp_v: 5784 case NEON::BI__builtin_neon_vuzpq_v: { 5785 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5786 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5787 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5788 Value *SV = nullptr; 5789 5790 for (unsigned vi = 0; vi != 2; ++vi) { 5791 SmallVector<Constant*, 16> Indices; 5792 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5793 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5794 5795 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5796 SV = llvm::ConstantVector::get(Indices); 5797 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5798 SV = Builder.CreateStore(SV, Addr); 5799 } 5800 return SV; 5801 } 5802 case NEON::BI__builtin_neon_vzip_v: 5803 case NEON::BI__builtin_neon_vzipq_v: { 5804 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5805 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5806 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5807 Value *SV = nullptr; 5808 5809 for (unsigned vi = 0; vi != 2; ++vi) { 5810 SmallVector<Constant*, 16> Indices; 5811 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5812 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5813 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5814 } 5815 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5816 SV = llvm::ConstantVector::get(Indices); 5817 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5818 SV = Builder.CreateStore(SV, Addr); 5819 } 5820 return SV; 5821 } 5822 case NEON::BI__builtin_neon_vqtbl1q_v: { 5823 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5824 Ops, "vtbl1"); 5825 } 5826 case NEON::BI__builtin_neon_vqtbl2q_v: { 5827 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5828 Ops, "vtbl2"); 5829 } 5830 case NEON::BI__builtin_neon_vqtbl3q_v: { 5831 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5832 Ops, "vtbl3"); 5833 } 5834 case NEON::BI__builtin_neon_vqtbl4q_v: { 5835 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5836 Ops, "vtbl4"); 5837 } 5838 case NEON::BI__builtin_neon_vqtbx1q_v: { 5839 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5840 Ops, "vtbx1"); 5841 } 5842 case NEON::BI__builtin_neon_vqtbx2q_v: { 5843 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5844 Ops, "vtbx2"); 5845 } 5846 case NEON::BI__builtin_neon_vqtbx3q_v: { 5847 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5848 Ops, "vtbx3"); 5849 } 5850 case NEON::BI__builtin_neon_vqtbx4q_v: { 5851 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5852 Ops, "vtbx4"); 5853 } 5854 case NEON::BI__builtin_neon_vsqadd_v: 5855 case NEON::BI__builtin_neon_vsqaddq_v: { 5856 Int = Intrinsic::aarch64_neon_usqadd; 5857 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5858 } 5859 case NEON::BI__builtin_neon_vuqadd_v: 5860 case NEON::BI__builtin_neon_vuqaddq_v: { 5861 Int = Intrinsic::aarch64_neon_suqadd; 5862 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5863 } 5864 } 5865 } 5866 5867 llvm::Value *CodeGenFunction:: 5868 BuildVector(ArrayRef<llvm::Value*> Ops) { 5869 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5870 "Not a power-of-two sized vector!"); 5871 bool AllConstants = true; 5872 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5873 AllConstants &= isa<Constant>(Ops[i]); 5874 5875 // If this is a constant vector, create a ConstantVector. 5876 if (AllConstants) { 5877 SmallVector<llvm::Constant*, 16> CstOps; 5878 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5879 CstOps.push_back(cast<Constant>(Ops[i])); 5880 return llvm::ConstantVector::get(CstOps); 5881 } 5882 5883 // Otherwise, insertelement the values to build the vector. 5884 Value *Result = 5885 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5886 5887 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5888 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5889 5890 return Result; 5891 } 5892 5893 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5894 const CallExpr *E) { 5895 SmallVector<Value*, 4> Ops; 5896 5897 // Find out if any arguments are required to be integer constant expressions. 5898 unsigned ICEArguments = 0; 5899 ASTContext::GetBuiltinTypeError Error; 5900 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5901 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5902 5903 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5904 // If this is a normal argument, just emit it as a scalar. 5905 if ((ICEArguments & (1 << i)) == 0) { 5906 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5907 continue; 5908 } 5909 5910 // If this is required to be a constant, constant fold it so that we know 5911 // that the generated intrinsic gets a ConstantInt. 5912 llvm::APSInt Result; 5913 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5914 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5915 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5916 } 5917 5918 switch (BuiltinID) { 5919 default: return nullptr; 5920 case X86::BI_mm_prefetch: { 5921 Value *Address = EmitScalarExpr(E->getArg(0)); 5922 Value *RW = ConstantInt::get(Int32Ty, 0); 5923 Value *Locality = EmitScalarExpr(E->getArg(1)); 5924 Value *Data = ConstantInt::get(Int32Ty, 1); 5925 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5926 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 5927 } 5928 case X86::BI__builtin_ia32_vec_init_v8qi: 5929 case X86::BI__builtin_ia32_vec_init_v4hi: 5930 case X86::BI__builtin_ia32_vec_init_v2si: 5931 return Builder.CreateBitCast(BuildVector(Ops), 5932 llvm::Type::getX86_MMXTy(getLLVMContext())); 5933 case X86::BI__builtin_ia32_vec_ext_v2si: 5934 return Builder.CreateExtractElement(Ops[0], 5935 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5936 case X86::BI__builtin_ia32_ldmxcsr: { 5937 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5938 Builder.CreateStore(Ops[0], Tmp); 5939 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5940 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5941 } 5942 case X86::BI__builtin_ia32_stmxcsr: { 5943 Value *Tmp = CreateMemTemp(E->getType()); 5944 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5945 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5946 return Builder.CreateLoad(Tmp, "stmxcsr"); 5947 } 5948 case X86::BI__builtin_ia32_storehps: 5949 case X86::BI__builtin_ia32_storelps: { 5950 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5951 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5952 5953 // cast val v2i64 5954 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5955 5956 // extract (0, 1) 5957 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5958 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5959 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5960 5961 // cast pointer to i64 & store 5962 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5963 return Builder.CreateStore(Ops[1], Ops[0]); 5964 } 5965 case X86::BI__builtin_ia32_palignr128: 5966 case X86::BI__builtin_ia32_palignr256: { 5967 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5968 5969 unsigned NumElts = 5970 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 5971 assert(NumElts % 16 == 0); 5972 unsigned NumLanes = NumElts / 16; 5973 unsigned NumLaneElts = NumElts / NumLanes; 5974 5975 // If palignr is shifting the pair of vectors more than the size of two 5976 // lanes, emit zero. 5977 if (ShiftVal >= (2 * NumLaneElts)) 5978 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5979 5980 // If palignr is shifting the pair of input vectors more than one lane, 5981 // but less than two lanes, convert to shifting in zeroes. 5982 if (ShiftVal > NumLaneElts) { 5983 ShiftVal -= NumLaneElts; 5984 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 5985 } 5986 5987 SmallVector<llvm::Constant*, 32> Indices; 5988 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5989 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 5990 for (unsigned i = 0; i != NumLaneElts; ++i) { 5991 unsigned Idx = ShiftVal + i; 5992 if (Idx >= NumLaneElts) 5993 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 5994 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 5995 } 5996 } 5997 5998 Value* SV = llvm::ConstantVector::get(Indices); 5999 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6000 } 6001 case X86::BI__builtin_ia32_pslldqi256: { 6002 // Shift value is in bits so divide by 8. 6003 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6004 6005 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6006 if (shiftVal >= 16) 6007 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6008 6009 SmallVector<llvm::Constant*, 32> Indices; 6010 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6011 for (unsigned l = 0; l != 32; l += 16) { 6012 for (unsigned i = 0; i != 16; ++i) { 6013 unsigned Idx = 32 + i - shiftVal; 6014 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6015 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6016 } 6017 } 6018 6019 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6020 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6021 Value *Zero = llvm::Constant::getNullValue(VecTy); 6022 6023 Value *SV = llvm::ConstantVector::get(Indices); 6024 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6025 llvm::Type *ResultType = ConvertType(E->getType()); 6026 return Builder.CreateBitCast(SV, ResultType, "cast"); 6027 } 6028 case X86::BI__builtin_ia32_psrldqi256: { 6029 // Shift value is in bits so divide by 8. 6030 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6031 6032 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6033 if (shiftVal >= 16) 6034 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6035 6036 SmallVector<llvm::Constant*, 32> Indices; 6037 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6038 for (unsigned l = 0; l != 32; l += 16) { 6039 for (unsigned i = 0; i != 16; ++i) { 6040 unsigned Idx = i + shiftVal; 6041 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6042 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6043 } 6044 } 6045 6046 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6047 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6048 Value *Zero = llvm::Constant::getNullValue(VecTy); 6049 6050 Value *SV = llvm::ConstantVector::get(Indices); 6051 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6052 llvm::Type *ResultType = ConvertType(E->getType()); 6053 return Builder.CreateBitCast(SV, ResultType, "cast"); 6054 } 6055 case X86::BI__builtin_ia32_movntps: 6056 case X86::BI__builtin_ia32_movntps256: 6057 case X86::BI__builtin_ia32_movntpd: 6058 case X86::BI__builtin_ia32_movntpd256: 6059 case X86::BI__builtin_ia32_movntdq: 6060 case X86::BI__builtin_ia32_movntdq256: 6061 case X86::BI__builtin_ia32_movnti: 6062 case X86::BI__builtin_ia32_movnti64: { 6063 llvm::MDNode *Node = llvm::MDNode::get( 6064 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6065 6066 // Convert the type of the pointer to a pointer to the stored type. 6067 Value *BC = Builder.CreateBitCast(Ops[0], 6068 llvm::PointerType::getUnqual(Ops[1]->getType()), 6069 "cast"); 6070 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6071 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6072 6073 // If the operand is an integer, we can't assume alignment. Otherwise, 6074 // assume natural alignment. 6075 QualType ArgTy = E->getArg(1)->getType(); 6076 unsigned Align; 6077 if (ArgTy->isIntegerType()) 6078 Align = 1; 6079 else 6080 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6081 SI->setAlignment(Align); 6082 return SI; 6083 } 6084 // 3DNow! 6085 case X86::BI__builtin_ia32_pswapdsf: 6086 case X86::BI__builtin_ia32_pswapdsi: { 6087 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6088 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6089 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6090 return Builder.CreateCall(F, Ops, "pswapd"); 6091 } 6092 case X86::BI__builtin_ia32_rdrand16_step: 6093 case X86::BI__builtin_ia32_rdrand32_step: 6094 case X86::BI__builtin_ia32_rdrand64_step: 6095 case X86::BI__builtin_ia32_rdseed16_step: 6096 case X86::BI__builtin_ia32_rdseed32_step: 6097 case X86::BI__builtin_ia32_rdseed64_step: { 6098 Intrinsic::ID ID; 6099 switch (BuiltinID) { 6100 default: llvm_unreachable("Unsupported intrinsic!"); 6101 case X86::BI__builtin_ia32_rdrand16_step: 6102 ID = Intrinsic::x86_rdrand_16; 6103 break; 6104 case X86::BI__builtin_ia32_rdrand32_step: 6105 ID = Intrinsic::x86_rdrand_32; 6106 break; 6107 case X86::BI__builtin_ia32_rdrand64_step: 6108 ID = Intrinsic::x86_rdrand_64; 6109 break; 6110 case X86::BI__builtin_ia32_rdseed16_step: 6111 ID = Intrinsic::x86_rdseed_16; 6112 break; 6113 case X86::BI__builtin_ia32_rdseed32_step: 6114 ID = Intrinsic::x86_rdseed_32; 6115 break; 6116 case X86::BI__builtin_ia32_rdseed64_step: 6117 ID = Intrinsic::x86_rdseed_64; 6118 break; 6119 } 6120 6121 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {}); 6122 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6123 return Builder.CreateExtractValue(Call, 1); 6124 } 6125 // SSE comparison intrisics 6126 case X86::BI__builtin_ia32_cmpeqps: 6127 case X86::BI__builtin_ia32_cmpltps: 6128 case X86::BI__builtin_ia32_cmpleps: 6129 case X86::BI__builtin_ia32_cmpunordps: 6130 case X86::BI__builtin_ia32_cmpneqps: 6131 case X86::BI__builtin_ia32_cmpnltps: 6132 case X86::BI__builtin_ia32_cmpnleps: 6133 case X86::BI__builtin_ia32_cmpordps: 6134 case X86::BI__builtin_ia32_cmpeqss: 6135 case X86::BI__builtin_ia32_cmpltss: 6136 case X86::BI__builtin_ia32_cmpless: 6137 case X86::BI__builtin_ia32_cmpunordss: 6138 case X86::BI__builtin_ia32_cmpneqss: 6139 case X86::BI__builtin_ia32_cmpnltss: 6140 case X86::BI__builtin_ia32_cmpnless: 6141 case X86::BI__builtin_ia32_cmpordss: 6142 case X86::BI__builtin_ia32_cmpeqpd: 6143 case X86::BI__builtin_ia32_cmpltpd: 6144 case X86::BI__builtin_ia32_cmplepd: 6145 case X86::BI__builtin_ia32_cmpunordpd: 6146 case X86::BI__builtin_ia32_cmpneqpd: 6147 case X86::BI__builtin_ia32_cmpnltpd: 6148 case X86::BI__builtin_ia32_cmpnlepd: 6149 case X86::BI__builtin_ia32_cmpordpd: 6150 case X86::BI__builtin_ia32_cmpeqsd: 6151 case X86::BI__builtin_ia32_cmpltsd: 6152 case X86::BI__builtin_ia32_cmplesd: 6153 case X86::BI__builtin_ia32_cmpunordsd: 6154 case X86::BI__builtin_ia32_cmpneqsd: 6155 case X86::BI__builtin_ia32_cmpnltsd: 6156 case X86::BI__builtin_ia32_cmpnlesd: 6157 case X86::BI__builtin_ia32_cmpordsd: 6158 // These exist so that the builtin that takes an immediate can be bounds 6159 // checked by clang to avoid passing bad immediates to the backend. Since 6160 // AVX has a larger immediate than SSE we would need separate builtins to 6161 // do the different bounds checking. Rather than create a clang specific 6162 // SSE only builtin, this implements eight separate builtins to match gcc 6163 // implementation. 6164 6165 // Choose the immediate. 6166 unsigned Imm; 6167 switch (BuiltinID) { 6168 default: llvm_unreachable("Unsupported intrinsic!"); 6169 case X86::BI__builtin_ia32_cmpeqps: 6170 case X86::BI__builtin_ia32_cmpeqss: 6171 case X86::BI__builtin_ia32_cmpeqpd: 6172 case X86::BI__builtin_ia32_cmpeqsd: 6173 Imm = 0; 6174 break; 6175 case X86::BI__builtin_ia32_cmpltps: 6176 case X86::BI__builtin_ia32_cmpltss: 6177 case X86::BI__builtin_ia32_cmpltpd: 6178 case X86::BI__builtin_ia32_cmpltsd: 6179 Imm = 1; 6180 break; 6181 case X86::BI__builtin_ia32_cmpleps: 6182 case X86::BI__builtin_ia32_cmpless: 6183 case X86::BI__builtin_ia32_cmplepd: 6184 case X86::BI__builtin_ia32_cmplesd: 6185 Imm = 2; 6186 break; 6187 case X86::BI__builtin_ia32_cmpunordps: 6188 case X86::BI__builtin_ia32_cmpunordss: 6189 case X86::BI__builtin_ia32_cmpunordpd: 6190 case X86::BI__builtin_ia32_cmpunordsd: 6191 Imm = 3; 6192 break; 6193 case X86::BI__builtin_ia32_cmpneqps: 6194 case X86::BI__builtin_ia32_cmpneqss: 6195 case X86::BI__builtin_ia32_cmpneqpd: 6196 case X86::BI__builtin_ia32_cmpneqsd: 6197 Imm = 4; 6198 break; 6199 case X86::BI__builtin_ia32_cmpnltps: 6200 case X86::BI__builtin_ia32_cmpnltss: 6201 case X86::BI__builtin_ia32_cmpnltpd: 6202 case X86::BI__builtin_ia32_cmpnltsd: 6203 Imm = 5; 6204 break; 6205 case X86::BI__builtin_ia32_cmpnleps: 6206 case X86::BI__builtin_ia32_cmpnless: 6207 case X86::BI__builtin_ia32_cmpnlepd: 6208 case X86::BI__builtin_ia32_cmpnlesd: 6209 Imm = 6; 6210 break; 6211 case X86::BI__builtin_ia32_cmpordps: 6212 case X86::BI__builtin_ia32_cmpordss: 6213 case X86::BI__builtin_ia32_cmpordpd: 6214 case X86::BI__builtin_ia32_cmpordsd: 6215 Imm = 7; 6216 break; 6217 } 6218 6219 // Choose the intrinsic ID. 6220 const char *name; 6221 Intrinsic::ID ID; 6222 switch (BuiltinID) { 6223 default: llvm_unreachable("Unsupported intrinsic!"); 6224 case X86::BI__builtin_ia32_cmpeqps: 6225 case X86::BI__builtin_ia32_cmpltps: 6226 case X86::BI__builtin_ia32_cmpleps: 6227 case X86::BI__builtin_ia32_cmpunordps: 6228 case X86::BI__builtin_ia32_cmpneqps: 6229 case X86::BI__builtin_ia32_cmpnltps: 6230 case X86::BI__builtin_ia32_cmpnleps: 6231 case X86::BI__builtin_ia32_cmpordps: 6232 name = "cmpps"; 6233 ID = Intrinsic::x86_sse_cmp_ps; 6234 break; 6235 case X86::BI__builtin_ia32_cmpeqss: 6236 case X86::BI__builtin_ia32_cmpltss: 6237 case X86::BI__builtin_ia32_cmpless: 6238 case X86::BI__builtin_ia32_cmpunordss: 6239 case X86::BI__builtin_ia32_cmpneqss: 6240 case X86::BI__builtin_ia32_cmpnltss: 6241 case X86::BI__builtin_ia32_cmpnless: 6242 case X86::BI__builtin_ia32_cmpordss: 6243 name = "cmpss"; 6244 ID = Intrinsic::x86_sse_cmp_ss; 6245 break; 6246 case X86::BI__builtin_ia32_cmpeqpd: 6247 case X86::BI__builtin_ia32_cmpltpd: 6248 case X86::BI__builtin_ia32_cmplepd: 6249 case X86::BI__builtin_ia32_cmpunordpd: 6250 case X86::BI__builtin_ia32_cmpneqpd: 6251 case X86::BI__builtin_ia32_cmpnltpd: 6252 case X86::BI__builtin_ia32_cmpnlepd: 6253 case X86::BI__builtin_ia32_cmpordpd: 6254 name = "cmppd"; 6255 ID = Intrinsic::x86_sse2_cmp_pd; 6256 break; 6257 case X86::BI__builtin_ia32_cmpeqsd: 6258 case X86::BI__builtin_ia32_cmpltsd: 6259 case X86::BI__builtin_ia32_cmplesd: 6260 case X86::BI__builtin_ia32_cmpunordsd: 6261 case X86::BI__builtin_ia32_cmpneqsd: 6262 case X86::BI__builtin_ia32_cmpnltsd: 6263 case X86::BI__builtin_ia32_cmpnlesd: 6264 case X86::BI__builtin_ia32_cmpordsd: 6265 name = "cmpsd"; 6266 ID = Intrinsic::x86_sse2_cmp_sd; 6267 break; 6268 } 6269 6270 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6271 llvm::Function *F = CGM.getIntrinsic(ID); 6272 return Builder.CreateCall(F, Ops, name); 6273 } 6274 } 6275 6276 6277 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6278 const CallExpr *E) { 6279 SmallVector<Value*, 4> Ops; 6280 6281 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6282 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6283 6284 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6285 6286 switch (BuiltinID) { 6287 default: return nullptr; 6288 6289 // vec_ld, vec_lvsl, vec_lvsr 6290 case PPC::BI__builtin_altivec_lvx: 6291 case PPC::BI__builtin_altivec_lvxl: 6292 case PPC::BI__builtin_altivec_lvebx: 6293 case PPC::BI__builtin_altivec_lvehx: 6294 case PPC::BI__builtin_altivec_lvewx: 6295 case PPC::BI__builtin_altivec_lvsl: 6296 case PPC::BI__builtin_altivec_lvsr: 6297 case PPC::BI__builtin_vsx_lxvd2x: 6298 case PPC::BI__builtin_vsx_lxvw4x: 6299 { 6300 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6301 6302 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6303 Ops.pop_back(); 6304 6305 switch (BuiltinID) { 6306 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6307 case PPC::BI__builtin_altivec_lvx: 6308 ID = Intrinsic::ppc_altivec_lvx; 6309 break; 6310 case PPC::BI__builtin_altivec_lvxl: 6311 ID = Intrinsic::ppc_altivec_lvxl; 6312 break; 6313 case PPC::BI__builtin_altivec_lvebx: 6314 ID = Intrinsic::ppc_altivec_lvebx; 6315 break; 6316 case PPC::BI__builtin_altivec_lvehx: 6317 ID = Intrinsic::ppc_altivec_lvehx; 6318 break; 6319 case PPC::BI__builtin_altivec_lvewx: 6320 ID = Intrinsic::ppc_altivec_lvewx; 6321 break; 6322 case PPC::BI__builtin_altivec_lvsl: 6323 ID = Intrinsic::ppc_altivec_lvsl; 6324 break; 6325 case PPC::BI__builtin_altivec_lvsr: 6326 ID = Intrinsic::ppc_altivec_lvsr; 6327 break; 6328 case PPC::BI__builtin_vsx_lxvd2x: 6329 ID = Intrinsic::ppc_vsx_lxvd2x; 6330 break; 6331 case PPC::BI__builtin_vsx_lxvw4x: 6332 ID = Intrinsic::ppc_vsx_lxvw4x; 6333 break; 6334 } 6335 llvm::Function *F = CGM.getIntrinsic(ID); 6336 return Builder.CreateCall(F, Ops, ""); 6337 } 6338 6339 // vec_st 6340 case PPC::BI__builtin_altivec_stvx: 6341 case PPC::BI__builtin_altivec_stvxl: 6342 case PPC::BI__builtin_altivec_stvebx: 6343 case PPC::BI__builtin_altivec_stvehx: 6344 case PPC::BI__builtin_altivec_stvewx: 6345 case PPC::BI__builtin_vsx_stxvd2x: 6346 case PPC::BI__builtin_vsx_stxvw4x: 6347 { 6348 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6349 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6350 Ops.pop_back(); 6351 6352 switch (BuiltinID) { 6353 default: llvm_unreachable("Unsupported st intrinsic!"); 6354 case PPC::BI__builtin_altivec_stvx: 6355 ID = Intrinsic::ppc_altivec_stvx; 6356 break; 6357 case PPC::BI__builtin_altivec_stvxl: 6358 ID = Intrinsic::ppc_altivec_stvxl; 6359 break; 6360 case PPC::BI__builtin_altivec_stvebx: 6361 ID = Intrinsic::ppc_altivec_stvebx; 6362 break; 6363 case PPC::BI__builtin_altivec_stvehx: 6364 ID = Intrinsic::ppc_altivec_stvehx; 6365 break; 6366 case PPC::BI__builtin_altivec_stvewx: 6367 ID = Intrinsic::ppc_altivec_stvewx; 6368 break; 6369 case PPC::BI__builtin_vsx_stxvd2x: 6370 ID = Intrinsic::ppc_vsx_stxvd2x; 6371 break; 6372 case PPC::BI__builtin_vsx_stxvw4x: 6373 ID = Intrinsic::ppc_vsx_stxvw4x; 6374 break; 6375 } 6376 llvm::Function *F = CGM.getIntrinsic(ID); 6377 return Builder.CreateCall(F, Ops, ""); 6378 } 6379 } 6380 } 6381 6382 // Emit an intrinsic that has 1 float or double. 6383 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6384 const CallExpr *E, 6385 unsigned IntrinsicID) { 6386 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6387 6388 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6389 return CGF.Builder.CreateCall(F, Src0); 6390 } 6391 6392 // Emit an intrinsic that has 3 float or double operands. 6393 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6394 const CallExpr *E, 6395 unsigned IntrinsicID) { 6396 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6397 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6398 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6399 6400 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6401 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6402 } 6403 6404 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6405 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6406 const CallExpr *E, 6407 unsigned IntrinsicID) { 6408 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6409 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6410 6411 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6412 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6413 } 6414 6415 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6416 const CallExpr *E) { 6417 switch (BuiltinID) { 6418 case R600::BI__builtin_amdgpu_div_scale: 6419 case R600::BI__builtin_amdgpu_div_scalef: { 6420 // Translate from the intrinsics's struct return to the builtin's out 6421 // argument. 6422 6423 std::pair<llvm::Value *, unsigned> FlagOutPtr 6424 = EmitPointerWithAlignment(E->getArg(3)); 6425 6426 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6427 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6428 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6429 6430 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6431 X->getType()); 6432 6433 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6434 6435 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6436 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6437 6438 llvm::Type *RealFlagType 6439 = FlagOutPtr.first->getType()->getPointerElementType(); 6440 6441 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6442 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6443 FlagStore->setAlignment(FlagOutPtr.second); 6444 return Result; 6445 } 6446 case R600::BI__builtin_amdgpu_div_fmas: 6447 case R600::BI__builtin_amdgpu_div_fmasf: { 6448 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6449 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6450 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6451 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6452 6453 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6454 Src0->getType()); 6455 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6456 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6457 } 6458 case R600::BI__builtin_amdgpu_div_fixup: 6459 case R600::BI__builtin_amdgpu_div_fixupf: 6460 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6461 case R600::BI__builtin_amdgpu_trig_preop: 6462 case R600::BI__builtin_amdgpu_trig_preopf: 6463 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6464 case R600::BI__builtin_amdgpu_rcp: 6465 case R600::BI__builtin_amdgpu_rcpf: 6466 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6467 case R600::BI__builtin_amdgpu_rsq: 6468 case R600::BI__builtin_amdgpu_rsqf: 6469 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6470 case R600::BI__builtin_amdgpu_rsq_clamped: 6471 case R600::BI__builtin_amdgpu_rsq_clampedf: 6472 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6473 case R600::BI__builtin_amdgpu_ldexp: 6474 case R600::BI__builtin_amdgpu_ldexpf: 6475 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6476 case R600::BI__builtin_amdgpu_class: 6477 case R600::BI__builtin_amdgpu_classf: 6478 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6479 default: 6480 return nullptr; 6481 } 6482 } 6483 6484 /// Handle a SystemZ function in which the final argument is a pointer 6485 /// to an int that receives the post-instruction CC value. At the LLVM level 6486 /// this is represented as a function that returns a {result, cc} pair. 6487 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6488 unsigned IntrinsicID, 6489 const CallExpr *E) { 6490 unsigned NumArgs = E->getNumArgs() - 1; 6491 SmallVector<Value *, 8> Args(NumArgs); 6492 for (unsigned I = 0; I < NumArgs; ++I) 6493 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6494 Value *CCPtr = CGF.EmitScalarExpr(E->getArg(NumArgs)); 6495 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6496 Value *Call = CGF.Builder.CreateCall(F, Args); 6497 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6498 CGF.Builder.CreateStore(CC, CCPtr); 6499 return CGF.Builder.CreateExtractValue(Call, 0); 6500 } 6501 6502 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6503 const CallExpr *E) { 6504 switch (BuiltinID) { 6505 case SystemZ::BI__builtin_tbegin: { 6506 Value *TDB = EmitScalarExpr(E->getArg(0)); 6507 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6508 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6509 return Builder.CreateCall(F, {TDB, Control}); 6510 } 6511 case SystemZ::BI__builtin_tbegin_nofloat: { 6512 Value *TDB = EmitScalarExpr(E->getArg(0)); 6513 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6514 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6515 return Builder.CreateCall(F, {TDB, Control}); 6516 } 6517 case SystemZ::BI__builtin_tbeginc: { 6518 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6519 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6520 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6521 return Builder.CreateCall(F, {TDB, Control}); 6522 } 6523 case SystemZ::BI__builtin_tabort: { 6524 Value *Data = EmitScalarExpr(E->getArg(0)); 6525 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6526 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6527 } 6528 case SystemZ::BI__builtin_non_tx_store: { 6529 Value *Address = EmitScalarExpr(E->getArg(0)); 6530 Value *Data = EmitScalarExpr(E->getArg(1)); 6531 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6532 return Builder.CreateCall(F, {Data, Address}); 6533 } 6534 6535 // Vector builtins. Note that most vector builtins are mapped automatically 6536 // to target-specific LLVM intrinsics. The ones handled specially here can 6537 // be represented via standard LLVM IR, which is preferable to enable common 6538 // LLVM optimizations. 6539 6540 case SystemZ::BI__builtin_s390_vpopctb: 6541 case SystemZ::BI__builtin_s390_vpopcth: 6542 case SystemZ::BI__builtin_s390_vpopctf: 6543 case SystemZ::BI__builtin_s390_vpopctg: { 6544 llvm::Type *ResultType = ConvertType(E->getType()); 6545 Value *X = EmitScalarExpr(E->getArg(0)); 6546 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6547 return Builder.CreateCall(F, X); 6548 } 6549 6550 case SystemZ::BI__builtin_s390_vclzb: 6551 case SystemZ::BI__builtin_s390_vclzh: 6552 case SystemZ::BI__builtin_s390_vclzf: 6553 case SystemZ::BI__builtin_s390_vclzg: { 6554 llvm::Type *ResultType = ConvertType(E->getType()); 6555 Value *X = EmitScalarExpr(E->getArg(0)); 6556 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6557 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6558 return Builder.CreateCall(F, {X, Undef}); 6559 } 6560 6561 case SystemZ::BI__builtin_s390_vctzb: 6562 case SystemZ::BI__builtin_s390_vctzh: 6563 case SystemZ::BI__builtin_s390_vctzf: 6564 case SystemZ::BI__builtin_s390_vctzg: { 6565 llvm::Type *ResultType = ConvertType(E->getType()); 6566 Value *X = EmitScalarExpr(E->getArg(0)); 6567 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6568 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6569 return Builder.CreateCall(F, {X, Undef}); 6570 } 6571 6572 case SystemZ::BI__builtin_s390_vfsqdb: { 6573 llvm::Type *ResultType = ConvertType(E->getType()); 6574 Value *X = EmitScalarExpr(E->getArg(0)); 6575 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6576 return Builder.CreateCall(F, X); 6577 } 6578 case SystemZ::BI__builtin_s390_vfmadb: { 6579 llvm::Type *ResultType = ConvertType(E->getType()); 6580 Value *X = EmitScalarExpr(E->getArg(0)); 6581 Value *Y = EmitScalarExpr(E->getArg(1)); 6582 Value *Z = EmitScalarExpr(E->getArg(2)); 6583 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6584 return Builder.CreateCall(F, {X, Y, Z}); 6585 } 6586 case SystemZ::BI__builtin_s390_vfmsdb: { 6587 llvm::Type *ResultType = ConvertType(E->getType()); 6588 Value *X = EmitScalarExpr(E->getArg(0)); 6589 Value *Y = EmitScalarExpr(E->getArg(1)); 6590 Value *Z = EmitScalarExpr(E->getArg(2)); 6591 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6592 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6593 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6594 } 6595 case SystemZ::BI__builtin_s390_vflpdb: { 6596 llvm::Type *ResultType = ConvertType(E->getType()); 6597 Value *X = EmitScalarExpr(E->getArg(0)); 6598 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6599 return Builder.CreateCall(F, X); 6600 } 6601 case SystemZ::BI__builtin_s390_vflndb: { 6602 llvm::Type *ResultType = ConvertType(E->getType()); 6603 Value *X = EmitScalarExpr(E->getArg(0)); 6604 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6605 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6606 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 6607 } 6608 case SystemZ::BI__builtin_s390_vfidb: { 6609 llvm::Type *ResultType = ConvertType(E->getType()); 6610 Value *X = EmitScalarExpr(E->getArg(0)); 6611 // Constant-fold the M4 and M5 mask arguments. 6612 llvm::APSInt M4, M5; 6613 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 6614 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 6615 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 6616 (void)IsConstM4; (void)IsConstM5; 6617 // Check whether this instance of vfidb can be represented via a LLVM 6618 // standard intrinsic. We only support some combinations of M4 and M5. 6619 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6620 switch (M4.getZExtValue()) { 6621 default: break; 6622 case 0: // IEEE-inexact exception allowed 6623 switch (M5.getZExtValue()) { 6624 default: break; 6625 case 0: ID = Intrinsic::rint; break; 6626 } 6627 break; 6628 case 4: // IEEE-inexact exception suppressed 6629 switch (M5.getZExtValue()) { 6630 default: break; 6631 case 0: ID = Intrinsic::nearbyint; break; 6632 case 1: ID = Intrinsic::round; break; 6633 case 5: ID = Intrinsic::trunc; break; 6634 case 6: ID = Intrinsic::ceil; break; 6635 case 7: ID = Intrinsic::floor; break; 6636 } 6637 break; 6638 } 6639 if (ID != Intrinsic::not_intrinsic) { 6640 Function *F = CGM.getIntrinsic(ID, ResultType); 6641 return Builder.CreateCall(F, X); 6642 } 6643 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 6644 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 6645 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 6646 return Builder.CreateCall(F, {X, M4Value, M5Value}); 6647 } 6648 6649 // Vector intrisincs that output the post-instruction CC value. 6650 6651 #define INTRINSIC_WITH_CC(NAME) \ 6652 case SystemZ::BI__builtin_##NAME: \ 6653 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 6654 6655 INTRINSIC_WITH_CC(s390_vpkshs); 6656 INTRINSIC_WITH_CC(s390_vpksfs); 6657 INTRINSIC_WITH_CC(s390_vpksgs); 6658 6659 INTRINSIC_WITH_CC(s390_vpklshs); 6660 INTRINSIC_WITH_CC(s390_vpklsfs); 6661 INTRINSIC_WITH_CC(s390_vpklsgs); 6662 6663 INTRINSIC_WITH_CC(s390_vceqbs); 6664 INTRINSIC_WITH_CC(s390_vceqhs); 6665 INTRINSIC_WITH_CC(s390_vceqfs); 6666 INTRINSIC_WITH_CC(s390_vceqgs); 6667 6668 INTRINSIC_WITH_CC(s390_vchbs); 6669 INTRINSIC_WITH_CC(s390_vchhs); 6670 INTRINSIC_WITH_CC(s390_vchfs); 6671 INTRINSIC_WITH_CC(s390_vchgs); 6672 6673 INTRINSIC_WITH_CC(s390_vchlbs); 6674 INTRINSIC_WITH_CC(s390_vchlhs); 6675 INTRINSIC_WITH_CC(s390_vchlfs); 6676 INTRINSIC_WITH_CC(s390_vchlgs); 6677 6678 INTRINSIC_WITH_CC(s390_vfaebs); 6679 INTRINSIC_WITH_CC(s390_vfaehs); 6680 INTRINSIC_WITH_CC(s390_vfaefs); 6681 6682 INTRINSIC_WITH_CC(s390_vfaezbs); 6683 INTRINSIC_WITH_CC(s390_vfaezhs); 6684 INTRINSIC_WITH_CC(s390_vfaezfs); 6685 6686 INTRINSIC_WITH_CC(s390_vfeebs); 6687 INTRINSIC_WITH_CC(s390_vfeehs); 6688 INTRINSIC_WITH_CC(s390_vfeefs); 6689 6690 INTRINSIC_WITH_CC(s390_vfeezbs); 6691 INTRINSIC_WITH_CC(s390_vfeezhs); 6692 INTRINSIC_WITH_CC(s390_vfeezfs); 6693 6694 INTRINSIC_WITH_CC(s390_vfenebs); 6695 INTRINSIC_WITH_CC(s390_vfenehs); 6696 INTRINSIC_WITH_CC(s390_vfenefs); 6697 6698 INTRINSIC_WITH_CC(s390_vfenezbs); 6699 INTRINSIC_WITH_CC(s390_vfenezhs); 6700 INTRINSIC_WITH_CC(s390_vfenezfs); 6701 6702 INTRINSIC_WITH_CC(s390_vistrbs); 6703 INTRINSIC_WITH_CC(s390_vistrhs); 6704 INTRINSIC_WITH_CC(s390_vistrfs); 6705 6706 INTRINSIC_WITH_CC(s390_vstrcbs); 6707 INTRINSIC_WITH_CC(s390_vstrchs); 6708 INTRINSIC_WITH_CC(s390_vstrcfs); 6709 6710 INTRINSIC_WITH_CC(s390_vstrczbs); 6711 INTRINSIC_WITH_CC(s390_vstrczhs); 6712 INTRINSIC_WITH_CC(s390_vstrczfs); 6713 6714 INTRINSIC_WITH_CC(s390_vfcedbs); 6715 INTRINSIC_WITH_CC(s390_vfchdbs); 6716 INTRINSIC_WITH_CC(s390_vfchedbs); 6717 6718 INTRINSIC_WITH_CC(s390_vftcidb); 6719 6720 #undef INTRINSIC_WITH_CC 6721 6722 default: 6723 return nullptr; 6724 } 6725 } 6726